SOLID STATE PHYSICS. VOL .

54

Quasiparticle Calculations in Solids

G. AULBUR.LARSJONSSON.
WILFRIED AND JOHN W . WILKINS
Department of Physics
Ohio State University
Columbus. Ohio

I . Many-Body Effects in Computational Solid State Physics . . . . . . . . . . . . . . 2

1. Introduction and Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2. Quasiparticles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3. Kohn-Sham Particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
I1. Quasiparticle Calculations in the GW Approximation . . . . . . . . . . . . . . . . 12
4. The Quasiparticle Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
5 . The Hedin Equations and the GWA . . . . . . . . . . . . . . . . . . . . . . . 18
6. Separation of the Self-Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
7. Determination of the Single-Particle Green Function . . . . . . . . . . . . . 28
8. Determination of the Dynamically Screened Interaction . . . . . . . . . . . 29
9. Early Quasiparticle Calculations . . . . . . . . . . . . . . . . . . . . . . . . . 42
10. Local-Field Effects and the Nonlocality of the Self-Energy . . . . . . . . . . 51
11. Energy Dependence of the Self-Energy . . . . . . . . . . . . . . . . . . . . . . 57
12. Core-Polarization Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
13. Self-Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
14. Vertex Corrections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
I11. GWA Calculations: Numerical Considerations . . . . . . . . . . . . . . . . . . . . 89
15. Different Implementations of the GWA . . . . . . . . . . . . . . . . . . . . . 90
16. Plane Waves: Numerical Details . . . . . . . . . . . . . . . . . . . . . . . . . 102
17. Parallel GWA Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
18. GWA Calculations for Five Prototypical Semiconductors . . . . . . . . . . 120
IV . Semiconductors and Insulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
19. Bulk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
20. Superlattices-Interfaces-Schottky Barriers . . . . . . . . . . . . . . . . . . . 140
21. Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
22. Defects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
23. Pressure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
24. Excitons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
25. Atoms and Molecules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
V. Metals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
26. Bulk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
27. Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
28 . Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
VI . GWA Calculations and Optical Response . . . . . . . . . . . . . . . . . . . . . . . 172

1
ISBN O-LZ-M)7754-I Copyright C)Zoo0 by Academic Press
ISSN 0081-1947/00$30.00 All rights of reproduction in any form reserved
2 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

29. Overestimation of Optical Constants within DFT . ... . . ... . ... .. 175
30. The “Scissors Operator” and its Limitations . . . . ... . . ... . ... .. 179
31. Local-Field Effects in Optical Response . . . . . . . ... . . . . . . . . . .. 184
32. Density-Polarization Functional Theory . . . . . . ... . . ... . ... .. 191
VII. Excited States within Density Functional Theory . . . . . ... . . ... . . . . .. 195
33. Functionals Based on Ground-State Densities . . . ... . . ... . ... .. 196
34. Functionals Based on Excited-State Densities . . . .... . . ... . ... .. 201
35. Time-Dependent Density Functional Theory . . . .... . . ... . ... .. 203
36. Monte-Carlo Calculations . . . . . . . . . . . . . . .... . . ... . ... .. 205
Appendix: Density Functional Theory . . . . . . . . . . . . . . .. . . . . . . . . ... .. 207
1. Universal Density Functionals . . . . . . . . . . . . .... .. ... . ... .. 208
2. The Kohn-Sham System . . . . . . . . . . . . . . . . .... . . . . . . . . . .. 209
3. The Band-Gap Discontinuity . . . . . ; . . . . . . . .... ..... .... .. 21 1
4. The Exchange-Correlation Hole . . . . . . . . . . . .... ..... . . . . .. 213
5. Coupling-Constant Averages . . . . . . . . . . . . . . . . . ..... .... .. 215
6. Local Approximations . . . . . . . . . . . . . . . . . . . . . . .... .... .. 217

1. Many-Body Effects in Computational Solid State Physics

1. INTRODUCTIONAND OVERVIEW

During the last decade, quasiparticle calculations have been used success-
fully to describe the electronic excited-state properties of solids such as
single-particle band structures and absorption spectra. Under the assump-
tion that electronic and ionic degrees of freedom can be decoupled, the
problem is that of N electrons in a solid described by the following
Hamiltonian: ’

Here ri is the coordinate of electron i and Yxl is an external potential that

accounts for the interaction with the nuclei. The last term in the above
equation is the Coulomb interaction between the electrons, which correlates
the electrons’ motion. As a consequence, an exact description of the
many-electron problem requires the solution of an equation with 3N
coupled spatial degrees of freedom. For macroscopic systems, the number of
electrons N is on the order of Avogadro’s number ( % loz3),so a solution to
the N-electron problem must be approximate.

We use atomic units throughout this article unless otherwise noted. In these units, energy is
measured in Hartree and h = e = me = 4m0 = 1; E~ is the permittivity of vacuum.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 3

A successful approximation for the determination of excited states is

based on the quasiparticle concept and the Green function method. The
Coulomb repulsion between electrons leads to a depletion of negative charge
around a given electron, and the ensemble of this electron and its surround-
ing positive screening charge forms a quasiparticle. The mathematical
description of quasiparticles is based on the single-particle Green function
G, whose exact determination requires complete knowledge of the quasi-
particle self-energy Z2 The self-energy Z is a non-Hermitian, energy-
dependent, nonlocal operator that describes exchange and correlation effects
beyond the Hartree approximation.
A determination of the self-energy can only be approximate, and a
working scheme for the quantitative calculation of excitation energies in
metals, semiconductors, and insulators is the so-called dynamically screened
interaction or the GW approximation (GWA).3,4In this approximation, the
self-energy C is expanded linearly in terms of the screened interaction W:

which explains the name of the approximation. The GWA for the compu-
tation of quasiparticle energies was proposed by in 1965. However,
not until the mid-eighties was the approach applied to large-scale, numerical
electronic structure calculation^.^^^ The resulting ab-initio band structures
compare favorably with experiment.
Several reviews of quasiparticle calculations in the GWA have been
published. An early review of bulk and surface calculations in the GWA was
done by Hybertsen and Louie.’ Bechstedt’ discussed the physics of the
GWA in relation to model approaches for the calculation of the dielectric
response as well as the self-energy. Godby’ reviewed quasiparticle calcula-
tions for jellium, simple metals, and semiconductors. Mahan“ examined
different GW approximations resulting from the inclusion of self-energy and

* The self-energy I:is related to G via Dyson’s equation; see Section 11.14, Eq. (2.5).
L. Hedin, Phys. Rev. 139, A796 (1965).
L. Hedin and S. Lundqvist, in Solid State Physics, vol. 23, eds. F. Seitz, D. Turnbull, and H.
Ehrenreich, Academic, New York (1969), 1 .
’ M. S. Hybertsen and S. G. Louie, Phys. Rev. Lett. 55, 1418 (1985).
R. W. Godby, M. Schliiter, and L. J. Sham, Phys. Rev. Lett. 56, 2415 (1986).
’ M. S. Hybertsen and S. G. Louie, Comm. Cond. Mat. Phys. 13, 223 (1987).
* F. Bechstedt, in ~esrkiipeiprooble~eelAdvances in Solid State Physics, vol. 32, ed. U. Rossler
Vieweg, Braunschweig/Wiesbaden (1992), 161.
R. W. Godby, “Unoccupied Electronic States,” Topics in Applied Physics, vol. 69, eds. J. E.
Inglesfield and J. Fuggle, Springer, New York (1992).
l o G. D. Mahan, Comm. Cond. Matt. Phys. 16,333 (1994).
4 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

vertex diagrams beyond the random phase approximation. Mahan and

Plummer " discussed many-body effects in photoemission spectra of simple
sp-bonded metals. Pollmann et al." concentrated on GWA calculations for
semiconductor surfaces. FaridI3 considered mathematical aspects of
quasiparticle calculations. Recently Aryasetiawan and G u n n a r ~ s o n 'fo-
~
cused on strongly correlated d and f electron systems and local-orbital basis
function^.'^
The aim of the present review is to discuss (1) the physics and extensions
of the GWA (Section 11), (2) numerical aspects of GWA calculations
(Section 111), (3) applications of the GWA to semiconductors and insulators
(Section IV) and metals (Section V), .and (4) the relevance of GWA
calculations to optical response (Section VI). Particular importance is given
to semiconducting and insulating systems in a plane-wave basis. We
compare all published first-principles GWA calculations for five pro-
totypical semiconductors (Si, Ge, GaAs, Sic, GaN) and show that dif-
ferences between published quasiparticle calculations for the lowest conduc-
tion-band state can be as large as 0.5 to 1.0eV. We also present for the first
time parallel algorithms both for reciprocal and real-space/imaginary-time
GWA calculations. In addition, Section VII gives a brief overview of
alternative methods to determine excited states within density functional
theory.
The remainder of this section introduces the quasiparticle concept and
defines the Green function and the spectral function (Section 1.2). The latter
is important since it can be related to photoemission experiments. A short
introduction to density functional theory is also necessary, since most
current quasiparticle calculations start from density functional theory wave
functions and energies, and measure their success by the degree of improve-
ment of excited-state properties over the corresponding density functional
description. Section 1.3 provides the basic equations and physics of density
functional theory that are relevant for quasiparticle calculations. A more
detailed overview of the basic concepts of density functional theory is given
in the appendix.

G. D. Mahan and E. W. Plummer, to appear in Handbook of Surfaces, vol. 2, eds. K. Horn

and M. Schemer.
l 2 J. Pollmann, P. Kriiger, M. Rohlfing, M. Sabisch, and D. Vogel, Appl. SurJ Sci. 104-105, 1
(1996).
l 3 B. Farid, to be published in Electron Correlation in the Solid Slate, ed. N. H. March, World
Scientific/Imperial College Press, London, UK.
l4 F. Aryasetiawan and 0. Gunnarsson, Rep. Prog. Phys. 61, 237 (1998).
We call basis sets that contain functions explicitly centered at atomic sites (e.g., LMTO,
Gaussian orbitals, etc.) local-orbital basis sets in this article.
 QUASIPARTICLECALCULATIONS IN SOLIDS 5

2. QUASIPARTICLES

Dejnition of Quasiparticles. The excitations of a system of strongly inter-

acting particles can often be described in terms of weakly interacting
quasiparticles. In a solid, an electron, or “bare” particle, repels the other
electrons via the Coulomb potential and, in effect, surrounds itself with a
positively charged polarization cloud. The positive screening charge and the
bare electron form a quasiparticle that weakly interacts with other quasipar-
ticles via a screened rather than the bare Coulomb potential. The quasipar-
ticle lifetime is finite since quasiparticles are only approximate eigenstates of
the N-electron Hamiltonian in Eq. (1.1). The residual interaction between
the quasiparticles leads to a complex energy whose imaginary part is
inversely proportional to the quasiparticle lifetime. That the quasiparticle
concept works well in solid state systems-in spite of strong interactions
between the bare particles -is demonstrated by the success of one-particle
theories such as density functional theory in the local density approximation
or GWA in the description of the structural and electronic properties of
solids.
The energy difference between the quasiparticle and the bare particle is
usually described by the self-energy which must account for all
exchange and correlation effects beyond the Hartree approximation. The
self-energy is a nonlocal, energy-dependent, and in general non-Hermitian
operator, whose properties will be discussed in more detail in Sections 11.10
and 11.11. An exact determination of the self-energy for real systems is not
possible, since it contains all the complexities of the many-body system.
Instead, practical approximations to the self-energy-such as the dynami-
cally screened interaction or GWA, which is the topic of the present review
(see Section 11)-must be used. The central equation that governs the
behavior of quasiparticles is the so-called quasiparticle equation. Neglecting
spin degrees of freedom, this equation can be written as

[ - 2 V2 + V, + V,,,
I S
Yi(r) + C(r, r’; Ei)Yi(r’)dr’= EiYi(r) (1.3)

Here V, is the electrostatic or Hartree potential of the electrons, that is, with
n as the electron density,

l6 The self-energy equals the energy of the bare particle interacting with itself via the
polarization cloud that the particle generates in the many-body system.
6 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

V,,, is the external potential from the ions, and Y iand E i are the quasi-
particle wave function and energy, respectively. We will come back to the
quasiparticle equation and define Yiand Eiin Section 11.4.
The Green Function. Quasiparticle properties such as energies, lifetimes,
and expectation values of single-particle operators, such as the density and
the total energy of a many-body system, are determined by the single-
particle Green function (see, for instance, standard textbooks such as Refs.
17, 18, 19, 20, 21, and 22). The Green function G is also called the
single-particle propagator. With IN, 0) as the ground state of the N-electron
Hamiltonian in Eq. (l.l), Y(rt) = exp(ifit)Y(r) exp( - ifit) as the fermion
annihilation operator23 in the Heisenberg representation, Y t(rt) as the
corresponding creation operator, and T as the time-ordering operator, the
single-particle Green function is defined as

G(rt, r'r') = -i ( N , OlT[Y(rt)Y t(r'tr)] IN, 0 )

For t > t' (t' > t), G describes the propagation of a particle (hole) added to
the many-body system described by fi, that is, G describes the dynamics of
the N + N f 1 excitations in an N-electron system. G is a function of only
six spatial degrees of freedom and hence much more manageable than the
N-electron wave function, which depends on 3N spatial degrees of freedom.
Many of the complexities of the ground-state wave function are eliminated
by taking the expectation values.
The imaginary part of the Green function determines the spectral func-
tion A,

A(r, r'; E ) = n- 'IIm G(r, r'; E)I,

which is closely connected to photoemission spectra. Quasiparticles are

" A. A. Abrikosov, L. P. Gorkov, and I. E. Dzyaloshinski, Methods of Quantum Field Theory

in Statistical Physics, Dover, New York (1975).
" A. L. Fetter and J. D. Walecka, Quantum Theory of Many-Particle Systems, McGraw-Hill,
New York (1971).
l 9 P. Fulde, Electron Correlations in Molecules and Solids, Springer Series in Solid-state
Sciences, vol. 100, Springer Verlag, Berlin (1991), Chap. 9.
' O J. C. Inkson, Many-body Theory of Solids, Plenum Press, New York (1984).

" G. D. Mahan, Many-Particle Physics, 2nd. ed., Plenum Press, New York (1993).
'' R. D. Mattuck, A Guide to Feynman Diagrams in the Many-Body Problem, Dover, New
York (1992).
2 3 %(rt) destroys an electron at point r and time t.
 QUASIPARTICLECALCULATIONS IN SOLIDS 7

identified with narrow peaks in the interacting spectral function which

contain a significant amount of spectral strength. The peak position and
width determine the quasiparticle energy and inverse lifetime; the area under
the peak equals the quasiparticle weight. Figure 1 shows a schematic picture
of an interacting and a noninteracting spectral function. The interacting
spectral function has an approximate pole at I? = B + ir. If Z is the quasi-
particle weight and $(E) a smooth function at I?, then the Green function
can be expressed as

Assuming Z to be real and neglecting the smooth background Im $(E), one

obtains the corresponding spectral function as

r
A(E) 2z n - l z (1.8)
( E - Elz + r2'

Fourier transformation of Eq. (1.7) into the time domain leads to an

EE'CL E
FIG. 1. Schematic representation of the spectral function A ( E ) (Eq. (1.8)) for a noninteract-
ing and an interacting many-body system. The differences between the two cases are (1) the
real part E of the quasiparticle energy E = E + iT is shifted with respect to the bare energy E;
(2) the quasiparticle acquires a finite lifetime ljr due to interaction compared to the infinite
lifetime of the noninteracting particle; and (3) the spectral weight Z (shaded area) of the
quasiparticle peak is less than unity due to redistribution of spectral weight into the incoherent
background (Im+(E)), whereas the spectral weight of the bare particle is unity. The chemical
potential is p.
8 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

exponential decay of G cc exp( - r t ) , which identifies r as the inverse

quasiparticle lifetime.24The smaller r is, the longer the quasiparticle lives
and the sharper the corresponding peak in A(E). The quasiparticle weight
Z equals the area under the L ~ r e n t z i a n it
; ~(1)
~ must be on the order of
unity for the quasiparticle to be clearly identifiable in the spectrum, and (2)
is less than unity since A(E) is positive definite and normalized to unity.
Connection to Experiment. Quasiparticle energies and lifetimes are mea-
sured experimentally by direct or inverse photoemission, which removes or
adds one electron to the system and corresponds directly to the definition
of the Green function in Eq. (1.5). In direct photoemission, a photon with
energy hv impinges on a sample. An electron from an occupied band with
energy E,, absorbs the photon energy and becomes a photoelectron whose
kinetic energy Ekinallows the determination of the initial-state (valence)
energy. From the schematic picture of the direct photoemission process in
Fig. 2 and measuring the valence- and conduction-band energies with
respect to the vacuum level, we obtain the energy balance Ekin= E,, + hv,
which gives EVE. Inverse photoemission, the complementary process to
direct photoemission, probes the energy of unoccupied states E,, by
injecting an electron into a solid. The electron loses its kinetic energy &in
via photon emission before it comes to rest at a point of lower energy in the
conduction band. With hv as the energy of the emitted photon, a simplified
energy balance of the process reads Ekin= Ec, + hv. This relation gives E,,
since Ekinand hv are measured. A schematic description of inverse photo-
emission is given in Fig. 2.
The photocurrent in photoemission experiments is closely related to the
single-particle spectral function. The intrinsic photoemission spectrum, that
is, the photoemission spectrum that takes only many-electron scattering into
account, is expected to be reliable for the determination of the quasiparticle
peak, or the quasiparticle energy.26*27.28 However, the intrinsic spectrum
neglects matrix element effects, phonon and defect scattering, the in-
homogeneous surface potential, and other complications. For the determi-
nation of, for example, the quasiparticle lifetime, these additional effects
must be included (see Refs. 26,29, and 30 and references therein).

24 Quasiparticle lifetimes in Si are discussed in connection with Fig. 15.

” Typical values of Z for semiconductors are listed in Table 9.
2 6 L. Hedin, Nucl. Instr. Meth. Phys. Res. A m , 169 (1991).

” L. Hedin, Int. J . Quant. Chem. 56,445 (1995).

’* S. Hufner, Photoelectron Spectroscopy, Springer Series in Solid-state Sciences, Springer,
Berlin (1996).
2 9 J. Fraxedas, M. K. Kelly, and M. Cardona, Phys. Rev. 843, 2159 (1991).

30 N. V. Smith, P. Thiry, and Y. Petro5, Phys. Rev. B47, 15476 (1993).

 QUASIPARTICLE CALCULATIONS IN SOLIDS 9

-P I

%
photon -> electron: photoelectrunspectroscopy
electron o photon: Inverse photocmlrJion

hv

c
\ VB

photoekctron inverse
spectroscopy photoemlulon
N->N-1 N->N+1

FIG.2. (a) Schematic representation of an (inverse) photoemission experiment; (b) schematic

representation of the excitation process in a photoemission (left) and inverse photoemission
(right) experiment. The energy of the incoming (outgoing) photon is hv; the electron kinetic
energy Eki,, E,,, is the band gap; E , the electronegativity; and E+ the photothreshold energy
of the sample. Direct photoemission measures the quasiparticle properties of occupied bands
and decreases the total number of electrons N in the system. An incoming photon ejects a
valence-band (VB) electron out of the sample. The electron energy is measured and the
valence-band energy can be obtained via E,,, = E,, + hv. Inverse photoemission measures the
properties of quasiparticles in unoccupied bands. An incoming electron of energy E,,, impinges
on the sample and loses energy via emission of a photon hv. The electron reaches an
energetically lower available conduction state leading to E,,, = E,, + hv.

PARTICLES
3. KOHN-SHAM

All GWA calculations start from a suitably chosen one-particle Hamiltonian

whose eigenfunctions and eigenvalues are used to construct the single-
particle propagator G, the screened interaction W and the self-energy C, as
will be detailed in Section 11. The independent-particle Hamiltonian of
 10 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

choice31 is a density functional (DFT) Hamiltonian in the local density

approximation (LDA).32v33Density functional theory describes ground-
state properties and in particular the ground-state density, in principle,
exactly, but it does not describe excited states accurately. The local density
approximation to DFT is a practical starting point for quasiparticle
calculations since (1) it describes the density of metals, semiconductors, and
insulators accurately; and (2) exchange and correlation are described by a
local potential. Point (1) suggests that LDA ground-state wave functions are
good approximations for quasiparticle wave functions. This is often the case
not only for ground states but also for excited states, as is further discussed
in Section 11.7. Point (2) leads to computationally efficient quasiparticle
calculations compared to quasiparticle calculations based on a nonlocal
independent-particle Hamiltonian such as Hartree-Fock (see Section 11.7).
The Kohn-Sham formulation of density functional theory34 maps the
problem of N interacting electrons onto a system of N noninteracting,
fictitious particles- the Kohn-Sham particles -which move in an effective
potential V,,. The Kohn-Sham potential V,, is constructed to ensure that
the ground-state density of the noninteracting, fictitious system equals the
ground-state density n,(r) of the interacting system. Let mi(r) denote an
orbital of the fictitious particles and E~ the corresponding eigenvalue of the
Kohn-Sham Hamiltonian. With V,,, and V, as defined earlier, the relevant
single-particle equations are3’

[ 2 1
N
- Vz + V, + V,,, + V,, mi = E ~ @ ~ , no =
i=l
IQil2. (1.9)

Here V, is the exchange-correlation potential, which is obtained as a

functional derivative of the exchange-correlation energy Ex,:

(1.10)

3 1 Other independent-particle Hamiltonians such as Hartree and Hartree-Fock or empirical

pseudopotential Hamiltonians have been used as well but to a much lesser extent. See, for
instance, Sections 11.7 and 11.9.
” P. Hohenberg and W. Kohn, Phys. Rev. 136, B864 (1964).
3 3 W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 (1965).
34 See the appendix for a brief review of the Kohn-Sham theory and other important concepts

of density functional theory.

3 5 In a crystalline system, i stands for the combination of a band index n, a spin index s, and

a wave vector k.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 11

and all potentials are evaluated at the ground-state density no@). The
exchange-correlation energy, which contains all Coulomb correlation effects
beyond the Hartree approximation and a part of the kinetic energy of the
interacting electrons, is not known explicitly for real systems.36 Practical use
of the Kohn-Sham equations requires good approximations to the ex-
change-correlation energy and, via Eq. (l.lO), to the exchange-correlation
potential.
The local density approximation replaces the inhomogeneous exchange-
correlation energy density per particle at a point r by the exchange-
correlation energy density per particle of a homogeneous electron gas,
U,h:m(r), evaluated at the local density. The total exchange-correlation
energy is then obtained as the integral over all local contributions:

(1.11)

The resulting exchange-correlation potential is local and energy indepen-

dent. Despite its simplicity, the local density approximation and extensions
that use gradient corrections to the local density -the generalized gradient
approximations -successfully describe ground-state properties of atoms,
molecules, and solids. This success and also several failures of LDA are
reviewed extensively in the literature; see, for instance, Refs. 37 and 38.
Structural properties of solids such as the lattice constant, the bulk modulus,
and the cohesive energy are generally determined to within a few percent of
the experimental value.
The Band-Gap Problem. In the derivation of Kohn-Sham theory,33 the
eigenvalues E~ in Eq. (1.9) enter as Lagrange parameters that ensure the
orthogonality of the orbitals (Di of the fictitious particles. As a consequence,
the E;S and (D;s must be considered as mathematical tools that contain no
relevant physical information besides the fact that the square of the
eigenfunctions sums up to the exact local ground-state density. In particular,
there is no formal justification that links the eigenvalues gi to the energy
dispersion of quasiparticles in a solid. A notable exception is the highest
occupied eigenvalue E ~ Sham
. and K ~ h identified
n ~ ~ E~ with the chemical
potential p of a metal. For semiconductors and insulators, Perdew, Parr,

36 For a formal expression of the exchange-correlation potential, see, for instance, L. J. Sham,
Phys. Rev. B 32, 3876 (1985), and references therein.
3 7 R. 0.Jones and 0. Gunnarsson, Rev. Mod. Phys. 61, 689 (1989).
38 Theory of the Inhomogeneous Electron Gas, eds. S. Lundqvist and N. H. March, Plenum
Press, New York and London (1983).
39 L. J. Sham and W. Kohn, Phys. Rev. 145, 561 (1966).
12 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Levy, and Balduz4' proved that E~ equals the negative of the ionization
energy.41
Perhaps the most prominent discrepancy between LDA band structures
and experiment is the fact that LDA underestimates the band gap of
semiconductors and insulators by about 0.5 to 2.0eV. In the case of Ge,
LDA leads to a semimetal rather than to an indirect-gap semiconductor.
Figure 3 demonstrates the underestimation of experimental band gaps in
LDA for all semiconductors and insulators for which ab-initio GWA
calculations have been reported. Figure 3 also shows that GWA calculations
largely correct the LDA band-gap underestimation and are in good agree-
ment with experiment. In spite of the bind-gap underestimation, LDA wave
functions are often good approximations to quasiparticle wave function^.'^
In the absence of quasiparticle calculations, LDA energies are routinely
used to interpret experimental spectra. LDA energy dispersions are often in
fair agreement with experiment, and in some'cases the LDA band gap can
be empirically adjusted to fit the experimental gap. This approach implies
an interpretation of the LDA exchange-correlation potential as an approxi-
mate self-energy that neglects nonlocal, energy-dependent, and lifetime
effects. Although the LDA band structure cannot claim quantitative accu-
racy for the determination of the electronic structure of solids, LDA
generally provides a qualitative understanding.

II. Quasiparticle Calculations in the GW Approximation

This section has four purposes. (1) The basic equations that govern the
dynamically screened interaction approximation are introduced (Sections
11.4 and 11.5) but not derived. We refer the reader to standard textbooks, for
example, Ref. 20, and the review articles by Hedin and Lundqvist4 and
Aryasetiawan and G u n n a r ~ s o n in
' ~ particular for a derivation of the Hedin
equations (Section 11.5). (2) Useful separations of the self-energy as well as
basis-set-independent details about the evaluation of the single-particle
propagator G and the screened interaction W are described in Sections 11.6,
11.7, and 11.8. (3) Section 11.9 gives a historical overview of early GWA
calculations and related approaches. (4) The physics of the self-energy
operator is analyzed in Sections 11.10, 11.11, 11.12, 11.13, and 11.14, with
Sections 11.13 and 11.14 focusing on the consistency of GWA calculations
and on extensions to the GWA, respectively. An overview of important
equations for the GWA and symbols used in this article is given in Tables
1 and 2.

40 J. P. Perdew, R. G. Parr, M. Levy, and J. L. Balduz, Phys. Rev. Lett. 49, 1691 (1982).
41 For a more detailed description, see the appendix.
 -2 0 2 4 6 a 1
Si; 42
Ge; 43
dlamond; 44
3c Sic; 42
UCI; 44
gcN.45
Bdf46
B P 46
BAS: 46
AIN; 47
AIP 48
A&; 43
AISb; 48
GaN; 42
GaP 48
GaAs; 43
GaSb; 48
InP 48 Expt., indirect gap
InAs: 48
InSb; 48 0 Expt.. direct gap

ZnS; 51
2% 51
ZnTe; 51
CdS; 52
CdSe; 51
CdTe; 51
%
! z
NIO; 55
CaCuO . 5 4
6:
u 56
Z h * ; 57
sno,; 58

FIG. 3. Comparison of characteristic direct and indirect LDA, GWA, and experimental energy gaps for all semiconductors and insulators for which
first-principles GWA calculations have been reported. GWA corrects most of the LDA band gap underestimation over more than one order of
magnitude in the experimental band gap. The values for MnO, ZnO, and CaCuO, are from model-GWA calculations, which are accurate to within
0.4eV. The discrepancy between GWA and experiment for LiO, results from the neglect of excitonic effects. The experimental value for BAS is tentative.
e
The references for the LDA, GWA, and experimental values are listed after the element symbols. W
14 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

TABLE1. DEFINITIONS
OF SYMBOLS
AND QUANTITIES

SYMBOL DEFINITION

(spatial time) Coordinate of first particle

Chemical potential, Fermi energy
Fermion annihilation or creation operator
Quantum number, band index n, and wave vector
k for solids
Single-particle energy
N-electron total energy for state i
ith quasiparticle energy
ith quasihole energy
Wave function of single-particle Hamiltonian
ith excited (i = 0 ground) state of N-electron system
Quasiparticle many-body amplitude
Quasihole many-body amplitude
Single-particle Green function
Spectral function
Bare Coulomb interaction
Electron density
Density fluctuation operator
Hartree potential
External perturbation potential
Total induced potential
Irreducible polarizability
Full polarizability
Dielectric matrix
Dynamically screened interaction
Non-Hermitian, nonlocal, energy-dependent self-energy
=exchange + energy-dependent correlation self-energy
=Coulomb hole + screened-exchange self-energy
Vertex function
Fluctuation potential
Energy of electron-hole excitation

42 W. G. Aulbur and J. W. Wilkins, unpublished.

43 E. L. Shirley, X. Zhu, and S . G. Louie, Phys. Rev. B56, 6648 (1997).
44 M. S. Hybertsen and S . G. Louie, Phys. Rev. B34, 5390 (1986).

4 5 J. L. Corkill and M. L. Cohen, Phys. Rev. 848,17622 (1993).

46 M. P. Surh, S. G. Louie, and M. L. Cohen, Phys. Rev. B43,9126 (1991).

47 A. Rubio, J. L. Corkill, M. L. Cohen, E. L. Shirley, and S . G. Louie, Phys. Rev. B48, 11810

(1993).
48 X. Zhu and S . G. Louie, Phys. Rev. B43, 14142 (1991).

49 A. Rubio and M. L. Cohen, Phys. Rev. 851,4343 (1995).

5 0 S. Massidda, R. Resta, M. Posternak, and A. Baldereschi, Phys. Rev. B 52, R16977 (1995).

0. Zakharov, A. Rubio, X. Blase, M. L. Cohen, and S . G. Louie, Phys. Rev. B50, 10780
(1994).
.52 M. Rohlfing, P. Kriiger, and J. Pollmann, Phys. Rev. Lett. 75, 3489 (1995).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 15

TABLE2. IMPORTANT
EQUATIONS
FOR THE GWA

Interacting Green function (Eq. (2.2)) G(r, r'; E) = 1Yi(r)YT(r')

E-Ei
-
=jcm A(r, r'; E')
dE'

Go(r, r'; E ) = C
mi(r)@f(r')
Independent-particle Green function
i E-Ei
(Eq. (2.4))

Dyson's equation (Eq. (2.5)) G(r, r'; E ) = Go(r, r'; E )

+ Sj ,,
Go(r, r,; E)X(r r,; E)G(r ,, r'; E)dr,dr,

Quasiparticle equation (Eq. (1.3)) [ - V2 + V,

I S
+ Kx, Yi(r)+ X(r, r'; Ei)Yi(r')dr'

First-order perturbation theory for E i

(Eq. (2.6))

Quasiparticle weight Zi (Eq. (2.7))

Dynamically screened interaction

approximation (GWA)
(Eqs. (2.1 1)-(2.13))

Independent-particle polarizability
0%. (2.21))

Dielectric matrix in RPA (Eq. (2.23))

Dynamically screened interaction

0%. (2.16))

C C V,(r) V:(r')Qi (r)@:(r')

occ
Energy-dependent correlation XC(r,r';E) =
contribution to the self-energy mfo E +E,,-E~-~S
(Eq. (2.25)) V, (r) V:(r')@i(r)@:(r')
Lo
""OSE

+ E-E,,-Ei+i6

Static Coulomb-hole self-energy XCoH(r,r') = 46(r -r')[W(r, r'; E = 0)- dr, r')]
(Eq. (2.32))
OEE

Static screened-exchange self-energy XSEX(r,r')= - 1 @i(r)@f(r')W(r,r'; E = 0)

(Eq. (2.33))

Note: Equation numbers within parentheses equal those in the text.

 16 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

4. THEQUASIPARTICLE
EQUATION

The physical relevance of the Green function6’ (Eq. (1.5)) can be made clear
by expressing G in terms of quasiparticle wave functions and energies via its
spectral function A. Consider a complete set of eigenstates of the many-body
Hamiltonian for a system with N + 1 or N - 1 particles. Let p be the
chemical potential and denote the quantum number of the (N + 1)-particle
or (N - 1)-particle states with i. With EN,i as the energy of the N-electron
system in state i (i = 0 for the ground state), one defines the quasiparticle
amplitude Yi(r) and the quasiparticle energy Ei as4
Yi(r) = (N,Ol’&)lN + l,i), Ei = EN+l,i- E N , O for Ei 2 p,
(2.1)
Yi(r) = ( N - 1, &r)IN, o ) , E~ = E ~- ,E,- ~ for E~ < p.

As mentioned above, the quasiparticle amplitudes correspond to a nonlocal,

energy-dependent, non-Hermitian Hamiltonian. They fulfill the complete-
ness relation but are unnormalized and linearly d e ~ e n d e n tIntroduction
.~ of
the complete set of eigenstates in the definition of G given in Eq. (1.5) and
Fourier transformation to energy space61 identifies the quasiparticle ener-
gies as the poles of the Green function,

The integral defines the interacting spectral weight function,

A(r, r’; E ) = 1Yi(r)Y?(r’)6(E - Ei),

i
(2.3)

and the integration contour C runs infinitesimally above the real ,Y-axis for
E‘ < p and infinitesimally below for E’ > p.

53 U. Schonberger and F. Aryasetiawan, Phys. Rev. B52,8788 (1995).

54 S. Massidda, A. Continenza, M. Posternak, and A. Baldereschi, Phys. Rev. B 55, 13494
(1997).
” F. Aryasetiawan and 0. Gunnarsson, Phys. Rev. Lett. 74, 3221 (1995).
56 S. Albrecht, G. Onida, and L. Reining, Phys. Rev. B55, 10278 (1997).
” B. Kralik, E. K. Chang, and S. G. Louie, Phys. Rev. B57,7027 (1998).
’13 M. Palummo, L. Reining, M. Meyer, and C. M. Bertoni, 2 n d International Conference on
the Physics of Semiconductors, vol. 1, ed. D. J. Lockwood, World Scientific, Singapore (1995),
161.
59 See Sections 11.7 and 11, and VI.30 for further discussions of this point.
6o Sections 11.4 and 11.5 follow Ref. 4 closely. However, we do not consider the spin degrees
of freedom explicitly.
We add infinitesimal convergence factors f is to Ei for Ei 2 p.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 17

Neglect of exchange and correlation effects beyond the Hartree approxi-

mation leads to the noninteracting single-particle Hamiltonian fro, whose
corresponding Green function Go(r, r’; E) describes the propagation of a
particle in a system of N + 1 noninteracting particles. With the complete set
of orthonormalized single-particle wave functions Oi(r) and real, indepen-
dent-particle energies q, we write the spectral representation of Go as
follows:

Go(r, r‘; E ) = 1Qi(r)@.T(r’)

E - E ~
Ao(r, r‘; E’)
E - E dE. (2.4)

The”independent-particle spectral function can be expressed either as in Eq.

(2.3), with Y i being replaced by mi, or in the basis of the orthonormalized
Qi. In the latter case, A’ reduces to a &function, A? = 6(E - q).
The time development of the interacting Green function is determined by
the quasiparticle self-energy via Dyson’s equation.” With G and Go as in
Eqs. (2.2) and (2.4), we have”

G(r, r’; E ) = Go(r, r’; E ) +

ss Go(r, rl; E)C(r,, r,; E)G(r,, r’; E)dr,dr,,

where C(r, r’; E ) is the nonlocal, energy-dependent, non-Hermitian self-

(2.5)

energy introduced in the previous section. As mentioned earlier, C accounts

for all exchange and correlation effects beyond the Hartree approximation.
The above equation can also be written symbolically as G - ’ = (Go)-’ - C.
Instead of determining the quasiparticle energies indirectly as poles of the
Green function, it is more convenient to obtain these energies as solutions
of the quasiparticle equation. The quasiparticle equation can be derived by
inserting the spectral representation of the interacting Green function into
Dyson’s equation and is given in Eq. (1.3). It is formally similar to the
single-particle equation of Kohn-Sham theory (Eq. (1.9)), but the solutions
of the quasiparticle equation, that is, the quasiparticle energy Ei and wave
function Y i ,are physically meaningful rather than mere mathematical tools
as is the case in DFT.
Whenever possible one chooses as a starting point a suitable independent-
particle Hamiltonian whose wave functions mi(r) are nearly identical with
the quasiparticle wave functions Yi(r), which choice allows the determina-
tion of quasiparticle energies via first-order perturbation theory. Assume
+
go= -V2/2 + V, V,,, + V,, as in Eq. (1.9) with (Qi1Yi)z 1. Then the
quasiparticle energy must be determined self-consistently from
18 WILFRIED G . AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

where the quasiparticle weight Zi is defined as

The linear expansion of the self-energy with respect to energy needed to

derive the approximate relationship in Eq. (2.6) is well justified, as will be
detailed in Section 11.1 1. Most current quasiparticle calculations determine
quasiparticle energies via Eq. (2.6) once a suitable approximation for the
self-energy has been found.

5. THEHEDINEQUATIONS
AND THE GWA

In principle, the exact self-energy can be obtained from a closed set of

integro-differential equations -the Hedin equation^^,^ -that, in conjunc-
tion with the Dyson equation, link the single-particle propagator G, the
self-energy X, and the screened interaction W to the irreducible polarizabil-
ity P and the vertex function r, which will be defined now. Consider the
application of a small perturbation SV,,, to the many-body system. The
irreducible polarizability P is defined as the change in the density n upon a
+
change in the total (external Hartree) field 6V = SV,,, + SV,:

641)
P(1,2) = -
SV(2) *
Here 1 is a short notation for a combined space and time coordinate.
Similarly, the vertex function r is given by the variation of the inverse Green
function with respect to a change in the total potential or, alternatively, by
the variation of the self-energy with respect to SV:

+
With the above definitions, 1 = (r ,t1 S), 6 > 0 infinitesimal, and u(l,2) as
+

bare Coulomb interaction, Hedin's equations are

Z(1,2) = i
s G(1,4)W(1+,3)r(4, 2; 3)d(3,4),

s
J
(2.10)
P(1,2) = - i G(2,3)G(4,2)r(3, 4; l)d(3,4),
 QUASIPARTICLE CALCULATIONS IN SOLIDS 19

From these equations, the quasiparticle self-energy can be determined

iteratively, as is shown schematically in Fig. 4,panel (a).
The simplest, consistent version of the Hedin equations sets the vertex
function to unity and expresses the self-energy as the product of the
self-consistent single-particle propagator G and the self-consistent dynami-
cally screened interaction W! The GWA is consistent in the sense that it is
a particle- and energy-conserving approximation- in other words, a con-
serying approximation in the Baym-Kadanoff ~ e n s e . ~The **~ ~ corre-
GWA
sponds to the first iteration of the Hedin equations; that is, higher-order
vertex corrections are not included (for a discussion of vertex corrections see
Section 11.14) and can be interpreted as the first-order term of an expansion
of the self-energy C in terms of the screened interaction. The equations

G. Baym and L. Kadano5, P h p . Rev. 124,287 (1961).

G. Baym, Phys. Rev. 127, 1391 (1962).

FIG. 4. (a) Schematic representation of the iterative determination of the self-energy Z using
Hedin’s equations (Eqs. (2.10)) in conjunction with Dyson’s equation (Eq. (2.5)). Entries in
boxes symbolize the mathematical relations that link C, G,r, P, and W Starting with Z = 0
leads to an RPA screened interaction W and subsequently to C = GWRPA.(b) Schematic
representation of the self-consistent determination of the self-energy in the GWA. P, U: and Z
are constructed starting from an LDA or Hartree-Fock independent-particle propagator.
Subsequently, Z updates the quasiparticle wave functions and energies and a new Green
function G is determined. This process is repeated until self-consistency is reached. Most
practical applications either determine only the quasiparticle energies self-consistently or do not
update quasiparticle energies and wave functions at all. Self-consistency of GWA calculations
is discussed in Section 11.13.
20 WILFRIED G. AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

governing the GWA are

C(l,2) = iG(l,2)W(1+,2), (2.11 )

W( 1, 2) = U( 1, 2) +
s
P(1, 2) = -iG(l, 2)G(2, 1).
W( 1,3)P(3,4)~(4,
2)d(3, 4), (2.12)

(2.13)

The dynamically screened-interaction approximation, or G WA, in principle

requires a self-consistent determination of the single-particle propagator G
and the screened interaction W as shown schematically in Fig. 4, panel (b).
In practice, such a determination is computationally expensive and has been
done only rarely as discussed in Section 11.13. Further approximations for
the determination of G and W are described in Sections 11.7 and 11.8.
The dynamically screened interaction W introduces energy-dependent
correlation effects. In particular, we analyze in Section 11.6 the self-energy in
terms of its bare-exchange and energy-dependent correlation or polariz-
ati01-1~~contribution. The bare-exchange contribution to C results from the
bare Coulomb interaction, u(r, r’) = l/lr - r’J,whereas the energy-dependent
correlation contribution results from W - u. Dynamic effects in the screen-
ing process are important since moving quasiparticles drag their polariz-
ation cloud behind them. As a consequence, dynamic screening is less
efficient than static screening and directly affects quasiparticle energies
(about a 20% effect, as noted, for example, in Ref. 44; see Section 11.1 1).
The dynamically screened interaction approximation can be applied to
weakly polarizable materials with a dielectric constant close to unity and to
strongly polarizable solids such as the semiconductors Si and GaAs with
dielectric constants of order ten. In the former case, the GWA reduces to
Hartree-Fock theory, which is known to work well for systems such as
atoms or large-band-gap, low-polarizability solids like rare-gas solids and
ionic crystal^.^' In the latter case, the GWA roughly halves the Hartree-
Fock gap and is close to experiment. The energy-dependent correlation
correction to Hartree-Fock theory systematically lowers conduction-band
energies and raises valence-band energies, as will be shown in detail in
Section 11.6.

64 Both names are used interchangeably in the literature. We will use the name “energy-

dependent correlation contribution” for the W - u contribution to the self-energy throughout

this article.
‘’ U. von Barth and L. Hedin, Nuouo Cimento 23B, 1 (1974).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 21

Limitations o f t h e GWA. Several important failures of the GWA have

been pointed out: (1) For core-level spectra in atoms, strong electron-
electron interaction breaks down the quasiparticle picture when single- and
double-core holes are nearly degenerate.66 (2) s .+ d promotion energies for
the second half of the iron series show large discrepancies with experiment
because of the onset of strong 3d electron pairs (see Ref. 67 and Section
IV.25a). (3) Exchange splittings in Ni are about a factor two larger than
experiment because of the existence of strong 3d hole pairs (see Refs. 68 and
69 and Section V.26~).(4) The GWA satellite spectrum is poor. In the case
of Ni, the 6-eV satellite is missing (Refs. 68, 69, 70, and 71) since the GWA
does not capture strong correlations between 3d holes. The GWA plasmon
satellite of a core electron is 50% too low at 1.50,,, where wpf is the plasmon
frequency, rather than at o ~ Even ~ for. the~homogeneous
~ electron gas, the
GWA yields a single plasmon satellite (Refs. 27, 73, 74, and 75) rather than
a series of shake-up peaks.76 The same observation holds for nearly-free-
electron metals such as Na and Al.77
The neglect of vertex corrections (see Section 11.14) such as short-ranged
particle-particle interactions -that is, ladder diagrams -causes the failures
of RPA-based GWA calculations. Improvements for absorption spectra can
be obtained by going beyond RPA-based GWA calculations and including
vertex corrections in the dielectric matrix and the self-energy, as is detailed
in Sections II.9b, 11.14, and IV.24. Improvement of satellite spectra and
exchange splittings has recently been obtained by ab-initio cumulant expan-
s i o n ~and~ ~by an ab-initio T-matrix approach.’l For a review of these
approaches, see, for instance, Ref. 14.

66 G . Wendin, Breakdown of the One-Electron Pictures in Photoelectron Spectra, Springer,

Berlin (1981), 24.
67 E. L. Shirley and R. M. Martin, Phys. Rev. B47, 15404 (1993).
6 8 F. Aryasetiawan, Phys. Rev, B 4 6 , 13051 (1992).
6 9 F. Aryasetiawan and U. von Barth, Phys. Scripta T45, 270 (1992).

70 A. Liebsch, Phys. Rev. B 2 3 , 5203 (1981).

7 ’ M. Springer, F. Aryasetiawan, and K. Karlsson, Phys. Rev. Lett. 80, 2389 (1998), and

references therein.
72 L. Hedin, B. I. Lundqvist, and S. Lundqvist, J . Res. Natl. Bur. Stand. Sect. A 74A, 417 (1970).
7 3 L. Hedin, Phys. Scripta 21, 477 (1980).
7 4 D. C. Langreth, Phys. Rev. E l , 471 (1970).

7 5 P. Minnhagen, J . Phys. C: Sol. State Phys. 8, 1535 (1975).

76 The term shake-up spectra corresponds to excitations created by a sudden change in a
quantum-mechanical system. Consider, for example, the creation of a core hole in a solid via
a photoemission process. The electron cloud will contract around the core hole to screen it,
which leads to excitation or ionization of the residual ion- the so-called shake-up and
shake-off processes.
7 7 F. Aryasetiawan, L. Hedin, and K. Karlsson, Phys. Rev. Lett. 77, 2268 (1996).
22 WILFRIED G . AULBUR, LARS JONSSON, AND JOHN W. WILKINS

The Dielectric Matrix. The screened interaction W can also be expressed

in terms of the inverse dielectric matrix E - ’ , which describes screening in a
solid when local fields due to density inhomogeneities and many-body
effects are taken into account. Rather than using the integral equation, Eq.
(2.12), W can be determined as a convolution of the inverse dielectric matrix
with the bare Coulomb interaction in real space:

W(r, r’; o)=

s E-

The definition of the spatial Fourier transform as

‘(r, r”;w)u(r”, r‘)dr”. (2.14)

(2.15)

with q as a vector in the reciprocal-space Brillouin zone and G,G’ as

reciprocal lattice vectors, allows the transformation of the above convol-
ution in real space into a simple multiplication in reciprocal space:

(2.16)

The off-diagonal elements, EGGs(q;a),G # G’, describe screening caused by

an inhomogeneous density distribution, that is, the local-field effects” (see
Sections 11.10 and VI.31).
Expressing the inverse dielectric matrix in terms of density fluctu-
a t i o n ~ allows
~ ~ . ~a ~simple interpretation of corrections to Hartree-Fock
theory by the energy-dependent correlation contribution to the self-energy
(see also next section). Let IN, 0) (IN, m)) be the ground (excited) state of
an N-electron system; A’, the density fluctuation operator:
A’(rt) = W ( r t ) 4+(rt) - ( N , OIW(rt)+(rt)lN, 0); (2.17)
n,(r) = ( N , mlA’(r)lN,0), a density fluctuation; and E, = EN,,- E,,,, an
electron-hole energy. The inverse dielectric matrix is given as27*79

E-’(T, r’; t ) = 6(r - r’)s(t) - i

J 0)dr”
u(r - r”)(N, OlT[A’(r”t)A’(r’O)]lN,

(2.18)

7 8 Screening etTects due to inhomogeneous density distributions are in this article simply called
local-field efects. Modifications of screening due to exchange and correlation beyond the
Hartree approximation are called many-body local-field efects.
79 Ref. 21, Chapter 5.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 23

or, after Fourier transform, as27

E- ‘(r, r‘; o)= 6(r - r’) +

s u(r - r”) C 2~, n, (r”)n%(r’)
,,,+o 0 - ( E m - i6)2

In actual calculations the time-ordered dielectric matrix rather than its

dr“. (2.19)

inverse is determined from the irreducible polarizability P. The irreducible

polarizability is connected to the dielectric matrix via

E(r, r’; 0)= 6(r - r’) -

s u(r, r”)P(r”, r’; w)dr”.

To describe screening in solids in the time-dependent Hartreee or random

(2.20)

phase approximation (RPA), P is replaced by the independent-particle

polarizability Po, which can be obtained via the Adler-Wiser formal-
ism.80981,82 If Qi(r) and E~ are as in Eq. (2.4) and fi are the corresponding
Fermi factors, the independent-particle polarizability is given by

(2.21)

In reciprocal space one finds for the independent-particle polarizability with

the factor two accounting for spin:

2
P&,(q; o)= - C (il exp(i(q
v ii’ + G ) r)li’)(i’l exp( - i(q + G ) .r’)li)
+ ( E ~ .- E~ -o + i6
and for the RPA dielectric matrix:

4n:
EgL!(q; 0)= 6 ~ -~1qs
~ + G 1 2 p:G(q; 4. (2.23)

To be consistent with the RPA-based GWA, the energies E~ should in

S. L. Adler, Phys. Rev. 126,413 (1962).

M. S. Hybertsen and S. G. Louie, Phys. Rev. B35, 5585 (1987).
’’ N. Wiser, Phys. Rev. 129, 62 (1963).
24 WILFRIED G . AULBUR, LARS JONSSON, AND JOHN W. WILKINS

principle be calculated from a Hartree Hamiltonian. However, an LDA

spectrum is generally used instead for convenience, as described further in
Section 11.7.

6. SEPARATION
OF THE SELF-ENERGY

The energy-dependent GWA self-energy is a product of the propagator G

and the screened interaction W in real space as a function of time (see Eq.
(2.1 l)), which turns into a convolution in frequency space:

2x
eiE”G(r,r’; E + E‘)W(r, r‘; E’)dE‘. (2.24)

This equation is used below in conjunction with Eq. (2.4) for the indepen-
dent-particle propagator to derive three possible, approximate (Go is used
instead of G ) ways of separating the full self-energy C into physically
meaningful pieces.
Energy-Dependent Correlation + Bare Exchange. The energy-dependent
correlation contribution Zc(E) describes self-energy effects beyond the
bare-exchange or Hartree-Fock contribution Cx.Subtracting the Hartree-
Fock exchange potential from the self-energy operator and using Eq. (2.19)
for the inverse dielectric matrix leads to an expression of Cc(E) in terms of
fluctuation potentials, V,(r) = 1u(r, r’)n,(r’)dr’, electron-hole energies, E , =
EN,, - EN,,, and single-particle energies, E ~ : ”

2x
exp(iE‘G)[ W ( r , r’; E’) - u(r - r’)]Go(r,r’;E + E’)dE‘

=?so
OCC
V, ( r )V,*(r’)@ (r)@T(r’)
E+E,-EEi-i6
V, ( r )V,* (r’)@ (r)@T(r’)
+YCc i ,+o E-&,-~~+i6 ’
(2.25)

The expectation value of Cc with respect to single-particle orbitals Ok= I k )

and 0,= Il) equals

Energy-dependent correlations decrease the Hartree-Fock band gap by

raising the valence-band energy and lowering the conduction-band energy,
 QUASIPARTICLE CALCULATIONS IN SOLIDS 25

as can be seen by considering the above equations. For a valence electron

in the highest occupied state, Iub), the largest contribution to
(ublCC(E = E J u b ) comes from the sum over the occupied states. For these
states IE,, - E J is small compared to the electron-hole energy E,, which is
on the scale of the Hartree-Fock energy gap. Hence, the polarization
contribution to the self-energy shifts the valence-band maximum upward in
energy. For the conduction-band minimum, Icb), the largest contribution to
(cblCC(E = E,,)lcb) comes from the unoccupied states-that is, the second
term in Eq. (2.26) -and hence leads to a reduction in the conduction-band
energy.
The correlation contribution to the self-energy of an electron in a highly
excited Rydberg state reduces via Eq. (2.26) to the classical Coulomb energy
of the Rydberg electron in the field of the induced core d i p ~ l e . ~ Assume
*~~.~’
that the density fluctuations react instantaneously to the presence of the
highly excited electron. The Fourier transform of the density fluctuations is
therefore independent of frequency, which translates to E - E~ = 0 in Eq.
(2.25). Summing over i and using completeness gives the following express-
ion for the self-energy:

CC(r,r‘; E = 0) z -6(r - r’) 1 K(r)J‘,*(r)

r n + ~ Em
a
--

2r4
6(r - r’), (2.27)

where c1 is the dipole polarizability of the ion core around which the
Rydberg electron cycles. The last approximate equality follows from a dipole
expansion of the Coulomb potential inside the expression of the fluctuation
potential in which one keeps only the lowest order, that is, dipole terms. The
above expression equals the classical Coulomb energy of the Rydberg
electron caused by the field of the induced dipole, which is adiabatically
switched on in the ion core. The GWA recovers the relevant classical limit
for this special case. Further examples whose essential physics is contained
in the GWA are (1) the energy loss per unit time of a fast electron in an
electron gas and (2) the self-energy shift of a core electron in a solid. These
limiting cases are detailed in Refs. 4, 26, and 27 and are not discussed here.
+
Coulomb Hole Screened Exchange (COHSEX). The COHSEX approx-
imation is a physically motivated separation of the self-energy into a
Coulomb-hole (COH) part and a screened-exchange (SEX) part whose
static limit (1) has been used extensively to correct Hartree-Fock band
structures; (2) produces direct band gaps to within 20% of experiment but
gives a less reliable account of indirect band gaps (see Section 11.9); and (3)
allows an examination of local-field effects on, and the energy dependence
of, the self-energy. Consider the convolution in energy space of G and W
given by Eq. (2.24). This convolution can be determined formally by
26 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

introducing the spectral function B(r, r’; E ) of the screened interaction W in

analogy to the spectral representation of the Green function in Eq. (2.2).
One can express the screened interaction in terms of B(r, r’; E ) - B is an
antisymmetric function of energy -as4
2E’B(r, r’; E‘)
W(r, r’; E ) = u(r, r’) + dE‘, (2.28)
E 2 - (E’ - i6)2
where 6 = 0’. The real part of the self-energy C can then be written as the
sum of a Coulomb-hole contribution CCoH arising from the poles of the
screened interaction and a screened-exchange contribution CSEXarising
from the poles of the Green function:
ReC + CSEX.
= CCoH (2.29)
Using the independent-particle Green function G o (Eq. (2.4)), one can show
that ( P stands for principal part)

CCoH(r,r’; E ) = 1Oi(r)@?(r’)P
i

OCC
:j dE
B(r, r‘; E’)
E-q-E’
(2.30)

CSEX(r,r’; E ) = - 1Oi(r)@?(r’)Re(W(r, r’; E - q)), (2.31)

i

which illustrates the interpretation of CSEXas a dynamically screened

exchange interaction.
The static COHSEX approximation assumes, in contrast to the GWA,
that the screened interaction is instantaneous, that is, that retardation effects
can be neglected. Hence, W is proportional to a &function in time and equal
to a constant in energy space. In the static COHSEX approximation, the
Coulomb-hole contribution to the self-energy equals the interaction of an
electron with the induced potential due to rearrangement of the other
electrons. The screened-exchange contribution equals the statically
screened-exchange energy of the quasiparticle. To see this, one takes the
limit of E - E~ + 0. In Eq. (2.31) this limit implies the neglect of E - E, in
comparison to the poles of the screened interaction W which are given by
the plasmon energies. This approximation should be reasonable for states
close to the Fermi energy, which is verified by actual calculations (see
Section II.9a). One finds

CCoH(r,r’) = +S(r - r’)[W(r, r’; E = 0) - u(r, r’)], (2.32)

 QUASIPARTICLE CALCULATIONS IN SOLIDS 27

With these equations, the interpretation of CCoHand CSEXis clear. The

factor 1/2 in CCoHresults from the adiabatic turn-on of the interaction.
Core + Vulence. Under the assumption that one can energetically and
spatially separate core electrons from valence electrons in a material, the
self-energy equals the sum of the bare core-exchange potential, the screened
core-polarization potential, and the valence electron self-energy,all of which
will be defined below. Separating the Green function, the screened interac-
tion, and the irreducible polarizability into core and valence contributions,

G = G, + G,, w = w, + w,, P = P, + P,, (2.34)

the self-energy X in the GWA can be expressed as

C = GW = G,W + G,W, + G,W, x G,W + G,W,P,W, + G,W,. (2.35)

Here, we assume that the core polarizability is negligible in comparison to

the valence polarizability, and we expand W in terms of the core polariza-
bility:

W = W , + W , P , W , + W , P , W , P , W , + - ~ ~ ~ WW,P,W,.
u+ (2.36)

The three terms in Eq. (2.35) can be identified as the exchange potential
from the core (G,W x G,u due to the large energy denominators involved,
compare Refs. 83,84, and 85), as the screened polarization potential from
the core ( G , W,P, W,), and as the self-energy of the valence electrons ( G , W,).
A local density approximation of the bare core-exchange and screened
core-polarization potential leads in general to small errors except for
deviations on the order of 0.3 to 0.4eV in the band gap of materials with
large, soft cores whose lowest conduction-band state is localized on the ionic
cores. The nonlocal bare core-exchange and the screened, energy-dependent
core-polarization potential are generally small, as discussed, for example, by
Hybertsen and Louie.44Estimates for atomic Na4 or solid AlS6 indicate that
both terms contribute approximately 1 eV, relative to the bottom of the
valence band, to the quasiparticle energies. In GWA calculations based on
the LDA the proper core-valence terms are replaced by an LDA exchange-
correlation potential, which leads to a much smaller error. For states that
are localized on large, soft cores, such as the r2, conduction-band state in

83 W. Brinkman and B. Goodman, Phys. Rev. 149, 597 (1966).

84 L. Hedin, Arkiu Fysik 30, 231 (1965).
” J. C. Phillips, Phys. Rev. 123,420 (1961).
86 G. Arbman and U. von Barth, J . Phys. F 5, 1155 (1975).
28 WILFRIED G . AULBUR, LARS JONSSON, A N D J O H N W. WILKINS

Ge, the local, energy-independent approximation to the core-valence ex-

change and correlation breaks down and leads to states that are sys-
tematially too low in energy. Better treatments of core-valence exchange and
correlation via (1) use of core-polarization potentials and (2) explicit
inclusion of core states in the valence band are described in Section 11.12.

7. DETERMINATION GREEN
OF THE SINGLE-PARTICLE FUNCTION

Most current GWA calculations do not attempt a numerically expensive,

self-consistent calculation of G and W but determine good approximations
for the single-particle propagator and the screened interaction separately.
That is, these calculations adopt a “best G, best W’ philosophy. Once the
“best” G and the “best” W are chosen the self-energy is determined via Eq.
(2.11) without further iteration. The main task is to find a single-particle
Hamiltonian fi0 whose wave functions and energies result in a good
single-particle Green function (Eq. (2.4)) and a good screened interaction
(Eqs. (2.14), (2.20), and (2.22)). The common choice for fi0 is LDA or
Hartree-Fock. Although this strategy does not correspond to a consistent
determination of self-energy corrections starting from Hartree theory,” it
has been applied with considerable success for the determination of band
structures of semiconducting and metallic materials (see Sections IV and V).
Calculated quasiparticle corrections to an LDA band structure generally
agree well with experiment and have the additional advantage that the LDA
potential is local -allowing numerically inexpensive calculations -and
that the LDA wave functions are close to quasiparticle wave functions.
Many-body corrections to the LDA Hamiltonian are determined by the
expectation value of the operator C(r, r’; E ) - Vx.(r)6(r - r’). An overlap
close to unity between LDA and quasiparticle wave functions has, for
example, been reported for Si.44*90As a consequence, the self-energy
operator is for all practical purposes diagonal in the LDA basis, which can
be motivated by Eq. (2.26). Diagonal matrix elements (klClk) contain sums
over positive definite matrix elements l(kl Vmli)I2, whereas off-diagonal
matrix elements (klCII), k # 1, contain sums over terms proportional to
(kl Vmli)(il V,*l1) -that is, over terms with varying phases- which poten-

The determination of Z starting from Hartree theory requires self-consistency -the very
complication that the “best G, best W” philosophy tries to avoid-since Hartree band
structures and wave functions for semiconductors and insulators are qualitatively wrong. See
Refs. 88 and 89.
R. Daling and W. van Haeringen, Phys. Rev. B40, 11659 (1989).
89 R. Del Sole, L. Reining, and R. W. Godby, Phys. Rev. B49,8024 (1994).
90 F. Gygi and A. Baldereschi, Phys. Rev. B34, 4405 (1986).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 29

tially cancel.*' Note, however, that overlaps small compared to unity have
been observed, for instance, in transition metals68 and transition-metal
oxides (see Refs. 50, 54, 55, and 91). The disadvantage of using an
independent-particle Hamiltonian in the LDA is that density functional
theory is nonperturbative. Systematic improvements of GWA calculations
based on LDA Hamiltonians are only feasible once self-consistency is
achieved.
In contrast to LDA, Hartree-Fock calculations provide a good single-
particle basis only for weakly polarizable materials and are numerically
expensive due to the nonlocality of the exchange kernel. Hartree-Fock
Hamiltonians were used as independent-particle Hamiltonians in the 1970s
and 1980s to study self-energy corrections in rare earth and ionic solids (see
Section II.9a). Current uses of Hartree-Fock wave functions and energies as
input for GWA calculations are limited to the study of trans-polyacetylene
(see Section IV.25b).

OF THE DYNAMICALLY
8. DETERMINATION SCREENED INTERACTION

The determination of the dynamically screened interaction and the dielectric

matrix in reciprocal space as a function of frequency is numerically expen-
sive and can be drastically simplified by (1) modeling EGG,(q; w = 0) via
model dielectric matrices, and (2) modeling the frequency dependence of
EGC'(q; w) via plasmon-pole models. The computational efficiency of model
dielectric functions and matrices is offset by a loss in accuracy, which limits
the use of model dielectric functions to systems whose quasiparticle energy
shifts are larger than the accuracy of the model. The approximation of the
imaginary part of the inverse dielectric matrix as a function of frequency by
a simple pole, which is the common approximation made in plasmon-pole
models (see Fig. 5), is less severe. As discussed below and in connection with
Table 21, plasmon-pole models are accurate to within a few hundredths of
an eV for states close to the Fermi level and to within a few tenths of an eV
for states whose energy is on the order of the plasmon energy of the given
material. This statement is true if the static dielectric matrix, which is an
input parameter to all plasmon-pole models, is determined from first-
principles using D F T rather than from model dielectric matrices.
This review focuses on the results of accurate first-principles calculations
of the self-energy Z since these calculations form a consistent framework for
the discussion of self-energy effects. Results using model dielectric functions

91 S. Massidda, A. Continenza, M. Posternak, and A. Baldereschi, Phys. Rev. Lett. 74, 2323

(1995).
30 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Om2 t
0.0

e
-0.2 -
-
4-0 ..
s

n
---- CPP Model
Walter and &hen
’.
3 -0.4 .

