KS Equations

Kohn-Sham potentials in density functional theory
Robert van Leeuwen
October 18, 1994
Contents
1 General introduction
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Overview of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 2.10 2.11 2.12 2.13 Key concepts and formulas . . . . . . . . . . . . . . . . . . . . Density matrices and density functionals . . . . . . . . . . . . . The pair-correlation function and the exchange-correlation hole E ective non-interacting particle models . . . . . . . . . . . . . The Hohenberg-Kohn theorem . . . . . . . . . . . . . . . . . . The Kohn-Sham equations . . . . . . . . . . . . . . . . . . . . . Scaling properties . . . . . . . . . . . . . . . . . . . . . . . . . . The coupling constant integration . . . . . . . . . . . . . . . . . Virial relations and the kinetic part of Exc . . . . . . . . . . . . Asymptotic properties of the exchange-correlation potential . . The Optimized Potential Model . . . . . . . . . . . . . . . . . . Approximate density functionals: LDA and GGA . . . . . . . . An integral equation for the exchange-correlation potential . . . . . . . . . . . . . . . . . . . 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
5 6
2 A review of density functional theory
9 10 13 15 18 20 22 25 29 30 31 34 38
3 A review of functional calculus
3.1 Introduction . . . . . . . . . . . . . . . 3.2 Banach spaces . . . . . . . . . . . . . . 3.2.1 De nition . . . . . . . . . . . . 3.2.2 Example . . . . . . . . . . . . . 3.2.3 The Banach spaces Lp and H 1 3.3 Operators and functionals . . . . . . . 3.4 Functional di erentiation . . . . . . . 3.4.1 The Frechet derivative . . . . . 3.4.2 The G^teaux-derivative . . . . a 3.4.3 Higher order derivatives . . . . 3.5 Variational calculus . . . . . . . . . . . 3.5.1 Extremal points of functionals 3.5.2 Convex functionals . . . . . . . 3.6 Functional integration . . . . . . . . . 3.6.1 The line integral . . . . . . . . 3.6.2 Integrability conditions . . . . 3.7 Taylor expansions . . . . . . . . . . . .
43
43 43 43 44 45 48 48 48 50 51 53 53 54 55 55 59 61
CONTENTS
4.1 4.2 4.3 4.4 4.5 4.6 4.7 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . Conditions on the electron density and external potentials Properties of the energy functional E v] . . . . . . . . . . The Hohenberg-Kohn functionals FHK ] and FEHK ] . . The Levy and Lieb functionals FLL ] and FL ] . . . . . G^teaux di erentiability of FL . . . . . . . . . . . . . . . a The Kohn-Sham equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Density functionals for Coulomb systems
63
63 63 67 68 73 75 78 81 82 86 88 93 98
5 An exchange-correlation potential with correct asymptotic behaviour

R.van Leeuwen and E.J.Baerends,Phys.Rev.A49:2421,1994 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Nonlocal exchange-correlation potentials : requirements . . . . . . . . 5.3 The nonlocal exchange and correlation potentials of Becke and Perdew 5.4 Constructing the potential from the density . . . . . . . . . . . . . . . 5.5 A model potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.6 Summary and conclusions . . . . . . . . . . . . . . . . . . . . . . . . .
81
6 Structure of the optimized e ective Kohn-Sham exchange potential and its gradient approximations 99
O.V.Gritsenko,R.van Leeuwen and E.J.Baerends,Int.J.Quant.Chem. (submitted) 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Properties of the exchange optimized e ective potential . . . . . . . . . . . 6.3 Comparison of the OPM, LDA, GGA and GEA exchange potentials . . . . 6.4 A gradient t to vx . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
99 100 102 108 114
7 Analysis of electron interaction and atomic shell structure in terms of local potentials 115
O.V.Gritsenko,R.van Leeuwen and E.J.Baerends,J.Chem.Phys. (accepted) 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 7.2 De nition of local potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 7.3 Structure of local potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 R.van Leeuwen, O.V.Gritsenko and E.J.Baerends,Z.Phys.D (submitted) 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 The exchange-correlation potential: separation into a long range and a short range part . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3 Functional derivative of the Kohn-Sham pair-correlation function . . . . . . . . . . . 8.4 The step structure in the exchange potential . . . . . . . . . . . . . . . . . . . . . . . 8.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.7 Appendix:The inverse density response function . . . . . . . . . . . . . . . . . . . . .
8 Step structure in the atomic Kohn-Sham potential
129
129 131 134 139 141 144 145
CONTENTS
9 A self-consistent approximation to the Kohn-Sham exchange potential
O.V.Gritsenko,R.van Leeuwen,E.van Lenthe and E.J.Baerends,Phys.Rev.A (submitted) 9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . mod 9.2 A model potential vresp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . mod 9.3 A self-consistent scheme with vresp . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4 A comparison of the self-consistent potentials . . . . . . . . . . . . . . . . . . . . . 9.5 Calculations with the accurate Slater potential . . . . . . . . . . . . . . . . . . . . 9.6 Calculations with the GGA approximation to vS . . . . . . . . . . . . . . . . . . . 9.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . R.van Leeuwen and E.J.Baerends,Int.J.Quant.Chem. (accepted) 10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2 Left-right correlation and the exact Kohn-Sham potential . 10.3 Analysis of Beckes and Perdews molecular potentials . . . . 10.4 Gradient corrected energy densities . . . . . . . . . . . . . . 10.5 Numerical procedure and results . . . . . . . . . . . . . . . 10.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . R.van Leeuwen and E.J.Baerends (to be submitted) 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Exchange and correlation in density functional theory 11.3 The in nite separation limit . . . . . . . . . . . . . . . 11.4 Numerical results . . . . . . . . . . . . . . . . . . . . . 11.5 An analysis of bond energy contributions . . . . . . . 11.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . .
149
. . . . . . . . . . . . . . . . . . . . . . . . . . 149 151 154 155 159 162 165
10 An analysis of nonlocal density functionals in chemical bonding

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
167
167 168 170 173 175 182 183 184 187 192 197 202
11 Exchange and correlation: a density functional analysis of the hydrogen molecule183

. . . . . . . . . . . . . . . . . .
12 Energy expressions in density functional theory using line integrals
R.van Leeuwen and E.J.Baerends,Phys.Rev.A (submitted) 12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2 Line integrals and path dependence . . . . . . . . . . . . . . . . . . . . . . 12.3 Exchange-correlation energy and the kinetic part: bounds from potentials 12.4 Constraints on vc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5 Invariance properties of potentials . . . . . . . . . . . . . . . . . . . . . . 12.6 Calculating molecular binding energies from potentials . . . . . . . . . . . 12.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
203
203 205 207 211 212 215 216
Samenvatting List of publications
219 223
CONTENTS
Chapter 1
General introduction
1.1 Introduction
The description of the physical properties of interacting many-particle systems has been one of the most important goals of physics during this century. The problem is to derive the properties of many-particle systems from the quantum mechanical laws of nature. This requires the solution of a partial di erential equation (the Schrodinger or Dirac equation ) of 3N spatial variables and N spin variables (for electrons) where N is the number of particles in the system. For atoms the number of electrons is in the range N 1 ; 100. Even small molecules have often more than 100 electrons and for a solid one has N 1023 . It is clear that the problem cannot be solved without making approximations somewhere along the line. However 'solving' the Schrodinger or Dirac equation is only part of the many-particle problem. The most important objective is to understand and predict the properties of many-particle systems and to calculate the several measurable quantities, such as the bonding energy, polarizability, conductivity,etc.. , rather than the wave function itself. One way to circumvent the complete solution of the many-particle problem is to construct model Hamiltonians containing only the physics one is interested in. A large number of model Hamiltonians is in use. One has for instance the Huckel method or tight-binding model, the Hubbard and Heisenberg Hamiltonians, the BCS-model etc.. to mention just a few (see for instance 1]). These approximate methods have been quite succesful in explaining a large range of physical phenomena ranging from bonding mechanisms in chemistry to ferromagnetism and superconductivity in solid state physics. However if one is looking for systematic improvements of these approximate models one still has to face the many-body problem. A systematic way of improvement can be achieved for nite systems such as atoms and molecules by variational methods or perturbation theory. A very simple method is the con guration interaction method. One uses an approximate wave function with many parameters (sometimes millions!) which are expansion coe cients of Slater determinants, and uses the variational principle to minimize the energy with respect to these parameters. The minimizing wave function then approximates the real wave function. The method is however not applicable to in nite systems. Also several variants of perturbation theory are available. A rather elegant formalism, which has been developed in the 50's, is the Greens function technique 2, 3, 4]. Instead of the wave function one uses the one- and two-particle Greens functions to calculate the measurable quantities. The method can in some cases also be applied to in nite systems, for instance quantum liquids or the electron gas. The disadvantage of the above methods is that their application when possible to inhomogeneous systems like atoms, molecules, solids and surfaces require a large computational e ort. A di erent formalism which is still exact and which can treat the systems mentioned with less 5
CHAPTER 1. GENERAL INTRODUCTION
computational e ort is the density functional method 5, 6, 7, 8, 9] . The formalism has been developed in the 60's by Hohenberg and Kohn 10] and Kohn and Sham 11]. The Kohn-Sham approach to density functional theory (DFT) allows an exact description of the interacting many-particle systems in terms of an e ective non-interacting particle system. The e ective potential in this noninteracting particle system (the Kohn-Sham system) can be shown to be completely determined by the electron density of the interacting system, and is for this reason called a density functional. In particular the ground state energy of the system is a density functional. Exact expressions for this functional are, due to the complicated nature of the many-body problem, not known. However over the years, due to a great amount of thoughtful work, more accurate ,more 'physical' and more practical approximations have continued to appear. The Kohn-Sham method has been used in solid state physics for about thirty years. By now, largely due to the development of increasingly accurate density functionals, the method has also gained a large popularity among quantum chemists, especially as it allows in many cases accurate treatments of molecular systems unattainable by the more traditional quantum mechanical methods. The review in the following chapter gives a basic introduction to density functional theory and the Kohn-Sham method. The choice of the subjects in this review has to a large extent been in uenced by the subjects of the thesis work and personal interest. It therefore treats among others exchange-correlation holes and pair-correlation functions, the OPM-model, scaling properties, and generalized gradient approximations. Two further chapters on the more mathematical aspects of density functional theory are included. The rst one is an introduction to functional calculus, the concepts of which are applied to density functional theory of Coulomb systems in the following section. This chapter contains a more rigorous discussion of some mathematical aspects of density functional theory providing a justi cation for the use of variational equations in DFT which assumes for instance di erentiability of functionals. From a physicists point of view these points are technicalities as all electron densities to be considered in nature are smooth and well-behaved and the corresponding energy functional should also be smooth and di erentiable, at least at the collection of physical densities. Although plausible, this still is an assumption. The fact that the statement is in fact true has been proved by mathematicians some 20 years after the Hohenberg-Kohn paper providing a sound basis for applications in DFT. The two chapters on functional analysis are hopefully useful for the reader interested in the functional analytical aspects of DFT.
1.2 Overview of the thesis

In this thesis we investigate the properties of the Kohn-Sham e ective potential for atomic and molecular systems. We in particular investigate the exchange-correlation potential and its approximations such as the local density approximation (LDA) and the generalized gradient approximations (GGA). We further propose improved approximations for the exchange and exchangecorrelation potential (shortly denoted as x-potential and xc-potential ) and discuss a method to obtain exchange-correlation energy expressions from approximate potentials using line integrals. In chapter 5 we discuss the properties of the xc-potential corresponding to the GGA's of Becke and Perdew which are compared to the exact xc-potential. The exact xc-potential is constructed from accurate atomic electron densities by a newly developed iterative method. From this comparison one can clearly see that the approximate GGA's lack some properties satis ed by the exact xc-potential, notably the long range Coulombic asymptotics which a ects the highest occupied orbital energies. To satisfy the asymptotic requirement a gradient expression is developed for the xc-potential with the correct asymptotics leading to greatly improved atomic and molecular ioni-
1.2. OVERVIEW OF THE THESIS
sation energies as determined from the highest occupied Kohn-Sham orbital. In chapter 6 we give a more extensive discussion of several gradient approximations to the exchange potential. These are Pade approximants which are constructed to satisfy the correct scaling properties, long range asymptotics and atomic shell structure of the exchange potential. The constructed Pade approximants lead to considerable improvements of the existing expressions for the GGA exchange potentials which do not exhibit the right long range asymptotics and which have an unphysical Coulombic divergence close to the atomic nucleus. The exchange energy is calculated from these Pade approximants using the Levy-Perdew relation. It turns out that to obtain good exchange energies this way a good description of the exchange potential in the core region is important. In chapter 7 we give an analysis of electron correlation and atomic shell structure in terms of local potentials. The analysis carried out leads to a splitup of the xc-potential into two terms, a long range and a short range part, both of which are relatively easy to approximate. An important question is how exchange and correlation e ects in uence the shape of the xc-potential as it provides insight in the way to construct more accurate density functionals. An important feature of the exchange and exchange-correlation potential is the atomic shell structure. The x- and xc-potential can be written as the sum of two potentials, a smooth long range attractive Coulombic potential and a short range repulsive step-like potential with clear atomic shell structure. The latter potential is constant within the atomic shells but changes rapidly at the atomic shell boundaries. The exchange-correlation energy is completely determined by the long range attractive part of the xc-potential. The atomic shell structure is further clearly re ected in the properties of the exchange and exchange-correlation hole and the conditional probability amplitude. Both quantities are related to the distribution of the other electrons when one electron, called the reference electron, is known to be at a certain position in space. The exchange hole and the conditional probability amplitude undergo rapid changes when the reference electron crosses atomic shell boundaries. As several of the constituents of the Kohn-Sham and xc-potential can be expressed in terms of the conditional probability amplitudes information on the structure of the Kohn-Sham potential can directly be obtained from the study of these probability amplitudes which contain all the exchange and correlation information of the system. For instance the peak structure at the atomic shell boundaries in the so-called kinetic potential which is a part of the functional derivative of the kinetic energy functional is directly related to changes in the conditional probabililty amplitudes when the reference electron crosses atomic shell boundaries. As discussed in chapter 10 the same potential describes the left-right correlation e ect in dissociating molecules. In chapter 7 we noted the step structure in the x- and xc-potential. In chapter 8 this step structure is derived by making an approximation in the functional derivative of the pair correlation function. The result is an equation for the exchange potential which has been derived earlier by Krieger, Li and Iafrate (KLI) in a di erent way. This potential turns out to be a very accurate approximation to the exact exchange potential. Another important result is an approximate expression for the inverse density response function as it occurs in several places within density functional theory (e.g. in the construction of xc-potentials from electron densities and in the calculation of excitation energies in time-dependent DFT). In chapter 9 we propose an approximation for the above discussed short range step-like and long range Coulombic part of the exchange potential. The long range part is approximated by a density gradient expression with the correct asymptotics derived from the Becke GGA for the exchange energy. The short range repulsive step-like part is written as a summation of orbital densities with coe cients that depend on the orbital energies. The analytic dependence of these coe cients on the orbital energies is determined by imposing the correct scaling relations, gauge invariance and
CHAPTER 1. GENERAL INTRODUCTION
reduction to the correct electron gas result for homogeneous systems. The resulting potential turns out to be a practical and accurate representation of the exchange potential and it moreover yields accurate (of Becke GGA quality ) values for the exchange energies. In chapters 10 and 11 we consider molecular systems. In chapter 10 we discuss the properties of the GGA's of Becke and Perdew for the case of molecular dissociation. The Becke-Perdew GGA is widely applied in quantum chemistry as it yields accurate ( with an accuracy of about 0.2 eV ) bond energies and geometries for molecules. In this chapter we try to obtain more insight in the reasons for this good performance of the gradient corrected functionals. We rst discuss the properties of the exact Kohn-Sham potential in the bond midpoint region. As has been noted before the Kohn-Sham potential in the bond midpoint region exhibits a peak structure related to the leftright correlation e ect. This peak structure arises from the kinetic part of the exchange-correlation functional and can be expressed in terms of the conditional probability amplitude. We show that this peak structure is also present in Becke-Perdew GGA however with a wrong functional dependence on the electron density. We furthermore explain why the Becke GGA for the exchange in general yields repulsive contributions to the bond energy and why the Perdew GGA for correlation yields an attractive contribution. This is related to the fact that density gradients at the saddle point in the electron density (the bond midpoint) are zero. Numerical results are presented for the dissociating nitrogen molecule and contributions from di erent regions of space are analyzed. In chapter 11 we discuss exact and approximate (the Becke-Perdew GGA) density functional results for the dissociating hydrogen molecule. The hydrogen molecule is chosen for the analysis as the exact DFT quantities needed for comparison with the approximate results can be calculated from accurate electron densities obtained from large con guration interaction calculations on this molecule. A disadvantage however is that this molecule is not a 'standard' molecule in the sense that is does not have core orbitals and therefore it lacks the Pauli repulsion contribution to the bond energy. The DFT results are further compared to the results of the Hartree-Fock approximation and the con guration interaction method which yields almost exact results. It turns out that the DFT results with GGA's included give a very accurate representation of the binding curve of the hydrogen molecule. The DFT curve deviates from the exact curve however in the dissociation limit, which can be cured by performing an unrestricted Kohn-Sham calculation. The DFT dissociation error is however much smaller than the error in the Hartree-Fock approximation. This is due to the fact that the LDA and GGA exchange-correlation hole is localized around the reference electron and resembles much more the true exchange-correlation hole than the delocalized Hartree-Fock exchange hole. The fact that there is still a dissociation error in DFT is due to the fact that the approximate density functionals are not invariant under spin rotations, which leads to problems for degenerate ground states (which are in practice solved by performing unrestricted calculations). We nally analyze the bond energy contribution of the Becke-Perdew GGA for certain density variations. In chapter 12 we discuss the problem how to calculate the exchange-correlation energy if only the exchange-correlation potential is known. From the exchange potential one can calculate the exchange energy using the Levy-Perdew relation which can be derived using the scaling properties of the exchange potential. For the correlation potentials such scaling relations are however not known. However as we show the correlation energy can be obtained from line integrals in the space of electron densities. Path dependence of the line integrals is discussed and integrability conditions are presented. We further discuss how several invariance properties of the xc-potential can be derived by selecting certain paths corresponding to certain symmetries (rotation, translation, scaling) and derive some inequalities for the correlation potential. We then show that the line integral formalism generalizes the transition state method of Ziegler for the calculation of molecular bond energies.
Chapter 2
A review of density functional theory

The evolution of a system of N particles moving in a external potential v(r) and having internal particle interactions w(r1 r2 ) is in many cases where relativistic e ects are not important to a very good approximation determined by the Schrodinger equation H (r1 : : : rN ) (r1 : : : rN t) = i@t (r1 : : : rN t) (2.1) where the Hamiltonian H is given by N 1 X 1X H (r1 : : : rN ) = ; 2 r2 + v(ri ) + 2 w(ri rj ) (2.2) i i i6=j In the above equations we neglected the internal degrees of freedom of the particles such as spin. For many-electron systems such as atoms, molecules or solids we have to take the electron spin into account and the stationary states of the system are determined from the time-independent Schrodinger equation H (r1 : : : rN ) (r1 1 : : : rN N ) = E (r1 1 : : : rN N ) (2.3) where i are the spin coordinates of the electrons and we require the wave function of the system to be antisymmetric under interchange of space and spin variables. The Hamiltonian can alternatively be written in second quantized notation as ^ ^ ^ ^ H =T +V +W (2.4) where Z ^ = ; 1 X ^+ (r)r2 ^ (r)dr T (2.5) 2 ^ V =
X Z
2.1 Key concepts and formulas
^+ (r)v(r) ^ (r)dr
(2.6) (2.7)
^ W =
We further denote xi = (ri i ). The action of the operators ^+ (r) and ^ (r) on a N -electron state N in coordinate space is de ned as ^+ (r) ]N +1 (x1 : : : xN ) = p 1 A (x ; xN +1 ) N (x1 : : : xN )] ^ (2.8) N +1 9
^+ (r) ^+0 (r0 )w(r r0 ) ^ 0 (r0 ) ^ (r)drdr0
10 where (x ; xi ) = (r ; ri ) and
CHAPTER 2. A REVIEW OF DENSITY FUNCTIONAL THEORY

(2.9) (2.10) (2.11)
^ (r) ]N ;1 (x1 : : : xN ;1 ) = pN N (x1 : : : xN ;1 x) ^ where A is the antisymmetrization operator de ned as ^ (A )(x1 : : : xN ) =

X
P 2SN
sign(P ) (xP (1) : : : xP (N ) )
where the summation runs over all permutations P of the group of permutations SN of N elements. The operator ^+ (r) adds a particle with spin at position r to the N -particle state N and operator ^ (r) removes a particle with spin from the N -particle state N . One can readily verify using the above de nitions that the Hamiltonians 2.3 and 2.4 are equivalent. One can furthermore derive the anticommutation relations h i h i ^ (r) ^ 0 (r0 ) = ^+ (r) ^+0 (r0 ) = 0 (2.12)
h
where A B ]+ = AB + BA. We now de ne the electron density as (r) = h j ^(r)j i = h j

X
+ +0 (r0 )i = ^ (r) ^ +
(r ; r0 )
(2.13)
^+ (r) ^ (r)j i = N X j (r x2 : : : xN )j2 dx2 : : : dxN
(2.14)
where dx denotes integration over the spatial variable r and summation over the spin variable . The electron density (r) is proportional to the probability per unit volume of nding an electron at position r.
2.2 Density matrices and density functionals

An important method of calculating ground state properties of many-particle systems is the variational method which is based on the variational principle which states in the case of fermion systems that for any anti-symmetric wavefunction : ^ E ] = h h jH j i i E0 (2.15) j ^ where E0 is the ground state energy of the system. This can be derived assuming H has a lowest eigenvalue which is almost always true for all (non-relativistic) cases of interest. This means that the ground state wavefunction is a stationary point of the energy functional E ], so we have " # ^ E 1 h jH j i ; E ] h j i 0 = (x : : : x ) = h j i (x1 : : : xN ) (x1 : : : xN ) 1 N
h i ^ = h 1 i H (x1 : : : xN ) ; E (x1 : : : xN ) j
(2.16)
2.2. DENSITY MATRICES AND DENSITY FUNCTIONALS
11
which is just Schrodingers equation as expected. ^ From the expression of the Hamiltonian 2.4 it follows that we can write the expectation value of H with state j i as 1 X Z r2 (r r0 )j 0 dr+X Z (r)v(r)dr+ 1 X Z ; 0 (r r0 )w(r r0 )drdr0 (2.17) ^ h jH j i = ; 2 r r=r 2 0 In the above expression we have de ned the one-particle density matrix as Z (r r0 ) = h j ^+ (r0 ) ^ (r)j i = N (r0 x2 : : : xN ) (r x2 : : : xN )dx2 : : : xN and the diagonal two-particle density matrix ; 0 as + ; 0 (r r0 ) = h j ^+ (r) ^Z 0 (r0 ) ^ 0 (r0 ) ^ (r)j i = N (N ; 1) j (r r0 0 x3 : : : xN )j2 dx3 : : : dxN (2.18)
(2.19)
A general M -particle density matrix (for an overview of the properties of density matrices see reference 12] ) is de ned as: 0 0 DM (r1 1 : : : rM M rM 1 : : : r0M M ) = h j ^+1 (r01 ) : : : ^+M (r0M ) ^ 1 (r1 ) : : : ^ M (rM )j i 0 0
N! Z = (N ; M )!
0 0 (r01 1 : : : r0M M xM +1 : : : xN ) (r1 1 : : : rM M xM +1 : : : xN )dxM +1 : : : dxN (2.20)
In systems were there are only two-particle interactions present we can restrict ourselves to the twoparticle density matrix. To calculate the energy of the system we then only need the one-particle density matrix and the diagonal two-particle density matrix which are related to D2 by ; 1 2 (r1 r2 ) = D2 (r1 1 r2 2 r1 1 r2 2 ) (2.21) Z X 0 (r1 r01 ) = D1 (r1 r01 ) = N 1 1 (2.22) ; 2 D2(r1 r2 2 r1 1 r2 2 )dr2 One could therefore wonder if one could derive a variational principle for density matrices instead of wavefunctions. This is readliy done by de ning ^ inf (2.23) E DM ] = !D h jH j i where the in mum of the expectation value is searched over all normalized anti-symmetric wavefunctions yielding DM (denoted ! DM ). This means that E DM ] is de ned on the set DM of M -particle density matrices obtainable from normalized antisymmetric wavefunctions, i.e DM = fDM j9 : ! DM h j i = 1 anti-symmetricg (2.24) One has the variational property: E DM ] E0 if DM 2 DM (2.25) An arbitrary function DM (x1 : : : xM x01 : : : x0M ) will in general not belong to the set DM . One might therefore ask what the necessary constraints are for a function DM to belong to DM . This problem, which is known as the N -representability problem for M -particle density matrices, has received considerable attention for the cases M = 1 and M = 2 13]. The reason for this is that
M
12
CHAPTER 2. A REVIEW OF DENSITY FUNCTIONAL THEORY
equation 2.17 gives an explicit expression for the total energy in terms of D2 and can therefore be applied for variational calculations. However to apply this formula one needs the constraints for D2 to belong to D2 which can then be incorporated in the calculation by means of Lagrange multipliers. If one minimizes without these constraints the energy expression one might get lower than E0 . However the N -representability problem for two-particle density matrices has not been solved until now. The N -representability problem for one-particle density matrices on the other hand has been solved 14] and we will therefore have a look at one-particle density matrix functionals. The energy for the one-particle density matrix can be written as 1 X Z r2 (r r0 )j 0 dr + X Z (r)v(r)dr + W ] E ] = ;2 (2.26) r r =r where W ] is de ned as ^ W ] = inf h jW j i (2.27) ! A very nice property of the functional W ] is that it is de ned independent of the external potential v. It can therefore in principle (but not in practice) be calculated once and for all independent of v. For practical applications one needs to know the necessary conditions for to belong to D1 . These conditions can be formulated as constraints on the eigenvalues i of . If we have
Z
(r1 r2 ) i (r2 )dr2 = i i (r2 )

Z
(2.28) 1g (2.29)
then the set D1 is equal to
D1 = f j
(r r)dr = N 0
The i are called the natural occupation numbers and the orbitals i are called the natural orbitals. The fact that 0 i 1 is a consequence of the Pauli-principle, any orbital can be occupied with maximally one electron with spin . For the application of the variation theorem one needs approximate expressions for W ]. One of the simplest expressions follows by considering a obtainable from a Slater determinant which yields 1 X Z w(r r ) (r r ) (r r ) ; (r r ) (r r )]dr dr W ]= 2 (2.30) 1 2 1 1 2 2 1 2 2 1 1 2 As one can prove 15] that for this approximate W ] the energy is always minimized by a obtainable from a Slater determinant wave function the variational equations corresponding to this W ] will be equal to the Hartree-Fock equations to be discussed in a later section. Some approximations which go beyond the one proposed here are known 16, 17] but the variational equations are not solved. An other approach for which approximations to the functionals are easier to obtain is the density functional approach. In the spirit of the lines described above one can de ne an energy density functional by
E ]=
(r)v(r) + FLL ]
(2.31) (2.32)
where the functional FLL introduced by Levy and Lieb is 18, 19] de ned as ^ ^ FLL ] = inf h jT + W j i !
2.3. THE PAIR-CORRELATION FUNCTION AND THE EXCHANGE-CORRELATION HOLE13

where the in mum is searched over all normalized antisymmetric wavefunctions giving electron density . The in mum can actually be shown to be a minimum 19]. As for any density integrating to N electrons and which satis es 0 one can easily construct a wavefunction (even a Slater determinant) which yields the densities in the set
N =f j
(r)dr = N g
(2.33)
satisfy the N -representability conditions. With the introduction of a Kohn-Sham system accurate approximate expressions for the functional FLL can be found. This leads in practice to the solution of a system of equations for noninteracting particles moving in an e ective external potential.
2.3 The pair-correlation function and the exchange-correlation hole

Important insight into the behaviour of interacting many-particle systems can be obtained by the analysis of correlation functions. Especially insightful are the pair-correlation function g (r1 r2 ) and the exchange-correlation hole function xc(r1 r2 ) Extensive discussions of the exchange-correlation hole function have been given by Slater 20] and McWeeny 21]. We de ne the function P 1 2 (r2 jr1 ) = ; 1 2 ((rr1 )r2 ) (2.34) 1 1 The function P 1 2 (r2 jr1 ) gives the probability to nd an electron with spin 2 at r2 if we know that there is an electron at r1 with spin 1 . We de ne the exchange-correlation hole function 1 xc 2 (r1 r2 ) by the equation 1 P 1 2 (r2jr1 ) = 2 (r2 ) + xc 2 (r1 r2 ) (2.35) As the presence of the reference electron at r1 reduces the probability to nd an electron near r1 the function xc represents a hole in the electron density (r2 ). It has the property
X Z
1 xc 2 (r1 r2 )dr2
= ;1
(2.36)
The hole therefore contains one electron. The electron repulsion energy between the electrons may now be written 1Z ^ (2.37) h jW j i = 2 j(rr1 ) (rr2j) dr1dr2 + Wxc 1; 2 ^ where W is the Coulombic repulsive potential between the electrons and the exchange-correlation energy is given by 1 1 X Z 1 (r1 ) xc 2 (r1 r2 ) dr dr = 1 X Z (r )vh (r )dr Wxc = 2 (2.38) 1 2 2 1 1 1 xc 1 1 jr1 ; r2 j 1 2 1
h where vxc (r1 ) is the potential of the exchange-correlation hole of the reference electron at r1 de ned as Z 1 xc 2 (r1 r2 ) dr dr h (r1 ) = X vxc 1 (2.39) jr ; r j 1 2
2
14
A REVIEW ON DENSITY FUNCTIONAL THEORY
h For neutral nite systems such as atoms the asymptotic behaviour of vxc at large distance r from the system decays Coulombically as ;1=r. If we move the reference electron to in nity then, for instance for an atom, the probability to nd an electron near the atomic nucleus is reduced. Another way of putting this is to say that the reference electron experiences the potential of the positive ion it leaves behind. Let us discuss the exchange-correlation hole function a bit further. The probability to nd another electron near the reference electron is reduced for two di erent reasons, one reason is the Pauli-principle which only a ects electrons with like spin, and the other reason is the Coulomb repulsion which is spin-independent. The Pauli principle forbids, even if there where no Coulomb interactions between the electrons, electrons with the same spin to occupy the same position in space. This is called the exchange e ect. The corresponding hole function x1 2 is called the exchange-hole function. For a non-interacting system of electrons the exchange-hole function is readily calculated. The wave function of such a system is just a Slater determinant 1 (2.40) D (r1 1 : : : rN N ) = p j 1 (r1 1 ) : : : N (rN N )j N! We nd 2 1 2 (r1 r2 ) = ; j 1 (r1 r2 )j (2.41) x 1 2 1 (r1 ) where
1 =
N1 X i
i (r1 1 ) i (r2 1 )
(2.42)
So we see that the hole for unlike spin electrons is zero as expected. If the reference electron at r1 is well-located within orbital i and the other orbital densities are small at r1 then 1 (r1 ) j i (r1 )j2 and we nd approximately 2 (2.43) x1 2 (r1 r2 ) ;j i (r2 2 )j 1 2 The exchange-correlation hole for an interacting system of electrons is much more di cult to h calculate. Note however that to calculate Wxc or the hole potential vxc we only need the spherical averaged part xc of the exchange-correlation hole, de ned as
xc
1 2(
r1 s) =
1 xc 2 (r1 r1 + s)d s
(2.44)
where d s is the integration volume of the angular variables for s. The exchange-correlation hole potential becomes Z 1 h (r ) = X 1 4 s2 xc 2 (r1 s) ds v (2.45)
xc 1 1
Another useful function for the description of electron correlation in many-electron systems is the pair-correlation function g 1 2 (r1 r2 ) de ned as 2 1 ( (2.46) g 1 2 (r1 r2 ) = ; (r2 )r1 rr )) 1 2( 2 1 Then we can write Z g 1 2 (r1 r2 ) ^j i= 1 X h jW (2.47) 1 (r1 ) 2 (r2 ) jr ; r j dr1 dr2 2 12 1 2
2.4. EFFECTIVE NON-INTERACTING PARTICLE MODELS
15
The function describes the screening of the interparticle interactions due to exchange- and correlation e ects, i.e. the 1=r12 potential is replaced by the screened potential g=r12 . The expression for Wxc then becomes 1 X Z (r ) (r ) (g 1 2 (r1 r2 ) ; 1) dr dr Wxc = 2 (2.48) 1 2 1 1 2 2 jr1 ; r2j 1 2 The pair-correlation functions satis es the sum rule
X Z
2(
r2)(g 1 2 (r1 r2 ) ; 1)dr2 = ;1

jr1 ; r2 j
2
(2.49)
And for the exchange-correlation hole potential we have Z 2 (r2 )(g 1 2 (r1 r2 ) ; 1) dr h (r ) = X v
xc 1 1
2
(2.50)
h From the sum rule we again nd that vxc has a Coulombic asymptotic behaviour.
2.4 E ective non-interacting particle models

Important insight in the properties of many-particle systems can be obtained from the study of e ective non-interacting particle models. In these models the interacting system of particles is replaced by a non-interacting system of particles in which the external eld is replaced by an e ective external eld which incorporates to some extent the interparticle interactions in an average way. Within the Kohn-Sham approach to DFT one can do this in a formally exact manner. In this section we will however discuss some approximate models, which by their relative ease in computation and by their physically appealling interpretation, have gained a considerable popularity. It turns out that from quite simple models such as the Huckel method in chemistry one can already understand many properties of chemical bonding. With the appearance of more powerful computers more sophisticated approximations, such as the Hartree-Fock approximation, have become quite popular in chemistry. In the Hartree-Fock approximation the ground state wave function is approximated by a Slater determinant 1 (2.51) D (r1 1 : : : rN N ) = p j 1 (r1 1 ) : : : N (rN N )j N! The corresponding energy expression is de ned as the expectation value of the Hamiltonian with ^ this Slater determinant wave function. The interparticle interaction W is taken to be Coulombic. 1 X Z (r )r2 (r )dr + Z (r)v(r)dr + 1 Z (r) (r0 ) drdr0 + E f g](2.52) E f i g] = ; 2 i x i i 2 jr ; r0 j where the electron density is given by XX (r) = j i (r )j2 (2.53) and the exchange energy by 1 Z j (r r0 )j2 drdr0 Ex f i g] = ; 2 jr ; r0 j
i
(2.54)
16 where
X 0 (r r0 ) = i (r ) i (r )

N
(2.55)
In order to obtain the equations for the e ective one-particle orbitals i we use the variational theorem and we minimize the functional
f i g] = E f i g] ;
ij
ij
i (r
) j (r )dr ; ij
(2.56)
where the Lagrange multipliers ij ensure orthonormality of the orbitals. The Euler-Lagrange equations are then obtained from
i (r
) =0
Z
yielding
i (r
) =0
ij
=0
Z
(2.57)
HF (r r0 ) (r0 i x
r0 ; 1 r2 + v(r) + jr0(; )rj dr0 2
i (r ) ;
)dr0 = i i (r )
(2.58)
where the one-particle energies i are obtained from a diagonalization of the matrix of Lagrange multipliers ij . The integral kernel HF is given by x 0) HF (r r0 ) = (r r (2.59)
x
jr ; r0 j
The above equations constitute the Hartree-Fock equations which is a set of integro-di erential equations for the orbitals i . Some of the features of the Hartree-Fock model become more clear if we rewrite these equations in a slightly modi ed form where we replace the integral kernel with a set of orbital dependent potentials 1 r2 + v(r) + Z (r0 ) dr0 + vHF (r ) (r ) = (r ) ;2 (2.60) i i i i jr ; r0 j where Z 0 (2.61) vHF (r ) = ; ni(r r ) dr0
i
jr ; r0j
and
0 0 ni (r r0 ) = i (r )j (r r)j)2 i (r ) i (r
(2.62)
The most important contribution of so we nd that approximately
to ni when i is an occupied orbital is the term i (r ) i (r0 ) (2.63) (2.64)
ni (r r0 ) j i (r0 )j2 viHF (r

)
Z
so we have
j i(r0 )j2 dr0 jr ; r0j
2.4. EFFECTIVE NON-INTERACTING PARTICLE MODELS
17
So we see that the term ni for occupied orbitals substracts out the density of orbital i . This is something which is intuitively clear. The e ective electron should move in the eld of the charge cloud of the other electrons, excluding the eld of its own charge cloud. For atomic systems with nuclear charge Z this implies that the asymptotic form of the total e ective Hartree-Fock potential for occupied orbitals is given by Z 0 0 2 vHF (r ) ; Z + (r ) ; j i (r )j dr0 ; (Z ; N + 1) (r ! 1) (2.65)
r jr ; r0j r which decays like ;1=r for neutral atoms. For unoccupied orbitals we have Z (r0 ) HF vTOT i(r ) ; Z + jr ; r0 j dr0 ; (Z ; N ) (r ! 1) r r
TOT i
(2.66)
which decays faster than Coulombic for neutral atoms. This last result implies that the virtual spectrum of a neutral atom within the Hartree-Fock approximation does not resemble the Rydberglike series characteristic for particles bounded by a Coulombic potential. The virtual orbitals (when bounded) of a neutral atom will be quite di use. This is of some importance when the virtual Hartree-Fock orbitals are used in more elaborate correlated methods such as con guration interaction expansions. The Hartree-Fock equations have the computational disadvantage that every e ective electron moves in a di erent potential. To simplify the equations one could therefore approximate the the HF-exchange energy locally by the HF-exchange energy of the homogeneous electron gas yielding (2.67)
i (r) = i i (r)
3 3 1 Z (r) 4 dr 3 1 Ex = ; 8 where = 2=3. The variational equations then become

"
1 r2 + v(r) + Z (r0 ) dr0 ; 4 ;2 jr ; r0j 3

i
3 8
1 3
(r
1 )3
(2.68) (2.69)
j i (r)j2 = (r)
This is a considerable simpli cation of the Hartree-Fock equations. The main advantage from a computational point of view is that one now does not have to calculate a nonlocal potential which requires large computational e ort. In actual applications for atoms, molecules and solids it turns out that the accuracy can be improved by replacing the = 2=3 of the electron gas by = 0:7. The method, originally developed by Slater 20], is known as the X -method and has extensively been used in solid state physics and quantum chemistry due to its simplicity and surprising accuracy which in many cases exceeds the accuracy of the Hartree-Fock approximation. The reason for this is among others, that the well-known molecular dissociation error in Hartree-Fock is considerably reduced in the X -approximation due to the localized character of the X -exchange hole which actually resembles more the true exchange-correlation hole than the exchange hole. A more extensive discussion of the feature in connection with the Local Density Approximation (LDA) is given in section of this chapter. The X -method shows that it is possible to obtain fairly accurate results within an e ective noninteracting particle model with a local potential depending only on the electron density . In the next two sections we will show that the X -model can be regarded as an approximation to an
18
exact non-interacting particle model, known as the Kohn-Sham system, containing a density dependent local potential. From the solution of the Kohn-Sham equations it is in principle possible to obtain the exact ground state density and energy. In practice of course we have to resort to approximations.
2.5 The Hohenberg-Kohn theorem

In the following we will demonstrate that the ground state energy of any interacting many-particle ^ system with a given xed interparticle interaction W is uniquely determined by its ground state density. Furthermore the knowledge of the ground state density determines the external potential v of the system (to within a trivial constant). An essential part of the proof of the latter ; v relation is the fact that when calculating the ^ expectation value of the Hamiltonian H with a certain state j i a term containing the electron density and the external potential explicitly separates out as ^ h jV j i =
Z
(r)v(r)dr
Z
(2.70) (2.71)
For the same reason one can prove from ^ h jW j i =

X
; 0 (r r0 )w(r r0 )drdr0
^ that the two-particle interaction W is (to within a constant ) a unique functional of the diagonal two-particle density matrix ; 0 . Another essential ingredient of the proof of the HK-theorem is the use of the variational theorem. We will derive the theorem for systems with non-degenerate ground states. The theorem is quite easily extended to the case of degenerate ground states. However to avoid unnecessary complication we will discuss the degenerate case in a later instant. ^ ^ An external potential V leads to a ground state density in an obvious way. First of all V leads to a ground state j i by solving the Schrodinger equation ^ H j i = Ej i (2.72) This provides us with a map C : V 7! from the set of external potentials V to the set of ground state wave functions . Secondly, for any ground state wave function one can calculate the corresponding electron density from (r) = h j ^(r)j i = N
X Z
j (r x2 : : : xN )j2 dx2 : : : dxN
(2.73)
which gives a second map D : 7! N from the set of ground state wave functions to the set of ground state densities N . The statement of the Hohenberg-Kohn theorem is then: The map DC : V 7! N is invertible. To prove this we rst show that two di erent external potentials ^ ^ V 6= V 0 + C (with C an arbitrary constant) lead to two di erent wave functions (that is di ering more than a phase factor). The proof is by contradiction. Suppose we have ^ ^ ^ (T + V1 + W )j 1 i = E1 j 1 i (2.74) ^ ^ ^ (T + V2 + W )j 2 i = E2 j 2 i (2.75)
2.5. THE HOHENBERG-KOHN THEOREM
19
^ ^ where V1 6= V2 + C where C is a constant and let us assume that j 1 i = j 2 i = j i, then by subtraction of these equations we have ^ ^ (V1 ; V2 )j i = (E1 ; E2 )j i (2.76) ^ ^ ^ ^ As V1 and V2 are multiplicative operators we must have V1 ; V2 = E1 ; E2 which is in contradiction ^2 + C unless j i vanishes in some region in space. This is however impossible by with V1 6= V the unique continuation theorem valid for all reasonably well-behaved potentials ( for instance for potentials not containing in nite barriers etc.). From the above contradiction we then conclude that our assumption was wrong so two potentials di ering more than a constant yield di erent ground state wave functions and hence the map C is invertable. A more rigorous discussion of the set of allowable potentials in DFT is presented in chapter 4 of this thesis. The second part of the proof consists of showing that two di erent ground state wave functions yield di erent ground state densities. The proof proceeds again by contradiction. Let us assume that j 1 i and j 2 i yield the same ground state density . Then we have Z ^ 1 j 1 i < h 2 jH1 j 2 i = h 2 jH2 + V1 ; V2 j 2 i = E2 + (r)(v1 (r) ; v2 (r))dr (2.77) ^ ^ ^ ^ E1 = h jH similarly we nd Z E2 < E1 + (r)(v2 (r) ; v1 (r))dr (2.78) Adding both inequalities then leads to the contradiction E1 + E2 < E1 + E2 (2.79) and therefore our assumption was wrong, j 1 i and j 2 i must yield di erent densities and the map D is invertible. Therefore by our previous result the map DC is also invertable and the ground state density uniquely determines the external potential. ^ Due to the invertability of the map D every ground state expectation value of an operator A is determined by the ground state density ^ A ] = h ]jAj ]i (2.80) In particular for the ground state energy we nd Z ^ j ]i = (r)v(r)dr + FHK ] Ev ] = h ]jH (2.81) where the functional FHK ] is de ned as ^ ^ FHK ] = h ]jT + W j ]i (2.82) ^ The functional FHK is universal in the sense that it does not depend on the external potential V . It can therefore in principle (but not in practice ) be calculated for all ground state densities . This is the most important conclusion of the HK-theorem, once approximations for FHK are known we can apply it to all electronic systems ranging from atoms and molecules to solids. The energy functional Ev satis es a variational property ^ ^ ^ ^ ^ ^ E0 = Ev0 0 ] = h 0 ]jT + V0 + W j 0 ]i < h ]jT + V0 + W j ]i = Ev0 ] (2.83) if 6= 0 . We therefore nd A more thorough discussion of the properties of FHK including a discussion of the extension of FHK to a functional FLL de ned on a larger set of densities is given in chapter 4 of this thesis.
E0 = min Ev0 ] = min 2N 2N
(r)v0 (r)dr + FHK ]
(2.84)
20
2.6 The Kohn-Sham equations

To apply the DFT formalism one obviously needs good approximations for the functional FHK or FLL . Accurate approximations have been obtained by means of the Kohn-Sham method 11]. A main advantage of this scheme is that it allows a straightforward determination of a large part of the kinetic energy in a simple way. Another advantage, from a more physical point of view, is that it provides an exact one-particle picture of interacting electronic systems. This then provides a rigorous basis for the one-particle arguments used in solid state physics and chemistry to explain and predict certain features of chemical bonding. ^ We introduce a non-interacting particle system with Hamiltonian Hs, ground state density and external potential vs . The ground state (assumed to be non-degenerate for the moment) is a Slater determinant of orbitals i which satisfy the equations ( we neglect spin for the moment) 1 ; 2 r2 + vs(r) i(r) = i i (r) (2.85) with
X
N i
j i (r)j2 = (r)
(2.86)
If we apply the Hohenberg-Kohn theorem to this non-interacting system we nd that there is at most one external potential vs to within a constant which generates ( one usually chooses this constant in such a way that vs ! 0 for jrj ! 1.) Therefore, for a given ground state density , all the properties of the system are determined. This is in particular true for the kinetic energy Ts ] and the total energy Evs ] given by
Ts ] =
and
N i
1 ;2
2 i (r1 )r i (r)dr
Z
(2.87) (2.88)
Evs ] = Ts ] +
(r)vs (r)dr
s 0 = (r) Evs ] ; s (r)dr = Tr) ] + vs (r) ; s (2.89) ( where we introduced the Lagrange multiplier s to satisfy the constraint that the density integrates to the correct number of electrons. Let us now go back to the interacting system where the energy functional reads
Z
The functional Ts ], which is called the Kohn-Sham kinetic energy, is just a particular case of the ^ functional FHK for the case W = 0. The ground state density of the system can now equivalently be obtained by the solution of the Euler-Lagrange equations
Ev ] =
(r)v(r)dr + FHK ]
(2.90) (2.91)
and let us de ne the exchange-correlation energy functional Exc ] as 1 Z (r1 ) (r2 ) dr dr ; T ] Exc ] = FHK ] ; 2 jr ; r j 1 2 s 1 2
2.6. THE KOHN-SHAM EQUATIONS
21
As FHK is only de ned for ground state densities of interacting systems and Ts is only de nend for ground state densities of non-interacting systems, we have implicitly assumed that for any ground state density of an interacting system there exists a non-interacting system with the same ground state density. This is the central assertion of the Kohn-Sham scheme. If we assume this assertion to be true then the Euler-Lagrange equation of the interacting system is given by Z Ts ] + v(r) + Z (r0 ) dr0 + Exc ] ; 0 = (r) Ev ] ; (r)dr = (r) (2.92) jr ; r0 j (r) If we de ne the exchange-correlation potential vxc by we nd using the fact that the equations for the interacting and the non-interacting system are solved for the same density that Z (r0 ) dr0 + v ( ] r) (2.94) v (r) = v(r) +
s
vxc ( ] r) = Exc) ] (r
(2.93)
to within a constant ( we can always choose the arbitrary constant in vxc in such a way that vxc ! 0 for jrj ! 1 and then ; s = 0). We therefore nd that we can nd the ground state density of the interacting system by solving the Kohn-Sham equations Z 1 (r0 ) ; 2 r2 + v(r) + jr ; r0 j dr0 + vxc( ] r) i(r) = i i(r) (2.95)
X
jr ; r0j
xc
N i
j i (r)j2 = (r)
(2.96)
The advantage of the above approach is that we now have reduced the problem of nding good approximations to the functional FHK to nding good approximations for the much smaller quantity Exc . ^ Note that the kinetic energy of the interacting system T ] = h ]jT j ]i with ] the ground state wave function of the interacting system with density , is not equal to the kinetic energy of the non-interacting system Ts ] with the same ground state density. It follows from Z ^j ]i + (r)vs (r)dr = h ]jHs j ]i ^ h ]jT
^ ]jHs j s ]i = Ts ] +
(r)vs (r)dr
(2.97) (2.98) (2.99)
^ where s ] is the Kohn-Sham determinant and Hs is the Kohn-Sham Hamiltonian, that T ] Ts ] and ^ Ts ] = minh jT j i !
where we minimize over all anti-symmetric and normalized wave functions yielding density . If we de ne the exchange-correlation part of the kinetic energy as Txc ] = T ] ; Ts ] (2.100)
22 we therefore nd

Txc ] 0
(2.101) (2.102)
So we have
Exc ] = Txc ] + Wxc ] Wxc ]

where
1Z ]i ; 2 j(rr1 ) (r2j) dr1 dr2 (2.103) 1 ; r2 Before we will look into approximations for the exchange-correlation energy functional we will rst discuss some exact scaling properties of density functionals which give useful constraints on approximate functionals.
Wxc ] = h
^ ]jW j
2.7 Scaling properties

We will now derive some scaling properties 22] of some density functionals which can be de ned in terms of the Kohn-Sham orbitals. Consider the Kohn-Sham equations 1 (2.104) ; 2 r2 + vs(r) i(r) = i i (r)
X
N i
j i (r)j2 = 1
(2.105)
and de ne the scaled coordinate r = r0 then r = 1r 0

r r
(2.106) (2.107) (2.108) (2.109) (2.110) (2.111) (2.112)
and we obtain ; 1 r20 + 2 vs( r) i ( r0) = 2 i i ( r0 ) 2 r If we de ne

i i
(r) = = 2i
3 2
i(
r)
(r) = 3 ( r) then we obtain 1 ; 2 r20 + 2 vs( r) i (r0 ) = i i (r0) r

X
N i
j i (r)j2 = (r)
2.7. SCALING PROPERTIES
23
We therefore conclude that if vs(r) generates ground state density (r) , which we will denote with vs ( ] r) , then vs (r) = 2 vs( r) generates ground state density (r). However as there is by the Hohenberg-Kohn theorem at most one potential which generates we must conclude that vs ( ] r) = 2 vs ( ] r) (2.113) to within a constant. We further obtain 3 (2.114) i ( ] r) = 2 i ( ] r) ]= 2 i ] This implies that the Kohn-Sham kinetic energy functional satis es Ts ] = 2 Ts ] Let us now de ne the exchange energy functional as 2 1Z Ex ] = ; 4 j jsr(r1 r2 )jj dr1 dr2 1 ; r2 where
i s (r1 r2 ) =
X
(2.115) (2.116) (2.117) (2.118)
N i
i (r1 ) i (r2 )
is the one-particle density matrix for the Kohn-Sham system. Using the above derived equations we nd
s(
] r1 r2 ) =
N i
N i
i(
] r1 ) i ( ] r2 ) = (2.119)
X 3 s ( ] r1 r2 ) = 3 i ( ] r1 ) i ( ] r2 ) =
This then yields Ex ] = Ex ] (2.120) The functionals Ts and Ex are special cases of homogeneously scaling functionals of the form A ] = kA ] (2.121) For these functionals we obviously have the property If we further denote
d A ]j = kA ] =1 d
(2.122) (2.123)
then on the other hand
a( ] r) = A ) ] (r d A ]j = Z =1 d A
(r) ]j =1 d d (r) j =1 dr
24
Z

= a( ] r)(3 (r) + r r (r))dr = ;
Z
(r)r ra( ] r)dr
(2.124) (2.125) (2.126) (2.127) (2.128)
where in the last step we performed partial integration. We therefore nd Z A ] = 1 a( ] r)(3 (r) + r r (r))dr In particular we nd for Ts and Ex that 1Z Ts ] = 2 vs( ] r)(3 (r) + r r (r))dr
Ex ] = vx ( ] r)(3 (r) + r r (r))dr

where the exchange potential vx is de ned as the functional derivative of Ex
x vx( ] r) = Er) ] (
This last relation between the exchange potential and exchange energy is often denoted as the Levy-Perdew relation 22]. The scaling property of the functional A ] further implies that a( ] r) = k a( ] r) (2.129) This we will derive as follows. From
A ] = a( ] r) (r)dr A ] = a( ] r)
(r ) = 3 ( r )
Z
(2.130) (2.131) (2.132) (2.133)

Z
(r)dr
A ]=
it follows that
Z
A ]
(r)dr = A ] = k A ] = k a( ] r) (r)dr =
k
Z
a( ] r)
k
Z
= 0=
Z
a( ] r) ( r)d r = a( ] r) ; k a( ] r)
a( ] r) 3 ( r)dr = k
(r)dr
a( ] r) (r)dr
(2.134) (2.135) (2.136) (2.137)
which yields
Z h
This relation is true for all and all variations (r)dr = 0 This then implies that a( ] r) = k a( ] r)
2.8. THE COUPLING CONSTANT INTEGRATION

Application of this relation to the functionals Ts and Ex yields 23]
25 (2.138) (2.139)
vs ( ] r) = 2 vs ( ] r) vx ( ] r) = vx( ] r) Ec ] = Exc ] ; Ex ]
The rst equation of the above we alreday derived above. For the correlation energy functional de ned as (2.140) (2.141) and its corresponding correlation potential
E vc ( ] r) = (rc) ]
no scaling relations are known. One can however prove the scaling inequalities
Ec ] < Ec ] Ec ] > Ec ]
<1 >1
(2.142) (2.143)
Proofs of the above relations and other scaling relations can be found in reference 22].
2.8 The coupling constant integration

The coupling constant integration technique 24, 25, 26] is a useful way of deriving relations among density functionals. It provides a way of expressing the exchange-correlation energy in terms of a coupling constant integrated pair correlation function g. ^ We de ne the Hamiltonian H by ^ ^ ^ ^ H =T +V + W (2.144) ^ ^ For = 1 we have the fully interacting system with external potential V = V =1 , having ground ^ in such a way that the ground state density remains state density . For 0 1 we de ne V ^ ^ unchanged. For = 0 we then obviously have V = Vs which is equal to the Kohn-Sham potential corresponding to density . If we denote the ground state for coupling constant by j i then by the Hellmann-Feynman theorem we obtain ^ dE ( ) = h j dH j i = h jW j i + d Z v (r) (r)dr ^ (2.145) d d d
Therefore
E (1) = E (0) +
= Ts ] + = Ts ] +
Z
^ h jW j id +
Z
(r)(v(r) ; vs (r))dr
(r)v(r)dr +
1 Z (r1 ) (r2 ) dr dr + E ] (r)v(r)dr + 2 jr ; r j 1 2 xc 1 2
^ h jW j id = (2.146)
26 where

(2.147) (2.148)
1 X Z 1 (r1 ) 2 (r2 ) (g (r r ) ; 1)dr dr Exc ] = 2 1 2 1 2 1 2 jr1 ; r2 j 1 2 Here we de ned the coupling constant integrated pair-correlation function g as
1 2
where g is the pair-correlation function corresponding to j i. The relation between Exc and the coupling constant integrated pair-correlation function is useful for constructing approximations to Exc by means of constructing approximations to g. Useful in this respect are relations between density scaling and coupling constant integration. Suppose that is the ground state eigenfunction ^ of H with a ground state density , so we can denote = ]. We have (we leave the spin indices out for notational convenience) 1X 1 4 ; 2 r2 + v (ri ) + 2 jr ; r j 5 (r1 : : : rN ) = E (r1 : : : rN ) i j i i6=j i If we de ne the scaled coordinate r = r0 we obtain 2 3 X 1 r2 + 2 v ( r0 ) + 1 X 4 5 ; 2 r0i ( r01 : : : r0N ) = 2 E ( r01 : : : r0N ) i 2 i6=j jr0i ; r0j j i
X 2 3
1 2(
r1 r2 )
(2.149)
(2.150) (2.151) (2.152)
If we take = 1= and de ne by
3N
So is a ground state wave function at full coupling strength yielding density 1= (r) = ;3 ( ;1 r). We therefore denote = =1 1= ]). So we nd that ](r1 : : : rN ) = 2 ( r1 : : : rN ) = 2 =1 1= ]( r1 : : : rN ) The diagonal two-particle density matrix at coupling strength is then 27] ; = =
1 2(
X
(r1 : : : rN ) = 2 ( r1 : : : rN ) Then satis es 2 3 X X 1 1 4 ; 2 r20i + ;2v ( ;1r0i) + 1 jr0 ; r0 j 5 (r01 : : : r0N ) = ;2E (r01 : : : r0N ) r 2 i6=j i j i
3N 3N
(2.153)
] r1 r2 ) =
Z
3 ::: N
](r1 1 : : : rN N )j2 dr3 : : : drN =
3 ::: N
X
3N j =1 1= ]( r1 1 : : : rN N )j2 dr3 : : : drN = 6

Z
3 ::: N
j =1 1= ]( r1 1 : : : rN N )j2 d r3 : : : d rN = 6; 1 2 ( 1= ] r1 r2 ) (2.154)
In particular we nd for the pair-correlation function at coupling constant that ; ( 1 ] r1 r ) 2 g 1 2 ( ] r1 r2 ) = ; 1(r () ] r(1r r)2 ) = 6 1 2 (r =) (r ) 2 = 1 1 2 2 1 1 2 2
2.8. THE COUPLING CONSTANT INTEGRATION

= ;
1= ] r1 r2 ) = g 1= 1 ( r1 ) 1= 2 ( r2 )
1 2( 1 2(
27 (2.155)
1= ]
r1 r2 )
So we can calculate the coupling constant integrated pair-correlation function g as
1 2(
] r1 r2 ) =
This is a useful relation as approximations for g ] are often based on models for full coupling strength. The coupling constant average can then be obtained from the above formula. We will further derive an expression for v in terms of potentials at full coupling strength. We will ^ rst show that if we de ne (with W being Coulombic) ^ ^ FLL ] = minh jT + W j i (2.157)
!
1 2(
1= ]
r1 r2 )d
(2.156)
where we search over all normalized anti-symmetric wave functions yielding density that
1= FLL ] = 2FLL 1= ] =1 where (r) = 3 ( r) and FLL = FLL . Equivalently one has FLL ] = 2 FLL 1= ]
(2.158) (2.159)
The above equations are readily proved (see Y.Wang in ref 13]). Every wave function yielding density can be written as (r1 : : : rN ) =
3N 2
( r1 : : : rN ) =
(r1 : : : rN )
(2.160)
This equation actually de nes and . Then yields density . Therefore we have ^ ^ ^ ^ FLL ] = min h T + W j i = minh jT + W j i = ! !
1= ^ ^ ^ ^ = minh j 2 T + W j i = 2 minh jT + 1 W j i = 2 FLL ] ! !
(2.161)
which proves the statement. The total energy functional for the system at coupling strength is given by
Ev ] = v (r) (r)dr + FLL ]
(2.162)
For = 1 we recover our usual energy functional for the fully interacting system. And for = 0 0 we have FLL ] = Ts ] and we obtain the energy functional of the non-interacting system which is the Kohn-Sham system. The variational equation for the density is given by (r)dr = F(LL ] + v (r) ; 0 = (r) Ev ] ; r) In particular for = 1 we nd
Z
(2.163) (2.164)
0 = F(LL ] + v(r) ; r)
28
By the de nition of v the variational equations are for all solved for the same density From the last two equations we then nd ( to within a constant ) that
v (r) = F(LL ] ; F(LL ] + v(r) r) r) We now split up FLL as FLL ] = Ts ] + 2

Z
(2.165) (2.166) (2.167) (2.168)
v (r) = v(r) + (1 ; )vH (r) + Exc) ; Exc) (r (r where we de ned the Hartree potential vH as
(r ) vH (r) = jr ; r0j dr0
Z
where the above equation de nes Exc ]. Then we can write
(r1 ) (r2 ) jr1 ; r2 j dr1 dr2 + Exc ]
From the known scaling property of Ts and the scaling property 2.159 it further follows that Exc ] = 2 Exc 1= ] (2.169) If we de ne the exchange energy for the system with coupling constant as 1 Z j s ( ] r1 r2 )j2 dr dr Ex ] = ; 4 (2.170) 1 2 jr1 ; r2 k where s is the Kohn-Sham one particle density matrix then it follows directly from the above de nition that Ex ] = Ex ] (2.171) If we furthermore de ne the correlation energy functional at coupling constant as Ec ] = Exc ] ; Ex ] (2.172) we nd Ec ] = 2 Ec 1= ] (2.173) So we obtain
Ex ] + 2 Ec 2 (2.174) (r) (r) 1= ] = vx ( ] r) + vc ( 1= ] r) and we nally get the following expression for v 28] v (r) = v(r) + (1 ; )(vx ( ] r) + vH (r)) + vc( ] r) ; 2 vc ( 1= ] r) (2.175) For = 1 we nd v = v and for = 0 we have v = vs . This equation has recently been used vxc ( ] r) = Exc) ] = (r
to derive exact perturbative expressions for the correlation functional in terms of the Kohn-Sham orbitals and eigenvalues, by performing an expansion in the coupling constant parameter 29]. How these expressions can be used in a Kohn-Sham calculation is discussed in the section on the Optimized Potential Model.
2.9. VIRIAL RELATIONS AND THE KINETIC PART OF EXC
29
2.9 Virial relations and the kinetic part of Exc

