You are on page 1of 18

Away from generalized gradient approximation: Orbital-dependent exchange-

correlation functionals
E. J. Baerends and O. V. Gritsenko

Citation: The Journal of Chemical Physics 123, 062202 (2005); doi: 10.1063/1.1904566
View online: http://dx.doi.org/10.1063/1.1904566
View Table of Contents: http://scitation.aip.org/content/aip/journal/jcp/123/6?ver=pdfcov
Published by the AIP Publishing

Articles you may be interested in


Hierarchy of model Kohn–Sham potentials for orbital-dependent functionals: A practical alternative to the
optimized effective potential method
J. Chem. Phys. 140, 18A535 (2014); 10.1063/1.4871500

Coexistence of 1,3-butadiene conformers in ionization energies and Dyson orbitals


J. Chem. Phys. 123, 124315 (2005); 10.1063/1.2034467

Orbital-dependent correlation energy in density-functional theory based on a second-order perturbation


approach: Success and failure
J. Chem. Phys. 123, 062204 (2005); 10.1063/1.1904584

A semiempirical generalized gradient approximation exchange-correlation functional


J. Chem. Phys. 121, 5654 (2004); 10.1063/1.1784777

Asymptotic correction of the exchange–correlation kernel of time-dependent density functional theory for long-
range charge-transfer excitations
J. Chem. Phys. 121, 655 (2004); 10.1063/1.1759320

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
THE JOURNAL OF CHEMICAL PHYSICS 123, 062202 共2005兲

Away from generalized gradient approximation: Orbital-dependent


exchange-correlation functionals
E. J. Baerends and O. V. Gritsenko
Section Theoretical Chemistry, Vrije Universiteit, De Boelelaan 1083, 1081 HV Amsterdam,
The Netherlands
共Received 19 August 2004; accepted 17 March 2005; published online 17 August 2005兲

The local-density approximation of density functional theory 共DFT兲 is remarkably accurate, for
instance, for geometries and frequencies, and the generalized gradient approximations have also
made bond energies quite reliable. Sometimes, however, one meets with failure in individual cases.
One of the possible routes towards better functionals would be the incorporation of orbital
dependence 共which is an implicit density dependency兲 in the functionals. We discuss this approach
both for energies and for response properties. One possibility is the use of the Hartree–Fock-type
exchange energy expression as orbital-dependent functional. We will argue that in spite of the
increasing popularity of this approach, it does not offer any advantage over Hartree–Fock for
energies. We will advocate not to apply the separation of exchange and correlation, which is so
ingrained in quantum chemistry, but to model both simultaneously. For response properties the
energies and shapes of the virtual orbitals are crucial. We will discuss the benefits that Kohn–Sham
potentials can offer which are derived from either an orbital-dependent energy functional, including
the exact-exchange functional, or which can be obtained directly as orbital-dependent functional.
We highlight the similarity of the Hartree–Fock and Kohn–Sham occupied orbitals and orbital
energies, and the essentially different meanings the virtual orbitals and orbital energies have in these
two models. We will show that these differences are beneficial for DFT in the case of localized
excitations 共in a small molecule or in a fragment兲, but are detrimental for charge-transfer excitations.
Again, orbital dependency, in this case in the exchange-correlation kernel, offers a solution. © 2005
American Institute of Physics. 关DOI: 10.1063/1.1904566兴

I. INTRODUCTION crude and simplistic modeling of the exchange-correlation


term in the total energy, using the electron gas as starting
The success of density functional theory1–4 共DFT兲 in the point 共in the LDA兲 or using refinements employing gradients
description of molecules has been a source of surprise for the of the density 共GGA兲. As we will see, one can understand
quantum chemical community. The theory in its earliest that these simple modelings, irrespective from the way they
manifestation in molecular quantum theory, the exchange-
have been constructed, have been successful: they avoided
only local-density approximation, or X␣ model, looked like
the essential error that the Hartree–Fock model is making.
a poor cousin of Hartree–Fock. So when it proved to do
This error is clear when one considers the statistical two-
better than Hartree–Fock in many cases, this did not sway
electron distribution, which is usually cast in terms of the
public opinion: one cannot 共and should not兲 prove by ex-
exchange-correlation hole: the decrease in probability to find
ample. We will in this paper try to remove the aura of mys-
tery: it is not a miracle that DFT, even with rather crude other electrons in the neigborhood of a reference electron,
approximations such as the local-density approximation compared to the 共unconditional兲 one-electron probability dis-
共LDA兲 or generalized gradient approximations 共GGAs兲, is so tribution. The essential error of Hartree–Fock arises from the
often better than Hartree–Fock. The simple truth is not that fact that in the Hartree–Fock 共HF兲 model the exchange-
LDA/GGA is particularly good, but that Hartree–Fock is correlation hole in a two-electron bond is not centered
rather poor. The Hartree–Fock mean-field theory is actually around the reference electron, but is too delocalized, having
remarkably good for atoms, where the central potential is a considerable amplitude on both atoms involved in the bond.
dominant feature that determines crucial properties such as This is a consequence of the fact that the hole is not gener-
the shell structure, and it is a meaningful approximation for ated by the physics of the Coulomb interaction, but by the
extended systems such as an electron gas. However, Hartree– self-interaction correction and the antisymmetry require-
Fock simply is such a bad model to describe the electron-pair ment. The LDA and GGA models incorporate, simply by
bond, which is the crucial thing to do right for a quantum using an electron-centered hole, an important part of the ef-
chemical theory, that it is not a good starting point for the fect of the interelectron Coulomb repulsion.
theoretical description of molecules, and it is not so difficult It has proved very difficult to improve on the LDA and
to do better. We will discuss the criticism against the GGA models. Recently the refinement of introducing orbital
Hartree–Fock model for electron-pair bonds in Sec. III. dependence, not just density and gradient dependence, into
Density functional theory has provided a seemingly the functionals has received much attention. This is indeed a

0021-9606/2005/123共6兲/062202/17/$22.50 123, 062202-1 © 2005 American Institute of Physics


This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-2 E. J. Baerends and O. V. Gritsenko J. Chem. Phys. 123, 062202 共2005兲

possible way forward in the endeavour to improve function- H

als and put them on a more transparent physical basis. We ␳共r1兲 = 2 兺 兩␾i共r1兲兩2 . 共2.2兲
will in Sec. II briefly mention the approaches that have been i

taken to solve the Kohn–Sham equations when orbital de- We use closed-shell systems throughout, with H = N / 2 dou-
pendence is introduced in the functionals. Then, in Sec. III, bly occupied orbitals. Both the total density and the indi-
we discuss the most generally used orbital-dependent func- vidual orbitals are used in DFT as the arguments in approxi-
tional, the exact-exchange 共EXX兲 functional. This functional mations to the unknown exact functional of the xc energy
uses a Hartree–Fock-type energy expression for the ex- Exc关␳兴. While in the early LDA Exc was modeled as a func-
change energy. The only difference is that Kohn–Sham orbit- tion of only the local density ␳共r1兲, Exc
LDA
关␳兴 = Exc
LDA
关␳共r1兲兴,6
als instead of HF orbitals are used. Although this orbital- the GGA employs, in addition, a “semilocal” characteristic
dependent functional is receiving much attention, in of the density, its gradient ⵱␳共r1兲, Exc GGA
关␳兴
particular, since it automatically incorporates the self- = Exc 关␳共r1兲 , ⵱␳共r1兲兴,
GGA 7–11
and next-generation functionals
interaction correction, the lack of which is believed to cause include explicit dependence on individual KS orbitals ␾i.
problems for the LDA and GGA functionals, we will argue Recently, the orbital-dependent exact-exchange 共EXX兲
that this functional cannot be expected to be better than the group of methods has received much attention in the litera-
Hartree–Fock model itself, and therefore to be actually ture. One may approximate the xc functional by just the
poorer than the LDA and GGA functionals. In Sec. IV we exchange-only form that appears in the well-known Hartree–
will discuss possible ways to incorporate both exchange and Fock 共HF兲 energy expression
correlation in orbital-dependent functionals. We will demon-


H H
strate that it is possible in this way to actually avoid the main ␾*i 共r1兲␾ j共r1兲␾i共r2兲␾*j 共r2兲
deficiency of the Hartree–Fock and EXX models, and obtain Ex关兵␾i其兴 = − 兺 兺 dr1dr2 .
i j r12
a proper description of electron correlation even in weak
electron-pair bonds. In Sec. V and VI we will discuss the 共2.3兲
rather more subtle role that orbital-dependent functionals
Since the Kohn–Sham potential is a local multiplicative po-
play in the calculation of response properties. In that case
tential vs共r兲, the optimization problem to obtain the orbitals
already the EXX functional offers important advantages, be-
that minimize the total energy with the exchange part 共2.3兲 is
cause its functional derivative, the effective potential, is con-
slightly different from Hartree–Fock. The additional con-
siderably improved over the LDA/GGA ones. It yields im-
straint of locality of the potential will result in orbitals that
proved orbital energies and orbital shapes for the virtual
may be slightly different from the HF ones, and therefore the
orbitals, hence improved excitation energies, and it is able,
EXX energy 共the expectation value of the determinant with
through the orbital-dependent response part of its potential,
EXX orbitals兲 will be slightly higher than the HF energy 共see
to generate the counteracting electric field that is required to
Sec. III for examples兲. The KS potential, with its exchange
obtain realistic 共hyper兲polarizabilities compared to the huge
part vx can be obtained in this case with the self-consistent
overestimation by LDA and GGA. Finally 共Sec. VI兲, the dra-
solution of the integro-differential equations of the so-called
matic failure of the LDA and GGA approximations for
optimized effective potential 共OEP兲 method.12 Talman and
charge-transfer excitations can be solved by the introduction
Shadwick12 derived these equations by considering explicitly
of an orbital-dependent exchange-correlation kernel.
the optimization problem of finding the local potential that
gives the orbitals yielding minimum total energy of the
single determinantal wave function, but they may also be
straightforwardly derived by applying the chain rule 共taking
II. ORBITAL-DEPENDENT EXCHANGE-CORRELATION the general case of an orbital-dependent xc functional兲
FUNCTIONALS IN DFT ␦Exc关兵␸i关␳兴其兴
vxc共r兲 =
␦␳共r兲
With the advent of the orbital-dependent exchange-
correlation 共xc兲 functionals modern DFT comes to explore
the full potentiality of the Kohn–Sham 共KS兲 theory.5 How- =兺
i
冕 ␦Exc ␦␸i共r1兲␦vs共r2兲
dr dr + c.c.
␦␸i共r1兲 ␦vs共r2兲␦␳共r兲 1 2
ever, orbital-dependent xc functionals are technically much
harder to treat than the density and gradient dependent func- ␦Exc
=−兺 G 共r ,r 兲␸ 共r 兲␹−1共r ,r兲dr1dr2
tionals. The KS orbitals ␾i are obtained from the one- i ␦␸i共r1兲 i 1 2 i 2 s 2
electron equations 共throughout the paper we consider closed-
shell systems兲 + c.c., 共2.4兲

再 冎
where we have used the orbital Green’s function Gi共r1 , r2兲
1 = 兺 j⫽i␸ j共r1兲␸ j共r2兲* / 共␧ j − ␧i兲 which describes the response of
− ⵜ2 + vs共r1兲 ␾i共r1兲 = ␧i␾i共r1兲, 共2.1兲
2 an orbital to a change of potential ␦vs: ␦␸i共r1兲
= −兰Gi共r1 , r2兲␸i共r2兲␦vs共r2兲dr2. Gi共r1 , r2兲 follows directly
with the local, state-independent potential vs, and the subset from first-order perturbation theory. The constraint that the
of occupied orbitals 兵␾ j其 determines the electron density ␳ of KS orbitals be solutions in a local KS potential vs共r兲 is ex-
the system plicitly applied in the optimized potential method, but is here
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-3 Orbital-dependent exchange-correlation functionals J. Chem. Phys. 123, 062202 共2005兲

introduced by the use of the chain rule in the functional the true differences 共␧ j − ␧i兲 for the occupied-occupied pairs
derivation in Eq. 共2.4兲. Multiplying by the density response are retained. The exchange potential vCEDA x obtained with
function ␹s共r , r3兲 = ␦␳共r兲 / ␦vs共r3兲 = −兺i=1
N/2
f i␸i共r兲*Gi共r , r3兲 this approximation is also expressed in terms of only the
⫻␸i共r3兲 + c.c. and integrating over r we obtain an integral occupied KS orbitals.24,25 It is partitioned naturally into
equation for vxc: physically meaningful components

