Home Search Collections Journals About Contact us My IOPscience

Ab initio tight binding

This article has been downloaded from IOPscience. Please scroll down to see the full text article.

2000 J. Phys.: Condens. Matter 12 R1

(http://iopscience.iop.org/0953-8984/12/2/201)

View the table of contents for this issue, or go to the journal homepage for more

Download details:
IP Address: 130.209.6.50
The article was downloaded on 18/04/2013 at 16:29

Please note that terms and conditions apply.

J. Phys.: Condens. Matter 12 (2000) R1–R24. Printed in the UK PII: S0953-8984(00)98759-9

REVIEW ARTICLE

Ab initio tight binding

A P Horsfield† and A M Bratkovsky‡

† Fujitsu European Centre for Information Technology, 2 Longwalk Road, Stockley Park,
Uxbridge, Middlesex UB11 1AB, UK
‡ Hewlett-Packard Laboratories, 3500 Deer Creek Road, Palo Alto, CA 94304-1392, USA

Received 2 June 1999, in final form 29 October 1999

Abstract. Empirical tight binding has proven to be very popular in recent years on account of
its computational efficiency and accuracy. However, it has limitations, notably the difficulties
associated with fitting parameters and improving models when the desired accuracy cannot be
achieved. In the light of this, a number of efforts have been made to derive tight-binding models
from first principles. Here are described a number of formalisms based on density functional theory
which span the range of approaches currently being used.

1. Introduction

Mathematical models are central to the interpretation of physical phenomena. The great
advantage of computer models is that they can be made very sophisticated, and so describe
rather accurately the phenomena we wish to understand. Indeed, the best quantum chemical
calculations rival experiment in the accuracy they can achieve. Computer simulations can be
carried out over the complete span of length scales from the cosmological to the sub-atomic.
Here we focus on the atomic length scale.
For the overwhelming majority of problems of interest which are best described at the
atomic level, achieving the most accurate account of some phenomenon requires that a balance
must be struck between two competing requirements: a large enough number of non-equivalent
atoms must be considered to remove effects due either to periodic boundaries or cluster surfaces;
the model used to describe the interactions between atoms must be precise enough to include
all the relevant features. The nature of the final compromise depends very sensitively on
the problem being studied, and so there can be no universal method. In this review we will
concentrate on those problems where a few hundred atoms are sufficient to describe the process
being simulated, but where an explicit account of the electrons is needed to describe the
interatomic interactions. It should be pointed out though that the rapid improvement in both
algorithms and performance of hardware are shifting ever higher the maximum number of
atoms that can be treated with the methods described below.

2. The role of quantum mechanical simulations

In the following, a basic understanding of total-energy quantum mechanical methods is

assumed. An introduction to methods appropriate to solids can be found in Ashcroft and
Mermin [1].
All interatomic interactions involve the motion of electrons. For some systems this can
be accounted for in a very simple way leading to simple models. Notable examples based on
0953-8984/00/020001+24$30.00 © 2000 IOP Publishing Ltd R1
R2 A P Horsfield and A M Bratkovsky

perturbation theory include: noble gases where the electrons are perturbed only slightly about
a very stable atomic ground state leading to dispersion forces accurately described by a pair
potential; nearly free-electron metals where most of the total energy can be described by a
uniform electron gas in the potential of pseudo-ions with the small residual interactions being
well described by a pair potential derived from second-order perturbation theory.
There are many other systems where simple perturbation theory is inadequate. A general
class of such systems is where strong covalent bonds are made and broken leading to large
redistributions of electron density. This redistribution leads to complicated non-local changes
in interatomic interactions which are most easily described by treating the electrons explicitly.
One example system is the carbon vacancy in titanium carbide [2]. The removal of the carbon
atom results in charge being distributed preferentially into some bonds over others. The
resultant atomic relaxation can only be understood using a many-centre analysis.
To describe electronic motion we must use quantum mechanics. Since exact solutions can
only be found for a very limited range of problems approximations must be made. Almost
all the effort in practice goes into constructing suitable numerical approximations. Which
approximation is chosen depends strongly on the problem being solved. There are two basic
decisions that always need to be made: the choice of theory (either the many-body Schrödinger
equation or density functional theory); the choice of basis set in terms of which to expand the
wavefunctions. Ab initio tight binding makes use of the Kohn and Sham [3] formulation of
density functional theory. This is chosen on account of the accuracy that has been achieved
consistently with rather simple approximations (notably the local density approximation).
Possible choices of basis set are described below.