Exact relations among density functionals can be derived from the general quantum mechanical virial theorem
X ^ 2hT i = h ri ri U (r1 : : : rN )i
where the brackets stand for expectation values with the ground state wave function where U represents the sum of all potentials ^ ^ ^ U =V +W =
X
i=1
(2.176)
N i
1 v(ri) + 2
i6=j
w(ri rj )
(2.177)
^ If W is Coulombic this yields ^ ^ 2hT i + hW i = If we de ne

Z
(r)r rv(r)dr ]i ]i (r)r rv(r)dr
(2.178) (2.179) (2.180) (2.181)
T ]=h W ]=h
Then we have
^ ]jT j ^ ]jW j
Z
2T ] + W ] =
T (r)dr = (r) + W ) + v(r) ; 0 = (r) Ev ] ; (r we nd by acting with r r on the above equation that
Z Z
One should note however that the above relation is not universal. For a given density the above relation is only satis ed for the particular external potential which generates the prescribed density . Universal relations can however be derived on the basis of the virial relation. From the variational relations of the energy functional (2.182) (2.183) (2.184)
T dr + Z (r)r r W dr = ; Z (r)r rv(r)dr (r)r r (r) (r)

Z
which by addition to the virial relation yields 22] 2T ] +
T drdr = ;W ] ; Z (r)r r W dr (r)r r (r) (r)
The above equation does not involve the external potential anymore and is universal. For the ^ special case of non-interacting particles we have W = 0 and T ] = Ts ]. In that case the above relation yields 2Ts ] +
Z
Ts dr = 2T ] + Z (r)r rv ( ] r)dr = 0 (r)r r (r) s s
(2.185)
30

T ] = Ts ] + Txc ]
1 W ] = Exc ] + Txc ] + 2
Z
which we already proved by scaling. The last two equations in combination with (2.186) (r1 ) (r2 ) jr1 ; r2 j dr1 dr2 (2.187)
then yields 30, 22]
Txc ] = (3 (r) + r r (r))vxc ( ] r)dr ; Exc ]
(2.188)
So we have now established a relation between the kinetic part of the exchange-correlation energy functional and the exchange-correlation energy and potential. If we further split up the exchangecorrelation potential vxc = vx + vc and use the Levy-Perdew relation for the exchange functional we obtain
Txc ] = (3 (r) + r r (r))vc ( ] r)dr ; Ec ]
(2.189)
The above relation provides a way to calculate approximate expressions of the kinetic contribution of the exchange-correlation energy from approximate correlation functionals. The equation can be rewritten as as can be checked by di erentiation. Further relations for Txc in connection with functional integration can be found in chapter 12 of this thesis.
Txc ] = dEc ]j =1 ; Ec ] d
(2.190)
2.10 Asymptotic properties of the exchange-correlation potential

In this section we will discuss the long-range properties of the exchange-correlation potential in nite systems such as atoms and molecules. For simplicity we will discuss the case of spin unpolarized systems. We can write the exchange-correlation energy in terms of the coupling constant integrated pair-correlation function g as 1 Z (r1 ) (r2 ) (g ( ] r r ) ; 1)dr dr Exc ] = 2 jr ; r j (2.191) 1 2 1 2 1 2 By functional di erentiation the expression for the exchange-correlation potential becomes (2.192) where the exchange-correlation screening part ( where g represents the sreening of the interparticle interactions due to exchange-correlation e ects) of vxc is de ned as Z (r1 )(g ( ] r1 r3 ) ; 1) dr = Z 1 xc(r3 s) 4 s2 ds (2.193) v (r ) =
xc scr 3 resp vxc (r) = vxc scr(r) + vxc scr(r)
and the screening response potential as Z (r1 ) (r2 ) g ( ] r1 r2 ) dr dr resp (r3 ) = 1 vxc scr 1 2 2 jr1 ; r2 j (r3 )
jr1 ; r3j
(2.194)
2.11. THE OPTIMIZED POTENTIAL MODEL
31
The potential vxc scr is just the potential of the coupling constant integrated exchange-correlation hole. Due to the fact that this hole integrates to one electron (or the sumrule property of g ) we nd that for neutral nite systems v (r) ; 1 (jrj ! 1) (2.195)
xc scr
This is therefore a long range Coulombic potential. If we pull an electron to in nity then it will just experience the Coulombic potential of the positive ion it leaves behind. For the calculation of the exchange-correlation energy we only need a good approximation for the screening potential vxc scr as we have 1 Z (r)v Exc ] = 2 (2.196) xc scr (r)dr resp The other part vxc scr of the exchange-correlation potential is short ranged and for atoms it shows a clear atomic shell structure. It is constant within the atomic shells but changes rapidly at the atomic shell boundaries. The function g= which describes the sensitivity of the exchange-correlation screening due to density variations shows a clear dependence on the atomic shell structure. An extensive discussion of this function is given in chapter of this thesis. As vxc scr is long-ranged and resp vxc scr is short-ranged ( decaying faster than Coulombic) we nd A more rigorous derivation of the above relation can be found in reference 31]. The asymptotic behaviour of vxc is important for density functional theories for excitation energies which have been developed, as one then needs a correct description of the virtual spectrum, such as the Rydberg-like series in atoms. Also density dependent properties such as polarizabilities are sensitive to a correct behaviour of the exchange-correlation potential. This is an important observation as most density functionals in use, such as the current LDA+GGA exchange-correlation functionals, do not satisfy this asymptotic relation for the exchange-correlation potential 32].
vxc ; 1 r
(r ! 1)
(2.197)
2.11 The Optimized Potential Model

Constructing approximate density functionals by using direct expressions in terms of the density itself can be quite di cult in some cases. This is for instance true for the kinetic energy functional Ts , for which direct Thomas-Fermi like approximations in terms of the density are much more inaccurate. It is moreover more di cult to represent properties such as the atomic shell structure in the exchange-correlation potential correctly be means of direct expressions in the density than expressions in terms of Kohn-Sham orbitals. It is therefore interesting to study energy expressions which depend on the density through the intermediate use of the Kohn-Sham orbitals and energies, like Ts and Ex . Such expressions have recently been proposed by Gorling and Levy 29]. For Ex we already had
Ex ] = Ex f i g]
For Ec we can write
(2.198) (2.199)
Ec ] = Ec f
i i g]
32 For practical applications we need f vx ( ] r) = Ex (r)i g]

i vc ( ] r) = Ec f (r) i g]

(2.200) (2.201)
i (r0 ) dr0 + c:c:
This can by means of partial di erentiation be written as
vx( ] r) = vc( ] r) =
N i
i (r0 )
Z
Ex
(r)
(2.202)
1
i=1
i (r0 )
Ec
i (r0 ) dr0 + c:c: + X @Ec (r) i @ i
(r)
(2.203)
From the above expressions we can see that we need to know the functional derivatives i = and i = . We will determine these in the following. The orbitals i satisfy the Kohn-Sham equations 1 ; 2 r2 + vs( ] r) i(r) = i i (r) (2.204)
X
N i
j i (r)j2 = (r)
(2.205)
From response theory it is then not di cult to derive 33] i (r) 0 0 vs (r0 ) = ;Gi(r r ) i(r )
(2.206) (2.207)
vs (r0 ) = j i (r )j Where Gi is the Greens function of the Kohn-Sham system de ned as

0 0 ) = X j (r) j (r ) Gi (r r i; j j =i 6
0 2
(2.208)
We can therefore calculate the density response function s of the Kohn-Sham system as N N X X (r r0 ) = v (r)) = j i (r)j2 = ;2 i(r)Gi (r r0 ) i (r0 ) (2.209) s 0 s (r0 i vs (r ) i The expressions for vx and vc can now be written as
vx ( ] r) = vc ( ] r) =
N i
Ex G (r r ) (r ) ;1(r r)dr dr + c:c: i 1 2 i 2 s 2 1 2 i (r1 ) Ec G (r r ) (r ) ;1 (r r)dr dr + c:c:+ 1 2 (r ) i 1 2 i 2 s 2
(2.210)
and similarly for the correlation potential

X
1
i
i 1
2.11. THE OPTIMIZED POTENTIAL MODEL

+
Z X
33 (2.211)
1 @E Z c
i
j i(r1 )j2 ;1(r1 r)dr1 s
In these equations ;1 is the inverse density response function determined by the equation s
;1 (r1 r3 ) s (r3 r2 )dr3 = (r1 ; r2 ) s
(2.212)
It must be emphasized however that ;1 is only determined to within a constant. This is because s a change of vs by a constant does not a ect the density. We have (r1 ) =
Z Z
s(r1 r2 ) vs (r2 )dr2
(2.213) (2.214)
If vs (r) = C is a constant then (r) = 0 and we nd

s (r1 r2 )dr2 = 0
The constant function is therefore an eigenfunction of s with eigenvalue 0. So if a certain ;1 s satis es equation 2.212 then ;1 + C will satisfy this equation as well. The constant can however s be xed by choosing a particular gauge for the potential vs , for instance vs (r) ! 0 (r ! 1). As s is explicitly known in terms of i and i the above equations for vx and vc together with the Kohn-Sham equations constitute a self-consistent set of equations which can be solved once approximate expressions for Ec f i i g] are known. Note that the problem is equivalent to minimizing a total energy expression Ev f i i g] under the constraint that the orbitals i obey an independent particle equation with a local potential, i.e. one has to nd the local potential that solves the Euler-Lagrange equation For this reason these equation are called the Optimized Potential Model (OPM) equations 34, 33, 35, 36]. The simplest case is the exchange-only OPM, for which we can put Ec = 0, yielding the x-only OPM equations 1 ; 2 r2 + vs( ] r) i(r) = i i (r) (2.216)
X
Ev f v (r)
s
i i g] = 0
(2.215)
N i
j i (r)j2 = (r)
Z
(2.217)
0
(r ) vs ( ] r) = v(r) + jr ; r0 j dr0 + vx( ] r) Z Ex 0 0 0 s (r r )vx (r )dr = v (r) s
(2.218) (2.219)
For solving these equations one needs to solve an integral equation (to be more precise a Fredholm integral equation of the rst kind) for the exchange potential vx . This requires inversion of the integral kernel s (after xing a gauge) 37, 36]. One can however avoid this problem if one makes explicit orbital dependent approximations for the inverse function ;1 . An explicit expression can s be derived on the basis of an approximation for the Kohn-Sham Greens function Gi proposed by
34
Sharp and Horton 34] and Krieger et al. 35]. One nds (see chapter 8 of this thesis) in the natural gauge where vs ! 0 (r ! 1) for ; ( for the spin unpolarized non-degenerate case) s
N ;1 2 2 ;1 (r1 r2 ) = ; (r1 ; r2 ) ; X ik j i (r1 )j j k (r2 )j s (r1 ) (r1 ) (r2 ) ik
(2.220)
where the coe cients ik are given by ;1 (2.221) ik = (I ; N )ik where I is the unity matrix and N is the matrix Z 2 2 Nik = j i (r)j (jr)k (r)j dr (2.222) where all the matrices are of dimension N ; 1. If one uses this approximation for ;1 one obtains s the following approximation for the OPM exchange potential 35]
KLI vx (r) = vS (r) + N ;1 X i
2 wi j i (r)j (r)
(2.223) (2.224)
wi = h ijvx ; vij i i
where vi is equal to the orbital dependent Hartree-Fock potential (only its self-consistent orbitals are di erent ). The potential vS is the Slater potential. These equations have been derived rst by Krieger, Li and Iafrate in a di erent way. This approximation turns out to be very accurate 38, 39, 40, 41] From these approximate equations for vx we can easily see the structure of the exchange potential. It consists of a part vS which has a Coulombic long range ;1=r behaviour and a part we will call the step potential. For atomic systems this step potential is constant within the atomic shells and changing rapidly at the atomic shell boundaries. It has therefore a step-like structure with the heigths of the steps approximately equal to the constants wi . A more extensive discussion of all these features can be found in chapter 8 of this thesis.
2.12 Approximate density functionals: LDA and GGA

The simplest and most widely used density functional approximation for the exchange-correlation energy is the Local Density Approximation (LDA). In this approximation one uses the exchangecorrelation energy density of the homogeneous electron gas ehom ( 0 ) dependent on the homogeneous xc density 0 and replaces this for the inhomogeneous system with density (r) by eLDA ( (r)) = ehom ( 0 )j 0 = (r) (2.225) xc xc The exchange-correlation energy functional then becomes
LDA Exc ] = eLDA ( (r))dr xc
Z
(2.226)
At rst sight this approximation seems to work only for systems with slowly varying densities, such as the weakly perturbed electron gas. However the approximation works surprisingly well even for systems which have very inhomogeneous electron densities such as atoms and molecules. This calls
2.12. APPROXIMATE DENSITY FUNCTIONALS: LDA AND GGA
35
for an explanation. A more detailed look on LDA can be obtained from considerations of the xc-hole and the paircorrelation function. We have
LDA (r r ) = 1 2 xc
h i (r1 ) ghom ( ] jr1 ; r2 j) ; 1
(2.227)
where ghom ] is the coupling constant integrated pair-correlation function of the homogeneous electron gas. By now accurate expressions for ghom ] are known. The simplest part is the exchange part of the pair-correlation function which is una ected by the coupling constant integration. It is given (in its spin unpolarized form) by
2 9 ; F ) 1 j hom gx ( ] r1 r2 ) = 1 ; 2 sin (kF (r1)jr1 ; r2 j) (k k(r(r)1jrjr;; rj23 cos (kF (r1 )jr1 ; r2j) (2.228) F 1 1 r2 )
F
where kF (r) is the local Fermi wave vector de ned as 3 3 k (r) = ( 3 ) 1 (r) 1 It satis es the scaling property
hom hom gx ( ] r1 r2 ) = gx ( 1= ] r1 r2 )
(2.229) (2.230)
where 1= (r) = ;3 ( ;1 r). The corresponding exchange functional as can be calculated from the above pair-correlation function is then Z 4 LDA ] = 3 ( 3 ) 1 3 Ex (r ) 3 d r (2.231) 4 which satis es the correct exchange scaling. Corresponding expressions exist for the correlation LDA ]. The rst important thing to note is that the LDA xc-hole is part of ghom ] and for Ec spherical around the reference electron where s = jr1 ; r2 j and secondly it satis es the sum rule
Z
LDA (r r ) = (r 1 2 xc 1 xc
s)
Z
(2.232)
1 LDA 2 xc (r1 s)s ds = ;1
LDA (r r )dr = 4 1 2 2 xc
The above relations provide some insight in the unexpected success of LDA in strongly inhomogeneous systems. The coupling constant averaged xc-hole in real inhomogeneous systems is not spherically symmetric. Therefore the LDA xc-hole cannot adequately approximate this xc-hole. However the spherically averaged hole is reasonably well reproduced within the local density approximation. Consequently the exchange-correlation energy which is determined by the spherically averaged xc-hole is reasonably well produced as well. The satisfaction of the sum rule 2.233 further implies that if the LDA xc-hole LDA (r s) has positive errors for some values of s, it must have xc negative errors for other values of s leading to a systematic cancelation of errors. An obvious way to go beyond the local density approximation is to extend the exchange-correlation functional with terms containing gradients of the electron density. The gradients then measure changes of the electron density and can therefore be expected to improve the local density approximation. Most of the gradient functionals developed are based on the response theory of the weakly
(2.233)
36
varying electron gas 42]. The approximations obtained in this way are called Gradient Expansion Approximations (GEA). This then leads to the expressions Z 2 E GEA ] = E LDA ] + (r ) dr + : : : (2.234)
x x
4 3
Z 2 EcGEA ] = EcLDA ] + C ( ) (r 4 ) dr + : : :
(2.235)
where is a constant and C ( ) is a function determined by response theory. However in practice, the success of LDA nothwithstanding, the GEA provides no systematic improvement over LDA as realistic densities in atoms and molecules do not vary slowly over space. An analysis of the gradient expansion of the xc-hole shows that although the short range part (near the reference electron) is improved by the gradient expansion, the long range part is considerably worsened 43]. If one however by hand corrects the long range part of the gradient corrected xc-hole and enforces the sum rule property one can obtain xc-energy functionals which give a considerable improvement over LDA for energetics. The corresponding approximations are known as Generalized Gradient Approximations or GGA's. The most widely used GGA's are the Becke GGA 44] for the exchange energy and the exchange- and correlation GGA's by Perdew 45] and Perdew and Wang 46, 47, 43, 48]. We will give a short account of the ideas behind the derivation of these GGA's. The Becke GGA correction to the LDA is of the form
GGA Ex
]=
(r) 3 f (x(r))dr
(2.236)
3 where x(r) = jr (r)j= (r) 4 and f is a function to be determined. The variable x is a dimensionless quantity introduced to satisfy the scaling property of the exchange functional. In order to obtain the GEA for weakly varying densities we must have the following small x behaviour of f f (x) ; x2 (x # 0) (2.237) h We can write the exchange energy in terms of the potential of the exchange hole vx as Z h Ex ] = 1 (r)vx (r)dr (2.238) 2 In the above ansatz we have h h 3 (2.239) vx (r) = vx LDA(r) + 2 (r) 1 f (x(r)) h h To satisfy the correct asymptotics for vx for nite systems ( vx ;1=r (r ! 1) ) we must have 49, 32] 1 x f (x) ; 6 ln x (x ! 1) ; (2r) 1 (r ! 1) (2.240) r which follows directly by inserting exponentially decaying densities. One of the simplest interpolations between the large and the small x behaviour of f (x) is then found by taking
f (x) = ;
The coe cient was tted by Becke to obtain the correct exchange energy of the noble gas atoms. A somewhat di erent approach to obtain GGA's has been used by Langreth and Mehl 50],
x2 1 + 6 x sinh;1 x
(2.241)
2.12. APPROXIMATE DENSITY FUNCTIONALS: LDA AND GGA
37
Perdew 45], and Perdew and Wang 43, 48]. One can write the spherical average of the coupling constant integrated xc-hole in momentum space as 50, 43] Z Z 1 (r k ) = (r s)eik s ds = 4 s2 sin (ks) (r s)ds (2.242)
xc xc
Its Fourier inverse is given by Z 1 (ks 4 k2 sinks ) xc(r k)dk (2.243) (r s) = (21 )3 xc 0 Langreth and Perdew 51] and Langreth and Mehl 50] have carried out an GEA wavevector analysis for the spherically averaged correlation hole within the random phase approximation (RPA) yielding
GEA (r c
ks
xc
k) =
p
LDA (r c
where kFT (r) = 2 kF (r)= is the Fermi-Thomas wavevector or inverse screening length and C is a local function of the electron density. From equation 2.242 we nd lim (r k) = 4 s2 xc (r s)ds = ;1 k!0 xc 0 In particular we have for the correlation hole in momentum space
c (r
Z
3 ))2 k) + k2 (r) C (1) (r (r4 exp (; k 2 (3r) C ((1)) k) C FT FT (r) 3
(2.244)
(2.245) (2.246)
k = 0) = 0
One can see from equation 2.244 that this property is violated by the GEA correlation hole (it is satis ed by the LDA correlation hole). To repair this feature Langreth and Mehl and Perdew propose a wavevector space cut-o . One replaces the gradient contribution to xc (r k) by zero for k < kc where kc = f~jr (r)j= (r) is a cut-o proportional to the inhomogeneity wavevector, i.e. one removes the long wave length part of the Fourier analysis of the xc-hole. This then leads to the following approximation for the correlation energy functional 45] Z ))2 EcGGA ] = EcLDA ] ; e; C ( ) (r (r4 dr (2.247) (r) 3 where (2.248) = 1:745f~C ((1)) jr7 j C 6 Other GGA approximations for exchange and correlation have been obtained by Perdew and Wang 43, 48, 47] using real space cut-o s of the xc-hole. We will describe the procedure for the exchange energy functional. The correlation functional can be treated in a similar manner. The GEA exchange hole in the second order gradient expansion has the following structure
GEA (r x
s) = kF (r)3 Ax(kF (r)s) + x(r)2 Bx(kF (r)s)

h
(2.249)
4 3
3 where kF (r) = (3 2 (r)) 1 is the local Fermi wavevector and x(r) = jr j= inhomogeneity parameter. In the GGA one takes
is a dimensionless (2.250)
GGA (r x
s) =
GEA (r x
s) (U (x) ; kF (r)s)
38

Z U (x) h i 4 s2 xc (r s)ds = 4 u2 Ax (u) + x(r)2 Bx (u) du 0 0
The Heaviside function provides a long range cuto of the exchange hole. The cut-o radius is chosen such that the sum rule is satis ed
;1 =
(2.251) (2.252)
One obtains with
GGA ] = Ex
(r) 3 F (x)dr 4 u Ax (u) + x(r)2 Bx (u) du

h i
F (x) =
U (x)
The functions U (x) and F (x) can be evaluated numerically and tted to an analytical expression for practical applications. A pleasing feature from a theoretical point of view is that the above approach does not contain any adjustable parameters which need to be tted to known exchange energies. Completely analogous to the exchange case one can also carry out a real-space cut-o procedure for the correlation hole leading to parameter free GGA's for the correlation energy. These GGA's turn out to yield accurate atomic correlation energies 52, 53, 54, 55]
(2.253)
2.13 An integral equation for the exchange-correlation potential

In this section we will establish some connections between density functional theory and Greens function theory. We will in particular derive an integral equation for the exchange-correlation potential in terms of the irreducible self-energy. This equation was rst derived by Sham and Schluter 56] and Sham 57]. We rst discuss the key concepts of Greens function theory. The one-particle Greens function is de ned as the following expectation value ^ iG(xt x0 t0 ) = (t ; t0)h (t)S j ^(x)e;iH (t;t0 ) ^+ (x0 )j (t0)S i (2.254) ^ (t0 ;t) ^(x0 )j (t) i ; (t0 ; t)h (t0 )S j ^+(x)e;iH S where is the Heaviside function and j (t)iS = e;iEtj > (2.255) is the ground state in the Schrodinger picture where E is the groundstate energy. The Greens function as de ned above has a nice physical interpretation. For t > t0 we add by the action of ^+ (x0 ) a particle to the ground state j (t0 )S i at spin-space point x0 . The in this way created ^ N + 1-particle state will then propagate under the in uence of the Hamiltonoperator H from t0 to ^+ (x)j (t)iS . The Greens function for t > t0 t where we take overlap with the N + 1-particle state therefore describes the transition amplitude that a particle which is added to the ground state of many-particle system will move in the time t ; t0 from point x0 to point x. In a similar manner we have for t0 > t a propagation of a N ; 1-particle state. So for t > t0 we are dealing with particle propagation and for t0 > t with hole propagation. The Greens function may be written in more compact notation by going to the Heisenberg picture iG(xt x0 t0 ) = (t ; t0)h j ^(xt)H ^+ (x0 t0)H j i ; (t0 ; t)h j ^+ (x0t0)H ^(xt)H j i (2.256) = h jT ^(xt)H ^+ (x0 t0 )H ]j i
2.13. AN INTEGRAL EQUATION FOR THE EXCHANGE-CORRELATION POTENTIAL 39

where we de ned ^ ^ ^(xt)H = eiHt ^(x)e;iHt (2.257) ^ ^+ (x)e;iHt ^ ^+ (xt)H = eiHt and the time-ordered product of two fermion operators as ^ ^ ^ ^ ^ ^ T A(t)B (t0 )] = (t ; t0 )A(t)B (t0) ; (t0 ; t)B (t0 )A(t) (2.258) The time-ordered expression is especially suitable for a perturbation expansion where the timeordering appears in a natural way. As G describes the time propagation of a N + 1-particle state and a N ; 1-particle state it contains information about the excitation spectrum of the system. This is made explicit by writing X N +1 iG(xt x0 t0 ) = (t ; t0) e;i(En ;E )(t;t0 )h j ^(x)j N +1 ih N +1 j ^+ (x0 )j i (2.259) n n
X N ;1 (t0 ; t) e;i(E ;En )(t0 ;t) h
N N where En +1 and En ;1 are the energies of the eigenstates j N +1 i and j N ;1 i of the N + 1- and n n N ; 1-particle systems. From this equation we can see that G only depends on x and x0 and the time di erence t ; t0 . We can therefore de ne the Fourier transform of G by
Z +1 G(x x0 !) = G(x x0 t ; t0)e;i!(t;t0 ) d! 2
j ^+(x0 )j
N ;1 ih N ;1 j ^(x)j n n
Using the following expression for the Heaviside step function ; Z +1 e;i! d! ( ) = lim 2 1i #0 ;1 ! + i we nd an expression for G which clearly reveals its analytical structure 0 X ( 0 nx nx G(x x0 !) = f! (; 1 )fn+xi1 ) + g!(;1 )gn (x1 ) n n+i n where fn and gn are de ned as Z p fn(x) = h j ^(x)j N +1 i = N + 1 (x1 : : : xN ) N +1 (x1 : : : xN x)dx1 : : : dxN n n
;1
(2.260) (2.261) (2.262) (2.263)
gn (x) = h
N ;1 j ^(x)j n
p i= N
N ;1 n
(x1 : : : xN ;1 ) (x1 : : : xN ;1 x)dx1 : : : dxN ;1 (2.264)
N N and where n = En +1 ; E and n = E ; En +1 correspond to the electron a nities ; n and the N N ionisation energies ; n. In the above derivation we assumed all the levels En +1 and En ;1 of the N ; 1 and N + 1 systems to be discrete. This is of course not true for a real electronic system such as an atom or a molecule where a continuous spectrum always exists. For an in nite system such a solid there is even no discrete spectrum at all. For these cases one has to replace the summation by an integral over the energies. This changes the analytical structure of G of equation 2.262 ,i.e. in the addition to the simple poles corresponding to the discrete levels we obtain branch cuts corresponding to the continuum states. If we want to nd approximations for G we need an equation of motion. Such an equation can be derived using the Heisenberg equation of motion ^ i@t ^(xt)H = ^(xt)H H ] (2.265)
40

(2.266)
which using the commutation relations 2.13 and the form of the Hamiltonian 2.4 gives Z 1 i@t ^(xt)H = ; 2 r2 + v(r) + w(r r0 ) ^+ (x0 t)H ^(x0 t)H dx0 ^(xt)H With the above equation and the de nition of G one nds 1 (i@t1 + 2 r2 ; v(r1 ))G(x1 t1 x01 t01 ) = 1 (x1 ; x01 ) (t1 ; t01 ) ; i
Z
w(r1 r2 )G2(x1 t1 x2 t1 x01 t01 x2 t+ )dr2 1
(2.267)
where t+ means t1 + where # 0. Here we de ned the two-particle Green function as 1 G2 (x1 t1 x2 t2 x01 t01 x02 t02 ) = (;i)2 h jT ^(x1 t1)H ^(x2 t2 )H ^+(x02 t02)H ^+(x01 t01)H ]j i (2.268) and where the time-ordered product of a arbitrary number of fermion operators has been de ned as ^ ^ ^ ^ ^ ^ T A1 (t1 )A2 (t2 ) : : : An(tn )] = sign(P )AP (1) (tP (1) )AP (2) (tP (2) ) : : : AP (n) (tP (n)) (2.269) where the permutaion P is chosen such that tP (1) > tP (2) > : : : > tP (n) . The two-partcle Greens function G2 has a similar physical interpretation as G as a transition amplitude but now with two particles added or removed from the system. Another important quantity is the self-energy de ned by the equation 1 (i@t1 + 2 r2 ; v(r1 ))G(x1 t1 x01 t01 ) = 1 (x1 ; x01 ) (t1 ; t01 ) ; i w(r1 r2 ) (x1 t1 x2 t2 )G(x2 t2 x01 t01 )dx2 dt2 dr2 The Fourier transform of the equation yields
Z
(2.270) (2.271)
! ; h(x1 )] G(x1 x01 !) ;
(x1 x2 !)G(x2 x01 !)dx2 = (x1 ; x01 )
If we de ne the Greens function G0 for a system of non-interacting particles by the equation ! ; h(x1 )] G0(x1 x01 !) = (x1 ; x01 ) (2.272) We obtain the equation
G(x1 x01 !) = G0 (x1 x01 !) + G0(x1 x2 !) (x2 x3 !)G(x3 x01 !)dx2 dx3
This equation is known as Dysons equation. We will write this more symbolically as G = G0 + G0 G In the same operator notation we have ! ;h ; ]G = 1 We now split up the self-energy in Hartree part and a remainder xc as follows (x1 x2 !) = (x1 ; x2 )vH (x1 ) + xc(x1 x2 !)
(2.273) (2.274) (2.275) (2.276)
2.13. AN INTEGRAL EQUATION FOR THE EXCHANGE-CORRELATION POTENTIAL 41

where vH is the repulsive potential of the electronic cloud. Let us now rewrite the Dyson equation as h i ! ; hs ; ~ G = 1 (2.277) where 1 hs (r) = ; 2 r2 + v(r) + vH (r) + vxc(r) (2.278) (2.279)
is the Kohn-Sham Hamiltonian and we de ned ~ as ~ (x1 x2 !) = xc(x1 x2 !) ; (x1 ; x2 )vxc (x1 )
One obtains an alternative Dyson equation in terms of the Kohn-Sham Greens function Gs G = Gs + Gs ~ G (2.280) where Gs satis es
! ; hs ] Gs = 1
(r) = ;iGs (x1 t x1 t+) = ;i d! Gs (x1 x1 !) 2 being equal to the diagonal Greens function of the fully interacting system, i.e.
Z Z (r) = ;i d! G(x1 x1 !) 2 From the Dyson equation it then follows that Z Z Z
(2.281)
By the de nition of the Kohn-Sham system the diagonal of Gs yields the ground state density of the system (2.282) (2.283) (2.284) (2.285) (2.286)
which by de nition of ~ yields the following integral equation for vxc .

Z
dx1 dx2 d! G(x x1 !) ~ (x1 x2 !)G(x2 x !) = 0 2 K (r1 r2 )vxc(r2 )dr2 = Q(r1 )

Z
where
K (r1 r2 ) = d! Gs(r1 r2 !)G(r2 r1 !) 2

xc (r1 r2
Z Z Z Q(r) = d! dr1 dr2 Gs(r r1 !) 2
!)G(r2 r !)
(2.287)
Which is an integral equation for the exchange-correlation potential. The equation can be solved self-consistently once an explicit expansion of G in terms of Gs is given. One of the simplest approximations is
G = Gs
(2.288)
42
xc(r1 r2 ) = ;i
Z

d! Gs(r1 r2 !) = ; s(r1 r2 ) 2 jr1 ; r2 j jr1 ; r2 j
(2.289) (2.290) (2.291) (2.292) (2.293)
yielding
Z K (r1 r2 ) = d! Gs(r1 r2 !)Gs(r2 r1 !) = s (r1 r2 ) 2
Q(r) = dr1 dr2 jr ; r j 1 2
s (r1 r2 )
d! G (r r !)G (r r !) s 2 2 s 1
These equations together with 1 ; 2 r2 + vs(r) 1 (r) = i i(r)

X
N i
j i (r)j2 = (r)
yield the x-only OPM equations. The integral equation for vxc can also be derived in an alternative way. One can express the exchange-correlation energy functional as Exc ] = itr(ln (1 ; ~ Gs) + ~ G) ; i xc ] (2.294) where in diagrammatic perturbation theory xc is expressed as a sum of so-called skeleton diagrams in terms of the Greens function G excluding two rst order graphs. For a derivation we refer to Sham and Schluter 56] and Sham 57]. Functional di erentiation of the above expression then yields the integral equation for vxc (see for example the references 58, 59]) The integral equation for vxc has been used in the so-called GW-approximation 2, 60, 61, 62] to estimate band gap corrections to the LDA 63] for semiconductors and insulators.
Chapter 3
A review of functional calculus

3.1 Introduction
In this review we present some aspects of nonlinear functional analysis 64, 65] .We thereby concentrate on di erential and integral calculus on function spaces. Of most practical importance is the nding of extrema of functionals. This is the central problem in the calculus of variations which is for instance widely used in the Lagrangian formalism in classical mechanics. In order to do calculus on more general spaces than the real numbers R and the complex numbers C those general spaces have to satisfy some requirements. Roughly said, they must look like the space of real numbers. The important features to be carried over to these spaces are its vectorspace structure (adding and multiplying), its topological structure (there is a distance between two points) and the completeness property (the real numbers form a continuum). General spaces with these properties are called Banach spaces. Special cases of Banach spaces are of course the real numbers R and the complex numbers C themselves.
3.2 Banach spaces

3.2.1 De nition
Many for theorems in ordinary calculus depend on three basic properties of the real numbers. Any extension of ordinary calculus to more general spaces should exhibit these basic properties. These are summed up in the following cryptic statements: a] You can add and multiply points b] There is a distance between points c] There is a continuum of points We will now discuss each statement separately. This then leads to the de nition of Banach space. First of all a Banach space B is a vectorspace which means that if x y 2 B and 2 R or C then:
x y 2B )x+y 2B x 2 B and 2 R or C ) x 2 B
43
44
CHAPTER 3. A REVIEW OF FUNCTIONAL CALCULUS
We thus just add and multiply as for the real numbers. Secondly a Banach space is a normed space which means that we have on B a function kk : B ! R 0 from B to the positive real numbers including zero with the following properties: k xk = j jkxk kx + yk kxk + kyk kxk = 0 , x = 0 These properties are a copy of the properties of the length of an ordinary vector in three-dimensional space. The introduction of a norm now allows us to de ne the distance between two points by the following formula: d(x y) = kx ; yk We have for instance: d(x y) = 0 , x = y So now we can talk about the nearness of points. This is a very important ingredient in calculus because now we can talk about the convergence of limits or about continuous functions which map nearby points to nearby points. The third important property of the real numbers is its completeness. This is best illustrated with an example. Consider the space of rational numbers Q embedded in the space of real numbers R. Both Q and R are normed vectorspaces but they di er in one property. Between two points in Q p 6 p there are points which do not lie in Q. We have for instance 2 2 5 3 ] but 2 is not an element of 2 Q. This is not the case for the real numbers. Between two real numbers there is no no-real number. The real numbers thus form a continuum. How is this property characterised mathematically ? p We continue our example. We can approach 2 2 R as close as we like with a sequence of rational p numbers xn 2 Q. The limit of this sequence is in R of course well-de ned and given by 2. This limit is however not de ned in Q as we walk out of the space of rational numbers into the space of real numbers. One can in Q however still notice 'convergence' as the di erence jxn ; xm j goes to zero if n and m go to in nity. Sequences with this property are called Cauchy sequences. Cauchy sequences do not always converge in Q but they do converge in R. This then nally leads to the de nition of complete spaces. A complete space is a space in which every Cauchy sequence converges to an elemant of the space. Now we can nally state the de nition of a Banach space. A Banach space is a complete, normed vectorspace.
3.2.2 Example
To make things more concrete we will now give an example of a Banach space. Consider the collection of continuous functions de ned on the interval a b] 2 R. This space we will denote by C a b]. This is a Banach space if the norm is appropriately chosen. One can check all requirements: a] Adding of two continuous functions or multiplying a continuous function by a real number yields again a continuous function. So C a b] is a vectorspace over the eld of real or complex numbers. b] We de ne the norm of a function by: kf k = xmaxb] jf (x)j (3.1) 2a
3.2. BANACH SPACES
45
This makes C a b] a normed space. One can easily check the requrements for the norm. c] We still have to check that C a b] with the given norm is complete space. This is always the most di cult part to check. Suppose we have a Cauchy sequence of continuous functions ffn g . Then for any point x 2 a b] : jfn(x) ; fm(x)j max jfn(x) ; fm(x)j = kfn ; fmk ! 0 (n m ! 1) This means that fn (x) for each x 2 a b] is Cauchy sequence in R and as R is complete this Cauchy sequence converges to some f (x) 2 R. This gives a limit function f (x). We still have to prove that this limit function is continuous. For xed n we have for all x 2 a b] : jf (x) ; fn(x)j = mlim jfm(x) ; fn(x)j xmaxb] mlim jfm(x) ; fn(x)j = !1 2 a !1 = max jf (x) ; fn(x)j = kf ; fn k
x2 a b] x2 a b]
Thus the sequence of continuous functions converges uniformly to f and hence its limit function f is continuous and f 2 C a b].
3.2.3 The Banach spaces Lp and H 1

Z
We now describe two types of Banach spaces which are relevant to quantum mechanics and density functional theory in particular. The rst Banach space of importance is the space of functions for which the Lebesque integral
jf (x)jp dx
(3.2)
exists where x is usually a n-dimensional real vector and p a positive real number. This space is usually denoted as Lp(Rn ). Another important space is the space of real functions for which the Lebesque integral
Z
jf (x)j2 + jrf (x)j2 dx
(3.3)
exists which called the Sobolev space H 1 (Rn ) ( the 1 in H 1 gives the order of the derivative). As both spaces are based on the concept of Lebesque integral we will give in the following a short account on the de nition of the Lebesque integral based on measure theory. In measure theory one is interested in the question how to de ne a volume element for more general spaces than the real numbers. Such a de nition is needed if one wants to perform integration on more general spaces Consider the space B and a family A of subsets of B containing the empty set . A positive measure m (which is intuitively something like a volume) is then de ned as a mapping from A into the extended real numbers m : A ! R+ f+1g which is countably additive for every disjoint family of subsets (a1 : : : an : : :) in A with union in A :
X m( 1 ai) = m(ai) i=1
and
i=1
(3.4) (3.5)
m( ) = 0
46
The space (B A m) endowed with measure m is called a measure space. The elements of A are called measurable subsets of B and m(a) is the measure of a 2 A. A property is said to hold almost everywhere if it holds for all points of B except possibly for points of a set a of measure m(a) = 0. As an example we take the space of real numbers R and take for A the collection of subsets of R generated by the open intervals through union and intersection (which makes R a topological space). We de ne the measure m on an open interval ]a b b > a by:
m(]a b ) = b ; a
We then have for instance for a < b < c :
(3.6) (3.7) (3.8)
m(]a c ) = c ; a = b ; a + c ; b = m(]a b ) + m(]b c )

On the other hand we have:
m(]a b ) + m(]b c ) = m(]a b ]b c ) = m(]a c nfbg)

and
m(]a c ) = m(]a c nfbg fbg) = m(]a c nfbg) + m(fbg) (3.9) Hence it follows that m(fbg) = 0 and so this measure for a single point is zero. By countable
addition it follows for example that the measure of the collection of rational numbers Q is zero:
1 1 fqi g) = X m(fqi g) = 0 m(Q) = m( i=1
i=1
(3.10)
In this measure two functions which di er on the set Q are almost everywhere equal as they di er on a set of measure zero. The measure we have de ned here on the real numbers is known as the Lebesque measure and is the basis of the Lebesque integration theory. For a function f : R ! R the Lebesque integral is de ned as follows. Divide the range of f into a nite number of small intervals and nd the set ai of all x 2 R for which f (x) is in the i-th interval. Assign a measure m(ai ) to the set ai . Let ki be some value of f (x) in the i-th interval and let fn be the step function equal to ki when x 2 ai . The Lebesque integral is then the limit when it exists of:
Z
fdm =
when the sequence of functions (fn ) tend to f in a sense we will make more precise in the following. In order to do this we need some further de nitions. A real function f : B ! R on the measure space (B A m) is said to be measurable when:
ki m(ai)
(3.11)
fxja < f (x) < bg 2 A 8a b 2 R (3.12) A function on (B A m) is called a step function or simple if it is zero except on a nite number n of disjoint sets ai 2 A of nite measure m(ai ) where the function is equal to a nite constant ki .
The integral of a step function is by de nition:
Z
fdm =
i=1
ki m(ai )
(3.13)
3.2. BANACH SPACES
47
For example if we denote the rational numbers on the real interval I = 0 1] by Q and we de ne the step function to be equal to two on Q and one elsewhere then the integral in the Lebesque-measure is given by:
Z
We nally give the de nition of the integral of a measurable function. We rst consider the case of a positive function.Let f be a positive real valued function on the measure space (B A m). Then we de ne the integral of f with respect to the measure m by:
Z
fdm = 2m(Q) + m(I nQ) = 0 + 1 = 1
(3.14)
as ranges over all step functions with 0 f . If this integral is nite then f is said to be integrable. Now an arbitrary real valued function can always be written f = f + ; f ; with f + and f ; positive functions. Then f is integrable if both f + and f ; are integrable and its integral is given by:
Z
fdm sup f
dmg
(3.15)
The integral on the real numbers R or Rn with respect to the Lebesque measure is called Lebesque integral. The Lebesque integral of two functions is equal if the two functions are equal almost everywhere. In the remainder of this section we will regard these functions as equal. More precisely we look at the equivalence class of functions which are equal almost everywhere. In a somewhat loose notation we will denote the equivalence class to which function f belongs with the same f . We now de ne Z Lp (R) = ff j jf (x)jpdx < 1g (3.17) where the integral is a Lebesque integral. For p 1 we can assign to this space a norm kkp which makes this space a Banach space
1
fdm
f +dm ;
f ;dm
(3.16)
kf kp =
jf (x)jpdx
(3.18)
This is a consequence of the Minkovski inequality kf + gkp kf kp + kgkp (3.19) for p 1. We further de ne the space L1 (R) as the space of (classes of ) measurable functions bounded almost everywhere with norm kf k1 = ess sup jf (x)j (3.20) where the essential supremum 'ess sup' is de ned as the smallest number M such that jf (x)j M almost everywhere. This a again a Banach space. An important inequality that can be proven for the Lp-spaces with 0 p 1 is the Holder inequality kfgk1 kf kpkgkq (3.21) with 1=p + 1=q = 1 with f 2 Lp and g 2 Lq . We further de ne the Sobolev space H 1 (R) as the Banach space
H 1 (R) = ff 2 L2 (R)j jf (x)j2 + jrf (x)j2 dx < 1
(3.22)
48 with corresponding norm
kf k =
Z h
jf (x)j2 + jrf (x)j2 dx
i1
(3.23)
( the derivative rf for f 2 L2 can be de ned by means of a Fourier transform ). We now have discussed the most important Banach spaces for applications in physics. We will in the next section discuss functions on Banach spaces which will be called operators or functionals.
3.3 Operators and functionals
Every map F : B1 ! B2 in which B1 and B2 are Banach spaces is called an operator. For the special case B2 = R this operator is called a functional. The spaces B1 and B2 are in general di erent spaces with di erent norms. We can consider for example the operator F : C a b] ! C a b] given by:
F f] = f2 F f] =
Z
(3.24) (3.25)
or the functional F : C a b] ! R given by: Special operators are the linear operators from B1 to B2 . Those operators have the property that if f g 2 B1 and 2 R then: F f] = F f] (3.26) F f + g] = F f ] + F g] The space of linear operators from B1 to B2 is also a vectorspace. On this space we can also introduce a norm by: (3.27) kF k = sup f kF f ]k2 g in which kk1 and kk2 are the norms on the spaces B1 and B2 . If the above operatornorm exists for an operator F then this operator is called bounded. The space of bounded linear operators from B1 to B2 is again a Banach space which is denoted by B1 B2 ]. This space will reappear in the next section when we want to de ne higher order derivatives. We now have obtained all the knowledge necessary to de ne integration and di erentiation on Banach spaces. We can now go on and prove most of the theorems of ordinary calculus by imitating the proofs. This then yields ordinary calculus as a special case of calculus on Banach spaces.
f 2B1 a b 2 f (x)dx
kf k1
3.4 Functional di erentiation

3.4.1 The Frechet derivative
In analogy with ordinary calculus on the vectorspace Rn we can de ne two types of derivatives. The rst is the Frechet-derivative, which is analogous to the total derivative in vector calculus, and the second is the Gateaux-derivative which is analogous to the directional derivative in vector calculus.We will start with the total derivative.
3.4. FUNCTIONAL DIFFERENTIATION
49
An operator F : DF ! B2 with DF B1 in which B1 and B2 are Banach spaces with norms kk1 and kk2 is called di erentiable in f 2 DF if there exists a linear operator:
F (f ) 2 B B ] (3.28) 1 2 f such that for f + h 2 DF : kF f + h] ; F f ] ; F (f ) h]k2 f lim 0 =0 (3.29) khk1 khk1 ! If F (f ) exists for all f 2 DF then this linear functional is called the Frechet-derivative of F . This f
de nition is analogous to the corresponding de nition of the total derivative in vectorcalculus in Rn in which case F (f ) is a n n-matrix working on the n-component vector h. f We will now give some examples. Consider again the operator F : C a b] ! C a b] de ned by F f ] = f 2 . Its Frechet-derivative is a linear operator F (f ) : B1 ! B2 de ned by: f This follows directly from the de nition of F : f lim
F (f ) h] = 2fh 2 C a b] f
(3.30)
2 k(f + h)2 ; f 2 ; F (f ) h]k f = lim kh kk = khk khk!0 khk!0 kh
1 1 = lim khk max jh2 (x)j lim khk ( max jh(x)j)2 = lim khk = 0 khk!0 x2 a b] khk!0 x2 a b] khk!0 Hence we have proven our statement. As a next example we will calculate the Frechet-derivative of the functional F : C a b] ! R de ned by:
F f] =
Its Frechet-derivative is given by the linear functional F (f ) : C a b] ! R given by: f We again check the de nition: Z b Z b 1 Zb 1 Zb lim khk j (f + h)2 dx ; f 2 dx ; 2fhdxj = lim khk h2 dx khk!0 khk!0 a a a a lim lim 0 bk;ka (xmax] jh(x)j)2 = khk!0 (b ; a)khk = 0 khk! h 2 a b
b 2 f (x)dx
(3.31)
F (f ) h] = Z b 2f (x)h(x)dx f a
(3.32)
which proves our statement. We now have de ned the derivative of an operator as a linear operator. This is in accordance with the view of regarding the derivative of a function as a linear approximation to that function. In order to calculate higher order approximations we must carry out a Taylor-expansion and de ne higher order derivatives. Before addressing this question we will now give another de nition of the derivative which is called the G^teaux-derivative. a
50
3.4.2 The G^teaux-derivative a
The G^teaux-derivative of an operator F : B1 ! B2 in the direction h 2 B1 is a linear operator a F 2 B1 B2 ] de ned by: f F (f ) h] = lim F f + th] ; F f ] (3.33) The limit should be taken in the norm kk2 of B2 ,thus one has to check: lim k F f + th] ; F f ] ; F (f ) h]k = 0
t!0
t!0
The existence of the G^teaux-derivative is guaranteed if F is also Frechet-di erentiable and in that a case they are equal. However existence of the Gateaux-derivative does not guarantee the existence of the Frechet-derivative. This is easy to imagine as F might be di erentiable in some directions but not in all directions, or those derivatives might not be equal. We will now give some examples of the Gateaux-derivative. Consider F : C a b] ! R de ned by:
(3.34)
F f] =
b 4
Then the Gateaux-derivative F (f ) : C a b] ! R is given by: f Z b 4 4 F (f ) h] = lim 1 ( (f + th) 3 dx ; Z b f 3 dx) =

b 4 4 h = tlim 1 f 3 (1 + 4 t f + O(t2 )) ; f 3 dx = !0 t 3
Z
f 3 (x)dx
(3.35)
t!0
3 f hdx Another important example is the derivation of the Euler-Lagrange equations. Consider the Banach space of continuously di erentiable functions on some subspace V of three-dimensional space, denoted by C 1 (V ). We can then de ne the functional F : C 1 (V ) ! R by: = tlim !0
a a
3
b4 1
3 3 f hdx + O(t) =
b4 1
F f] =
in which L is a local function of f and rf . The functions f and rf are supposed to disappear on the edge of V . Then the Gateaux-derivative of F is given by: F (f ) h] = lim 1 Z L(f + th rf + trh) ; L(f rf )dr =
L(f rf )dr
(3.36)
@ L h + @ L rhdr = Z ( @ L ; r @ L )hdr (3.37) @ rf @ rf V @f V @f If we are looking for extreme values of F then F (f ) h] must be zero for all variations h. In that f
=
Z
= tlim 1 !0 t
L t @ L h + t @ rf rh + O(t2)dr = @f
t!0
case we must have:
which yields the famous Euler-Lagrange equations.
@L ; r @L = 0 @f @ rf
(3.38)
3.4. FUNCTIONAL DIFFERENTIATION
51
We will now adress the question of higher order derivatives. We note that if F : B1 ! B2 then F (f ) is a linear operator. It is an element of the space B1 B2 ]. So we can de ne an operator f F : B1 ! B1 B2 ] which assigns to each element f 2 B1 the functional derivative F (f ) in f . f f Thus:
3.4.3 Higher order derivatives
operator:
F : f 7! F (f ) 2 B B ] 1 2 f f As the space B1 B2 ] is also a Banach space we can di erentiate

2F f 2 (f ) : B1 ! B1 B2 ]
(3.39)
F f
again and obtain a linear (3.40)
thus 2fF (f ) 2 B1 B1 B2 ]]. This is called the second order derivative of F . We can now start 2 allover again and de ne an operator: This operator can then be further di erentiated and so on. If we go on like this , this procedure nally gives our nth -order derivative which is a functional:
nF fn n : f 7! fF (f ) 2 |B1 B1 {z 1 : : : B1 B2 ] : : :]]] B n } n
Z
2F 2F 2 : f 7! f 2 (f ) 2 B1 B1 B2 ]] f
(3.41)
(3.42)
Let us make this more concrete with an example. Take the functional F : C a b] ! R de ned by:
F f] =
The functional derivative is a linear functional F (f ) : C a b] ! R given by: f We will denote this functional by 0 2f 0 . We can now de ne the operator F : C a b] ! C a b] R] f by: This assigns to each function f a linear functional. We now can di erentiate this operator again. This yields again a linear operator which is an element of C a b] C a b] R]] de ned by: The proof of this statement is not di cult. If we denote the norm on the space C a b] R] by kk3 then we have: 2 lim 1 k F (f + h) ; F (f ) ; F (f ) h]k =
khk1 !0 khk1
b 2 f (x)dx
(3.43)
F (f ) h] = Z b 2f (x)h(x)dx f a F f ] =0 2f 0 f
(3.44)
(3.45)
2F 0 0 f 2 (f ) h] = 2 h
(3.46)
f2
52
0 00 0 0 0 0 0 0 = lim k 2(f + h) k;k2f ; 2 h k3 = lim kkhkk3 = 0 h1 khk1 !0 khk1 !0 1 2F As f 2 assigns to each point f 2 C a b] a linear operator working on a point h 2 C a b] which in its turn assigns a linear operator working on a point k 2 C a b] we can also regard 2fF as an 2 operator which assigns to each point f 2 C a b] a bilinear operator on C a b]:
de ned by:
2F 2F 2 : f 7! f 2 (f ) 2 C a b] C a b] R] f
(3.47) (3.48) (3.49)
This is completely general. We can equivalently view n F as an operator: fn whichs assigns to each f 2 B1 a multilinear operator on the product space B1 : : : B1 which is also a Banach space: nF n (f )(h1 : : : hn ) 7! fF (f ) h1 : : : hn ] 2 B2 (3.50) n fn
nF fn
Z 2F 0 h0 k] = b 2h(x)k(x)dx f 2 (f ) h k] = 2 a
: B1 ! |B1 :{z: B1 B2 ] : }
n
This is analogous to the case of vector calculus in which the nth -order derivatives are multilinear tensors. In our example we have as a rst order derivative the linear functional:
and as a second order derivative the bilinear functional: Z b 2F (f ) h k] = 2h(x)k(x)dx (3.52) f2 a Those operators are often given by their integral kernels when it is possible to regard the functional derivative as a linear integral operator. In the case of our example we have:
F (f ) h] = Z b 2f (x)h(x)dx f a
(3.51)
where the integral kernels are given by: F f (x) = 2f (x)
F (f ) h] = Z b F h(x)dx f a f (x) Z bZ b 2F 2F (f ) h k] = h(x)k(y)dxdy f2 a a f (x) f (y)
(3.53) (3.54) (3.55) (3.56)
In general we can write: Z nF n (f ) h1 : : : hn ] = f (x ) : :F f (x ) h1 (x1 ) : : : hn (xn )dx1 : : : dxn (3.57) fn 1 : n This completes our discussing of the higher order derivatives. We can now go on to de ne Taylor sequences, but before we do that we will rst give extremal points of functionals and give an overview on functional integration.
f (x) f (y) = 2 (x ; y)
2F
3.5. VARIATIONAL CALCULUS
53
3.5 Variational calculus
3.5.1 Extremal points of functionals
One of the most important applications of functional calculus concerns the determination of extremal points. One usually looks for local or global minima of functionals. Important applications can be found in almost all areas of physics. In the following we will discuss local minima of a functional F : B 7! R from a Banach space B to the real numbers. All the results derived can easily be applied to the case of local maxima as they are local minima of the functional ;F . We rst state a de nition. A functional F has a local minimum at f0 2 B when there exists a neigbourhood V (f0 ) = ff 2 B jkf ; f0 k < g such that
F f0 ] F f ] F (f ) = 0 f 0
8f 2 V (f0)
(3.58)
Let F be Frechet di erentiable ( G^teaux di erentiable in every direction h 2 B ). Then a necessary a condition for F to have a local minimum at f0 is (3.59) For a proof we de ne g : R 7! R by g(t) = F f0 + th] for h 2 B . As f0 is a local minimum we have g(0) = F f0 ] and g(t) F f0] . Therefor g : R 7! R has a local minimum at t = 0 and therefore dg=dt(0) = 0 and we nd So F= f (f0 ) = 0. The following theorem is useful for practical applications. Let F : B 7! R be twice di erentiable. A su cient condition for f0 to have a minimum at f0 2 B is that 0 = dg (0) = F (f0 ) h] dt f
8h 2 B
(3.60)
F (f ) = 0 f 0
(3.61) 0
and
k(t) = F f0 + th] + (1 ; t) F (f0 + th) h] f Then k(1) = F f0 + h] and k(0) = F f0 ] because the derivative at f0 vanishes dk (t) = (1 ; t) 2 F (f + th) h h] dt f2 0
Which yields
To prove this we de ne k : R 7! R by
2F f 2 (f ) h h]
8f h 2 B
(3.62) (3.63)
(3.64)
F f0 + h] ; F f0 ] = k(1) ; k(0) =
and we nd that F f0 + h] F f0 ] , and therefore f0 must be a local minimum.
2 (1 ; t) fF (f0 + th) h h]dt 0 2
(3.65)
54
3.5.2 Convex functionals
An important class of functionals for variational calculations is the set of convex functionals. A functional F : B 7! R is convex when for f1 f2 2 B and 0 1 2 1 with 1 + 2 = 1 we have F 1 f1 + 2 f2 ] 1F f1 ] + 2 F f2] (3.66) It is readily veri ed that an equivalent de nition is
The importance of convexity for variational problems can be inferred from the following statement. If F : B 7! R is convex and F is twice di erentiable then To prove this we rst note that due to the convexity of F we have for 0 t 1 F f0 + t(f ; f0)] = F tf + (1 ; t)f0] tF f ] + (1 ; t)F f0] and therefore F f0 + t(f ; f0)] ; F f0 ] F f ] ; F f ]
2F f 2 (f ) h h]
i=1
i fi ]
i=1
i F fi ]
i=1
i=1
(3.67)
8f h 2 B
(3.68) (3.69) (3.70) (3.71) (3.72)
Adding both inequalities then yields
F (f ) f ; f ] F f ] ; F f ] 0 0 f 0 Similarly by interchanging f and f0 we nd F (f ) f ; f ] F f ] ; F f ] 0 0 f
Taking the limit t # 0 and using the di erentiability of F we nd
F (f ) ; F (f ) f ; f ] 0 (3.73) 0 f f 0 This inequality means that F= f is a monotonously increasing functional. If we take f = f0 + th and take the limit t # 0 and use the fact that F is twice di erentiable we have 2 (3.74) 0 lim 1 F (f0 + th) ; F (f0 ) h] = fF (f0 ) h h] 2 t#0 t f f which proves our statement. The converse, which we will not prove here, is also true. If F is twice di erentiable and the second derivative is always larger than or equal to zero then F is convex. From the results of the previous section we can now conclude that if F= f (f0 ) = 0 for a twice di erentiable convex functional then f0 is a local minimum of F . F (f ) f ; f ] 0 f 0
We will now discuss one other quantity which is useful to prove di erentiability of convex functionals, which is the tangent functional or subgradient. Suppose that F : B 7! R is a convex functional which is di erentiable at f0 then
F f ] ; F f0 ]
8f 2 B
(3.75)
3.6. FUNCTIONAL INTEGRATION
55
This equation which has been derived above is equivalent to equation 3.71. The above equation is a special case of the more general equation
F f ] ; F f0 ] L f ; f0] 8f 2 B (3.76) where L : B 7! R is a bounded linear functional. If for a convex functional F such a linear functional exists at f0 then F is called subdi erentiable at f0 . The functional L is called a subgradient or a
tangent functional. One can prove the following statement. For a proof see 66]. If there is for a convex functional F a unique tangent functional at f0 2 B then F is G^teaux di erentiable at f0 a and
L = F (f0 ) f
(3.77)
G^teaux di erentiability of a convex functional is therefore equivalent to the uniqueness of a tangent a functional. This can be illustrated with the following example. Let the Banach space B be the real numbers B = R and the function F : R 7! R de ned as F (x) = jxj. This is a convex function. The linear functions L1 (x) = x and L2 (x) = ;x are both subgradients at the point x = 0. Therefore there is no unique subgradient at x = 0 and F is not di erentiable at x = 0. For x > 0 there is the unique subgradient L1 and for x < 0 there is a unique subgradient L2 . Therefore F is di erentiable for x 6= 0.
3.6 Functional integration

3.6.1 The line integral
As a motivation for the de nition of line integrals in Banach spaces we will rst give a familiar example of line integration in ordinary vector calculus. Suppose we have a curve ~ : R ! R3 in three-dimensional space starting at ~ = ~ (0) and ending in ~ = ~ (1). Suppose we also have a scalar a b function F : R3 ! R and we de ne F (t) = F (~ (t)) then it follows: From this formula we can deduce:
dF (t) = rF (~ (t)) d~ (t) dt dt
(3.78)
as:
Z Z ~ ) ; F (~ ) = 1 dF (t)dt = 1 rF (~ (t)) d~ (t)dt F (b a (3.79) dt 0 dt 0 In general we can for a given vector eld ~ calculate the line integral along curve ~ which is de ned v Z Z
~ d~ v
~ (~ (t)) d~ (t)dt v dt 0
(3.80)
In general the outcome of this integral is path dependent.Only if there exists some scalar function F with rF = ~ then the outcome of the integral is path independent. In that case we have a v conservative vector eld. Let us generalise this to Banach spaces. Consider a functional F : B ! R where B is Banach space with norm kk. We can then de ne for a given f h 2 B the function k : R ! R by:
k(t) = F f + th]
(3.81)
56 Then:
dk (t) = lim k(t + s) ; k(t) = s!0 dt s = s!0 F f + th + sh] ; F f + th] = F (f + th) h] lim s f This thus yields the Gateaux-derivative in point f + th. This then gives: Z 1 Z 1 F (f + th) h]dt F f + h] ; F f ] = k(1) ; k(0) = dk (t)dt = 0 dt 0 f
So we have derived the formula:
(3.82) (3.83)
F fb ] ; F fa ] =
This de nes the line integral of F 2 B B R]] along the path (t) = fa + t(fb ; fa) 2 B . This can f be generalised to arbitrary elements v 2 B B R]]. Suppose we have an operator v : B ! B R] then the integral of v from fa to fb is de ned by:
Z
(f + t(fb ; fa )) fb ; fa ]dt 0 f a
1 F
fb fa
v f ]df
(By v(f ) h] we mean v in point f 2 B working on h 2 B .) This integral can also be de ned for more general paths by splitting up the general path in small straight pieces on which the integral is de ned and then taking the limit of an in nite number of in nitesimal pieces. This then leads to the following de nition:
Z
v(fa + t(fb ; fa )) fb ; fa ]dt
(3.84)
v f ]df
where : R ! B is a di erentiable path in Banach space. Using this de nition we can prove the following statements just as for ordinary line integrals:
Z
v( (t)) d (t)]dt dt v f ]df

