PHYSICAL REVIEW B VOLUME 23, NUMBER 10 15 NIAY 1981

Self-interaction correction to density-functional approximations for many-electron systems

J. P. Perdew
Department of Physics and Quantum Theory Group, Tulane University, Netv Orleans, Louisiana 70118

Alex Zunger
Solar Energy Research Institute, Golden, Colorado 80401
and Department of Physics,University of Colorado, Boulder, Colorado 80302
(Received 31 October 1980)
The exact density functional for the ground-state energy is strictly self-interaction-free (i.e., orbitals demonstrably
do not self-interact), but many approximations to it, including the local-spin-density (LSD) approximation for
exchange and correlation, are not. We present two related methods for the self-interaction correction (SIC) of any
density functional for the energy; correction of the self-consistent one-electron potenial follows naturally from the
variational principle. Both methods are sanctioned by the Hohenberg-Kohn theorem, Although the first method
introduces an orbital-dependent single-particle potential, the second involves a local potential as in the Kohn-Sham
scheme. We apply the first method to LSD and show that it properly conserves the number content of the exchange-
correlation hole, while substantially improving the description of its shape. We apply this method to a number of
physical problems, where the uncorrected LSD approach produces systematic errors. We find systematic
improvements, qualitative as well as quantitative, from this simple correction. Benefits of SIC in atomic calculations
include (i) improved values for the total energy and for the separate exchange and correlation pieces of it, (ii)
accurate binding energies of negative ions, which are wrongly unstable in LSD, (iii) more accurate electron densities,
(iv) orbital eigenvalues that closely approximate physical removal energies, including relaxation, and (v) correct long-
range behavior of the potential and density, It appears that SIC can also remedy the LSD underestimate of the band
gaps in insulators (as shown by numerical calculations for the rare-gas solids and CuCl), and the LSD overestimate
of the cohesive energies of transition metals. The LSD spin splitting in atomic Ni and s-d interconfigurational
energies of transition elements are almost unchanged by SIC. We also discuss the admissibility of fractional
occupation numbers, and present a parametrization of the electron-gas correlation energy at any density, based on
the recent results of Ceperley and Alder.

I. INTRODUCTION
n(r) =Q n „(r),
A major problem of solid-state theory and quan-
tum chemistry is to understand the behavior of where
many electrons interacting via Coulomb's law:

is the density of an orbital with quantum numbers

n and a, and a=+-'(i) or ——, '(w) is the spin. In
(All equations are in atomic units, if=m = e'= ) I. this approximation, the total interelectronic ener-
In the earliest quantum-mechanical theory, Thom- gy is given as a sum of direct and exchange con-
tributions:
as and Fermi replaced the expectation value ( „)
by the direct Coulomb energy, a functional of the
electron number density n(r):
where the exchange energy is
n(gin(r )
[8 ] y g
X da X
j
(2)

As early as 1934, Fermi and Amaldi' observed the

failure of Eq. (2) to vanish for one-electron sys-
The o. '= ct terms in Eq. ( t) (the only ones retained
tems due to the spurious self-interaction inherent
in the Hartree approximation) constitute a self-
in it, and proposed the first and crudest version
exchange energy:
of self-interaction correction:
self-exchange .= — g U[n, ] (8)

which correctly cancels the self-Coulomb energy

where N is the number of electrons in the system. in U[n] on an orbital-by-orbital basis. Table I
The mean-field Hartree-Fock (HF) approximation (Ref. 2 shows how much of the exchange energy
introduced orbitals g„(r) with occupation numbers in atoms is simply self-exchange.
f, obeying Fermi statistics: Although the HF approximation produces only a

5048
 SELF-IHTKRACTIOX CORRECTION TO DENSITY-FUNCTIONAI . ..

TABLE L. Hartree-Fock total energy, exchange energy, and interelectronic exchange

energy (which omits self-exchange) for lighter inert gas atoms (from Bef. 2). The correla-
tion energy is from Bef. 62. Energies in eV.

Total energy Exchange

He 2 -77.9 -27.9 0.0 -1.1

Ne 10 -3497.8 -329.5 -60.3 -10.4
Ar 18 -14 334.7 -821.3 -203.5 -19.9
Kr 36 -74 883.4 -2561.9 -908.3

very small relative error in the total energy of The quantitative successes of the LSD approxi-
atoms, it is not suitable for most solid-state and mation have been impressive, particularly for
many molecular applications. The total energy is molecular bonding '0'x metallic magnetism x' x4
dominated by high-density inner-shell electrons cohesion, "*" and the surface electronic properties
that are well described by Hartree-Fock theory, of metals"*" aIld semi-conductors. "-'0 Neverthe-
but one is more usually interested in the low-den- less, there remain some worrisome failures of
sity valence electrons, for which correlation (ne- LSD, some of which have been revealed only re-
glected in HF theory) may be as important as ex- cently. These include the facts that:
change. In addition, the long range of the Cou-
lomb interaction produces unrealistic features in
"
is too Lose when compared with the exact value, ""
(i) While the LSD total energy of a metal surface

the HF energy eigenvalues, e.g. , vanishing density the LSD energy for atoms is Ioo high. Further-
'
of states at the Fermi level in metals, unphysi- more~ the lowest-older correction to the LSD ex-
cally large band gaps in insulators (sometimes by
'
a factor of 2-5), etc. While correlation effects
change-correlation energy predicted by many-body
theory, i.e. , the density-gradient correction, " is
can be addressed through configuration-interaction positive and so its inclusion can only worsen the
corrections to HF, the complexity of these cor- agreement between the calculated energies of
rections, their remarkable sensitivity to the choice atoms and experiment.
of basis functions, and the increase in effort re- (b) For atoms" the magnitude of the exchange
quired with the decrease in spacing between ener- is consistently underestimated by 10-15% in LSD.
gy levels, preclude application to large systems. The magnitude of the correlation energy is over-
The density-functional theory' ' provides an al- estimated by as much as 100-200 %%.
ternative to this approach. This theory which is (iii) The experimentally stable negative ions
exact in principle includes correlation explicitly. (e.g. , H, 0, F ) are predicted to be unstable in
in the total energy and one-body potential; typ- LSD 25, 26
ically in practice exchange and correlation are (iv) .Self-consistent LSD band-structure calcula-

approximation. "
treated together in the local-spin-density (LSD)
This essentially statistical ap-
proximation works fairly well even for one-elec-
tions systematically underestimate the one-elec-
tron energy gaps of insulators by as much as 40%
(e.g. , Refs. 27-31). More generally, the LSD one-
tron systems': In the hydrogen atom, the direct electron energy eigenvalues are not close to phy-
(self-interacting) Coulomb energy U[n] is large sical removal energies from bound states. There
(3.5 eV), but about 93% of this spurious energy is are often large deviations from experiment when
cancelled by the LSD exchange-correlation ener- I SD energy eigenvalues are identified with the
gy. Unlike HF theory where the self-interaction positions of surface states, deep defect levels in
is eomPLetely canceJed because the direct elec- solids, core bands in solids, etc.
trostatic and exchange operators have the same (v) The long-range behavior of the LSD one-elec-
kernel 1/~r —r' ~, LSD achieves only a Partial tron potential for ions of charge Q is -Q/r, rather
cancellation, assisted largely by spin-polarization than the electrostatically correct limit of -(@+1)/
and correlation effects. Hence, in LSD a residue Among other problems, this leads to an erron-
of spurious self-interaction remains as the price eous description of charged point-defect states in
to be paid for a simple, local one-electxon poten- solids
tial. Clearly, this self-interaction vanishes for (vi) The LSD calculated spin splitting of the en-
orbitals delocalized over extended systems. We ergy bands in ferromagnetic metals fe. g. , Ni (Ref.
wil. l see, however, that self-interaction l.eads to a 32)] is often much larger than the observed value.
number of systematic errors for finite systems (vii) LSD total energies unduly favor the d" 's'
and localized states in extended systems. configuration over the d" 's' configuration in sd'
 5050 J. P. PERDE% AND ALEX ZUNGER

transition atoms. "

In addition, the LSD ordering expectation value. Then the usual variational
of s and d levels in the d" 's' configuration of the principle ((B}~ the ground-state energy E) implies
3d elements Sc to Mn is reversed relative to HF. that for any choice of trial spin densities n& and n&
Much of the criticism of the LSD approximation representing N electrons,
has been directed toward its incomplete descrip-
tion of many-body correlation effects. In this Q [n [, n &]+ „cPr v(r)n (r) —2)u J d'r B(r)g([n, (r}~ &,
paper we show that many of the problems indicated
above are interrelated and that a simple, phy-
sically transparent one-body correction to LSD— .
where n=S~, n, Furthermore, the true ground-
the subtraction of spurious self-interaction ac- — state spin densities are those which make Eq. (11)
counts for and corrects most (but not all) of the an equality. (This is the generalized Hohenberg-
qualitative and quantitative discrepancies. Empir- Kohn' theorem. Note that the argument does not
ical methods used previously for correcting some require a nondegenerate ground state, nor does
of these discrepancies (e. g. , use of adjustable ex- it require that the trial spin densities be v repre-
change coefficient a, empirical gradient terms, sentable. )
etc. ) become unnecessary. A preliminary version
of our work has already appeared. " The variational principle
culate n&, n&, and 8, if the
(11) may be used to cal-
functional dependence
of Q on n& and n& is known. It is convenient to
II. FORMALISM break Q up into three parts as follows:
A. Spin4ensity-functional theory T[n[, n]]+ U[n]+&„,[n&, n]],
Q[n)], n&]= (12)
We begin with a review of spin-density-functional where T[n&, n, ] is the "noninteracting" kinetic ener-
theory, ' ' including a synthesis of recent develop- gy defined below, U[n] is the direct Coulomb en-

of Levy. "
ments, following the simple and elegant approach
This formally exact theory will then
ergy of Eq. (2), and E„,[n&, n&] is the exchange-
correlation energy.
serve to introduce self-interaction corrections in The spin densities can always be written (nonun-
Sec. IIB. iquely) in terms of orthonormal orbitals [t),(r)
Consider a system of N electrons with the Ham-
iltonian
and occupation numbers in the interval 0& f, & 1:
(13)
a=i+ V'„+ Qv(r, .) —2([[ QB(r, }s,', (9)
S t
We define
where v(r) and B(r) are external scalar and mag-
netic fields (the latter coupling to electron spin r[n~, n, ]=mm(L(„, (d„~ — v'ld —',
)) (44)
s,'), and T is the kinetic energy operator. Let
n&(r) and n& (r) be up- and down-spin number densi- where the minimum is over all possible f„and
ties, and define the universal functional" g, yielding the given spin densities. (For further

Q[n&, n]] = min(T+ V„), (10)

Following Janak, "
discussion of this definition, see Appendix A. )
the minimization of the left-
which searches the set of all N-particle antisym- hand side of Eq. (11) can now be accomplished in
metric wave functions producing the given spin two steps:
densities n, (r) = (n, (r )}, and delivers the minimum
I
(1) Assume a fixed set of f,
and minimize

f
4'
gf 'v'] =4,) +U„[,n] ~ z„[
(4„]—— ~. , nr]+ d r n2 Vfd rBn(v)L nn(r)
(r) (n) —

[with n, given by Eq. (13)] with respect to the g, (r), (17)

subject to the normalization constraint on the or- with an effective one-body potential
bitals. The resulting Euler equation is
v', s(r) = v(r) —2@crB(r) + u([n]; r)

+ v„'.([n„n,]; r), (18)

where the direct Coulomb potential is
f„ is a Lagrange multiplier. This leads
directly to the Kohn-Sham' one-electron self-con- , ,' n(r'}
u([n]; r) = 6n — = d'r (19)
sistent equations: r U[n] I r —r'I '
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. . .

and the exchange-correlation potential is is the density at r' of the hole around an electron
at r. +, is the ground state of a fictitious system
v„',([nk, n, ]; r) = ——,Z„,[nk, kk, . (20)
in which the electron-electron interaction is
] A/If'-r'I and the spin densities are those of the
real system (A. = 1). The A integration appears be-
The significance of the eigenvalue can be seen q, cause of the definitions of T[n&, n&] (e (7})and II[n]
by an argument of Slater' and Janaksv: Differenti- +&„,[nk, n&] (e (V„)); the integrand of Eq. (25) at
ation of Eq. (15) with respect to the occupation the limit X=O is the exchange hole. From the def-
f
number ~~ (wkth all ok'bitals fkxed) ykelds inition (25), one gets

cA 4*,(r) [ -'v'+ v', kk(r)]y, (r)

d'~'p(r, r )=-1, (26)

(21} i.e. , the electron and its hole together constitute

a neutral object.
Since E has been minimized with respect to nor- Spin-density-functional calculations would yield
malization-preserving variations of the orbitals, exact results if the exact E„,[n&, n&] were known
it follows that in any infinitesimal, normalization- and used. In practice one often makes the local-
preserving variation of the occupation numbers spin-density (LSD) approximation"
and orbitals,
E (2V)

where &„,(n&, n&) is the exchange-correlation energy

per particle of an electron gas with uniform spin
(2) Now minimize E with respect to the f „sub-

=¹
ject to the particle-conserving constraint
This is equivalent to the unconstrained" min-
Q, f, densities g&, g&. In this way the results of many-
body calculations for the homogeneous system,
which have recently become quite reliable, 4'4' can
imization of 8-
pN, where p, is the chemical po- be folded into calculations for inhomogeneous sys-
tential [= —electronegativity]":
tems. The functional derivative (20) gives the
5(g ik~)=g(&. . P}5/. . O. (23) LSD exchange-correlation potential
v„"""(r) = p'(n, (r), n, (r)), (28)
From Eq. (23) we immediately find an "aufbau
principle"'": orbitals with &, & p have
and those with q ik have ) =0. f
= 1,
Fk'actiokkal oc-
f, where ik„',(n&, n„) = s[ns„,(n&, n, )]/&n.
By construction the LSD approximation becomes
.
cupation is found only for q, .
= p, If it turns out exact when the spin densities vary slowly enough
on the scale of the local Fermi wavelength and
that all occupation numbers are O or 1, then the
ground-state spin densities of the real system are screening length, a formal validity condition that
also ground-state spin densities of a fictitious sys- is seldom satisfied in real systems. For xeasons
tern of noninteracting electrons moving in the po- reviewed in Sec. G C, the I SD approximation is
tential v', s(r), as was assumed in the original still reasonably successful for real systems, while
Kohn-Sham' derivation of Eq. (17). This assump- inclusion of the next term in the density-gradient
tion may not be universally true, and in the pres- expansion does not give the systematic improve-
ent formulation it is unnecessary. If this assump- ment that might naively be expected.
tion is true, and if in addition the interacting
ground state is adiabatically connected to the non- B. Self-interaction correction
interacting one, then the exchange-correlation en- From our definitions of the functionals Q, T, and
ergy may be written as half the sum of the elec- E„„ it follows at once that the exchange-correla-
trostatic interaction of each electron with its pos- tion energy of a single, fully occupied orbital must
itively charged exchange- correlation hole"'": exactly cancel its self-dixect Coulomb energy,
r') l.e.
0

, + , kk(r)P(r
y

E d, ,
t (24)
. .
fI[n. ]+ E„,[n. . 0] = 0, (29)
where where n, , =f.. .
Iy. I' and f„.
=1. (Note that a
single orbital is fully spin polarized, and note fur-
p(r, r') = dik(+, [n(r) —n(r)]
I
ther that the orbital density n,
does not have to be
a possible ground-state density. ) Although Eg.
x [n(r ) —n(r )]/&(&) —~(r —r }I+„& (29) is hardly surprising, this is the first proof
(25) of it to our knowledge. By a suitable continuation
 5052 J. P. PERDEW AND ALEX ZU1VGER

of E„,to noninteger numbers of electrons, we can Since the LSD approximation for exchange is
also satisfy Eq. (29} for f,
& 1. (Physical reasons
for doing so are presented in Appendix B.) Fur- 33, [ 1, i]= —0.9305fd'3[ ]"( )+ )'( )], (35)
ther insight can be gained if we define the ex-
change energy E„[n&,n&] as the right-hand side of the rough cancellation between U[n, ] and
Eq. (7), which is a functional of the spin densities E, [n„, o] for a single occupied orbital is evi-
because the Kohn-Sham orbitals and occupation dent. The spurious residue of this incomplete
numbers are. (This definition is obviously the cancellation may, however, be non-negligible.
correct one for a single, fully occupied orbital. )
Then we find directly from the definitions of U and &., -O. i6 a'rn~', r . (36)
E, that
"
.
U[n. ]+ Z„[n. .. O] = O. (30)
Since (34) is usually close to an equality,
(36). We see that, in the absence of correlation,
so is