冕 vCEDA 共r1兲 = vhole


x 共r1兲 + vresp 共r1兲, 共2.7兲
CEDA
x
vxc共r兲␹s共r,r3兲dr
the Slater potential vhole
x of the Fermi 共exchange兲 hole 共see


CEDA
below兲 and the response potential vresp ,
=兺 vi共r1兲␸*i 共r1兲Gi共r1,r3兲␸i共r3兲dr1 + c.c., 共2.5兲 H H
␾*i 共r1兲␾ j共r1兲
i CEDA
vresp 共r1兲 = 兺 兺 sCEDA , 共2.8兲
i
ij
j ␳共r1兲
with vi共r1兲 = 关1 / ␸i共r1兲*兴␦Exc / ␦␸i共r1兲. Several methods of so-
lution of Eq. 共2.5兲 have been explored. Talman and Shadwick which contains the characteristic orbital-shell “step” struc-
used inversion of ␹s on a numerical grid in one-dimensional ture with “diagonal” contributions from the orbital densities
共atomic兲 calculations. In principle, these equations can be 兩␾i兩2 共a similar structure is also present in the KLI potential兲
solved by using basis set expansions, as pointed out by as well as with additional “off-diagonal” contributions from
Görling.13–15 Validation and application of this approach has the occupied-occupied orbital products ␾i␾*j . The orbital step
been done by Bartlett and co-workers.16,17 The function ␹s structure follows from the fact that in the region, where a
can be inverted on a subspace of its eigenfunctions, such that particular orbital density 兩␾i兩2 brings a dominant contribution
the eigenfunctions with zero 共or near zero for practical pur- to the total density ␳, the potential 共2.8兲 is close to the cor-
poses兲 eigenvalues are excluded. To obtain vx of high quality, responding constant 共“step height”兲 sCEDA ii . Note, that the
a large basis is required, in particular, the KS basis should same potential vCEDA
x has been obtained by Della Sala and
include all occupied and a large number of unoccupied or- Görling26 within the so-called localized Hartree–Fock 共LHF兲
bitals ␾i. The numerical stability of this inversion becomes a approximation, in which the assumption is made that the
problem, especially for molecular applications, see Ref. 17 occupied KS orbitals are only unitary transforms of the oc-
for a detailed discussion. Recently, in Refs. 18 and 19, vx has cupied HF orbitals, i.e., the HF and KS determinants are
been constructed by using the Fermi–Amaldi potential, identical. For closed-shell systems, the derivation based on
which is a simple functional of ␳共r1兲 with the correct long- this LHF assumption leads to exactly the vCEDA x of Eq. 共2.7兲.
range asymptotics, vx → −1 / 兩r1兩, 兩r1兩 → ⬁, as a constant part For open-shell systems the LHF assumption cannot be used,
of the potential 共using, e.g., the LDA ␳兲, and expanding the and in that case a variant of the CEDA approximation has
unknown remainder in a finite basis set for which the coef- been proposed in Ref. 27. Krieger et al. have already derived
ficients are to be determined. Since the Fermi–Amaldi part a method28 to improve on the KLI approximation and de-
already takes care of the asymptotic part of the potential, the velop the full EXX solution. A numerical procedure for this
remainder can be expanded in Gaussian basis functions with further refinement of the KLI/CEDA potential has been de-
finite extent. The expansion coefficients are obtained in an veloped by Kümmel and Perdew,29 which also employs only
iterative optimization procedure. The EXX method of occupied KS orbitals. It updates vKLI/CEDA
x with a correction
Görling,20 which also uses basis set expansion, does not re- built from the products of the occupied orbitals ␾i and the
quire the inversion of ␹s. Instead, vx is obtained as the po- first-order perturbation theory corrections to ␾i for the per-
tential of the exchange charge density ␳x, turbing potential 共vix − vx兲, where vix is the HF exchange po-
tential for ␾i.
vx共r1兲 = 冕 ␳x共r2兲
r12
dr2 , 共2.6兲 We conclude that several methods are available to obtain
solutions of the optimized potential method for orbital-
dependent functionals. We will not go into further technical
from the solution of the corresponding Poisson’s equation details.
with a basis set expansion of ␳x.
Good-quality approximations to EXX have been ob-
III. A CRITIQUE OF THE EXCHANGE-ONLY
tained with the Green’s function Gi being broken up in a
APPROACH „EXX… IN DFT
local 共␦-function兲 part and a remainder that is approximated.
An example is the admittedly crude Sharp and Horton ap- Although it would be technically possible to deal with
proximation for Gi,21 which does not preserve the proper all kinds of orbital-dependent functionals, one may observe
orbital structure of the response function ␹s. Surprisingly, in the literature primarily application of the exact-exchange
this approximation yields the good-quality exchange poten- functional of Eq. 共2.3兲. 共Notable exceptions are the work by
tial vKLI
x of the Krieger, Li, and Iafrate 共KLI兲 method,22,23 Bartlett and co-workers,30,31 who use many-body perturba-
which is expressed in terms of only the occupied KS orbitals. tion theory and coupled-cluster methods to obtain more ad-
Recently, a more physically reasonable common energy de- vanced functionals, and our own work,32 to be discussed in
nominator approximation 共CEDA兲 has been applied to the the following section, which uses functionals inspired by
Green’s function Gi.24 According to CEDA, the orbital en- density matrix functional theory.兲 Two advantages of the
ergy differences for the occupied-unoccupied pairs in Gi are EXX functional are often mentioned: 共a兲 the self-interaction
approximated with the mean energy ⌬␧ ˜ , ␧c − ␧i ⬇ ⌬␧
˜ , while correction is incorporated in this functional: 共b兲 it is natural
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-4 E. J. Baerends and O. V. Gritsenko J. Chem. Phys. 123, 062202 共2005兲

TABLE I. Correlation contributions 共in eV兲 to the various energy compo-


nents of the H2 molecule at equilibrium distance 共2.4 bohrs兲 and at elon-
gated distances.

corr corr corr


Etotal Tcorr Vel-nuc Wel-el

H2共R = Re兲 −1.1 eV +1.3 −0.5 −1.9


H2共R = 5.0 bohrs兲 −3.9 +8.9 −8.5 −4.3
H2共R = 10.0 bohrs兲 −6.3 +7.9 −8.6 −5.6

tron from around the right nucleus and nothing around the
left nucleus. The reference electron should see the right
nucleus unscreened by density of the other electron, which
will reside around the left nucleus. It can then build a den-
sity, which is like a H atom density, with correct behavior
FIG. 1. Fermi hole, Coulomb hole, and total hole in the hydrogen molecule
around the nucleus and with the correct exponential decay.
at R = 1.4 共equilibrium distance兲 and 5.0 bohrs. The reference electron is This is of course in agreement with reality, where dissocia-
always placed at 0.3 bohr to the left of the right H atom. tion will be into two H atoms. The total hole has indeed
precisely this required shape. The Fermi hole, however, re-
to break up the total problem of electron-electron interaction moves half of an electron from around the left and from
in the large 共due to self-interaction correction兲 exchange part around the right nucleus each, so it leaves half of an electron
and the small correlation part. The hope is of course that it around the right nucleus, which will partly screen that
will be simpler to find accurate density functionals for the nucleus. The wave function for the HF/EXX electron will
small correlation energy. We wish to argue that, on the con- therefore become too diffuse. This is very clearly illustrated
trary, starting with EXX is actually making life much more with the breakdown of the correlation contribution to the
difficult. It undermines the very basis of the success of DFT, total energy in contributions to the kinetic energy, the
which has been the breaking away from the quantum chem- electron-nuclear attraction energy, and the electron-electron
istry paradigm of dividing up the full electron-electron inter- interaction energy in Table I. A too diffuse orbital, with too
action problem into two steps: first Hartree–Fock 共or low gradients, will lead to a too low kinetic energy, and
exchange-only兲 and next the remaining correlation problem. indeed the correlation correction to the kinetic energy, Tcorr
That this is indeed the case is masked by the DFT terminol- = T − TEXX, is strongly positive. Also, the correlation correc-
ogy which still uses the names “exchange” 共e.g., Becke ex- tion to the density will make it more compact around the
change, Perdew exchange兲 and “correlation” 关e.g., Perdew– nucleus, hence will make the electron-nuclear energy more
Wang correlation, Lee–Yang–Parr 共LYP兲 correlation兴. Let us, negative. We note that the correlation corrections in the one-
however, consider the simple electron-pair bond. One should electron energies T and Ven are at long distance individually
corr
be aware that, while correlation is indeed much smaller than larger than the two-electron correlation correction Wel-el .
exchange in for instance 共heavy兲 atoms, this is not true in the They are large compared to the total bond energy 共
electron-pair bond. This can be illustrated with the very −4.7 eV兲 already at Re, and become very large compared to
simple picture given in Fig. 1. Figure 1共a兲 displays the the much smaller bond energies at longer bond distance. This
Fermi, Coulomb, and the total xc holes for the prototype substantiates the well-known fact that the Hartree–Fock
electron-pair bond, the H2 molecule, at the equilibrium H–H 共EXX兲 model is a poor zero-order approximation for the
distance R共H–H兲 = 1.4 bohrs,33,34 while in Fig. 1共b兲 the holes electron-pair bond. It is also extremely size inconsistent, in
are plotted for the stretched H2 at R共H–H兲 = 5.0 bohrs. In the the sense that when we introduce a second H atom in the
H2 case HF and EXX-KS are identical, since the HF operator universe, remote from a given H atom, the HF/EXX model
in this case is just the local potential that is obtained by totally fails to deliver the sum of the energies of two H
correcting the full Hartree potential of the total density atoms. It is interesting to note that when we add the Cou-
␳共r2兲 = 2兩␴g共r2兲兩2, VH共r1兲 = 兰␳共r2兲 / r12dr2, with the self- lomb hole 共or correlation hole兲 to the Fermi hole, we obtain
interaction correction from the potential at r1 due to the a total hole that is localized at the nucleus where the refer-
Fermi hole, containing −1 electron, ␳hole x 共r2兩r1兲 = −兩␴g共r2兲兩 .
2 ence electron is located. This will provide the right potential,
In this case the Fermi hole is completely independent of the and therefore the right density and kinetic energy. The local-
position r1 of the reference electron, and is delocalized over ized hole does not require any knowledge of where the other
the two centers between which the bond exists. As becomes H atom is located, it is the unphysical breaking up of this
particularly apparent at long bond distance, the resulting po- total hole into a delocalized exchange part and an equally
tential in which the HF/EXX electron moves is very defi- delocalized Coulomb part that complicates the electron-
cient. When the reference position is chosen, as indicated in electron interaction treatment and requires knowledge of the
Fig. 1, close to the nucleus at the right, the other electron is position of the second nucleus in order to build accurate
most likely not close to the same nucleus, but will with great holes. Actually, this example illustrates that the division in
probability be somewhere around the left nucleus. In other exchange and correlation holes creates one of the largest
words, the hole should remove density integrating to −1 elec- problems of DFT.35 The problem basically is that the total
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-5 Orbital-dependent exchange-correlation functionals J. Chem. Phys. 123, 062202 共2005兲

TABLE II. Components of the LDA, BLYP, CEDA, HF, KS, and CI energies 共in atomic units兲 for CO at the
equilibrium bond distance R共C–O兲 = 2.13 bohrs. The difference of the exact 共CI兲 quantity with respect to the
model quantity is also given. For the exchange energy WX the KS value is the benchmark. The LDA and BLYP
WX values have been calculated from the corresponding exchange functionals. Eel is the sum T + Ven + WH
+ WX; the difference of the exact 共CI兲 energy with respect to Eel is by definition the correlation energy. The Ec
entries in the last row under LDA and BLYP are calculated with the corresponding DFT functionals and can be
compared with the required Ec numbers in the row above.