‘
n *
3
v - 0 2 4 8 8 1 0
W M

-
W I
n
0 w 0.6 I 1
Y
3 GPP i
w 0.4
.- 1
f-cw
-2
8-8-cur 0.2

0.0

-0.2

-0.4
---* CFT Model
-0.6
0 4 8 12 16
-0.4
10 M 90
0 40 50
fiw (eV)
fi w’ (ev)

FIG.5. Left panel: Numerical results obtained using the empirical pseudopotential technique
for various elements of the real ( E , ) and imaginary ( E J parts of the dielectric matrix EGG(q;CO)
as a function of frequency for Si (solid line). The real part of the Hybertsen-Louie plasmon-pole
model is plotted for comparison (dashed line). The model replaces the peaked structure in E ~ ( C O )
by a &function, is constructed to describe the limits w -P 0 and w + co correctly, and breaks
down for intermediate frequencies. Right panel: Real part of the inverse dielectric function of
Si derived from the numerical, empirical-pseudopotential-based results by Walter and Cohen9’
in comparison to the Hybertsen-Louie plasmon-pole model. The average behavior of the
inverse dielectric function is captured rather well by the plasmon-pole model for energies below
the plasmon energy of Si (up,= 16.7 eV). (Adapted from Ref. 44.)
 QUASIPARTICLE CALCULATIONS IN SOLIDS 31

will be used as supplementary material or in cases where no first-principles

calculations exist.
Similar to model dielectric functions, models for the self-energy C or the
band-gap correction A are based on well-founded physical insights into the
effects of correlation in solids, but their accuracy is often limited, leading to
semiquantitative results. A review of simplified GWA calculations based on
models for either the dielectric function or the self-energycan be found in Ref. 14.

a. Model Dielectric Functions

Models for the static dielectric matrix must describe (1) plasmon excitations,
which dominate screening for small reciprocal lattice vectors, and (2)
electron-hole excitations, which fulfill the f-sum rule at large reciprocal
lattice vectors where screening is less effective. Plasmons are collective
coherent excitations of the electron gas which result from screening and can
be visualized as macroscopic density fluctuations. In Fig. 6 areas in
(9, o)-space in which the imaginary part of the dielectric function of the

92 J. P. Walter and M. L. Cohen, Phys. Rev. B5,3101 (1972).

..........................

t .. ............

$?
wE
I ubz
Excitatiom...

I
4, 2kF
- Reciprocal lattice vector q --+-

FIG. 6. Spectrum of excitation energies versus wave vector transfer q for a homogeneous,
interacting electron gas. The upper (lower) boundary of the single-particle excitations is
O.S[(k, + q)2 - kb] ( O . S [ ( - k , + q)2 - kb]). In simple plasmon-pole models this spectrum is
replaced by a single mode wp,(q)indicated by the dotted line. The real plasmon mode (full line)
is the dominant excitation for small q-vectors (0 ,< q < 4,). Beyond qc, the plasmon is heavily
damped (dashed line). (Adapted from Ref. 93.)
32 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

interacting homogeneous electron gas is nonzero are shown schematically.

Besides the continuum of particle-hole excitations, the imaginary part is
nonzero at the plasmon energy o,,(q).For small wave vectors q, 0 d q d q,,
where q, is a cut-off vector, the plasmon is the dominant excitation and
exhausts the f-sum rule to more than 90%.93 Close to q, the plasmon
oscillator strength drops rapidly to zero and particle-hole excitations
become important. For large wave vectors ( q 2 2k,) the spectrum converges
towards the independent-particle spectrum since screening becomes less and
less effective. A good model should therefore interpolate between the
plasmon (0 d q d 4,) and the particle-hole ( q 2 2k,) excitation channels.
Model dielectric matrices for real crystals must also capture important
features that determine screening, such as density inhomogeneities and the
existence of an energy gap in the excitation spectrum. Nonzero off-diagonal
matrix elements are needed to account for local-field effects (see Sections
11.10 and VI.31). Also, the dynamical response of an electron to fluctuations
in the density n(r) from its mean value is qualitatively different in a metal
and a semiconductor or insulator. In a metal the Coulomb potential of a
particle is screened very effectively by the electron gas, and the long-range
Coulomb potential is turned into a short-range interaction that depends on
the local electron density. In a semiconductor or insulator the Coulomb
potential is not completely screened by the electron gas. For example,
consider the case of an additional electron localized in a bond (i.e., an
N + 1- rather than the original N-electron system). This electron will induce
dipoles in neighboring bonds,94 which will in turn interact with and screen
the electron. This screening mechanism is nonlocal in real space.9s Clearly,
model dielectric matrices must account for the gap in the excitation
spectrum and the qualitatively different screening of insulators.
In the context of COHSEX calculations (see Section II.9), static model
dielectric functions are often used. Fair agreement between experimental
and calculated direct gaps is achieved due to a significant error cancellation
between the neglect of local-field effects and the neglect of the energy
dependence of dielectric screening.
The model dielectric function used by Lundqvist9' for the determination
of self-energy corrections of jellium and simple metals (e.g., potassium; see,

93 A. Overhauser, Phys. Rev. 8 3 , 1888 (1971).

94 Ref. 19, 56-57.
9 5 There are simplified tight-binding models that sum all dipole contributions to get an

efective screening potential that depends only on the local electron density; see, for instance,
Refs. 96 and 97.
96 W. Hanke and L. J. Sham, Phys. Rev. B38, 13361 (1988).
97 P. A. Sterne and J. C. Inkson, J. Phys. C Sol. State Phys. 17, 1497 (1984).

98 B. I . Lundqvist, Phys. Kondens. Muter. 6, 206 (1967).

 QUASIPARTICLE CALCULATIONS IN SOLIDS 33

for instance, Ref. 99) (1) reduces to a single plasmon in the limit q 0; (2) -+

describes independent-particle excitations for large q; (3) reduces to

Thomas-Fermi screening for intermediate q; and (4)fulfills important sum
rules like the f-sum rule. L u n d q v i ~ t ' ssingle-plasmon-pole-model
~~ dielectric
function is
Wpzl
&(q,0)= 1 - (2.37)
wz + - w2(q) '

where wpl = (47cn0(0))"2 is the plasmon frequency of the system and no(0)is
the G = 0 component of the unperturbed density. With uF as Fermi velocity,
the plasmon dispersion is approximated by w2(q)= +
4 q 2 / 3 +(q2/2)2.
The continuous spectrum of density fluctuations is substituted by a single
plasmon pole, and electron-hole excitations in particular are neglected. In
the case of the homogeneous electron gas, L u n d q v i ~ t ~ ~ *shows ' ~ ~ ~that
'~'
electron-hole excitations are of minor importance for the determination of
the self-energy C.
In inhomogeneous semiconductors such as Si, quasiparticle energies
determined using plasmon-pole models (e.g., Ref. 42) differ, by no more than
50meV for the direct gap at r, from those determined by taking the full
frequency dependence of &GG'((I; w) into a c c o ~ n t . ' Table
~ ~ *21 ~ shows
~ ~ that
the valence-band width in Si is about 0.3 eV smaller when the full frequency
dependence is taken into account (compare calculations by Fleszar and
Hankelo2 and Rieger et a1.1°3 with plasmon-pole model calculations). The
effect of particle-hole excitations on quasiparticle energies is not significantly
larger than typical numerical uncertainties of GWA calculations (see Sec-
tions 111.15, 111.16, and 111.18).
A modification of the Lundqvist model for s e m i c o n d u ~ t o r(1)
s~~ reduces
~
to the static dielectric constant for q -+ 0; (2) reproduces free-electron
behavior at large q; (3) describes a modified Thomas-Fermi screening at
intermediate q ; and (4) allows for the analytic evaluation of the static
Coulomb-hole contribution to the self-energy. Capellini et u1.'04 suggested
the following expression for the dielectric function of a semiconductor:

99 M. Schreiber and H. Bross, J . Phys. F : Met. Phys. 13, 1895 (1983).

loo B. I. Lundqvist, Phys. Kondens. M a t . 6, 193 (1967).
B. 1. Lundqvist, Phys. Kondens. M a t . 7, 117 (1968).
A. Fleszar and W. Hanke, Phys. Rev. B 56, 10228 (1997).
103
M. M. Rieger, L. Steinbeck, I. D. White, H. N. Rojas, and R. W. Godby, Cond-Mat/
9805246.
G . Cappellini, R. Del Sole, L. Reining, and F. Bechstedt, Phys. Rev. B 47, 9892 (1993).
34 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

where the Thomas-Fermi wave vector k,, depends on the electron density
n and ~(0) is the static RPA dielectric constant. The parameter a is taken as
a fitting parameter to optimize agreement with first-principles dielectric
constants and turns out to be approximately constant for small- and
medium-gap semiconductors.104As is shown in Fig. 7 for Si and GaAs, the
model by Cappellini et al. gives a better description of the full RPA dielectric
function than the Levine-Louie model, which is introduced below. However,
in contrast to the Levine-Louie model, the model of Cappellini et al. does
not reduce to the RPA dielectric function in the case of metallic screening.

\. ..-

FIG. 7. Static dielectric function E(q. w = 0) as a function of the reciprocal lattice vector q
for Si and GaAs. The model dielectric function of Cappellini et (solid line) compares
better with RPA results of Walter and C ~ h e n "(closed
~ boxes for q along the (111) direction,
open boxes for q along the (100) direction) than the Levine-Louie model dielectric function
(dashed line).lo6 RPA results by Baldereschi and Tosatti (stars)'" for large q seem to favor
the Levine-Louie model over the model by Cappellini et al. The parameters a and ~(0) are the
same as in Eq. (2.38). (Adapted from Ref. 104.)
 QUASIPARTICLE CALCULATIONS IN SOLIDS 35

For an application of the model in the context of GWA see, for example,
Ref 108.
COHSEX calculations for insulators in the 1970s and early 1980s (see
Section II.9a) often relied on the Penn dielectric function'09 and its
variations, which describe the dielectric response of an isotropic, three-
dimensional insulating electron gas. Perm'" used the Ehrenreich-Cohen
formula"0 for the dielectric response function. With Egapas an average
optical gap adjusted to fit E ( q = 0) with the experimental dielectric constant,
E F as the Fermi energy, and k, as the Fermi vector, the numerical results
can be fitted with an interpolation formula:"'*''2

(2.39)

where the factor F is given by

(2.40)

Modifications of the Penn model based on a more complicated band

structure have been suggested"'*"2 and applied to static COHSEX calcu-
lations (see Section II.9a).
In the same spirit as the Penn model, the Levine-Louie dielectric func-
tion106 modifies the RPA dielectric function by an ad-hoc introduction of
an energy gap in the spectrum of the homogeneous electron gas. This
model dielectric function (1) interpolates smoothly between the screening
properties of a metal and an insulator; (2) reproduces well the numerical
results of Walter and Cohen"' for the static diagonal dielectric matrix in
Si, as shown in Fig. 7; (3) fulfills important sum rules such as the f-sum rule;
(4) has the correct long-wavelength behavior; and (5) neglects local-field
effects. Since semiconductors and insulators cannot absorb light with
energies below the fundamental band gap, Levine and Louie introduced the
band gap in an ad-hoc fashion into the imaginary part of the RPA dielectric

J. P. Walter and M. L. Cohen, Phys. Rev. 8 2 , 1821 (1970).

Z. H. Levine and S. G. Louie, Phys. Rev. B25, 6310 (1982).
lo' A. Baldereschi and E. Tosatti, Phys. Rev. B 17, 4710 (1978).
lo* F. Bechstedt, R. Del Sole, G. Cappellini, and L. Reining, Sol. Srure Comm. 84,765 (1992).
log D. R. Penn, Phys. Rev. 128, 2093 (1962).
H. Ehrenreich and M. H. Cohen, Phys. Rev. 115, 786 (1959).
J. L. Fry, Phys. Rev. 179, 892 (1969).
'I2 N. 0. Lipari, J . Chem. Phys. 53, 1040 (1970).
36 WILFRIED G. AULBUR, LARS JONSSON, A N D J O H N W. WILKINS

function via

sgn(o-) = sgn(o). (2.41)

This approach results in a diagonal model dielectric function that depends

on only two parameters: the average electron density given by the electron
gas parameter rs, which determines the plasmon frequency opl,and the
lowest excitation frequency ogap.In the long-wavelength limit ( q 0) one -+

finds

of,
E(q -+ 0,o) = 1 + oiap
- o2.
(2.42)

For a known dielectric constant the above equation at o = 0 defines ogap.

To include local-field effects -in particular the variation of the screening
hole depth as a function of the location of an added electron-and the
correct symmetry of the screening potential under exchange of r and r’,
Hybertsen and Louie’I3 considered a screening potential that is the sum of
the local Levine-Louie potentials evaluated at r and r’ (see Eq. (2.45) below).
They evaluated the homogeneous Levine-Louie screened potential,

at the local density determined by r,(r’) and used the Levine-Louie gap
parameter to ensure correct long-range screening:

Here ~ ( 0is) the static dielectric constant and an input parameter to the
model. Hybertsen and Louie’ explicitly symmetrized the screening poten-
tial:

WHL(r,r’) = l[WLL(r- r‘; rs(r’)) + WLL(r’- r; rs(r))], (2.45)

‘I3 M. S. Hybertsen and S. G. Louie, Phys. Rev. B37, 2733 (1988).

‘I4 M. Rohlfing, P. Kriiger, and J. Pollmann, Phys. Rev. B48, 17791 (1993).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 31

which leads to the following expression in reciprocal space:

[EE&] + ;l
(q ;o = O)u(q G ) = - u(q G ) + J [E~J - '(Iq + GI; rs(r'))e'(G-G).dr'
r'

1.
-

+ u(q+G')
s [ ~ ~ ~ ] - l ( l q + Gr,(r))e'(G'-G)'rdr
'I;

The diagonal part of the Hybertsen-Louie dielectric matrix is given as the

(2.46)

average over the local Levine-Louie screening response at different points in

the crystal. The Hybertsen-Levine-Louie model has been applied to a
variety of bulk semiconductors and interfaces (see Sections IV. 19 to IV.22).
It generally reproduces results of full RPA calculations for quasiparticle
energies to within 0.1 to 0.4 eV for states close to the band gap.' 1 3 s 1 l 4

b. Plasmon-Pole Models
The imaginary part of the important elements of the dielectric matrix has a
peaked structure as a function of frequency, which plasmon-pole models
approximate by a &function characterized by two parameters: the effective
strength and the effective frequency of the plasmon excitation. Together with
the independent-particle propagator (Eq. (2.4)), whose frequency depend-
ence is straightforward, plasmon-pole models allow an analytic evaluation
of the convolution of G and W in frequency space (Eq. (2.24)). As a
consequence, the numerically expensive integration in Eq. (2.24) is avoided.
Plasmon-pole models give a good description of both the low-energy
behavior of the dielectric matrix -via reproduction of the static limit -and
the high-energy behavior of the dielectric matrix -via reproduction of the
first frequency moment. As a consequence, energy integrations of the
screened interaction are sufficiently accurate for the determination of quasi-
particle energies and effective masses." The effective strength and fre-
quency of the plasmon excitation are determined by forcing the model to
reproduce the static dielectric matrix in the zero-frequency limit and by
using Johnson's sum rule.' l 7 The latter sum rule connects the first
6+1

l 5 N . H. March, Electron Correlation in Molecules and Condensed Phases, Plenum Press, New
York (1996), 132.
D. L. Johnson, Phys. Rev. B 9 , 4475 (1974); M. Taut, J. Phys. C: Sol. State Phys. 18, 2677
(1985); and Ref. 44.
'I7 The use of Johnson's sum rule in conjunction with nonlocal pseudopotentials is not
justified, as pointed out for instance in Refs. 44 and 118. However, band-gap errors due to this
procedure are only about 15 meV for Si, as reported in Ref. 119.
G . E. Engel and B. Farid, Phys. Rev. B46, 15812 (1992).
R. T. M. Ummels, P. A. Bobbert, W. van Haeringen, Phys. Rev. B57 11962 (1998).
38 WILFRIED G. AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

frequency moment of the full polarizability x = 6n/61/,,,,'20

(2.47)

with the ground-state density no(G)of the crystal,

x
X&!AQ) = - 5 (9 + G ).(4 + G')no(G- G'). (2.48)

In RPA, x is related to the irreducible polarizability P = Po in matrix

notation via x(q, o)= P(q, 0)[1 - u(q)P(q, o)]-'.
Quasiparticle energies in jellium calculated with the RPA dielectric
function or with the Lundqvist plasmon-pole r n ~ d e l ~ ~are ~ 'very
~ ~ close
~ ' ~ '
(see previous subsection), which indicates that the neglect of electron-hole
excitations in plasmon-pole models is justified for energies close to the Fermi
energy. Similar observations hold for semiconductors, as discussed above.
The plasmon-pole assumption, that the only possible quasiparticle decay
mechanism -described by the imaginary part of the dielectric matrix -is
due to quasiparticle scattering off plasmon excitations, results in finite
quasiparticle lifetimes only at plasmon frequencies. Plasmon-pole models
fail to give good descriptions of systems whose imaginary self-energy is
large.' l 5 In particular, low-lying valence states and hence the valence band
width are not described accurately. For example, Table 21 shows that taking
into account the full frequency dependence of the dielectric matrix leads to
a valence-band narrowing of about 0.3 eV in Si compared to a plasmon-pole
model calculation (compare the frequency-dependent results of Refs. 102
and 103 with results using plasmon-pole approximations such as Ref. 42; see
also Ref. 89).
Hybertsen and approximated each matrix element of the dielec-
tric matrix by a plasmon-pole model, which results in (1) N 2 plasmon-pole
parameters ( N = size of the dielectric matrix), (2) a nontrivial extension of
the model to systems without inversion symmetry,'2' and (3) unphysical
solutions with imaginary plasmon-pole energies for some off-diagonal
matrix elements. Rather than concentrate on the details of the Hybertsen-
Louie plasmon-pole model, we consider two alternative plasmon-pole
models suggested by von der Linden and Horsch'22 and Engel and

G. E. Engel and B. Farid, Phys. Rev. B47, 15931 (1993).

S.B. Zhang, D. Tomanek, M. L. Cohen, S. G. Louie, and M. S. Hybertsen, Phys. Rev. 840,
3162 (1989).
12* W. von der Linden and P. Horsch, Phys. Rev. 837,8351 (1988).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 39

Farid.' 2o Both models give identical expressions for the expectation value
of the self-energy operator provided that appropriately scaled plasmon-pole
eigenvalues and eigenvectors are introduced (see below and Sections 111.15
and 111.16).
Von der Linden and Horsch. Von der Linden and Horsch'22 considered
an N-parameter plasmon-pole model derived from the eigenvalue decompo-
sition of the symmetrized dielectric matrix E under the assumption that the
energy dependence of E is contained in its eigenvalues only. Define the
following symmetrized, Hermitian dielectric matrix,

(2.49)

whose inverse 2-l has the eigenvalue d e c o m p ~ s i t i o n , ' ~ ~ * ' ~ ~

where p numbers the real, positive eigenvalues &'(a) and the correspond-
ing eigenvectors IOw(o)). Assume that the frequency dependence of the
inverse dielectric matrix is solely contained in the eigenvalues (lOw(o)) =
IOw(o = 0))) and is of the

where the plasmon pole strength, zw = 1 - h,'(O), is determined by com-

parison with the static dielectric matrix. With the definition of the scaled
plasmon-pole eigenvectors as

(2.52)

and the "first-moment" matrix,

2
LGG'(q) = (q $. G ) .(q + G')no(G- G') = --
71
xg&(q), (2.53)

one can express the plasmon-pole frequencies using Johnson's sum rule' l6

lZ3 R. Hott, Phys. Rev. B44, 1057 (1991).

40 WILFRIED G. AULBUR, LARS JBNSSON, A N D JOHN W. WILKINS

as

(2.54)

Subsequently, the energy integration of Eq. (2.24) can be done and the result
is given in Sections 111.15 and 111.16 for a plane-wave basis.
Engel and Farid. Engel and Farid'a' derived an N-parameter plasmon-
pole model whose eigenvalues and eigenvectors are frequency dependent by
explicitly constructing an approximation X to the full polarizability x, which
approximation is exact in the static and the high-frequency limits. With X,
x, and L-' (the inverse of the first-moment matrix L in Eq. (2.53)) as
matrices in the reciprocal lattice vectors G and G', the approximation X to
the full polarizability x can be expressed as"'

The Engel-Farid plasmon-pole model can be obtained by diagonalizing the

above equation, which is equivalent to the solution of the following
generalized eigenvalue problem:

(2.56)

The eigenvectors x, are normalized as follows:

(2.57)

Defining the scaled eigenvectors y, as

(2.58)

one finds for the spectral representation of X

(2.59)
 QUASIPARTICLE CALCULATIONS IN SOLIDS 41

and its inverse

These equations define the Engel-Farid plasmon-pole model. For future

reference in the determination of the self-energy using the Engel-Farid
plasmon-pole model, we define scaled plasmon-pole eigenvectors as

(2.61)

An interpretation of the plasmon-pole eigenvalues as plasmon energies, as

suggested by Engel and Farid, leads to an agreement between theory and
experiment to within 10% for plasmon energies o,(O) and to within 30%
for the small q-vector dispersion coefficient a, o,(q) z o,(O) +
alql', in the
case of the semiconductors Si, Ge, GaAs, Sic, and GaN. This is seen in
Table 3. March and Tosi131 showed that the plasmon frequencies of a
system are the zeros of the determinant of the inverse full polarizability x- ':

(2.62)

TABLE3. THEORETICAL AND EXPERIMENTAL PLASMON ENERGIES

AND DISPERSION COEFFICIENTS
FOR Si, Ge, GaAs, Sic, AND GaN. THETHEORETICAL VALUESOF ENGELAND FARIDFOR Si'*O
AND BACKES et a[. FOR SicL3' ARE LISTED
I N PARENTHESES.
AGREEMENTBETWEEN THE DIFFERENT
CALCULATIONS IS WITHIN A FEW PERCENT. THE THEORETICAL
AND EXPERIMENTAL PLASMON
ENERGIES AGREE TO WITHIN 10% AND THE DISPERSION COEFFICIENTS TO WITHIN 30%.

%(O) a

THEORY' EXPT.~ THEORY' EXPT.~

Si 15.7 (15.9) 16.7' 0.36 (0.34) 0.41

Ge 14.2 15.9-16.5 0.38 [lo01 0.38'
0.35 [l 1 13
0.25 [ l l O ]
GaAs 14.1 15.7 0.49 0.44'
0.61 [loo]
0.57 [llO]
0.53 [lll]'
Sic 22.5 (22.4) 22.1
GaN 19.4

"Ref. 42 except for values in parentheses; bRef. 124 for ~ ~ ( and

0 ) Ref. 125 for a unless
otherwise noted; 'Ref. 126; 'Ref. 127; 'Ref. 128; /Ref. 129.
42 WILFRIED G . AULBUR, LARS JONSSON, AND JOHN W. WILKINS

This observation in conjunction with Eq. (2.60) led Engel and FaridI2’ to
interpret the eigenenergies of Eq. (2.56) as plasmon energies and their small
q-vector dispersion coefficient as plasmon dispersion coefficient.

9. EARLYQUASIPARTICLE
CALCULATIONS

This section reviews the history of GWA calculations and outlines some
related approaches. As in the section on the model dielectric matrices, we
concentrate on work done on semiconductors and insulators. We focus our
survey on the time span between 1970 and 1986. Before 1970 most of the
GWA work was done on the homogeneous electron gas. The only work on
semiconductors is the paper by Brinkman and Goodman.83 These papers
are thoroughly reviewed in Refs. 4 and 132. In 1985, first-principles
quasiparticle calculations were done by Hybertsen and Louie’ and in 1986
by Godby, Schluter, and Sham.6 First-principles GWA calculations are
discussed at length starting in Section 11.10.

a. Static COHSEX Calculations

General Considerations. The error cancellation between the neglect of
dynamical effects in the static COHSEX approximation and the neglect of
local-field effects due to the use of a diagonal model dielectric function- the
common characteristic of all quasiparticle calculations on bulk insulators in
the 1970s and early 1980s-results in direct quasiparticle band gaps that
agree well with experiment, as shown in Table 4. Indirect band gaps,
however, differ significantly from experimental values. In the case of sili-
the first few conduction bands at r are at 3.24 and 4.41 eV in
comparison to experimental values of 3.4 and 4.2 eV. The indirect band gap
is 0.68 eV and hence underestimates the experimental value of 1.17 eV by
about 0.5 eV. As a consequence, the static COHSEX approximation cannot
replace the dynamically screened interaction approximation if one is inter-
ested in the full band structure.

H. Raether, Excitation of Plasmons and Interband Transitions by Electrons. Springer Tracts

in Modem Physics, vol. 88, Springer-Verlag, Berlin (1980), Table 5.1.
’’’ Ref. 124, Table 7.4.
J. Stiebling and H. Raether, Phys. Rev. Lett 40,1293 (1978).
12’ H. Watanabe, J. Phys. SOC.Jpn. 11, 112 (1956).
12* C. von Festenberg, Z . Phys. 214,464 (1968).
12’ R. Manzke, J. Phys. C: Sol. State Phys. 13, 911 (1980).
I 3 O W. H. Backes, P. A. Bobbert, and W. van Haenngen, Phys. Rev. B 51,4950 (1995).
13’ N. H. March and M. P. Tosi, Proc. R SOC.Lond. ,4330, 373 (1972).
”’ L. Hedin and B. I. Lundqvist, J . Phys. C Sol. State Phys. 4, 2064 (1971).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 43

TABLE4. COMPARISON OF TECHNICAL ASPECTSOF COHSEX CALCULATIONS BY LIPARIet a/.

REFS. 133, 134, 135, AND 136, -,I3’ BRENFX,’~*.’~~BARONIet aL,140.141 AND GYGIAND
BALDERESCHl,90 WITH GWA CALCULATIONSBY HYBERTSEN AND LOUIE’ AND GODBY, SCHLmR,
AND SHAM.^ COREELECTRONS ARE EITHERTREATEDON AN LDA-PSEUDOPOTENTIAL (PSP)
OR HARTREE-FOCK (HF, NEGLECTOF CORE-VALENCE POLARIZATION) LEVEL,OR TREATED
ON THE S m LEVELAS THE VALENCE ELECTRONS. MOST COH_SEX CALCULATIONS START
FROM A HARTREE-FOCK INDEPENDENT-PARTICLE HAMILTONIAN H o , WHICH IS APPROPRIATE
CONSIDERING THE S ~ DIELECTRIC
L CONSTANTS OF THE MATERIALS. GWA CALCULATIONS USE
LDA AS AN INDEPENDENT-PARTICLE HAMILTONIAN INSTEAD.AGREEMENT WITH EXPERIMENT FOR
THE DIRECTBANDGAPIS IN GENERAL WITHINA FEW TENTHSOF AN eV. VALUES FOR THE BAND
GAP AND THE DIELECTRIC CONSTANT ARE TAKENFROM THE ORIGINALPAPERSWITH THE
EXCEPTION OF THE DIELECTRICCONSTANT OF LiFI4’ AND NaBr AND NaC1.’43 OTHER
ABBREVIATIONSUSED ARE EPSP = EMPIRICAL P S P MB = MIXEDBASIS;(0)PW = (ORTHO-
GONA~IZED) PLANEWAVE;GO = GAUSSIAN ORBITAL;LCAO = LINEARCOMBINATION OF
ATOMICORBITALS.

DIRECTGAP (eV)
Ehhp/E;:; (MATERIAL,
EXP.
FEATURE CORE I? BASIS DIELECTRIC CONSTANT)

LiparP HF HF MB/OPW 13.7/14.3 (ArJ.67)

7.5/7.1-7.4 (NaBr,2.60)
8.6118.75 (NaQ2.33)
Kane“ EPSP EPSP PW NIA
Brenef valence HF LCAO 7.617.3- 7.6 (C,5.70)
17.9113.6 (LiF, 1.96)
Baroni valence HF GO + OPW 5.2414.99 (LiH,3.61)
14.62p4.15 (Ar, 1.67)
Gygi” LDAIPSP LDA PSPIPW 3.24/3.4 (Si,l1.7)
Hybertsenb LDAIPSP LDA PSPIPW 3.313.4 (Si,l1.7)
Godby* LDAIPSP LDA PSPIPW 3.2713.4 (Si,ll.7)

“Static COHSEX calculation neglecting local fields and based on a model dielectric function.
bGWAcalculation including energy dependence (plasmon-pole model in Ref. 5) and local fields,
and based on first-principles dielectric matrix.

Hartree-Fock calculations offer a convenient single-particle basis for

low-polarizability solids-the focus of much of the early work in the
COHSEX approximation -but fail to give a good description of polariz-
able solids such as Si for which empirical pseudopotential or LDA wave
functions and energies work better. Table 5 shows that in the case of Si,90
self-consistent COHSEX and LDA densities are very similar whereas
Hartree-Fock calculations concentrate too much density in the bonds.
Core-valence exchange and correlation potentials can be determined
either via effective (e.g., core-polarization) potentials or via treatment of the
core electrons on the same level of approximation as the valence electrons.
44 WILFRIED G. AULBUR, LARS JONSSON, A N D J O H N W. WILKINS

TABLE5. SELECTEDFOURIER OF THE VALENCEDENSITYI N Si OBTAINED

COEFFICIENTS BY
HARTREE,LDA, DIAGONALCOHSEX (STATIC,N o LOCAL FIELDS),AND HARTREE-FOCK
CALCULATIONS, USINGTHE SINGLEMEAN-VALUE P O ~ TECHNIQUE.
T ALL VALUESARE GIVENI N
ELECTRONS PER UNIT CELL. THEDEGREEOF DENSITYLOCALIZATION IS LOWESTIN THE HARTREE
AND HIGHESTIN THE HARTREE-FOCK CALCULATION.THELDA DENSITY LIES BETWEENTHESE
T w o EXTREMESAND IS CLOSEST
TO THE DENSITY
OBTAINED FROM A SELF-CONSISTENT COHSEX
CALCULATION.(ADAPTEDFROM REF. 90).

G HARTREE LDA COHSEX HARTREE-FOCK

(1,1,1) - 1.42 - 1.67 - 1.59 - 1.17

(2,2,0) 0.14 0.13 0.14 0.16
(3,1,1) 0.30 0.36 0.35 0.42
(2,2,2) 0.20 0.31 ' 0.29 0.43
(4,O 30) 0.34 0.40 0.38 0.43

Explicit treatment of core electrons using a linear-combination-of-atomic-

orbitals basis' 3 8 * 139 or a basis set of localized, contracted Gaussian-type
orbital^'^^^'^^ -which allow the analytic evaluation of screened-exchange
two-electron integrals -has been the method of choice in static COHSEX
calculations. Alternatively, unscreened core-core and core-valence exchange
potentials from Eq. (2.35) are taken into account in static COHSEX
calculations, but self-energy contributions due to core-valence polarization
are neglected (see Refs. 133, 134, 135, and 136).
Recent GWA calculation^^*^ extend earlier COHSEX work by using a
nondiagonal, frequency-dependent'44 first-principles dielectric matrix for
the determination of quasiparticle corrections to the single-particle spec-
trum. The treatment of core electrons is (1) at LDA l e ~ e l , (2)
~ . implicit
~ via
core-polarization potential^,^^ and (3) explicit by treating core electrons on
the same level as valence electron^.^^*^^*'^^ The consequences of the GWA

133 A. B. Kunz and N. 0. Lipari, Phys. Rev. B4, 1374 (1971).

134 N. 0. Lipari and W. B. Fowler, Sol. State Comm. 8, 1395 (1970).
N. 0. Lipari and W. B. Fowler, Phys. Rev. B2, 3354 (1970).
N. 0. Lipari and A. B. Kunz, Phys. Rev. 8 3 , 4 9 1 (1971).
13' E. 0. Kane, Phys. Rev. B5, 1493 (1972).
13' N. Brener, Phys. Rev. B 11, 929 (1975).
1 3 9 N. Brener, Phys. Rev. E l l , 1600 (1975).

S. Baroni, G. Grosso, and G. Pastori Parravicini, Phys. Rev. B29, 2891 (1984).
14' S. Baroni, G. Pastori Parravicini, and G. Pezzica, Phys. Rev. B32, 4077 (1985).
1 4 2 Handbook of Optical Constants ofsolids, ed. E. D. Palik, Academic Press, Orlando (1985).

1 4 3 M. E. Lines, Phys. Rev. B41, 3383 (1990).

144 Hybertsen and Louie used a plasmon-pole model; Godby and collaborators determined

the full frequency dependence of the screened interaction.

1 4 5 F. Aryasetiawan and 0. Gunnarsson, Phys. Rev. B54, 17564 (1996).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 45

extensions compared to static COHSEX calculations are discussed in the

following sections.
Applications. Work on bulk insulators was done by Lipari and collabor-
ators (NaCI, NaBr, solid Ar; Refs. 133, 134, 135, and 136) and Kane (Si; Ref.
137) in the early seventies; by Brener in the mid-seventies (LiF, diamond; Refs.
138 and 139);by Louis in the late seventies (AX, A = Li’, Na’, X = F-, C1-,
Br-, I - ; Ref. 146); and by Baroni and collaborators (solid Ar, LiH; Refs. 140
and 141) as well as by Gygi and Baldereschi (Si; Ref. 90) in the mid-eighties.
Calculations on potassium and other metals were, for instance, reported
by Bross and collaborators (see Ref. 99 and references therein). The starting
Hamiltonian for potassium is defined by a “parametrized Thomas-Fermi-
like” approximation to the static ~ e l f - e n e r g y : ~ ~ . ’ ~ ’

Modeling the frequency dependence of the screened interaction by Lund-

qvist’s plasmon-pole model9* (see Section 11.8) and core-valence polariz-
ation using the dipolar core polarizability of potassium leads to
quasiparticle energies that agree to within 20% with e ~ p e r i m e n t The
.~~
inclusion of core polarization reduces self-energy shifts by 0.1 to 0.2 eV.99
Other applications of the static COHSEX approximation include Gad-
zuk‘s determination of polarization energies of core holes in atoms or
molecules that are embedded in or adsorbed on the surface of a free electron
gas (Ref. 148 and references therein). H o d g e ~ used
’ ~ ~ a dynamic COHSEX
approximation to explore the image potential of a charged particle trap-
ped at the surface of a metal, and Cooper and Linderberg15’ used the
COHSEX approximation within the Pariser-Parr-Pople model for the
description of n-electron systems in hydrocarbons. Theoretical issues, such
as the importance of vertex corrections, have been addressed by several
author^.'^'*^^^*'^^ We review these works in more detail in Section 11.14,
where we discuss vertex corrections to the GWA.

146 E. Louis, Phys. Rev. B20, 2537 (1979).

14’ This Harniltonian can be derived within the framework of DFT; see Section VII.33. The
notation in the equation is the same as used elsewhere in this article.
14’ J. W. Gadzuk, Phys. Rev. B 14, 2267 (1976).
14’ C. H. Hodges, J. Phys. C: Sol. State Phys. 8, 1849 (1975).
I. L. Cooper and J. Linderberg, Molec. Phys. 25, 265 (1973).
1 5 1 L. W. Beeferrnan and H. Ehrenreich, Phys. Rev. B 2 , 364 (1970).

P. Minnhagen, J. Phys. C Sol. State Phys. 7, 3013 (1974).

H. Suehiro, Y. Ousaka, and H. Yasuhara, J. Phys. C: Sol. State Phys. 17, 6685 (1984).
46 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Several alternative or more general approaches for the determination

of quasiparticle energies have been suggested in the literature. The approaches
by Hanke and collaborator^,'^^^'^^^^^^ by Fulde and collaborators,"7~'58~'59
by Pickett and Wang,'60*'61 and by Sterne and Wang16' are discussed in
some detail in the paragraphs below, since they either extend RPA-based
GWA calculations or correspond to interesting alternative pictures for 'the
description of exchange and correlation in solids. Other approaches, such as
the electronic polaron model by K ~ n z or ' ~ by
~ H e r m a n ~ o n , 'can
~ ~ be
identified as limiting cases of GWA calculation^.^^ Their physics is contained
in the general discussion of the GWA approach, and we do not discuss these
references separately.

b. Excitonic Efects
An extension of the COHSEX approximation' 54,1 that includes excitonic
correlations in the irreducible polarizability gives a better description of
photoabsorption experiments-in which photons create interacting electron-
hole pairs- than do RPA-based approaches. In contrast to RPA-based GWA
calculations whose fundamental excitations are noninteracting electron-hole
pairs and plasmons, GWA calculations based on a description of screening in
the time-dependent Hartree-Fock approximation include the short-range,
attractive interaction between electrons and holes via ladder diagrams.
The determination of the influence of excitonic correlations on density
fluctuations and hence on the dielectric matrix requires the solution of the
equation of motion of the two-particle Green function, that is, the Bethe
Salpeter equation.' 5 4 * 1 5 5 ~ 15 6 The inverse dielectric matrix is given in terms
of the density-density correlation function,'65 and the latter can be written
as a two-particle Green function:'67

W. Hanke and L. J. Sham, Phys. Rev. Lett. 33, 582 (1974).

W. Hanke and L. J. Sham, Phys. Rev. B 12,4501 (1975).
1 5 6 G. Strinati, H. J. Mattausch, and W. Hanke, Phys. Rev. B25,2867 (1982).
"' W. Borrmann and P. Fulde, Phys. Rev. B35, 9569 (1987).
S. Horsch, P. Horsch, and P. Fulde, Phys. Rev. B28,5977 (1983).
S. Horsch, P. Horsch, and P. Fulde, Phys. Rev. B29, 1870 (1984).
W. E. Pickett and C. S. Wang, Phys. Rev. E M ,4719 (1984).
16' W. E. Pickett and C. S. Wang, Int. J. Quant. Chem. Quant. Chem. Symp. 20, 299 (1986).
P. A. Sterne and C. S. Wang, Phys. Rev. B37, 10436 (1988).
A. B. Kunz, Phys. Rev. B 6 , 606 (1972).
164 J. Hermanson, Phys. Rev. B 6, 2427 (1972).

16' To be consistent with Hanke et aL's notation we do not consider density fluctuation
operators A'(rt) = h(rt) - (A(rt)) as in Eqs. (2.17) and (2.18), but density operators h(rt). For
the definition of the physically relevant retarded dielectric matrix, the two approaches are
equivalent, since c-numbers always commute. For the time-ordered dielectric matrix, the
approaches differ only in the limit w + 0. The formulation using the density fluctuation
operator avoids a divergence due to the bosonic nature of the density-density correlation
function (see, for instance, Refs. 18, 558ff., and 166).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 47

( N , OlT[A(rt)n*(r’O)](N,0) = - iG(rt, r’0, r’O+, rt’). (2.64)

Let Go( 1, l’, 2,2’) be the noninteracting electron-hole propagator,

GO(l,l’,2,2’) = G(l’, 2’)G(2, l), (2.65)

and I be the irreducible electron-hole interaction, whose approximation is

shown in Fig.

Here - W,, is a static approximation to the screened interaction between an

electron and a hole and u is the unscreened Coulomb repulsion. Choosing
W,, = 0 leads back to the RPA results. Solution of the Bethe-Salpeter
equation -

G(1, l’,2,2’) = GO(l,l’, 2,2’)

+ GO(l, l’,3,3‘)1(3, 3’,4,4’)G(4,4‘,2,2’)d(3,3’, 4,4‘) (2.67)

-determines the influence of excitonic correlations on the frequency

166 H. Stolz, Einfihrung in die Vielelektronentheorie der Kristalle, Bertelsmann Universitats-

verlag, Diisseldorf (1975), Chap. 6.

x
Generally, a two-particle Green function is defined as
G ( r , t , , rzt,, r3t3,r4f4) = ( - i ) * ( N , OlTC~(r,tl)Y(r2t2)Yt(r4r4)~t(r3r3)llN,
0).
16’ W. Hanke and L. J. Sham, Phys. Rev. B21,4656 (1980).

,IITs= 3’
4‘

_ - _ _ - - _Coulomb
- repulsion
‘VVL Coulomb attraction
FIG. 8. The irreducible electron-hole interaction in the time-dependent Hartree-Fock ap-
proximation is given by the sum of a screened Coulomb attraction and an unscreened Coulomb
repulsion. The first term generates an infinite sum of ladder diagrams, whereas the second term
generates an infinite sum of polarization bubbles (compare Eq. (2.66)). (From Ref. 168.)
48 WILFRIED G. AULBUR, LARS JbNSSON, A N D JOHN W. WILKINS

dependence of the dielectric matrix (via Eqs. (2.64) and (2.18)) and on the
quasiparticle spectrum (via Eqs. (1.3) and (2.1 1)).
The inclusion of vertex corrections in the irreducible polarizability but not
in the self-energy is motivated by two observations. (1) An exact Ward-
identity-based relation between the self-energy and the vertex function is
expected to be better observed when the irreducible polarizability is deter-
mined in the time-dependent Hartree-Fock rather than the time-dependenl
Hartree approximation.' 5 6 This Ward-identity-based relation results from
charge conservation. (2) Hanke and collaborators suggested that the evalu-
ation of the GWA self-energy with an interaction screened using the
time-dependent Hartree-Fock approximation is more appropriate than an
evaluation of C using an RPA-screened interaction. The choice of screening
in the time-dependent Hartree-Fock approximation can be motivated a$
follows: Choosing C = 0 in Fig. 4 as the starting point for an iterative
solution leads to l-(1,2; 3) = 6(1,2)6(1,3) and E = E~~~ (see Eqs. (2.10) and
(2.23)). The next iteration yields C(1,2) = iG(1,2)W(1+,2) and subsequent13

x G(4,6)G(7,5)r ( 6 , 7 ; 3)d(4,5,6,7)

= 6(1,2)6(1,3) +i
s W(1+,2)G(1,6)G(7,2)r(6,7; 3)d(6,7),
(2.68
where contributions due to SW/SC are neglected since they are difficult tc
handle and their contributions are estimated to be negligible for Si anc
diamond.'69 Using the above expression for the vertex function in thc
definition of the polarizability as well as the dielectric matrix, one finds tha
the dielectric matrix includes not only RPA bubbles, corresponding tc
6(1,2)6(1,3) in the above equation, but also an infinite series of screenec
ladder diagrams resulting from the second term in Eq. (2.68). Note tha
Mahan" used Ward identities to justify the simultaneous inclusion of iden
tical vertex corrections in C and P in contrast to the approach suggested b!
Hanke et al. Mahan's approach is discussed in more detail in Section 11.14
Inclusion of excitonic effects via the solution of the Bethe-Salpete
equation for several semiconductors leads to (see Refs. 156, 169, 170, 171

1 6 9 W. Hanke, N. Meskini, and H. Weiler, in Electronic Structure, Dynamics, and Quantun

Structural Properties of Condensed Matter, eds. J. T. Devreese and P. Van Camp, Plenum, Nev
York (1985), 113.
170 H. J. Mattausch, W. Hanke, and G. Strinati, Phys. Rev. B27,3735 (1983).
"' N . Meskini, H. J. Mattausch, and W. Hanke, Sol. State Comm. 48, 807 (1983).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 49

and 172): (1) band gaps and valence band widths in good agreement with
experiment, (2) an increase in oscillator strength at the El peak of the
frequency spectrum of E by 50- loo%, in good agreement with experiment,
and (3) a larger El peak intensity than E , peak intensity for Si, in contrast
with experiment (see Fig. 39 for a definition of the E l and E , peaks in the
imaginary part of the Si dielectric function). Practical calculations use a
matrix representation of the Bethe-Salpeter equation with respect to semi-
empirical tight-binding valence and conduction bands, which are expressed
in a Gaussian-orbital basis

c. Local Approach
Local approaches to the correlation problem in solids (for a review see Ref.
19) are based on the physical picture that adding an electron or hole to an
insulator leads to a polarization of neighboring bonds whose induced
dipoles will act on and change the energy of the original particle. A local
description of correlation effects based on quantum chemistry methods
starts from a Hartree-Fock wave function and includes local correlations in
the form of one- and two-particle1 7 3 excitations in the many-body wave
function I Y ) via a projection operator exp($:

I'r) = exp(~)l%). (2.69)

The operator s
depends on parameters that characterize ground- and
excited-state correlations and that can be determined variationally.
Bond polarizations are interatomic correlation effects that dominate
dielectric screening and hence quasiparticle energy shifts of occupied and
unoccupied states in covalent semiconductors. These interatomic correlation
effects can be described by projector methods based on a minimal basis set
of Gaussian orbital^.'^^.'^^,'^^ Other important effects, like the relaxation
of electronic orbitals in the neighborhood of an added electron or the
change of ground-state correlations due to the presence of an extra electron,
require a larger basis set. Each of these two partially canceling effects is
estimated to amount to band-gap corrections on the order of 1 to 2eV
based on molecular calculations.157~158~159
Local projector methods determine ground-state properties of a variety
of semiconductors with better accuracy than LDA,' 74.1 7 5 - 17 6 but applica-

17* G. Strinati, H. .I.

Mattausch, and W. Hanke, Phys. R K KLett 45, 290 (1980).
173 Excitonic correlations are not included, in contrast to Refs. 156 and 169.
M. Albrecht, B. Paulus, and H. Stoll, Phys. Rev. B56, 7339 (1997).
"' B. Paulus, P. Fulde, and H. Stoll, Phys. Rev. B51, 10572 (1995).
176 B. Paulus, P. Fulde, and H. Stoll, Phys. Rev. B54, 2556 (1996).
50 WILFRIED G. AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

tions to excited states are limited and remain difficult due to the extended
nature of these ~ t a t e s . ' ~ The
~ . ' ~calculations
~ scale as the number of basis
functions to the fifth power,'79 compared to standard LDA and GWA
calculations that scale as the number of basis functions to the third and
fourth power, respectively. Applications of projector methods to excited-
state proper tie^'^^*'^^*'^^ have been reported for Si, Egap= 3.0eV (3.3-
3.4 eV experiment); diamond, Egap= 7.2 eV (7.3 eV experiment); and Ge,
Egap= 0.44eV (0.9 eV experiment).

d. Quasiparticle Local Density Approximation (QPLDA)

The self-energy is a short-ranged operator39 and -as a ground-state
property -a functional of the ground-state density.39 For systems with
slowly varying density, a possible approximation of I: is given by the
self-energy I:,,of the homogeneous electron gas evaluated at the local
density:39

C(r, r'; E ) % I:,,(p(r); E - p - p,,(n(r)); n(r))6(r - r'). (2.70)

Here p(r) is a local momentum and the chemical potentials of the in-
homogeneous (p) and the homogeneous @,,) system are introduced to line
up the Fermi levels. Calculations on metals using this approximation are
reported by, among others, Arbman and von BarthS6 for A1 and by
Sacchetti"' for Cu.
For practical calculations [Refs. 160, 161, 162, and 1811 in semiconduc-
tors the homogeneous self-energy is evaluated in the GWA with the further
assumptions that (1) the Levine-Louie dielectric function describes screen-
ing, and (2) the Green function equals the Green function of a fictitious,
insulating free-electron gas. Although this ansatz is guided by physical
intuition, it lacks rigorous justification86*'60 and may elude systematic
improvements.'60 Advantages of the QPLDA are that (1) it is not limited
to minimal basis sets, in contrast to the early work of Hanke et u1.154*155,156
and Fulde et a1.,'57*'58*159 and (2) it leads to only a 30% increase in
computation time compared to LDA and is therefore very efficient com-
pared to the other methods discussed above.
Important physics of the many-body self-energy is not captured in the
QPLDA for two reasons: (1) the energy-dependence of the QPLDA self-

177 J. Grifenstein, H. Stoll, and P. Fulde, Phys. Rev. BS5, 13588 (1997).
'" B. Paulus, private communication.
179 G. Stollhoff and P. Fulde, Z . Phys. B29,231 (1978).
F. Sacchetti, J. Phys. l? Met. Phys. 12, 281 (1982).
"' C. S. Wang and W. E. Pickett, Phys. Rev. Lett. 51, 597 (1983).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 51

energy increases the LDA band gaps and achieves satisfactory agreement
with experiment; and (2) local-field effects are not taken into account. In
more accurate GWA calculations static, inhomogeneous screening effects
open the LDA band gap of semiconductors, and dynamic effects reduce the
gaps to within a few tenths of an eV of the experimental values (see next
sections). Nevertheless, QPLDA band gaps generally agree with experiment
to within 0.4 eV. For example, the direct band gap of Si is 2.99 eV compared
to experimental values of 3.3 to 3.4eV, and the indirect gap equals 0.81 eV
compared to an experimental value of 1.17 eV.162

10. LOCAL-FIELD
EFFECTS AND THE NONLOCALITY
OF THE SELF-ENERGY

The response of an inhomogeneous density distribution to an external

electric field gives rise to microscopic fields that vary on the length scale of
a bond length. These so-called local fields are described by the off-diagonal
matrix elements of the dielectric matrix (Eq. (2.15)) whose diagonal elements
account for homogeneous screening (see Section VI.31 for a detailed
discussion). Local-field effects are often important in semiconductors and
insulators but rarely in metals,
In contrast to local operators-which depend only on one real-space
variable r and act multiplicatively on a wave function -nonlocal operators
such as the self-energy depend on two independent real-space variables and
act on wave functions via a nontrivial real-space integral; that is, they probe
the wave function not only for r = r’ but for all r’ within their range. Figure
9 shows schematically that fX(r, r’; E)cD(r’)dr‘will be large and negative for
the highest valence-band wave function cDo and much smaller in absolute
value for the lowest conduction-band wave function cDc of Si. The extra node
of the conduction-band wave function leads to canceling positive and
negative contributions to the integral and contributes to the discontinuous
jump of self-energy corrections at the band gap (see below).
The combined effect of local fields and the nonlocality of the self-energy
operator on the band gap of solids is significant and cannot be clearly
separated into two distinct contributions. For instance, local-field effects
modify the range of the self-energy. Despite the interdependence of local-
field effects and the self-energy nonlocality, the next two paragraphs discuss
separately the dominant influence of local fields and of nonlocality on
self-energy corrections.

a. Local Fields
A real-space analysis of the static screening potential, Wscr(r,r’) = W(r, r‘;
E = 0) - v(r, r’), around an extra electron centered at a bond center or at
52 WILFRIED G. AULBUR, LARS JBNSSON, AND JOHN W. WILKINS

2- ....,....,....,.-..-

r' (BOND LENGTHS)

FIG.9. Plot of the self-energy Z(r, r'; E = midgap) of Si where r is chosen to lie at the bond
center and r' is varied along the [111] direction. Atomic positions are indicated by closed
circles, and Yu,(Yc) denotes the real part of the highest-lying valence (lowest-lying conduction)
band wave function close to the r point calculated in the LDA. The range of nonlocality of
the self-energy is about one bond length. It follows from the figure that (Y,,lZIYv) is large and
negative, whereas (Y,lZ(Y,) has large positive and negative contributions due to the non-
locality of the self-energy and the nodal structure of the conduction band wave function. Hence,
the self-energy pulls the valence band deeper in energy with respect to the conduction band
and the band gap opens up in comparison to the LDA. (Taken from Ref. 6.)

an interstitial site44 shows that (1) local fields contribute more than a third
of the screening potential in the region near the center of the bond; (2) the
local-field contribution is much smaller (roughly one-tenth) in the low-
density, interstitial regions; (3) local fields are responsible for the anisotropy
of the screening potential; and (4) local fields are short-ranged and their
effect is negligible beyond roughly one bond length. Figure 10 shows the
results for WScr evaluated at E = 0 in analogy to the static COHSEX
approximation to separate the energy dependence of C from local-field
effects.
The screening potential due to an electron at site r evaluated at the same
site determines the Coulomb-hole contribution to the self-energy, which
consequently is constant when local fields are neglected and much deeper in
the bonding region than in the interstitial region when local-field effects are
 QUASIPARTICLE CALCULATIONS IN SOLIDS 53

FIG.10. Contour plot in the Si (110) plane of the screening potential W"' (in Rydberg) in
response to a single electron (indicated by a cross) at (a) a bond-centered or (c) an interstitial
site. Panels (b) and (d) show the corresponding local-field (LF) contributions. Local fields (1)
contribute between 1/3 (bond center) and 1/10 (interstitial) to the total screening potential; (2)
are responsible for the asymmetry of the screening potential; and (3) are short-ranged and
become negligible after one bond length. (Adapted from Ref. 44.)

taken into account. Local fields are more important for valence than for
conduction bands since the valence-band density is concentrated in the
bonding region and the conduction-band density is concentrated in the
interstitial region. The Coulomb-hole and screened-exchangecontributions
to the self-energy can be expressed in terms of the screening potential WScr
as follows (see Eqs. (2.32) and (2.33):

1
F o H ( r , r') = -6(r - r')Wscr(r, r), (2.71)

1,
2
SEX Wscr(r, r') +-
Ir - r'( (2.72)
54 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

where the exchange density n, is defined as

occ

n,(r, r') = C Qi(r)Qf(r'). (2.73)

i

Neglect of local fields leads to homogeneous screening; Wscr(r,r) is constant

and does not contribute to dispersion within a band gap or a given band.
Based on the above remarks about the strength of WScrin bonding and
antibonding regions, one expects both an increase in the fundamental gap due
to CCoHupon inclusion of local-field effects and a strong local-field effect on
the expectation value of CCoHfor occupied but not for unoccupied states. Both
expectations are confirmed by actual calculations, as shown in Tables 6 and 7.
Local-field effects are somewhat weaker for the screened-exchange contri-
bution to the self-energy, since XSEX is dominated by the bare Coulomb
interaction when Ir - r'I is less than a typical bond length, that is, within
the range of the local-field effects. Table 7 shows that for occupied states
local-field effects in XSEX amount to about a third of the local-field effects in

Landolt-Bornstein, Numerical Data and Functional Relationships in Science and Technol-

ogy, eds. K.-H. Hellwege, 0. Madelung, M. Schulz, and H. Weiss, New Series, Group 111, vol.
17a, Springer-Verlag, New York (1982).
lE3 G.Baldini and B. Bosacchi, Phys. Stat. Solid. 38, 325 (1970).

TABLE6. DIRECTAND INDIRECTBANDGAPSFOR THE COVALENTLY BONDED SEMICONDUCTORS

Si AND Ge AND FOR THE INSULATORS DIAMOND AND LiCI, USINGLDA, THE STATIC COHSEX
APPROXIMATIONWITHAND WITHOUTLOCALFIELDS (LF), AND THE GWA IN COMPARISONWITH
EXPERIMENT. THELDA BANDGAP UNDERESTIMATES EXPERIMENT BY AT LEAST0.7 TO 3.4eV.
THE COHSEX APPROXIMATION WITHOUT LOCAL FIELDS LEADS(WITHTHE EXCEPTION OF si) TO
AN INCREASE IN THE FUNDAMENTAL BANDGAPBUT STILL UNDERESTIMATES THE E X P E R I ~ A L
VALUES.THE INCLUSION OF LOCAL FIELDS LEADS TO A FURTHER BAND-GAPINCREASE AND TO
AN OVERE~TIMATION ~ ABOUT0.4 TO 1.1eV. DYNAMICEFFECTS
OF E X P E R IBY (GWA) REDUCE
THIS OvwEsnMATION AND REPRODUCE EXPERIMENT TO WITHIN0.1 TO 0.3eV. NOTE THAT
RELATIVISTIC EFFECTSAND CORE ELECTRONSCAN HAVE A LARGE INFLUENCE ON
SEMICONDUCTOR GAPS,AS SHOWNIN TABLE33. (FROMREF44.)

COHSEX

LDA No LF LF GWA EXPT.'

Diamond 3.9 5.1 6.6 5.6 5.48

Si 0.52 0.50 1.70 1.29 1.17
Ge 0.07 0.33 1.09 0.75 0.744
LiCl 6.0 8.2 10.4 9.1 9.4b

"Ref. 182 unless otherwise noted; bRef. 183.

 QUASIPARTICLE CALCULATIONS IN SOLIDS 55

TABLE7. STATIC COH AND SEX CONTRIFIUTIONS TO THE VALENCE-BAND-

MAXIMUM AND CONDUCTION-BAND-MINIMUM SELF-ENERGIES, WITHAND
WITHOUTLOCAL FIELDS,FOR Si IN COMPARISON TO THE GWA SELF-
ENERGY. LOCAL-FIELD
EFFECTSAFFECT OCCUPIED STATES SIGNIFICANTLY
AND LEADTO ABOUT AN 1.6-eV DECREASE IN ZcoH FOR l-!z5m WHEREAS
TO LOCAL
ZSEXIS LESSSENSITIVE FIELDS(0.6 eV FOR rz,,,0.3 eV FOR X , J .
DYNAMICCORRECTIONS S m THE CONDUCTION AND VALENCE SELF-
ENERGIES UPWARDBY ABOUT 10 TO 20%. IN ABSOLUTETERMS AND FOR
THE TOTALSELF-ENERGY,THE DYNAMIC CORRECTION IS LARGER FOR THE
VALENCE BANDSAND LEADSTO A BAND-GAPDECREASE. (ADAPTED FROM
REF.44.)

COHSEX

Si No LF LF GWA
xCOH

r25v - 8.72 - 10.30 - 8.41

XI, - 8.72 - 8.70 - 7.40
zSEX

ri5" - 4.44 - 3.85 - 3.56

XI, -2.37 - 2.08 - 1.65

XCoH and that local-field effects for unoccupied bands in XSEX are even
smaller. Since local-field effects in X S E X and CCoHare of opposite sign and
amount to at least several tenths of an eV, they must be included in the
determination of both terms to ensure quantitatively accurate results. The
strong effect of local fields on XCoH for the valence band with respect to the
conduction band, and the relative insensitivity of XSEX to local fields, leads
to a band-gap increase of 0.8 to 4.4eV compared to the LDA or the static
COHSEX approximation without local fields.
Table 6 demonstrates that:44 (1) the LDA gives band gaps that are at least
0.5 to 2.0eV smaller than experiment (see also Fig. 3); (2) the COHSEX
approximation without local fields gives band gaps that are in better
agreement with experiment, although significant deviations from experiment
remain; and (3) local fields dramatically open up the band gap and require
inclusion of dynamic screening to achieve quantitative agreement with
experiment.
Local-field effects on the self-energy shifts in simple metals are less than
the numerical uncertainties of GWA calculations (see Section V.26b and Ref.
184) since their density is relatively homogeneous compared to covalently
bonded semiconductors.

J. E. Northrup, M. S. Hybertsen, and S. G. Louie, Phys. Rev. E39, 8198 (1989).

56 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

b. Nonlocality of the Self-Energy

The screened-exchange contribution to the self-energy is proportional to the
nonlocal exchange density (Eq. (2.72)) and hence sensitive to the nodal
structure of wave functions, which results in a discontinuity of the expecta-
tion value of CSEXupon crossing the band gap.185 This discontinuity,
discussed in connection with Fig. 9 and Table 7, shows a 0.3-eV sensitivity
to the inclusion of local-field effects.’86 The Coulomb-hole self-energy (Eq.
(2.71)) does not experience any nonlocal effects upon crossing the band gap
because it is a local operator.
Figure 11 shows that the nonlocality of the self-energy X(r,r’;E) of
common semiconductors in real space is,’dominated by a spherical hole

R. W. Godby, M. Schluter, and L. J. Sham, Phys. Rev. B37, 10159 (1988).

Ia6 The discontinuity of ZSEXequals 1.77 eV with and 2.07 eV without local fields for Si.

FIG. 11. Contour plots of self-energy Z(r, r’; E = midgap) in eV a.u.-3 for r fixed at the bond
center and r’ shown in the (110) plane for (a) Si, (b) GaAs, (c) AIAs, and (d) diamond. For
silicon, the corresponding plots with r fixed at the tetrahedral interstitial site are also shown
(panel (e)). For comparison, the self-energy operator of jellium with rr = 2 (the average density
of silicon) is shown (panel (f)). Godby, Schliiter, and Sham suggested modeling the approxi-
mate form of the nonlocality of the self-energy (not its depth) using the results from jellium
calculations. (From Ref. 185.)
 QUASIPARTICLECALCULATIONS IN SOLIDS 57

centered around r = r’ whose shape-but not its strength or depth-is

roughly independent of the location of r and can be approximated by the
corresponding hole of the jellium self-energy.’” The nonlocality hole of the
self-energy mirrors the nonlocality of the short-ranged screened interaction
W rather than the long-ranged single-particle propagator G. The average
radius of the first nodal surface of the nonlocality in jellium is 2r, (2.1 A for
Si). The extent of the nonlocality is on the scale of a typical wavelength of
a conduction-band wave function, and the hole accounts for more than 99%
of the self-energy expectation value in Si. The nonlocality of the self-energy
contributes significantly to the step-like structure in the self-energy correc-
tions, at the band gap shown in Fig. 12, panels (a) and (b) for Si.IE5Panel
(c) of the same figure shows that this step-like structure is absent in jellium,
which indicates that although the overall form (not the depth) of the
nonlocality hole is well reproduced by jellium at the appropriate density, the
interaction between the wave functions and the nonlocality is not captured
in the homogeneous electron gas.185
Evaluation of the self-energy operator of jellium at the local density of Si
leads to an underestimation of the self-energy nonlocality radius by about
-30% in the bonding region and an overestimation by about 80% in the
interstitial region, as shown in Table 8. This result questions the validity of
basic assumptions in the QPLDA.18’ In addition, the nonlocality hole is too
anisotropic, and it therefore seems unlikely that QPLDA (see Section II.9d)
can give quasiparticle energies of an accuracy comparable to the full GWA
results.