Defining the correlation energy as E, = E„—E„, we 6, is positive and the self-interaction correction
find from Eqs. (29) and (30) that to the LSD total energy [Eq. (32)] is negative.
Furthermore, the correction becomes more sig-
z, [n, ]= o. nificant as the localization of the orbital increases:
Equations (29), (30), and (31) form the basis for For a Is hydrogenic orbital, n,
(r) = Z'e 'x"/7[ and
our self -interaction corrections. Eq. (36} becomes 6,
~ 0.047Z; as Z increases,
Suppose now that we have some approximation correlation becomes less important relative to
E'„",~ "[n&, n&] (such as LSD, the gradient expansion, exchange and the inequality gets closer to a rel-
pseudopotential LSD screening, etc. ) which does ative equality. On the other hand, for an electron
not satisfy Eq. (29) but is useful otherwise. We delocalized over a volume 0, is of order 0-'
n,
define our self-interaction corrected (SIC) version so 5, is of order 0-'', which vanishes as 0-.
of this approximation by (In a metal the sum over orbitals contributes
"
Z„'. =Z„",'""[n„n, ]- g 6... (32)
another factor of 0, so the self-interaction cor-
rection to the total energy is of order 0'~'. ) It
follows that, for densities varying slowly enough
where over space, the LSD and self-interaction-cor-
rected LSD (SIC-LSD) approximations both yield
6, = U[n, ]+E„[n „0] (33) the exact energy per electron.
In order to have a self-consistent self-interac-
is the self-interaction of orbital o,'o. Unlike some
other schemes which have been previously pro- tion correction, we now seek the orbitals T])„, and
posed (Sec. IV), this correction may be applied to occupation numbers f,
which minimize E of Eq.
(16}, with E„, replaced by E„, . (Here we use a
any spin-density functional for E„„when applied
to the exact one the correction vanishes. bar to denote quantities which are self-consistent
In order to get some feeling for the behavior and with a self-interaction-corrected potential; else-
where our notation will not be so explicit. ) Follow-
magnitude of the self-interaction correction, we
pause to estimate it within LSD. Gadre et gl. " ing the procedure of Eq. (16), we obtain a one-
electron Schrodinger equation like Eq. (17},
have proved the general inequality [for n(r) ~ Oand
&= fd rn(r)]:
[ ', .
&'+v.f'f(r}]-V-. (r)= ~.". 't. .(r},
rr[3] 3 093dd f-d' . d'( )'. (34) but with an orbital-dependent potential:

"
v, ff(r) =(v(r) —2paB(r)+n([n]; r)+ v„', ""([n[dn[];&)] —(v([n, ]; r)+ v[,"~ ""([n „0];r)]. (38)
The self-interaction correction to the potential [the second curly brace in Eq. (38}]appears as a. natural
consequence of the correction to the energy. For a one-electron system, Eq. (38) correctly reduces to
the external potential v(r) —2i], oz(r). Equations (32) and (37) form the basis of the SIC approach.
When applied to an electron gas of uniform density n, Eq. (37) admits plane-wave orbitals as self-con
sistent solutions, for which the SIC-LSD total energy per electron is exact. Conceivably, however, there
might be another set of self-consistent solutions to Eq. (37), with localized orbital densities adding up to
a uniform total density, for which the SIC-LSD energy per electron lies helot the exact value. Since the
orbitals cannot be localized to a region of radius less than r, = (3/4m'n}'~', any resulting contamination of
the total energy from self-interaction correction should be minor if it occurs at all.
The energy eigenvalue &s[c of the SIC single-particle Eq. (37) is
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. ..

q', o = ( q, .
~
—-', v'+ v(r) —2 poB(r) + u([n]; r) + v'„,'""'"([ni, n}];r) ~ goo&
(39)
—(P„~u([n, ]; r)+v„,""'"([n „0],r) ~g„),
while the uncorrected energy eigenvalue is
"
q'}'('""= (g„~ —-', V'+ v(r) —2&oB(r)+u([n]; r)+v'„, "x([ni, n&]; r) ~g„) . (40)
I
The first terms in Eqs. (39) and (40) are identi- actly via Eq. (7), leaving only correlation to be ap-
cal except that in Eq. (39) the SIC orbitals T}}„(r) proximated by
are used instead of the uncorrected orbitals g, (r).
Hence, SIC has not only a direct effect on the en-
ergy eigenvalue [the last term in Eq. (39)], but
@sIc QxPPxox[n n ] g QaPPxox[n
0] (42)

also an indirect effect (change in energy due to However, it is not clear that such an unsymmetri-
modification of the orbitals). Note that the direct cal treatment of exchange and correlation is de-
change is made of two contributions having oppo- sirable, since it would reproduce the well known
site signs: The self-Coulomb correction is nega- pathologies of the Hartree-Fock eigenvalues for
tive while the self-exchange-correlation correc- valence electrons in metals. '
tion is positive. Much of the success of the LSD
approximation applied to localized configurations C. Exchange-correlation hole
is due to the fact that in many systems the indir-
ect SIC effect is small and the two contributions Gunnarsson and Lundqvist" have argued that the
LSD approximation (27) gives reasonable results
to the direct term show large cancellation. The
SIC eigenvalues q»~ still have the significance of
for realistic systems that are formally outside
its domain of validity because it exactly satisfies
Eq. (22), and the aufbau principle still applies,
the sum rule Eq. (26) on the number content of the
as can be seen by repeating the arguments of Sec.
exchange-correlation hole. The density-fluctua-
II A.
tion wave-vector analysis of Langreth and Per-
A simple approximate relationship exists be-
dew ' supports this conclusion, and also shows
tween the orbital energies of the LSD and SIC-LSD
approximations for exchange if one neglects the
(at the risk of oversimplification) that the gradient
correction does not give the expected improvement
difference between the orbitals 7(}, and (i.e. , g, over LSD because it fails to satisfy this sum rule.
self-consistency effects). We find from Eqs. (39)
and (40) that
The LSD exchange-correlation hole" is easily
deduced from Eqs. (24) and (27):
—(2V(n„}—0.9305(-}Jl d'inc(r}) p~sn(r, r')= px(n& (r), n&(r); ~r' —r ~), (43)
—0.94 d'rn ', (r) . (41) where p}}(n&,n&', ft) is the hole in a homogeneous
4
electron gas with spin densities n~ and n&-. Note
Comparison of Eq. (41) with Eq. (36) shows how that the LSD hole around the electron at r is spher-
the self-interaction error of the LSD total energy ically symmetric, i.e. , Eq. (43 depends on the ob-
gets magnified in the LSD orbital energy. servation point r' only through r' —r~ This ~

This SIC formalism can be regarded as a densi- spherical symmetry is incorrect, but introduces
ty-functional approximation corrected for self-in- no. error into E„„since the Coulomb interaction
teraction, or equivalently as the Hartree approxi- 1/ ~r' —r samples only the spherical average of
~

mation corrected by an approximation for exchange the hole4'.

and correlation between different orbitals. Like
the Hartree approximation, it introduces an orbit- E„;——— d xn r dR 4pR p r, r+
al-dependent effective potential [Eq. (38)) which
makes the orbitals slightly nonorthogonal. If de- Since p„satisfies the sum rule (26), so obviously
sired, the optimum set of orthogonal orbitals does p~».
could of course be found by introducing off-diagon- The hole in SIC-LSD can also be found by in-
al Lagrange multipliers into Eq. (37), or by other spection of Eqs. (24), (27), and (32):
means. 44 As discussed in Sec. IIIB, we have found ps'c(r, r') = p}}(nt(r),n&(r); ~r' —r ~)
for atoms that orthogonalization of the orbitals ob-
tained from the one-electron Schrodinger equation
has only negligible effect on the total energy.
—Q .
™" [n, (r')+p„(n, (r), 0;tP'-r ~)],
Clearly, in an orbital-functional scheme such (45)
as the present one, exchange could be treated ex- from which we find
 J. P. PERDE% AND ALEX ZUNGER

d'y'p~&(r r'} = — Qf n (r) n(r) . (46)

The right-hand side of Eq. (46) reduces to if -I

the occupation numbers f,
are restricted to the
values 0 or 1, which will be the case under the +
I 00

$a
conditions for which the sum rule (26) was de- $a
50 ~0
rived; otherwise, the right-hand side of Eq. (46) ~0

lies between 0 and -1.

Note that p s'c(r, r ) is no Q
I
0

spherically symmetric about the point r, even

when the orbital densities n„,(r) have been spheri-
..- "f~
I

-0.2 -O. I O. l
calized; it depends on n,
(r') as well as on ~r' —r ~. R (a.u. )
%'e will show below that the SIC-LSD exchange I
'
I
Ijl/$
hole agrees much better with the exact one then
the LSD hole does. (
r=0. 39au. SIC
The exchange-correlation hole is not known very [
a. ccurately, even for the uniform electron gas. t
However, the exact exchange hole may be found
from Eq. (7):

p"(r, r')=- P gf, .(r)4 .(r') ' n(r).

g*. (47)
~~
0

Note that for orthogonal orbitals we find 0

I
(a.u. ) R
d'r'p"(r, r')=- Qf„n, (r) n(r), (48) FIG. 1. Exchange hole about an electron located at
distance y from the nucleus in the neon atom. The full'
the same as the result (46) for the SIC-LSD curves are exact, while the dashed and dotted curves
scheme. Note further that represent the SIC-LSD and LSD approximations,
respectively. Parts (a) and (b) are for two different
w I
p"(r, r) = -[ni2(F) + n', (r)]/n(r), (48) values of r, Note that the observation point r =r+8
lies on a line with the electron and the nucleus.
an exact result which is also satisifed by the LSD
and SIC-LSD approximations for exchange. From
that SIC-LSD gives a much more realistic de-
Eq. (47) the exact exchange hoie is strictly non- scription of the shape of the ho1.e, which peaks up
positive, a condition which is also satisfied by
strongly when the observation point r' is near the
LSD but not by SIC-LSD. (The SIC-LSD hole is
nucleus, than does LSD. Figure 2 shows that SIC
positive only at large R, as shown in Fig. 1, so
also improves —R(p(r, r+ R))~q [the spherical av-
this error has little effect on the total energy. )
erage which by Eq. (44) yields the exchange ener-
In order to evajuate the exchange hole in LSD
and SIC-LSD, we need to know the exchange hole gy] when the electron at r is close to the nucleus
.

of the homogeneous electron gas. Using plane

[Fig. 2(a)], but not necessarily when the electron
is further out [Fig. 2(b)]. The improved descrip-
waves as orbitals in Eq. (47), we find
tion of the exchange-correlation hole in the SIC
p„"(n„n,;It ) = ', I,
[(1+ C)n, F(—'-ft) scheme will show up as a substantial improvement
in the atomic exchange, correlation, and total en-
+ (1 —g)n, r(a,'ft)], (60)
ergies, presented in Sec. III.
where the relative spin polarization is t = (n&-n, )/
n, the Fermi wave vector is k~ = (6v'n, and }', D. Eigenvaloes and removal energies
s&nx —x cosx Much of the utility of the Hartree-I ock theory
I"~xp= 9 0
X3 lies in Koopmans's theorem, which states that the
change in HF total energy due to removal of an
The SIC scheme describes exchange exactly in
electron from an unrelaxed orbital o.o is simply
the He atom, since two electrons of opposite spin
related to the HF eigenvalue:
experience only self-exchange. Our results for
nEHF HF
the exchange hole in the Ne atom are shown in +ff i uflre1 (52)
Figs. 1 and 2. (Similar figures for the exact and
LSD hole have been given by Gunnarsson, Jonson, Relaxation effects upon removal energies can be
and Lundqvist, ")
The orbitals employed were the calculated in the HF theory by performing two
analytic HF orbitals of Bagus. 46 Figure I shows independent self-consistent-field (SCF) calculations
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. ..
4 to localized states (deep defects in insulators
O and semiconductors, surface states, core excita-
4 tion energies in solids and molecules, etc, ), re-
laxation effects Z, are significant. We will dem-
tlL onstrate, however, that for atoms the negative of
2 the SIC-LSD energy eigenvalue, -q»c, approx-
imately equals the relaxed excitation energy, due
Q to a remarkable near cancellation between the
negative relaxation energy Z, and the positive
~ og ~ ~ ~ ~y
non-Koopmans correction II,
in SIC. Such an
effective cancellation does not appear in the LSD
O. I 0.2 0.3 0.4 approach. This cancellation allows the formula-
R '(a. u. )
tion of a physically correct one-electron theory
0.4— / based on the calculation of g~'c for the ground state.
r=O. 59au. g
Although many band calculations (pseudopoten-
0.3 tial or all-electron} now rely on one version of
density-functional theory or another, density-
+ functional theorists have always viewed the eigen-
0.2 values with suspicion. They appear first as purely
formal Lagrange multipliers, unredeemed by any
Q
O. l
Koopmans's theorem. The relationship between
removal energy and eigenvalue can be found, how-
I ever, by integrating Eq. (21):
0 0.5
R (a.u. )
I.O j.
df „e,(f, )~ (54)
FIG. 2. Spherical average of the neon exchange hole
.
Here it is the relaxed energy difference or physi-
times g for (a) y =0.09 a.u. and (b) r =0.39 a.u. The
full, dashed, and dotted curves are the exact, SIC-LSD
cal removal energy that is obtained, but the eigen-
value must be calculated, using the relaxed orbit-
and LSD results, respectively. In part (a), the SIC-LSD
curve is almost indistinguishable from the exact one. als, for occupation numbers f,
between 0 and 1.
Clearly, the use of Eq. (54) would undermine the
for the inital and final states ("&SCF"method); simplicity of band theory, which seeks a descrip-
the resulting energy difference is used to define tion of excitations in terms of differences between
the orbital relaxation energy Zfg,": fixed energy levels. Equation (54} was first dis-
covered within XQ. theory by Slater, ' who pro-
d, EsF -=ZE"" (53}
~ ~

posed approximating it via the "transition state "~'4'

The relaxation energy Z"", is negative, since re- of half-occupancy:
laxation lowers the energy of the final state rel-
ative to a non-self-consistent estimate for the ion-
ized system.
.
~E., ~, = g. (,'). (55)

The advantage of having Koopmans's theorem is, Obviously, evaluation of Eq. (54) or even Eq. (55)
of course, that fairly accurate removal energies can be difficult, since the "transition state" may
can be found from a single calculation of the lack the symmetry of the initial state.
eigenvalues for the N-electron system, without Contrary to popular prejudice, Eq. (54) is also
also calculating for the N-1 electron system as true in HF theory once we admit fractional oc-
in a 4SCF approach. (If the relaxation of the or- cupation numbers [with the exchange energy de-
bitals is small, of order g, the relaxation energy fined by Eq. (7)]. An approximation to the integral
is only of order q'. ) of Eq. (54), which avoids the need to calculate for
Koopmans's theorem has been a cornerstone of fractional occupation numbers, is
energy-band theory for bulk solids, surfaces, de-
fects, and interfaces, in which much of the inter-
".(1)+a",(0)].
a. E", ~„,= --,'[e, (56)
est is directed toward the energy eigenvalues. We
will see below that Koopmans's theorem does not Hedin and Johansson, "
Equation (56) was derived in a very different way by
'
using the HF polarization
hold in the LSD approximation which is frequently potential" .
used in electronic-structure calculations. The In order to have an analog of Koopmans's theo-
non-Koopmans correction to it, denoted as II
is positive. Furthermore, in many applications
„ rem, one would need
occupancy) in place of
c,
(1) (the eigenvalue at full
e,
(~) in Eq. (55). However,
 5056 J. P. PERDEW AND ALEX ZUNGER

in the X& and LSD approximations, depends c, (i) For a one-electron system, is d'rnite'.
clearly II,
strongly on f,
and no analog of Koopmans's theo-
rem is possible. This result is an artifact of the
zero as it must be, while II~, = 0.78
(ii) When the orbital density n, contributes little
f
self-interaction error of the local approximation, to the total spin density may be expandedn„S,
and not a general feature of density-functional as
theory. Even in the hydrogen atom, the LSD
eigenvalue varies dramatically as f„
is reduced s„,=-; e', , „(. o)("-)*+o(".-'),
,
from 1 to 0, while the eigenvalue would be inde-
pendent of f„
for the exact density-functional sat-
isfying Eq. (29). In a self-interaction-free calcu-
lation for a system with more than one electron, or simply neglected. In this case Koopmans's
the eigenvalue a,
will depend only indi~ectly on theorem does not hold for localized orbitals in
either LSD or SIC-LSD, since the "non-Koopmans"
f~„ through the reflection back onto orbital no of
corrections are positive. Using Eq. (34) for esti-
its self-consistent effect on the other orbitals,
i.e. , through physical (rather than spurious) re- mates, we find that in general
laxation effects. Physical (interelectron) relaxa-
tion effects are easily described: %hen an elec- H~, = 1.09 d'snag+ S~, , (64)
tron is removed from an atomic orbital o'0, i.e. ,
when f„-0, the remaining orbitals are drawn
II~, = 0.3l d'x n~+g'+ S~, . (65)
closer to the nucleus, which becomes more effec-
tively screened. This makes c„(2)~ e„(1), i.e.,
As in HF theory, the physical removal energies
in a self-interaction-free theory one expects that
include a relaxation energy : Z,
(57) LSD LSD LSD
+Ee.e ~rel: ++ac ~umel+ ~ue
Since the screening effect from the physical re- LSD LSD LSD
laxation is often small, the inequality (57) is often Nty CLQ + Z NV (66)
close to an equality, yielding a powerful if approx- and similarly
imate analog of Koopmans's theorem.
Closed-form expressions for the change in un- &EN, ~,g
= —E~,
SIC SIC
+ II~~ SIC
+ Z~~ . (67)
relaxed total energy can be found by evaluating en-
ergy expressions such as Eq. (15) with fixed orbi-
While Z,
cannot be calculated analytically, its
scaling with simple energy terms in known:
tals for f„=1 or 0, and taking the difference.
Gopinathan" has found by numerical calculations
Neglecting correlation terms for simplicity (in-
that
clusion of correlation terms poses no difficulty),
we find that the LSD and SIC-LSD approximations
for exchange yield the simple results:
Z."'.= a, U [n. .], (68)
with k, = -0.
44 for atoms. (Better approximations
(58) with a shell-dependent k, are possible but will not
SIC SIC SIC
concern us here. ) Equation (68) reduces via Eq.
~

