PHYSICAL REVIEW B VOLUME 45, NUMBER 23 15 JUNE 1992-I

&ccnrate and simple analytic representation of the electron-gas correlation energy

John P. Perdew and Yue Wang'
Department of Physics and Quantum Theory Group, Tulane Uniuersity, New Orleans, Louisiana 70118
(Received 31 January 1992)
We propose a simple analytic representation of the correlation energy c, for a uniform electron gas, as
a function of density parameter r, and relative spin polarization g. Within the random-phase approxi-
mation (RPA), this representation allows for the r, behavior as r, ~ ~. Close agreement with nu-
merical RPA values for s, {r„0), s, {r„ 1 ), and the spin stiffness a, (r, ) = t) s, (r„(=0)/g't, and recovery
of the correct r, lnr, term for r, ~O, indicate the appropriateness of the chosen analytic form. Beyond
RPA, different parameters for the same analytic form are found by fitting to the Green's-function Monte
Carlo data of Ceperley and Alder [Phys. Rev. Lett. 45, 566 (1980)], taking into account data uncertain-
ties that have been ignored in earlier fits by Vosko, Wilk, and Nusair (VWN) [Can. J. Phys. 58, 1200
(1980)] or by Perdew and Zunger {PZ) [Phys. Rev. B 23, 5048 {1981)].While we confirm the practical ac-
curacy of the VWN and PZ representations, we eliminate some minor problems with these forms. We
study the g-dependent coefficients in the high- and low-density expansions, and the r, -dependent spin
susceptibility. We also present a conjecture for the exact low-density limit. The correlation potential
p, (r„g) is evaluated for use in self-consistent density-functional calculations.

I. INTRODUCTION sentation is definitely more accurate than the Vosko-

Wilk-Nusair (VWN-RPA) and Cole-Perdew' (CP)
An analytic representation of the uniform electron-gas forms, which assume p =1. Beyond RPA, we At to the
correlation energy e(r„g) is an essential ingredient of Green's-function Monte Carlo results of Ceperley and
most local' and nonlocal density functionals. An accu- Alder, as in the earlier parametrizations of Vosko, Wilk,
rate representation is needed, especially for the descrip- and Nusair (VWN) or Perdew and Zunger (PZ). While
tion of delicate magnetic effects. Here c, is the correla- we confirm the practical accuracy of the VWN and PZ
tion energy per electron; the density parameter r, and rel- forms, we also avoid some of the minor problems of those
ative spin polarization g are par ametrizations.
The PZ and CP forms for s, (r„g) display an artificial
r, =[3/4tr(n&+n{ )]' ', discontinuity of second and higher derivatives with
respect to r, at r, =1. In the high-density limit (r, ~O),
g=(n —n )/(n +n {),
& & &
(2) these forms do not recover the exact spin stiffness a, (r, ).
where n and n & are the up- and down-spin electron den- Furthermore, we note the exact relations
&

sities. (All equations are expressed in atomic units:

e =Pi= m =
l. ) The correlation energy s, has the small- co(g)=co (g), (6)

( expansion RPA(g)

E, (r„g) =E, (r„O)+ —,'a, (r, )g + (3)

where x2=0. 024179 hartree is the second-order ex-
the small-r, or high-density expansion change constant. ' The PZ form relies upon an older and
less precise constant (xz =0. 023 hartree).
s, (r„g)=co(g) lnr, c&(g)+c2(g—
)r, lnr, The VWN and VWN-RPA forms are analytically
complicated and nontransparent. In the high-density
c3(g)r, + . — limit, they are not constrained to reproduce the exact
' '
ci(0) and ci(1) (although these constants are fitted in a
and the large-r, or low-density expansion
least-squares sense). Furthermore, they artificially make
—do(g) di(g) c2(0)=c2(1)=0. For low electron densities, the VWN
2p 1/2 (5) form produces an unphysical maximum of the spin sus-
p
S S
ceptibility enhancement g/g&& at r, =50 (Sec. III). Final-
The exact correlation energy has p =1. The same ex-
ponent has been assumed ' '
to hold within the
ly, we note that the VWN fit does not employ a11 of the
Ceperley-Alder data (/=0 and I; r, = 1, 2, 5, 10, 20, 50,
random-phase approximation (RPA), but we have recent- 100), but only the data for r, ~ 10. Although the data set
ly found" that the correct RPA exponent is p = —, '. for r, &10 is less precise, its known uncertainty can be
This discovery has led us to reconsider the analytic taken into account.
representation of E, (r„g). Within the RPA, our repre- Our parametrization avoids these problems, and pro-

