Density Functional Theory and the Band Gap

Problem
JOHN P. PERDEW
Depanmenr of Physics and Quantum Theory Group, W a n e Uniwrsiry.
New Orleans, Louisiana 70118, USA

Abstract
How can the fundamental band gap of an insulator be predicted? As a difference of ground-state energies.
the fundamental gap seems to fall within the reach of density functional theory, yet the predicted gaps from
band structure calculations within the local density approximation (LDA) are about 40% too small. It is argued
here that even the exact Kohn-Sham potential v,,A[), which generates the exact density in a self-consistent-
field calculation, generates a band structure which underestimates the gap. Within the context of the band gap
problem. several recent developments in the density-functional theory of manyelectron systems an:
reviewed: ( I ) The Langreth-Mehl approximation to the Kohn-Sham exchange-correlation energy and
potential. based upon the Langreth-Perdew wavevector analysis of the density gradient expansion. This
functional leads to more accurate ground-state energies and densities than those of the LDA with little change
in the calculated band structures of solids. (2) The derivative discontinuity of the exchangecorrelation
energy. which is responsible for substantial underestimation of the fundamental gap by even the exact Kohn-
Sham potential. (3) The self-interaction correction, which yields accurate gaps in insulators only by virtue of
its orbital-dependcru potential. (4) The density response function of the uniform elcctron gas, which suggests
that the DA gives a good estimate of the exact Kohn-Sham potential for a semiconductor with a weak
periodic potential. In short, several very different (but admittedly approximate) numerical calculations
suggest that most of the error in the LDA fundamental gap would persist in the gap of the exact Kohn-Sham
band structure. This error would persist in any attempt 10calculate the gap from LDA total energy differences
for clusters of increasing size.

1. Introduction and Summary

The fundamental gap of an electronic system is the difference between the ionization
energy I and electron affinity A . In any system, finite or infinite, the quantity of I-A
plays the role of “chemical hardness” [ 1). In a metallic crystal, 1-A is zero. In an
insulating crystal, 1-A is the fundamental band gap or conductivity gap, the least
excitation energy for a‘separated electron-hole pair.
The band theory of solids aims to calculate the ground-state density n(c) and the
single-particle spectrum, especially the fundamental gap. The density can be calculated
as for a hypothetical system of noninteracting electrons moving in the effective
potential vef&):

International Journal of Quantum Chemistry Quantum Chemistry Symposium 19. 497-523 ( 1YR6)
0 1986 by John Wiley & Sons. Inc CCC 0360 R832/86/010497-27U)4 00
498 PERDEW

wheref, = O(k - ei)is a fermion occupation number, and

That effective potential vefdrJ which generates the true density is known as the Kohn-
Sham [2] potential. As discussed in Section 2, the Kohn-Sham potential is intimately
related to the ground-state energy. Thus one might hope that the Kohn-Sham band
structure would predict the true fundamental gap, which is a difference of ground-state
energies. This paper discusses the failure of that expectation, the reason for this failure,
and an alternative approach to the calculation of the fundamental gap. Instead of a full
review of the band gap problem, this is the story of how several separate avenues of
research have lined up to point in the same direction.
The organization of the paper is as follows:
Section 2 outlines the Hohenberg-Kohn-Sham density functional theory [2,3] and
the local density approximation [2], the de fucto basis for much of modern band theory.
Within the local density approximation to the Kohn-Sham potential, the fundamental
gaps of insulators and semiconductors are typically underestimated by 40%.
However, this observation leaves open the possibility that improved approximations
to the Kohn-Sham potential might correct the gap. This possibility is explored in
Section 3, which reviews the Langreth-Mehl [4] approximation based upon the
Langreth-Perdew [ 5 ] wavevector analysis of the density gradient expansion. The
Langreth-Mehl approximation corrects most of the error of the local density
approximation in the total energy and electron density, but corrects little of the error in
the fundamental gap.
This observation suggests that even the exact Kohn-Sham potential might not yield a
band-structure with the correct fundamental gap. The reason for this failure of
expectation, the derivative discontinuity of the exact exchange-correlation energy, is
discussed in Section 4. This surprising nonanalyticity was discovered by Perdew, Pan;
Levy and Balduz [ 6 ] ,and applied to the band gap problem in simultaneous papers by
Perdew and Levy [7] and by Sham and Schluter [8]. It is conjectured in Section 4 that
the exact Kohn-Sham band structure normally provides a lower bound on the
fundamental gap.
The spin-density generalization [9,10] of density functional theory is introduced in
Section 5, and applied to an instructive model system: an expanded lattice of hydrogen
atoms. The true energy cost to remove an electron from this system is the same,
whether the electron is removed from a localized Wannier orbital or from a delocalized
Bloch orbital. But the local spin density approximation is contaminated by a self-
interaction error which is much worse in the Bloch picture than it is in the Wannier
picture.
As an alternative to the Kohn-Sham method, the self-interaction corrected (SIC)
method [ 1 1,121 offers an orbital-dependent potential which generates accurate electron
densities and fundamental gaps in large-gap insulators. Section 6 reviews the self-
interaction correction to the local spin density approximation, within the formalism of
Perdew and Zunger [ 1 I]. Special attention is directed to the band-theory formulation of
 DENSITY FUNCTIONAL THEORY 4w

SIC by Heaton, Harrison, and Lin [12], in which each Bloch orbital sees its own
individual periodic potential. When this orbital-dependent is replaced by an orbital-
independent Kohn-Sham potential, chosen by the method of Norman and Koelling 1131
to optimize the SIC total energy, the good agreement between slc-calculated and
measured gaps is destroyed. This observation suggests that only an orbital-dependent
potential v‘,drJ, or a nonlocal one ve&r,’), can simultaneously generate the electron
density and the fundamental gap.
A metal may be distinguished from an insulator either by its vanishing fundamental
gap or by its perfect screening. In Section 7, the theory of static dielectric screening is
reviewed, and a physical argument is presented for the equivalence of these two criteria
for metallic behavior.
Recently, Antoniewicz and Kleinman [ 141 have suggested that the exact Kohn-Sham
potential might, after all, yield the observed fundamental gap for semiconductors.
Their suggestion is based upon the strong difference between the exact dielectric
response for the uniform electron gas and that of the local density approximation, both
within the exchange-only version of density functional theory. In Section 7, it is shown
that recent estimates of the correlation effect upon the dielectric response function
suppress this difference, making it unlikely that the exact Kohn-Sham potential
(including correlation) could yield the observed gaps in semiconductors.
The failure of local potentials vef&) in band theory is not unanticipated. More than
twenty years ago, Herman found a serious underestimation of the fundamental gap in
relativistic local-potential calculations for germanium [ 15,441. It has been known for
some time that band theory requires in principle an energy-dependent, nonlocal self-
energy [16,17] U(K,L‘;E). Recent progress [ 18-24] toward the evaluation of the self-
energy will not be reviewed here.
Fourteen years ago, Kane [25] pointed out that any potential ve&) which fits the
observed cyclotron masses (valence and conduction band shapes) in silicon yields a
fundamental gap of 0.6 eV, which is closer to the local density gap [26] of 0.5 eV than
to the true gap [25] of 1.1 eV. At the same time, Lipari and Kunz [27] reached similar
conclusions for the alkali halides. Eleven years ago, Mott [28] made a prescient
statement which could serve as the theme of this paper:
“But often it will not be sufficient, if we are to obtain even qualitatively the right
separation between occupied and empty bands, to take the same potential V(E) for the
valence and conduction bands. Particularly in a tight-binding situation, as for a d-band,
it would be a poor approximation to do so. An electron in the valence band of (say)
solid argon sees the field of Ar+ , while an electron in the conduction band sees that of
Ar. This should introduce a separation between the two bands of order I-A, where I is
the ionization potential and A the electron affinity. This is a term depending on
correlation, and if a one-electron formulation with the same function V(r) for electrons
inboth bands gives good agreement with experiment for the energy gap, this must be
something of an accident.”