Z
(3.85)
v f ]df = ;
1
;
2
(3.86)
1+ 2
v f ]df +
v f ]df =
v f ]df
(3.87)
where ; is with reverse orientation thus with parameter t running from 1 to 0. Further is 1 + 2 the combined path obtained by rst walking along path 1 and subsequently walking along path 2 . If the integrals are path independent the outcome of the integrals can only depend on the endpoints of the path and we can denote the integral by:
Z
v f ]df =
fb fa
fb fa
v f ]df
Z
(3.88)
where fa = (0) and fb = (1). In that case the above statements give:
Z
v f ]df = ;
fa fb
v f ]df
(3.89)

Z
57 (3.90)
fc fa
v f ]df +
fb fc
v f ]df =
fb fa
v f ]df
Suppose we have an operator v 2 B B R]] for which the line integral is path independent. We will prove for the case of a straight path (the particular path is not important as the integral is path independent) that if we de ne the functional F : B ! R by
F f] =
with v 2 B B R]] then
fa
v k]dk =
v(fa + t(f ; fa)) f ; fa ]dt
(3.91) (3.92)
The proof is analogous to the corresponding proof for the Riemann integral.We calculate: F f + th] ; F f ] = 1 Z f +th v k]dk = 1 Z t v(f + sh) h]ds (3.93) t t t Here we used the path independence property:
Z
F (f ) = v(f ) 2 B R] f
f +th
Then it follows: Z t j F f + th] ; F f ] ; v(f ) h]j = j 1 (v(f + sh) h] ; v(f ) h])dsj
fa
v k]dk ; t
f fa
v k]dk =
f +th
v k]dk
(3.94)
1 Z 1 jv(f + sh) h] ; v(f ) h]jds
t 0
t 0
1 max fjv(f + sh) h] ; v(f ) h]jtg = t s2 0 t] (t ! 0) (3.95)
= max fjv(f + sh) h] ; v(f ) h]jg ! 0

s2 0 t]
Thus we have:
and hence we have proven our statement. Thus path independency for line-integrals of v implies that v(f ) is the functional derivative of some functional F : B ! R We will now give some examples. Suppose we have an operator v : C a b] ! C a b] R] de ned by: 4 1 (3.96) v f ] = 3 f 3 2 C a b] R] The action of this operator on a function h 2 C a b] is de ned as: Z b 1 3 v(f ) h] = 4 f 3 h] = 4 f 1 (x)h(x)dx (3.97) 3 a 3 Then the integral of v from fa to fb is given by: Z f Z 1 b 1 4 v k]dk = 3 (fa + t(fb ; fa )) 3 fb ; fa ]dt = fa 0
F (f ) h] = lim F f + th] ; F f ] = v(f ) h] t!0 f t
58 = = =
Z

b a
Z
dx 4 (fb ; fa ) 3
3 dt(fa + t(fb ; fa)) 1 =
b a b
3 1 4 3 0 dx 3 (fb ; fa ) 4 f ; f (fa + t(fb ; fa )) 4 ]1 = a b
dx(fb3 ; fa3 ) = F fb ] ; F fa ]
Z
where F : C a b] ! R is de ned by:
F f] =
b 4 a
f 3 (x)dx
(3.98)
One can check that the functional derivative of F yields v(f ) and hence the result is path independent. Our nal example in this section also involves gradients. De ne the functional F : C 1 (V ) ! R in which C 1 (V ) is the space of continuous di erentiable functions on some subset of three-dimensional space R3 by: Z 2 F f ] = (rf ) dr (3.99)
V
The functions f and rf are supposed to disappear on the edge of V . Using the Euler-Lagrange equations 3.37 we nd for the functional derivative: 2 2 v(f ) = F (f ) = @ ( (rf ) ) ; r @ ( (rf ) ) =
3 f4
4 @f f 3
4 @ rf f 3
2 2 2 2 f 4 8 = ; 4 (rf ) ; r (2 r4 ) = ; 3 (rf ) ; 2 r 4f + 3 (rf ) = 7 7 7 3 f3 f3 f3 f3 f3 2 2 = 4 (rf ) ; 2 r 4f 2 C 1 (V ) R] 3 7
f3
f3
(3.100)
We will now integrate this derivative from 0 2 C 1 (V ) to f 2 C 1(V ). Then we must calculate: Z f Z 1 Z Z 1 2 4 )2 v k]dk = v(tf ) f ]dt = drf dt 3 (trf 7 ; 2 tr f = 4 0 0 V 0 (tf ) 3 (tf ) 3 4 (rf )2 Z 1 dtt; 1 ; 2 r2 f Z 1 dtt; 1 ) = 3 3 = drf ( 3 7 4 V f3 0 f3 0
Z
= =
dr(2 (rf4 ) ; 3 r 1f ) = V f3 f3
(r f ) 2 = F f ] 4
dr (rf4 ) ; 3 V f3
f drr ( r1 ) = V f3
(3.101)
f3
which recovers our starting functional.
59
3.6.2 Integrability conditions
In the calculation of line integrals it is often useful to know whether a given line integral is path independent because in that case we can try to deform the path to a contour for which the integral is easy to calculate without changing the outcome of the integral. We thus like to have an easy to verify criterion which tells us whether a given line integral is path independent. For everyday three-dimensional vector elds such a criterion is easy to give. The line integral along the eld is path independent if this eld is rotationless. Mathematically this means that its curl should vanish. This condition is readily derived. Suppose we have a vector eld v : R3 ! R3 . If the line integral along this vector eld is path independent then this vector eld should be the gradient of some scalar function f : R3 ! R thus: v = rf (3.102) Or in terms of vector components: vi = @if (3.103) If f is twice di erentiable then it follows that: @k vi ; @i vk = @k @if ; @i @k f = 0 (3.104) and hence: r v=0 (3.105) In the derivation of this condition we used the fact that we could interchange the di erentiation with respect to di erent variables. In general it is true for a n-times di erentiable function f : R3 ! R that: @i1 @i2 : : : @in f = @ip(1) @ip(2) : : : @ip(n) f (3.106) where p is an arbitrary permutation of the numbers 1 2 : : : n. A direct generalisation of this statement to n-times di erentiable operators F : B1 ! B2 on Banach space would be: where hi 2 B1 . Using the de nition of di erentiation one can prove this statement. In particular we have for the bilinear operator 2fF (f ) 2 B1 B1 B2 ] that: 2 for h k 2 B1 . If it is possible to view this operator as an integral operator then the integral kernel should be symmetric: = f (y) Ff (x) (3.109) f (x) f (y) Suppose we have in point f 2 B1 a linear operator v(f ) : B1 ! B2 which is the functional derivative of an operator F : B1 ! B2 , thus:
2F 2 2F 2 (f ) h k] = fF (f ) k h] 2 f2
nF f n (f ) h1 h2 n : : : hn ] = fF (f ) hp(1) hp(2) : : : hp(n) ] n
(3.107)
(3.108)
v(f ) = F (f ) 2 B1 B2 ] f
(3.110)
60
If we let this operator work on an element h 2 B1 and we take the derivative in direction k 2 B1 then: v (f ) h] k] = lim v(f + tk) ; v(f ) h] =
2 2 2 = tlim 1 ( fF (f + tk) h] ; F 2 (f ) h]) = fF (f ) h k] 2 !0 t 2 f
t!0
(3.111)
Thus using the symmetry of the second order derivative (equation 3.108) we can conclude that if v 2 B1 B1 B2 ] is the derivative of a twice di erentiable operator F : B1 ! B2 then: Using the integral kernel notation: v (f ) h] k] = Z v(x) h(x)k(y)dxdy (3.113) f f (y ) we must have for the integral kernels: v(x) = v(y) (3.114) f (y) f (x) This is a necessary condition for the path independence for the line integral of v. We now give some examples. Consider a functional F : C 2 a b] ! R on the space C 2 a b] of twice continuously di erentiable functions on the interval a b]. The functions are also required to vanish at the endpoints of the interval so we are really working in a subspace of C 2 a b] . Let the functional F be de ned by: Z b 1 df F f ] = 2 ( dx (x))2 dx (3.115) a Its functional derivative v(f ) = F (f ) 2 C 2 a b] R] is given by: f This is easily derived using the Euler-Lagrange equations 3.37. Its action on a function h 2 C 2 a b] is given by: (3.117)
v f
v (f ) h] k] = v (f ) k] h] f f
(3.112)
2 v(f ) = d f 2 C 2 a b] R] dx2
(3.116)
d2 f h] = Z b d2 f (x)h(x)dx dx2 a dx2 As v(f ) 2 C 2 a b] R] is the functional derivative of functional F the operator
should be a symmetric linear operator. This is easily checked: v (f ) h] k] = lim 1 Z b ( d2 (f + tk) ; d2 f )h(x)dx =
2 C 2 a b] C 2 a b] R]]
t!0
b d2 k
dx2
Z
dx2
dx2 (x)h(x)dx = ;
b a
dk (x) dh (x)dx dx dx
(3.118)
In the last step we used partial integration and the fact that h(a) = h(b) = 0. In this derivation all the limits are taken in the norm sense. A more careful derivation using norms shows that this last
3.7. TAYLOR EXPANSIONS
61
result is indeed correct. This last formula is obviously symmetric in h and k. If we use the integral kernel notation we have: v(x) = d2 (x ; y) (3.119) f (y) dx2 which is also symmetric. As a nal example we consider on the same space an operator v(f ) 2 C 2 a b] R] with a nonsymmetric derivative.Let v(f ) be de ned by:
v(f ) h] =
b a
Its derivative is given by: v (f ) h] k] = lim 1 Z b (( d (f + tk))2 ; ( df )2 )h(x)dx =
df ( dx (x))2 h(x)dx
t!0
(3.120)
dx
dx
b a
df dk 2 dx (x) dx (x)h(x)dx dx dy
(3.121)
This formula is not symmetric in h and k.We can also write this as: v (f ) h] k] = Z b Z b df (x) d (x ; y) h(x)k(y)dxdy
So the integral kernel is given by: v(x) df d (3.123) f (y) = dx dy (x ; y) This kernel is nonsymmetric so any line integral along v will be path dependent. This ends our discussion of functional integration. There is one important topic left which is very useful in the approximation of functionals and that is the de nition of Taylor-sequences. We will deal with this question in the next section.
a a
(3.122)
3.7 Taylor expansions

k(t) =
1 (1 ; t)n n=0 n!
X Z
In this section we will derive the Taylor formula for operators F : B1 ! B2 . For given f h 2 B1 de ne the function k : R ! B2 by:
N nF f n (f + th) |h
:{z: h}] :
n
(3.124)
we now will use the following formula for the Riemann integral:
k(1) ; k(0) =
in order to calculate dk (t) we must calculate: dt nF d ( (f + th)) = lim 1 ( n F (f + th + sh) h : : : h] ; n F (f + th) h : : : h]) = s!0 s f n dt f n fn
0 dt
1 dk
(t)dt
(3.125)
62
n+1 n

F F = f n+1 (f + th) |h :{z: h}] h] = f n+1 (f + th) |h :{z: h}] : :
n+1 n+1
(3.126)
It thus follows that:
N dk (t) = X ;1 (1 ; t)n;1 n F (f + th) h : : : h] dt fn n=1 (n ; 1)!
1 (1 ; t)n n+1 F (f + th) h : : : h] = 1 (1 ; t)N N +1 F (f + th) h : : : h] f n+1 N! f N +1 n=0 n!

X
(3.127)
We further have:
k(1) ; k(0) = F f + h] ;
X
Using equation 3.125 for the Riemann integral this yields: 1 n F (f ) h : : : h] + Z 1 1 (1 ; t)N N +1 F (f + th) h : : : h]dt(3.129) F f + h] = F f ] + n! f n f N +1 0 N! n=1
N
1 n=0 n!
X
nF f n (f ) h
: : : h]
(3.128)
If F : B1 ! B2 is in nitely di erentiable and if the restterm: Z 1 N +1 F 1 RN (f h) = N ! (1 ; t)N f N +1 (f + th) h : : : h]dt 0 goes to zero for N ! 1 ,thus if: lim kR (f h)k2 = 0 N !1 N then we have within the convergence radius: 1 X 1 nF F f + h] = F f ] + n! f n (f ) h : : : h] n=1
(3.130) (3.131) (3.132)
In the integral kernel notation we have: 1 Z n X 1 (3.133) F f + h] = F f ] + n! f (x ) : :F f (x ) h(x1 ) : : : h(xn )dx1 : : : dxn 1 : 2 n=1 We now have derived most formulas needed in practical application of functional calculus. This nal chapter therefore ends our review of functional calculus. A more rigorous discussion on all the discussed topics can be found in 64, 65].
Chapter 4
Density functionals for Coulomb systems

4.1 Introduction
In this section we will discuss the several functional analytical properties of density functionals for electronic systems with Coulombic interparticle interactions. To derive continuity and di erentiability of the density functionals we restrict ourselves to a certain set of external potentials, mathematically denoted as L1 (R3 ) + L3=2 (R3 ) with contains the physically important Coulomb potential as well as any nite sum of Coulomb potentials relevant to molecular systems. The discussion in this section is largely based on the mathematical papers by Lieb 19] and Englisch and Englisch 67, 68]
4.2 Conditions on the electron density and external potentials

The electron density corresponding to a normalized N -electron wave function (r) = N
X Z
is de ned as (4.1)
We rst put some constraints on the wave function. First of all because of the probability interpretation one likes the wavefunction to be normalizable to one, so we require k k < 1 where the norm is de ned as
1 ::: N
j (r 1 r2 2 : : : rN N )j2 dr2 : : : drN
k k=
Secondly because of the superposition principle in quantum mechanics one requires that also some in nite linear combinations exist and are normalizable. More precisely we rst de ne
M=
X
1 ::: N
j (r1 1 : : : rN N )j2 dr1 : : : drN
(4.2)
with k m k = 1 and
X
i=1
am
(4.3)
i=1
jai j2 = 1
63
(4.4)
64
CHAPTER 4. DENSITY FUNCTIONALS FOR COULOMB SYSTEMS

(4.5)
One can easily check this to be a Cauchy sequence lim k ; M k ! 0 (M ! 1) M !1 M +1
Then we would like the limit function to exist, i.e. we would like that our space of normalizable functions contains some limit functions such that lim k ; M k = 0 (4.6) M !1 so we require the above Cauchy sequence to converge, more compactly we like our space of functions to be complete with respect to the square integral norm. It can be proven that the smallest space which is complete with respect to the above norm is the space L2 of functions for which the Lebesque integral of j j2 exists. For a de nition of the Lebesque integral we refer to the previous chapter. A third requirement on wavefunctions is that their kinetic energy expectation value must be nite. The expectation values of the potential energy operators we will discuss later on. So we require that N 1 X X Z jr (r : : : r )j2 dr : : : dr < 1 T ]= 2 (4.7) i 1 1 N N 1 N i=1 1 ::: N
N
Here r is de ned almost everywhere (a precise de nition of r for a L2 -function can be given using Fourier transforms ). This implies that
X Z
and so we now nd that 2 All these constraints on have consequences for the constraints on the electron density . First of all because 2 L2 (R3N ) we nd from equation 4.1 that 2 L1 (R3 ) which means that the Lebesque integral of the electron density is nite
Z
1 ::: N
j (r1 1 : : : rN N )j2 +
i=1 H 1(R3N ).
jri (r1 1 : : : rN N )j2 dr1 : : : drN < 1
(4.8)
(r)dr < 1
(4.9)
We now will show that the niteness of the kinetic energy implies that must also be in L3 (R3 ) which together with the previous results implies that 2 L1 (R3 ) \ L3 (R3 ). If we consider a wave function 2 H 1 (R3N ) which yields density then
r (r) = N
Z
Using the Schwarz inequality
1 ::: N
(r 1 : : : rN N )r (r 1 : : : rN N )dr2 : : : drN + c:c:
(4.10)
j f (x1 : : : xN )g (x1 : : : xN )dx1 : : : dxnj2

Z
jf (x1 : : : xN )j2 dx1 : : : dxN
jg(x1 : : : xN )j2 dx1 : : : dxN
(4.11)
we nd
jr (r)j2
X Z
4N 2
1 ::: N
(r 1 : : : rN N )j2 dr2 : : : drN
1 ::: N
jr (r 1 : : : rN N )j2 dr2 : : : drN
4.2. CONDITIONS ON THE ELECTRON DENSITY AND EXTERNAL POTENTIALS
65
1 N X Z jr (r : : : r )j2 dr : : : dr = 8 (r) 2 (4.12) 1 N N 2 N 1 ::: N So we nd that Z 2 1Z q TW ] = 2 jr (r)j2 dr = 1 (r ((rr))) dr T ] < 1 (4.13) 8 p Therefore (r) 2 H 1 (R3 ). If we now use the Sobolev inequality for functions f 2 H 1 (R3 )
Z
1 ( 2 ) 4 Z jrq (r)j2 < 1 3 (4.15) 3 so we conclude that 2 L3 (R3 ) which together with 2 L1 (R3 ) implies 2 L1 (R3 ) \ L3 (R3 ). Some de nitions are useful for further discussions in this section. We de ne the sets
Z
we nd that
4 jrf (r)j2 dr 3( 2 ) 3
jf (r)j6 dr
1 3
(4.14)
3 (r)dr
SN = f j (r) 0 p 2 H 1 (R3)
0 2 L3 (R3 )
(r)dr = N g
(4.16) (4.17)
1 2
and
RN = f j (r) (r)dr = N g The set RN is convex, that is, if 1 and 2 2 RN then = 1 1 + 2 2 2 RN with 0
and 1 + 2 = 1 . This follows from
Z
(r)dr = 1
1 (r)dr + 2
2 (r)dr = N
(4.18) (4.19) (4.20) (4.21) (4.22) (4.23)
( 1 r 1 + 2 r 2 )2 This yields (r )2 4 ( 1 (rp 1 )2 + 2 (rp 2 )2 ) and we nd

Z
k 1 1 + 2 2 k3 1k 1 k3 + 2 k 2k3 < 1 The set SN is also convex. This follows from r = 1 r 1 + 2r 2 = 2 1 p 1rp 1 + 2 2 p 2rp 2 = 1 r 1 + 2 r 2 p p and = 2 and from the Schwarz inequality with = 2
1 1 1 2 2 2 ( 2 + 2 )((r 1 )2 + (r 2 )2 ) 1 2
and
(p )2 dr
p p 1 ( 1 )2 dr + 2 ( 2 )2 dr < 1
Z Z
So 2 SN . In particular it follows that the von Weiszacker kinetic energy functional 4.13 is a convex functional, i.e. (4.24) TW 1 1 + 2 2 ] 1TW 1 ] + 2 TW 2 ]
66
We will now prove the following statement. For any electron density 2 SN there is a Slater determinant wavefunction 2 H 1 (R3N ) which yields this density (i.e. with satis es relation 4.1 ). The proof is by explicit construction. We write r = (r1 r2 r3 ) and de ne 2 Z r1 dv Z +1 dv Z +1 dv (v v v ) f (r ) = (4.25)
N ;1 1 ;1 2 ;1 3 1 2 3 The function f is monotonously increasing with f (;1) = 0 and f (+1) = 2 . We de ne the

orbitals n by
s
n (r) = n (r1 r2 r3 ) =
for n = 0 : : : N ; 1. Then j (r)j2 = (r)

n
(r1 r2 r3 ) exp (inf (r )) 1 N
(4.26) (4.27)
and for n 6= m we have

Z
So the orbitals n form an orthonormal set which sum to the prescribed density and therefore the corresponding Slater determinant wavefunction will yield the same density. After having discussed the density and the expectation value of the kinetic energy operator we will now discuss the expectation value of the potential energy operators. Because 2 H 1 (R3N ) it follows 19] that for ^ 1X 1 (4.29) W = 2 jr ; r j j i6=j i we have ^ h jW j i < 1 (4.30)
Z +1 df i(n;m)f (r1 ) dr 1Z (r) n (r)dr = N (r)ei(n;m)f (r1 ) dr = 21 1 m dr1 e ;1 = 2 i(n1; m) (ei(n;m)f (+1) ; ei(n;m)f (;1) ) = 0
(4.28)
Also a lower bound for this quantity can be proven. If yields density we have Z Z ^ j i > 1 (r1 ) (r2 ) dr1 dr2 ; C 4 (r)dr 3 h jW (4.31) 2 jr1 ; r2 j ^ with C = 1:68 69]. We now turn to the expectation values of the external potential V . From the condition
Z
(r)v(r)dr < 1
(4.32)
and
This follows directly from

Z
2 L1(R3 ) \ L3 (R3) we can deduce some constraints on the external potential v(r). If 1 (R3 ) then the above integral exists for bounded potentials,i.e. for potentials v 2 L1 (R3 ). 2L
(r)v(r)dr ess sup jv(r)j
Z
(r)dr < 1
(4.33)
4.3. PROPERTIES OF THE ENERGY FUNCTIONAL E V ]
67
2 if v 2 L1(R3 ). If 2 L3 (R3 ) then the above integral exists for potentials in the set L 3 (R3 ). This follows from the Holder inequality kfgk1 kf kpkgkq (4.34) 3 with 1=p + 1=q = 1. If v 2 L 2 (R3 ) then we nd using this inequality
Z
(r)v(r)dr = k vk1
k k3 kvk 32 < 1
(4.35)
^ The most general set of potentials for which the expectation value h jV j i exists is therefore the set 3 3 L 2 (R3 ) + L1(R3 ) = fvjv = u + w u 2 L 2 (R3 ) w 2 L1 (R3 )g (4.36) This set is a Banach space with norm (4.37) kvk = inf fkuk 32 + kwk1 jv = u + wg
r where r = jrj and
An important potential in this Banach space is Coulombic potential as it can be written as 1 = u(r) + w(r) (4.38)
; w(r) = (r r 1) (4.39) where is the Heaviside function, (x) = 0 if x 0 and (x3 = 1 if x > 0. One can readily ) 3 2 (R3 ) and w 2 L1 (R3 ). The Banach space L 2 (R3 ) + L1 (R3 ) does not contain check that u 2 L
external potentials that go to in nity as jrj ! 1. This choice precludes some physically interesting potentials such as the harmonic oscillator potential. These potentials can be handled with the methods to be disussed but then one has to put additional restrictions on such that the integral of v makes sense.
; u(r) = (1 r r)
4.3 Properties of the energy functional E v
3 For the external potentials in the set L 2 (R3 ) + L1(R3 ) we de ne the total energy functional E v] as ^ E v] = inf h jH j i (4.40)
3 for all v 2 L 2 (R3 ) + L1 (R3 ) and 0 1 2 1 and 1 + 2 = 1. This follows directly from the variational principle. If is a wave function corresponding to the in mum in equation 4.40 then ^ ^ ^ ^ ^ ^ ^ ^ ^ E v] = h jT + V + W j i = 1h jT + V1 + W j i + 2h jT + V2 + W j i (4.42) 1 E v1 ] + 2 E v2 ]
where 2 H 1 (R3N ) and where the wave function is normalized to one k k = 1. We will prove some properties of this functional. First of all we have that E v] is concave, that is E 1 v1 + 2v2 ] 1 E v1 ] + 2 E v2 ] (4.41)
68
where we used 1 + 2 = 1. A simple application of the above relation is the following. Consider a molecular Hamiltonian with N atoms of positive nuclear charge Zi at positions Ri .
v(r) = ;
X
if we de ne the total nuclear charge
i=1 jr ; Ri j
Zi
(4.43)
Z=
and de ne i = Zi =Z then we nd using the concavity property of E v] that
i=1
Zi
N
(4.44)
E v] = E
=
N
i=1
; jr ;i Z j ] R
i
So we conclude that the total electronic energy of a molecule with total nuclear charge Z is always larger than or equal to the total energy of an atom with the same nuclear charge Z and the same number of electrons. For instance the electronic energy of the hydrogen molecule is larger than the electronic energy of the helium atom. A second property of E v] is that it is monotonously decreasing, that is, if v1 (r) v2 (r) for all r (almost everywhere) then E v1 ] E v2 ]. This follows again from the variational property. If is a wave function corresponding to the in mum in equation 4.40 then ^ ^ ^ ^ ^ ^ ^ ^ E v2 ] = h jT + V2 + W j i = h jT + V1 + W j i + h jV2 ; V1j i Z E v1 ] + (r)(v2 (r) ; v1 (r))dr E v1 ] (4.46) We further state without proof a third property of E v], which is : E v] is continuous3 in the 3 L 2 + L1 -norm. So if we have a sequence of external potentials fvig in the Banach space L 2 + L1 2 with norm kk converging to v 2 L 3 + L1 then lim kv ; vk = 0 ) ilim jE vi ] ; E v]j = 0 (4.47) i!1 i !1 The proof of this can be found in the paper by Lieb 19].
i=1
i E ; jrj ] = E ; jrj ]
i E ; jr ; R j ] i i=1
(4.45)
4.4 The Hohenberg-Kohn functionals FHK and FEHK

]
We will now discuss the Hohenberg-Kohn functional which has been introduced by Hohenberg and Kohn in their well-known paper. First we will prove the Hohenberg-Kohn theorem. Suppose that 1 and 2 2 H 1 (R3N ) are non-degenerate ground state wave functions corresponding to external 3 potentials v1 and v2 2 L1 + L 2 with corresponding electron densities 1 and 2 . Then if v1 6= v2 + C where C is constant then 1 6= 2 . As a rst step we have that if v1 6= v2 + C that 1 6= 2 . This follows by contradiction. Suppose 1 = 2 = then by subtraction of the Hamiltonian for 1 and 2 we nd that (v1 ; v2 ) = (E1 ; E2 ) (4.48)
4.4. THE HOHENBERG-KOHN FUNCTIONALS FHK ] AND FEHK ]
69
If v1 ; v2 is not constant in some region then must vanish in this region for the above equation 2 to be true. However if v1 v2 2 L1 + L 3 then cannot vanish on an open set (a set with nonzero measure) by the unique continuation theorem. So we obtain a contradiction and we nd 1 6= 2 . So di erent potentials ( di ering more than a constant) give di erent wavefunctions. In the following we will denote 1 = v1 ] and 2 = v2 ]. These di erent wave functions also yield di erent densities. This follows again by contradiction. If 1 = 2 = then ^ ^ ^ ^ ^ ^ E v1 ] = h v1 ]jT + V1 + W j v1 ]i < h v2 ]jT + V1 + W j v2 ]i ^ ^ ^ = h v2 ]jT + V2 + W j v2 ]i + = E v2 ] + Likewise
Z Z
(r)(v1 (r) ; v2 (r))dr (4.49) (4.50) (4.51) (4.52)
(r)(v1 (r) ; v2 (r))dr

Z
E v2 ] < E v1 ] +
(r)(v2 (r) ; v1 (r))dr
3 where we only consider ground state densities from potentials in the set L1 + L 2 . The set AN is a subset of the previously de ned set SN . The densities in the set AN we will call pure state v-representable densities, shortly denoted as PS-V-densities. From the Hohenberg-Kohn theorem there is a unique external potential v (to within a constant) and a unique ground state wave function ] (to within a phase factor) which yields this density. On the set of PS-V densities we can therefore de ne the Hohenberg-Kohn functional FHK as ^ ^ FHK ] = h ]jT + W j ]i (4.53) we can then de ne the energy functional Ev as
Adding the both inequalities yields the contradiction E v1 ] + E v2 ] < E v1 ] + E v2 ] So we conclude that 1 6= 2 , which proves our statement. We now de ne the set AN as AN = f j comes from a non-degenerate ground stateg
Ev ] =
(r)v(r)dr + FHK ]
(4.54) an arbitrary other
If 0 is a ground state density corresponding to external potential v0 and ground state density then Z ^ ^ ^ Ev0 ] = (r)v0 (r)dr + FHK ] = h ]jT + V0 + W j ]i ^ ^ ^ h 0 ]jT + V0 + W j 0 ]i = Ev0 0 ] Therefore
(4.55) (4.56)
E v] = 2A inf
(r)v(r)dr + FHK ]
70
For application of the above formula we have to known FHK on the set AN . As the set AN is di cult to determine we need an extension of FHK to a larger better known set. We will however rst discuss the above functional. The functional FHK has some mathematical inconveniences. First of all its domain AN is not convex, that is, if 1 2 2 AN then not necessarily 1 1 + 2 2 2 AN with 0 1 2 1 and 1 + 2 = 1. We will prove this in a later instant. The functional FHK is however convex, that is, if 1 2 2 AN and if 1 1 + 2 2 2 AN with 0 1 2 1 then FHK 1 1 + 2 2 ] 1FHK 1 ] + 2 FHK 2] (4.57) This is readily proved. Suppose that the ground state densities 1 2 1 1 + 2 2 2 AN correspond to the external potentials v1 v2 and v. Then
Z
(r)v(r)dr + FHK ] = ^ ]jH j ]i = 1 h
(r)v(r)dr + h ]i + 2 h
^ ^ ]jT + W j ^ ]jH j ]i
Z
]i =
=h
^ ]jH j
Z
^ ^ ^ ^ 1 h 1 ]jT + W j 1 ]i + 2 h 2 ]jT + W j 2 ]i + ( 1 1 (r) + 2 2 (r))v(r)dr = 1 FHK 1 ] + 2 FHK 2 ] + (r)v(r)dr (4.58)
and we obtain the convexity of FHK . The functional FHK is de ned on the set AN . This is not a convex set. There are convex combinations of PS-V ground state densities which are not in AN . An example of this is a convex combination of densities corresponding to a q-fold degenerate ground state multiplet =
X
which in general is not in AN . We will demonstrate this at the end of this section. We can however readily extend our functional FHK to this type of densities. We de ne X X BN = f = i i j i comes from the same v for all i (4.60) i=1 0 i 1g We call these densities ensemble v-representable densities or E-V-densities. We will now extend the functional FHK to the set BN of ensemble v-representable densities. The corresponding ensemble functional we will denote by FEHK . For a degenerate ground state multiplet f i ig with q-fold degeneracy corresponding to some external potential v we de ne the ensemble density matrices ^ X i j i ih i j D=
i=1 q
X
i=1
i i
i=1
i=1
(4.59)
the corresponding ensemble density is given by ^ (r) = TrD ^(r) (4.62) where ^(r) is the density operator. We now prove the following extension of the Hohenberg-Kohn ^ ^ theorem to arbitrary degenerate ground states. Suppose D1 and D2 are ground state ensemble density matrices belonging to v1 and v2 resp., with corresponding densities 1 and 2 . If v1 6= v2 + C
i=1
i=1
(4.61)
4.4. THE HOHENBERG-KOHN FUNCTIONALS FHK ] AND FEHK ]
71
with C a constant then 1 6= 2 . The proof is analogous to the proof of the non-degenerate case. First the sets of ground state ensemble density matrices corresponding to the two di erent potentials v1 and v2 are disjoint. Suppose v1 generates the ground state multiplet A1 = fj i i i = 1 : : : q1 g and v2 generates the ground state multiplet A2 = fj i i i = 1 : : : q2 g. All the wave functions within these multiplets may without loss of generality be chosen orthonormal. Then none of the wave functions in the sets A1 and A2 are equal. This follows from the same argument as used in the proof of the Hohenberg-Kohn theorem for the non-degenerate case. In particular, as the sets A1 and A2 are only de ned to within a unitary transformation no j i i in A2 is a linear combination of the j i i in A1 . This then implies that two ground state ensemble density matrices constructed from the ground states in A1 and A2 are di erent ^ X i j i ih i j 6= X i j i ih i j = D2 ^ D1 =
i=1 i=1
P P
q1
q2
(4.63)
where i = i = 1. This follows for instance by taking the inproduct on both sides with j m i as the j i i are not linear combinations of the j i i. ^ ^ ^ ^ ^ ^ ^ ^ Secondly, if H1 = T + V1 + W and H2 = T + V2 + W then ^ ^ ^ ^ TrD1 H2 > TrD2 H2 (4.64) This follows directly from ^ ^ X i h i jH2 j i i > X i h i jH2 j i i = ^ ^ TrD1 H2 =
i=1 i=1 q1 q1
1 X
i=1
i E v2 ] = E v2 ] =
2 X
i=1
^ ^ ^ 2 h i jH2 j i i = TrD2 H2
(4.65)
^ ^ Now we can show that D1 and D2 yield di erent densities. We proceed again by reductio ad absurdum. Suppose 1 = 2 = . Using the last result we nd ^ ^ ^ ^ ^ ^ E v1 ] = TrD1 H1 = TrD1 (H2 + V1 ; V2 ) = ^ ^ = TrD1 H2 + = E v2 ] +
Z Z
^ ^ (r)(v1 (r) ; v2 (r))dr > TrD2 H2 +
(r)(v1 (r) ; v2 (r))dr = (4.66) (4.67)
(r)(v1 (r) ; v2 (r))dr

Z
Likewise we have
E v2 ] > E v1 ] +
(r)(v2 (r) ; v1 (r))dr
which added to the last inequality leads to the contradiction E v1 ] + E v2 ] > E v1 ] + E v2 ] (4.68) ^ ^ Therefore D1 and D2 must give di erent densities, which proves the theorem. Within the set of ensemble ground state density matrices corresponding to the same potential how^^ ever, two di erent density matrices can yield the same density. The energy TrDH for those di erent
72
density matrices is however the same. For every E-V-density we can therefore unambiguously de ne ^ ^ ^ FEHK ] = TrD ](T + W ) (4.69) ^ where D ] is any of the ground state ensemble density matrices corresponding to . We can now de ne an extension of the energy functional Ev to the set of E-V-densities Z ^ ^ Ev ] = (r)v(r)dr + FEHK ] = TrD ]H (4.70) Similarly as for FHK we easily can prove The functional FEHK is an extension of FHK as we have FEHK ] = FHK ] if 2 AN (4.72) This follows directly from the fact that for a non-degenerate ground state j ]i corresponding to ^ we have D ] = j ]ih ]j ,so ^ ^ ^ ^ ^ FEHK ] = TrD ](T + W ) = h ]jT + W j ]i = FHK ] (4.73) We can furthermore prove that FEHK is convex by the same proof as for FHK . Nothing is however known on the the convexity of the set of E-V-densities BN which constitute the domain of FEHK . As we will now demonstrate the subset of PS-V-densities AN of BN is not convex. More precisely we will now show that there are E-V-densities which are not PS-V-densities. As any E-V-density is a convex combination of PS-V densities this then demonstrates the non-convexity of AN . Consider an atom with total angular momentum quantum number L > 0 which has a 2L + 1degenerate ground state. The external potential v is the Coulomb potential. The degeneracy is due to the fact that the Hamiltonian of the system is invariant with respect to rotations. The ground state wave-functions then transform among one another according to a 2L + 1-dimensional unitary representation of the rotation group. We assume that there is no accidental degeneracy. If we denote the ground state wave functions by fj i ]i = j i i i = 1 : : : 2L + 1g and the corresponding electron densities by i then the following convex combination
2X L+1 = 2L 1 1 i +
i=1
E v] = 2B inf
(r)v(r)dr + FEHK ]
(4.71)
(4.74)
is invariant under all rotations and therefore spherically symmetric. However the i are not spherical. In fact not any of the densities corresponding to linear combinations of the ground states j i i is spherically symmetric. As the j is obtained from j j i which by a unitary transformation can be obtained from any other j i i and the external potential is invariant under rotations we nd that
Z
i (r)v(r)dr =
j (r)v(r)dr =
(r)v(r)dr
(4.75)
for all 0 i j 2L + 1. Let us now suppose that is generated from a ground state wave function ], then this wave function is not a linear combination of the j i i otherwise would not be spherically symmetric. Then we nd Z Z ^ ^ ^ (r)v(r)dr + FHK ] = (r)v(r)dr + h ]jT + W j ]i = h ]jH j ]i
4.5. THE LEVY AND LIEB FUNCTIONALS FLL ] AND FL ]

1 h > i=1 2L + 1 =
2X L+1 2X L+1
73 1 Z (r)v(r)dr = i ]i + i i=1 2L + 1 (4.76)

2X L+1
^ i ]jH j
1 h i ]i = i=1 2L + 1
2X L+1
^ ^ i ]jT + W j
Z 1 F ] + (r)v(r)dr HK i i=1 2L + 1
This then gives 1 (4.77) 2L + 1 FHK i ] i=1 But we already knew that FHK was convex on the set of ground state densities which leads to a contradiction and therefore we must conclude that is not a ground state density of any potential. The density is however a convex combination of ground state densities corresponding to the same external potential and therefore by de nition an ensemble v-representable density. We therefore have constructed a E-V-density which is not a PS-V-density. Therefore AN is a real subset of BN and moreover AN is not convex.
FHK ] >
2X L+1
4.5 The Levy and Lieb functionals FLL and FL

]
The functionals FHK and FEHK have the unfortunate mathematical di culty that their domains of de nition AN and BN ,although they are well-de ned, are di cult to characterize, i.e. it is di cult to know if a given density belongs to AN or BN . Although there are reasons to assume that all reasonably well-behaved (that is, twice di erentiable, bounded and positive ) densities belong to BN this remains until now an unproven statement. It is therefore desirable to extend the domains of de nition of FHK and FEHK to an easily characterizable (preferably convex) set of densities. This can be achieved using the constrained search procedure introduced by Levy. We de ne the Levy-Lieb functional FLL as ^ ^ FLL ] = inf h jT + W j i (4.78)
!
where the in mum is searched over all normalized anti-symmetric N -particle wave functions in H 1(R3N ) yielding density . As shown earlier such a density is always in the convex set SN which is again a subspace of L1 \ L3 . One can furthermore show, as has been done by Lieb 19], that the in mum is always a minimum, i.e. there is always a minimizing wave function. Let us discuss some properties of FLL . The functional FLL is an extension of the Hohenberg-Kohn functional FHK which was de ned on AN to the larger set SN , i.e
FLL ] = FHK ] if 2 AN
Z
(4.79)
This is readily derived. Suppose is some ground state density corresponding to some external potential v and ground state wave function ] then (r)v(r)dr + FHK ] = h
Z
^ ]jH j
^ ]i = inf h jH j i = !
Z
^ ^ (r)v(r)dr + inf h jT + W j i = !
(r)v(r)dr + FLL ]
(4.80)
74

Ev ] =
Z
We de ne a corresponding energy functional (r)v(r)dr + FLL ] (4.81) If 0 is the ground state density corresponding to v with corresponding ground state wave function 0 ] then ^ ^ Ev ] = inf h jH j i h 0 ]jH j 0 ]i = Ev 0 ] (4.82) !
Minimizing Ev over the set SN therefore yields the ground state density 0 corresponding to external potential v. The functional FLL has however one inconvenient property, it is not convex. We take the example of the previous section where we presented a density which did not correspond to a ground state wavefunction. It was however a convex combination of degenerate ground state densities i with corresponding ground states j i ]i corresponding to an external potential v. Then we nd Z ^ (r)v(r)dr + FLL ] = inf h jH j i
!
L+1 1 2X h > 2L + 1 i=1
^ i ]jH j
N
Z L+1 1 2X F (r)v(r)dr i ]i = 2L + 1 LL i ] + i=1
(4.83)
and we nd
which proves the non-convexity of FLL . This is somewhat unfortunate as convexity is an important property which can be used to derive di erentiability of functionals. We will therefore now de ne a di erent but related functional with the same domain SN which is also convex. This is the Lieb functional FL de ned as ^ ^ ^ FL ] = înf TrD(T + W ) (4.85) where the in mum is searched over all N -particle density matrices ^ X i j ih j X i = 1 2 H 1 (R3N ) D=
i=1 i=1 D!
FLL ] > 2L1 1 FLL i ] + i=1

X
(4.84)
(4.86)
^ which yield the given density (r) = TrD ^(r). One can also for this case prove the in mum to be a minimum, i.e. there is a minimizing density matrix. This functional is an extension of FEHK to the larger set SN , that is FL ] = FEHK ] if 2 BN (4.87) This follows directly from the fact that if 2 BN then there is a potential v which generates a ^ ground state ensemble density matrix D ] which yields . So
Z
(r)v(r)dr + FEHK ] =
Z
^ ^ ^ ^ ^ (r)v(r)dr + TrD ](T + W ) = TrD ]H

Z
^^ = înf TrDH =
D!
^ ^ ^ (r)v(r)dr + înf TrD(T + W ) =

D!
(r)v(r)dr + FL ]
(4.88)
4.6. GATEAUX DIFFERENTIABILITY OF FL

We can again de ne an energy functional
75 (4.89)
Ev ] =
(r)v(r)dr + FL ]
which by a similar proof as for FLL assumes its minimum at the ground state density corresponding to potential v. We further have the following relations FL ] = FLL ] if 2 AN (4.90) and FL ] < FLL ] if 2 BN and 62 AN (4.91) The rst relation follows from the fact that is the density is a pure state v-representable density then the minimizing density matrix for FL is a pure state density matrix. The second relation follows from the fact that if is an ensemble v-representable density there is a ground state ensemble ^ density matrix D ] for which we have
Z
(r)v(r)dr + FL ] =
^ ^ ^ ^ (r)v(r)dr + TrD ](T + W ) = h i jH j i i
(4.92)
where j i h is any of the ground states in the degenerate ground state multiplet. Any wave function yielding density can not be a linear combination of these ground state wave-functions otherwise would be pure state v-representable. Therefore its expectation value with the Hamiltonian must be larger, i.e ^ ^ h i jH j i i < inf h jH j i = !
Z
(r)v(r)dr + FLL ]
(4.93) =
which proves our statement. We will now demonstrate another important property of FL , which is its convexity. If 1 1 + 2 2 then we have ^ ^ ^ ^ ^ ^ 1 FL 1 ] + 2 FL 2 ] = 1 ^ inf TrD1 (T + W ) + 2 ^ inf TrD2 (T + W ) =
D1 ! 1 D2 ! 2
^ ^ ^ ^ ^ ^ ^ = ^ înf Tr( 1 D1 + 2 D2 )(T + W ) înf TrD(T + W ) = FL ] D1 D2 ! 1 2 D!
(4.94)
We therefore now have established that FL is a convex functional on a convex space. This is important information which enables one to derive the G^teaux di erentiability of the functional a FL at the set BN of ensemble v-representable densities. We will discuss this feature of FL in the next section.
4.6 G^teaux di erentiability of FL a

We now turn to the question of the di erentiability of the functional FL . A convenient property of FL which it shares with the functional FLL is that it is de ned on a convex set SN . This means that whenever 0 2 SN that also 0 + t( ; 0 ) 2 SN when 0 t 1. In view of the de nition of the G^teaux derivative a lim FL 0 + t( ; 0 )] ; FL 0 ] = FL ( ) ; ] (4.95)
t!0
76
this is a convenient property. If the functional derivative FL = ( 0 ) exists then it is a linear functional on the space L1 \ L3 . These linear functionals can be identi ed with the Banach space of 3 potentials L 2 + L1 . The potential is then regarded as the bounded linear functional with assigns to each density the value of the integral of v. More mathematically formulated, the dual space of 3 L1 \ L3 ( which is the space of bounded linear functionals on L1 \ L3 ) is the space L 2 + L1 19]. If we use the convexity of FL we nd for 0 t 1 that FL 0 + t( ; 0)] = FL t + (1 ; t) 0 ] tFL ] + (1 ; t)FL 0 ] (4.96) and it follows that FL 0 + t( ; 0 )] ; FL 0] F ] ; F ] (4.97) If we assume G^teaux di erentiability at density 0 then we nd a FL ] ; FL 0 ] lim FL 0 + t( ;t 0 )] ; FL 0 ] = FL ( 0 ) ; 0 ] t#0
L 0
(4.98)
Conversely, from the results of section 3.5.2 , we know because of the convexity of FL that if there is a unique linear continuous functional (which is called a tangent functional or a subgradient) L : L1 \ L3 7! R such that FL ] ; FL 0 ] L ; 0 ] (4.99) is satis ed for all 2 SN then FL is G^teaux di erentiable at 0 with a
FL ( ) = L 0
(4.100)
To prove the di erentiability of FL at 0 it is therefore su cient to prove the existance of a unique continuous tangent functional at 0 . We will 3 prove the following statement. The functional rst 3 2 + L1 at 0 2 SN if there is a v 2 L 2 + L1 such FL has a continuous tangent functional ;v 2 L that Z E v] = 0 (r)v(r)dr + FL 0 ] (4.101) is satis ed. This is readily derived. Suppose the above relation is satis ed for some v. We will show ;v to be tangent functional. We have Z Z ^^ FL 0 ] + 0 (r)v(r)dr = E v] înf TrDH = FL ] + (r)v(r)dr (4.102) Therefore
D!
FL ] ; FL 0 ] ; v(r)( (r) ; 0 (r))dr = L ; 0 ]

Where we de ne the linear functional L : L1 \ L3 7! R by
(4.103) (4.104)
L : 7! ; v(r) (r)dr
3 2 This functional is continuous. If we split up v 2 L 2 + L1 as v = u + w where u 2 L 3 and w 2 L1 then we have Z Z jL ; 0 ]j ju(r)jj ; 0 jdr + jw(r)jj ; 0 jdr kuk 32 k ; 0 k3 + kwk1 k ; 0k1 (4.105)
4.6. GATEAUX DIFFERENTIABILITY OF FL
77
So if ! 0 in the norm on L1 \ L3 then jL ; 0 ]j goes to zero, which proves the continuity of L. We will now show that if ;v is a tangent functional then equation 4.101 is satis ed. If equation 4.103 is satis ed then
E v]
inf FL ] + 2S
N
Z
(r)v(r)dr
inf F ] + 2SN L 0
0 (r)v(r)dr =
= FL 0 ] +
0 (r)v(r)dr
E v]
(4.106)
which proves our statement. 3 We will now show that if 0 2 BN then FL has a unique tangent functional ;v 2 L 2 + L1 at 0 . (Unique means here that v is determined to within an overall constant). Therefore FL is G^teaux a di erentiable on the set of ensemble v-representable densities. From the previous results it follows that to prove this we must show that for every 0 2 BN there is a unique potential v (to within a constant) which satis es equation 4.101. If 0 2 BN then there is an external potential v and a ^ ground state ensemble density matrix D 0 ] yielding density 0 . Then obviously ^ ^ E v] = TrD 0 ]H =
Z
0 (r)v(r)dr + FL 0 ]
(4.107) (4.108)
X
We must now show its uniqueness. Suppose that we have a v 6= v + C which satis es ~
E v] = ~ E v] = ~
v 0 (r)~(r)dr + FL 0 ]
^ ^ ^ ~^ v 0 (r)~(r)dr + TrD 0 ](T + W ) = TrH D 0 ] = ~ i h i jH j i i
^ ^ ^ As we have FL 0 ] = TrD 0 ](T + W ) we can also write

Z
(4.109)
~ where H is the Hamiltonian corresponding to external potential v and j i i are ground states ~ ~ corresponding to potential v. However as the j i i are not ground states for H we obtain X ~ E v] = ~ ~ (4.110) i h i jH j i i > E v ]
i
which is a contradiction and therefore v must be unique. We therefore have proven the di erentiability of FL on the set of ensemble-v-representable densities. We can further prove that FL is not di erentiable at the non-E-V-densities. As proven by Lieb, for any density 2 SN there is a ^ density matrix D ] which minimizes the constrained search for FL . Let us write X ^ D ]= (4.111) i j i ih j Suppose that is not E-V-representable, i.e. 62 BN . Since is not ensemble v-representable there is at least one j i i which cannot be a ground state for a potential v. Therefore
i
E v] <
^ i h i jH j i i =
(r)v(r)dr + FL ]
(4.112)
It follows that equation 4.101 can never be satis ed for non-E-V-densities, which yields that FL is not di erentiable at the non-E-V-densities.
78
Let us summarize the results of this section. The functional FL is di erentiable at the ensemble v-representable densities and nowhere else. This means that the Euler-Lagrange equation is well-de ned on the set BN . Here is a Lagrange multiplier to ensure the correct normalization of the density. The energy functional is minimized for the ground state ensemble densities corresponding to external potential v. As the Euler-Lagrange equation is well-de ned for these densities, they are found as the solution of the above Euler-Lagrange equation. This then puts the variational equations within density functional theory on rigorous grounds.
FL + v(r) ; = 0 (r)
(4.113)
4.7 The Kohn-Sham equations

Very important in any practical application of density functional theory are the Kohn-Sham equations. The Kohn-Sham system is introduced as the non-interacting system of particles which yield the same density as the fully interacting system of particles. If we consider a system non-interacting ^ particles then W = 0. For this system our functionals FHK FEHK FLL and FL then only contain the kinetic energy operator. We therefore denote these functionals as THK TEHK TLL and TL and de ne ^ THK ] = h ]jT j ]i (4.114) ^ ^ TEHK ] = TrD ]T ^ TLL ] = inf h jT j i ! ^^ TL ] = înf TrDT
D!
(4.115) (4.116) (4.117)
The functionals TLL and TL are de ned on the same set SN as FLL and FL for the interacting 0 system. The functionals THK and TEHK are now de ned on the sets A0N and BN of PS-V and E-V densities for the non-interacting system. The set A0N is not convex, and convexity is unknown for 0 the set BN . The various convexity or non-convexity properties of the functionals FHK FEHK FLL and FL carry directly over to the functionals THK TEHK TLL and TL . For TL we can similarly to FL prove that it is Gateaux di erentiable on the set of non-interacting ensemble v-representable 0 densities BN . Let us now de ne the exchange-correlation functional Exc L 1 Z (r1 ) (r2 ) dr dr ; T ] Exc L ] = FL ] ; 2 jr ; r j 1 2 L (4.118) 1 2 The functional Exc L is de ned on the set SN . As FL is di erentiable on BN and TL is di erentiable 0 on BN the exchange-correlation functional Exc L is di erentiable on the intersection of both sets 0 BN \ BN . 0 Suppose now that an external potential in a non-interacting system vs generates density 2 BN then the variational equation corresponding to the energy functional of the non-interacting system
Evs ] = TL ] +
(r)vs (r)dr
(4.119)
4.7. THE KOHN-SHAM EQUATIONS

is given by
79
where s is a Lagrange-multiplier to ensure the correct normalization of the density. As is an enesemble v-representable ground state we also have 1 ; 2 r2 + vs(r) i(r) = i i (r) (4.121) X 2 ^ (r) = TrD ]^(r) = (4.122) i j i (r)j
i
TL + v (r) ; (r) s
s=0
(4.120)
On the other hand if is also in BN and is generated by an external potential v then is obtained from the solution of the Euler-Lagrange equation
0 with again a Lagrange multiplier for the density normalization. So if 2 BN \ BN we nd TL + v(r) + Z (r0 ) dr0 + Exc L ; = 0 (4.124) (r) jr ; r0j (r) If we denote vxc L(r) = Exc L = (r) we obtain (to within a constant) Z (r0 ) vs (r) = v(r) + jr ; r0 j dr0 + vxc L(r) (4.125)
FL + v(r) ; = 0 (r)
(4.123)
We therefore can nd the ground state density for the interacting system by the solution of the equations Z (r0 ) 1 (4.126) ; 2 r2 + v(r) + jr ; r0 j dr0 + vxc L(r) i (r) = i i(r)
X ^ (r) = TrD ]^(r) = i j i (r)j2
(4.127)
which constitute the Kohn-Sham equations. Note however that the above equations are only 0 valid for densities in BN \ BN . Englisch and Englisch have however proved that for interacting ensemble v-representable densities the above Kohn-Sham equations always have a solution, which 0 means BN = BN . It further means that TL is di erentiable for all densities which are ensemble v-representable with respect to the interacting system. This then puts the Kohn-Sham equations on rigorous grounds.
80
Chapter 5
An exchange-correlation potential with correct asymptotic behaviour

In this work we analyse the properties of the exchange-correlation potential in the Kohn-Sham form of density functional theory, which leads to requirements for approximate potentials. Ful lment of these requirements is checked for existing gradient-corrected potentials. In order to examine the behaviour of approximate potentials over all space we compare these potentials with exact KohnSham potentials calculated from correlated densities using a newly developed iterative procedure. The main failures in the existing gradient corrected potentials arise in the asymptotic region of the atom where these potentials decay too fast and at the atomic nucleus where the potentials exhibit a Coulomb-like singular behaviour. We show that the main errors can be corrected by a simple potential in terms of the density and its gradients leading to considerably improved one-electron energies compared to the local density approximation. For Be and Ne it is shown that the electron density is improved in the outer region.
5.1 Introduction
In the past few years there has been considerable progress within density functional theory 5] in the calculation of properties of electronic systems ranging from solids to atoms and molecules. This progress is due to the introduction of gradient-corrected density functionals 44, 45, 43, 50] that give an overall improvement to the exchange-correlation energies of the local density approximation (LDA). For instance the atomisation energies of a standard set of molecules are improved by an order of magnitude compared to LDA 70], thereby correcting the overbinding of LDA. There are also succesful applications of nonlocal corrections in transition metal chemistry 71] and solid state and surface physics 47]. (Although the gradient corrected potentials are still local, we follow the conventional nomenclature of "nonlocal corrections" to distinguish from the LDA.) However several other features are not improved by the present day nonlocal corrections. This is especially the case for properties that are sensitively dependent on the behaviour of the exchange-correlation potential. One can for instance prove rigorously 31, 72] that the eigenvalue of the highest occupied KohnSham orbital represents the ionisation energy of the system. However, typical errors in LDA for this quantity are 5 eV. This same error prevents the calculations of bound state solutions for negative ions as LDA gives an unbound outer electron with positive eigenvalue. The origin of this error can be traced to the incorrect asymptotic decay of the LDA exchange-correlation potential. This potential has an exponential decay into the vacuum as can directly be seen from the expontial decay 81
82
AN EXCHANGE-CORRELATION POTENTIAL
of the density itself. On physical grounds however (see for a proof 31]) the outer electron should experience the mean eld of the ion it leaves behind, i.e. a potential that decays Coulombically like ;1=r. The LDA electron is therefore too weakly bound and for negative ions even unbound. This breakdown of LDA in the outer region of the atom or molecule is also re ected in the exchangecorrelation energy per particle xc(r) which can equivalently be seen als the potential due to the exchange-correlation hole and which has an asymptotic decay like ;1=2r. The LDA in this case gives again an exponential decay. In the electron gas this quantity xc is usually expressed in terms of the Wigner-Seitz radius rs representing the mean electronic distance which is proportional to 1 ; 3 . If the local density approximation is applied to the outer regions of atoms and molecules rs grows exponentially and loses its meaning as a mean interelectronic distance which should grow linearly. If one believes (for intuitive physical reasons) that the mean electronic distance determines xc then the bad representation of this quantity by LDA explains the failure of LDA in this region. One might now wonder if the present day nonlocal gradient corrections give any improvement for this asymptotic failure of LDA. Somewhat surprisingly, this is not the case. Although they give large improvements in energies they give little improvement in the asymptotic behaviour of the exchange-correlation potential. This is immediately apparent from the fact, undoubtedly noted by many DFT practitioners, that the gradient corrected potentials yield almost no improvement in the LDA eigenvalues, which are generally in error by 5-6 eV. In this paper we discuss, apart from the fairly well known asymptotic behaviour, other requirements which are to be ful lled by the exact exchange or exchange-correlation potential. Those requirements apply to limiting or special situations (r ! 1, r # 0, transition regions from one atomic shell to the next, limit of homogeneous electron gas) and also comprise invariance conditions (translational, rotational). They are useful in constructing approximate Kohn-Sham potentials. In section 5.3 we use the formulated requirements to examine some of the presently used nonlocal functionals, in particular the Becke correction for exchange and the Perdew correction for correlation. In order to investigate the potentials at arbitrary ~, we need the exact Kohn-Sham potential over all space. In section 5.4 r we discuss a general procedure to construct the Kohn-Sham exchange-correlation potential from a given electron density . The procedure is simple and is applicable to both atoms and molecules and to systems with an arbitrary number of electrons. This procedure is used to generate exact Kohn-Sham potentials from very accurate (highly correlated) densities of Be and Ne. This a ords detailed insight in the strengths and weaknesses of existing approximations. As a rst step towards improved potentials, we propose in section 5.5 a model Kohn-Sham potential which exhibits the correct asymptotic behaviour and also displays atomic shell structure. This potential gives a large improvement over the LDA eigenvalues and is in fact capable of yielding good ionisation energies as determined from the highest occupied Kohn-Sham orbital energy. It also improves the asymptotic decay of the electron density. In section 5.6 we present a summary and conclusions.
5.2 Nonlocal exchange-correlation potentials : requirements

In this section we will discuss some of the requirements that approximate exchange-correlation potentials have to satisfy. In connection with this we shall demonstrate in the next section which of these requirements are lacking with some of the presently used nonlocal potentials. Scaling. The exchange energy within density functional theory may be de ned as 23]:
Ex ] =
(r) x (r)dr
(5.1)
NONLOCAL EXCHANGE-CORRELATION POTENTIALS:REQUIREMENTS
83
in which the the potential of the exchange-hole x (r) or equivalently the exchange energy density per electron is de ned as: 1 Z j s (r r0 )j2 dr0 (5.2) x (r) = ; 2 (r) jr ; r0 j Here s is the Kohn-Sham one-particle density matrix, constructed from a determinant of KohnSham orbitals. The exchange functional of eq. 5.1 is a functional of the density as the Kohn-Sham orbitals of which the one-particle density matrix s is composed are uniquely determined by the density (an explicit scheme for doing this is presented in section 5.4). The exchange functional satis es the following scaling relation 22]:
Ex ] = Ex ]
in which is the following scaled density: (r) = 3 ( r)
(5.3) (5.4)
The exchange potential which is de ned as the functional derivative of the exchange functional satis es the following scaling relation 23, 22]:
vx ( ] r) = vx( ] r)
Using the scaling relation 5.3 one can prove the Levy-Perdew relation 22]:
(5.5) (5.6)
Ex ] = ;
(r)r rvx (r)dr
Any approximate exchange potential should satisfy equation 5.5. It is then possible to de ne an approximate exchange energy using the Levy-Perdew relation 5.6. We will return to this relation between potential and energy below. There are no known scaling relations for the correlation energy functional de ned as:
the coupling strength integration method 73, 24, 25, 26]
Ec ] = Exc ] ; Ex ] (5.7) in which Exc is the Kohn-Sham exchange-correlation energy which can be de ned for instance using
Asymptotic behaviour. An approximate functional or potential also has to satisfy some requirements with respect to asymptotic behaviour. First of all it follows from equation 5.2 using the sum rule property of the one-particle density matrix s that: 1 (r ! 1) (5.8) x (r) ; 2r Thus the potential of the exchange hole has a Coulombic asymptotic behaviour. The exchange potential has to ful l a similar type of relation 35, 49]: vx (r) ; 1 (r ! 1) (5.9)
The asymptotic behaviour of the potential due to the (coupling strength integrated) Coulomb hole is not known. We can only say that the correlation potential must decay faster than a Coulombic
84
potential at in nity. This follows from the asymptotic ;1=r decay of the total Kohn-Sham potential 74, 72]: v (r) ; 1 (r ! 1) (5.10)
xc
and the similar behaviour of the exchange potential of eq. 5.9. This result is consistent with the fact that the Coulomb hole of the coupling strength integrated two-particle density matrix integrates to zero electrons.
Weak inhomogeneity. Further known properties of the exchange functional follow from the gradient expansion of the weakly varying electron gas 10] or from the semi-classical expansion of the Kohn-Sham one-particle density matrix 75]. This gives the following approximate nonlocal correction to the exchange energy (for the spin unpolarised case): Z 2 (5.11) E ] = ; (r ) dr
x
4 3
This is the nonlocal correction used in the X -approximation 76]. For weakly varying densities there is a known gradient expansion for the correlation energy. The nonlocal correction is (up to second order) given by: Z 2 E ] = C ( ) (r ) dr (5.12)
c
4 3
in which C ( ) is a local function of the electron density 42].