Energy LDA BLYP EXX 共CEDA兲 HF KS CI

T / Ts 111.951 113.181 112.395 112.641 112.881 113.185


⌬CI +1.234 +0.004 +0.790 +0.544 +0.304

Ven −310.170 −311.520 −310.651 −310.879 −311.256 −311.256


⌬CI −1.086 +0.264 −0.605 −0.377 0.00

WH 76.204 76.391 76.251 76.262 76.399 76.399


⌬CI +0.195 +0.008 +0.148 +0.137 0.00

WX −12.064 −13.475 −13.296 −13.331 −13.319 共−14.089兲


⌬KS −1.255 +0.156 −0.023 +0.012 共WXC兲

Eel = T + Ven + WH + WX −134.079 −135.423 −135.301 −135.307 −135.295


⌬CI = Ec = Etotal
CI
− Eel −1.682 −0.338 −0.460 −0.454 −0.466

Ec 共DFT兲 −0.950 −0.486


Etotal 135.029 −135.909 −135.761
⌬CI −0.732 +0.148

density ␳ around each H nucleus is identical at large H–H treatments will start to be somewhat different. There may be
distances to that in a free H atom. However, the exchange several pairs of nonbonding electrons in core shells and lone
and correlation energies in the stretched H2 are very different pairs, which account for almost all of the energies of various
from those in two isolated H atoms. In particular, the corre- kinds, such as kinetic energy, electron-nuclear potential en-
lation energy is zero in individual H atoms, in contrast with ergy, etc. The exchange-only approximation is not so poor
the strong left-right nondynamical correlation in the dissoci- for these orbitals, so this approximation may appear better
ating H2 molecule. It is virtually impossible to devise func-
justified. However, for the electron-pair bonds the exchange-
tionals that use only local information 共local density and de-
only approximation is as bad as in the prototype case of the
rivatives of the density兲 and still recognize the position of the
other H atom and build correctly the individual exchange and H2 molecule, and therefore the errors are large compared to
correlation holes, the shape of which is determined by the the bond energy. We illustrate these statements, and make a
positions of the nuclei. comparison between HF and KS-EXX for the CO molecule
The LDA and GGA total holes are localized around the 共Tables II and III兲. These tables compare various components
reference electron, which is a definite advantage. Figure 1 in of the HF and EXX energies of the prototype molecule CO,
Ref. 36 shows plots of the LDA and GGA exchange holes in calculated at the equilibrium bond length R共C–O兲
H2 around reference positions some 0.18 Å to the left and to = 2.13 bohrs 共Table II兲 and at elongated bond length R
the right of the left nucleus. These pictures illustrate the per- = 2.8 bohrs 共Table III兲, with the energy components of an
fect spherical symmetry of the LDA hole around the refer- accurate KS solution for CO. In the following we just use the
ence position, and for the somewhat more realistic GGA hole acronym “EXX” to denote the Kohn–Sham exchange-only
the displacement of the maximum depth of the hole in the
results, which have been obtained with the CEDA approxi-
direction opposite the density gradient, i.e., towards the
mation, since they are very close approximations to the full
nearby nucleus. In the dissociating H2 molecule, this prop-
erty of LDA and GGA to localize the exchange hole around EXX data. The accurate KS solution has been obtained from
the reference electron makes that the hole more closely ap- the correlated density ␳ of the multireference configuration
proximates the total hole rather than the exchange hole. interaction 共MRCI兲 method with the iterative local updating
However, the LDA and GGA exchange holes appear to be procedure of van Leeuwen and Baerends 共LB兲.37 This
too shallow in the dissociating H2 molecule to describe with method is based directly on the Hohenberg–Kohn 共HK兲 theo-
good numerical accuracy the strong left-right correlation.32 rem. The HK theorem compares the Hamiltonians Ĥ and Ĥ⬘,
LDA produces a very large error 共ca. 70 kcal/ mol!兲 at large which are characterized by different external potentials v and
R共H–H兲, although not as large as the HF method. GGA re-
v⬘, and have different ground-state wave functions ⌿ and
duces the error in the dissociation limit compared to LDA,
⌿⬘. If one adds the two inequalities, following from the
but still yields a large error in this case 共ca. 45 kcal/ mol, see
also Refs. 32 and 35 and the following section兲. variation theorem,
In bonds between many-electron atoms the HF and EXX
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-6 E. J. Baerends and O. V. Gritsenko J. Chem. Phys. 123, 062202 共2005兲

TABLE III. Components of the LDA, BLYP, CEDA, HF, KS, and CI energies 共in atomic units兲 for CO at the
stretched bond distance R共C–O兲 = 2.8 bohrs. The difference of the exact 共CI兲 quantity with respect to the model
quantity is also given. For the exchange energy WX the KS value is the benchmark. The LDA and BLYP WX
values have been calculated from the corresponding exchange functionals. Eel is the sum T + Ven + WH + WX; the
difference of the exact 共CI兲 energy with respect to Eel is by definition the correlation energy. The Ec entries in
the last row under LDA and BLYP are calculated with the corresponding DFT functionals and can be compared
with the required Ec numbers in the row above.

Energy LDA BLYP EXX 共CEDA兲 HF KS CI

T / Ts 111.023 112.257 111.437 111.662 111.977 112.270


⌬CI +1.247 +0.013 +0.833 +0.608 +0.293

Ven −298.862 −300.239 −299.216 −299.430 −299.777 −299.777


⌬CI −0.915 +0.462 −0.561 −0.347 0.00

WH 71.038 71.241 71.071 71.045 71.073 71.073


⌬CI +0.035 −0.168 +0.002 +0.028 0.00

WX −11.752 −13.168 −12.973 −13.027 −12.980 共−13.822兲


⌬KS −1.228 +0.188 −0.007 +0.047 共WXC兲

Eel = T + Ven + WH + WX −128.553 −129.909 −129.738 −129.750 −129.707


⌬CI = Ec = Etotal
CI
− Eel −1.703 −0.347 −0.518 −0.506 −0.549

Ec 共DFT兲 −0.935 −0.472

Etotal −129.489 −130.381 −130.256


⌬CI −0.767 +0.125

具⌿兩Ĥ⬘兩⌿典 ⬎ 具⌿⬘兩Ĥ⬘兩⌿⬘典 series of atoms and molecules.37–43 Since the KS orbitals are
obtained in the iterative process, the KS kinetic energy Ts
⇒T+W+ 冕⬘v ␳dr ⬎ T⬘ + W⬘ + 冕 ⬘⬘v ␳ dr
and the KS exchange energy Ex become available. The tables
also contain the corresponding energy contributions for the
LDA and GGA-BLYP 共Refs. 7 and 10兲 energies. All calcu-
and lations have been performed in the correlation-consistent po-
具⌿⬘兩Ĥ兩⌿⬘典 ⬎ 具⌿兩Ĥ兩⌿典 larized core-valence quadruple zeta added 共cc-pCVQZ兲 basis
set44 using the ATMOL package45 and its DFT extension.46
⇒ T⬘ + W⬘ + 冕 v␳⬘dr ⬎ T + W + 冕 v␳dr,
Of course, since the KS density is identical to the CI
one, the electron-nuclear attraction energy Ven and the Har-
tree energy of the electrostatic electron repulsion WH are
one obtains “exact” in the KS case. It is enlightening to see how much

冕⬘ 冕 冕 ⬘⬘ 冕
HF
the electron-nuclear HF energy Ven differs from the exact
v ␳dr + v␳⬘dr ⬎ v ␳ dr + v␳dr one. Although in this case the total HF number is percentage-
wise not so bad, due to the large contributions of nonbonding
or 共particularly core兲 electrons, the absolute error of −0.377 at


Re 共−0.347 hartree at R = 2.8 bohrs兲 is of the same order as
⌬v⌬␳dr ⬍ 0, the total bond energy of ca. 0.40 hartree. The errors are also
given in the table, always as the difference of the benchmark
where ⌬␳ = ␳⬘ − ␳ is the difference between the densities be- CI value with the value for a particular case; in the Hartree–
longing to the potentials v and v⬘ = v + ⌬v. The HK theorem Fock case: ⌬CI共Ven兲 = Ven CI HF
− Ven HF
. The fact that the HF Ven is
concludes that if ⌬v is nonzero, also ⌬␳ must be nonzero, not negative enough shows that the HF orbitals and density
i.e., a change in the potential v must induce a change in the are too diffuse. This is also reflected in the HF kinetic energy
density ␳. The relation 兰⌬v⌬␳dr ⬍ 0, however, not only tells term, which is less positive by 0.544 hartree 共0.608 hartree at
us that there is a unique mapping between density and po- R = 2.8 bohrs兲 compared to the CI kinetic energy. 共In the re-
tential, but also that, if a target density ␳⬘ differs in a small mainder of this parargaph numbers in parentheses will refer
region from a trial density ␳ obtained with a trial potential v, to the elongated bond distance, Table III.兲 The electron-
one may approach the target density more closely using an electron repulsion energy WH is, as expected for a diffuse
HF
update ⌬v on the potential v which is negative if ⌬␳ is density, too low. Especially in the case of Ven and THF, the
positive and vice versa. So using a ⌬v proportional to −⌬␳ correlation errors are not much smaller than the overall cor-
will help to approximate v⬘. This affords an iterative proce- relation energy of −0.454 hartree 共−0.506 hartree兲. One
dure to determine the potential that approaches the target would expect these errors to persist in the EXX determinan-
density ␳⬘ arbitrarily closely, which has been applied to a tal wave function. As a matter of fact, they are still consid-
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-7 Orbital-dependent exchange-correlation functionals J. Chem. Phys. 123, 062202 共2005兲

erably larger in the EXX case: −0.605 hartree Re兲 of TCI with respect to TBLYP is not so relevant, the large
共−0.561 hartree兲 for EXX
Ven and +0.790 hartree difference of −0.30 a.u. of TKS with TBLYP demonstrates the
共+0.833 hartree兲 for T . So the added constraint that the
EXX
too contracted nature of the BLYP orbitals, and is mirrored in
CI BLYP
potential be local makes the orbitals and density even more the +0.264 difference Ven − Ven . The WX numbers illustrate
diffuse than they already are at the HF level. It is interesting the well-known ca. 10% error of the LDA exchange energy
to observe that these large effects in the individual energy and the considerable correction towards the exact exchange
terms are not reflected in the variationally determined total provided by the GGA model. We use the KS exchange en-
energy: the energy of the EXX determinant is only 0.006 ergy as the exact-exchange energy; it is the same exchange
hartree 共0.012 hartree兲 higher than the HF energy. This is to energy expression as the HF one, but it is evaluated with the
be compared to the large total electronic energies 共sum T KS orbitals instead of the HF ones. There remains a sizable
+ Ven + WH + WX兲 EEXX = −135.301 hartree 共−129.738 hartree兲 error in the BLYP exchange energy 共0.156 at Re兲. The tables
and EHF = −135.307 hartree 共−129.750 hartree兲, and the cor- also contain the total electronic energies calculated using the
relation energies of EEXX c = −0.460 hartree 共−0.518 hartree兲 LDA or GGA density functionals for the exchange energy,
and EHFc = −0.454 hartree 共−0.506 hartree兲. i.e., Eel is in all cases the sum T + Ven + WH + WX. The differ-
Comparing next to the determinant with the KS orbitals, ence of the CI energy with respect to this Eel energy is by
we note that these orbitals and hence also the density are definition the correlation energy. The correlation energy Ec is
contracted by the correlation effects embodied in the KS overestimated by the well-known ca. 100% by the LDA
model, so the KS energies Ts , Ven, and WH are all substan- functional, at least when compared to the standard correla-
tially larger 共in absolute magnitude兲 than the corresponding tion energy of ca. −0.45 a.u. at Re. However, it is interesting
HF and EXX energies. The terms Ven and WH are actually to note that the other energy terms in the LDA energy are so
exact, but Ts is not yet equal to the exact T, it is still a sizable poor that actually the correlation error of the LDA calcula-
0.304 hartree 共0.293 hartree兲 too low, about half of the error tion is, with −1.682 a.u., much larger, and the LDA correla-
of the HF case. Note, that in DFT the energy EKS of the KS tion energy of −0.950 a.u. is too low! The BLYP correlation
determinant is taken as the reference for the correlation en- energy is much better, although the agreement is by no
CI
ergy, Ec = Etotal − EKS. Since the lowest energy of the single means perfect. The BLYP correlation energy of −0.486 a.u.
determinant is the HF energy EHF, the absolute value of the at Re is fairly close the standard correlation energy 共KS
DFT correlation energy Ec must be larger than the conven- model兲 of −0.466 a.u., but it is rather too strongly negative
tional correlation energy EHF CI HF
c = Etotal − E , though the actual compared to the remaining correlation error of EBLYP el of
difference 共0.012 hartree for the equilibrium CO distance, −0.338 a.u. The implication is that the total energy of the
0.043 hartree for R = 2.8 bohrs兲 is much smaller than the dif- BLYP model 共−135.909 at Re, −130.381 at R = 2.8 bohrs兲 is
ferences between HF and KS in the individual kinetic and too negative compared to the CI energy. The 共empirical兲 ex-
potential energy terms. This is, of course, related to the fact act nonrelativistic total energy will be lower than the CI total
that for the HF orbitals a variational minimum in the total energy, but is not known in the case of CO. It is not clear
energy is obtained, so that distortion to KS orbitals will not therefore how closely the BLYP total energy happens to be to
affect the total energy greatly. this empirical total energy.
We note that the numbers at the elongated bond length We observe that the BLYP exchange energy is consis-
give the same overall picture as those at Re, with maybe a tently too negative compared to the KS exchange energy. It
slight increase of the errors. The CO molecule is not excep- has been observed that a more consistent interpretation of the
tional, essentially the same conclusions have been drawn LDA and GGA models is obtained if we conclude from the
from a similar comparison for other diatomics such as N2 , F2, fact that the LDA and GGA exchange holes are localized
and Li2.41 In summary, the HF and EXX models significantly around the reference position so that these model holes in
distort, compared to the KS model, the density and orbitals, fact do not represent exchange, but the sum of exchange and
and therefore the separate energy terms, but there is error nondynamical correlation. In a number of cases where rea-
cancellation when adding these terms up. One might also sonable estimates of the nondynamical correlation energies
take the opposite view: the KS model changes the orbitals were available, it has been shown41 that the GGA exchange
and density considerably from the HF and EXX models to energy indeed represents with good quantitative accuracy the
bring the separate energy terms much closer to the exact sum of the KS exchange energy and the nondynamical cor-
ones, without being punished very much in the total energy relation energy. In its turn, the GGA correlation energy ac-
of the KS determinantal wave function. ccurately approximates the dynamical correlation energy,
Tables II and III also contain the corresponding data for which can be obtained as the sum of the total correlation
the LDA and the BLYP GGA models. Apparently, these mod- energy minus the nondynamical correlation energy. The total
els do not provide one-electron potentials that lead to accu- correlation energy to be used here is not the one obtained
rate densities and orbitals. The total density and the orbitals from the CI calculations in Table II, since it deviates 共due to
are much too diffuse for LDA 共too low TLDA and WH LDA
, not basis set deficiency and the necessarily limited CI兲 too much
negative enough Ven 兲, but it appears that the BLYP ap-
LDA
from the conventional correlation energy, defined as the dif-
proximation overcorrects, already at Re but particularly for ference of the exact total energy and the Hartree–Fock en-
the longer bond distance, towards too contracted orbitals and ergy. The “empirical correlation energy” can be obtained
density 共too large energy components兲. Note that TBLYP from the “empirical” total energy established for a number of
should be equal to TKS, so the small 0.004 a.u. difference 共at molecules.47–49 Unfortunately, the empirical total energy is
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-8 E. J. Baerends and O. V. Gritsenko J. Chem. Phys. 123, 062202 共2005兲