3. Empirical tight binding

In order to understand the interest in ab initio tight binding it is necessary to go back to its
precursor, empirical tight binding [4–7]. This is the simplest quantitative quantum mechanical
model. Its simplicity allows analytic results to be produced for a number of systems; thus it
has been used extensively in the past to provide qualitative understanding of a wide range
of electronic phenomena. Recently it has been rediscovered as a quantitative total-energy
method, often being combined with molecular dynamics. The main reasons for this are: it is
a quantum mechanical model and thus allows for a proper description of electronic motion;
it is very simple and thus can be implemented very efficiently; it is a real-space method and
thus can be used with linear scaling algorithms; it is a parametrized model and thus can give
remarkably high accuracy for some systems.
But it also has major limitations: fitting the parameters is often a very lengthy business;
constructing models for systems with more than one type of atom is usually much more
difficult than creating models for a single atom type; when the model breaks down it is very
hard to decide how to improve the model. A number of attempts have been made to produce
more accurate models [8–17], and the success of Hartree–Fock-based empirical methods such
as CNDO [18–22] suggests that for some systems improved empirical methods have value.
However, the fundamental weaknesses remain.

4. Ab initio tight binding

The strengths of empirical tight binding are so clear that they provide a strong motivation to
overcome its limitations. Since the limitations are related to fitting parameters and extending
the model, the natural way to proceed is to derive a tight-binding model from first principles.
 Ab initio tight binding R3

The decisions made during the derivation will define the limitations of the model. If greater
accuracy or speed are required, different decisions can be made. In a first-principles description
every term has a clear definition; thus evaluating terms for mixed systems should be no harder
than for single-element systems.
Empirical tight binding is efficient for several reasons: the basis set is minimal, thus
minimizing the time spent on diagonalizing the Hamiltonian matrix; the integrals are all given
by formulae that are rapid to evaluate; the range of interaction between atoms is short (the
orbitals are localized in space), thus allowing the construction of the Hamiltonian to be carried
out in a time that scales linearly with the number of atoms. We would like to retain these
properties in any first-principles formalism. There is no problem in principle with constructing
a localized minimal basis set. However, in general the integrals will not be able to be represented
by simple functions, but they can be evaluated once, and stored in tables that can be interpolated
later on.
We now take a brief look at the fundamental theory underlying ab initio tight binding,
and then look at a number of practical implementations. A comparison of the methods, using
the isolated vacancy in silicon as an example, is given at the end of the review. However, it
is important to note that the method which is most appropriate will depend strongly on the
application. Thus cited applications of the methods should be referred to in order to determine
which is the best for a given problem.

4.1. The Harris–Foulkes functional

A fundamental decision underpinning every ab initio tight-binding model is to make it non-
self-consistent (though this constraint can be relaxed, as discussed below). The Harris–Foulkes
functional [23–25] (UH F ) is very similar to the Kohn–Sham functional, except that it is defined
entirely in terms of an input charge density (nin ) (whereas the Kohn–Sham functional is defined
in terms of both an input and an output charge density):
X Z
1 nin (Er )nin (Er 0 ) 1 X ZI ZJ
UH F [nin ] = fi εi − dEr dEr 0 +
i
2 |Er − rE0 | 2 I 6=J |REI − REJ |
Z
+ Exc [nin ] − dEr vxc [nin ; rE]nin (Er ) (1)
P
where εi is an eigenvalue of the effective Hamiltonian ĥ = T̂ + I vI + vxc + vH a , fi is the
corresponding single-particle state occupancy, ZI is the charge on ion I , REI is the position
of ion I , Exc is the exchange and correlation functional, vxc [nin ; rE] = δExc [nin ]/δnin (Er ), T̂
is the electron kinetic
R energy operator, vI is the interaction potential for an electron and ion
I , and vH a (Er ) = dEr 0 nin (Er 0 )/|Er − rE0 |. The set of terms following the sum of eigenvalues is
called the double-counting term. It corrects for the fact that part of the potential the electrons
move in is generated by the electrons themselves. The eigenvalues are found by solving the
equation
ĥψi = εi ψi (2)
where ψi is a single-particle wavefunction.
Given the original motivation for considering the electrons explicitly, namely that there
are often cases where large charge transfers occur when bonds are made or broken, we need
to justify the use of a non-self-consistent scheme in which we work with a fixed input charge
density. The justification is that the error in the total energy is second order in the difference
between the input charge density and the self-consistent charge density [26]. Provided the first-
order terms dominate over all others, this is a good approximation. However, the electrostatic
R4 A P Horsfield and A M Bratkovsky

terms are second order in the density, so if there is significant charge transfer leading to long-
ranged internal fields, errors may occur.
This functional has been tested on a wide range of systems, and has been found to
be surprisingly accurate. In particular, Polatoglou and Methfessel [27] looked at the bulk
properties of Be, Al, V, Fe, Si, and NaCl. They found that the bulk modulus and lattice
constant were well described in each case (even in ionic NaCl), though the energy was less
well converged. Finnis [26] found that it is important that the input charge density be contracted
relative to the free-atomic charge density. Using the contracted density he was able to obtain
well converged results even for the surface and vacancy in aluminium.
One notable set of systems where it fails is transition metals [28]. The problem here is
that the electronic configuration in the atom is quite different from that in the solid, even in
the neighbourhood of the core. One cycle of self-consistency greatly improves the results.
Having discussed the general underlying theory, we now look at some specific
implementations.