2. Overview of Density Functional Theory

Hohenberg and Kohn [3] have demonstrated the existence of a functional E,[n]
which, when minimized over trial densities n(r_)integrating to N electrons, yields the
500 PERDEW

exact ground-state energy E and density for N electrons subject to an external potential
v(c). The Euler equation for this variational principle is

or
6E,/8n(c) = p . (4)

The Lagrange multiplier p is evidently the chemical potential:

Kohn and Sham [2] have shown how to implement this variational principle in
practical calculations: Divide E,[n] into pieces,

where TJn]is the ground-state kinetic energy for hypothetical noninteracting electrons
of density n ( ~ )V[n]
, is the classical repulsion

and Exc[n] is the exchangecorrelation energy. The true n(c) is found from the self-
consistent solution of the one-electron equations (1) and (2). The effective potential of
the Kohn-Sham equation (2) is

where

Thus the many-electron problem has in principle been reduced to self-consistent-field

form.
The Kohn-Sham orbital energies ei have at least a formal meaning. Janak [29]
(following arguments given by Slater [30]) proved that

where

E = Xfi(+il- +
‘/2v2(+;) ~ i 3 v(c)
r n(c) + ~ [ n +] E J ~ ]. (1 1)

Q. (10) holds for any choice of the occupation numbers cf;:}; the orbitals {Jli} are to be
optimized or relaxed for each choice. The proof of Eq. (10) requires only self-
 DENSITY FUNCTIONAL THEORY 50 I

consistency, and so is valid for the exact E,[n] or for any approximation to it. From Eiq.
(10).Jan& [29]also derived an “aufbau” principle: If the occupation numbersfi are
permitted to take any value between 0 and 1, then E“ is minimized whenfi = 1 for all
ei< p, and 0 for all ei> p, where p is the chemical potential. For fixed electron number
N, E“ at its minimum is just the ground-level energy E(N). Fractional values offi and N
arise in ensemble averages for open systems [6,29,31].
Let the orbital energies of the N-electron system be ordered as

Then, for the J-electron system (J = any positive integer), the first ionization potential
and electron affinity are respectively (from Eq. (10))
I
I(J) = E ( J - 1) - E(J) = -Jdje,{J- 1 +f), (13)
0

A(J) = I(J+ 1) = E(J) - E(J+ 1) = -idf~,+,(J+f) . (14)

According to Eqs. (13) and (14).the fundamental gap of the J-electron system is
determined by the highest partly-occupied Kohn-Sham orbital energy of the N-electron
ground-level, for J- l<NCI+ 1. This is the only Kohn-Sham orbital energy which
carries any exact physical meaning, since I? of Eqs. (10) and ( 1 1) has an exact physical
significance only when the “aufbau” principle is obeyed. In a crystal (J+w), the
removal or addition of one electron can change the ground-state density n ( ~ only )
infinitesimally. If this infinitesimal density change is associated with an infinitesimal
change in the Kohn-Sham potential v,rf([n]r_), and hence in the Kohn-Sham orbital
energies, then by Eqs. (1 3) and (14)the fundamental band gap will be predicted exactly
by the Kohn-Sham band-structure of the neutral (J-electron) system:

This interesting argument, in essence, was presented by Williams and von Barth
[32,33],and was foreshadowed in discussions by Slater and Wood [34]and by Worth
and Trickey [35].The argument is almost correct; what it overlooks is the possibility
that the Kohn-Sham potential, and specifically 6ExJ6n(~), might shift [6]by a constant
C when N increases through the integer J . This possibility will be considered in
Section 4.
The exchange-correlation energy Exc[n] can be interpreted as the electrostatic
interaction between the electron density at each point L and the density p x c ( ~ , ~of’ )the
exchange-correlationhole around L:
502 PERDEW

In ECq. (17),the density-density correlation function must be evaluated for all coupling
strengths A in the range 0 < A < 1 . For each A, there is a fictitious system of electrons
with interparticle interaction A/[: - ~ ‘ 1in the presence of external potential &([),
constructed to hold the ground-state density n ( i ) at its true value. A = 1 describes the
real electronic system, while A = 0 describes the Kohn-Sham noninteracting system.
(The exchange energy EJn] is defined by Eqs. (16) and (17) with (. . .)* replaced by
(. . .)o, and the correlation energy Ec[n]is defined as Ex, - Ex; these definitions differ
slightly from the traditional ones of quantum chemistry.) For fixed integer electron
number [ 3 7 ] ,or for fluctuating electron number in an infinite system [38,39],the hole
obeys the sum rule

Around a given electron, the average density is depleted by one electron.

In practice, the exchange-correlation functional E J n ] must be approximated. A
widely-used choice is the local density approximation 121 (LDA):

where cx,(n) is the exchange-correlation energy per panicle of a uniform electron gas
with density n. The LDA becomes exact only in the limit of slowly-varying density. Its
formal validity condition is [ 4 ]

where 4 is the inhomogeneity wavevector

and kF is the local Fermi wavevector

The LDA exchange-correlation hole,

where p:, ( n ; - [ ’ I ) is the hole in a uniform electron gas of density n , obeys 1371 the
sum rule of ECq. (18).
Kohn-Sham theory in general, and the LDA in particular, are the defacto foundation
for much of modern band theory. The LDA exchange-correlation potential,

where px,(n) = d(ne,,.)/an, is easy to construct, and the LDA density is reasonably
 DENSITY FUNCTIONAL THEORY 503

TABLE
I. Fundamental gaps of insulators, from the band structure within the local density approximation
(LDA) and from experiment. The "LDA" value for antiferromagnetic NiO was actually obtained from the local
spin density approximation. The second "LDA" values for Ge and GaAs include relativistic effects. The
"experimental" value for BN may be an underestimate due to point defects in the crystal (eV).
~~

Crystal LDA Exper. Crystal IDA Exper.

Ne" 11.4 21.4 C" 4. I 5.5

Ar" 8.3 14.2 Si" 0.56 1.17
KP 7.0 11.6 Ged 0.65 0.76
Lip 9.9 14.2 GC' 0.09 0.76
NaCI" 5.3 9.0 GaAs' 0.71 1.5
CSCI' 5 .O 8.3 GaAs' 0.25 I .5
MgO" 4.5 7.7 CuAlSf 2.1 3.5
CaS' 2.6 5.4 CuAlSe; I .7 2.1
Bas' 1.8 3.9 CuGaS; I .3 2.4
ZnSb 2.3 3.8 CuGaSe: 0.5 I .7
1.8 2.8 CuInS: 0.0 1.5
GaPb I .8 2.4 CuInSer 0.0 1 .0
AIP I .6 2.5 CdSh 2.0 2.6
NiO' 0.3 4.3 BN' 8.7 6.4
CUCI' 2.0 3.4
a Ref. 40.
Ref. 41.
Ref. 42.
Ref. 43.
Ref. 44.
'Ref. 45.
Ref. 46.
Ref. 47.
' Ref. 48.
* Ref. 49.

accurate. However, in LDA the fundamental band gaps of insulators and semiconductors
are typically underestimated by 40% (Table I) [40-49].
The LDA exchange potential, which dominates over correlation in Eq. (24), is
kX(n)= - ( 3 d 2 7 ~(37~*n)"~,
) where a equals [2] 213. In Slater's Xa method [30], the
correlation term is dropped and the coefficient a is treated as an adjustable parameter.
Khan and Callaway [50]found that self-consistentXa band-structure calculations for
neon and argon could reproduce the observed fundamental gaps with an empirical
coefficient a = 1.25. Tnckey, Ray and Worth [51] have observed that this exchange
scaling seriously worsens the calculated band widths and cohesive properties. Almost
surely, it worsens the calculated electron density as well.
Beside the Langreth-Mehl approximation of Section 3 and the self-interaction
correction of Section 6, there have been several attempts to improve upon the LDA. One
of them is the weighted density approximation 152-541:
504 PERDEW

TABLE
11. Fundamental gaps of insulators, from the band structure within the local density (LDA) and
weighted density (WDA) approximations, and from experiment (eV).