45 13 244 1992 The American Physical Society

45 ACCURATE AND SIMPLE ANALYTIC REPRESENTATION OF. . . 13 245

vides estimates of the g-dependent coefficients in the cally from RPA integrals. On the basis of this study,
high-density expansion of Eq. (4). Previous estimates of they proposed the spin-interpolation formula we will use:
these coeScients by Perdew and Zunger and by
Aguilera-Navarro,
stricted to (=0
Baker, and de Llano have been re-
and 1, and have not taken into account
E, (r„g}=s,(r„O}+a,(r, ) „(1—
( )
g }
the uncertainties of the Monte Carlo data. We also dis- + [E,(r„ 1 ) —E, (r„O}]f(g)g (8)
cuss the g-dependent coefficients of the low-density ex-
pansion (5). Finally, we report results for the spin suscep-
tibility enhancement y/yo as a function of r„where [(1+/) i +(1 —g) —2] (9)
F0=2. 589X10 /r, is the spin susceptibility for nonin- (24r3 2)
teracting electrons.
Note that f
(0) =0, (1)= 1, and f "(0)= l. 709 921.
(The PZ spin interpolation, taken from Ref. 2, amounts
f
II. PARAMETRIZED CORRELATION ENERGY
f
to a, (r, ) = "(0)[E,(r„ 1 ) — c., (r„O) ]. ) Instead
rather complicated analytic VWN form for s, (r„O),
of the
Vosko, Wilk, and Nusair made a careful study of the g
dependence of s, (r„g), which they evaluated numeri- s, (r„ 1 ), and —a, (r, ), we use the simpler form
I

G(r„A, a„p„pz, p3, p4, p)= —2A(1+a, r, ) ln 1+

1
(10)
2A(fj, r,'"+P,r, +I33r,'"+P4r,'+')
I

The parameters A, P, , and P2 are chosen to match the ex- Carlo e(r„O) and s, (r„l), and the uncertainties b, s, .
act high-density expansion (4), for which co and c1 are Thus we perform a "chi-square fit by minimizing "
known:
G(r„. . . ) —a (r, )
A =Cp, (19)
ba(r, }

P, =
1
exp( —c, /2co ), (12) i.e., we fit within the uncertainty for each r, . By using
2cp
more information than VWN, we hope to obtain a more
P2=2AP1 . (13) accurate representation of the correlation energy. The re-
sulting s, (r„ 1) is very close (within 0.2 mRy) to that of
We easily find VWN, but s, (r„O) and a, (r, ) show differences over the
C2= Acx) (14) range 0. 5 & r, & 100 as great as 1.0 and 2.2 mRy, respec-
tively.
= —2A a, ln(2AP1)— 132 133 The beyond-RPA spin stiffness a, (r, ) requires special
c3 (15)
131 P1 treatment because of the lack of exact or Monte Carlo
values to fit. We follow a proposal of VWN, fitting
The parameter p =' —,(RPA, Ref. 11) or 1 (beyond RPA) a, (r, ) to
expansion (5), for which
E, (r„ 1 ) —
fixes the low-density
s, (r„O)
=a1/P4 (16) (20)
do sRPA( r 1) sRPA( 0)
d1 a lf 3 /f 4 ' (17) at r, =2, 5, 10, and 20. However, we note that the spin
Since the exact RPA 10(g) is independent" of g, p =1 for susceptibility depends sensitively upon r, at large values
the RPA a, (r, ). The parameters a1, p3, and p4 are ad- of r, To fix t. he low-density behavior, we also fit a, (r, ) at
justed to give a "best fit" to numerical data for r, =2, 5, r, =75 to a value which makes the spin susceptibility
10, 20, 50, and 100. enhancement
Within RPA, we known s, (r„O), E, (r„l), and 1/3 2/3
—a, (r, ) exactly from the VWN numerical calculation. x = 1— —
r
+3
9~
4
r, a, (r, ) (21)
So we simply adjust a„p3, and p4 to minimize the sum of Xp
the squared errors