OF THE SELF-ENERGY
11. ENERGYDEPENDENCE

Table 6 shows that the dynamically screened interaction approximation

reduces the 0.8- to 2.2-eV overestimation of direct band gaps in the static
COHSEX approximation to a discrepancy of 0.1 to 0.3 eV between theory
and e ~ p e r i m e n t A
.~~
quantitative prediction of quasiparticle energies re-
quires a correct treatment of the energy dependence of the self-energy. A
simple interpretation of the dynamical effects can be obtained by realizing
that the electron drags a polarization cloud behind it. Screening and the
associated screening energy will depend on the velocity of the electron. The
strongest screening and therefore the lowest energy is obtained in the static
case. Including energy dependence will lead to a screening energy that is not
as low and hence remedies the overestimation of the self-energy in the
COHSEX approximation.
The band-gap reduction results from the linear behavior of the self-energy
as a function of energy close to the quasiparticle energy. Figure 13 shows
58 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

2 - (2) 51
c

0
f;
=
W
-
w
-2

-5 0 I
OW OUASIPARTICLE
EN€RCr ( r V )
FIG. 12. (a) The real part of the matrix elements of the GWA self-energy operator of silicon
(Z(E))and the LDA exchange-correlation potential (V’D”) plotted against the quasiparticle
energy E. (b) The differences (Z(E)- VkD”)(essentially the quasiparticle energy correction)
and (Z(0) - V k D A )of silicon plotted against the quasiparticle energy. The energy dependence
of the self-energy must be taken into account to obtain the correct dispersion for the
quasiparticle energies. (c) For comparison, (Z(E)- V;””) and (Z(0) - V i D ” ) are plotted
against the quasiparticle energy E for jellium at the average density of silicon (r, = 2). The
Fermi energy E , is aligned with the middle of the band gap. Both are relatively featureless and
 QUASIPARTICLE CALCULATIONS IN SOLIDS 59

TABLE8. RANGE rnloe OF THE SELF-ENERGY NONLOCALITY EVALUATED

AT A BONDCENTERED(BC) AND AN INTERSTITIAL (I) SITEIN Si, GaAs,
AlAs, AND DIAMOND COMPARED TO THE CORRESPONDING VALUE OF THE
JELLIUM SELF-ENERGY EVALUATED AT THE FERMIENERGY AND AT (1)
THE AVERAGEDENSITY OF THE SEMICONDUCTOR n(rs), AND (2) THE LOCAL
DENSITY n(r) AT A POINTr IN THE SEMICONDUCTOR. rnlocEQUALS THE
SPHERICALLY AVERAGED DISTANCETO THE FIRST ZERO IN THE
OSCILLATORY FUNCTIONZ(r, r’; E = MIDGAP) WITH r FIXEDAT A BC OR I
SITE. THE SELF-ENERGY OF JELLIUM AT THE AVERAGEDENSITY OF THE
SEMICONDUCTOR UNDERESTIMATESTHE RADIUS rnloc OF THE FULL
CALCULATION ON AVERAGE BY ABOUT 10%. THE “LOCAL-DENSITY”
JELLIUM CALCULATIONS SEVERELY UNDERESTIMATE (OVERESTIMATE) rnlorIN
HIGH (LOW) DENSITY REGIONS (-30% ON AVERAGEFOR A BC SITE,
+ 80% ON AVERAG~FOR AN 1 SITE). (ADAPTED FROM REF. 185.)
rnrocIN JELLIUM (a.u.)
r“l0C

MATERIAL SITE (a.u.) nlrJ n(r)

Si BC 4.1 3.9 2.8

I 4.7 3.9 8.8
GaAs BC 4.2 4.0 2.8
I 4.3 4.0 9.0
AlAs BC 4.3 4.0 2.8
I 4.6 4.0 9.0
C BC 2.8 2.7 2.2
I 3.5 2.7 4.7

that the expectation values of the self-energy operator in diamond, Si, Ge,
and LiCl have a negative slope of about -0.2.44 This slope translates into
nearly constant quasiparticle weights Ziof about 0.8, as shown in Table 9
(Eq. (2.7)), and to about a 20% reduction in the quasiparticle shifts via Eq.
(2.6). However, the quasiparticle weights Ziare close enough to unity that
quasiparticles are well-defined excitations for energies close to the band gap,
although dynamical effects are nonnegligible. This statement is no longer
valid in d and f electron systems, as discussed in Section V.26~.The curves
of X ( E ) bend upward for hole and downward for electron states, which

clearly do not share the sharp discontinuities present in (a) and (b). (d) The real parts of the
matrix elements of the bare-exchange (Hartree-Fock) self-energy operator of Si (Z,) and the
statically screened exchange self-energy operator ( X s s x ) . ( VhD”) is subtracted as in (b). The
nonlocality of Z, has too large a range due to the neglect of screening which leads to a jump
in the self-energy corrections at the Fermi level that is too large in comparison with experiment.
(e) The real parts of the matrix elements in silicon and jellium (r, = 2) of the frequency
derivative of the self-energy ( d Z ( o = midgap)/aw). (Taken from Ref. 185.)
60 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Energy (eV)
FIG. 13. Matrix elements of the electron self-energy operator evaluated in the GWA drawn
as a function of energy for selected states near the band gap are displayed for (a) diamond, (b)
Si, (c) Ge, and (d) LiCI. The quasiparticle energies of these states are indicated on the energy
axis. The self-energy shows an approximately linear dependence on energy close to the
quasiparticle energy for all states and all materials considered. The slope of (Z(E))is roughly
material and state independent and amounts to about -0.2. This indicates that dynamical
effects are important for a quantitative description of quasiparticle energies in solids. (Taken
from Ref. 44.)
 QUASIPARTICLE CALCULATIONS IN SOLIDS 61

TABLE9. QUASIPARTICLE WEIGHTZ, (EQ.(2.7)) FOR THE VALENCE-BAND MAXIMUM (VBM)

AND THE CONDUCTION-BAND MINIMUM(CBM) AT r, X, AND L FOR Si, Ge, GaAs, Sic,
GaN, DIAMOND, AND LICI. Zi FOR THE VBM AND CBM OF DIAMOND AND LiCl IS TAKEN
FROM REF. 44; ALL OTHER VALUES ARE FXOM REF. 42. FOR THE SP-BONDED MATERIALS
SHOWN IN THE TABLE,ziIS EQUAL TO 0.8 TO WITHIN 10% AND IS THEREFOREAPPROXIMATELY
A MATERIAL- AND STATE-INDEPENDENT CONSTANT.NOTE, HOWEVER,THAT COVALENTLY
BONDEDSEMICONDUCTORS, WHICH HAVELARGERDIELECTRIC CONSTANTS, SHOW A STRONGER
DYNAMIC RENORMALIZATION THAN DO IONICINSULATORS.

MATERIAL k-POINT VBM CBM MATERIAL k - m w VBM CBM

Si r 0.79 0.79 Sic r 0.82 0.82

X 0.78 0.80 X 0.82 0.85
L 0.78 0.80 L 0.82 0.83
Ge r 0.79 0.79 GaN r 0.82 0.85
X 0.77 0.80 X 0.83 0.87
L 0.78 0.80 L 0.82 0.86
GaAs r 0.79 0.80 C 0.86 0.86
X 0.78 0.81 LiCl 0.83 0.87
L 0.79 0.81

agrees with calculations of the energy dependence of the self-energy of

j e l l i ~ mCore
. ~ electrons can qualitatively change the energy dependence of
the self-energy, as discussed for the case of Ag in Section V.26a.
Besides band-gap reduction, the energy dependence of C is important in
changing the dispersion of bands, which follows from the comparison of the
expectation value ( C ( E ) - V,,) with the expectation value evaluated at the
mid-gap energy (C(0) - Kc).1s5For Si, Fig. 12 panels (a)-(c) show thatls5
the expectation value ( C ( E ) ) has a pronounced state dependence, which is
largely canceled by subtracting the expectation value of the exchange-
correlation potential; that ( C ( E ) - V $ P A ) is dominated by a jump at the
energy gap and a flat dispersion for both occupied and unoccupied
states; that (Z(0) - V$?’) shows a larger jump at the band gap than
( C ( E ) - V$PA) and a linear dispersion for both occupied and unoccupied
states; and that, with the exception of the jump at the Fermi energy, the
energy dependence can be modeled by the energy dependence of the jellium
self-energy. Hence, the effect of the energy dependence of the self-energy,
besides the reduction in the band gap, is a significant change-from linear
to flat -in the dispersion of the bands.
Self-energy corrections in small- and medium-gap semiconductors align
the theoretical and experimental band gaps by an approximately rigid
shift -the so-called scissors shift. Although the scissors-shift approach
works well in Si, it breaks down in wide-band-gap materials such as Sic and
62 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

0.5
r ‘
1 2 3 4 5 6 7 8 9
I
10
Wave Vector

FIG. 14. Self-energy correction A(k) = (Z(E)- V;””) to the direct band gap at the point
and at 10 special k-points in the fcc Brillouin zone for Si, SIC, and GaN. The standard deviation
from the average self-energy correction (the “scissors” shift) amounts to 0.03 eV for Si, 0.13 eV
for SIC, and 0.17eV for GaN. This corresponds to 4.5%/10.4%/11.7% of the average self-
energy correction (0.73 eV/1.20 eV/1.47 eV for Si/SiC/GaN). Although the scissors approxi-
mation works well for covalently bonded Si, it clearly breaks down for more ionic materials
such as Sic and GaN.

GaN, as shown in Fig. 14. The increased dispersion in the self-energy

correction for S i c and GaN may indicate that the LDA wave functions
approximate less well the quasiparticle wave functions as the band gap
increases. A review of applications of and extensions to the scissors-shift
approach for the determination of optical response functions is given in
Section VI.30. A related observation is that the deviation of the highest
occupied eigenvalue in LDA from the highest occupied DFT eigenvalue,
which equals the highest occupied quasiparticle eigenvalue (see Section 1.3),
increases with increasing ionicity of the material under c ~ n s i d e r a t i o nThis
.~~
deviation amounts to about 0.1 eV for covalent semiconductors such as Si
 QUASIPARTICLE CALCULATIONS IN SOLIDS 63

and Ge. For more ionic materials such as diamond and LiCI, the deviation
is as large as 1.5 eV.
A simplified model of the self-energy of an insulator approximates its
nonlocality and energy dependence by the corresponding quantities of the
jellium self-energy evaluated at the local density in the i n s u l a t ~ r . Figure
'~~
12, panel (e) shows that the variation of the self-energy with respect to
energy can be modeled by results obtained from the homogeneous electron
gas. Assuming that the energy dependence of the self-energy factors out
completely and considering the observations in the previous section, Godby,
Schliiter, and Sham' 8 5 suggested the following physically appealing ap-
proximation for the self-energy:

where g(lr - r'l) and h ( E ) are functions that describe the nonlocality and
energy dependence of the self-energy of jellium. The factor (f(r) + f(r'))/2
accounts for local-field effects. To the best of our knowledge applications of
this formula have been limited to model calculations.

12. CORE-POLARIZATION
EFFECTS

Accounting for the hybridization of valence orbitals with semicore or-

b i t a l ~ and
' ~ ~ the relaxation of the latter (e.g., d orbitals in II-VI materials)
in calculations leads to better agreement with experiment for the structural
properties of II-VI (Refs. 188, 189, 190, and 191) and of some III-V (GaAs,
G a N Refs. 192, 193, and 194) materials. In general, the equilibrium lattice
parameters are increased and the cohesive energies reduced. Density dis-
tributions are changed to such a degree that bond and interface dipoles can
change ~ i g n . ' ~ In ~ .general,
' ~ ~ semicore states should not be treated via a
frozen-core or pseudopotential approach.

I S 7 We define semicore orbitals loosely as d and f orbitals that have a significant overlap with
and are close in energy to valence electrons.
IS8 G. E. Engel and R. J. Needs, Phys. Rev. B41, 7876 (1990).
'13' V. Fiorentini, M. Methfessel, and M. Schemer, Phys. Rev. B 47, 13353 (1993).
A. Nazzal and A. Qteish, Phys. Rev. B 53, 8262 (1996).
S.-H. Wei and A. Zunger, Phys. Rev. B 37, 8958 (1988).
G. B. Bachelet and N. E. Christensen, Phys. Rev. 831, 879 (1985).
N. E. Christensen and I. Gorczyca, Phys. Rev. B 50,4397 (1994).
A. Garcia and M. L. Cohen, Phys. Rev. B 47, 6751 (1993).
W. G. Aulbur, Z. H. Levine, J. W. Wilkins, and D. C. Allan, Phys. Rev. B51,10691 (1995).
D. Cociorva, W. G. Aulbur, and J. W. Wilkins, unpublished.
64 WlLFRlED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Semicore states also significantly affect the electronic and optical proper-
ties of solids, since, for example, (1) band gaps decrease upon inclusion of
semicore states as valence states by about 0.5 eV;'91*'92 (2) the calculated
dielectric constant increases by 10-50% upon inclusion of semicore
state^,'^'*'^^*'^^ and (3) for higher-order optical response functions, such as
the coefficient of second-harmonic generation, inclusion of semicore states
can give an effect sufficiently large to switch their sign.'9s-200Since quasipar-
ticle calculations depend crucially on an accurate description of screening,
a careful treatment of core states is necessary in 11-VI semiconductors, some
111-V materials, alkali metals, alkali earths, and noble metals.
The LDA description of exchange and correlation between valence,
semicore, and same-shell core electrons (e.g., Cd 5s, Cd 4d, and Cd 4p, 4s;
Ref 52) is based on a local, energy-independent functional of the total
density of the system. However, exchange and correlation effects between
core and valence electrons are nonlocal and energy dependent. Dynamic
correlation effects occur when fluctuating core dipoles interact with fluctuat-
ing valence densities. These correlations modify the effective interaction
between valence electrons. Rather than interacting via the bare Coulomb
potential u, valence electrons experience an interaction potential,
W, = EC u, where EC describes the screening by core dipoles.20'
The errors due to the LDA can be eliminated either by a core-polariz-
ation-potential (CPP) approach20'*202*203- which is computationally effi-
cient, since core electrons can still be treated on a frozen-core or
pseudopotential level -or by treatment of core electrons as valence elec-
t r o n ~ . ' ~ * ~The
~ , 'most
~ ' important error introduced by the pseudopotential
approximation is the total neglect of core relaxation, that is, changes in the
core orbitals due to a change in the chemical environment (e.g., crystal-field
effects) and to hybridization of core and valence orbitals. These effects must
be estimated in the C P P approach, whereas they are included on an LDA
level when core electrons are treated explicitly. Also, the semicore d states
of 11-VI materials energetically overlap the valence band and hence cannot
be eliminated via a core-polarization potential. Explicit inclusion of the d
electrons in the valence band amounts to a significant increase in computa-
tional cost for plane-wave-based quasiparticle calculations204 such as the

19' H. R. Philipp and H. Ehrenreich, Phys. Rev. 129, 1550 (1963).

19* T. Tomoyose, J . Phys. SOC.Jpn. 63, 1149 (1994).
19' S. H. Wemple and M. DiDomenico, Jr., Phys. Rev. B 3 , 1338 (1971).
,O0 B. F. Levine, Phys. Rev. B 7 , 2591 (1973).
E. L. Shirley and R. M. Martin, Phys. Rev. B47, 15413 (1993).
,02 W. Miiller, J. Flesch, and W. Meyer, J . Chem. Phys. 80, 3297 (1984).

203 W. Miiller and W. Meyer, J . Chem. Phys. 80, 3311 (1984).

,04 Only one plane-wave calculation has been reported so far for ZrO, that includes the Zr
4s, 4p, and 4d core and semicore states in addition to the 5s electrons as valence electrons (see
Ref. 57).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 65

one used in Ref. 43. Other approaches that describe semicore states via only
nonlinear core corrections205have been shown to give band structures in
good agreement with experiment for 11-VI materials.51 However, this
agreement must be considered coincidental, since it relies upon error
cancellation between core-relaxation and core-valence exchange and polar-
ization effects.

a. Core-Polarization Po ten tial

Dynamical intershell correlation between semicore and valence electrons
leads to induced polarization of the ion cores by the valence electrons. It can
be taken into account in the valence-electron Hamiltonian by adding the
energy contribution of all induced core dipoles in the electric field of the
valence electrons and of the core dipoles, excluding s e l f - i n t e r a c t i ~ n . ~ ~ ~ + ~ ~
Core-polarization functions are approximated by static, atomic polarizabil-
ities aJ ( J = Jth core at location RJ).201*202One obtains in a point dipole
picture,"

(2.75)

where Ej is the electric field at R,

(2.76)

and ri is the location of the ith valence electron and 2, the atomic number
of the Jth core. The modification of the original valence Hamiltonian Vcpp
can now be written aszo1

(2.77)

Besides the standard -a/(2r4) term (t;

see below), the electric fields due
to the ions and valence electrons introduce (1) an additional term f i - I in
the ion-ion interaction of the total energy of the system; (2) an additional
local potential k - I ,which is felt by every electron; and (3) an additional

'05 S. G. Louie. S. Froven. and M. L. Cohen. Phvs. Rev. B26. 1738 (1982).
66 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

interaction V,- between two valence electrons due to core-polarization

A

effects.z0’~202~z03 The first two effects are easy to incorporate into standard
band structure calculations. The contributions due to the valence electrons
need further modifications, since valence electrons cannot be described as
point charges. For instance, the single-electron potential for one valence
electron far away from a core of polarizability a reduces correctly to the
classical result - 0 1 / ( 2 r ~ ) .Ho
’ ~ ~wever, valence electrons can penetrate the
core where the classical result diverges. Accordingly, a cut-off function has
to be i n t r o d u ~ e d , and
~ ~ ~a ’typical
~ ~ choice is

(2.79)
The sum over 1 in the above equations is a sum over angular momenta, and
is the corresponding projection operator. For each atomic species one has
to determine the parameters 01, r I , and Fe-,. Shirley et al.43used the experi-
mental core polarizabilityZo7 and varied the rl (1 = 0, 1, 2) to obtain the
correct removal energy for one valence electron of angular momentum 1
outside the ion core (rl = rz for 12 3) and Fe-e = 0.5(r0 rl). +
The two-electron potential k-eis an effective interaction between two
valence electrons due to core-polarization effects. This term screens the bare
Coulomb interaction, which screening has to be taken into account in
quasiparticle calculations. The effective interaction between valence elec-
trons can be expressed as4943

(2.80)

where { x J } are the self-consistent density-response functions of isolated

cores.43 The effective interaction W, replaces the bare Coulomb interaction
in the determination of the valence-valence self-energy in the GWA and in
the plasmon-pole sum
Shirley et al.43 used this modified GWA approach and evaluated the
self-energy corrections using LDA wave functions and energies. Self-consist-
ency in the quasiparticle energies is achieved and core relaxation effects are

’06 M. Born and W. Heisenberg, Z . Phys. 23, 388 (1924).

’07 Calculations of a in the time-dependent Hartree-Fock approximation, that is, including
excitonic effects, have been reported as well (Refs. 67 and 201).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 67

added a posteriori by comparison of LDA all-electron and pseudopotential

calculations.
Core-polarization-potential calculations give accurate quasiparticle ener-
gies for materials with semicore states that do not energetically overlap with
the valence bands and, in particular, describe Ge correctly as an indirect
rather than a direct emi icon duct or.^^ Table 10 shows that for Ge and GaAs,
the fundamental band gaps are significantly improved compared to GWA
calculations that treat core-valence interactions on an LDA level.2l o

208 "Crystal and Solid State Physics", in Landolt-Bornstein, Numerical Data and Functional
Relationships in Science and Technology, vol. 17a, ed. by 0. Madelung, Springer, Berlin (1984).
2 0 9 As cited in Ref. 43.

210 Note that the correct ordering of the lowest conduction bands in Ge can also be achieved

in calculations that have no self-interaction errors (Refs. 211 and 212). The energies of the
lowest-lying conduction bands in these methods deviate significantly from experiment (,!IcB at
L - r - X = 0.10 eV - 0.12 eV - 0.53 eV (Ref. 21 l), 1.01 eV - 1.28 eV - 1.34 eV (Ref. 212) versus
experimental values of 0.74 eV - 0.90 eV - 1.2 eV given in Table 10.

TABLE10. BANDENERGYDIFFERENCES IN Ge AND GaAs IN ev. R e x r ~ T sFOR LDA FULL-

POTENTIALCALCULATIONS, QUASIPARTICLE
(QP) CALCULATIONS WITH AN LDA TREATMENT OF
CORE-VALENCEINTERACTIONS (CVI), QUASIPARTICLECALCULATIONS WITH A CORE-
POLARIZATION-POTENTIAL-BASED(CPP) TREATMENTOF CORE-VALENCE INTERACTIONS, AND
EXPERIMENTARE SHOWN.ALL RESULTSINCLUDECORE-RELAXATION EFFECTS(ADDELI A
POSTERIORI FOR THE QUASIPARTICLECALCULATIONS). FUNDAMENTAL BAND GAPS IN THE
D ~ R E N TAPPROACHES ARE PRINTED IN BOLDFACE.OVERALL AGRE~MENT BETWEEN
EXPERIMENT AND QUASIPARTICLECALCULATIONS IS IMPROVED UPON INCLUSION OF CORE-
POLARIZATION IN PARTICULAR,
EFFECTS. Ge IS PREDICTED TO BE AN INDIRECT RATHERTHAN A
DIRECT GAPSEMICONDUCTOR. (ADAPTED FROM REF. 43.)

QP

CVI IN CVI IN
QUANTITY LDA LDA CPP ExF-T."

-0.26 0.53 0.85 0.89

0.55 1.28 1.09 l.lOb
-0.05 0.70 0.73 0.744
0.60 0.58 0.36 0.36
GaAs
'8" --t '6r 0.13 1.02 1.42 152
'SO x6c 1.21 2.07 1.95 2.01
'8" + L6c 0.70 1.56 1.75 1.84
L6c --* x 6 c 0.51 0.52 0.20 0.17
x,, + x7c 0.21 0.26 0.33 0.40

"Ref. 208 unless noted otherwise; bRef. 209.

 68 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

b. Explicit Treatment of Core Electrons

Local-orbital basis function^'^ such as Gaussian orbitals (Refs. 52, 140, 141,
213, and 214) or LMT0'45*215allow the explicit treatment of core electrons
on the same level of approximation as the valence electrons. They include
relaxation and hybridization effects on the LDA level and exchange and
dynamic screening effects of core and semicore electrons on the GWA level.
The advantage of local-orbital basis functions compared to plane-wave
methods is the drastically reduced number of basis functions.
Including semicore d states as valence electrons in non-self-consistent
GWA calculations for 11-VI materials leads to disappointing
First, the LDA d levels are shifted upwards in energy, that is, away from
experiment; second, the self-energy operator is nondiagonal in the basis of
LDA states. Unphysical self-interactions in LDA -that is, incomplete
cancellations between Hartree and exchange potentials -shift the d bands
up in energy and lead to a too strong hybridization with the anion p bands.
As a consequence of the unphysical p-d mixing, the LDA states are not good
approximations to quasiparticle states and the self-energy is nondiagonal.
An iterative determination of quasiparticle wave functions and energies in
principle eliminates the self-interaction error contained in the wave func-
tions and energies. However, self-consistency alone is not sufficient, since the
exchange interaction with the 4s and 4p electrons must be taken into
account as well.52
The localized semicore d states strongly overlap with the wave functions
of the same-shell s and p electrons, which leads to large exchange energies
between the corresponding state^.^' Treatment of the s, p , and d electrons
as valence electrons changes the non-self-consistent self-energy operator
qualitatively5' in three ways: (1) quasiparticle energy corrections lower the
semicore d states and lead to improved agreement with experiment; (2) the
self-energy becomes diagonal in the LDA basis; and (3) in the case of CdS,
the exchange between 4s and 5s electrons increases the band gap by 0.85 eV.
The energy of the semicore 4d level is unchanged (E4d= - 7.4 eV) regardless
of whether 4d, 4d and 4p, or 4d, 4p, and 4s electrons are treated as valence
electrons." This suggests that hybridization between cation d and anion p
orbitals is not affected significantly by the s and p core electrons. Remaining
self-interaction errors in the wave function may cause errors in the non-self-
consistent GWA calculations on the order of about 1 eV.

'I1 M. M. Rieger and P. Vogl, Phys. Rev. B52, 16567 (1995).

212 M. Stadele and A. Gorling, private communication.
'13 M. Rohlfing, P. Kruger, and J. Pollmann, Phys. Rev. B56, R7065 (1997).
'14 M. Rohlfing, P. Kriiger, and J. Pollmann, Phys. Rev. B57, 6485 (1998).
'I5 F. Aryasetiawan, Physica B 237-238, 321 (1997).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 69

As shown earlier, nonlocal exchange and inclusion of screening effects due

to shallow core states are important and affect quasiparticle proper-
ties.43.197*198Exchange can have qualitative effects on the energy depend-
ence of the self-energy, as demonstrated by Horsch, von der Linden, and
Lukas216 and discussed in Section V.26a. Moreover, the results of Ref. 52
clearly illustrate the qualitative importance of exchange for the determina-
tion of band structures. The qualitative accuracy of the results in Ref. 52 is
less certain, however, because of technical difficulties that are detailed in
Section 111.15a.
Approximate self-consistent GWA calculations improve the position of
the d bands in comparison to experiment,213 although (1) the discrepancy
between theory and experiment for energy gaps is increased upon self-
consistency -that is, the improvement of quasiparticle energies is not
uniform; (2) vertex corrections have not been included although they may
be important; and (3) self-consistency increases discrepancy with experiment
for all other known calculations (see the next section) with the exception of
total energies and charge conservation. The most drastic example of point
(1) is CdS:213The 4d level equals -7.2 eV in LDA,’17 -9.1 eV in a partially
self-consistent GWA calculation, and -9.2/-9.5 eV in experiment, as
shown in Table 11. The corresponding values for the band gap are 2.45 eV,
3.21 eV, and 2.50/2.55 eV. Partial self-consistency leads to an overestimation
of the direct band gap by about 0.7eV. Vertex corrections (point (2))
increase calculated RPA core polarizabilities by about 30%67and may be
important for semicore levels as well.
Semicore states in ZnSe, GaAs, and Ge have been determined within
standard GWA by Aryasetiawan and G u n n a r ~ s o n . ’As ~ ~summarized in
Table 11, standard GWA calculations describe semicore levels within a
5- 15% range. The remaining discrepancies with experiment are partially
due to the lack of self-consistency, vertex corrections, and self-interaction
errors in the LDA wave functions.

13. SELF-CONSISTENCY

General Remarks. Most current GWA calculations have three characteris-

tics: (1) the Green function is calculated in an appropriate single-particle
basis following the “best G, best W ” philosophy (Hartree-Fock, LDA, or in
some cases empirical pseudopotentials122~1 (2) the screened interac-
239224);

tion is determined in RPA using the chosen single-particle wave functions

216 P. Horsch, W. von der Linden, and W.-D. Lukas, Sol. State Comm. 62, 359 (1987).
2’7 The quoted value in Ref. 52 is slightly diBerent and equals -7.4 eV.
70 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

11. CALCULATED
TABLE BINDING
ENERGIES
IN eV OF THE SEMICORE
STATESCd 4d IN
Cumc CdS, Ge 3d IN Ge, AND Si 2p IN Si. THEZEROOF THE ENERGYSCALEIS GIVEN
BY THE RESPECTIVEVALENCE-BAND MAXIMUM. THE EXPERIMENTAL DATAARE
WEIGHTED AVERAGESOF SPIN-ORBIT-SPLIT LEVELSAND ARE T m AS QUOTED IN
REF.213. THES ~ LGLDA S AND GQPSCORRESPOND TO GWA CALCULATIONS USING
AN LDA PROPAGATOR OR A PROPAGATOR CALCULATED m A RENORMALIZED
SPECTRALFUNCTION (APPROXIMATELY SELF-CONSISTENTGWA CALCULATION; SEE
REF. 213). ALL VALUESARE TAKEN FROM REF.213 mTHE FOLLOWING EXCEPTIONS:
IN THE CASEOF Ge WE ALSO LISTLDA, GLDAWA, AND SLATER-TRANSITION-STATE
(STS) RESULTS OBTAINED BY AN LMTO FORCdS, WE LISTTHE
LDA A M ) SELF-INTERACTION- AND RELAXATION-CORRECTED PSEUDOPOTENTIAL
(SIRC-PSP) &.SULTs OF REF. 218.

GWA ’

LDA GLDA GQPS Em.

Cd 4d -7.2 -8.1 -9.1 -9.2,’ -9.5’

-6.8” -9.7“Vb
Ge 3d -24.6 -27.7 - 30.0 - 29.5,’ -29.7h
-24.4‘ -28.5‘ -30.1‘*’
Si 2p - 89.4 -94.8 - 100.4 -99.0,’ - 100.0*

“Ref. 218; bSIRC-PSP ‘Ref. 145; dSTS; ‘Ref. 219; fRefs. 220 and 221; 8Ref. 222;
hRef. 223.

and energies; and (3) in the majority of cases the frequency dependence of
the dielectric response is further approximated by plasmon-pole models.
This approach, which we call the “standard” GWA, or GoWoA in what
follows, works well in practical applications and generally leads to good
agreement with experiment for quasiparticle energies. However, from a
principle point of view it has several shortcomings.
A consistent approach to GWA calculations evaluates the RPA polariz-
ability and the GWA self-energy using Hartree wave functions and energies
and iterates the calculation until self-consistency in G and W is achieved.

’18 D. Vogel, P. Kriiger, and J. Pollmann, Phys. Rev. B54, 5495 (1996).
A. P. J. Stampfl, P. Hofmann, 0.SchalT, and A. M. Bradshaw, Phys. Rev. B 55,9679 (1997).
220 L. Ley, R. A. Pollak, F. R. McFeely, S. P. Kowalczyk, and D. A. Shirley, Phys. Rev. B9,
600 (1974).
’” N. G. Stoffel, Phys. Rev. B28,3306 (1983).
2 2 2 Zahlenwerte und Funktionen aus Naturwissenschajien und Technik, eds. A. Goldmann and
E. E. Koch, Landolt-Bornstein, New Series, Group 111, vol. 23a, Springer-Verlag,Berlin (1989).
223 W. Monch, in Semiconductor Surfaces and Interfaces, eds. by G. Ertl, R. Gomer, and D.
Mills, Springer Series in Surface Sciences, vol. 26, Springeer-Verlag,Berlin (1993).
224 W. von der Linden and P. Horsch, Phys. Scripta 38,617 (1988).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 71

Non-self-consistent results for quasiparticle energies obtained starting from

Hartree theory differ from, for example, LDA-based “best G, best W”
approaches, since Hartree wave functions and energies are qualitatively
different from the corresponding LDA quantities.88 Self-consistencyelimin-
ates the dependency of the final results on the starting wave functions and
energies, as shown explicitly for a 1D semiconducting wire by de Groot,
Bobbert, and van H a e ~ i n g e n . ~ ~ ’
An approximate procedure for incorporating self-consistency into a
standard GWA calculation is to shift the energy spectrum of the indepen-
dent-particle Green function so that the Fermi levels of the interacting
system and the noninteracting system are aligned.3.75This approximation
(1) leads to an increase of fundamental band gaps by 0.1 to 0.2eV in
semiconductors and insulator^;^^ (2) improves the satellite spectrum of a
Hubbard cluster significantly;z26(3) leads to improved charge conservation
in GWA calculations;2z7and (4)is necessary for systems like NiO to obtain
reasonable agreement with e ~ p e r i m e n t A . ~ slightly
~ modified version for
insulators (see, for instance, Ref. 46) shifts the independent-particle valence
(u) and conduction (c) bands in a linear, isotropic manner until they agree
with the corresponding quasiparticle energies:
E,k = (1 + A,)E,k + B,, E,k = (1 + Ac)Eck + B,. (2.81)
Here A iand Bi,i = u, c, are fitting parameters determined via the band gaps
at two high-symmetry points. Note, however, that Ricezz8advised against
using a shifted self-energy since the energy dependence of the self-energy is
largely canceled by other self-consistencyeffects. In constrast to the energies,
the independent-particle wave functions are left unchanged and not replaced
by the interacting spectral function. The corresponding assumption that no
significant quasiparticle weight is transferred to the satellite spectrum
contradicts experiment since 10-50% of the total spectral weight resides in
the incoherent background and the plasmon satellites.
Fleszar and Hanke”’ examined the quasiparticle energy shifts in Si as a
function of energy (- 15 eV < E < 70 eV) for a non-self-consistent GWA
calculation that (1) takes the frequency dependence of the screened interac-
tion fully into account and (2) accounts for finite lifetime effects, that is, the
finite imaginary part of the self-energy.In most GWA calculations, point (1)
is treated approximately via plasmon-pole models and point (2) is neglected.
Figure 15 (bottom panel) shows that the neglect of the imaginary part of the

’” H. J. de Groot, P. A. Bobbert, and W. van Haeringen, Phys. Rev. B52,11OOO (1995).

226 T. J. Pollehn, A. Schindlmayr,and R. W. Godby, J . Phys. Cond. Mat. 10, 1273 (1998).
’*’A. Schindlmayr,Phys. Rev. B56,3528 (1997).
”* T. M. Rice, Ann. Phys. 31, lOO(1965).
72 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

u
J

1.6

5 1
Y
4 1.8
41B 1
4

0.8

4.6

-1
- 1 0 0 i o m 5 o a m 6 0 7 0
ELDA(ev)

FIG. 15. Top panel: Silicon quasiparticle energy shifts E,, - EL,, as a function of the LDA
energy E,,,. The overall shape of the energy shifts is similar to the shape of quasiparticle
energy shifts in jellium4~z29(not shown) and Ag (see Fig. 42 and Section V.26a). Bottom panel:
Diagonal matrix elements of the imaginary part of the Si self-energy with respect to LDA
orbitals calculated at the energy of the corresponding quasiparticle peak. The shape of the
imaginary Si self-energy is similar to the shape of the non-self-consistent decay rate of jellium
(see Fig. 18). In both figures, the full circles are the result of a standard (RPA) GWA
calculation, the empty circles have the (time-dependent) LDA vertex function included (GWT
calculations, see Section 11.14). The self-energy is calculated along the L-T-X line. (From Ref.
102.)
 QUASIPARTICLE CALCULATIONS IN SOLIDS 73

self-energy is only justified in the immediate vicinity of the band gap. As can
be seen in Fig. 15 (top panel), quasiparticle energy shifts arelo2 linearly
decreasing for occupied states, roughly constant for conduction band states
with 0 IE < 25 eV, and roughly linearly increasing for energies E 2 25 eV.
Equation (2.81) does not capture the behavior of quasiparticle corrections
over an extended energy range. The overall shape of the energy dependence
of quasiparticle corrections in Si is reminiscent of the behavior in j e l l i ~ m ~ ~ ~
and Ag (see Section V.26a and Ref. 216).
A truly self-consistent GWA calculation not only eliminates the depend-
ency of results on the initial starting point of the iteration but is also charge
and energy conserving in a Baym-Kadanoff ~ e n s e . ~ It' . has~ ~ five disadvan-
tages: (1) the self-consistent fundamental band gap of Si ( x 1.9 eV) exceeds
experiment (1.17eV) by as much as LDA (x0.5eV) falls below experi-
ment;230 (2) the band width of simple metals is larger than independent-
particle results and contradicts experiment; 2 3 0 * 2 3 (3) plasmon-pole
19232

peaks in the spectral function are strongly suppressed, in contrast to

experiment (Refs. 230, 231, 232, and 233); (4) the screened interaction W
does not fulfill the f-sum rule and is only a mathematical tool to determine
the self-energy-that is, it has no physical meaning;231*232 and ( 5 ) self-
consistent GWA calculations are computationally very demanding. These
points will be discussed in more detail below.
The increasing discrepancy between theory and experiment upon full
self-consistency clearly indicates that missing physics in the form of vertex
corrections must be included in dynamically screened interaction calcula-
tions and that inclusion of self-consistency alone is not justified due to the
increase in computational cost and the loss of predictive power. The success
of standard GWA calculations seems to result from error cancellation
between the lack of vertex corrections and the omission of self-consistency.
Applications. Most fully self-consistent GWA calculations have been
reported for model systems: a one-dimensional semiconducting wire,225a
Hubbard cluster,234and the homogeneous electron Self-consist-
ent calculations for real solids were reported very recently by Schone and
E g u i l ~ z . ~Partially
~' or approximate self-consistent calculations for the
same model ~ y ~ t e m and~for ~the ~semiconductors
~ * ~ ~ Si ~and*C' ~l 9 have~ ~

229 K. W.-K. Shung, B. E. Sernelius, and G. D. Mahan, Phys. Rev. B36,4499 (1987).
230 W. D. Schone and A. G. Eguiluz, Phys. Rev. Lett. 81, 1662 (1998).
2 3 1 B. Holm and U . von Barth, Phys. Rev. B57, 2108 (1998).
232 E. L. Shirley, Phys. Rev. B54, 7758 (1996).
2 3 3 A description of (multiple) satellites that improves upon the GWA can be obtained by
cumulant expansions (Ref. 77) or by a T-matrix approach (Ref. 71). For a review. see Ref. 14.
234 A. Schindlmayr, T. J. Pollehn, and R. W. Godby, Phys. Rev. B58, 12684 (1998).
2 3 5 A. G. Eguiluz and W.-D. Schone, Mol. Phys. 94, 87 (1998).
74 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

been reported as well (see below). In particular, the calculations by Shir-

ley232 and by von Barth and Holm'" determined the propagator self-
consistently but calculated the screened interaction from the independent-
particle wave functions and energies. We call this approach the GWOA
approach, in contrast to the non-self-consistent GoWoA and fully self-
consistent GWA approaches.
Results. Fully self-consistent GWA calculations in jellium find spectral
weight functions whose plasmon satellite weight is significantly reduc-
ed,231*235 as shown in Fig. 16. This trend is in clear contradiction to
experiments on simple metals, which show a narrow plasmon satellite at wpl
below the quasiparticle energy and smaller plasmon satellites further down
in energy. Note that the spectral function of Ref. 235 has discernible
plasmon peaks rather than a featureless background, as in the spectral
function of Ref. 231. This difference may result from technical differences in
the calculations (finite T versus T = 0 approach, fitting of the spectral
function versus numerical evaluation, Loss of plasmon spectral
weight translates into an increase in quasiparticle weight, as shown in Fig.
17. Also, the broad and featureless plasmon spectrum of Ref. 231 reduces
inelastic quasiparticle-plasmon collisions. Hence, the quasiparticle decay
rate of self-consistent GWA calculations shown in Fig. 18 does not increase
strongly due to the onset of plasmon production -which is what is
physically expected-in contrast to the GOW'A and the GW'A decay rates.
Figure 19 shows that the self-consistent GWA band width of j e l l i ~ m ~ ~ '
is larger rather than smaller than the independent-electron band
~ i d t h . ~The ~ ~latter
* ~is ~already
~ . ~too~ large
~ in comparison to the
experimental band width of simple metals. To explain the discrepancy
between theory and experiment, several combinations of vertex correc-
tions' 84.232.240 and the inclusion of surface effect^^^'.^^' have been sugges-
ted, as discussed in the next section and Section V.26b. A generally accepted
explanation for this discrepancy has not yet emerged.
Total energy calculations of jellium in the energy-conserving self-consist-
ent GWA agree with Monte-Carlo data to within 25% (I, = 2) and 1%
(I, = 4).231Schindlmayr, Pollehn, and G ~ d b noted y ~ ~ in this
~ context that
the total energy of a finite, two-leg Hubbard ladder is always raised by

236 H. J. de Groot, R. T. M. Ummels, P. A. Bobbert, and W. van Haeringen, Phys. Rev. 854,
2374 (1996).
237 U. von Barth and B. Holm, Phys. Rev. 854,8411 (1996).
A. G. Eguiluz, private communication.
239 B. Holm and U. von Barth, private communication.
240 G. D. Mahan and B. E.Sernelius, Phys. Rev. Left. 62,2718 (1989).
241 K. W.-K. Shung and G. D. Mahan, Phyx Rev. Left. 57, 1076 (1986).
 75

-20 10 0 -10 -20

-ray (ev)
FIG. 16. Fully self-consistent one-particle spectral function A(k, o) for the homogeneous
electron gas evaluated at k = k,, r, = 4, T = 0 (Ref. 231, upper panel), and at k = 0.99kF,r, = 5,
finite T (Ref. 235, lower panel). The results in the upper panel are compared with a partially
self-consistent spectral function obtained from a GW'A calculation.237 Those in the lower
panel are compared with results obtained after the first and third iteration toward self-
consistency. Both calculations show that self-consistency (1) sharpens and increases the weight
of the quasiparticle peaks (see also Fig. 17), and (2) reduces the spectral weight of the plasmon
satellite peaks. The two self-consistent calculations differ in the shape of the plasmon peaks.
76 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

FIG. 17. The quasiparticle weight Z , for jellium at rs = 4 determined by a fully self-
consistent GWA calculation (solid line), a partially self-consistent GW’A calculation (dashed
line), and a standard G’W’A calculation (dotted line). Self-consistency leads to a systematic
increase in quasiparticle weight and to a loss of structure due to the suppression of plasmon
satellites in the fully self-consistent result. (From Ref. 231.)

self-consistent calculations relative to standard GWA calculations. This is

true not only for RPA-based GWA calculations but also for GWA calcula-
tions that include vertex corrections and for a Bethe-Goldstone approach
based on the T - m a t r i ~ . ’Depending
~~ on the approximation used and the
parameters of the model, the self-consistent energy can even be in worse
agreement with experiment than the non-self-consistent energy. Schin-
dlmayr, Pollehn, and G ~ d b y ’interpreted ~~ this fact as an indication that
the excellent total energies obtained for the electron gas with self-consistent
GWA may be fortuitous. The qualitative changes in the Hubbard-ladder
spectral function at self-consistency (broadening of high-energy plasmons,
sharpening of quasiparticle features) agree with the findings of Holm and
von BarthZ3land of Eguiluz and S ~ h O n e . ’ ~ ~
The effects of self-consistency for real materials are qualitatively similar to
those observed in j e l l i ~ r n . ’ A
~ ~self-consistent GWA c a l ~ u l a t i o n ’overes-
~~
timates the valence band width of potassium by more than 1 eV (2.64eV
compared to 1.60 eV in experiment, 2.21 eV in LDA, and 2.04 eV in standard
GWA). Self-consistency increases the quasiparticle weight by 20% (0.72
compared to 0.60 non-self-consistently). Results for Si are equally dra-
 QUASIPARTICLE CALCULATIONS IN SOLIDS 77

FIG.18. Decay rate r, of jelliurn at r, = 4 determined by a fully self-consistent GWA

calculation (solid line), a partially self-consistent GW’A calculation (dashed line), and a
standard G’W’A calculation (dotted line). The two latter results show a region with a much
larger decay rate (shorter lifetime) caused by inelastic collision with plasmons. This physically
correct feature is absent from the fully self-consistent GWA result. (From Ref. 231.)

ma ti^:'^' The direct band gap is 4.02 eV in self-consistent GWA compared

to 3.4 eV in experiment (2.57 eV in LDA, 3.27 eV in standard GWA), and
the indirect band gap is 1.91 eV compared to 1.17eV experimentally
(0.53 eV in LDA, 1.34eV in standard GWA). The self-consistent occupied
band width (13.1 eV) is marginally in better agreement with experimentz4’
(12.5 f 0.6 eV) than the result of a standard GWA calculation (11.65eV,
11.93 eV in LDA).
Total Energy and Density. Early total-energy calculations for jelliumz4’
and bulk Si244based on the Galitskii-Migdal were reported by
Levin ef al.243and Farid ef al.744 In addition, Rieger and G ~ d b y recently
’~~
determined the density of Si and Ge in the GWA using the real-space/

242 For a discussion of the experimental value of 12.5 eV, see Section III.18a.
243 Y. Levin, C.D. Wu, and Y . Bar-Yam, Comp. Mat. Sci. 3, 505 (1995).
244 B. Farid, R. W. Godby, and R. J. Needs, in 20th International Conference on the Physics
of Semiconductors, eds. E. M . Anastassakis and J. D. Joannopoulos, vol. 3, World Scientific,
Singapore (1990), 1759.
245 V. Galitskii and A. Migdal, Zh. Eksp. n o r . Fiz. 34, 139 (1958) [Sox Phys. JETP 7, 96
(1958)l.
246 M.M. Rieger and R. W. Godby, Phys. Rev. 858, 1343 (1998).
78 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

4.2-

Y -
-0.4
e
ur' 4.6 -

-1.4 - ' I
0.0 02 0.4 0.6 0.8 1.0 12
WkF
FIG. 19. Quasiparticle dispersion Ek for a fully self-consistent GWA calculation (solid line),
a partially self-consistent GW'A calculation (dashed line), and a standard G'W'A calculation
(dotted line), and for noninteracting electrons (dashed-dotted line). Only the standard G'W'A
calculation leads to band-width narrowing in comparison to the free-electron result. However,
this observed band-width narrowing is too small to explain experimentally observed narrow
band widths in simple metals. Self-consistency increases the band width in comparison to free
electrons and hence leads to increased deviation from experiment. Data shown correspond to
r, = 4 and are taken from Ref. 231.

imaginary-time algorithm. In the latter work, the Hartree potential of Eq.

(1.3) is iterated to self-consistency in the presence of a non-self-consistent
GWA self-energy. With Go as the LDA Green function, AV, as the change
in the Hartree potential due to a change in density, and A as a constant shift
that aligns the Fermi level of the noninteracting and interacting Green
functions (see Eq. 2.81), the interacting Green function can be obtained via
Dyson's equation247(Eq. (2.5)):
G(io) = Go(io) + Go(io)[C(io) + AV, - V, - A] G(io), (2.82)
from which the density follows via
3 ro
n(r) = - do ImG(r, r; a). (2.83)
R -00

247 All quantities are to be interpreted as matrices with respect to plane-wave coefficients (see
Section III.15b).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 79

Both equations must be solved self-consistently, since a change in the

density yields a change in the Hartree potential. Within this approach248
Rieger and Godby found that (1) density is conserved to within 0.3% in Si
and 0.05% in Ge; (2) LDA and GWA structure factors differ by less than
0.1YOand agree well with experiment; and (3) a self-consistent Hartree
potential leads to band structure changes of less than 0.1 eV.

14. VERTEXCORRECTIONS

As follows from Eq. (2.9), vertex corrections to the self-energy describe the
linear response of the self-energy to a change in the total electric potential
of the system. Vertex corrections in dielectric screening describe exchange
and correlation effects between an electron and other electrons in the
screening density. For example, if screening is provided by a spin-up
electron, other spin-up electrons cannot get too close and are less likely to
help with the screening." As a consequence, the screening is weakened and
the interaction is strengthened. Such short-ranged vertex corrections im-
prove the description of quasiparticles and of low-energy satellites. Other,
long-ranged, vertex corrections improve the description of the plasmon
satellites, as discussed in detail in Refs. 14 and 249.
Vertex corrections increase correlation functions such as the density-
density response function, whereas self-energy insertions in the Green
function, that is, self-consistency corrections, reduce correlation functions.
DuBois250*25'was the first to notice that, within the context of the GWA,
vertex-correction diagrams and self-energy diagrams cancel each other to a
large degree. Because of this cancellation, discrepancies between experiment
and self-consistent GWA calculations, such as those described in the
previous section, are to be expected. In principle, a consistent treatment of
vertex corrections and self-consistency is required for a quantitative descrip-
tion of experiment. M a h a n " ~suggested
~ ~ ~ on the basis of the work of Baym
and Kadanoff and Ward,253 that a consistent procedure to include
62963

vertex corrections in the polarizability P can be obtained as follows. The

vertex diagrams associated with a self-energy Z -X governs the equation of
motion of the interacting Green function G- that should be included in the

248 Special care must be taken to account for core charges that are missing in the pseudopoten-
tial approach of Ref. 246.
249 C. Verdozzi, R. W. Godby, and S. Holloway, Phys. Rev. Lett. 74, 2327 (1995).

D. F. DuBois, Ann. Phys. 7, 174 (1959).

D. F. DuBois, Ann. Phys. 8, 24 (1959).
2 5 2 G. D. Mahan, Int. J . Mod. Phys. 8 6 , 3381 (1992).

2 5 3 J. C. Ward, Phys. Rev. 78, 182 (1950).

80 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

determination of correlation functions are obtained by attaching an external

interaction line to the internal Green function lines of the self-energy
diagrams. As a consequence, a consistent conserving approximation for, for
example, the inclusion of ladder diagrams in P requires the use of the
self-consistent single-particle propagator G rather than the independent-
particle propagator As discussed below, most of the self-consistent or
vertex-corrected calculations are at variance with Mahan’s suggestion in one
way or another.
Early work on vertex corrections in the homogeneous electron gas is not
reviewed here (for a general review, see Ref. 10). The present focus is on
work related to the effect of vertex corrections255on the electronic proper-
ties of insulator^^^^"^ as well as on recent work on the homogeneous
electron gas and simple metals (see Refs. 232, 240, 256, 257, and 258).
Generally, vertex corrections have been studied for excitations close to
the Fermi level. Little is known about vertex corrections of high-energy
excitations. This is unfortunate, since the physics of high- and low-energy
excitations can differ significantly. An example is the energy dependence of
the self-energy of Ag discussed in Section V.26a.
Two possibilities to include vertex corrections beyond the standard GWA
approach are being studied in the literature: (1) an iterative evaluation of
Hedin’s equations (Eq. (2.10)) starting from Hartree theory,z59 and (2) an
expansion of the self-energy C and the polarizability P to order n and n - 1
in the screened interaction N respectively, with n = 2 being the case studied
so far (Refs. 3, 119, 232, and 236). Method (1) reformulates the Bethe-
Salpeter equation for the vertex function and has been applied to a
two-dimensional, 3 x 3 Hubbard cluster. Method (2) is the method of
choice for most of the current work on vertex corrections although the
expansion of C in terms of W is only asymptotically or conditionally
~onvergent,’~’and Minnhagen75.’52 found unphysical results-a negative
spectral density- by extending the GWA approach for the self-energy of
the homogeneous electron gas to second order in W Method (2) is based
on the expectation that higher-order contributions to C and P are negli-
gible3 or are canceled by self-consistency effects, which is empirically
supported by the good agreement between standard GWA calculations and
experiment. In what follows, we discuss both methods starting with method

2 5 4 This assumes a bare interaction between the electrons. Screened interactions have propa-
gators in internal lines, which lead to additional vertex corrections (Ref. 10).
2 5 5 For a discussion of vertex corrections see also B. Farid, Phil. Mag. B76, 145 (1997).
2 5 6 M. Hindgren and G O . Almbladh, Phys. Rev. B56, 12832 (1997).
’” J. E. Northrup, M. S. Hybertsen, and S. G. Louie, Phys. Rev. Lett. 59, 819 (1987).
”* M. P. Surh, J. E. Northrup, and S. G. Louie, Phys. Rev. B38, 5976 (1988).
’” A. Schindlmayrand R. W. Godby, Phys. Rev. Lett. 80, 1702 (1998).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 81

(1). In addition, we discuss approaches that include vertex corrections on an

LDA level as well as techniques that combine vertex-correction and self-
consistency diagrams.
Iterative Solution of Hedin's Equations. Schindlmayr and G ~ d b y ' ~ ~
solved Hedin's equations iteratively starting from Hartree theory. The key
feature of their approach is the replacement of the implicit Bethe-Salpeter
equation for the vertex function (compare Eq. (2.10), n = order of the
iteration)

x P+
')(6, 7; 3)d(4, 5, 6, 71, (2.84)

by an expression that depends only on the self-energy of the nth iteration,

the vertex function of the first iteration, and G('!

x F1)(6,7; 3)d(4, 5, 6, 7), (2.85)

A further advantage of Eq. (2.85) is that the functional derivative 6ZC'")/6G'o)

can in principle be calculated at all levels of iteration.259For n = 1, a vertex
correction r(')is obtained that contains not only terms to second order in
the RPA screened interaction but also terms of zeroth and first order.
Numerical results for a 3 x 3 Hubbard model suggest that an iterative
evaluation of Hedin's equations may improve upon the plasmon satellite
spectrum compared to standard GWA calculations.
Vertex Corrections to Second Order in the Interaction. The second-order
contribution in the bare Coulomb interaction to the self-energy and the
energy gap of Si is negligible compared to the first-order contribution.88 In
a bond-orbital approximation199260 the second-order terms can be written
as the sum of two partially canceling contributions:260 (1) the partial
blocking of ground-state correlations within a bond upon addition/removal
of a particle, and (2) exchange enhancements of the polarization cloud of the
electron (hole) for the bonds adjacent to the bond under consideration. This
suggests that second- and higher-order effects have a negligible impact on
band gaps and can be safely neglected.
R. Daling, P. Unger, P. Fulde, and W. van Haeringen, Phys. Rev. B 43, 1851 (1991).
82 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

However, vertex corrections evaluated with a screened rather than a bare

interaction lead to a band-gap narrowing261 of -0.26 eV at the r point of
Si, which amounts to -40% of the standard GWA quasiparticle correction
of about 0.7 eV.262Although higher-order corrections are included in the
self-energy, they are neglected in the polarizability, in conflict with Mahan's
suggestion." The second-order contribution partially cancels the additional
band-gap increase of 0.7 eV due to self-con~istency.~~~ Effects due to the
consistent use of vertex corrections not only in the self-energy but also in
the polarizability are discussed in the paragraph Ertex Corrections and
Partial Self-Consistency below.
LDA Vertex Corrections. Extension of a standard GWA (GWRPA)
calculation for SiS9by including LDA mean-field vertex corrections in the
polarizability (GWK,) or in the polarizability and the self-energy (GWT)
shows that relative energies are modified by including vertex corrections
only in the polarizability and that cancellations between vertex corrections
to the polarizability and to the self-energy occur for relative but not for
absolute energy shifts. The GWT approximation uses an LDA exchange-
correlation potential263as zeroth-order self-energy approximation and can
be expressed in a GWA-like form using an effective electron-electron
interaction tt and a dielectric matrix P9

X ( 1 , 2) = iG(1, 2)@(1, 2), (2.86)

where

tt= u[1 - P O ( U + Kxc)]-I= U / P ,

(2.87)

Table 12 shows that these vertex corrections included only in the screening
reduce band gaps by 0.1 to 0.2 eV and valence band widths by about 0.5 eV
in Si. Subsequent inclusion of vertex corrections in the self-energy causes a
band-gap increase that cancels the effect of the vertex corrections in the
screening. Note the qualitative difference between this band-gap increase"
and the band-gap decrease obtained by including dynamical vertex correc-
tions in X,119*262
as discussed in the paragraph Vertex Corrections to Second

261 We take the corrected value of Ref. 119 rather than the original value of 0.12 eV quoted
in Ref. 262.
262 P. A. Bobbert and W. van Haeringen, Phys. Rev. B49, 10326 (1994).
263 Note that because of the use of the LDA, the diagrammatic structure of these vertex
corrections is not well defined.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 83

TABLE12. QUASIPARTICLE CORRECTIONS (GWA MINUSLDA ENERGY VALUES IN

eV) FOR Si CALCULATED WITHINGWRPA (NO VERTEX CORRECTIONS FOR THE
DIELECTRIC MATRIX E AND SELF-ENERGY Z), GWK, (VERTEX CORRECTIONS IN E
BUT NOT IN Z), AND GWT (VERTEX CORRECTIONS M E AND Z; CORRECTIONS
LISTEDIN FIRSTF m Rows, LDA GAP = 0.52 eV). THE LAST Row IS AN
ABSOLUTEQUASIPARTICLE CORRECTION TO THE VALENCE-BAND MAXIMUM. LDA
VERTEXCORRECTIONS IN THE POLARIZABILITY REDUCEFUNDAMENTAL BANDGAPS
BY ABOUT 0.1 eV TO 0.2eV AND THE VALENCE-BAND WIDTHBY ABOUT 0.5 eV.
TAKINGVERTEX CORRECTIONS INTO ACCOUNT FOR BOTHTHE POLARIZABILITY AND
THE SELF-ENERGY DOES NOT HAVEA SIGNIFICANT EFFECTON RELATIVE ENERGY
DIFFERENCFSCOMPARED TO RPA-BAs~, GWA, BUT SHIFTS THE ABSOLUTE
ENERGY SCALEBY ABOUT 0.4 eV. (ADAPTED FROM REF.89.)

GWRPA GWKX GWT

Direct gap at r 0.64 0.56 0.65

Direct gap at X 0.78 0.57 0.73
Direct gap at L 0.68 0.58 0.72
Valence bandwidth - 0.56 - 1.01 -0.48
Minimum gap 0.63 0.59 0.66
Valence-band maximum -0.36 -0.44 0.01

Order in the Interaction above. Absolute energies between GWRPA and

GWT are shifted by about 0.4 eV. Whereas in GWRPA the highest occupied
state is shifted by - 0.36 eV, it is shifted by only + 10 meV in GWT. Since
the highest occupied state is believed to be well described by LDA, GWT
could turn out to be a better starting point, for example, for valence-band
offset calculations at interfaces.
Vertex Corrections and the Alkali Metal Band Width. In jellium and
simple metals, band-width corrections due to vertex corrections to the pol-
arizability P and to the self-energy C cancel, leading to a reproduction of
RPA results, as shown in Fig. 20, and an overestimation of the experimental
band width. The discrepancy with experim.ent can be eliminated by (1)
inclusion of vertex corrections only in the polarizability-that is, use of a
GWK,, approach -and self-consistent determination of quasiparticle ener-
gies;'84*258(2) careful treatment of surface and finite lifetime effects (Refs.
229,240,241, and 265); and (3) inclusion of vertex corrections in both P and
C in a partially self-consistent calculation.232 Approach (1) is motivated
by Ward-identity-like arguments by Strinati et al.156(see Section II.9b).
Mahan and c o l l a b ~ r a t o r sand ~ ~Ting,
~ ~ ~Lee,
~ ~ and
~ ~ ~Quinn266
~ empha-

264 P. Vashishta and K. S . Singwi, Phys. Rev. B6, 875 (1972).

265 K.W.-K. Shung and G. D. Mahan, Phys. Rev. B38,3856 (1988).
266 C. S. Tin& T. K. Lee, J. J. Quinn, Phys. Rev. Leu. 34, 870 (1975).
84 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

-1.0
0 1 2 5 4 5 6
rs

FIG. 20. Band narrowing relative to the independent-particle band width versus electron
density parameter rr for the homogeneous electron gas and several approaches to the
determination of the self-energy. A positive value corresponds to an actual narrowing of the
bands. The two dashed lines correspond to a standard (RPA) GWA calculation (in which
vertex corrections are neglected in the dielectric function E and the self-energy C) and a GWA
calculation in which vertex corrections are included only in E (I'= 1). The solid line
corresponds to a calculation that includes vertex corrections in both E and Z using a
Vashista-Singwi many-body local-field factor.264This curve is universal in the sense that it is
hardly a5ected by a change of the many-body local-field factor. The r = 1 calculation leads to
band narrowing that is about twice as large as the RPA band narrowing for metallic densities
(rs z 2 - 6) and compares favorably with experiments for simple metals. Adding consistent
vertex corrections to the self-energy cancels the effect of vertex corrections in the dielectric
screening, giving results in close agreement with RPA. (From Ref. 240.)

sized that a consistent way of handling vertex corrections requires the

inclusion of identical vertex corrections in P and C. Both contributions
cancel to a large degree as shown in Fig. 20, leading to a band width nearly
identical to the one obtained in RPA and too large in comparison with
experiment. Vertex corrections are neglected in approach (2).229 The
required band-width narrowing is obtained from a treatment of the in-
homogeneous surface potential and the finite imaginary part of the self-
energy only. Shirley232 combined results from partially self-consistent
GWOA calculations with static vertex corrections in P and dynamic second-
 QUASIPARTICLE CALCULATIONS IN SOLIDS 85

order corrections in 2, and s ~ g g e s t e d ~that

~ ~ dynamic
* ~ ~ ’ vertex corrections
for C are more appropriate than the static corrections considered by Mahan
et al. In spite of the different physics contained in the respective approxi-
mations, all three approaches agree with the experimental band widths (for
example, 0.7 eV band-width narrowing in sodium compared to the free-
electron result; see Section V.26b for more details). We estimate that missing
self-consistency effects could increase the band width of all of the above
calculations by 0.4 to 0.7 eV based on the results of Holm and von BarthZ3’
for fully self-consistent GWA calculations in jellium with r, = 4 (average Na
density). This suggests that the interpretation of alkali metal photoemission
spectra is a difficult, not yet fully understood, problem.268
Vertex Corrections and Partial Self-Consistency. Table 13 shows that the
inclusion of dynamic vertex corrections for C to second order in WZ3’
virtually cancels the results of partial self-consistency. Vertex corrections
decrease the quasiparticle weight and the band width and lead to a Fermi
energy in worse agreement with the exact solution for the homogeneous

’‘’ E. L. Shirley, private communication.

2 6 8 For a critical discussion of Mahan’s photoemission theory for the determination of the

alkali metal photoemission spectra, see, for instance, Ref. 11.

TABLE13. QUASIPARTICLE WEIGHTOF JELLIUMAT k = 0 AND k = k, ( k , = FERMIWAVE

VECTOR),BAND WIDTH w , AND FERMIENERGYE, FOR A NON-SELF-CONSISTENT GWA
CALCULATION, A PARTIALLY SELF-CONSISTENT
GW’A CALCULATION (DATAIN PARENTHESES),
AND A GW’A CALCULATION THAT INCLUDES DYNAMICVERTEX CORRECTIONSTO SECOND ORDER
IN W IN THE SELF-ENERGY(DATAUNDER“HIGHERORDER”)EVALUATED AT SEVERAL VALUES
OF r,. EXACTVALUES FOR E , ARE GIVENAS WELL.THE SCREENED INTERACTION CONTAINS A
STATICVERTEXCORRECTION. SELF-CONSISTENCY
IMPROVESE , EXCEPTFOR rs = 2, WHERETHE
PLASMON-POLE MODELOF REF.232 IS ONLYMARGINALLY VALIDBUT INCREASES Z AND w IN
CONTRADICTION TO EXPERIMENTS ON SIMPLE
METALS.DYNAMIC VERTEXCORRECTIONS ALMOST
COMPLETELY
CANCEL SELF-CONSISTENCY
EFFECTS.
(ADAPTED FROM REF.232.)