Crystal LDA WDA Exper.

Sin 0.56 0.71 1.17

Ge’ 0.52 0.60 0.74
GaAsb 0.50 0.56 I .5
a Ref. 55 (uniform electron-gas input).
Ref. 45.

where the weighted density i i ( r ) is chosen to satisfy the sum rule of Eq. (18).
Semiconductor fundamental gaps from WDA band structures are only slightly more
realistic than those from LDA (Table 11) [45,55].
A systematic correction to the LDA, in the slowly-varying limit, is provided by the
gradient expansion approximation 131 (GEA):

The gradient coefficient Cxc(n) is known 15,561. Its exchange piece is C,=
- 1.667 x while its correlation piece is C,- +4.2 X at high and metallic
densities. For redistic density variations, the GEA apparently gives no improvement in
total energy over the LDA [57]. This misfortune is explained and corrected in Section 3.
In fact, it can be shown [4] that the gradient term in Eq. (26) gives the right
correction to the LDA only when

where k , is the local Fermi-Thomas screening wavevector

For physical electron densities, Eq. (20) is often satisfied reasonably well, while
Eq. (27) is not.

3. Langreth-Mehl Approximation to the Kohn-Sham Potential

In order to understand and correct the errors of the LDA and GEA, Langreth and
Perdew [36] proposed the wavevector analysis

where S(k) is the structure factor of the inhomogeneous electronic system, averaged
over the coupling strength A and over the direction of the wavevector k:
 DENSITY FUNCTIONAL THEORY 505

S(k) varies from 0 at k=O to 1 as k-a. Eqs. (29) and (30) follow from Eqs. (16) and
(17) upon Fourier analysis of the Coulomb interaction l/lr-r’l. In Eq. (29), the
exchangecorrelation energy is decomposed into contributions from dynamic density
fluctuations of various sizes k - I .
In the local density approximation,

NSLDA(k) = J a r n([) S,(n([);k) (31)

where $,(TI;&) is the structure factor of a uniform electron gas of density n. The LDA is
exact [36] for &%-2kF, in agreement with the intuition that a localized density fluctuation
sees only the local density. The LDA breaks down for k42kF, but the breakdown is not
too serious because of the small weight given to this region of phase space in three
dimensions [36]. In fact QDA(k) is proportional to kz at small k , while the exact S(k) is
also proportional to k2 (with a different coefficient), except for surface energy
contributions proportional to llcl in semi-infinite systems [4,5,36].
The gradient expansion approximation of Eq. (26) provides a useful correction to the
LDA at intermediate wavevectors [4,5,58]:

e’kp 2
NSGEA(k) = NSLDA(k) + 5d3r 3zxc(kF,O;k)
4T
$,F

where z,, = z, + z,. The exchange term is

1 11 1
Z,(kF,O;k) = --{-4Y~(l-y) -k - 9- @ - I ) -k 9
-8’@-1)} (33)
2kF
where y = W2kF. The wavevector analysis of the correlation term was performed by
Rasolt and Geldart [ 5 8 ] and by Langreth and Perdew [5]. The result within the random
phase approximation (RPA) [5] has been parametrized by Langreth and Mehl [4]:

4v5
Z, ( ~ = -exp( - 2 v 5 Wk,)
k,O;k) (34)
km

Unlike the LDA, the gradient term in Eq. (32) is seriously wrong in the k-0 limit,
tending to a large positive constant and not to zero.
The problem is that, for the GEA to be valid, q of Eq. (21) must be small compared to
all relevant wavevectors, including k. As a simple remedy, Langreth and Mehl [4]
replaced z,(kF,O;k) of Q. (34) by the cut-off form

wheref=O. 15. With this correction, they found, after integration over the wavevector
k.
506 PERDEW

where

F = blVnl/n7’6 (37)

and a = 0.0021435, b = 1.745f. Since the gradient term was derived within RPA, the
LDA term is also to be evaluated within RPA. The Langreth-Mehl approximation to the
exchange-correlation part of the Kohn-Sham potential is then

where 5 = Vn(r_).
Truly impressive results have been achieved with the Langreth-Mehl (LM)approxi-
mation of Eqs. (36) and (38). It gives about the right negative correction to the LDA
total energies of closed-shell atoms, and the right positive correction to the LDA surface
energies of metallic crystals. The shell removal energies of atoms are also improved, as
are the electron densities of the lighter atoms, where accurate densities are known from
configuration-interactionstudies. Unlike the LDA potential, the LM potential of Eq. (38)
displays the cusp-like structures between atomic shells which are found in the exact
exchange potential [59].
Figure 1 shows the deviations from LDA of the LM [60] and exact [61] electron
densities in the neon atom, as well as the deviation [60]from LDA of the self-consistent
LM effective potential veff([n];r_).
In band structure calculations for crystals, the LM approximation yields only small
corrections to the LDA. In ttie transition metals Cu and V, the valence bands experience
an s-d shift of only a tenth of an electron volt [62]. In the large-gap insulators Ne and
Na C1 (Table 111) [63], the fundamental gap eJ+ I(J)- EAJ) in the Kohn-Sham band
structure is increased, but only slightly-the discrepancy with the true fundamental
gap I-A is barely reduced. In the semiconductor Si, the band gap discrepancy also
persists [a].
These results suggest that even the exact Kohn-Sham band structure would fail to
predict the fundamental gap. However, this conclusion is not firm because, as Langreth
and Mehl [4] point out, the LM approximation breaks down in regions of very low
density where q22kF. In a semiconductor or insulator, the unoccupied orbitals of
interest may have nonnegligible components in these regions. In Sections 4, 6 and 7,
stronger evidence will be presented to show that the exact Kohn-Sham band structure
fails to predict the fundamental gap.
 DENSITY FUNCTIONAL THEORY 501

:
- II neon atom
/ effective potential
I
II
-2 - I
I
I
II
I
,,.,I I 1 1 , 1 1 1 1 1 I [Link] 1 I 1 1 1 1

0.00 -

0.04-

0.00-

-0.04 - neon atom

Figure I . (a) Deviation of the self-consistent Kohn-Sham potential vefc(r)in the neon atom
from its form in the local density approximation (LDA). LM: Lengreth-Mehl approximation of
Eq. (38). w - S I C :optimized effective potential for the self-interaction correction, from Eq.
(69). The electron gas correlation-energy input is the Ceperley-Alder parametrization of Ref.
I 1 for the LDA and SIC calculations, and the random phase approximation for the LM
calculation. To set the scale, (r) = 0.16, 0.91 and O.% for the Is, h,and 2p orbitals
rcsptctively. Note the inter-shell bumps in the potentials. (Figure from M. R. Norman.
Ref. 60.) (b) Deviation of the self-consistent radial distribution 4nr2n(r) in the neon atom
from its form in the LDA. The "exact" density is taken from the careful configuration-
interaction study (150 configurations) of Ref. 61. The SIC and LM approximations to the
density behave as might be expected: SICis better at very large and very small r , while LM
gives a better account of the depletion of the density in the inter-shell region.
508 PERDEW

TABLE111. Fundamental gaps of insulators, from the band structure within the local density (LDA) and
Langreth-Mehl (LM)approximations, and from experiment (eV).

crystal LDA LM Exper.

Ne' 11.5 12.7 21.4

NaCI' 5.6 6.0 9.0
Ref. 63.