not available for CO, so we cannot check the finding of Ref. mations from ab initio theory.31 Also, expressions based on
41 for N2 , F2, and Li2 that the GGA correlation energy is random-phase approximation or Green’s functions can be
indeed close to the dynamical part of the total correlation used.53–57 However, we have argued that one should not take
energy. Hartree–Fock as the starting point. Since modeling of the
One can argue that here is the real origin of the success total exchange-correlation hole appears to be easier than
of DFT in its density/gradient forms. In reality 共maybe not modeling of the exchange and correlation holes separately, it
intentionally by the originators of LDA and GGA兲 the DFT should be feasible to develop orbital-dependent expressions,
approximations have not tried to solve the total electron- which represent exchange and correlation simultaneously.
electron interaction problem by doing first exchange and We, therefore, highlight here a development, which does not
then correcting with the correlation energy. Instead, DFT has use perturbation theory, but obtains an expression for the
broken away from that partitioning of the electron-electron exchange-correlation hole that is valid for both dynamic cor-
interaction, and first treats exchange plus nondynamical cor- relation 共where perturbation theory might work兲 and nondy-
relation in one step 共by what is, unfortunately, called the namic or near-degeneracy correlation 共where perturbation
“exchange functional”兲 and in the next step obtains the dy- theory typically fails兲. This approach makes use of recent
namical correlation 共called “correlation”兲 from expressions developments in the theory of functionals of the one-electron
based on the correlation energy of the electron gas. The no- density matrix35,46,58 共see also Refs. 59–64兲. In the latter
tion that even the exchange-only LDA model 共i.e., X␣兲 al- theory the exchange-correlation hole surrounding a reference
ready contains left-right correlation has been widespread 共cf. electron at position r1 is written as the square of an xc hole
the terminology exchange-correlation potential by Slater50兲 amplitude ␸xc hole
: ␳xc
hole
共r2兩r1兲 = −兩␸xc
hole
共r2兩r1兲兩2. The concept of
and has, e.g., been elaborated on by Cook and Karplus.51 It a hole amplitude has been introduced earlier by Luken for
should be noted, though, that Becke has recently argued in the pure exchange hole.65,66 When we write the exchange
favor of restoring the traditional steps of first doing exchange energy 共for a closed-shell system with all doubly occupied
and next the correlation.52 orbitals, ␥ is the one-electron density matrix兲,


Though we have been critical about the exact-exchange
1
functional, since one cannot expect it to do any better than Ex = − ␥共r1,r2兲␥共r2,r1兲dr1dr2
the Hartree–Fock model, we should point out that in one 4
important respect the Kohn–Sham exact-exchange model
differs from Hartree–Fock. The Kohn–Sham potential is a
local potential, which is the same for the occupied and the
=
1
2
冕 冕 ␳共r1兲
␳hole
x 共r2兩r1兲
r12
dr2 共4.1兲

virtual orbitals. The Hartree–Fock model effectively uses a which defines the exchange hole as
different potential for the virtual orbitals. The virtual elec-
trons in the HF model see a potential due to all N electrons, 1 ␥共r1,r2兲␥共r2,r1兲
␳hole
x 共r2兩r1兲 = − , 共4.2兲
not that of N − 1 electrons, as the HF electrons in occupied 2 ␳共r1兲
orbitals do, and the Kohn–Sham electrons in both occupied
we find immediately from the fact that the one-electron den-
and virtual orbitals. The orbital energies of the virtual orbit-
sity matrix is, in the case of a closed-shell single determinant
als therefore are considerably different in the Kohn–Sham
wave function,
and the Hartree–Fock models. When the exact-exchange
functional is used in the Kohn–Sham model, the optimized H

potential remains a single local potential for all states, occu- ␥共r1,r2兲 = 兺 2␸i共r1兲␸*i 共r2兲, 共4.3兲
pied and virtual. The difference of the virtual orbital energies i=1

with respect to HF therefore remains. For excitation energy and that ␳hole can be written as the square of an amplitude,
x
calculations in the time-dependent DFT, where the orbital
energies and orbital shapes play an important role, the dif- 冑2␸*i 共r1兲
N/2 N/2
冑2␸ j共r1兲 *
␳hole =−兺 ␸i共r2兲 兺
ference between HF and Kohn–Sham EXX will be very im- x 共r2兩r1兲
i=1
冑␳共r1兲 j=1
冑␳共r1兲 ␸ j 共r2兲
portant. We will return to this point in Sec. V.
⬅ − 兩␸hole
x 共r2兩r1兲兩 .
2
共4.4兲
IV. A „VIRTUAL… ORBITAL-DEPENDENT XC
FUNCTIONAL The amplitude ␸hole
x ,
also called the Fermi amplitude, can in
general be expanded in a complete one-electron basis,
The KS EXX approximation cannot be regarded to be ␸hole
x 共r2兩r1兲 = 兺ci␸i共r2兲. The expansion coefficients will de-
the next improvement beyond GGA, which has now been pend on the reference position, and in the exchange case they
awaited for a long time but has not materialized yet. Still, have the simple form
such improvement can come from explicitly orbital-
dependent functionals. The requirement should be that such 冑2␸*i 共r1兲
ci共r1兲 = for occupied orbitals
orbital-dependent functionals incorporate correlation effects. 冑␳共r1兲
Such functionals can be based on perturbation theoretic ex-
共4.5兲
pressions, such as the second-order Görling–Levy perturba-
ci共r1兲 = 0 for virtual orbitals.
tion expression, which is the DFT equivalent of the well-
known Møller–Plesset one, cf. the OEP-MBPT共2兲 work of This expression helps to understand the properties of the
Bartlett and co-workers,30 or on still more advanced approxi- Fermi hole. When at a reference point a specific orbital ␸i
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-9 Orbital-dependent exchange-correlation functionals J. Chem. Phys. 123, 062202 共2005兲

makes a predominant contribution to the density, the hole guments Arias and Csanyi63兲. Following these density matrix
will have the shape of 共the negative of兲 that orbital density. functional theory arguments, a similar expression has been
This explains why the exchange hole has the shape of a core proposed recently in Refs. 32 and 35 for the amplitude of the
orbital density when the reference point is inside that orbital, coupling constant integrated xc hole in DFT, with all H oc-
or a lone pair shape when the hole is in that region. In the cupied ␾ j and a finite number Nv of virtual KS orbitals ␾a
case of H2 the summation reduces to just one orbital, the being used instead of the NOs 共M = H + Nv兲,
occupied ␴g, and with ␳共r1兲 = 2兩␴g共r1兲兩2 we obtain

␸hole
冑2␴g共r1兲 ¯␸xc
hole
共r2兩r1兲 = 兺
M
冑 wi *
␾ 共r1兲␾i共r2兲,
␳共r1兲 i
共4.11兲
x 共r2兩r1兲 =
冑2兩␴g共r1兲兩2 ␴g共r2兲 = ␴g共r2兲, i

共4.6兲 M M
冑wiw j
␳hole
x 共r2兩r1兲 = − 兩␴g共r2兲兩 , =−兺兺
2
¯␳xc
hole
共r2兩r1兲 ␾*i 共r1兲␾i共r2兲␾ j共r1兲␾*j 共r2兲.
i j ␳共r1兲
which is in agreement with our earlier observation that this
hole is independent of the reference postion r1 and extends 共4.12兲
over the two nuclei whatever the distance between the two This yields the following orbital-dependent potential of the
nuclei. hole共BB兲
xc hole v̄xc 共BB—Buijse–Baerends兲
It is possible to extend this simple representation of the
hole to the full exchange-correlation case.58 The amplitude is
again expanded in a complete set, for instance, in the set of
hole共BB兲
v̄xc 共r1兲 = − 冕 ␳xc
hole
共r2兩r1兲
r12
dr2
natural orbitals 共NOs兲, M M
冑wiw j
␳xc
hole
共r2兩r1兲 =− 兩␸xc
hole
共r2兩r1兲兩2 , =−兺兺 ␾*i 共r1兲␾ j共r1兲
i j ␳共r1兲
共4.7兲
␸xc
hole
共r2兩r1兲 = 兺 ci共r1兲␹i共r2兲.