GWA HIGHERORDER
EXACT
EF z(o),Z(k,) W EF EF

2 0.6,O.S 11.9 0.10 0.5.0.8 11.6 0.08 0.10

(0.6,O.g) (13.3) (0.13)
3 0.5,0.7 5.0 - 0.05 0.5.0.7 5.0 -0.06 -0.04
(0.6,0.7) (6.0) ( - 0.04)
4 0.5,0.6 2.6 - 0.09 0.5,0.6 2.7 -0.09 -0.08
(0.6,0.7) (3.3) ( - 0.08)
5 0.4,0.6 1.6 -0.10 0.4.0.6 1.7 -0.09 -0.08
(0.5,0.7) (2.3) ( - 0.08)
86 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

electron gas. Overall, the results agree with those of a non-self-consistent

GWK,, calculation that takes vertex corrections for E into account via the
following many-body local-field factor f ( q ):

(2.88)

where K J n ) is defined as in Eq. (2.87) and where Vxc(n)is the exchange-

correlation potential of a homogeneous electron gas with density n and
Fermi momentum k, .269
Inclusion of lowest-order vertex and self-consistency effects beyond the
RPA polarizability for Si and diamond”’ leads to (1) a cancellation of the
constant contributions to the polarizability at k + O of vertex and self-
consistency corrections, (2) a significant modification of the dielectric
constant, (3) an increase in fundamental band gaps compared to RPA
screening by 0.1 eV (0.2 ev) for Si (diamond), (4) a significant increase in the
absolute value of the diagonal polarizability matrix elements compared to
RPA by vertex corrections, and ( 5 ) a significant decrease in the same matrix
elements by self-consistency effects, leading to large cancellations and an
overall increase in the matrix elements by < 15% compared to RPA. The
diagrams that are calculated for the dielectric polarizability and the self-
energy are shown in Fig. 21. The solid lines in these diagrams correspond
to a single-particle propagator constructed from LDA wave functions and
energies. The wavy lines correspond to the screened interaction whose
frequency depedence is modeled by the Engel-Farid plasmon-pole m0de1.~”
The results of adding the lowest-order vertex (v) and self-energy (SC)
corrections to the RPA polarizability are shown graphically in Fig. 22. Both
the vertex and the self-consistency corrections to the RPA polarizability are
finite for k + 0 and only their sum tends to zero proportional to k2. This
proportionality for the head of the static polarizability is necessary to ensure
the correct screening behavior of a semiconductor in the long-wavelength
limit’ 19,272 and provides a formal motivation for an inclusion of both effects

269 The local-field factor f(q) is obtained by fitting the dielectric constant to quantum
Monte-Carlo data, Ref. 270. A similar approximation has been used by Hindgren and
Almbladh, Ref. 256, who showed it to be very reliable.
”* C. Bowen, G. Sugiyama, and B. J. Alder, P h p . Rev. SSO, 14838 (1994).
This approximation affects the self-energy corrections to the lowest conduction and highest
valence band by less than 20 meV, as shown in Table 21. Also, de Groot et al. (Ref. 236) studied
the lowest-order vertex and self-consistency corrections to the RPA polarizability and the
GWA self-energy for a 1D semiconducting wire. We do not discuss that model system but
concentrate on the results for Si.
272 W. Kohn, Phys. Rev. 110,857 (1958).
 + 0' i
+ Qsc4 SCI SC2 SC3 SC4

+o V +% 2
xU
FIG.21. Left panel: RPA potarizability and corrections to first order in the screened interaction W SCl-SC4 denote corrections due r
to selfenergy insertions in the Green function, that is, due to lowest-order self-consistency effects, V corresponds to a ladder diagram
that is the lowest-order vertex correction. The solid directed line denotes the LDA Green function. The cross denotes - V ,' and the
wavy line denotes the RPAdynamically-screened interaction. Right panel GWA self-energy plus lowest-order self-consistency
(SC1-SC4)and vertex (V)corrections. The diagrams SC3 and SC4 are selfenergy corrections due to lowest-order corrections to the
valence density. The wavy line denotes the dynamically screened interaction, which can be evaluated either in RPA or in
+ +
RPA V SC.The notation is otherwise identical to the one used in the left panel. (From Ref. 119.)
88 WILFRIED G. AULBUR, LARS JQNSSON, AND JOHN W. WILKINS

0.03

0.01

4 - 0 4

-0.-

0.0 1.0 2 0 3
.0 4.0 6.0 6.0
4-07
0.0 4.0
a
a.0 e.0 at.0 4.0 6.0
1- [Zrua] 1- [2rr/orl

FIG. 22. Cancellation between vertex (V) and self-consistency (SC) corrections to the
diagonal elements of the RPA static polarizability PGc(k;w = 0) (in Rydberg am) for silicon
and diamond. Lowest-order vertex and self-consistency corrections as considered in Ref. 119
largely cancel, leading to an overall enhancement of the RPA polarizability by < 15%. The
constant contributions of the vertex and self-consistency corrections to the head element of the
RPA polarizability matrix for k + 0 cancel, which ensures the correct screening behavior in
semiconductors. (From Ref. 119.)

in the polarizability. The effect of the vertex and self-consistency corrections

on the dielectric constant within the approximations by Ummels et ul.'19
amounts to -20% in the case of Si and - 5 % in the case of diamond273
and is therefore at least as important as local-field effects due to an
inhomogeneous density distribution.
Inclusion of vertex and self-consistency corrections in the self-energy leads
to an increase of fundamental band gaps by 0.4 eV (0.3 eV) in Si and 0.7 eV
(0.5 eV) in diamond using RPA + V + SC (RPA) screening.' Band gaps
determined using different combinations of vertex and self-consistency
corrections for the polarizability and the self-energy are shown in Table 14.
The influence of V + SC corrections to the self-energy operator leads to a
significant increase in fundamental band gaps. The results of Ummels et

273 Silicon: E ~ 12.8,~ E ~~ =~ ~ 10.4,

+ ~P P+ = ~ 11.4,
~ 11.7;
= diamond cRP"=5.6, E ~ ~ ~ + ~
5.3, E~~~ = 5.5, 5.7.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 89
TABLE 14. INFLUENCEOF LOWEST-ORDER VERTEX (V) AND SELF-CONSISTENCY (SC)
CORRECTIONS TO THE SELF-ENERGY (GWA VERSUS GWA + V + SC) AND TO THE DIELECTRIC
SCREENING (RPA VERSUS RPA + V + SC) ON THE FUNDAMENTAL BANDGAPSAT r, X, AND L
IN Si AND AT r IN DIAMOND. LOWEST-ORDER CORRECTIONS TO THE DIELECTRIC SCREENING LEAD
TO A SMALLINCREASE IN FUNDAMENTAL BANDGAPS( x 0.1 eV IN Si, x 0.2 eV IN DIAMOND).
LOWEST-ORDER CORRECTIONS TO THE SELF-ENERGY LEADTO A SUBSTANTIAL INCREASEIN THE
FUNDAMENTAL BANDGAP( x 0.3 eV IN Si, x 0.5 eV IN DIAMOND),RESULTING IN A DISCREPANCY
OF ABOUT 0.3 eV IN Si AND 1eV IN DIAMOND BETWEEN CALCULATION AND EXPERIMENT.Wmm
THE APPROXIMATION OF THE CALCULATION (LDA GREENFUNCTION, PLASMON-POLE MODEL,
ETC.) THE EFFECTS OF VERTEX AND SELF-CONSISTENCY CORRECTIONS DO NOT CANCEL. ENERGIES
ARE IN eV. (ADAPTED FROM REF. 119.)

Si DIAMOND
METHOD
(SELF-ENERGY/SCREENING) r X L r
LDA 2.5 3.4 2.6 5.5
GWA/RPA 3.3 4.2 3.4 1.6
GWA/RPA+V+SC 3.4 4.3 3.4 1.8
GWA + V + SC/RPA 3.6 4.5 3.1 8.1
GWA+V+ SC/RPA+V+SC 3.1 4.6 3.8 8.4
EXPT.”.~ 3.4 4.3 3.5 1.3

“Ref. 182; bRef. 274.

al.’ l 9 suggest that the cancellation between lowest-order vertex and self-
consistency contributions is far from complete if one determines these
corrections to the GWA self-energy using a noninteracting LDA Green
function. In fact, the V contribution cancels the SC contribution to the
self-energy by only 35% (45%) in the case of Si and only - 10% (15%) in
+ +
the case of diamond for RPA V SC (RPA) screening. In the latter case,
even the term “cancellation” is therefore inappropriate. ’’
111. GWA Calculations: Numerical Considerations

This section (1) describes various reciprocal-space and real-space/imagin-

ary-time implementations of the GWA; (2) gives numerical details for an
implementation of the GWA in a plane-wave basis; (3) describes parallel
algorithms of GWA calculations in reciprocal space as well as in real
space/imaginary time; and (4) compares GWA calculations for five proto-
typical semiconductors.

274 D. E. Aspnes and A. A. Studna, Phys. Rev. B27,985 (1983).

 90 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

15. DIFFWENT
IMPLEMENTATIONS
OF THE GWA

Quasiparticle calculations can be performed either in reciprocal space as a

function of frequency or in real space as a function of imaginary time. Both
approaches are described in this section; subsection a contains the reciprocal-
space approach; subsection b, the real-space approach. The reciprocal-space
method is widely used and has been implemented using plane waves as well as
a variety of local-orbital-basis sets. Subsection a contains a comprehensive
comparison of the advantages and weaknesses of the different basis sets.

a. Reciprocal-Space Approach
Early calculation^^.^ in the dynamically screened potential approximation
determined the expectation value of the self-energy operator using a plane-
wave basis set in reciprocal and frequency space, since this allows both a
straightforward evaluation of matrix elements that occur in the self-energy
and a systematiccontrol over convergence. In addition, a GWA calculation in
reciprocal and frequency space parallels the experimentalsituation in (inverse)
photoemission, which determines the band structure of a solid as a function of
the reciprocal wave vector k and of the frequency w of the quasiparticle.
a. 1. PLANE WAVES. Pseudopotentialsin conjunction with a plane-wave basis
set are widely used in computationalcondensed matter theory because of their
ease of use and systematic convergence proper tie^."^ Mostly, plane-wave-
based GWA calculations are applied to sp-bonded bulk solids and to their
interfaces, surfaces, defects, and clusters, as discussed in Sections IV and V.
The two main disadvantages of plane-wave basis sets are that (1) the
number of plane waves Npw increases with the system volume V and the
energy cut-off E,,, needed for a converged description as N,, x (1/21)VE:i?
(in Hartree atomic units); and (2) plane waves have no direct physical
interpretation in contrast to local-orbital basis sets discussed below. Recip-
rocal-space GWA calculations scale as N $ , (see below), which makes plane-
wave-based GWA calculations prohibitively expensive for large- V systems
such as complicated defect structures and for large-E,,, systems such as d
and f electron materials.
Detailed Formulas. A Bloch wave function of wave vector k and band n
can be expressed in terms of its Fourier components cnk(G),where G is a
reciprocal lattice vector, as
1
@nk(r) =- 1 c,,(G) exp(i(k + G) * r). (3.1)
J V G

*’’ W. E. F’ickett, Comp. Phys. Rep. 9, 115 (1989).

 QUASIPARTICLE CALCULATIONS IN SOLIDS 91

Define the following matrix element between occupied valence (n = u) and

unoccupied conduction ( n = c) states:

(3.2)

The independent-particle polarizability given by Eq. (2.22) is

where the factor 2 accounts for spin degeneracy.

The frequency convolution in Eq. (2.24), together with the independent-
particle Green function (Eq. (2.4)) expressed in terms of Bloch wave
functions and the Fourier component of the screened interaction WGG.(q;w),
allows the determination of the matrix elements of the self-energy with
respect to Bloch states @mk = Im,k) and @,, = I/, k) as the following sum
over occupied and unoccupied states:

c
1 occ+unocc BZ
(4k I W ) I L k) = -
v n
1 2 ME&,
q GG'
q)CM"d(kk, q>l*

exp(iw6) WGG~(q9 0)
(-27Ci) E +0- - i6Sgn(P - E,k-q)
Enk- q
do.

(3.4)
Here, the matrix elements M;.(k, q) generalize Eq. (3.2) to arbitrary states
n and m.
The remaining frequency integration is usually performed analytically
using the plasmon-pole models presented in Section II.8b rather than
numerically as in Refs. 102, 103, 185, and 276. The plasmon-pole model
(PPM) is accurate to within a few tens of meV for states close to the Fermi
level and to within a few tenths of an eV for low-lying valence states, as is
demonstrated in Table 21. Splitting the self-energy into a bare exchange part
C x and an energy-dependent correlation contribution C c ( E ) , as detailed in

276 H. N. Rojas, R. W. Godby, and R. J. Needs, Phys. Rev. Left. 74, 1827 (1995).
92 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Section 11.6 gives

(3.5)

(3.6)
where the matrix elements are defined as277

P;m(k, 4) = c CM&rn(k,:q)I*@,(G),
G
(3.7)

and ci = 27tzpq0,(q) for the von der Linden-Horsch model and ct = 1 for the
Engel-Farid plasmon-pole model.
The N & scaling of the reciprocal-space GWA algorithm follows from
Eqs. (3.2), (3.6), and (3.7). The construction of the matrix elements in Eq.
(3.2) scales as the number of plane waves used to describe the LDA wave
functions. The matrix elements in Eq. (3.7) are determined in N , operations,
where N , is the number of plane waves used to describe the plasmon-pole
eigenvectors. The self-energy itself (Eq. (3.6)) must be determined via a sum
over LDA bands and a sum over plasmon-pole bands. The number of LDA
and plasmon-pole bands as well as N , and N p , increase linearly with
system size and are all of the same order. Hence, the overall algorithm scales
as the size of the system or the number of plane waves to the fourth power.
An approximate evaluation of the ~elf-energy"~via Taylor expansions
around a given set of energies reduces the GWA scaling to N;w and allows
larger GWA calculations based on plane waves. Other appro ache^^^^.^^^
permit the determination of the density-density response function with an
O ( N ; , ) effort, but have not been extended to the calculation of the
self-energy.
a.2. LOCAL-ORBITAL BASISSETS. The number of basis functions of local-
orbital basis sets,' s which is needed to describe bulk semiconductors, is
smaller by about one order of magnitude than the corresponding number
of plane waves.281 Although the overall scaling of the reciprocal-space

277 The scaled plasmon-pole eigenvectors for the von der Linden-Horsch and the Engel-Farid
plasmon-pole models are defined in Eqs. (2.52) and (2.61), respectively.
2 7 8 L. Reining, G. Onida, and R. W. Godby, Phys. Rev. B56,R4301 (1997).
''' X. Blase, A. Rubio, S. G. Louie, and M. L. Cohen, Phys. Rev. BS2, R2225 (1995).
A. A. Quong and A. G. Eguiluz, Phys. Rev. Lett. 70, 3955 (1993).
The description of bulk Si requires, for example, 40-60 Gaussian orbitals compared to
about 450 plane waves (Ref. 114).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 93

GWA algorithm still scales as the number of basis functions to the fourth
power, the strong reduction of the number of basis functions results in
significant savings in computation time and allows the study of materials
that are not accessible in plane-wave-based approaches such as N i 0 5 5 or
Si,H, clusters.’” Unfortunately, the matrix element construction is com-
putationally more intense in a local-orbital than in a plane-wave basis, which
partially reduces gains due to the smaller number of basis function^."^
Linearized Augmented Plane Waves (LAP W ) . LAPW is an all-electron
method that has been applied to GWA calculations in Si’85 and Ni6’ In Si, 45
basis functions per Si atom are needed, which corresponds to a reduction by a
factor of five compared to plane-wave-based calculations. Although LAPW
calculations allow systematic convergence, they do not seem to offer significant
computational savings compared to plane-wave-based calculations, which
may have inhibited a widespread use of LAPW for GWA calculations.’86
Linearized Mufin-Tin Orbitals (LMTO). LMTO is an all-electron
method that has been used, for instance, for GWA calculations of transition-
metal oxide5’ and the Compton profile of alkali It is exten-
sively reviewed in Ref. 14. GWA calculations for transition-metal oxides are
prohibitively expensive for plane-wave-based methods, since an accurate
description of the transition-metal d and oxygen 2 p electrons requires a very
large energy cut-off and several thousand plane waves. In LMTO, the
description of transition-metal oxides uses only 50- 100 basis functions per
atom.55
Aryasetiawan, Gunnarsson, and collaborator^'^ used LMTO in the
atomic sphere approximation and included so-called combined correc-
t i o n ~ ’ ’ ~only in the energies, not in the wave f ~ n c t i o n s . ’This
~ ~ approach
has two disadvantages. First, omission of the combined-correction term
leads to discontinuous wave functions in the interstitial region and in the
overlapping region of the muffin-tin sphere^.*^^*^^' This results in errors in

282 M. Rohlfing and S. G. Louie, Phys. Rev. Lett. 80, 3320 (1998).
283 In the case of bulk Si, Gaussian-orbital-based calculations take as much time as
plane-wave-based calculations (Ref. 284). For the Si(OO1) (2 x 1) reconstructed surface, a
speed-up of a factor of five compared to plane waves is achieved (Ref. 284).
284 M. Rohlfing, P. Kriiger, and J. Pollmann, Phys. Rev. B52, 1905 (1995).

2 8 5 N. Hamada, M. Hwang, and A. J. Freeman, Phys. Rev. B41, 3620 (1990).

2 8 6 F. Aryasetiawan and 0. Gunnarsson, Phys. Rev. B49, 16214 (1994).
2 8 7 Y. Kubo, J . Phys. Soc. Jpn. 65, 16 (1996).
2 8 8 Y. Kubo, J . Phys. Soc. Jpn. 66, 2236 (1997).
2 8 y 0. K. Andersen, Phys. Rev. B 12, 3060 (1975).

F. Aryasetiawan, private communication.

291 H. L. Skriver, The LMTO method: Mujin-Tin Orbitals and Electronic Structure, Springer
Series in Solid-state Sciences, Vol. 41, Springer, Berlin (1984).
2 y 2 M. Alouani, private communication.
94 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

the determination of the matrix elements of the momentum operator and

hence, for instance, in the dielectric properties and subsequently the self-
energy operator. Second, the atomic sphere approximation assumes a
spherically symmetric potential. This shape approximation can affect prop-
erties such as the dielectric constant. Christensen and G o r ~ z y c a , 'for~~
instance, determined the dielectric constant of GaN within the atomic
sphere approximation, including combined-correction terms, to be 4.78 -a
20% underestimation compared to the full-potential LMTO value of
5.96.293
An estimate of the influence of the above approximations on quasiparticle
energies is difficult. Insight can be gained by comparing all-electron LMTO
results for GaAsZ9' with pseudopotential results that include core-polariz-
ation potentials and core-relaxation effects,43 as well as with experiments.
Shirley and collaborators obtained quasiparticle corrections of 1.29/0.74/
1.05 eV for the lowest conduction band at r, X , and L, compared to values
of 1.19/1.59/1.42 eV using scalar relativistic LMT0.290This can be inferred
from Table 23. LMTO quasiparticle corrections are larger than the best
plane-wave results by 0.1 to 0.8 eV. In addition, LMTO in the atomic sphere
approximation overestimates the low-lying experimental conduction-band
energies by 0.3 to 1.1 eV, again as shown in Table 23. Note that the LMTO
calculation of Ref. 290 chooses a trial energy around which it makes an
expansion. This trial energy is chosen to accurately describe the valence and
semicore states of GaAs with less emphasis placed on unoccupied states. The
occupied states are indeed accurately described in LMTO, while the
description of the unoccupied states is substantially worse. Choosing several
reference energies can improve the description of unoccupied states.290Note
that Kubo's calculation^^^^.^^^ avoid the above mentioned problems by
evaluating the GWA based on a full-potential LMTO calculation.294
Gaussian Orbitals. Baroni and c o l l a b ~ r a t o r s ' ~use
~ ~ 'a~ ~contracted
Gaussian-orbital basis set to determine the electronic properties of core and
valence electrons in the COHSEX approximation (see Section 9). More
recently, Rohlfing, Kruger, and P01lman~~" 14,284 developed an approach
for a GWA calculation that is based on pseudopotentials in conjunction
with an uncontracted Gaussian-orbital basis set.
The merit of Gaussian-orbital-based GWA calculations is the reduced
number of basis functions and hence their greater efficiency and applicability
compared to plane-wave calculations, as exemplified by (1) a factor of five
speed-up compared to a plane-wave calculation of the Si(OO1) (2 x 1)

293 M. Alouani and J. Wills, as presented in Ref. 542.

294 In contrast to most other GWA calculations, Kubo also used a linear tetrahedron method
for the evaluation of Brillouin zone integrals.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 95

reconstructed ~ u r f a c e ; ” (2)
~ an explicit treatment of the Cd n = 4 core and
semicore electron^;^' and (3) the inclusion of excitonic effects in Si,H,
clusters.’
The main disadvantage of Gaussian orbitals is that while Gaussian basis
sets can systematically converge, constructing them is difficult. Table 15
demonstrates the problems in establishing the energies of low-lying conduc-
tion states of small-gap and medium-gap semiconductors. For bulk
semiconductors, 20 Gaussians per atom are suggested to be adequate for
convergence of quasiparticle energies of low-lying conduction and valence
states to within 0.1eV.”4 While this is true for Si”4 and for large-gap
materials with strongly localized densities such as diamond, Sic, and
GaN,295good convergence for more “metallic,” small- and medium-gap
semiconductors such as Ge and GaAs requires at least 30 Gaussians per
atom. Tables 15 and 22 show that in the case of Ge (1) the lowest
conduction-band energies at r, X , and L for a well-~onverged’~~ plane wave

295 M. Rohlfing, diploma thesis, University of Munster, Germany (1993).

296 The LDA band structure of Ref. 42 is calculated at 10 special k-points and a 10-Hartree
cut-off. Comparison with a calculation using 60 special k-points and an energy cut-off of 18
Hartree leads to shifts in the eight lowest eigenvalues at r, X , and L by less than 0.03 eV (rZc:
0.05 eV).

TABLE15. CONVERGENCE OF LDA AND QUASIPARTICLE ENERGIESFOR SELECTEDSTATES I N Ge

AND GaAs FOR A GAUSSIAN-ORBITAL-BASED METHOD,”4’295 IN COMPARISON TO A PLANE-
WAVE (Pw) FORBOTHMATERIALS AT LEAST60 GAUSSIANORBITALS (GO) ARE
NECESSARY TO OBTAIN A CONVERGED LDA BAND STRUCTURE. THEGWA BAND STRUCTURE
MAY REQUIREMO RE THAN60 GAUSSIANS FOR COMPLETE CONVERGENCE. A 40-GAUSSIAN BASIS
SETIS USEDIN THE LITERATURE,’l 4 BUT THAT SETIS INSUFFICIENTTO OBTAINBANDSTRUCTURES
ACCURATETO WITHIN0.1 eV FOR SMALL-AND MEDIUM-GAP SEMICONDUCTORS. NOTE THAT
IMPROVING CONVERGENCE OF THE GAUSSIAN-ORBITAL CALCULATION MOVESTHE RESULTS
SYSTEMATICALLY TOWARD THE PLANE-WAVE RESULTS.ALL ENERGIESARE IN eV.

PW 60 GO 40 GO

LDA G WA LDA G WA LDA GWA

Ge Xl, 0.70 1.15 0.79 1.36 1.03 1.74

LI, 0.12 0.61 0.18 0.76 0.33 0.98
L;, 7.07 7.56 7.46 8.1 1 8.33 9.20
GaAs 0.47 1.16 0.52 1.21 0.57 1.32
Xl, 1.42 2.00 1.47 2.14 1.80 2.65
x3c 1.62 2.24 1.66 2.37 1.85 2.72
Ll, 0.97 1.62 0.98 1.66 1.13 1.92
Ll, 7.74 8.40 8.08 8.86 8.88 9.92
96 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

and a 40-Gaussian-orbital calculation differ by up to 0.3 eV in LDA and up

to 0.6 eV in GWA; (2) higher-lying states, such as the L;, state, differ by up
to 1.3 eV in LDA and 1.6 eV in GWA; and (3) the Gaussian-orbital results
converge systematically towards the plane-wave results when 60 Gaussians
rather than 40 Gaussians are used. Similar conclusions can be drawn from
Tables 15 and 23 for GaAs. In particular, the difficulty in describing low-
lying conduction states at X and L with only 40 Gaussians may result from
an enhanced d character of these states.297
The problem ofproperly treating d states near the valence band is illustrated
by the work on CdS5’ As pointed out in Section 11.12,semicore d states modify
dynamic screening in a ~ o l i d . ~ ~ ~In~ addition,
’ ~ ’ ~ ’ the exchange interaction
between all electrons of a shell ( n = 4 in CdS) affects the energies of semicore and
valence electron states5’ if the corresponding wave functions overlap strong-
l ~ . ’ ~Quantitative
’ results for CdS are difficult to obtain since CdS seems to be
sensitive to the pseudopotential used. As shown in Table 16, treating the Cd 4d
electrons’96 as core electrons, via nonlinear core correction^,'^^ or as valence
electrons gives LDA energy gaps of 1.72 eV, 1.60 eV, and 0.84 eV, respectively,
for scalar-relativistic Troullier-Martins pseudopotentials in the Kleinman-
Bylander form.304Semi-relativistic,nonlocal ab-initio pseudo potential^^ give ’
an energy gap of 1.36 eV when nonlinear core corrections are used (E,,, = 12.5
Hartree); and scalar relativistic Bachelet-Hamann-Schliiter pseudopoten-
tials301give an energy gap of 1.65 eV when the d electrons are treated as core
electrons and of 0.3 eV when the d electrons are treated as valence electrons.305
Using the scalar-relativisticBachelet-Hamann-Schliiter Cd pseudopotential in
Refs. 52 and 302 gives an LDA energy gap of 2.15 eV and 2.18 eV, respectively,
when the d electrons are treated as core electrons. An energy gap of 0.78 eV is
obtained when they are treated as valence electron^.^' The deviation in the
fundamental LDA energy gap between the calculations of Refs. 51,196, and 301
and those of Refs. 52 and 302 using different Cd’ pseudopotentials amount to
-+

at least 0.5 eV and are on the order of the self-energy shifts observed upon
inclusion of 4p and 4s electrons as valence electron^.^'
Good quasiparticle band gaps are obtained by treating the core and
semicore electrons in II-VI materials via nonlinear core correction^.^'
However, the success of this approach is based on a fortuitous cancellation
of errors due to the neglect of wave-function relaxation and hybridization
and to the neglect of dynamic screening by semicore electrons and exchange
interaction with core and semicore electrons.

’” S. G. Louie, Phys. Rev. B22, 1933 (1980).

298 This observation agrees with the determination of the influence of core charges on the
self-energy of Ag (Ref. 216) and on the self-energy of Ni (Ref. 68), as detailed in Section V.26.
2 9 9 As in Ref. 182, vol. 17b.
300 M. Cardona, M. Weinstein, and G. A. Wol5, Phys. Rev. 140, A633 (1965).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 97

TABLE16. LDA AND GWA (BOLDFACE) ENERGYGAPS AND LDA LATTICECONSTANTS

(PERCENT DEVIATION FROM EXPERIMENT IN PARENTHESES) IN COMPARISON WITH EXPERIMENT
FOR CUBIC CdS. COCIORVAet al. 196 USED SCALAR-RELATIVISTIC TROULLIER-MARTINS
PSEUDOPOTENTIALS IN THE KLEINMAN-BYLANDER FORM (E,,, = 25 HARTREE FOR d ELECTRONS IN
CORE,45 HARTREE FOR d ELECTRONS I N VALENCE). CHANGet USEDSCALAR-RELATIVISTIC
BACHELET-HAMANN-SCHL~ER PSEUDOPOTENTIALS (E,,, = 6.75 HARTREE FOR d ELECTRONS IN
CORE, 9.5 HARTREEFOR d ELECTRONS IN VALENCE) A N D SHOWED THAT THE NEGLECT OF
RELATIVISTIC EFFECTS(NUMBERS IN PARENTHESES) INCREASES FUNDAMENTAL BANDGAPSBY
ABOUT 0.3 ev. ZAKHAROV et al.” USEDSEMIRELATIVISTIC, NONLOCAL PSEUDOPOTENTIALS WITH
NONLINEAR CORECORRECTIONS (NLCC)”’ (E,,, = 12.5 HARTREE). ROHLFING et AND
FLESZAR302 USED SCALAR-RELATIVISTIC BACHELET-HAMANN-SCHL~ER PSEUDOPOTENTIALS.
THEDIFFERENCE OF ABOUT 0.5 eV IN THE FUNDAMENTAL LDA BANDGAPBETWEEN DIFFERENT
Cd2+ CALCULATIONS IS OF THE SAMEORDER AS THE GWA BAND-GAPINCREASE UPON
INCLUSION OF 4s AND 4p ELECTRONS AS VALENCE ELECTRONS IN GWA CALCULATIONS.
INCLUSION OF THE 4d ELECTRONS AS VALENCEELECTRONS LEADSTO A REPRODUCTION OF THE
EXPERIMENTAL LATTICE CONSTANT TO WITHIN-3.6%52 AND -0.7%.196 THISDIFFERENCE IN
THE EQUILIBRIUM STRUCTURE LEADSTO DIFFERENCES IN THE FUNDAMENTAL BANDGAPOF 0.2
TO 0.3 eV.196-303
THELDA AND GWA ENERGY GAPSARE DETERMINED AT THE EXPERIMENTAL
LATTICE CONSTANT IN ALL CALCULATIONS.
________~~ ~

E,,,(W a d 4

d ELECTRONS COCIORVA CHANG ZAKHAROV ROHLFING FLESZAR COCIORVA ROHLFINC

~~

Core 1.72 1.65 (1.93) 2.15 2.18 5.05 5.05

3.70 3.83 (-13.2%) (-13.2%)
NLCC 1.60 136 5.33
2.79 ( - 8.4%)
Valence 0.84 0.3 (0.67) 0.78 5.78 5.61
1.50 (-0.7%) (-3.6%)
Expt.“ 2.50 5.818
2.55b

“Ref. 299 unless noted otherwise; bRef. 300.

b. Real-Spacellmaginary-Time Approach
The basic idea of the space-time approach is to choose the representation
(e.g., reciprocal space and frequency or real space and time) that minimizes
the computations necessary to evaluate the basic GWA quantities: G , E, W

301 K. J. Chang, S. Froyen, and M. L. Cohen, Phys. Rev. B28,4736 (1983).

302 A. Fleszar, private communication.
3 0 3 M. Rohlfing, P. Kriiger, and J. Pollmann, private communication.
304 The energy cut-offs used in the calculations are 25 Hartree, 25 Hartree, and 45 Hartree,

respectively.
305 The energy cut-off used for the core case is 6.75 Hartree and 9.5 Hartree for the valence
case. Neglect of scalar relativistic effects increases the gap by about 0.3 eV to 1.93 eV (core)
and 0.67 eV (valence) (Ref. 301).
98 WILFRIED G . AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

and C.For example, the determination of the self-energy in reciprocal and

frequency space, Eqs. (3.5) and (3.6), involves numerically expensive convol-
utions that lead to an overall scaling of the algorithm proportional306 to
N& and to N:-the square of the number of mesh points used for the
representation of the frequency dependence of the screened interaction. In
real space and as a function of time, the self-energy is a simple product
(compare Eq. (2.1 1)):

C(r, r’; t) = iG(r, r’; t) W(r, r’; T), (3.8)

which eliminates two convolutions in reciprocal space and one in the fre-
quency domain, and leads to an algorithm whose dominant parts scale as
O(N$,) and O ( N , ) . Similarly, the dielectric matrix and the screened
interaction require convolutions in real space (Eqs. (2.20) and (2.14)) but
multiplications in reciprocal space (Eqs. (2.23) and (2.16)). A representation
change from reciprocal to real or frequency to time space using Fast Fourier
Transforms (FFTs) and vice versa scales as N,, log(N,,) or N , log(N,)
and is computationally very efficient. This allows the efficient determination
of C, E, and W as products rather than as convolutions.
The evaluation of operators in real space and the extensive use of FFTs
require operators that are short-ranged in real space and quantities that can
be represented by an equidistant numerical grid. Relevant quantities in the
GWA, namely, G, C, and Po, are proportional to Ir - r’l-’ for Ir - r’I -, co
and indeed short-ranged (compare for example Figs. 9 and 11). The screened
interaction decays proportionally to Ir - r’1-l for Ir - r’I -+ co, but this
long-ranged tail can be taken into account e~plicitly,”~ allowing the use of
a finite cut-off for Ir - r’I in numerical work. Although G, X, and W have
branch cuts on the real-frequency axis, they can be analytically continued
to the imaginary-frequency (and imaginary-time) axis along which they are
much smoother and can be accurately discretized by an equidistant Fourier
grid, as shown in Fig. 23.
The central task of the real-space/imaginary-time algorithm is the same
as that of the reciprocal-space/frequency-space algorithm: determination of
G -, Po -, E W -, X and subsequent calculation of quasiparticle shifts and,
if desired, the solution of Dyson’s equation. Here, we will describe in detail
the construction of G and Po and the Fourier transformation of Po, since
their construction is used to illustrate the parallel real-space/imaginary-time
algorithm in Section 111.17. For further details and convergence studies we
refer the reader to the work by Rieger et

306 All real-space/imaginary-time algorithms implemented so far are based on pseudopoten-

tials and use plane waves as basis functions.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 99

2
h

2
v

L3 o
.-
0
-
9
-2

20

-z
A
3
10

0
-
9 0

-10

-60 -40 -20 0 20 40 60

0 (W
FIG. 23. Energy-dependent correlation contribution to the self-energy of Si as a function of
imaginary (upper panel) and real (lower panel) frequency. The expectation value of ZC has
significantly less structure along the imaginary axis compared to the real axis, allowing the use
of Fast Fourier Transforms with equidistant frequency and time grids to switch between
imaginary time and frequency. (From Ref. 103.)

b.1. THEGREEN FUNCTION.Analytic continuation of the spectral represen-

tation (Eq. (2.4)) of the noninteracting Green function from real to imagin-
ary energy and Fourier transformation results in3’’
i XF@nk(r)@n*k(r’)exp(q,kr),r > 0
G’(r, r’; ir) = (3.9)
- i X:Pcc@nk(r)@:k(r’) exp(Enkr), r < 0,
In this approach, r denotes a point in the irreducible part of the real-space

307 The index i has been replaced by the band index n and the reciprocal vector k in the first
Brillouin zone. Spin is disregarded. The Fermi energy, which is at zero in this subsection, lies
in the band gap; that is, if' < 0 and E;Y z 0.
 100 WILFRIED G . AULBUR, LARS JONSSON, AND JOHN W. WILKINS

unit cell while r’ denotes a point in the “interaction cell” outside of which
Go is set to zero. Denote as unk(r) the periodic part of the Bloch wave
function Onk(r),and define for a given r a vector r” in the real-space unit cell
and a real-space lattice vector L, such that308

r’ = r + r” + L = r + x; x = r” + L. (3.10)

Then
occ(unocc)
Go(r, r + x; ir) = i sgn(.s) C u,,k(r)t&(r + r”) exp( - ik .x) exp(snkc).
nk
(3.1 1)

Note that T is positive or negative depending on whether the sum is over

occupied or unoccupied states, respectively. Finally, write the sum over
k-points in the full Brillouin zone (BZ) as a sum over k-points in the
irreducible Brillouin zone (IBZ) and over symmetry operations S that
generate the k-points in the full BZ. One obtains the following formula for
Go, which is implemented in the real-space/imaginary-time program:

c c c unk(S-’r)z&(S-’(r + r”))
occ(unocc) IBZ
Go@, r + x; ir) = isgn(T)
n k S

x exp( -iSk.x) exp(cnkr). (3.12)

Assuming an interaction cell that is of constant size for a given material, the
determination of the independent-particle propagator scales as N , x Nband
x N,, where N , is the number of frequencies or times considered; Nband, the
number of bands in the summation of the above equation; and N,, the
number of r-points in the real-space unit cell. Since Nband and N , increase
linearly with the size of the system, the overall scaling of the method is
proportional to the system size squared and hence of order O(N2,,).
Convergence considerations. Convergence parameters important for the
real-space approach are (1) the number of bands in Eq. (3.12), (2) the real-
space grid spacing Ar and the size of the interaction cell, and (3) the time
spacing AT and the maximum sampling time rmax.Convergence to within
20meV for Si requires 145 bands, a Ar = 0.32 a.u., an interaction cell
radius309 Rmax= 18 a.u., a time spacing AT = 0.15 a.u., and T~~~ z 20 a.u.

308 The r and r‘ meshes are not offset, which consequently requires special treatment of the
Coulomb divergency; see Ref. 103.
309 Assume a spherical interaction cell for simplicity.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 101

This corresponds to 170 r-points in the irreducible wedge of the real-space

unit cell and to Nr.% (471/3)(R~~JAr)~ % 0.75 x lo6 points in the interac-
tion cell.310The propagator Go(r, r’; r) requires 0.8 MW storage for every r
and each of the r m a x / A =
~ 130 time points. Storage requirements for the
algorithm are therefore large despite the finite range of the interaction cell
and may become a bottleneck of the calculation for larger systems.311
b.2. THE INDEPENDENT-PARTICLE POLARIZABILITY. The independent-par-
ticle p ~ l a r i z a b i l i t y ~can
~ ’ be determined in real space and imaginary time
by an O(N,,) operation (compare Eq. (2.13)):
Po(r, r’; iT) = - iGo(r, r’; ir)Go(r, r’; -ir). (3.13)
However, the dielectric matrix and the screened interaction require convol-
utions in real space (Eqs. (2.20) and (2.14)) but multiplications in reciprocal
space (Eqs. (2.23) and (2.16)). It is easier to construct these quantities in
reciprocal space and as a function of imaginary frequency, which requires a
Fourier transformation of Po(r, r’; ir). The spatial Fourier transforms
require special care since Po is not translationally invariant in the 6-
dimensional space spanned by r and r‘. In the following equations, we do
not exhibit the time/frequency dependence for simplicity and define Po(k, r,
r’)-where both r and r’ are in the real-space unit cell-as
.,
IC(r1

Po(k, r, r’) =
R
exp( - ik. (r - r’ - R))Po(r, r’ + R). (3.14)

The sum over lattice vectors R is limited to vectors in the interaction cell of
r (ZC(r)). The quantity Po(k, r, r’) is periodic in r and r’ separately, and its
Fourier transform can be determined by two successive 3-dimensional
F F T s . ~ The Fourier transform of the independent-particle polarizability is
given as (V, is the unit cell volume)314

P&(k) = v,1 fvc fvc exp( -iG.r)Po(k, r, r’)exp(iG’.r’)drdr’. (3.15)

3 ’ 0 Materials such as GaN with larger density variations than Si will require an even smaller
Ar to obtain similar convergence. The range of the interaction cell should not increase,
although the dielectric constant of GaN is less than half of the dielectric constant of Si since
the short-range Green function, not the long-range screened interaction, determines the range.
311 L. Steinbeck, private communication. Note that Steinbeck, Rieger, and Godby recently

developed algorithms that reduce memory requirements by an order of magnitude.

3 1 2 We discuss the case of an RPA-based GWA calculation for simplicity.
3 1 3 This so-called “mixed-space’’ representation of the nonlocal operators in the GWA was
introduced by Godby et al. (Ref. 185) and discussed in detail by Blase et al. (Ref. 279).
314 A slightly different notation was used in Ref. 103.
102 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Once Po is determined in reciprocal space and imaginary time, the

determination of the dielectric matrix and the screened interaction reduces
to matrix multiplications and inversions that scale as the third power of the
matrix size. The actual computation is not dominated by this part of the
algorithm since the dimension of the dielectric matrix can be chosen to be
much smaller than the number of points in the real-space unit cell.
The self-energy must be calculated in real space and imaginary time to
avoid convolutions. Therefore, the screened interaction W must be Fourier-
transformed from reciprocal space and imaginary frequency to real space
and imaginary time, with careful consideration of the long-ranged part of
W103 Expectation values of the self-energy operator (@'nkIC(i~)l@nk) can
then be determined after Fourier transformation of C from imaginary time
to imaginary frequency. The analytic structure of G and W dictates that the
self-energy has poles in the second and fourth quadrant of the complex
plane. One can therefore analytically continue the self-energy from the
negative (positive) imaginary-frequency axis to the negative (positive) real-
frequency axis without crossing any branch cuts. This analytic continuation
is achieved by fitting the self-energy along the imaginary-frequency axis to
a multipole function.'03

DETAILS
16. PLANEWAVES:NUMERICAL

This section has a threefold purpose: (1) Detailed information on the use of
symmetry and the integration of the Coulomb divergence in actual GWA
implementations (see subsections a and b) is presented. (2) The effects of
different numerical parameters on quasiparticle energies in plane-wave-
based GWA calculations are specified (see subsections c and d). (3) Details
about a particular GWA implementation are provided (see subsections e
and f).

a. Use of Symmetry
The special k-points technique3' reduces by symmetry the summation over
a uniform mesh of k-points needed to integrate the independent-particle
polarizability (Eq. (3.3)) and the self-energy (Eqs. (3.5) and (3.6)) over the
Brillouin zone to the summation over a smaller set of special k-points. With
S as an element of the little group L, of the external wave vector,316 that is,
Sq = q, and T~ as its nonsymmorphic translation, the matrix elements of Eq.

315 H. J. Monkhorst and J. D. Pack, Phys. Rev. B 13, 5188 (1976).

316 Umklapp processes are neglected since they are only important for k-points on the
Brillouin zone boundary.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 103

(3.2) can be written ass1

ME(S-'k, q) = exp(i(q + S G ) . Q M % ( k , 9). (3.16)

With BZq as the irreducible part of the Brillouin zone defined by the little
group of q, the static independent-particle polarizability (Eq. (3.3)) is given
by8'

where the contribution from each k-point in the BZq is multiplied by an

appropriate weight. The use of symmetry reduces considerably the computa-
tional demand for q-points along high-symmetry directions. For instance,
256 k-points in the full zincblende (cubic) Brillouin zone can be reduced to
40 for q parallel to qx and to 60 for q parallel to qL. Care must be taken to
add up the phase factors in the above equation correctly.
The external wave vector k for Cx and Cc(E) determines the Brillouin
zone BZ, that is relevant for the integration of the self-energy in reciprocal
space. Application of the elements of the little group k to Eqs. (3.5) and (3.6)
leads to a cancellation of phase factors and the following expressions for C x
and Cc(E):

(3.19)

where ct = 2nzPqo,(q) for the von der Linden-Horsch and ct = 1 for the
Engel-Farid plasmon-pole model.

b. Integration of the Coulomb Singularity

The expectation value of the self-energy contains sums over umklapp
processes whose G = 0 contribution has an integrable l/lqI2 divergency for
q+O. Consider the case317 of Cx and let f(q) be a smooth function that

317 The case of Z C ( E ) is analogous and will not be discussed here.

104 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

diverges as l/lqlz for q -,0 and whose integral over the Brillouin zone is
known analytically. Rewrite the expectation value of C x as9'

(3.20)

The term in square brackets is no longer singular at q = 0 and can be

integrated using special points. The last term is integrated analytically.
Possible choices for f(q) for fcc symmetry are (a = lattice constant,
V, = unit cell v~lurne)~'
(a/a2
f(q) = 3 - cos(aqx/2)cos(aqJ2) -cos(aq$2) cos(aqz/2)- cos(aqz/2)cos(aqx/2)'

(3.21)

and with 0 as the step function and b = lqLl = fix/u,"'"

Choosing either Eq. (3.21) or (3.22) for f(q) leads to deviations in the
valence and low-lying conduction quasiparticle energies of Si and Ge at
r, X , and L of less than 30meV. Functions f(q) appropriate for sym-
metries other than fcc have been suggested by Wenzien, Cappellini, and
Be~hstedt.~~

c. Convergence
Quasiparticle band structures can vary by a few hundreds to a fev tens of
meV because of differences in (1) the number of q-points used in the
integration of Eqs. (3.5) and (3.6); (2) the energy cut-off used in the
expansion of the LDA wave functions; (3) the number of single-particle and
plasmon-pole bands kept in the summations in Eq. (3.6); (4) the number of
G-vectors kept in the summation in Eq. (3.7); and ( 5 ) the degree of
convergence of the plasmon-pole band structure. To determine reliable

318 W. H. Backes, private communication.

319 B. Wenzien, G. Cappellini, and F. Bechstedt, Phys. Rev. 851, 14701 (1995).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 105

quasiparticle band structures, all of these numerical cut-offs have to be

checked systematically. Convergence with respect to the numerical cut-offs
mentioned in items 1 through 4 is generally smooth,320as can be inferred
from Figs. 24-27 and Table 17. Details of the calculations are given in the
captions. A plot of the logarithm of the relative deviation of the quantity to

320 Similar conclusions hold for the convergence properties of the plasmon-pole parameters,
point (3,which are not presented here.
' 2 1 C. M. Bender and S. A. Orszag, Advanced Mathematical Methods for Scientists and

Engineers, McGraw-Hill, New York (1978), 369.

-1

Si
-2
K risV=o.o

c
.-0
.-iii -3
5
al
U
>
.-c
al
-m
c
E
0
0 -4
-

-5

-6
20 40 so
number of k-points in IBZ

FIG.24. Convergence of valence and conduction band energies at the point in Si as a

function of the number of k-points used in the integration of the self-energy (Eqs. (3.18) and
(3.19)). Convergence is smooth, with the equivalent of 10 k-points in the IBZ being necessary
to converge the low-lying and r2, to better than 1%. The top of the valence band is set
to zero in each case. The converged value for each energy is determined by Shanks iteration.321
All other numerical cut-offs are given in Table 18.
106 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Si

-5
number of reciprocal lattice vectors in sum

FIG. 25. Convergence of valence and conduction band energies at the r point in Si as a
function of the number of G-vectors in the internal sum over G-vectors in Eq. (3.7).
Convergence is smooth, and about 100 G-vectors are sufficient to converge the energies to less
than 1%. The top of the valence band is set to zero in each case. The converged value for each
energy is approximated by the energy calculated with 283 G-vectors corresponding to a cut-off
in reciprocal space of 3.9 Hartree. All other numerical parameters are identical to those in
Table 18.

be converged E from its converged value Em as a function of the conver-

gence parameter N is roughly linear.322 In other words, with 01 being a
constant
log,, (y) x -uN 3 E = E,(1 - (3.23)

322 Note that fitting schemes for convergence plots are not unique, as discussed, for example,
in S. Wei, D. C. Allan, and J. W. Wilkins, Phys. Rev. B46, 12411 (1992) (see Fig. 3 and
discussion thereof).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 107

-1

-2

c
.-0
.-z -3
a

-
V
a
.->
m
-
g!
c
0
-4
-
rn
0

-5

-6
40 80 120
number of plasmon pole bands

FIG. 26. Convergence of valence- and conduction-band energies at the r point in Si as a

function of the number of plasmon-pole bands included in the sum of Eq. (3.19). Convergence
is smooth, and 40 plasmon-pole bands are sufficient to converge results to within 1YO.
The top
of the valence band is set to zero in each case. The converged value for each energy is
approximated by the energy calculated with 130 plasmon-pole bands. The sum over reciprocal
G-vectors is cut off at 2.7 Hartree. All other numerical parameters are identical to those in
Table 18.

The numerical cut-offs used for Si as well as cut-offs resulting from

convergence studies of the other four materials considered in Section 111.18
are summarized in Table 18.
Convergence properties differ from material to material and have to be
repeated for every material under consideration. Convergence studies as a
function of the energy cut-off and the number of k-points are costly, since a
new plasmon-pole band structure has to be generated each time. Often
convergence studies can be limited to two and ten special k-points. An
appropriate energy cut-off can be determined by choosing the cut-off large
108 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

-1

-2

s
.-
iii -3
.-
U
a,
a,
.-
-m
c

p!

i
c

- -4

-5
Y

R
"25 75 125 1 J
number of bands in sum over states

FIG. 27. Convergence of valence- and conduction-band energies at the r point in Si as a

function of the number of bands included in the sum over states of Eq. (3.19). Convergence is
smooth with the exception of the Y2,state, which is converged with as little as 40 bands, 60
bands are necessary for good convergence ( < 1 YO)of all energies. The top of the valence band
is set to zero in each case. The converged value for each energy is approximated by the energy
calculated with 300 bands. The maximum number of bands available at an energy cut-off
of 8.5 Hartree is about 310. The sum over reciprocal G-vectors is cut off at 2.7 Hartree. All
other numerical parameters are identical to those in Table 18.

enough to converge LDA valence-band and low-lying conduction-band

energies to within 50 meV. For example, for Si an increase in energy cut-off
from 8.5 to 18 Hartree changes the LDA energies of the valence and the four
lowest conduction bands at r, X , and L by less than 80meV; the corre-
sponding quasiparticle energies (shown in Table 17), by less than 0.1 eV.

d. Choice of Pseudopotentials and Plasmon-Pole Models

Differences in pseudopotentials can influence LDA band structures by about
0.1 eV for states close to the band gap. Hybertsen and collaborator^^*^^ used
 QUASIPARTICLE CALCULATIONS IN SOLIDS 109

TABLE17. CONVERGENCE OF VALENCE- AND CONDUCTION-BAND ENERGIES AT THE r POINT IN

si AS A FUNCTION OF THE ENERGYCUT-OFF USED TO DETERMINE THE SELF-CONSISTENT
GROUND-STATE DENSITY,AS WELL AS THE WAVE FUNCTIONS USED TO CALCULATE THE
INDEPENDENT-PARTICLEPOLARIZABILITY AND THE EXPECTATION VALUES OF THE SELF-ENERGY
OPERATOR. DEVIATIONS
BETWEENTHE DIFFERENT EIGENVALUES ARE LESSTHAN 0.08eV. ALL
CALCULATIONAL PARAMETERS ARE GIVEN IN TABLE 18 EXCEPT FOR THE CUT-OFF IN RECIPROCAL
SPACE FOR THE DETERMINATIONOF Z, WHICHIS 2.7 a.u. FOR THE PRESENT STUDY. THECUT-OFF
ENERGYE,,, IS GIVEN IN HARTREE.

E,,, r,c Y25" r; 5c r2, l-k

8.5 - 11.90 0.00 3.25 3.86 8.32

12.0 -11.88 0.00 3.17 3.80 8.26
18.0 - 11.88 0.00 3.23 3.83 8.33

TABLE18. PARAMETERS USED TO DETERMINE THE SELF-CONSISTENT

DENSITY n(r), THE INDEPENDENT-PARTICLE POLARIZABILITY Po (SEE EQ.
(3.17)), AND THE SELF-ENERGY Z (SEE EQS. (3.18) AND (3.19)) FOR s i , Ge,
G a A s , Sic, AND GaN. ALL WAVE-FUNCTION COMPONENTS ARE KEPTIN
THE CALCULATION. THEVON DER LINDEN-HORSCH PLASMON-POLE MODEL
AND THE GYGI-BALDERESCHI SCHEMEFOR THE INTEGRATIONOF THE
COULOMB SINGULARITY ARE USED. FOR THE LATTERSCHEME, THE
PLASMON-POLE BAND STRUCTURE IS DETERMINED AT q = O.00125qx,
USING THE EQUIVALENT OF 60 SPECIAL k-POINTS IN THE IRREDUCIBLE
BRILLOUIN ZONE. TROULLIER-MARTINS PSEUDOPOTENTIALS WITH
NONLINEAR CORECORRECTIONS FOR GaAs AND GaN ARE USEDTO MODEL
THE IONIC POTENTIAL. THE ENERGYCUT-OFFE,,, IS IN HARTREE;THE
CUT-OFF VALUEG,,, FOR SUMSOVERRECIPROCAL LATTICEVECTORSIS IN
ATOMIC UNITS.N,, EQUALS THE NUMBER OF UNOCCUPIED BANDSFOR P o
AND THE NUMBER OF OCCUPIED AND UNOCCUPIED BANDSFOR Z. Nkpt IS
THE EQUIVALENT NUMBEROF k-POINTS IN THE IRREDUCIBLE BRILLOLIIN
ZONE OBTAINED USING THE FULLSPACE-GROUP SYMMETRYOF THE
CRYSTAL, AND N,,, IS THE NUMBER OF PLASMON-POLE BANDSKEPT.ALL
CALCULATIONS ARE DONEAT THE EXPERIMENTAL LATTICECONSTANT.

Si Ge GaAs SIC GaN

10 28 28 10 10
8.5 10 10 25 25
10 10 10 10 10
8.5 10 10 25 25
3.4 3.4 3.1 4.5 4.5
146 146 146 146 196
10 28 10 10 10
8.5 10 10 25 25
3.0 3.0 3.0 4.5 4.5
60 60 60 100 80
60 100 100 150 150
110 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Kerker pseudo potential^;^^^ Godby et u1.,6*185Rohlfing et u1.,52*114and

Backes et dL3' used Bachelet-Hamann-Schliiter pseudo potential^;^^^ and
Ref. 42 used soft Troullier-Martins pseudo potential^.^^ Deviations for
well-converged LDA and GWA calculations from different authors of about
0.1 to 0.2eV are to be expected. Table 19 estimates the influence of the
pseudopotential choice on GWA calculations by comparing quasiparticle
energies obtained with a Troullier-Martins p s e u d o p ~ t e n t i awith
l ~ ~ ~energies
obtained for a generalized norm-conserving p s e u d o p ~ t e n t i a l Deviations
.~~~
between the LDA and GWA calculations using both pseuodopotentials do
not exceed 0.2 eV for both the valence and low-lying conduction bands.
Corresponding deviations for self-energy shifts are 60 meV or less.
Table 20 compares Si quasiparticle energies obtained using the von der
Linden-Horsch and the Engel-Farid plasmon-pole models and shows
that the choice of different models changes quasiparticle energies by 0.1 eV
or less. This agrees with the findings of Northrup et who found
only small differences in the band structure of nearly-free-electron metals

323 G. P. Kerker, J . Phys. C 13, L189 (1980).

324 G. B. Bachelet, D. R. Hamann, and M. Schliiter, Phys. Rev. B26,4199 (1982).
325 N. Troullier and J. L. Martins, Phys. Rev. B43, 8861, 1993 (1991).
3 2 h These pseudopotentials are based on the Teter '93 pseudopotentials (M. Teter, Phys. Rev.
B48, 5031 (1993)) and were provided by D. C. Allan. They do include core charges, but they
do not take hardness into account, which leads to a much lower plane-wave kinetic energy
cut-off. In the case of Si we have for instance E,,, = 11 Hartrees for pseudopotentials without
hardness conservation, versus E,,, = 25 Hartrees for pseudopotentials with hardness conserva-
tion.

TABLE19. COMPARISON OF LDA AND GWA ENERGIES AND THE QUASIPARTICLE SHIFTAT THE r
POINT IN si FOR A TROULLIER-MARTINS PSEUDOPOTENTIAL325 (TM) AND A GENERALIZED
NORM-CONSERVING326(GNC) PSEUDOPOTENTIAL.DEVIATIONS IN THE GWA BANDSTRUCTURE
ARE LARGEST (0.18eV) FOR THE r2,STATE,WHICH IS THE MOST DIFFICULTLOW-LYING
CONDUCTION-BAND STATETO CONVERGE. THECHOICEOF THE PSEUDOPOTENTIAL AFFECTSTHE
QUASIPARTICLE SHIFTSSIGNIFICANTLY LESSTHAN THE ABSOLUTEENERGIES (60-meV MAXIMUM
DEVIATION).ALL NUMERICAL PARAMETERS ARE GIVENIN TABLE 18 EXCEPTFOR G,,,, WHICH
EQUALS 3.1 a.u. r~ THE PRESENT STUDY.ALL ENERGIES ARE IN eV.

PSP

TM LDA - 11.98 0.00 2.52 3.15 7.64

GWA - 11.92 0.00 3.19 3.82 8.29
GWA -LDA 0.06 0.00 0.67 0.67 0.65
GNC LDA - 12.07 0.00 2.54 3.27 7.70
GWA - 12.01 0.00 3.21 4.00 8.36
GWA-LDA 0.06 0.00 0.67 0.73 0.66
 QUASIPARTICLE CALCULATIONS IN SOLIDS 111

TABLE20. COMPARISON OF QUASIPARTICLE ENERGIESCALCULATED WITH THE VON DER LINDEN-

HORSCH'" (LH) AND ENGEL-FARID'~' (EF) PLASMON-POLE MODELSFOR HIGH-SYMMETRY
POINTSIN Si. THE TYPICAL DEVIATION BETWEEN THE Two SETSOF QUASIPARTICLE ENERGIES
CLOSETO THE BAND GAPIS ON THE ORDEROF A FEW TENS OF meV. THELARGEST DISCREPANCY
OF O.lOeV OCCURSFOR THE LOW-LYING r,"STATE,FOR WHICH THE PLASMON-POLE MODELIS
AN INSUFFICIENT
APPROXIMATION. ALLCALCULATIONAL DETAILSARE IDENTICALTO THOSE GIVEN
IN TABLE18 EXCEPTFOR G ,,,, WHICH EQUALS3.1a.u. IN THE PRESENT STUDY.MINOR
DIFFERENCESWITH THE LH Si BANDSTRUCTURE ENERGIESIN TABLE21 RESULT FROM A BETTER
CONVERGENCE IN THE LATTER.ALL ENERGIESARE IN eV.

LH EF LH EF LH EF

rl" -11.92 -11.82 XI, -1.92 -7.86 L;, -9.67 -9.59

r
25v 0.00 0.00 X4, -2.98 -2.96 L,, -7.15 -7.01
r15c 3.19 3.21 XI, 1.26 1.30 L;, -1.26 -1.25
r;, 3.82 3.81 X,, 10.66 10.60 L,, 2.08 2.10
rlc 8.29 8.28 L,, 4.08 4.10
L;, 8.17 8.18

using the Hybertsen-Louie model compared to an earlier calculation by

Lundqvist9' that used a different plasmon-pole model.

e. Object Orientation
The common computing paradigm in most scientific applications is pro-
cedure-oriented programming. The central task of a software engineer using
the procedure-oriented paradigm is to identify the procedures, that is, the
data manipulations, required to solve a problem and the corresponding best
algorithms. The structure element of a procedure-oriented code is the
procedure or subroutine with its input and output data. This way of
programming leads in general to many global and few local data. The global
data are either passed explicitly from subroutine to subroutine or-as in
Fortran -via COMMON blocks. This violation of the software engineering
principle of data can lead to severe problems once a complex
piece of software has to be debugged, modified, or extended beyond the
realm of its initial use.328

3 2 7 B. Stroustrup, The C++ Programming Languuge, 2nd ed., Addison-Wesley, New York
(1991) 14ff.
3 2 8 For example, it took a graduate student at the University of Illinois, Urbana-Champaign,
two years to include d basis functions into an O ( N ) tight-binding code. The 20,000 lines of O ( N )
code were originally implemented for s and p functions only (D. Drabold, private communica-
tion).
112 WILFRIED G. AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

An object-oriented programming paradigm329 conceives the task of

solving a computational problem as identifying interacting classes or ob-
jects. A class is an abstract or a user-defined data type. It comprises data
and all operations necessary to manipulate it. For instance, the abstract data
type VECTOR is the ensemble of the operations that create, destroy, and
manipulate (scalar and vector multiplication, addition, etc.) a vector, an
integer corresponding to the vector length, and a pointer to its first element.
Classes consist of a header file that declares the class type- that is, its
data members and member functions-and a file containing the actual
implementation of the class functions. A software user should be able to use
a class efficiently in a particular application, by looking only at the header
file. Details of the representation of the data as well as algorithms used for
a particular numerical problem are hidden from the user. Moreover,
unauthorized user access to class data is prohibited. This principle of data
or information hiding reduces the use of global variables and leads to the
design of reusable, local program modules with encapsulated data.
Classes as program modules should be written in a reusable and general
way. For instance, a class VECTOR should not be limited to vector-vector
operations but should also allow matrix-vector operations. The principle of
inheritance can increase the reusability of code considerably. The basic idea
is that common features of two different classes are incorporated into a base
class whose properties are then inherited by the so-called derived classes. For
example, consider a base class HUMAN with private data AGE and WEIGHT
and member functions EAT, DRINK, and SLEEP. A derived class STUDENT
would contain all the functionalities of HUMAN but, on top of them, the
member functions ATTEND -LECTURE, DO -HOMEWORK, and GOOF -OFF.
The reciprocal-space GWA code of Ref. 42 is written in C + + in an
object-oriented style. Data abstraction at a low level is achieved through
extensive use of an in-house C + + library that defines the abstract data
types VECTOR and MATRIX.330At a higher level, the central classes of the
code are symmetry,selfenergy,polarizability,and wavefunc-
t ion. To describe data encapsulation and code reusability, we consider
symmetry and wavefunction as examples. The class symmetry can
currently handle zincblende materials such as GaAs and elemental semicon-
ductors such as Si. However, it provides the full functionality (apply a
particular symmetry operation, find its inverse and apply the inverse,
determine the little group of a vector, etc.) irrespective of the space group

3 2 9 M. A. Ellis and B. Stroustrup, The Annotated C++ Reference Manual, Addison-Wesley,

New York (1990).

3 3 0 This library was written and is maintained by W. Wenzel, University of Dortmund,
Germany, and M. M. Steiner, Ohio State University.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 113

under consideration. Treating a system with, for example, wurtzite symme-

try therefore requires only hardcoding of the corresponding symmetry
operations in the implementation of symmetry (about 150 lines out of a
total of 10,000 lines of code). The class wavefunction contains the
k-point at which the wave function is defined, the number of bands and
plane waves kept, and the wave-function coefficients and energies. More-
over, it contains basic operations such as time reversal, convolution, and
Fourier transform. These basic operations are written in a general manner
such that they do not apply only to the construction of the matrix elements
of the independent-particle polarizability (see Eq. (3.2)) from the wave
functions. They can be used as well for the plasmon-pole eigenvectors (also
of type wavefunc tion) and the construction of the self-energy expectation
values (see Eq. (3.7)).

f. Eficiency
Since C + + is less well established in the scientific community than is
Fortran, an interface with existing Fortran codes such as low-level BLAS
subroutines is a must. Efficient code results, for example, from interpreting
wave functions as matrices rather than as linear arrays and by performing
convolutions as in Eqs. (3.2) and (3.7) via BLAS3 matrix multiplications. As
a consequence, major portions of the code run at about 270 MFlops on a
CRAY-YMP8 (peak performance 333 MFlops). The overall code perform-
ance is 170 MFlops and the code uses memory very efficiently according to
the standard CRAY performance tools. About 20% of the program is
documentation to make it easy to maintain and use.
A good test of the reliability of a code is a “standardized” test calculation
and comparison with an independently written code. To eliminate subtle
differences in pseudopotentials, parametrization of exchange-correlation
potentials, and the like, which plague input to self-energy calculations using
different LDA codes, one can generate wave functions and energies using an
empirical Si p ~ e u d o p o t e n t i a instead.
l ~ ~ ~ The corresponding wave functions
and energies are, in addition, easy to generate. Results based on empirical
pseudopotentials between two independent codes from Refs. 42 and 302
agree to within nine significant digits.