4. Derivative Discontinuities of the Energy

The derivation of Eq. (15) from Eqs. (13) and (14) must be re-examined. The
infinitesimal density change due to addition or removal of one electron in an infinite
crystal can only be associated with a change in the Kohn-Sham potential v,f'([n];c)
which is an infinitesimal plus a possible finite constant C which may appear [6-81 as
the electron number N increases through the integer J . Thus, if 6 is a positive
infinitesimal,

The true fundamental gap equals the gap in the exact Kohn-Sham band-structure plus a
possible finite constant C.
For example, consider a monatomic crystal in the separated-atom (infinite lattice
constant) limit. The Kohn-Sham orbitals for the crystal are tight-binding linear
combinations of the Kohn-Sham atomic orbitals. Suppose first that the underlying
atom has an open subshell, so that cJ+ I(J)= EAJ). Nevertheless, the crystal is a Mott
insulator with a fundamental gap I(J)-A(J) equal to that of the underlying atom; all of
the fundamental gap comes from the positive constant C. Next suppose that the
underlying atom is closed-shell, so that the atomic electron affinity is zero. Even
though the neutral atom cannot bind an extra electron, it has an unoccupied Kohn-
Sham bound orbital with energy eZ+l ( Z - 6)<0 (because the atomic veff([n];~) tends to
- I/r as r+w). It follows from Eq. (40) that again D O . The open-shell example is
from Perdew, Pam, Levy and Balduz [6], while the closed-shell was presented by von
Barth [33] as a possible objection to the Williams-von Barth [32,33] expression (15).
Intimation of a correction to the exact Kohn-Sham gap in an insulator can also be found
in Ref. 65.
Eqs. (13) and (14), which are valid in finite systems as well as infinite ones, refer to
fractional electron number. The meaning of fractional N has been explained in Ref. 6.
The ground level for N electrons, with N between the integers J - 1 and J , is an
 DENSITY FUNCTIONAL THEORY 509

ensemble or statistical mixture of the (J- 1)- and J-electron ground states, with
probabilitiesp =J - N and 1 - p = N - J +
1 respectively. For J - 1S N G , the ground-
level energy and density are

E(N) = (J-N) E ( J - 1) + (N-J+ 1) E(J) (42)

As a function of N, the energy E(N) is a linkage of straight-line segments with possible

slope discontinuities at integer values of N. The chemical potential p. according to Eqs.
(3,(10) and (42) is

As N increases through the integer J , the chemical potential p. = 6EJ6n(r_)and the

highest partly-occupied Kohn-Sham orbital energy jump discontinuously from one
physical value (minus the ionization energy) to another (minus the electron affinity).
The fundamental gap is thus the discontinuity [6-81 in the functional derivative:

With the energy functional EJn] decomposed as in Eq. (6), it is clear that the
discontinuity of Eq. (46) can arise only from the kinetic energy Ts[n] and the exchange-
correlation energy EJn]. In the Kohn-Sham noninteracting system, with fixed
effective potential veff([)l,, only the kinetic energy contributes [7,8]:

Therefore, the difference between the true fundamental gap and that from the Kohn-
Sham orbital energies is [7,8]

c = 6E,c/W)l,+s - 6Exc/6n(rj(,-s f
(48)

Now the noninteracting kinetic energy TJn] is a clever mathematical construct rather
than a physical reality; a situation in which it displays a derivative discontinuity, while
the exchangecorrelation energy does not, can arise only by the purest accident.
Therefore C,the discontinuity of 6EXJ6n(r),is expected to be nonzero in all real finite
systems [6,7] and real infinite insulators and semiconductors [7,8]. In a Mott insulator
[28,66], where 6Ts/6n([)and 6Ex/6n([)are both continuous, all of the gap [6,39] arises
from the discontinuity of the correlation potential 6Ec/6n([).
At first it seems very peculiar that the Kohn-Sham exchange-correlation potential
) jump by a constant C. Shouldn’t 6EXJ6n([)tend to zero far outside an
6 E x J 6 n ( ~can
electronic system? Yes it should, but sometimes one must go far outside to reach this
limit. Consider for example a large but finite insulating crystal. For the neutral system,
with N =J- 6 electrons (6-*0+), the exchange-correlation potential 6ExJ6n(c)
5 10 PERDEW

effectively goes to zero at some microscopic distance outside the surface. But, when an
extra electron is added to the conduction band, 6ExJ6n([)does not go to zero until E is
so far outside the surface that the density n(J + I ;[) is dominated by the density of the
one extra electron, and not by that of the J original electrons. For a full discussion of
this and related questions, see Ref. 39.
A discontinuity in 6F/6n(rJ (where F is any functional) must not be confused with a
discontinuity in aF/aN. The latter can arise without the former, as a consequence of the
discontinuity in dn(r)/dN. Note however that the discontinuity in a n ( ~ ) / a Nis caused by
that of 6EJ6n(~).
Derivative discontinuities of the energy manifest themselves not only in open
electronic systems but also in closed ones with fixed integer electron number. In the
latter situation, certain “unconventional” number-conserving density variations [67]
6 n ( ~ about
) the ground-state density lead to energy variations 6E, of order (tin[,and
not of order as might have been expected.
A formal expression for the gap correction C of Eq. (41) may be developed. The
quantities - 1 0 and -A(J) are eigenvalues of a one-electron Schrdinger equation
like Eqs. (2) and (8). but with a nonlocal, energydependent self-energy [16,17]
&,(I-,E’;z) in place of 6ExJ6n(c). If the difference Zxc-6E,J6n is regarded as a
perturbation on the Kohn-Sham equation (2), the correction C to the gap in the Kohn-
Sham eigenvalues may be found by perturbation theory. To first order in this difference
WI,

The second expectation value vanishes, so

where JIJ+ is the lowest unoccupied Kohn-Sham orbital. Eq.(50) is not quite exact,
because there are terms of higher order in Zxc- 6ExJ6n, which have been formally
summed by Sham and Schliiter [8]. These higher-order terms could be very small. To
leading order in e2, they corresponded to the difference between the Hartree-Fock
orbitals and the exact Kohn-Sham exchange-only orbitals, a difference which is
almost negligible in atomic calculations [59]. (Unlike the derivation in Ref. 8, the
present derivation of Eiq. (50) applies even to finite systems.)
Return to Eq. (49). In a normal insulator, there is a finite difference between $AE)
and a,+ l ( ~ ) In
. a Mott insulator [28,66] this difference is infinitesimal, and C arises
entirely from a discontinuous change in 2 , across the Fermi level. As discussed
previously, the latter discontinuity is a pure correlation effect.
The local density and Langreth-Mehl approximations display no discontinuity of
Instead they average over it, yielding for an open-shell J-electron system
6 E X J6n (~).
reasonable estimates [7,39] of the averaged exchange-correlation potential
 DENSITY FUNCTIONAL THEORY 51 1

except in very low-density regions of space (where the exact 6Ex,/6n(~)lJ- 8 tends
[7,68] asymptotically to zero).
It is conjectured here that, for static external potential v([), normally C20. In the
Kohn-Sham system of noninteracting electrons, the total energy is certainly a
semiconvex function of the electron number N (i.e., the ionization energy is never
smaller than the electron affinity). It seems intuitively right that the total energy of
interacting (repelling) electrons should be no less convex than that of noninteracting
electrons of the same density. (Possible exceptions, with “negative electronic U,” are
certain transition-metal impurities in semiconductors [69].)
If this conjecture is true, then every crystal with a static external potential and a
nonvanishing fundamental gap in its exact Kohn-Sham band structure is an insulator.
The exact Kohn-Sham band structure then predicts an upper bound (or possibly an
exact value) for the volume per unit cell at which metallization occurs under pressure.
Band structure calculations within the local spin density approximation have been
performed for highly-compressed monatomic hydrogen [70], xenon [7 11, neon (721,
Ba S , Ba Se and Ba Te [73]. The calculated metallization volumes for Ba Se and Ba Te
exceed the observed ones by 2 to 16%.