Translational and rotational invariance. Two requirements which must be ful lled by any density functional representing a physical quantity are translational and rotational invariance. The question of translational invariance for instance arises naturally in the discussion of the asymptotic ;1=r behaviour of the energy densities and the exchange or exchange-correlation potentials, where r should not refer to the distance from the arbitrary origin of the coordinate system. Denoting a physical quantity by A, translational invariance means: A 0] = A ] (5.13) with 0 (r) = (r + R) ~ (5.14) ~ in which R is an arbitrary translation vector. This means physically that A should not change when we translate our coordinate system. If we de ne:
then it follows that: ~ a( 0 ] r) = a( ] r + R) (5.16) This does not hold for arbitrary a( ] r) but it does hold if a is the functional derivative of a translational invariant functional A. This equation is easily proved if one uses:
a( ] r) = A ) ] (r
(5.15)
A ]=
(r)a( ] r)dr =
~ ~ (r + R)a( ] r + R)dr =
~ (r)a( ] r + R)dr
(5.17)
NONLOCAL EXCHANGE-CORRELATION POTENTIALS:REQUIREMENTS

Comparing this equation with:
85 (5.18)
A 0] =
0 (r)a( 0 ] r)dr
~ yields equation 5.16 as both variations should be equal for any variation 0 (r) = (r + R). We will now take A = Ex and a = vx . Noting the explicit r dependence in the Levy-Perdew relation between Ex and vx one might wonder whether this relation is translationally invariant. The exact exchange potential of course satis es the translational invariance equation 5.16. Then it follows from the Levy-Perdew relation 5.6, if we insert 0 , that: ~ Ex 0 ] = Ex ] + R
Z Z
(r)rvx ( ] r)dr
(5.19)
Translational invariance requires the last term in this equation to be zero. As this should be true ~ for any vector R we obtain (after carrying out a partial integration):
vx (r)r (r)dr = 0
Z
(5.20)
Translational invariance thus gives an additional condition on the potential. The above formula is a special case of a more general result which follows directly from equation 5.13:
~ A ] = A 0] = A + R
Z
+ O(R2 )] = A ] +
~ As this equation should be valid for any translation vector R it follows that: A r (r)dr = 0 (r)
A R r (r)dr + O(R2 ) ~ (r)
(5.21) (5.22)
This equation has also been noted without proof in reference 23]. Equation 5.20 is then obtained by taking A = Ex in the last equation. In particular for A = Exc we obtain:
Z
vxc (r)r (r)dr = 0
(5.23)
For the case of rotational invariance we require equation 5.13 to be valid for: 0 (r) = (Rr) (5.24) in which R is a rotation operation within three-dimensional coordinate space. For the functional derivative of A we then have: a( 0 ] r) = a( ] Rr) (5.25) This equation is proven in the same way as equation 5.16. If we take a = vx and use equation 5.25 we see that the Levy-Perdew equation 5.6 is already rotationally invariant. In general rotational invariance gives the following constraint on the functional derivative of A:
Z
which can be proved by performing an in nitisimal rotation instead of a translation in equation 5.21. This gives :
Z
A r r (r)dr = 0 (r)
(5.26)
vxc (r)r r (r)dr = 0
(5.27)
86
We have now summarised some properties of the exact exchange-correlation potential, which are at the same time requirements to be ful lled by approximate potentials. We will concentrate in the remainder of this paper on the potential rather than on the energy. One of the reasons is that the potential, being a unique, local function of the position ~, is more easily obtained, analysed r and modelled (see below). Of course, even if one has obtained a good model potential that very closely approximates the exact Kohn-Sham potential, the exchange-correlation energy still has to be determined. This problem will be adressed in a subsequent paper.
5.3 The nonlocal exchange and correlation potentials of Becke and Perdew
Using the conditions formulated in the previous section we shall now discuss some of the currently used exchange-correlation functionals and potentials that include nonlocal corrections. These are the potentials derived from Becke's nonlocal exchange functional 44] and Perdews nonlocal correlation functional 45]. First we discuss Becke's nonlocal exchange functional 44] which in spin polarised form is given by:
Ex ] =
4 3( 4 3
r)f (x (r))dr
(5.28)
in which x = jr j= by:
is a dimensionless quantity and is a spin index. The function f is given
(5.29) 1 + 6 x sinh;1 (x) The form of equation 5.28 is chosen such that the exchange functional satis es the scaling relation 5.3. The function f is chosen such that the potential of the exchange hole x (r) or equivalently the exchange energy density per electron behaves asymptotically as;1=2r (eq. 5.8). To enforce relation 5.8 the function f in equation 5.28 must satisfy the following asymptotic relation: 1 f (x) ; 6 logx(x) (r ! 1 ) x ! 1) (5.30) which is easily veri ed if exponentially decaying densities are inserted. If one also wants to obtain the gradient expansion result of equation 5.11 for slowly varying densities f must satisfy:
f (x) = ;
x2
f (x) ; x2
(x # 0)
(5.31)
The form chosen by Becke is one of the simplest interpolations of f between these two limits and therefore his energy functional satis es the important requirements of correct r ! 1 and x ! 0 behaviour as well as translational and rotational invariance. However, the potential of the Becke energy expression decays asymptotically like 77, 49]:
Becke vx (r) rk2 (r ! 1) in which k is some constant instead of the exact 35, 49]:
(5.32) (5.33)
vx (r) ; 1 r
THE NONLOCAL EXCHANGE AND CORRELATION POTENTIALS..
87
In the following we will prove that any functional of the form of equation 5.28 satisfying the asymptotic relation 5.30 will not satisfy asymptotic relation 5.33. Such a proof has also been given by Engel et al. 49] and is only presented here for reasons of clarity and completeness. We will however also use the following short proof to make some statements about the behaviour of the exchange potential near the atomic nucleus for exchange functionals of the form of equation 5.28 For simplicity we use the spin unpolarised form but everything goes exactly through for the spin polarised case by just adding a spin index everywhere in the derivation. This is due to the fact that the exchange energy in the spin polarised case is just a simple sum of contributions of and spins in which both contributions have the same structure. We will take the functional of the form:
Ex ] =
with
4 3(
r)f (x(r))dr
(5.34) (5.35)
x = jr4 j
3
The second derivative of the function f is assumed to exist. The functional derivative of Ex of 5.34 is then given by: Ex = v (r) = 4 1 (f ; x df + x2 d2 f ) 3 (r) x 3 dx dx2 +
X
ij
@i @i@j @j ( df ; x d2 f ) ; r2 df jr j3 dx dx2 jr j dx
(5.36)
We insert for the density the exponential (r) = Ne; r . This exponential is exact in two regions of the atom: near the atomic nucleusp where we have = 2Z with Z the nuclear charge and for the outer asymptotic region where = 2 ;2 with the chemical potential (negative of the ionisation energy). This gives: 2 2 df df 3 vx (r) = 4 1 (f ; x dx + x2 d f ) ; x d f + 2 dx (5.37) 2 2 r 3 dx dx In order to satisfy the asymptotic behaviour of x (equation 5.8) f must satisfy the asymptotic relation 5.30. If we now insert the large x behaviour of f (equation 5.30) into equation 5.37 we get retaining only the terms of lowest order in 1= log x: 1 (5.38) vx (r) ; 31r log x ; 31r 1 ; 1 ; 1 r12 (r ! 1) log 3 We thus see that the exchange potential has a ; ;1 =r2 behaviour instead of a Coulombic behaviour. Hence it follows that the requirements of correct asymptotic behaviour of x (relation 5.8) and of vx (relation 5.33) are incompatible for exchange functionals of the form 5.34. From equation 5.37 we can also draw the following conclusion concerning the behaviour for r ! 0: df If dx jx=x(~ ) 6= 0 then: 0
vx (r) k r
(r # 0)
(5.39)
88 in which the constant k is given by:
df k = 2 dx jx=x(~ ) 0
(5.40)
This is an unphysical behaviour of the exchange potential near the nucleus. In practice however (that is for the Becke functional) the constant k is very small compared to the nuclear charge (a typical value for k is 0.02 for the noble gases). We will now discuss some properties of the Perdew nonlocal correlation functional which is of the form 45]:
Ec ] = f (
)e;g( )jr j (r 4 ) dr
3
(5.41)
In this formula f is a local function of and which are spin densities and g is a local function of the total density. The form of this functional is adapted from the correlation part of the LangrethMehl functional in such a way that the functional reduces to the gradient expansion expression of equation 5.12 for slowly varying densities. The corresponding potential decays exponentially to in nity. This is not a bad feature as the correlation potential should decay faster than Coulombic although maybe one should expect a correlation potential decaying like ; =r4 in which is the polarisability of the system. This term occurs in the exact exchange-correlation potential 74] and describes the polarisation of the system by an asymptotic electron, which is clearly a correlation e ect. Near the nucleus the Perdew potential also has a Coulombic singular behaviour. The origin of this unphysical singularity can be traced, like in the Becke potential, to terms in the potential that contain the Laplacian of the density, which due to the Slater-like behaviour of the atomic density near the nuclues leads to a Coulombic potential. These terms also occur in potentials of other nonlocal density functionals such as the one from Langreth and Mehl 50] and in the Generalized Gradient Approximation (GGA) of Perdew and Wang 46]. These functionals consequently also su er from this same de ciency.
5.4 Constructing the potential from the density

We have demonstrated that the most widely used gradient corrected exchange and correlation potentials su er from incorrect asymptotic behaviour at r ! 0 and r ! 1. In order to investigate these potentials over the whole r range one needs the exact Kohn-Sham potential at arbitrary position. In this section we present an iterative scheme of obtaining the Kohn-Sham potential from a given electron density . This scheme is di erent from most other schemes used thus far 78, 74, 79] and in particular is not limited to two-electron systems but is applicable to systems with any number of electrons. Recently during the writing of this paper we have noticed that work along similar lines has been carried out 80]. However our procedure is somewhat simpler and as we tested has the same convergence rate. In contrast to reference 80] in which calculations within a basis set were performed we use a completely numerical approach obtaining a basis set free representation of the potential. We start from the Kohn-Sham-equations: 1 (; 2 r2 + vs (r)) i (r) = i i (r) (5.42)
CONSTRUCTING THE POTENTIAL FROM THE DENSITY

in which vs is the Kohn-Sham-potential. The Kohn-Sham-orbitals are required to satisfy:
X
89
N i
j i (r)j2 = (r)
(5.43)
where N is the number of electrons in the system. Multiplying equation 5.42 by i and summing over i gives after dividing by :
N X vs (r) = (1r) 1 i (r)r2 i (r) + ij i (r)j2 i 2
(5.44)
We now de ne an iterative scheme using this equation. We want to calculate the potential corresponding to the density . Suppose that at some stage in the iteration we have calculated orbitals o with eigenvalues o and density o and potential vo . In the next step we de ne the new potential: i i
N X vn (r) = (1r) 1 i 2 o = ((rr)) vo (r) o i
(r)r2 o (r) + o j o (r)j2 i i i (5.45)
Using this potential we calculate new orbitals and a new density and de ne in the same way a new potential. This procedure is continued until the density calculated from the orbitals is the same as the given density. In practice until: o max j1 ; ((rr)) j < (5.46) r with a given threshold. To achieve convergence one should take care to keep the prefactor in the last term of equation 5.45 in each iteration within an acceptable range: o 1 ; < ((rr)) < 1 + (5.47) for example with = 0:05. We noticed in the application of this procedure that one has to make sure that the potential is set to zero in in nity otherwise one might fail to converge.This is due to the fact that one can add an arbitrary constant to both the potential and the one electron eigenvalues without changing the density. This may also play a role in the practical application of the procedure of reference 80] where explicit use is made of the one electron eigenvalues in the potential construction procedure. The xing of the potential in in nity is however easily carried out in our program in which di erential equations are solved numerically. In that case boundary conditions at in nity and at the nucleus immediately x the solution. The scheme is not guaranteed to converge as there are densities which are not non-interacting v-representable (however many densities which are not v-representable by one determinant are v-representable by a linear combination of determinants 81] which can be accomplished in our scheme by using fractional occupation numbers). However if the procedure converges then its limit is unique as guaranteed by the Hohenberg-Kohn theorem applied to a non-interacting electron system 10].
90
We have calculated exact potentials for the beryllium and the neon atoms as for these atoms accurate densities from con guration interaction calculations are available. Both densities have been published by Bunge and Esquivel 82, 83] in a large basis of Slater functions. The use of Slater functions is advantageous because they give a much better representation of the density tail in the outer regions of the atom than for instance Gaussian functions. For the calculation of the potentials from these densities we used a modi cation of the Herman-Skilman atomic program 84] based on numerical integration. For the results of table 1 we have used our atomic and molecular density functional package based on Slater Type Orbitals (STO's) with which it is possible to carry out self-consistent calculations using the Becke-Perdew potential. In the local density approximation we use for the correlation potential the VWN-parametrisation of the electron gas data 85]. For the open shell atoms discussed in this article we performed spherically averaged spinpolarised calculations. The exact exchange-correlation potentials are displayed in gure 5.1a and gure 5.2a for Be and Ne respectively and r times these potentials in gs.5.1c and 5.2c. Both potentials have the same structure, a characteristic peak between the atomic shells (in our case between the K and the L-shell) and a Coulombic asymptotic behaviour. These features are most clearly displayed in the plots of rV . The appearance of the intershell peak has been observed before 74, 79] and can be understood from the work of Buijse et al. 86]. In ref. 86] it has been observed that an important contribution to the Kohn-Sham potential is the socalled kinetic potential Vkin de ned in terms of the conditional amplitude : (x (x2 : : : xN jx1 ) = p 1 : : : xN ) (5.48) (x1 )=N
Vkin (x1 ) =
Z (; 1 r2 ) dx2 : : : dxN = + 1 jr1 j2 dx2 : : : dxN 2 1 2
(5.49)
The conditional amplitude (x2 : : : xN jx1 ) describes the system of N ; 1 electrons with positions x2 : : : xN when one electron is known to be at position x1 and is the amplitude connected with the conditional probability of nding the other electrons when one electron is known to be at x1 . Vkin makes a signi cant (positive) contribution to the Kohn-Sham potential at those positions x1
of the reference electron where the conditional amplitude changes rapidly, so that r1 is large. As discussed in ref. 86], this is the case when x1 crosses the border region between two atomic shells, since the exchange hole is localized within one atomic shell if the reference position is in that shell but "jumps" to the next shell when the reference position crosses the border (see refs. 87, 88, 89]). The intershell peak of Vxc is therefore an exchange e ect and re ects the fact that the atomic shell structure has its origin in the antisymmetry principle. The peak in the potential helps to build the shell structure in the total density by reducing the density in the intershell region. It is obvious from the gures that the LDA potentials almost completely lack this important feature of the exact potential. This is particularly clear from the pronounced appearence of peaks in the di erence plots of (VXC ; VLDA ) and r(VXC ; VLDA ) in gs. 5.1b,d and 5.2b,d. The LDA potential also has a wrong asymptotic behaviour for r ! 1, as is evident from the fact that rVLDA does not go to ;1 but to 0. The exact potential has a much improved asymptotic behaviour. However, the quality of the exact potential we generate depends on the quality of the correlated wavefunction and density on which it is based. The asymptotic region is notoriously di cult to describe accurately, since wavefunctions are almost invariably obtained from energy optimization algorithms which have a strong bias towards improving the energetically important inner region. In the case of Neon we observe that for large r (r 3 bohr) there is a spurious minimum in the curve for the exact VXC ( g.5.2c) which we ascribe to inaccuracy of the correlated density at such large r. It is nevertheless
CONSTRUCTING THE POTENTIAL FROM THE DENSITY
91
Figure 5.1: Exchange-correlation potentials of the beryllium atom.
92
Figure 5.2: Exchange-correlation potentials of the neon atom.
5.5. A MODEL POTENTIAL
93
clear that the "exact" rVXC approaches ;1 much better than the LDA potential does. For Be it is not evident that the calculated "exact" VXC su ers from poor accuracy of the asymptotic behaviour of the CI density, but we do feel that some suspicion is warranted concerning the (too?) slow approach by rVXC to the limiting value of ;1 ( g.5.1c). This suspicion is aggravated by the strange minimum between 6 and 8 bohr in the curve of r(VXC ; VLDA ) in g.5.1d. In panels a and c of gures 5.1 and 5.2 we have also added Beckes gradient-corrected exchange potential 44] and Perdews correlation potential 45] to the LDA potential. The plots demonstrate that these potentials make rather small corrections to the LDA potential. This makes it understandable that the one-electron energies shown in table 1, which are rather poor for LDA when compared to the experimental ionisation energies, do not improve when the Becke-Perdew non-local corrections are included in the SCF calculation. Panels b and d of the gures display directly the comparison between the non-local corrections to the potential and the di erence (VXC ; VLDA ) to which they should be equal. It can be seen that the nonlocal Becke-Perdew potential behaves singularly near the atomic nucleus. This potential has in this region the Coulombic singular behaviour discussed earlier. At somewhat larger distances it crosses the horizontal axis and gives a small positive peak which is located at the right spot in the intershell region. For Be this does not lead to good agreement with (VXC ; VLDA ) in the intershell region, but for Ne VBECKE ;PERDEW quite nicely approximates (VXC ; VLDA ) just at the position of the intershell peak. At larger distances the Becke-Perdew potential is almost zero and there is no correction to the LDA potential, in agreement with the failure to give improvement of the LDA eigenvalues. The large deviation in the asymptotic region between the Becke-Perdew potential and the exact nonlocal potential is most clearly exhibited in panels d: the exact nonlocal corrections correctly tend to a constant which approximates ;1, whereas the Becke-Perdew potential tends to zero. The panels d also demonstrate that the Becke-Perdew potential multiplied by the radial distance does not tend to zero at the atomic nucleus but to a nite value, indicating the Coulombic behaviour of this potential in this region.
5.5 A model potential

In this section we will make a rst step towards the construction of model potentials that more closely resemble the exact Kohn-Sham potential. The advantage in modeling potentials instead of energy expressions is the fact that potentials are uniquely de ned by the exact density. The quality of model potentials can therefore be judged by comparing to exact potentials calculated from accurate densities. In modeling the potentials one should incorporate the general features of atomic shell structure and asymptotic Coulombic behaviour. One might also wish the potential to satisfy some scaling properties. To incorporate the shell structure we use for our model potential the dimensionless 4 parameter x = jr j= 3 . This parameter is proportional to the length of the gradient of the local Wigner-Seitz radius and can be interpreted as the change in mean electronic distance (at least in regions with slowly varying density where rs ;1=3 is meaningful, i.e. in the regions where LDA is applicable). As an illustration of the behaviour of this parameter we display both this parameter and r2 (r) in gure 5.3 for the krypton atom. Whereas the electron density (r) is monotically decreasing in an atom, r2 (r) exhibits the shell structure. We observe that x also oscillates with the atomic shells, with maxima approximately at the in ection points of r2 (r), and thus seems a suitable parameter to model the shell structure of the exact potential. For the asymptotic behaviour of the potential we then have to take into account that x behaves for an exponentially decaying
94
Figure 5.3: The electron density times the radial distance squared and the parameter x = jr j= for the krypton atom.
4 3
density e; r asymptotically as ;1=3 , i.e. increases exponentially (see g. 5.3). We thus choose our nonlocal correction to the LDA potential of the form
1 vxc (r) = 3 (r)f (x(r))
(5.50) (5.51)
This form of the potential scales like an exchange potential 22, 23]:
vxc ( ] r) = vxc( ] r)
We do not know the scaling behaviour of the correlation part of the potential but as the major part of the potential comes from the exchange we take the above form as an approximation. For systems with small density variations (small values of x) we want the nonlocal correction potential also to be small. We therefore require f (0) = 0. We further know that 31, 72]: v (r) ; 1 (r ! 1) (5.52)
r This means that f must asymptotically satisfy: 1 (x ! 1) f (x) ; 3 logx(x) x2
xc
(5.53)
Our problem of nding a smooth interpolation between these limiting situations is similar to the one that Becke faced for the exchange energy density. Inspired by his solution we take:
f (x) = ;
1 + 3 x sinh;1 (x)
(5.54)
A MODEL POTENTIAL
ATOM H He Be Ne Ar Kr Xe ION F; Cl; Br; I; MOLEC N2 F2 CO HF 0.622 0.551 HF 0.500 0.918 0.309 0.850 0.591 0.524 0.457 HF LDA 0.234 0.571 0.206 0.490 0.381 0.346 0.310 LDA NL(BP) MODEL EXPT 0.280 0.585 0.209 0.496 0.380 0.344 0.308 0.440 0.851 0.321 0.788 0.577 0.529 0.474 0.128 0.140 0.140 0.139 0.557 0.607 0.529 0.500 0.903 0.343 0.792 0.579 0.517 0.446 0.125 0.133 0.124 0.112 0.573 0.582 0.515
95
NL(BP) MODEL EXPT
-0.097 -0.099 -0.022 -0.023 -0.008 -0.009 +0.005 +0.004 LDA 0.328 0.339 0.334 0.322 0.334 0.336
NL(BP) MODEL EXPT
Table 5.1: ionisation energies and electron a nities from the highest occupied Kohn-Sham orbital This gives the nal form of our model potential for the nonlocal corrections. In spin polarised form:
vxc (r) = ;
1 3(
In order to check if with this potential we have captured the main features of the Kohn-Sham potential which the LDA potential is lacking, i.e. shell-structure and a Coulombic asymptotic behaviour, we compare to exact potentials in gs. 5.1 and 5.2. The parameter in our model potential was tted in such a way that our model potential resembled as closely as possible the di erence between the exact and the LDA potential for the beryllium atom. For this procedure we choose the beryllium atom instead of the neon atom because its density appears to be the most accurate of the two. This leads to a value of of 0:05. The model potential for neon in gure 5.2 uses the same parameter . The model potentials in gures 5.1 and 5.2 have been calculated selfconsistently using our density functional program package. As can be seen from gure 5.1 for the beryllium atom our potential is in reasonable agreement with the exact potential. In particular gs. 5.1c,d show that the intershell peak is fairly well represented and the asymptotic behaviour is essentially correct in that r(VLDA + VMODEL) approaches ;1. Concerning the remaining di erence, there is obviously room for improvement, although we suspect that part of the di erence between our model potential and the exact potential in the asymptotic region is due to the possible inaccuracy in the latter mentioned before. The model potential does clearly improve upon the Becke-Perdew potential. For the neon atom ( gure 5.2) the asymptotic behaviour of the model potential appears to be quite good (assuming that the minimum of the "exact" potential between 5 and 6 bohr is incorrect). However, in the inner region of the neon
2 r) 1 + 3 x x ; 1 ( x ) sinh
(5.55)
96
atom the model potential is not so well behaved. Although it does exhibit an intershell peak, the approximation of this feature to the true intershell peak is rather poor. As a matter of fact the Becke-Perdew potential gives a better approximation to the exact nonlocal potential precisely at the intershell peak. Nevertheless, considering the whole r range we can say that for both the beryllium and the neon atom our simple model potential gives a considerable improvement of the LDA potential, especially in the asymptotic region were the Becke-Perdew potential gives almost no correction. We might now ask whether these improvements are re ected in the quality of the eigenvalues. In order to investigate this question we calculated for several atoms and molecules the eigenvalue of the highest occupied Kohn-Sham orbital. This eigenvalue should be equal to the ionisation energy of the system (or electron a nity for negative ions). We have done this for both LDA, LDA with the Becke-Perdew potential added and for LDA with our model potential ( = 0:05) added. The results for H and a number of atoms with noble gas con gurations are presented in table 5.1. In this table we also compare with the self-interaction free eigenvalues of the HartreeFock approximation. From the table we can see that the LDA eigenvalues have a large discrepancy with experiment, with a mean absolute error of 5.40 eV. We can also see that these values are not improved by inclusion of the Becke-Perdew potential. The model potential on the other hand gives a considerable improvement compared to the LDA eigenvalues, with a mean absolute error of 0.56 eV. A nice feature of the model potential is also that it yields bound state solutions for the negative ions. We see that the improvements are not restricted to atoms but also occur for molecules. Table 5.2 shows results for the alkali and alkaline earth atoms and ions. The same quality is obtained as for the noble gas atoms and ions. Does the improvement of the asymptotic behaviour achieved by the model potential, apart from showing up clearly in the one-electron energies, also have observable consequences for the density? ln tables 5.2,5.3 and 5.4 we investigate a number of moments of orbital and total densities. Table 5.2 demonstrates that the radial extent of the highest occupied orbital in the alkali and alkaline earth atoms is signi cantly a ected by the asymptotic correction introduced by the model potential. Maximum errors of the LDA potential amount to 9% and 18% for hri and hr2 i respectively of the alkalis and 6% and 13% for the alkaline earths. The rst members of the two series, Li and Be, show comparitively small errors of 0.2% and 0.3% for Li and 1% and 2% for Be. The errors are much smaller for the ions, probably since they possess very tight closed shells. Although the e ects are rather small for Be, making this atom perhaps not a good test case, we happen to have a very accurate CI density available for this atom and therefore we compare in table 5.3 for Be the moments of the total density as obtained from various calculations. Also the Hartree-Fock data are shown in this table, in order to see how much of the e ect of correlation is taken into account by the LDA or model potential. Judging by this example, for hrp i with p positive the LDA potential corrects the Hartree-Fock result in spite of the wrong asymptotic behaviour of this potential, the model potential gives improvement over LDA but in general only something like 50% of the di erence between Hartree-Fock and exact is covered by the model potential. For hr;1 i and hr;2i there is no improvement over the Hartree-Fock results. For Neon (table 5.4) it is seen that for hrp i with p positive the model potential considerably improves the LDA result. With respect to the di erence with Hartree-Fock the picture is di erent from that for Be: the LDA potential does not correct the Hartree-Fock results but gives as expected a too di use density. For hr;1 i and hr;2 i there is again no improvement over the Hartree-Fock results. The LDA potential appears to give a too di use, the model potential a too contracted density in the inner region.
A MODEL POTENTIAL
ATOM Li < 2s > Na < 3s > K < 4s > Rb < 5s > Cs < 6s > Be < 2s > Mg < 3s > Ca < 4s > Sr < 5s > Ba < 6s > ION Li+ < 1s > Na+ < 2p > K+ < 3p > Rb+ < 4p > Cs+ < 5p > Be2+ < 1s > Mg2+ < 2p > Ca2+ < 3p > Sr2+< 4p > Ba2+ < 5p >
97
< r > < r2 >

3.822 3.995 4.839 5.125 5.674 2.621 3.137 3.991 4.339 4.880 0.585 0.803 1.428 1.720 2.087 0.421 0.684 1.260 1.555 1.912 17.345 18.753 26.962 30.041 36.481 8.263 11.572 18.304 21.468 26.909 0.468 0.839 2.421 3.428 4.959 0.240 0.599 1.866 2.775 4.129
LDA
0.12 0.11 0.10 0.09 0.08 0.20 0.17 0.14 0.13 0.12 2.19 1.34 0.92 0.80 0.69 4.81 2.46 1.58 1.34 1.14
< r > < r2 >

3.815 3.703 4.428 4.664 5.170 2.595 2.985 3.771 4.079 4.588 0.576 0.795 1.430 1.713 2.078 0.417 0.681 1.263 1.553 1.909 17.410 16.256 22.739 25.036 30.475 8.118 10.528 16.437 19.066 23.897 0.453 0.821 2.428 3.404 4.920 0.235 0.592 1.879 2.772 4.123
MODEL
EXPT 0.19 0.21 0.18 0.18 0.16 0.32 0.29 0.24 0.23 0.21 2.65 1.70 1.15 1.01 0.88 5.45 2.88 1.85 1.58 1.35 0.20 0.19 0.16 0.15 0.14 0.34 0.28 0.22 0.21 0.19 EXPT 2.78 1.74 1.16 1.00 0.85 5.66 2.95 1.87 1.60 1.35
< r > < r2 >
LDA
< r > < r2 >
MODEL
Table 5.2: expectation values of < r > and < r2 > for the highest occupied orbitals with corresponding eigenvalue
< r ;2 > < r ;1 > < r > < r 2 > < r 3 > < r 4 >
HF LDA MODEL EXACT 57.618 56.766 57.837 57.597 8.409 8.339 8.446 8.427 6.129 6.091 6.022 5.975 17.319 17.019 16.704 16.284 63.151 61.475 59.796 56.946 270.656 261.641 250.937 233.167
< r5 >
1325.49 1276.385 1198.763 1085.87
Table 5.3: density moments for the beryllium atom
< r;2 > < r;1 > < r > < r2 > < r3 > < r4 > < r5 >
HF LDA MODEL EXACT 414.890 411.916 416.406 414.967 31.113 30.998 31.275 31.110 7.891 8.016 7.895 7.935 9.372 9.853 9.477 9.545 15.967 32.500 80.692 14.799 28.576 66.239 14.941 29.006 67.904
Table 5.4: density moments for the neon atom
98
5.6 Summary and conclusions

In this work we have formulated a number of conditions for approximate exchange-correlation potentials. Potentials of gradient-corrected density functionals that are currently used with much success to calculate atomic exchange and correlation energies as well as bond energies of molecules do not obey some of these conditions. In order to make comparisons over all space to (almost) exact Kohn-Sham potentials we devised an iterative scheme to obtain the corresponding KohnSham potential from a given density. The scheme is applicable to both atoms and molecules and is not limited to systems with few electrons. In this paper highly accurate CI densities for Be and Ne have been used to generate accurate Kohn-Sham potentials. Comparison to potentials derived from existing gradient-corrected functionals demonstrated that these potentials cover only a small part of the di erence between the LDA potential and the exact one. This is at rst sight a little surprising: how can we have large improvements in energies and almost no improvement in the potential? Several explanations may be advanced. First of all the approximate nonlocal functional might "oscillate" around the exact functional giving a good approximation of the energies but a bad approximation of its functional derivative. Other de ciencies are inherent to the derivation of the functionals. For instance both the nonlocal exchange-correlation functional of Langreth and Mehl 50] and the Perdew-Wang generalised gradient expansion 46] use a long-range cut-o of the exchange-correlation energy density. Langreth and Mehl use a low-k (large distance) cut-o in the momentum distribution of the exchange-correlation energy and Perdew and Wang perform a real-space cut-o in the exchange-correlation hole. This neglect of this asymptotic region re ects itself in the potential which thereby loses its asymptotic Coulombic behaviour. However the Becke functional shows that even a correct behaviour of the exchange hole potential does not garantee a good behaviour of the exchange potential. A model potential has been presented which corrects some of the de ciencies of the current potentials, notably the asymptotic behaviour. The improved asymptotic behaviour shows up very clearly in the one-electron energies. The error in the highest occupied LDA eigenvalue (which should represent in the exact case the ionisation energy) is reduced by almost an order of magnitude by using the model potential. The model potential also corrects signi cantly the higher moments of the density, notably for di use outer orbitals. Because of this correction of the highest occupied orbitals, the model potential may be useful in the calculation of highly excited di use states and in general for density dependent properties such as dipole and quadrupole moments and derivatives thereof (infrared intensities). In spite of some success, the present model potential is clearly de cient in some respects, notably the behaviour in the atomic intershell region. It will therefore require further improvement. This problem will be addressed in a subsequent paper of this series. We feel that the most signi cant success of the nonlocal corrections to LDA apply to bond energies of molecules. In particular the potential derived from the Becke gradient correction to the exchange energy density displays very interesting structure in molecules 90] that may explain the success of the Becke energy expression for bond energies. We feel that further modeling of exchange-correlation potentials should not only try to optimize the potential for atoms but should take into account the special e ects of chemical bonding.
Acknowledgements: We would like to thank Gijsbert Wiesenekker for his help in the programming of the iterative scheme of section 5.4.
Chapter 6
Structure of the optimized e ective Kohn-Sham exchange potential and its gradient approximations
An analysis of the structure of the optimized e ective Kohn-Sham exchange potential vx and its gradient approximations is presented. The potential is decomposed into the Slater potential vS and the response of vS to the density variations vresp, the latter exhibits the distinct atomic shell structure. The peaks of vresp are interpreted as a consequence of the Fermi hole localization. It is the approximation of vresp that raises the main problem for the gradient approaches. A direct t of vx is made with the gradient-dependent Pade approximant form that possesses proper asymptotic and scaling properties and reproduces its shell structure.
6.1 Introduction
The optimized potential model (OPM) 34, 33, 91, 35] provides a direct way to calculate local potential v(r) of the Kohn-Sham equations (Hartree atomic units will be used throughout the paper) 1 ; 2 r2 + v(r) i(r) = i i(r) (6.1) Within the OPM the optimized e ective potential (OEP) v(r) is de ned as a potential of (6.1) with the eigenfunctions i (r), whose Slater determinant minimizes the expectation value of the total Hamiltonian H of a many-electron system ^ h jH j i (6.2)
v(r) can be subdivided in the usual way v(r) = vext (r) + vH (r) + vx(r) where vext (r) is the external potential, vH (r) is the Hartree potential of electrons
(r ) vH (r) = jr ; r0j dr0
Z
(6.3) (6.4)
99
100
N
STUCTURE OF THE OPTIMIZED..
with the electron density (r) built from the occupied orbitals i (r) (r) =
X
and vx (r) is called the exchange optimized e ective potential. The important features of atomic vx such as its asymptotic behaviour at short and long distances from a nucleus 33, 57, 31, 35, 92, 49] and the shell structure 33, 50, 93, 35, 94] were established in the OPM. The correct long-distance asymptotics was proved for the vx of the OPM in the pioneer work of Talman and Shadwick 33]. On the other hand, an interpretation of the shell structure, the "peaks" of vx in the atomic intershell regions remained an open problem. Another problem is nding approximations to vx . Even for atomic systems the rigorous OPM calculations are rather involved 93, 35, 36], because they require the simultaneous self-consistent solution of the oneelectron Kohn-Sham equations and an integral equation for vx . Owing to this, it is desirable to develop an e ective approximation to vx within the density functional theory (DFT). Recent progress of the DFT is due to the generalized gradient approximation (GGA) 50, 51, 45, 95, 44, 43, 47, 96] that gives a remarkable improvement to the exchange-correlation energies of the local density approximation (LDA) 10, 11, 97, 98, 99]. It was shown 92, 49, 36, 94, 32], however, that the standard GGA exchange potentials do not possess the proper asymptotic behaviour and shell structure. The corresponding gradient corrections add little to the LDA exchange potential. In this paper we analyze a structure of the exchange OEP vx and its gradient approximations. In Section 6.2 properties of vx are discussed. vx is decomposed into the Fermi hole or Slater potential vS and the response of vS to the density variation vresp and the peaks of atomic vx are interpreted as a consequence of the Fermi hole localization. In Section 6.3 the atomic OEP and its components vS and vresp are compared with those of the LDA, GGA and the pioneer gradient expansion approximation (GEA) of Herman et al. 76, 100]. In Section 6.4 a gradient t to vx is presented, which possesses the proper asymptotic and scaling properties and reproduces the shell structure of the OEP. The results of atomic calculations with the tted potentials are compared with those of the OPM, LDA, GGA and GEA. The comparison is made for the form of potentials, orbital eigenvalues and also for the exchange energies, the latter have been calculated for the tted potential via the Levy-Perdew relation 22].
i=1
j i (r)j2
(6.5)
6.2 Properties of the exchange optimized e ective potential

The exchange OEP vx(r) possesses the following scaling property 22, 23] where (r) = 3 ( r). For nite systems it has the asymptotic form 33, 57, 31, 35] vx (r) ! ; j1j (jrj ! 1) r
vx ( ] r) = vx( ] r)
(6.6) (6.7)
Using 6.1,6.2 the OPM provides a radial integral equation 33, 91] for the atomic OEP v(r) (r is the radial coordinate), which can be solved numerically. Figure 1 illustrates the typical features of the OEP exchange part vx for Kr extracted from 91]. One can see from this Figure the niteness of vx at the nucleus
vx (r) = const
(r = 0)
(6.8)
PROPERTIES OF THE EXCHANGE OPTIMIZED EFFECTIVE POTENTIAL
101
Figure 6.1: Exchange optimized e ective potential for Kr and its shell structure, i.e. the non-monotonous dependence of vx on r with the local maxima ("peaks") in the regions between the atomic shells. In order to trace the origin of the intershell peaks of vx , we use its de nition as the functional derivative of the exchange energy Ex ] vx(r) = Exr)] (6.9) ( that holds true in the OPM and express the exchange part of 6.2 as an integral of the Fermi hole or Slater potential vS 101] 1Z (6.10) Ex = 2 (r)vS (r)dr
0 vS (r) = ; fr(r r 0) dr0 j ;rj
Z
(6.11)
In equation 6.11 f (r r0 ) is the Fermi hole density 101, 102] built from the occupied Kohn-Sham orbitals i (r)
XX f (r r0 ) = (r);1 i (r) i (r0 ) j (r) j (r0 )
N N
i=1 j =1
(6.12)
The expression 6.10 formally looks like an energy of the electronic interaction with an external potential. In reality, being an internal potential of the electronic system, vS has the non-zero
102
response vresp to density variations and, according to equations 6.9-6.11, vx can be expressed as the arithmetical mean of vS and its response vresp 1 (6.13) vx(r) = 2 vS (r) + vresp(r)]
vresp(r) =
0
Z Z 0 f (r0 r 00 ) dr 00 0 ) vs (r ) dr0 = ; (r0 )dr0 (r (r) (r) jr0 ; r 00 j
(6.14)
r with vs((r)) and f (r(rr) ) being the linear response functions 6]. 1 In Figure 2 OEP vx 91] and its components 2 vS and 1 vresp are plotted for Ca, Kr and Cd. The 2 relative contribution of vresp to vx ranges from 0.5, its asymptotic value at long distances r, and from 0:35 within the shell regions to the near zero values in the intershell regions. In all cases vS is a monotonous function of r, while vresp possesses the distinct peaks in the intershell regions and, hence, the shell structure of vx originates mainly from vresp. From Figures 2a-c and eq.6.14 the shell 00 structure can be interpreted in terms of the di erent response of the Fermi hole density f (r2 r ) to the perturbations (r) of the electron density in 1 the shell and intershell regions. At the intershell points vx is closer to 2 vS and vS looks in fact like an external potential which has, by de nition, zero response to the density perturbations. In the sense of integral contribution to 6.14, intershell points appear as the points outside the area of the Fermi hole 6.12, which determines vresp. But this is true, because, owing00 to an exchange e ect, 00the Fermi repulsion of electrons with like spins, for various pairs r0 and r the Fermi hole f (r0 r ) is mainly localized within the shell regions 102, 103]. One can expect (and we can see it from Figure 2) the smaller response of f (r0 r 00 ) when (r) is perturbed somewhere outside its localization region, in particular, at the intershell points. To sum up, the intershell peaks of vx can be interpreted as a consequence of localization of the Fermi hole and successful approximations to vx should re ect this feature. It is interesting to note, that the shell structure of the exchange potential is only the secondary e ect of the Fermi repulsion. The main e ect is contained in the Pauli potential 1 104], which appears as a kinetic contribution to the total potential of the Euler equation for 2 (r) and has much more pronounced intershell peaks 105, 106, 107]. The analysis of the shell structures in the Euler equation will be given elsewhere 108].
0 00
6.3 Comparison of the OPM, LDA, GGA and GEA exchange potentials
In the recent papers 92, 49, 35, 94] OEP vx has already been compared with various DFT approximations. In this paper the comparison is made not only for vx , but also for its components vS and vresp. All the standard DFT schemes approximate the exchange energy density x( r)
approx ] = Ex
Z
approx( x
] r)dr
(6.15)
and, according to equations 6.10 and 6.15, the corresponding approximation to vS can be simply related to approx x approx approx (6.16) vS (r) = 2 x (r() ] r)
COMPARISON OF THE OPM, LDA, GGA AND GEA POTENTIALS
103
Figure 6.2: Exchange optimized e ective potential and its components for a) Ca, b) Kr and c) Cd
104
Figure 6.3: Exchange potentials of OPM, LDA, GEA and GGA for a) Ca, b) Kr and c) Cd

approx while vresp can be obtained by the functional di erentiation of equation 6.16. LDA LDA LDA In the LDA all the potentials vx , vS and vresp have the same form 1 1 LDA LDA 3 3 vx (r) = ;2( 83 ) 1 3 (r) vS (r) = ;3( 83 ) 1 3 (r) 1 LDA 3 vresp (r) = ;( 83 ) 1 3 (r) LDA In the GEA 76, 100] vS is modi ed with the density gradient correction 1 GEA LDA LDA 3 vS (r) = vS (r) ; 0:015( 43 ) 3 1 (r) = vS (r) + vS (r) ~GEA 2 = 2 3 ( r4 (r) )2 3 (r)
105
(6.17) (6.18) (6.19)
GEA and in vresp the additional Laplacian term appears 1 GEA LDA LDA 3 vresp (r) = vresp (r) ; 0:005( 43 ) 3 1 (r)( ; 6 ) = vresp (r) + vresp (r) ~GEA (6.20) 2 2 = 2 3 r 5 (r) (6.21) 3 (r) 3 LDA GEA LDA 3 3 ~GEA (6.22) vx (r) = vx (r) ; 0:01( 43 ) 1 1 (r)( ; 2 ) = vx (r) + vx (r) All the gradient corrections in eqs.6.18,6.20,6.22 contain the empirical parameter determined in 76, 100]. GEA In the GGA 44] vS is further modi ed with the correction function f ( ) in order to provide the GGA correct long-distance asymptotics of vS GGA LDA vS (r) = vS (r) + f ( )~S (r) vGEA (6.23)
3 0:56( 43 ) 1 f( ) = 1 1 + 0:0252 2 sinh;1 1 2
(6.24)
GGA and in vresp the function f ( ) appears together with its rst and second derivatives 2 1 1 df GGA LDA vresp (r) = vresp (r) + f ( )~resp (r) ; 0:03( 43 ) 3 3 (r) d ( + 2 ) ; d f ] vGEA d2
(6.25) (6.26) (6.27)
3 df GGA LDA 3 3 vx (r) = vx (r) + f ( )~x (r) ; 0:015( 43 ) 1 1 (r) 2 d ( + 2 ) ; 3 d f ] vGEA 2d 2

4 = 2 3 r (r5)r (r) 3 (r)
Eq. 6.24 contains an empirical parameter determined in 44]. Being the functional derivatives of LDA GEA GGA the corresponding exchange energy functionals, vx , vx , vx and also their components have the proper scaling property 6.6.
106
Figure 6.4: Slater potentials of OPM, LDA, GEA and GGA for a) Ca, b) Kr and c) Cd
107
Figure 6.5: Response potentials of OPM, LDA, GEA and GGA for a) Ca, b) Kr and c) Cd
108
In Figures 3-5 OEP vx 91] and its components vS and vresp are compared for Ca, Kr and Cd with the corresponding potentials of LDA, GEA and GGA. The potentials were calculated with the OPM densities . All the approximate total exchange potentials fail to reproduce the correct LDA GEA asymptotics 6.7 of the OEP. At long distances vx falls o exponentially, vx diverges (See Figure 3b where the potentials for Kr are represented in logarithmic scale for larger distances), GGA LDA vx decreases similarly to vx and only at very long distances it is proportional to ; r12 49]. At LDA is nite, while the gradient potentials vGEA and vGGA diverge, being proportional the nucleus vx x x to ; 1 92, 32]. r LDA Both LDA and GGA potentials do not possess the distinct shell structure. vx is an everywhere GGA exhibits a peak only for the outermost intershell regions of monotonous function of r and vx GEA Ca and Cd, being a monotonous function of r for the rest of the space. Contrary to this, vx possesses clear shell structure. Amplitudes of the peaks obtained within the GEA di er from those of the OPM, but the locations of the peaks are nearly the same as in the OPM (See Figure 3a-c). approx originate mainly from those between The abovementioned discrepancies between vx and vx approx , while the correspondence between vS and vapprox is de nitely better. All considvresp and vresp S ered approximations to vS are nite at the nucleus and approximate rather closely vS at short and intermediate distances r (See Figure 4). In particular, the GGA provides especially good approxiGGA mation for vS . By the construction 44], vS has the correct asymptotics 6.7 for r ! 1 and it is very close to vS for all distances. Hence, the well-known success of the GGA in calculation of the integral characteristics, such as Ex , can be attributed to its locally very good approximation to vS LDA GEA LDA and, which is the same, to the exchange energy density x . vS and vS are not so good: vx GEA diverges at long distances and both potentials exhibit appreciable falls of exponentially and vS discrepancies with vS at r 2a:u: However, all the approaches show more pronounced de ciencies in approximating vresp (See Figure approx 5). Due to the presence of the unbalanced Laplacian contributions to vresp , both GEA and GGA diverge at the nucleus, which causes the abovementioned divergence of the corresponding total exchange potentials. All the approaches fail to reproduce the correct asymptotics vresp(r) ! ; 1 r LDA GEA GGA for r ! 1: vresp decreases exponentially, vresp diverges and vresp yields the positive asymptotics, GGA GGA vresp (r) ! 1 , that cancels the correct asymptotics of vS and produces the incorrect asymptotics r GGA LDA of the total exchange GGA potential vx . In the intershell regions vresp is a monotonous function GGA has the peaks, but for all intershell regions, with the exception of the outermost one, of r, vresp they are much smaller than those of the OPM. Contrary to the LDA and GGA, GEA reproduces GEA the peaks of vresp, but with substantial deviations of amplitudes for vresp and vresp peaks. One can see from the above analysis and discussion, that the methods to obtain vx as functional derivative of the current approximate expressions for Ex do not meet the high demand to provide satisfactory approximation not only for vS , but also for its functional derivative. The alternative way of a direct tting of vx to a gradient function with a due account of the accurate asymptotic and scaling properties is presented in the next section.
6.4 A gradient t to vx
To t vx , we use the Pade approximant approach, which has been used previously for the tting of x 49, 109, 85, 110, 94]. The following Pade approximant form is considered as a tting potential
f vx
1 1 1 n( m( f LDA 2 vx (r) = vx (r) ; 2 3 3 (r) Pn+1) ( 1 sinh;1 2 ) ; Pm+1) (j j) ]
(6.28)
A GRADIENT FIT TO Vx
where Plk(y) (x) is a k=l]-Pade function of degrees k and l 111]
y y y k(y) (x) = a0 + a1 x + ::: + ak xk Pl 1 + by x + ::: + by xl 1 l 1 n( 2 simpliest Pade function Pn+1) ( 1 sinh;1 2 )
109
(6.29)
The with n = 0 is the Becke's f ( ) (eq.6.24), which has also been used in 32] as a model of the exchange-correlation potential. The whole expression 6.28 can be considered as the Pade approximant modi cation of the GEA exchange potential 6.22. It appears that with the following restriction for the coe cients an and bn+1 (6.30) the Pade approximant form eqn.6.28 of arbitrary orders n and m possesses the scaling property 6.6, accurate long-distance asymptotics 6.7 and niteness 6.8 at the nucleus. The forms 6.28 with n m = 0 ; 2 have been used to t vx with the coe cients determined from f the least squares minimization of the weighted di erence vx (r) ; vx (r)] for the closed-shell atoms Be, Ne, Mg, Ar, Ca, Zn, Kr, Cd XX f fwA (ri ) vx(ri ) ; vx (ri )]g2 = min (6.31) In equation 6.31 the inner summation is over the radial mesh points ri , the outer one is over atoms and weighting functions wA (ri ) of the following form have been chosen 2 A wA (r ) = pi ri (ri) (6.32)
1 i
A Ex dA 2 A A (ri ) = pi ri 3 (ri ) + r dr jr=ri ] w2 A Ex A i
bn+1 = 3an
(6.33)
where pi are the weighting parameters. The radial densities are inserted in equation 6.32 to provide weights to the points ri according to their contributions to the sum of the orbital energies of the Kohn-Sham equations. The weights in eqn.6.33 represent the corresponding contributions to the A atomic exchange energy Ex , which can be calculated for a given vx via the Levy-Perdew relation 22]
Ex =
3 (r) + rr (r)]vx (r)dr
(6.34)
A Energies Ex are inserted in equations 6.32, 6.33 in order to make an uniformly representative functional for both light and heavy atoms. The minimized value of the functional (eqn. 6.31) decreases considerably with increasing of degrees n and m of the t of eqn. 6.28 from the least n = 0, m = 0 to n = 0, m = 1, but higher degrees appear to produce little e ect. As a result, the following simple particular case of the form of eqn. 6.28 has been chosen as a tting potential
(6.35) 1 + 3a0 sinh;1 2 f The optimized form of vx , naturally, depends essentially on the actual weighting function wA (ri ) chosen. To illustrate this, in Figure 6 the non-local potential correction vnl of the OPM (a di erence LDA between vx and vx ) is compared for Ca, Kr and Cd with the corresponding corrections for
1 2 1 2
1 1 f LDA vx (r) = vx (r) ; 2 3 3 (r)
a0
+ ; (a10 + ba1 2) ]
110 ORBITAL 1s 2s 2p 3s 3p 4s OPM 144.50 15.26 12.43 1.77 1.08 0.195 144.60 15.20 12.37 1.73 1.05 0.144
f vx 1

145.11 15.36 12.54 1.77 1.09 0.165
f vx2
146.32 15.68 12.89 1.85 1.18 0.206
f vxc
GGA 144.38 15.05 12.26 1.67 0.99 0.116
LDA 143.85 14.98 12.22 1.66 0.98 0.111
Table 6.1: Orbital energies for Ca (with the opposite sign and in a.u.) calculated within OPM, f f f LDA, GGA and with the tting potentials vx 1 , vx 2 and vxc
f f the GEA, GGA and for two ts vx 1 and vx 2 of the form of eqn. 6.35 with the di erent sets of 1 = 0:0123, a 1 = 0:0087, a 1 = 0:0004, b 1 = 0:0011 for vf 1 and the optimized coe cients: a0 x 0 1 1 f a02 = 0:0204, a02 = 0:0086, a12 = 0:0005, b12 = 0:0022 for vx2 . For both ts the radial mesh fri g of f Numerov method 112] with the adaptive step size has been used. vx 1 has been optimized for the A (ri ) (eqn. 6.32) with all the weighting parameters pi set equal, i.e. pi = 1, weighting function w1 f A while vx 2 has been optimized for the weighting function w2 (ri ) (eqn. 6.33) with the mesh weights of the Numerov quadrature being used as weighting parameters pi. Due to a large concentration A of the mesh points near the nucleus, the function w1 (ri ) provides the greater contribution of the f A inner radial points in equation 6.31 as compared to w2 (ri ). As a result, vnl1 reproduces the shell structure of the OPM potential better than any other potential presented (See Figure 4) and is f close to vnl in the core region, while vnl2 goes closer to vnl in the valence region. f f It is interesting to note that, as regards to a0 value, vx 2 is placed in between vx 1 and the model f of 32]. The latter is a particular case of eqn.6.35 with a = a = exchange-correlation potential vxc 0 1 b2 = 0 and with a0 = 0:05 chosen to reproduce the rst ionisation potential of Be atom. As it was f shown in 32], vxc yields a good estimate of ionisation energies for various atomic and molecular systems and it goes close to the accurate Kohn-Sham potential in the outer valence regions. The abovementioned features of the potentials are re ected in Tables 1-3 where the orbital energies i of eq.6.1 calculated with various exchange potentials are presented for Ca, Kr and Cd. LDA underestimates electron exchange and its orbital energies are smaller (in absolute magnitude) than those of the OPM. GGA yields energies, which are close to the LDA ones for valence and outer core orbitals, and only for the deep core orbitals the GGA energies grow de nitely greater. In all f cases (with the only exceptions of 2s- and 2p- orbitals of Cd) vx 1 provides an improvement to the f GGA orbital energies. Typical i errors vary within 0:05 ; 0:15 a.u. for vx 1 , which means small f relative errors for the core orbital energies and quite appreciable errors for the valence ones. vx 2 provides further improvement for the valence and outer core energies, though its deep core orbital f f energies are worse than those of vx 1 . vxc gives the best energies of the highest occupied orbitals for Ca and Kr, while the deep core energies are de nitely overestimated. For Cd the highest orbital f f energy for vxc deviates more from OPM than vx 2 does, but it is quite close to its accurate ionisation f , being an approximation to the exchange-correlation rather than exchangeenergy 0.330 a.u. vxc only potential, should indeed yield the ionisation energy. Note that, by the construction (eqn.6.35), all tted potentials have the accurate long-distance asymptotics (eqn.6.7). One can see, that this property by itself cannot guarantee high quality of the calculated energies of the highest occupied f orbitals that depends critically on the vx behaviour at shorter distances in the outer valence region. nl In Table 6.4 the non-local corrections Ex of the OPM to the LDA exchange energies are compared for the closed-shell atoms Be, Ne, Mg, Ar, Ca, Zn, Kr, Cd with those of the approximate schemes.
111
Figure 6.6: Non-local corrections of OPM, GEA, GGA and of the tting potentials to the LDA exchange potential for a) Ca, b) Kr and c) Cd
112
Figure 6.7: Dimensionless arguments and for Kr
ORBITAL 1s 2s 2p 3s 3p 3d 4s 4p
OPM 511.09 66.64 60.27 9.65 7.37 3.30 0.97 0.501
511.21 66.52 60.17 9.41 7.15 3.10 0.83 0.358
f vx 1
512.08 66.76 60.42 9.50 7.24 3.20 0.87 0.399
f vx2
514.22 67.32 61.05 9.68 7.44 3.41 0.96 0.480
f vxc
GGA 510.87 66.35 60.02 9.30 7.05 3.02 0.78 0.305
LDA 509.89 66.21 59.94 9.25 7.02 3.01 0.77 0.300
Table 6.2: Orbital energies for Kr (with the opposite sign and in a.u.) calculated within OPM, f f f LDA, GGA and with the tting potentials vx 1 , vx 2 and vxc
ORBITAL 1s 2s 2p 3s 3p 3d 4s 4p 4d 5s OPM 943.01 137.00 127.68 25.56 21.76 14.75 3.70 2.50 0.56 0.267 943.10 137.12 127.82 25.49 21.71 14.71 3.61 2.40 0.47 0.208
f vx 1
113 944.24 137.41 128.14 25.65 21.86 14.86 3.65 2.44 0.52 0.250
f vx2
947.13 138.18 129.00 25.98 22.21 15.21 3.74 2.54 0.63 0.321
f vxc
GGA 942.70 136.94 127.67 25.35 21.61 14.64 3.56 2.35 0.43 0.170
LDA 941.38 136.76 127.56 25.31 21.57 14.62 3.54 2.34 0.42 0.168
Table 6.3: Orbital energies for Cd (with the opposite sign and in a.u.) calculated within OPM, f f f LDA, GGA and with the tting potentials vx 1 , vx 2 and vxc ATOM Be Ne Mg Ar Ca Zn Kr Cd OPM 0.360 1.121 1.402 2.339 2.614 4.112 5.274 7.419 GGA 0.345 1.102 1.387 2.287 2.597 4.215 5.241 7.376 0.175 0.908 1.233 2.023 2.457 4.651 5.649 8.400
f vx 1
0.193 1.183 1.611 2.663 3.238 6.464 8.024 11.776
f vx 2
0.367 2.147 2.921 5.034 6.047 12.193 15.760 22.758
f vxc
GEA 0.826 2.275 2.766 4.409 4.915 7.611 9.351 12.859
Table 6.4: Non-local corrections to the LDA exchange energies (with the opposite sign and in a.u.) calculated within OPM, GGA, GEA and via the Levy-Perdew relation with the tting potentials f f f vx 1 , vx 2 and vxc
nl For the tted potentials Ex are calculated from the Levy-Perdew relation (eqn.6.34) and for GGA and GEA they are calculated directly from the corresponding energy expressions (eqn.6.15). f nl Though non-variational, vx 1 yields a reasonable estimation of Ex for the atoms considered. Its quality is worse, than the highly superior quality of the GGA 44] energies, but better than that of f nl GEA 76, 100]. vx 2 also yields somewhat better Ex values than the GEA, though in this case for nl f all atoms heavier than Ne the calculated Ex values are substantially overestimated. vxc provides nl for Mg and heavier atoms. A comparison of the results for tted the largest overestimation of Ex potentials shows, not unexpectedly, that, in order to provide a reasonable estimate for the total exchange energy, it is more important to have a better t for the inner region than for the outer one. A natural question arises whether it possible to construct a potential that would combine the f f f advantages of vx 1 , vx 2 and vxc and would provide a close t to vx at all distances. Our experience gives a negative answer to this question for the case of the Pade approximant form (eqn. 6.28). The reason for this is the concerted oscillatory behaviour of the dimensionless arguments (eq. 6.19) and (eq. 6.21) as functions of r (See Figure 7 for Kr as a typical example). and are the natural parameters of the gradient expansion for x and vx and their utilization ensures the accurate scaling property (eqn. 6.6) for various gradient approximations. However, because of their oscillations, there always exist core and valence regions where both and have a similar behaviour, so that
114
the gradient approximations always produce concerted, related to each other non-local corrections for these regions. Due to this, the local exibility of the t (eqn. 6.28) is substantially restricted and better tting for the inner part of vx produces at the same time worse tting for the outer part and vice versa. As a result, the increase of a0 and removal of the -dependent part of (eqn.6.35), f f f when passing on from vx 1 through vx 2 to vxc, brings better energies of the highest occupied orbitals nl but worse Ex values. f f From the above discussion vx 1 seems to be the more optimal t to vx as compared with vx 2 . The former provides systematical improvement for the orbital energies of the GGA, while keeping reasonable estimation of the exchange energies and it can be used in the Kohn-Sham equations as a simple approximation to vx and also as a starting point for more re ned approximations. For such a re nement one can consider (in addition to the standard and ) other gradient parameters that should have distinctly di erent behaviour in the core and valence regions. Using these parameters f f for the interpolation between vx 1 in inner regions and vxc in the outer ones, one can hope to construct uniformly close t to vx , or indeed to vxc if required. The corresponding work is in progress.
6.5 Conclusions
In this paper an analysis of the structure of the optimized e ective Kohn-Sham exchange potential and its gradient approximations has been presented. The potential has been decomposed into the Slater potential and its response to the density variations, the latter exhibits the distinct atomic shell structure. The intershell peaks of the potential originate from the smaller response of the Fermi hole density, which is localized in the shell regions, to the density perturbations at the intershell points. It has been shown, that the approximation of the response potential is a major problem for the gradient approach. A direct t of the potential has been made with the the Pade approximant form that possesses proper asymptotic and scaling properties and reproduces shell structure of the OEP.
Acknowledgements: This investigation was supported in part by the Netherlands Foundation for Scienti c Research (NWO).
Chapter 7
Analysis of electron interaction and atomic shell structure in terms of local potentials
The Kohn-Sham potential vs of an N-electron system and the potential veff of the Euler-Lagrange equation for the square root of the electron density are expressed as the sum of the external potential plus potentials related to the electronic structure, such as the potential of the electron Coulomb repulsion, including the Hartree potential and the screening due to exchange and correlation, a potential representing the e ect of Fermi-Dirac statistics and Coulomb correlation on the kinetic functional, and additional potentials representing "response" e ects on these potentials. For atoms several of these potentials have distinct atomic shell structure: one of them has peaks between the shells, while two others are step functions. In one of those step functions the steps represent characteristic shell energies. Examples of the potentials extracted from the optimized potential model (OPM) are presented for Kr and Cd. Correlation potentials, obtained by subtracting the exchange potential of the OPM from (nearly) exact Kohn-Sham potentials, are discussed for Be and Ne.
7.1 Introduction
Local one-electron potentials are key ingredients of density functional theory (DFT) 5] The potential vs (r) of the familiar Kohn-Sham one-electron equations 11] (Hartree atomic units will be used throughout the paper) 1 (; 2 r2 + vs ( ] r)) i (r) = i i (r) (7.1) and the e ective potential veff ( ] r) of the Euler-Lagrange equation for the square root of the electron density 113, 114, 115, 116, 107] q q 1 (; 2 r2 + veff ( ] r)) (r) = (r) (7.2) accumulate various e ects of the electron interaction in many-electron systems. They provide an informative picture of the electron interaction in terms of potential barriers and wells in physical 115
116
ANALYSIS OF ELECTRON INTERACTION..
space. Evidently, the complete e ect of electron correlation has been folded into these local potentials. In this paper we will analyze the relationship between the potentials and electron correlation. This analysis leads to a decomposition of the potentials in various components, which have been calculated within the (exchange only) optimized potential model 33] for the atoms Kr and Cd. For Be and Ne the correlation part of the exchange-correlation potential has been obtained from nearly exact vs corresponding to the electron density as given by highly correlated wavefunctions. A distinct structure (peaks, steps) is observed in the various potentials, that can be rationalized in terms of aspects of the electron correlation. The observed structure will be very helpful when one tries to model the potentials accurately and e ciently by judiciously chosen functionals of the density and its derivatives.
7.2 De nition of local potentials