g ~G(r„. . . ) a(r, )~— diverge at a density (r, =77. 5) slightly lower than the
density (r, =73) at which the electronic spins spontane-
ously polarize (from /=0 to g= 1 ). This susceptibility
The resulting fit is remarkably accurate, with a maximum behavior, which is consistent with the approximate valid-
error over the range 0. 5 & r, & 100 of only 0.2 mRy, com- ity of assumption (20), is very similar to that found in
pared to 1.6 mRy for VWN-RPA. (2000 mRy= 1 har- RPA, where y/yp diverges at r, =19, and the ferromag-
tree. ) This level of accuracy gives us confidence in the netic transition occurs at r, =17. It is also similar to the
analytic form (10). behavior in the Hartree-Fock approximation (a, =O),
Beyond RPA, we know the Ceperley-Adler Monte where g/gp diverges at r, =6.03 and the ferromagnetic
13 246 JOHN P. PERDEW AND YUE WANG 45

TABLE I. Parameters of the fit to Eq. {10), and of the resulting high-density [Eq. {4)] and low-
density [Eq. (5)] expansions. Asterisks indicate parameters that are constrained to exact values from
Ref. S. (Energies are in hartree. )
RPA( r 0) RPA(r 1) ~RPA(» ) c, (r„1) —a, (r, )
0.75 0.75 1.00 1.00 1.00 1.00
A =co 0.031 091 0.015 545 0.016 887 0.031 091 0.015 545 0.016 887
a& 0.082 477 0.035 374 0.028 829 0.213 70 0.205 48 0. 111 25
Pi 5.148 6 6.486 9 10.357 7.595 7 14.118 9 10.357
P2 1.648 3 1.308 3 3.623 1 3.587 6 6. 197 7 3.623 1
133 0.236 47 0.151 80 0.479 90 1.638 2 3.366 2 0.880 26
P4 0.206 14 0.082 349 0. 122 79 0.492 94 0.625 17 0.496 71

c) 0.070 823 0.049 778 0.035 475 0.046 644 0.025 599 0.035 475
C2 0.002 56 0.000 55 0.000 49 0.006 64 0.003 19 0.001 88
C3 0.009 36 0.002 30 0.003 59 0.010 43 0.003 84 0.005 21
d 0.400 1 0.429 6 0.234 8 0.433 5 0.328 7 0.2240
d] 0.459 0 0.791 8 0.917 7 1.440 8 1.769 7 0.396 9