4.2. Atomic-like orbital-based tight binding

A small number of atomic orbitals are capable of giving good convergence. There are two
factors responsible for this. The first is that, by construction, the orbitals have the correct form
near the ion cores. The second is more general. The correct wavefunctions will minimize
the energy of the system. Thus we can make a linear combination of atomic orbitals, and
adjust the coefficients so as to minimize the energy. This gives us a best-fit approximation
with the property that any errors in the final energy will vary as the square of the error in the
wavefunctions. Thus the energy is rather insensitive to errors in the wavefunction.
There are three formalisms that will be discussed here that begin by expanding the
wavefunctions in terms of atomic orbitals φI α , where α is an orbital index covering principal
and angular quantum numbers. We begin by describing the formalism of Sankey and Niklewski
[29], followed by the variant due to Horsfield [30], and finally the two-centre formalism of
Frauenheim [31].

4.2.1. Formalism of Sankey and Niklewski. Conceptually this method is straightforward. It

is a non-self-consistent linear combination of atomic orbitals (LCAO) method in which the
integrals are evaluated prior to a simulation, and then obtained from tables by interpolation
during a simulation. There are, however, a number of important technical points that make the
method viable, and hence interesting. P
If we expand the wavefunctions as ψi = I α CI(i)α φI α , and define
Z
hI α,Jβ = dEr φI α ĥφJβ
Z (3)
OI α,Jβ = dEr φI α φJβ

then equation (2) can be transformed into the following matrix equation:
X (i)
X (i)
hI α,Jβ CJβ = εi OI α,Jβ CJβ . (4)
Jβ Jβ

The method then consists of a choice of input charge density, basis set and ways to evaluate
and tabulate the integrals given in equation (3) (the Hamiltonian and overlap matrices) and
equation (1) (the double-counting terms).
The orbitals are taken to be atomic-like,
P and the input charge density is taken as a sum of
atomic-like charge densities (Nint (Er ) = I nI (Er ), where nI is the atomic-like charge centred
 Ab initio tight binding R5

on site I ). We have already seen from the work of Finnis [26] that it is important to compress
the charge density relative to the free atom. There are also two problems associated with taking
the orbitals from the free atom. In the first place it makes calculations slow since the orbitals
are long ranged and many neighbours must be considered when constructing the Hamiltonian
and overlap matrices. The second problem is that, according to the virial theorem for particles
that interact with potentials (V ) that vary with distance as 1/r, the electronic kinetic energy
(T̂ ) should increase as atoms bond to form a condensed system (hT̂ i = − 21 hV i). Both these
problems can be overcome very elegantly by taking the orbitals (and also the charge density)
from a confined atom. This puts the atom into a slightly excited state. In the method of Sankey
and Niklewski the atom is confined by forcing the orbitals (but not their derivatives) to go to
zero at some radius (5 Bohr radii for silicon). This is equivalent to confining the atoms in an
infinitely deep spherical square well potential (see figure 1).

s orbital p orbital
0.15

0.10
Ψ

0.05

0.00
0 2 4 0 2 4 6
r (Bohr radii) r (Bohr radii)

Figure 1. A comparison of two sets of orbitals for a minimal basis set for silicon. The full lines
correspond to the orbitals used by Sankey and Niklewski, and the broken line to those used by
Horsfield. The qualities of the basis sets are essentially the same. However, the basis set of Sankey
and Niklewski has a discontinuity in its first derivative at the cut-off radius.

R
The overlap integrals (OI α,Jβ = dEr φI α φJβ ) and kinetic energy integrals (TI α,Jβ =
R
dEr φI α T̂ φJβ ) clearly consist of one-centre (I = P R two-centre (I 6= J ) terms.
J ) and
The electrostatic part of the double counting (− 21 I J dEr dEr 0 nI (Er )nJ (Er 0 )/|Er − rE0 | +
1 P E E
2 I 6=J ZI ZJ /|RI − RJ |) also consists of one- and two-centre terms.
The Hartree potential and the local part of the atomic pseudopotential can be combined
to form a short-ranged neutral-atom potential
( Z )
X n ( E0 )
r X (N A)
vI (Er ) + drE0
I
v (NA)
(Er ) = = vI (Er ). (5)
I |Er − rE0 | I
R6 A P Horsfield and A M Bratkovsky

A typical matrix element involving this potential is

Z XZ (N A)
dEr φI α (Er )v (NA) (Er )φJβ (Er ) = dEr φI α (Er )vK (Er )φJβ (Er ). (6)
K

If I = J = K we have a one-centre integral. If I 6= J , but K = I or K = J , then we have a

two-centre integral. Otherwise we have a three-centre integral.
The one-centre integrals can be evaluated once and stored. The two-centre integrals can
be tabulated as a function of separation on a one-dimensional grid, with rotations being taken
into account by means of Slater–Koster [32] tables. The three-centre integrals are tabulated
as a function of three variables (see figure 2): the bond length (r), the distance between the
bond centre and the site on which the potential appears (s), and an angle (θ ). The tables are
created for a specific geometry, and the integrals for other geometries are obtained by means
of rotations. The integrals are three dimensional, and are performed in reciprocal space as this
allows two out of the three integrals to be performed analytically, leaving a one-dimensional
integral to be performed numerically.

z z
r
θ
θ s
r s

x x

a) b)
Figure 2. The geometry used to define variables in terms of which the three-centre tables are
constructed. Panel (a) is for the method of Sankey and Niklewski, and panel (b) is for the method
of Horsfield.