5. Spin-Density Functional Theory and the Expanded Lattice of Hydrogen

Atoms
In spin-density functional theory [9,10], the density n ( ~ is) replaced as basic variable
by the separate “up” and “down” spin densities n , ( E ) , n (E):

E,[n,,n,l = TSln,,n,l + V[nl + U n l + Ex,[n,,n,l , (52)

where V[n] = Jd3rv(g)n(~). The Kohn-Sham equation (2) then involves a spin-
dependent effective potential

v:fr(L) = V(E) + u([nI;c) + 6ExJ6nu(E) . (53)

The density-functional expressions of Sections 1 and 2 have straightforward spin-

density analogs. In particular, there is a local spin density (LSD) approximation:

ExYD[ n ,n I = Jd3rn(E) Gxc ( n , ( E b (L) 1 (54)

6ExYD/6nu(r_)= F.“,(n,(L),n,(E)) 9 (55)

where ~ , , ( n , , n l ) is the exchange-correlation energy per particle of a uniform electron

,,
gas with up- and down-spin densities n n I , and p:,(n, , n , ) = d(nex,,)/dn,. An accuratc
parametrization of ex,(n ,n,) may be found in Ref. 11. In a spin-unpolarized system,
LSD reduces to LDA.
While the Kohn-Sham band structures of the Mott insulators [28,66] all look
“metallic” within density functional theory, they sometimes look “insulating” within
spindensity functional theory. Thus the LSD band structures [74] of the antiferromag-
netic Mott insulators MnO and NiO display small nonzero fundamental gaps, while
512 PERDEW

those of FeO and COO do not. This situation probably persists within the exact Kohn-
Sham spindensity functional theory, provided that the functionals are defined from
constrained searches over pure states rather than ensembles [75].
The simplest example of a Mott insulator, a lattrice of hydrogen atoms in the limit of
infinite lattice constant, provides an instructive case study of the ionization problem
from a total-energy viewpoint. In the separated-atom limit, the properties ofthe crystal
may be deduced from free-atom properties. When the atoms are hydrogen atoms, the
many-body problem also reduces to the one-body problem. To keep the following
discussion as simple as possible, replace the Coulomb interaction llr everywhere by
exp( - r/A)/r,and let the range A tend to infinity only after the lattice constant tends to
infinity.
Let each hydrogen atom be centered in its own unit cell of a Bravais lattice {R},with
N unit cells ( N + a ) . Let each electron have spin "up" due to a weak external magnetic
field. The Kohn-Sham orbitals can be chosen as site-localized Wannier functions
ww(c) = w& - &I), or equivalently as delocalized Bloch functions bk,(rJ = 2e"'k
W&)/*. All occupied orbitals have the same exact Kohn-Sham orbital energy
E = - 13.6 eV. The orbital densities are related in the following way:

for L in the u t h cell

0 otherwise .

The exact energy needed to remove an electron from either a Wannier or a Bloch
orbital is
hE" = AEb = -E = ATs + AV, (57)
or 13.6 eV. When an electron is removed from the crystal, the hole so created remains
on one site for a long (macroscopic) time before it tunnels to a neighboring site.
However, the hole is still delocalized over the entire crystal on the infinite-time
average.
Consider the detailed calculaton of AE"' and AEb, according to Eq. (52):

AE"' = E,[n - I w , , , ~-~ E,[n]

] + AV"' + AU"' + m,,
= A G (58)

AEb = E,[n - [&, I 2 - E,[n] = AT$ + AF + AUb + AE:, . (59)

It is not hard to see, from the definitions of T,(n]and V ( n ] , that A T = A C and

A P = AV! The quantity

AU"' + W c= - { W ) W , , , ~+
~ IE&%,l2,O1} (60)

is zero because an electron does not interact with itself. The quantity

obtained via Eq. (56) from the Taylor expansion of Eq. (59), is zero for the same
reason.
 DENSITY FUNCTIONAL THEORY 513

Now consider what is found in the LSD approximation. The right-hand side of
Eq. (60) becomes a small nonzero number ( -0.6 eV), while the right-hand side of
Eq. (61) becomes a somewhat bigger number (-6.3 eV): The self-interaction error of
the LSD total energy is magnified in the LSD potential. Thus in LSD the energy AEWpLSD
needed to create a localized hole is 13.0 eV, not far from the exact value of 13.6 eV.
However, the energy needed to create a delocalized hole is

or 7.3 eV, far less than 13.6 eV. The LSD total energy of the crystal with one hole is
minimized by delocalizing the hole, although the only realistic LSD estimate of the
ionization energy is obtained by localizing the hole.
These results persist qualitatively in real crystals (or at least in large-gap insulators),
and account for the LSD band-theory underestimation of the fundamental gap. In
particular, it is easy to see from Eq. (10) that Eq. (62) still relates the LSD eigenvalue at
the top of the valence band to the LSD total-energy cost to create a delocalized hole, and
that both quantities contain a serious self-interactionerror. In fact, Zunger and Freeman
[76] found that the fundamental gap in Li F could be estimated accurately within LDA,
not from the LDA bandstructure but from the LDA total energy in a self-consistent
calculation with a localized hole. This kind of self-consistent solution may be a local
extremum of the LDA total energy. The global minimum, however, delocalizes the hole
over the volume of the crystal (or over its surface, if more than one hole is present).
Carlsson [ a ] has presented a very interesting total-energy analysis of the LDA band
gap problem, and an approximate remedy based upon a modification of the long-range
behavior of the LDA expression for p x c ( ~ , ~of’ ) Eq. (17). Harrison [77] has discussed a
screened intraatomic Coulomb repulsion as a correction to the LDA band gap.
Finally, it is instructive to compare the local density and Hartree-Fock approxima-
tions. In each, a realistic estimate of the ionization energy of an insulating crystal may
be found from the total-energy cost to create a localized hole self-consistently,
including the dielectric polarization of the medium around the hole (i.e., to create an
“electronic polaron” [78]). Again in each, the orbital energy at the top of the valence
band, relative to the vacuum level, gives the corresponding total-energy cost to create a
delocalized hole. In LDA,the energy of the crystal with one delocalized hole does not
lie high enough above the energy of the neutral crystal, leading to band-structure
underestimation of the fundamental gap. In the Hartree-Fock approximation, the
energy of the crystal with one delocalized hole lies too high above the energy of the
neutral crystal, leading in part to band-structure overestimation of the fundamental gap.
(For example, the Hartree-Fock band gap of silicon (791 is 9.4 eV-eight times bigger
than the true gap.) As Kunz et a1 [80] have pointed out, the equivalence of Slater
determinants constructed in the Wannier and Bloch pictures holds only for a filled
band; in the presence of a hole, the Slater determinant in the Wannier picture can have
the lower Hartree-Fock energy.

6. Self-Interaction Correction to the Local Spin Density Approximation

The LSD self-interaction error discussed in Section 5 may be excised by the method
of self-interaction correction (SIC) of Perdew and Zunger [ 1 11:
514 PERDEW

E:;" = E x ! D [ ~ , , n ,-] Z, { U [ n , ] + Ex!D[n;,O]}, (63)

where n,(c) = ~ ~ ) I J J ~is( [the

) [ ~i-th orbital density. Minimization of the SIC energy

subject to the constraint of orbital normalization,

with respect to the orbitals IJJ~(I),
implies the Euler equation

This is a single-particle problem like Eqs. (2) and (8). but with 6Ex,./6n(~) replaced by
its LSD approximation ~ " c . = ~ ( (c),n,
n, (I)) plus an orbital-dependent correction

Avi;SIC
(c) = - 4 I n i l ; ~ )- P!xc(ni(c),O) . (66)

Like LSD, SIC is exact for the uniform electron gas-an old conjecture [ 1 I ] which
Norman 1811 has recently proved.
In self-consistent calculations for atoms I1 11. SIC yields improvements over LSD in
the total energy, the separate exchange and correlation energies, the binding energies of
negative ions (which find no self-consistent solution within LSD), and the long-range
behavior of the density and effective potential.
The SIC exchange-correlation hole obeys the sum rule I1 11

Jd3r p,S:C(c,e') = - Zf,n,(c)/"(E) . (67)

The right-hand side of Eq. (67) equals - 1 in a finite system with fixed integer electron
number (allJ's 0 or l), or in the ground-state of an infinite system with fluctuating
electron number. In a finite system with fluctuating (i.e., fractional) electron number,
the right-hand side of Eq. (67) is between 0 and - 1, as is the sum rule for the exact
exchange-correlation hole [38,39]. Thus SIC gives a significantly better account of
fractional electron number than LSD does. For a detailed review of SIC,see Ref. 82.
The SIC orbital energies cfIc equal aE'slc/afi, so SIC theory obeys the "aufbau
principle." Moreover, in an atom is often approximately linear infi. so that - E;'"
is a good approximation to m-fIClrcl, the relaxed SIC removal energy from orbital i:

- EfIC 5 AEf'Cl,, . (68)

Thus SIC provides an analog of Koopmans' theorem which includes intraatomic

relaxation.
Unlike the effective potential vefr(E) in Kohn-Sham theory, the SIC effective potential
v''ff(c) is orbital-dependent. Because it is both orbital-dependent and self-interaction-
free, the SIC potential can to some extent mimic the nonlocal, energy-dependent self-
energy [ 16,171 Z([,c';e). Thus in SIC calculations for atoms, Eq. (68) is close to an
 DENSITY FUNCTIONAL THEORY 515

TBLE IV. Energy eigenvalues for the Is t and Is 1 orbitals, and gap in the orbitals energies, for the
hydrogen atom (eV).