GWA CALCULATIONS
17. PARALLEL

The use of local-orbital basis functions or recent, real-space algorithms can

cut down on the computational cost of quasiparticle calculations, as

331 M. L. Cohen and T. K. Bergstresser, Phys. Rev. 141, 789 (1966)

114 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

discussed in Section 111.15. Another way to speed up quasiparticle calcula-

tions is the use of massively parallel computers. This section describes
parallel algorithms for reciprocal-space and real-space quasiparticle calcu-
lations.

a. Reciprocal-Space Approach
The basic idea for a parallel determination of the independent-particle
polarizability (Eq. (3.17)) and the self-energy (Eqs. (3.18) and (3.19)) is to
distribute the conduction bands over n p e s processing elements (PEs)
during the course of the calculation. The principle will be explained for the
case of the independent-particle polarizabiliiy; the case of the self-energy can
be handled analogously. Let PEO denote the PE that does all the 1 / 0
between the file system and the parallel machine; Nvblcbthe number of
valence/conduction bands; N, the number of G-vectors; and N , , the
number of plane waves.
For each k-vector PEO reads the energies and the G-vectors. Then it reads
the first N,,/npes conduction bands (assume mod(Ncb/npes)= 0 for
simplicity) and sends the information about k-point, energies, G-vectors,
and the first block of conduction bands to PE1. Subsequently, PEO reads
the second block of conduction bands and sends it together with the k-point,
energies, and G-vectors to PE2, and so on. Once all conduction bands are
distributed, PEO reads the valence-band information for a point k - q in the
BZ, and broadcasts that information to all PEs. Then all PEs -including
PEO -calculate the following expression:

(3.24)

Finally, the results of this calculation are added up via a single-node

accumulation on PEO. For a graphical description of the algorithm see Fig.
28. The important features of this parallel implementation are as follows.

1. Storage requirements. The storage needed for complex wave functions

on each PE is proportional to ( N u b+ N,,/npes)N,,. The memory
requirements per node for wave functions increase approximately
linearly with system size since N,, << N,, and since an increasing
number of processors can compensate for increasing system size. For
the matrix elements M g the storage requirements are proportional to
N,,(N,,/npes)N,, which leads to a quadratic memory increase per
node with system size. The use of massively parallel computers leads
to a gain in memory per node by one order in system size for both
wave functions and matrix elements.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 115

mp
...over
....q-points
...............
loop
.. ..over
.....k-points
.... .
.. ..
...

'
'
I
I read CBs i*ncb/npes.
' I
' I
' II

from disk and send to

$ 1
PEi

read VB for k-point in

BZq and broadcast to
all PEs

calculate

single-node accumulation
of partial sums a t PEO to
calculate Plw(k.q)

FIG. 28. Schematic diagram of a parallel calculation of the independent-particle polarizabil-

ity in reciprocal space (Eq. (3.24)). The same scheme applies to the calculation of the self-energy
in reciprocal space (Eqs. (3.18) and (3.19)). PEO manages disk IjO and distributes the valence
and blocks of the conduction-band wave functions over all PEs. Each PE calculates its partial
contribution to the independent-particle polarizability. The partial contributions are summed
up at the end of the calculation via a single-node accumulation on PEO.

2. Scalability. The increase in the volume of the system by a factor of two

leads to twice as many conduction bands. The corresponding matrix
elements can be calculated efficiently by using twice as many proces-
sors, leading to a scalable algorithm. Moreover, scalability applies to
memory usage as well, as discussed under storage requirements.

3. Load balance. The load balance is excellent. Not counting the 1/0 over-
head, all PEs have exactly the same workload if mod(Ncb,npes)=O.
If mod(N,,,npes) # 0, then the last n p e s - 1 PEs calculate the
116 WILFRIED G . AULBUR, LARS JONSSON, AND JOHN W. WILKINS

matrix elements M E for N,,/npes conduction bands, while PEO

calculates M g for the remaining conduction bands. This distribution
of workload compensates partially for the extra work that PEO has to
do to assemble P&,(q,O) at the end of the calculation (compare
Fig. 28).

4. Communication and portability. The algorithm is coarse-grain parallel

and portable. PEO reads and sends the input information and wave
function data to all other processors. These calculate the matrix
elements for every k-point and send the results of their calculation back
to PEO. PEO then assembles the total polarizability. Therefore, for a
fixed k-point, communication occurs only at the beginning and at the
end of the computation. The use of either MPI or PVM communica-
tion primitives in the coarse-grain parallel code ensures portability to
different platforms.332The use of symmetry as described in Eq. (3.17)
does not require any interprocessor communication since the matrix
elements for all G-vectors, and hence for their images under symmetry
operations of the little group of q, are created and used locally on each
processor.

5. Object orientation-Use of C + + . The central object in the code is

wavef unc t ion. As described in Section III.17e, this class comprises
the k-point at which the wave function is defined, the number of bands
and plane waves kept, and the wave function coefficients and energies.
Sending a wave function w v f n c t n from processor my-pe to proces-
sor y o u r -pe requires only the following statements:

The actual sending of the data and the communication primitive used in
this process are hidden from the user. This feature makes a change
from PVM to, for example, MPI easy, since the change is done in only
one function. Switching between PVM and MPI can be conveniently
achieved via preprocessor directives.

During the course of the code development the parallelization scheme

changed twice. Parallelization schemes used are (1) the embarrassingly
parallel approach of distributing k-vectors over P E s , (2)
~ ~a ~parallelization

332 Fully optimized PVM executables are no longer being created for most new platforms.
333 W. G. Aulbur and J. W. Wilkins, Bull. Am. Phys. Soc. 40,No. 1,251 (1995).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 117

over reciprocal G - v e c t o r ~ and

, ~ ~ ~(3) a parallelization over conduction
bands. Changing the code to accommodate new parallelization schemes
profited immensely from the flexibility that an object-oriented programming
style offers compared to a procedure-oriented programming style.

b. Real-Spaceflmaginary- Time Approach

For the reciprocal-space GWA calculation, scalability can be attained by
parallelizing over the conduction bands. In real space, the r-vectors in the
irreducible real-space unit cell are an appropriate variable for a scalable
parallel i m p l e m e n t a t i ~ nThe
. ~ ~ ~number of vectors in the irreducible real-
space unit cell is directly proportional to system size and hence the number
of atoms in the system. A system that increases its size by a factor of two
can therefore be handled by using twice as many processors. The important
operations in the program are the assembly of the Green function (Eq.
(3.12)) and of the independent-particle polarizability (Eq. (3.13)) in real
space and imaginary time, and the subsequent transformation of the
independent-particle polarizability to reciprocal space and imaginary fre-
q ~ e n c y . ~ lAll
* subsequent operations in the code are either the exact
inversion (for instance, one-to-all broadcast and single-node accumulation)
or slight modifications of these basic operations. It is sufficient to discuss
parallelism only up to the construction of the screened interaction in
reciprocal space and imaginary frequency, which is shown schematically in
Figs. 29 and 30.
Using managed 110, PEO starts by reading input information about the
lattice, k-points, LDA energies, etc., and by broadcasting it to all other PEs.
Then, in a small serial portion of the program, PEO reads the Fourier
coefficients for each band and FFTs them to obtain the periodic parts of the
Bloch wave functions unk(r)and U:k(r’). These wave functions are written to
direct-access files. For u,k(r) the record index is the F F T index of the
r-vector in the real-space unit cell. The record contains unk(r)for all bands
and all k-points. For U n k ( r ’ ) the record index is the number of the special
k-point, and the record contains all r’-points in the real-space unit cell for
all bands.
Next an external loop over imaginary time starts in which the nonin-
teracting propagator at imaginary times f i and ~ for all r-vectors in the
irreducible real-space unit cell is determined. Each P E has a list of r-vectors
for which it calculates Go(r, r’; fi T ) for a particular time i.s and all r’ in the

334 This parallelization scheme was later adapted for the parallelization of the real-space/
imaginary-time code; see below.
335 This algorithm was developed by W. G. Aulbur in collaboration with L. Steinbeck, M. M.
Rieger, and R. W. Godby.
118 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

read u ,#I for all bands

and all k-points for PEI

read and broadcast u*Jr’) for

one particular k-point. all bands
and all r’

0
calculate C (r. r’. tit)

0
wrlte P 11. r’, +ato dlsk
for each PEI

dlstrlbutePo ( 1 , r’. +Id

over PEs

do 3D FFT to obtaln
4 , Polk. r. G’. 4
1
, I,
redlstrlbute using MPILGATHER
do 3D FFT to obtain
P o (k. G. G’. 10

write @lk. G. 0’.Ul to dlsk

FIG. 29. Schematic diagram of the communication structure of the parallel implementation
of the real-space GWA code. We list all communication and calculational steps needed to
determine the independent-particle polarizability Po in real space and imaginary time, and to
transform it to reciprocal space. The distribution of the unlr(r)and of Po(r,r‘; ir) over all PEs is
described in the text. The redistribution step using MPI -GATHER is detailed in Fig. 30. The
Fourier transformation from imaginary time to imaginary frequency and assembly of the
screened interaction via parallel matrix operations is also described in the text.

interaction cell according to Eq. (3.12). For example, if the number nviuc
of r-points in the irreducible real-space unit cell is 55 and the number of
processors is 4, then PEO calculates Go for the first 13 r-vectors, PE1 for the
following 14, and so on.
To calculate Go at a point ri, each PE needs u,,k(ri) for all bands and all
k-points, as well as U,*k(r’) for all r’, all bands, and all k-points. Let rank
denote the number of a PE. For each i and each node, PEO reads the
corresponding wave function information unk(ri)in real space and sends it
to PE rank.At the end PEO reads the information that it needs for its own
 QUASIPARTICLE CALCULATIONS I N SOLIDS 119

of G-vectors

First block of r-vectors

in irred. real-space cell PEO
Images under space
group symmetries

Second block of r-vectors

PE 1

Third block of r-vectors PE2

MPI-GATHER

Fourth block of r-vectors ,

I
I
I
I

PE3

FIG. 30. Communication structure for the redistribution of the independent-particle polar-
izability over PEs. After the first 3D FFT each PE has the independent-particle polariz-
ability Po(k, r, G’,ir) for one particular k-point, a subset of r-vectors in the irreducible
real-space unit cell as well as their images under symmetry operations of the space group, all
G’-vectors, and one particular time i t As indicated by the arrows in the figure, MPI -GATHER
is used to assemble Po for all r-vectors and a block of G’-vectors on each PE (first block on
PEO, second on PEI, and so on). After the MPI _GATHER step described in the figure, Po is
reordered according to its r argument to have standard FFT ordering, and the second 3D FFT
is performed on each node. The different length of the r-blocks in the picture corresponds to a
different number of r-points on each PE. The total number of G‘-vectors kept after performing
the first FFT is nG and n p e s is the number of PEs used. In the above example
mod(n,.,npes) = 2.

calculations. Subsequently in a loop over k-points, PEO reads U,k(r’) for a

particular k-point and broadcasts this wave function for all bands and all r’
in the real-space unit cell to all PEs, as shown in Fig. 29. After each PE has
finished the loop over k-points and has calculated its noninteracting
+
propagator at fiz, it sends the product Go(ri, r’; iz)Go(ri, r’; - ir)- that
is; the independent-particle polarizability up to a constant factor -to PEO
and PEO writes the result to disk. This communication is the inverse
operation of the distribution of U,k(ri) described above. A graphical descrip-
tion of this communication step is given in Fig. 29.
120 WILFRIED G. AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

In the next step, PEO reads the independent-particle polarizability for

different vectors ri in the irreducible real-space unit cell and distributes
Po(ri,r’; iz) over the different processors in an analogous way as for the
distribution of unk(ri).Within an external loop over special k-points, each PE
performs the first 3D FFT over r’ on node, turning Po(ri,r’;iz) into
Po(k, ri, G’; iz).
Two points are important for the 3D FFT: (1) Only those G’-vectors in
Po(k, ri, G’; iz) with IG’I < G,,, are kept. In the case of Si, G,,, = 3.4 Hartree,
which corresponds to 169 G’-vectors. (2) After the first 3D FFT each PE
contains all G’-vectors and its chunk of r-vectors in the irreducible real-space
unit cell as well as their images under symmetry operations of the space group
of the crystal. Since the distribution of the r-vectors is not contiguous in any
coordinate, parallel FFTs to perform the next 3D FFT cannot be used.
Instead, a “block-column’’ gather operation using the MPI -GATHER function
is performed, which is illustrated in Fig. 30. The basic idea is to assemble all
r-vectors and a block of G’-vectors on one PE such that a local 3D FFT on
Po(k, r, G’; iz) can be done after reordering the r-vectors in FFT order.
Although the Green function is constructed with close to perfect load
balancing by parallelizing over the r-vectors in the irreducible real-space unit
cell, the load balancing for the construction of the independent-particle
polarizability in reciprocal space is less favorable since now the symmetry
of the r-vectors in the irreducible real-space unit cell is important as
well. However, larger systems have many r-vectors in the irreducible
real-space unit cell, and low symmetry and differences in symmetries should
average out.
After all of these steps, one obtains Po(k, G, G’; iz) for all G-vectors and
a block of G’-vectors, which is written to disk via PEO for all k-points and
all times ir. The next step is the Fourier transform with respect to imaginary
time. To do this Fourier transform, PEO reads quadratic submatrices in
(G,G’) of Po(k, G, G’; iz) for all times iz and distributes these submatrices
over a quadratic grid of processors. The time Fourier transform is done
locally. Further matrix operations such as the construction of the RPA
dielectric matrix and of the screened interaction are done using ScaLapack
routines for the multiplication and inversion of distributed matrices.

18. GWA CALCULATIONS

FOR FIVE
PROTOTYPICAL
SEMICONDUCTORS
This section summarizes all quasiparticle calculations for five
prototypical semiconductors and is intended to serve as a reference for

336 For GaAs, we include unpublished data provided by F. Aryasetiawan (Ref. 290). For Ge,
GaAs, and SIC, we include unpublished 60-Gaussian orbital data by Rohlfing, Kriiger, and
Pollmann (Ref. 303).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 121

readers with a detailed interest in the subject. In addition, the accuracy of

quasiparticle calculations is assessed by comparison of independent calcula-
tions for several materials and is shown to be state- and material-dependent
and often significantly larger than the commonly quoted kO.1 eV. All
quasiparticle calculations yield good results for Si, suggesting that tests of
quasiparticle approaches cannot be limited to this
Relativistic effects and the treatment of core electrons have significant
effects on LDA band gaps, as demonstrated in Table 33. These effects have
to be considered in standard GWA calculations to achieve good agreement
with experiment, as demonstrated by the work of Shirley et ~ 1 . ~ ~ 9 ~ ~

a. Silicon
Silicon is a prototypical semiconductor that has been studied in seven inde-
pendent GWA calculations based on (1) pseudopotentials and plane waves
(Refs. 5,6,42,102, and 103), (2) pseudopotentials and Gaussian orbitals,114and
(3) LAPW.285 In contrast to all other calculations, the pseudopotential
calculations of Rieger et d 1 0 3 and of Fleszar and HankeIo2 avoid the
plasmon-pole approximation by explicitly computing the frequency-dependent
interaction.
Comparison of both the LDA and GWA energy variations between the
different calculations for the lowest conduction-band states of Si listed in
Table 21 reveals
1. All GWA calculations for the lowest conduction band lie within 0.2 eV
of the range of experimental energies observed.
2. Between all calculations, the variations in both LDA and GWA
energies of the lowest conduction band are 0.03 eV and 0.14 eV at r,
0.05 eV and 0.30 eV at X , and 0.1 eV and 0.17 eV at L, respectively.
3. Comparison of plane-wave calculations without an update of the Green
function, disregarding354Ref. 6, gives variations for the lowest conduc-
tion-band energy of 30 meV and 40 meV at r, 50 meV and 40 meV at X ,
and 50 meV and 50 meV at L, for LDA and GWA, respectively.
4. An update of the energies in the Green function (see Section 11.13)
increases the fundamental gap by about 0.1 eV.
Four main conclusions can be drawn from the above observations:
1. LDA and GWA calculations for the lowest direct band gaps in the
prototypical semiconductor Si are accurate to within 0.1 eV for LDA
and 0.1 to 0.3 eV for GWA.

33' This point will be discussed in more detail in Sections VI.29 and 30.
338 E. L. Shirley and S. G. Louie, Pbys. Rev. Lett. 71, 133 (1993).
122 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

TABLE21. COMPARISON OF Si LDA (ROMAN), GWA (BOLDFACE), AND EXPERIMENTAL (ITALIC)

ENERGIESAT THE r, X, AND L POINTS FOR ALL QUASIPARTICLE CALCULATIONS AVAILABLE IN THE
LITERATURE.QUASIPARTICLE ENERGIES OF DIFFERENT CALCULATIONS FOR BOTHCONDUCTION-
AND VALENCE-BAND STATESVARYBETWEEN 0.1 AND 0.6 eV BECAUSE OF DIFFERENTTECHNIQUES,
DIFFERENT DEGREES OF CONVERGENCE, ETC. FOREASEOF COMPARISON, THE LDA AND GWA
VALENCE-BAND TOPSARE SETTO ZERO.THE SUBSTANTIAL ABSOLUTE ENERGYS m s OF THE
GWA VALENCE-BANDMAXIMUM WITH RESPECTTO THE LDA VALENCE-BAND MAXIMUM ARE
GIVENIN BRACKETS IN THE ri5"COLUMN. UNLESSNOTEDOTHERWISE, ALLCALCULATIONS USE
PSEUDOPOTENTIALS (PSP) AND PLANEWAVES,A PLASMON-POLE MODEL (PPM), AND
DETERMINE SELF-ENERGY S m s IN FIRST-ORDER PERTURBATION THEORY (PT). CALCULATIONS
THAT Do NOT USE A PLASMON-POLE MODELTAKETHE FREQUENCY DEPENDENCE OF THE
SCREENED INTERACTION FULLY INTO ACCOUNT. THIS FREQUENCY DEPENDENCE LEADSTO A
VALENCE-BAND-WIDTH DECREASE OF ABOUT 0.3 eV BUT AFFECTS THE LOWESTDIRECTGAPSBY
ONLYA FEW TENSOF meV. QUASIPARTICLE ENERGIES DETERMINED WITH AN UPDATED GREEN
FUNCTION (GF)- WHOSEENERGIES ARE GIVENBY THE QUASIPARTICLE ENERGIES OBTAINED
IN THE FIRSTITERATION-AREGIVENIN PARENTHESES. IN THE CASEOF REF 102, WE GIVETHE
RESULTS OF STANDARDGWA AND GWT (IN BRACKETS)CALCULATIONS (SEE TEXT).
THE CALCULATION OF SHIRLEY et a1.353 IS NOT INCLUDED IN THIS TABLESINCETHE GWA
VALUESWITHOUTCOREPOLARIZATION AGREEWITH THE RESULTS OF HYBERTSEN et al.44 THIS
IS BECAUSE CORE POLARIZATION IS A SMALLEFFECTIN Si. GO STANDSFOR GAUSSIAN
ORBITALS; FLAPW FOR FULLPOTENTIAL LINEARAUGMENTEDPLANEWAVE.ALLENERGIES ARE
GIVENIN eV.

REF. I-1" r25u r15c r;, COMMENTS

Expt. -12.5(6)" 0.00 3.34-3.36' 4.15(5)'

-12.4(15)~ 3.05' 4.1'
-11.4' 4.21(21f
-11.24
A u l b ~ r ~ ~ - 11.96 2.56 3.18
-11.90 [ -0.651 3.25 3.86
(- 12.04) (3.30) (3.92)
Rohlfing'14 - 11.91 2.57 3.24 PSP/GO
- 11.95 [ -0.401" 3.36 3.89
HamadazS5 - 11.95 2.55 3.17 FLAPW
- 12.21 [ -0.311 3.30 4.19
H y b e r t ~ e n ' . ~ ~ - 11.93 2.57 (4.08) Update GF
- 11.84 [ -0.201" 3.27
(- 12.04) (3.35)
Godby6 2.51 3.56 2nd order PT
C0.071 3.30 4.27
Rieger"' -11.89 2.58 3.28 No PPM
-11.57 [ - 0.49'1 3.24 3.94
(- 11.58) [( - 0.53')] (3.32) (4.02) Update G F
Fleszar O2 - 11.93 2.55 3.26 No PPM
-11.57 [-0.631 3.23 3.96
[- 11.721 [ -0.061 C3.221 [3.99]
 QUASIPARTICLE CALCULATIONS IN SOLIDS 123

TABLE
21. Continued

XI" X4" x4,

Expt. -2.9'
-2.5(3)"
-3.3(2)'
Aulbur4' - 7.82 - 2.87 0.66 9.99
- 7.90 - 2.96 1.31 10.72
( - 7.99) ( - 2.99) (1.37) (10.79)
- 7.77 - 2.78 0.65 10.03
- 7.95 - 2.93 1.43 10.76
Ha~nada~'~ - 7.82 - 2.84 0.65
-8.11 -3.03 1.14
( -2.99) (1.44)
C3.531"
C4.271"
- 7.78 - 2.82 0.61 10.1 1
- 7.67 - 2.80 1.34 10.54
( - 7.68) ( - 2.81) (1.42) (10.63)
Fleszar O' - 7.77 - 2.83 0.65
- 7.57 - 2.83 1.35
[ - 7.681 [-2.881 c 1.341

Expt. - 9.3(4)" -6.8(2)" -1.2(2)e 2.06(3) 3.9(1)'

-6.4(4)' -1.5' 2.40 (15) li 4.15 (10) Ir
Aulb~r~~ - 9.62 - 7.00 - 1.21 1.46 3.36 7.55
-9.65 -7.13 - 1.25 2.13 4.13 8.23
( - 9.76) (-7.21) (- 1.26) (2.18) (4.19) (8.30)
- 9.58 - 6.94 - 1.17 1.47 3.32 7.77
-9.70 - 7.14 - 1.25 2.19 4.25 8.56
- 9.63 - 6.98 -1.19 1.43 3.35
- 9.92 - 7.31 - 1.26 2.15 4.08
( - 9.79) (-7.18) - 1.21 1.51 3.37
- 1.26 2.18 4.14
(- 1.27) (2.27) (4.24)
Godby6 - 1.22 1.53 3.37
- 1.19 2.30 4.11
Rieger"' -9.57 - 6.96 - 1.17 1.46 3.33 7.71
- 9.39 -6.86 - 1.17 2.14 4.05 8.29
( - 9.40) ( - 6.88) (- 1.17) (2.22) (4.14) (8.39)
Fleszar O' - 9.58 - 6.96 - 1.19 1.50 3.33
- 9.35 -6.78 - 1.20 2.18 4.06
[ - 9.471 [ -6.883 [- 1.221 C2.181 [4.05]

"Ref. 339; 'Ref 340; 'Ref. 341; dRef. 342; 'Ref. 343; IRef. 344; @Ref.345; *Ref, 346; 'Ref. 347;
'Ref. 348; 'Ref. 349; 'Ref. 311; '"Ref. 350, gap at X ; "estimated from Fig. 1 of Ref. 284; "estimated
from Fig. 1 of Ref. 44; PRef. 351; 4Ref. 352.
124 WILFRIED G. AULBUR, LARS JdNSSON, AND JOHN W. WILKINS

2. The accuracy of converged plane-wave calculations for the same

transitions as above is 50meV in LDA and 50meV in GWA and
comparable to chemical accuracy.
3. Plasmon-pole approximations do not affect the lowest direct band gaps
in Si by more than 50meV.
4. GWA improves significantly upon the LDA description of excited
states and gives agreement with experiment within the experimental
and computational uncertainties.

Conclusions 1,2, and 4 are state and material dependent. Take the F2, state
in Si, which is particularly difficult to converge, as an example. Here, the
deviations between all calculations are 0.39 eV in LDA and 0.41 eV in GWA.
The deviations for well-converged plane-wave calculations are 0.1 eV in
both LDA and GWA, well outside chemical accuracy. In addition, materials
that are less “benign” than Si- that is, in which, for example, core electrons
or relativistic effects are relevant -exhibit larger variations in LDA and
GWA energies, as will be shown below. Exceptions to conclusion 4 will be
discussed in Sections IV and V.
The plasmon-pole approximation is expected to break down for quasipar-
ticle energies of the order of the Si plasmon energy, wpl = 16.7 eV (see Table
3). Comparing plane-wave calculations, with the exception of Ref. 6, for the
lowest valence-band energies at r, X,and L shows that the plasmon-pole

339 L. Ley, S. P. Kowalcyzk, R. A. Pollak, and D. A. Shirley, Phys. Rev. Lett. 29, 1088 (1972),
as presented by J. R. Chelikowsky and M. L. Cohen, Phys. Rev. 8 10, 5095 (1974).
340 W. D. Grobman and D. E. Eastman, Phys. Rev. Lett. 29, 1508 (1972).
341 M. Welkowsky and R. Braunstein, Phys. Rev. 8 5 , 497 (1972).

J. E. Ortega and F. J. Himpsel, Phys. Rev. B47,2130 (1993).

343 W. E. Spicer and R. C. Eden, in Proceedings of the Ninth International Conference on the
Physics of Semiconductors, Moskau, 1968, ed. by S. M. Ryvkin, Nauka, Leningrad (1968), vol.
1, 61.
344 D. E. Aspnes and A. A. Studna, Sol. State Comm. 11, 1375 (1972).
345 Ref. 341, as presented by F. Szmulowicz, Phys. Rev. 823, 1652 (1981).
346 R. Hulthen and N. G. Nilsson, Sol. State Comm. 18, 1341 (1976).
347 A. L. Wachs, T. Miller, T. C. Hsieh, A. P. Shapiro, and T.-C. Chiang, Phys. Rev. B32,2326
(1985), as presented in Ref. 44.
348 F. J. Himpsel, P. Heimann, and D. E. Eastman, Phys. Rev. 824, 2003 (1981).
349 D. Straub, L. Ley, and F. J. Himpsel, Phys. Rev. Lett. 54, 142 (1985).
350 R. W. Godby, M. Schliiter, and L. J. Sham, Phys. Rev. 836, 6497 (1987).
3 5 1 D. H. Rich, T. Miller, G. E. Franklin, and T. C. Chiang, Phys. Rev. 839, 1438 (1989).
3 5 2 J. R. Chelikowsky,T. J. Wagener, J. H. Weaver, and A. Jin, Phys. Rev. B40,9644 (1989).
353 E.L. Shirley, X.Zhu, and S. G. Louie, Phys. Rev. Lett. 69, 2955 (1992).
354 Ref. 6 used 169 plane waves, which corresponds to an energy cut-off of about 5.5 Hartree
that is significantlylower than the 8.5-Hartree cut-off used in the other calculations.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 125

approximation leads to an artificial widening of the valence band width.

The widening amounts to about 0.25eV and shows very little variation
( <20 meV) between r, X, and L. Taking the full frequency dependence into
account improves agreement with experiment at the L point but leads to
disagreement with experiment at the r point if the experimental value of
12.5eV is used, as shown in Table 21. The accuracy of the commonly quoted
experimental value of 12.5 f 0.6 eV has been q u e s t i ~ n e d . ~Recent~.~~~
experiments find a valence band width at r of 11.4eV3" and ll.2eV,352in
good agreement with the calculations by Fleszar and Hanke102and Rieger
et al.' O 3
Fleszar and Hanke"' compared standard GWA calculations for Si with
GW'T calculations (see, for example, Section 11.14) that included LDA
vertex corrections in the dielectric matrix and the self-energy. Inclusion of
vertex corrections (see also Section 11.14, Eqs. (2.86) and (2.87)) (1) shifts
absolute energies by about 0.5 eV, (2) affects relative energies of valence and
conduction bands close to the Fermi level by less than 50 meV, and (3) leads
to larger shifts of up to 0.15 eV for the relative energies of low-lying valence
states.
Absolute energies depend on the choice of the exchange-correlation
potential,44 the degree of convergence,'02 and the choice of the self-energy
approximation (GWA versus GWT; Refs. 89 and 102), as discussed in
Section 11.14. As a result, absolute energies fluctuate from +0.07eV6 to
-0.65eV for the risvstate. Absolute energies agree to within 0.16eV
(- 0.49 eV to - 0.65 eV) for well-converged, plane-wave-based GWA calcu-
lations whose LDA exchange correlation is based on the Ceperly-Alder data
for the homogeneous electron gas.355 For these calculations, the main
self-energy correction is in the valence band and not in the conduction
band."'
In density functional theory the highest occupied valence energy is
given and, under the assumption that LDA is a good approxi-
mation to DFT, one would expect small GWA corrections to the highest
occupied LDA energy. Indeed, GWA c a l c ~ l a t i o n sbased~ * ~ ~on an LDA
calculation using RPA correlation356or GWT calculations consistent with
the Ceperley-Alder exchange correlation' O' shift the highest occupied state
of Si by only 0.1 to 0.2 eV. These calculations may be the most appropriate,
for example, for the determination of valence-band offsets.357Note that the
self-energy corrections to the highest occupied LDA state increase with an

"' D. M. Ceperley and B. I. Alder, Phys. Rev. Lett. 45, 566 (1980).
356 U. von Barth and L. Hedin, J. Phys. C 5, 1629 (1972).
357 The difference between relative energies in GWA and GWT for the lowest direct band gaps
is less than 0.02 eV (Ref. 102), as can be seen in Table 21.
 126 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

increasing energy gap and ionicity of the material and exceed, for example,
a negative correction of 1 eV in the case of LiCl even with the use of RPA
~orrelation.~~
Figure 31 compares the results of an LDA and GWA calculation with
photoemission and inverse photoemission experiments. The size of a typical
experimental resolution is 0.27 eV in energy and 0.1 8- in momentum.342
In numerical calculations the momentum is well defined, while the energy
uncertainty is estimated to be about kO.1 to 0.2eV from the above
discussions. The agreement between theory and experiment is good along

15

10

z- 5

zi
>
B
9
s
0
-m
e
>r
F
w
E -5

-10

-1 5
L A r A X
Wave vector

FIG. 31. Comparison of LDA and GWA band structures along L - r - X with photoemis-
sion and inverse photoemission experiments for Si. Agreement between theory and experiment
is within the experimental and theoretical uncertainties. Data are taken from Ref. 342 (full
diamonds), Ref. 347 (open circles), Ref. 348 (full triangle), Ref. 351 (open triangles), and Ref.
358 (open diamonds). The typical experimental resolution is 0.27 eV in energy and 0.1 k'in
momentum (compare Ref. 342). The theoretical error in energy is estimated to be about +0.1
to 0.2 eV (compare Table 21).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 127

the T L direction. The lowest conduction band shows the largest discrepancy
between theory and experiment and seems to lie slightly higher than the
calculation. Along the T X direction agreement between experiment and theory is
satisfactory. In part due to the large experimental momentum uncertainty,
theory and experiment still agree to within the respective uncertainties. However,
the lowest valence band as well as the first two conduction bands -the A, and
the A; bands -show less dispersion than the theoretical prediction.

b. Germanium and Gallium Arsenide

Ge and GaAs are more difficult theoretically due to the importance of the
extended core and relativistic effects. The core can influence the valence
energies in several ways: core relaxation, core-valence exchange, core-
valence correlation, and core polarization. As shown in Section 11.12, an
improved treatment of exchange and correlation effects due to core electrons
beyond LDA increases the direct band gaps of Ge and GaAs obtained in
standard GWA calculations by 0.32 eV and 0.40 eV, respectively. This leads
to good agreement with experiment (Ge (GaAs): 0.85 (1.42) eV from theory
versus 0.89 (1.52)eV from e ~ p e r i m e n t ~As
~ ) .shown in Table 22, scalar-
relativistic GWA calculations for Ge that treat core-valence exchange and
correlation on an LDA level deviate significantly from experiment (e.g.,
direct gap of 0.48 eV in contrast to 0.89 eV in experiment), and they agree
to within 0.1-0.3 eV with each other if an update of quasiparticle energies44
and insufficient convergence (Ref. 114; see Section III.15a) are accounted for.
An update of the energy spectrum of the Green function is bound to affect
the l-‘z5”- r;, gap significantly because the second iteration deals with a
semiconductor rather than a semimetal.
The results for GaAs are summarized in Table 23. Due to the neglect of
core-valence interactions the experimental band gap at r is underestimated
by about 0.4 eV in standard GWA calculations.353The quasiparticle ener-
gies vary by 0.3 to 1.5 eV between the different calculations due to different
approximations, as detailed in the caption of Table 23 and discussed in
Sections III.l5a, 111.16, and III.18a. In particular, deviations of 0.65 eV and
0.3 eV with the 40-Gaussian-orbital calculation of Rohlfing et u Z . ” ~ for the
lowest conduction-band state at X and L reduce to 0.14eV and 0.04eV
when 60 Gaussian orbitals are used, as shown in Table 15. Also, devia-
tions between pseudopotential results that include core-polarization and
core-relaxation effects43 and scalar-relativistic, all-electron LMTO calcula-
t i o n are ~ ~ to result from the neglect of combined corrections in the
~ ~likely
LMTO wave functions and from the atomic sphere approximation.

358 D. Straub, L. Ley, and F. J. Himpsel, Phys. Rev. B33, 2607 (1986).
128 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

TABLE22. Ge ENERGIES AT HIGH SYMMETRY POINTS;THE NOTATION IS THE SAME AS IN

TABLE21. BOTHGWA VALENCE- AND CONDUCTION-BAND ENERGIESVARY BY 0.1 TO 1.6eV
B m D ~ ~ E R ECALCULATIONS
NT BECAUSE OF A D~PFERENT TRE~TMENT OF COREVALENCE
POLARIZATIONAND OF CORERELAXATION, A USEOF UPDATED QUASIPARTICLE ENERGIESRATHER
THAN LDA ENERGIES IN THE GREENFUNCTION, AND DIFFERENTDEGREES OF CONVERGENCE (SEE
TABLE15). ALL CALCULATIONS USE PLASMON-POLE MODELS. THECHOSEN BASISSET IS PLANE
WAVES EXCEPTFOR THE GAUSSIAN ORBITALCALCULATION BY ROHLFING et THE
QUASIPARTICLE ENERGIES OF HYBWTSEN AND AND SHIRLEY, ZHU,AND L ~ u I USE
E~~
UPDATED GREENFUNCTIONS. THE DATAIN PARENTHESES INCLUDE CORERELAXATIONAND
CORE-POLARIZATION POTENTIALS,WHICH ARE NECESSARY TO DETER^ THE QUASIPARTICLE
GAPIN QUANTITATIVE AGREEMENTWITH EXPERIMENT. HYBERTSEN AND L o r n INCLUDE SPM-
ORBIT(SO) COUPLING TO FIRSTORDER. ALL ENERGIESARE IN eV.

REF. r1. G" r;, r15c COMMENTS

Expt." - 12.6 0.00 0.90 3.25(1)

- 12.9(2)b 0.8Y 3.006'
3.206'
Aulbur4' - 12.74 0.00 - 0.05 2.59
- 1259 0.00 0.48 3.10
Rohlfing114 - 12.19 0.00 0.01 2.53
- 1284 0.00 0.65 3.21
Hybert~en~~ - 1286 0.00 0.71 3.04 Update GF, SO 1st
- 0.30 3.26 order
Shirley43 - 0.26 Update GF, w/o (w)
053 core polarization
(0.85)

XI"
Expt.' - 9.3(2)b -3.15(20) 1.3(2)
-3.5(2)'
Aulbur4' - 8.88 - 3.02 0.70 9.50
-887 - 3.08 1.15 10.04
Rohlfing114 -8.91 - 3.02 1.03 9.54
- 9.06 - 3.16 1.74 10.19
Hybert~en~~ -9.13 - 3.22 1.23
Shirley43 0.55
1.28
(1-09)

Expt." -10.6(5) -7.7(2) -1.4(3) 0.744 4.3(2) 7.8(6)

4.2(1)' 7.9(1)'
Aulbur4' - 10.61 -7.51 - 1.37 0.12 3.77 7.01
- 10.60 -7.63 - 1.39 0.61 4.30 756
Rohlfing114 - 10.71 - 1.63 - 1.40 0.33 3.80 8.33
- 1082 - 7.81 - 1.47 0.98 457 9.20
Hybertsen" - 10.89 -7.82 - 1.61 0.75 4.33 7.61
- 1.43 4.43
Shirley4, -0.05
0.70
(0.73)

"Ref. 182 unless otherwise noted; 'Ref. 347;'Ref. 359;"Ref. 349.

 QUASIPARTICLE CALCULATIONS IN SOLIDS 129

TABLE23. GaAs ENERGIES AT HIGHSYMMETRY POINTS, SAMENOTATION AS IN TABLE21. THE

LDA AND GWA ENERGIES AT r, X,AND L FOR THE LOWEST CONDUCTION BAND VARY BETWEEN
THE LISTED CALCULATIONS BY 0.4 TO 0.6 eV IN LDA AND 0.3 TO 1.0 eV IN GWA BECAUSE OF (1)
INCLUSION OF CORERELAXATION AND POLARIZATION IN REFS.43 AND 290, (2) INSUFFICIENT
CONVERGENCE (REF.114; SEE TABLE15), (3) USE OF THE ATOMIC SPHERE APPROXIMATION AND
NEGLECT OF COMBINED AND (4) UPDATE OF GREENFUNCTIONS IN REFS. 43 AND
121 IN CONTRASTTO REFS. 42, 114,290, AND 360. THE EFFECTSOF THESE APPROXIMATIONS ARE
DISCUSSEDIN SECTIONS 111.15 AND 111.18 IN THE TEXT. ALL CALCULATIONS USE PLANE WAVES
UNLESSNOTEDOTHERWISE. FORAN EXPLANATION OF THE COMMENTS IN THE TABLE, SEE TABLES
21 AND 22. ALL ENERGIES ARE GIVENIN eV. SPIN-ORBIT SPLIITNGIS NEGLECTED IN THE
EXPERIMENTAL DATA.MOSTOF THE EXPERIMENTAL DATAARE AT ROOMTEMPERATURE.

REF. TI" r15. TIC COMMENTS

Expt." -13.21 0.00 1.52 4.61

A~lbur~~ - 12.62 0.00 0.47 3.80
- 12.46 0.00 1.16 4.47
Rohlfing' l4 - 12.69 0.00 0.51 3.73 PSP/GO
- 1269 0.00 132 4.60
ZhangIz' 0.00 0.38 SO 1st order;
0.00 1.29 update GF
-0.34
-0.34
G~dby~~' 0.00 0.56 3.70 SO 1st order;
0.00 1.47 4.52
-0.34
- 0.34

Shirley43 0.13 w/o (w) core polarization

1.02
(1.42)
A r y a s e t i a ~ a n ~ ~ ~ - 12.85 0.00 0.04 3.93 LMTO
- 1297 0.00 1.23 5.61

Expt." -10.86 -6.81 -2.91 1.90 2.47

Aulbur41 - 10.29 -6.78 -2.57 1.42 1.62 10.19
-10.12 -6.98 -2.68 2.00 2.24 10.93
Rohlfing'14 -10.37 -6.19 -2.56 1.80 1.85 10.33
- 10.27 -7.16 -2.71 2.65 2.72 11.20
Zhang ' -2.11 1.32
- 2.79 2.05
- 2.79
- 2.87

G~dby~~' -2.66 1.38 1.55

-2.13 2.08 230
- 2.13
-2.80
 130 WILFRIED G. AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

TABLE
23. Continued

Shirley43 1.21
2.07
(1.95)
AryasetiawanZgo - 10.37 -7.01 -2.72 1.29 1.59 10.34
-10.50 -6.88 -2.45 2.88 3.38 11.50

Expt." -11.35 -6.81 -1.41 ' 1.74 5.45b 8.6b

Aulbur4' -11.01 -6.56 -1.09 0.97 4.68 7.74
-10.84 -6.74 -1.12 1.62 5.38 8.40
Rohlfing114 -11.08 -6.59 -1.10 1.13 4.67 8.88
-11.02 -6.91 -1.17 1.92 5.65 9.92
Zhanglzl -1.16 0.88
- 1.19 1.69
- 1.37
- 1.40

GodbyJ6' - 1.07 1.04 4.57

- 1.11 1.82 5.41
- 1.28
- 1.32

Shirley43 0.70
1.55
(1.75)
Aryasetiawan'" -11.11 -6.84 -1.12 0.72 4.66 8.26
-11.27 -6.59 -1.01 2.14 6.51 9.51

"Ref. 361 unless noted otherwise; bRef. 342.

c. Silicon Carbide and Gallium Nitride

Numerical and experimental data on the quasiparticle band structure of Sic
and GaN are scarce. To the best of our knowledge only three GWA
calculations for S i c have been published so far: one based on Gaussian
orbitals"4 and two based on plane waves.42~130 Table 24 demonstrates that
agreement between the three calculations is unsatisfactory. In particular, the

359 D. E. Aspnes, Phys. Rev. B 12, 2297 (1975).

360 R. W. Godby, M. Schliiter, and L. J. Sham, Phys. Rev. B35,4170 (1987).
361 as in Ref. 182, vol. 22a.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 131

TABLE24. CUBICS i c ENERGIESAT HIGHSYMMETRY POINTS,SAMENOTATION AS IN TABLE 21.

AVERAGEDIFFERENCES BETWEEN THE DIFFERENTCALCULATIONS OF THE LDA AND GWA BAND
STRUCTURESAND THE QUASIPARTICLE SHIITSARE 0.19 eV, 0.37 eV, AND 0.29 eV, RESPECTIVELY,
AND SIGNIFICANTLY LARGERTHAN THE COMMONLY ACCEPTED UNCERTAINTY OF 0.1 eV IN
QUASIPARTICLEENERGIES. RESULTS OF PLANE-WAVE CALCULATIONS4'" 30 AND OF A GAUSSIAN-
ORBITAL ARE COMPARED. ALL ENERGIES ARE GIVEN IN ev. COMPARISON WITH
EXPERIMENT IS GIVENIN TABLE 25.

REF. rl" rl5" TIC r15c

Aulbur4' - 15.34 0.00 6.21 7.16

- 16.08 0.00 7.19 8.18
Rohlfing'14 - 15.44 0.00 6.41 7.16
- 16.44 0.00 7.35 8.35
Backes'30 - 15.07 0.00 6.51 7.24
- 16.13 0.00 7.81 8.66

XI" X3" X,"

Aulbur4' - 10.22 -7.82 -3.20 1.31 4.16 13.78

- 10.96 -8.44 -3.53 2.19 5.23 15.23
Rohlfing''4 -10.31 -7.89 -3.22 1.31 4.33 14.05
-11.24 -8.64 -3.62 2.34 5.59 15.78
Backes' 30 -10.12 -7.10 -3.06 1.22 4.14 13.93
-11.19 -8.38 -3.42 2.37 555 16.05

Aul bur4' -11.71 -8.56 -1.06 5.33 7.12 9.90

-12.46 -9.19 -1.21 6.30 8.25 11.32
Rohlfing'14 -11.80 -8.63 -1.06 5.46 7.20 10.31
-12.75 -9.42 -1.21 6.53 8.57 12.04
Backesl3' -11.57 -8.45 -0.98 5.46 7.12 10.14
-12.65 -9.15 -1.11 6.76 8.68 12.08

average deviation in the LDA band structure between all three approaches
is 0.19 eV (maximum 0.41 eV for the L,, state); the average deviation in the
quasiparticle energies is 0.36 eV (maximum 0.82 eV for the X,, state); and
the average deviation in the quasiparticle shifts is 0.29 eV (maximum 0.67 eV
for the X,, state). It is not understood what causes the significant differences
in the LDA and quasiparticle band structures of these three independent
calculations.362 All three, however, give a good account of experimentally

''' Determination of self-consistent energy eigenvalues at r, X, and L with 40 Hartree rather

than 25 Hartree changes the LDA energies of Ref. 42 by less than 0.01 eV for the lowest 10
states. Convergence tests indicate that the corresponding quasiparticle energies are converged
on the level of 0.1 eV.
 132 WILFRIED G. AULBUR, LARS JONSSON, AND J O H N W. WILKINS

accessible interband transitions and energies, as shown in Table 25. Average

deviations from experiment excluding the L,, state range from 0.22 eV4' to
0.24 eV (Ref. 114, 0.20 eV if 60 Gaussians are used) to 0.29 eV,130
All three calculations place the L,, state between 6.30 and 6.76eV, in
marked contrast to experiment (4.2 eV). The difference between theory and
experiment is well beyond theoretical and experimental uncertainties. The
agreement of the three GWA calculations for the L,, state strongly suggests
that the experimental 4.2-eV transition cannot be interpreted as an indirect
transition between quasiparticle-like r15, and L,, states.l14
GaN is a very important material for optical devices such as blue-light-
emitting diodes and blue lasers.365However, two obstacles impede theoreti-
cal p r o g r e ~ s : ' * The
~ * ~Ga
~ ~ 3d electrons interact strongly with the N 2s
electrons, and GaN band structures are sensitive to the choice of the N
pseudopotential. In GaN the 3d electrons in LDA lie at the bottom of the

W. R. L. Lambrecht, B. Segall, M.Yoganathan, W. Suttrop, R. P. Devaty, W. J. Choyke,

J. A. Edmond, J. A. Powell, and M.Alouani, Phys. Rev. BSO, 10722 (1994).
364 R. G. Humphreys, D. Bimberg, and W. J. Choyke, Sol. State Comm. 39, 163 (1981).

TABLE25. COMPARISON OF CUBICS i c QUASIPARTICLE ENERGIESFOR PLANE-WAVE CALCULA-

GAUSSIAN
T I O N S ~ ' . ' ~AND
~ A ORBITAL APPROACH"^ WITH EXPERIMENTAL ENERGIESAT HIGH-
SYMMETRY POINTS.AGREEMENT WITH EXPERIMENT IS GOODFOR ALL THREE GWA CALCULA-
TIONS.AVERAGE DEVIATIONS ARE 0.22 eV,4* 0.24 eV (REF. 11% 0.20 eV IF 60 GAUSSIANSARE
USED), AND 0.29eV.13' INCREASING THE CONVERGENCE OF THE GAUSSIAN-ORBITAL-BASED
CALCULATION BY KEEPING 60 (REF.295; NUMBERSIN PARENTHESES) RATHERTHAN 40 (REF.114)
GAUSSIANSMom THE QUASIPARTICLE ENERGIES CONSISTENTLY TOWARDSTHE CONVERGED
RESULTS OF THE PLANE-WAVE CALCULATION OF REF.42. E X P ~ M E N TVALW
AL ARE TAKEN
FROM REF. 361 UNLESSO~OWWISE NOTED.ALL ENERGIES ARE GIVENIN eV.
~~ ~~

EXPT.' ALJLBIJR~~ RoHLmc114 BACKES130

x3, - Xlc 3.10 3.04 3.25 (3.09) 3.18

L3" - XI, 3.55 3.40 3.55 (3.49) 3.48
X5" - X I , 6.0 5.72 5.96 (5.83) 5.79
X5" - x 3 c 8.3 8.76 9.21 (8.92) 8.97
L3" - L3c 9.7 9.46 9.78 (9.70) 9.79
L3" - LIC 7Sb 7.51 7.74 (7.64) 7.87
r1. 7.4b 7.19 7.35 (7.29) 7.81
rl5C 7.75 8.18 8.35 (8.42) 8.66
X5" - 3.4b -3.53 - 3.62 (- 3.52) - 3.42
XI, 2.39, 2.417' 2.19 2.34 (2.31) 2.37
LIC 4.2 6.30 6.53 (6.46) 6.76
K, 4.00

"Ref. 361 unless otherwise noted; bRef. 363; 'Ref. 364.

 QUASIPARTICLE CALCULATIONS IN SOLIDS 133

valence band and hybridize with the s bands. In LDA, d electrons are
underbound in comparison to experiment, leading to a strong, unphysical
s-d hybridization. A proper treatment of the Ga 3d electrons requires a
self-consistent determination of the wave functions and the inclusion of
exchange with the 3s and 3p electrons, as explained in Section III.15a. A
computationally cheaper, surprisingly accurate approach to the determina-
tion of the quasiparticle band gap is the treatment of the 3d electrons via
nonlinear core corrections,205 which has been adopted in the plane-wave
calculations of Refs. 42,47, and 367. Variations in the quasiparticle energies
of the lowest-lying conduction band for the three different calculations
amount to about 0.2 to 0.7 eV and are significantly larger than the generally
quoted GWA uncertainty of f O . l eV. In part, these variations may result
from the choice of different pseudopotentials, as specified in Table 26, and
the use of a model dielectric matrix in Ref. 47 compared to RPA dielectric
matrices in Refs. 42 and 367.

IV. Semiconductors and Insulators

This section gives a comprehensive overview of applications of quasiparticle

calculations to semiconductors and insulators. These applications are
grouped according to bulk materials, superlattices and interfaces, surfaces,
defects, pressure dependence, and excitons. The final section discusses
quasiparticle calculations for atoms and molecules.

19. BULK

The technical details of GWA calculations in bulk semiconductors and

insulators and results for five prototypical semiconductors were extensively
discussed in Section 111. Figure 3 shows an overview of-to the best of our
knowledge -all GWA calculations for semiconductors and insulators pub-
lished so far. This section is therefore limited to three examples of the
relevance of GWA to describe the electronic structure of materials: (1)
band-gap narrowing in n-type Si, (2) transition-metal oxides, and (3) solid
C,, and related systems.

365 S. Nakamura and G . Fasol, The Blue Laser Diode: GaN Based Light Emitters and Lasers,
Springer-Verlag,Berlin and New York (1997).
366 A. F. Wright and J. S. Nelson, Phys. Rev. B50, 2159 (1994).
M. Palummo, L. Reining, R. W. Godby, C. M. Bertoni, and N. Bornsen, Europhys. Lett.
26, 607 (1994).
134 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

TABLE26. GaN ENERGIES AT HIGH-SYMMETRY POINTS, SAME NOTATION AS IN TABLE 21. THE
DIFFERENTCALCULATIONS OF THE LDA AND GWA ENERGIES FOR BOTHTHE VALENCE-BAND
MAXIMUM AND THE CONDUCTION-BAND MINIMUM DIFFERBY 0.1 TO 0.4 eV AND BY 0.2 TO 0.7 eV,
RESPECTIVELY. TAKINGALL STATESINTO ACCOUNT LEADSTO A VARIATION OF 0.1 TO 0.8eV
FOR LDA AND 0.1 TO 1.3eV FOR GWA. ALL CALCULATIONS ARE PSEUDOPOTENTIAL-PLANE-
WAVE-BASED AND USEA FIRST-PRINCIPLES RPA DIELECTRIC MATRIX, WITH THE EXCEPTION OF
RUBIOet aL4' WHO USEDA MODELDIELECTRIC MATRIX.THE MODELDIELECTRIC MATRIXIS
USUALLY ACCURATE TO WITHIN0.1 TO 0.4 eV AND CANNOT ACCOUNT FOR ALL THE DIFFERENCES
IN THE DATA.FURTHER DEVIATIONS MAY RESULTFROM THE DIFFERENT CHOICE OF THE
NITRCGENPSEUDOPOTENTIAL. BHS STANDS FOR BACHELET-HAMANN-SCHLmR324 PSEUDO-
POTENTIALS AS IMPLEMENTED BY STUMPF, GONZE AND sCHEFFLER;370 HSC STANDS FOR
HAMANN-SCHL~ER-CHIANG PSEUDOPOTENTIALS;3 AND TM STANDS FOR TROULLIER-MARTINS
PSEUDOPOTENTIALS.32s'372 ALL ENERGIESARE IN ev.

REF. rl" rls" rlc r* 5e COMMENTS

Expt. 0.0 3.2"

3.3*
Aulbur4' - 15.5 0.0 2.0 10.5 E ~ TM~ ~ ,
- 16.7 0.0 3.1 12.0
Palurn~no~~~ 0.0 1.9 10.2 E~'", BHS
0.0 2.8 11.3
Rubio4' - 16.3 0.0 2.1 10.6 , update GF, HSC(Ga),
E'""''
- 17.8 0.0 3.1 12.2 TM(N)

Aulbur4' -12.4 -6.1 -2.4 3.3 6.7 11.9 14.2

-13.5 -6.8 -2.7 4.4 8.1 14.0 16.1
Pal~rnrno~~~ -6.2 -2.6 3.2 6.6
-6.8 -2.9 4.0 7.7
Rubio4' -13.0 -6.5 -2.8 3.2 6.9 12.2 14.6
-14.8 -6.9 -3.0 4.7 8.4 14.5 16.7

Aulbur4' -13.2 -6.8 -0.8 4.8 8.9 10.5

-14.3 -7.6 -0.9 6.1 10.8 11.9
Palu~nrno~~~ -7.0 -0.9 4.7 8.9 10.3
-7.5 -1.0 5.7 10.4 11.3
R ~ b i o ~ ~ - 13.8 -7.4 -1.0 5.0 9.1 10.6
-15.5 -7.8 -1.1 6.2 11.2 12.3

"Ref. 368; "Ref. 369.

 QUASIPARTICLE CALCULATIONS IN SOLIDS 135

a. Band-Gap Narrowing in Si
Two competing processes influence the band-gap modification of Si upon
n-type doping:373(1) Electrons fill up conduction bands to some new Fermi
level and effectively increase the photoexcitation gap of valence electrons. (2)
The conduction electrons form a low-density, “metallic” gas of carriers and
increase screening of the electron-electron interaction, which reduces the
band gap. A quantitative understanding of these effects on the band gap of
doped semiconductors is important since it affects the performance of
semiconductor devices.
Quasiparticle calculation^^^^*^ 75 have so far considered band-gap nar-
rowing in doped Si and find that (1) LDA is inadequate for the description
of band-gap narrowing, (2) the modification of screening due to excess
carriers dominates band-gap narrowing in the GWA calculation, and (3) the
energy dependence of the intrinsic dielectric matrix -omitted in all model
calculations (for a review of model calculations see Ref. 375) -modifies
band-gap narrowing by up to a factor of two for common dopant concen-
trations in Si. A direct comparison of GWA ~ a l c u l a t i o n s with ~~~*~~~
experiment376*377-378 is inappropriate since the self-energy calculations do
not consider electron-donor scattering.379With AVH and AVxc as changes in
the electrostatic and LDA exchange-correlation potentials upon introduc-
tion of additional electrons in the system, the LDA contribution to band-
gap narrowing can be calculated in first-order perturbation theory as

AEnk =
s @:k(r)[AVH(r) + AVxc(r)]@nk(r)dr~
3 6 8 T. Lei, T. D. Moustakas, R. J. Graham, Y. He, and S. J. Berkowitz, J. Appl. fhys. 11,4933
(4.1)

(1992); T. Lei, M. Fanciulli, R. J. Molnar, T. D. Moustakas, R. J. Graham, and J. Scanlon,

Appl. fhys. Lett 59, 944 (1991); C. R. Eddy, T. M. Moustakas, and J. Scanlon, J. Appl. fhys.
73, 448 (1993).
369 M. J. Paisley, Z. Sitar, J. B. Posthill, and R. F. Davis, J . Vac. Sci. Echnol. A 7 , 701 (1989);
Z . Sitar, M. J. Paisley, J. Ruan, J. W. Choyke, and R. F. Davis, J. Mat. Sci. Leu. 11, 261 (1992).
3 7 0 R. Stumpf, X. Gonze, and M. Schemer, Research Report of the Fritz-Haber Institute,
Berlin, Germany, April 1990.
3 7 1 D. R. Hamann, M. Schliiter, and C. Chiang, Phys. Rev. Leu. 43, 1494 (1979).
3 7 2 N. Troullier and J. L. Martins, Sol. State Comm. 7 4 , 613 (1990).
3 7 3 Ref. 20, 210.
3 7 4 A. Oschlies, R. W. Godby, and R. J. Needs, fhys. Rev. B45, 13741 (1992).
375 A. Oschlies, R. W. Godby, and R. J. Needs, fhys. Rev. B51, 1527 (1995).
3 7 6 P. E. Schmid, fhys. Rev. B 2 3 , 5531 (1981).
3’7 D. D. Tang, I E E E Trans. Electron. Deu. 27, 563 (1980).
3 7 8 J. Wagner, Phys. Rev. B 3 2 , 1323 (1985).
3 7 9 Charge neutrality is achieved by considering a homogeneous, positive background charge
(Refs. 374 and 375).
136 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

which even for as high electron concentrations as n = 4 x lo2’ cm-3 affects

the band gap by less than 5 meV.375
In a GWA calculation contributions to band-gap narrowing result from
a change in the screening of the electron-electron interaction due to the
presence of additional charge carriers, and a change in the pole structure of
the Green function in the doped case. In first-order perturbation theory and
neglecting the unimportant Hartree contribution, the quasiparticle band-
gap change equals

where (* denotes convolution in frequency space)

AX(r, E ) = Gdoped * Wdoped - Gintrinsic * wintrinsic

- Gdoped * A W + AG * w n t r i n s i c . (4.3)

The first term describes the change in the screening of the electron-electron
interaction and turns out to be the dominant contribution to band-gap
narrowing, equaling about -0.1 eV at a carrier concentration of n =
lo2’ cm-3 and about -0.25 eV at n = lo2’ ~ m - The ~ . effect of screening
becomes stronger with increasing carrier concentrations. The second term
stems from the changed pole structure of the Green function and also
increases with increasing concentration. However, its absolute contribution
to the band-gap narrowing is negligible and amounts to less than 10meV
for concentrations smaller than cm-j.
First-principles GWA calculations identify the neglect of frequency de-
pendence of the intrinsic dielectric matrix as an unphysical assumption
common to many model calculations of carrier-induced band-gap narrow-
ing.37 5 The frequency dependence of the intrinsic dielectric matrix increases
band-gap narrowing by about 30% of the dominant GdoPedAW contribution
at n = lo2’ cmV3and by about a factor of two at n = 1021~ m - and ~ ,
should not be neglected. Other effects, such as local fields in the intrinsic
dielectric matrix or the description of the conduction-band states either as
LDA states or as plane waves, are unimportant since the main contributions
to the modification of screening arise from the long-wavelength component.

b. Transition-Metal Oxides
The large discrepancy between local spin density (LSDA) functional calcu-
lations and experiment indicates that strongly correlated transition-metal
 QUASIPARTICLE CALCULATIONS IN SOLIDS 137

oxides pose a challenge to perturbation theory calculations such as the

GWA. Transition-metal oxides such as NiO are antiferromagnetic insula-
tors whose 3d onsite repulsion is about 8 eV and thus on the same scale as
the 3d band width, which is indicative of strong correlations between the 3d
electrons. The band gap in NiO is a charge transfer gap; that is, if a hole
with a small binding energy is created on the Ni site, it is filled with a high
probability by an oxygen 2 p electron. Experimentally, the hole in the highest
valence band has a strong oxygen 2 p character, in contrast to local spin
density calculations that show a predominant Ni 3d character. In addition,
the LSDA severely underestimates the band gap and magnetic moment in
transition-metal oxides such as NiO (Eizp = 0.2 eV versus Eiig = 4.0 eV;
pLSDAM lpB versus pexp= 1 . 7 - 1 . 9 ~ ~since
) the LSDA is unable to properly
describe localized states in these materials.
Accurate treatment of the transition-metal d and oxygen 2 p states is
difficult and has limited the number of transition-metal-oxide quasiparticle
calculations. Aryasetiawan and Gunnarssod5 reported an LMTO study of
NiO, and Massidda et a1.54*91*3a0 used a model-GWA approach3a1 to
determine the band structure of MnO, NiO, and CaCuO,. The large
discrepancy between the LSDA and the experimental gap, for instance,
in the case of NiO suggests that the initial LSDA system and, in par-
ticular, the LSDA wave functions are not good approximations to the
quasiparticle system and wave functions. This suspicion is confirmed in
actual calculation^.^^^^^ Self-consistency is generally used to construct a
better basis set than the LSDA basis set. Aryasetiawan and G u n n a r ~ s o n ~ ~
used an approximate self-consistency scheme implemented via a modifica-
tion of the LSDA one-particle Hamiltonian, whereas Massidda et
expanded the quasiparticle wave functions in a basis of occupied and
unoccupied LSDA wave functions and determined the expansion coefficients
self-consistently.
The basic physics behind the model-GWA calculation of Massidda et al.
is the separation of the screened interaction W into a short-ranged part
wIEG
that describes the screening of an inhomogeneous electron gas and a
long-range part 6 W that accounts for the incomplete screening in insula-
t o r ~ The . ~ ~short-ranged
~ contribution WIEGis approximated by the
local, energy-independent Kohn-Sham exchange-correlation p~tential.~’
The long-ranged screened interaction 6 W must decay as l/lr - r’l for large
Ir - r‘l and accounts for the increase of, for example, the NiO LSDA gap of
0.3 eV by 3.4 & 0.4 eV.54

380 S. Massidda, A. Continenza, M. Posternak, and A. Baldereschi, Physica B237-238, 324

(1997).
381 F. Gygi and A. Baldereschi, Phys. Rev. Lett. 62, 2160 (1989).
138 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Table 27 summarizes GWA results for transition-metal oxides and shows

that fundamental energy gaps are in fair agreement with experiment and
magnetic moments are in good agreement with experiment. Both quantities
show significant improvement over LSDA and their accuracy is comparable
to or better than LDA + U or self-interaction-corrected calculations. The
inability of RPA-based GWA to reproduce the satellites in photoemission
spectra is attributed to the lack of short-ranged correlations in GWA and is
reminiscent of the failure to reproduce the 6-eV satellite in Ni (see Sections
11.5 and V.26~and Ref. 68). While self-consistency in GWA calculations
increases the oxygen 2 p character of the highest valence band,54 that
increase is not enough to eliminate the doininance of the Ni 3d state and to
achieve agreement with e ~ p e r i m e n t . ~ ~

c. C,, and Related Systems

Quasiparticle calculations describe the electronic structure of solid c60
well (see Ref. 394 for an overview of experimental and theoretical results for
the C,, band structure). Estimates of on-molecule and nearest-neighbor-

TABLE27. ENERGY GAPS(eV) AND MAGNETIC MOMENTS (pB) FOR TRANSITION-METAL OXIDES
OBTAINED IN ab-initio GWA (ARYASETIAWAN AND G U N N A R ~ ~ MODEL
~ N ~ ~ GWA
), (MASSIDDA
AND cOLLABORATORS50'54),A N D LOCALSPIN DENSITY (LSDA) FUNCTIONAL CALCULATIONS IN
COMPARISON TO EXPERIMENT, LDA + U CALCULATIONS, AND SELF-INTERACTION-CORRECTED
(SIC) CALCULATIONS. GWA AGREEMENT WITH EXPERIMENT IS BETTER THAN OR AT LEASTAS
GCOD AS IN LDA+U OR SIC. BECAUSE OF THE LARGEGWA CORRECTIONS TO LSDA r~ THE
TRANSITION METALS, IN THE GWA WAVEFUNCTIONS
SELF-CONSISTENCY IS REQUIREDTO GET
GOODAGREEMENTWITH EXPERIMENT. VALUES IN PARENTHESES FOR THE MAGNETIC MOMENTS
INCLUDE THE ORBITALCONTRIBUTION.(ADAFTED FROM REF. 54.)