The electrostatic integrals are easy to tabulate because the total potential can be expressed
as a linear combination of spherical single-site quantities. However, the integrals involving the
exchange and correlation potential and energy are more difficult to handle because the functions
involved have a strongly sub-linear dependence on density (roughly n1/3 ). In the Sankey–
Niklewski method this functional dependence is exploited by approximating the density in some
region by an optimized constant value, for which the integrals are easy to evaluate. Consider
the integral of the exchange and correlation potential in the local density approximation
Z
dEr φI α (Er )vxc (nin (Er ))φJβ (Er )
Z Z
0
≈ dEr φI α (Er )vxc (n̄)φJβ (Er ) + dEr φI α (Er )vxc (n̄)(nin (Er ) − n̄)φJβ (Er ) + · · ·
0
= vxc (n̄)OI α,Jβ + vxc (n̄)(nI α,Jβ − n̄OI α,Jβ ) + · · · (7)
R
where nI α,Jβ = dEr φI α (Er )nin (Er )φJβ (Er ). The optimum constant value for the density (n̄) is
chosen so as to make the second term zero (n̄ = nI α,Jβ /OI α,Jβ ). Care has to be taken when
the overlap goes to zero. Dipole and quadrupole fluctuation corrections are included.
 Ab initio tight binding R7

The method has been applied to such systems as amorphous silicon [33–39], silicon
clusters [40], silicon surfaces [41, 42], various carbon structures [43–51], a number of
multicomponent problems such as GeSe [52] and Si–C alloys [53], as well as many others
[54–70].

4.2.2. Formalism of Horsfield. The formalism of Horsfield [30] is rather similar to that of
Sankey and Niklewski. The important differences are the choice of basis set, and the way in
which the exchange and correlation integrals are handled.
A minimal basis set was found to be inadequate for an accurate description of fluoro-
carbons, but a double-numeric basis set gave good agreement with accurate density functional
calculations. The orbitals were taken from the neutral atom and a positively charged ion [71].
As with the formalism of Sankey and Niklewski the atomic calculations were performed in a
confining potential. The potential had the form r 6 . The square-well potential was not used
because it produces a discontinuity in the first derivative of the wavefunction at the cut-off radius
(see figure 1). The integrals were all performed in real space using partition functions [71].
The perspective taken when evaluating the exchange and correlation integrals is rather
different from that of Sankey and Niklewski. The key point is no longer the sub-linear
dependence of the functionals on density, but rather the localized character of the confined
atomic charge densities. For the exchange and correlation potential integrals, this localization
allows us to write a many-centre expansion of the form
Z
dEr φI α (Er )vxc [nin ; rE]φJβ (Er )
Z
≈ dEr φI α (Er )vxc [nI + nJ ; rE]φJβ (Er )
X Z
+ dEr φI α (Er ){vxc [nI + nJ + nK ; rE] − vxc [nI + nJ ; rE]}φJβ (Er ) + · · · .
K(6=I,J )
(8)
As for the electrostatic terms we have one-, two-, and three-centre (and higher) terms. These
are added to the electrostatic and kinetic energy terms to create a single set of tables (see
figure 2 for the geometry used for the three-centre integrals). It was found to be necessary
to carry out an additional numerical integral for the on-site exchange and correlation term
since the few-centre approximation is not accurate enough in this case. However, the integral
requires only a few points, and so is fast to evaluate. This method gives very good agreement
with accurate self-consistent calculations for molecules [30].

4.2.3. Formalism of Frauenheim. It could be argued that this is not strictly a first-principles
method in that it requires the fitting of a pair potential, but it does make use of tabulated
integrals for the hopping integrals and overlap matrix, and has been used very successfully;
thus it is considered here. The form used for the total energy is
X 1X
UF rauenheim = fi εi + φI J (|REI − REJ |) (9)
i
2 I 6=J
where φI J is a short-ranged repulsive pair potential. This form can be justified in terms of
screening arguments [72].
As for the previous two formalisms the orbitals are obtained from atomic calculations
with the atoms being confined by a localizing potential, the form used being r 2 [73]. Once
the orbitals are chosen, the overlap and Hamiltonian matrices can be generated. This method
R8 A P Horsfield and A M Bratkovsky

eliminates two terms from the Hamiltonian matrix that are present in the previous formalisms,
namely the crystal-field terms and the three-centre integrals [31]. This can again be justified in
terms of pseudopotential and screening arguments [74]. The potential used to evaluate the two-
centre hopping integrals is the sum of the two spherical atomic potentials. The on-site terms
in the Hamiltonian are taken from the free atom (not confined). Using this Hamiltonian and
overlap matrix the band-structure energy can be evaluated. The pair potential is generated by
subtracting the band energy from first-principles total-energy calculations for certain selected
structures. The absence of three-centre terms makes this method faster and more economical
with its use of memory as compared with the previous two methods.
This formalism has been applied widely. Materials that have been studied using this
formalism include carbon [31], silicon [75], boron nitride [76], germanium [77], gallium
arsenide [78], Sin Nm clusters [79], and gallium nitride [80].