transition is found at r, = 5. 45. cal difFerence from the VWN (Ref. 5) and PZ (Ref. 7) for-
Table I shows the parameters of the fit. The asterisk mulas.
indicates parameters which have been constrained to ex- From these equations, we estimate the g-dependent
act values. Numerical results for s, (r„g) from the coeScients of the high-density expansion:
present Perdew-Wang (PW) parametrization within RPA
are compared to VWN-RPA (Ref. 5} and exact-RPA co(g) =0.03109 —0. 009 88f(g)(1 —P)
(Ref. 5) values in Table II of Ref. 11, while beyond-RPA —0. 015 55f (f)g (23)
PW results are compared to VWN (Ref. 5) and PZ (Ref.
7) values in Table III of the same reference. c, (g) =0.046 64 —0. 020 75f (g)(1 —P)
Our numerical RPA calculations for r, up to 10 show —0. 021 05f ( g) g (24)
that the low-density limit of E, (r„O) is
0. 40lr, ~ +0.4— 6lr„as the parameters of Table I sug- etc. Table II compares cc(g) from Eq. (23) to the exact
gest. An alternative and preferred set of parameters for cc(g), which we have recently evaluated, ' and also
(r„l) may be constructed from those for E, (r„O) displays c, (g) from Eq. (24).
via the Misawa spin-scaling relation' In Table III, we compare the coefficients c2(0), c3(0),
RPA( ] eRPA(2 —4/3 cz(l), and c3(1) obtained here (PW) with those from PZ
1) () ) (22)

The result is a, =0.032731, P&=6. 4868, Pz=1. 3083,

(Ref. 7) and from Aguilera-Navarro,
Llano, and with exact values. '' Baker, and de
The close agreement
=
p3 0. 118 24,
=
p4 0. 081 81; c 2 = 0. 000 51, c 3 = 0. 002 33, between the PZ and Aguilera-Navarro values may be a
do=0. 4001 d& =0. 5782. consequence of the fact that neither of those fits takes
The exact g dependence in the high-density limit could into account the uncertainties in the Monte Carlo data,
be recovered by replacing Eq. (9) by which are largest at small r, . ' The P% coeScients are
significantly closer to the exact values for c2,' exact values
f(g) = [I(g) —I ]/[I( I —1],
)
TABLE II. The g-dependent coefficients co{/) and c, {gl of
where I(g) is the analytic function of Eq. (32) of Ref. 14, the high-density expansion (4). (Energies are in hartree. ) The
making f"(0)= 1.086 28. However, this change would percent error of Eq. (23), relative to exact values from Ref. 11, is
not necessarily improve the fit for r, ~0. 5, where Eq. (9) also shown, as is the percent error of Eq. (24), relative to exact
actually gives a better description of the RPA correlation values from Ref. 19.
energy. For the sake of the low-density limit (discussed co{/) c, ( )
at the end of Sec. III}, we retain Eq. (9). [Eq. {23)] [Eq. {24)]
0.0 0.0311 (0.0% ) 0.0466 (
—0. 6%)
III. DISCUSSIQN OF RESULTS BEYOND RPA —0.0% ) 0.0459 —0. 6%)
0.2 0.0308 ( (

Equations (8) —(10) and Table I generate what may be 04 0.0297 (

—0.0% ) 0.0438 —
( 0. 6%)
the most accurate available representation of the correla- 0.6 0.0277 —
( 0. 1%) 0.0400 —
( 0. 3%)

tion energy c, (r„g). The corresponding correlation po- 0.8 0.0239 (

—2. 0%) 0.0343 (+0. ) 8%%uo

tential p, (r„g) is presented in Appendix A. We recom- 0.9 0.0206 (— 5. 0% ) 0.0305 (+2.6%)

0.999 0.0156 (
—2. 1%) 0.0257 (+2.6%)
mend the use of these formulas in density-functional and
1.0 0.0155 (0.0%) 0.0256 (— 0. 5%)
other calculations, although they may show little practi-
45 ACCURATE AND SIMPLE ANALYTIC REPRESENTATION OF. . . 13 247

TABLE III. More coefBcients of the high-density expansion (4). Note that cz =C~=O for the
VWN form of Ref. 5. (Energies are in hartree. )
=0
RPA
C2 C2 C3 cC2RPA C2 C3

pz' 0.0021 0.0020 0.0116 0.0005 0.0007 0.0048

Aguilera-Navarro, 0.0036 0.0118 0.0014 0.0049
Baker, and
de Llano
PW' 0.0026 0.0066 0.0104 0.0006 0.0032 0.0038
Exact 0.0027 0.0092' 0 0005' 0 0031'
' Reference 7 (RPA: Ref. 10).
b
Reference 8.
' Present work.
" Reference 15.
' Appendix B.