)unrel ~&fy ~any (59) (34) to

where the "non-Koopmans" corrections are
—0.48
"."=U[n. ,]+ S. .
11. . (60)
Z LsD ~~ d3y n (69)

By comparing Eq. (69) with Eqs. (64) and (65), we

11 ', = -e Jl d'rn, c„(n „0)+S,, (61) see that Z~, can typically cancel less than half of
Ii~n, so that the physical removal energy &E,
is poorly approximated by the orbital energy
In contrast, in SIC-LSD. there is an effec-
S,= d'x n, — n, e„n, —n „0 tive cancellation between Z, and II, leading to a

.
-n. e„(n., O)+ ~n e„(n„O)]. (62)
generalized Koopmans scheme in which -& ',
forms a good approximation to the relaxed ioniza-
All quantities in Eqs. (58)-(62) are to be evaluated tion energy. Ironically, in the HF theory, which
at full occupancy. The analogs of Eqs. (58) and is the only case where Koopmans's theorem holds,
no such generalized scheme can exist since there
"
(60) within Xn theory were derived by Gopina-
than, and less directly by Slater and Wood. ' is no "non-Koopmans" term to counteract relaxa-
Note that by Eq. (52), II", =
—0. tion effects. As a result, a 4SCF approach is of-
Two cases are simple enough to be instructive: ten unavoidable in HF.
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. .. 5057

From the relations (64)-(69) one can deduce the tells us that an electron which ventures out to
ordering of terms large r will not "see" its own charge, and so will
experience the potential -(Q+ 1)/r; this conclusion
(70)
is also supported by a detailed analysis of the be-
If Z~, is similar to Z~, , it will slightly overcancel havior of the exchange-correlation hole. "
The
II sic so SIC-LSD potential (38), unlike the LSD potential,
clearly has the correct long-range behavior.
(71)
This is necessary for the successful description
[in agreement with Eq. (57)]. These relationships of negative ions (Sec. IHE).
are illustrated graphically in Fig. 3. Numerical. After the discussion of Sec. IID, there can be
examples are given in Sec. III G. The striking re- little doubt that the eigenvalues of density-func-
sult is that the negative of the SIC orbital energy tional theory have physical significance. 'This is
forms a good approximation io the relaxed re- especially true of the eigenvalue c „of the least-
moval energy, in contrast to hotly LSD and HF. bound occupied orbital for an N-electron system
Finally, we observe that a simple relationship in its ground state, which by the arguments sur-
exists between the unrelaxed energy differences rounding Eq. (57) should be almost the negative of
in the LSD and SIC-LSD approximations for ex- the first ionization potential. (Here the initial and
change when the difference between the LSD and final states are ground states of the N and N-1
SIC-LSD orbitals is neglected. From Eq. (41) and electron systems, and the relaxation effects asso-
from Eqs. (58)-(61) we find that ciated with removal of this outermost electron
(V2)
should be small. ) a,
„controls the long-range be-
havior of the electron density. For example, in
where 5,
is the spurious self-interaction of Eqs. an ion of charge Q the one-electron potential tends
-
(33) and (36). In general, however, higher-order to -Z"/v as v ~ (where Z*=Q+ 1 in a self-inter-
terms (i.e. , the self-interaction effects on the action-free theory), and the solution of the one-
orbitals) should not be neglected. electron Schrodinger equation tends to
(73)
E. Long-range behavior of the potential and density
in an atom; excited states where P =Z~/n —1 and & = (-2a, „)~'. In fact if we
"
define o. = (2I)~', where I is the first ionization
The LSD one-electron potential for an ion of "
potential, we find from Eq. (57) that o. & o. and
charge Q has erroneous long-range behavior due hence from Eq. (73)
to self-interaction, as is evident from inspection
of Eqs. (18) and (28): Far from the nucleus the [n(f')]~'- kra'e (74)
LSD exchange-correlation potential (28) decays
exponentially along with the density, leaving only
-
where O' = Z* &/' I, in agreement with the rigor-
ous asymptotic upper bounds of Morrell, Parr,
the long-range potentials v(r) and u([n]; r) which
sum to Qlr Ho-wever. , classical electrostatics
and Levy" and 'Tal. "
Figure 4 illustrates the long-range behavior of
the SIC potential v,",, (x) [Eq. (38)]. We have plotted
the difference between the SIC-LSD potential and
I.3
the LSD potential, both multiplied by x, for the
- —bEI Cu atom in the 3d"4s' configuration. At large x,
lUhfel

I.2—
—
bE) this quantity correctly approaches -1. Note,
fu HEI '~ it a (unrel however, that the self-consistently calculated
E X
"bE,Irel
ce Iree SIC potentials differ from the LSD potential not
only in their asymptotic limit but also at finite
to set the scale we define the quantities R„and
R~ which denote the muffin-tin radius of metallic
0.9— Cu and the Pauling tetrahedral radius, respective-
I SIC I LSD ly, while (r)„, indicates orbital moments. Clearly
0.8
the SIC potentials deviate substantially from the
FIG. 3. Relationships between the orbital-energy LSD potential at distances from the origin which
eigenvalue e, the unrelaxed energy difference nE ( „~~~,
are chemically and physically important.
aud the relaxed energy difference nE upon removal
~'
„, We close this section with a comment on excited
of an electron, in the HF, SIC-LSD, and LSD approxi-
mations. In the example shown here, the electron is states. The Kohn-Sham equations (17) have solu-
removed from the 3s orbital in atomic Ar. The re- tions that look like states of particle-hole excita-
laxation energy g has been taken from HF calculations. tion. Although several density-functional theorems
 5058 J. P. PKRDK% AND ALEX ZUlVGKR

0.1- ergy as a functional of orbitals and occupation

RM RT numbers, the SIC energy will be a functional of the
p.p
density if the orbitals and occupation numbers
-0.1 are. The needed functional relationship is gener-
-0.2 ated by the following rule: Given any pair of trial
0$ spin densities n~(r) and n ~(r), consider all possi-
-0.3
ble occupation numbers
~ 1 and normalized orbitals
f, P,
in the range 0~ f„,
(r) satisfying Eq.
I (13). (The orbitals do not have to be orthogonal. )
-0.6. The correct f,and g, (r) corresponding to these
trial spin densities are the ones which minimize
-0.7
-0.8-
-0.9-
-1.0'
0 2 3 5
with n, given by Eq. (5). Since the one-electron
1
Schrodinger equation with the orbital-dependent
Distance (a. u;) potential (38) results from minimization of the en-
FIG. 4. Orbital-dependent self-interaction correction
to the LSD one-electron potential in atomic Cu. The ergy over all orbitals, the Hohenberg-Kohn vari-
cusps arise from nonanalyticity of n 3 (r) at nodes of ational principle (11) is valid in the SIC formal-
orbital o (T. All curves tend to —1 at large y. ism of Sec. II B. (For further discussion, see
Appendix A. )
have been proved about excited states, ' '"~' This self-consistent SIC formalism does not fit
there is usually little a Priori justification to in- into the Kohn-Sham scheme because, unlike the
Kohn-Sham scheme, it does not define a "kinetic-
terpret the Kohn-Sham solutions as descriptions
of physical excitations. Nevertheless, only a energy" functional of the density which is indepen-
purist would discard these Kohn-Sham solutions. dent of the choice of E„'»"". As a result two unus-
While each excitation ought to require a separate ual and possibly unlikable features emerge:
self-consistent calculation, in practice one can of- (1) The minimizing orbitals that solve the one-
ten just transfer an electron from an orbital oc- electron Schr'odinger equation are slightly but an-
cupied in the ground state to a "virtual orbital" noyingly nonorthogonal. (3) These orbitals belong
(one of the unoccupied eigenfunctions of the to an orbital-dependent potential which can break
ground-state self-consistent one-electron Hamil- the symmetry of the system. In a crystal, for ex-
tonian). In view of our discussion of long-range ample, the core orbitals which minimize the SIC-
behavior, the potential used to calculate the vir- LSD total energy are localized, so that the one-
tual orbital ought to be self-interaction free. If electron potential may not be periodic even when
we have an orbital-dependent potential as in Sec. the spin densities are, and Bloch's theorem (with
IIB, the natural way to achieve this is as follows: all its computational utility) is lost. (Note, how-
For excitation of an electron from state Q, to ever, that although individual orbitals may not
state c.'&, compute the virtual orbital (r) using
the potential v, ff(r). The HF analog of this pre-
g, obey this theorem, the total wave function can
easily be constructed to have a Bloch periodicity. )
scription was suggested by Hunt and Goddard. " While in principle as well as in practice one may,
in a perfectly legitimate way, solve problems
such as core excitations in molecules or solids
F. Relationship of SIC to the Hohenberg-Kohn
using the SIC scheme with nonorthogonal orbitals
and Kohn-Sham theorems
and broken symmetries, it may also be useful to
In this section we will show that the self-consis- seek other methods.
tent SIC formalism of Sec. II 8 is protected by the These unlikable features could be avoided by
Hohenberg-Kohn theorem of Eq. (11) and so has incorporating self-interaction corrections within
the same theoretical sanction as other self-consis- the Kohn-Sham scheme, in which the kinetic-en-
tent density-functional schemes. We will also ergy density functional has the independent defini-
show that this formalism does not fit into the tion of Eq. (14). The simplest scheme which does
Kohn-Sham scheme of Eq. (17) (e.g. , it has an or- this is the following ':
(1) To each pair of trial
bital-dependent potential), although it can be made spin densities n, (r) there corresponds a pair of
to fit by minor alterations. (These conclusions model potentials v', «(r) which generate these spin
apply with equal force to the self-consistent Har- densities via Eqs. (13) and (17). [v,'«(r) need not
tree formalism. ) be the self-consistent Kohn-Sham potentials but
Since the SIC formalism specifies the total en- may be some assumed model potentials. ] Solve
 SELF-INTERACTION CO'RREGTION TO DENSITY-FUNCTIONAL. .. 5059

Eq. (17) using these model potentials to find the Fock and Kohn-Sham total energies are invariant
orbitals tt„,(r), which are now functionals of the under such a transformation. Since the orbital
spin density. For small systems such as atoms densities n,
(r) change under the transformation,
and ions, the Kohn-Sham orbital densities n„(r) the SIC total energies are not invariant.
can be substituted directly into the SIC energy ex- To get a size-consistent self-interaction-cor-
pression. (2) This expression can then be mini- rected Kohn-Sham theory, insert between steps
mized by varying the spin densities; in practice, (1) and (2) above another step: (1') Perform a
however, one varies the model potentials v', «(r). unitary transformation of the Kohn-Sham fully
If the model potentials have sufficient variational occupied (molecular or Bloch) orbitals to a second
flexibility, they will approach a self-consistent set of (more localized) orbitals which will gener-
solution. 'This method of potential variation was ate the orbital densities to be used in step (2).

tions by Rose and Shore. "

suggested as a practical scheme for LSD calcula-
[The minimization dis-
The right unitary transformation might be the one
which minimizes the SIC energy expression.
cussed here is constrained to the "Kohn-Sham" Since the spin densities and the kinetic energy re-
space of orbitals belonging to orbital-independent main invariant, so does each piece of the energy
potentials, so the resulting self-interaction-cor- (15) excePt the self-interaction correction, so
rected total energy will be greater than the value this choice amounts to finding the equivalent orbi-
found by the method of Sec. II 8, Eq. (37), which tals that carry maximal self-interaction error.
does not impose this constraint. ] Within SIC-LSD, the transformed orbitals for core
'The scheme just defined is a Kohn-Sham for- electrons in solids will be localized, correspond-
malism which is self-interaction free in the sense ing roughly to the localized hole that is left behind
that it satisfies Eq. (29). Unfortunately, the self- in x-ray ionization, while loosely bound conduc-
interaction error reappears in the form of a size- tion electrons may have delocalized transformed
consistency problem: The theory will be useful orbitals.
for an isolated atom, but not for a lattice of atoms Finally, although the exact realization of this
in the limit of infinite lattice constant, since in the scheme would be tedious, it is probably unneces-
latter case the Kohn-Sham orbitals will be atomic- sary. LSD already gives reasonable spin densities
like Bloch functions and the correction per atom and one-electron potentials, and it is not hard to
will be zero. As mentioned in Sec. II 8, there will imagine an approximate way to incorporate self-
in this latter case be a correction to the total en- interaction corrections within a self -consistent
ergy proportional to the volume of the system to orbital-independent potential, e.g. , by using a
the & power, which looks at first like a contribu- weighted average of the SIC orbital-dependent po-
tion to the surface energy, but is not; it is simply tentials of Eq. (38) [the orbitally-averaged-poten-
a shadow of the size-consistency problem. This tial (GAP) method]:
example, and another discussed in Appendix B,
show how subtle the self-interaction problem can v', P, " (r) =+ n~„(r)v„', (r) Q n~, (r),
be: Satisfaction of Eq. (29) is a necessary condi-
tion for solving it but not a sufficient one. where p is some positive constant. Furthermore,
The reason for the size-consistency problem is the variational principle suggests that accurate
that the self-interaction correction depends on the energies could be obtained without optimizing the
choice of orbital representation. For example, unitary transformation of the orbitals: It should
the core electrons in a solid could be represented' often suffice to compare the Kohn-Sham orbitals
either by localized Wannier orbitals which carry with any unitarily equivalent but more localized
self-interaction, or by delocalized Bloch orbitals orbitals. Such localizing transformations are al-
which do not. Stated differently, the self-inter- ready familiar in the Edmiston-Ruedenberg"
action 6~, depends on the number of atoms which method for molecules and the Wannier transfor-
share the orbital P«, as illustrated by the SIC mation for filled bands in solids. '
band-structure calculation for solid Ar presented
in Sec. IIH. III. NUMERICAL RESULTS
Thus, the key to resolution of the size-consis- A. Ingredients of the calculation
tency problem lies in unitary transformation of
the fully occupied orbitals. Such a transformation We have performed numerical calculations using
leaves invariant a Slater determinant and any one- the SIC-LSD formalism of Sec. IIB. The para-
electron property that can be evaluated from a metrized electron-gas correlation energies E,
Slater determinant, such as the spin density [Eq. (n~, n&) are described in Appendix C. They are
(13)], the kinetic energy [Eg. (14)], or the ex- based on Ceperley's" accurate Monte Carlo cal-
change energy [Eq. (7)]. As a result, the Hartree- culations for low and metallic densities, which
 5060 J. P. PERDKW AND ALEX ZUNGER

we have matched smoothly to the correct high- TABLE II. Orbital eigenvaluese„&, total energy E,
density limit. As shown in Appendix C, other and separate exchange (E„) and correlation (E )
parametrized correlation energies commonly used energies for the Kr atom, calculated self-consistently
calculations are in error at in three schemes. Energies in eV. (Throughout the
in density-functional
paper the calculations were done in atomic units. 'We
the high densities which are important in atoms. then used the conversion 1 a.u. = 27.21 eV. )
For atoms and ions, we solve Eg. (37) with the
state-dependent potential of Eq. (38) using a direct SIC-LSD SIC-LSD
predictor-corrector numerical integration. At LSD (not orthogonalized) (orthogonalized)
each iteration we have trial spin densities (n &(r),
n &(r)] as well as trial spin-orbital densities -13 876.6 -14127.4 -14129.5
in both sets -1803.6 -1851.9 —1852.5
{n„„(r)].We achieve self-consistency 2s
-1695.0
to a relative accuracy of 10 ' or a relative accura- e2p -1633.1 -1695.8
—253.46 -269.46 —269.54
cy of 10 ' in the energy (whichever comes first). -192.81 -209.02 -209.11
With the self-consistent orbitals we calculate the -83.63 -101.27 -101.34
total energy and its individual components [e.g. , E'4~ -22.30 -27.76 -27.76
Eqs. (32) and (33)] as well as various moments of C4p -9.40 -13.95 -13.95
The same calculation is repeated in each case
for the LSD approximation. E -74 830.57 -75 015.47 -75 015.47
Following the custom of LSD calculations, we
E„-2409.06 -2632.90 -2631.75
E -88.40 -53.26 -53.25
have sphericalized the orbital densities n,
(r) be-
fore evaluating the potential and total energy for
atoms. This "central-field" approximation is per-
haps most severe for the 2p state, which makes a are shown diagramatically in Fig. 5 for the 3d
sizable contribution to the total self-interaction elements in the d" 's' configuration. White areas
correction. However, for a hydrogenic 2p state indicate the self-direct contribution, dashed areas
(with m, = 0) the self-direct Coulomb energy indicate the self-exchange part, while dotted

sphericalized orbital density. "

U[n, ] is only 4% larger than it would be for the areas denote the self-correlation part. Asterisks
mark the SIC eigenvalues c„,', while open circles
The eigenfunctions of the orbital-dependent one- denote the LSD eigenvalues &„"„. The self-inter-
electron Hamiltonian are very slightly nonortho- action terms are clearly of similar magnitude to
gonal. Except where otherwise noted, we have the energy eigenvalues themselves. The self-cor-
performed a Schmidt orthogonalization after each relation is usually a small correction. For the
iteration: The 2s orbital is orthogonalized to the more localized 3d orbitals, SIC effects are signi-
1s then 3s is orthogonalized to 1s and 2s, etc. ficantly larger than for the more diffuse 4s orbi-
tals. 'The SIC contributions to the orbital ener-
gies follow the same trend as the energy eigen-
B. Size of self-interaction terms
values in going from the left- to the right-hand
'Table II compares the orbital eigenvalues, total side of the row in the Periodic 'Table. The large
energy and separate exchange and correlation en- (but incomplete) cancellation between the self-
ergies for the Kr atom calculated self-consistently direct and self-exchange-correlation terms is ev-
in three schemes: LSD, SIC-LSD without ortho- ident.
gonalization, and SIC -LSD with orthogonalization.
The separate orbital contributions to the self-
interaction correction -6„, for Kr are shown in TABLE III. Orbital self-interaction corrections —6„&
Table III. The sum -Z, 2(21+ 1)5„,= -185.1 eV is Eqs. (32) and (33)J to the total energy in the Kr
t'see
just the difference between the LSD and SIC total atom. (SIC-LSD orbitals. ) Energies in eV.
energies of 'Table II, showing that the difference
between LSD and SIC-LSD orbitals has little ef-
fect on the total energy. Note that even in an atom
as large as Kr, 44% of the self-interaction correc- lg -40.86
2g -5.05
tion to the total energy is carried by the 1s orbi- 2p -9.80
tals. It is clear from Table II that SIC has a pro- 3s -1.12
found effect on all the orbital energies, core as 3p 1 33
well as valence, and that orthogonalizing the 3d -2.31
SIC-LSD orbitals is a small effect. 4g -0.18
'The various self-interaction contributions to —0.17
the spin-up valence orbital energies c„I, [Eq. (39)]
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. ..