4.3. Muffin-tin orbital-based tight binding

As stated above, we wish to make our basis set localized in real space and as small as possible.
Making use of orbitals obtained from calculations of an atom in a confining potential is one way
of approaching this problem. A less ad hoc approach is to construct the basis set from solutions
to an exactly solvable problem that is as close as possible to the one we are interested in. The
muffin-tin potential [81] provides a popular and successful solvable problem for close-packed
systems. The potential for a solid is approximated by a series of non-overlapping atomic-
like spherical potentials, and a constant potential between the spheres (a two-dimensional
representation of the regions looks rather like a muffin tin). Schrödinger’s equation can be
solved exactly in both regions. These solutions are matched at the boundaries of the spheres
to produce muffin-tin orbitals (MTOs). One can further reduce the effect of the interstitial
regions by working with space-filling, overlapping atomic spheres instead of MT spheres (so-
called atomic spheres approximation, ASA). Unfortunately these MTOs are very long ranged.
However, a unitary transformation can be applied to these long-ranged orbitals to render them
short ranged, and hence suitable for tight binding [82]. The formalism is somewhat intricate,
so the relevant theory has been reproduced in the appendix.

4.3.1. Nearly orthogonal tight-binding LMTOs. In the linear MTO method, MTOs are used
to construct a basis set which is (i) energy independent, (ii) exact to linear order in energy, and
(iii) rapidly convergent. The latter means that for the valence electrons it is generally sufficient
to retain one orbital per site per orbital quantum number L (=lm), where l and m are the orbital
and azimuthal quantum numbers respectively. Typically nine standard LMTOs (χ 0 ) per site
(corresponding to an spd basis set) produce sufficient accuracy for most transition metals [83].
This is a minimal basis set, which is what we would like for tight binding.
The LMTOs represent a variational basis constructed from the solutions of the Schrödinger
equation within the muffin-tin spheres φ(ν , r), where ν is selected within the region of
energies occupied by the valence electrons. If we consider an arbitrary energy , then the
LMTOs have an error of order ( − ν )2 within the muffin-tin (MT) spheres and of order
( − ν )1 in the interstitial region (not belonging to any of the muffin-tin spheres). To improve
the accuracy of the method one usually selects overlapping MT spheres whose volume is equal
to the volume of the respective Wigner–Seitz spheres and neglects completely the contribution
of the interstitial regions (so-called atomic sphere approximation, ASA, or LMTO-ASA).
The MTOs are long ranged with tails that decay with distance from the origin r as 1/r l+1 .
(Thus an s-MTO decays as 1/r, and p- and d-MTOs decay as 1/r 2 and 1/r 3 respectively.)
These are not suitable for real-space calculations. However, these tails are formally analogous
 Ab initio tight binding R9

to a Coulomb field produced by a superposition of electric multipoles placed at atomic sites,