In this way one defines a trial two-electron density matrix,


⫻ 冕 dr2
␾i共r2兲␾*j 共r2兲
r12
共4.13兲

BB
and the xc energy functional Exc
⌫trial共r1,r2兲 = ␳共r1兲␳共r2兲 − ␳共r1兲␳xc
hole
共r2兩r1兲. 共4.8兲
BB
Exc 关兵␾ j其,兵␾a其兴
One can derive from symmetry properties and sum rules of
the two-electron density matrix 关such as permutation symme- M M

兺 兺 冑w i w j
1
try in r1 and r2; integration of ⌫ over r2 should yield 共N =−
2 i j
− 1兲␳共r1兲; integration of the hole density should yield −1兴
that the coefficients should be
冑ni␹*i 共r1兲
⫻ 冕 dr1dr2
␾*i 共r1兲␾i共r2兲␾ j共r1兲␾*j 共r2兲
r12
. 共4.14兲
ci共r1兲 = 共4.9兲
冑␳共r1兲 , Here wi关␳兴 is the xc orbital weight, which governs the in-
volvement of the occupied/virtual orbital ␾i in the xc func-
where ni are the NO occupation numbers. Evidently, not just
tional. The weights wi关␳兴 are the important novel density
the heavily occupied orbitals, which resemble the occupied
functionals. We note that the link with the EXX functional
Hartree–Fock orbitals, enter, but also the weakly occupied
can be made when we restrict the summations to just the H
ones. The latter are not similar to virtual Hartree–Fock orbit-
occupied orbitals, and take for the weights wi the occupation
als, but are much more compact. In the case of dissociating
H2, the wave function reduces to 共1 / 冑2兲兩␴g共r1兲¯␴g共r2兲兩
numbers 2.0.
− 共1 / 冑2兲兩␴u共r1兲¯␴u共r2兲兩, and the hole amplitude becomes
By giving weights to the virtual orbitals we can incorpo-
rate the effect of correlation, as is clear from the discussion
冑ng␴g共r1兲 冑nu␴u共r1兲 of the density matrix theory. The weights have to take into
␸xc ␴g共r2兲 +
共r2兩r1兲 =
冑␳共r1兲 ␴u共r2兲.
hole
冑␳共r1兲 account that the full coupling constant integrated hole in-
cludes both kinetic energy and electron-electron potential en-
共4.10兲 ergy effects of electron correlation and also they have to
compensate for the approximate form of the BB functional
One can easily verify that this representation of the total
共4.4兲. We use overbars in ¯␳xchole hole
and v̄xc to indicate that we
exchange-correlation hole obeys the required localization on
are dealing with the coupling constant integrated quantities.
the site where the reference position is located 共in the limit
In Ref. 32 the functional form of wi has been approximated
where ng = nu = 1.0兲: if r1 is close to nucleus A共r1 苸 ⍀A兲,
共r2兩r1 苸 ⍀A兲 = 共1 / 冑2兲关␴g共r2兲 + ␴u共r2兲兴 = 1sA共r2兲,
with the Fermi-type distribution
␸xc
hole
and
when r1 is close to nucleus B, ␸xc 共r2兩r1 苸 ⍀B兲 = 共1 / 冑2兲
hole
2
⫻关␴g共r2兲 − ␴u共r2兲兴 = 1sB共r2兲. wi = , 共4.15兲
1 + exp关f共␧i − ␧F兲兴
Effectively, a representation of the two-electron density
matrix in terms of the one-electron density matrix has been where ␧i are the KS orbital energies in Eq. 共1.1兲 and ␧F is the
obtained 共see for further generalization of the symmetry ar- Fermi level parameter.
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-10 E. J. Baerends and O. V. Gritsenko J. Chem. Phys. 123, 062202 共2005兲

The various OEP approaches13,14,16–18,20,67,68 which have ployed; however, the contributions of the derivatives
until now been applied mostly to the EXX functional involv- ␦wi关␳兴 / ␦␳共r1兲 to the potential vxc
BB-CEDA
were neglected.
ing only occupied orbitals, can, in principle, be extended to These derivatives contribute only to the response potential
treat functionals, which depend on both occupied and virtual BB-CEDA
vresp , so that they do not affect the leading term of
orbitals. In Ref. 32 a generalized CEDA 共Ref. 24兲 has been BB-CEDA
vxc
hole共BB兲
, the xc hole potential v̄xc . On the other hand, a
employed to develop a variational, self-consistent KS functional dependence wi关␳兴 could, in principle, be devised
method with the BB functional 共4.14兲 and 共4.15兲. Within in such a way that its derivative would cancel the asymptotic
CEDA, all terms of the exact KS Green’s function 共2.5兲 with divergence in the potential 共4.20兲. This divergence is a seri-
M occupied and virtual orbitals included in Eqs. 共4.14兲 and ous problem of functionals of virtual orbitals. It is immedi-
共4.15兲 are treated rigorously, while in the remaining terms ately apparent from Eq. 共4.20兲 that asymptotically the den-
the common energy denominator ⌬␧ ˜ is used. The resultant sity ␳共r1兲 in the denominator, which decays as the highest
CEDA Green’s function has the following form occupied orbital density, will tend to zero faster than the
M ⬁ numerator if virtual orbitals appear in the numerator. The
␾ j共r1兲␾*j 共r2兲 1
GCEDA
i 共r1,r2兲 =兺
ji ␧ j − ␧i
+
⌬␧
兺 ␾a共r1兲␾*a共r2兲
˜ a⬎M
potential will therefore exhibit a divergence when r → ⬁. The
divergence of the potential has been highlighted in Refs. 70
␦共r1 − r2兲 1
M and 71 for the second-order Görling–Levy 共GL2兲 perturba-
=
⌬␧
˜

⌬␧
兺 dij␾ j共r1兲␾*j 共r2兲,
˜ j
tion theoretic Exc. There has been some debate about the
question whether the divergence is an inherent property of
共4.16兲 every functional that includes virtual orbitals. According to
Niquet et al. this is not the case. In, for instance, the GL2
˜ 共1 − ␦ij兲兴 / 共␧ j − ␧i兲. The function 共4.16兲
where dij = 关␧ j − ␧i − ⌬␧ expression for Exc, and indeed other perturbation theoretic
is to be inserted in the master OEP equation, which in this expressions based on the Green’s function, one can show that
general case has the following form
the optimized potential will exhibit proper −1 / r decay for

冕 r → ⬁ due to cancellation of divergent terms in the part of the


H
2 兺 ␾*i 共r1兲 vxc共r2兲Gi共r1,r2兲␾i共r2兲dr2 + c.c. potential containing derivatives with respect to the orbitals
i
against divergent terms from derivatives with respect to or-


M bital energies. This is a quite subtle effect. In the investigated
= 兺 ␾*i 共r1兲 i
vxc 共r2兲Gi共r1,r2兲␾i共r2兲dr2 + c.c., 共4.17兲 functionals it does not arise from term-by-term cancellation
i
of expressions with finite summations over virtual orbitals,
i
where the nonlocal potentials vxc are defined as but builds up when complete summations are carried

冋 册
through. This may make it difficult to devise an orbital en-
1 ␦Exc关兵␾i其兴 ergy dependence of the functional 共in our case through the
i
vxc 共r2兲 = . 共4.18兲
␾i共r2兲 ␦␾*i 共r2兲 orbital dependency of wi兲 in such a way that the cancellation
is obtained when using a finite number of virtual orbitals.
BB
Inserting Exc of Eq. 共4.14兲 in Eq. 共4.18兲 and GCEDA
i in Eq.
Alternatively, since the exact potential is known to exhibit
共4.17兲, and then solving Eq. 共4.17兲 for the potential vxc, we
the −1 / r asymptotic behavior, one may constrain in the op-
obtain the BB-CEDA xc potential in the characteristic CEDA
timization the asymptotic behavior of the potential to this
form
form. This is effectively the strategy of Wu, Ayers, and Yang
BB-CEDA
vxc 共r1兲 = v̄xc
hole共BB兲
共r1兲 + vresp
BB-CEDA
共r1兲. 共4.19兲 in Ref. 72 to corect the wrong asymptotic behavior of the
LDA and GGA potentials. In the same vein, in Ref. 32 the
The main xc effect is represented with the first term of Eq.
density in the denominators of the potentials 共4.15兲 and
共4.19兲, the potential 共4.13兲 of the xc hole 共4.12兲. In particular,
hole共BB兲 共4.20兲 was replaced with the function ˜␳共r1兲 = 兺iM wi兩␾i共r1兲兩2,
having the component v̄xc , the BB-CEDA potential rep-
i.e., the xc weights wi were interpreted as effective occupa-
resents correctly the effect of strong left-right correlation in
tion numbers in an effective density ˜␳共r1兲. It is expected that
the important special case of stretched H2 considered above.
˜␳共r1兲 is actually a close approximation to ␳共r1兲, except for
The second term of Eq. 共4.19兲 is the response potential
the crucial difference in how these two densities tend asymp-
M M
␾*i 共r1兲␾ j共r1兲 totically to zero.
BB-CEDA
vresp 共r1兲 = 兺 兺 sBB-CEDA . 共4.20兲 The 共virtual兲 orbital-dependent xc energy functional Exc BB
i
ij
j ␳共r1兲
共4.14兲 and its xc potential ␯xc
BB-CEDA
共4.19兲 have been applied
The diagonal terms corresponding to occupied orbitals rep- in Ref. 32 in self-consistent KS calculations of the potential
resent the characteristic step structure of the response part of curve of the H2 molecule. Figure 2 compares the total energy
the KS potential,23,69 which is responsible for the well- curve obtained with this method with those calculated with
known “bumps” in the KS potential at atomic shell bound- HF and full configuration interaction 共FCI兲 methods as well
aries. The BB-CEDA potential has a similar orbital structure as with LDA and GGA-BP BP—Becke 88–Perdew 86 共Refs.
to that of the exchange-only CEDA potential 共2.7兲 and 共2.8兲, 7 and 9兲 functionals. The most dramatic is the HF failure due
which is now extended to virtual orbitals. to the neglect of the Coulomb correlation, which was dis-
In the implementation of Ref. 32 a particular param- cussed above. The HF curve goes higher than other curves at
etrized form of the xc weight functional 共4.15兲 was em- nearly all H–H distances and at R共H–H兲 = 10 bohrs the error
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-11 Orbital-dependent exchange-correlation functionals J. Chem. Phys. 123, 062202 共2005兲

V. MOLECULAR RESPONSE PROPERTIES WITH


ORBITAL-DEPENDENT XC FUNCTIONALS:
POLARIZABILITIES
We have seen that orbital-dependent xc functionals can
improve greatly over LDA and GGA functionals for the en-
ergy. Another important field of application of orbital-
dependent functionals is in response properties, calculated
with time-dependent DFT, where sometimes the purely den-
sity dependent functionals exhibit fundamental weaknesses.
The orbital dependence 共which implies an implicit density
dependence兲 can introduce features in the density functionals
that would be very hard to obtain in an explicitly density
dependent form.
Until now, the EXX model and its close approximations
have been applied within time-dependent DFT 共TDDFT兲 to
the calculation of various molecular response properties,
FIG. 2. Total energy of the hydrogen molecule as a function of internuclear such as 共hyper兲polarizabilities and electron excitation ener-
distance for the Hartree–Fock approximation 共HF兲, the local density ap-
gies, see Refs. 19, 25, and 73–75. While the EXX model has
proximation 共LDA兲, the Becke88–Perdew86 generalized gradient approxi-
mation 共BP兲, a full CI calculation, and the Buijse–Baerends 共BB兲 exchange- deficiences for total and bond energies, similar to Hartree–
correlation functional. Fock, for response properties this close resemblance between
KS-EXX and HF need not necessarily apply in view of the
of the HF total energy is EHF − EFCI = 148 kcal/ mol and the difference between the shapes and energies of virtual orbitals
the error of the dissociation energy is DHF FCI
e − De in the two models. In the present section we will discuss
= 123 kcal/ mol. 共hyper兲polarizabilities, and find that KS-EXX is actually
Spin- and symmetry-restricted LDA and GGA-BP also close to Hartree–Fock. It is of course not necessary to use the
produce too high energy of the dissociating H2 共see Fig. 2兲, functional derivatives of improved energy functionals in or-
although not as high as HF. In particular, near the equilib- der to improve on the deficient orbital energy spectrum of
rium distance the LDA curve is close to the HF one, while at LDA/GGA. The KS orbital energy spectrum may also be
larger R共H–H兲 it goes in between the HF and FCI curves. At modeled directly by introducing functionals, possibly orbital
R共H–H兲 = 10 bohrs the LDA total energy error is ELDA dependent, for the KS potential itself. Recently, a relatively
− EFCI = 69 kcal/ mol and the LDA dissociation energy error simple orbital-dependent potential with statistical averaging
is DLDA
e − DFCI
e = 46 kcal/ mol. The gradient correction lowers of 共model兲 orbital potentials 共SAOP兲 共Ref. 76兲 has been ap-
the GGA-BP energy compared to the LDA one and the BP plied to the response properties, including excitation ener-
curve goes below and almost in parallel to the LDA curve. gies, of small prototype molecules77,78 and to the calculation
Near the equilibrium the BP curve is close to the FCI one, of electron excitations of practically important porphyrin
however, the BP energy is still much too high at long dis- systems and their complexes.79,80 Excitation energies will be
tances. The BP total energy error at R共H–H兲 = 10 bohrs is discussed in the following section.
EBP − EFCI = 44 kcal/ mol and the BP dissociation energy error A remarkable effect of the full KS response theory,
is DLDA
e − DFCI
e = 47 kcal/ mol, even slightly larger than the which requires orbital mechanisms for its explanation and
LDA one. orbital-dependent functionals for its adequate description, is
In contrast to the LDA and GGA curves, the potential the field-counteracting effect of the xc potential of molecular
curve of the self-consistent 共virtual兲 orbital-dependent BB chains24,25,81,82 and extended dielectric systems83–88 in an ex-
method, taken with the two-parameter Fermi-distribution ternal electric field, compared to the LDA or GGA approxi-
functional 共4.15兲, excellently reproduces the FCI curve, both mation. Specifically, a characteristic ultranonlocal term 关not
curves practically coincide at all H–H distances 共see Fig. 2兲. describable by a local-density 共or density-gradient兲 func-
The average absolute error of BB is only 0.72 kcal/ mol and tional兴 is induced in the xc potential by an applied uniform
the maximal error is 2.4 kcal/ mol at R共H–H兲 = 3 bohrs. The electric field, which spans the entire system and counteracts
important feature of the functional 共4.14兲 and 共4.19兲 is its the field. The orbital-dependent EXX exchange potentials
correct behavior for dissociating H2. Due to this, the BB 共also in their KLI and CEDA approximations兲 possess this
dissociation energy error is only 0.20 kcal/ mol. We conclude field-counteracting behavior19,25 and Fig. 3 displays the KLI
that, in contrast to the EXX model, the BB functional is able and CEDA potentials together with their exchange hole
to treat the electron-electron correlation all the way from the 共Slater兲 and response components 共2.7兲 calculated for the
mostly dynamical correlation situation at Re to the purely prototype hydrogen chain H18 with alternating H–H bonds of
nondynamical correlation at R → ⬁. It is achieved by a mod- 2 and 3 bohrs in an electric field of 0.004 a.u. 共the dotted line
eling of the xc functional that does not separate exchange in Fig. 3 indicates the external potential generating this
and correlation but treats them on equal footing, in the trad- field兲. The Slater 共exchange-hole兲 potential 共indistinguish-
tional statistical physical manner by modeling directly the able for KLI and CEDA in Fig. 3兲 exhibits only periodical
conditional probability of finding an electron in the neigbor- variation, representing the individual H2 units. The
hood of a reference electron. exchange-hole potential does not exhibit any field counter-
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-12 E. J. Baerends and O. V. Gritsenko J. Chem. Phys. 123, 062202 共2005兲