TABLE 1V. Hesults of exchange-only calculations for

p P /y/F//F/'/// P//// ////

//
//////
the Ne atom. Here the SIC-LSD orbitals were not or-
thogonallzed, Energies in eV and distances in a.u.

HF
LSD Pef. 59)
0 ~~s -822. 7 -888.4 -891.7
o 0
o o o
~2s -34.4 -44.2 -52.5
6'2p -12.1 -22. 0 -23.1
-3469.0 -3506.2 -3497.8
7 3469.0 3506.2 3497.8

+ -10 ca sc Ti V Cr Mn Fe Co Ni Cu 0.159 0.158 0.158

~ 15- 0.906 0.908 0, 892
&~&2p 0.990 0.965 0.965
C"-
LLI
10-
p P // /// // ///
// ///// &&'&2m
0.034
1.005
1.326
0.034
1.007
1.238
0.033
0.967
1.228
/ &~'&2p

o
0
0 0
0 tion effects were not included in this calculation.
o
The results are compared with the HF results of
Ref. 59. Note that, for the d" 's' configuration,
LSD predicts that the Sd energies are above the
4s energies for the elements Sc to Mn, in con-
trast to the HF results. The SIC scheme, on the
other band, shows'a reasonable agreement with the
HF results. Self-interaction corrections may have
profound effects on the hand structures of the cor-
PIG. 5. Breakdown of the orbital energiee in the SIC- responding elemental metals and the energy levels
LSD formalism for spin-up 4s and 3d orbitale of transi-
tion atoms in the d" ~ sy configuration. %hite area:
self-Coulomb; dashed: self-exchange; dotted: self-
correlation. Open circles: LSD eigenvalues; asterisks:
SIC eigenvalues. d-e one-electron energy gape
Ispinupj
a HF results fd" s j
C. Exchange-only calculations compared to Hartree-Pock o HF results ld" s

Exchange-only calculations are performed by LSD

setting the electron-gas correlation energy z,
(ni, n~) to zero. Table IV shows a comparison of
exchange-only orbital eigenvalues, total and kine-
tic energies, and orbital expectation values (r)
"
and (r') for the Ne atom in LSD, SIC-LSD, and
Hartree Fock. All three approximations satisfy
slc-
LSD
the virial theorem E= -T. The LSD orbital val-
ues of (r) (and also (x '), not shown) are already
remarkably close to Hax tree-Pock values, so -10
self-interaction correction provides little im-
provement in them. The same is true of (r') ex-
cept for the loosely bound 2p orbital, which is too
diffuse in I SD. On the other hand, the total energy -'55; I I I I I I
and even more so the orbital eigenvalues have Sc Ti V Cr Mn Fe Co Ni Cu
been significantly improved by SIC. PIG. 6. Difference between d and s orbital energiee
Figure 6 shows the splitting inenel gybetween (spin-up) in LSD, SIC-LSD, and spin-unpolarized HP
the spin-up M and 4s eigenvalues in the SIC and (Ref. 59) calcu1ations for the 3d elements in the cP st
LSD approximations for the Sd elements. Correla- and cP 2 s~ configurations.
 5062 J. P. PERDEW AND ALEX ZUNGER

of transition-element impurities in solids. " TABLE V. Exchange energies of atoms from ex-
Table V compares the exchange energies of change-only calculations. Here the SIC-LSD orbitals
atoms calculated in these three approximations. were not orthogonalized. Energies in eV.
LSD underestimates the magnitude of the exchange EHF
energy by 10-15%, while SIC-LSD gives values in Atom E xSK'. -LSD (Ref. 2)
better agreement with HF. (Note that SIC-LSD for
exchange only and HF are the same approximation H -6.9 -8.5 -8.5
in the case of atomic H and He. ) He -23.2 -27.9 -27.9
It was once hoped that LSD exchange energies Ne -297.6 -337.8 -329.5
would be improved the first correc-
by including Ar -755.8 -842.4 -821.3
tions from the density-gradient expansion (e.g. , Kr -2407.5 -2632.0 -2561.9
the X&P method, Ref. 61). For exchange alone
(as opposed to exchange and correlation), the
a priori gradient correction is negative and so values are higher still, with an error about twice
improves the LSD exchange energy, but the im- that of HF. The SIC-LSD values, on the other
provement is slight for all but the lightest atoms; hand, are lower than experiment (because the
the coefficient of the gradient term must be arti- magnitude of the exchange energy is slightly over-
ficially magnified by a factor of 5 to get agree-
ment with Hartree-Fock energies. "
It might be
estimated, as shown in Table V), with an error
magnitude far less than that of HF for the lighter
interesting to reexamine the convergence of the atoms but comparable to HF for Ne and Ar.
gradient expansion for exchange once each term The conventional correlation energy E,'"" is de-
has been self-interaction corrected. fined" as the difference between the experimental
total energy and the spin-restricted HF value.
D. Total and correlation energies of atoms compared The exact spin-density-functional theory defines
with experiment the correlation energy E, as E —T —J vn —U —E„,
The "experimental" total energy of an N-elec- with E„given by Eq. (7). The difference between
tron atom is the sum of the ionization potentials these two definitions arises only from the differ-
ence between the spin-restricted HF and Kohn-
(76) Sham orbitals, so for spin-unpolarized atoms the
variational principle implies the close inequality
corrected for relativistic and reduced-mass ef- Econv~ E (77)
fects. In Table VI we compare experimental val- C C

ues determined this way by Veillard and Clemen- Table VII compares the LSD and SIC-LSD cor-
ti" with LSD and SIC-LSD values, which now and relation energies [the latter evaluated from Eq.
hereafter include correlation. The HF values" (42)] with E "". Although the inequality (77) is sa-
are of course higher than experiment, and the LSD tisfied, it seems clear that essentially all of the

TABLE VI. Calculated ground-state total energies of atoms compared with experiment
(eV). The uncertain Lamb-shift corrections (Ref. 62) were not included in the experimental
numbers.

Spin-restricted
gHF Expt.
Atom (Ref. 59) @SIC LSD (Bef. 62)

H -13.6 -13.0 -13.6 -13.6

He -77.9 -77.1 -79.4 -79.0
Li -202.2 -199.8 -204.2 -203.5
Be -396.5 —393.0 -399.8 -399.1
B -667.4 -662. 5 -672.0 -670.8
N -1480.2 -1472.7 -1488.9 -1485.3
p -2704.9 -2696.6 -2720.7 -2713.5
Ne -3497.8 -3488.9 -3517.6 -3508.1
Na -4404. 2 -4392.6 -4426. 1 -4414.7
Mg -5431.5 -5418.3 -5456.4 -5443.2
Al -6581.5 -6566.0 -6608.8 —6594.0
P -9271.0 -9251.1 -9303.7 -9285.1
Ar -14 334.7 -14 310.5 -14 378.3 -14 354.6
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. .. 5063

TABLE VII. Calculated correlation energies of atoms cept for such fine details, the LSD and SIC-LSD
compared with the "conventional" experimental values correlation energies for all atoms and ions may
Ec~" (eV). The uncertain Lamb-shift corrections {Ref. be described by a simple analytic model. '4
62) were not included in the experimental values.
We have established that the SIC-LSD functionals
Econv for the total and correlation energies of neutral
C
EaC Tc"LSD (Ref. 62) atoms are rather accurate. Even greater accur-
acy might be found if the orbital densities were
H -0.6 -0.0 -0.0 not sphericalized. However, an alternate ap-
He -3.0 -1.5 -1.1 proach promises to establish a new standard of
Be -6.0 -3.1 -2.6 theoretical accuracy: Levy, Clement, and Tal"
Ne -19.9 -11.4 -10.4 have shown how to get the total energy of an atom
Mg -23.9 -13.6 -11.6 of atomic number Z from a knowledge of the first
Ar' -38.4 —22. 3 -19.9
ionization potentials, electron affinities, and (r )
for neutral atoms with atomic number ~Z. Using
difference between E, and E,"""arises from errors experimental ionization potentials and affinities,
and HF values of (r '), they obtained remarkably
in our approximations. LSD overestimates the
magnitude of the correlation energy by 100-200% accurate total and correlation energies. All the
while SIC-LSD overestimates it by only -20%%uq. inputs needed for this calculation could be com-
Tong" ascribed the LSD overestimate of the puted to excellent accuracy in the SIC-LSD form-
magnitude of the atomic correlation energy to the alism.
difference in polarizability between systems with
discrete energy levels like atoms and those with
continuous energy levels like the electron gas.
E. Binding energies of negative ions
Since the self-interaction correction arises only The negative ions are a sensitive test of any
for localized orbitals, and localized orbi. tais imply many-electron theory, since a large fraction of
discrete energy levels, there is perhaps no con- their binding energy is due to correlation. In LSD
flict between his argument and the correctness of the Kohn-Sham equations do not admit self-con-
the SIC-LSD correlation energies. sistent negative-ion solutions": The eigenvalue of
Figure 7 shows the LSD and SIC-LSD correlation the extra electron iterates to a positive value be-
energies E, for neutral atoms as a function of cause of the incorrect long-range behavior of the
N=Z. Note the change of slope that occurs in the one-electron potential discussed in Sec. II E. (If,
first transition series at Cr, as electrons of anti- however, the extra electron is artificially confined
parallel spin start to be added to the d shell. Ex- to a region near the atom, the LSD total energy is

O.O 0

'3
O -20
-I.O—
C9
K
4J
X
LLI

x
O
~ -2.0—
LLI
K
O
-60
O

-5.0- -80

0 4 8 l2 l6 20 24 28 52 56
ATOMIC NUMSKR Z
FIG. 7. LSD and SIC-LSD correlation energies of neutral atoms, as a function of atomic number g.
 5064 J. P. PKRDKW AND ALEX ZUNGER 23

losoex than for the neutral atom, and the difference I I I

is a rough estimate of the true binding energy. 6) He, Density response to the
correlation potential
This kind of misbehavior cannot occur in a self- 0.06—
interaction-free theory like HF or SIC-LSD: Be-
cause of the inequality (57), the eigenvalue of the
extra electron will always be negative if binding 0.04 —Exact (CI)
is allowed by the total energy. -"-
S1C

Table VIII shows HF (Refs. 66 and 67) and SIC-

LSD binding energies of negative ions (calculated
~I
0.0Z
LSD

from total energy differences) in comparision with

experiment. ' The HF values grossly underesti- 0.00
mate the binding, and in some cases predict an
instability for the experimentally stable negative -0.02
!.5
I I I

ions, while the SIC-LSD values are remarkably 0.5 I.O

close to experiment, giving us additional confi- r (a.u. )
dence in the SIC-LSD description of correlation. FIG. 8. Density response to:he cor relation potential
in the He atom, in LSD, SIC-LSD, and configuration-
F. Density response to the correlation potential interaction (CI) calculations. (The exact or CI curve
has been sketched from a figure in Ref. 69.)
In Fig. 8 we show the density response to the
correlation potential, i. e. , the difference between
the electron density calculated self-consistently G. Eigenvalues and removal energies
with and without correlation, for the He atom in
As discussed in Sec. IID, the orbital eigenval-
LSD and SIC-LSD. The effect of the correlation
ues (multiplied by -1) should approximate the
potential is to build up some extra density near relaxed energy changes due to electron removal in
the nucleus by depleting the low-density tail (the
SIC-LSD, but not in LSD. This conclusion is
region r & 1 a. u. ); the resulting relative change
supported by Table IX, which compares the eigen-
in the density is less than 2% in the high-density
value of the least-bound occupied orbital to the
region (r & I a. u. ), and the relative change may
measured first-ionization potential" in a series of
be even smaller in larger atoms where correlation
atoms. On the average, the SIC-LSD eigenvalues
is less important. Just as we found that the SIC-
agree with experiment better than the HF eigen-
LSD correlation energy is about half that of LSD in
values do. They seem to satisfy the inequality
Sec. IIIC, we now find that the SIC-I SD correla- (57) almost as an equality, even in He where re-
tion potential is about half that of LSD.
laxation effects should be largest. In contrast,
In Fig. 8 we have also plotted the difference be-

"
tween the configuration-interaction (CI) and HF
densities in He, which can be regarded as the ex- TABLE IX. Comparison of outer-orbital eigenvalues
act density response to the correlation potential. em~ with measured first ionization potentials I&. (e ~
Although the SIC-LSD density is clearly more is the least negative eigenvalue of the occupied orbitals. )
realistic than the LSD density, both approxima- Energies in eV.
tions miss the CI tendency to deplete the density
HP
in the intermediate region (x=0. 5 a. u. ). Smith, -~max I)
(Ref. 59) ~LSD ~SIC-LSD (Ref. 71)
Jagannathan, and Handler" have pointed out how max ~max

hard it is for approximations based on the uniform

H 13.6 7.3 13.6 13.6
electron gas to get this behavior right. Li 5.3 3.2 5.4 5.4
Na 5.0 3.1 5.1 5.1
K 40 2.6 4.3 4.3
TABLE VIII. Binding energies of negative ions in eV
(or electron affinities of the corresponding neutral 15.4 8.3 14.9 14,5
atoms). 10.7 6.3 10.0 10.5
HF Expt. 6.5 4.0 6.7 6.8
(Refs. 66 and 67) SIC-LSD P, ef. 68) 6.7 4. 6 7.1 7.4
H —0.33 0.7 0.75 He 25.0 15.5 25.8 24. 6
0 -0.54 1.6 1.46 Ne 23.1 13.5 22. 9 21.6
F 1.36 3.6 3.40 Ar 16.1 10.4 15.8 15.8
Cl" 2.58 3.8 3.62 Kr 14.3 9.4 14.0 14.0
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. .. 5065