and thus they may be screened (made short ranged) by appropriate unitary transformations
of the initial basis [83] to form a tight-binding representation. We shall denoted the screened
LMTO by χ α .
By screening the linearized MTOs, one can produce the most localized basis set χ α ,
most advantageous for real-space calculations. These orbitals are non-orthogonal, but can
be transformed into nearly orthogonal (though somewhat longer-ranged) orbitals χ γ . A cor-
responding nearly orthogonal tight-binding Hamiltonian H γ can then be generated and used
in much the same way as an orthogonal empirical tight-binding Hamiltonian. A summary of
the underlying theory is given in the appendix. Here the central results are presented.
A brute-force method to transform the most localized orbitals χ α into orthogonal ones is
to apply the Löwdin transformation
1 1
H= √ Hα √ ≡ h(O α )−1/2 χ α |−∇ 2 + v|χ α (O α )−1/2 i (10)
O α Oα
where Hα ≡ hχ α |−∇ 2 + v|χ α i is the standard LMTO Hamiltonian, v is the electron potential,
and O α ≡ hχ α |χ α i is the matrix of overlap integrals between the screened LMTOs χ α . We
would like to avoid doing this transformation by careful choice of our LMTO representation.
Consider the representation where α = γI l , i.e. where we use site-dependent screening.
In this representation the parameter P̈ α (εν ) = 0, as follows from equations (A.47), (A.52),
and (A.53). Thus we have
O γ = 1 + hγ phγ + i hκ γ |κ γ ii ≈ 1
(11)
H γ ≈ εγ + hγ + hγ εν phγ + i hκ γ |−∇ 2 + v|κ γ ii ≈ εγ + hγ .
Therefore, we can easily construct a nearly orthogonal representation which is sufficiently
accurate for applications using the recursion method. To get the expression for hγ we
have to use the fact that the first term in the overlap (A.53) is the major one, so that
(O α )−1/2 ≈ (1 + σ α hα )−1 and we can introduce (see equation (10)) the nearly orthogonalized
LMTOs
|χ γ i = |χ α i(1 + σ α hα )−1
and using this to calculate the Hamiltonian matrix, we get
H γ ≡ hχ γ |−∇ 2 + v|χ γ i = (1 + hα σ α )−1 H α (1 + σ α hα )−1 .
Substituting here the expression for H α from equation (A.52), we get equation (11) with
hγ ≡ (1 + hα σ α )−1 hα = hα (1 + σ α hα )−1 = hα − hα σ α hα + · · · (12)
α γ
so we can use the most-localized h for constructing the nearly orthogonal H . We should
keep in mind that, in principle, the γ -representation is longer ranged than the most-localized
α-representation.
The actual computational procedure starts from the transformation S 0 → S α ; then one
calculates the potential parameters for a trial potential (or density) and evaluates hγ from
equation (12). One then proceeds to calculate H γ and spectral functions (the electron density
of states, etc) needed to reach a self-consistent solution to the electronic problem. This
is especially useful in combination with the recursion method [84, 85] for large disordered
solids [86], where the k-space method breaks down.
We now briefly consider a few examples of the applications of this method. It has proven
very useful in studies of amorphous magnetism, such as that in Fe1−x Bx metallic glasses [87]
and Al–Mn [88, 89]. In the latter case it was possible to make a prediction of an unusual
transition when the paramagnetic crystalline Al6 Mn becomes magnetic on being melted. The
R10 A P Horsfield and A M Bratkovsky

real-space recursion TB-LMTO method was extended to treat arbitrary non-collinear magnetic
E Ib
structures (i.e. with arbitrary directions of the exchange potential 21 1 σE , where b
σE are the Pauli
matrices) which made it possible to establish the existence of a random magnetic order in Al–
Mn liquids in [89], and investigate complex magnetism and magnetization direction switching
in ultrathin Co/Cu films [90,91]. In the latter case the convergence was achieved to better than
0.0004 electrons au−3 for electron densities and about 0.003◦ for spin directions. The number
of iterations necessary for such a self-consistency was very large, about 1000. Additional
examples are mentioned in e.g. [106, 108].

4.3.2. Effective-medium-theory-based tight binding. In this model the total energy is given
by [92]
" # " #
X X X
UEMT B = ec (sI ) + Eas − eas (sI ) + E1el − e1el (sI ) . (13)
I I I

A central quantity in this expression is sI , which is the neutral-atom radius. This is the radius
of the smallest sphere centred on site I which is charge neutral. The function ec (sI ) gives the
energy per atom in a reference system (the diamond structure for the case of silicon) with the
same neutral-atom radius. This is the large term in the expression. The following two pairs of
terms are corrections to this term which take  intoP account the
 details of the environment. The
pair of terms (the atomic sphere terms) Eas − I eas (sI ) is the difference in a combined
electrostatic and exchange and correlation term between the real system (Eas ) and the reference
system. The term for the real system is expressed as a density-dependent pair potential. The
final pair of terms is the difference in the one-electron energies between the real system and the
reference system. A two-centre TB-LMTO Hamiltonian is used to evaluate the one-electron
energies for the real system. This method has been applied to a number of problems involving
silicon [93–95].

5. Introducing self-consistency

One of the main limitations of the formalisms described above is the absence of charge self-
consistency. This is important for many systems, and so a number of attempts have been made
to include self-consistency into the tight-binding models. The schemes range from simple
monopolar corrections to full self-consistency with no shape approximations for the charge
density.
Elstner et al [96] introduced self-consistency into the formalism of Frauenheim. We begin
with this approach because it can be applied to any tight-binding formalism (empirical as well
as ab initio). Self-consistency is introduced by means of the addition of the following term to
the tight-binding expression for the total energy:
1X
1E = qI qJ γI,J (14)
2 I,J
R
where γI,J = dEr dEr 0 FI (|Er − REI |)FJ (|Er 0 − REJ |)/|Er − rE0 |. The function FI (r) is a spherical
charge density, normalized to 1, and is taken to have an exponential form. Gaussians could
also be used, as they makePthe algebra simpler [97]. The charges qI are given by Mulliken
population analysis (qI = α,Jβ,i CI(i)α fi CJβ (i)
OJβ,I α − ZI ). Minimizing the total energy with
(i)
respect to the expansion coefficients CI α gives the self-consistent Hamiltonian
1
hI α,Jβ = h(0)
I α,Jβ + (wI + wJ )OI α,Jβ (15)
2
 Ab initio tight binding R11

where h(0)
P I α,Jβ is the non-self-consistent Hamiltonian, and where the energy shift wI =
J J I,J . The force on an atom is given by
q γ
X (wI + wJ ) ∂OI α,Jβ 1X ∂γI,J
FEK = FEK(0) − CI(i)α fi CJβ
(i)
− qI qJ (16)
I α,Jβ,i
2 E
∂ RK 2 I,J ∂ REK