To analyze the structure of vs and veff , we begin with the electronic energy expression of the constrained-search DFT 18] for a trial density
Ev ] =
^ ^ (r)vext (r)dr + min h jT + W j i !
(7.3)
^ ^ where vext is the external (usually, Coulombic) potential, T and W are operators of the kinetic energy and electron repulsion and the minimum is searched over all antisymmetric normalized wavefunctions , which yield . One can express E in terms of local potentials 1 Z q (r)r2 q (r)dr+ Ev ] = TW ] + (T ] ; TW ]) + Vext ] + W ] = ; 2 Z Z 1Z (r)vkin ( ] r)dr + (r)vext (r)dr + 2 (r)vcond ( ] r)dr (7.4) Here TW ] is the von Weizsackerp kinetic energy functional which is N times the kinetic energy of the normalized "density orbital" =N , and the explicit e ect of the Fermi statistics and electron interaction on the true kinetic functional T ], i.e. the di erence T ] ; TW ], is represented with the local potential vkin. According to 86], the latter potential can be de ned via the conditional probability amplitude 113] (x x (s1 ~ 2 ::: ~ N jr1 ) = p~ 1 ::: ~ N ) x x (7.5) (r1 )=N of the (non-degenerate) function , which minimizes the energy functional (7.3) for the prescribed density (f~ i g = f~i si g, f~i g are the space and fsi g are the spin variables) x r r Z 1 vkin( ] r1 ) = 2 jr1 (s1 ~ 2 ::: ~ N jr1 )j2 ds1d~ 2 :::d~ N : x x x x (7.6) (s1 ~ 2 ::: ~ N jr1 ) embodies all e ects of electron correlation (exchange as well as Coulomb) in x x that its square is the probability distribution of the remaining N ; 1 electrons associated with positions ~ 2 :::~ N when one electron is known to be at r1 . vkin can be interpreted as a measure of x x how strongly the motion of the reference electron at r1 is correlated with the other electrons in the system, in the sense that it re ects the magnitude of change in with changing r1 .
DEFINITION OF LOCAL POTENTIALS

is expressed through the conditional density 86] cond ( ] r1 r2 ) = g( ] r1 r2 ) (r2 ) = = (N ; 1)
Z Z
117
vcond describes the e ective Coulomb repulsion of the reference electron by the other electrons. It
(s1 ~ 2 ::: ~ N jr1 ) (s1 ~ 2 ::: ~ N jr1 )ds1 ds2 d~ 3 :::d~ N x x x x x x
(7.7) (7.8)
vcond ( ] r1 ) =
r (r2 ) g(jr ] ;1r rj2 ) dr2

1 2
and can be considered as the potential of a interparticle Coulomb interaction screened due to exchange-correlation e ects, with the pair-correlation function g(r1 r2 ) being the screening factor. Evidently, vcond can be broken up into the Hartree potential and a screening potential due to the exchange-correlation hole density (r2 )(g(r1 r2 ) ; 1) surrounding the reference electron at r1 :
Z Z r) vcond ( ] r1 ) = jr (;2r j dr2 + (r2 )(g(r ];r1 r2 ) ; 1) dr2 = vH ( ] r1 ) + vscr ( ] r1 )(7.9) j 1 r2 j 1 2 The potentials vH and vscr are long-ranged, having asymptotic N=r resp. ;1=r behaviour.
resp Here the potential vscr is an integral of the linear "response" of g, g(r(1r3r)2 ) Z (r1 ) (r2 ) g( ] r1 r2 ) dr dr resp( ] r3 ) = 1 (7.11) vscr 1 2 2 jr1 ; r2 j (r3 ) This potential is a measure of the sensitivity of the pair-correlation function g to density variations. These density variations are to be understood in the following way. If we change the ground state density to + (v-representable), then of course Ev + ] will be higher than Ev ] as we do not change the external potential vext . By the Hohenberg-Kohn theorem a wavefunction + ] can be obtained as the ground state wavefunction corresponding to a unique external potential v + v ( which will minimize Ev+ v ). For this system we of course also have a corresponding Kohn-Sham system and a pair-correlation function g( + ] r1 r2 ). So the derivative of g in the resp above potential vscr may be regarded as linear response in the above sense of g to density changes caused by potential changes v. resp vkin is the response of the potential vkin to density variation Z resp( ] r ) = vkin (r2 ) vkin ((r ]) r2 ) dr2 (7.12) 1 1 resp resp Like vscr and vkin, the potentials vscr and vkin can be related directly to the electron correlation in the ground state. Their sum can according to ref. 86] be expressed in terms of the conditional amplitude: resp resp vscr ( ] r1 ) + vkin ( ] r1 ) = vN ;1 ( ] r1 ) Z N = (s1 ~ 2 ::: ~ N jr1 )H N ;1 (s1 ~ 2 ::: ~ N jr1 )ds1 d~ 2 d~ 3 :::d~ N ; E0 ;1 x x x x x x x (7.13)
Minimization of (7.4) with respect to leads to the Euler-Lagrange equation (7.2) with the following expression for veff veff ( ] r) = vext + vH ( ] r) + vscr ( ] r) + vkin( ] r) resp resp +vscr ( ] r) + vkin ( ] r) (7.10)
118
where H N ;1 is the hamiltonian of electrons 2 ; N (interacting) in the given external potential. The value of vN ;1 at position r is the energy expectation value of the system of N ; 1 electrons described by the conditional amplitude (s1 ~ 2 ::: ~ N jr) minus the ground state energy of the x x N ; 1 electron system. An expression for the Kohn-Sham potential vs can be derived from equation (7.10) for veff and a suitable expression for the Pauli potential vP 117, 115, 116, 107] , the di erence between veff and vs . The formula of 115] for the Pauli potential is useful. It may be interpreted with the help of the relation derived in 86] between the Pauli potential and the Kohn-Sham determinantal wavefunction s , which may be built from the Kohn-Sham orbitals i (r) of eq.(7.1): N vP ( ] r) = veff ( ] r) ; vs( ] r) = vs kin( ] r) + vs ;1 ( ] r) (7.14) N where vs kin and vs ;1 are the analogues of (7.6) and (7.13) but now constructed with the conditional amplitude s of the determinant s . Due to the simple one-determinantal nature of s , vs kin can be expressed straightforwardly in terms of the Kohn-Sham orbitals (cf. 115]):
N 1 Z jr (s ~ ::: ~ jr )j2 ds d~ :::d~ = 1 X r i (r1 ) 2 vs kin( ] r1 ) = 2 1 s 1 x2 xN 1 1 x2 xN 2 1 1 2 (r1 ) i=1
(7.15)
It has been argued 86] that from the relation between vkin resp. vs kin and the conditional amplitude resp. s , one may expect large peaks in the kinetic potentials at shell boundaries due to the special behaviour of the exchange hole, a point to which we will return below. N The second contribution to vP can also be obtained easily from the one-electron nature of Hs ;1 and the one-determinantal s:
N vs ;1 ( ] r1 ) =
Z
=
N Es 0;1 =
Z
N Here Es 0;1 is the energy of the N ; 1 Kohn-Sham system in the following sense: N ;1 x s (~ 2
N ;1 (s ~ x x s 1 x2 s (s1 ~ 2 ::: ~ N jr1 )Hs N 2 X ; i j i(r1 )j (r1 ) i=1
N ::: ~ N jr1 )ds1 d~ 2 d~ 3 :::d~ N ; Es 0;1 x x x x
(7.16)
: : : ~ N ) HsN ;1 (~ 2 : : : ~ N ) s ;1 (~ 2 : : : ~ N )d~ 2 : : : d~ N x x x N x x x x
(7.17)
N where s ;1 is the N ; 1-particle Kohn-Sham determinant with one electron removed and where HsN ;1 is the Kohn-Sham Hamiltonian Hs with one electron removed ( the Kohn-Sham potential vs ( ] r) is however still determined by the N -electron density so this is not strictly the KohnSham Hamiltonian of the ion):
HsN ;1 (~ 2 : : : ~ N ) = x x
N N = Es 0 ; Es 0;1 is equal to the highest occupied one-electron Kohn-Sham energy, M , which in N N turn is equal to the negative of the exact rst ionisation energy, M = E0 ; E0 ;1 . resp The components of vP are the Kohn-Sham analogues of the components vkin and vkin of veff (eq.7.10), de ned in terms of the exact wavefunction in eqs. (7.6) and (7.12). To demonstrate this we note that
i=2
1 r ; 2 r2 + vs( ] ~i ): i
(7.18)
Ts ] = TW ] +
(r)vs kin ( ] r)dr = TW ] + TP ]
(7.19)
DEFINITION OF LOCAL POTENTIALS
119
and since vP = TP = 115] resp vP ( ] r) = vs kin( ] r) + vs kin( ] r) (7.20) where Z resp ( ] r ) = vs kin (r2 ) vs kin((r ]) r2 ) dr2 1 1 resp with vN ;1 , and indeed vN ;1 will not have a response term Comparison to eq. 7.14 identi es vs kin s s resp corresponding to the term vscr in vN ;1 (eq.7.13) since the Kohn-Sham system of electrons is noninteracting. Combining now equation 7.20 with equation 7.10 for veff one obtains for vs resp vs( ] r) = vext (r) + vH ( ] r) + vscr ( ] r) + vscr ( ] r) resp resp +(vkin ( ] r) ; vs kin( ] r)) + (vkin ( ] r) ; vs kin( ] r)): (7.21) resp resp It is to be expected that the potentials vkin and vs kin, and the potentials vkin and vs kin, will be rather similar in atoms with predominantly exchange e ects and only dynamical (no neardegeneracy) correlation, in which case the exchange-correlation potential vxc = vs ; vext ; vH resp would be dominated by vscr + vscr . As a matter of fact, since the exchange-correlation energy Exc ] can be written as the sum of a kinetic part
Txc ] = T ] ; Ts ] =
(r)(vkin ( ] r) ; vs kin( ] r))dr
(7.22)
and an interparticle part 1 Z dr dr (r1 ) (r2 ) (g( ] r r ) ; 1) Wxc ] = 2 (7.23) 1 2 jr ; r j 1 2 1 2 the exchange-correlation potential can be split into a kinetic part that corresponds to the second resp line of equation 7.21 and a two-particle interaction part that corresponds to vscr + vscr :
xc xc (7.24) vxc ( ] r1 ) = Er ) = Tr ) + Wrxc) = txc( ] r1 ) + wxc( ] r1 ) (1 (1 (1 Z txc ( ] r1 ) = vkin( ] r1 ) ; vs kin( ] r1 ) + dr2 (r2 )( vkin ((r ]) r2 ) ; vs kin((r ]) r2 ) ) resp resp = vkin ( ] r1 ) ; vs kin( ] r1 ) + vkin ( ] r1 ) ; vs kin( ] r1 )
(7.25) (7.26) (7.27)
Z r 1Z r wxc( ] r1 ) = dr2 (r2 )(gj(r ]; 1 r2 ) ; 1) + 2 dr2 dr3 j(rr2) (rr3j) g( ](r 2) r3 ) r2 j ; 3 1 2 1 resp = vscr ( ] r1 ) + vscr ( ] r1 )
If the Kohn-Sham wavefunction s is regarded as a good approximation to the exact ground state wavefunction then it follows from the above formula that txc is very small and the major part of the exchange-correlation potential then originates from the second term wxc . In systems with strong near-degeneracy correlation, however, such as dissociating molecules, a one-determinantal wavefunction cannot be close to the exact wavefunction and in fact there are then regions (such as the bond-midpoint region 86]) where vkin deviates strongly from vs kin.
120
Figure 7.1: Radial density p(r) = 4 r2 (r) for Kr and Cd.
7.3 Structure of local potentials

We now proceed to investigate the relation between certain characteristics of the various local potentials de ned above, and therefore of vs and veff , and the atomic shell structure. The latter is described in a compact form with the radial Euler-Lagrange equation q q 1 d2 (7.28) (; 2 dr2 + veff (r)) p(r) = p(r) where p(r) = 4 r2 (r) is the radial density. Owing to (7.28), the atomic problem is e ectively 1 reduced to the one-dimensional problem of a single particle in the state p 2 (r) bound to the potential veff (r) with the energy . veff ; vext represents the repulsive barrier of the electron potential, which prevents localization of the electron density in the inner region and produces the true atomic density distribution p(r). Figure 1 displays the functions p(r) for Kr and Cd with the characteristic peaks in the shell regions. In Figures 2,3 the components of our partitioning of the potentials veff and vs are presented for Kr and Cd. All local potentials are extracted from the optimized potential model (OPM) 33, 91]. The solution of the OPM is equivalent to that obtained by the minimization of (7.3) with respect to with the restriction on the functions to be Slater determinants. As a result, veff in the 1 Euler-Lagrange equation for p 2 is approximated as resp resp veff (r) vext (r) + vOPM cond(r) + vOPM kin(r) + vOPM kin(r) + vOPM scr(r) (7.29) where all local potentials are constructed as de ned above with OPM being the optimal wavefunction of the OPM. The one-determinantal approach is well justi ed for atoms, so that in this case
STRUCTURE OF LOCAL POTENTIALS
121
resp Figure 7.2: veff ; vext and its components vcond , vkin and vkin . All potentials are derived from the optimized potential model, OPM 33]. a) Kr and b) Cd and c) Be
122
resp Figure 7.3: Demonstration of the step character of the screening-response potential vscr . a) Kr and b) Cd
(7.29) should be close to the exact expression (7.10). One can see from the Figures that partitioning (7.29) represents veff ; vext as a sum of the monotonous potential of Coulomb repulsion vOPM cond and three additional potentials. It is remarkable, and also pleasing in view of the desirability of accurate modelling, that the potential vcond is so smooth. The other potentials have distinct shell structure and have a very similar form for both represented atoms. Speci cally, vOPM kin has the resp peaks in the intershell regions that we anticipated, while vOPM kin is a step function, with the steps representing the energetical characteristics of the individual shells (see below). It is a quite intriguresp ing observation that the potential vOPM scr , which is particularly important for vs , also exhibits very clear step function behaviour (see g. 3). The shell structure of the potentials nds its interpretation in the formulas for the local potentials presented in this paper. Considering rst vOPM kin, it follows from eq. (7.15) that this potential is a measure of the change of the conditional probability amplitude as a function of position of the reference electron. The structure of the kinetic potential can readily be determined from physical considerations using the interpretation of the conditional probability amplitude. If the reference electron at r is positioned well within one atomic shell then as a consequence of the Pauli principle the probability will be large that the other electrons will have positions within the other atomic shells. If the position of the reference electron crosses an atomic shell boundary and moves into another atomic shell then the electrons in this new shell have to switch to another shell and the conditional probability amplitude will have to describe a large probability for a di erent distribution of the electrons over the atomic shells. So the changes in s as a function of the position r of the reference electron will be the largest at the atomic shell boundaries and we therefore expect vs kin to have local maxima at the atomic shell boundaries which is indeed what we nd. This explanation ts in with what is known about the behaviour of the exchange hole 87, 118, 119]. The exchange hole can be described by the Fermi amplitude or orbital, the square of which gives
123
Figure 7.4: The local exchange potential of the OPM as the sum of the screening or Slater (Fermi resp hole) potential vscr and its response vscr . a) Kr and b) Cd
124 in a closed shell system the hole density 87]:

N i (r1 )
Z

1 1 2 gs ( ] r1 r2 ) = 1 ; 2 jf (r(rr) )j 2
Z
f (r1 r2 ) =
i=1
1 2(
r1 )
i (r2 )
(7.30) (7.31)
As was described in 87], for various positions r within an atomic shell, or generally within a localized orbital, f (r1 r2 ), as a function of r2 , is localized within the shell (localized orbital) region and insensitive to the reference electron position r1 , but changing the reference position r1 through the small intershell interval produces a sharp change of f (r1 r2 ), its "jump" from one localization region to another. This may already be seen from the fact that orbitals i (r2 ) will get a large weight in the expansion of f (r1 r2 ), when the reference position r1 is such that the coe cients j i (r1)j=p (r1) are large, i.e. when r1 is in the region where i has a large amplitude (see 119]). Since the form of vOPM kin is determined by the rate r1 f (r1 r2 ) of the change of the Fermi orbital f (r1 r2 ) with changing r1 , the characteristic peaks of vOPM kin in the intershell regions (See Figure 2) re ect the maximal mobility of f (r1 r2 ) in these regions 86]. We now turn to the behaviour of vs kin near the atomic nucleus. Some properties in this region can be deduced from equation 7.15. The electron density has the following behaviour near the atomic nucleus 120]: (r) = (~ )(1 ; 2Zr + O(r2 )) (r # 0) 0 (7.32) where Z is the nuclear charge. The behaviour of the Kohn-Sham orbitals in the nuclear region can be deduced from the fact that close to the nucleus the only important potentials in the radial KohnSham equations are the nuclear potential ;Z=r and the centrifugal potential l(l + 1)=2r2 where l is the orbital angular momentum quantum number of the orbital. Inserting a series expansion of the orbitals i in the Kohn-Sham equation then yields: (7.33) i (r) = cl rl + O(rl+1 ) (r # 0) and in particlar for s-orbitals: 0 (7.34) i (r) = i (~ )(1 ; Zr + O(r2 )) (r # 0) By inserting the above expansions for the density of equation 7.32 and for the orbitals of equations 7.33 and 7.34 into equation 7.15 for vs kin it follows that the only terms which give a nite contribution to vs kin(~ ) are the terms containing orbitals with angular momentum quantum num0 ber l = 1,i.e. terms containing p-orbitals. So we can conclude that vs kin is in general not zero at the atomic nucleus except for the case that no p-orbitals are occupied in the Kohn-Sham wavefunction in which case vs kin(~ ) = 0. This has been illustrated in gure 2 where we nd that vs kin(0) is 0 zero for the case of the beryllium atom (no p-orbitals occupied) whereas it is nonzero for the case of cadmium and krypton. For comparison we have drawn the plot for the beryllium atom on the same scale as the one given in reference 107]. We further note that the proof given in reference 86] that the kinetic potential vkin of the exact ground state wave function is zero at the atomic nucleus is incorrect. The error is in the step from equation (A1) to (A2) in ref. 86]. The correct treatment of this step is given by Bingel 120]. For atoms vOPM kin (cf. eq. 7.15) represents the dominant part of vkin, so that going beyond the OPM, one can expect analogous intershell peaks of vkin due to the specially large mobility of the
vs kin( ] r1 ) = 1 jr1 s(s1 ~ 2 ::: ~ N jr1 )j2 ds1 d~ 2 :::d~ N = 2 jr1f (r1 r2 )j2 dr2 x x x x 1 2
125
conditional amplitude in these regions. It should however be remembered that in systems with strong nondynamical correlation, such as dissociating molecules, vkin will show additional structure, not present in vs kin, notably in the bond region 86].
resp Turning now to the kinetic response potential vOPM kin (See Figure 2), we note that its stepped form follows from its expression (7.16). Writing this in the form resp vOPM kin(r) =
X
2 ( ; i ) j i (r)j (r ) i=1
N
(7.35)
it is clear that for r within shell i, j i (r)j2 (or the sum over j i (r)j2 with the same i ) dominates over all other contributions, i.e. is approximately equal to (r), so eq.(7.35) is expected to describe a step function with step height in shell i equal to ; i , which is the energy needed to excite an electron to the "Fermi" level . resp The sum vOPM kin + vOPM kin, the Pauli potential vP , may be interpreted as a contribution to the barrier veff due to the e ect of the Fermi-Dirac statistics on the kinetic functional and its resp functional derivative 106]. vscr also brings a repulsive contribution to the barrier veff , but being resp of the order of a few a.u. for r < 1, this contribution is small as compared to those from vkin , vkin, though not negligible. There is a clear correlation between the maxima in veff due to the resp peaks and steps of vkin and vkin and the minima in the radial density, cf. 117]. As is immediately evident from equations (7.15) and (7.16), for all distances vP is a non-negative function 115]. One can see from Figure 2 that vcond is long-ranged (vcond has an asymptotic (N ; 1)=r behaviour), whereas both the kinetic potential and the kinetic response potential are short-ranged. At short radial distances vP (r) brings the dominant contribution to veff . We now focus on the Kohn-Sham potential and its constituents ( gures 3 and 4). In the OPM we are dealing with the situation that the Kohn-Sham potential does not have a kinetic contribution (cf. discussion following eq.(7.21)). If we subtract out from vOPM the Hartree potential, the exchange potential of the OPM is obtained
resp vOPM x = vOPM ; vext ; vOPM H = vOPM scr + vOPM scr
(7.36)
The two contributions are displayed in gure 4. It is interesting to observe that the screening potential is quite smooth. In this case vscr is purely due to the Fermi hole, there is of course no Coulomb hole in OPM . vscr is then equal to the Slater potential 101], the average Hartree-Fock exchange potential (using the OPM orbitals). The small intershell peaks 33] that are present in resp vx are evidently built in by the superposition of the stepped form of vscr on the smooth vscr . This observation may be expected to considerably facilitate accurate modelling of vx . The stepped form resp of vscr ( gure 3) does not follow immediately from the expressions given in this paper. A detailed analysis of the response of the screening factor, gs (r1 r2 )= (r3 ), is called for but will clearly be more involved than the treatment of the other potentials and is outside the scope of the present paper. Since we have restricted ourselves sofar to the OPM one-determinantal wavefunction, no information has been obtained on the e ect of dynamical (Coulomb) correlation. For some light atoms such as Be and Ne extensive con guration interaction calculations 82, 83] have yielded highly accurate electron densities, from which corresponding Kohn-Sham potentials have been generated 79, 80, 32]. The di erence vs ; vext ; vH is the exchange-correlation potential vxc = Exc = . The exchange
126
Figure 7.5: The correlation potential as the di erence between the exchange-correlation potential and the local exchange potential of the OPM. (a) Be and (b) Ne
127
part may be de ned as vOPM x, but an alternative and perhaps more consistent de nition of the exchange part of vs would be in terms of the exchange hole of the Kohn-Sham determinant s , so that the correlation part of vs would be vcorr ( ] r) = vs( ] r) ; vext(r) ; vH ( ] r) ; vx( ] r) resp resp = vscr ( ] r) + vscr ( ] r) + vkin ( ] r) + vkin ( ] r) resp resp ; (vs scr ( ] r) + vs scr( ] r) + vs kin( ] r) + vs kin( ] r)) (7.38) This may also be written as a sum of the kinetic part of vxc plus the correlation part wc , de ned as: wc( ] r) = wxc( ] r) ; vx( ] r) (7.39) Thus we have: vcorr ( ] r) = wc( ] r) + txc( ] r) (7.40) Probably vx is very close to vOPM x. Figure 5 shows the correlation part of vs as vs ; vext ; vH ; vOPM x for Be and Ne. Although we used a slightly di erent de nition of the correlation part of vs than Aryasetiawan and Stott 79], who took vx ( HF ] r) = vs( HF ] r) ; vext(r) ; vH ( HF ] r) where HF is the Hartree-Fock density, for the exchange part of vxc rather than vOPM x, our curves of vcorr are very similar to those of ref. 79]. It is notable that vcorr is quite small relative to vx . This may be due to a cancellation of the contributions from di erent terms in vcorr (7.38). For instance, to the extent that the OPM determinantal wavefunction may be identi ed with the Hartree-Fock determinantal wavefunction, the correlation contribution to the kinetic energy will R be (vkin ; vOPM kin)d~. This is known to be positive and equal to the absolute value of the r total correlation energy on account of the virial theorem, so the di erence vkin ; vOPM kin must be R predominantly positive. On the other hand, (1=2) (vscr ; vOPM scr )d~ is the contribution to the r correlation energy due to the electron-electron interaction. For a system like Ne this is negative and close to twice the (negative) correlation energy 119], so vscr will be predominantly more negative than vOPM scr . This is an expected e ect of the Coulomb hole that a ects vscr but not vOPM scr. At the border region between the K and L shells, where the exchange hole is not so strongly localized around the reference position as when the latter is within a shell, the e ect of the Coulomb hole is probably particularly noticable. This may explain the negative minimum in vcorr at the K/L shell boundary. It should be kept in mind that there are several contributions to vcorr , not only the negative Coulomb hole contribution vscr ; vs scr . Ref. 86] demonstrates that in the outer part of the (1s)2 shell of He the positive contributions of the response potentials (vN ;1 ) and the kinetic potential outweigh the negative Coulomb hole contribution vcond ; vHF vscr ; vs scr . Only at the inner part of the shell the Coulomb hole is so strongly negative that it causes vcorr to turn negative. This pattern is repeated in the K and L shells of Be and Ne, although in Be the total vcorr never becomes positive. This may be related to the well-known strong near-degeneracy correlation and therefore more signi cant Coulomb hole contribution in Be. The present paper analyzes the relationship between electron correlation and local one-electron potentials featuring in DFT, and establishes the connection with the atomic shell structure. The results of this analysis are also useful for the accurate and e cient approximation of vs and veff in resp DFT 121, 32]. In this respect, the stepped shell structure found for vscr is of special interest in
resp vx ( ] r) = E(srx = vs scr ( ] r) + vs scr ( ] r) )
(7.37)
128
our opinion. Further elucidation of the origin of this special structure will be undertaken, but it is clear that, now that it has been recognized, this simple step behaviour considerably facilitates the accurate modelling of the exchange-correlation Kohn-Sham potential. The corresponding work is in progress. Acknowledgements: This investigation was supported in part by the Netherlands Foundation for Scienti c Research (NWO).
Chapter 8
Step structure in the atomic Kohn-Sham potential

In this work we analyze the exchange-correlation potential vxc within the Kohn-Sham approach to density functional theory for the case of atomic systems. The exchange-correlation potential is written as the sum of two potentials. One of these potentials vxc scr is the long-range Coulombic potential of the coupling constant integrated exchange-correlation hole which represents the screenresp ing of the two-particle interactions due to exchange-correlation e ects. The other potential vxc scr contains the functional derivative with respect to the electron density of the coupling constant integrated pair-correlation function representing the sensitivity of this exchange-correlation screening to density variations. An explicit expression of the exchange-part of this functional derivative is derived using an approximation for the Greens function of the Kohn-Sham system and is shown to resp display a distinct atomic shell structure. The corresponding potential vx scr has a clear step structure and is constant within the atomic shells and changes rapidly at the atomic shell boundaries. Numerical examples are presented for the Be and Kr atoms using the Optimized Potential Model (OPM).
8.1 Introduction
The Kohn-Sham approach to density functional theory (DFT) 5] is an exact scheme to obtain the ground state properties of electronic many-particle systems by solving the problem of a system of noninteracting particles moving in the eld of an e ective Kohn-Sham potential vs ( ] r) which is a functional of the electron density . An important constituent of this potential is the exchange-correlation potential vxc( ] r) which is de ned as the functional derivative of the exchange-correlation functional Exc ]. With the appearance of so-called generalized gradient approximations (GGA's) 44, 45, 43, 49] increasingly accurate approximations for this exchangecorrelation functional have been applied with considerable succes in the calculations of properties of electronic systems ranging from atoms and molecules 71, 70, 122] to solids and surfaces 47]. However the corresponding exchange-correlation potential still has some de ciencies 49, 32]. In this paper we will perform an analysis of the exchange-correlation potential and in particular the exchange potential of the Kohn-Sham system. These potentials have two important characteristics. First of all the exchange potential vx and the exchange-correlation potential vxc have an asymptotic Coulombic ;1=r behaviour in nite systems such as atoms and molecules 31, 49, 32, 35]. Secondly, they display the atomic shell structure in atoms which is characterized by peaks at the atomic shell 129
130
CHAPTER 8. STEP STRUCTURE IN THE ATOMIC KOHN-SHAM POTENTIAL
boundaries 49, 35, 32]. In this paper we show that vx and vxc can be written as the sum of two potentials. One of these potentials is monotonous and has a long range Coulombic behaviour. The other is short range and displays a clear step structure being constant within atomic shells and changing rapidly at the atomic shell boundaries 108]. The step structure of the latter potential can be derived using an approximation of the Greens function of the Kohn-Sham system as we will demonstrate in the remainder of this paper. The required features of vxc and vx , i.e. the the Coulombic asymptotics and the atomic shell structure are not well-presented by most of the approximate exchange and exchange-correlation potentials. The required asymptotic Coulombic ;1=r behaviour in nite systems such as atoms and molecules is not reproduced with the current local density (LDA) and generalized gradient approximations. As a result of this a too low absolute value for the highest occupied Kohn-Sham orbital, which should be equal to the ionisation energy of the system 31], is obtained. Furthermore also the required atomic shell structure is not correctly reproduced with the LDA and GGA approximations 121]. The GGA potentials even have a wrong Coulombic behaviour near the atomic nucleus 92, 32]. An other approach leading to better exchange-correlation potentials is the so-called weighted density approximation (WDA) 123, 124, 125] which is based on approximate pair-correlation functions. In this approach the exchange-correlation potential is split up as the sum of two potentials, one containing the coupling constant integrated pair-correlation function which we will call the screening potential vxc scr and one containing the functional derivative of this function with respect to resp the electron density, which we will call the screening response potential vxc scr . The exchange part of this last potential is the main subject of this paper. Due to the fact that the approximate pair-correlation function is required to satisfy the sum rule and must integrate to one electron the corresponding exchange-correlation potential has a Coulombic asymptotic behaviour. However as all current approximate pair-correlation functions are not symmetric under interchange of the electron coordinates the Coulombic behaviour is in general not ;1=r as required but ;c=r with c some constant. Most of the approximate exchange and exchange-correlation potentials do not exhibit the atomic shell structure. Only a few attempts have been made to incorporate atomic shell structure in a WDA scheme 126, 127] employing the idea that electron interactions within one shell can be treated within a local density type of approximation but not the interaction between electrons in di erent atomic shells. This physical picture is re ected in the step potential, exchange and correlation e ects do not change over regions which lie well within an atomic shell but they do change considerably when we move from one atomic shell to another. A very good approximation to the exchange part vx of vxc has recently been proposed by Krieger et al. 35]. This approximation displays the atomic shell structure, has the correct asymptotics and also satis es the requirement of integer preference. This approximation was derived using an approximation for the Green's function of the Kohn-Sham system 34, 35]. In this paper we will use a similar approximation to derive an explicit expression of the functional derivative with respect to the electron density pair-correlation function of the Kohn-Sham system. Using this approximation resp one can derive an expression for the screening response part vx scr of the exchange potential. This response part is shown to display a clear atomic step structure, being constant within the atomic shells and changing rapidly at the atomic shell boundaries. The paper is divided as follows. In section II we will give a short account of the potentials we want to analyze. In section III we will derive the expression for the functional derivative of the Kohn-Sham pair-correlation function using an approximation for the Green's function of the Kohnresp Sham system. In section IV we derive how the step structure in vx scr arises from this functional
8.2. THE EXCHANGE-CORRELATION POTENTIAL: SEPARATION INTO A LONG RANGE AND A SHORT
derivative. In section V we will present some numerical results for some atoms and in section VI we present our conclusions and discuss implications for future approximations.
8.2 The exchange-correlation potential: separation into a long range and a short range part
Within the constrained search approach to spinpolarized density functional theory the energy functional is de ned as 5, 128, 67]
Ev f g] =
(r)v(r)dr + FL f g]
P
(8.1) is the sum of the spin densities and the universal (8.2)
where the total electron density = functional FL is de ned as: ^ ^ ^ FL f g] = ^ inf trfD(T + W )g
D!f g
^ In this functional the in mum of the expectation value of the kinetic energy operator T and the in^ is searched over all N -particle density matrices D which integrate ^ terparticle interaction operator W to the prescribed spin densities f g. The functional FL is usually split up as: 1 Z (r1 ) (r2 ) dr dr + E f g] FL f g] = TL f g] + 2 jr ; r j 1 2 xc (8.3) 1 2 where TL is de ned as: ^^ TL f g] = ^ inf trfDT g (8.4) As all functionals except the exchange-correlation functional Exc now are de ned equation 8.3 actually de nes Exc . Minimalization of the energy functional Ev leads to the well-known KohnSham equations: 1 (; 2 r2 + vs ( f g] r)) i (r) = i i (r) (8.5) where the spin density is given by: (r) =
X
D!f g
where the exchange-correlation potential vxc is de ned as: vxc ( f g] r) = Exc fr) g] (8.8) ( The exchange-correlation potential can be further analyzed in terms of the coupling constant integrated pair-correlation function g 1 2 127, 129] de ned as:
and the Kohn-Sham potential vs is split up as: Z (r vs ( f g] r) = v(r) + jr ; 1r) j + vxc ( f g] r)

1
fi j i (r)j2
(8.6) (8.7)
2 ( f g] r1 r2 ) =
1 2(
f g] r1 r2 )d
(8.9)
132 where g
1 2 1 2(
CHAPTER 8. STEP STRUCTURE IN THE ATOMIC KOHN-SHAM POTENTIAL

is de ned as:
g
with: ;
f f g] r1 r2 ) = ; 1 2 ((r ) g] (rr1 )r2 ) 1 1 2 2

X Z
(8.10) (8.11)
2 ( f g] r1 r2 ) =
3 ::: N
Df g (r1 1 r2 2 : : : rN N )dr3 : : : drN
Here Df g is the diagonal part of the density matrix which minimizes the functional FL for the ^ prescribed spin densities f g and in which the interparticle operator W is multiplied by the constant . The case = 1 corresponds to the ground state density matrix of the fully interacting system and the case = 0 corresponds to the density matrix of the Kohn-Sham system. A useful relation relating the paircorrelation function at coupling strength to the pair correlation function of the fully interacting system is proven by Levy 27]:
g
where:
1 2(
f g] r1 r2 ) = g 1 2 ( f 1= g] r1 r2 )
(8.12) (8.13) (8.14) (8.15) (8.16)
1= (r) = ;3 ( ;1 r)
The exchange-correlation energy in terms of g is given by: 1 X Z 1 (r1 ) 2 (r2 ) (g ( f g] r r ) ; 1)dr dr Exc f g] = 2 1 2 1 2 1 2 jr1 ; r2j 1 2 By functional di erentiation it follows that vxc can be split up into two terms:
resp vxc ( f g] r1 ) = vxc scr ( f g] r) + vxc scr ( f g] r)
where the screening potential is de ned as: Z X 2 (r2 ) (g v ( f g] r ) =

xc scr 1
jr1 ; r2 j
1 2(
f g] r1 r2 ) ; 1)dr2
2 3(
and the screening response potential as: 1 X Z 2 (r2 ) 3 (r3 ) g resp vxc scr 1 ( f g] r1 ) = 2 jr2 ; r3j 2 3
f g] r2 r3 ) dr dr 2 3 1 (r1 )
(8.17)
The screening potential is equal to the potential of the coupling constant integrated exchangecorrelation hole and represents the screening of the interparticle Coulomb potential between electrons at r1 and r2 with spin 1 and 2 by exchange-correlation e ects with screening factor equal to g 1 2 ( f g] r1 r2 ). Due to the fact that the exchange-correlation hole integrates to one electron the screening potential has a long-range Coulombic behaviour: (8.18) v ( f g] r) ; 1 (r ! 1)
xc scr
Physically this means that if we move one electron away from a nite system such as an atom it just experiences the potential of the ion it leaves behind.
THE EXCHANGE-CORRELATION POTENTIAL
133
resp We will carry out an analysis of the exchange part of vxc scr and vxc scr . We split up g into an exchange and a correlation part:
g 1 2 ( f g] r1 r2 ) = gs 1 ( f g] r1 r2 ) 1 2 + gc 1 2 ( f g] r1 r2 ) where gs is the pair correlation of the Kohn-Sham system:

2 gs ( f g] r1 r2 ) = 1 ; j s(r(r)1 r(2r)j ) 1 2
(8.19) (8.20)
where
s
(r1 r2 ) =
fi
(r1 ) i (r2 )
(8.21)
the one-particle density matrix for the Kohn-Sham system. The pair correlation function gs is una ected by the coupling strength integration as it satis es the following scaling property:
gs ( f g] r1 r2 ) = gs ( f 1= g] r1 r2 )
(8.22)
This is also apparent from the fact that the coupling constant integration keeps the density constant and therefore does not in uence the Kohn-Sham noninteracting system. The exchange-correlation potential vxc = vx + vc can now be written as an exchange and a correlation part with corresponding screening and screening response potentials:
resp vx ( f g] r) = vx scr ( f g] r) + vx scr ( f g] r)
(8.23)
and
resp vc ( f g] r) = vc scr ( f g] r) + vc scr ( f g] r) (8.24) where the potentials vx scr and vc scr and their responses are de ned as in equations 8.16 and 8.17 with g replaced by gs and gc . The potential vx scr has a Coulombic asymptotic behaviour due to the fact that the exchange hole integrates to one electron. There is no Coulombic term in vc scr as
coupling constant integrated Coulomb hole integrates to zero electrons. In gure 1 we plot some of the described potentials for the case of the beryllium atom. The exact exchange-correlation potential vxc is calculated from an accurate CI (Con guration Interacresp tion) density using the procedure described in reference 32]. The potentials vx ,vx scr and vx scr are calculated within the optimized potential model (OPM) 33, 35, 36] and are probably very close to their exact values which can be obtained from the solution for vx of the OPM integral equation 33, 35, 36] by insertion of the exact Kohn-Sham orbitals instead of the OPM orbitals. OPM resp We further plotted vxc ; vx scr which can be regarded as an approximation to vxc scr + vc scr . Note resp the very clear step structure in vx scrOPM . This potential is almost constant within the 1s-shell and drops rapidly to zero at the atomic shell boundary between the 1s and the 2s shell at a radial OPM distance of about 1 bohr. As can be seen from the graph of vxc ; vx scr this step structure is somewhat smoothed by correlation e ects but as these e ects are less important than the exchange e ects the step structure is still clearly visible. resp The atomic structure of the potential vx scr is the main subject of this paper. This requires an analysis of the functional derivative gs 0 (r1 r2 )= (r3 ) of the Kohn-Sham pair-correlation function. Such an analysis will be carried out in the next section.
134
STEP STRUCTURE IN THE ATOMIC KOHN-SHAM POTENTIAL
Figure 8.1: The exchange-correlation potential and the OPM exchange potential with corresponding screening and screening response parts for the beryllium atom
8.3 Functional derivative of the Kohn-Sham pair-correlation function

resp In order to understand the structure of the vx scr potential we must calculate the functional derivative of the Kohn-Sham pair-correlation function. This function describes the sensitivity of the exchange screening between two electrons at r1 and r2 to density changes at point r3 . One property of this function is readily derived. As gs (r1 r1 ) = 0 for any electron density it immediately follows that: gs 0 (r1 r1 ) = 0 (8.25) (r3 )
This puts a constraint on approximate functional derivatives of gs . In general from the de nition of gs it follows that:
gs 0 (r1 r2 ) = ;( (r )
3
1 s 0 (r1 r2 ) + c:c:) s 0 (r1 r2 ) 0 (r1 ) 0 (r2 ) (r3 )

0
;( (r1 (; r3 ) + (r2 (; r3) )(gs (r1 r2 ) ; 1) r1 ) r2 )
(8.26)
From constraint 8.25 it follows that the functional derivatives of s must contain deltafunctions in order to cancel the deltafunctions in the second part of the above equation for r1 equal to r2 .
FUNCTIONAL DERIVATIVE OF..
135
We now must calculate: N0 s 0 (r1 r2 ) = X fi 0 ( i 0 (r1 ) 0 (r2 ) + i 0 (r1 ) i 0 (r2 ) ) (8.27) (r3 ) (r3 ) i (r3 ) i We therefore have to calculate the functional derivative of the Kohn-Sham orbitals with respect to the density. This can be done using: Z i 0 (r1 ) = X i 0 (r1 ) vs 00 (r4 ) dr (8.28) (r3 ) vs 00 (r4 ) (r3 ) 4 00 where vs is the Kohn-Sham potential. The derivative of i with respect to the Kohn-Sham potential is known in terms of the Kohn-Sham orbitals and one-electron energies: i 0 (r1 ) = ;G (r r ) (r ) 0 (8.29) i 1 4 i 4 vs (r4) where Gi is the following Greens function: X j (r1 ) j (r4 ) (8.30) Gi (r1 r4 ) = ; The other functional derivative is equal to the inverse density response function which we only need to know for equal spins: vs (r4 ) = ;1 (r r ) (8.31) s 4 3 (r3 ) So: (r1 ; r2 ) =
Z
j 6=i
(r1 r4 ) ;1 (r4 r2 )dr4 s
(8.32) (8.33)
However as a density variation (r) determines the potential variation vs (r) only up to constant the derivative vs = and hence ;1 is only de ned up to constant. This arbitrariness can be lifted s for instance by specifying ;1 as an integral operator acting in the space of functions orthogonal s to the constant function. The constant function is an eigenfunction of the integral operator s with zero eigenvalue. This follows from the relation:
s
where s is the density response function: N X (r2 r4 ) = v ((r2 )) = ; fi i (r2 )Gi (r2 r4 ) i (r4 ) + c:c: s s r4 i
(r1 ) =
(r1 r2 ) vs (r2 )dr2
(8.34)
and from the fact that a constant potential variation vs (r) = C does not produce a density variation. Thus we have:
Z
(r1 r2 )dr2 = 0
(8.35)
This also immediately follows from equation 8.33 using the fact that Gi projects on the space orthogonal to i . From the above relation 8.35 it clearly follows that equation 8.32 has no unique solution for ;1 as adding a constant to ;1 gives another solution for equation 8.32. However s s
136
choosing a speci c constant (which amounts to choosing a gauge for the potential) cannot in uence the calculations. Keeping this in mind, we may write: Z i 0 (r1 ) = ; 0 G (r r ) (r ) ;1 (r r )dr (8.36) i 1 4 i 4 s 4 3 4 (r3 ) and it follows that:
s
Z N (r1 r2 ) = X ;f (r2 ) Gi (r1 r4 ) i (r4 ) ;1 (r4 r3 )dr4 + (1 $ 2) i i s (r3 ) i
(8.37)
As one can readily verify the left hand sides of the above equations indeed do not change when adding a constant to ;1 . To nd an explicit expression for the above functional derivatives we s must nd an expression for the inverse density response function ;1 . In order to do this we make s the following approximation to the Greens function (see Sharp and Horton 34], Krieger et al. 35] ): Gi (r1 r2 ) = 1 ( (r1 ; r2 ) ; i (r1) i (r2 )) (8.38) ~i where ~i is some mean energy di erence. Just as the exact Greens function of equation 8.30 the above approximate Greens function projects orbital i to zero which garantees condition 8.35. This approximate Greens function yields the following expression for the density response function s :
s
(r1 r2 ) = a (r1 ) (r1 ; r2 ) ; b (r1 r2 )

X
(8.39) (8.40)
where
a (r1 ) =
and
; 2f~i j i (r1 )j2 i

N
b (r1 r2 ) =
; 2f~i j i (r1 )j2 j i (r2 )j2

i
(8.41)
The inverse of this function is derived in the appendix using an approximation which xes the arbitrary constant (i.e. the gauge of potential vs ) so that vs ! 0 in in nity. It is given by:
N ;1 j (r )j2 j (r )j2 i 2 k 3 ;1 (r2 r3 ) = (r2 ; r3 ) + X s ik a (r )a (r ) a (r2 ) 2 3 ik
(8.42)
where the ik are given by: 2fi ;1 ik = ; ~ (I ; N )ik i for i k = 1 : : : N ; 1 and matrix N is given by: 2fk Z j i (r2 )j2 j k (r2 )j2 dr Nik = ; ~ 2 a (r2 ) k
(8.43)
(8.44)
FUNCTIONAL DERIVATIVE OF..
137
for i k = 1 : : : N ; 1. With the above equation for the inverse density response function and the approximate equation for the Greens function we nd using equation 8.36 that: Z i 0 (r1 ) = ; 0 ( (r ; r ) ; (r ) (r )) (r ) ;1 (r r )dr 1 4 i 1 i 4 i 4 s 4 3 4 (r3 ) ~i
j i (r4 )j2 ;1(r4 r3 )dr4 = ; ~ i (r1 ) s i The last integral can be worked out as:
Z
;1 (r1 r3 ) ; s
(8.45)
2 N ;1 j i (r4 )j2 ;1(r4 r3 )dr4 = j ai (r3 )j + s (r3 )

X
jk
jk
j i (r4 )j2 j j (r4 )j2 dr j k (r3 )j2 4 a (r ) a (r4 ) 3

(8.46) (8.47)
jk
= where
N ;1 X k
j k (r3 )j2 ik a (r ) 3
;1
ik = (I ; N )ik
We now de ne:
N N X r k (r ) 2 X Ki (r1 r3 ) = ; 1 ( (a1 (; r3 ) ; ik j a (r3)j + ~i r3 ) 3 k jk
j j (r1 )j2 j k (r3 )j2 ) a (r1 )a (r3 )
(8.48)
where we for notational convenience de ne ik and ik to be zero for i or k equal to N . This is a real function ( Ki = Ki ). We then have: i 0 (r1 ) = (r )K (r r ) 0 (8.49) i 1 i 1 3 (r3 ) and similarly: i 0 (r1 ) = (r1 )Ki (r1 r3 ) 0 (8.50) i (r3 ) This formula can be used to calculate the functional derivative of all explicitly orbital dependent functionals. Using the above expressions we nd:
s 0 (r1 r2 )
(r3 ) =
fi
X
(r1 )(Ki (r1 r3 ) + Ki (r2 r3 )) i (r2 )
(8.51)
We thus have: j s 0 (r1 r2 )j2 = 2 (r3 )
ij
fi fj
(r1 ) i (r1 )(Ki (r1 r3 ) + Ki (r2 r3 )) i (r2 ) j (r2 )(8.52)
Due to the 0 term we have: gs 0 (r1 r2 ) = gs (r1 r2 ) (r3 ) (r3 )
(8.53)
138 where

N gs (r1 r2 ) = ;2 X f f i j (r ) j
ij
(r1 ) i (r1 )(Ki (r1 r3 ) + Ki (r2 r3 )) i (r2 ) j (r2 ) (r1 ) (r2 ) (8.54)
;( (r1 (; r3 ) + (r2 (; r3) )(gs (r1 r2 ) ; 1) r1 ) r2 )
If we insert the expression 8.48 for the function Ki in the above equations we obtain: gs (r1 r2 ) = S (r r r ) + D (r r r ) (8.55) 1 2 3 1 2 3 (r3 ) where S is a part we will call, for reasons to be explained in the next section, the steplike part and D the deltafunction part given by: N N X 2 j (r1 ) i (r1 ) i (r2 ) j (r2 ) X 2 j k (r3 )j2 S (r1 r2 r3 ) = ;fi fj ~ ik a (r ) (r1 ) (r2 ) i 3 ij k + and:
X
ij
fi fj
N 2 j (r1 ) i (r1 ) i (r2 ) j (r2 ) X j k (r1)j2 + j k (r2 )j2 j l (r3 )j2 (8.56) kl a (r ) ~i (r1 ) (r2 ) a (r2 ) a (r3 ) 1 kl
" #
As we made an approximation for the Greens function in the derivation of gs 0 (r1 r2 )= (r3 ) one might ask whether this function satis es constraint 8.25. One can easily verify that S (r1 r1 r3 ) is not equal to zero unless one uses the additional approximation of Krieger et al. 35] that ~i = ~ independent of i. In this approximation the expressions for the functions h and S simplify to: 2 1 (8.59) h (r1 r2 ) = ; (r ) j s(r(r)1 r(2r)j ) = gs (r1 (r2 ) ; 1 r1 ) 1 1 2 and N N X X j k (r3 )j2 S (r1 r2 r3 ) = sr(r1 r2 ) ) i (r1 ) i (r2 ) 2 ik ( 1 ) (r2 i (r3 ) k
2N ; j s(r(r)1 r(2r)j ) 1 2 kl
X "
D (r1 r2 r3 ) = (r1 ;r3 ) h (r1 r3 ) ; gs (r1 (r2 ) ; 1 + (r2 ;r3) h (r2 r3 ) ; gs (r1 (r2 ) ; 1 (8.57) r1 ) r2 ) where h is de ned as: N (r ) (r ) 2 X f h (r1 r2 ) = a (r ) fi ~ j j (r1 )r i ) (r1 ) i 2 (rj ) 2 (8.58) (1 1 ij i 2
kl
j k (r1 )j2 + j k (r2)j2 j l (r3 )j2 (r1 ) (r2 ) (r3 )

#
(8.60) (8.61)
where:
ik = (I ; M
);1
8.4. THE STEP STRUCTURE IN THE EXCHANGE POTENTIAL
139
for i k = 1 : : : N ; 1 and zero otherwise. Matrix M is de ned as: Z 2j k 2 Mik = fk j i (r)j (r) (r)j dr (8.62) for i k = 1 : : : N ; 1. The deltafunction part in this approximation D unlike S does not satisfy D (r1 r1 r3 ) = 0, however it satis es this constraint in an integral sense:
Z
D (r1 r1 r3 )dr1 = 0
(8.63)
Some properties of the function S are readily derived from equation 8.60. Suppose that electrons at r1 and r2 are well-separated in di erent atomic shells with small overlap. In that case the one-particle density matrix s (r1 r2 ) will be small and consequently the function S will also be small. If on the other hand the two electrons are close together within the same atomic shell then, because S is exactly zero for r1 = r2 , the function S will also be very small. We can therefore expect that the largest contribution to S (r1 r2 r3 ) for a xed position r3 is obtained if one of the electrons is well within one atomic shell and the other electron is at an atomic shell boundary. The behaviour of S as a function of r3 is determined by factors of the form j k (r3 )j2 = (r3 ) which is approximately constant if r3 is within atomic shell k. The contribution of this factor to the total function S is determined by the constants ik describing the coupling of the density perturbation in shell k with an electron in shell i. These constants are the largest if i = k. Plots of the function S will be presented in section V for the case of the beryllium atom.
8.4 The step structure in the exchange potential

resp We now will derive an expression for the potential vx scr using the expression for the functional dervative of gs derived in the previous section. We will for generality not yet make the additional approximation ~i = ~ . We have: 1 Z (r1 ) (r2 ) gs (r1 r2 ) dr dr resp vx scr (r3) = 2 (8.64) jr1 ; r2 j (r3 ) 1 2 Inserting the expression for the functional derivative of gs we nd: Z (r1 ) (r2 ) S (r r r )dr dr + 1 Z (r1 ) (r2 ) D (r r r )dr dr (8.65) resp (r ) = 1 vx scr 3 2 1 2 3 1 2 2 1 2 3 1 2 jr1 ; r2 j jr1 ; r2j The part containing S yields: N 1 Z (r1 ) (r2 ) S (r r r )dr dr = X wl j l (r3 )j2 (8.66) 1 2 3 1 2 2 jr1 ; r2j a (r3 ) l where !# N Z f f j (r1 ) (r1 ) i (r2 ) (r2 ) " N X X j k (r1 )j2 + j k (r2 )j2 dr dr (8.67) i j i j wl = ;2 il + kl a (r ) 1 2 ~i jr1 ; r2j a (r2 ) 1 ij k
So the S term leads to a steplike function given by equation 8.66. Within shell l this function is almost constant and equal to the value wl given by equation 8.67. If we now denote: N Ex = ; 1 X Z f i (r2 ) k (r2 ) dr (r ) 1 (8.68) vi (r1 ) = f (r ) (r ) 2 k 1 k jr1 ; r2j i i 1 i 1 i (r1 ) k
140
which is equal to the orbital dependent potential within the Hartree-Fock approximation (except for the fact that we do not use Hartree-Fock orbitals) and further de ne P N X 2fi vi (r1 )j i (r1 )j2 = N 1i vi (r1 ) i (r1 ) i ~ W (r1 ) = ; ~ (8.69) PN 1 i (r1 ) a (r1 ) i i i ~i which is similar to the Slater potential 101] except for the energies ~i and where (8.70) i (r1 ) = fi j i (r1 )j2 then we can write: N N N ;1 2f X X X wl = 2f~i il vi + Wi = ; ~i (I ; N );1(Wi ; vi ) (8.71) il il i i i i i where
vi = vi (r1)j i (r1 )j2 dr1

Z
(8.72) (8.73)
and
Wi = W (r1 )j i (r1)j2 dr1 resp Now the deltafunction part of vx scr gives:
1 Z (r1 ) (r2 ) D (r r r )dr dr = W (r ) ; v (8.74) 1 2 3 1 2 3 x scr (r3 ) 2 jr1 ; r2j This is the di erence of two long range potentials both having a Coulombic ;1=r behaviour and it therefore decays faster than Coulombic. In the approximation ~i = ~ this di erence is exactly zero. So we nd: N ;1 2 resp (r1 ) = X wi j i (r1 )j + W (r1 ) ; vx scr (r1 ) vx scr (8.75) a (r1 ) i This gives for the exchange potential:
resp vx (r1 ) = vx scr (r1 ) + vx scr (r1 ) = W (r1 ) + N ;1 w X i i
j i (r1 )j2 a (r1 )

X
(8.76)
So we nd from this equation:
vx i ; Wi =
N ;1 X l
N ;1 X l
wl
j i (r1 )j2 j l (r1 )j2 dr = ; 1 ~i 1 a (r1 ) 2 fi
On the other hand it follows from equation 8.71 that
wl Nli
(8.77)
Nlk wl = 2f~k (Wk ; vk ) + wk k
(8.78) (8.79)
So we obtain:
1 vx i ; Wi = ;(Wi ; vi ) ; 2 f~i wi i
8.5. RESULTS
and we have:
141 (8.80)
(vx i ; vi ) i (r1 ) (8.81) 1 i (r1 ) ~i The rst term in this equation is the term derived by Sharp and Horton 34]. The second term was derived by Krieger et al. 35] in the additional approximation ~i = ~ . This additional approximation then leads to the exchange potential of Krieger, Li and Iafrate 35]:
So our nal exchange potential is:
wi = ; 2f~i (vx i ; vi )
i
vx (r1 ) = W (r1 ) +
N ;1 1 i ~i PN i
KLI vx (r1 ) =
If we multiply equation 8.81 by N ( ~i );1 i (r1 ) and integrate over r1 one can easily show that i our approximate vx satis es: vx m = vm (8.83) where m = N corresponds to the highest occupied Kohn-Sham orbital m . This equation is exactly valid within the so-called optimized potential model (OPM) 33, 36] exchange potential and KLI also for the approximate exchange potential vx 35] which follows directly from our derivation by putting ~i = ~ .
P
; vi (r1 ) i (r1 ) + NX1 (vx i ; vi ) i (r1 ) (r1 ) (r1 ) i
(8.82)
8.5 Results
In this section we present some numerical results obtained from the optimized potential model OPM (OPM) 33, 35, 36]. The OPM-orbitals and exchange potential vx ( f g] r) = vx ( f OPM g] r) are obtained by putting the correlation functional equal to zero and minimizing the exchange-only Kohn-Sham energy functional. This leads to an integral equation for the exchange potential which is equal to the functional derivative of the exact exchange functional evaluated at the OPM electron density OPM . OPM In gure 2 we plot the function gs (r1 r2 ) ; 1 for the case of the beryllium atom. As in the Be atom only s-shells are occupied this function only depends on the radial distance r1 = jr1 j and r2 = jr2j of electrons 1 and 2 from the atomic nucleus and not on the angle between vectors r1 and r2. This is a convenient feature for analysis. As Be is a closed shell atom gs is equal for up and down spin gs" = gs# . From gure 2 we can see that gs (r1 r2 ) ; 1 is close to ;1 if r1 are r2 within the same atomic shell. This is an e ect of the Pauli-principle, the probability that two electrons of the same spin are close together is small. If the two electrons are in di erent atomic shells ( the boundary between the 1s-shell and 2s-shell is at a radial distance of about one bohr ) then gs (r1 r2 ) ; 1 is close to zero. In gure 3 we plot the S part of equation 8.60 of the functional derivative gs (r1 r2 )= (r3 ) as a function of r2 and r3 = jr3 j for the Be atom where r1 = 0:1 bohr which is well within the 1s-shell. As we can see from this gure S is small when the other electron at radial distance r2 is either situated in the 1s-shell or in the 2s-shell. This function only becomes large when electron 2 crosses the boundary between the 1s- and the 2s-shell at a radial distance from the atomic nucleus of about 1 bohr. However this function is then only large when r3 is also located within the 1s-shell
142
Figure 8.2: The screening factor gs (r1 r2 ) ; 1 as a function of the radial distance r1 and r2 to the atomic nucleus of electron 1 and 2 for the beryllium atom
Figure 8.3: The functional derivative gs (r1 r2 )= (r3 ) as a function of the radial distance r2 and r3 to the atomic nucleus. Electron 1 is located at a distance r1 = 0:1 bohr from the atomic nucleus
8.5. RESULTS
143
Figure 8.4: The exchange potential with corresponding screening and screening response parts as calculated within the optimized potential model (OPM) and within the Krieger-Li-Iafrate (KLI) approximation for the beryllium atom
Figure 8.5: The exchange potential with corresponding screening and screening response parts as calculated within the optimized potential model (OPM) and within the Krieger-Li-Iafrate (KLI) approximation for the krypton atom
144
and within this shell S is almost constant as a function of r3 clearly displaying the step structure. OPM resp We now turn to the exchange potential. In gure 4 we plot vx scr and vx scrOPM and their KLI resp sum and the corresponding ones vx scr and vx scrKLI from selfconsistent solution of equation 8.82 KLI and vresp KLI are de ned as the rst and the second for the beryllium atom. The terms vx scr x scr term of equation 8.82. In gure 5 we present the same quantities for the krypton atom. As we OPM KLI can see from these gures the vx scr and vx scr are so close that they can not be distinguised on resp resp the scale presented. Also the step potentials vx scrOPM and vx scrKLI are very close, the largest di erence being constant within the 1s-shell. The most striking di erence between the OPM and KLI exchange potential is the smoothing of the intershell peak at the atomic shell boundaries in the KLI exchange potential. This di erence is not the most important di erence from an energetical point of view, as for the energy the atomic core region is the most important. This is most easily seen from the Levy-Perdew relation 22]:
Ex f g] =
vx (r)(3 (r) + r r (r))dr
(8.84)
which relates the exchange potential to the exchange energy. This relation is not exactly satis ed by the KLI exchange potential as an approximation is made to the functional derivative of Ex but the error is within 1% 35]. For approximate exchange potentials based on separate approximations resp for the vx scr and the vx scr part it is therefore important to obtain accurate values for the steps resp in vx scr in the atomic core region.
8.6 Conclusions
In this paper we analyzed the structure of the atomic Kohn-Sham potential. The exchangecorrelation potential was written as the sum of two terms, one term containing the coupling constant integrated pair-correlation function which represents the long-range potential of the exchangecorrelation hole, and one term, containing the functional derivative of the coupling constant integrated pair-correlation function, which is short ranged and displays a distinct atomic shell structure. An explicit expression for the exchange part of this functional derivative was derived using an approximation for the Greens function of the Kohn-Sham system. Properties of this function are analyzed and plots are presented for some atomic systems. It is shown that gs (r1 r2 )= (r3 ) is small when electrons at r1 and r2 are close together or when they are in the middle of the same or di erent atomic shells. This function is the largest when one of the electrons at r1 or r2 crosses an atomic shell boundary and as a function of r3 it is proportional to i (r3 )= (r3 ) which is the electron density of atomic shell i in which point r3 is located divided by the total electron density. As a function of r3 we therefore see a steplike behaviour, gs = (r3 ) is constant within the atomic shells and changes rapidly at the shell boundaries. This behaviour is induced in the short range part of the exchange potential. The other part of the exchange potential is monotonous, has a Coulombic long range behaviour and does not show any distinct atomic shell structure 108]. The natural splitting of the exchange and exchange-correlation potential in these two parts has important implications for obtaining accurate approximations for the Kohn-Sham potential. Di erent strategies can be developed to approximate the long and the short range part separately. This can improve one-electron energies, the ionisation energy and density dependent quantities such as dipole moments. A full discussion of the this is deferred to a subsequent paper 130].
8.7. APPENDIX:THE INVERSE DENSITY RESPONSE FUNCTION
145
8.7 Appendix:The inverse density response function
We will derive the expression for the inverse density response function ;1 . We rst split up ;1 s s as: ;1 (r2 r3 ) = (r2 ; r3 ) + c (r2 r3 ) (8.85) s a (r2 ) where c is a function to be determined. For this function we nd using equation 8.39 the following equation: (r1 ; r3 ) =
Z
(r1 r2 ) ;1 (r2 r3 )dr2 = s (8.86) (8.87) (8.88) (8.89) (8.90) (8.91)
b (r1 r3 ) + a (r )c (r r ) ; Z b (r r )c (r r )dr (r1 ; r3 ) ; a (r ) 1 1 3 1 2 2 3 2 3