ARYASETIAWAN MASSIDDA

COMPOUND LSDA GWA LSDA GWA LDA+U" SICb EXPT.

ENERGY
GAPS(eV)
MnO 1.o 4.2 3.5 3.98 3.8-4.2"'
NiO 0.2 5.5 0.3 3.7 3.1 2.54 4.3'
4.0r
CaCuO, 0.0 1.4 2.1 1 9
MAGNETICMOMENTS(pB)
MnO 4.29 4.52 4.61 4.49 4.58h
(4.52) (4.49)
NiO 1.o 1.6 1.12 1.56 1.59 1.53 1.77'
(1.83) (1.80) 1.64'
1.90h
CaCuO, 0.42 0.66 0.51'

"Ref. 382; bRefs. 383 and 384; 'Ref. 385; dRef. 386; 'Ref. 387; IRef. 388; @Ref.389; hRef. 390;
'Ref. 391; 'Ref. 392; 'Ref. 393.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 139

molecule Hubbard-U parameters of 1.0 and 0.5 eV, respectively, are of the
same order as the measured band width of about 0.5eV of the highest
occupied (Hlu) and two lowest unoccupied (T,, and 7J multiband com-
plexes. In spite of these intermediate to strong electronic correlations, GWA
calculations for face-centered cubic C,, describe the fundamental energy
gaps quantitatively we11.338*394,395 Self-energy corrections double the LDA
fundamental gap of 1.04eV to a quasiparticle gap of 2.15 eV that compares
well with a gap of 1.85 & 0.1 eV obtained in microwave conductivity
experiments396 or of 2.3-2.7 eV deduced from direct and indirect photo-
emission experiment^.^^'^^^^.^^^ The H,-T,, peak-to-peak distance in the
experimental density of states equals 3.5-3.7 eV3979398,399 and is reproduced
reasonably well by a quasiparticle value of 3.0eV, which corrects an LDA
result of 1.6- 1.7 eV. Quasiparticle calculations for perfectly ordered, crystal-
line C , , lead to about a 30% increase in the LDA band width, as shown in
Fig. 32, resulting in GWA band widths of 0.9 eV, 0.7 eV, and 0.8 eV for H , ,
TI,, and Tlg, respectively. Since LDA overestimates the experimental band
width, further physics is needed to explain the experimental band-width
narrowing within a quasiparticle framework.
Four causes for a lack of dispersion in angle-resolved photoemission
spectra of the H , , T,,, and Tlg bands338*394 are (1) the multiband nature of
the system, (2) orientational disorder, (3) integration over reciprocal lattice

V. I. Anisimov, J. Zaanen, and 0. K. Andersen, Phys. Rev. B44, 943 (1991).

383 A. Svane and 0. Gunnarsson, Phys. Rev. Lett. 65, 1148 (1990).
3 8 4 Z. Szotek, W. M. Temmerman, and H. Winter, Phys. Rev. B47, 4029 (1993).

I. A. Drabkin, L. T. Emel'yanova, R. N. Iskenderov, and Y.M. Ksendzov, Pis. Tverd Tela

(Leningrad) 10, 3082 (1968) [Sov. Phys. Sol. State 10, 2428 (1969)l.
386 J. van Elp, J. L. Wieland, H. Eskes, P. Kuiper, G. A. Sawatzky, F. M. F. de Groot, and
T. S. Turner, Phys. Rev. B44, 6090 (1991).
387 A. Fujimori and F. Minami, Phys. Rev. B30, 957 (1984).
3 8 8 S. Hufner, J. Osterwalder, T. Riesterer, and F. Hulliger, Sol. State Comm. 52, 793 (1984).
389 Y. Tokura, S. Koshihara, T. Arima, H. Takagi, S. Ishibashi, T. Ido, and S. Uchida, Phys.
Rev. B41, 11657 (1990).
390 A. K. Cheetham and D. A. 0. Hope, Phys. Rev. B27, 6964 (1983), and references therein.
3 9 1 B. E. F. Fender, A. J. Jacobson, and F. A. Wegwood, J . Chem. Phys. 48, 990 (1968).

392 H. A. Alperin, J. Phys. SOC.Jpn. Suppl. B 17, 12 (1962).

393 D. Vaknin, E. Caignol, P. K. Davies, J. E. Fischer, D. C. Johnston, and D. P. Goshorn,
Phys. Rev. B39, 9122 (1989).
394 S. G. Louie and E. L. Shirley, J . Phys. Chem. Solids 54, 1767 (1993).
395 Undoped C,, is a band insulator since, for example, direct photoemission creates only one
hole in an otherwise filled band. This hole can travel freely without having to pay the large
Coulomb interaction of Li z 0.5 to 1.OeV (0.Gunnarsson, private communication).
396 T. Rabenau, A. Simon, R. K. Kremer, and E. Sohmen, 2. Phys. BW,69 (1993).
397 R. W. Lof, M. A. van Veenendaal, B. Koopmans, H. T. Jonkman, and G. A. Sawatzky,
Phys. Rev. Lett. 68, 3924 (1992).
140 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

4 j -

0
L r x w
FIG.32. Band structure of the face-centered cubic Fm3 structure of solid C,, as obtained in
LDA (a) and GWA (b). GWA leads to a band gap in good agreement with experiment and to
a 30% increase in band width of the T,g,TI, and If, bands compared to LDA. (From Ref. 338.)

vectors perpendicular to the sample surface, and (4) finite experimental

resolution. calculation^^^^*^^^ of the photoemission spectra that are based
on a Slater-Koster parametrization of the quasiparticle energies and that
include orientational disorder of the C,, molecules agree with experiment.
These calculations prove that a lack of dispersion of the H,,TI,, and Tlg
band complexes of disordered molecular C,, solid in experiment cannot be
interpreted as a sign of strong electron-electron correlation^.^^^*^^^
Further Applications. Band-width renormalization in A3C,,
(A = K, Rb) due to the T,, plasmon was examined by Gunnarsson within
the framework of the GWA.400 Quasiparticle calculations on periodically
repeated BN sheets with interlayer distances varying from 5.5 8, to 13.5 8,
can be used to deduce the band structure of BN nanotubes via zone folding,
as shown by Blase et al.401*402The layered structures have a calculated
indirect band gap of about 5.5 eV between the top of the valence band at K
and the bottom of the conduction band at r, which is relatively independent
of the interlayer distance.

BARRIERS
20. SUPERLATTICES-INTERFACES-SCHOTTKY

a. Superlattices
Simple superlattices consist of periodically repeated units of n layers of a

398 T. Takahashi, S. Suzuki, T. Morikawa, H.Katayama-Yoshida, S. Hasegawa, H.Inokuchi,

K. Seki, K. Kikuchi, S. Suzuki, K. Ikemoto, and Y. Achiba, Phys. Rev. Lett. 68, 1232 (1992).
399 J. H. Weaver, P. J. Benning, F. Stepniak, and D. M. Poirier, J. Phys. Chem. Solih 53,1707
(1992).
400 0. Gunnarsson, J . Phys. Cond. Mat. 9, 5635 (1997).

401 X. Blase, A. Rubio, S. G. Louie, and M. L. Cohen, Europhys. Lett. 28, 335 (1994).
402 X. Blase, A. Rubio, S. G. Louie, and M. L. Cohen, Phys. Rev. B51,6868 (1995).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 141

material A stacked onto m layers of a material B and are designated by

A,,B,. The superstructure of these artificial compound semiconductors gives
rise to a variety of new physical phenomena such as confinement, built-in
strain, Brillouin zone folding, and two-dimensional electron gas behav-
ior.403,404*405Variation of superlattice characteristics permits fabrication of
semiconductor devices with custom-tailored electronic properties. Molecu-
lar beam epitaxy (MBE) allows the growth of superlattices with monolayer
precision.404 Quasiparticle calculations for superlattices A,, B, are identical
to bulk calculations with the sole exception that the unit cell contains n
layers of material A and m layers of material B and is therefore a factor
(n + m)/2 larger than the unit cell of the corresponding binary semiconduc-
tor AB.
The first application of GWA for a semiconductor superlattice was for
Si,Ge,,4°6*407 a superlattice based on the indirect-gap semiconductors Si
and Ge, since this structure could potentially be a pseudo-direct-gap material
suitable for optoelectronic applications and is compatible with Si-based chip
technology. However, for Si substrates GWA finds two indirect band edges at
0.85 and 0.95 eV in comparison to values of 0.78 and 0.90 eV obtained in
photocurrent experiments.407Direct, zone-folded transitions are predicted at
1.24, 1.34, 1.76, and 1.86eV and explain features observed in electro-
reflectance measurements at 1.1-1.25 and 1.8 eV.,07 For a Ge substrate,
Si,Ge, is predicted to be approximately a direct-gap semiconductor.
Confinement of states in either the Si or the Ge region of the superlattice
should not have a significant quantitative effect on the expectation value of
the self-energy operator since the self-energy effects in Si and Ge are very
similar, as shown in Section 111.18. Figure 33 shows the qualitative agree-
ment between quasiparticle shifts for bulk Si, tetragonally strained Ge, and
Si,Ge, on a silicon substrate. This agreement occurs after all quasiparticle
energies have been aligned at the top of the valence band.
Further Applications. Band-gap variation and variation of direct-indirect
band-gap transitions due to a change in superlattice period or in concentra-
tion in the case of ordered alloys have been studied in GaN/AlN,,O*
GaAs, -,N,, and AlAs, -,N, systems.49At the interface of (GaAs),(AlAs),,

403 G. Bastard, Wave Mechanics Applied ro Semiconducfor Hererostrucfures, Halsted Press,

New York (1988).

404 J. H. Davies, The Physics of Low-Dimensional Semiconductors, Cambridge University

Press, Cambridge (1998).

405 T. P. Pearsall, “Strained-Layer Superlattices: Physics,” Semiconductors and Semimetals, vol.
32, eds. R. K. Williardson and A. C. Beer, Academic Press, New York (1990).
406 M. S. Hybertsen and M. Schliiter, Phys. Rev. 836, 9683 (1987).

407 M. S. Hybertsen, M. Schliiter, R. People, S. A. Jackson, D. V. Lang, T. P. Pearsall, J. C.

Bean, J. M. Vandenberg, and J. Bevk, Phys. Rev. B 37, 10195 (1988).

408 A. Rubio, J. L. Corkill, and M. L. Cohen, Phys. Rev. 849, 1952 (1994).
 142 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

0
0 si (CUBIC)
0 Ge(TETR1
Si4Ge4 (001)
0.5

I
Y

4
P

-0.5
-4
I I
-2 0
I
:
I
2 4
I
E''(eV)
FIG. 33. The calculated self-energy correction EQP - E~~~ is plotted against the quasiparticle
energy EQP for a Si,Ge, superlattice, cubic Si, and tetragonally distorted Ge. No significant
quantitative difference between the self-energy corrections in, the superlattice and the bulk
materials is found since self-energy effects in Si and Ge are very close. All values are aligned
with respect to the valence-band edge. (From Ref. 406.)

(001) (n = 1,2) super lattice^,^^^*^^^*^^^ electrons in excited states whose

density is located on cation sites accumulate in the GaAs rather than the
AlAs region of the superlattice, because of the relative repulsive character of
the A1 versus the Ga pseudopotential. This local bonding effect causes the
GWA direct gap in 1 x 1 and 2 x 2 superlattices to be lower in energy than
the pseudodirect gap, in contrast to effective-mass model calculations.
Effective-mass calculations miss the local bonding effect and consequently
show an inverted ordering of bands.41'
b. Interfaces
An interface between two dissimilar semi-infinite semiconductors A and B
is characterized by the difference in the valence-band energy between the
two bulk regions, the valence-band offset AEu, and the corresponding
conduction-band offset AEc,413 as shown in Fig. 34. Interfaces play an
essential role in heterojunction devices, and the band offsets AEu and AEc
determine their transport proper tie^.^'^ The qualitative effects of interface

,09S. B. Zhang, M. S. Hybertsen, M. L. Cohen, S. G. Louie, and D. Tomanek, Phys. Rev. Lett.
63,1495 (1989).
S.B. Zhang, M. L. Cohen, S. G. Louie, D. Tomanek, and M. S. Hybertsen, Phys. Rev. B41,
10058 (1990).
S. B. Zhang, M. L. Cohen, and S. G. Louie, Phys. Rev. B43,9951 (1991).
4'2 X. Zhu, S. B. Zhang, S. G. Louie, and M. L. Cohen, Phys. Rev. Lett. 63,2112 (1989).
For a review, see, for instance, M. S. Hybertsen, Mat. Sci. Eng. B 14,254 (1992).
See, for instance, F. Capasso and G. Margaritondo (eds.), Heterojunction Band Discontinu-
ities: Physics and Applications, North Holland, New York (1987).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 143

FIG. 34. Schematic representation of the volume-averaged dipole density n(z) (dashed line),
the Hartree potential V ( z ) (solid line), and the interface conduction-band offset, AEc, and
valence-band offset, AE,,, of two semiconductors A and B. In the semiconductor bulk regions,
V ( z ) equals the average Hartree potential ( which serves as an energy reference for the
bulk valence-band maximum ( E u , A ( Band J bulk conduction-band minimum (Ec.,qnJ of each
semiconductor. The dipole potential Vdipolc equals the difference between the average Hartree
potentials. Egap,AfBl is the energy gap of the bulk semiconductors.

orientation, defects, polarity, and strain on band offsets can be studied using
density functional theory. However, a more accurate prediction of valence-
and conduction-band offsets requires a quasiparticle approach.413 Since the
sum of the valence- and conduction-band offsets equals the known band-gap
discontinuity between the semiconductors A and B, AE,,, = AE” AE,, +
self-energy corrections of only the valence-band offset are discussed in this
subsection.
Within a few monolayers of the sharp interface, charge density rear-
rangement leads to the formation of an interface dipole that causes a
potential step between the bulk materials.41 The dipole contribution can
be combined with the bulk-derived valence-band edge positions to obtain a
band offset for a particular interface. Let the volume-a~eraged,~’electros-
tatic (Hartree) potential ( VH)A(B) define an absolute energy level with
respect to which the quasiparticle valence-band edges EV,A(B) are measured.
The (LDA) dipole potential is given as V&:c) = ( V H ) y D ”- ) (VH)LjD”).
Denote the difference in the quasiparticle corrections to the bulk LDA band
structure of the semiconductors A and B as AX. The valence-band offset is
144 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

then given as
AEv = &,A +
- E ~ . B hipole =
ELDA
U.A
- ELDA
v,B + AX + Vdipole. (4~~)
Under the assumption that LDA describes the interface dipole correctly,
that is, hipole = Vk;ie, and ignoring possible complications related to
absolute energy shifts in GWA,416AEv is determined from an LDA supercell
calculation, which gives Ef,3A- Ef,$A and the interface dipole, and from
bulk GWA calculations, which determine AX. This implies that interface
GWA calculations are computationally less demanding than quasiparticle
calculations of superlattices since they require consideration of only the bulk
rather than the generally much larger superlattice unit cell. Neglect of the
difference between the real and the LDA interface dipole potential is
considered to be a good approximation, since LDA describes semiconductor
densities to within a few percent.
Density distributions at semiconductor interfaces such as GaAs-A1As are
smooth because of the similar average valence electron densities and
screening properties of the constituent materials (average valence electron
density = 0.177 electrons/A3 (GaAs), 0.176 electrons/A3 (AlAs); dielectric
constant = 10.8 (GaAs), 8.2 (AlAs)). More inhomogeneous systems such as
metal-vaccum interfaces require a self-consistent treatment of quasiparticle
wave functions, as discussed in Section V.28.
Quasiparticle corrections to the valence-band offset at a GaAs-A1As
i n t e r f a ~ e ~amount
l ~ . ~ ~to~about 0.1 eV or 30% of the valence-band offset
of 0.41 eV determined in LDA. Self-energy corrections are crucial to obtain
a theoretical valence-band offset to 0.53 f 0.05 eV, in good agreement with
experimental values of 0.53-0.56 eV.418*419
As in the case of GaAs-AlAs, Table 28 shows that quasiparticle effects

415 After performing an in-plane average of, for example, the potential v(z) =
l/sJsV(x,y,z)dxdy (S is the in-plane area of the surface unit cell), an additional “running”
average over the extension of a unit cell in the growth direction (z) is performed. With a as the
lattice constant in the z direction, one defines V(z) = I/a J:?$z T(z’)dz’, which reduces to a
constant in the bulk and shows a smooth transition at the interface. For a discussion of these
kinds of averages, see, for instance, S. Baroni, R. Resta, A. Baldereschi, and M. Peressi, in
Spectroscopy of Semiconductor Microstructures, eds. G. Fasol, A. Fasolino, and P. Lugli,
NATO AS1 Series B, vol. 206, Plenum Press, New York (1989), 251.
416 The absolute energy level in GWA can be shifted by convergence, the choice of the
exchange-correlation potential, and the choice of vertex corrections, as detailed in Section 11.14.
The valence-band offset is determined with the implicit assumption that the difference in
self-energy corrections is physically meaningful and independent of technical details provided
that both semiconductors are treated on the same level of approximation.
417 S. B. Zhang, D. Tomhek, S. G. Louie, M. L. Cohen, and M. S. Hybertsen, Sol. State Comm.
66, 585 (1988).
418 P. Dawson, K. J. Moore, and C. T. Foxon, in Quantum Well and Superlattice Physics,
Proceedings of the SPIE 792, eds. G. H. Dohler and J. N. Schulman, SPIE, Washington
(1987), 208.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 145

28. VALENCE-BAND
TABLE OFFSETM eV AT THE NONPOLAR,
ZINCBLENDE
GaN/AlN( 110) INTERFACE, USINGTHE AVERAGELATTICECONSTANT OF
AlN AND GaN AND PSEUDOPOTENTIALS WITH THE Ga 3d ELECTRONS
TREATED EITHERAS CORE OR AS VALENCE ELECTRONS'96OR LMT0.420.421
INCLUSION OF THE Ga 3d ELECTRONS AS VALENCE ELECTRONS LEADSTO
AGREEMENT BETWEEN THE THREE LDA CALCULATIONS AND CHANGES THE
SIGN OF THE INTERFACED1POLE.lg6 THE SELF-ENERGY CORRECTION OF
0.25 eV EQUALS ABOUT 30% OF THE LDA VALENCE-BAND OFFSET AND HAS
BEEN ESTIMATED FROM BULK CALCULATIONS THAT TREATTHE Ga 3d
ELECTRONS VIA NONLINEAR CORE CORRECTIONS'96 (SEE SECTION III.15a
FOR A DISCUSIONOF THE TREATMENT OF d ELECTRONS IN GWA). THE
EXPERIMENTAL VALUESREFeR TO WURTZITE INTERFACES.

LDA

Ga 3d CORE Ga 3d VALENCE GWA

Cociorva' 96 0.62 0.93 1.18

Albanesi4'' 0.85
Ke4" 0.81
Expt. 0.70 f 0.24"
1.36f0.07'

"Ref. 422; *Ref. 423.

increase the valence-band offset at a zincblende GaN-AlN (110) nonpolar

interface by about 30%, or 0.25 eV.Ig6The size of the self-energy correction
for the nonpolar interface as well as the large discrepancy between experi-
mental valence-band offsets for the polar wurtzite GaN-AlN (OOO1) interface
(0.70 & 0.24 eV422versus 1.36 f 0.07 eV423in comparison to a theoretical
value of 1.18 eV) suggests that self-energy corrections are important in
understanding GaN-AlN interfaces. Treatment of the Ga 3d electrons as
valence rather than core electrons increases the LDA valence-band offset by
about 50% and switches the sign of the LDA interface dip01e.I~~ For
practical purposes, one can treat the Ga 3d electrons via nonlinear core
corrections205in the GWA-that is, one can rely upon error cancellation
between the neglect of wave function relaxation and the neglect of screening
and exchange (see Section III.15a for a discussion of the GWA treatment of
d electrons). This approach gives good quasiparticle energies in comparison
with experiment even for 11-VI semiconductors.51

419 D. J. Wolford, in Proceedings of the 18th International Conference on the Physics of

Semiconductors, World Scientific, Singapore (1987), 1115.
420 E. A. Albanesi, W. R. L. Lambrecht, and B. Segall, J. Vac. Sci. Technol. B 12,2470 (1994).
421 S.-H. Ke, K.-M. Zhang, and X.-D. Xie, J. Appl. Phys. 80,2918 (1996).
422 G. Martin, A. Botchkarev, A. Rockett, and H.Morkw, Appl. Phys. Lett. 68, 2541 (1996).
423 J. R. Waldrop and R. W. Grant, Appl. Phys. Lett. 68,2879 (1996).
146 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Quantitatively accurate predictions of transport properties require an

accuracy of quasiparticle calculations that is better than the relevant energy
scale of about 25meV set by room temperature. As discussed in Section
111.18a, only a few independent quasiparticle calculations agree to within
50 meV for quasiparticle energies or even quasiparticle energy corrections.
Differences on the order of a few tenths of an eV between quasiparticle
calculations and also between theory and experiment are the norm, indica-
ting that the determination of self-energy corrections to band offsets pushes
the accuracy limits of the GWA. Nevertheless, a consistent improvement
over LDA of calculated valence-band offsets compared to experiment can
be achieved.41
Further Applications. An overview of valence-band offset calculations in
the GWA and explicit calculations for the In,X, -,As-In,Y, - y Z (X = Ga,
Al; Y = Ga, Al; Z = P, As) system is given in Ref. 413. These alloy
calculations are based on the virtual crystal approximation, which may not
capture local bond-length fluctuations in the real alloy that could system-
atically change the interface dipole.

c. Schottky Barriers
At a metal/insulator interface such as Al/GaAs(l lo), alignment of the metal
and semiconductor Fermi levels leads to the creation of a 100-1000-A-thick
space charge layer on the semiconductor side, which, via Poisson’s equation,
bends the valence and conduction bands. Electrons (holes) that flow from
the metal into the n-type (p-type) semiconductor must overcome the
so-called Schottky barrier if the metal Fermi level is pinned inside the
semiconducting gap E, (for a recent review of Schottky barriers see Ref.
424). The n-type and p-type Schottky barriers B, and B, are defined in terms
of the conduction-band minimum E, and the valence-band maximum E, of
the bulk semiconductor and the bulk metal Fermi energy E, as

B, = E, - E ,
B, = E , - E ,
* B, + B, = E,. (4.5)

In analogy with Eq. (4.3) and neglecting density rearrangements beyond

LDA, we obtain with AX as the quasiparticle correction to E, - E , and
E , - E,, respectively,
Bn(,) = Bkg: + AX (4.6)

where B t i t is the LDA Schottky barrier.

424 J.-G. Li, Mat. Chem. Phys. 47, 126 (1997),and references therein.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 147

In contrast to the LDA Schottky barriers, the bulk quasiparticle correc-

tions do not depend upon the atomic structure of the interface and equal
0.22 eV in the case of Al/GaAs(l 10).42s*426*427 This correction amounts to
20-60% of the LDA barrier depending on the particular interface structure
chosen and establishes the importance of many-body corrections.
Self-energy effects lead not only to corrections to the Schottky barrier
height but also to semiconductor band-gap narrowing in the 1-lo-A
vicinity of the metal-semiconductor i n t e r f a ~ e . ~Classically,
~ ” ~ ~ ~ an electron
in the semiconductor experiences an additional energy-lowering, electro-
static potential whose effect can be described by the interaction of the
electron with its image charge in the metal. With E as the dielectric constant
of the semiconductor and z as the electron’s distance from the interface, the
energy lowering of the conduction band equals 1/(4~z).Analogously, the
valence band is bent upward by the same amount. Quasiparticle calcula-
tions for Al/GaAs(llO) give a significant band-gap reduction of 0.4eV
compared to a bulk GWA gap value of 1.1 eV (core-polarization effects are
omitted; see Refs. 425, 426) and suggest that the narrowing of the gap is
mainly due to the bending of the conduction band rather than to the equal
distribution of the gap between the valence and conduction bands as in the
classical case.42s,426
21. SURFACES
The theoretical determination of surface structures and reconstructions
relies on two techniques:428 (1) calculation of surface-state bands for
proposed surface geometries and comparison with spectroscopic data, and
(2) total energy minimization over some set of geometries to find optimal
positions for atoms near the surface. While the second approach is a reliable
tool because of the accuracy of LDA in determining densities, the first
approach suffers from the LDA band-gap underestimation, which translates
into similar albeit smaller errors for surface states. Reliable empirical
corrections to LDA are difficult to construct since quasiparticle corrections
for surface states can show strong dispersion rather than simple “scissors-
shift”-like behavior, as discussed below.
LDA surface-state energies disagree with experiment4” since (1) band
gaps between empty and occupied surface-state energies are too small, (2)
the dispersion of LDA surface band states is too small in some cases, too

425 J. P. A. Charlesworth, R. W. Godby, R. J. Needs, and L. J. Sham, Mat. Sci. Eng. B 14, 262
(1 992).
426 J. P. A. Charlesworth, R. W. Godby, and R. J. Needs, Phys. Rev. Lett. 70, 1685 (1993).
4 2 7 R. J. Needs, J. P. A. Charlesworth, and R. W. Godby, Europhys. Lett. 25, 31 (1994).
4 2 8 M. S. Hybertsen and S. G. Louie, Phys. Rev. B38,4033 (1988).
148 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

large in others, and (3) the placement of occupied surface-state energies is

in some cases too high by 0.5 to 1.0eV relative to the bulk valence-band
maximum. Two missing physical effects must be considered to improve
upon these LDA failures.428 First, the inclusion of local fields in the
dielectric matrix is crucial for the quasiparticle approach since these local
fields describe the strongly inhomogeneous screening (bulk-like versus
vacuum) at the surface. Second, the nonlocality of the self-energy operator
is more sensitive to the localization properties of surface states than the only
density-dependent LDA. This nonlocality leads to a modified dispersion of
the quasiparticle energies throughout the surface Brillouin zone.
Quasiparticle corrections to the LDA surface band gap lead to improved
agreement with experiment and depend strongly on the character of the
occupied and unoccupied surface states.429 The occupied and unoccupied
surface states of the GaAs(ll1) (2 x 1) surface430 derive their character
from the corresponding bulk valence and conduction bands and are local-
ized on the As and Ga atoms, respectively. Consequently, the quasiparticle
corrections to the surface states are substantial and amount to a roughly
rigid shift by 0.7eV of the unoccupied surface states with respect to the
occupied surface states, which leads to a quasiparticle surface gap of
2.0 eV,430in good agreement with the experimental value of 1.9 eV.431The
quasiparticle corrections for surfaces states of the Ge(ll1):As (1 x 1)428,432
surface are different from the bulk corrections for Ge valence and conduc-
tion bands since (1) the occupied surface state is an As lone pair, that is, a
filled As dangling bond, and not derived from bulk Ge states, and (2) the
unoccupied surface states derive their character from both Ge valence and
conduction bands. The bulk self-energy corrections amount to -0.1 eV for
occupied and +0.6 eV for unoccupied states. In contrast, the self-energy
shift of the As lone pair equals about +0.1 eV and that of the unoccupied
surface state is intermediate between the bulk Ge occupied and unoccupied
shifts and equals +0.45 eV.428,432
Quasiparticle corrections often improve the dispersion of occupied and
unoccupied LDA surface states and agree well with experiment. Self-energy
corrections can either increase or decrease the surface-state dispersion, as
shown by the following examples. For the Ge(l1l):As (1 x 1) s ~ r f a ~ e ~ ~ ~
quasiparticle corrections increase the band width of the occupied As lone-
pair band by 0.5eV (LDA 1.18eV, GWA 1.64eV), in excellent agreement
with experiment (1.62 eV; Refs. 433,434, and 439, and modify the unphysi-

429 X. Blase, X. Zhu, and S. G. Louie, Phys. Rev. B49,4973 (1994).

430 X.Zhu, S. B. Zhang, S. G. Louie, and M. L. Cohen, Phys. Rev. Lett. 63,2112(1989).
431 J. van Laar, A. Huijser, and T. L. van Rooy, J. Vuc. Sci. Technol. 14,894 (1977).
432 M.S. Hybertsen and S. G. Louie, Phys. Rev. Lett. 58, 1551 (1987).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 149

cal, flat LDA dispersion into a parabolic dispersion at in the surfacer

Brillouin zone, as shown in Fig. 35. For the Si(ll1):H (1 x 1) s ~ r f a ~ e ~ ~ ~
the occupied surface band a’ between M and I? has an LDA band width of
0.42 eV but none in GWA and experiment. Note that unreconstructed
(1 x 1) surfaces that result from passivation with As or H are excellent
systems for quasiparticle calculations since the complications of surface
reconstructions are eliminated and supercell calculations can be based on a
small 1 x 1 unit cell rather than 2 x 1, 4 x 2, or even larger unit cells.
Quasiparticle calculations bring the alignment of occupied states with
respect to the valence-band maximum into agreement with experiment.
Large self-energycorrections of -0.5 to -0.8 eV of occupied surfaces states
are obtained for strongly localized surface states of the Si(ll1):H (1 x 1)
surface and are shown in Table 29. Strongly localized LDA states are
underbound since the Hartree and exchange self-interactions cancel only

433 R. D. Bringans, R. I. G. Uhrberg, R. Z. Bachrach, and J. E. Northrup, Phys. Rev. Left. 55,
533 (1985).
434 R. D. Bringans, R. I. G. Uhrberg, R. Z. Bachrach, and J. E. Northrup, J. Vuc. Sci. Technol.
A 4 , 1380 (1986).
4 3 5 R. D. Bringans, R. I. G. Uhrberg, and R. Z. Bachrach, Phys. Rev. 834,2373 (1986).

436 K. Hricovini, G. Giinther, P. Thiry, A. Taleb-Ibrahimi, G. Indlekofer, J. E. Bonnet,

P. Dumas, Y. Petroff, X.Blase, X. Zhu, S. G. Louie, Y. J. Chabal, and P. A. Thiry, Phys. Rev.
Lett. 70, 1992 (1993).

5 ,

--
% -I

P: -8
bl

-
P r it
FIG. 35. Left panel: LDA and quasiparticle (QP) Ge( 1 1 1):As (1 x 1) surface-state energies in
comparison to the projected bulk quasiparticle band structure (shaded) along high-symmetry
directions in the surface Brillouin zone. Quasiparticle corrections (1) open up the gap between
r;
the surface states; (2) eliminate the unphysical, flat LDA dispersion at and (3) improve the
position of the occupied surface bands with respect to the valence-band maximum in
comparison to experiment. Right panel: Calculated occupied quasiparticle surface-state ener-
gies compared to angle-resolved photoemission data (Ref. 433). Agreement between theory and
experiment is within the quoted theoretical uncertainty of kO.1 eV. (Adapted from Ref. 432.)
150 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

TABLE29. LDA, GWA, AND EXPERIMENTAL ENERGIESOF

Si( 11 l):H( 1 x 1) SURFACESTATESAT I? AND (eV) WITH RESPECT
TO THE VALENCE-BAND MAXIMUM.THEGWA CORRECTS LDA
ENERGIES BY -0.5 TO -0.8eV. LDA UNDERBINDS STRONGLY
LOCALIZEDSTATESBECAUSE OF UNPHYSICALSELF-INTERACTIONS,
WHICH ARE ELIMINATED IN THE GWA. (ADAPTEDFROM REF.429.)

k-POINT LDA GWA EXPT.~

I? - 3.22 - 3.82 - 3.80

- 4.29 -4.76 -4.78
- 7.85 - 8.47 - 8.64
M - 3.87 -4.63 -4.76

“Ref. 436.

partially in LDA. These self-interaction errors are largely eliminated in

GWA, leading to the large self-energy corrections mentioned above.
GWA calculations cannot always eliminate the discrepancies between
L D A and experiment in the dispersion and position of occupied surface
states. A case in point is the free Ge(001) (2 x 1) surface4j7 whose band
structure is shown in Fig. 36. Two conclusions follow from Fig. 36: (1) a
significant discrepancy of up to 0.8 eV remains between quasiparticle calcu-

437 M. Rohlfing, P. Kriiger, and J. Pollmann, Phys. Rev. B54, 13759 (1996).

r J , K J’ r [oto] J;

FIG. 36. GWA (solid line), LDA (dashed line), and experimental photoemission data
(diamonds: Refs. 438, 439, 440, circles: Refs. 441, 442) for occupied and unoccupied states of
the clean Ge(001) (2 x 1) surface. Clear discrepancies with experiment exist for the dispersion
of the unoccupied Ddown state, of the occupied D,, state between JK/2 and J’, and for the
dispersion of the occupied states at I-. The projected bulk GWA band structure is shadowed,
and energies are measured with respect to the valence-band maximum. Open symbols represent
weak features in the experimental spectra. (From Ref. 437.)
 QUASIPARTICLE CALCULATIONS IN SOLIDS 151

lations and experiment along the JK/2 to J' direction in the surface Brillouin
zone, and (2) the 0.6-eV experimental band width of the Ddownstate is
strongly overestimated in quasiparticle theory (z1.0eV, Ref. 437). Similar
discrepancies exist for the sulphur-terminated Ge(001):S (2 x 1) surface.437
Here the energies of the occupied quasiparticle D surface band and the
occupied B surface band are underbound by 0.9 and 0.4eV, respectively,
compared to experiment.
Quasiparticle calculations for surface states help explain and analyze
surface reconstructions and can be especially valuable when LDA total
energy minimizations lead to contradicting results with very different
electronic properties. For example, the Pandey n-bonded-chain model is
favored for the explanation of the C( 111) (2 x 1) surface reconstruction, but
the precise position of the atoms in the surface has been a subject of
discussion. Vanderbilt and suggested a surface geometry of slightly
buckled but undimerized chains based on a linear-combination-of-atomic-
orbitals LDA calculation. Iarlori et a1.444found a dimerized surface with no
buckling based on LDA molecular dynamics simulations. Model GWA
calculations by KreB, Fiedler, and B e ~ h s t e d t ,whose
~ ~ ~ results
, ~ ~ ~are shown
in Fig. 37, support the dimerized model. The quasiparticle calculations show
that the Vanderbilt-Louie model leads to a metallic surface and that the
model by Iarlori et al. opens up a surface gap of 1.7 k 0.3 eV, in good
agreement with the experimental value of about 2.0 eV.447,448

438 L. Kipp, R. Manzke, and M. Skibowski, Surf: Sci. 269/270, 854 (1992).
439 L. Kipp, R. Manzke, and M. Skibowski, Sol. State Comm. 93, 603 (1995).
440 M. Skibowski and L. Kipp, J. Electron Spectrosc. Relat. Phenom. 68, 77 (1994).

4 4 1 E. Landemark, R. I. G. Uhrberg, P. Kriiger, and J. Pollmann, Surt Sci. Lett. 236, L359

(1990).
442 E. Landemark, C. J. Karlsson, L. S. 0. Johansson, and R. I. G. Uhrberg, Phys. Rev. B49,
16523 (1994).
443 D. Vanderbilt and S. G. Louie, Phys. Rev. B29, 7099 (1984).
444 S. Iarlori, G. Galli, F. Gygi, M. Parinello, and E. Tosatti, Phys. Rev. Lett. 69, 2947 (1992).
445 C. KreB, M. Fiedler, and F. Bechstedt, Europhys. Lett. 28, 433 (1994).

446 C. KreB, M. Fiedler, W. G. Schmidt, and F. Bechstedt, Surf. Sci. 331-333, 1152 (1995).

447 S. V. Pepper, Surf. Sci. 123, 47 (1982).

448 Further applications: Several other surfaces have been considered in first-principles GWA

calculations, such as the Ge(ll1) (2 x 1) surface (Ref. 449); the Ge(001):H (2 x 1) surface (Ref.
437); the Si(l1l):As (1 x 1) surface (Refs. 428 and 450); the Si(ll1) (2 x 1) surface (Ref. 451); the
Si(OO1) (2 x 1) surface (Refs. 284 and 452); the Si(001) c(4 x 2) surface (Ref. 453); and the
p-SiC(001) (2 x 1 ) surface (Ref. 454).
449 X. Zhu and S. G. Louie, Phys. Rev. B43, 12146 (1991).

4 5 0 R. S. Becker, B. S. Swartzentruber, J. S. Vickers, M. S. Hybertsen, and S . G. Louie, Phys.

Rev. Left. 60, 1 16 (1988).

4 5 ' J. E. Northrup, M. S. Hybertsen, and S. G. Louie, Phys. Rev. Lett. 66, 500 (1991).

4 5 2 M. Rohlfing, P. Kriiger, and J. Pollmann, Phys. Rev. B52, 13753 (1995).

453 J. E. Northrup, Phys. Rev. B47, 10032 (1993).

152 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

r J K J’ r J K J’
FIG. 37. Quasiparticle band structure of the C(111) (2 x 1) surface for slightly buckled,
undimerized rr-bonded chains (left and dimerized but unbuckled r-bonded chains
(right Only the latter case (b) leads to semiconducting rather than metallic behavior
and a surface gap of 1.7 eV445in good agreement with the experimental gap of about 2.0 eV.447
The projected bulk quasiparticle band structure is shown in black. (Adapted from Ref. 445.)

22. DEFECTS

Defects in semiconductors affect the transport, electronic, and optical

properties by introducing defect levels within the band gap. An LDA
description of defect levels is insufficient since the underestimation of the
bulk band gap by 30-100% translates into a corresponding uncertainty for
the defect level position. Also, quasiparticle shifts depend on the character of
the defect level and cannot be estimated from the shifts of the bulk band
edges in a simple way. Indeed, we have discussed examples of strongly
dispersive and state-dependent quasiparticle shifts at semiconductor surfa-
ces, in transition metals, and even in large-gap bulk insulators (Sections
111.18, IV.19, IV.20, IV.21, and V.26).
Quasiparticle calculations can determine defect levels accurately but are
computationally challenging. To date, applications are limited to the F-
center defect or halogen vacancy in LiC14” and an oxygen vacancy in
zirconia, ZrO,.” We confine the discussion to the halogen vacancy in LiCl
since this system is an electronically and structurally simple point defect.
This neutral vacancy contains one bound electron in a highly localized

454 M.Sabisch, P. Kriiger, A. Mazur, M.Rohlfing, and J. Pollmann, Phys. Reo. B53, 13121
( 1996).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 153

(within one to two lattice constants) 1s state and shows structural relaxations
relative to the ideal crystal structure that are mostly confined to nearest-neighbor
alkali-metal at0ms.4~~ In this case, a theoretical description of the F-center using
a supercell approach with relatively small unit cells seems appropriate.
The fundamental electronic F-center transitions are (1) transitions from
the localized 1s electron to the conduction band critical points at L, A, and
X, and (2) an intra-defect transition from the localized 1s to a localized 2 p
level. For the first transition, the experimental values of 4.5 eV, 5.0 eV, and
5.8eV for the L, A, and X transitions are well reproduced by the quasi-
particle values 4.5 eV, 5.0 eV, and 5.7 eV, which correct LDA results of
1.8 eV, 2.2 eV, and 2.8 eV by an almost constant shift of 2.8 eV.455For the
second, because of the localized nature of the 1s and 2 p orbitals involved in
the 1s -,2p intra-defect excitation, excitonic effects must be included in the
determination of the quasiparticle transition energy. These effects lower the
transition energy by about 1eV, leading to a quasiparticle value of 3.4 eV-
a 1.0-eV correction to the LDA value of 2.4eV-in good agreement with
experiments ranging from 3.1 to 3.3 eV.455
Besides excitonic effects, quasiparticle defect calculations are challenging
because of problems related to finite-size effects, the treatment of the
localization and spin of the defect electron, the possible need for a self-
consistent treatment of quasiparticle defect states, and so forth. In the case of
the 2p defect state even the determination of the corresponding LDA state is
difficult since the band gap underestimation in LDA causes this state to be
resonant and hybridize with bulk bands. For a detailed discussion of technical
problems occurring in GWA defect calculations, we refer the reader to Ref. 455.

23. PRESSURE

Quasiparticle calculations can describe isostructural metal-insulator transi-

tions due to pressure-induced overlap of conduction and valence bands.
Examples discussed below include solid Xe,456*457 solid molecular hydro-
gen,457*458 and diamond.459 Other possible scenarios for metal-insulator
transitions include structural transformations, as is the case for Si under
and simultaneous magnetic and metal-insulator transitions, as

455 M. P. Surh, H. Cacham, and S. G. Louie, Phys. Rev. 851, 7464 (1995).
456 H. Chacham, X. Zhu, and S. G. Louie, Europhys. Lerr. 14,65 (1991).
457 H. Chacham, X. Zhu, and S. G. Louie, Phys. Rev. 846,6688 (1992).
4s8 H. Chacham and S. G. Louie, Phys. Rev. Leu. 66,64 (1991).
4s9 M. P. Surh, S. G. Louie, and M. L. Cohen, Phys. Rev. 845, 8239 (1992).
460 For a review, see F. Siringo, R. Pucci, and N. H. March, High Press. Res. 2, 109 (1989).
154 WILFRIED G. AULBUR, LARS JBNSSON, AND JOHN W. WILKINS

TABLE30. THE PRESSURE DERIVATIVESOF Si BANDGAPS(eV/Mbar) AT

EQUILIBRIUM r~ LDA, GWA, AND EXPERIMENT.INPARTICULAR,
THE DATA
OF ZHU et al.462 SHOWTHAT LDA PRESSURE DERIVATIVES
ARE IN GOOD
AGREEMENT WITH EXPERIMENT AND WITH GWA. THE THEORETICAL
RESULTS OF REF. 463 ARE ACCURATE TO WITHIN f0.3eV/Mbar.463
FROM REFS.462 and 463.)
(ADAPTED

z H U4 6 2 GODBY463

STATE EXPT. LDA GWA LDA GWA

0.85 X , , - 1.6" -1.41 -1.32 -1.3 - 1.8

- 1.44.b
XI, -1.73 -1.68 -0.9 - 1.6
LIC 3.95 4.06
r15c 0.59 0.53 0.5 0.6
r;, 11.9 12.2

"Ref. 182; bRef. 464.

in NiI,.461 To date, quasiparticle calculations for structural or magnetic

metal-insulator transitions have not been attempted.
The simplest estimate of the metal-insulator transition pressure, at which
band overlap occurs, is given by the negative ratio of the energy gap EBap
and the variation of the energy gap with respect to pressure, dE,,,/dP.
Assuming linear behavior, a theory that predicts Egap and dE,,,/dP in
agreement with experiment will predict correct transition pressures. Table
30 indicates that the derivative of energy gaps with respect to pressure
evaluated at the equilibrium volume is rather well described in LDA. The
success of LDA in the determination of dE,,,/dP can be understood in terms
of a simple two-band semiconductor which shows that the
screened exchange term is not very sensitive to pressure and that the
dependence of the self-energy on the density and hence on pressure is mainly
due to its Coulomb-hole term. The Coulomb-hole term can be approxi-
mated by a local potential, for instance, an LDA exchange-correlation
potential (see Section 11.10). Note that with dE,,,/dP being rather well
described in LDA, the LDA band-gap underestimation leads to a systematic
underestimation of the metal-insulator transition pressure.

4 6 1 M. P. Pasternak, R. D. Taylor, A. Chen, C. Meade, L. M. Falicov, A. Giesekus, R. Jeanloz,

and P. Yu,Phys. Rev. Lett. 65, 790 (1990).
462 X.Zhu, S. Fahy, and S. G. Louie, Phys. Rev. B39,7840 (1989).
463 R. W. Godby and R. J. Needs, Phys. Rev. Lett. 62, 1169 (1989).
464 B. Welber, C. K. Kim, M. Cardona, and S. Rodriguez, Sol. State Comm. 17, 1021 (1975).

465 L. Brey and C. Tejedor, Sol. State Comm. 55, 1093 (1985).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 155

The increased band gap in GWA calculations compared to LDA leads

to a higher transition pressure in better agreement with experiment.
For example, the transition pressure of solid Xe457is 128 GPa in compari-
son to 104 GPa in LDA and experimental values of 132 & 5 GPa466 and
150GPa.467Note that spin-orbit coupling must be taken into account in
the cal~ulations.4~'An LDA calculation without spin orbit predicts a
transition pressure of 123 GPa and hence agrees with experiment, although
it does not describe the physics correctly. The high-pressure behavior of
diamond is of particular interest because of the widespread use of diamond-
anvil cells. In diamond, the minimum band gap decreases under application
of anisotropic pressure along the [OOl] direction, and the pressure coeffi-
cient for the minimum gap is very sensitive to the degree of anisotropy of
the stress.459The experimental geometry and hence the stress anisotropy in
diamond-anvil cells is not fully known, which may explain the underestima-
tion of the experimental diamond metallization pressure of 700-900 GPa468
by GWA calculations459that predict a metallization pressure of 400 GPa
for an idealized geometry.
The self-energy correction to the LDA band structure of solid molecular
hydrogen shows a nonlinear density dependence and is larger for molecular
hydrogen at equilibrium than for molecular hydrogen under large pressure
(at high densities).457Figure 38 depicts the density dependence of the band
gap of solid molecular hydrogen for LDA, Hartree-Fock, and GWA. For
low densities (atmospheric pressure) the self-energy correction to LDA is
large since this regime is dominated by exchange and since GWA is close to
Hartree-Fock, which is known to be accurate for molecular solids. For high
densities, that is, large pressures, the electron density of solid molecular
hydrogen becomes more uniform and consequently the corrections to LDA
become smaller. Since the corrections to the LDA band structure are
nonconstant, it follows that dE,,/dP will be different in LDA and quasi-
particle calculations. Note that LDA calculations for solid, molecular
hydrogen are difficult457since (1) the exact experimental structure at metal-
lization is unknown; (2) the large zero-point motion energy of the H,
molecules makes accurate total energy calculations difficult since it is larger
by an order of magnitude than the differences in electronic energies between
different molecular orientations; and (3) in contrast to the total energy the
minimum band gap is very sensitive to the orientation of H, molecule
vibration. Disorder in the vibration direction of the H, molecules increases

466 K. A. Goettel, J. H. Eggert, I. F. Silvera, and W. C. Moss,Phys. Rev. Lett. 62,665 (1989).
467 R. Reichlin, K. E. Brister, A. K. McMahan, M. Ross, S. Martin, Y. K. Vohra, and A. L.
Ruoff, Phys. Rev. Lett. 62,669 (1989).
468 A. L. Ruoff, H. Luo, and Y. K. Vohra, J. Appl. Php. 69,6413 (1991).
156 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

a
E
6
rn

Density (mol/cma)

FIG. 38. Minimum band gap of onentationally ordered solid molecular hydrogen in the hcp
structure as a function of density and hence pressure for Hartree-Fock (HF), GWA (GW), and
LDA. GWA is close to Hartree-Fock for low densities and close to LDA for high densities.
Lines are drawn as a guide to the eye. (From Ref. 457.)

the minimum band gap and is (1) negligible at zero pressure,"69 and (2)
crucial at high densities exemplified by a metal-insulator transition pres-
sure of 151 GPa for hcp H,, whose molecules are perfectly aligned along
the c-axis and of 300 GPa for orientationally disordered H2.457Experimen-
tally, metallization of solid hydrogen-defined by a finite T + 0 DC con-
ductivity-remains to be observed (Refs. 470, 471, 472, and 473). Optical
experiments have failed to detect evidence for gap closure below about
200 GPa,"70.471 and a metal-insulator transition pressure of more than
300 GPa seems likely.473

469 The LDA and GWA gaps for hcp H, with all molecules aligned along the c-axis are 8.4
and 15.4 eV. When disorder is included via a virtual crystal approximation, the gaps change
to 8.3 and 15.3 eV compared to an experimental value of 14.5+ 1 eV (Ref. 457).
470 N. H. Chen, E. Sterer, and I. F. Silvera, Phys. Rev. Left. 76, 1663 (1996).
471 R. J. Hemley, H.-K. Mao, A. F. Goncharov, M. Hanfland, and V. Struzhkin, Phys. Rev.
Left 76, 1667 (1996).
472 R. J. Hemley and N. W. Ashcroft, Physics Today 51.26 (1998).
473 M. Ross, Phys. Rev. 854, R9589 (1996).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 157

24. EXCITONS

Interacting electron-hole pairs are created in absorption spectroscopy, and

their electron-hole attraction leads to the following modifications of the
optical absorption spectra of semiconductors and insulators: (1) the energies
of low-lying excited states are decreased; and (2) the corresponding oscillator
strength is increased. Semi-empirical tight-binding results of Hanke, Sham,
and collaborators (see Section II.9b) for Si show about a 1-eV shift of the
energies of low-lying excited states to lower energies and an oscillator strength
increase of the El peak at about 3.5eV by 50%, in good agreement with
experiment. However, the strength of the theoretical El peak is predicted to be
larger than that of the 4.2 eV E, peak, in contrast to experiment.
Recent implementations that include excitonic correlations extend the
Hanke-Sham approach by using first-principles rather than semi-empirical
wave functions and band energies. Applications to bulk semiconductors and
wide-gap insulators (Si, Ge, GaAs, diamond, Li,O, MgO, LiF; Refs. 56,474,
475, 476, and 477) confirm Hanke and Sham’s conclusions regarding the
shift of low-lying excited states to lower energies and find in addition the
correct ordering (El < E,) of the two main peaks in the Si absorption
spectrum. Figure 39 compares theoretical absorption spectra of Si and
diamond with and without excitonic correlations with experimental re-
s u l t ~and ~ shows
~ ~ that
, ~ good
~ ~agreement between theory and experiment
is obtained and that excitonic effects are more important in large-gap
insulators than in small- and medium-gap semiconductors, since screening
is less effective in the former. Core-hole excitons have been studied in LiF,
NaF, KF, graphite, diamond, and h-BN.479
Excitonic binding energies in clusters such as Na, (Ref. 480) or Si,H,
(Ref. 282) amount to a few eV compared to binding energies of a few meV
in bulk, since screening is inefficient compared to the bulk and electrons
and holes are confined in a small region. In contrast to bulk excitons, which
at low energies can be described by singly excited electronic states, excitons in
clusters require the consideration of several excited electronic states.262*480
Note also that dynamic screening of the electron-hole interaction results in
smaller changes (a few tenths of an eV) in the final excitation energies.282

474 S.Albrecht, L. Reining, R. Del Sole, and G. Onida, Phys. Rev. Lett. 80, 4510 (1998).
475 L. X. Benedict, E. L. Shirley, and R. B. Bohm, Phys. Rev. B57, R9385 (1998).
476 L. X. Benedict, E. L. Shirley, and R. B. Bohm, Phys. Rev. Lett. 80,4514 (1998).
477 M. Rohlfing and S. G. Louie, Phys. Rev. Lett. 81, 2312 (1998).
478 Handbook of Optical Constants of Solidr, ed. E. D. Palik, Academic Press, Boston (1991).
479 E. L. Shirley, Phys Rev. Lett. 80, 794 (1998).
480 G. Onida, L. Reining, R. W. Godby, R. Del Sole, and W. Andreoni, Phys. Rev. Lett. 75,
818 (1995).
158 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

2 2.5 3 3.5 4 4.5 5 5.5 6

Eaasy (eV)

2 S , ] , ,
;‘ ‘ a

E2 E,

6 8 10 12 14 16 1% 20 6 8 10 12 14 16 18 20
hiD (eV) (eV)
FIG.39. Imaginary part of the dielectric function E ~ ( w )= Im(&(o)) as a function of ho for
Si (upper panel, adapted from Ref. 474) and diamond (lower panel, adapted from Ref. 475).
For Si, the dots correspond to experiment274 and the solid line includes local fields and
electron-hole attraction in the Hamiltonian; the long-dashed line includes only local fields, the
short-dashed line only RPA screening with quasiparticle shifts. The low-energy experimental
peak is commonly referred to as the E , peak, and the high-energy experimental peak is referred
to as the E , peak. For diamond, the left panel contains theoretical results, neglecting the
electron-hole interaction (dashed lines) and experimental results from Refs. 142 and 478 (solid
lines). The right panel contains theoretical results including the electron-hole interaction
(dashed line) and the same experimental results (solid line). Excitonic effects improve agreement
between theory and experiment significantly for peak position and height and are larger in
large-gap materials such as diamond than in medium-gap semiconductors such as Si.

25. ATOMSAND MOLECULES

This section describes quasiparticle calculations of atoms and molecules.

Atoms are applications of GWA to “zero-dimensional’’ systems. GWA
calculations for molecules have been reported only for quasi-one-dimen-
 QUASIPARTICLE CALCULATIONS IN SOLIDS 159

sional and quasi-two-dimensional systems. Understanding whether GWA

captures the most important physics of these systems also gives im-
portant insights into the usefulness of quasiparticle calculations in
strongly correlated solids such as NiO and high-T, cuprates, where cor-
relations are often strongly localized, that is, “atomic-like.’’ Extensions
to quasiparticle calculations to include vertex corrections valid for
atoms and molecules may therefore be of relevance for solids and are
the focus of this section. Other aspects, such as the physics of core-valence
correlations in atoms and their relevance for solids were discussed in
Section 11.12.

a. Atoms
Assessment of the usefulness of GWA for atoms requires the comparison of
theoretical and experimental results for some key quantities, such as the first
and second ionization potentials and transfer energies.67The GWA ioniz-
ation potentials of major group elements such as B, Al, and Ga deviate on
average only 3% from experimental values, compared to a 12% deviation in
H a r t r e e - F ~ c k No
. ~ ~ systematic improvement compared to the local spin
density approximation (LSDA)481*482 is obtained. For s + p promotion
energies the error is 20% in GWA, 40% in Hartree-Fock, and 5% in LSDA,
as shown in Table 31. In absolute terms, deviations between GWA and
experiment vary between 0 and 0.8eV. Hence, taking the dynamically
screened interaction into account leads to a significantly improved agree-
ment with experiment compared to Hartree-Fock but not compared to
LSDA. Absolute errors exceed chemical accuracy by more than one order
of magnit~de.~’
Figure 40 shows the s + d promotion energy in the iron series and
demonstrate^^^.^^^ that (1) GWA agrees with experiment quantitatively for
the first half of the iron series, in contrast to Hartree-Fock and the LSDA;
(2) GWA reproduces experimental trends qualitatively for the second half
of the iron series, in contrast to Hartree-Fock, but is inferior by about a
factor of three to the LSDA; and (3) GWA cannot compete in accuracy with

481 A. Gorling, private communication.

482 J. P. Perdew, J. A. Chevary, S. H. Vosko, K. A. Jackson, M. R. Pederson, D. J. Singh, and
C. Fiolhais, Phys. Rev. E 4 6 , 6671 (1992).
4 8 3 W. C. Martin and W. L. Wiese, in Atomic, Molecular, and Optical Physics Handbook, ed.

G . W. F. Drake, AIP, Woodbury, N Y (1996), 135-153.

484 A. A. Radzig and B. M. Smirnov, Reference Data on Atoms, Molecules, and Ions, Springer

Series in Chemical Physics, vol. 31, Springer-Verlag, Berlin (1985).

4 8 5 F. R. Vukajlovic, E. L. Shirley, and R. M. Martin, Phys. Rev. B 43, 3994 (1991).
486 D. M. Bylander and L. Kleinman, Phys. Rev. E41,7868 (1990).
160 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

TABLE 31. FIRST AND SECONDIONIZATION AND S~-PROMOTION

POTENTIALS ENERGIESIN GWA,
HARTREE-FOCK (HF), LOCALSPIN DENSITYFUNCTIONALTWRY (LSDA), AND EXPERIMENT
FOR THE GROUP-111ELEMENTS BORON, ALUMINUM,
AND GALLIUM.GWA IMPROVES UPON
HARTREE-FOCK BUT ERRORSOF ABOUT 20% IN THE s + p PROMOTION ENERGIESREMAIN
COMPARED TO AN ERROROF ABOUT 5% IN LSDA. IMPROVEMENT OF GWA IONIZATION
POTENTLUS o m LSDA IS MARGINAL.ALL ENERGIES
ARE IN eV. (ADAPTEDFROM REF. 67.)

PROMOTION
IONIZATION
POTENTLUS ENERGY

S Z p ( 2 P ) +s2( 1s) sZ( 1s) +42s) S Z p ( 2 P ) +sp(3P) S Z p ( 2 P ) +spZ(4P)

B
HF 7.5 23.8 10.4 2.0
GWA 8.5 24.9 : 12.0 2.8
LSDA" 8.6, 8.6b 24.6 12.7 3.2
Expt.' 8.3 25.2 12.9 3.6
A1
HF 5.3 17.8 8.5 2.1
GWA 6.0 18.9 10.0 3.0
LSDA" 6.0, 6.p 18.9 10.7 3.6
Expt.' 6.0 18.8 10.6 3.6
Ga
HF 5.3 19.0 9.5 2.9
GWA 6.0 20.0 11.2 4.1
LSDA" 6.1 20.8 12.3 5.0
Expt.' 6.0 20.5 12.0 4.8

"Ref. 481 unless otherwise noted; bRef. 482; 'as cited in Ref. 67; for further reference, see Refs.
483 and 484.

the results of a screened-exchange local spin density approach4" for the

entire iron series. The breakdown of GWA for the second half of the iron
series stems from the formation of strongly correlated pairs of electrons with
opposite spin.67 These strong, localized two-body correlations are not
included in the standard dynamically screened interaction approximation,
and their inclusion would require vertex corrections.

b. Molecules
Besides "zero-dimensional" atoms, quasiparticle calculations have been
done for quasi-one-dimensional trans-polyacetylene (trans-(CH)J4" -the
simplest prototype of a conducting polymer -and a quasi-two-dimensional

"' This approach modifies the LSDA by considering a nonlocal, screened Hartree-Fock
operator as part of the exchange-correlation potential (Ref. 486). This technique includes
nonlocality effects of the quasiparticle self-energy but omits its energy dependence, in contrast
to GWA.
488 E. C. Ethridge, J. L. Fry, and M. Zaider, Phys. Rev. B53, 3662 (1996).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 161

0.._.__.
0HF
D-OGWA
4-4 Exp.
B - - V LSD R ..a,
......
A- - 4 LSDSX Q\

f '\ '0

Ce Sc Ti V Cr Mn Fe Co Ni Cu
Element of iron series
FIG.40. The neutral s - + d promotion energies (eV) for the iron series elements from
experiment, Hartree-Fock (HF), and GWA6' and from the local spin-density approximation
(LSDA) and the screened-exchange local spin-density approximation (LSDSX)486from Ref.
485. GWA improves upon Hartree-Fock for all elements and upon the local spin-density
approximation for the first half of the series, but cannot compete in accuracy with the
screened-exchange local spin-density results. The figure combines data from Refs. 67 and 485.

hypothetical polysilane c0mpound.4~~ The polysilane compound consists of

planar Si layers stacked in the (1 11) direction and terminated by hydrogen.
Replacement of 50% of all hydrogen in polysilane by OH groups gives
siloxene, a compound suggested to be responsible for visible luminescence
in porous silic0n.4~~ These calculations are discussed below. Calculations on
BN sheets have been mentioned in connection with the quasiparticle
calculations of solid C,, (see Section IV.19c).

489 C. G. Van de Walle and J. E. Northrup, P h p . Rev. Lett. 70, 1116 (1993).
490 M. S. Brandt, H. D. Fuchs, M. Stutzmann, J. Weber, and M. Cardona, Sol. State Comm.
81, 307 (1992).
162 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Duns-Polyacetylene. In trans-(CH),,488 GWA leads to a 68% reduction

of the initial Hartree-Fock band gap of 5.83 eV to predict a gap of 1.86 eV,
in good agreement with the experimental range of gap energies from 1.4 to
1.8 eV (see Refs. 491, 492, 493, 494, 495, 496). This statement holds only if
quasiparticle energies are determined self-consistently. Self-consistency in
the quasiparticle energies is necessary to achieve quantitative agreement
with experiment, since the first iteration reduces the Hartree-Fock gap only
to 3.1 eV.488 The theoretical electron-energy-loss spectrum shows a low-
energy plasmon peak whose location and linear dispersion with respect to
A ~second
the reciprocal wave vector k agree well with e ~ p e r i m e n t . ~ ~ ,~~',~~~
plasmon peak at 13.2eV deviates significantly from the experimental peak
at 22.5eV. This discrepancy may be due to the limited Gaussian basis set
used in Ref. 488, which cannot describe high-energy states accurately, and
to experimental difficulties related to the production of high-quality crystal-
line t r a n ~ - ( C H ) , . ~ ~ ~
Polysilane. A two Si-layer polysilane compound489 exhibits strong
quantum confinement effects with strong quasi-direct optical transitions.
The direct GWA gap at r equals 3.0eV and is only 0.2eV larger than the
indirect gap. Self-energy corrections of 0.9 eV are a substantial fraction of
the band gap. Siloxene is formed by substitution of OH for H for 50% of
all hydrogen atoms in the planar polysilane. Siloxene has a direct quasipar-
ticle gap of 1.7 f 0.3 eV489with strong optical transition^.^^' The siloxene
band gap corresponds reasonably well to emission in the red, which
emission is observed in porous silicon.498

491 C. R. Fincher, Jr., M. Ozaki, M. Tanaka, D. Peebles, L. Lauchlan, A. J. Heeger, and A. G.

MacDiarmid, Phys. Rev. B20, 1589 (1979).
492 J. Fink and G. Leising, Phys. Rev. B34, 5320 (1986).

493 H. Fritzsche, N. Niicker, B. Scheerer, and J. Fink, Synth. Met. 28,D237 (1989).
494 D. Moses, A. Feldblum, E. Ehrenfreund, A. J. Heeger, T.-C. Chung, and A. G. MacDiar-
mid, Phys. Rev. B26,3361 (1982).
495 J. J. Ritsko, E. J. Mele, A. J. Heeger, A. G. MacDiarmid, and M. Ozaki, Phys. Rev. Letr.
44, 1351 (1980).
496 H. Zscheile, R. Griindler, U. Dahms, J. Frohner, and G. Lehmann, Phys. Stat. Solidi. B 121,

K161 (1984).
497 The same quasiparticle correction is assumed to be valid for the polysilane and the siloxene
compound based on the similarity of the highest occupied and lowest unoccupied state in the
two materials. This approximation is crude, since quasiparticle shifts are determined by a sum
over a range of occupied and unoccupied states.
498 For recent reviews on porous silicon, see, for instance, P. D. J. Calcott, Mat. Sci. Eng. B
51, 132 (1998); A. G. Cullis, J. Appl. Phys. 82,909 (1997); P.M. Fauchet, J. Lumin.70, 294
(1996); B. Hamilton, Semicond. Sci. Technol. 10,1187 (1995); G. C. John and V. A. Singh, Phys.
Rep. 263, 93 (1995); and M. H. Ludwig, Crit. Rev. Sol. State Mat. Sci. 21, 265 (1996).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 163

V. Metals

Quasiparticle calculations for metals have been reported for bulk (Section
V.26), clusters (Section V.27), and surfaces (Section V.28), and are reviewed
in detail in Ref. 14. This section concentrates on a few important aspects
such as the importance of core-valence exchange for the energy-dependence
of the self-energy, the band width of alkali metals, and quasiparticle
corrections in Ni. The reader is referred to Ref. 14 for a more detailed
discussion of some of the quasiparticle applications. Regarding surfaces in
particular, we discuss only the recent work of White et ~ . A1~ (111)~
1 on an ~
metal surface since the extensive work on jellium surfaces was reviewed in
Ref.’ 14.