where FEK(0) is the force evaluated using the expression for non-self-consistent tight binding.
Formally this expression ignores any contribution from exchange and correlation.
However, the main contribution to self-consistency is from the on-site term, and the decay
rates of the functions FI are adjusted to give an on-site value (γI,I ) that agrees with either
experiment or first-principles calculations, and thus includes correlation.
There have been two attempts to include self-consistency into the formalism of Sankey
and Niklewski within the spherical charge approximation [55,61]. Here we will only consider
the later method [61]. Consider silicon. It has s and p valence electrons. Self-consistency is
achieved by varying the number of s and p electrons contributing to the charge density on each
site. This allows both for promotion of electrons from s to p, and for the net accumulation of
charge. The number of s and p electrons is determined from the output wavefunctions using
the projection onto the Löwdin orbitals [98]. All the integrations are carried out exactly as
before, with one exception. When there is a net charge on a site there will be a long-ranged
electrostatic field. Integrals involving this field are treated within a dipole approximation.
The formalism of Horsfield also includes approximate self-consistency. It assumes
spherical input charges, but allows them to vary as nI (Er ) = n(0) I (Er ) + qI 1I (Er ), where n(0)
I
is the charge density from the neutral atom. The perturbing charge density 1I is spherical,
normalized to one, and constructed from partially occupied orbitals. The Hamiltonian
matrix elements are assumed to vary linearly with qI , and the double-counting term to vary
quadratically. For the electrostatic terms this is the correct behaviour, but for the exchange and
correlation terms this is approximate. The long-ranged electrostatic fields are handled using
a low-order Gaussian expansion for the orbitals which allows analytic expressions to be used.
The values of qI are found by maximizing the total energy. This follows from the fact that the
Harris–Foulkes functional is maximized by the correct input charge density [99] (unlike the
Kohn–Sham functional which is minimized).
The formalism of Lin and Harris [99] is self-consistent from the beginning. The main
difference as compared with the tight-binding formalisms described above is that analytic
functions are used for the orbitals and input charge density, allowing all the integrals except
those involving exchange and correlation to be expressed in closed form. A quadratic approx-
imation is made for the exchange and correlation energy. The charge on each site is found by
maximizing the total energy.
All the above self-consistent schemes make the approximation that the input charge
density is a sum of atom-centred spherical charge densities. Ordejon et al [100] relaxed
this approximation and represent the difference between the sum of atomic densities and the
self-consistent density on a uniform mesh. They thus turned the formalism of Sankey and
Niklewski into a fully self-consistent LCAO method.

6. Comparison of methods

In order to give some idea of the relative efficiencies and accuracies of the methods described
above, calculations of the formation energy of the relaxed isolated vacancy in silicon were
performed. All the calculations use a 64-atom cell with one atom removed. Only the 0
point is used for sampling the Brillouin zone. The basis set is always minimal (s and p).
R12 A P Horsfield and A M Bratkovsky

The calculations were performed using the program Plato (package for linear combination of
atomic-type orbitals).
The times we consider are for the evaluation of one total energy and set of atomic forces,
and we will measure them relative to the time for non-self-consistent orthogonal empirical tight
binding. The methods considered are: empirical orthogonal tight binding using the parameters
of Bowler et al [101]; the non-orthogonal tight binding of Frauenheim; the ab initio tight
binding of Horsfield; and full linear combination of atomic orbitals (LCAO) with many of
the integrals evaluated during the calculation. With the exception of the full LCAO method,
self-consistency is imposed using the monopole method of Elstner et al [96]. From table 1 we
can draw several general conclusions.

Table 1. Results of calculations of the formation energy of the relaxed vacancy of silicon. The times
are measured relative to orthogonal tight binding. Note the following abbreviations: TB = tight
binding, SC = self-consistent, AI = ab initio, LCAO = linear combination of atomic orbitals,
DNP = double numeric with polarization. The times in curly brackets for the ab initio tight binding
correspond to the use of the Chebyshev expansion for the three-centre integrals, whereas the other
numbers correspond to the use of linear interpolation.

Method Formation energy (eV) Time

Orthogonal TB 3.2 1
SC orthogonal TB 3.2 13
Non-orthogonal TB 3.1 2
SC non-orthogonal TB 3.4 19
AI TB 4.3 72 {180}
SC AI TB 4.4 87 {210}
AI LCAO 3.8 44
SC AI LCAO 4.1 110
DNP AI LCAO 3.9 880
DNP SC AI LCAO 3.2 2100