Approximation CAI -6) 1 + 6) €,(I + @ - € , ( I -6)

(1s t) (Isl) (gap)
LDA - 6.4 - 6.4 0.0
LSD -7.3 -2.3 5.1
SIC - 13.6 - 2.0 11.6
exact Kohn-Sham - 13.6 -0.8 12.8

-
equality for all the occupied orbitals [ 111. Within exact Kohn-Sham theory, by contrast,
-ei = AEilRl is an exact equality [6,33] for the highest-occupied orbital i, but the
orbital energy is not very close to the removal energy for the deeper orbitals [83,84].
Can the SIC orbital energies predict the fundamental gap? Table IV, for the hydrogen
atom, suggests that they might. While the gap in the orbital energies is zero in LDA, and
too small in ED, it is about right in SIC. Table V, for the neon atom, tells a similar tale:
The LSD and Langreth-Mehl orbital energies underestimate the gap, while the SIC
orbital energies predict the gap with good accuracy.
The ability of the SIC orbital energies to predict the gap is mostly a consequence of
the orbital-dependence of Eq. (66). Norman and Koelling [13] have constructed an
optimized effective potential (OEP) for SIC. That orbital-independent potential whose
+;
orbitals minimize the SIC energy is obtained from the solution of

where

The OEP-SIC effective potential may be the best approximation we have to the exact

TABLE
V. Fundamental gap for the neon atom, from the orbital energy eigenvalues in various approximation
and from experiment (with A = 0). LSD: local spin density approximation. LM: Langreth-Mehl approximation.
SIC: self-interaction correction. OEP-SIC: Norman and Koelling's optimized effective potential version of
SIC(eV).

LSD. LMb OEP-SIC' SIC' Exper.'

13.5 15.7 17.5 22.9 21.6

'Ref. 13.
Ref. 60.
516 PERDEW

TABLE
V1. Fundamental gaps of insulators, from the band structure within the local spin density (LSD) and
self-interaction corrected (SIC)approximations, and from experiment. The SIC numbers are based on afomic
estimates of the self-interaction correction (eV).

Crystal LSD SIC Exper.

Ne‘ 11.2 21.1 21.4

Ar’ 8.3 14.1 14.2
w 6.8 11.7 11.6

a Ref. 1 1 .

Kohn-Sham potential in the neon atom (Fig. I ) . Table V shows that the OEP-SIC orbital
energies seriously underestimate the fundamental gap.
In atomiclike insulators, such as the rare-gas solids, the self-interaction correction to
the fundamental gap for an atom may be added to the LSD band gap of the solid,
yielding an estimate of the true solid state gap [ I I] (Table VI). (Improved gaps from a
different brand of self-interaction correction were found by Kunz et al. [80].)
For these atomiclike insulators, SIC has an immediate intuitive appeal. It asserts that
an electron in the valence band of argon, like a valence electron in the neutral Ar atom,
sees an Ar+-like potential in each unit cell, while an “excess’’ electron in the
conduction band sees an Aro-likepotential in each cell, as demanded by Mott [28]. LSD,
on the other hand, forces an Aro-like potential for both bands [ 1 I ] .
However, it is not so clear how SIC should be applied in more general classes of
solids. In fact a straightforward application of Eqs. (63) and (66) to the Bloch orbitals
yields no correction to LSD in the infinite-volume limit.
Perdew and Norman [83] addressed this problem by constructing an energy-
dependent effective potential which simulates the self-interaction correction in atoms:

where ii = Vn/)Vnl is a unit vector, and

is the local density of occupied one-electron states. The SSIC potential for a given band
in a solid is actually constructed from 4.(71), with ~ ( E , E )replaced by its integral over
energies in the band. With Bloch functions +i(~), this scheme involves a band-
dependent periodic correction to the LSD periodic potential. Fundamental gaps from
self-consistent SIC band structures [63,72] of Ne and Na C1 are shown in Table VII.
The simulated self-interaction correction (SSIC) has also been applied [85] to the
semiconductor Si, where it increases the fundamental gap from its LDA value of 0.6 eV
to 2.0 eV-an overshoot of the true gap (1.2 eV). It’s hard tosay whether this is a
failure of s i c itself, or merely of its simulation in SSIC. At any rate, there is evidently a
need for a better-founded implementation of SIC in solids.
 DENSITY FUNCTIONAL THEORY 517

TABLE VII. Fundamental gaps of insulators, from the band structure within the local spin density (LSD) and
self-interaction corrected (SIC) approximations, and from experiment. The self-consistent SIC band structure
calculations follow the methods of Ref. 63 and Ref. 12 (eV).

Crystal LSD SIC Exper.

Ne' 11.5 20.2 21.4

NaCI' 5.6 9.2 9.0
AP 1.9 13.5 14.2
LiClb 6.0 10.6 9.9
a Ref. 63.
Ref. 12.

This need is met, at least for large-gap insulators, by the SIC band-structure scheme
of Heaton, Harrison and Lin [12]. They first introduce off-diagonal Lagrangian
multipliers to insure orthogonality of the SIC orbitals, and then interpret the solutions of
the resulting single-particle equations,

as site-localized Wannier orbitals w,& = w,(c - E"). It is these Wannier orbitals which
are employed to evaluate the SIC exchange-correlation energy (63). The corresponding
delocalized Bloch orbitals,

bnk(c) = N - 2" e'k'k w, (c - Eu), (74)

obey a single-particle equation like (73), but with the band- and wavevector-dependent
periodic correction to the LSD periodic potential

where

is a weighting factor which sums over sites u to unity.

Heaton, Harrison and Lin [ 121 have used these equations to generate self-consistent
SIC band structures E ~ E ,for
. ~ Ar and Li CI. The resulting fundamental gaps, displayed in
Table VII, show about the same degree of improvement over LSD as do those of SSIC
[63]. However, Lin's method is preferable to SSIC because it is variationally based and
its SIC bands reduce exactly to the original SIC orbital energies in the separated atom
limit.
An essential step in the evaluation of Eq. (75) is the construction of the wannier
orbitals from the Bloch orbitals:
518 PERDEW

The summation over the Brillouin zone is tedious, and can be replaced by a simpler
sum [ 121 in atomiclike insulators. In semiconductors, where this replacement is not
justified, no SIC calculation has yet been performed. However, the SIC method has been
generalized and applied successfully to molecules [86].
In fact, application of SIC to semiconductor band structure would be fully justified
only after certain fundamental questions were answered: (1) Should there be some
correction to the bottom of the LSD conduction band, [ 12,401 in addition to that for the
top of the LSD valence band? (2) While the Wannier orbital densities can be used in the
SIC exchange-correlation energy of Eq. (63) when all bands are either full or empty,
what orbital densities should be used for partly-filled bands? This question is important
not only for metals, but also for insulators with extra electrons or holes. (3) Does the
SIC total energy of Eq. (64)properly account for the extra-atomic relaxation effects that
occur when an electron is added to or removed from a semiconductor? (4) Away from
the separated-atom limit, how do the orbital energies E Z ~ , . from
~ Q. (75) relate to SIC
total-energy differences?