for c3 are unknown. the more physical behavior of the PW y/yo. The PZ,
We turn now to lower densities. Besides the correla- VWN, and PW values are all consistent with measured
tion energy s, (r„g}, the total energy per electron in- spin susceptibilities for the alkali metals's (3 (6).
(r,
cludes noninteracting kinetic In summary, we have presented a simple and useful an-
2/3 alytic form for the correlation energy s, (r„g). Tests
3
t, (r„g) =
9m.
[( 1 + g)5/3+( I g)5/3]/2 within the random-phase approximation show errors less
10r, 4 than 0.2 mRy for s, (r„O), s, (r„l), and a, (r, ) over the
range 0. 5&r, &100. Another positive indicator is the
(25)
correct value found for cz (Table III). Beyond RPA,
and exchange our representation also seems highly accurate, and may
1/3 be improved through parameter reoptimization when
s„(r„g)=— 4ar, 4
[(1+g) + (1 —g) ]/2
more precise Monte Carlo data (including correlation en-
ergies for 0(g(1}
become available. An appealing ave-
nue is to demand that Eq. (10) reproduce the exact do,
(26) leaving only two adjustable-fit parameters (at and P3).
contributions. The total energy s(r„O) minimizes at The exact d„, (g) is presumably 0.89593 from Eq. (27)
(see Table IV). We can achieve this low-density limit
r, =4. 19, close to the density of metallic sodium. The
ferrotnagnetic transition (from /=0 to /=1) in our pa- with do(/=0) =0.437 76, do(/= 1) =0.318 67, and
rametrization occurs at r, =73, close to the transition do( — a, ) =0.203 63. In fact, with further development of
density estimated by Ceperley and Alder (r, =75+5}. the high- and low-density limits, it may be possible to
construct e, (r„g) accurately with no Monte Carlo input.
Using the expression'
Recently, Hoffman' has made an "exact" numerical
0. 89593 evaluation of c, (g) of Eq. (4), with results that are rather
1.325 0. 365
rS
+ r3/2 r2S
(27) similar to those of our analytic representation (Table II).
S However, his values for /=0 and 1 [c, (0)=0.046921,
for the energy of the body-centered-cubic Wigner crystal, c, (1)=0.025 738 hartree], are slightly (0.5%) larger than
we find a second transition to the crystalline state at the values from Ref. 5 that we have used in our Table I.
r, =98, close to the density estimated by Ceperley and
Alder (r, = 100+20). TABLE IV. Spin dependence of the exchange-correlation en-
In the low-density limit (r, ~ao }, the energy of the ergy e„,(r„g)~ d„,(g)lr, in — the low-density limit r, ~Do,
fluid phase in our parametrization is — d„,(g)/r„where from Eq. (28). The exact d„,(g) is probably independent of g.
(Energies are in hartree. )
d„,(g) =0.4582[(1+/) +(1 —g) ]/2
d „,(g)
+0.4335 —0. 1310f(g)+0.0262f (g)P, (28)
0.0 0.892
which depends very weakly upon g (Table IV). The exact 0.2 0.891
d„,(g) is probably" independent of g. 0.4 0.890
Table V shows the spin susceptibility enhancement of 0.6 0.889
Eq. (21). Note the peculiar maximum of the VWN y/yo 0.8 0.891
(as pointed out to us by Clougherty) at r, = 50, well below 0.9 0.896
1.0
the ferromagnetic transition (at r, = 80 in VWN}, and also 0.906
13 248 JOHN P. PERDEW AND YUE WANG 45

TABLE V. Spin susceptibility enhancement of Eq. (21). The

Bc,(r„g) a, (r, )
RPA values (from our present parametrization) agree with the =4/'f(g) c, (r„ 1 ) —c, (r„O)— "0
numerical RPA values from Ref. 5.