FIG. 3. The CEDA 共drawn line兲 and


the KLI 共broken line兲 total exchange
potential 共bottom兲 and the response
共top兲 and Slater 共middle兲 components
are plotted along the main axis for the
chain H18 in an electric field of 0.004
a.u. 共the external field-induced poten-
tial is displayed with a dotted line in
the panel of the response components兲.

CEDA
acting effect. In the sense that one could consider incorpora- region makes the potential vresp higher in this region. In
tion of the full exchange hole potential 共the hole integrates to other words, while the total density ␳ does not directly feel
−1 electron兲 equivalent to avoiding a self-interaction error, the polarizations of occupied orbitals in opposite directions
the absence of a counteracting field effect does not appear to in an electric field, the weighted density 兺H CEDA
i sii 兩␾i共r1兲兩2
arise from a self-interaction error. However, it has been ar- produces under this polarization a field-counteracting term
gued by Mori-Sanchez et al.19 that the wrong scaling of the 共5.1兲, which is responsible for the field-counteracting behav-
LDA exchange energy with effective occupation number ior of the total exchange potential. This basic “weighted-
when an electronic charge density is fractionally distributed density” mechanism is valid for both KLI and CEDA, which
over noninteracting or weakly interacting sites89,90 can be possess diagonal terms with 兩␾i兩2, and the additional CEDA
related to the wrong polarizability of the chain of weakly off-diagonal contributions increase further its field-
interacting H2 units. In contrast to LDA and GGA, the KLI counteracting term compared to KLI 共see Fig. 3兲.
and CEDA response potentials counteract the external field- Effective reduction of an external applied potential by
induced potential, i.e., they become more positive at the the induced field-counteracting terms in the KS xc potential
down-field end of H18. This determines the overall field- is required in order that DFT methods yield an adequate
counteracting effect of the KLI and CEDA exchange poten- description of the response properties of molecular chains.
tials. From the upper part of Fig. 3 one can clearly see that LDA and standard GGA potentials all lack the abovemen-
CEDA develops a larger field-counteracting response poten- tioned compensation structure and, in the absence of the
tial than KLI. As was established in Refs. 24, 25, 81, and 82 counteracting effect, they produce a dramatic overestimation
the field-counteracting behavior of the CEDA 共and KLI兲 re- of 共hyper兲polarizabilities of molecular chains.25,81,91 Table IV
sponse potential originates from its orbital step structure presents the static longitudinal polarizabilities ␣ of the hy-
共2.8兲 and from the characteristic pattern of the polarization of
the orbitals of the valence “band” of a molecular chain in an TABLE IV. Comparison of the LDA, GGA, KLI, CEDA, EXX, HF, MP2,
external electric field. The dominant term of the change MP4, and CCSDT static longitudinal polarizabilities ␣ 共in atomic units兲 of
␦vresp
CEDA
of the potential 共2.8兲 appears to be that containing linear hydrogen chains Hn.
changes ␦兩␾i兩2 of individual orbital densities
H
␣ H6 H12 H18
␦兩␾i共r1兲兩2
␦vresp
CEDA
共r1兲 ⬃ 兺 sCEDA . 共5.1兲 LDA 72.7 210.6 367.3
i
ii
␳共r1兲 BLYP 68.98 194.80 334.85
BP 69.18 196.68 339.10
But then, a fundamental feature of the orbital response to an
PW 68.96 195.84 337.47
external-field potential, which can easily be established with KLI 60.50 157.15 260.66
perturbation theory 共see Ref. 82兲, is that the densities of CEDA 59.28 149.35 244.24
lower occupied orbitals in Eq. 共5.1兲 are localized more in the EXXa 227.97
down-field region, while higher occupied orbitals are local- HF 56.38 137.64 222.31
ized more in the up-field region. These individual changes in MP2b 54.16 134.00 217.42
opposite directions hardly change the total density ␳, how- MP4b 51.59 126.90 205.39
ever, they produce a field-counteracting change ␦vresp CEDA
. The CCSDTb 50.54 123.63 199.65
explanation is that lower orbitals ␾i possess larger steps a
Reference 19.
sCEDA
ii in Eq. 共5.1兲, so that their localization in the down-field b
Reference 94.
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-13 Orbital-dependent exchange-correlation functionals J. Chem. Phys. 123, 062202 共2005兲

drogen chains H6 , H12, and H18 calculated within the finite- 97–100, the energy ␧i of an occupied orbital can be inter-
field approach with KLI, CEDA, and HF. They are compared preted approximately as 共minus兲 the vertical ionization po-
with those calculated with LDA and GGA Becke–Lee–Yang– tential 共VIP兲 Ii of the corresponding primary ionization, −␧i
Parr 共BLYP兲,7,10 Becke–Perdew 共BP兲,7,9 and Perdew–Wang ⬇ I共␸−1 i 兲. Remarkably, for accurate KS solutions 共see the dis-
共PW兲 共Refs. 92 and 93兲 methods. The ␣ values obtained in cussion in Sec. III on how these can be obtained兲 it is found
Ref. 94 with the correlated Møller–Plesset 共MP2 and MP4兲 for a series of prototype molecules that the quality of this
and with the coupled cluster 共CCSDT兲 methods serve as a approximation appears, in the case of valence orbitals, to be
reference. Although the effects are much more dramatic for an order of magnitude better than for the HF orbitals97,99 共0.1
the second hyperpolarizability ␥ 共more than a factor 10 over- eV versus 1.0 eV兲, and for the KS highest occupied molecu-
estimation by LDA, reduced to a factor 1.2 for EXX-CEDA兲, lar orbital 共HOMO兲 it becomes an exact equality.101–104
we take the longitudinal polarizability ␣ as an example since To illustrate this point, Table V displays valence orbital
comparison can be made to the recently19 obtained accurate energies of the accurate KS solution obtained in Ref. 99 for
EXX ␣ for H18. One can see from Table IV that the better the prototype molecules with the LB procedure37 from the cor-
quality of the method, the lower is the calculated ␣ value. In related CI density as well as the HF orbital energies, the
particular, LDA greatly overestimates ␣, the LDA error in- EXX orbital energies 共from Ref. 67兲, and the GGA-BP or-
creases with the length of the chain. While for H6 the LDA ␣ bital energies. They are compared with the experimental
value is 1.4 times higher than that of CCSDT, for H18 this VIPs.105–109 One can see from Table V that only inclusion of
ratio grows to 1.84. ␣ values of different GGAs are close to both exchange and correlation in the KS potential ␯xc can
each other and GGAs improve only slightly upon LDA. Con- provide a high quality of the DFT analog of Koopmans’
trary to this, the orbital-dependent KLI and CEDA dramati- theorem. Indeed, the orbital energies of the accurate KS xc
cally decrease ␣ values upon LDA/GGAs, with CEDA pro- potential excellently reproduce the experimental VIPs with a
ducing consistently lower ␣ than KLI. The CEDA typical absolute average difference 共AAD兲 of only 0.1 eV.
approximation to the EXX lowers the LDA overestimation of The quality of Koopmans’ theorem for EXX is much worse,
184% to 122%, the accurate EXX ␣ lowers this to 114%. it is similar to that of the HF Koopmans’ theorem. Just as HF,
Still, just as in the case of the ground-state total energy con- EXX tends to overestimate VIPs with the Koopmans’ esti-
sidered in the previous sections, one is not gaining anything mate. Only for the FCN molecule the EXX AAD of 0.42 eV
compared to Hartree–Fock. Indeed, the HF ␣ values are the is substantially smaller than the HF one of 1.85 eV. For CO,
lowest among the exchange-only methods 共see Table IV兲. N2, and H2O the EXX AAD values of 0.95, 1.30, and 1.02
Furthermore, ␣ values of the correlated MP2, MP4, and eV, respectively, are similar to the corresponding HF differ-
CCSDT methods are consistently lower than the HF ones. So ences, and for the HF molecule the EXX deviation of the
we can conclude that an orbital-dependent functional is re- energies of the two highest occupied MOs appears to be
quired to avoid the large errors of the LDA and GGA meth- higher than the Hartree–Fock one.
ods for 共hyper兲polarizabilities in this particular case, but While HF and EXX occupied orbital energies overesti-
DFT within the EXX model does not offer any advantage mate VIPs, GGA-BP consistently underestimates them 共see
over straight Hartree–Fock. Here, again, further improve- Table V兲. For a particular molecule, the upward shift of the
ment beyond the exchange-only approach is required. It BP ␧i with respect to the vertical −Ii is remarkably uniform.
could come from orbital-dependent functionals, which com- For each molecule we also present for the BP HOMO this
bine the effects of both exchange and correlation in their shift 共␧H BP
+ IH兲. For the other orbitals the energies shifted
orbital structure. Natural candidates for this role are func- downwards by 共␧H BP
+ IH兲 are given, as well as the corre-
tionals, which depend explicitly on both occupied and virtual sponding AADs. Note, that for CO, N2 , H2O, and FCN the
KS orbitals, such as the BB functional presented in the pre- upward shift of the GGA-BP HOMO orbital energies varies
ceding section, but the performance of such functionals for within the rather narrow margins of 4.83–5.30 eV, and only
共hyper兲polarizabilities has not yet been studied. for the HF molecule the BP shift is higher at 6.44 eV. One
can see from Table V that the 共␧H BP
+ IH兲-shifted BP orbital
VI. MOLECULAR RESPONSE PROPERTIES WITH energies reproduce the valence VIPs very well, with the
ORBITAL-DEPENDENT XC FUNCTIONALS:
AAD values being close to those of the accurate KS solution.
EXCITATION ENERGIES
Unlike the abovementioned similarity of the Koopmans’
The orbital energies, both of occupied and of virtual or- type interpretation for occupied orbitals, there is an impor-
bitals, are key quantities in the calculation of response prop- tant qualitative difference between virtual HF and KS orbit-
erties. This is the case, in particular, for electron excitations als. The energies of virtual HF orbitals represent approxi-
calculated with time- dependent DFT 共TDDFT兲. In TDDFT mately the energy of an added electron in the field of all
the zero-order estimate of the excitation energy ␻k is just the other N electrons, in particular, the HF lowest unoccupied
energy difference of the occupied ␾i and virtual ␾a ground- molecular orbital 共LUMO兲 energy represents, within the
state orbitals associated with the excitation95 Koopmans approximation, the molecular electron affinity. In
contrast to this, as observed earlier, it is a basic feature of the
␻0k = ␧a − ␧i . 共6.1兲
KS solution 共2.1兲 of an N-electron system that all occupied
Actually, the energies of occupied HF and KS orbitals have and virtual orbitals are obtained with the same local, state-
the same physical meaning. By virtue of Koopmans’ independent KS potential vs, so that they all “feel” the effec-
theorem96 and its DFT analog established recently in Refs. tive field of N − 1 electrons. As a result, the energies of the
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-14 E. J. Baerends and O. V. Gritsenko J. Chem. Phys. 123, 062202 共2005兲

TABLE V. Comparison of the KS, HF, EXX, and GGA-BP orbital energies −␧i 共eV兲 with experimental VIPs 共the corresponding references are indicated in
the table兲. AAD are the average absolute differences between the KS orbital energies and VIPs. For BP the deviation of ␧HBP from −IH, i.e., the sum 共␧HBP
+ IH兲 for the HOMO, is given in parentheses; for all lower orbitals, next to the orbital energies, the energies shifted up by this amount are given. AADs for
the shifted orbital energies are also included in the table.