7. Dielectric Screening
So far in this paper, the criterion for metallic behavior has been a vanishing
fundamental gap. An alternative criterion for metallicity is perfect screening: When an
infinitesimal external charge 6 N + (of either sign) is introduced into a metal, which is in
equilibrium at zero temperature with an electron reservoir, the metal draws S N ,
electrons from the reservoir to neutralize the external charge. By energy arguments,
these two critera are equivalent.
The response of the electron density to an infinitesimal external perturbation & ( I ) is

The inverse of the density response function is [87] the second functional derivative of
the Hohenberg-Kohn energy of Eq. (3):

where xs is the density response function of the Kohn-Sham noninteracting system and

due to a change in
The change in the exchange-correlation potential vX&) = 6EX,/6n(r_)
density is evidently
 DENSITY FUNCTIONAL THEORY 5 19

Q. (78) can also be cast in the more familiar form [36]

An insulator is distinguished from a metal by the condition Jd3r8n(r_)= 0, or

Eq. (84) should follow mathematically from Eq. (79) and from the nonzero derivative
discontinuity, Eq. (46), for an insulator. From Eq. (82), it follows that for an insulator

where

and fl is the volume of the crystal.

For the uniform electron gas, Eq. (83) Fourier transforms into (21

where x,(Q) is the Lindhard function, and Eq. (82) becomes

Expansion of KJQ) in powers of Q,

K,,(Q) = K,,(O) + YX;,(O)Q2 + ... , (89)

amounts to a density-gradient expansion of the exchange-correlation potential. The

leading term in (89), K,,(O), is the LDA term.
Recently Antoniewicz and Kleinman [ 141 have calculated K,(Q) exactly within the
exchange-only version of density-functional theory. They find that K,(Q)/Kx(0)
increases smoothly from 1 to -2 in the range 0CQ52kF. Thus, in a uniform electron
gas perturbed by a weak external potential 6v(Q), the exact Kohn-Sham exchange
potential 6v,(Q) can be 50-100% stronger than it is in LDA.
Antoniewicz and Kleinman [14] then propose a model for a semiconductor: a
uniform electron gas perturbed by a weak periodic external pseudopotential. The
important reciprocal lattice vectors for the diamond and zincblende semiconductors fall
in the range O<QS2kF. The exact exchange potential is enhanced by about 50% over
the LDA exchange potential in this range. This enhancement, corresponding roughly to
520 PERDEW

0.5- uniform electron gas

Figure 2. Exchangecorrelation response function K J Q ) of Eq. (88) for the uniform

electron gas with r, = 2, according to the estimate of Ref. 89. The Q-independent behavior of
the local density approximation (LDA) is indicated by the dashed line. Vertical arrows indicate
the ( I I I ) , (200). (220) and (31 1 ) reciprocal lattice vectors of the zincblende structure,
according to Ref. 14.

a = 1 in place of a = 2/3 in an Xa band-structure calculation, is large enough to bring

the calculated band gaps into agreement with experiment. Antoniewicz and Kleinman
therefore suggest that the exact Kohn-Sham potential can yield the observed
fundamental gap in a semiconductor, despite the LDA underestimate of the gap.
The exact Kxc(Q) for the uniform electron gas, including correlation as well as
exchange, is not known. However, Taylor [88] has observed that the Geldart-Taylor
estimate of Kxc(Q)/Kxc(0) does not deviate much from unity over the range O<Q52kF.
This behavior is also found in more recent estimates of Kx,(Q). For example, Utsumi
and Ichimaru [89] have estimated the “local field correction” C ( Q ) in the uniform
electron gas with a high degree of internal consistency, and presented an analytic
parametrization of it from which

has been constructed in Figure 2. Thus, when correlation is included, the Antoniewicz-
Kleinman model of a semiconductor actually suggests that the exact Kohn-Sham
potential underestimates the true gap by about as much as the LDA does.

8. Conclusions
(1) The fundamental gap in the exact Kohn-Sham bandstructure of an insulator is not
equal to the true gap I-A. ( 2 ) The difference arises because the exchange-correlation
potential 6 E x c / 6 n ( changes
~) discontinuously, by a finite constant, as the conduction
band begins to fill. ( 3 ) Several very different (but admittedly approximate) numerical
calculations suggest that most of the error in the gap of the local density approximation
would persist in the gap of the exact Kohn-Sham bandstructure.
 DENSITY FUNflIONAL THEORY 52 1

Acknowledgments
Thanks for preprints are due to many authors including G. B. Bachelet, A. E.
Carlsson, W. A. Harrison, L. Kleinman, C. C. Lin and M.R. Norman. Special thanks
to Mike Norman for calculating some of the input to Figure 1, and to the organizers of
the 25th Anniversary Sanibel Symposia for their invitation to lecture. This research
was supported in part by the National Science Foundation, Grant No. DMR80-16117.
Note. Some of the questions raised at the end of Section 6 have been addressed
recently by M.R. Pederson, R. A. Heaton and C. C. Lin (unpublished), and by R. A.
Heaton and C. C. Lin, Bull. Am. Phys. SOC. 30,225,(1985).

Bibliography
[I] R. G. Parrand R. G. Pearson, J. Am. Chem. SOC. 105,7512 (1983).
(21 W. Kohn and L. J. Sham, Phys. Rev. 140, A1 133 (1%5).
[3] P. Hohenberg and W. Kohn, Phys. Rev. 9, B864 (1964).
[4] D. C. Langreth and M. J. Mehl, Phys. Rev. B 28, 1809 (1983).
(51 D. C. Langreth and J. P. Perdew, Phys. Rev. B 21, 5469 (1980).
161 J. P. Perdew R. G. Pam, M. Levy and J. L. Balduz, Jr.. Phys. Rev. Lett. 49, 1691 (1982).
[7] J. P. Perdew and M. Levy, Phys. Rev. Lett. 51, 1884 (1983).
[8] L. J. Sham and [Link], Phys. Rev. Lett. 51, 1888 (1983).
(91 U. von Barth and L. Hedin, J. Phys. C 5, 1629 (1972).
[lo] A. K. Rajagopal and J. Callaway, Phys. Rev. B 7, 1912 (1973).
[I I] J. I? Perdew and A. Zunger, Phys. Rev. B 23,5048 (1981). and references therein.
[I21 R. A. Heaton, J. G. Harrison and C. C. Lin, Phys. Rev. B 28,5992 (1983); R. A. Heaton and C. C.
Lin, 1. Phys. C 17, 1853 (1984).
[I31 M. R. Norman and D. D. Koelling, Phys. Rev. B 30, 5530 (1984).
(141 P. A. Antoniewicz and L. Kleinman, Phys. Rev. B (to appear).
(151 E Herman, private communication.
(161 L. I. Sham and W. Kohn, Phys. Rev. 145, 561 (1966).
[I71 L. Hedin and S. Lundqvist, Solid State Phys. 23, 1 (1969).
(181 E Sacchetti, J. Phys. F 12, 281 (1982).
1191 A. H. MacDonald, M.W.C. Dharma-wardana and D. 1. W. Geldart, 1. Phys. F 10, 1719 (1980).
[20] W. E. Pickett and C. S. Wang. Phys. Rev. B 30,4719 (1984).
[21] G. Strinati, H. J. Mattausch, and W. Hanke. Phys. Rev. B 25,2867 (1982).
(221 S. Horsch, P. Horsch, and P. Fulde, Phys. Rev. B 28,5977 (1983).
(231 P. Steme and J. Inkson, J. Phys. C17. 1497 (1984).
[24] W. Hanke. T. Golzer. and H. J. Mattausch, Solid State Commun. 51, 23 (1984).
[25] E. 0. Kane, Phys. Rev. B 4, 1910 (1971); Phys. Rev. B 5, 1493 (1972).
[26] D. R. Hamann, Phys. Rev. Lett. 42, 662 (1979).
[27] N. 0. Lipari and A. B. Kunz, Phys. Rev. B 3,491 (1971).
(281 N. F. Mott, Metal-Insulator Transitions (Taylor and Francis, London, 1974).
(291 J. E Janak, Phys. Rev. B 18,7165 (1978).
[30] J. C. Slater, The Self-consisrent Field for Molecules and Solids (Mdjraw-Hill, New York, 1974); J.
W. D. Connolly, in Semiempirical Merhods of Elecrronic Structure Calculation. Part A: Techniques, G.
A. Segal. Ed. (Plenum, New York, 1977).
[31] S. B. Trickey, Chem. Phys. Lett. 21,581 (1973).
(321 A. R. Williams and U. von Barth, in Theory of the Inhomogeneous Electron Gas.S . Lundqvist and
N. H. March, Eds. (Plenum, New York, 1983).
(33) U. von Barth, in Electronic Srrucrure of Complex Materials. NATO Advanced Studies Institutes Series
B, Physics, Vol. 113, P. Phariseau and W. Temmermann, Eds. (Plenum, New York 1984).
522 PERDEW