the LSD eigenvalues only capture about 60/p of the To what extent can these conclusions carry over
ionization potential. into the self-interaction-corrected Kohn-Sham
Qualitatively similar conclusions apply when the scheme of Sec. IIF, in which all orbitals see the
ionization process leaves the atom in an excite same potential? In an attempt to answer this
state. Table X displays a detailed comparison question, we have also calculated the orbital
between ground-state orbital eigenvalues and re- eigenvalues for the Ar atom using the ox'bitally
laxed energy differences due to hole creation in averaged potential (OAP) of Eq. (75) (with expon-
Ar. First consider the relaxed energy differences ent p = 1), and the results are displayed in Table
&E~ „& in comparison with measured removal ener- X where they can be compared to the eigenvalues
gies": The SIC-I SD values show the best overall obtained from the orbital-dependent potential of
agreement with experiment, while the HF values Eq. (38). There is fair agreement between the
are somewhat in error for the outer orbitals (due two sets of eigenvalues, suggesting that our con-
to neglect of correlation) and the LSD values are clusions can in fact carry over.
slightly in error for the inner orbitals (due to
self-interaction). Note in particular the serious
H. Band structure of nonmetals
error of the HF approximation for the Ss hole in
Ar, where &E" ~«~ is 4 eV greater than the ex- Turning now to solids, we note that LSD band
'
perimental removal energy. Bagus has attribu- calculations have had some striking successes in
ted this error to an anomalous inc~ease in the relating the band eigenvalues and related spectral
magnitude of the correlation energy upon forma- functions to observed excitation. However, in
tion of the hole. Most of this increase is accoun- many of the calculations for insulators, if the cal-
ted for within the LSD and SIC-LSD total energy culations are pushed to full self-consistency and if
calculations. additional approximations such as artificial ex-
Next compare the orbital eigenvalues to the change scaling, muffin-tin approximations, em-
measured removal energies in Table X: The HF pirical charge-transfer factors, or empirical
eigenvalues overestimate the removal energies pseudopotentials are avoided, anomalously small
from inner orbitals (due to neglect of relaxation optical gaps E, are frequently obtained in LSD
and correlation), while the LSD eigenvalues ser- [e.g. , see Refs. (27)-(31) and Table XI]. Such
iously underestimate the removal energies from agomalies normally do not occur in semiempixical
all orbitals. In contrast the SIC-LSD eigenvalues
yield a good estimate of the removal energies. used for screening, "
pseudopotential calculations, even though LSD is
since agreement with the ex-
Finally, note that the self-interaction-free theor- perimental band gaps is usually forced through
ies (HF and SIC-LSD) satisfy the inequality (57}. empirical fitting of the pseudopotential. However,
In Table X we also show an estimate of the un- in the first principles p-seudopotential approach"
relaxed energy difference in SIC-LSD, based on such systematic errors persist. This failure of
the assumption that the relaxation energy p»~ of first-principles band-structure theory with local
Eq. (67) is the same as Z"r of Eq. (53). Note LSD-type potentials occurs, in fact, for chemical-
that in SIC-LSD the negative of the eigenvalue falls
between the unrelaxed and relaxed energy differ-
ly broad classes of nomnetals, ranging from van
der Waals solids such as Ne [E~c=11.2 eV, "
ences in agreement with Eq. (71), and lies closer Ep'=2l. 4 eV (Ref. 27}], through covalent solids
to the relaxed energy difference except for the 1s like Si(E""=0. 6 eV, "'"' '""
E'*"=1.17 eV} to
hole. ionic solids like LiF (E",so =10.5 eV, E'""=14.2 "
TABLE X. Comparison of ground-state orbital eigenvalues e», relaxed energy differences
hE~„&, and experimental electron removal energies in atomic Ar. The SIC-LSD column also
shows an estimate of the unrelaxed energy differences d E] ~,
and eigenvalues calculated
&

with the orbitally averaged potential (OAP). Energies in eV.

SIC-LSD
Hole HF (Ref. 46) LSD -~ni Expt.
-&, &&I ~t
g
&&I ~, g (Oltp) -d E(~, (Ref. 46)

1s 3227.0 3195.0 3097.0 317g.O 3227.0 3197.0 3220.0 3195.0 3206, 0

2s 335.0 325.0 294.0 310.0 322.0 322. 0 316.0 311.0
2P 260.0 24g. O 230.0 248.0 263.0 260.0 257.0 251.0 249.0
3s 34.8 33.2 24. 0 29.9 31.6 32.3 30.2 30.1 29.2
3P 16.1 14.8 10.4 15.7 17.2 18.2 15.8 15.8 15.8
 5066 J. P. PERDEW AND ALEX ZUNt" ER

eV) and transition-metal oxides like Coo (predic- the LSD approximation] using exactly the same lo-
ted erroneously by LSD to be metallic' ). While cal exchange-correlation functionals which were
for years such anomalies were often disguised by employed in the corresponding band-structure ca1.—
computational devices which "fixed" the gaps, it has
become clearer in the last few years" "'~'""'"" culations (Refs. 29 and 30). Table XI clearly
shows that the self-interaction of this orbital ac-
that they reflect a fundamental limitation of the counts for almost 100% of the LSD error in the in-
LSD band-structure model. In what follows we sulating gap of the solid. This is also confirmed
will present evidence (see also Refs. 31, 75, and through an actual SIC band-structure calculation
76 for our earlier corrections to LSD band gaps for solid Ar, described below.
due to SIC) that these discrepancies arise largely A similar argument applies to ionic solids such
(but not entirely) from self-interaction errors, as the alkali halides, in which the valence band is
and that the SIC scheme can form the basis for constructed predominantly from the anion p orbi-
a new and improved band theory. tals. Interestingly, in the past, the discrepancies
It is not difficult to see why self-interaction cor- between the small LDS band gaps and the larger
rections should significantly alter the results of experimental values in these materials were "re-
band theory for atomic crystals like solid Ar
(Ref. 77): Although an electron should see an "
medied" by employing an artifically large ex-
change parameter n ~ 1. This device restored an
Ar'-like potential in the valence band and an Ar'- apparent agreement with experiment since the val-
like potential in the conduction band, LSD actually ence states, characterized by fairly localized an-
forces an Ar'-like potential for both bands. The ion p orbitals, were lowered in energy more than
difference between the two corresponding orbital the diffuse conduction-band cationlike s orbitals
energies has both a kinetic component W (inclu- by the artificial enhancement of the (negative) ex-
ded in band theory) and an intra-atomic self-direct change potential. From our present discussion it
component II (neglected in LSD but not in SIC). is clear„however, that such an empirical ap-
For localized electronic states one expects U 8 W. proach does not treat the true physical origin of
We show in Table XI, along with the LSD band- the discrepancy.
structure error &E (=E",' —E;*"') for rare-gas It is easy to see why such an approach is not a
solids, the difference ~»c between the LSD and panacea: One could adjust n to make the self-in-
SIC-LSD eigenvalues of the outermost atomic or- teraction correction to the total energy zero, by
bital forming the valence-band maximum (e.g. , 2p, choosing in Eq. (33)
3p, 4p for Ne, Ar, and Kr, respectively). The cor- &&" =-U[n ]/-,'E""[n.„o],
rection &sl& is expected to be accurate for the val-
ence band when the latter is largely atomiclike and or to zero the self-interaction correction to the
unhybridized. Thi's is indeed the case for rare- orbital energy, by choosing in Eq. (39)
gas solids having narrow p-like valence bands
which are well separated from the lower s-like
~, ) (y„, lu([n„„];r) I „) '
band. We have calculated &stc =Ega —e ' [Eqs. '(y„. v'„""([n„., 0; r) y„.)
—, i i

(39) and (40), where e'&p'" of Eq. (40) is taken as

Not only are n",' 0 n'2, ' (so that fixing the band gaps
may spoil the total energy), but also these ex-
TABLE XI. Self-interaction corrections 4&zc to the change parameters are state dependent (so fixing
LSD band-structure error 4E~ in the band gaps of rare- one gap may spoil another one).
gas solids and the core ionization in solid LiF. Energies A recent example of this situation is the CuC1
in eV. band structure. The experimental band gap (3.4
eV) was obtained in an Xo' band-structure" calcu-
Solid
lation using + =1, while the correct LSD treatment
incorporating o = 3 for exchange (plus local corre-
Ne
Ar
21.4 ~
14 2
11.2
8.3 -5.9
9.9
5.8 "
lation) yielded &, =2.0 eV. The reason for this
Kr 11. 6~ 6.8 -4.9 4.9 is that the top of the valence band in this material
Xe 9.3' 4.2 is composed largely of the localized Cu 3d orbitals
LiF (Li ls) 64.4' 47.2' -17.2 18.0 (with 25%%up admixture of Cl p orbitals) which carry
LiF (F2@) 37 6" 301' -7.5 7.9 a considerable amount of self-interaction whether
& =1 or 3. The correction due to this self-interac-
~Reference 27.
~References 28 and 31. tion term has been estimated" from simple atomic
c Reference 29. models and found to lead to the experimentally cor-
d Reference 30. rect band gap only if o' =-', (plus correlation) is
'Reference 31. used, while for & =1 the corrected" band gap is
 SELF-INTR RAC TION CORRECTION TO DKN SITY-F UNCTIONAL. . . 5067

-4.0 eV. It is gratifying that an artificial enhance- overlaps with the lower (oxygen 2P-derived) val-
ment of the exchange parameter is no longer ence bands, local-density band-structure calcula-
needed to obtain the correct band gap in such ma- tions (e.g. , Ref. 74) indicate a large gap (up to 8
terials, since the use of + =1 spoils the agreement eV) between the two groups of bands. One expects
between the calculated ground-state bulk properties then that self-interaction corrections will play a
(equilibrium lattice parameter, cohesive energy, central role in reconciling this result with experi-
and bulk modulus) and experiment, while the use ment: The d-based valence bands, being far more
of & = 3 (plus correlation} produces the correct localized spatially than the 2p band, would receive
results. ~ a larger self-interaction correction (and orbital
It is interesting to contrast this situation with relaxation correction) and move down in energy
the one encountered in HF band-structure studies. relative to the p band, reducing if not closing the
HF calculations (e.g. , Ref. 4) predict band gape spurious LSD gap.
for insulators that are too large, often by as much LSD band structures are known to exhibit differ-
as a factor of 2-5 relative to experiment (due to ences with optical experiments not only at thresh-
lack of correlation and the inclusion of spurious old but also in the interband region, a few eV
self-interaction in the conduction band}. On the threshold. Such differences usually show up as
other hand, LSD band-structure calculations show 0.1-0.5 eV shifts of the calculated peaks in a, (~)
band gaps which are 20-40 Vp too seawall (among relative to experiment. These differences have
other reasons due to spurious self-interaction in encouraged theoretical work (e.g. , Refs. 82-84)
the valence band). Both the correlated HF results which emphasizes many-body corrections to the
(e.g. , Ref. 4) and the self-interaction-corrected spectra, in particular screened electron-hole and
LSD results match the experimental gaps reason- exchange effects. These corrections produce a
sonably well. It would seem, therefore, that LSD desired shift in the peaks (e.g. , a 0.2-eV shift out
forms a better starting point than HF to calculate of the 0.1-0.5 eV shift needed). However, such
band structures, as the needed corrections to the approaches completely ignore self-interaction cor-
band structure are far smaller than in HF while rections (a fundamental one-body effect} which lead
the bulk ground-state properties are described in to a far larger shift at threshold and may well re-
LSD at least as well (and usually better) than in sult in substantial shifts of a, (v) above threshold
HF. due both to a direct SIC modification in the single-
It is likely that SIC could explain additional dis- particle energies and to the change in orbitals
crepancies between LSD calculations and experi- which enter the screened exchange calculations.
ment for localized states in. extended system. For For example, the LSD band structure of Si pre-
example, the LSD-calculated positions of the lo- dicts a gap of 0.8 eV [Refs. 73(b} and 73(c)], com-
calized surface states in the relaxed GaAs(110) pared with the observed value of 1.17 eV. Many-
surface ' are systematically too high in energy by electron corrections" amount merely to a =0.2-
0.2-0.3 eV relative to experiment, when the ener- eV shift in the interband region and practically do
gy of the state is measured in both cases with re- not affect the value of the band gap.
spect to the Bloch-type valence-band maximum. We also show in Table XI the error of the LSD
SIC is expected to lower these states relative to band-structure prediction of the energy of core
the Bloch states. Similarly, nonempirical LSD electrons, relative to vacuum, in the ionic insula-
band-structure calculations on tetrahedrally tor LiF; again self-interaction correction of the
eigenvalue removes most of the discrepancy. Al-
(e.g. , BN, CdS, GaAs) "
bonded heteropolar semiconductors and insulators
as well as alkali and ternately, but with much greater effort, the cor-
alkali-earth halides' often indicate a systemati- rect answer Could be obtained within LSD by a
cally too narrow valence-band width. Inspection 4SCF calculation of the relaxed total energy dif-
of the wave functions involved shows that at the ference in the solid" upon removal of the core
bottom of the band (predominantly of cation s-like electron. The significance of self-interaction
character) the wave function is more localized terms in narrow-band solids has been emphasized
than it is at the valence-band maximum (predomi- by Zunger and Freeman" who have evaluated them
nantly anion p-like character). Self-interaction directly in a soLid-state 4SCF approach.
corrections would then tend to lower the valence- All band-structure techniques which employ in-
band minimum more than the valence-band maxi- scribed atomic spheres, and solve a radial atom-
mum, leading to a la. rger ba.ndwidth. iclike Schrodinger equation inside such spheres,
A somewhat similar situation occurs in the Mott lend themselves naturally to incorporating a sim-
insulators FeO, CoO, MnO, and NiO: While pho- plified SIC. One has to solve an SIC-.LSD single-
toemission studies indicate that the upper (transi- particle equation within the sphere (rather than an
tion-element 3d-derived) valence band strongly LSD equation) much as in the atomic case (except
 5068 J. P. PERDEW AND ALEX ZUNGER

for the modified Wigner-Seitz boundary conditions tons in solids, localized surface and interface
at the sphere boundary), while the more diffuse states, f-band materials and valence fluctuations,
wave-function components outside the sphere can and deep defect levels, and for the description of
be treated in the standard LSD form. The wave metal-nonmetal transitions.
functions are normalized over the entire unit cell.
This approach constitutes a generalization of the I. The "Fermi Statistics" problem in Fe and Co
classical Wigner-Seitz treatment of metallic sodi-
um" to include both exchange and correlation in A peculiar behavior occurs in LSD calculations

a self-consistent fashion. Such a technique would for atomic Fe and Co: If the occupation numbers
incorporate the major SIC on the localized states are restricted to integer values (0 or 1), then
and lead to improved band-structure
the orbital configuration which minimizes the total
energies and
wave functions. energy violates the aufbau principle. For ex-
A self-interaction-corrected self-consistent ample, in Fe the calculated lowest-energy con-
band-structure calculation has been performed" figuration is the observed one d's', but in this
for the valence (3p) and conduction bands of solid configuration the eigenvalue of the occupied 4sf
Ar in the fcc structure. A pseudopotential appro- orbital lies above the eigenvalue of the four unoc-
cupied 3db orbitals. Slater' argUed that the mini-
priate to the 3p states, and a mixed basis of Gaus-
mizing configuration must then have fractional
sians and plane waves, were employed. The re-
sulting pseudo-orbitals g,.-„( r) were used to find the
occupation (d"s"
in this case").
Fractional occupation is not allowed in the ori-
pseudodensity by summing over the first Qrillouin
ginal Kohn-Sham scheme, as Janak" and Harris"
zone and over the three upper valence bands as
have pointed out. From the discussion of Sec.
BK 3
n„(r)=2 +Q
k /=1
~P, ;(r)~'. IIA, we know that fractional occupation is now
permitted in prie pie. The more practical prob-
lem of finding a realistic approximation for the
If there are M atoms per unit cell (where conven-
exchange- correlation energy under fractional
tionally M=1), then the integral of n~, (r) over the
occupation remains; this problem is not solved
unit cell equals 6M. The averaged orbital density
by our scheme of self-interaction correction ex-
n~, (r)/6M was used to calculate &, [Eq. (33)j as cept in certain idealized cases, as discussed in
well as the self-interaction-corrected potential Appendix B.
v, f'f(r) [Eq. (38)]. Since n~, (r)/6 approaches the The SIC-LSD scheme of Sec. IIB evades the
atomic orbital density ~g,'~(r) in the limit of in-
~
Fermi statistics problem in the following way:
finite lattice constant, the resulting band structure When this scheme is applied to Fe in the d's' con-
for M =1 has the correct atomic limit. Prelimi- figuratiori, the eigenvalue for the occupied dk or-
nary results indicate that at the observed. lattice bital lies below that of the occupied s4 orbital,
constant, the calculated band gap is 13.8 eV, with while the eigenvalue of the unoccupied dk orbitals
the Kohn-Sham exchange and 15.0 eV with the He-
lies above s4, so the aufbau principle is satis-
din-Lundqvist exchange-correlation (8.2 eV with- fied. In the self-interaction-corrected Kohn-Sham
out SIC) which compares favorably with both the
scheme defined- in Sec. II F, all the dk orbitals
experimental value (14.2 eV) and the atomic esti- will be degenerate and we expect that the Fermi
mate of Table XI. The valence-band width in-
statistics problem may reappear.
creases from the LSD value of 1.3 eV to the SIC-
LSD value of 2.4 eV. Note, however, that the
J. Cohesive energies of transition metals
success of this approach rests on the intuitive
choice M =1; if the calculation were repeated with The cohesive energy E„„isthe small (1-'6 eV)
larger unit cells, the orbital density n~, (r)/6M difference of two large numbers: the energy of the
would carry zero self-interaction in the limit of free atom and the energy per atom of the solid.
large M. This approach clearly exemplifies the Although LSD-calculated total energies of atoms
size-consistency problem described in Sec. IIF. and solids are in error due to self-interaction,
More applications of the SIC band-structure
method will have to be attempted in the future
their difference yields a reasonable estimate for
the cohesive energy in most metals. In K for "
(either within the practical schemes described example, where the LSD self-interaction error of
here and in Sec. IIF or by other schemes) before the atomic energy is 73 eV, the LSD cohesive
an efficient computational method will be clearly energy is only 1 eV, in good agreement with ex-
identified. One would expect such a method to be periment. The explanation is of course that nearly
particularly useful for band-structure calculations all the self-interaction error is carried by the
of nonmetals as well as for the calculation of the ionic core, which is essentially the same for the
electronic structure of core excitations and exci- solid as for the atom, whereas the self-interaction
 SKLF-INTERACTION CORRECTION To DKNSITY-FUNCTIoggL. ..
error for the K 4s orbital is much smaller than
the cohesive energy.
The 3d transition metals are a notable exception valence
to this rule. Figure 9 shows a comparison of LSD
(Ref. 12) and measured" cohesive energies across "'d valence,
N&5',.
the 3d series. Note that LSD gives accurate co-
hesive energies at both ends of the series, but
wllere onl)I VRlellce orbltais contribute Il 8 ~

overestimates them by as much as 70/q near the P"I"(core) = 5„",'m(core)]. The solid-state self-
middle.
interaction 5& need not be evaluated in the present
As we cross the transition sex'ies from Ca to
estimate, but is in principle to be calculated from
CU, the 3d band is progressively fiBed, and the
sultabi)I localized (Wannler-tiIpe) fllIlc'tlolls fol'
orbitals contract in response to the increasing
the valence bands. Vfe take the 6 values from Fig.
nuclear charge. As a result the LSD self-inter-
action 6„, of each orbital grows progressively, f " '
9, gL from Janak et al. and 5„'", (valence) from
as shown in Fig. 10. The 3d orbitals are xathex' Fig. 10, and calculate Q&&& P,."'~(valence) from Eq.
(V9). From this we get the ratio between the
localized and each carries a self-interaction
valence self-interaction in the solid and in the
(0.2 to 1.2 8V per electron) that is large on the
scale of the cohesive energy, but they are not
inert core orbitals until we reach Cu; in fact, they
hybridize strongly with the 4s orbitals. As we
move to the post-transition elements, the 3d or-
bitals carry even mox'e self-interaction but they
are largely chemically inert when the solid is 0.00
formed from atoms; this leads to an effective
cancellation of SIC core terms in the cohesive
eQergy x'ende11Qg LSD a good appl oxlmatlon. It
seems likely, then, that the s-d valence shell, car-
x'les a gl eater LSD self-1ntex'act1on ex'1 ox' 1Q the
transition atom than it does in the solid, leading
to the LSD ovex'estimation of the cohesive energy.
Let us define the error of the LSD cohesive ener- -0.6