So we nd: ) 1 Z c (r1 r3 ) = a b(r(r)1a r(3r ) + a (r ) b (r1 r2 )c (r2 r3 )dr2 1 3 1
If we further de ne: d (r1 r3 ) = a (r1 )a (r3 )c (r1 r3 ) we have Z r ( d (r1 r3 ) = b (r1 r3 ) + b (r1 a2 )(d )r2 r3 ) dr2 r2 If we insert the expression 8.41 for b we nd:
d (r1 r3 ) =
where
i (r3 ) =
Z
; 2f~i j i (r1 )j2 j i (r3)j2 +

i
; 2f~i j i (r1 )j2 i (r3 )

i
If we insert into the above formula the expression 8.90 for the function d we obtain a system of N equations for the functions i :
i
j i (r2 )j2 d (r2 r3 ) dr 2 a (r2 )

Nik (j
(r3 ) =
2fk Z j i (r2 )j2 j k (r2 )j2 dr Nik = ; ~ (8.93) 2 a (r2 ) k A general solution to the equations 8.92 can now be found. As the functions j i j2 are only nite in number the most general expression for the i is a linear combination of N of these functions:
i
where
(r3 )j2 + k (r3 ))
(8.92)
(r3 ) =
ik j k
(r3 )j2
(8.94)
146
N
This immediately gives an equation for the matrix ik :

ik = Nik +
X
Nil
lk
(8.95)
or if we denote the matrices by and N : = N (I ; N );1 = (I ; N );1 N = (I ; N );1 ; I (8.96) The above equations assume that the inverse of I ; N exists. This is necessary to nd a unique solution for the coe cients ik . However as discussed the inverse ;1 is only de ned up to a s constant and, as discussed below (I ; N );1 does not exist, we cannot specify ;1 uniquely unless s we make a special choice for this arbitrary constant. We return to this point after we have obtained our nal expression for ;1 . So we nd: s
d (r1 r3 ) =
where the ik are given by: 2fi (8.98) ik = ; ~ ( ik + ik ) i where ik is one of the solutions of equation 8.95. We then have our nal expression for the inverse density response function which is however still not unique:
N 2 2 ;1 (r2 r3 ) = (r2 ; r3 ) + X j i (r2 )j j k (r3 )j ik a (r )a (r ) s a (r2 ) 2 3 ik
ik
ik j i
(r1 )j2 j k (r3 )j2
(8.97)
(8.99)
Now we adress the question of the invertability of I ; N . From equation 8.93 it follows that:
X
so:
k
X
Nik = 1
(I ; N )ik = 0
(8.100)
From this equation it follows that the columns of matrix I ; N are linearly dependent and so this matrix has no inverse. Consequently the approximate density response function s has no unique inverse. As discussed this is due to the fact that inverse is only determined up to an arbitrary constant. From the equation of the inverse density response function 8.99 one can see that ;1 (r1 r2 ) goes to a constant if r1 or r2 goes to in nity. This constant is determined by the s orbital density of the highest occupied Kohn-Sham orbital j N j2 as it has the slowest exponential decay of all orbital densities. We can therefore x this constant by requiring that: lim ;1 (r r ) = 0 jr1 j!1 s 1 2 (8.102) This property can be satis ed by restricting the sum over the orbital densities in the equation 8.41 for the function b to the rst N ; 1 orbitals. In this approximation s has an inverse and the
(8.101)
8.7. APPENDIX:THE INVERSE DENSITY RESPONSE FUNCTION
147
derivation of this appendix can be carried out in a similar way. One nds for ;1 expression 8.99 s in which N is replaced with N ; 1 with matrix ik given by: 2fi ;1 (8.103) ik = ; ~ (I ; N )ik i which is a matrix of dimension N ; 1. Note that the approximation used here is equivalent to the approximation used by Krieger for the second part of the exchange potential in equation 8.82 where the summation is restricted to N ; 1 orbitals. This then leads to the ful lment of the exact constraint 8.83 by the approximate exchange potential.
148
Chapter 9
A self-consistent approximation to the Kohn-Sham exchange potential

A scheme of approximation of the Kohn-Sham exchange potential vx has been proposed, making use of a partitioning of vx into the long-range Slater vS and the short-range response vresp components. mod The model vresp has been derived from the dimensional arguments, which possesses the proper short-range behaviour and the characteristic atomic-shell stepped structure. When combined with mod OPM the accurate vS , vresp provides an excellent approximation to vx . With the GGA approximation mod to vS , vresp provides an e cient DFT approach which, for the rst time, ts closely the form of the accurate exchange potential and yields reasonably accurate exchange and total energies as well as the energies of the highest occupied orbital.
9.1 Introduction
E cient approximation of the Kohn-Sham exchange potential vx remains one of the important problems of the density functional theory (DFT) 35, 49, 131, 32, 121]. vx is de ned in the DFT as a functional derivative of the exchange energy Ex ] with respect to the density (r) vx(r) = Exr)] (9.1) ( In the one-electron Kohn-Sham equations (Hartree atomic units will be used throughout the paper) ; 1 r2 + vext (r) + vH (r) + vx(r) + vc(r) i(r) = i i (r) (9.2) 2 vx represents the local e ect of exchange, the dominant part of the electron correlation, while the external vext and the Hartree vH potentials represent the electrostatic interaction and vc is the e ective potential of the electron Coulomb correlation. Figure 1 illustrates the typical behaviour of vx with the exchange potential of the optimized potential model (OPM) 33, 91, 37] for Mg. vx is nite at the nucleus position vx (r) = const (r = 0) (9.3) it has the Coulombic asymptotics 33, 35, 57, 31] (jrj ! 1) (9.4) v (r) ! ; 1
x
jrj
149
150
A SELF-CONSISTENT APPROXIMATION..
Figure 9.1: The OPM exchange potential and its components for Mg and it possesses the shell structure, i.e. a non-monotonous dependence on r with the small peaks in the regions between the atomic shells 35, 33, 50, 93]. Taken together, these features of vx present considerable di culties for its e cient approximation. Recent progress of the DFT is due to the generalized gradient approximation (GGA) 50, 95, 44, 43, 47, 96] that gives a remarkable improvement to the exchange-correlation energies of the local density approximation (LDA) 11, 97, 98, 99]. It was shown 49, 32, 121, 92, 94], however, that the standard GGA exchange potentials do not possess the proper asymptotic behaviour and shell structure. The corresponding gradient corrections add little to the LDA exchange potential. For the analysis 108] and approximation 35, 121] of vx it appears very useful to separate the Slater potential vS 101] Z r r r v (~ ) = (~2 ) gx ( ] ~1 ~2 ) ; 1] d~ r r (9.5)
S 1
as an individual part of vx . In equation 9.5 gx is the exchange pair-correlation function, which can be expressed in terms of the occupied Kohn-Sham orbitals i (r) N N 1 X X i (r1 ) i (r2 ) j (r1 ) j (r2 ) gx ( ] r1 r2 ) = 1 ; 2 (9.6) (r1 ) (r2 ) i=1 j =1
j~1 ; ~2 j r r
9.2. A MODEL POTENTIAL VRESP
151
The separation of vS follows naturally from the expression of Ex in terms of vS 1 Z (r)v (r)dr Ex = 2 (9.7) S Di erentiation of equation 9.7 with a proper account of 9.5 represents vx as a sum of vS and the additional potential vresp, an integral of the linear "response" of g, g(r(1r3r)2 ) 6]
vx (r) = vS (r) + vresp(r) (9.8) The potentials vS and vresp have rather distinct, characteristic behaviour 121, 108] and Figure OPM OPM 1 clearly illustrates this with the vS and vresp components of the OPM exchange potential OPM 33, 91, 37]. vS is an attractive potential, which accumulates the Coulombic asymptotics vx (eqn. 9.4) of the total potential vx vS (r) ! ; j1j r
(jrj ! 1) (9.9)
It is rather smooth and does not display the pronounced shell structure. Contrary to this, vresp is repulsive and short-range. Remarkably enough, it exhibits very clear step function behaviour 108]. As it has been shown in 132], the steps of vresp originate from the corresponding stepped structure of the "response" g(r(1r3r)2 ) as a function of r3 . One can see from Figure 1 that the abovementioned small intershell peaks of vx are evidently built in by the superposition of the stepped form of vresp on the smooth vS . This specialized behaviour makes desirable the modelling of vx with the direct individual approxmod imation of vS and vresp. In this paper a model vresp is derived from the dimensional arguments, mod which represents vresp as the statistical average of the orbital energy contributions. vresp possesses the properties and closely reproduces the behaviour of the accurate potential. When combined with mod OPM the accurate functional vS , vresp provides an excellent approximation to vx . A more practical mod approximation is obtained with vresp and the GGA 44] approximation to vS . Contrary to the standard GGA schemes, the latter combination provides both the proper form and eigenvalues of the one-electron potential and the reasonable estimate of atomic Ex and the total energies Etot .
mod 9.2 A model potential vresp
As a starting point for our model we use an approximate equation of Krieger, Li and Iafrate OPM (KLI) 35] for the OPM exchange potential vx . OPM neglects the electron Coulomb correlation and its Kohn-Sham equations have the form 33, 35] 1 ; 2 r2 + vext (r) + vH (r) + vx(r) i (r) = i i(r) (9.10) The ground-state OPM wavefunction is the Slater determinant of the eigenfunctions i of 9.10, OPM OPM OPM so that the components vS and vresp of vx have the form 9.5 and 9.8 with gx built from OPM these functions. In 35] the very accurate approximation to vx was de ned with the following equation
KLI vx (r) = vS (r) +
X
2 wi j i(r)j (r) i=1

N
(9.11)
152
In equation 9.11 vS is the exact functional 9.5,9.6 and the second term is the statistical average of the orbital contributions wi , the latter being the di erence between the expectation values of the potential 9.11 and the Hartree-Fock exchange operator vxi for the orbital i
OPM wi = j i (r)j2 vx (r) ; vxi (r)]dr
Z N X vxi(~1 ) = ; (1~ ) r (~1 ) r j i r1 j =1 Z
(9.12)
2
r r r i (~2 ) j (~2 ) d~
j~1 ; ~2 j r r
(9.13)
OPM Note, that for the highest occupied orbital N the expectation values of vx and vxi are equal 35], thus providing the zero value of the corresponging parameter wN wN = 0 (9.14) mod The second term of equation 9.11 o ers a promising form for the model vresp . Because of the integral kernel 9.13, straightforward evaluation of 9.12 requires laborious calculation of the twoelectron integrals with the orbitals i . However, with the suitable approximation for wi one can develop an e cient model N 2 X mod (9.15) vresp (r) = wi j i(r)j (r) i=1 wi can be calculated, for example, as the orbital expectation values of some local potential vw
wi = j i (r)j2 vw (r)dr
(9.16)
chosen as the best local approximation of the di erence vx (r) ; vxi (r)]. In this paper an alternative approach is presented. We propose to use the model 9.15 with the orbital contributions wi being approximated by a function of the orbital energies of eq.2. The form of this function is chosen to mod provide the gauge invariance, the proper scaling and the short-range behaviour of vresp . According to the gauge invariance requirement, addition of a constant to the eigenvalues i should not alter wi values. To satisfy this requirement, we choose wi to be a function of the di erence ( ; i) wi = f ( ; i ) (9.17) where is the Fermi level of a given system, which is equal to the one-electron energy of the highest occupied orbital, = N . The exchange potential 9.1 and its components vS and vresp have the following scaling property vx ( ] r) = vx( ] r) (9.18) where (r) = 3 ( r) (9.19) while i has the scaling property 2 (9.20) i ] = i (r)] To provide 9.18, the function f from eq.24 should scale as follows f ( 2( ; i )) = f ( ; i ) (9.21)
A MODEL POTENTIAL vresp
153
and so we nd the square root of ( ; i ) to be the proper scaling function f p (9.22) wi = f ( ; i ) = K ] ; i By the de nition, 9.22 satis es the condition 9.14. Owing to this, the highest occupied orbital N mod does not contribute to the numerator of 9.15, thus providing the short-range behaviour of vresp
mod vresp (r) = K ]
X
vresp = kF (9.24) 2 where kF is the Fermi wavevector 3 kF = (3 2 ) 1 (9.25) mod Putting vresp of eq.30 to be equal to 9.24, one can calculate Kg . For the homogeneous electron gas
the Kohn-Sham orbitals and eigenvalues of eq.2 are given by k (r) = p1 ei~ r
~ k
As a result, our model potential 9.23 possesses the gauge invariance, the proper scaling and the short-range behaviour. K ] in equation 9.23 is a numerical coe cient, which can be determined from the homogeneous electron gas model. For the gas of a density the exact vresp of eq.9 has the form
p ; j i(r)j2 i (r) i=1

N
(9.23)
where V is a volume of the system and = k2 + vx ] + vc ] The Fermi level is given by

~ k F = k2 + vx ] + vc ]
(9.26) (9.27) (9.28) (9.29)
Inserting the above expression in equation 9.23, we obtain

X q mod k2 ; k2 vresp = pKg 2 V j~ j<kF F k
A replacement of the sum in equation 9.29 by an integral yields

Z k q F mod 2 vresp (r) = p Kg 3 kF ; k24 k2 dk = 2(2 ) 0 4 K Kg kF Z 1 p1 ; x2x2 dx = 3 p g k = p
2 2(2 )2 0 16 2 F From equations 9.24 and 9.30 the K value is de ned by p 8 2 0:382 Kg = 3 2 which is valid for the homogeneous electron gas of an arbitrary density.
(9.30) (9.31)
154
mod 9.3 A self-consistent scheme with vresp
We propose to use the model 9.23 within the self-consistent scheme, in which the Kohn-Sahm equations 9.10 are solved with the following approximate exchange potential mod mod vx (r) = vS (r) + vresp (K r) (9.32) where vS is the Slater potential 9.5 or its suitable approximation. The resulting Kohn-Sham orbitals are used to calculate the total energy, with the exchange energy being calculated from equation 9.7 with the self-consistent potential vS . The electron-gas value Kg of equation 9.31 can be chosen as mod the universal parameter of vresp for all systems. Another option is to determine K self-consistently from the requirement, that the Levy-Perdew relation 22]
Ex =
mod 3 (r) + rr (r)]vx (r)dr
(9.33) (9.34) (9.35)
should yield the same value of Ex as eq. 7 with the potential vS Z 1Z Ex = 3 (r) + rr (r)] vS (r) + Ksc ]R(r)]dr = 2 (r)vS (r)dr
R(r) =
p ; j i(r)j2 i (r) i=1

N
From the requirement 9.34 it follows the expression for K ]
I Ksc ] = ; I1
2
Z Z
(9.36) (9.37) (9.38)
I1 = I2 =
5 (r) + rr (r)]v dr S 2
3 (r) + rr (r)]R(r)]dr
In order to develop an e cient DFT scheme with 9.32, an appropriate approximation to vS of eq.9.5 is needed. One can use, for example, the weighted-density approximation (WDA) 125, 133, 123, 126] for the pair-correlation function gx , which guarantees the correct asymptotics 9.9 of vS . In this paper we use a more practical model of vS , which is obtained from the GGA of Becke 44, 121]. In 44] the exchange energy density x ( r)
Ex =
x(
r)dr
(9.39) (9.40) (9.41)

1 2
of the second order of gradient expansion 4 approx( r) = ; 3 (r) 3 ( 3 ) 1 + a ] 3 0 x 2 8 2 = 2 3 ( r4 (r) )2 3 (r) is modi ed with the correction function f of the argument 1 1 4 GGA ( r) = ; 3 (r) 3 ( 3 3 + a f ( 1 sinh;1 2 ) ] 0 2 x 2 8 sinh;1
1 2
(9.42)
9.4. A COMPARISON OF THE SELF-CONSISTENT POTENTIALS

mod vx (S ) mod vx (GGA)
155
ATOM
Be Ne Mg Ar Ca Zn Kr Sr Cd Xe 0.305 0.342 0.384 0.365 0.389 0.381 0.381 0.397 0.388 0.386 0.803 0.518 0.536 0.463 0.478 0.440 0.440 0.453 0.432 0.428
Table 9.1: Values of the parameter Ksc obtained within the self-consistent scheme for potentials mod mod vx (S) and vx (GGA) 1 (9.43) 2 1 + 6a0 sinh;1 1 Using equations 9.7 and 9.39, one can derive from relation 9.42 the corresponding approximation for vS GGA 1 1 GGA 3 3 vS (r) = 2 x (r() r) = ; 4 (r) 3( 83 ) 1 + 2a0 f ( 2 sinh;1 2 ) ] (9.44) This potential has the proper scaling 9.18, the correct asymptotics 9.9 and for the tted value a0 = 0:0042 it yields rather accurate estimate of Ex via the integral 9.7. The results of the selfconsistent calculations within the proposed scheme will be discussed in the next sections.
1 2 f ( 1 sinh;1 2 ) = 1 2
9.4 A comparison of the self-consistent potentials

In order to test the proposed self-consistent scheme, the comparative exchange-only atomic calculations have been performed for the closed-shell atoms Be, Ne, Mg, Ar, Ca, Zn, Kr, Sr, Cd, Xe. The exchange potentials vx compared can be subdivided into two groups, depending on whether GGA the accurate functional vS 9.5 or its GGA approximation vS 9.44 is used as the Slater potential mod in 9.32. The rst group includes the single vS (i.e. in this case vresp is neglected in 9.32), vx (S ) , mod OPM which is a sum of vS and vresp calculated in both variants with Ksc and Kg , and vx . The second GGA mod GGA mod group includes the single vS , vx (GGA) , which is a sum of vS and vresp calculated with either GGA Ksc or Kg , and the total exchange potential vx of the standard GGA approximation 92, 44]. The results are also compared with those obtained with the KLI potential 9.11. GGA To investigate a quality of the GGA approximation 9.44 to vS , in Figure 2 vS~ calculated selfGGA and vGGA obtained with the neglect of vresp are compared with consistently as a part of vx S OPM OPM vS calculated as a part of vx and vS obtained with the neglect of vresp for Ne and Mg. It is OPM interesting to note, that the neglect of vresp does not in uence on the Slater part of the exchange OPM potential, so that vS and vS are hardly distinguished from each other on the scale presented. GGA GGA Contrary to this, the neglect of vresp makes vS visibly more attractive than the corresponding GGA GGA part vS~ of vx . GGA One can also see from Figure 2 the appreciable local deviations of vS from vS . For both Ne and Mg the former is more attractive than the latter within the regions of 1s; and 2s; shells and less attractive in the intershell region. At longer distances (not shown here) all the potentials presented have the same Coulombic asymptotics 9.9. GGA The local deviations of vS from vS clearly manifest themselves in Table 1 where the atomic Ksc mod(S ) and vmod(GGA) are presented. For all atoms (with the exception of the lightest Be values for vx x and Ne) Ksc values obtained with the accurate functional 9.5 are rather close to the electron-gas GGA constant Kg = 0:382. However, substitution of vS for vS leads to a great overestimation of the
156
Figure 9.2: Slater potentials calculated self-consistently as the parts of the OPM and GGA exchange potentials and also obtained with the neglect of vresp. a) Ne and b) Mg
A COMPARISON OF THE SELF-CONSISTENT POTENTIALS

ATOM Be Ne Mg Ar Ca Zn Kr Sr Cd Xe 14.572 128.545 199.612 526.812 676.752 1777.834 2752.043 3131.533 5465.114 7232.121
resp resp resp OPM ;Etot vKLI vmod (Ksc) vmod (Kg ) neglect of vresp
157 11 44 79 109 146 258 288 324 419 450
0 0 1 2 2 4 4 4 6 7
0 1 2 3 4 6 5 7 6 12
0 0 2 4 4 6 5 7 6 11
OPM Table 9.2: Comparison of overestimates (in mHartrees) of the OPM total energies Etot 40, 36] (the latter are given in Ha calculated self-consistently with the exact functional vS and various approximations to vresp GGA integral I1 of 9.37. As a result, Ksc value for Be calculated with vS is 2.6 times as large as that calculated with vS . The overestimation decreases rapidly with the increasing atomic number. Still, GGA even for Xe Ksc for vS is about 1.1 times as large as that for vS . The analysis of the contributions to I1 shows, that the major part of the abovementioned overestiGGA mation comes from the region close to the nucleus where vS exhibits the largest deviations from vS (See Figure 2). The success of the GGA approximation is due to the fact, that the exchange enGGA ergy integrals 9.7 with vS are very close to those with vS , the typical error is only about 0.1show that not a high local quality of GGA approximation, but an incredibly precise cancellation of local errors for the integral 9.7 provides this success. For the integrals of the type 9.37 associated with the Levy-Perdew relation this balance is destroyed, which leads to the overestimated Ksc values (See Table 1). mod To analyse a local quality of the proposed model 9.23, vresp(S ) with the parameter Ksc and mod OPM KLI vresp(GGA) with Kg are compared for Ne and Mg in Figure 3 with vresp , vresp with the paramGGA , the latter has been obtained eters wi from 9.12 and also with the corresponding potential vresp GGA from the total exchange potential vx of the GGA 92, 44] by the subtraction of its Slater part GGA vS (eq.9.44). In this case GGA gives considerably worse approximation than in the case of the Slater potentials discussed above. Due to the inclusion of the uncompensated Laplacian terms, GGA vresp has incorrect Coulombic divergence at nucleus, being proportional to ; 1 32, 92]. At large r GGA distances it has incorrect Coulombic decay, being proportional to 1 . Furthermore, vresp does not r display the shell structure at intermediate distances, thus exhibiting large local deviations from OPM vresp . OPM Contrary to this, the simple model 9.23 provides a good t to vresp and to a more complicated mod mod KLI approximation vresp . Both potentials vresp(S ) and vresp(GGA) have the proper short-range behaviour OPM and they reproduce well a height and a width of the individual shell steps of vresp , the largest di erences being constants within 1s-shell. OPM In Figure 4 various self-consistent approximate exchange potentials are compared with vx . The OPM and the former can be considered as Slater potential vS is everywhere more attractive than vx a satisfactory approximation to the latter only at larger distances where both potentials have the same Coulombic asymptotics. At r < 1 a.u. the neglect of the repulsive potential vresp leads to the
158
OPM Figure 9.3: Comparison of vresp with various approximate response potentials. a) Ne and b) Mg
9.5. CALCULATIONS WITH THE ACCURATE SLATER POTENTIAL
159
substantial overestimation of the exchange e ect. The GGA 44] o ers a more balanced approximation to vx , though the abovementioned defects GGA GGA GGA of its components vS and vresp clearly manifest themselves in the total potential vx . In 12 and it is not attractive enough in the outer particular, it has incorrect long-range asymptotics ; r GGA region. However, due to the divergence of its component vresp at the nucleus and the overattracGGA in the region close to the nucleus, vGGA strongly overestimates the exchange tive character of vS x GGA e ect in this region. At intermediate distances vx smooths away the clear shell structure of OPM vx . The self-consistent scheme of Section produces potentials with a higher local quality of approximod OPM mod mation to vx . One can see from Figure 4 that vx (S ) with the components vS and vresp is OPM mod OPM at any distance and in the an excellent approximation to vx . vx (S ) is very close to vx major intervals they are even hardly distinguished from each other. The main di erence is that OPM the approximation 9.23 smooths away the small peaks of vx . However, this di erence does not seem the very important one, since these peaks appears to be a minor detail of the stepped shell OPM structure of vx (See Figures 1,4). mod(GGA) with the components vGGA and vmod shows a worse local quality of approximation, vx resp S GGA GGA mod mainly due to the defects of vS discussed above. Nevertheless, the replacement of vresp for vresp improves the short- and long-range asymptotics of the approximate potential and makes more clear mod its shell structure. As a result, vx (GGA) is nite at the nucleus and has the proper Coulombic asymptotics at longer distances (See Figure 4).
9.5 Calculations with the accurate Slater potential

Tables 2,3 present the total Etot and exchange Ex atomic energies calculated self-consistently with OPM KLI mod vS and various potentials vresp, namely, with vresp , vresp , vresp , the latter has been calculated in both variants with Ksc and Kg , and also with the neglect of vresp. In all these approaches the energies are calculated with the same functional 9.5-9.7, so that the corresponding energy differences are caused exclusively by the di erences in orbitals and densities generated with various exchange potentials. It is of no surprise, that the OPM Etot values 40, 36, 94] are always the least ones presented in Table 2. The OPM potential is de ned within the variational method and so it provides the true minimum of Etot calculated within the one-determinantal approach with the exchange functional 9.5-9.7. The inclusion of the proper vresp is of importance for a quality of calculated Etot . The neglect of vresp leads to the considerable errors, the error of Etot obtained with the single vS increases with the increasing atomic number and reaches 0.45 a.u. for Xe. OPM One can see from the Table, that the model 9.23 provides an excellent approximation to vresp as regards the total energy. It is not too sensitive to variations of the parameter K and calculations with either Ksc or Kg yield the same energies as a more complicated KLI approximation. Etot mod values obtained with vS and vresp are only by a few milliHartrees higher than those of the OPM. The corresponding error increases (though non-monotonously) with the increasing atomic number, but even for Xe it is about 0.01 a.u. It means that the addition of 9.23 considerably improves a quality of the Kohn-Sham orbitals and density, which becomes very close to those of the rigorous OPM. Because of the overattractive character of vS , calculations with the neglect of vresp yield too negative Ex (See Table 3). Addition of the approximate repulsive potentials vresp overcompensates this OPM 37] (the only exception is error and makes the self-consistent Ex values more positive than Ex
160
OPM Figure 9.4: Comparison of vx with various approximate exchange potentials. a) Ne and b) Mg
CALCULATIONS WITH THE ACCURATE SLATER POTENTIAL
161
ATOM Be Ne Mg Ar Ca Zn Kr Sr Cd Xe
OPM vresp vresp(Ksc ) vresp(Kg ) neglect of vresp ;Ex KLI mod mod
2.666 12.107 15.992 30.182 35.209 69.647 93.875 101.974 148.963 179.173
-1 9 10 8 5 78 65 58 123 119
2 5 11 41 40 55 88 59 150 218
14 30 9 63 30 60 90 14 123 199
-49 -200 -305 -436 -525 -1030 -1101 -1177 -1536 -1591
OPM 37] Table 9.3: Comparison of di erences (in mHartrees) of the OPM exchange energies Ex (the latter are given in Hartrees) an self-consistently with the exact functional vS and various approximations to vresp
resp resp resp ATOM ; OPM vKLI vmod (Ksc ) vmod (Kg ) neglect of vresp N Be 309 0 1 6 -17 Ne 851 2 21 30 -61 Mg 253 1 5 5 -31 Ar 591 2 18 21 -47 Ca 196 1 10 10 -29 Zn 293 1 -14 -14 -64 Kr 523 1 20 20 -44 Sr 179 1 12 10 -30 Cd 265 0 2 1 -65 Xe 456 1 23 22 -40
Table 9.4: Comparison of di erences of the OPM energies of the highest occupied orbital OPM (in N mHartrees) 40, 36] and those, calculated self-consistently with the exact functional vS and various approximations to vresp
162
KLI Ex for Be obtained with vresp ) and much more close to the latter. A comparison of Tables 2 and 3 reveals another type of compensation. In all cases the errors of Ex are compensated with those of the opposite sign in other parts of the total energy, so that Etot errors are much smaller than those of Ex . In most cases the Ex errors of the model 9.23 are somewhat larger (though of the
same magnitude) than those of KLI. The most important one-electron energetical characteristic of the Kohn-Sham theory is the energy N of the highest occupied orbital N . For the exact Kohn-Sham potential N is equal to minus the ionisation energy Ip of the system 134, 31], and in the exchange-only case N has the same meaning through the Koopmans's theorem 135, 35]. Table 4 represents N values obtained with vS and various vresp. It follows from the Table, that in the case of the orbital energies N the qualitative trends are the same as in the case of the exchange energies discussed above. In spite of its correct asymptotics 9.9, the single vS always has too negative N values. KLI mod Addition of vresp compensates this error. By the construction, both vresp and vresp decay exponentially in the region of N location, because N does not contribute into the numerators of 9.23 and the second term of 9.11, while contributing into the density in the denominators. The resulting mod KLI exponential tails of vresp and vresp produce the compensating repulsive contribution to N . Admod even has an overcompensating e ect, because the approximation 9.23 always (with dition of vresp the only exception of Zn) overestimates a value of the parameter wN ;1 for N ;1 , the next to the mod KLI highest occupied orbital. As a result, vresp becomes more positive than vresp in the outer valence region (See Figure 3) and yields more positive N values than those of KLI (the only exception is N of Zn), the latter are virtually the same as the OPM N . However, the corresponding errors are not large and in both variants with Ksc and Kg vary within 0.01 - 0.03 a.u. mod The present results show, that the self-consistent scheme with vS and vresp can be used as a very good approach to the OPM. The variants with Ksc and Kg yield results of the same quality, so one can use a more simple variant with the universal electron-gas parameter Kg for all systems. Because of the exact functional vS , the scheme requires calculation of the two-electron integrals with the orbitals i and the required computational time per iteration is approximately the same as KLI mod in the case of KLI. However, the replacement of vresp for vresp greatly accelerates the convergence of the self-consistent procedure. The ratio of iterations before the convergence in KLI and in the present scheme varies within 6-12 for the noble-gas atoms from Ar to Xe and within 12-18 for the alkaline-earth atoms from Mg to sr. As a result, the proposed scheme takes about an order of magnitude of the computational time as small as KLI. Still, to develop a practical DFT scheme, one should approximate not only vresp, but also vS . The results of calculations with the GGA approximation to vS will be presented in the next section.
9.6 Calculations with the GGA approximation to vS

Tables 5,6 represent Etot and Ex values obtained with the GGA approximation 9.44 to vS with and GGA mod without the response potentials vresp and vresp . All these energies are calculated with the same GGA standard GGA energy functional, for which the variationally derived vresp yields the minimal Etot . GGA OPM However, because of the approximate nature of this functional, Etot are not bound to Etot from below and in the most cases the former are too negative (See Table 5). On the other hand, similarly to Table 2, calculations with the neglect of vresp yield too positive Etot values. In the latter case the corresponding error increases monotonously with the increasing atomic number, the only exception is Zn.
CALCULATIONS WITH THE GGA APPROXIMATION TO VS
163
14.572 6 128.545 -45 199.612 -21 526.812 11 676.752 -2 1777.834 -287 2752.043 -60 3131.533 -48 5465.114 -79 7232.121 2
resp resp resp OPM ;Etot vGGA vmod (Ksc) vmod (Kg ) neglect of vresp
34 -30 3 33 27 -272 -39 -19 57 35
12 -29 -5 26 15 -268 -41 -23 60 25
21 35 78 141 149 33 269 298 364 463
OPM Table 9.5: Comparison of di erences (in mHartrees) between the OPM total energies Etot 40, 36] GGA and those, calculated self- vS and various approximations to vresp
OPM vresp vresp (Ksc ) vresp(Kg ) neglect of vresp ;Ex GGA mod mod
2.666 12.107 15.992 30.182 35.209 69.646 93.876 101.974 148.963 179.174
14 20 20 59 27 -76 76 61 126 167
44 -36 25 93 97 -138 95 91 189 314
-10 -122 -92 -13 -39 -303 -83 -128 -27 100
-73 -359 -406 -520 -606 -1375 -1277 -1332 -1688 -1714
OPM 37] Table 9.6: Comparison of di erences (in mHartrees) between the OPM total energies Ex GGA and those, calculated self-consistent vS and various approximations to vresp
164
resp resp resp ATOM ; OPM vGGA vmod (Ksc ) vmod (Kg ) neglect of vresp N Be 309 128 29 8 -15 Ne 851 396 128 98 9 Mg 253 104 19 6 -30 Ar 591 249 57 43 -23 Ca 196 80 15 6 -34 Zn 293 102 -19 -25 -74 Kr 523 218 42 33 -31 Sr 179 72 12 5 -38 Cd 265 95 -6 -13 -82 Xe 456 187 35 28 -34
Table 9.7: Comparison of di erences of the OPM energies of the highest occupied orbital OPM (in N GGA mHartrees) 40, 36] and those, calculated self-consistently with vS and various approximations to vresp
mod Addition of vresp considerably reduces the error, the only exceptions are Zn in both variants with OPM Ksc and Kg and Be in the variant with Ksc . Etot for Be obtained with Kg is much closer to Etot than that obtained with Ksc, because of the overestimation of Ksc in the GGA discussed above. In general, however, both variants yield the very similar Etot values, which are of the same accuracy as those of the standard GGA. The absolute error, induced mainly by the GGA approximation to vS , vary largely for di erent atoms, the typical value is of a few centiHartrees and the maximal (and the exceptionally large) one is 0.3 a.u. for Zn. GGA In the complete analogy with the case of vS , calculations with the single vS yield too negative Ex values (See Table 6). Addition of the approximate vresp considerably compensates the correGGA mod sponding error. In the case of vresp and vresp with Ksc this leads even to overcompensation and mod for the most atoms Ex are too positive, while in the case of vresp with Kg the compensation is not enough, thus producing too negative Ex (the only exception in the latter case is Xe). In general, all the schemes with the approximate vresp yield Ex values of a comparable accuracy. GGA mod The self-consistent scheme with vS and vresp shows a de nite advantage over the standard GGA in calculation of N (See Table 7). As it was indicated in 36, 32] and one can see this from the Table, the GGA greatly underestimates the absolute magnitude of N . Due to the incorrect GGA asymptotics of its response part (See Figure 3), j N j values of vx are about twice as small as compared with the OPM values and are very close to the LDA ones. On the other hand, the single GGA vS overestimates j N j (the only exception is Ne), though the corresponding errors are considerably smaller. mod Addition of vresp with Ksc to vS overcompensates this e ect and produces comparable errors of the GGA mod opposite sign. The N values of the potential (vS + vresp ) obtained with Kg are in most cases the best approximate ones and the closest to OPM . N GGA mod It follows from the above analysis that the self-consistent scheme with vS and vresp provides the same accuracy for the total and exchange energies as the standard GGA scheme 44] and considerably improves the form and the eigenvalue N of the one-electron potential. Bearing in mind mod GGA mod high quality of the presented results, we propose vx with the components vS and vresp as an e cient approximation to vx . The variant with Kg provides a better t to the accurate vresp and it is also simpler than that with Ksc , so the electron-gas value Kg can be recommended as the universal parameter for the many-electron calculations.
9.7. CONCLUSIONS
165
9.7 Conclusions
In this paper a scheme of approximation of the Kohn-Sham exchange potential vx has been promod posed, making use of a partitioning of vx into the Slater and response potentials. The model vresp has been derived from the dimensional arguments, which possesses the proper short-range behaviour mod and the characteristic atomic-shell stepped structure. When combined with the accurate vS , vresp OPM . With the GGA approximation to vS , vmod provides provides an excellent approximation to vx resp an e cient DFT approach which, for the rst time, ts closely the form of the accurate exchange potential and yields reasonably accurate exchange and total energies as well as the energies of the highest occupied orbital. Still, we have to mention appreciable errors, which are introduced with the GGA approximation to vS . For example, the typical error of the calculated total energies is increased from milliHartrees to GGA santiHartrees with the replacement of vS for vS . So, in order to provide a better practical DFT scheme, one should improve, rst of all, a quality of vS approximation. In particular, the present GGA approximation shows considerable deviations from the accurate vS in the region close to the mod nucleus. A promising way of vS re nement is to construct it not as a function of and r , but as some function of the orbital densities j i j2 and their gradients. mod To test a quality of the developed vx approximation, the exchange-only atomic calculations have been performed in this work. Our main goal, however,is to apply this approximation for mod the exchange-correlation molecular calculations. For this purpose one can use vx either as an appr independent exchange part of the approximate exchange-correlation Kohn-Sham potential vxc or mod as the basic functional form for approximation of the total vxc. Within the former approach vx is inserted in the Kohn-Sham equations 9.2 together with some approximation for the Coulomb correlation potential vc, the latter is to be constructed independently. Then, the exchange energy mod is calculated via 9.7 with the self-consistent density and vS , while the correlation energy is calculated with the independent functional, which corresponds to vc . An alternative approach is based on the fact, that vx is the dominant part of vxc and the exchange energy density x is the dominant part of xc. To approximate the total vxc and xc , one can use the mod mod same models vS and vresp with other parameters chosen to t the available accurate exchangecorrelation Kohn-Sham potentials 32, 74, 78, 79, 80]. The work along both abovementioned lines is in progress. Acknowledgements: This investigation was supported in part by the Netherlands Institution Fundamenteel Onderzoek der Materie (FOM).
166
Chapter 10
An analysis of nonlocal density functionals in chemical bonding

In this work we carry out an analysis of the gradient corrected density functionals in molecules that are used in the Kohn-Sham density functional approach. We concentrate on the special features of the exchange and correlation energy densities and exchange and correlation potentials in the bond region. By comparing to the exact Kohn-Sham potential it is shown that the gradient corrected potentials build in the required peak in the bond midplane, but not completely correctly. The gradient corrected potentials also exhibit wrong asymptotic behaviour. Contributions from di erent regions of space (notably bond and outer regions) to nonlocal bonding energy contributions are investigated by integrating the exchange and correlation energy densities in various spatial regions. This provides an explanation why the gradient corrections reduce the LDA overbinding of molecules. It explains the success of the presently used nonlocal corrections, although it is possible that there is a cancellation of errors, too much repulsion being derived from the bond region and too little from the outer region.
10.1 Introduction
In the last few years considerable progress has been achieved within density functional theory 5] in the calculation of molecular 70, 71, 122] and solid state 47, 136, 137] properties by the use of nonlocal density functionals. The use of these nonlocal corrections gives an overall improvement of the results obtained by the local density approximation (LDA). This is especially true for energy quantities such as bonding energies in molecules. The successes are not restricted to simple covalent molecules (cf. the G1 and G2 sets 70, 122]) but also occur for more complicated cases where the conventional ab-initio Hartree-Fock method fails notoriously such as in multiply bonded molecules and particularly in the eld of transition metal chemistry 71]. Although numerical calculations clearly show that the nonlocal corrections do work, not much understanding has been gained in why and how they work. There are several questions left unanswered. Why do the nonlocal corrections always reduce the bonding energies of molecules? What do molecular Kohn-Sham potentials look like and do the nonlocal corrections give a good representation of these potentials? How can the correlation e ects (left-right correlation) in dissociating molecules be incorporated in gradient corrected density functionals and their potentials? Why does the Becke correction for exchange, that has been devised to yield good (Hartree-Fock) exchange energies in atoms, change LDA bond energies not towards Hartree-Fock but (close) to exact bond energies? An understanding of such 167
168
AN ANALYSIS OF NONLOCAL DENSITY FUNCTIONALS
questions might lead to the construction of still more improved density functionals which is not only of theoretical but also of practical importance. In the literature little attention has been paid to the question why the nonlocal corrections improve the energies. It has been argued 70, 138] that the improved bond energies due to the Becke functional 44] are due to the improvement of the asymptotic behaviour of the exchange energy density. However in a clear paper by Engel et al. 49] it was shown for the case of atoms that the correct behaviour of the Becke exchange energy density is only reached at very large distances where it has almost no e ect on the energy. Moreover these authors constructed a functional which did not satisfy the correct asymptotic property but which yielded even better atomic exchange energies. Whether the correct asymptotic behaviour of the Becke functional is responsible for the improved bond energies is therefore an open problem which we will investigate more closely in this paper. The paper is divided in the following way. In section 10.2 we discuss the properties of the exact molecular Kohn-Sham potential based on calculations of the hydrogen molecule. We then discuss how the left-right correlation shows up in the potential. In section 10.3 we discuss the properties of the potentials of two widely used nonlocal functionals namely those of Becke 44, 139] and of Perdew 45]. It is shown that these potentials are characterised by a peak in the bond midpoint region. In section 10.4 we analyse the exchange and correlation energy densities of the Becke and the Perdew functionals and give an explanation for the fact that the nonlocal corrections reduce the bonding energy in molecules. In section 10.5 we then give a numerical analysis of the bonding contributions of the nonlocal functionals. This is done by integrating the exchange and correlation energy densities of Becke and Perdew over speci c spatial regions of some diatomic molecule. We nally present a summary and conclusions in section 10.6.
10.2 Left-right correlation and the exact Kohn-Sham potential

In order to improve on existing density functionals it is very helpful to know how molecular KohnSham potentials look like. Especially the knowledge on the correlation potential is very limited. There is however one molecular system for which the correlation potential can be accurately calculated and in which correlation is very important. This sytem is the dissociating hydrogen molecule. For a general molecular system with N electrons the Kohn-Sham equations are given by: 1 (; 2 r2 + vKS ( ] r)) i (r) = i (r) (10.1) where the electron density is given by the sum of orbital densities: (r) =
X
N i
j i (r)j2
nuc
(10.2)
The Kohn-Sham potential vKS is usually split up in the following way: Z (r0 ) dr0 + v ( ] r) KS ( ] r) = v (r) + v
The rst term in this equation is the Coulomb attraction of the atomic nuclei and the second term is the repulsion of the electronic cloud. The last term, the one we are interested in, is the exchangecorrelation potential which is the functional derivative of the exchange-correlation functional. For a
jr ; r0 j
xc
(10.3)
LEFT-RIGHT CORRELATION AND THE EXACT KOHN-SHAM POTENTIAL
169
spin restricted two-electron system such as the hydrogen molecule in its ground state the occupied p orbitals can explicitly be expressed in the electron density as i = =2 . Substituting this in the Kohn-Sham equations then yields an explicit expression of the Kohn-Sham potential expressed in the exact ground state density: 2 2 1 (10.4) vKS (r) = ; 8 (r 2) + 1 r + 4 For a two-electron system is the highest occupied Kohn-Sham orbital and therefore equal to the ionisation energy of the system. The above equation gives us the possibility to calculate the Kohn-Sham potential from a correlated electron density which can for instance be obtained from an accurate con guration interaction (CI) calculation. This has been done by Smith et al. 140] for the helium atom. A di erent approach to obtain the Kohn-Sham potential has been used by Buijse et al. 86] who carried out a thorough investigation of this potential for the hydrogen molecule at various distances. For comparison in the next section we stress some points of their results. The basic quantity in the analysis of Buijse et al. is the conditional probability amplitude (r2 : : : rN jr1 ). de ned by: (r (r2 : : : rN jr1 ) = p1 : : : rN ) (10.5) (r1 )=N where is the ground state wavefunction of the system. The amplitude gives a description of the (N ; 1)-electron system when one of the electrons is known to be at position r1 . The amplitude squared gives the probability that the other electrons are at positions r2 : : : rN if one electron (the reference electron) is known to be at position r1 . All the correlation e ects of the system are contained in this quantity. These correlation e ects especially show up in the case of the hydrogen molecule when we move the position of the reference electron along the bond axis from one atom to the other atom. Due to the Coulomb correlation of the electrons we have that if one electron is known to be at a given atom then the probability amplitude is large that the other electron will be at the other atom. This changes rapidly when the reference electron crosses the bond midpoint region because in that case the other electron has to switch quickly from one atom to the other atom. A quantity which measures this change in conditional amplitude (the so-called left-right correlation e ect) as we move our reference electron is the so-called kinetic potential de ned by: 1Z vkin(r1 ) = 2 jr1 j2 dr2 : : : drN (10.6) This potential is clearly positive de nite. The origin of the name of this potential stems from the fact that it contributes to the kinetic energy of the system which is given by: 1 Z (r )2 dr + Z (r)v (r)dr T=8 (10.7) kin This follows directly from the de nition of the kinetic energy and formula 10.5. Note that the rst term in the above formula is the von Weiszacker kinetic energy TW ] which for a two-electron system is equal to the Kohn-Sham kinetic energy Ts ]. The term with vkin therefore contributes to the Kohn-Sham exchange-correlation energy of the system. As has been shown in reference 86] vkin is also an important constituent term of the Kohn-Sham potential. By numerical calculation the following properties of the kinetic potential were observed 86]. First of all this potential is peaked on the bond midpoint. As a result of this the total exchangecorrelation potential is also peaked in this region. This peak in vkin is not surprising from its de nition. The changes in the conditional amplitude as we move our reference electron is the largest
170
in the bond midpoint region as explained before. Secondly the peak in the kinetic potential gets higher when we dissociate the molecule. This is also what one expects since the left-right correlation e ects grow stronger as we dissociate. This can also be seen from the well-known Hartree-Fock error for a dissociating system. The Hartree-Fock one-determinant wavefunction upon dissociation no longer gives a right description of the division of the electrons over the atomic fragments by putting too much weight on ionic terms 17]. Knowing the behaviour of the Kohn-Sham potential in the bonding region of molecules one might wonder whether these peculiar properties are well represented by the presently used nonlocal gradient corrections. We will show in the next section that both the Becke and Perdew potentials exhibit peaks in this region in cases of strong left-right correlation. The Becke potential exhibits a positive peak in this region although the Becke functional is purely derived as an exchange correction and not a correlation correction. The Perdew potential however, which is derived as a correlation correction, exhibits a negative peak although not large enough to cancel against the peak in the Becke potential. The peaks in both potentials are also shown to grow upon dissociation of the molecule.
10.3 Analysis of Beckes and Perdews molecular potentials

In this section we will analyse the properties of both Beckes exchange potential and Perdews correlation potential. We will start with Beckes exchange potential. The Becke gradient correction to the exchange energy which belongs to the family of Generalised Gradient Approximations (GGA) 43, 49] is de ned as:
GGA Ex
]=
4 3 f (x
)dr
4
(10.8)
in which is a spin index and x = jr j= 3 . The function f is in the case of the Becke functional de ned as:
f (x) = ;
The constant is given by = 0:0042. This function is chosen so as to satisfy some scaling, GGA asymptotic and weak inhomogeneity properties 32]. The potential corresponding to Ex is given by (for convenience we leave the spin index out, it simply can be added to the nal result): 2 df x GGA 3 vx (r) = Er) = 4 1 (f (x) ; x dx + x2 d f )+ ( 3 dx2
X
x2 1 + 6 x sinh;1 x
(10.9)
ij
@i @i @j @j ( df ; x d2 f ) ; r2 df jr j3 dx dx2 jr j dx
(10.10)
We will now analyse this potential in the vicinity of the bond midpoint region. In that case the gradients are small and our parameter x is therefore also small. We therefore have:
f (x) ; x2 + O(x4) df 3 dx ;2 x + O(x )
(x # 0) (x # 0)
(10.11) (10.12)
hence:
AN ANALYSIS OF BECKES AND PERDEWS MOLECULAR POTENTIALS

and
171 (10.13) (10.14)
Inserting this in the potential formula 10.10 yields:

2 GGA vx (r) = 2 r 4 + O(jr j2 )
3
d2 f dx2
;2 + O(x2 )
(x # 0)
This yields in the bond midpoint rM , where r = 0, for the Becke potential:
B vx (rM ) = 0:0106 r 4
3
(10.15)
for the spin unpolarised case ( = 2 ). In this derivation we only used the property f (x) x2 (x # 0). This requirement originates from the gradient expansion for slowly varying densities of the exchange energy, for which the gradient correction in the so-called Gradient Expansion Approximation (GEA) is given by: Z Z 2 4 3 x2 dr E GEA ] = ; (r ) dr = ; (10.16)
x
4 3
This is the gradient correction used by Herman et al. in the X -method 76]. The corresponding potential is given by: 2 2 GEA (10.17) vx (r) = 2 r 4 ; 4 (r 7 ) 3 3 3
4 which is in accordance with equation 10.14. We will analyse the r2 = 3 term a bit further. In order to do this we consider the case of a homonuclear diatomic molecule. In the dissociation limit the molecular density can be well approximated by the sum of atomic densities = A + B of atom A and B . If we use the asymptotic relation A B Ne; rA B (r ! 1) where rA B is the distance to atom A B we obtain 2 (r ! 1) (10.18) r2 = 2 ; 2 ; 2
and for the height of the peak in the potential in the bond midpoint at large internuclear distance RAB of the atoms:
GGA vx (rM )
rA
rB
AB
K ! 1 (R ! 1) AB 1
3
(10.19)
where K is a positive constant. We thus conclude that the peak increases at large distances and even has the unphysical behaviour of becoming in nite in the limit of in nite separation of the atoms. However at these distances the density decreases faster ( e; r ) than the peak increases so the e ect of this erroneous behaviour of the GGA exchange potential in the bond midpoint region on the molecular density may still be very small at large bond distances. At shorter bond distances we cannot make any de nite conclusions on the behaviour of the Becke (and Perdew) potential from the above analysis. For instance from formula 10.14 we can deduce that the sign of the peak depends on the sign of the Laplacian. In the weak bonding limit this sign is positive, however at smaller distances this sign depends strongly on the type of bond considered
172
which has extensively been studied in reference 102]. We will next discuss the bond midpoint properties of Perdews correlation potential. The Perdew correlation energy functional is de ned by 45] : Z 4 2 ; ( jr j) 1 GGA ] = 3 dr (10.20) Ec d( ) C ( ) x e where = ;
0
is the spin polarisation and: (10.21)
The function C ( ) has been obtained from the gradient expansion for slowly varying densities of the correlation energy of the electron gas 42] and is usually expressed in the Wigner-Seitz radius 3 rs = ( 4 3 ) 1 as: are given by 0:023266 7:389 10;6 8:723 0:472. We nally have: ( jr j) = 1:745f~C ((1)) jr7 j (10.23) C 6 ~ where f = 0:11. The form of this function is obtained from the wavevector analysis of Langreth and Mehl 50]. For slowly varying densities the Perdew functional reduces to the gradient expansion result for the correlation energy (we consider the spin unpolarised case for which d( ) = 1 ): The constants
2 :002568 + s C ( ) = 0:001667 + 10+ r + +r2 rs 104 rr3 +
s s s
5 1 3 3 d( ) = 2 1 (( 1 + ) 5 + ( 1 ; ) 3 ) 2 2 2
(10.22)
EcGEA
] = C ( ) 3 x2 dr
4
(10.24)
Both this functional and the Perdew functional have the same values for their functional derivatives in the bond midpoint region where r = 0. We have:
2 GGA vc (rM ) = ;2C ( ) r 4
3
(10.25)
In the dissociation limit where the bond midpoint densities are small we have C ( ) 0:001667 and thus:
GGA vc (rM ) = ;0:0033 r 4
3
C (0) =
(10.26)
This has the same form as Beckes potential in the bond midpoint region, however with a di erent sign. We therefore nd a negative peak in the Perdew potential which increases in absolute value when the molecule dissociates. However, the total nonlocal potential which is the sum of Beckes and Perdews potentials still has a positive peak in the bond region since the negative peak in Perdews potential is not large enough to cancel the positive peak in Beckes potential. We have for the total nonlocal exchange-correlation potential:
NL Becke Perdew (r ) = 0:0073 r vxc (rM ) = vx (rM ) + vc M 4
3
(10.27)
10.4. GRADIENT CORRECTED ENERGY DENSITIES
173
So in the bond region in the dissociation limit we have a repulsive Becke potential and an attractive Perdew potential. We can now ask the question how this behaviour of the Becke-Perdew potential compares with the exact potential. Let us compare equation 10.27 in the case of a two-electron system with the exact Kohn-Sham potential of equation 10.4 in the bond midpoint region. In the bond midpoint we have: 1 2 vKS (rM ) = 4 r + (10.28) In contrast to the potential in equation 10.27 the bond midpoint value of the exact Kohn-Sham potential does not go to in nity if (RAB ! 1) but to the positive constant ; (see also 86]) which p can easily be derived from formula 10.18 using = 2 ;2 . This shows that the Becke-Perdew potential cannot give a good description of the bond midpoint behaviour of the exact potential at very large bond distances. However for intermediate bond distances it can give an improvement of the LDA potential but this improvement is only of qualitative nature as the height of the peak in the Kohn-Sham potential in the LDA+Becke-Perdew approximation has a di erent functional dependence on the density than the exact Kohn-Sham potential. In section 10.5 we will present some numerical examples of the molecular Becke and Perdew potentials. However we will rst address the question of bond energies and analyse the exchange and correlation energy densities in the bonding region.
10.4 Gradient corrected energy densities

The positive peak in the potential already indicates that adding the Becke-Perdew nonlocal corrections to the local density calculations may in weak bonding cases reduce the bonding energy of molecules, i.e. give a positive contribution to E de ned as the molecular energy minus the atomic energies. We will try to identify the origin of the positive contribution due to the nonlocal corrections by considering the di erence between the molecular energy density and the sum of atomic energy densities. The exchange and correlation energy densities can be de ned as: 1 (10.29) x (r) = 2 (r)Vx (r) and 1 (10.30) c (r) = 2 (r)Vc (r) where Vx and Vc are the potentials due to the coupling constant integrated Fermi and Coulomb holes 24] which should not be confused with the potentials vx and vc which are the functional derivatives of the exchange and correlation functionals. There are however other de nitions possible of Vx and Vc . One could for instance de ne Vx to be the potential due to the Fermi hole of the optimised potential model (OPM) 33, 36] and de ne Vc as the di erence between this potential and the potential due to the coupling constant integrated exchange-correlation hole. In either de nition the potential Vx goes asymptotically like ; 1 due to the fact that the Fermi hole integrates to one r electron. The exchange potential potential vx has a similar asymptotic decay. The contribution of the nonlocal corrections to x and c we will in the following denote by ex and ec .
174
The contribution to the bonding energy E of a molecule of the nonlocal corrections to the exchange and correlation functionals can be written as:
NL Ex =
Z
ex (r)dr ec (r)dr
X
(10.31) (10.32)
and
EcNL =
where and
ex (r) = eM (r) ; x ec (r) = eM (r) ; c eM x eM c
A
X
eA(r) x eA (r) c
(10.33) (10.34)
in which and denote the gradient corrections to the molecular exchange and correlation energy densities.The functions eA and eA denote the gradient corrections to the atomic exchange x c and correlation energy densities where the atoms are situated at their molecular positions. In equations 10.33 and 10.34 the sum is taken over all atoms in the molecule. We will analyse the energy density di erence functions ex and ec for the Becke and Perdew functional. In the bond midpoint we have that the gradient of the molecular density is zero, r M = 0, and in that case the Becke and Perdew energy density functions are given by:
ex (rM ) =
and
4 A3
A
X
xA2 >0 1 + 6 xA sinh;1 xA
(10.35) (10.36)
ec (rM ) = ;
We will show in the next section for the example of N2 that ex does have a positive peak in the bond midpoint region and ec has a negative peak in the same region. Suppose that the main contributions to the functions ex and ec are found in the bond midpoint region of the molecule. In that case we would have:
NL Ex =
Z
C ( A) d( A )
A 4 xA 2 e; ( A j A j) < 0 3
ex (r)dr > 0 ec (r)dr < 0
(10.37) (10.38)
and
EcNL =
So we can infer that in the bond midpoint region the Becke functional gives a positive correction to the bond energy and the Perdew functional gives a negative contribution. The main reason for the improvement of the bonding energy of molecules by the nonlocal corrections is then the replacement of a region of nonzero gradients in the sum of atomic exchange-correlation densities by a region of zero gradients in the exchange-correlation energy density of the molecule.
10.5. NUMERICAL PROCEDURE AND RESULTS
175
There is also another region from where we may expect a positive contribution from gradient corrections to the LDA bond energy, namely the outer region where the density decays exponentially to zero. It is well known that the LDA exchange-correlation energy density also decays exponentially to zero, whereas it should behave asymptotically like ; =2r. From any gradient correction that builds in the correct asymptotic behaviour, as does the Becke correction, one may expect a positive contribution to the bond energy from the asymptotic region. In that region, where
ex (r) ; 2r (r ! 1) we have for ex that: ex (r) ; 2r +

P
(10.39) (r ! 1) (10.40)
2r = ; 2r
in which = M ; A A is the deformation density of the molecule. If the deformation density is negative in the outer regions of the molecule then: This then contributes positively to Ex. As may be inferred from elementary considerations regarding the relation between the exponential decay of molecular and atomic densities and the rst ionisation energies, is usually negative in the outer region of the molecule, but it is also quite small in that region. On the other hand the region is very large, so it is di cult to make a quantitative estimate of this e ect. If the e ect is signi cant we expect the Becke functional to provide a positive contribution from the outer region, but maybe not large enough since it attains the asymptotic behaviour at too large distances 49, 32]. To analyse the properties of the functions ex and ec and determine from which region in the molecule the bonding energy correction originates (bond midpoint region or outer asymptotic region) we numerically integrated the functions ex and ec in di erent spatial regions. The procedure and results will be discussed in the next section.
ex (r) = ; 2r > 0
(r ! 1)
(10.41)
10.5 Numerical procedure and results

In this section we will carry out a numerical investigation of the energy density di erence functions ex and ec . In order to do this we performed several self-consistent density functional calculations on some dissociating diatomic molecules using the Becke and Perdew nonlocal corrections. The bonding contribution analysis is carried out using integration schemes in two di erent types of grid. The rst scheme uses an integration in prolate spheroidal coordinates in which we integrated ellipsoidal regions around the molecular bond midpoint and in which the atomic positions coincide with the foci of the ellipsoid. If the atoms A and B are situated on the z -axis at positions z = a then we can de ne the coordinates u v by:
x = a sinh (u) sin (v) cos ( ) y = a sinh (u) sin (v) sin ( ) z = a cosh (u) cos (v)
(10.42)
176
As the functions ex and ec are invariant for rotations around the z -axis the -integration just yields a factor 2 . We de ne further a variable p = cos ( ). The Jacobian in this coordinate system is given by:
J (u p) = a3 sinh (u)(cosh2 (u) ; p2) Ik (R) = 2

Z
(10.43) (10.44)
We then want to calculate the following functions:

U
where k stands for exchange k = x and correlation k = c and in which R = a sinh (U ) is the minor axis of the ellipsoid. These functions which represent the bonding energy contribution of the exchange and correlation energies within an ellipsoid around the bonding axis are calculated using Gauss-Legendre integration in the variables u and p. The second integration scheme uses a cylindrical coordinate system de ned by:
du
1 dpJ (u p) ek (u p) ;1
x = r cos ( ) y = r sin ( ) z = z
(10.45)
If the atoms are again situated at positions z = a on the z -axis then r represents the distance to the line through the bonding axis which is the z -axis and describes the rotation angle around this axis. For this coordinate system we want to calculate the following functions:
Ik (Z ) =
The subindex k again stands for exchange and correlation k = x c and the Jacobian J is given by J (r z) = 2 r. The functions Ik are again obtained using Gauss-Legendre integration. Using these two integration schemes it is possible to locate the regions in the exchange- and correlation energy density di erences ex and ec which have the largest contributions to the bonding energy. Using the elliptic integration we can compare the contribution of the region outside some ellipsoid to the contribution of the inside and thus determine the importance of the outer asymptotic region. Using the integration in cylindrical coordinates we can determine the importance of the bond midpoint region. In the following we carry out an analysis of the Becke and Perdew functionals in the nitrogen molecule N2 . The nitrogen molecule is a typical case of a molecule where the Becke correction gives a lowering of the bonding energy and where the Perdew correction gives a smaller increase in the bonding energy. Most molecules (for instance all rst and second row diatomic molecules) exhibit this behaviour. A notable exception is the hydrogen molecule H2 for which both the Becke and Perdew corrections have a di erent sign. For the nitrogen molecule an LDA calculation at the experimental bond distance of 2.07 bohr yields a bonding energy of 11.5 eV which is to be compared to the experimental bonding energy of 9.9 eV. A self-consistent calculation at the same bond distance using the Becke correction gives a bonding energy of 10.0 eV and a Becke-Perdew calculation NL NL yields 10.2 eV. We therefore have Ex = 1:5eV and Ec = ;0:2 eV at equilibrium distance. In these calculations we used a triple-zeta Slater Type Orbital (STO) basis set plus polarisation and as atomic reference we used spin unrestricted nitrogen atoms in which we occupied (according to Hunds rule) the 2p-shell with three electrons with the same spin. The atomic exchange energy density is then given by a sum of di erent contributions from - and -spin electrons. Our results
;Z
dz
dr ek (r z)J (r z)
(10.46)
NUMERICAL PROCEDURE AND RESULTS
177
Figure 10.1: Becke and Perdew potential around the bond midpoint region for the dissociating nitrogen molecule at bondlengths of 3.0,5.0,7.0 and 9.0 bohr
178
are in substantial agreement with the completely numerical results of Becke 141]. In gures 10.1a and 10.1b we display the Becke and Perdew potential of the nitrogen molecule along the bond axis around the bond midpoint region for several bond distances. In gures 10.1c and 10.1d we display the same potentials along an axis perpendicular to the bond axis through the bond midpoint. As we can see from these gures the Becke peak is larger than the Perdew peak at each distance and the peaks are increasing in absolute value as we dissociate the molecule. We can also see that the width of the peak in the direction perpendicular to the bond axis is much larger than the width of the peak along the bond axis. The total potential (Becke+Perdew) therefore builds a repulsive wall between the atoms. In gures 10.2a an 10.2b we display the exchange and correlation energy density di erence functions ex and ec for the Becke and Perdew functional along the bond axis for the equilibrium distance of 2.07 bohr and the larger distance of 3.0 bohr. In gures 10.2c and 10.2d we display the same functions along an axis through the bond midpoint perpendicular to the bond axis. As we can see from gures 10.2a and 10.2c the function ex is positive in the bond midpoint region as was expected from equation 10.35. This positive region arises because the gradient of the molecular density is zero at the bond midpoint. As the gradient of the sum of atomic densities is also zero in the bond midpoint this positive region in ex is not the result of density changes due to bond formation but merely a result of the topological properties of the molecular density. In general ex is positive whenever the gradient of the molecular density vanishes. Due to the A1-symmetry of the molecular density this often happens in points which belong to the invariant manifold of the molecular symmetry group. In gures 10.2a and 10.2b we can also see peaks in ex around the nuclei. As the gradient of the sum of atomic densities in the region of one atom is close to the gradient of the density of the atom itself in this region these peaks can only be explained by density changes in the atomic region due to bond formation. This point is illustrated in gure 10.3a which displays the deformation density due to bond formation along the bond axis. We can see from this gure that the gradients of the deformation density are the largest around the nuclei. We have therefore identi ed two possible contributions to the Becke correction to the bonding energy. There is a contribution from the region around the nuclei which arises from density changes induced by bond formation and there is a contribution from the bond midpoint region which arises from the shape of the molecular density. The relative importance of these regions will be discussed below. Having discussed the properties of ex it is clear that the shape of ec as displayed in gures 10.2b and 10.2d can be explained in a similar way. The negative peak at the bond midpoint is due to the zero gradient in the bond midpoint (see equation 10.36) and the peaks closer to the nuclei are caused by density changes in the atomic regions. After having discussed the local properties of ex and ec in di erent regions of space we will now investigate how much the various regions contribute to the bonding energy at equilibrium bondlength using the integration schemes described in the beginning of this section. In gure 10.4a we display the functions Ix and Ic of equation 10.44 as a function of the length R of the minor axis of the ellisoid which has the atoms positioned at the foci. As we can see from this gure half of the nonlocal correction to the bonding energy due to exchange is obtained for R = 1 bohr. The major axis then has a length of 1.44 bohr which therefore extends 0.40 bohr beyond the atoms. More than about 80% of the bonding energy contribution is obtained at R = 1:5 bohr with major axis of length 1.82 bohr which extends 0.79 bohr beyond the atoms. The ellipsoid is then just enclosing the valence region of the atoms. We therefore conclude that it is mainly from the inner region of the molecule and not from the asymptotic region that the Becke correction originates, the precise ratio of the contributions depending of course on the admittedly somewhat arbitrary
179
Figure 10.2: The exchange and correlation energy density di erences of Becke and Perdew around the bond midpoint region
180
Figure 10.3: The deformation density of the nitrogen molecule at the equilibrium bondlength of 2.07 bohr de nition of these regions. Let us rst consider the asymptotic region somewhat more closely. The relatively small contribution from the asymptotic region, where the expression of equation 10.41 for ex holds, may be related to gure 10.3b which displays the deformation density along an axis through the bond midpoint perpendicular to the bond axis. From this gure we see that when the deformation density is negative (which is the case in the outer asymptotic region) it is very small and therefore leads to a small asymptotic ex . However, the asymptotic contribution is not negligible and we may even wonder whether the division of the exchange energy over the inner and outer region of the molecule given by the Becke functional re ects the true situation as there are clear indications that the Becke functional underestimates the importance of the outer region. It has been demonstrated by Engel et al. 49] for the case of atoms that the exchange energy density of Becke reaches its Coulombic behaviour much further from the atomic nucleus (beyond 10 bohr) than the exact exchange energy density which reaches this behaviour just past the outermost orbital. Also the Becke exchange potential has a much too small contribution in the asymptotic region 49, 32] which is re ected in the fact that the LDA one-electron energies are not improved by the Becke potential 32]. The consequences of the too fast decay of the Becke exchange potential for the asymptotic contribution to the exchange energy can be inferred from the following de nition of exchange energy density. We de ne the Levy-Perdew exchange energy density eLangrethPerdew:wavevector 22] in terms of the x exchange potential vx and the density by: eLP (r) = vx (r)(3 (r) + r r (r)) (10.47) x This exchange energy density integrates to the exchange energy, which follows from a scaling property of the exchange functional 22]. (10.48) As the LDA+Becke exchange potential is much smaller in the outer region of the molecule than the exact exchange potential it gives a too low contribution from this region. As a corollary, the
Ex ] = eLP (r)dr x
181
Figure 10.4: Contributions of the Becke and Perdew functional to the bonding energy in the elliptic (a) and cylindrical (b) coordinate system as a function of the minor elliptic axis (a) and the Z coordinate (b) at the equilibrium bondlength of 2.07 bohr roughly 20% of the total positive nonlocal contribution to the bond energy that comes from the outer region (in spite of the small in that region) is certainly too low. Let us next consider the inner region of the molecule. In this region there are two contributions to the exchange energy, there is a contribution from the bond midpoint region and a contribution from the atoms. To determine the relative importance of the atomic regions and the bond midpoint region we integrated ex using the cylindrical coordinate system. In gure 10.4b the functions Ix and Ic of equation 10.46 are plotted as a function of Z . The integration regions are thus slabs of thickness 2Z perpendicular to the bond axis centered around the bond midpoint. For Z = 1:035 bohr the slab is just touching the nuclei. We can see from gure 10.4b that at Z = 0:5 bohr which is about halfway the distance from the bond midpoint to the nucleus about half of the Becke correction to the exchange energy is obtained. If we denote the region Z > 0:5 bohr (somewhat arbitrarily) as atomic region then we can conclude that about half of the Becke correction is due to a replacement of high atomic gradients by a region of low molecular gradients (equation 10.35) around the bond midplane and half is due to changes in the density on the atoms as a result of bond formation. Note that the peaks in ex around the atoms in gure 10.2 are only seemingly more important than the bond midpoint peak in ex as they are only restricted to a small area around the nucleus whereas the bond midpoint peak also has a considerable extension in the direction perpendicular to the bond axis as demonstrated in gure 10.2c. About the same conclusions as for Beckes exchange correction can be drawn for Perdews correlation correction. As can be seen from gure 10.2b the function ec has a somewhat larger extension than ex into the outer region of the molecule. However from gures 10.4a and 10.4b we still can conclude
182
that the bond midpoint region (Z < 0:5 bohr) contributes about half of the nonlocal correlation correction to the bonding energy. We may wonder, as we did for the asymptotic region, whether the nonlocal corrections from the inner region are given correctly by the Becke-Perdew functional. The analysis of section 10.3 indicates that the Becke-Perdew potentials are overestimated in the bond midpoint region and therefore probably also the positive contribution to the bond energy. This ts in with the suggestion above that the contribution from the asymptotic region may be underestimated, since the total Becke-Perdew correction yields quite accurate bond energies. The conclusion that the contribution from the inner region is overestimated is supported by the fact that for solid state bulk systems (which of course exhibit no asymptotic region) the Becke correction often overshoots and leads instead of the overestimation by LDA to an underestimation of the cohesive energy 136, 137].
10.6 Conclusions
In this work we analysed the working of some gradient corrected density functionals in the case of chemical bonding. From an analysis of the exact Kohn-Sham potential based on the work of Buijse et al. 86] it follows that the exact exchange-correlation potential is peaked in the bond midpoint region. This fact can be related to left-right correlation in the dissociating molecule. It is shown that the Becke-Perdew potential also exhibits a peak in the bond midpoint region but with a wrong functional dependence on the density which makes this peak go to in nity at in nite separation of the atoms instead of going to a constant. Then we showed that one important reason that gradient corrected density functionals reduce the bonding energy in overbinded LDAmolecules is that these functionals replace a region of high gradients in the sum of atomic exchangecorrelation energy densities by a region of low gradients in the molecular exchange-correlation energy density around the bond midpoint. This region of zero gradients arises purely due to the topological properties of the molecular density and is not related to density changes upon bond formation. We further argued that probably the relative importance of the bond midpoint contribution as compared to the asymptotic contribution to the total nonlocal correction is not represented completely correctly by the Becke-Perdew exchange-correlation energy densities. It appears that the positive contribution from the bond region is overemphasised particularly in weakbonding (near dissociation) situations, in keeping with the overestimation of the bond midpoint peak in the potential in weak bonding situations. On the other hand the contribution from the asymptotic region is probably underestimated, in keeping with the too fast decay to zero of the Becke-Perdew potentials 32]. In order to improve upon existing density functionals and to obtain further insight in the relation between electron correlation and Kohn-Sham potentials we feel one should closely investigate the relation between the structure of the exchange-correlation potential (shell- and bond midpoint peaks, asymptotic Coulombic behaviour) and conditional amplitudes. The kinetic potential in atoms for instance measures the in-out correlation e ect and induces some structure in the atomic correlation potential, just as it re ects the left-right correlation in the peak at the bond midpoint. It will obviously be very helpful to have exact molecular Kohn-Sham potentials available not just for the two-electron H2 system but also for general many-electron molecules. This may be feasible with the procedure proposed in ref. 32]. From there one might endeavour to construct very accurate exchange-correlation potentials from the density and its derivatives. Work along this line is in progress.
Chapter 11
Exchange and correlation: a density functional analysis of the hydrogen molecule

In this work we analyze the exact quantities Ts Ex and Ec from density functional theory (DFT) for the dissociating hydrogen molecule.These quantities obtained from con guration interaction calculations are compared with the approximate quantities within the local density approximation (LDA) and within LDA with nonlocal corrections (LDA+NL). We also make a comparison of the densities obtained within LDA and within the Hartree-Fock (HF) approximation. We further analyze the atomic fragments in the dissociation limit within the LDA and Local Spin Density (LSD) approximation. We conclude that LDA+NL gives a very good approximation to the exact exchange and correlation energies at equilibrium distance, but a not so good approximation for these quantities at large bond distances due to a incomplete cancelation of the Hartree potential by the exchange-correlation potential in the atomic fragments in the limit of in nite separation. This error can be corrected by using the LSD and LSD+NL approach. We also demonstrate that the present correlation functionals do not give a proper account of the correlation energy in the dissociation limit, an error which is partly compensated by an overestimation of the exchange energy providing a good total exchange-correlation energy. We will however show that conventional de nition of correlation energy within DFT is not a useful one in describing the physics of molecular dissociation but one should rather consider the total exchange-correlation energy to describe the physical situation. We analyze how several types of density changes due to bond formation a ect the various LDA and LDA+NL exchange and correlation energies, explaining the sign of the corrections to the bond energy.
11.1 Introduction
Density functional theory 5] has become a very successful method in the calculation of electronic properties of systems ranging from atoms to molecules 70, 122] and solids 47]. Especially in the last few years after the the introduction of the so-called Generalized Gradient Approximations (GGA) 44, 45, 43, 46, 50] or nonlocal (NL) corrections density functional theory has gained wide popularity within the eld of quantum chemistry through its ability to yield very good bond energies which can compete with those obtained from other correlated quantum chemical methods. It has for instance been widely applied to the eld of transition metal chemistry 71] where many systems 183
184
EXCHANGE AND CORRELATION..
can not be studied by more conventional methods for practical reasons. Observing this large eld of application within chemistry it would be very desirable to investigate whether the physics of for instance bond breaking is correctly described by the presently used density functionals. Such an investigation would not only contribute to a better understanding of the physics behind the phenomena in question but might also lead to the development of more accurate density functionals. For this reason we calculated the some of the quantities which are used in DFT exactly for the case of dissociating hydrogen molecule and compare them with the approximate ones obtained from LDA and LDA+NL. The hydrogen molecule is one of the rst molecules used as a test for the LSD method 142, 143]. Its Kohn-Sham potential has been constructed and thorougly analyzed 86]. An analysis of the exact exchange-correlation energies has however not been presented. Bond breaking is a specially interesting testcase for the presently used functionals as here correlation e ects (especially left-right correlation e ects) are large. We further want to analyze how the various exchange and correlation functionals (LDA and GGA) contribute to the bond energy. In order to understand this we also studied the atomic fragments in the in nite separation limit. A question we like to answer in this paper is why the Becke-Perdew nonlocal exchange-correlation correction as a rule gives a positive (repulsive) contribution for the bond energy (as in all homonuclear rst and second row diatomic molecules) with the exception of the hydrogen molecule which is the subject of this paper. In order to answer this question we analyze how the exchange and correlation contributions to the bond energy are changed by certain density variations resulting from bond formation. The paper is divided as follows. In section 11.2 we give a short account of the exact quantities we want to calculate. In section 11.3 we discuss some properties and give some numerical results for the atomic fragments in the dissociation limit in order to understand the role of the atomic error in bond breaking. In section 11.4 we discuss the exact results for the DFT quantities as a function of the bond distance obtained from accurate calculations. In section 11.5 we analyze the bond energy contribution of several exchange and correlation functionals and analyze how they are in uenced by density variations. In section 11.6 we nally present our conclusions.
11.2 Exchange and correlation in density functional theory

In this section we de ne some of the quantities we will calculate in section 11.4. The energy functional Ev within density functional theory is usually de ned by the constrained search approach of Levy and Lieb 67, 68]:
Z
Ev ] =
(r)v(r) + FL ]
(11.1) (11.2)
where the functional FL is de ned as: ^ ^ ^ FL ] = min TrfD(T + W )g

^ D!
which minimizes the trace value of the kinetic energy T and the interparticle interaction W with the ^ set of N -particle density matrices D yielding density . The above formulation is important for the ^ consideration of degenerate ground states where D can be a linear combination of density matrices corresponding to pure states within the ground state multiplet. This is of importance for instance when a molecule dissociates into open shell atoms with degenerate ground states. Moreover Ev is also invariant under unitary transformations within the multiplet of degenerate wavefunctions. We will return to this point again in section 11.3 when we discuss the in nite separation limit of the
EXCHANGE AND CORRELATION IN DENSITY FUNCTIONAL THEORY
185
dissociating hydrogen molecule. The functional FL is universal in the sense that it can in principle be calculated for each density independent of the external potential v. From its de nition it is also clear that FL is invariant under rotations and translations of the density , providing important constraints for approximate exchange and correlation functionals 32]. Within the Kohn-Sham approach to density functional theory the functional FL is usually split up as follows: Z F ] = T ] + (r1 ) (r2 ) dr dr + E ] (11.3)
L L
in which TL is the Kohn-Sham kinetic energy which is the kinetic energy of a noninteracting particle system with the same density as the interacting system ,de ned by constrained search as: ^^ TL ] = min TrfDT g (11.4)
^ D!
jr1 ; r2 j 1 2
xc
If the ground state Kohn-Sham wavefunction can be represented by a pure state density matrix corresponding to a single Kohn-Sham determinant we will denote TL by Ts . We will do this for instance for the hydrogen molecule. As we now have de ned all functionals except Exc equation 11.3 actually de nes Exc. The Euler-Lagrange equations obtained by minimizing Ev with respect to the density can be written in the form of the well-known Kohn-Sham equations. For future reference we give the spinpolarized equations with spin index : 1 (11.5) (; 2 r2 + vs (r)) i (r) = i i (r) where (r) =
X
(r) =
XX
i=1
fi j
(r)j2
(11.6)
where fi are the orbital occupation numbers (in general fractional for degerate ground states) and the vs is the Kohn-Sham potential: Z (r1 ) dr + v ( ] r) v (r) = v(r) + (11.7)
s
where the exchange-correlation potential vxc is the functional derivative of the exchange-correlation functional Exc with respect to the density . We further de ne the exchange functional as: X 1 Z j s (r1 r2 )j2 dr dr Ex " # ] = ; 2 (11.8) jr1 ; r2 j 1 2 where s is the Kohn-Sham density matrix:
s
jr ; r1 j 1
xc
(r1 r2 ) =
fi
(r1 ) i (r2 )
(11.9)
This is for a xed density ensemble, i.e. for xed occupation numbers fi , a density functional 67, 68] as the Kohn-Sham orbitals and therefore also the Kohn-Sham one-particle density matrix are uniquely determined by the ensemble density ( an explicit construction procedure has been given in 32] ). The Kohn-Sham one-particle density matrices corresponding to di erent ground state ensemble densities are di erent and therefore they lead in general to di erent exchange energies.
186 The exchange potential vx (r) = Ex = asymptotic behaviour 49, 35]:

(r) for spherical atoms has the following long range
(r ! 1) (11.10) vx (r) ; fm r where fm is the occupation of the highest occupied Kohn-Sham orbital. From this formula it is clear that it depends on the occupation fm of the highest occupied Kohn-Sham orbital to
which amount the self-energy is being cancelled. For nondegenerate ground states one always has fm = 1 and for this case the self-energy is always being cancelled. This has important implications for dissociating molecules which dissociate into open shell fragments that have degenerate ground states as will be discussed in the next section. For closed shell systems we can simplify the above exchange functional as:
2 1Z Ex ] = ; 4 j jsr(r1 r2 )jj dr1 dr2 1 ; r2
(11.11)
where s =
. Finally the correlation energy functional is de ned as: (11.12)
Ec ] = Exc ] ; Ex ]
Note that the de nition of exchange-correlation energy in other quantum chemical methods ,which we will denote by Wxc, di ers from the above de nition in density functional theory, the relation being:
Wxc ] = Exc ] ; Txc ]

where Txc is given by:
(11.13) (11.14)
Txc ] = T ] ; Ts ]
and where T is the kinetic energy functional which for nondegenerate ground states is de ned as the expectation value of the kinetic energy operator with the wavefunction which minimises FL within constrained search, which is just the ground state kinetic energy if we insert the ground state density. Note that T ] is only de ned for nondegenerate ground state densities as otherwise could be generated by two di erent degenerate ground state wavefunctions, corresponding to the same total energy but having di erent expectation values with the kinetic energy operator 5]. For a density corresponding to a nondegerate ground state Wxc is just the expectation value of ^ the ground state wavefunction with the interparticle interaction operator W minus the Coulomb repulsion energy. It is readily shown using equations 11.2 and 11.4 that Txc ] 0 so:
Exc ] Wxc ]
(11.15)
We have now de ned all the quantities we need for our analysis of the dissociating hydrogen molecule. In one of the following sections we will compare the above de ned functionals Ts Ex and Ec calculated from accurate Con guration Interaction calculations with those of the Local Density Approximation (LDA) and with the Generalized Gradient Approximation (GGA) for several bond distances.
11.3. THE INFINITE SEPARATION LIMIT
187
11.3 The in nite separation limit

For the case of a hydrogen molecule the Kohn-Sham equations (in the spin restricted case) reduce to the following di erential equation involving only the electron density: q q (; 1 r2 + vs ( ] r)) (r) = (r) (11.16) 2 The Kohn-Sham kinetic energy is for a two-electron system given by: 1 Z (r )2 dr Ts ] = 8 (11.17) The exchange functional in this case is then given by: Z ) Ex ] = ; 1 j(rr1 ; (rr2j) dr1 dr2 (11.18) 4 1 2 and the corresponding exchange potential : Ex = ; 1 Z (r1 ) dr vx ( ] r) = (r) (11.19) 2 jr ; r1 j 1 which is just minus onehalf times the Hartree potential. So the Kohn-Sham potential becomes: 1 Z (r vs ( ] r) = v(r) + 2 jr ; 1r) j dr1 + vc( ] r) (11.20) 1 If we neglect the correlation potential vc then we just obtain the Hartree-Fock equations or the optimized potential model 33, 35] equations for this two-electron system. This is of course a bad approximation for cases where correlation e ects are large such as in the dissociation limit where left-right correlation e ects are large. In this limit the Kohn-Sham potential can be approximated by the sum of atomic Kohn-Sham potentials and the Kohn-Sham potential becomes: vs ( ] r) vs( A] r RA) + vs ( B ] r RB ) (11.21) where the atomic nuclei are on positions RA and RB . The last step in this approximation is only valid in the regions of the atoms and not in the bond midpoint region. In this region the Kohn-Sham potential vs is small in the dissociation limit and approximating this by the sum of atomic densities might be a bad approximation. This is re ected by the fact that a peak structure is develops in the Kohn-Sham potential 86, 90] indicating strong left-right correlation e ects. In the region of each atom we have : 1Z vs ( ] r RA ) = vA (r) + 2 jrA (rr1 )j dr1 + vc( ] r RA) (11.22) ; 1 where vA is the potential of the atomic nucleus. For the hydrogen molecule the atoms are just the hydrogen atoms which are one electron systems for which the Kohn-Sham potential should of course just be the nuclear potential vA .It follows that the correlation potential vc should be equal to minus onehalf the Hartree potential and thus be equal to the exchange potential in this limit. The Hartree-Fock approximation is obtained by neglecting this correlation potential and therefore the potential is too repulsive making the electron too weakly bound and the density too di use. This is a well-known error of the Hartree-Fock approximation which is usually solved by doing a spin unrestricted Hartree-Fock calculation. Let us now turn this Hartree-Fock problem for a more general homonuclear diatomic molecule. Suppose we consider dissociating closed shell diatomic
188
molecule and suppose that in the dissociation limit (where we can neglect the overlap of the atomic p orbitals) we have M bonding molecular orbitals i = ( A + B )= 2 and their corresponding i i p antibonding combinations i = ( A ; B )= 2 ,where A B are the atomic orbitals of atoms A i i i and B , and N ; M molecular orbitals without corresponding antibonding combinations. Then the one-particle density matrix of the molecule in the region of atom A becomes: 1 A (r ) A (r ) i 1 i 2 i=M +1 2 and the corresponding atomic exchange energy becomes: (r2 ) +
X
M A (r1 r2 ) = X A (r1 ) A i i i=1

Z
(11.23)
j A (r1 r2)j2 dr dr jr1 ; r2 j 1 2 The atom A in this limit has N ; M degenerate orbitals with occupation number
A X 1 Ex = ; 2
(11.24)
1 in the 2
outermost atomic shell. The corresponding one-particle density matrix can not be represented by a single determinant but it can be written as:
NA A (r1 r2 ) = X wi A (r1 r2 ) (11.25) i i where iA is the one-particle density matrix correponding to the i-th degenerate con guration in P the NA -fold degenerate ground state multiplet and NA wi = 1. The Hartree-Fock solution for the i
atom A in the in nite separation limit can be obtained by minimizing the energy: Z X 1Z E FHF = ; 2 r2 A(r1 r2 )j1=2 dr1 dr2 + (r1)vA (r1 )dr1 + 2 1 Z (r1 ) (r2 ) dr dr ; X 1 Z j A (r1 r2 )j2 dr dr 2 jr1 ; r2 j 1 2 2 jr1 ; r2j 1 2
(11.26)
where A is given by 11.23. This procedure we will call fractional Hartree-Fock (FHF). We will present numerical results for this procedure for the case of the hydrogen atom in section 11.4. Application of this FHF method leads to an incomplete cancellation of the self-energy for open shell atoms, which is the cause of the Hartree-Fock error for dissociating molecules. Before we go further into this point we rst discuss the behaviour of the Local Density Approximation in this limit. Within LDA the exchange energy is given by:
Ex " # ] =
;cx
(r) 3 dr
(11.27)
where the constant cx is given by 3 ( 6 ) 1 . For the spin-restricted case we obtain: 4 3 Z 1 3 (11.28) Ex 2 1 ] = ; cx (r) 4 dr 2 3 21 The LDA correlation potential is much smaller than the exchange potential so we will neglect this potential for the moment and make a comparison with the Hartree-Fock approximation. Doing a restricted LDA calculation for an open shell atom is similar to doing an FHF calculation, one
THE INFINITE SEPARATION LIMIT
189
Figure 11.1: Minus the Hartree potential and the exchange-correlation potential for the hydrogen atom from LDA and LSD calculations as a function of the distance from the atomic nucleus
Figure 11.2: Minus half of the Hartree potential and the exchange-correlation energy density for the hydrogen atom from LDA and LSD calculations as a function of the distance from the atomic nucleus
190 LDA LDA ( e ) -0.2335 Ts T HF 0.4250 0.5000 ENe -0.9210 -1.0000 EC 0.2828 0.3125 Etot -0.4457 -0.4416 Etot +NL -0.4650 -0.4629 Ex -0.1931 -0.2127 Ex+NL -0.2357 -0.2588 Ec -0.0394 -0.0414 Ec +NL -0.0161 -0.0166 Exc -0.2325 -0.2541 Exc+NL -0.2518 -0.2754 H-atom

LSD LSD ( e ) -0.2690 0.4666 0.5000 -0.9656 -1.0000 0.2984 0.3125 -0.4787 -0.4776 -0.4999 -0.4996 -0.2564 -0.2680 -0.2968 -0.3097 -0.0217 -0.0221 -0.0025 -0.0024 -0.2781 -0.2901 -0.2993 -0.3121 KS -0.5000 0.5000 -1.0000 0.3125 -0.5000 -0.1563 -0.3125 FHF -0.2295 0.3578 -0.8437 0.2565 -0.3577
-0.1563 -0.1283
Table 11.1: Energies of the atomic fragment in the in nite separation limit just fractionally occupies each of the degenerate LDA Kohn-Sham orbitals in the outermost atomic shell. The exchange-only LDA Kohn-Sham potential for atom A is then given by: Z ( 3 vs ( A ] r) = vA(r) + jrA;rr1 )j dr1 ; 4 cx A(r) 1 (11.29) 3 21 3 1 Also this exchange-potential gives an incomplete cancelation of the Hartree potential, however the cancelation is better than in the FHF case. More important, as has been pointed out by Gunnarsson et al. 144] and shown in gure 11.1, is the fact that the di erence between the LDA exchange correlation potential and the Hartree potential is an almost constant function. A constant shift (of about 0.2 a.u. ) in the potential a ects the eigenvalue but not the density. Note that this constancy is approximately independent of the nuclear charge as both the Hartree and the exchange potential (which is the main part of the exchange-correlation potential) have a similar scaling property with respect to the nuclear charge. If we look at the energy expressions in LDA we see a much better cancelation of the exchangecorrelation energy and the electron repulsion than for the corresponding potentials. This has been displayed for in gure 11.2 where we plot the LDA ( and LSD to be discussed later on ) exchangecorrelation density xc and minus half of the Hartree potential ;vH . The most important di erence between a LDA and an Hartree-Fock calculation is therefore the di erence in energy expressions for the exchange energy leading to a better cancelation of the self-energy of the atoms in the in nite separation limit. The importance of this di erence in energy expressions is apparent from table 11.1 in which we display the several energy terms of the hydrogen atomic fragments in the in nite separation limit for the self-consistent LDA ,LSD, FHF approximations and for the exact Kohn-Sham values. To show the dependency on the quality of the density we also included the LDA and LSD results calculated from the exact density e . As we can see from this table both LDA and HF make an error in the electron nuclear attraction energy and in the kinetic energy which are almost equal but of opposite sign, which leaves a small error. Now we note the importance of the di erent energy expressions for the exchange energy. The LDA exchange energy is roughly equal to the exact exchange-correlation energy, whereas the LDA correlation energy is much smaller than the exact correlation energy, which leads to a rather good value of the exchange-correlation energy and a cancellation of the electron repulsion energy. This then gives a good approximation of the total atomic energy. The HF exchange energy is however only half of the electron repulsion energy
THE INFINITE SEPARATION LIMIT
191
and therefore gives a bad estimate of the total atomic energy. The Hartree-Fock results for the dissociating hydrogen molecule can be improved (an in fact will give the exact limit for this special case) by performing an unrestricted Hartree-Fock (UHF) calculation in which one breaks the spin symmetry and localizes one spin up electron on one atom and a spin down electron on the other atom. A similar procedure can be carried out in density functional calculations using a spin density functionals. A better description than LDA for the dissociated atom in the in nite separation limit is obtained in the Local Spin Density (LSD) approximation. This method has been thorougly investigated by Gunnarsson and Lundqvist 127]. In spin density functional theory one uses di erent potentials for the spin up and spin down electrons. As in UHF this leads to a violation of the spin rotational symmetry. Using LSD one can for instance solve the Kohn-Sham equations for the hydrogen atom containing only one electron with quantum number Sz = 1. Performing a spin rotation which leads to a di erent mixture of spin densities will give a di erent LSD energy. We will now investigate this problem more closely. The exact ground state solution for the hydrogen atom A at in nite separation ,as obtained from singlet ground state wavefunction of H2 , is given by: q (r s) = A (r)(c1 (s) + c2 (s)) (11.30) p p with c1 = 1= 2 and c2 = i= 2. The ground state energy should not change when we perform an unitary transformation on the vector (c1 c2 ) and the energy should therefore not depend on the spinpolarisation 145, 146] of the two degenerate spin states. In general ,if we have an electron P density which is a linear combination of densities i = i corresponding to pure eigenstates within one M -fold degenerate ground state multiplet, i.e: (r) =
P X
M i
wi i(r)
(11.31)
with M wi = 1, then any density obtained from the above density by a unitary transformation i within the space of degenerate states should yield the same ground state energy. This is an immediate consequence of the de nition of the functional FL . This means for the example of the 1 hydrogen atom that the spinrestricted density which is a linear combination ( w1 = w2 = 2 in equation 11.31) of the degenerate pure spin up and spin down densities ( (c1 c2 ) = (1 0) and (0 1) in equation 11.30 ) should yield the ground state energy for the hydrogen atom. As the invariance property for unitary transformations within the degenerate ground state multiplet is not ful lled for the presently used approximate density functionals they do not decribe properly the dissociation of closed shell molecules into open shell atoms with degenerate ground states. These atomic fragments in the dissociation limit are spinrestricted atoms with fractional occupation ( as in the FHF case, see equation 11.23 ) of the degenerate orbitals. This is also true in the exact Kohn-Sham case for which the corresponding density then corresponds to an ensemble ( see equation 11.31 ) of pure state multiplet densities which gives the exact ground state energy of the atom when inserted in the energy functional Ev . The total energies of the dissociated atomic fragments in the LDA case however are too large, but smaller than the Hartree-Fock ( FHF ) atoms in this limit. As an unrestricted LSD calculation lowers the atomic energy this yields an improvement of the total atomic energy although such a correction should not be needed from a theoretical points of view. From table 11.1 we can see the LSD improvements in the density re ected in the good estimates for electron nuclear attraction energy and the kinetic energy, leading to a very good estimate of the total energy especially if the nonlocal corrections are included. After studying the in nite separation limit we will in the next section study some exact density functional quantities as a function of the bond distance.
192
Figure 11.3: Electron-electron repulsion energy as a function of the bond distance
11.4 Numerical results

In order to calculate exact Kohn-Sham quantities Ts Ex and Ec for the hydrogen molecule we performed some accurate Con guration Interaction (CI) calculations at several bond distances. We compare these accurate results with Hartree-Fock and self-consistent density functional calculations within the Local Density Approximation. The nonlocal exchange functional of Becke 44] and the nonlocal correlation functional of Perdew 45, 147] where added on top of the LDA calculation (using the VWN-correlation potential 85] ) in a post-scf manner. The e ect of not including the Becke and Perdew functionals in a selfconsistent manner is negligible 139] as the in uence of these potentials on the density is very small. We rst look at some density dependent quantities. In gure 11.3 we plotted the Coulomb repulsion energy of the electronic cloud given by: 1 Z (r1 ) (r2 ) dr dr EC ] = 2 jr ; r j 1 2 (11.32) 1 2 calculated from the HF, CI and LDA densities. We can see from this gure that the Coulomb repulsion energy approaches much earlier the 1=R(R ! 1) behaviour, which is just the repulsion of point charges with charge one, than the HF and LDA densities. That the LDA graph is closer to the CI graph than the HF graph is in accordance with the result of the previous section that the HF-density density is the most di use one. This fact is supported by gure 11.4 which plots the electron nuclear attraction energy:
ENe ] =
(r)v(r)dr
(11.33)
NUMERICAL RESULTS
193
Figure 11.4: Electron nuclear attraction energy as a function of the bond distance where v is the Coulomb potential of the atomic nuclei. We again can see that the correct asymptotic behaviour ;2 ; 2=R(R ! 1) is reached much earlier by the CI density, the worst density again being the HF-density. Another measure of the quality of the electron density is the kinetic energy T and the Kohn-Sham kinetic energy Ts . In gure 11.5 we plotted the kinetic energies T CI and T HF and the Kohn-Sham kinetic energies Ts and TsLDA , where Ts is calculated from the CI-density. We can see from this gure that the graphs of T CI and Ts are very close at all bond distances, the di erence being the exchange-correlation part Txc of the kinetic energy. The LDA kinetic energy TsLDA has a large deviation from Ts at large bond distances which approaches quickly the atomic error in this quantity. The HF kinetic energy at these bond distances then deviates even more from T CI again re ecting the di useness of the density and the bad representation of the atomic fragments. Note however that at large distances the LDA and HF errors in the electron nuclear attraction energy and the kinetic energy are approximately equal and of opposite sign, giving a good approximation for the sum of those quantities. In the vicinity of the equilibrium bond distance (1.401 bohr) the errors in the HF and LDA densities are small and in that case we have a good estimate of the energy terms described above. In gure 11.6 we plotted the total energies for the CI, HF, LDA and LDA+NL with the nonlocal corrections of Becke and Perdew added. We also calculated the expection value of the LDA KohnSham determinant, which is just the HF expression for the energy but with LDA orbitals inserted. We will denote the corresponding graph by HFLDA. We see that the curves HF, HFLDA and LDA are very close around the equilibrium distance. Also the CI and LDA+NL are very close in this bond distance region. At large bond distances all the curves deviate from the CI curve, the worst ones being HF and HFLDA (the HFLDA a little higher than HF because the LDA determinant
194
Figure 11.5: Kinetic energy and Kohn-Sham kinetic energy as a function of the bond distance
Figure 11.6: Total energy as a function of the bond distance
NUMERICAL RESULTS
195
Figure 11.7: Exchange-correlation energy as a function of the bond distance
Figure 11.8: Exchange energy as a function of the bond distance
196
does not minimize the HF energy expression). The fact that the LDA curve is so much closer to the CI curve than the HFLDA curve is mainly due to a di erent expression for the exchange energy. The LDA exchange energy gives a better cancelation of the electron repulsion energy in the atomic fragments than the HF and HFLDA exchange energy which cancel only half of the electron repulsion energy. As the nonlocal corrections give an additional energy lowering the best curve at all distances is given by the LDA+NL curve. To analyze the several exchange and correlation functionals more closely we plotted in gure 11.7 LDA LDA the exchange-correlation energies Exc Wxc Exc and Exc +NL as a function of the bond distance. From this plot we can see that indeed Exc Wxc the di erence being the exchange-correlation part Txc of the kinetic energy. At large bond distances Exc and Wxc approach ;EC thereby canceling the electron repulsion energy. Due to incomplete cancelation of this term within LDA and LDA+NL exchange-correlation energies we expect the largest errors in these quanties at large separation. We LDA LDA can see from gure 11.7 that this is indeed the case. Both Exc and especially Exc +NL give a very good approximation to the exact Kohn-Sham exchange-correlation energy Exc around the LDA LDA equilibrium distance. At larger distances both Exc and Exc +NL deviate from the exact Exc the di erence approaching fast the atomic error in these quantities. We will now investigate the separate behaviour of the exchange and correlation energies. In gHF LDA LDA ure 11.8 we plot the exchange energies Ex Ex Ex and Ex +NL as a function of the bond distance. We can see the same trends as for the total exchange-correlation energies. We have a very LDA good description of the exact exchange energy Ex at equilibrium distance especially by Ex +NL LDA and large errors at large distance. The Ex +NL exchange will of course approach two times its value for the spinrestricted atoms in the large distance limit. However one should realize that the de nition of exact exchange as de ned in equation 11.8 is not a useful one to describe the physics of bond breaking as the corresponding exact exchange potential does not contain the complete selfenergy correction of the atomic fragments. This means that the correlation potential which has to correct this should contain part of the self-energy which gives obviously a rather unphysical description of the correlation energy. As discussed this not only a feature of bond breaking but a general problem in the calculation of open shell systems within a spin restricted formalism. Let us look LDA more closely at the correlation energies. In gure 11.9 we plot the correlation energies Ec Ec LDA+NL . We can see that E LDA and E LDA+NL have a very slow variation with respect to the and Ec c c LDA bond distance. The Ec represents the VWN parametrization of electron gas correlation energy data giving a correlation energy which about a factor two too large in the molecule at equilibrium distance and in the atomic fragments. Addition of the Perdew gradient correction to the correlation energy gives very good values for the correlation energy for the molecule at equilibrium distance and in atoms and is also almost constant as a function of the bond distance. This is not the case for the correlation energy Ec as de ned in equation 11.12 .The graph of Ec approaches just as Ex 1 LDA LDA the value ; 2 EC at large bond distances and becomes much larger than Ec and Ec +NL . This LDA is due to the fact that Ec has to cancel half of self-energy. The approximate functionals Ec , LDA which is tted to electron gas data, and the Ec +NL , which is tted to atomic data, are of course unable to imitate this behaviour. In view of the inconvenient de nition for exchange and correlation in the dissociation limit it is better to regard the total exchange-correlation energy. The total exchange-correlation energy of an atom is a quantity which automatically cancels the self-energy, as the exchange-correlation hole always will integrate to one electron, also for degenerate ground state wavefunctions. However Exc is in general still di erent for di erent ground state ensemble densities but due to the correct inclusion of the self-energy these di erences are smaller than for the exchange or correlation alone. These di erences in Exc can of course not be avoided as di erent degenerate ground state wavefunctions
11.5. AN ANALYSIS OF BOND ENERGY CONTRIBUTIONS
197
Figure 11.9: Correlation energy as a function of the bond distance will in general have di erent expectation values for any other operator than the total energy operator. From the above considerations it is clear that for the case of dissociating molecules it is much easier to nd appropriate approximations for Exc than for Ex and Ec as de ned in equations 11.8 and 11.12. We can therefore conclude that LDA+NL gives a very good description of the properties of the molecule ( including exchange and correlation energies ) at equilibrium distance. The remaining errors for LDA and LDA+NL are due to errors in the Coulomb repulsion and in the exchangecorrelation energy and the molecular error is quickly approaching the atomic errors in this quantity. These is however not a problem in practical calculations as a spin unrestricted calculation with one spin density on one atom gives a good description of the total atomic energy and one can get good binding curves at all distances using spin unrestricted Kohn-Sham calculations similar to UHF calculations. This is however unsatisfactory from a theoretical point of view as ,due to unitary transformations among the degenerate ground states of the exact density functional Ev , spin restricted and spin unresticted calculations on the hydrogen atom should give the same energy. In the next section we will investigate how and why several LDA and nonlocal (GGA) in uence the bond energy as they do and how this is related to density changes.
11.5 An analysis of bond energy contributions

In this section we will investigate how the various exchange and correlation energy contributions (LDA and GGA) to molecular bonding energy are a ected by certain density changes. We therefore construct the molecular density of the hydrogen molecule in two steps. First , which is the sum
198 H2 LDA 1.107 -3.602 1.297 -1.137 -1.178 -0.559 -0.648 -0.089 -0.041 -0.648 -0.689 0.863 -3.220 1.128 -1.091 -1.126 -0.488 -0.566 -0.089 -0.046 -0.577 -0.612 1.116 -3.556 1.257 -1.109 -1.150 -0.547 -0.636 -0.093 -0.045 -0.640 -0.681

KS 1.140 -3.647 1.321 -1.172 -0.039 -0.700 HF 1.125 -3.631 1.316 -1.133
Ts T HF ENe EC Etot Etot +NL Ex Ex+NL Ec Ec +NL Exc Exc+NL
-0.661 -0.658
Table 11.2: Energies of hydrogen molecule at equilibrium distance (R=1.401 bohr) of atomic densities, is made. For the hydrogen molecule this density is given by: ~ (r) = 1 (e;2jrj + e;2jr;Rj )
(11.34)
~ where ~ and R are the positions of the atomic nuclei. The bond energy contribution of a quantity 0 E corresponding to the density we will denote by E , so we have: E =E ];E
A] ; E B ]
(11.35) to real molecular density which we will approximate (11.36)
In the second step there is a change from by a density contraction:
~ ~ (r R) = 3 ( r R) E =E ];E ]
which is just a scaling transformation. The bond energy contribution caused by this second density change we will denote by E : (11.37) A third contribution to the bond energy is caused by going from spin restricted to spin unrestricted atoms. This energy change is not caused by a change in the spatial electron density but by a change in the electron spin densities. This third energy change we will denote by Es . We will now discuss how the two described density changes and the spin density change a ect the bond energy contributions of the LDA and GGA exchange and correlation functionals.
LDA exchange. The bond energy contribution Ex for this case is given by:
3 3 3 Ex = ;cx (( A (r) + B (r)) 4 ; A(r) 4 ; B (r) 4 )dr < 0
Z
(11.38) (11.39)
This follows directly from the fact that: (x + y) 3 > x 3 + y 3

4 4 4
AN ANALYSIS OF BOND ENERGY CONTRIBUTIONS
199
if x y > 0. For our model density we have Ex = ;1:70 eV. In order to estimate the bond energy contribution E we use the scaling property of the exchange functional 22]. powers of ; 1 yields:
~ ~ Ex (r R)] = Ex (r R)] (11.40) ~ where R is the relative position vector of the atomic nuclei. An expansion of equation 11.40 in Ex = ( ; 1)(Ex ~ ]+R
Z
vx ( ] r)rR (r)dr) + O(( ; 1)2 ) ~
(11.41)
~ If we neglect the second term (assuming that Ex (r R)] varies much faster with respect to than ~ ) and we obtain: with respect to R Ex
( ; 1)Ex ]<0 (11.42) The optimal with minimizes the LDA+GGA functional is given by = 1:140 and therefore we obtain to a good approximation:
Ex = 0:140Ex ]
(11.43)
Using the above formula we nd Ex = ;1:86 eV. An actual calculation of Ex ] gives Ex = ;1:60 eV. Finally there is the correction by going from spin unrestricted to spin restricted 1atoms. As can be seen from formula 11.27 this lowers the LDA exchange energy by a factor 2 3 . We therefore obtain the following positive correction to the bond energy: 3 3 3 (11.44) Es x = ;(2 1 ; 1)(Ex A ] + Ex B ]) = 2(2 1 ; 1)( 3 )4 ( 32 ) 1 = 3:01 eV 4 In general we can say that both formation of and contraction of the density give a negative contribution to the bond energy contribution of the LDA exchange energy and that the spin unrestricted corrections give a positive contribution.
GGA exchange. We consider in particular the Becke exchange correction. However our conclusions remain valid for a broad class of nonlocal exchange corrections of the Becke form. The Becke exchange energy correction is given by:
Ex " # ] =
where
(r) 3 f (x )dr
(11.45) (11.46) (11.47)
f (x) = ;
with = 0:0042 and x (r) = jr (r4)j (r) 3 Spin-unrestricted we have: 1 Z (r) 4 f (2 1 x)dr 3 3 Ex ] = 1 23
x2 1 + 6 x sinh;1 x
(11.48)
200
The bond energy correction Ex due to the Becke exchange is given by: Z 3 3 3 3 3 3 Ex = 11 ( (r) 4 f (2 1 x ) ; A(r) 4 f (2 1 xA ) ; B (r) 4 f (2 1 xB ))dr (11.49) 23 The main contribution to the integrand of the above integral arizes in the bond midpoint region of the molecule where the molecular gradients are small (in the bond midpoint they are exactly zero ). In that case it is easily seen that we have: Ex > 0 (11.50) Within our model we nd Ex = 0:38 eV. If we now contract the density then we have considerable changes in the density at the atomic nuclei. As the Becke exchange functional scales according to equation 11.40 we can use the same arguments as for the LDA exchange and use equation 11.42 in order to calculate Ex . Both an exact calculation and the approximate formula 11.42 give Ex = ;0:30 eV. The atomic spin unrestricted correction to the Becke exchange is given by: Z 3 3 (11.51) Ex s = 2 A (r) 4 (f (xA) ; 11 f (2 1 xA ))dr < 0 23 We nd Ex s = ;0:24 eV. So we can say in general that both formation of the density and the spin unrestricted corrections give a positive contribution to the bond energy. The density contraction leads to a negative contribution.
LDA correlation. Due to the rather complicated nature of the Vosko-Wilk-Nusair expression for the LDA correlation energy it is di cult to predict the change in this expression due to the described density variations. We nd by numerical evaluation that Ec = ;0:17 eV, Ec = ;0:11 eV and Ec s = ;1:05 eV. GGA correlation. We consider the Perdew correlation correction which is given by: Z Ec " # ] = d(1 ) C ( )g( jr j)dr where 3 g( jr j) = 4 x2e; ( jr j) with
(11.52) (11.53) (11.54)
x = jr4 j
3
where = ( " ; # )= is the spin polarization and: 1 5 1 3 d( ) = 2 3 (( 1 + ) 5 + ( 1 ; ) 3 ) 2 (11.55) 2 2 The function C ( ) is a local function of the density and has been obtained from the gradient expansion for slowly varying densities of the correlation energy of the electron gas. The function is given by: ( jr j) = 1:745f~C ((1)) jr7 j (11.56) C 6
AN ANALYSIS OF BOND ENERGY CONTRIBUTIONS
201
~ where f = 0:11. This function is obtained from the wavevector analysis of Langreth and Mehl. For the spin unpolarized case we have (d=1) :
4 Ec ] = C ( ) 3 x2e; ( jr j)dr
Z
(11.57)
B
We then have: Z E c = (C ( ) g (
jr j) ; C ( A)g( A jr Aj) ; C ( B )g(
jr B j)dr
(11.58)
As with the Becke exchange correction the main contribution to the above integral arises from the molecular bond midpoint region where the density gradients of the molecule are small ( the function g is exactly zero in the bond midpoint region ) so we nd: Ec < 0 (11.59) Within our model we nd Ec = ;0:17 eV. The Perdew correlation functional does not scale like the exchange functional. However due to the fact that the function C is almost constant as a 4 function of the density and the importance of the term 3 x2 it is to a good approximation true (as we checked for several values of ) that (at least for atomic and molecular densities): ~ ~ Ec (r R)] Ec (r R)] (11.60) which yields just as for the exchange case: Ec ( ; 1)Ec ] > 0 (11.61) A calculation using the above formula (using the optimal value = 1:140 ) yields Ec = 0:16 eV. An actual calculation gives Ec = 0:13 eV. We nally consider the atomic spin corrections. For spin polarized hydrogen atom we have an spin 1 polarisation = 1 and d( ) = 2 3 which means that the Perdew correlation energy for the spin 3 unpolarized hydrogen atom is a factor 2 1 times this energy for the spin polarized atom (assuming that the spatial densities are the same). So we nd: 3 Ec s = 2(2 1 ; 1)Ec A ] > 0 (11.62) Within our model we nd Ec s = 0:28 eV. So for the Perdew correction the formation of gives a negative contribution to the bond energy and both the density contraction and spin polarization corrections give a positive contribution to the bond energy. As can be seen from the above analysis the actual sign of the bond energy contribution of the several exchange and correlation energies to the bond energy depends on some opposing e ects. However as a rule the Becke exchange correction is positive and lowers the bond energy and the Perdew correlation correction is negative and increases the bond energy. This is for instance true for all rst and second row homonuclear diatomic molecules. A notable exception being the hydrogen molecule which is the subject of this paper. However the hydrogen molecule is a rather special case due the absence of core electrons so that density changes on the nuclei , which a ect E are rather important. Also the spin corrections for the atoms are rather large because the spin polarized hydrogen atoms are an extreme case of spin polarization with = 1. Those corrections are smaller for atoms with a less extreme spin polarisation. This means that in applications to other molecules the most important correction to the bond energy is the term E which is positive for the Becke exchange correction and negative for the Perdew correlation correction which con rms the actual trend observed.
202
11.6 Conclusions
In this work we analyzed the exact exchange and correlation energy Ex and Ec as well as the KohnSham kinetic energy Ts as a function of the bond distance for the case of a dissociating hydrogen molecule. These quantities are compared with the correspnding ones from the local density approximation and nonlocal corrections. The LDA+NL gives a very good account for these quantities at equilibrium separation and gives a binding curve close to the exact one. At large separation the LDA and LDA+NL give a not so good description of the binding curve due to errors in the total atomic energies. These errors are caused by the fact the presently used approximate density functionals for the case of degenerate ground states are not invariant under unitary transformations within the ground state multiplet. Correct dissociation can however be obtained through the construction of approximate density density functionals which give a good total ground state energy for open shell atoms within a spin restricted formalism. The construction of such density functionals remains a future challenge. It is further shown that LDA plus nonlocal correlation is unable to describe the left-right correlation by underestimating this correlation energy. This error is partly corrected by a corresponding overestimation of the exchange energy. The fact that the Becke-Perdew correction in the hydrogen molecule is negative instead of the the usual positive correction is shown to be related to special property of the hydrogen molecule of having no core electrons which make density contraction e ects important. Another important fact in this respect is the fact the spin corrections for the atomic fragments in this molecule have a large e ect due to the extreme polarization of the polarized atoms.
Chapter 12
Energy expressions in density functional theory using line integrals

In this paper we will adress the question of how to obtain energies from functionals when only the functional derivative is given. It is shown that one can obtain explicit expressions for the exchange-correlation energy from approximate exchange-correlation potentials using line integrals along paths within the space of densities. The path dependence of the results is discussed and criteria for path independence are given. Derivations are given of upper and lower bounds to the exchange-correlation energy in terms of the exchange-correlation potential at the beginning and the end point of a certain path. We further express the kinetic part Txc of the exchange-correlation energy in terms of a line integral and derive a constraint on approximate correlation potentials. We show how to use the line-integral formalism to derive the requirements which exchange-correlation potentials must ful l in order to make the exchange-correlation functional satisfy some symmetry property such as rotational and translational invariance and scaling properties. Finally we will discuss how to obtain bonding energies of molecules from approximate potentials.
12.1 Introduction
Density functional theory 5] has in the last few years been succesfully applied with increasing accuracy to systems ranging from atoms and molecules 71, 70, 122] to surfaces and solids 47]. Especially the introduction of the so-called generalized gradient expansion approximations (GGA) 44, 50, 46, 43, 45] for the exchange-correlation energy has constituted a great improvement over the local density approximation (LDA) in the calculation of for instance molecular bond energies. In view of the accuracy of the GGA energy expressions for the exchange-correlation functional and their improvement of the LDA energy expressions it is surprising that the GGA exchange-correlation potential gives little improvement over the LDA exchange-correlation potential 49, 32] and has a small in uence on the density. For this reason calculating the GGA corrections from an LDA density yields almost the same results as an inclusion of the GGA potentials in a self-consistent calculation 139]. This fact is also re ected in the LDA+GGA eigenvalues which are very close to the LDA eigenvalues which are much too small for nite systems such as atoms and molecules and also for surfaces. As the eigenvalue of the highest occupied Kohn-Sham orbital has been proven to be equal to the ionisation energy of the system 31, 72] the LDA gives too weakly bound electrons. This latter fact is due to the LDA exchange-correlation potential having exponential decay instead of the correct Coulombic decay. One way to improve the one-electron energies and the density (and 203
204
ENERGY EXPRESSIONS IN DENSITY FUNCTIONAL THEORY
related quantities such as dipole moments) is to directly approximate the exchange or exchangecorrelation potential instead of the exchange or exchange-correlation functional. Over the years several potentials (not only within density functional theory) have been proposed. Well-known is the average Hartree-Fock exchange potential proposed by Slater 101, 148]. A so-called optimized e ective potential (OEP) and an approximation to this was suggested by Sharp and Horton 34] and by Talman and Shadwick 33]. This potential received considerable attention within the context of exchange-only density functional theory or the optimized e ective potential model (OPM) and was shown to give one-electron energies close to the Hartree-Fock ones. An approximation to this potential was recently proposed by Krieger et al. 35] and Gritsenko et al. 121]. Also within the weighted density approximation (WDA) method approximate exchange potentials with the correct asymptotic behaviour have been used 149, 150, 124]. A more crude way to improve the asymptotic behaviour of the exchange potential was given by Latter 151] within the context of the Thomas-Fermi model which due to its simplicity has found widespread use in atomic structure calculations. Much less work has been done regarding the full exchange-correlation potential. An approximate exchange-correlation potential has been proposed by Harbola and Sahni 131] . This exchange-correlation potential vxc (r) has been de ned as the work done by bringing an electron from in nity to point r against the force of the electric eld of the exchange-correlation hole. However to insure path-independence for non-spherical systems like molecules one must only consider the irrotational part of this electric eld 152]. A gradient approximation to vxc has recently been proposed in reference 32] which was shown to give very good ionisation energies as calculated from the highest occupied Kohn-Sham orbital. In view of the fact that correct one-electron energies, the asymptotic Coulombic behaviour and the atomic shell structure can be reasonably obtained by approximate potentials it is an interesting question of theoretical and practical importance whether we can calculate good exchange and correlation energies from these approximate potentials. This is also of importance since it is possible to test the quality of approximate exchange-correlation potentials by comparing them with accurate ones constructed from accurate densities. It is for instance an interesting question how certain features in the exchange-correlation potential (such as the bond midpoint peaks in molecules 86, 90] related to the left-right correlation e ect ) contribute to the exchange-correlation energy. This requires an understanding of the relation between potential and energy expression. A further understanding of the features displayed in the Kohn-Sham potential might then also lead to more improved expressions for the exchange-correlation energy. If we want to assign some energy expression for instance to the model potentials considered above we immediately run into some theoretical di culties because none of the potentials considered, with the exception of the OPM potential, is a functional derivative of some energy density functional. However for approximate exchange potentials which satisfy the exchange scaling property it is still possible to assign an exchange energy using the Levy-Perdew relation 121, 35]. However such a relation is not available for the exchange-correlation potential due to the unknown scaling properties of the correlation functional. In section 12.2 of this article we will show how to use line integrals to calculate the exchange-correlation energy from an exchange-correlation potential. We also discuss some criteria of path independence.In section 12.3 we discuss some bounds on the exchange-correlation energy in terms of the exchange-correlation potential and we give an line integral expression of the kinetic part Txc of the exchange-correlation energy. In section 12.4 we derive some constraints on the correlation potential. In section 12.5 we derive some properties with respect to rotation, translation and scaling of the exchange-correlation energy from given properties with respect to these symmetries of the exchange-correlation potential. In section 12.6 we will adress the question how to calculate molecular bond energies from given potentials. Finally in section 12.7 we will present
12.2. LINE INTEGRALS AND PATH DEPENDENCE

our conclusions.
205
12.2 Line integrals and path dependence

In this section we will discuss the problem of obtaining energies from approximate potentials. The exchange-correlation energy for a system with Coulombic two-particle interactions is within density functional theory usually de ned as: 1Z Exc ] = F ] ; Ts ] ; 2 dr1 dr2 j(rr1 ) (rr2j) (12.1) 1; 2 where the universal functional F is as usual de ned by constrained search 18, 128] over all antisymmetric wavefunctions yielding density : ^ ^ F ] = min h jT + W j i (12.2) ! ^ ^ where T is the kinetic energy operator and W is the interparticle interaction operator. The KohnSham kinetic energy Ts ] is the kinetic energy of a noninteracting particle system with the same density as the interacting system de ned by: ^ ^ Ts ] = min h s jT j s i = h s ]jT j s ]i (12.3) !
s
Exc 2 ] ; Exc 1 ] = dt dExc = dt 0 Z 1 Z Z 1 Z ( ( = dt dr Exc) (t)] d dtt) = dt drvxc( (t)] r) d dtt) (r 0 0 in which vxc is the functional derivative of Exc : v ( ] r) = Exc
Z
where we search over all Slater-determinants yielding density and s ] is the Kohn-Sham determinant. The exchange energy functional is usually de ned as 22]: 1Z ) ^ Ex ] = h s ]jW j s ]i ; 2 dr1 dr2 j(rr1 ; (rr2j) (12.4) 1 2 and the correlation functional by: Ec ] = Exc ] ; Ex ] (12.5) Suppose we have a parametrisation (t) of a set of electron densities starting at (0) = 1 and ending in (1) = 2 . Suppose further that we have an exchange-correlation energy functional Exc ]. Then we have:
1
(12.6)
(12.7) (r) If we take for instance the straight path from zero to de ned by (t) = t and we have Exc 0] = 0 (which must be the case for any physically acceptable exchange-correlation functional) then d = dt and it follows that:
xc
Exc ] = dr (r)"xc ( ] r)
(12.8)
206 in which "xc is de ned as:

"xc ( ] r)
Z
This thus yields an explicit expression of the exchange-correlation energy in terms of the exchangecorrelation potential. It should be remembered however that in this expression vxc is a functional derivative. Suppose however that we only have an expression of vxc for instance in terms of the density and its gradients. Then we can de ne the line integral of vxc along a path in the space of densities as: Z Z 1 Z (12.10) v dt drv ( (t)] r) d (t)
xc
dtvxc( t ] r)
(12.9)
As we can see from the above formulas the outcome of this line integral does not depend on the chosen path (except for the endpoints of course) whenever vxc is the functional derivative of some functional Exc . If one uses approximate potentials to obtain energy expressions this way one has to make sure that the result does not depend on the path . One therefore needs some criteria to determine whether vxc yields path independent line integrals. This is for instance the case when: vxc (r) ; vxc (r0) = 0 (12.11) (r0 ) (r) which is a vanishing curl condition which is equivalent to:
xc
dt
Exc (12.12) (r) (r0 ) = (r0 ) (r) To prove this we only have to show that a vxc satisfying the above relation is the functional derivative of some functional Exc .Suppose we de ne the functional Exc by integrating vxc along the straight path (t) = t : Exc ]
Z
2 Exc
dr (r)"xc ( ] r)
(12.13)
in which "xc is de ned as in equation 12.9. If vxc satis es condition 12.11 then vxc is the functional derivative of the above de ned functional Exc and hence path independent. This is readily shown by di erentiation of the above equation 12.13. This yields: Exc = " ( ] r) + Z dr0 (r0 ) "xc ( ] r0 ) = (r) xc (r) = "xc ( ] r) +
Z
Z ] 0 dt dr0 (r0 ) vxc ( tr) r ) (
(12.14)
Now using condition 12.11 yields:
Exc = " ( ] r) + Z 1 dt Z dr0t (r0 ) vxc ( t ] r) = (r) xc (t (r0 )) 0 Z 1 t = "xc ( ] r) + dtt dvxc (dt ] r) = "xc ( ] r)+ 0 Z 1 Z 1 d dt dt (tvxc( t ] r)) ; dtvxc ( t ] r) = vxc( ] r) 0 0
(12.15)
12.3. EXCHANGE-CORRELATION ENERGY AND THE KINETIC PART: BOUNDS FROM POTENTIALS207
Hence vxc is the functional derivative of Exc and therefore path independent. We will now give another criterium for path independence related to equation 12.11. In order to do this we recall Stokes theorem in vector calculus which states for a vector eld ~ in n-dimensional v space:
@v ( @xi ; @vk )dxi dxk (12.16) k @xi i ik O in which is a closed curve which is the boundary of surface O. A generalisation of this theorem to function spaces would be: I Z 0 r v (r)dr = ( v((r0)) ; v((r )) ) (r) (r0 )drdr0 (12.17) r O in which in the rst integral the variations are restricted to be on the path and in the second integral on the surface O. From this equation we can see that if relation 12.11 is satis ed then the integral of v along any closed path is zero and hence the line integral of v is path independent. On the other hand if the line integral is zero for any closed path then we obtain the integrability condition 12.11.
X I
vi dxi =
12.3 Exchange-correlation energy and the kinetic part: bounds from potentials
The present formalism can be used to obtain energy expressions from approximate potentials from line integrals. In order to obtain sensible results from such a calculation one has to make sure that any approximate vxc is either a functional derivative (for instance by requiring the integrability condition 12.11 ) or by requiring that the approximate potential is a good approximation to the exact potential for every density along the integration path. This favors in practical applications some paths over other paths. For instance the path de ned by (t) = t has the disadvantage of not conserving the particle number which can therefore be fractional which gives theoretical problems if one wants to assign a potential to the corresponding density. Another more appealing choice of path is the following: (t) = t3 (tr)
Z Z
(12.18)
If we let the path parameter t run from 0 to 1 then we are integrating from 0 to . The most important feature of this path is that it is particle number conserving, thus: (t)dr = (r)dr = N (12.19) (12.20) along the path. For hydrogen-like atoms for instance we have: 3 (t) = (tZ ) e;tZr
in which Z is the nuclear charge. So if we let t approach zero then the density (t) becomes increasingly di user and approaches zero in every point of space but keeping its normalization. Our main task is now to construct for the case of many electron systems exchange-correlation
208
potentials which approximate the exact exchange-correlation potential along this path. Application of equation 12.6 for this path leads to the following expression for the exchange-correlation energy:
Exc ] =
=
Z
1 0
dt drvxc( (t)] r)(3t2 (tr) + t3 r rtr (tr)) =
Z dt dr 1 vxc( (t)] 1 r)(3 (r) + r r (r)) (12.21) t t 0 in which we performed the substitution tr ! r. Before we go on to discuss the exchange-correlation
potential let us rst discuss the simpler case of the exchange-potential alone. Suppose we have an approximate exchange potential vx for instance of the form: ~
1 vx (r) = 3 (r)f (x y) ~
(12.22) (12.23)
where
x(r) = jr4 (r)j

3(
r)
and
2 y(r) = r 5 (r)
3(
r)
(12.24)
are dimensionless functions. This approximate potential satis es the following well-known scaling property 23] of the exchange potential:
vx ( t ] r) = tvx( ] tr) (12.25) in which t = (t) is just the path of equation 12.18 (in the following we will for notational convenience use t instead of (t) for this particular path). Let vx denote the exact exchange
potential being the functional derivative of the exact exchange functional. The di erence between ~ the exact exchange energy Ex ] and the approximate exchange energy Ex ] using the line integral of equation 12.21 and scaling property 12.25 is then given by: ~ Ex ] ; Ex ] =
Z
vx ; vx = dr(vx ( ] r) ; vx ( ] r)(3 (r) + r r (r)) ~ ~
(12.26)
For a correctly scaling approximate exchange potential the di erence in the two line integrals thus only depends on the di erence between the two potentials at the endpoints of the path and can be made arbitrarily small by better ts of vx to vx at this endpoint. This can in practice be done for ~ instance by tting to the so-called Optimised E ective Potential (OPM) 35, 121]. To show that if vx is not a functional derivative we can make the di erence in the line integral of ~ vx and vx arbitrarily large we consider the following path ~
~ In which R is an arbitrary vector in three-dimensional space. We have: d~ (t) = 3t2 (tr+(1;t)R)+t3 r r ~ ~ ~ (tr+(1;t)R);t3 R r dt
~ tr+(1;t)R
~ ~(t) = t3 (tr + (1 ; t)R)
(12.27)
~ tr+(1;t)R
~ (tr+(1;t)R)(12.28)
EXCHANGE-CORRELATION ENERGY AND THE KINETIC PART
209
Integrating along this path (which can easily be checked to be particle number conserving ) yields using the fact that both the exact vx and the approximate potential vx satisfy some translational ~ invariance property (see equation 12.67 of section 12.5) we nd:
Z
~ We see that this expression can be made arbitrarily large by choosing an arbitrarily large R (which amounts to choosing a di erent path ) if the di erence between vx and vx is not zero. ~ Things are more complicated if we want to calculate the total exchange-correlation energy due to the fact that we do not know the scaling properties of the correlation potential. This means that we cannot calculate the exchange-correlation energy directly from knowledge of vxc ( ] r) alone, we must know vxc along some path (t) in density space. However some useful inequalities can be derived from the knowledge of vxc( ] r). Averill and Painter 30] and Levy and Perdew 22] have derived the following useful relation: Exc ] = drvxc( ] r)(3 (r) + r r (r)) ; Txc ]
where
Z
~ vx ; vx = dr(vx (r) ; vx (r))(3 (r) + r r (r)) ; R ~ ~

~
dr(vx(r) ; vx (r))r (r) (12.29) ~
(12.30) (12.31)
Txc ] = T ] ; Ts ]
is the kinetic part of the exchange-correlation energy. The kinetic energy T ] is as usual de ned by: ^ T ] = h ]jT j ]i (12.32)
^ where T is the kinetic energy operator and ] is the anti-symmetric wavefunction yielding density and minimising the universal functional F ]. As can be inferred directly from equation 12.6 using the scaling path 12.18 it follows from equation 12.30 that equivalently: T ] = dExc (t)] j ; E ] = dEc (t)] j ; E ] (12.33)
xc
which is a useful relation that can be used to obtain the kinetic energy part of the exchangecorrelation energy from approximate exchange-correlation or correlation functionals such as used in LDA and GGA. Relation 12.30 is also useful to obtain an upper bound to the exchange-correlation energy from knowledge of the exchange-correlation potential. From the de nitions for T ] and Ts ] we can deduce the well-known result 5]: ^ ^ Ts ] = h s ]jT j s ]i h ]jT j ]i = T ] (12.34) and thus Txc ] 0 and we nd the following inequality:
dt
t=1
xc
dt
t=1
Exc ]
drvxc( ] r)(3 (r) + r r (r))
(12.35)
which provides an upper bound to Exc. A similar inequality can be obtained for the correlation functional Ec (with corresponding correlation potential vc being the functional derivative of Ec ) by substracting the Levy-Perdew relation 12.26 for the exchange from equation 12.30 with yields:
Ec ] = drvc( ] r)(3 (r) + r r (r)) ; Txc ]
drvc ( ] r)(3 (r) + r r (r)) (12.36)
210
This gives an upper bound to the correlation energy when the correlation potential is known. To provide a lower bound to Ec in terms of potentials we use the following relation valid for systems with Coulombic interparticle interactions proved by Levy 27]: lim 1 E ] = ;b ] (12.37) where b ] is a positive functional satisfying the following inequality: ^ Txc ] ; Ec ] b ] h s ]jW j s ]i where s ] is the Kohn-Sham determinant. Relation 12.37 immediately leads to: Z ^ ;b ] = lim 1 E ] = dEc t ] j = drv ( ] r)(3 (r) + r r (r)) where
t!0 t c t t!0
c t
(12.38) (12.39) (12.40)
dt
t=0
Combining relations 12.38 and 12.33 then gives: 1 ( dEc t ] j + dEc t ] j ) = Z dr 1 (^ ( ] r) + v ( ] r))(3 (r) + r r (r))(12.41) Ec ] 2 dt t=0 c dt t=1 2 vc So we have now derived an upper and a lower bound to the correlation energy functional in terms of the correlation potential. Adding the Levy-Perdew relation for the exchange potential on both sides of the inequalities gives corresponding bounds for the exchange-correlation energy: Z 1v dr 2 (^xc( ] r) + vxc( ] r))(3 (r) + r r (r)) Exc ]
Z
vc ( ] r) = lim0 1 vc ( ] r ) ^ !
drvxc( ] r)(3 (r) + r r (r))
(12.42)
where vxc = vx + vc (using vx = vx which is due to exchange scaling property of vx ). So upper and ^ ^ ^ lower bounds to the exchange-correlation energy can be given from the knowledge of the exchangecorrelation potential at the beginning (t = 0) and the end (t = 1) of the scaling path. We rst will now give an expression for the kinetic part of the exchange-correlation energy. From relation 12.30 and equation 12.6 we nd:
Txc ] = dr(vxc( ] r) ; vxc( ] r))(3 (r) + r r (r))

where
(12.43) (12.44)
vxc ( ] r) =
This gives an explicit expression for Txc in terms of the exchange-correlation potential. It can therefore directly be calculated from approximate expressions of vxc or from the knowledge of vxc along the scaling path. As a result of the scaling property of the exchange potential the above equation can be further split up as follows. Z 1 Z 1 v ( ] r) = dt 1 v ( (t)] r ) + dt 1 v ( (t)] r )
xc
dt t vxc( (t)] r ) t 0
1 1
12.4. CONSTRAINTS ON VC
= vx ( ] r) + where
Z
211 (12.45) (12.46)
dt t vc ( (t)] r ) = vx ( ] r) + vc ( ] r) t 0
1 1
vc ( ] r) =
This leads to another expression for Txc in terms of the correlation potential: Z Txc ] = dr(vc ( ] r) ; vc ( ] r))(3 (r) + r r (r)) (12.47) From Txc ] 0 it follows further that: Z Z drvc ( ] r)(3 (r) + r r (r)) drvc ( ] r)(3 (r) + r r (r)) (12.48) which provides a further constraint on approximate correlation potentials. Further constraints are derived in the next section.
dt t vc( (t)] r ) t 0
1 1
12.4 Constraints on vc
We now will derive some constraints on the correlation potential. Levy has proven the following useful relation 27] which is valid for systems with repulsive Coulombic interparticle interactions: lim E ] > ;1 (12.49) t!1 c t Using relation 12.49 it then follows directly from: Exc t ] = tEx ] + Ec t ] (12.50) that Ex ] = tlim 1 Exc t ] (12.51) !1 t Equation 12.49 immediately leads to a constraint on the correlation potential. If we use the scaling path we nd the following line integral: Z Z E ] = dt dr 1 v ( ] r )(3 (r) + r r (r)) (12.52)
c
Application of equation 12.49 then immediately gives: Z 1 Z dt dr 1 vc ( t ] r )(3 (r) + r r (r)) > ;1 (12.53) t t 0 which puts constraints on approximate correlation potentials. The above constraint is for example not satis ed by the LDA correlation potential due to the logarithmic divergence of the LDA correlation energy as function of the scaling parameter t. We can also write equation 12.51 in line integral form giving: Z Z (12.54) E ] = lim 1 dt dr 1 v ( ] r )(3 (r) + r r (r))
x
Inserting vxc = vx + vc in the above equation and using the scaling property of vx we nd the following constraint on vc: Z Z lim 1 dt dr 1 vc( t ] r )(3 (r) + r r (r)) = 0 (12.55) !1 0 t t The above constraint which is weaker than constraint 12.53 follows also directly from constraint 12.53. The correlation potential of the local density approximation for instance satis es the above equation 12.55 but not constraint 12.53.
!1
xc
212
12.5 Invariance properties of potentials

The above formalism of line integrals provides an easy way to obtain conditions on the energy functional in the case that the exchange-correlation potential has some invariance or symmetry property. Suppose for instance that we vary the densities along our path by varying our path parameter t but that the potential vxc has some symmetry property under such changes. Using equation 12.6 we then can deduce some properties of Exc . In the following we will apply this idea to translation, rotation and scaling properties of Exc . We de ne a path (t) by: (t) = (R(t)r) (12.56) where R(t) is a rotation in threedimensional space which rotates the vector r around a vector ~ ! by an angle t. If the functional Exc is invariant under rotations we have that Exc (t)] = Exc ] is constant as a function of t and we nd: dExc (t)] j = Z dr Exc ( )] d (t) j = 0= dt t= (r) dt t= = drvxc ( ( )] r)(~ R( )r) rR( )r (R( )r) ! For = 0 in particular we nd: 0 = drvxc ( ] r)(~ r) r (r) = ! = drvxc ( ] R( )r)(~ R( )r) rR( )r (R( )r) !
Z Z Z
(12.57)
(12.58)
The above equations 12.57 and 12.58 yield the same result for all densities and all rotations R( ). We thus must have:
vxc ( (R( )r)] r) = vxc ( (r)] R( )r)
(12.59)
So if we insert in vxc the rotated density then we obtain the same value in point r as vxc with the original density in the rotated point R( )r. This result and some other results in this section have been presented elsewhere 23, 32] and are only presented here for clarity and to demonstrate the usefulness of the density path formalism in deriving properties of the potential and functional. The line integral of equation 12.6 o ers the possibililty to make statements about the energy functional when we know properties of the potential. We can now prove the following statement. If a potential vxc satis es the rotation symmetry property of equation 12.59 and if vxc is a functional derivative of some functional Exc then Exc is rotationally invariant. To prove this we insert the path of equation 12.56 in equation 12.6 which gives:
Exc (R( )r)] ; Exc (r)] =

=
Z Z
dt drvxc( (t)] r)(~ R(t)r) rR(t)r (R(t)r) = !

Z
dt drvxc( ] r)(~ r) r (r) = ~ ! !
drvxc ( ] r)r r (r)
(12.60)
INVARIANCE PROPERTIES OF POTENTIALS
213
in which we in the second step used the rotation symmetry property of equation 12.59 and carried out a substitution R(t)r ! r. If we insert = 2 in the above formula 12.60 then we are integrating along a closed path and we obtain:
I
vxc = 2 ! ~
drvxc ( ] r)r r (r)
(12.61)
As vxc is a functional derivative the loop integral should be zero. This is also obvious from the left hand side of equation 12.60 as the 2 -rotated density is equal to the starting density and we obtain: Z 0 = drvxc ( ] r)r r (r) (12.62) for any density . If we insert this equation back into equation 12.60 then we obtain: Exc (R( )r)] = Exc (r)] (12.63) which proves our statement. We will now carry out a similar derivation for the translation properties of the potential. If we de ne the path (t) by: ~ (t) = (r + tR) (12.64) ~ in which R is an arbitrary translation vector. Now suppose that Exc is translationally invariant then we have that Exc (t)] = Exc ] is constant as a function of t and : Z ( 0 = dExc (t)] jt=1 = dr Exc) (1)] d dtt) jt=1 = dt (r
~ ~ = drvxc ( (1)] r)R r (r + R)

Similarly:
Z
(12.65)
~ 0 = dExc (t)] jt=0 = drvxc ( ] r)R r (r) dt

Z
~ ~ ~ = drvxc ( ] r + R)R r (r + R)
(12.66)
As the above equations 12.65 and 12.66 yield the same result for all densities and translation ~ vectors R it follows that: ~ ~ vxc ( (r + R)] r) = vxc( (r)] r + R) (12.67) Thus vxc with the translated density inserted yields the same value in point r as vxc with the ~ original density inserted in point r + R. This result has been derived elsewhere 32] and is only presented here for the sake of clarity. Using the line integral of equation 12.6 we now however also prove the opposite statement under some restrictions. If a potential satis es relation 12.67 and if vxc is a functional derivative of some functional Exc which is bounded from above or below ( The exact exchange-correlation functional is bounded from above as Exc 0 ) then this functional is translational invariant. If we use the path of equation 12.64 and insert it in equation 12.6 we obtain:
Z 1 Z ~ ~ ~ Exc (r + R)] ; Exc (r)] = dt drvxc ( (t)] r)R r (r + tR)
(12.68)
214

Z 1 Z ~ ~ ~ ~ Exc (r + R)] ; Exc (r)] = dt drvxc ( ] r + tR)R r (r + tR) =
If we use the translation symmetry property 12.67 we have:

0
1 0
~ ~ dt drvxc( ] r)R r (r) = R
drvxc( ] r)r (r)
(12.69)
~ in which we carried out a substitution r + tR ! r which makes the t-integration trivial. The right ~ hand side of this equation can be made arbitrarily large (both positive and negative) by making R arbitrarily large. As Exc is bounded from above or below this righthand side of the equation must therefore be zero which yields:
Z
drvxc ( ] r)r (r) = 0
(12.70) (12.71) (12.72)
and therefore one has:
~ Exc (r + R)] = Exc (r)]

We can carry out a similar analysis for scaling properties. We then de ne a path: (t) = t3 (tr)
Suppose we have an energy functional E with functional derivative v which satis es the following scaling relation:
E (t)] = tnE ] with n some positive integer. If we di erentiate the above equation in t = 1 we obtain: Z ( nE ] = dE dt(t)] jt=1 = dr E ) (t)] d dtt) jt=1 = (r
= drv( ] r)(3 (r) + r r (r)) Di erentiating the same expression in t = gives: Z n n;1 E ] = dE dt(t)] = drv( ( )] r)(3 2 ( r) + 3 r r r ( r))
t=
Z
(12.73)
(12.74)
(12.75) (12.76)
This yields:
On the other hand from equation 12.74 it follows: 1 Z drv( ] r)(3 (r) + r r (r)) = E ]=
1 E ] = n dr 1n v( ( )] r)(3 3 ( r) + 4 r r r ( r))
Z
1Z = n drv( ] r)(3 3 ( r) + 4 r r r ( r))
(12.77)
12.6. CALCULATING MOLECULAR BINDING ENERGIES FROM POTENTIALS
215
As the outcome of the integrals in equation 12.76 and 12.77 is equal for any density and scaling parameter it follows:
v( 3 ( r)] r) = n v( ] r)
(12.78)
Thus result has been derived before 23] but is presented here to motivate our following statement. If a potential v satis es the scaling relation 12.78 and is the functional derivative of some functional E with E 0] = 0 then this functional satis es the scaling property 12.73.If we use the line integral of equation 12.6 and insert the path of equation 12.72 we obtain:
E ( )] ; E 0] = E ( )] ; E 0] =
n
Z
If we now use the scaling property 12.78 we have after a substitution:

Z
dt drv( (t)] r)(3t2 (tr) + t3 r rtr (tr)) dt drtn;1v( ] r)(3 (r) + r r (r)) =
Z
(12.79)
= n drv( ] r)(3 (r) + r r (r)) Hence we have:
(12.80)
E ( )] ; E 0] = n(E (1)] ; E 0]) (12.81) In this equation E 0] is just an integration constant. For instance if we add to the functional 12.73 E 3 ( r)] = nE (r)]
(12.82)
an arbitrary constant then it will satisfy the above equation 12.81 and its potential will still satsify 12.78. So if we set E 0] = 0 then we obtain: which proves our statement.
12.6 Calculating molecular binding energies from potentials

One of the most succesful applications of density functional theory has been the calculation of molecular binding energies. This is largely due to the development of gradient corrected functionals. However the potentials corresponding to these functionals are not much improved 32] . An improvement of the existing gradient corrections might be sought in the development of better potentials as they can be compared to exact potentials obtainable from accurate densities. In this section we will discuss how to obtain binding energies from approximate potentials. Suppose we have a diatomic molecule consisting of atom A and atom B . The total density of the molecule is given by M = + where = A + B is the sum of atomic densities and is the deformation density of the molecule. The binding energy contribution of the exchange-correlation energy is then given by: ];Exc A ];Exc B ])(12.83) The rst term between brackets we will call E1 and the second term between brackets we will call E2 . This division is useful in connection with the application of the transition-state method for bond energies developed by Ziegler and Rauk 153] to E1 and is helpful to obtain numerical
Exc = Exc
M ];Exc A ];Exc B ] = (Exc M ];Exc
])+(Exc
216
stability for Coulombic terms analogous to E2 in schemes that rely on numerical integration 154]. The rst term E1 can in practice accurately be calculated using the linear path 154, 155]: (t) = + t( M ; ) = + t (12.84) which yields:
E1 =
This procedure is equivalent to the transition state method 153] as is evident when using the Simpson rule for the t-integration: Z 1 1 E1 = dr (r)( 6 vxc( ] r) + 2 vxc( T ] r) + 6 vxc ( M ] r)) (12.86) 3 1 where T = ( 2 ) = 1 ( + M ) is the so-called transition state density. For practical applications 2 of the above method with approximate exchange-correlation potentials it is necessary that the approximate potential gives a good description of the real exchange-correlation potential in the region where the deformation density is the largest. The other part of the binding energy E2 is easily obtained by numerical integration if explicit exchange-correlation energy expressions are known. However if no energy expressions are known we have to calculate also this term from the potential.One way to do this is is just to calculate the total energies of the atom and the molecule using equations 12.8 and 12.9 and to calculate the di erence. This procedure has however some disadvantages. First of all along the path (t) = t the number of particles is not conserved which presents some di culties from a theoretical point of view. Secondly if one makes an approximate expression for the potential v for some N-electron system, for instance by some tting procedure to some known exact potential, one makes unknown errors for systems with a di erent number of electrons. It is therefore of some advantage to use particle number conserving paths. An other way of calculating the energy di erence E2 is to choose the following path: ~ (t) = A (r) + B (r + tR) (12.87) in which we let t run from 1 to 1. This path is particle number conserving. If we have = (1) then:
dt drvxc( (t)] r) (r) = dr (r)
dtvxc( (t)] r)
(12.85)
E2 =
vxc = dr
Application of the above formula puts some severe restrictions on approximate exchange-correlation potentials. This approximate potential should give a good description of the exact vxc at all bond distances. For instance the bond midpoint peak in vxc in dissociating molecules 86, 90] must also be described by this approximate vxc . This might be feasible for instance in a gradient expansion using Laplacians of the density in any approximate vxc . However care should be taken for potentials which are not functional derivatives as in that case the value of E2 will depend on the dissociation path taken.
~ ~ vxc ( (t)] r)R r B (r + tR)
(12.88)
12.7 Conclusions
In this work we adressed the question how to obtain exchange-correlation energies from approximate exchange-correlation potentials. This is of some theoretical and practical importance as many approximate exchange- and correlation potentials have been proposed. It is shown how one can
CONCLUSIONS
217
use line integrals to express the exchange-correlation energy in terms of the exchange-correlation potentials. It was further shown how to derive symmetry properties of the exchange-correlation energy functional from the corresponding properties of the exchange-correlation potential. We further obtained some upper and lower bounds to the exchange-correlation energy which can be calculated if the exchange-correlation potential is known in two limiting cases, at the beginning and the end of the scaling path. We showed how to express the kinetic part Txc of the exchange-correlation energy in terms of line integrals over the exchange-correlation vxc or correlation potential vc and derived some constraints on the correlation potential. We nally adressed the problem of calculating the exchange-correlation part of molecular binding energies from approximate potentials. The constraints and inequalities derived in this article might prove useful to the development of more accurate exchange-correlation potentials improving the LDA and GGA potentials. Work along this line is in progress.
218
Samenvatting
In dit proefschrift wordt de dichtheidsfunktionaaltheorie (DFT) onderzocht en worden nieuwe benaderingen voorgesteld voor de Kohn-Sham potentiaal. De dichtheidsfunktionaaltheorie, die in de jaren zestig is ontwikkeld door Hohenberg en Kohn en Kohn en Sham, geeft een methode om de grondtoestandseigenschappen van veeldeeltjessystemen zoals atomen, molekulen en vaste sto en te berekenen. In de dichtheidsfunktionaaltheorie zijn de grondtoestandsverwachtingswaarden van de kwantummechanische operatoren die de fysische observabelen voorstellen dichtheidsfunktionalen, d.w.z. het zijn funkties die gede nieerd zijn op de oneindigdimensionale ruimte van elektronendichtheden. Een belangrijke dichtheidsfunktionaal is de grondtoestandsenergiefunktionaal. Deze funktionaal wordt geminimaliseerd voor de grondtoestandselektronendichtheid van het systeem. Minimalisering van de energiefunktionaal leidt dan tot een Euler-Lagrange-vergelijking voor de elektronendichtheid. In de Kohn-Sham aanpak van dichtheidsfunktionaaltheorie krijgt deze variationele vergelijking de vorm van een onafhankelijk deeltjesmodel dat bekend staat als het Kohn-Sham systeem. Het wisselwerkende systeem van elektronen wordt dus vervangen door een niet-wisselwerkend systeem van Kohn-Sham elektronen die bewegen in het veld van een e ektieve potentiaal. Deze Kohn-Sham potentiaal is een funktionaal van de elektronendichtheid. Tengevolge van de ingewikkelde wiskundige de nitie van de KohnSham potentiaal zijn hiervoor geen exakte analytische uitdrukkingen in termen van de dichtheid bekend. De onbekende term in de Kohn-Sham potentiaal is de exchange-korrelatie potentiaal, die de funktionele afgeleide is naar de elektronendichtheid van de exchange-korrelatie funktionaal. Voor praktische toepassingen worden er voor deze funktionaal benaderingen gebruikt. De historisch gezien meest gebruikte benadering is de lokale dichtheidsbenadering. Deze benadering, die gebaseerd is op berekeningen aan het elektronengas, blijkt een simpele en tegelijkertijd verrassend nauwkeurige benadering. De nauwkeurigheid van de lokale dichtheidsbenadering blijkt echter in belangrijke mate te worden vergroot door de zogenaamde gegeneraliseerde gradientbenaderingen, die naast de lokale elektronendichtheid ook afhangen van de gradient van de elektronendichtheid. Om deze laatste reden worden de gradientfunktionalen ook wel nietlokale funktionalen genoemd. De gradientfunktionalen blijken echter niet op alle punten de lokale dichtheidsbenadering te verbeteren. De verbetering geldt vooral voor energietermen, zoals de bindingsenergie van molekulen. De verbetering voor lokale grootheden zoals de exchange-korrelatie potentiaal of de elektronendichtheid is veel minder. Dit betekent bijvoorbeeld ook dat grootheden die direkt van de elektronendichtheid afhangen, zoals polariseerbaarheden, weinig verbetert worden. In dit proefschrift stellen we een aantal benaderingen voor voor de exchange- en exchange-korrelatie potentiaal die de potentialen van de gegenariseerde gradientbenaderingen aanzienlijk verbeteren, terwijl energietermen met gelijke nauwkeurigheid uitgerekend worden. In hoofdstuk 5 van dit proefschrift bekijken we voor een aantal atomen de gradientbenaderingen van Becke en Perdew voor de exchange-korrelatie potentiaal en vergelijken deze met bijna exakte exchange-korrelatie potentialen berekend uit nauwkeurige elektronendichtheden. Om dit te 219
220
SAMENVATTING
doen gebruiken we een nieuw ontwikkelde methode om de Kohn-Sham potentiaal te berekenen uit een gegeven elektronendichtheid. Vervolgens wordt een nieuwe gradientbenadering voorgesteld die in tegenstelling tot de bekende gradientbenaderingen het goede langeafstandsgedrag van de xcpotentiaal weergeeft. De Kohn-Sham eigenwaarden en de ionisatie-energie (die gegeven wordt door de energie van de hoogst bezette baan van het Kohn-Sham systeem) worden hierdoor aanzienlijk verbeterd. In hoofdstuk 6 worden verschillende nieuwe gradientbenaderingen voorgesteld voor de Kohn-Sham exchange potentiaal. Dit zijn Pade-approximanten in termen van variabelen die dimensieloze samenstellingen zijn van de dichtheid en de gradient en de laplaciaan van de dichtheid. Op deze wijze wordt automatisch aan de goede schalingseigenschappen voldaan. De vorm van de Padeapproximant wordt verder zo gekozen dat het langeafstandsgedrag van de potentiaal goed wordt beschreven. De voorgestelde Pade-approximanten blijken de eigenwaarden en de vorm van de potentiaal te verbeteren. De exchange energien die worden berekend met behulp van de Levy-Perdew relatie zijn echter minder goed. Deze energien worden vooral bepaald door de vorm van de potentiaal in het kerngbied van het atoom. Een betere t van de potentiaal in dit gebied verslechtert echter het langeafstandsgedrag van de potentiaal hetgeen de te beperkte vorm van de Pade-approximant aantoont. In hoofstuk 7 analyseren we verschillende delen van de Kohn-Sham potentiaal. De vraag hierbij is hoe de e ekten van het Pauli-principe en de e ekten van de Coulomb-afstoting tussen de elektronen de vorm van de Kohn-Sham potentiaal bepalen. In de uitgevoerde analyse wordt de exchange-korrelatie potentiaal opgesplitst in twee termen, een aantrekkende langedrachtsterm en een repulsieve kortdrachtsterm, die beide relatief eenvoudig te benaderen zijn. Een belangrijk kenmerk van de exchange-korrelatie potentiaal is zijn atoomschillenstruktuur. Deze komt tot uiting in de kortedrachtspotentiaal die een stapvorm heeft. De potentiaal is konstant binnen de atoomschillen en verandert snel op de schilovergangen. De atoomschillenstruktuur komt ook duidelijk tot uiting in de eigenschappen van het exchange- en exchange-korrelatie gat en de konditionele waarschijnlijkheidsamplitude. Beide grootheden beschrijven de verdeling van de andere elektronen wanneer bekend is dat een gegeven elektron, die wij het referentie-elektron noemen, zich op een bepaalde plaats in de ruimte bevindt. Het exchange gat en de konditionele waarschijnlijkheidsamplitude ondergaan grote veranderingen wanneer het referentie-elektron een schilovergang passeert. Omdat verschillende delen van de Kohn-Sham potentiaal direkt in de konditionele waarschijnlijkheidsamplitudes kunnen worden uitgedrukt kan belangrijke informatie worden verkregen over de Kohn-Sham potentiaal door de bestudering van deze waarschijnlijkheidsamplitudes. Bijvoorbeeld de piekenstruktuur op de atoomschilovergangen in de zogenaamde kinetische potentiaal, die een deel is van de funktionele afgeleide van de kinetische energie funktionaal, is direkt gerelateerd aan de veranderingen van de waarschijnlijkheidsamplitude op de schilovergangen. Zoals besproken wordt in hoofdstuk 10 beschrijft dezelfde kinetische potentiaal het links-rechts korrelatie-e ekt in dissocierende molekulen. In hoofdstuk 8 leiden we de stappenstruktuur van de exchange-korrelatie potentiaal die in hoofdstuk 7 was gekonstateerd af. We maken hierbij een benadering voor de funktionele afgeleide naar de elektronendichtheid van de paarkorrelatiefunktie. Een analytische uitdrukking voor deze funktionele afgeleide van de paarkorrelatiefunktie, die verantwoordelijk is voor de stappenstruktuur in de exchange-korrelatie potentiaal, was nog niet eerder afgeleid. De bijbehorende exchange potentiaal is echter op een andere wijze eerder afgeleid door Krieger, Li en Iafrate (KLI). Deze potentiaal blijkt een heel goede benadering voor de exakte Kohn-Sham exchange potentiaal. Een ander belangrijk resultaat is een benaderde uitdrukking voor de inverse dichtheidsresponsfunktie die in verschillende gebieden binnen de dichtheidsfunktionaaltheorie een belangrijke rol speelt (zoals in de konstruktie
SAMENVATTING
221
van potentialen uit dichtheden en in de berekening van energien van aangeslagen toestanden in tijdsafhankelijke dichtheidsfunktionaaltheorie). In hoofdstuk 9 stellen we benaderingen voor voor het stapgedeelte en het langedrachtsgedeelte van de exchange-korrelatie potentiaal. Het langedrachtsgedeelte wordt benaderd door een gradientuitdrukking met het goede langeafstandsgedrag afgeleid uit Beckes gradientfunktionaal voor de exchange energie. De repulsieve stappotentiaal wordt geschreven als een som van orbitaaldichtheden met voorfaktoren die afhangen van de orbitaalenergien. Het analytisch verband tussen deze faktoren en de orbitaalenergien wordt afgeleid door het opleggen van schalingsrelaties, ijkinvariantie en reduktie tot het goede elektronengasresultaat voor homogene systemen. De resulterende potentiaal blijkt een praktische en nauwkeurige weergave te zijn van de exakte exchange potentiaal en leidt bovendien tot goede (van dezelfde kwaliteit als Beckes gradientbenadering) exchange energien. In hoofdstuk 10 en 11 bekijken we molekulaire systemen. In hoofdstuk 10 onderzoeken we de eigenschappen van de Becke en Perdew gradientfunktionalen voor het geval van molekulaire dissociatie. De Becke-Perdew gradientfunktionaal wordt nu binnen de kwantumchemie op grote schaal toepast in molekulaire berekeningen omdat het goede geometrien en ( met een nauwkeurheid van ongeveer 0.2 elektronvolt of minder ) bindingsenergien geeft. We proberen meer inzicht te verkrijgen in de werking van deze gradientfunktionalen in chemische binding. We bekijken hiervoor eerst de eigenschappen van de exakte Kohn-Sham potentiaal in het bindingsmiddelpuntgebied. Zoals eerder opgemerkt vertoont de Kohn-Sham potentiaal in het bindingsmiddelpunt een piekstruktuur die gerelateerd is aan het links-rechts korrelatie-e ekt. Deze piekstruktuur vindt zijn oorsprong in het kinetische gedeelte van de exchange-korrelatie funktionaal en kan worden afgeleid uit het gedrag van de konditionele waarschijnlijkheidsamplitude. We laten zien dat er ook een piekstruktuur aanwezig is in de exchange-korrelatie potentialen van de Becke-Perdew gradientfunktionaal, echter met een verkeerde funktionele afhankelijkheid van de dichtheid. We laten verder zien waarom de Becke gradientfunktionaal voor de exchange in het algemeen een repulsieve bijdrage levert aan de bindingsenergie en waarom de Perdew korrelatiefunktionaal gewoonlijk een aantrekkende bijdrage levert. Dit heeft te maken met het feit dat dichtheidsgradienten in de zadelpunten van de elektronendichtheid (het bindingsmiddelpunt) nul zijn. Numerieke resultaten worden geanalyseerd aan de hand van dichtheidheidsfunktionaalberekeningen aan het stikstofmolekuul. In hoofdstuk 11 bekijken we exakte en benaderde dichtheidsfunktionaalresultaten voor het dissocierende waterstofmolekuul. Het waterstofmolekuul is gekozen voor deze analyse omdat hiervoor de exakte dichtheidsfunktionaalgetallen kunnen worden berekend uit nauwkeurige elektronendichtheden verkregen met grote kon guratie interaktie berekeningen. Een nadeel is echter dat het waterstofmolekuul niet een 'standaard' molekuul is in de zin dat er door de afwezigheid van gevulde kernbanen geen Pauli-repulsie optreedt bij het vormen van de binding. De dichtheidsfunktionaalgetallen worden verder vergeleken met resultaten verkregen in de Hartree-Fock benadering en met de bijna exakte kon guratie interaktie resultaten. Het blijkt dat DFT met gradientfunktionalen een zeer nauwkeurige weergave geeft van de bindingskurve van het waterstofmolekuul. De kurve wijkt echter af van de exakte kurve in de dissociatielimiet, hetgeen echter gekorrigeerd kan worden door het uitvoeren van een onbeperkte Kohn-Sham berekening (Engels:unrestricted KohnSham). De DFT dissociatiefout is echter veel kleiner dan die in de Hartree-Fock benadering. Dit komt omdat het exchange-korrelatie gat in de lokale dichtheidsbenadering en de gegeneraliseerde gradientbenadering gelokaliseerd is rond het referentie-elektron en hierdoor veel meer lijkt op het exakte exchange-korrelatie gat dan het gedelokaliseerde Hartree-Fock exchange gat. Dat er nog steeds een dissociatiefout in DFT bestaat is het gevolg van het feit dat de benaderde dichtheidsfunktionalen niet invariant zijn onder spinrotaties hetgeen to problemen leidt voor gedegenereerde
222
SAMENVATTING
grondtoestanden zoals de open-schil fragmenten van een dissocierend molekuul. In hoofdstuk 12 onderzoeken we het probleem hoe de exchange-korrelatie energie te berekenen indien alleen de exchange-correlatie potentiaal bekend is. Uit de exchange potentiaal kan de exchange energie berekend worden met behulp van de Levy-Perdew relatie die kan worden afgeleid uit de schalingseigenschappen van de exchange funktionaal. Voor korrelatiefunktionalen zijn dergelijke schalingseigenschappen niet bekend. Zoals we aantonen kan de korrelatie-energie echter wel worden afgeleid uit de korrelatiepotentiaal door middel van lijnintegralen in de ruimte van elektronendichtheden. Wegafhankelijkheid van deze lijnintegralen wordt besproken en integreerbaarheidsvoorwaarden worden gegeven. We bespreken verder hoe verschillende invariantie-eigenschappen van de exchange-korrelatie potentiaal kunnen worden afgeleid door het kiezen van paden die overeenkomen met bepaalde symmetrieen (rotatie,translatie,schaling) en leiden ongelijkheden af voor de korrelatiepotentiaal. We laten vervolgens zien hoe het lijnintegraalformalisme de overgangstoestandmethode van Ziegler voor de berekening van molekulaire bindingsenergien generaliseert.
List of publications
R.van Leeuwen and E.J.Baerends Phys.Rev. A49:2421,1994 see chapter 5 An exchange-correlation potential with correct asymptotic behaviour R.van Leeuwen and E.J.Baerends Int.J.Quant.Chem. accepted see chapter 10 An analysis of nonlocal density functionals in chemical bonding O.V.Gritsenko,R.van Leeuwen and E.J.Baerends Int.J.Quant.Chem. submitted see chapter 6 Structure of the optimized e ective Kohn-Sham exchange potential and its gradient approximations R.van Leeuwen, E.van Lenthe, E.J.Baerends and J.G.Snijders J.Chem.Phys. 101:1272,1994 Exact solutions of regular approximate relativistic wave equations for hydrogen-like atoms R.van Leeuwen and E.J.Baerends Phys.Rev.A submitted see chapter 12 Energy expressions in density functional theory using line integrals O.V.Gritsenko,R.van Leeuwen and E.J.Baerends J.Chem.Phys. accepted see chapter 7 Analysis of electron interaction and atomic shell structure in terms of local potentials E.van Lenthe,R.van Leeuwen,E.J.Baerends and J.G.Snijders in New Challenges in Quantum Chemistry (eds).R.Broer,P.J.C.Aerts and P.J.Bagus,Groningen University Press 1994 Regular relativistic two-component Hamiltonians E.van Lenthe,R.van Leeuwen,E.J.Baerends and J.G.Snijders Int.J.Quant.Chem. submitted Regular relativistic two-component Hamiltonians R.van Leeuwen,O.V.Gritsenko and E.J.Baerends Z.Phys.D submitted see chapter 8 Step structure in the atomic Kohn-Sham potential 223
224 O.V.Gritsenko,R.van Leeuwen,E.van Lenthe and E.J.Baerends Phys.Rev.A submitted see chapter 9 A self-consistent approximation to the Kohn-Sham exchange potential
LIST OF PUBLICATIONS
R.van Leeuwen and E.J.Baerends to be submitted see chapter 10 Exchange and correlation: a density functional analysis of the hydrogen molecule
Bibliography
1] 2] 3] 4] 5] 6] 7] 8] 9] 10] 11] 12] 13] 14] 15] 16] 17] 18] 19] P.Fulde. Electron Correlations in Molecules and Solids. Springer-Verlag, Berlin, 1991. L.Hedin and S.Lundqvist. Solid State Physics, 23:1, 1969. E.K.U.Gross and E.Runge. Vielteilchentheorie. B.G.Teubner, Stuttgart, 1986. E.K.U.Gross, E.Runge, and O.Heinonen. Many-Particle Theory. Hilger, Bristol, 1991. R.M.Dreizler and E.K.U.Gross. Density Functional Theory: An Approach to the Quantum Many-Body Problem. Springer-Verlag, Berlin, 1990. R.G.Parr and W.Yang. Density Functional Theory of Atoms and Molecules. Oxford University Press, New York, 1989. R.M.Dreizler and E.K.U.Gross (eds.). Density Functional Theory. NATO ASI Series. Plenum Press,New York, 1994. R.M.Dreizler and J.da Providencia (eds.). Density Functional Methods in Physics. NATO ASI Series B123. Plenum Press,New York, 1985. S.Lundqvist and N.H.March (eds.). Theory of the Inhomogeneous Electron Gas. Plenum Press,New York, 1983. P.Hohenberg and W.Kohn. Phys. Rev., 136B:864, 1964. W.Kohn and L.J.Sham. Phys. Rev., 140A:1133, 1965. R.McWeeny. Rev. Modern Phys., 32:335, 1960. R.Erdahl and V.H.Smith (eds.). Density Matrices and Density Functionals. Reidel, Dordrecht, 1987. A.J.Coleman. Rev. Modern Phys., 35:668, 1963. E.H.Lieb. Phys. Rev. Lett., 46:457, 1981. A.M.K.Muller. Phys. Lett., 105A:446, 1984. M.A.Buijse. Electron correlation. PhD thesis, Vrije Universiteit, Amsterdam, The Netherlands, 1991. M.Levy. Proc. Natl. Acad. Sci. USA, 76:6062, 1979. E.H.Lieb. Int. J. Quantum Chem., 24:243, 1982. 225
226
BIBLIOGRAPHY
20] J.C.Slater. Quantum Theory of Molecules and Solids, volume 4. McGraw-Hill, 1963. 21] R.McWeeny. Methods of Molecular Quantum Mechanics. Academic Press, 1992. 22] M.Levy and J.P.Perdew. Phys. Rev., A32:2010, 1985. 23] Hui Ou-Yang and M.Levy. Phys. Rev. Lett., 65:1036, 1990. 24] O.Gunnarson and B.I.Lundqvist. Phys. Rev., B13:4274, 1976. 25] D.C.Langreth and J.P.Perdew. Solid State Commun., 17:1425, 1975. 26] D.C.Langreth. Phys. Rev. Lett., 52:2317, 1984. 27] M.Levy. Phys. Rev., A43:4637, 1991. 28] A.Gorling and M.Levy. Phys. Rev., B47:13105, 1993. 29] A.Gorling and M.Levy. Phys. Rev., A50:196, 1994. 30] F.W.Averill and G.S.Painter. Phys. Rev., B24:6795, 1981. 31] C.O.Almbladh and U.von Barth. Phys. Rev., B31:3231, 1985. 32] R.van Leeuwen and E.J.Baerends. Phys. Rev., A49:2421, 1994. 33] J.D.Talman and W.F.Shadwick. Phys. Rev., A14:36, 1976. 34] R.T.Sharp and G.K.Horton. Phys. Rev., 90:317, 1953. 35] J.B.Krieger, Yan Li, and G.J.Iafrate. Phys. Rev., A45:101, 1992. 36] E.Engel and S.H.Vosko. Phys. Rev., A47:2800, 1993. 37] J.D.Talman. Computer Phys. Commun., 54:85, 1989. 38] Yan Li, J.B.Krieger, M.R.Norman, and G.J.Iafrate. Phys. Rev., B44:10437, 1991. 39] J.B.Krieger, Yan Li, and G.J.Iafrate. Phys. Rev., A46:5453, 1992. 40] Yan Li, J.B.Krieger, and G.J.Iafrate. Phys. Rev., A47:165, 1993. 41] Yan Li, J.B.Krieger, and G.J.Iafrate. Chem. Phys. Lett., 191:38, 1992. 42] D.J.W.Geldart and M.Rasolt. Phys. Rev., B13:1477, 1976. 43] J.P.Perdew. Physica, B172:1, 1991. 44] A.D.Becke. Phys. Rev., A38:3098, 1988. 45] J.P.Perdew. Phys. Rev., B33:8822, 1986. 46] J.P.Perdew and Wang Yue. Phys. Rev., B33:8800, 1986. 47] J.P.Perdew, J.A.Chevary, S.H.Vosko, K.A.Jackson, M.R.Pederson, D.J.Singh, and C.Fiolhais. Phys. Rev., B46:6671, 1992.
BIBLIOGRAPHY
227
48] J.P.Perdew in. Electronic Structure of Solids 1991. (eds.)P.Ziesche and H.Eschrig, Akademie Verlag, Berlin, 1991. 49] E.Engel, J.A.Chevary, L.D.Macdonald, and S.H.Vosko. Z. Physik, D23:7, 1992. 50] D.C.Langreth and.J.Mehl. Phys. Rev., B28:1809, 1983. 51] D.C.Langreth and J.P.Perdew. Phys. Rev., B21:5469, 1980. 52] A.D.Becke. J. Chem. Phys., 97:9173, 1993. 53] J.P.Perdew. Int. J. Quantum Chem. Symp., 27:93, 1993. 54] T.Zhu, C.Lee, and W.Yang. J. Chem. Phys., 98:4814, 1993. 55] T.Zhu, C.Lee, and W.Yang. J. Chem. Phys., 99:4239, 1993. 56] L.J.Sham and M.Schluter. Phys. Rev. Lett., 51:1888, 1983. 57] L.J.Sham. Phys. Rev., B32:3876, 1985. 58] K.Schonhammer and O.Gunnarsson. J. Phys., C20:3675, 1987. 59] W.Hanke, N.Meskini, and H.Weiler in. Electronic Structure, Dynamics, and Quantum Structural Properties of Condensed Matter, volume B121 of NATO ASI Series. (eds.)J.T.DeVreese and P.van Camp, Plenum Press,New York, 1985. 60] L.Hedin. Phys. Rev., 136:A796, 1965. 61] L.Hedin, A.Johansson, B.I.Lundqvist, S.Lundqvist, and V.Samathiyakanit. Arkiv for Fysik, 39:97, 1968. 62] E.L.Shirley and R.M.Martin. Phys. Rev., B47:15404, 1993. 63] R.W.Godby, M.Schluter, and L.J.Sham. Phys. Rev. Lett., 56:2415, 1986. 64] H.Jeggle. Nichtlineare Funktionalanalysis. Studienbucher Mathematik. B.G.Teubner, Stuttgart, 1979. 65] Y.Choquet-Bruhat, C.DeWitt-Morette, and M.Dillard-Bleick. Analysis, Manifolds and Physics. North-Holland Publishing, Amsterdam, 1978. 66] I.Ekeland and R.Temam. Convex Analysis and Variational Problems. North-Holland Publishing, Amsterdam, 1976. 67] H.Englisch and R.Englisch. Phys. Stat. Sol., 123:711, 1983. 68] H.Englisch and R.Englisch. Phys. Stat. Sol., 124:373, 1983. 69] E.H.Lieb and S.Oxford. Int. J. Quantum Chem., 19:427, 1981. 70] A.D.Becke. J. Chem. Phys., 96:2155, 1992. 71] L.Fan and T.Ziegler. J. Chem. Phys., 95:7401, 1991. 72] M.Levy, J.P.Perdew, and V.Sahni. Phys. Rev., A30:2745, 1984.
228 73] A.D.Becke. J. Chem. Phys., 88:2, 1988.
BIBLIOGRAPHY
74] C.O.Almbladh and A.C.Pedroza. Phys. Rev., A29:2322, 1984. 75] E.K.U.Gross and R.M.Dreizler. Z. Physik, A302:102, 1981. 76] F.Herman, J.P.van Dyke, and I.B.Ortenburger. Phys. Rev. Lett., 22:807, 1969. 77] C.Lee and Z.Zhou. Phys. Rev., A44:1536, 1991. 78] A.G. Gorling. Phys. Rev., A46:3753, 1992. 79] F.Aryasetiawan and M.J.Stott. Phys. Rev., B38:2974, 1988. 80] Yue Wang and R.G.Parr. Phys. Rev., A47:R1591, 1993. 81] M.Levy. Phys. Rev., A26:1200, 1982. 82] R.O.Esquivel and A.V.Bunge. Int. J. Quantum Chem., 32:295, 1987. 83] A.V.Bunge and R.O.Esquivel. Phys. Rev., A33:853, 1986. 84] F.Herman and F.Skilman. Atomic Structure Calculations. Prentice-Hall Inc., 1963. 85] S.H.Vosko, L.Wilk, and M.Nusair. Can. J. Phys., 58:1200, 1980. 86] M.A.Buijse, E.J.Baerends, and J.G.Snijders. Phys. Rev., A40:4190, 1989. 87] W.L.Luken and D.N.Beratan. Theoret. Chim. Acta, 61:265, 1982. 88] V.W.Maslen. Proc. Phys. Soc. (London), A69:734, 1956. 89] G.Sperber. Int. J. Quantum Chem., 6:881, 1972. 90] R.van Leeuwen and E.J.Baerends. Int. J. Quantum Chem., 1994. Accepted. 91] K.Aashamar, T.M.Luke, and J.D.Talman. At.Data.Nucl.Data.Tabl., 22:443, 1978. 92] G.Ortiz and P.Ballone. Phys. Rev., B43:6376, 1991. 93] Y.Wang, J.P.Perdew, J.A.Chevary, L.D.Macdonald, and S.H.Vosko. Phys. Rev., A41:78, 1990. 94] E.Engel and S.H.Vosko. Phys. Rev., B47:13164, 1993. 95] A.D.Becke. J. Chem. Phys., 84:4524, 1986. 96] C.Lee, W.Yang, and R.G.Parr. Phys. Rev., B37:785, 1988. 97] L.Hedin and B.I.Lundqvist. J. Phys., C4:2064, 1971. 98] U.von Barth and L.Hedin. J. Phys., C5:1629, 1972. 99] D.M.Ceperley and B.J.Alder. Phys. Rev. Lett., 45:566, 1980. 100] F.Herman, I.B.Ortenburger, and J.P.Van Dyke. Int. J. Quantum Chem. Symp., 3:827, 1970.
BIBLIOGRAPHY
101] 102] 103] 104] 105] 106] 107] 108] 109] 110] 111] 112] 113] 114] 115] 116] 117] 118] 119] 120] 121] 122] 123] 124] 125] 126] 127]
229
J.C.Slater. Phys. Rev., 81:385, 1951. R.F.W.Bader. Atoms in Molecules: An Quantum Theory. Clarendon Press, Oxford, 1990. R.Daudel, R.F.W.Bader, M.E.Stephens, and D.S.Borrett. Can. J. Chem., 52:1310, 1974. N.H.March. Phys. Lett., A113:476, 1986. N.H.March and R.F.Nalewajski. Phys. Rev., A35:525, 1987. L.J.Bartolotti and P.K.Acharya. J. Chem. Phys., 77:4576, 1982. Q.Zhao and R.G.Parr. Phys. Rev., A46:2337, 1992. O.V.Gritsenko, R.van Leeuwen, and E.J.Baerends. J. Chem. Phys. Accepted. A.Isihara and E.W.Montroll. Proc. Natl. Acad. Sci. USA, 68:3111, 1971. S.H.Vosko and L.D.Macdonald in. Condensed Matter Theories. (eds.)P.Vashishta,R.K.Kalia and R.F.Bishop, Plenum Press,New York, 1987. G.A.Baker. Essentials of Pade Approximants. Academic Press, 1975. J.W.Cooley. Math. Comp., 15:363, 1961. G.Hunter. Int. J. Quantum Chem., 29:197, 1986. B.M.Deb and S.K.Ghosh. Int. J. Quantum Chem., 23:1, 1983. M.Levy and H.Ou-Yang. Phys. Rev., A38:625, 1988. A.Holas and N.H.March. Phys. Rev., A44:5521, 1991. R.F.Nalewajski and P.M.Kozlowski. Acta Physica Polonica, A74:287, 1988. R.Daudel, R.F.W.Bader, M.E.Stephens, and D.S.Borett. Can. J. Chem., 52:1310, 1974. M.A.Buijse and E.J.Baerends in. Electronic density functional theory of molecules,clusters and solids. (ed.)D.E.Ellis, Kluwer, Dordrecht, 1993. W.A.Bingel. Z.Naturforschg., 18A:1249, 1963. O.V.Gritsenko, R.van Leeuwen, and E.J.Baerends. Int. J. Quantum Chem. Submitted. B.G.Johnson, P.M.W.Gill, and J.A.Pople. J. Chem. Phys., 98:5612, 1992. O.Gunnarsson, M.Jonson, and B.I.Lundqvist. Phys. Rev., B20:3136, 1979. M.D.Glossman, L.C.Balbas, A.Rubio, and J.A.Alonso. Int. J. Quantum Chem., 49:171, 1994. O.V.Gritsenko, N.A.Cordero, A.Rubio, L.C.Balbas, and J.A.Alonso. Phys. Rev., A48:4197, 1993. O.Gunnarsson, M.Jonson, and B.I.Lundqvist. Solid State Commun., 24:765, 1977. O.Gunnarsson and B.I.Lundqvist. Phys. Rev., B13:4274, 1976.
230 128] 129] 130] 131] 132] 133] 134] 135] 136] 137] 138] 139] 140] 141] 142] 143] 144] 145] 146] 147] 148] 149] 150] 151] 152] 153] 154] 155]
BIBLIOGRAPHY
E.Lieb. Int. J. Quantum Chem., 24:243, 1982. U.von Barth. Phys. Rev., A20:1693, 1979. O.V.Gritsenko, R.van Leeuwen, E.van Lenthe, and E.J.Baerends. Phys. Rev., A. Submitted. M.K.Harbola and V.Sahni. Phys. Rev. Lett., 62:489, 1989. R.van Leeuwen, O.V.Gritsenko, and E.J.Baerends. Z. Physik, D. Submitted. J.A.Alonso and L.A.Girifalco. Phys. Rev., B17:3735, 1978. J.P.Perdew, R.G.Parr, M.Levy, and J.L.Balduz. Phys. Rev. Lett., 49:1691, 1982. T.Koopmans. Physica, 1:104, 1933. A.Garcia, C.Elsasser, J.Zhu, S.G.Louie, and M.Cohen. Phys. Rev., B46:9829, 1992. G.Ortiz. Phys. Rev., B45:11328, 1992. V.Tschinke and T.Ziegler. Theoret. Chim. Acta, 81:65, 1991. L.Fan and T.Ziegler. J. Chem. Phys., 94:6057, 1991. D.W.Smith, S.Jagannathan, and G.S.Handler. Int. J. Quantum Chem. Symp., 13:103, 1979. D.R.Salahub and M.C.Zerner. The Challenge of d and f Electrons: Theory and Computation. ACS Symposium Series. American Chemical Society,Washington DC, 1989. O.Gunnarsson, P.Johansson, S.Lundqvist, and B.I.Lundqvist. Int. J. Quantum Chem. Symp., 9:83, 1975. O.Gunnarsson and P.Johansson. Int. J. Quantum Chem., 10:307, 1976. O.Gunnarsson, B.I.Lundqvist, and J.W.Wilkins. Phys. Rev., B4:1319, 1974. O.Gunnarsson, J.Harris, and R.O.Jones. J. Chem. Phys., 67:3970, 1977. B.I.Dunlap. Chem. Phys., 125:89, 1988. J.P.Perdew. Phys. Rev., B34:7406, 1986. O.L.Malkina, O.V.Gritsenko, and G.M.Zhidomirov. Chem. Phys. Lett., 171:191, 1990. H.Przybylski and G.Borstel. Solid State Commun., 49:317, 1984. H.Przybylski and G.Borstel. Solid State Commun., 52:713, 1984. R.Latter. Phys. Rev., 99:510, 1955. M.K.Harbola, M.Slamet, and V.Sahni. Phys. Lett., A157:60, 1991. T.Ziegler, A.Rauk, and E.J.Baerends. Theoret. Chim. Acta, 46:1, 1977. E.J.Baerends. Bond energy evaluation. Technical report, Vrije Universiteit, Amsterdam, The Netherlands, 1988. P.M.Boerrigter. PhD thesis, Vrije Universiteit, Amsterdam, The Netherlands, 1987.

KS Equations

Uploaded by

Copyright:

Available Formats

KS Equations

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

KS Equations

Uploaded by

Copyright:

Available Formats

Kohn-Sham potentials in density functional theory

Robert van Leeuwen

October 18, 1994

2 A review of density functional theory

3 A review of functional calculus

4 Density functionals for Coulomb systems

5 An exchange-correlation potential with correct asymptotic behaviour

99 100 102 108 114

8 Step structure in the atomic Kohn-Sham potential

9 A self-consistent approximation to the Kohn-Sham exchange potential

10 An analysis of nonlocal density functionals in chemical bonding

11 Exchange and correlation: a density functional analysis of the hydrogen molecule183

12 Energy expressions in density functional theory using line integrals

Samenvatting List of publications

CHAPTER 1. GENERAL INTRODUCTION

1.2 Overview of the thesis

1.2. OVERVIEW OF THE THESIS

CHAPTER 1. GENERAL INTRODUCTION

A review of density functional theory

2.1 Key concepts and formulas

^+ (r) ^+0 (r0 )w(r r0 ) ^ 0 (r0 ) ^ (r)drdr0

CHAPTER 2. A REVIEW OF DENSITY FUNCTIONAL THEORY

^ (r) ]N ;1 (x1 : : : xN ;1 ) = pN N (x1 : : : xN ;1 x) ^ where A is the antisymmetrization operator de ned as ^ (A )(x1 : : : xN ) =

sign(P ) (xP (1) : : : xP (N ) )

where A B ]+ = AB + BA. We now de ne the electron density as (r) = h j ^(r)j i = h j

^+ (r) ^ (r)j i = N X j (r x2 : : : xN )j2 dx2 : : : dxN

2.2 Density matrices and density functionals

2.2. DENSITY MATRICES AND DENSITY FUNCTIONALS

0 0 (r01 1 : : : r0M M xM +1 : : : xN ) (r1 1 : : : rM M xM +1 : : : xN )dxM +1 : : : dxN (2.20)

CHAPTER 2. A REVIEW OF DENSITY FUNCTIONAL THEORY

(r1 r2 ) i (r2 )dr2 = i i (r2 )

then the set D1 is equal to

2.3. THE PAIR-CORRELATION FUNCTION AND THE EXCHANGE-CORRELATION HOLE13

2.3 The pair-correlation function and the exchange-correlation hole

A REVIEW ON DENSITY FUNCTIONAL THEORY

2.4. EFFECTIVE NON-INTERACTING PARTICLE MODELS

r2)(g 1 2 (r1 r2 ) ; 1)dr2 = ;1

2.4 E ective non-interacting particle models

A REVIEW ON DENSITY FUNCTIONAL THEORY

r0 ; 1 r2 + v(r) + jr0(; )rj dr0 2

The most important contribution of so we nd that approximately

to ni when i is an occupied orbital is the term i (r ) i (r0 ) (2.63) (2.64)

ni (r r0 ) j i (r0 )j2 viHF (r

j i(r0 )j2 dr0 jr ; r0j

2.4. EFFECTIVE NON-INTERACTING PARTICLE MODELS

3 3 1 Z (r) 4 dr 3 1 Ex = ; 8 where = 2=3. The variational equations then become

1 r2 + v(r) + Z (r0 ) dr0 ; 4 ;2 jr ; r0j 3

A REVIEW ON DENSITY FUNCTIONAL THEORY

2.5 The Hohenberg-Kohn theorem

For the same reason one can prove from ^ h jW j i =

j (r x2 : : : xN )j2 dx2 : : : dxN

2.5. THE HOHENBERG-KOHN THEOREM

E0 = min Ev0 ] = min 2N 2N

(r)v0 (r)dr + FHK ]

A REVIEW ON DENSITY FUNCTIONAL THEORY

2.6 The Kohn-Sham equations

2.6. THE KOHN-SHAM EQUATIONS

(2.97) (2.98) (2.99)

A REVIEW ON DENSITY FUNCTIONAL THEORY

Exc ] = Txc ] + Wxc ] Wxc ]