GGA-BP EXXa HF KS Expt.

Molecule MO −␧i −␧i −␧i −␧i


COb 5␴ 9.18; 共4.83兲 15.14 15.12 14.01 14.01
1␲ 11.95; 16.78 17.56 17.42 16.80 16.91
4␴ 14.27; 19.10 20.80 21.94 19.37 19.72
AAD 5.08; 0.25 0.95 1.28 0.15

N 2c 3␴g 10.39; 共5.19兲 17.24 17.27 15.57 15.58


1␲u 11.72; 16.91 17.68 16.72 16.68 16.93
2␴u 13.60; 18.79 20.24 21.21 18.77 18.75
AAD 5.21; 0.02 1.30 1.45 0.09

HFd 1␲ 9.75; 共6.44兲 17.70 17.69 16.18 16.19


3␴ 13.61; 20.04 21.14 20.92 19.90 19.9
AAD 6.51; 0.07 1.37 1.26 0.00

H 2O e 1b1 7.32; 共5.3兲 13.85 13.85 12.62 12.62


3a1 9.46; 14.76 15.82 15.89 14.73 14.74
1b2 13.19; 18.49 19.31 19.35 18.33 18.55
AAD 5.34; 0.04 1.02 1.06 0.08

FCNf 2␲ 8.83; 共4.82兲 13.64 13.68 13.65 13.65


7␴ 9.68; 14.50 14.91 16.38 14.72 14.56
1␲ 14.02; 18.84 20.26 22.16 19.74 19.3
6␴ 17.43; 22.25 22.98 25.31 22.83 22.6
AAD 5.02; 0.22 0.42 1.85 0.21
a
Reference 67.
b
Reference 105.
c
Reference 106.
d
Reference 107.
e
Reference 108.
f
Reference 109.

KS virtual orbitals represent an “excited” electron interacting that this problem does not occur in the KS-EXX model. Even
with N − 1 electrons, rather than an extra electron interacting though the wrong energies for excitations to Rydberg orbitals
with N electrons. They have therefore much lower orbital are not caused by the lack of proper asymptotic decay of the
energies than the virtual HF orbitals, and are much more LDA/GGA potentials, this deficiency of the these potentials
contracted. In KS-EXX calculations this will remain true, does have some adverse effects. It causes the Rydberg orbit-
since there is, in contrast to HF, still a single local optimized als to have the wrong shape. In Ref. 78 a detailed account
effective potential for both the occupied levels and the unoc- has been given for the case of the Rydberg excitations in N2,
cupied ones. where the oscillator strengths of these excitations are se-
Due to this physical nature of the virtual orbital energies, verely underestimated 共by factors of more than 50兲 in LDA/
the KS virtual-occupied orbital energy differences are, in GGA calculations. This is simply caused by the too tight
principle, much better approximations to the real excitation nature of the Rydberg orbitals in these calculations, which is
energies than in the case of HF. In order to reproduce this caused by the too rapid, exponential decay to zero of the
feature of the exact KS orbitals in approximate calculations, potential. As a consequence, also the spectral analysis of, for
the shape of the potential has to approach the exact one suf- instance, the polarizability reveals a wrong distribution over
ficiently closely. In this respect the LDA/GGA potentials are the various excitations. We are not aware of such an analysis
very deficient. This is already apparent from the large upshift in the case of KS-EXX calculations, but since its KS poten-
of the occupied orbital energies noted before. Such an upshift tial possesses the correct Coulombic asymptotics vx共r兲
cannot occur for the Rydberg orbitals, which are typically at → −1 / 兩r兩, 兩r兩 → ⬁, we expect also the oscilator strengths to
high energy, already close to the energy zero. Excitations to be meaningful. The abovementioned crucial features of the
such orbitals will become much too low. Often this failure potential—proper behavior in the bulk molecular region in
for the Rydberg excitations in the LDA and GGA methods is order to get correct orbital energies and proper −1 / r decay to
ascribed to the lack of proper −1 / r decay of the potential in get correct orbital shapes—are incorporated in the refined
those cases. That is not the true origin: it is the wrong posi- approximate xc potentials37,110–112 which have been intro-
tion 共not negative enough兲 of the occupied levels with re- duced. The SAOP potential76,77 is an orbital-dependent func-
spect to the energy zero which causes this problem. We note tional for the KS potential, which exhibits reasonably accu-
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-15 Orbital-dependent exchange-correlation functionals J. Chem. Phys. 123, 062202 共2005兲

ALDA
f xc is used. The proper treatment in TDDFT of long-range
ALDA
charge-transfer excitations requires a correction to f xc . It
is possible to formulate an orbital-dependent f xc which incor-
porates the required divergence at large R.115

VII. SUMMARY
We have argued that both Hartree–Fock and the use of
the exact-exchange functional in Kohn–Sham DFT are very
deficient models for the description of chemical bonding.
FIG. 4. Schematic illustration of the positions of the exact Kohn–Sham
orbital energies in relation to the ionization energies and electron affinities They both use a two-particle probability distribution which is
of the neutral donor 共D兲 and acceptor 共A兲 systems and of the negative A−. wrong, as can be seen from the shape of the “hole” a refer-
The magnitude of the gap ⌬ between LUMO energy and electron affinity −A ence electron creates around itself: the hole is not centered
is indicated for D and A, as well as the equality of −AA and ␧−a . around the reference electron, but is delocalized over the
atoms forming the bond. From this wrong hole shape one can
rate occupied orbital energies97 due to shape corrections of understand deficiencies of this model, such as too diffuse
the potential in the bulk molecular region, and in addition orbitals and density, hence wrong kinetic energy and
has −1 / r asymptotic behavior. Both excitation energies and electron-nuclear attraction energy. The true xc hole is sub-
oscillator strengths, and hence the spectral structure of re- stantially more localized around the reference electron. That
sponse properties, are therefore realistic.78 is why rough localized model holes, like those of GGA,
It is interesting to note that for a particular type of exci- which approximate this total hole, are so successful. Maybe,
tations the abovementioned advantageous characteristic fea- we should not say that GGA is very accurate, but rather that
ture of the KS orbital spectrum turns, actually, to a serious HF 共and EXX兲 is so inaccurate that it proved not to be very
problem of TDDFT. These are excitations associated with difficult to do considerably better even with a rough hole
molecular long-range charge transfer 共CT兲 between an elec- modeling like the one used in the LDA and GGA models.
tron donor D and electron acceptor A.113–115 With a weak It is possible to devise orbital-dependent functionals that
interfragment interaction at the large D-A separation R the do describe the shape of the exchange-correlation correctly,
corresponding excitation energy ␻CT becomes the difference both for the dynamical correlation at equilibrium bond length
between VIP ID of the donor and the electron affinity AA of and for the nondynamical correlation at elongated bond
the acceptor corrected with the electron-hole interaction lengths. We have discussed the functional introduced in Refs.
−1 / R 35 and 58, and discussed its application to dissociating H2.32
Any functional, including the EXX functional, that rem-
␻CT ⬇ ID − AA − 1/R, 共6.2兲 edies the major deficiencies of the self-consistent orbitals
and at large separations R the difference 共ID − AA兲 becomes and orbital energies of LDA and GGA, which are caused by
the dominant term of Eq. 共6.2兲. Figure 4 presents a schematic the too weak attractive nature in the bulk molecular region
picture of the relevant KS orbital energies of the donor D, and the wrong asymptotic behavior of the corresponding po-
acceptor A, and anionic acceptor A− fragments. In this case, tential, will improve response properties as calculated with
as follows from the abovementioned rigorous KS theory, a time-dependent DFT. We have discussed such improvement
fair estimate of ␻CT would be produced with the energy dif- for the polarizability, which involves the orbital dependency
ference 共␧−a − ␧d兲 of the orbital energies of the singly occupied of the EXX model rather directly. It is also possible to use
MO of A− and the HOMO of D. But then, the energy differ- orbital-dependent functionals for the KS potential directly,
ence 共␧a − ␧d兲 of the LUMO of A and HOMO of D, which is such as the SAOP potential,76,77 which are easy to use and
the zero-order approximation 共6.1兲 of the excitation in TD- remedy the deficiencies of the LDA/GGA potentials. Excel-
DFT, yields a poor estimate of ␻CT. It has been pointed out lent response properties are obtained. We have finally indi-
by Dreuw et al.114 that in the asymptotic limit of negligible cated that the long-standing problem of wrong charge-
overlap between the donor orbital 共HOMO of D兲 and accep- transfer transition energies in TDDFT can also be remedied
tor orbital 共LUMO of A兲 the modifications due to the cou- by invoking orbital dependence, this time in the exchange-
pling matrix disappears when the standard adiabatic LDA correlation kernel.
ALDA
kernel f xc is used, and the TDDFT result reduces to the 1
R. G. Parr and W. Yang, Density Functional Theory of Atoms and Mol-
zero-order value 共␧a − ␧d兲. This value is much too low com- ecules 共Oxford University Press, New York, 1989兲.
pared to 共␧−a − ␧d兲 because of the essentially positive and large 2
R. M. Dreizler and E. K. U. Gross, Density Functional Theory: An Ap-
difference ⌬A = ␧−a − ␧a between the energies of the extra and 3
proach to the Many-Body Problem 共Springer, Berlin, 1990兲.
excited electrons on the orbital ␾a 共see Fig. 4兲. For small W. Kohn, A. D. Becke, and R. G. Parr, J. Phys. Chem. 100, 12974
共1996兲.
prototype molecules the evaluation of the difference ⌬A from 4
E. J. Baerends and O. V. Gritsenko, J. Phys. Chem. A 101, 5383 共1997兲.
the accurate KS solution99 obtained from the correlated ab 5
W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 共1965兲.
initio CI density with the LB procedure37 produces ⌬A values
6
Theory of the Inhomogeneous Electron Gas, edited by S. Lundqvist and
N. H. March 共Plenum, New York, 1983兲.
in the range 7–10 eV, and this is, basically, the error of the 7
A. Becke, Phys. Rev. A 38, 3098 共1988兲.
TDDFT with the zero-order 共␧a − ␧d兲. As mentioned earlier, 8
J. P. Perdew, K. Burke, and Y. Wang, Phys. Rev. B 54, 16533 共1996兲.
9
this zero-order result is obtained when the standard xc kernel J. P. Perdew, Phys. Rev. B 33, 8822共E兲 共1986兲.
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-16 E. J. Baerends and O. V. Gritsenko J. Chem. Phys. 123, 062202 共2005兲