Following the argument of the preceding para- -$.0-

graph, we equate 6 to the difference between the -1.2

LSD self-interaction per atom in the gas and in the
IC
LU

-20-

40—
a
gp
O
CP
hJ

Cl
CO ~~
Du
td
0'= Ca Sc Ti V Cr Mn Fe Co Ni Cu

t~~ l ) ) l l I l ) l l i
FIG. 10. LSD self-interaction error of the atomic
Li No K Ca Sc Ti V Cr Mn Fe Co Ni Cu Zn Go
orbitals in the 3d transition series. The self-interaction
FIG. 9. Percent error in the LSD cohesive energies ~o~~g~ggo@ -g„& is plotted here for the majority-spin
for some alkali metals and the sd metals. The LSD orbitals. This correction is small (and in some cRses
and experimental values were taken from Refs. 12 and . positive) for the 4g orbitals, but large and negative for
88, respectively. the 3d orbitals.
 5070 J. P. PEROEW AND ALEX ZUNGKR

atom. We find that the valence self-interaction I I l I I I I l

in the solid is negligible for the free-electron 0.0S

metals Li and Na [e.g. ZP' 68"'~ (valence) equals u 0.06 I}t-

0. 00 and 0. 01 eV for Li and Na, whereas the co- LLI

0.04—
hesive energies are 1.6 and 1.1 eV, respectively], LLI

and it varies from -8% of the atomic valence 0.0R—

X
self-interaction at the beginning of the 3d series Q 0
(Ti), through 55% at the center (Mn} to 86% at
the end of this series (¹).
Such a residual SIC
for d electrons in the 3d metals seems physically
~~

L
-O. OR—
-0.04—
reasonable and points to the possibility that self- -0.06—
4J
interaction effects should be considered in transi- ~ -0.08— d 8 stable
tion-metal physics. An g priori calculation of 5& I I I I I I I I

(valence} in the solid using a suitably localized Ca Sc Ti V Cr Mn Fe Co Ni Cu

wave-function representation seems very desir- FIG. 11. s-d interconfigurational energies of transi-
able. Note from Eq. (33) that 6,
is basically the tion-metal atoms (calculated as total energy differen-
screened "Hubbard U" (intrasite repulsion term). ces) compared with experiment (crosses, from Ref. 33).
Our -analysis above suggests that the Hubbard re- The break in the curve at Cr occurs because of the
pulsion may not be screened out to zero even in Hund's-rule special stability of the d configuration.
The LSD values shown here are about 0.2 eV closer
elemental transition metals.
to experiment than those of Ref. 33, due to our use of
the accurate Ceperley correlation.
K. s-d interconfigurational energies of transition elements

The interconfigurational energy is the total ener- another.

gy difference between the atomic d" 's' and d" s' Self-interaction correction neither improves
orbital configurations. Harris and Jones" have nor worsens the LSD interconfigurational ener-
pointed out that LSD predicts the trends of this gies. These LSD values are in fact remarkably
quantity across the 3d transition series in good "persistent": The same values are found in LSD
agreement with spectroscopic measurements (Fig.
ll), although there is an absolute error of about "
calculations whether the density is spherical. ized
or not, and the same values are also obtained
1 eV in the LSD results which unduly favors the
d-rich configuration, leading to the wrong pre- (83) below. ""
from the "weighted density" approximation of Eq.

dicted ground state in Ti, V, and Co. The same

error seems to lead to an incorrect description L. Spin splitting in Fe, Co, and Ni
of s-d charge transfer in transition-metal dimers. '9
It seems at first thought that self-interaction The LSD approximation has been successful in
correction should seoxsen the calculated intercon- predicting the spin susceptibilities of the para-
figurational energies, since from Fig. 10 the cor- magnetic metals, "'"
which are strongly enhanced
rection is negligible for a 4s orbital but substan- by exchange and correlation. LSD also correctly
tially negative (-0.3 to -1.
2 eV) for a 3d orbital. predicts that the ground states of Fe, Co, and Ni
Thus the first-order effect of the self-interaction are ferromagnetic, with calculated magnetic mo-
correction is to favor the d-rich configuration even ments in good a,greement with experiment. ' Ex-
more than LSD does. However, a glance at Fig. cept for the weak ferromagnet Ni, the LSD band
11 shows that this expectation is not met, and that structures of the ferromagnetic metals are in
the self- consistent SIC- LSD interconfigurational reasonable agreement with the results of angle-
energies are essentially the same as those of LSD. resolved photoelectron spectroscopy. ~2 For ex-
For estimates of the correction to the total ener- ample, the ratio of theoretical to experimental oc-
gy or to the one-electron energy eigenvalues, the cupied d-band width is about 1.1, 1.2, and 1.45
self-interaction correction can be treated as a for Fe, Co, and Ni, respectively, while the ratio
first-order perturbation on the LSD solution (i.e. ,
approximating SIC orbitals by LSD orbitals), but
of theoretical to experimental spin splitting is
about 1.0, 1.2, and 2.2 for the same metals. "
this approach fa, ils for the interconfigurational The LSD overestimate of the spin spbtting in
energy. %'hat happens is that the first-order ef- Ni by a factor of two led us to wonder if self-inter-
fect lowers the energy of d" 's' relative to d" s', action correction would change the LSD spin split-
but the higher-order effects lower the energy of ting. The analog in our atomic calculations of the
the more polarizable d" 2s' configuration. relative solid-state band-structure spin splitting is &3~~
to d" 's', and these effects essentially cancel one —&,~&, the difference in eigenvalues of the minority
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. .. 5071

(0)- and majority (0)-spin 3d electrons. Figure 12 rections are not the same in the solid transition
shows this quantity as a function of atomic con- metals as in the corresponding free atoms, so
figuration in LSD and SIC-LSD; the measured these atomic calculations may not be very predic-
solid-state spin splittings are also shown for tive of solid-state effects.
comparison.
Note that self- interaction correction increases IV. BRIEF REVIEW OF RELATED WORK
the spin splitting, although the increase for Ni When we separately began to work on self-inter-
is only marginal. The reason is very simple: The action corrections, we knew about the work of
majority spins see a more attractive exchange- Lindgren. ' Since then we have learned of other
correlation potential than the minority spins do, related studies. Here we attempt a brief critical
and so have more localized orbitals. The self- review of the papers we have found, in relation to
interaction correction lowers the eigenvalues of our own work. Except where otherwise indicated,
all occupied d orbitals, but the eigenvalues of the this related work was directed toward removing
more localized majority-spin orbitals get lowered the self-interaction error of the one-electron
more, leading to an increase in the spin splitting. potential/ in the local-density approximation, in
[This argument, due to L. Kleinman and K. Mednick contrast to our own work which gives a general
(private communication}, is based on treating the scheme for self-interaction correction of gny
self- interaction correction to the one- electron po- spin-density- functional approximation for the
tential as a first-order perturbation on the LSD total energy as well as the potential.
eigenvalue, a treatment which gives numerical Lattere' observed that the Slater n'"(r) exchange
results in rather good agreement with our self- potential approaches zero exponentially far from
consistent ones. However, the success of this the center of a neutral atom, while the Hartree-
approach should not be interpreted to mean that Fock exchange potential goes to -1jv, and pro-
the self-interaction correction has no effect on the posed switching over from the former to the latter
orbitals of transition-metal atoms; in fact, at self- form at the value of x where they cross, a scheme
consistency it contracts the 3d orbitals and expands followed by Herman and Skillman~~ in their atomic-
the 4s orbitals non-negligibly. J structure calculations. From the present work,
These atomic calculations suggest that self- however, it is clear that self-interaction effects
interaction correction couM slightly worsen the modify the potential even before it reaches its
calculated spin splittings in solid Fe, Co, and Ni. asymptotic -1/x value (see the difference in orbi-
However, as we observed in our discussion of the tal-dependent potentials in Fig. 4 for small x}.
cohesive energies (Sec. III J), self-interaction cor- Cowan". showed that Latter's procedure can intro-
duce errors into calculated ionization energies,
and proposed a scheme in which the self-Coulomb
potential u([n, ]; r) was subtracted from the one-
electron potential. seen by orbital Qt, o, while the
4— interelectronic exchange potential was parame-
CQ
trized in a complicated way to give a best fit to
the Hartree-Fock orbital energies. This scheme
gave the correct long-range behavior of the po-
L,
CO
tential, and the orbital, s gave a good total energy
4J
C9
Ni— when inserted in the Hartree-Fock energy function-
z
I: ~ Fe
al (not the density-functional expression).
O
X
4l
~ Co
Liberman"-proposed partitioning the density of
an atom into a sphere containing N-1 electrons
and the rest of space containing one electron. The
total energy then involved local exchange only
dn o d
n-I
s
I dn-I 2 within the sphere and electrostatic interaction
between the two regions. This scheme corrected
CONF I GURATION
the self-interaction error for the outermost elec-
FIG. 12. Spin splitting in atomic Fe, Co, and Ni as tron, so that a ~SCF calculation gave binding ener-
a function of atomic configuration. The difference bet- gies of negative ions as good as the Hartree-Fock
ween the energy eigenvalues of minority- and majority- values or better, but of course it left the large
spin occupied 3d orbitals. is plotted' here. (The LSD
self-interaction error of the inner- shell el.ectrons
values for Fe and Co in the d"s configuration are un-
certain due to convergence difficulties. ) The observed uncorrected.
(Ref. 32) spin splittings in the solid, indicated by Slater and Wood' pointed out that a large fraction
horizontal arrows, are shown for comparison. of the total exchange energy in atoms is just self-
 J. P. PERDEW AND ALEX ZUNGER

exchange. However, they rejected any attempt to essentially the exact Kohn-Sham potential for ex-
treat self-exchange exactly because, unlike the change in atoms. In a comparative study of various
total exchange energy, it is not invariant under spatially local exchange potentials v„(r), Talman,
unitary transformation of the occupied orbitals. Ganas, and Green' found that those which mini-
We discussed this fact, and how it may be ex- mized the Hartree-Fock energy either exactly or
ploited, in Sec. IIF. in a restricted variational sense gave orbital ener-
Lindgren's" "work is perhaps closest to our gies &, in better agreement with measured re-
own. Lindgren and Rosen ' calculated the inter- moval energies and excitation energies than the
electronic exchange energy of a uniform electron Hartree- Fock eigenvalues.
gas confined to a finite volume by periodic bound- Brandow' ' pointed out that a local-exchange Xn
ary conditions, and found a correction to the local band-structure model fails for Mott insulators
exchange energy which becomes such as CoO, MnO, and NiO essentially due to
the failure of the theory to subtract the self-inter-
—
F fd'rn, (r)a, (n„(r), 0) (80) action terms from the potential. He then proposed
to construct a d-band Hubbard Hamiltonian based
for small electron numbers N, and never deviates on a linear combination of atomic orbitals fit to
much from this for large N. Using Eq. (77) and Xa band structure with the self-interaction terms
the exact self-Coulomb correction -Q,
U[n, ] as properly subtracted. This model Hamiltonian
corrections to the local exchange energy, they correctly describes the energy levels and magnetic
variationally derived an orbital-dependent potential properties of these systems when empirical values
which is the same as the one we would find from for the self-Coulomb terms are used.
Eq. (38) within the SIC scheme for local exchange Gopinathan" devised a scheme in which the
Otlly '. self-Coulomb correction was treated exactly,
as in Lindgren's work, while the interelectronic
v„"(r) = p'„(nt(r), ) n~(r)) exchange was approximated as
—u( [n, ]; r) —p, &(n, (r), 0) . (81)
In atomic calculations, this scheme gave eigen-
values &, much closer to Hartree-Fock than the x fn, (r) —n„(r)], (82)
local exchange eigenvalues. Curiously, the total which is roughly in the form of an interaction be-
energies obtained directly from the self-interac- tween the orbital density n,
(r) and the remaining
tion-corrected density functional for the energy spin density [n,(r) —n, (r)]. The variationally
were never reported. Instead, the Schrodinger derived potential had the correct long-range be-
equation with the exchange potential of Eq. (81) havior, and the eigenvalues and hyperfine contact
was used merely as a wave-function generator for densities were close to Hartree-Fock values.
HF calculations. In separate work, Lindgren and Again the self-interaction-corrected density func-
Schwarz" showed that the dominance of the self- tional for the energy served only as an interme-
exchange over interelectronic exchange in small diate step in the derivation of the potential and was
atoms was the reason for the decrease in the op- not itself evaluated.
timum n value of Slater's Xe theory from -0.8 A Kohn-Sham density functional for the energy
for small atoms to -&, the proper LSD value for which is self-interaction-free [to the extent of
exchange, in large atoms. satisfying Eq. (29) for fully occupied orbitals] has
Szasz, Berrios-Pagan, and McGinn" proposed
an approximation for the atomic energy as an ex- "
been proposed by Gunnarsson, Jonson, and Lund-
qvist, and independently by Alonso and Girifal-
'
plicit functional of the density, including a simple co.' This is the "weighted density" approxima-
self-interaction correction in the spirit of the tion, in which the exchange-correlation hole (Secs.
Fermi-Amaldi correction. ' IIA and IIC) is approximated as
Schrieffer briefly discussed the self-interac-
tion error of Xn and LSD total energies, and gave , p„(n(r)) fr' —r I)
(83)
an argument qualitatively similar to the one of
Eq. (41) to show how this error gets magnified where p„(n, ~
r' —r ~) is
the hole in a homogeneous
in the orbital eigenvalues &,. electron gas of density n, . The weighted density
Talman and Shadwick"' showed how to construct n at each r is determined to satisfy the sum rule
an orbital- independent, spatially local exchange (26). This approximation obviously becomes ex-
potential v„(r) which minimizes the Hartree- Fock act in the limit of slow spatial. variation of the
total energy. This potential naturally turns out to density, like LSD. Applied to a one-electron sys-
be self-interaction free, and can be regarded as tem, the sum rule (26) can only be satisfied for
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. . . 5073

n =0 (the Wigner crystal limit) and then p(r, r') (xn' ', and that an appropriate choice of the pro-
=-n(r'), which is exact;" thus our SIC scheme portionality constant, slightly' different from the
gives no correction to the weighted density ap- standard Wigner value, elminates the self-interac-
proximation. The weighted density approximation tion in LSD for a variety of site-localized orbital. s
has been tested in non-self-consistent calculations (exponential, Gaussian, etc. }