26. BULK

a Core- Valance Exchange

Inverse photoemission spectra of transition and noble metals for energies up
to 70eV above the Fermi energy E , show peaks that can be identified with
similar structures in a LDA density of states.500 However, for energies
larger than about 10 eV the theoretical peak positions underestimate experi-
ment by an amount that grows linearly with energy (prefactor of about 0.05
to 0.10; Ref. 500). Since GWA self-energy corrections within the jellium
model decrease as the inverse square root of the energy E, rather than
increase linearly with E, improvements beyond the jellium model must be
considered. ’ s 5

The nonlocal exchange interaction between valence electrons (e.g., 5s

electrons in Ag) and electrons in closed shells (e.g., the 4s, 4p, and 4d
electrons in Ag) is neglected in jellium-based GWA calculations, even
though it contributes about 1- 10eV to the conduction-band electron
self-energy for the Ag 5s electron,216 as shown in Fig. 41. As shown in
both Fig. 41 and Fig. 42, nonlocal exchange of the Ag 5s electron with
the Ag 4s, 4p, and 4d electrons216is essential to obtain qualitative agree-
ment with experiment and leads to a self-energy that (1) is approximately
constant for E < lOeV, (2) increases linearly with energy for
lOeV < E < 70eV, and (3) decreases as E - ’ for E 2 70eV.’16 More than

499 I. D. White, R. W. Godby, M. M. Rieger, and R. J. Needs, Phys. Rev. Lett. 80,4265(1998).
500 W. Speier, R. Zeller, and J. C. Fuggle, Phys. Rev. B32, 3597 (1985).
501 G. Materlik, J. E. Miiller, and J. W. Wilkins, Phys. Rev. Left. 50, 267 (1983).
 164 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

c -
0

-2

0 20 40 00 80 100
Energy (eV) '

FIG.41. Self-energy contribution to the energy of the Ag 5s photoelectron arising from the
exchange interaction with electrons in the filled valence shell (n = 4) for different values of the
angular momentum 1 (dashed lines). Comparison of the sum of the s, p, and d contributions
(dashed-dotted line) with the solution that takes the effect of dynamic screening into account
(solid line) shows that exchange is dominant. Energies are measured from the Fermi level.
(From Ref. 216.)

qualitative or semi-quantitative agreement with experiment cannot be

expected since Ref. 216 models (1) dielectric screening by fitting the
(2) the disper-
experimental electron energy loss function Im[l/e(q = 0; o)];
sion of plasmon poles; and (3) the 4s, 4p, and 4d electrons using atomic
Roothaan-Hartree-Fock wave function^.^^^*^^^

b. Alkali Metals
The weakness of the effective crystal potential in alkali metals seems to
permit a nearly-free-electr~n-model~~~ description of the conduction elec-
trons and seems to offer an experimentally accessible system whose many-
body corrections are weak, allowing the use of perturbation theory such as
GWA, and not complicated by the effects of a complex band structure.

'02 E. Clementi and C. Roetti, Atomic Data and Nuclear Data Tables 14, 177 (1974).
Figure 42 shows that screening of the nonlocal exchange part leads to moderate quanti-
tative changes in the results, as does the inclusion of lifetime effects, that is, the imaginary part
of the selfenergy.
N. W. Ashcroft and N. D. Mermin, Solid Srare Physics, Holt, Rinehart, and Winston, New
York (1976) 2858.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 165

0 50 100 150
Energy (eV)
FIG.42. Self-energy of a 5s photoelectron in Ag relative to the Fermi energy as a func-
tion of energy. The different curves are for (1) the bare exchange due to the 5s conduction
electrons (dotted line), (2) the screened exchange due to the 5s conduction electrons (dashed
line), (3) the screened exchange due to the 5s conduction electrons and including the effects of
a finite imaginary part of the self-energy (dash-dotted line), (4) same as in (3) but adding the
bare exchange with the core electrons (solid line a), and ( 5 ) same as in (3) but adding the
screened exchange with the core electrons (solid line b). Experimental data from Ref. 500 are
given as squares. Inclusion of exchange with the core electrons changes the energy dependence
of the self-energy qualitatively. Screening and a nonzero imaginary part of the self-energy affect
the results quantitatively. (From Ref. 216.)

However, the experimental determination of the occupied band width Aw of

Na via photoemission505~506 contradicts nearly-free-electron theory in two
important ways: (1) the measured band width is about 0.6 eV smaller than
the nearly-free-electron value of 3.2 eV, and (2) sharp, nondispersive peaks
in the energy gap can be identified for photon energies of about 35eV,
which peaks have no equivalent in nearly-free-electron theory. Self-
energy corrections in GWA for jellium at rs = 3.95 (average density of Na)
narrow the band width by only 0.3 eV, raising concern about the validity of
GWA for this simple, weakly correlated solid. Several improvements and
extensions to GWA are considered in the literature that all reproduce
experimentally observed Aw’s but correspond to different physics. Differen-
ces between these approaches were discussed in Section 11.14. Here, we give
a short summary of the basic assumptions and results of the different
techniques.

E. Jensen and E. W. Plummer, Phys. Rev. Len. 55, 1912 (1985).

L-W. Lyo and E. W. Plummer, Phys. Rev. Leu. 60,1558 (1988).
166 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN w. WILKINS
Self-energy effects and surface-induced momentum-nonconserving excita-
tions each account for about half of the observed Na band-width correc-
tion.Zz9*z41~26’ The latter effect also provides a possible explanation for the
experimentally observed nondispersive peaks around 35 eV in the photo-
emission spectrum shown in Fig. 43. The inclusion of surface effects is
important, according to the work of Mahan et u1.,229~241~265 since the
mean-free path of the photoelectrons is only 5-6A in Na; that is, photo-
electrons probe only the first two to three layers of the material. Mahan et
ul. used Rayleigh-Schrodinger perturbation theory -suggesting that it is
more accurate than GWA for alkali metals -and the surface potential by
Lang and KohnSo7(see Section 11.14). ,

507 N. D. Lang and W. Kohn, Phys. Rev. B1, 4555 (1970).

I I I I I

a0 26.0 360 46.0 66.0 06.0 76.0

Photon energy (eV)

FIG. 43. Photoemission peak positions as a function of photon energy for the Na (110)
surface. The quasiparticle (QP) bands include the real and imaginary part of the self-energy and
are represented by solid curves. The QP bands are narrower than the nearly-free-electron bands
(NFE, dashed curves) by 0.37 eV. The full theory (solid dots) also includes surface effects and
reproduces the sharp peaks in the “gap” at photon energies of about 35 eV. Experimental data
are denoted by crosses and taken from Ref. 505. (From Ref. 229.)
 QUASIPARTICLE CALCULATIONS IN SOLIDS 167

A “best G, best W ’ approach that goes beyond standard RPA-based

GWA calculations by (1) including LDA vertex corrections in the dielectric
screening but not in the self-energy and (2) determining the quasiparticle
energies self-consistently leads to agreement between theoretical and experi-
mental band ~ i d t h s , ’ ~as~shown . ~ ~ in~ Table
* ~ ~32.~ Point (1) alone leads
to a band-width reduction of 0.57 eV in Na; points (1) and (2) together lead
to a band-width reduction of 0.71 eV, compared to a standard GWA value
of 0.31 eV, and reduce the band width from the LDA value of 3.16eV to
2.45eV.’” This approach does not address the existence of a nearly
dispersionless peak in the photoemission spectrum. Northrup et
u1.184*257.258motivated the inclusion of vertex corrections in the screening
in terms of a better fulfillment of a Ward identity related to charge
conservation.’ s6 Their approach contradicts the arguments of Mahan’ and
others.266
Results of partially self-consistent GWA calculations of the band-width
reduction for j e l l i ~ mcan
~ ~be~ extended to alkali metals and give good

’08 H. J. Levinson, F. Greuter, and E. W. Plummer, Phys. Rev. B27,727 (1983).

R. S. Crisp and S. E. Williams, Philos. Mug. 5, 1205 (1960).
’I0 E. W. Plummer, Phys. Scr. T17,186 (1987).
’I1 The results of Refs. 257 and 184 differ by 70 meV, leading to 2.52 eV for the quasiparticle
band width as used in Table 32.

TABLE32. COMPARISON BETWEEN EXPERIMENTAL AND THEORETICALOCCUPIEDBANDWIDTH

(eV) OBTAINED WITH (1) VERTEXCORRECTIONS IN THE DIELECTRICMATRIX E ONLY AND USEOF
SELF-CONSISTENT QUASIPARTICLE ~ ~ ’PARTIAL
E N E R G I E S ; ~ *(2) ~~ SELF-CONSISTENCY, STATIC
VERTEXCORRECTIONS IN E, AND DYNAMIC VERTEXCORRECTIONS IN THE SELF-ENERGY 2;’” AND
(3) NONZEROIMAGINARY PART OF I: AND SURFACE EFFECTS.229~265 AGREEMENT WITH
EXPERIMENT IS GENERALLY GOOD. HOWEVER,SELF-CONSISTENCY EFFECTSNOT TAKENINTO
ACCOUNT IN THE ABOVECALCULATIONS MAYWIDENTHE Na BANDWIDTHBY AN ESTIMATED
0.4 TO 0.7 eV BASEDON RESULTSOF REF. 231, AND SHIFTTHE THEORETICAL RESULTSTOWARD
THE NEARLY-FREE-ELECTRON AND LDA BANDWIDTHVALUES. THERESULTS FOR REF. 265 ARE
PRESENTED AS AN AVERAGE OVER THE RANGEOF BANDWIDTHSGIVEN. r, IS THE WIGNER-SEITZ
RADIUSOF AN ELECTRON GASWITH IDENTICAL AVERAGEDENSITY AS THE SIMPLE METALS.

NORTHRUP
et Ul.184’258 SHIRLEYz3’ SHUNG et

r, LDA GWA GWA RSPT EXPT.

A1 2.1 10.0 10.2 10.6“

Li 3.3 3.5 2.9 3.1 3.0‘
Na 4.0 3.2 2.5 2.7 2.5 2.65 f 0.05‘
K 4.9 2.3 1.6 1.9 1.5 1.4d

“Ref. 508; ’Ref. 509; ‘Ref. 506; dRef. 510.

 168 WILFRIED G. AULBUR, LARS JoNSSON, AND JOHN W. WILKINS

agreement with experiment if (1) static vertex corrections similar to the

Hubbard local-field factor are included in the dielectric function, and ( 2 )
dynamical vertex corrections to second order in the screened interaction W
are included in the self-energy. In contrast to Mahan,” Shirley suggested
that a dynamical vertex correction to the self-energy in conjunction with
a static vertex correction in the screening is more appropriate’” for
the determination of Aw. Results of Shirley’s c a l ~ u l a t i o n ’are
~ ~ given in
Table 32.
Estimates based on the partially and fully self-consistent GWA calcula-
tions shown in Fig. 19 for the band-width reduction Aw of jellium at rs = 4
suggest that (1) non-self-consistent results for Aw increase by about 0.7 eV
upon inclusion of self-consistency,and (2) partially self-consistent results for
Aw increase by about 0.4 eV upon inclusion of full self-consistency. This
estimate suggests that all published results for Aw miss physical effects that
are as large as the difference between the Jensen-Lyo-Plummer experimen-
tal datasos*s06and nearly-free-electron theory.’I4 Consequently, a quanti-
tative analysis and unified physical understanding of the Na photoemission
experiments require further work, in particular the establishment of a
consistent set of vertex corrections, self-consistency, surface, and lifetime
effects.’ ’
c. d and f Electron Metals
The LDA band structure of the transition metal Ni has four discrepancies
with experiment? (1) The experimental 3d band width of 3.3 eV is about
30% smaller than its LDA value of 4.5eV; (2) the experimental exchange
splitting of 0.25-0.30 eV is about half of the LDA value; ( 3 ) a 6-eV satellite
in the photoemission spectrum is absent in LDA; and (4) the bottom of the
3d band cannot be described by sharp LDA excitations since these states
have quasiparticle widths of about 2 eV, indicating strong interactions
between 3d electrons. Quasiparticle calculations of the d electron metal Ni
and the f electron metal Gd can be found in Refs. 68 and 516, respectively,
and are extensively reviewed in Ref. 14. Here, the discussion is limited to a

’I2 The appropriate choice of vertex corrections is a topic of current discussions (Refs. 232 and
267).
’13 Note that both Shirley and Northrup er al. use plasmon-pole models, which may affect the
accuracy of their calculated valence band widths.
’14 See also Ref. 11 for a critical discussion of the theoretical determination of the alkali metal
bandwidth.
’” Further Applications: Recently, Kubo reported Compton profile studies of Li and Na using
full-potential LMTO (Refs. 287 and 288).
’I6 F. Aryasetiawan and K. Karlsson, Phys. Rev. 854, 5353 (1996).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 169

short description of the successes and failures of GWA in the case of Ni6'
Quasiparticle calculations for Ni lead to a significant improvement of the
LDA band structure and in particular reduce the 3d band width by about
1 eV, in agreement with experiment. Experimental quasiparticle lifetimes are
reproduced as well. Self-energy shifts in Ni are strongly state dependent
because the LDA exchange-correlation potential is a better approximation
for the self-energy of, for example, free-electron-like s states than for the
self-energy of strongly correlated d states. The different character of the
s-like and d-like states is also reflected in the quasiparticle weight, which is
about 0.7 for s-like but only about 0.5 for d-like states.
The 6-eV satellite is not reproduced in GWA, and the exchange splitting
is only marginally improved since both features result from strong hole-hole
interactions that require the inclusion of vertex corrections. A common
explanation of the photoemission process starts with the creation of a 3d
hole, which introduces a strong perturbation because of its localized nature
and excites another 3d electron to an empty state just above the Fermi level.
The two holes scatter repeatedly and form a bound state at 6eV. A
first-principles T-matrix calculation7' includes the neglected hole-hole in-
teraction via ladder diagrams and leads to a 6-eV satellite and an improved
exchange splitting.

27. CLUSTERS
The total energy per atom of, for example, Na clusters as a function of the
number of Na atoms N is a smooth function except for small kinks at
N = 8, 18, 20, 34, and so on (for a review on metal clusters, see, for instance,
Ref. 517). This shell structure is reminiscent of the behavior observed in
nuclei and atoms and results from the fact that the electrons in the alkali
cluster can be reasonably well described by an effective one-particle, spheri-
cal potential. The properties of the metal clusters-such as their total
energy and hence their stability, their ionization potential, and their electron
affinity-change in an abrupt way whenever one shell of electrons is filled
up, leading to an abrupt change as a function of cluster size. The spherical,
one-particle potential of choice is the so-called jellium-sphere-background
model in which the positive ion cores in the cluster are replaced with a
constant-density sphere. The sphere radius is given by T , N ' / ~with , r, being
the Wigner-Seitz radius corresponding to the average metal density.
As shown in Fig. 44, LDA calculations of the ionization potential and
electron affinity do not reflect the experimental shell structure, that is, the
dependence on cluster size of these q u a n t i t i e ~ . ~This
~ ~ *discrepancy
'~~ arises

517 W. A. De Heer, W. D. Knight, M. Y. Chou, and M. L. Cohen, Sol. State Phys. 40,93(1987).
 170 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

21
0 10 20 50 40
Number of atoms per cluster, n

4
J'/>! ............x......)<
X....''

-/
0 1
0 10 20 50 40
Number of atoms per cluster, n

FIG. 44. The absolute value of the quasiparticle energies of the highest-occupied (top) and
lowest-unoccupied (bottom) states in potassium as a function of cluster size in LDA (open
circles), GWA (filled circles), and experiment (triangles, data from Refs. 517, 518, 519, 520, and
521). The crosses represent LDA total-energy differences (ASCF) between positive (KJ) and
neutral (top) clusters or between neutral (K,,) and negative (Kn-, bottom) clusters. The LDA
calculation is not possible for small (n = 2 and 8) clusters. GWA tracks experiment and the
ASCF calculations closely for the highest-occupied states in contrast to LDA. The same is true
for the lowest-unoccupied state although the improvement over LDA is less dramatic. (Adapted
from Ref. 523.)
 QUASIPARTICLE CALCULATIONS IN SOLIDS 171

from LDA’s incomplete cancellation between the Hartree and exchange

self-interaction -the so-called self-interaction error -and increases with
decreasing system size, which conceals the strong size dependence of the
ionization potential in p a r t i c ~ l a r . ~ ’ ~Self-interaction
,~’~ errors lead to an
underbinding of the occupied and an overbinding of the unoccupied states
in strongly localized systems.
The size dependence of the ionization potential and the electron affinity
are well described in GWA because self-interaction errors are small in
standard GWA.524 Remaining quantitative differences with experiment are
attributed to the jellium-sphere-background model and finite temperature
e f f e ~ t s . ~ ’ ’ ~The
~ ’ ~ quasiparticle results for the ionization potential are
systematically above the corresponding LDA results and eliminate the LDA
underbinding of occupied states because of self-interaction errors. Similarly,
the absolute value of the GWA electron affinity is below the corresponding
LDA value since GWA corrects for the overbinding of unoccupied states
within LDA.

28. SURFACES
In classical electrostatics, an electron at a location z outside a metal surface
induces a surface charge and in turn experiences an attractive image poten-
tial ym(z) whose asymptotic form for large z is given as - 1/(4(z - zo)),
where zo is the effective edge of the metal. On a microscopic level, the
rearrangement of charges at the metal surface is due to long-range exchange
and correlation effects, which are absent in LDA because of the exponential,
ather than inverse power-law decay of the LDA exchange correlation
potential outside the metal surface.525This severe LDA failure leads to a
poor description of surface states and to an absence of image states and
resonances in LDA. Discrepancies between LDA results and the results of
surface-sensitive experimental techniques such as low-energy electron dif-

518 P. Fayet, J. P. Wolf, and L. Woste, Phys. Rev. B33, 6792 (1986).
’I9 A. Herrmann, E. Schumacher, and L. Woste, J . Chem. Phys. 68, 2327 (1978).
5 2 0 M. M. Kappes, M. Schar, P. Radi, and E. Schumacher, J . Chem. Phys. 84, 1863 (1986).
5 2 1 K. M. McHugh, J. G. Eaton, G. H. Lee, H. W. Sarkas, L. H. Kidder, J. T. Snodgrass,
M. R. Manaa, and K. H. Bowen, J. Chem. Phys. 91,3792 (1989).
5 2 2 S. Saito, S. B. Zhang, S. G. Louie, and M. L. Cohen, Phys. Rev. B40, 3643 (1989).
5 2 3 S. Saito, S. B. Zhang, S. G. Louie, and M. L. Cohen, J. Phys. Cond. Mar. 2, 9041 (1990).
5 2 4 What self-interaction errors there are in GWA arise from the input wave functions and
energies used for the construction of the self-energy operator. Errors in the wave functions are
small since the overlap between quasiparticle and LDA wave functions is larger than about
95% (Refs. 522 and 523). Errors in the energies can be eliminated by a self-consistent
determination of quasiparticle energies.
5 2 5 N. D. Lang and W. Kohn, Phys. Rev. B7, 3541 (1973).
172 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

fra~tion,’~~*’~’scanning tunneling m i c r o s ~ o p y , ’ ~ and

~ * ’ inverse
~~ and two-
photon photoemission experiment^^^^*'^^ are therefore expected.
The nonlocal, energy-dependent, many-body effects that cause the asym-
ptotic inverse power-law behavior of the image potential are contained in
the GWA self-energy. To obtain a local, state-dependent potential Koc(r)
from the nonlocal self-energy that can be compared with the local image
potential and the LDA exchange correlation potential Vk:A, one defines532

Koc(r) =
s X(r, r’; E)Y(r’)dr’/Y(r).

For an Al(111) surface, Fig. 45 shows that (1) outside the metal surface
(5.1)

&(r) has the correct asymptotic, image-like behavior, and (2) inside the
metal surface self-energy corrections to LDA for states close to the Fermi
energy are small, resulting in nearly identical values of Koc(r) and V:FA(r).
Essential for the above successful applications of the dynamically screened
interaction approximation is a self-consistent evaluation of quasiparticle
states.499The difference between the exponential LDA decay and the GWA
inverse power-law decay leads to an increased weight of the quasiparticle
states in the near-surface region, demonstrated in Fig. 46. Note that results
similar to the ones discussed here for the Al( 111)surface were obtained earlier
than those of Ref. 499 by Deisz, Eguiluz, Hanke, and c ~ l l a b o r a t o r s ~ ~ ~ ~ ’ ~ ~ ~ ’
for the case of a jellium surface. This work was reviewed in Ref. 14.

VI. GWA Calculations and Optical Response

For semiconductors and insulators, the lessons learned from the GWA
regarding the band structure have led to a new level of accuracy in
calculations of optical response. The reason is the crucial role played by the

526 P. J. Jennings and R. 0. Jones, Phys. Rev. B34,6695 (1986).

”’ J. Rundgren and G. Malmstrom, Phys. Rev. Lett. 38, 836 (1977).
528 G. Binnig, N. Garcia, H. Rohrer, J. M. Soler, and F. Flores, Phys. Rev. B 30,4816 (1984).
529 G. Binnig, K. H. Frank, H. Fuchs, N. Garcia, B. Reihe, H. Rohrer, F. Salvan, and
A. R. Williams, Phys. Rev. Lett. 55, 991 (1985).
530 U.Hofer, I. L. Shumay, C. Re&, U. Thomann, W. Wallauer, and T. Fauster, Science 277,
1480 (1997).
5 3 ’ F. Passek and M. Donath, Phys. Rev. L e r r . 69, 1101 (1992).

5 3 2 J. J. Deisz, A. G. Eguiluz, and W. Hanke, Phys. Rev. Lett. 71,2793 (1993).

533 J. Deisz and A. G. Eguiluz, J . Phys. C o d . Mat. 5, A95 (1993).
5 3 4 A. G. Eguiluz, M. Heinrichsmeier, A. Fleszar, and W. Hanke, Phys. Rev. Lett. 68, 1359
(1992).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 173

0.00

-0.10

-0.20

-0.30

I
-0.401 . ’ ” . ’ . ’ ’
-10.0 -5.0 0.05.0 10.0 15.0
Distance outside surface (a.u.)
FIG.45. Surface averaged effective local potential (Eq. (5.1)) at the Al(111) surface compared
with the exponentially decaying LDA exchange correlation potential Vkfl”. The local potential
calculated from the GWA self-energy is virtually identical to Vkfl” in the A1 region, and crosses
over to the classical image form (best fit shown) in the vacuum, in contrast to KtflA.(From
Ref. 499.)

4.00 I 1
I .

-5.0 0.0 5.0 10.0 15.0

Distance outside surface (a.u.)

FIG. 46. Surface-state quasiparticle wave function (full line) of the Al(111) surface at r
1.66 eV below the vacuum level in comparison to its LDA counterpart (dashed line). Since the
local potential resulting from the self-energy decays as l/z, z being the distance from the surface,
rather than exponentially as does the LDA exchange-correlation potential, the quasiparticle
state has weight transferred into the vacuum relative to the LDA state. The quasiparticle state
is obtained by an iterative solution of the energy-dependent quasiparticle equation. (From Ref.
499.)
174 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

band gap in determining the optical polarizability of nonconducting ma-

terials. Efforts to correct the too small LDA band gaps are therefore directly
connected to efforts to accurately calculate optical response: dielectric
constant, optical absorption, second-harmonic generation, and so on. In
Section VI.29, we discuss in detail the overestimate of optical constants in
LDA calculations. In Section VI.30 we introduce the widely used scissors-
operator a p p r o a ~ h , in ~ ~which
~ ~ ~the
~ ’self-energy corrections to the LDA
band structure are approximated by a rigid shift of all the conduction-band
energies while the valence-band energies are unaffected.
The inhomogeneous electron density in crystalline materials generates
electric fields that vary over distances of the order of the lattice constant.
These so-called local fields80,82must be considered when accurate calcula-
tions of optical response are performed. Typically, as shown below, the
local-field effects are of order 5-10% in the dielectric constant and 10-30%
in the second-harmonic coefficient. In Section VL3 1, we outline the theory
of local-field effects in optical response and review the numerical results for
a range of semiconductors and insulators.
Finally, in Section V1.32 we discuss the possibility of obtaining correct optical
response coefficientsfrom density functional theory without the introduction of
quasiparticle corrections. Recent work by Gonze, Ghosez, and GodbyS3’
suggests that previous optical response calculations neglected an important
exchange-correlation effect specific to systems with a band gap. The Kohn-
Sham electrons are acted upon not only by the physical electric field- the
external optical field plus the induced Hartree field -but also by a fictitious
field derived from the induced change in the exchange-correlation potential.
This fictitious exchange-correlation electric field counteracts the unphysical
Kohn-Sham energy spectrum to produce the correct induced density or
polarization.
The discussion below has some overlap with the previous discussion
concerning the calculation of the dielectric matrix for use in GWA calcula-
tions. However, the focus is different. Previously, we were concerned only
with methods to obtain a dielectric matrix suitable for calculations of the
screened Coulomb interaction in the GWA. In particular, a prominent role
was played by the need to perform an integration over all frequencies (see
Eq. (2.24)). In the discussion below, we are concerned with ways to calculate
optical response coefficients with high accuracy (within 5% of experiment).
In particular, we discuss calculations of the linear and nonlinear electric
susceptibilities in the long-wavelength limit.

535 Z. H. Levine and D. C. Allan, Phys. Rev. Lett. 66,41 (1991).

536 Z. H. Levine and D. C. Allan, Phys. Rev. B43,4187 (1991).
537 X. Gonze, P. Ghosez, and R. W. Godby, Phys. Rev. Lett. 14, 4035 (1995).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 175

29. OVERESTIMATION
OF OPTICAL WITHIN DFT
CONSTANTS

In terms of the induced macroscopic polarization density P, the linear and

second-order susceptibilities x and x‘’) are defined by

where E is the total internal electric field. The scalar product symbolizes the
fact that in general the susceptibilities are tensors. However, in many
common materials only a few tensor components are independent -for
crystals with cubic symmetry in particular, x = x,, = xyy = x,, are the only
nonzero components. We use a simple scalar notation for all tensor
properties below, and only when needed do we add explicit functional
dependencies on time, position, frequency, or momentum.
In an insulator with cubic symmetry, the linear susceptibility, in the
independent-particle approximation, can be written as (I/ is the total
volume)538

x=v4 c I(ckIxl~k)l’
c,u,k &ck - &uk
9 (6.2)

where c, u represent summations over conduction and valence bands, k

represents a summation over the first Brillouin zone, and E ~ , ” are the
corresponding single-particle energies. The matrix element is of the position
operator x in some direction in real space.
Any electronic band structure can be used in Eq. (6.2), but the most
common choice is an LDA band structure. Another possible choice would be,
for example, one based on empirical pseudo potential^^^^ for which the
pseudopotential parameters have been adjusted to reproduce an experimental
band structure. Since this section focuses on the influence of GWA self-energy
corrections in optical response, we assume LDA wave functions throughout.
A major problem with LDA calculations of the dielectric constant,
E = 1 + 4nx, is the underestimation of the band gap. It leads to a roughly
equal relative overestimation of x and therefore E, as shown in Fig. 47.
However, in these sp-bonded materials the error due to the use of LDA wave
functions seems less severe. This conclusion is supported by the fact that
GWA calculations give reasonable corrections to the gap without changing
the wave functions dramatically. For higher-order response, the error due
to the too small gap becomes even worse, since for each order the power of

538C. Aversa and J. E. Sipe, Phys. Rev. B52, 14636 (1995).

539M. L. Cohen and J. R. Chelikowski, Electronic Structure and Optical Properties of
Semiconductors, Springer Series in Solid-state Sciences, vol. 75, Springer, New York (1989).
176 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Ge

30 -
Se
00
GaAS

J
x
0It20
-
I
0
a AlAs
9"
F
si
GaP 0
0

0
10 - AIP

0 warn
sic 0
0 OC Bh
GaN
0 0
OO 5 10 1
Direct band gap (eV)

FIG. 47. Percent deviation of the LDA dielectric constants from experiment for 13 semicon-
ductors and insulators. The numerical and experimentaldata are from Refs. 536, 540, 541, 542,
and 543). The calculations use scalar relativistic pseudopotentials.

the energy differences increases by one in the denominator of the formulas

for the response coefficients. Hence, the primary focus in optical-response
theory of semiconductors over the last few decades has been on the best way
to correct the LDA band gap without worrying about changes to the wave
functions. However, before discussing how gap corrections can be introduc-
ed with a reasonable computational effort, we want to discuss in some detail
how large the LDA error actually is.
When the results of a particular publication are considered, it is important
to note exactly how the LDA band structure was obtained. In Table 33, we
 QUASIPARTICLE CALCULATIONS IN SOLIDS 177

TABLE33. DIRECT LDA BANDGAP (IN eV) AT r WITH DIFFERENT LEVELSOF APPROXIMATION:
NR -NONRELATIVISTIC; SR-PP -SCALARRELATIVISTIC PSEUDOPOTENTIAL(WITH ATOMIC 3d
CORE);FR-PP- FULLY RELATIVISTICPSEUDOPOTENTIAL;
FR-CR -FULLY RELATIVISTICWITH
CORE RELAXATION. THE TAFIULATEDVALUES REFLECT MEDIANSOF A MULTITUDE OF PUBLISHED
VALUES.TYPICALLY, A PARTICULAR REFERENCEWILL HAVEVALUESWITHIN0.15eV OF THE
TABULATED ONES.HOWEVER,DESPITEVARIATIONS IN ABSOLUTEVALUESTHE RELATIVETRENDS
AGREEBETWEEN DEWRENT REFERENCES.

E,,(eV)

NR SR-PP FR-PP FR-CR EXPT.

Si 2.7a*b 2.Wd 2.W' 2.6'~' 3.4"

Ge O.Tb 00 -0.1'*' -0.3'" 0.9"
GaAs l.lo*b 0.6d.g.h.i 0 5c.e.g.h 0.2e.J.#.l-" 1.5"

"Ref. 544; bRef, 545; 'Ref. 114; dRef. 185; 'Ref. 43; 'Ref. 546; ORef. 192; "Ref. 121; 'Ref. 541; 'Ref.
547; 'Ref. 548; 'Ref. 549; "Ref. 550; "Ref. 551.

show for Si, Ge, and GaAs how different approximations affect the LDA gap.
The table shows, from left to right, the importance of scalar relativistic effects
with frozen core states, spin-orbit interactions (fully relativistic),and relaxation
and hybridization of core states. In GaAs, all these effects are important if we
aim for an accuracy of 0.1 eV for band-gap predictions. With a 1.5-eV band
gap, a 0.1-eV error corresponds to a 5-10% error in the dielectric constant.
The commonly used scalar relativistic pseudopotentials typically give a gap of
0.4-0.5 eV in GaAs (including a 0.1-eV spin-orbit correction), which is about
0.3 eV larger than a fully relativistic all-electron calculation. In contrast to
GaAs, the band gap of silicon is not sensitive to the level of approximation,
which shows that silicon is not a good test case for the evaluation of accurate
methods. Note also the large effects in Table 33 for Ge and GaAs from the
neglect of relativistic effects even for these relatively light atoms.

'40 J. Chen, Z. H. Levine, and J. W. Wilkins, Phys. Rev. B50, 11514 (1994).
541 Z. H. Levine and D. C. Allan, Phys. Rev. 844, 12781 (1991).
542 J. Chen, Z. H. Levine, and J. W. Wilkins, Appl. Phys. Lett. 66, 1129 (1995).
'43 H. Zhong, Z. H.Levine, D. C. Allan, and J. W. Wilkins, Phys. Rev. 848, 1384 (1993).
'44 C. S. Wang and B. M. Klein, Phys. Rev. B24,3393 (1981).

"' M.-Z. Huang and W. Y.Ching, Phys. Rev. B 47, 9449 (1993).
546 M. Alouani and J. Wills, Phys. Rev. B 54, 2480 (1996).
547 S. Bei der Kellen and A. J. Freeman, Phys. Rev. B54, 11187 (1996).

'48 S.-H. Wei and A. Zunger, Phys. Rev. B39, 3279 (1989).
549 H. Krakauer, S.-H. Wei, B. M. Klein, and C. S . Wang, Bull. Am. Phys. SOC.29,391 (1984).
''O B. I. Min, S. M d d a , and A. J. Freeman, Phys. Rev. 838, 1970 (1988).
'" Crystd and Wid State Physics, Landholt-Bornstein, Numerical Data and Functional
Relation&@ in Sciawv and Technology, vol. 17% ed. 0. Madelung, Springer, Berlin (1984).
178 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

This emphasis on the level of approximations made is equally important

when the general accuracy of the GWA is evaluated. Table 34 compares the
underestimation of the most accurate LDA gaps (the difference between the
last two columns in Table 33) with the gap corrections from various
approximate GWA schemes: first-order perturbation theory, self-consistent-
ly updated band energies, and inclusion of core polarization. The most
accurate values (last column) show that GWA corrections including core-
polarization effects added onto fully relativistic, all-electron LDA energies
do indeed come within 0.1eV of experiment. However, other levels of
approximation yield different conclusions. For example, pseudopotential
calculations that neglect spin-orbit interactions (Table 33, column 2) com-
bined with perturbative GWA corrections (Table 34, column 2), which are
commonly used, typically underestimate the band gap by 0.1-0.3 eV for Si,
Ge, and GaAs.
The use of LDA theoretical lattice constants calculated within a
pseudopotential approximation can yield gaps half an eV larger than the
gaps obtained at the experimental lattice constant. For example, one
calculation552found a gap of 1.0eV in GaAs with a theoretical lattice
constant obtained with a scalar relativistic pseudopotential. That approach
gave a 1.7% too small lattice constant (5.55A), which generated a 0.5-eV
increase in the LDA band gap compared to calculations at the experimental
lattice constant (5.65A). Similarly, F i ~ r e n t i n found
i ~ ~ ~that the band gaps
at the theoretical lattice constant from pseudopotential calculations are
increased by 0.02 eV in Si, 0.45 eV in Ge, 0.61 eV in GaAs, and 0.27 eV in

’” A. Dal Corso, F. Mauri, and A. Rubio, Phys. Rev. B53, 15638 (1996).
’” V. Fiorentini, Phys. Rev. B46, 2086 (1992).

TABLE 34. GWA CORRECTION (INev) TO THE LDA DIRECT GAPAT

r WITH DIR+”T LEVELSOF APPROXIMATIONIN THE GWA
CALCULATION: PT-PERTURBATION THEORY WITH LDA BAND
STRUCTURE; UE -SELF-CONSISTBNTLY UPDATED BANDENERGIES;
CP -INCLUSION OF CORE-POLARIZATION EFFECTS.THE FIRST
COLUMNIS THE LDA GAP ERRORGIVENBY THE DIFFERENCE
BETWEEN THE LASTTwo COLUMNS IN TABLE 33. THEGWA VALUES
ARE FROM TABLES21,22, AND 23 IN SECTION111.18.
AGWA
4 , P 8.P

LDA PT UE CP
Si -0.8 0.7 0.7 0.7
Ge - 1.2 0.6 0.8 1.1
GaAs - 1.3 0.7 0.9 1.3
 QUASIPARTICLE CALCULATIONS IN SOLIDS 179

AlAs, compared to gaps obtained at the experimental lattice constants.

Such large overestimations of the LDA band gap can lead to the
conclusion that LDA gaps are good enough without self-energy corrections.
However, such a claim is countered by the fact that fully relativistic LDA
all-electron calculations give lattice constants in much better agreement with
experiment and much smaller gaps. For example, Alouani and Wills546
obtained a theoretical lattice constant in GaAs of 5.62A and a direct gap of
0.29 eV. Therefore, conclusions concerning the need for quasiparticle correc-
tions based on pseudopotential calculations at the theoretical lattice con-
stant must be considered with scepticism. The inclusion of gradient
corrections (GGA) to LDA does not alter this conclusion, although GGA
typically overestimates the lattice ~ o n s t a n t . ’ ~ ~ . ~ ~ ~

30. THE“SCISSORS
OPERATOR”
AND ITS LIMITATIONS

Ideally, one would like to add an energy- and momentum-dependent

self-energy to the denominator in Eq. (6.2), but this is in general too
computationally expensive. An often used simplification, called the “scissors-
operator approach”, is to add a constant, energy- and momentum-indepen-
dent shift A to the conduction-band energies, leaving the valence energies
and all wave functions ~ n c h a n g e d .That
~ ~ ~is,. a~ term
~~

is added to the LDA Hamiltonian. In the next section, we investigate the

consequences of doing this. But we first discuss what value to choose for A.
The discussion above concerning the band-gap problem suggests that one
should choose A = - EkF. However, such a choice often leads to an

overcorrection of the too high LDA dielectric constant. This overcorrection can
be understood by considering the band and k-space summations in Eq. (6.2).
The most important regions in k-space when calculating E are those where
valence and conduction bands are approximately parallel, leading to a large
joint density of states. The effect of the joint density of states on the dielectric
constant is best seen when E is written in terms of the imaginary part ~ ~ ( 0 )
via the Kramers-Kronig relation:

554 A. Dal Corso, S. Baroni, R. Resta, Phys. Rev. B49, 5323 (1994).
555 A. Dal Corso, A. Pasquarello, A. Baldereschi, and R. Car, Phys. Rev. B 53, 1180 (1996).
180 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

where

The position matrix elements are the same in LDA and in the scissors-
operator approach, since the wave functions are the same by construction.
The only effect of the scissors shift A in the expression for E ~ ( w )is that A is
added inside the 6 function. As shown in Fig. 39 and in Figs. 48 and 49, for
most semiconductors the dominant structure in &'(a) is not the region
around the band gap but the one that consists of a couple of peaks a few
eV higher up in energy.
To shift the main peaks in ~ ' ( 0to) their experimental positions requires
a A smaller than the gap mismatch. Because of both band-structure and

50

40
n 30
3
YII
w- 20

10

-0
n
2 4 8 8 1 0
Energy [eV]

0 2 4 6 8 1 0
Energy [eV]

FIG. 48. Imaginary part ~ ~ (of 0dielectric

) functions for GaAs and GaP from Ref. 556. The
solid curve was obtained with a FLAPW (full-potential linearized augmented plane-wave) band
structure with a scissors shift chosen to fit the experimental band gap. This choice of scissors
shift puts the dominant peaks in the absorption spectra about 0.5 eV too high compared to
experiment (dashed lines-GaAs, Ref. 197; Gap, Ref. 274).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 181

Energy (OW EnWY (*v)

FIG. 49. Effects of quasiparticle weights and vertex corrections on the imaginary part of the
dielectric constants for diamond and silicon (from Ref. 564): short dashed line-experiment
(diamond, Ref. 197; Si, Ref. 565); dotted line- LDA, long-dashed line-self-energy corrected
without quasiparticle-weight correction; dash-dotted line -self-energy corrected with weight
factor Z , Z , for each matrix element; solid line- self-energy corrected including vertex
corrections. The close agreement between the solid and long-dashed curve shows that the
strong effect of the weight factor Z,Z, is almost canceled by vertex corrections. Since the dotted
LDA curves look almost like solid lines, note that the LDA curves are farthest to the left and
have the highest peak in both materials.

many-electron effects, the mismatch between the LDA and the experimental
regions of strong optical absorption is not related to the band-gap mismatch
in an obvious way. Figure 48 shows the results by Hughes and SipeSs6for
GaAs and Gap. They used a scissors shift to fit the LDA band gaps to the
experimental gaps, which puts the dominant absorption peaks about 0.5 eV
too high. As discussed in Sections II.9b and IV.24, inclusion of excitonic
effects will strongly improve the absorption spectrum in regard to both peak
position and the peak heights (see Refs. 154, 155, 156,474, 475, and 476). A
systematic first-principles argument for what A to choose is therefore hard
to find and the scissors-shift approach must be considered largely empirical.
Early calculations found good agreement with experiment by using LDA
gaps plus an averaged GWA correction for A, which suggests a close to
ab-initio method for optical calculations.535~536 Table 35 shows the results
of Levine and Allan for the dielectric constant of Si, Ge, GaAs, Gap, AlAs,
and Alp. Leaving the discussion of local-field corrections to the next section,
we see that the scissors approximation brings the too large LDA values to
within 5% of experiment. These good results are obtained in large part
because of the particular approximation made to the band structure. For

'" J. P. L. Hughes and J. E. Sipe, Phys. Rev. E 53, 10751 (1996).

 182 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

TABLE35. DIELECTRIC CONSTANTSFOR SMALL-AND MEDIUM-

GAPSEMICONDUCTORS IN LDA AND WITHIN THE SCISSORS-SHIFT
APPROACH WITHOUT(No LF) AND WITH (LF) LOCAL-FIELD
CORRECTIONS.536'54'THE VALUES INCLUDING LOCAL-FIELD
CORRECTIONS ARE ALL WITHIN5% OF EXPERIMENT.
____ ~~ ~

LDA N o LF LF EXPT.

Si 13.5 11.7 11.2 11.4"

Ge 22.0 16.6 16.0 15.3"
GaAs 13.7 11.7 11.2 1O.Sb
GaP 10.4 9.3 8.8 9.0b
AlAs 9.5 8.7 8.1 8.2b
AlP 8.3 7.8 ' 7.2 7.4b

"Ref. 557; bRef. 558.

GaAs, Levine and Allan found good agreement with experiment by using
the average GWA correction of 0.8 eV from the semirelativistic
pseudopotential calculation by Godby, Schluter, and Sham."' This correc-
tion is added to a semirelativistic pseudopotential LDA band structure
(neglecting Ga 3d core relaxation and spin-orbit corrections). Their LDA
gap is 0.5 eV at the experimental lattice constant, which yields a 1.3-eV gap
after the scissors correction. This gap is 0.2eV too small compared to
experiment, but this underestimation of the gap is what is needed to get the
dominant contribution from the strong peaks in the LDA absorption
spectrum to end up close to the experimental peak structures.
For large band-gap materials (E,,, > 4 eV), GWA-based scissors-oper-
ator corrections frequently lead to an underestimation of E that is as large
as the overestimation by LDA. This effect was first seen by Chen et al. for
diamond and Table 36 shows the results of Chen et al. for diamond,
Sic, and GaN. Here the LDA values again overestimate the dielectric
constant but by only 4-6%. The average GWA corrections to the direct
gaps are of the order of 1-2 eV, which when used as scissors corrections give
dielectric constants well below experiment. Similar results were obtained by
Adolph et aLS6l and by Gavrilenko and B e ~ h s t e d t , ~who
~ ' concluded that
to obtain an E in agreement with experiment a shift of OSA,,, is needed in
Si; 0.45A,,, in Sic; and 0.2A,,, in diamond.
Another important consideration is the effect on the optical response of
the quasiparticle weight Z that multiplies the single-particle Green function
(see Eqs. (1.7), (1.8), (2.7), and Section 11.11). Within the independent-

55' H. H. Li, J . Chem. Phys. Re$ Data 9, 561 (1980).

"* B. Monemar, Phys Scripta 3, 193 (1971).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 183

TABLE36. DIELECTRIC CONSTANTS

FOR LARGE-GAPINSULA-
TORS INLDA AND WITHINTHE SCISSORS-SHIFTAPPROACH.
THE C AND Sic VALUESARE FROM REF. 540, AND THE
GaN VALUESARE FROM REFS.542 (LDA) AND 560 (Scrss).
THE SCISSORSSHIITS A, DERIVEDFROM AVERAGED GWA
CORRECTIONS, ARE SHOWN WITHIN PARENTHESIS. THE
GWA-BASEDSCISSORS-SHIFT APPROACHSYSTEMATICALLY
OVERCORRECTS THE SOMEWHAT Too LARGELDA VALUES.

LDA Scrss (A) EXPT.

C 5.9 5.2 (1.8 eV) 5.7"

Sic
GaN
6.9
5.5
6.1 (1.2 eV)
4.8 (1.5 eV)
6.5"
*
5.35 0.20b

"Ref. 182; bRef. 559.

quasiparticle approximation, the susceptibility in Eq. (6.2) should be multi-

plied by the product of the electron and hole quasiparticle weights Z,Z,,
leading to a drastic reduction of about a factor of two in the dielectric
and even larger reductions in higher-order response. In
practice, the multiplication by Z,Z, is seldom made in optical response
calculations, which accounts for their success.
This practice of disregarding the reduction due to quasiparticle weights
has recently been put on a more solid foundation by the calculations of
Bechstedt et ~ 1 . They
' ~ ~ showed that when vertex corrections in the form of
excitonic electron-hole interactions are considered the weight factor Z , Z , is
replaced by a combined weight factor Z,, N 0.9. Hence the correction due
to 2, and 2, is almost canceled by the effect of vertex corrections, which are
also neglected in the GWA-based scissors-operator approach. Figure 49
shows the imaginary part of the dielectric function of diamond and silicon
obtained by Bechstedt et al. from LDA, GWA with and without the Z,Z,
correction, and GWA plus vertex corrections. A similar cancellation be-
tween quasiparticle weight and vertex corrections for interband transitions
in metals was discussed by Beeferman and Ehrenreich.' 5 1

559 A. S. Barker and M. Ilegems, Phys. Rev. B7, 743 (1973).

560 J. Chen, Calculation of Linear and Nonlinear Optical Susceptibilities in Wide Gap Semicon-
ductors, Ph. D. thesis, Ohio State University at Columbus, September 1996.
561 B. Adolph, V. I. Gavrilenko, K. Tenelsen, F. Bechstedt, and R. Del Sole, Phys. Rev. B53,

9797 (1996).
5 6 2 V. I. Gavrilenko and F. Bechstedt, Phys. Rev. B55,4343 (1997).

563 R. Del Sole and R. Girlanda, Phys. Rev. B54, 14376 (1996).
564 F. Bechstedt, K. Tenelsen, B. Adolph, and R. Del Sole, Phys. Rev. Lett. 78, 1528 (1997).
184 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

In summary, to choose A from the gap mismatch leads to corrections too

large in most materials, with increasing error as the band gap increases;The
conclusion is that we lack a first-principles method to choose a correct
scissors shift. However, there is no doubt that some scissors shift is needed,
since the best LDA band structures yield gap values well below the
experimental gaps. Most authors today use a scissors shift in optical
response calculations, but the method of choosing the shift varies, and the
shift is largely an empirical way to include quasiparticle and other many-
electron effects.

EFFECTSIN OPTICAL
31. LOCAL-FIELD RELSPONSE

In semiconductors and insulators, the inhomgeneity of the density distribu-

tion leads to internal electric fields that vary over distances comparable to
the lattice constant. As shown below, such local fields give rise to about 10%
corrections in E and 10-30% corrections in x ( ~ )and
, can lead to dominant
corrections in properties such as the optical rotatory power.
a. Definitions and Gauge
In the notation of Section 11, an external potential Kxt(q;a)of wave vector
q and frequency o yields a total internal electrical potential Kot given by

Kot.c(q; 0 ) = &lo1 K x t h ; 0). (6.6)

Note that only the G = 0 component of the inverse dielectric matrix E;,$
appears on the right-hand side of Eq. (6.6), since the external optical field
has no local-field component.
It is important to distinguish the dielectric matrix EGG,(q; a) from
the dielectric function E(q; a).The dielectric function E(q; a) is defined by
E(q; a)= &(q; a)/K:,t,o(q;a). When local fields are disregarded, we have
E(q; o)= b ( q ; o),but when local-field effects are included, the dielectric
function is given by the inverse of the G, G = 0 component of the inverse
dielectric matrix:

E(% a)= l/CGil(q; 011. (6.7)

For independent electrons in the time-dependent Hartree or random
phase approximation (RPA) the dielectric matrix is that given in Section 11.5
in terms of the independent-particle polarizability Po:
 QUASIPARTICLE CALCULATIONS IN SOLIDS 185

where P&(q; o)is given by the Adler-Wiserso*82formula:

2
(ilexp(i(q + G)*r)Ii')(i'(exp(-i(q + G ) - r ) l i )
Pgd.(q; o)= -
v1
ii'

A(1 -.&,I A 4 -A)

+ o + i6
+

E~ - E ~ , E~-
, E~ - o + i6
which is an extension to include local-field effects of the Ehrenreich and
Cohen formula for the dielectric constant in a solid."'
In the scissors-operator approach, the conduction-band energies in the
denominator of the Adler-Wiser formula for P&(q; o) are shifted by A.
Without local fields, the static susceptibility x, defined by E = 1 + 47q, is
obtained by taking the limit 4 + 0 in Eq. (6.8): x = - limqd0P&(q,O)/q2.
In numerical implementations, two alternative forms of the Adler-Wiser
formalism have been used: (1) the length-gauge formulation of Sipe and
C O W O ~ ~ based ~ ~ onS matrix
~ ~ ~elements
* ~ ~of ~the position operator, which
was used in Eq. (6.2); and (2) the velocity-gauge formulation developed by
Levine and Allan (see Refs. 535, 536, 541, 567, and 568). These two
formulations differ by the way the matrix elements of type
(cklexp(iq.r)luk) are evaluated.
The length-gauge formulation can be obtained by an expansion in q of
the exponential exp(iq-r):

(cklexp(iq * r) luk') = iq. (cklrl uk)6k,k,. (6.10)

The q-independent term in the expansion, (ckluk'), is zero because of the

orthogonality of orbitals from different bands. When Eq. (6.10) is used in
the Adler-Wiser formula, we obtain the expression for x given in Eq. (6.2),
if we identify the x-direction with the direction along q and if we anticipate
that the intraband matrix elements do not contribute when all the band-
indices and k vectors are summed over.
The direct expansion in Eq. (6.10) is valid only for interband matrix
elements, since the position operator is not well defined for intraband matrix
elements. A detailed derivation of Eq. (6.2) requires a more careful treatment
of intraband matrix element^.^^^.^^^ Further, when the Hamiltonian has
nonlocal components, because of pseudopotential terms or self-energy

'" A. D. Papadopoulos and E. Anastassakis, Phys. Rev. 843,5090 (1991).

566 J. E. Sipe and E. Ghahramani, Phys. Rev. B48, 11705 (1993).
567 Z. H. Levine, Phys. Rev. B42, 3567 (1990).
Z. H. Levine, Int. J. Quunt. Chem. S28, 411 (1994).
186 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

corrections, additional complications arise. These points have been dis-

cussed extensively in recent
In the velocity-gauge formulation, the exponential is not expanded in
terms of q. Instead the matrix element (cklexp(iq r)luk),which is taken
between the full Bloch states integrated over all space, is reformulated in
terms of the periodic part of the Bloch states and the integration limited to
the volume R of a unit cell in the crystal:

1 1
- (ck(exp(iq.r)luk) = - (ckluk
V R
+ q)&+,,k* (6.11)

Now the expansion in q can be performed by k - p perturbation theory,

which yields

Instead of a position operator, the velocity operator v k = i[& r] = V k H k

appears, and each matrix element also yields an extra energy denominator.
Hence, in the velocity gauge the formula for x will contain two velocity
matrix elements and three energy denominators.
When the scissors-operator approach is used, it is crucial to recognize the
different origins of the three energy denominators in the velocity gauge.
Only one energy denominator should be shifted by A; the other two should
remain unshifted, and the velocity matrix element should be taken with
respect to the LDA Hamiltonian without the shift. Since the wave functions
+
are unchanged, the overlap (ckluk q) must be the same in LDA and in
the scissors-operator approach. This requires

(6.13)

where Hi = H t D A+ A&lck)(ckl. Eq. (6.13) must be true for all u, c, and

k, so in the velocity gauge the velocity matrix elements in the scissors-shift
approach are given by the LDA velocity matrix elements scaled by
(&uk - &ck + A)/(Euk - %k)*

b. Numerical Results
The numerical results below for local-field effects in semiconductors and
insulators are all based on an extension to the Adler-Wiser expression in Eq.

569 0.Pulci, G. Onida, A. I. Shkrebtii, R. Del Sole, and B. Adolph, Phys. Rev B55,6685 (1997).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 187

(6.9), which extension treats the local fields in the time-dependent local-
density approximation570 rather than in RPA (time-dependent Hartree
approximation). In this extension, exchange and correlation effects in the
induced electron density are calculated self-consistently within LDA. For
further details, see the references cited in the following discussion.
Columns two and three in Table 35 show typical local-field corrections to
the dielectric constant. Invariably a reduction of E by 5-10% is obtained.
These results are consistent with the effect of local-field corrections on the
absorption spectrum shown in Fig. 39. The weight in the dominant peaks
in E ~ ( W ) shifts somewhat to higher frequencies, causing a reduction in E.
In second-order response, the local fields give a larger correction -gen-
erally reducing x(') but occasionally increasing it. Table 37 shows the effect
of both the scissors correction and the local fields on x(') for those of the
previously discussed materials that lack inversion symmetry, which is a
prerequisite for a nonzero second-order response. The LDA values are too
large by as much as a factor of two, with the severest errors for the smaller-
gap materials. The local-field corrections are of the order of 10-20% and
are negative except for the zzz component in Sic, which shows a 13%
increase. Despite the limited experimental data, we can see a trend similar
to that found for &-the scissors-operator approach works well for the
smaller-gap materials, but an overcorrection is obtained for the larger-gap
materials, here exemplified by Sic and GaN.

''O A. Zangwill and P. Soven, Phys. Rev. Lett. 45, 204 (1980).

TABLE37. SECOND-HARMONIC COEFFICIENT d = x(')/2 IN pm/V IN

LDA AND WITHINTHE S c r s s o ~ s - S mAPPROACHWITHOUT (No LF)
AND WITH (LF) LOCAL-FIELD THE SECOND COLUMN
SHOWS WHICH TENSORCOMPONENT IS BEINGTABULATED. TIil3 LOCAL-
FIELDCORRECTION IS TYPICALLY OF ORDER - 15%, BUT IS 4- 13%
(ABSOLUTEVALUE) IN THE s i c ZZZ COMPONENT.

cow. LDA NoLF LF Em.

GaP XYZ 65 46 38 37 f 2"

GaAs XYZ 205 106 95 81 k 5"
AIP XYZ 21 17 14
AlAs XY* 35 24 22
SIC xxz 6.6 4.9 4.4
zzz -4.1 -2.7 -3.1
GaN xxz - 3.2 -27 -2.1 f2.6b
ZZZ 5.4 4.2 3.5 k5.4b

"Ref.571; bRef.572.
188 WILFRIED G . AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Hughes and Sipe556made an important observation: It is not clear that

the experimental values for x(') are actually correct. Hughes and Sipe found
that their theoretical values for the second-harmonic coefficient with a
scissors-correctedband structure were quite far from experimental values in,
for example, GaAs and Gap. However, they found close agreement with
recent measurements of the linear electro-optic coefficient given by
x'*)(o;o,0) in contrast to the second-harmonic coefficient ~'')(2o;o,a).The
second-harmonic (SH) coefficient gives the polarization at 2w induced by a
field at o,while the electro-optic (EO) coefficient gives the polarization at
o from a field at o and a static field. Formally, these two coefficients should
be equal in the static limit, while the experimental values disagree by almost
a factor of two. For example, in GaAs x ~ = 162~ _+ 1 0, ~ m ~/ V ~ ~ ' ~
and xg&xpl=99.8 pm/V,556*574 while some of the theoretical values are
172pm/V ( p s e ~ d o p o t e n t i a l ) ,96.5
~ ~ ~pm/V (FLAPW -full-potential lin-
earized augmented plane wave),556and 104.8 pm/V (LMTO -linearized
muffin-tin orbitals).576All of these calculations were done with scissors-
shifted LDA band structures. Clearly, both more experimental and theoreti-
cal work is called for.
The theory for local-field corrections to nonlinear response was correctly
worked out only recently by Chen et a1.573Earlier work had invariably
assumed that only the linear local fields are important in nonlinear response.
However, Chen et al. showed that the nonlinear local fields are of equal
importance. For example, for second-harmonic generation we must consider
local fields both at the driving frequency o and at the second harmonic 2 0 .
It can also be shown analytically that for scalar potentials the dominant
term of the linear local field is exactly twice the dominant term of the
nonlinear local field, leading to a 50% increase in the local-field correction
compared to previously published results. The typical correction due to
local fields is therefore about - 15% for second-harmonic generation, but
with large variation in strength and an occasional change of sign.
Besides the work of Chen et a1.,573several other authors discuss nonlinear
response in semiconductors, although Chen et al. were the only ones to
consider the full effect of local-field corrections. The plane-wave pseudo-

5 7 1 B. F. Levine and C. G. Bethea, Appl. Phys. Lett 20,272 (1972), as revised by D. A. Roberts,

IEEE J. Quantum Electron. 28, 2057 (1992).

5 7 2 J. Miragliotta, D. K. Wickenden, T. J. Kistenmacher, and W. A. Bryden, J. Opt. Soc. Am.

B 10, 1447 (1993).

573 J. Chen, L. Jonsson, J. W. Wilkins, and Z. H. Levine, Phys. Rev. B56, 1787 (1997).
5 7 4 S. Adachi, GaAs and related Materials: Bulk Semiconducting and Superlattice Properties,

World Scientific, Teaneck, NJ (1994).

5 7 5 Z. H. Levine, Phys. Rev. B49,4532 (1994).
5 7 6 S. N. Rashkeev, W. R. L. Lambrecht, and B. Segall, Phys. Rev. B57, 3905 (1998).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 189

potential method, used by Chen et al., was developed by Levine and Allan
(Refs. 536, 541, 567, and 568), who showed how to include a scissors
shift in the velocity gauge. Their method is mostly applied in the long-
wavelength limit,540*542 although frequency dependence below the gap can
be obtained.575
Sipe and coworker^^^^,^^^ made a detailed comparison between the
velocity-gauge and length-gauge approaches and found that a simpler
formulation than the Levine-Allan approach can be obtained when working
in the length gauge. The length-gauge formulation was used by Hughes and
Sipe for GaAs and and by Hughes, Wang, and Sipe for GaN and
They calculated both the real and imaginary part of x ( ~ for )

frequencies up to well above the absorption threshold within a scissors-

shift approach using an LDA FLAPW (full-potential linearized aug-
mented plane wave) band structure. However, they included no
local-field corrections. The formulation of Aversa and Sipe was also imple-
mented by Rashkeev et al.576using an LMTO (linearized muffin-tin orbital)
band structure. They too studied the frequency-dependent x") and added
BN and S i c to the materials studied by Hughes et al. The article by
Rashkeev et al. also includes a detailed discussion of the strengths and
weaknesses of the scissors-shift approach. Adolph and Bechstedt 78 cal-
culated the frequency-dependent x ( ~ for ) Gap, GaAs, InP, InAs, and S i c
with a plane-wave-pseudopotential method at the theoretical lattice
constants. They used scissors shifts for the 111-V compounds but discussed
a more sophisticated momentum- and band-dependent self-energy shift
in Sic.
The optical response calculations described so far have been done by
sum-over-states methods, but an alternative exists. Dal Corso et ul.552used
the so-called 2n + 1 theorem579 to derive a different and more efficient
numerical method for second-order response within time-dependent density
functional theory. The 2n + 1 theorem in this context expresses second-
order response functions (derivable from third-order derivatives of the total
energy) in terms of first-order changes in the wave function. The calculations
by Dal Corso et al. were performed at theoretical lattice constants far from
the experimental ones, which strongly affected their values. Nevertheless,
their method should lead to the same results as others at the experimental
lattice constants.
The relatively modest local-field corrections we have seen up to this point
are not a general rule. Jonsson et ~ 1calculated
. ~ the~local-field
~ corrections

577 J. P. L. Hughes, Y. Wang, and J. E. Sipe, Phys. Rev. B55, 13630 (1997).
578 B. Adolph and F. Bechstedt, Phys. Rev. B57, 6519 (1998).
579 X. Gonze and J.-P. Vigneron, Phys. Rev. B39, 13120 (1989).
190 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

TABLE 38. LOCAL-FIELD(LF) CORRECTIONS FOR SELENIUM AND a-QUARTZ (sic),) IN

DIELECTRICCONSTANT E , SECOND-HARMONICCOEFFICIENT d, AND OPTICAL ROTATORYPOWER
MODESTLOCAL-FIELD
p.543*583 CORRECTIONSFOR E DO NOT IMPLY MODESTCORRECTIONS IN
THE OTHER OPTICALPROPERTIES. NOTE ALSOTHAT FOR BOTH d AND p THE SIGN OF THE
LOCAL-FIELD CORRECTION DIFFERSBETWEEN Se AND QUARTZ.FORTHE SECOND-HARMONIC
COEFFICIENTd, THERE ARE MOREEXPERIMENTAL RESULTS THAN QUOTEDIN THE TABLE.FOR
SELENIUMTHESE EXPERIMENTALVALUESVARY SUBSTANTIALLY, WHILE FOR QUARTZTHE
DIFFERENT VALUES ARE IN GOODAGREEMENT. WE REFER TO REF. 543 FOR A MORE DETAILED
DISCUSSIONAND REFERENCES. THE QUOTEDEXPERIMENTAL VALUES FOR THE ROTATORYPOWER
ARE THE ZERO-FREQUENCY LIMITSGMN IN REF. 583, WHICH WERE EXTRAPOLATED FROM THE
DATAOF REFS 584 AND 585 (Se) AND REF. 586 (QUARTZ).