XI'XO
RPA pz' PW' + f'(g) Pc, (r„ 1) —g c, (r„O)
0.5 1.08 1.08 1.08 1.08
a, (r, )
2
1 1.16
1.31
1.15
1.27
1.15
1.31
1.15
1.30
+(1 —P) "0 (A3)
3 1.48 1.39 1.46 1.44
4 1.66 1.51 1.62 1.58 where
5 1.86 1.62 1.79 1.72
6 2.08 1.73 1.97 1.86 4 [(1+()i/3 (1 g)1/3]
—2) (A4)
10 3.44 2. 15 2.83 2.43 (2
20 —26.77 3. 12 6.71 4. 16
30 —3.64 3.93 17.87 6.95 Finally, we obtain Bc, ( r„O) Ii)r„Bc,(r„ 1 ) /dr„and
50 —1.60 5.69 282.41 24. 52 da—
, (r, )/dr, from Eq. (10):
60 —1.31 6.72 83.29 63.01 I

75 —1.07 13.00 20. 15 930.61 = —2 A a, ln(1+ 1/Q, )— (A5)

' Reference 7.
Reference 5. where
' Present work. = —2/1 (1+a,r, ),
Qo (A6)

Q, =2~ (P, r,'"+P,r, +P, r,'"+P,r~+'), (A7)

Q I = /1 [P, r,
We have checked and confirmed the numerical part of
Hoffman's calculation, i.e. , the integral ( ln(R ) ) .
'
+2P2+3P3r, '/ +2(p+1)P4ri']
. . (A8)

By generalizing the derivation of Ref. 21 to (%0, or by

combining Eq. (34) of Ref. 22 with Eq. (Al), we find the
ACKNOWLEDGMENTS interacting or correlation contribution to the kinetic en-
ergy per electron:
One of us (J.P.P. ) acknowledges correspondence with
(1+() (r„g)
A. D. Becke, D. P. Clougherty, G. G. Hoffman, and S. H. t, (r„g) = )+ 3
, (r„g—
4c. p,
Vosko. This work was supported in part by the National
Science Foundation under Grant No. DMR88-17866.
—g)
+ (1 p, (r„g) . (A9)
2
APPENDIX A: CORRELATION POTENTIAL The simplest version of the virial theorem,
AND INTERACTING KINETIC ENERGY
t, + c + c,, = —( t, +t, ), .

The correlation potential for electrons of spin 0. is is obeyed at the equilibrium density (r, =4. 19 and (=0).
p, ( r„g) = 8[n c, (r„g) ]/Bn
APPENDIX B: EXACT COEFFICIENT
OF r, Inr, IN EQ. (4)
Carr and Maradudin' found c2 (0)=0.0027 hartree
—(g —sgno )
Bc, (r„g)
(A 1) and cz(0) — cz (0) =0.0065 hartree. We have evaluated
the numerical integral of their Eq. (15) to higher pre-
cision, with the result
where sgno. is + 1 for o = 1 and —1 for cr = $. From Eq.
(8), c~ (0) =0.002 709 5 hartree, (Bl)
and the numerical integral of their Eq. (22), with the re-
'dc, (r„g) Bc, (r„O) Bc,(r„ 1—
)
[1 f (g)0']+ f (0)0' sult

c2(0) —c~ (0) =0.006 519 7 hartree . (B2)

da, (r, ) f(g) The latter integral is greatly simpli6ed by our analytic ex-
(A2)
dr, f"(0) pression for R ' "(iu), which Carr and Maradudin left in
the form of a two-dimensional integral [Eq. (A5) of Du-
and Bois ']:
45 ACCURATE AND SIMPLE ANALYTIC REPRESENTATION OF. . . 13 249

R'"(iu)=4[(1+3u ) —(2+3u )u tan '(u)]/(1+u ) . change contribution

(B3) Et„(r„l)=e»(2 ~
r„0),
To find the RPA coefficient for /= 1, we use Eq. (22) with gives
the result c (1)—c (1)= [c (0) —c" (0)]/2 ~
(B5)
cRPA( 1) —cRPA(())/27/3 (B4) as in Ref. 5.
We hope to generalize these results to arbitrary g in fu-
A different spin-scaling relation' for the second-order ex- ture work.