"
for atoms. It reduces the -10% error of the LSD
exchange energy to -4%, and greatly improves
Zunger and Cohen'o' applied a self- consistent
self-interaction correction, analogous to the pre-
the exchange energy density. ' The error in the sent one, to. the local-density potential, including

by a shell-partitioning
"
integrated exchange energy can be reduced further
scheme, at the cost that
correlation, and displayed the improvements in
the eigenvalues as well as the valence-electron
the approximatioq ceases to be a density function- total energies that result for a number of mono-
al. . Unfortunately,
estimated by about a factor of 2, as in LSD. "
the correlation energy. is over- valent metallic atoms. They also proposed that
the self-interaction correction can be profitably
Also, as in LSD, the orbital eigenvalue for the wedded to the pseudopotential idea: In a simple
negative-ion H turns out to be incorrectly posi- metal the valence electrons are delocalized and
tive, ~' while the ionization potentials of atoms carry no sel. f-interaction, so they can be treated
(calculated from relaxed energy differences) are in LSD; the self-interaction corrections to the
"
somewhat worse than in LSD. Thus this scheme, core orbitals can then be absorbed into a first-
which in comparison with SIQ-LSD is rather more principles electron- ion pseudopotential. This
sophisticated, usually gives less satisfactory nu- program has now been implemented xxo
merical results. Perdew"' described the present version of self-
Bryant and Mahan' ' used a form of self-interac- interaction correction to LSD, and tested it in
tion correction in the potential to calculate soft non-self-consistent calculations for small atoms.
x-ray absorption and emission in metals. To find This work revealed the-accuracy of the SIC-LSD
the potential for core orbital no, they subtracted energy functional itself (not just the one-eLectron
the self-Coulomb potential u([n„]; r) from the LSD potential), and of the separate exchange and cor-
potential (ignoring self-exchange correlation), and relation contributions to it.
remarked that. this gave a small correction to the
shape of the orbital which translated into a large
V. CONCLUSIONS
improvement in the ionization cross section near
threshold. From the present study (cf. Fig. 5) The Hohenberg-Kohn-Sham" density-functional.
it is clear, however, that the exchange and cor-
relation contribution to the self-interaction cor- "
theory, as generalized by von Barth and Hedin'
and Levy, is a formally exact prescription for
rection for the eigenvalue [last term in Eq. (39)] the ground-state energy, electron density, and
is very large and of the opposite sign to the Cou- spin polarization of any many-electron system.
lomb term included by Bryant and Mahan. Fractional-occupation numbers are allowed in
Stoll, Pavlidou, and Preuss'" suggested that the principle, although further work on the exchange-
correlation energies of atoms and molecules could correlation functional is needed before they can be
be estimated by the equation used in practice. The exact density unctional
for the energy is self-interaction free, which
E, = d'rn(r &, nt r, n& r) means in particular that it satisfies Eq. (29).
Approximate density functionals [e.g. , the LSD
(84} of Eq. (27)] can be made to satisfy Eq. (29) by a
simple scheme of self- interaction correction
Here the first term is the LSD approximation for (SIC), Eq. (32). Correction of the self-consistent
the total correlation, and the next term subtracts one-electron potential follows naturally from the
out the LSD approximation for parallel-spin cor- variational principle. The most straightforward
relation. They argued that the correlation be- self- consistent formalism incorporating self-
tween electrons of parallel spin is large in the interaction correction is that of Eq. (3V). Like
homogeneous electron gas but not in atoms and the Hartree approximation which it resembles,
molecules. Stoll"' has pointed out that Eq. (84) it is a density-functional theory in the sense of
effectively subtracts out the spurious self-cor- the Hohenberg-Kohn' theorem, although it does
relation, as in the present work. not fit into the Kohn-Sham scheme.
Shore, Zaremba, Bose, and Sander' ' applied In self-interaction-free approximations (such as
density-functional theory to the Wigner crystal- Hartree, Hartree-Fock, or SIC-LSD), there is an
lization of a low-density electron gas. They ob- inequality [Eq. (5V)J relating the eigenvalue of an
served that in the low-density limit e„,(n~, n&} occupied orbital to the relaxed energy difference
 J. P. PERDEW AND AI EX ZUNGER

upon removal of the electron from this orbital. D. Ceperley for sending us his numerical data on
This inequality is often close to an equality, and the electron-gas correlation energies prior to
amounts to an "extended Koopmans's theorem", publication, to L. Kleinman for correcting our
less precise but possibly more powerful then the earlier misconception about the spin splitting in
or iginal. metallic Ni, and to M. Levy for comments on the
When applied to LSD, the self-interaction cor- manuscript and collaboration on Appendix A. One
rection properly satisfies the sum rule of Eq. of us (J.P. ) acknowledges the support of the Na-
(26) for systems without fractional occupation tional Science Foundation (Grant No. DMR 78-
numbers, and improves the shape of the exchange- 12398) .
correlation hole. We have applied this SIC-LSD
scheme in atomic calculations, with results that APPENDIX A: COMMENTS ON THE DEFINITION
are far more accurate than those of LSD for a OF THE KINETIC ENERGY
modest increase of computational effort. (In com-
Here we will present an alternate definition of
parison with HF, SIC-LSD is both simpler gnd
T[n~, n&] which is equivalent to that of Eq. (14).
more accurate. ) Calculations of this type were
We will then discuss the nature of the "minimiza-
recently shown'" to generate accurate a priori
tion" which appears in the Kohn-Sham formalism
pseudopotentials to describe the interaction be-
of Eq. (17) and in the Hartree-type formalism
tween valence electrons and ions in more extended
of Eq. (37).
systems such as molecules and solids. The SIC
scheme can also be applied directly to the all-
If the occupation numbers f,
which yield the
minimum in Eq. (14) are all 0 or 1, then the de-
electron problem in molecules and solids, if care
finition (14) obviously reduces to the original
is taken to avoid the "gremlins" discussed in Sec.
definition of Kohn and Sham'. T[n~, n&] is the
II F.
Benefits of applying self-interaction ground-state kinetic energy of a system of non-
correction
to the LSD approximation in atomic calculations in- interacting electrons with spin densities n~(r) and
clude: (i) improved values for the total energy n&(r). If some of the minimizing occupation num-
bers are fractional, then Eq. (14) is equivalent
and for the separate exchange and correlation
to a more general definition. Levy" originally
pieces of it, (ii) accurate binding energies of
proposed the generalization
negative ions, which otherwise could only be cal-
culated by heavy CI calculations, (iii) orbital T[nt, n(] = min(T), jA1)
eigenvalues that closely approximate physical
A
removal energies, including orbital relaxation when T is the N-electron kinetic-energy operator
effects (in sharp contrast to the LSD eigenvalues, and the minimum is over all antisymmetric N-
which are almost useless for quantitative pur- electron wave functions which yield the given spin
poses), (iv) correct long-range behavior of the densities. Valone'" has proposed that the search
one-electron potential and density, and (v) im- in Eq. (10) be made over statistical mixtures of
proved overall description of the density. Our antisymmetric N-electron wave functions; if we
calculations also suggest that self-interaction cor- apply this search also in Eq. (Al), we find that
rection will remedy the LSD underestimate of the the definitions of Eqs. (14) and (Al) are equivalent,
band gaps in insulators (as shown by numerical since by Coleman's theorem'" the necessary and
results for the rare-gas solids and ionic crystals) sufficient conditions for a Hermitian one-particle
and the LSD overestimate of the cohesive energy density matrix n(r, r') to arise from some statis-
of the 3d transition metals. However, self-inter- tical mixture of N-fermion wave functions are
action correction is not the cure for gll the ills that its eigenvalues (occupation numbers) lie in
of LSD, and in particular it leaves the spin split- the interval 0 ~f, & 1, and Jd'rn(r, r) =Q,
f„,=N.
ting in Ni and the s-d interconfigurational ener- We now consider the nature of the "minimiza, -
gies of the transition elements unimproved and es- tion" in Eq. (14). If the trial orbitals g, are con-
sentially unchanged. strained to be orthogonal, then T[nt, n&] is a true
minimum of Z„f, (P, --,' V' g„). Note, how-
~ ~

ever, that the Euler equation (17) is derived with-

ACKNOWLEDGMENTS
out imposing the orthogonality constraint. Thus if
One of us (J.P.) acknowledges valuable discus- the trial orbitals are not constrained to be ortho-
sions with L. Cole, D. Langreth, J. Rose, gonal, T[n ~, n~] will be the lowest-lying extremum
H. Shore, and S, Vosko. We are grateful to of Z, f,(P, ~--,-c' ~ g,
) such that no orbital is
G. Baraff, L. Hedin, D. Koelling, I. Lindgren, more than singly occupied, and this extremum is
and H. Stoll for pointing out a number of important not necessarily a minimum. (It might, for ex-
references to us. Special thanks are due to ample, be a saddle point. ) The extremizing or-
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. .. 5075

bitals, being solutions of Eq. (17), are of course and this can be described by giving half-occupancy
orthogonal. to the spin-up and spin-down orbitals on each
These observations become more than academic proton. This situation is not properly described
in the Hartree-type theory of Sec. IIB, when Eq. by our scheme of self-interaction correction,
(14} no longer applies and the extremizing orbitals which properly eliminates the interaction of each
are no longer orthogonal. In this case, "minimi- orbital with itself but improperly retains an inter-
zation over the orbitals" really means extremiza- action between the two fractionally occupied or-
tion. However, "minimization over the spin den- bitals on the same proton.
sities" retains its unqualified meaning.
APPENDIX C PARAMETRIZED CORREL-ATION
ENERGIES
APPENDIX B: FRACTIONAL OCCUPATION
NUMBERS Here we describe the parametrization we used
XVe need to define the energy functional for frac- for e,(nt, n&}, the correlation energy per electron
tional occupation for two distinct but related rea- of a homogeneous electron gas, and explain why
sons: (1) According to Eq. (54), the relaxed energy we believe it is preferable to other parametriza-
difference upon removing an electron from orbital tions.
no is the integral of the orbital eigenvalue e,(f, )
Ceperley" has calculated the energy of a uni-
over the interval 0(f, & 1. From this point of form electron gas over a wide range of densities.
He used Monte Carlo techniques to sample a cor-
view, the most natural continuation of F.„, is one
which satisfies Eq. (29}, since in that case the related wave function for electrons in a finite
eigenvalue of an isolated orbital over the range volume subject to periodic boundary conditions,
0(f, ( 1 will be the total energy at full occu- and extrapolated the energy per electron to infi-
nite volume. Let us define
pancy (f, =1}. Then for many-electron systems
the whole dependence of &, on f, will come from
physical relaxation effects. (2) There may be
n =(4mr', /8) ', (Cl}
systems in which the energy minimizes with frac- g =(n, —n, )/n. (C2)
tional occupation numbers at the Fermi level, so Then Ceperley's parametrization of the correlation
here the continuation of E„, must be a physical one. energy for z, )1 is
Fractional occupation numbers arise whenever
an electron hops between degenerate orbitals. &,'=r;/(1+ p'«, + p.'r, ), (C8)
Let us consider a model system composed of N where i = U (unpolarized, ) =0) or P (polarized,
protons equally spaced around a ring, and one l' =1). The corresponding correlation potential is
electron. The solution of the Kohn-Sham equations
is a molecular orbital in which the electron is p
f — ]
shared equally by all the protons. As the proton
separation goes to infinity, one can also define
1s hydrogenic orbitals on each proton which are
, (1+ ~P,
c
'«, '
+ , P,'r. )— (C4)
(] + p1~r + plr }
arbitrarily close to being solutions of the Kohn-
Sham equations; these orbitals have fractional For atomic calculations we also need &, at high
densities (r, (1) and arbitrary polarizations
occupation 1/N. (The electron stays a long time
on one proton before hopping to another. } The
(
0 g & 1. The leading behavior of the high-density
energy must be the same as it was for an isolated expansion is
hydrogen atom, which it will be if Eq. (29) is satis- &,'=A. , le, +B,. +C,.x, lnx, + D,.x, , (C5)
fied for fractional occupation.
As a second example, consider two protons and and the corresponding correlation potential is
two electrons. For small interproton separations, p, , —
'=A, lnr, +(B, -', 2,.)+ ', C, r, lnr, -
the Kohn-Sham solution of lowest energy will put
both electrons, with opposite spins, into the same
+ ', (2D,. —C,.)r, . - (C6}
molecular orbital. As the separation increases, The constants A~ and BU are known from the cal-
a point may be reached where a spin-split solu- culation of Gell-Mann and Brueckner, "'
tion, with the spin-up electron on proton 1 and the
spin-down electron on proton 2, has the lower 4 =0.0311, B =-0.048 (C7)
energy. '" This situation is degenerate with one (atomic units). To find A~ and B~, note that in
in which the spin-down electron is on proton 1 the high-density l.imit
and the spin-up on proton 2. The physical situation —
is that the spin on each proton flips up and down, ~ Ci ~ Ci, RPA + ~ g2
c (C8)
 J. P. PKRDKW AND ALEX ZUNGKR

where &,' ~" is the correlation energy in the ran- TABLE XIII. Correlation energy per electron of the
dom- phase approximation and unpolarized and polarized uniform electron gas. CA:
parametrized Ceperley-Alder, HPA: numerical random-
&~ = 0.023 phase approximation. r~ in eV.

is the second- order exchange constant (indepen- Unpolarized Polarized

dent of spin potarization). Using Misawa's'" CA HI A CA It PA
RPA scaling relation
0.01 -5.21 -5.82 -2.68 —3,32
4. P, RPA(
r) — e Ui RPA(r /24/3)
~
(C10) 0.1 -3.30 -3.89 1 72 + -2.34
0.5 -2.07 -2.65 -1.10 -1.68
me find 1 -1.62 —2.14 -0.86 -1.41
Z =0.01555, a =-0.0269. 2 -1.23 —1.68 -0.66 -1.15
3 -1.01 -1.44 -0.55 -1.01
We used Elis. (C3) and (C4) for r, ) 1, and Elis. -0.87 -1.27 -0.47 -0.91
(C5) and (C6) for r, &1. The coefficients C, and 5 -0.77 —1,16 -0.42 -0.84
chosen to match &,'and p, ,' at r, =l, and
6 —,0.69 -1.06 -0.38 -0.79
D,. were 10 -0.51 -0.84 -0.29 -0.64
'till'lied [Tills ls all llltel'-
oil't 'to tie 1'atilel' smaU. 20 -0.31 —0.58 -0.19 -0.47
polation procedure and not a precise determina- 50 -0.16 -0.35 -0.10 -0.29
tion of the 0(r) terms in the high-density expan- 100 -0.09 -0.23 -0.06 -0.20
sion. ]
For intermediate spin polarizations 0 & g & 1,
me used a standard interpolation formula, first were performed only for x, ~ 10, we adjusted his
proposed by von Barth and Hedin, a in which the parameters to give a better match to the high-
correlation energy has the same polarization de- density limit. This adjustment successfully pre-
pendence as the exchange energy: dicted the results of the more recent calculation
of Ceperley and Alder" for y, & 1.
e,(r„f) = e (r)+f(t)[~ (r,) —e,'(r, )], (C12) Ceperley and Alder4' have made a stochastic
)I.'(r, &) = u.'( )r+f(K) [I1.'(r, ) —)1.'(r, )] (C13) sampling of an exact solution of the Schrodinger
equation for the electron gas. Table XII also
+ [a P(r,) —«,U(r, ) ] [sgn(o) —1] —, gives the parameters me have determined by a

where
0.05 ) I —0.8
(C15) NIQ tlf

A more realistic interpolation formula, based on 0

the g dependence of the HPA correlation energy,
has recently been given by Vosko, %ilk, and Nu- 0.0a
Sair 117

The parameters we used are given in Table

XII. For the unpolarized electron gas, where
Ceperley's4' calculations mere performed for all
y, ~ 1, me used Ceperley's original parameters, o.oi— ~~
4~
tg
but for the polarized gas, where his calculations Q - —OP
I-
O

TABLE XII. Parameters used to fit the r~ dependence /Gunnaraaon- Lundqvist

of the correlation energy in an unpolarized (U} and / l ) I

polarized (P) uniform electron gas. Values of A and B 0 2

are given in Eqs. (C7) and (Cll). Parameters in a.u. DENSITY PARAMETERIs
Ceperley-AMer Ceperley FIG. 13. Correction to the RPA correlation energy
Parameter U P U P of the unpolarized electron gas as a function of r~. The
parametrized correlation energies considered are those
-0.1423 -0.0843 -0.1471 -0.0790 of Ceperley and Alder (Bef. 42, solid curve), Moruzzi,
1.0529 1.3981 1.1581 1.2520 Janak, and Williams (Bef. 12, dashedcurve), Gunnarsson
0.3334 0.2611 0.3446 0.2567 and Lundqvist (Bef. 40, dash-dotted curve), and signer
0.0020 0.0007 0.0014 0.0001 (Bef. 118, dotted curve). Ceperley values (Bef. 41, open
-0.0116 -0.0048 -0.0108 -0.0046 circles) are shown for comparison. The arrow shows
the exact high-density limit of the correction.
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. . . 5077

0.05 —0.8 values should be preferred in future work. Vosko,

Wilk, and Nusair"' have independently parame-
a trized the Ceperley-Alder results in a different
way, but for the r, values shown in Table XIII
K their numbers never deviate from ours by more
0.02 than 1.3/p relatively or by more than 0.0013 a. u.
CL ~ absolutely.
uJ
One way to assess the accuracy of the correla-
o.e & "
tion energy is to plot &,' —&, "vs x„as sug-
Moruzzi et al. gested by Vosko, %ilk, and Nusair. "'As ~ -0,
O
0.0I— this difference correctly approaches the second-
ir —0.2 order exchange constant (C9) in our parametriza-
O
tions, but not in some other parametrizations
hl Gunnarsson- that have often been used in density-functional
K Lundqvist
O calculations. In Figs. 13 and 14, we compare the
O I
high-density behavior of our parametrization of
0 4 Ceperley-Alder correlation with the parametriza-
DENSITY PARAMETER I'
s tions of signer, "' Gunnarsson and Lundqvist, '
FIG. 14. Same as Fig. 13, but for the polarized and Moruzzi, Janak, and Williams. '2 [The latter
electron gas. employs the Hedin- Lundqvist" parametrization
of the Singwi et gl. calculation" for the unpolar-
ized gas, and then uses the RPA scaling relation
careful fit to their numbers. For the densities (C10) to estimate &, for the polarized gas. ]
they considered (v, =l, 2, 5, 10, 20, 50, 100), our Recently Zabolitzky'" has obtained a numerical
parametrization never deviates from their results upper bound on the correlation energy of the in-
by more than 0.0001 a. u. In Table XIII we compare finite uniform electron gas; his results for 1&x,
the parametrized Ceperley-Alder correlation ener- &20 are greater than the Ceperley-Alder" values
gies with numerical RPA correlation energies. "' by about 5/a for the unpolarized gas, and about 15%
The difference between the Ceperley-Alder values for the polarized gas. Thus his results are con-
and Ceperley values we used in our calculations sistent with the Ceperley-Alder results, and
can be gleaned from Figs. 13 and 14; although the closer to them than to most other calculated cor-
difference is very small, the Ceperley-Alder relation energies.