Se 9.0 7.9 7.3 f 1.1" 78 111 97+25' 21 -55 k56k30

a-quartz 2.42 2.30 2.35b 0.35 0.33 0.34' 0.7 5.6 4.6f0.1

"Ref. 142; *Ref. 580; 'Ref. 581; dRef. 582.

to the optical rotatory power of a-quartz and selenium and found that the
local fields dominate the response. The optical rotatory power is the ability
of crystals with a chiral structure to rotate the plane of polarization of light
passing through. Table 38 shows the effect of local fields on the dielectric
constants, the second-order susceptibilities, and the rotatory powers of
selenium and a-quartz. For E in both materials and for x") in quartz, the
corrections are modest reductions. For x(2) in Se there is a 33% positive
correction. Finally, in the rotatory power the local fields dominate the
response -in a-quartz the local-field correction increases the rotatory
power by a factor of eight, while in Se the local fields change the sign and
more than double the absolute value.
The calculations for the rotatory power cannot be made with the scalar
theory of optical response described above, since the physical effect is a
rotation of polarization and requires a vector description. Therefore, a
generalization to vector response in terms of induced currents and vector

F. F. Martens, Ann. Phys. (Leipzig) 6, 603 (1901), cited in D. E. Gray, American Institute
of Physics Handbook, 2nd ed., McGraw-Hill, New York (1963).
G. W. Day, Appl. Phys. Lett. 18, 347 (1971).
5 8 2 B. F. Levine and C. G. Bethea, Appl. Phys. Lett. 20, 272 (1972).
583 L. Jonsson, Z. H. Levine, and J. W. Wilkins, Phys. Rev. Lett. 76, 1372 (1996).
584 W. Henrion and F. Eckart, Z. Naturforsch. 19A, 1024 (1964).

5 8 5 J. E. Adams and W. Haas, in The Physics of Selenium and Tellurium, ed. W. C. Cooper,
Pergamon, New York (1969), 293.
5 8 6 A. Carvallo, C. R Acad. Sci. 126, 728 (1898), cited in D. E. Gray, American Institute of
Physics Handbook, 2nd ed., McGraw-Hill, New York (1963).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 191

potentials has to be made.543*583*587 H owever, for wavelengths much longer

than the lattice constant one can use a hybrid scheme that treats the slowly
varying fields by vector theory but uses scalar theory for the local fields (see
Refs. 583, 588, 589, 590), thereby greatly simplifying the treatment and
avoiding the use of time-dependent current-density functional theory.

32. DENSITY-POLARIZATION
FUNCTIONAL
THEORY

In the static limit, the induced response to a weak perturbing potential can
be obtained from a ground-state calculation and should therefore be
obtained exactly by Kohn-Sham perturbation theory. This fact has always
been the strong, and just, argument made by opponents to the scissors-
operator approach, which artificially adds a nonlocal potential to a theory
that should be able to give the correct low-frequency optical response by
using an appropriate Kohn-Sham potential. However, there is no doubt that
LDA does not suffice, because of the band-gap problem, and there are
indications that even in exact Kohn-Sham theory a large gap mismatch
exists (see Appendix, Section 3c).

a. Divergence of K,,
A solution to this apparent paradox was recently presented by Gonze,
Ghosez, and G ~ d b y . ’ ~They
’ pointed out that in an infinite sample the
exchange-correlation potential in Kohn-Sham theory can be divergent in the
long-wavelength limit. In a system with a gap, within which a finite electric
field can exist and a finite polarization can be induced, great care has to be
applied when the limit of infinite sample size is taken within density
functional theory. The discovery of Gonze, Ghosez, and Godby has impor-
tant consequences for many situations where bulk polarization occurs (see
Refs. 591, 592, 593, 594, 595, 596, and 597), not least for ferroelectrics, but
here we discuss only the necessary revision of linear response theory.

58’ H. Zhong, Z. H. Levine, D. C. Allan, and J. W. Wilkins, Phys. Rev. Lett. 69, 379 (1992);

70, 1032(E) (1993).

5 8 8 R. Del Sole and E. Fiorino, Sol. State Comm. 38, 169 (1981).

5 8 9 W. L. Mochan and R. G. Barrera, Phys. Rev. B32,4984, (1985); 4989 (1985).

S. T. Chui, H. Ma, R. V. Kasowski, and W. Y. Hsu, Phys. Rev. B47, 6293 (1993).
5 9 1 R. Resta, Phys. Rev. Lett. 77, 2265 (1996).
”* W. G. Aulbur, L. Jonsson, and J. W. Wilkins, Phys. Rev. 854, 8540 (1996).
593 X. Gonze, P. Ghosez, and R. W. Godby, Phys. Rev. Lett. 78, 294 (1997).
594 R. M. Martin and G. Ortiz, Phys. Rev. B56, 1124 (1997).

5 9 5 P. Ghosez, X. Gonze, and R. W. Godby, Phys. Rev. B56, 12811 (1997).

5 9 6 D. Vanderbilt, Phys. Rev. Lett. 79, 3966 (1997).
5 9 7 G. Ortiz, I. Souza, and R. M. Martin, Phys. Rev. Lett. 80, 353 (1998).
192 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Let us first state unequivocally that there are no flaws in the original work
by Hohenberg and Kohn3’ and Kohn and Sham.33Their work addressed
finite systems. The only question is how the limit of infinite extent should
be taken. This limit can be approached in two ways: (1) The infinite-volume
limit can be taken for the finite-size theory in a weak constant field, with
careful consideration of surface effects; or (2) the long-wavelength limit can
be taken in the infinite-sample results. Here we adopt the latter approach
and focus on the long-wavelength limit of optical response in the type of
bulk formulations we use in the previous sections.
In the appendix, we define the exchange-correlation potential Vxc(r)=
&Ex,[n] /6n(r), which is the additional potential besides the physical poten-
tial felt by the Kohn-Sham electrons. By “physical potential” we mean the
electrical potential that would act on a weak test charge inside the sample.
The physical potential consists of the external potential and the Hartree
potential. The fictitious Kohn-Sham electrons are, in addition, acted upon
by the exchange-correlation potential, which by construction is necessary to
yield the same density in the Kohn-Sham system as for the electrons in the
real, interacting system.
In linear response, the exchange-correlation potential can be expanded to
first order in the density change. The exchange-correlation kernel K,,, which
is the response function giving the induced exchange-correlation potential
in terms of the induced density, is given by the second functional derivative
of the exchange-correlation energy:

6V,,(r) =
s Kxc(r,r’)&n(r’)dr’; r’) =
KXE(r,

where no is the ground-state density. For the Kohn-Sham electrons this

(6.14)

induced potential must be added to the induced Hartree potential. The total
induced potential acting on the Kohn-Sham electrons is therefore

(6.15)

When local-field effects are disregarded, we can write in momentum space:

(6.16)

The key point of the work by Gonze, Ghosez, and Godby is that the
exchange-correlation kernel in an insulator should be expected to have a
 QUASIPARTICLECALCULATIONS IN SOLIDS 193

l/q2 divergence just like the Coulomb potential in the Hartree term. This
assertion has been shown to be true in model calculations.595 In a different
language, the l/q2 divergence of K,, means that the Kohn-Sham electrons
feel a macroscopic potential different from the physical electrical potential.
In LDA, K,, goes to a constant as q + 0 and is unimportant compared
to the l/q2 divergence of the Coulomb potential. This finiteness of K,, in
LDA is due to the metallic character of a homogeneous electron gas, which
is used to compute the LDA kernel. Hence, despite its success for many
material properties, LDA completely fails to describe the divergence in K,,
expected in insulators.
The discussion above concerning potentials and density can be recast in
terms of electric fields and polarization. Instead of the long-wavelength
component of the density, we can consider the induced polarization P and
a corresponding exchange-correlation electric field felt by the Kohn-Sham
electrons besides the physical field. In general, for longitudinal, scalar
potentials we can define E = - V V = - iqV and 6n = - V . P = -iq.P.
Then an induced exchange-correlation potential, SV,, = K,,6n a 6n/q2, is
equivalent to an exchange-correlation electric field directly proportional to
the induced polarization, Ex,a P. The latter formulation has given this
version of Kohn-Sham theory for infinite insulators the name “density-
polarization functional theory.”
The problem with the band-gap mismatch between the quasiparticle and
the Kohn-Sham spectrum has now been put in a different light than in our
discussion in the previous sections, where the focus was on how to adjust
the gap. Concerning optical response at long wavelengths, the Kohn-Sham
picture is very different from the quasiparticle picture. The quasiparticles are
driven by the physical electrical field and need a correct band structure to
give results close to experiments. On the other hand, the fictitious Kohn-
Sham electrons inherently will not have the correct band structure and must
therefore be driven by a fictitious additional electric field that exactly
compensates for the spectral differences and ensures that the correct density,
or polarization, is induced as guaranteed by density functional theory in the
static limit.

b. Real Materials
The above discussion demonstrates only a qualitative way in which density-
functional theory can give correct response, but there is no recipe yet that
yields a numerical expression for the exchange-correlation field. However, as
discussed by Aulbur et the fact that this field is proportional to the
induced polarization allows us to gain some important insight about its
strength and qualitative behavior.
194 WILFRIED G . AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Let us define the parameter y to be the assumed constant of propor-

tionality between the exchange-correlation electric field and the polariz-
ation, Ex, = yP.From the definition of the susceptibility, x, P = xE,, where
E, is the physical (Hartree + external) field, we obtain Ex,= yxE,. Further,
in the Kohn-Sham picture we have P = xKs(E, + EJ. That is, the Kohn-
Sham electrons are driven by the physical field plus the exchange-correla-
tion field, and the induced polarization is given by their response function
xKS, which is the independent-particle susceptibility given by the Kohn-
Sham band structure. By construction, the polarization must be the same in
both pictures. This yields

Hence, by elimination of E, from the second equality:

Because of the too small gap in the Kohn-Sham spectrum, we expect

xKS> x and therefore y to be negative. Further from the assumption that
the main difference between the quasiparticle spectrum and the Kohn-Sham
spectrum resides in the energy differences, not the wave functions, we obtain
from Eq. (6.2) y N - A/(xEgap).An induced exchange-correlation potential
approximately proportional to the gap mismatch was also discussed by
Godby and Shams9* in an analysis of band bending at doped Schottky
(metal-semiconductor) interfaces.
Aulbur et ~ 1 . considered
~ ~ ’ the consequences of Eq. (6.18) and its extension
to nonlinear response for most semiconductors and insulators for which data
exist. They showed that y is remarkably material independent and of order
-0.25 in most small- and medium-gap materials, as shown in Fig. 50. For
large-band-gap materials the value for y goes down to about -0.1. These
numbers were obtained by using the experimental values for x and the LDA
susceptibility for xKS.The latter approximation is motivated by the observa-
tion that LDA, being derived from a metallic system, completely ignores any
effect of Ex, but describes ground-state properties well. Hence, it is reasonable
to assume as a first approximation that xLDais close to xKS.
In summary, the Kohn-Sham electrons in density functional theory are
fictitious particles that, besides the physical potential, are acted upon by a
fictitious exchange-correlation potential. Formally, this extra potential de-
rives from the divergence of the exchange-correlation kernel in exact
Kohn-Sham theory for infinite insulators. This divergence is absent in LDA.

598 R. W. Godby and L. J. Sham, Phys. Rev. B 49, 1849 (1994).

 QUASIPARTICLE CALCULATIONS IN SOLIDS 195

0.1 1
\

\ \
\ AIN
\
0.0 \
n
k
\
-0.1
I
<

>2 -0.2 \ \ \

Sic w. >
II
&

-0.3 --- Ge\

seC&- --
-0.4
3 4 5 6
Average bond length (bohr)

FIG. 50. The y parameter defined by E, = yP and approximated by y = I/xLDA- l/xexp,for

11 semiconductors and insulators (from Ref. 592). The solid line is a linear fit to the data. The
broken curves show how much y varies when xLoA is changed by * 5 % . For small- and
medium-gap materials with a bond length above 5 bohr, the value for y is almost material
independent. For larger-gap materials, y gradually becomes lower in absolute value as the bond
length decreases.

All calculations of optical response in semiconductors and insulators to date

have neglected the existence of a macroscopic exchange-correlation poten-
tial. When this extra potential is considered in optical-response calculations,
agreement with experiment can be obtained without adjustments to the
energy spectrum.

VII. Excited States within Density Functional Theory

Within density functional theory, there are several methods to directly

calculate excited states. Some of these methods yield band gaps in semicon-
ductors without need for quasiparticle theory. This section gives a short
description of these alternatives to GWA calculations.
To structure the presentation, we classify the methods according to the
density on which the density functionals are based. For time-independent
196 WILFRIED G. AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

functionals either the ground-state density or the excited-state density can be

used to construct a density functional theory for excited states. These two
cases are discussed in Sections VII.33 and VII.34, respectively. Time-depend-
ent density functional theory, outlined in Section VII.35, can also be used to
find excited-state properties by analysis of various response functions.
Finally, Monte-Carlo methods, as discussed in Section VII.36, can yield
excited states and generate highly accurate densities for the purpose of
detailed analysis of density functional theories.
Most of the work with these new methods has been done for atoms and
small molecules rather than for extended systems. For a review of results for
atoms and molecules, see the recent article by nag^.'^^ Here, we give a brief
description of the ideas behind the methods and focus the discussion on the
band gaps in semiconductors, even though that aspect represents only a
small fraction of the published work.
The material in this section deals with extensions to excited states of
conventional density functional theory, which is constructed to give only the
ground-state density and energy. In the main text, we assume that the reader
is familiar with standard Hohenberg-Kohn and Kohn-Sham theory. How-
ever, for the reader who needs a reminder, we give a short review in the
appendix of the density functional concepts central to the discussion below.

BASEDON GROUND-STATE
33. FUNCTIONALS DENSITIES

The ASCF Method. With a slight extension of the original Hohenberg-

Kohn formulation the theory is valid not only for ground states but for any
state that is the lowest of its symmetry class.600 The excitation energies
between two such states of different symmetry can therefore be calculated as
an energy difference between two separate calculations. This approach is
called the ASCF method,37 since a difference between two self-consistent
energy calculations is taken.
The band gap in a semiconductor, with N electrons in the fully occupied
valence orbitals, is equal to the difference between the energies of the highest
occupied Kohn-Sham orbital in the (N + 1)-electron ground state and the
highest occuped Kohn-Sham orbital in the N-electron ground state:
Egap= E ~ + ~ -, EN,N, ~ + where
~ the notation EN,M refers to the Mth Kohn-
Sham orbital in the N-electron system. The difference between the energies
EN+ l , N + and E N , N + (lowest unoccupied orbital in the N-electron system)
is, by definition, equal to the discontinuity AXc of the Kohn-Sham po-

599 A. Nagy, Phys. Rep. 298, l

(1998).
6oo 0. Gunnarsson and B. I. Lundqvist, Phys. Rev. B 13, 4274 (1976).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 197

tential on addition of an extra electron to the N-electron system:

+
Egap= E N , N + - EN,N A,.. (see the appendix, Section 3).
Since the discontinuity Axc is believed to be of the order eV in semicon-
ductors, practical applications of the ASCF method require a density
functional that captures the discontinuity. Local-density approximations,
including gradient corrections, therefore cannot be used to calculate the
band gap in semiconductors. When a single electron is added to a macro-
scopic system, the density does not change; therefore, local-density approxi-
mations give E ~ + - E ~ , 1.~ +In practice, only nonlocal, orbital-
dependent functionals give rise to a potential discontinuity.
Exact Exchange. Functionals based on an exact treatment of exchange
yield Kohn-Sham potentials that include the exchange part Ax of
Ax. = Ax + Ac. Krieger, Li, and Iafrate (KLI)601 showed how the Kohn-
Sham potential for exact exchange can be calculated in practice in finite
systems. Earlier, Sharp and Horton602 and Talman and Shadwick603
showed in principle how the exact-exchange potential -the so-called opti-
mized effective potential (OEP) -can be derived, but KLI demonstrated
approximations that lead to computationally simpler equations with little
loss of accuracy. Exact-exchange methods (EXX) for extended systems were
developed by Bylander and Kleinman,604~605*606 who used the simplified
KLI version, by Kotani and and by Gorling.608 Stadele et aL609
applied the formally exact formulation of Gorling to several semiconduc-
tors.
Table 39 collects the exact-exchange values for the direct band gaps of Si,
Ge, diamond, and GaAs. For completeness, we also include the correspond-
ing experimental, LDA, GWA, and Hartree-Fock gaps and gaps obtained
from other methods discussed below. The values related to exact-exchange
methods are in rows 6-8: EXX-KS are the Kohn-Sham gaps E N , N + - E ~ , N ;
+
EXX+A, are the ASCF gaps E N , N + ~- EN,N Ax; and KLI-KS are the
Kohn-Sham gaps obtained with the approximation suggested by Krieger,
Li, and Iafrate. The Kohn-Sham gaps of Stadele et al. for Si and C agree
well with experiment, while the KLI values of Bylander and Kleinman and

601 J. B. Krieger, Y. Li, and G. J. Iafrate, Phys. Rev. A 45, 101 (1992); 46,5453 (1992); Y. Li,

J. B. Krieger, and G. J. Iafrate, Phys. Rev. A47, 165 (1993).

6 0 2 R. T. Sharp and G. K. Horton. Phys. Rev. 90, 317 (1953).

6 0 3 J. D. Talman and W. F. Shadwick, Phys. Rev. A 14,36 (1976).

604 D. M. Bylander and L. Kleinman, Phys. Rev. B52, 14566 (1995).

6 0 5 D. M. Bylander and L. Kleinman, Phys. Rev. B54, 7891 (1996).

606 D. M. Bylander and L. Kleinman, Phys. Rev. B55,9432 (1997).
'07 T. Kotani and H. Akai, Phys. Rev. B 5 4 , 16502 (1996).
6 0 8 A. Gorling, Phys. Rev. A 53, 7024 (1996).
M. Stadele, J. A. Majewski, P. Vogl, and A. Gorling, Phys. Rev. Lett. 79, 2089 (1997).
198 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

TABLE39. DIRECTBANDGAPSIN eV AT r FOR DENSITY FUNCTIONAL

METHODS THAT GO BEYOND LDA COMPARED TO EXPERIMENT,GWA,
AND HARTREE-FOCK: EXX-KS-KOHN-SHAM O R B ~ A GAP L WITH
EXACT EXCHANGE; EXX + A,-FuLL BAND GAP WITH EXACT
EXCHANGE INCLUDINGTHE GAP DISCONTINUITYAx; KLI-KS-
KOHN-SHAM GAPWITH THE SIMPLIFTED EXACT-EXCHANGE METHOD OF
KRIEGER, LI, AND IAFRATE;sX-LDA-GENERALIZEDNONLOCAL
KOHN-SHAM FORMULATION WITH SCREENED-EXCHANGE AND LOCAL-
DENSITYAPPROXIMATION OF CORRELATION; FRITSCHE -FmscHE's
GENERALIZED DENSITYFUNCTIONAL THEORYFOR EXCITEDSTATES;
QMC -QUANTUM MONTE-CARLO BAND-STRUCTURE METHOD.FOR
Si, Ge, AND GaAs, THE EXPERIMENTAL AND LDA VALUESARE FROM
COLUMNS 5 AND 4 IN TABLE33, WHILE THE GWA VALUESARE THE suhl
OF THE LDA GAP AND THE GWA CORRECTION IN THE LASTCOLUMN
OF TABLE34. FOR DIAMOND,THE VALUES ARE FROM TABLE 14
COMPLEMENTED WITH THE EXPERIMENTAL VALUE (6.5eV) QUOTED
IN REF. 614 TO ILLUSTRATE THE RATHERLARGEEXPERIMENTAL
UNCERTAINTY. THE HARTREE-FOCK GAPSARE FROM REF. 615 FOR Si,
Ge, AND C, AND FROM REF.616 FOR GaAs.

METHOD Si Ge C GaAs

Expt. 3.4 0.9 6.5 1.5

7.3
LDA 2.6 -0.3 5.5 0.2
GWA 3.3 0.8 7.6 1.5
HF 8.7 7.9 14.6 9.1
EXX-KS 3.3" 1.6' 6.3"
2.9' 5.9'
EXX + Ax 9.6" 15.6"
KLI-KS 2.9' 1.3' 1.9'
sX-LDA 3.4' 0.3' 1.1'
Fritsche 3.7J 6.5'
3.2# 6.4O
QMC 3.Ih
3.9'

"Ref. 609; 'Ref. 607; 'Ref. 604, 'Ref. 605; 'Ref. 610; 'Ref. 611 with
Gaussian exchange-correlation hole; #Ref. 61 1 with Lorentzian ex-
change-correlation hole; hRef. 612- Diffusion MC; 'Ref. 613 -Varia-
tional MC.

the EXX values of Kotani and Akai, who used a nonrelativistic KKR band
structure within the atomic-sphere approximation, deviate by up to 0.7 eV
from experiment. The ASCF gaps, which include the band-gap discontinuity
of the exchange potential, are close to the Hartree-Fock values and
overestimate the experimental gaps by 5-8 eV. The good agreement between
experiment and the EXX Kohn-Sham gaps, together with the large values
of Ax (5- 10eV), imply that Ac is equally large but negative.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 199

Generalized Kohn-Sham Schemes. A different solution to the band-gap

problem was proposed by Seidl et a1.610 The Kohn-Sham equations are
generalized so that more of the gap is given by the Kohn-Sham gap and less
by the discontinuity. Formally, this redistribution is achieved by modifying
the original partition of the total-energy functional into kinetic, Hartree,
external, and exchange-correlation functionals. Part of the exchange-corre-
lation energy is added to the kinetic energy before the functional is defined
through the constrained search approach. In such a formulation, the
Kohn-Sham potential is nonlocal.
One generalized Kohn-Sham theory discussed by Seidl et al. is based on
exact inclusion of exchange. In the original Kohn-Sham formulation, for
nondegenerate ground states, the kinetic-energy functional is defined by
minimizing the kinetic energy operator ? over all Slater determinants @ that
yield the given density n:

To include exchange, Seidl et al. defined a new functional Go that, besides

also includes the Coulomb interaction ?caul, but with the minimum still
taken over Slater determinants @ with N orbitals ai:

where E , is the Hartree energy and E x is the exchange energy:

@: (r)@y (r‘)Qj(r)mi(r’)
drdr’. (7.3)
i<j

Whereas in the standard formulation, as shown in the appendix, the

exchange-correlation energy is defined by E,,[n] = F[n] - E,[n] - Torn],

‘lo A. Seidl, A. Gorling, P. Vogl, J. A. Majewski, and M. Levy, Phys. Rev. B 53,3764 (1996).
‘” L. Fritsche and Y. M. Gu, Phys. Rev. B48,4250 (1993).
‘12 A. J. Williamson, R. Q. Hood, R. J. Needs, and G. Rajagopal, Phys. Rev. B57, 12140
(1998).
‘13 P. R. C. Kent, R. Q. Hood, M. D. Towler, R. J. Needs, and G. Rajagopal, Phys. Rev. B57,
15293 (1998).
‘14 0. Madelung (ed.), Semiconductors-Busic Data, Springer, New York (1996).
‘15 A. Svane, Phys. Rev. B35,5496 (1987).
616 R. Padjen, D. Paquet, and F. Bonnouvrier, Int. J. Quunt. Chem. Symp. 21,45 (1987).
200 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

where F[n] is the Hohenberg-Kohn functional, Seidl et al. defined a new

functional G,[n] = F[n] - Go[n]. By construction, G, is expected to be
dominated by correlation, even though some exchange and kinetic energy
remain in G,.
The “exact”-exchange generalized Kohn-Sham equations are similar to
the Hartree-Fock equations but not identical to them because of the
formally exact treatment of correlations. However, if G , is neglected, the
Hartree-Fock equations are obtained. Note, therefore, that with this gener-
alization the Kohn-Sham equations are nonlocal. We also know that in this
case the Kohn-Sham band gap will typically be 5-10eV larger than
experiment, since this is the result obtained in Hartree-Fock calculations.
However, by using screened exchange, Seidl et al. found better agreement
with experiment, even though the gaps in Ge and GaAs are now half an eV
too small. In screened exchange, an extra factor, exp(-kk,FIr - r’l), is
introduced in the exchange integral in Eq. (7.3), where k,, is the Thomas-
Fermi screening constant for the average density. In Table 39, we have
included the sX-LDA gaps of Seidl et al., which are calculated with the
screened-exchange functional and a local density approximation for G,.
Coupling-Constant Perturbation Theory. The excited-state theory of Gorl-
ing6’ gives all excited-state properties as functionals of the exact ground-state
density of the N-electron system. In the standard formulation of Kohn-Sham
theory, a formal connection between the ground state in the noninteracting
Kohn-Sham system and the exact ground state is obtained if the Coulomb
coupling constant ez is adiabatically switched off. See the appendix, Section 5,
for further details. During the switching, the density is kept fixed by the help of
a coupling-constant-dependent potential chosen so that the density is
independent of the coupling constant and equal to the exact ground-state
density. In the limit of no coupling this external potential is by construction
equal to the Kohn-Sham potential for the noninteracting system.
In the method of GOrling6l7 the adiabatic connection is assumed to exist
not just between the ground states of the Kohn-Sham and the interacting
system, but also between all of the excited states of the two systems.
Perturbation theory to infinite order in the coupling constant then gives all
the excited-state properties of the interacting system as functions of the
complete set of Kohn-Sham orbitals and energies, including the orbitals
unoccupied in the ground state. These orbitals and energies are all deter-
mined by the Kohn-Sham potential and are in that sense functionals of the
exact ground-state density, since the Kohn-Sham potential is uniquely
determined by the ground-state density.
In the Kohn-Sham system, the excited states are given by all possible Slater
determinants that can be formed from the complete set of single-particle

6’7 A. Coding, Phys. Rev. A 54, 3912 (1996).

 QUASIPARTICLE CALCULATIONS IN SOLIDS 20 1

Kohn-Sham orbitals. By construction, the ground-state determinant yields the

correct ground-state density, while the excited Kohn-Sham determinants yield
densities and energies that are not equal to the exact excited-state energies and
densities. However, by the assumed adiabatic connection between the
Kohn-Sham and interacting systems, the exact energies and densities are
obtained as an infinite perturbation series in the coupling constant. For
example, in the coupling-constant expansion of the band gap the zeroth-order
term is the Kohn-Sham gap, the first-order correction is the exchange
contribution, and the higher-order terms are corrections due to correlation. No
calculations for semiconductors using this method have been published to date,
and the convergence properties of the perturbation expansion are not known.

34. FUNCTIONALS
BASED ON EXCITED-STATE
DENSITIES

In the appendix, the universal functional F [ n ] that appears in the Hohen-

berg-Kohn theory as refined by Levy618*619 and Lieb620 is defined by a
constrained minimization of the kinetic and Coulomb energies over all
N-particle wave functions Y that yield the given density n:

That is, that functional is defined for a very large set of densities and, in
particular, for any well-behaved excited-state density. The total energy E [ n ]
is obtained by adding the energy of the interaction with the external
potential: E [ n ] = F [ n ] + 1T/ext(r)n(r)dr.
What is the relationship between the exact energy Ei of the’ ith excited
state and the energy E [ n i ] obtained from the exact density,of the ith excited
state? From the minimization in the constrained search in Eq. (7.4), it
follows that the functional value of E[ni] must be lower or’equal to the exact
excited-state energy: E [ n i ] < E i . E r n , ] is by construction the lowest energy
any wave function can have for the given density ni; however, the wave
function that minimizes E [ n i ] does not have to be equal to the excited-state
wave function, even though they both yield the same density.
Perdew and Levy6” showed that the equal sign in the relation E [ n i ] < Ei
holds if and only if the density ni yields an extremum for E [ n i ] . However,
all excited states do not yield an extremum for ECn,]. Perdew and Levy took

618 M. Levy, Proc. Nut. Acad. Sci. U S A 76, 6062 (1979).

‘I9 M. Levy, Phys. Rev. A 26, 1200 (1982).
‘’O E. H. Lieb, in Physics as Natural Philosophy, eds. A. Shimony and H. Feshbach, MIT Press,
Cambridge (1982); Int. J. Quant. Chem 24, 243 (1983); in Density Functional Methods in
Physics, NATO AS1 Series B123, eds. R. M. Dreizler and J. da Providencia, Plenum, New York
(1985).
J. P. Perdew and M. Levy, Phys. Rev. B31, 6264 (1985).
202 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

as an example two noninteracting electrons in an attractive Coulomb

potential, and showed that, with the ground-state configuration ls’, the
excited-state configuration 2s’ yields an extremum of E [n] while the
configuration ls2s does not. That is, they found E[n,,,] = EZS2and E [ n l s z s ]
< Elszs.There is no mechanism in general to identify which excited state
can be obtained from the ground-state functional, so no practical approach
is obtained from these considerations.
Theophilou and collaborators622~623 demonstrated that excited-state en-
ergies can be obtained from an extension of density functional theory to the
sum of the lowest-lying states. For the ground state, the Rayleigh-Ritz
variational principle states that the ground-state wave function is the one
that minimizes the expectation value of the Hamiltonian. An extension of
this principle says that the sum of the N lowest energies is obtained as the
minimum of the sum of the expectation values of the Hamiltonian fi with
respect to N orthogonal states:

Based on this extended variational principle a density functional theory can be

constructed in complete analogy with the standard Hohenberg-Kohn formu-
lation. The density obtained from this theory is the sum of the densities of the
N lowest eigenstates, and the energy is the sum of the corresponding energies.
The excited-state energy of a given state can be found with this extended
formulation if the energies of all the lower-lying states are known. For
example, to obtain the energy of the first excited state, the ground-state
energy is first calculated and then the sum of the ground-state and
excited-state energies. From these two calculations, the excited-state energy
is obtained by subtraction. This so-called ensemble density functional theory
can be generalized to arbitrary mixtures of states, that is, sums of states with
unequal ~ e i g h t s .Ho ~ wever,
~ ~ * the
~ ~functional
~ itself is different for every
mixture, and it is not clear how to obtain these functionals in practice. This
method has been used for atoms and molecules599but, as far as we know,
not for semiconductors.
Valone and C a p i t a n F used a variational formulation for the oper-
ator (fi - U)’, where U is a free parameter, to construct a functional
RJn, U, V,,,] that, when minimized, yields the exact excited-state energy

622 A. K. Theophilou, J. Phys. C 12, 5419 (1979).

623 N. Hadjisavvas and A. Theophilou, Phys. Rev. A 32, 720 (1985).
624 E. K. U. Gross, L. N. Oliveira, and W. Kohn, Phys. Rev. A 37,2805 (1988).
625 E. K. U. Gross, L. N. Oliveira, and W. Kohn, Phys. Rev. A37, 2809 (1988).
626 S. M. Valone and J. F. Capitani, Phys. Rev. A 23, 2127 (1981).
 QUASIPARTICLECALCULATIONS IN SOLIDS 203

closest to U . A scan of U over a range of energies can then in principle yield

all excited states in that range. However, as indicated in the functional
dependence of R,, this functional is implicitly and nontrivially dependent on
the external potential and is not universal as is the Hohenberg-Kohn
functional. The potential dependence makes this theory impractical for
application to real systems.
F r i t s ~ h e ~ ~ formulated
’.~~* a generalized Kohn-Sham theory for excited
states based on a formal partitioning of any excited state Y i into a sum
Y i = m i + qi, where mi is a Slater determinant that yields the exact
excited-state density ni, while the remainder Pi gives no contribution to ni.
By adiabatic switching of the coupling constant, a mapping between an
excited state in a noninteracting Kohn-Sham system and an excited state in
the interacting system can be obtained. During the switching of the coupling
constant the density is kept fixed to the exact excited-state density of a given
excited state in the interacting system. The Kohn-Sham potential in this
formalism is therefore different for each excited state, and the noninteracting
system is in general in an excited state described by a Slater determinant
with some excited-state orbitals occupied. This Slater determinant is identi-
fied with the wave function mi in the partitioning above, while the second
part qi remains unknown.
The formalism of Fritsche can be used to derive an expression for the
band gap in semiconductors, which for silicon and diamond yields direct
band gaps in good agreement with experiment. However, the expression for
the band gap depends explicitly on the exchange-correlation energy density
and the exchange-correlation potential. To obtain numerical values,
Fritsche and Gu61 modeled the shape of the equal-spin exchange-correla-
tion hole (see the appendix, Section 4), from which the energy density and
potential can be obtained, and neglected the opposite-spin correlation hole.
In Table 39, we have included the gap values for silicon and diamond
obtained by assuming either a Gaussian or Lorentzian shape of the
exchange-correlation hole.

35. TIME-DEPENDENT
DENSITY THEORY
FUNCTIONAL

For finite systems, a practical method for obtaining excited-state energies is

to find the poles in the frequency-dependent linear-response functions.629
Within density functional theory this approach requires an extension of

”’ L. Fritsche, Phys. Rev. B33, 3976 (1986).

L. Fritsche, Physica B 172, 7 (1991).
629 M. Petersilka, U. J. Gossman, and E. K. U. Gross, Phys. Rev. Lett. 76, 1212 (1996).
204 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Kohn-Sham theory to time-dependent external potentials. Runge and

Gross630showed that such a generalization can be obtained under general
circumstances. However, the obtained density functionals depend on the
initial state, and the time dependence of the external potential must be such
that it can be expanded in a Taylor series in time. For a review of
time-dependent density functional theory see, for example, Ref. 63 1.
As for the time-independent case, the exchange and correlation effects in
linear response are contained in an exchange-correlation kernel K,(q; a),
which is now frequency dependent. For numerical applications some ap-
proximation must be chosen for K,,(q; o).A useful kernel at low frequencies
and for slowly varying densities is given by the adiabatic local-density
a p p r o ~ i m a t i o n , ~in’ ~which the frequency dependence is neglected and the
static LDA kernel is used. However, Vignale and K ~ h n ~showed ~ ’ that
local-density approximations fail at finite frequencies. In particular, for
interacting electrons in a harmonic potential exact constraint^^^^.^^^ on the
form of K,,(q; o)are obtained, which constraints can only be satisfied by
nonlocal kernels. Thus, a central issue in time-dependent density functional
theory is to find improved approximations to K,,(q; a).
Many of the theories described in Section VII.33 and VII.34 can be
generalized to include time dependence, each of them corresponding to a
particular approximation to K,(q; o).Examples are the nonadiabatic
local-density kernel of Gross and K ~ h n which , ~ ~violates
~ the harmonic-
potential constraints, and exact-exchange The coupling-con-
stant perturbation theory of Gorling can also be extended to time-depend-
ent potentials, from which theory formally exact correlation can be
systematically included in K,,(q; o).638
For semiconductors, as we saw in Section VI.32 concerning density-
polarization functional theory, the divergence of K,,(q; o)as q + 0 plays a
central role in optical response. This aspect of K,,(q; o)has yet to be fully
understood in the static limit, and nothing is known at present concerning
the finite frequency aspects of this problem. The calculations of optical
coefficients in Section VI.3 1 are obtained by using time-dependent density

630 E. Runge and E. K. U. Gross, Phys. Rev. Leu. 52, 997 (1984).
631 E. K. U. Gross and W. Kohn, Adv. Quantum Chem. 21, 255 (1990).
632 G. Vignale and W. Kohn, Phys. Rev. Lett. 77, 2037 (1996).
6 3 3 J. F. Dobson, Phys. Rev. Lett 73, 2244 (1994).
6 3 4 G. Vignale, Phys. Rev. Lett. 74, 3233 (1995).

6 3 5 E. K. U. Gross and W. Kohn, Phys. Rev. Lett. 55, 2850 (1985).

6 3 6 C. A. Ullrich, U. J. Gossman, and E. K. U. Gross, Phys. Rev. Lett. 74, 872 (1995).
637 S. J. A. van Gisbergen, F. Kootstra, P. R. T. Schipper, 0. V. Gritsenko, J. G. Snijders, and
E. J. Baerends, Phys. Rev. A 57, 2556 (1998).
6 3 8 A. Gorling, Phys. Rev. A 57, 3433 (1998).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 205

functional theory with a static LDA kernel to describe the effect of exchange
and correlation on the local fields. The need for a scissors-operator correc-
tion to obtain a reasonable absorption spectrum demonstrates that this
simple kernel is inadequate for the description of semiconductor gaps.
Vignale and K ~ h n developed
~ ~ ’ a time-dependent formalism based on
functionals of both density and current. They showed that in this formalism
local approximations are more promising. Further, the theory can be cast
in a form where the electron gas is described by a continuum fluid model in
which the viscosity and elasticity coefficients can be obtained from par-
ameters of the homogeneous electron gas.639,640On e of the successes of this
method is a derivation of the damping of collective modes in an in-
homogeneous electron gas. Applications to quantum wells give dampings of
the correct order of magnitude, although these applications also reveal some
fundamental difficulties.641

36. MONTE-CARLO
CALCULATIONS

In quantum Monte-Carlo approaches, statistical sampling methods are used

to calculate matrix elements -variational Monte C a r 1 0 ~ ~ -or
’ to propa-
gate the imaginary-time Schrodinger equation by numerically simulated
diffusion - or Green function Monte C a r 1 0 . ~In~both
~ of these
methods, a real-space many-electron wave-function representation allows
the generation of highly accurate wave functions, including strong correla-
tion effects. In variational Monte Carlo, a parametrized form of the wave
function is chosen and the total energy is then minimized by numerical
integration of the matrix elements of the Hamiltonian. In diffusion Monte
Carlo, the mathematical equivalence between the imaginary-time Schrodin-
ger equation and a real-time diffusion equation is used to propagate the
Schrodinger equation by simulated numerical diffusion of a collection of
fictitious particles (“walkers”) whose density corresponds to the wave
function. Propagation of the imaginary-time Schrodinger equation exponen-
tially projects out the ground state from any initial state that is not
orthogonal to the ground state. In this way, an exact numerical wave
function for the ground state can be obtained from a reasonably accurate
initial guess. Often a variational Monte-Carlo calculation generates an

G. Vignale, C. A. Ullrich, and S. Conti, Phys. Rev. Lett. 79, 4878 (1997).
640 S. Conti, R. Nifosi, and M. P. Tosi, J. Phys. Cond. Mat. 9, L475 (1997).
641 C. A. Ullrich and G. Vignale, Phys. Rev. E58, 15756 (1998).
642 W. L. McMillan, Phys. Rev. 138, A442 (1965).
643 D. Ceperley, G. Chester, and M. Kalos, Phys. Rev. E16, 3081 (1977).
644 M. H. Kalos, Phys. Rev. A 128, 1791 (1962).
206 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

initial wave function for a more accurate, and more computationally

expensive, diffusion Monte-Carlo calculation.
Application of diffusion Monte Carlo is straightforward only for bosons,
while for fermions the antisymmetry of the wave function is difficult to
handle. If the antisymmetry is not strictly enforced, the numerical simulation
will in general lead to a bosonic ground state even for a fermionic initial
wave function. The common way to avoid this problem is to force the wave
function to keep the nodal structure of the initial wave function. This
“fixed-node” approximation355ensures that the obtained wave function is
antisymmetric. However, the exact ground state cannot be found this way,
since it is generally impossible to predict the correct nodal structure. The
wave function found is instead the one with the lowest energy of all states
with the given node structure. In practice, the nodal structure of the initial
wave function is given by a Slater determinant of single-particle orbitals
obtained from, for example, a Hartree-Fock or LDA calculation. Despite the
fixed-node approximation, diffusion Monte-Carlo calculations usually gen-
erate highly accurate wave functions.
Diffusion Monte Carlo can be used to obtain the lowest energy eigenstate
of a given symmetry. Mostly indirect band gaps in solids have been
c o m p ~ t e d , ~ ~ since
’ . ~ the
~ ~difference
~ ~ ~ ’ in total crystal momentum between
the excited state and the ground state ensures their orthogonality. However,
Williamson et d 6 1 2 showed that, although in principle diffusion Monte
Carlo is not useful for calculations of direct gaps, in practice such calcula-
tions can be done within the fixed-node approximation.
To obtain the direct gap at r, for example, the highest valence-band
orbital in the ground-state wave function is replaced by a conduction-band
orbital. Without the fixed-node approximation this initial wave function
would again generate the ground state. However, within the fixed-node
approximation this does not happen, since the nodal structure is not the
same in the initial state and the ground state because of the difference
between the valence- and conduction-band orbitals. Williamson et al.
obtained a 3.7-eV gap in silicon (see Table 39) by implementing this form
of diffusion Monte Carlo.
The fact that Monte-Carlo methods can generate highly accurate densities
makes these methods particularly helpful in analyzing detailed features of
density functional theory. For example, Hood et used variational

645 L. Mitas and R. M. Martin, Phys. Rev. Lett. 72, 2439 (1994).
646 L. Mitas, Comp. Phys. Comm. %, 107 (1996).
647 L. Mitas, Physicu B 237-238, 318 (1997).

648 R. Q. Hood, M. Y. Chou, A. J. Williamson, G. Rajagopal, R. J. Needs, and W. M. C.

Foulkes, Phys. Rev. Lett. 78, 3350 (1997).

 QUASIPARTICLE CALCULATIONS IN SOLIDS 207

Monte Carlo to calculate the exact form of the exchange correlation hole in
silicon, including the dependence on coupling constant. Other examples are
the work by Knorr and Godby 649 and Engel et a1.,650who analyzed the
Kohn-Sham potential for model semiconductors with the help of exact
densities obtained from diffusion Monte-Carlo simulations. Also, Kent et
uL613 used variational Monte Carlo to calculate both density matrices and
the full band structure of silicon (see Table 39 for their value of the gap).

ACKNOWLEDGMENTS

It is a pleasure to acknowledge helpful comments from M. Alouani,

F. Aryasetiawan, J. Chen, A. G. Eguiluz, A. Fleszar, R. W. Godby,
A. Gorling, 0. Gunnarsson, P. Kruger, M. Rohlfing, L. J. Sham, E. L.
Shirley, J. E. Sipe, and U. von Barth. Also, we thank F. Aryasetiawan and
0. Gunnarsson, A. Gorling, and M. Rohlfing, P. Kruger, and J. Pollman
for allowing us to use unpublished data; A. Fleszar, A. Gorling, and
L. Steinbeck for a critical reading of the manuscript; A. G. Eguiluz, B. Farid,
and G. D. Mahan for sending us material prior to publication; and J. Chen
for providing pseudopotentials. WGA gratefully acknowledges continued
collaboration with L. Steinbeck, M. M. Rieger, and R. W. Godby on the
real-space/imaginary-time GWA.

Appendix: Density Functional Theory

This appendix gives a short overview of density functional theory. The

purpose is to remind the reader of those concepts and ideas that are built
upon in the main text of this review. We shall neither attempt a complete
review of density functional theory nor follow a historical sequence of
presentation. For further review see, for example, Refs. 651, 652, 653, 654,
and 655.

649 W. Knorr and R. W. Godby, Phys. Rev. B50, 1779 (1994).

G. E. Engel, Y. Kwon, and R. M. Martin, Phys. Rev. B51, 13538 (1995).
6 5 1 Density Functional Methods in Physics, eds. R. M. Dreizler and J. da Providencia, NATO

AS1 Series B123, Plenum Press, New York (1985).

"' R. M. Dreizler and E. K. U. Gross, Density Funcrional Theory, Springer, New York (1990).
653 Density Functional Theory, eds. E. K. U. Gross and R. M. Dreizler, NATO AS1 Series
B 337, Plenum Press, New York (1995).
6 5 4 R. G . Parr and W. Yang, Density Functional Theory of Atoms and Molecules, Oxford
University Press, New York (1989).
6 5 5 Density Functional Theory of Many-Fermion Systems, ed. S . B. Trickey, Advances in
Quantum Chemistry, vol. 21, Academic Press, San Diego (1990).
208 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

1. UNIVERSAL FUNCTIONALS
DENSITY

The goal of density functional theory, as introduced by Hohenberg and

K ~ h n , ~is' to find the exact ground-state density and total energy of a
system of N interacting electrons in an external potential V,,,. The electrons
are assumed to interact pairwise via the Coulomb interaction. The Hamil-
tonian I? for this system is

where ri is the coordinate for the ith electron. To simplify the notation, we
denote the first term in Eq. (A.l), the kinetic-energy operator, by 2 and the
second term, the Coulomb-interaction operator, by pcOul.These two terms
are the same in all N-electron systems considered; only the external
potential differs between systems. This invariance of the kinetic and
Coulomb operators in the Hamiltonian is the basis for the idea of universal
functionals.
Following the formulation by Levy61* p 6 1 9 and Lieb,620 we define the
density functional,

where the minimization is over all N-electron wave functions Y that yield
the given electron density n(r). That is, F [ n ] is the lowest expectation value,
or energy, of the Hamiltonian ? + pcoulthat can be obtained by any
N-electron wave function, given that the density is equal to n. This
functional is universal in the sense that it does not include any dependence
on the external potential. The total energy functional E [ n ] for a specific
system is obtained by addition of the interaction with the external potential:

The above formulation is valid for an integer number of electrons. However,

the definition of the functionals can be extended to noninteger electron
numbers by the inclusion of mixtures between states with different total
number of electrons in the minimization in Eq. (A.2).40*656For example,

656 N. D. Mermin, Phys. Rev. 137, A1441 (1965).

 QUASIPARTICLE CALCULATIONS IN SOLIDS 209

fractional electron numbers between N and N +

1 are obtained from
+
mixtures of N - and ( N 1)-electron wave functions.
When a fractional total number of electrons is allowed, the ground-state
energy and density are obtained by minimizing E [ n ] under the constraint
that I,n(r)dr = N :

where p is a Lagrange multiplier. In terms of density functional derivatives,

we can therefore write

where n,(r) is the ground-state density. It can further be shown that if N is

a noninteger pN is the chemical potential of the N-electron ~ y s t e m . ~ For
”
integer N , the chemical potential is in general undefined and p N as a function
of N is discontinuous.
According to the theory of Hohenberg and Kohn3’ there is a unique
ground state belonging to each external potential. This one-to-one mapping
between the density and the external potential means that the exact
ground-state wave function is also a functional of the density. A given
density corresponds to a particular external potential, and that potential
gives, for nondegenerate ground states, a unique ground-state wave function
of the Hamiltonian in Eq. (A.1). Since the ground-state wave function is a
functional of the density, so are all expectation values of this wave function.

2. THEKOHN-SHAM
SYSTEM

Kohn and Sham33 showed how a practical computational method can be

constructed from the density functionals of Hohenberg and Kohn. The idea
is to associate the physical system of N interacting electrons with a fictitious
system of N noninteracting electrons. We call this fictitious system the
“Kohn-Sham system” and the noninteracting electrons the “Kohn-Sham
electrons.” The connection between the two systems is that they are defined
to have exactly the same ground-state density and chemical potential.
The external potentials in the two systems are not the same. On the
contrary, the external potential in the Kohn-Sham system -the Kohn-
Sham potential -must be precisely that single-particle potential that yields
a ground-state density for the Kohn-Sham system that is equal to the
ground-state density of the interacting system.
210 WILFRIED G. AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

For nondegenerate ground states, the wave function of the Kohn-Sham

electrons is a Slater determinant of the N lowest orbitals Oi of the
Kohn-Sham equations:

where the notation VKs(no;r) means that V,, is a local single-particle

potential that is an implicit functional of the exact ground-state density no.
The Kohn-Sham equations must be solved self-consistently so that both
equations in Eq. (A.6) are fulfilled.
Formally, the Kohn-Sham potential can be obtained from the functional
F [ n ] . The first step is to define another universal functional To[n]:

where Y is any N-electron wave function that yields the density n. That is,
To is the lowest kinetic energy any wave function can give under the
constraint that it must yield the density n. The total-energy functional of the
Kohn-Sham system is

When the minimization in Eq. (A.7) is extended to include mixed states and
fractional occupation, T,,[n] has a well-defined functional derivative, and we
obtain, in analogy with Eq. (AS),

Comparing Eqs. (A.9) and (AS), we see that VKSmust be defined by

It is customary to split the universal functional F[n] - T,[n] into a Hartree

piece,

(A. 11)
 QUASIPARTICLE CALCULATIONS IN SOLIDS 21 1

and a remainder Ex,[n], which is called the “exchange-correlation energy”:

FCnl - ToCnl = EHCnl + ExcCnI. (A.12)
The Kohn-Sham potential can then be expressed in terms of E,,[n] as

where V, is the ground-state Hartree potential,

(A.14)

and V,, is the so-called exchange-correlation potential,

(A.15)

In systems with a fractional number of electrons, the Kohn-Sham orbitals

have fractional occupation number^.^" Specifically, a continuous change
from the (N - 1)-electron system to the N-electron system is obtained by a
variation between 0 and 1 of the occupation number of the highest occupied
orbital in the N-electron system. One of the most important conclusions
from such a formulation is that the energy E N , N of the highest occupied
Kohn-Sham orbital in the N-electron system (integer N) is equal to the
chemical potential p i :40*658
EN,N = Pi, (A.16)
where p i = pN-&,6 infinitesimal. That is, by the chemical potential for
integer N we mean, by definition, the chemical potential of a system with a
fractional occupation of the Nth Kohn-Sham orbital. The notation E ~ , ~
refers to the energy of the Mth Kohn-Sham orbital in the N-electron system.

3. THEBAND-GAP
DISCONTINUITY

Let us first define the ionization energy I and the electron affinity A for the
N-electron system:
I = E N - 1 - EN; A = EN - EN+l, (A.17)

657J. F. Janak, Phys. Rev. B18, 7165 (1978).

J. P. Perdew, in Density Functional Methou5 in Physics, eds. R. M . Dreizler and J. da
Providencia, NATO AS1 Series B123, Plenum Press, New York (1985), 265.
212 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

where EM denotes the ground-state energy of the M-electron system. In a

macroscopic semiconductor with N electrons in the fully occupied valence
bands, we can associate I with minus the chemical potential p; of the
N-electron system, and A with minus the chemical potential pi+1 of the
(N + 1)-electron system (one (fractional) electron in the lowest conduction
state). Hence, from Eq. (A.16) we obtain:

I = -p-N - - E N,N; A = -pi + I = -EN+~,N+~. (A.18)

is given by the difference between I and A:

The quasiparticle band gap EBap

EBap= I - A = & N + I , N + I - E N , N = A x c + EN,N+I - E N ~ N , (A.19)

where we have defined Axc = cN+l , N + - E N , N + 1, that is, the energy differ-
ence between the highest occupied Kohn-Sham orbital in the (N + 1)-
electron system and the lowest unoccupied Kohfi-Sham orbital in the N-
electron system. This difference is in general finite659*660 and believed to be
of the order of eV in semiconductors and insulators.6 The last two terms in
Eq. (A.19), E N , N + I - E N , N , give the band gap in the N-electron Kohn-Sham
system.
As the notation suggests, Axc is due to the exchange-correlation potential
V,, in Eq. (A.15) and is caused by a discontinuous jump in Vxcon addition
of an electron:659*660v661

where the explicit r dependence indicates that the difference between the
exchange-correlation potentials in the (N + 1)-electron system and the
N-electron system is a position-independentconstant Axe. Note that since the
potential only changes by a constant, there is no difference between the
+
Kohn-Sham wave functions in the (N 1)-electron system and the N-
electron system.
The discontinuity Axc originates solely from the exchange-correlation
potential. The external potential is by construction unchanged; the Hartree
potential has a simple, direct dependence on the density, which only changes
infinitesimally, while the discontinuity in the kinetic energy To is exactly
equal to the Kohn-Sham gap E N , N + - E N , N . This discontinuity in the kinetic
~~~~

659 J. P. Perdew and M. Levy, Phys. Rev. Lett. 51, 1884 (1983).
660 L. J. Sham and M. Schliiter, Phys. Rev. Lett. 51, 1888 (1983).
6 6 1 Strictly, similarly to the chemical potential, the exchange-correlation potentials at integer

N must also be defined by their limits as the electron number approaches N from below in a
system with fractional occupation of the Nth Kohn-Sham orbital.
 QUASIPARTICLE CALCULATIONS IN SOLIDS 213

energy is due to the qualitative difference between the lowest conduction-

band orbital and the highest valence-band orbital. However, the definition
of the kinetic energy in terms of the squared gradient of the orbital means
that, since the orbitals in the (N + 1)-electron system and the N-electron
system are the same, the discontinuity in the kinetic energy is the same
whether we calculate it with the orbitals in the N-electron system or from
+
the lowest conduction-band orbital in the (N 1)-electron system and the
highest valence-band orbital in the N-electron system.662
Numerical implementations of Kohn-Sham theory, using LDA or GGA
(see Section A.6), typically obtain Kohn-Sham band gaps about 0.5-2 eV
smaller than the experimental gaps, see Fig. 3, Section 1.3. Even though
some of this discrepancy can be due to the LDA or GGA approximations,
there is evidence that a significant part is due to Axc.6 This inability of
Kohn-Sham theory to give band gaps close to experiment is often referred
to as the “band-gap problem” and is the main reason that quasiparticle
calculations are needed as a complement to Kohn-Sham theory when
unoccupied bands are considered.

4. THEEXCHANGE-CORRELATION
HOLE

In terms of a normalized N-electron wave function Y(rl, r2,. . . ,r,), we

define the density n(r) by

n(r) =N
I IY(r, rz,. . . ,rN)12dr2... dr,.

Similarly, we can define a pair-distribution function n(r, r’) by

(A.21)

n(r, r’) = N(N - 1)

J” IY(r, r’, r3,. . . ,rN)12dr3... d r N . (A.22)

Note that the total integral over both coordinates in n(r, r’) is equal to twice
the number of unique electron pairs.
The exchange-correlation hole is described by the function g(r, r’), which
is defined through
n(r, r’) = n(r)n(r’)(l + g(r, r’)). (A.23)

Formally, by applying the theory for fractionally occupied orbitals to the noninteracting
Kohn-Sham system, one can prove that the discontinuity in the kinetic energy To as the
electron number passes through N is exactly equal to the Kohn-Sham band gap. The intuitive
argument in the text relying on the infinitesimal difference between the Kohn-Sham orbitals in
the N- and (N + 1)-electron systems is not necessary.
214 WILFRIED G. AULBUR, LARS JONSSON, A N D JOHN W. WILKINS

If we put g = 0, the pair-distribution function is equal to the product of the

densities at the two points r and r’, which implies that the electrons are
uncorrelated. However, g cannot be equal to zero for all r and r’. The
integral of n(r, r’) over r’ yields ( N - l)n(r), while the integral of n(r)n(r’)
yields Nn(r). Hence, we arrive at the important sum rule,

That is, if one electron is taken to be at r, the total number of remaining

electrons in all of space is exactly N - 1, not N . The sum rule for g subtracts
out this one electron from the density distribution n(r’)(l + g(r, r’)) as a
function of r‘ for fixed r.
Though the sum rule prevents g from being zero in all of space, g typically
approaches zero when r and r’ are far apart, which means that the electrons
are uncorrelated at large distances. The exchange-correlation hole is there-
fore often of finite extent, and for a given electron in r it describes the
redistribution of the other electrons in the vicinity of r caused by exchange
and Coulomb repulsion. For example, Ref. 648 presents a Monte-Carlo
calculation of the exchange-correlation hole in silicon and considers the
coupling-constant dependence discussed in the next section.
For electrons with spin 1/2, each spatial coordinate is associated with a
spin index that has two values +(1/2) or ‘‘7” and “1”. The density then has
two nonzero components and the pair-distribution function is a two-by-two
matrix in the spin indices. The sum rule in Eq. (A.24) is thus modified to
n

(A.25)

(A.26)

and the sum rules for the other two g components are obtained by
interchange of 7 and 1 in Eqs. (A.25) and (A.26).
These two relations can be understood by extending the motivation for
the sum rule for spinless particles to fermions with N f and N1 particles of
each spin, respectively. If the particle in r is assumed to have spin up, there
are N T - 1 more particles with spin up in the distribution described by
nf(r’)(l + gTT(r,r’)), which gives Eq. (A.25). However, there are still N1
particles of opposite spin left in the distribution nT(r‘)(l + gT1(r,r‘)), which
gives Eq. (A.26).
 QUASIPARTICLE CALCULATIONS IN SOLIDS 215

The pair-distribution function plays a unique role in an interacting

electron system because the exact Coulomb energy of the system is equal to
the Coulomb energy of the distribution n(r, r’):

From this equation we see directly that g = 0 gives the Hartree energy,
while,the exchange and correlation effects are described by g. For a system
of noninteracting electrons described by a single Slater determinant, there
are no correlation effects, only exchange. The sum rule for g describes the .
important role of exchange to subtract out the self-interaction piece from
the Hartree term, which piece is unphysically included in the Hartree term
when defined in terms of an integral over the total densities. For spin
one-half electrons, gfT describes both exchange and Coulomb correlation
among the equal-spin electrons, and the sum rule for gTTensures that the
self-interaction is subtracted out, while gT1describes Coulomb correlation
among electrons with opposite spin.

5. COUPLING-CONSTANT
AVERAGES

A formal relationship between the physical, interacting electron system and

the noninteracting Kohn-Sham system can be obtained by a coupling-
constant integration in which the electron-electron interaction is multiplied
by a parameter h that is varied between 0 and l.37 In addition-and this
additional constraint is what makes the procedure useful -we require that
the electron density should be the same for all h, which is achieved by
introducing a fictitious single-particle potential K(r).663 The h dependence
of V,(r) is implicitly given by the requirement that nJr) = n,(r) for all h.
The Hamiltonian A, for a given h is

(A.28)

By construction, we must recover the interacting system for h = 1 and the

663 J. Harris and R. 0. Jones, J. Phys. F4, 1170 (1974).

216 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

Kohn-Sham system for h = 0, which implies

v,=,= K,,; v,=, = V K S . (A.29)

Denoting the ground-state wave function for a given h by Y,, the total
ground-state energy E , is

E, = (Ynl~IY,) +~ ~ , l ~ ~ c o u l K +
( r ,b o~ ( W .
l ~ (A.30)

From the Hellmann-Feynman theorem, dE,/dh = (Y,ldH,/dhlY,), we

have

(A.31)

which after integration over h from 0 to 1, together with the relations in Eq.
(A.29), yields

But E L = O= ToCnol + f v VKsWn0(r)dr and E x e l = F[no3 f v Kx,(r)no(r)dr, +

so the single-particle potentials can be eliminated from Eq. (A.32), and we
obtain a relation between the universal functionals F [ n o ] and To[no] at the
ground-state density no:

FCnol = Torn01 + ~ ~ ~ l ~ c o u l l ~ * ~ ~ ~

We now define the average exchange-correlation hole, ij = f: g,dh, and

reach the important conclusion that the exchange-correlation energy E x , in
Kohn-Sham theory is equal to the Coulomb energy of the coupling-
constant-averaged pair-distribution function:

(A.34)

In the interacting system the functional F[nO] is equal to the sum of the
 QUASIPARTICLE CALCULATIONS IN SOLIDS 217

kinetic energy, the Hartree energy, and the Coulomb energy of the ex-
change-correlation hole g, while the Kohn-Sham system is described by the
kinetic energy To[nO],the Hartree energy, and the Coulomb energy of the
averaged exchange-correlation hole @.Hence, the effect of the coupling-
constant average is to incorporate the difference between To and the total
physical kinetic energy into the Kohn-Sham exchange-correlation functional
E,. The Kohn-Sham construction is made specifically to obtain a set of
single-particle equations by which the exact density can be found. However,
the partitioning of F[n] into a sum of To[n],E H [ n ] , and a remainder E,,[n]
is not a physical partitioning, so the exchange-correlation energy contains
both ,kinetic and exchange-correlation energy.

6. LOCALAPPROXIMATIONS

For practical implementations of the Kohn-Sham scheme, an approxi-

mation must be chosen for the exchange-correlation energy E,,[n] or the
exchange-correlation potential 6E,, [n]/6n. The most commonly used ap-
proximation is the local density approximation (LDA),33 in which the local
exchange-correlation energy per electron is approximated by the exchange-
correlation energy per electron of a homogeneous electron gas with the local
density. In general, we can define an exchange-correlation energy per
electron Uxc(r) by

(A.35)

In LDA, it is assumed that U is a function ( N . B . not functional) of the local

density; that is, u L D A ( r ) = ULDA(n(r)).
There are several parametrizations -
for example, Refs. 664 and 665 -for ULDA(n)based on accurate calculations
for the homogeneous electron gas3’ Historically, Slater’s density functional
theory666 based on local exchange preceded the formal foundation in
Kohn-Sham theory. Even earlier, the first de fucto density functional theory
was the Thomas-Fermi equation, in which the kinetic energy is also treated
in a local a p p r ~ x i m a t i o n . ~ ~ ~ . ~ ~ ~
The local density approximation is by construction exact for homogene-
ous systems and is the zeroth-order term in an expansion of ExC[n]in terms

664 S. H. Vosko, L. Wilk, and M. Nusair, Can. J. Phys. 58, 1200 (1980).
665 J. P. Perdew and A. Zunger, Phys. Rev. B23, 5048 (1981).
666 J. C. Slater, Phys. Rev. 81, 385 (1951); ibid. 82, 538 (1951).
667 E. Fermi, Z. Phys. 48, 73 (1928).
668 L. H. Thomas, Proc. Curnb. Phil. SOC.23, 542 (1927).
218 WILFRIED G. AULBUR, LARS JONSSON, AND JOHN W. WILKINS

of the spatial gradients of n. Attempts to include more terms in the gradient

expansion turned out to be difficult. A straightforward gradient expansion
(GEA)33 yields a s" that violates the exchange-correlation-hole sum rule,
which leads to severe problem^.^^^.^^' More sophisticated so-called gener-
alized gradient approximations (GGA) (Refs. 669,670,671, and 672) assume
that U is a function of the density and its gradient Vn-that is, UGGA(r)=
UGGA(n(r), Vn(r)) -but without limiting U G G , to be an actual gradient
expansion of U(r). Instead, the various GGA exchange-correlation energies
that have been proposed are specifically constructed to obey the important
sum rule for J and several other exact relations that it has been shown a
good U(r) should fulfill. However, although GGA often yields better results
than LDA-for example, for atomic total energies and cohesive energies in
crystals, there are several cases in which GGA gives worse results than
LDA -for example, for bulk moduli and phonon f r e q u e n c i e ~ . For ~~~*~~~
lattice constants, both LDA and GGA give about 1% errors, but LDA gives
too small values while GGA gives too large v a l ~ e s . ~ ~ ~ * ~ ~ ~

669 D. C. Langreth and J. P. Perdew, Phys. Rev. B21,5469 (1980).

670 J. P. Perdew, Phys. Rev. 833.8822 (1986); ibid. 34,7406(E) (1986).
671 J. P. Perdew, in Density Fwrerbnul l'kory, eds. E K. U. Gross and R. M. Dreizler, NATO
AS1 Series 9337, Plenum Press, New York (1995), 51.
6 7 2 K.Burke, J. P. Perdew, and M. E m r h d , J. Chem Phys. 109,3760 (1998).
673 1.-H. Lee and R. M. Martin, Phys. Rev. B56,7197 (1997).