10
C. Lee, W. Yang, and R. G. Parr, Phys. Rev. B 37, 785 共1988兲. 62
A. Holas, Phys. Rev. A 59, 3454 共1999兲.
11 63
F. A. Hamprecht, A. J. Cohen, D. J. Tozer, and N. C. Handy, J. Chem. G. Csanyi and T. A. Arias, Phys. Rev. B 61, 7348 共2000兲.
Phys. 109, 6264 共1998兲. 64
K. Yasuda, Phys. Rev. A 63, 032517 共2001兲.
12
J. D. Talman and W. F. Shadwick, Phys. Rev. A 14, 36 共1976兲. 65
W. L. Luken, Int. J. Quantum Chem. 22, 889 共1982兲.
13
A. Görling, Phys. Rev. A 46, 3753 共1992兲. 66
W. L. Luken and J. C. Culberson, Theor. Chim. Acta 66, 279 共1984兲.
14
A. Görling and M. Levy, Phys. Rev. A 50, 196 共1994兲. 67
S. Hamel, P. Duffy, M. E. Casida, and D. R. Salahub, J. Electron Spec-
15
A. Görling and M. Levy, Int. J. Quantum Chem., Symp. 29, 93 共1995兲. trosc. Relat. Phenom. 123, 345 共2002兲.
16
S. Ivanov, S. Hirata, and R. J. Bartlett, Phys. Rev. Lett. 83, 5455 共1999兲. 68
S. Hamel, M. E. Casida, and D. R. Salahub, J. Chem. Phys. 116, 8276
17
S. Hirata, S. Ivanov, I. Grabowski, R. J. Bartlett, K. Burke, and J. D. 共2002兲.
Talman, J. Chem. Phys. 115, 1635 共2001兲. 69
O. V. Gritsenko, R. van Leeuwen, and E. J. Baerends, J. Chem. Phys.
18
W. Yang and Q. Wu, Phys. Rev. Lett. 89, 143002 共2002兲. 101, 8955 共1994兲.
19 70
P. Mori-Sanchez, Q. Wu, and W. Yang, J. Chem. Phys. 119, 11001 E. Engel, A. Faco Bonetti, S. Keller, I. Andrejkovics, and R. M. Dreizler,
共2003兲. Phys. Rev. A 58, 964 共1998兲.
20
A. Görling, Phys. Rev. Lett. 83, 5459 共1999兲. 71
A. Facco Bonetti, E. Engel, R. N. Schmid, and R. M. Dreizler, Phys. Rev.
21
R. T. Sharp and G. K. Horton, Phys. Rev. 90, 317 共1953兲. Lett. 86, 2241 共2001兲.
22
J. B. Krieger, Y. Li, and G. J. Iafrate, Phys. Rev. A 45, 101 共1992兲. 72
Q. Wu, P. Ayers, and W. Yang, J. Chem. Phys. 119, 2978 共2003兲.
23 73
R. van Leeuwen, O. V. Gritsenko, and E. J. Baerends, Z. Phys. D: At., S. Hirata, S. Ivanov, I. Grabowski, and R. J. Bartlett, J. Chem. Phys. 116,
Mol. Clusters 33, 229 共1995兲. 6468 共2002兲.
24
O. V. Gritsenko and E. J. Baerends, Phys. Rev. A 64, 042506 共2001兲. 74
Y.-H. Kim and A. Görling, Phys. Rev. Lett. 89, 096402 共2002兲.
25 75
M. Grüning, O. V. Gritsenko, and E. J. Baerends, J. Chem. Phys. 116, T. Hupp, B. Engels, and A. Görling, J. Chem. Phys. 119, 11591 共2003兲.
6435 共2002兲. 76
O. V. Gritsenko, P. R. T. Schipper, and E. J. Baerends, Chem. Phys. Lett.
26
F. Della Sala and A. Görling, J. Chem. Phys. 115, 5718 共2001兲. 302, 199 共1999兲.
27
F. Della Sala and A. Görling, J. Chem. Phys. 118, 10439 共2003兲. 77
P. R. T. Schipper, O. V. Gritsenko, S. J. A. van Gisbergen, and E. J.
28
J. B. Krieger, Y. Li, and G. J. Iafrate, Phys. Rev. A 46, 5453 共1992兲. Baerends, J. Chem. Phys. 112, 1344 共2000兲.
29
S. Kümmel and J. P. Perdew, Phys. Rev. Lett. 90, 043004 共2003兲. 78
M. Grüning, O. V. Gritsenko, S. J. A. van Gisbergen, and E. J. Baerends,
30
S. Grabowski, S. Hirata, S. Ivanov, and R. J. Bartlett, J. Chem. Phys. J. Chem. Phys. 116, 9591 共2002兲.
116, 4415 共2002兲. 79
A. Rosa, G. Ricciardi, E. J. Baerends, and S. J. A. van Gisbergen, J. Phys.
31
R. J. Bartlett, I. Grabowski, S. Hirata, and S. Ivanov, J. Chem. Phys. 121, Chem. A 105, 3311 共2001兲.
1 共2004兲. 80
A. Rosa, G. Ricciardi, and E. J. Baerends, J. Phys. Chem. 105, 5242
32
M. Grüning, O. V. Gritsenko, and E. J. Baerends, J. Chem. Phys. 118, 共2001兲.
7183 共2003兲. 81
S. J. A. van Gisbergen, P. R. T. Schipper, O. V. Gritsenko, E. J. Baerends,
33
M. A. Buijse, E. J. Baerends, and J. G. Snijders, Phys. Rev. A 40, 4190 J. G. Snijders, B. Champagne, and B. Kirtman, Phys. Rev. Lett. 83, 694
共1989兲. 共1999兲.
34 82
M. A. Buijse and E. J. Baerends, in Electronic Density Functional Theory O. V. Gritsenko, S. J. A. van Gisbergen, P. R. T. Schipper, and E. J.
of Molecules, Clusters and Solids, edited by D. E. Ellis 共Kluwer Aca- Baerends, Phys. Rev. A 62, 012507 共2000兲.
83
demic, Dordrecht, 1995兲, p. 1. R. W. Godby and L. J. Sham, Phys. Rev. B 49, 1849 共1994兲.
35
E. J. Baerends, Phys. Rev. Lett. 87, 133004 共2001兲. 84
X. Gonze, P. Ghosez, and R. W. Godby, Phys. Rev. Lett. 74, 4035
36
V. Polo, J. Gräfenstein, E. Kraka, and D. Cremer, Theor. Chem. Acc. 共1995兲.
109, 22 共2003兲. 85
R. Resta, Phys. Rev. Lett. 77, 2265 共1996兲.
37
R. van Leeuwen and E. J. Baerends, Phys. Rev. A 49, 2421 共1994兲. 86
R. M. Martin and G. Ortiz, Phys. Rev. B 56, 1124 共1997兲.
38 87
O. V. Gritsenko, R. van Leeuwen, and E. J. Baerends, Phys. Rev. A 52, D. Vanderbilt, Phys. Rev. Lett. 79, 3966 共1997兲.
1870 共1995兲. 88
G. Ortiz, I. Souza, and M. Martin, Phys. Rev. Lett. 80, 353 共1998兲.
39 89
O. V. Gritsenko, R. van Leeuwen, and E. J. Baerends, J. Chem. Phys. M. Sodupe, J. Bertran, L. Rodriguez-Santiago, and E. J. Baerends, J.
104, 8535 共1996兲. Phys. Chem. A 103, 166 共1998兲.
40
O. V. Gritsenko and E. J. Baerends, Phys. Rev. A 54, 1957 共1996兲. 90
Y. Zhang and W. Yang, J. Chem. Phys. 109, 2604 共1998兲.
41 91
O. V. Gritsenko, P. R. T. Schipper, and E. J. Baerends, J. Chem. Phys. B. Champagne, E. A. Perpete, S. J. A. van Gisbergen, E. J. Baerends, J.
107, 5007 共1997兲. G. Snijders, C. Soubra-Ghaoui, K. Robins, and B. Kirtman, J. Chem.
42
P. R. T. Schipper, O. V. Gritsenko, and E. J. Baerends, Phys. Rev. A 57, Phys. 109, 10489 共1998兲.
1729 共1998兲. 92
J. P. Perdew, Physica B 172, 1 共1991兲.
43 93
P. R. T. Schipper, O. V. Gritsenko, and E. J. Baerends, Theor. Chem. Acc. J. P. Perdew, J. A. Chevary, S. H. Vosko, K. A. Jackson, M. R. Pederson,
99, 329 共1998兲. D. J. Singh, and C. Fiolhais, Phys. Rev. B 46, 6671 共1992兲.
44
D. E. Woon and T. H. Dunning, J. Chem. Phys. 103, 4572 共1995兲. 94
B. Champagne, D. H. Mosley, M. Vracko, and J. M. Andre, Phys. Rev. A
45
V. R. Saunders and J. H. van Lenthe, Mol. Phys. 48, 923 共1983兲. 52, 178 共1995兲.
46 95
M. A. Buijse, Ph.D. thesis, Vrije Universiteit ter Amsterdam, 1991. M. Casida, in Recent Advances in Density Functional Methods, edited by
47
E. R. Davidson, S. A. Hagstrom, S. J. Chakravorty, V. M. Umar, and C. D. P. Chong 共World Scientific, Singapore, 1995兲, Vol. 1.
F. Fischer, Phys. Rev. A 44, 7071 共1991兲. 96
T. Koopmans, Physica 共Amsterdam兲 1, 104 共1933兲.
48 97
S. J. Chakravorty, S. R. Gwaltney, E. R. Davidson, F. A. Parpia, and C. D. P. Chong, O. V. Gritsenko, and E. J. Baerends, J. Chem. Phys. 116,
F. Fischer, Phys. Rev. A 47, 3649 共1993兲. 1760 共2002兲.
49
A. Savin, H. Stoll, and H. Preuss, Theor. Chim. Acta 70, 407 共1986兲. 98
O. V. Gritsenko and E. J. Baerends, J. Chem. Phys. 117, 9154 共2002兲.
50
J. C. Slater, Quantum Theory of Molecules and Solids 共McGraw-Hill, 99
O. V. Gritsenko, B. Braida, and E. J. Baerends, J. Chem. Phys. 119, 1937
New York, 1974兲, Vol. 4. 共2003兲.
51
M. Cook and M. Karplus, J. Phys. Chem. 91, 31 共1987兲. 100
O. V. Gritsenko and E. J. Baerends, J. Chem. Phys. 120, 8364 共2004兲.
52
A. D. Becke, J. Chem. Phys. 119, 2972 共2003兲. 101
J. P. Perdew, R. G. Parr, M. Levy, and J. L. Balduz, Phys. Rev. Lett. 49,
53
F. Furche, Phys. Rev. B 64, 195120 共2001兲. 1691 共1982兲.
54
M. Fuchs and X. Gonze, Phys. Rev. B 65, 235109 共2002兲. 102
M. Levy, J. P. Perdew, and V. Sahni, Phys. Rev. A 30, 2745 共1984兲.
55
S. Kurth and J. P. Perdew, Phys. Rev. B 59, 10461 共1999兲. 103
C. O. Almbladh and A. C. Pedroza, Phys. Rev. A 29, 2322 共1984兲.
56 104
F. Aryasetiawan, T. Miyake, and K. Terakura, Phys. Rev. Lett. 88, C. O. Almbladh and U. von Barth, Phys. Rev. B 31, 3231 共1985兲.
166401 共2002兲. 105
D. W. Turner, C. Baker, A. D. Baker, and C. R. Brundle, Molecular
57
Y.-M. Niquet, M. Fuchs, and X. Gonze, Phys. Rev. Lett. 90, 219301 Photoelectron Spectroscopy 共Wiley-Interscience, New York, 1971兲.
共2003兲. 106
P. Baltzer, M. Larsson, L. Karlsson, B. Wannberg, and C. Göthe, Phys.
58
M. A. Buijse and E. J. Baerends, Mol. Phys. 100, 401 共2002兲. Rev. A 46, 5545 共1992兲.
59
A. M. K. Müller, Phys. Lett. 105A, 446 共1984兲. 107
G. Bieri, A. Schmelzer, L. Asbrink, and M. Jonsson, Chem. Phys. 49,
60
S. Goedecker and C. Umrigar, Phys. Rev. Lett. 81, 866 共1998兲. 213 共1980兲.
61
J. Cioslowski and K. Pernal, J. Chem. Phys. 111, 3396 共1999兲. 108
C. R. Brundle, M. B. Robin, N. A. Kuebler, and H. J. Basch, J. Am.
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34
062202-17 Orbital-dependent exchange-correlation functionals J. Chem. Phys. 123, 062202 共2005兲

Chem. Soc. 94, 1451 共1972兲. J. Chem. Phys. 114, 652 共2001兲.
109 113
L. Asbrink, A. Svensson, W. von Niessen, and G. Bieri, J. Electron Spec- M. E. Casida, F. Gutierrez, J. Guan, F.-X. Gadea, D. R. Salahub, and J.-P.
trosc. Relat. Phenom. 24, 293 共1981兲. Daudey, J. Chem. Phys. 113, 7062 共2000兲.
110
M. E. Casida, C. Jamorski, K. C. Casida, and D. R. Salahub, J. Chem. 114
A. Dreuw, J. L. Weisman, and M. Head-Gordon, J. Chem. Phys. 119,
Phys. 108, 4439 共1998兲. 2943 共2003兲.
111
D. J. Tozer and N. C. Handy, J. Chem. Phys. 109, 10180 共1998兲. 115
112 O. V. Gritsenko and E. J. Baerends, J. Chem. Phys. 121, 655 共2004兲.
M. Grüning, O. V. Gritsenko, S. J. A. van Gisbergen, and E. J. Baerends,

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
146.189.194.69 On: Sat, 20 Dec 2014 15:07:34

You might also like