'Present address: Department of Chemistry, University of ~2L. Onsager, L. Mittag, and M. J. Stephen, Ann. Phys.
North Carolina, Chapel Hill, NC 27599. (Leipzig) 18, 71 (1966).
W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 (1965). S. Misawa, Phys. Rev. 140, A1645 (1965).
~U. von Barth and L. Hedin, J. Phys. C 5, 1629 (1972). ' Y. %ang and J. P. Perdew, Phys. Rev. B 43, 8911 {1991).
3J. P. Perdew, Phys. Rev. B 33, 8822 (1986). The present (PW) W. J. Carr, Jr. and A. A. Maradudin, Phys. Rev. 133, A371
representation of e, (r„g}has been used by J. P. Perdew and (1964).
Y. Wang (unpublished), and by J. P. Perdew, J. A. Chevary, ' L. Hedin, Phys. Rev. 139, A796 (1965).
S. H. Vosko, K. A. Jackson, M. R. Pederson, D. J. Singh, ~7R. A. Coldwell-Horsfall and A. A. Maradudin, J. Math. Phys.
and C. Fiolhais {unpublished). 1, 395 (1960); W. J. Carr, R. A. Coldwell-Horsfall, and A. E.
4D. Singh, D. P. Clougherty, J. M. MacLaren, R. C. Albers, and Fein, Phys. Rev. 124, 747 (1961); F. Herman and N. H.
C. S. Wang, Phys, Rev. B 44, 7701 (1991). March, Solid State Commun. 50, 725 (1984).
~S. H. Vosko, L. Wilk, and M. Nusair, Can. J. Phys. 58, 1200 See Fig. 2 of D. Y. Kojima and A. Isihara, Phys. Rev. B 11,
(1980). 710 (1975), or Fig. 6 of J. P. Perdew and J. W. %'ilkins, ibid.
M. Gell-Mann and K. A. Brueckner, Phys. Rev. 106, 364 7, 2461 (1973).
(1957); B. S. Shastry, Phys. Rev. B 17, 385 (1978). G. G. Hoffman, Phys. Rev. B. 45, 8730 (1992).
~J. P. Perdew and A. Zunger, Phys. Rev. B 23, 5048 (1981). According to S. H. Vosko (private communication), the value
This form for s, (r„g}has been used by A. Zunger, J. P. Per- of (In(R}) reported by Gell-Mann and Brueckner (Ref. 6)
dew, and G. L. Oliver, Solid State Commun. 34, 933 (1980). was numerically refined in Ref. 5 to the same value found in
V. C. Aguilera-Navarro, G. A. Baker, and M. de Llano, Phys. Ref. 19. However, Ref. 5 employed the numerical value for 5
Rev. B 32, 4502 (1985). used by Gell-Mann and Brueckner, while Ref. 19 evaluated
D. M. Ceperley and B. J. Alder, Phys. Rev. Lett. 45, 566 this integral analytically.
(1980). M. Levy and J. P. Perdew, Phys. Rev. A 32, 2010 (1985).
L. A. Cole and J. P. Perdew, Phys. Rev. A 25, 1265 (1982). A. Gorling, M. Levy, and J. P. Perdew (unpublished).
Y. %'ang and J. P. Perdew, Phys. Rev. B 44, 13 298 (1991). D. F. DuBois, Ann. Phys. (N. Y.) 7, 174 (1959).