[34] J. C. Slater and J. H. Wood, Int. J. Quantum Chem. S 4, 3 (1971).

(351 S. B. Trickey and J. P. Worth, Int. J. Quantum Chem. S 11,529 (1977).
[36] D. C. Langreth and J. P. Perdew, Solid State Commun. 17, 1425 (1975); Phys. Rev. B 15, 2884
(1 977).
(371 0. Gunnarsson and B. 1. Lundqvist, Phys. Rev. B 13,4274 (1976).
[38] I. P Perdew and M. Levy, in Many-Body Phenomena ar Surfaces. D. C. Langreth and H. Suhl, Eds.
(Academic, New York, 1984).
[39] J. P. Perdew, in Demiry FunctioMl Methods in Physics, R. M. Dreizler and J. da Providencia, Eds.
(Plenum, New York, 1985).
[40] A. E. Carlsson, Phys. Rev. B, April 15, (1985).
(411 C. S. Wang and B. M. Klein, Phys. Rev. B 24, 3393 (1981).
[42] G . A. Sawatzky and J. W. Allen, Phys. Rev. Lett. 53,2339 (1984).
(431 A. Zunger and M. L. Cohen. Phys. Rev. B 20,4082 (1979).
(441 G. B. Bachelet and N. E. Christensen, Phys. Rev. B 31. 879 (1985).
(451 F. Manghi, G. Riegler, C. M. Bertoni, C. Calandra and G. B. Bachelet, Phys. Rev. B 28,6157 (1983);
erratum 29, 5966 (1984).
[46] J. E. Jaffe and A. Zunger, Phys. Rev. B 28, 5822 (1983).
[47] A. Zunger and A. J. Freeman, Phys. Rev. B 17,4850 (1978).
[48] A. Zunger and A. J. Freeman. Phys. Rev. B 17, 2030 (1978).
[49] A. Zunger and M. L. Cohen, Phys. Rev. B 20, I189 (1979).
[SO] M. A. Khan and J. Callaway, Physics Letters A 76, 441 (1980).
[51] S. B. Trickey, A. K. Ray, and J. P. Worth, Phys. Stat. Sol. (b) 106,613 (1981).
[52] 0. Gunnarsson, M. Jonson, and B. 1. Lundqvist, Phys. Rev. B 20,3136 (1979).
[53] J. A. Alonso and L. A. Girifalco, Phys. Rev. B 17, 3735 (1978).
[54] 0. Gunnarsson and R. 0. Jones, Phys. Scr. 21. 394 (1980).
[55] M. S. Hybertsen and S. G. Louie, Phys. Rev. B 30. 5777 (1984).
[56] D. J. W. Geldart and M. Rasolt, Phys. Rev. B 13, 1477 (1976).
[57] J. P. Perdew, D. C. Langreth, and V. Sahni, Phys. Rev. Lett. 38, 1030 (1977).
[58] M. Rasolt and D. J. W. Geldart, Phys. Rev. B 21, 3158 (1980).
[59] K. Aashamar, T. M. Luke, and J. D. Talman, Phys. Rev. A 19.6 (1979).
[60] M. R. Norman, private communication.
[61] D. R. Beck, Int. J. Quantum Chem. 16, 345 (1982).
[62] M. R. Norman and D. D. Koelling, Phys. Rev. B 28,4357 (1983).
[63] M. R. Norman and J. P. Perdew, Phys. Rev. B 28,2135 (1983).
[64]U. von Barth and R. Car, unpublished work cited in Ref. 33.
[65] J. C. Inkson, Many-Body Theory ofSolids (Plenum, New York. 1984). p. 270.
[ a ] B. H. Brandow, Adv. in Phys. 26,651 (1977).
(671 I. P. Perdew and M. Levy, Phys. Rev. B 15 (to appear).
[68] M. Levy, J. P. Perdew, and V. Sahni, Phys. Rev. A 30,2745 (1984).
[69] A. Zunger and V. A. Singh, Bull. Am. Phys. SOC.30,302 (1985).
[70] L. M. Sander, H. B. Shore, and J. H. Rose, Phys. Rev. B 24,4879 (1981).
[71] M. Ross and A. K. McMahan, Phys. Rev. B 21, 1658 (1980); A. K. Ray, S. B. Trickey. R. S.
Widman, and A. B. Kunz, Phys. Rev. Lett. 45, 933 (1980); A. K. Ray, S. B. Trickey, and A. B.
Kunz, Solid State Commun. 41, 351 (1982).
[72] J. Hama, Phys. Lett. A 105, 303 (1984).
[73] A. E. Carlsson and J. W. Wilkins, Phys. Rev. B 29, 5836 (1984).
[74] K. Terakura, A. R. Williams, T. Oguchi and J. Kubler, Phys. Rev. let^. 52, 1830 (1984).
[75] J. P. Perdew and M. Levy, unpublished.
[76] A. Zunger and A. J. Freeman, Phys. Rev. B 16, 2901 (1977).
[77] W. A. Harrison, Phys. Rev. B 31, 2121 (1985).
 DENSITY FUNCTIONAL THEORY 523

[78] W. B. Fowler, Phys. Rev. 151.657 (1966); A. B. Kunz, Phys. Rev. B 6,606 (1972).
I791 R. Dovesi, [Link], and G. Angonoa. Phys. Rev. B 24,4177 (1981);J. C. Phillips and L. Kleinman,
Phys. Rev. 128, 2098 (1%2).
[80] A. B. Kunz, R. S. Weidman, J. Boettger and G. Cochran, Int. J. Quantum Chem. S 14, 585 (1980).
[81] M.R. Norman, Phys. Rev. B 28,3585 (1983).
[82] J. P. Perdew, in Local Density Approximations in Quantum Chemistry and Solid State Physics. J. P.
Dahl and I. Avery. Eds. (Plenum, New York, 1984).
[83] 1. P. Perdew and M.R. Norman, Phys. Rev. B 26,5445 (1982). For a comment by S. B. Trickey and a
reply, see Phys. Rev. B 30,3523 and 3525 (1984).
[84] C.-0. Almbladh and A. C. Pedroza, Phys. Rev. A 29,2322 (1984).
[85] M.R. Norman, Phys. Rev. B 29, 2956 (1984).
[86] M.R. Pederson, R. A. Heaton. and C. C. Lin, J. Chem. Phys. 80, 1972 (1984).
[87] K. F. Freed and [Link]. J. Chem. Phys. 77, 3% (1982).
[88] R. Taylor, J. Phys. F 8, 1699 (1978); D. J. W. Geldart, [Link], and R. Taylor, Solid State Commun.
10,279 (1972).
[89] K. Utsumi and S. Ichimaru. Phys. Rev. A 26, 603 (1982).

Received May 20, 1985