E. Fermi and E. Amaldi, Accad. Ital. Rome 6, 119 O. Gunnarsson, J. Phys. F 6, 587 (1976); J. Callaway,
(1934). D. Laurent, and C. S. Wang, in Transition Metals,
J. C. Slater and J. H. Wood, Int. J. Quantum Chem. edited by M. J. G. Lee, J. M. Perg, and E. Fawcett
4, 3 (1971). (The Institute of Physics, London, 1977), p. 41.
N. W. Ashcroft and N. D. Mermin, Solid State Physics S. H. Vosko, J. P. Perdew, and A. H. MacDonald,
(Holt, Rinehart and Winston, New York, 1976). Phys. Rev. Lett. 35, 1725 (1975).
A. B. Kunz, Phys. Rev. B 12, 5890 (1975). A. Zunger and A. J. Freeman, Phys. Rev. B 15, 47
J. C. Slater, The Self-Consistent Field for Molecules (1977); 15, 5049 (1977); 17, 2030 (1978); Int. J.
and Solids (McGraw-Hill, New York, 1974). Quantum Chem. S11, 539 (1977); Bef. 31.
P. Hohenberg and W. Kohn, Phys. Rev. 136, B864 N. D. Lang, Solid State Phys. 28, 225 (1973).
(1964). R. Monnier and J. P. Perdew, Phys. Bev. B 17, 2595
W. Kohn and L. J.
Sham, Phys. Rev. 140, A1133 (1965). (1978); B. Monnier, J. P. Perdew, D. C. Langreth,
J. and J. W. Wilkins, Phys. Rev. B 18, 656 (1978); J. P.
'
U. von Barth and L. Hedin, Phys. C 5, 1629 (1972).
Also A. K. Rajagopal and J.
Callaway, Phys. Bev. B 7, Perdew and B. Monnier, J. Phys. F 10, L287 (1980).
1912 (1973). J. A. Appelbaum and D. R. Hamann, Rev. Mod. Phys.
O. Gunnarsson, B. I. Lundqvist, and J. W. Wilkins, 48, 3 (1976).
iSG. P.
Phys. Rev. B 10, 1319 (1974). Kerker, K. M. Ho, and M. L. Cohen, Phys. Rev.
O. Gunnarsson, J. Harris, and R. O. Jones, J. Chem. Lett. 40, 1593 (1978).
Phys. 67, 3970 (1977). A. Zunger, Phys. Bev. B 22, 959 (1980).
B. I. Dunlap, J. W. Connolly, and J. R. Sabin, J. 21
D. C. Langreth and J. P. Perdew, Solid State Commun.
Chem. Phys. 71, 4993 (1979). 17, 1425 (1975); Phys. Rev. B 15, 2884 (1977).
V. L. Moruzzi, J. F. Janak, and A. B. Williams,
2r
D. C. Langreth and J. P. Perdew, Solid State Commun.
Calculated Electronic ProPerties of Metals {Perga- 31, 567 (1979); Phys. Rev. B 21, 5469 (1980).
mon, New York, 1978). B. Y. Tong and L. J. Sham, Phys. Rev. 144, 1 {1966).
 J. P. PKRDK% AND ALEX XUNGKR

D. J. W. Geldart and M. Rasolt, Phys. Rev. 8 13, 1477 A 15, 2135 (1977).
(1976). OM. M. Morrell, R. G. Parr, and M. Levy, J. Chem.
K. Schwax'z, Chem. Phys. Ldtt. 57, 605-(1978). Phys. 62, 54e (19V5).
H. B. Shore, J. H. Rose, and E. Zaremba, Phys. Rev. Y. Tal, Phys. Rev. A 18, 1781 (1978); 21, 2186(E)
5

8 15, 2858 (1evv). (198O).

U. Rossler» in Rare gas 8g/ids, . edited by M. L. Klein ~2A. K. Theophilou, J.
Phys. C 12, 5419 (1979).
and J. A. Venables (Academic, New York, 1976), p. J. Katriel, J.
Phys. C 13, L375 (1980).
505. +S. M. Valone and J. F. Capitani, Phys. Bev. A 23, 2127
K. Hamx'in, G. Johansson, U. Gelius, C. Nordling, (1981).
and K. Siegbahn, Phys. Scr. 1, 277 {1970). 5~W. J. Hunt and W. A. Goddard, Chem. Phys. Lett. 3,
S. B. Trickey and J. P. Worth, Int. J. Quantum Chem. 414 {1e69).
s11, 529 (1evv). 56J. H. Rose and H. B. Shore, Solid State Commun. 17,
M. Boring, Int. J. Quantum Chem. Ss, 451.(1974).
- 32V (19V5).
3 A. Zunger and A. J. Freeman, Phys. Rev. 8 16, 2901 C. Edmiston and K. Ruedenberg, J. Chem. Phys. 43,
(1977) (and references therein); Phys. Lett. 57A; 453 S97 {1965).
(1976). 5
We are grateful to Lee Cole for performing this cal-
D E. Eastman F. J. Hlmpsel, and J- A. Knapp, Phys culation.
Rev. Lett. 44, 95 (1980). ~9C. Froese Fischer, The Hartree-I'os Method for
J. Harris and R. O. Jones, J. Chem. Phys. 68, 3316 A. toms (Wiley, New York, 1977).
(19V8). 60L. A. Hemstreet, Phys. Rev. 8 11, 2260 (1975).
A. Zunger, J. P. Perdew, and G. L. Oliver, Solid 6
F. Herman, J. P. Van Dyke, and I. B. Ortenburger',
State Commun. 34, 933 (1980). In Fig. 1 of this paper, Phys. Rev. Lett. 22, 807 (1969); F. Herman, I. B.
an error occurred in plotting the LSD splitting between Ortenburger, and J. P. Van Dyke, Int. J. Quantum.
the spin-up and spin-down potentials; the correct Chem. S3, 827 (1970); I. B. Ortenburger and F. Her-
splitting is 10 times smallex' than the one plotted. We Computationa/ Methods in Band Theory,
would like to thank L. Kleinman and K. Mednick for edited by P. M. Marcus, J. F. Janak, and A. R.
drawing our attention to this mistake. The exchange W'illiams (Plenum, New York, 1971), p. 469.
splitting in the SIC scheme for magnetic elements is A. Veillard and E. Clementi, J. Chem. Phys. 49, 2415
discussed in Sec III L in the present paper. (196S).
M. Levy, Proc. Natl. Acad. Sci. USA 76, 6062 (1979). 638. Y. Tong, Phys. Rev. A 4, 1375 (1971).
In Eq. (14) we are pursuing a hint" in the unpublished @J. P. Perdew, E. B. McMullen, and A. Zunger, Phys.
first draft of Ref. 35, to the effect that the kinetic Bev. A (in press).
energy can be defined as a functional of the density
6
M. Levy, S. C. Clement, and Y. Tal, in Chemical
in a way that admits fractional occupation numbers. Applications of Atomic and Mo/ecu/ar E/ectrostatic
See also M. Levy, Bull. Am. Phys. Soc. 24 (4), ab- Potentia/s, edited by P. Politzer and D. G. Truhlar
stract ED6 (1979). (Plenum, New York, 1981) (in press).
J. F. Janak, Phys. Rev. 8 18, 7165 (1978). 66E. Clementi and C. Roetti, At. Data Nucl. Data Tables
SThe remaining constraints 0 ~ f
a 1 make Eq. (23) 14, 1vv (1ev4).
@L. C. Green, M. M. Mulder, M. N. Lewis, and J. W.
an inequality. A possible objection to relaxing the
constraints, f,
=N is that, although T, U, and
were defined for noninteger numbers of electrons,
s, 6
Woll, Phys. Rev. 93, 757 (1954).
H. Hotop and W. C. Lineberger, J; Phys. Chem. Ref.
E~ was defined at Q-'E-U and Q was defined only for Data 4, 539 (1975).
69
integer numbers of electrons. The answer to this S. Jagannathan, Ph. D. thesis, University of Georgia,
objection is that any continuation of E~, differentiable Athens, Ga. , 1979 (unpublished).
in the range 0 ~ f ~ 1, will imply Eq. (23). D. W. Smith, S. Jagannathan, and G. S. Handler, Int.
J. Quantum Chexn. S13, 103 (1979).
R. G. Parr, R. A. Donnelly, M. Levy, and W. E. Palke,
J. Chem. Phys. 68, 3801 (1978); L. J. Bartolotti, S. R. ~~C. E. Moore, Natl. Bur. Stand. (U. S.) Bef. Data Ser.
Gadre, and R. G. Parr, J. Am. Chem. Soc. 102, 2945 34, 1 (1970).
(19so). ~2M. L. Cohen, M. Schluter, J. R. Chelikowsky, and
O. Gunnarsson and B. I. Lundqvist, Phys. Rev. 8 13, S. G. Louie, Phys. Rev. 8 12, 5575 (1975); J. A.
4274 (1976). Appelbaum and D. R. Hamann, Phys. Rev. 8 8, 1977
D. M. Ceperley, Phys. Rev. 8 18, 3126 (1978). {19V3}.
42D. M. Ceperley and B. J. Alder, Phys. Rev. I ett. 45, 73(a) A. Zunger and M. A. Rather, Chem. Phys. 30, 423
566 0.98O). (1978); (b) A. Zunger and M. L. Cohen, Phys. Rev.
R. S. Gadre, L. J. Bartolotti, and N. C. Handy, J. 8 20, 4082 (1979); (c) A. Zunger, J. Vac. Sci. Technol.
Chem. Phys. 72, 1034 (1980). 16, 1337 (1979).
+For example, F. E. Harris, Int. J. Quantum Chem. T. M. Wilson, Int. J. Quantum. Chem. S2, 269 (1968);
13, 189 (1978). 3, 757 {1970).
7a'
O. Gunnarsson, M. Jonson, and 8. I. Lundqvist, Phys. A. Zunger and M. L. Cohen, 'Phys. Rev. 8 20, 1189
Rev. 8 20, 3136 (1979). O. eve).
46P. S. Bagus, Phys. Rev. 139, A619 (1965). 76A. Zunger and A. J. Freeman, Phys. Rev. 8 17, 4850
4~A. R. Williams, R. A. de Groot, and C. B. Somxners, O. evs}; 8 1v,
2o3o (19vs).
J. Chem. Phys. 63, 628 (1975). YT
M«t» Meta/-Insulator Transitions (Taylor and
L. Hedin and A. Johansson, J. Phys. 8 2, 1336 {1969). Francis, London, 1974), p. 128.
M. Gopinathan, J. Phys. 8 12, 521 (1979); Phys. Rev. 8W. P. Menzel, C. C. Lin, D. Fouquet, E. E. Lafon,
 SELF-INTERACTION CORRECTION TO DENSITY-FUNCTIONAL. .. 5079

and R. C. Chancy, Phys. Rev. Lett. 30, 1313 (1973); I. Lindgren and K. Schwarz, Phys. Rev. A 5, 542
R. C. Chancy, E. E. Lafon, and C. C. Lin, Phys. Rev. (19V2).
B 4, 2734 (1971); G. E. Laramore and A. C. Sviten- L. Szasz, I. Berrios-Pagan, and G. McGinn, Z.
dick, ibid. 7, 3615 (1973); D. M. Drost and J. L. Fry, Naturforsch. 309, 1516 (1975).
ibid. 5, 684 (1972). J. R. Schrieffer, J. Vac. Sci. Technol. 13, 335 (1976).
L. Kleinman and K. Mednick, Phys. Rev. B 20, 2487 J. D. Talman and W. F, Shadwick, Phys. Rev. A 14,
(1979). 36 (19V6).
A. Zunger, Phys. Rev. B 21, 4785 (1980); J. Ihm and J. D. Talman, P. S. Ganas, and A. E. S. Green, Int,
M. L. Cohen, Solid State Commun. 29, 711 (1979); J. Quantum Chem. S13, 67 (1979).
Phys. Hev. B 21, 1527 (1980); J. P. Worth and S. B. B. H. Brandow, Adv. Phys. 26, 651 (1977).
Trickey, ibid. 11, 3658 (1975); A. Zunger and A. J. J. A. Alonso and L. A. Girifalco, Phys. Rev. B 17,
Freeman, ibid. 15, 5049 (1977); A. Zunger and M. L 3V35 (19V8).
Cohen, ibid. 19, 568 (1979); Ref. 29. G. W. Bryant and G. D. Mahan, Phys. Rev. B 17, 1744
R. T. Pool, J. Liesegang, R. C. C. Leckey, and J. G. (1978).
Jenkin, Phys. Rev. B 11, 5190 (1975). 6H. Stoll, C. M. E. Pavlidou, and H. Preuss, Theoret.
W. Hanke and L. J. Sham, Phys. Rev. Lett. 33, 582 Chim. Acta 149, 143 (1978); H. Stoll, E. Golka, and
(1974); 43, 387 (1979). H. Preuss, ibid. 55, 29 (1980).
.

W. Schafer and M. Schrieber, Solid State Commun. TH. Stoll, private communication.
32, 591 (1eve). H. B. Shore, E. Zaremba, J.
H. Rose, and L. Sander,
M. Bennett and J. C. Inkson, J. Phys. C 10, 987 (1977). Phys. Rev. B 18, 6506 (1978).
E. Wigner and F. Seitz, Phys. Rev. 43, 804 (1933); A. Zunger and M. L. Cohen, Phys. Rev. B 18, 5449
46, 509 (1934). (19V8).
A. Zunger (unpublished results). A. Zunger, Phys. Rev. B 22, 649 (1980).
J. Harris, Int. J. Quantum. Chem. S13, 189 (1979). J. P. Perdew, Chem. Phys. Lett. 64, 127 (1979).
C. Kittel, Igtroduction to Solid State Physics, 5th ed. S. M. Valone, J. Chem. Phys. 73, 1344 (1980).
(Wiley, New York, 1975), p. 74 (from data furnished A. J.Coleman, Rev. Mod. Phys. 35, 668 (1963).
by L. Brewer). ~~40. Gunnarsson, P. Johansson, S. Lundqvist, and
J. Harris and R. O. Jones, J. Chem. Phys. 70, 830 B. I. Lundqvist, Int. J. Quantum Chem. S9, 83 (1975).
(1979). ~~M. Gell-Mann and K. A. Brueckner, Phys. Rev. 106,
O. Gunnarsson and R. O. Jones (unpublished). 364 (1e5v).
O. Gunnarsson and R. O. Jones, Phys. Scr. 21, 394 S. Misawa, Phys. Rev. 140, A1645 (1965).
(1980). S. H. Vosko, L. Wilk, and M. Nusair, Can. J. Phys.
I. Lindgren, Int. J. Quantum Chem. 5, 411 (1971). 58, 1200 (1980).
R. Latter, Phys. Rev. 99, 510 (1955). E. P. Wigner, Phys. Rev. 46, 1002 (1934); Trans.
4F. Herman and S. Skillman, Atomic Structure Calcu- Faraday Soc. 34, 678 (1938).
latioms (Prentice-Hall, Englewood Cliffs, 1963).
~ OL.
Hedin and B. I. Lundqvist, J. Phys. C 4, 2064
R. D. Cowan, Phys. Rev. 163, 54 (1967). (1ev1).
6D. A. Liberman, Phys. Rev. 171, 1 (1968). K. S. Singwi, A. Sjolander, M. P. Tosi, and R. H.
I. Lindgren and A. Rosen, Technical Report GIPR-034, Land, Phys. Rev. B 1, 1044 (1970).
Institute of Physics, Chalmers Technical University, ~2~J. G. Zabolitzky (unpublished).
Goteborg, Sweden (unpublished).