First, we consider the cost of simulations in terms of computer time: introducing three-
centre terms makes the simulations much slower; compared with the non-self-consistent
calculations, self-consistency is cheap for ab initio tight binding, but expensive for everything
else; full LCAO can be cheaper than ab initio tight binding (this is because the three-centre
integrals do not have to be evaluated one at a time in the full LCAO calculations).
Second, we consider the accuracies. For the four results that use the same basis set
(ab initio tight binding and the full LCAO method), we see that no one of the approximations
is obviously better than the others. The four results based on two-centre tight binding show
remarkable agreement with each other, but are all about 1 eV smaller than the corresponding
ab initio results. In table 1 are also given results obtained with a well converged basis set
(two sets of s and p orbitals, and one set of d orbitals) using the full LCAO method. The self-
consistent result (3.2 eV) is in remarkably good agreement with the two-centre tight-binding
results. This follows from the fact that the two-centre models are fitted to well converged
self-consistent results.
The accuracies cited above have been relative to other atomic-type orbital methods. The
result from well converged plane-wave calculations is about 3.6 eV [102], which is rather
larger than the most accurate result obtained from the self-consistent LCAO calculation with
the large basis set. The origin of the discrepancy is the k-point sampling. Using only the 0
point is insufficient. If the non-self-consistent orthogonal tight-binding simulation is repeated
 Ab initio tight binding R13

with a 3 × 3 × 3 mesh of k-points, the vacancy formation energy is found to increase from
3.2 eV to 3.8 eV.
For the case of the vacancy it is clear that the two-centre tight-binding models are faster
and more accurate than the ab initio minimal-basis-set methods. This is because the fitting
procedures used to generate the models were appropriate for this problem. However, the
two-centre models are not as transferable as the ab initio models. This can be clearly seen
from the phase diagram of silicon. The non-orthogonal two-centre model of Frauenheim gives
a description of the close-packed phases of silicon that is rather less good than that of the
more open structures [75], whereas the ab initio model of Sankey and Niklewski [29] gives a
rather better description. It should also be remembered that the motivation for developing the
ab initio methods was to provide a scheme that can be improved systematically and which can
be applied easily to multicomponent systems. To see both factors coming into play in the case
of hydrocarbons and fluorocarbons, see Horsfield [30].

Appendix. Derivation of the tight-binding LMTO method

In the muffin-tin approximation we have spherically symmetric potentials vI (rI ) within atomic
spheres (AS) of radius sI centred on sites I , and a constant potential (v0 = 0) in the interstitial
region. Thus the total potential (v(Er )) is given by
X
v(Er ) = vI (rI ) (A.1)
I

where rI = |Er − REI |, and is the distance from the centre of the site at REI . Orbitals with
eigenvalue ε (ϕl (ε, r)) and energy derivatives ϕ̇L (ε, r) = dϕl (ε, r)/dε are found from
[∇ 2 + ε − vI (rI )]ϕl (ε, r) = 0 (A.2)
with the energy measured in Ryd, and the distances in Bohr radii.
An orbital with arbitrary energy ε can be approximated inside the muffin-tin sphere by
 
ϕl (ε, r) = ϕl (εν , r) + (ε − εν )ϕ̇(εν , r) + O (ε − εν )2 . (A.3)
These are linearized muffin-tin orbitals. This approximation is accurate for energies spanning
about 1 Ryd around εν .
We shall denote by |ϕ(ε, r)i the solutions terminated (that is, set to zero) outside their
respective atomic spheres. The orbitals are normalized to unity within the AS:
Z SI
hϕ|ϕi ≡ dr r 2 ϕ 2 (ε, r) = 1
0 (A.4)
hϕ|ϕ̇i = 0
where the second orthogonality condition follows directly from normalization.
In the interstitial region (where v = v0 ) the Schrödinger equation reduces to the Helmholtz
wave equation
(∇ 2 + κ 2 )ϕ(ε, rE) = 0 (A.5)
where κ 2 = ε − v0 .
A complete orbital is constructed from a part inside the AS (the head) and a part that lies
outside (the tail). For the tail region one selects a solution of (A.5) in terms of the Bessel
functions Hl (κ 2 ) at given energies εν , which are generally different from the values used for
the head. The simplest, but very effective, choice is κν2 = 0. In this case equation (A.5) reduces
to the Laplace equation, with the LMTO tail in the interstitial region being proportional to
 l+1
w
KI0lm (ErI ) ≡ Ylm (ErI ) (A.6)
rI
R14 A P Horsfield and A M Bratkovsky

where w is a scaling length, such as the average Wigner–Seitz radius. The envelope function
K 0 has the form of an electrostatic potential produced by a 2l pole at I . It is regular everywhere
except at RI , and can be expanded about any site I 0 6= I [103]:
X
KI0L (ErI ) = − |JI00 L0 (ErI 0 )iSI00 L0 I L (A.7)
L0

in terms of the regular solutions of the Laplace equation

(rI /w)l
JI0L (ErI ) ≡ YL (ErI ). (A.8)
2(2l + 1)
The ket-vector notation is used to denote the orbital terminated beyond the atomic sphere it
is centred on. We shall make use of the convention that the orbitals, which depend explicitly
on the direction rEI , contain the spherical function YL (ErI ). Otherwise, only the radial part
depending on the absolute value of rI is used. The expansion coefficients S 0 are the so-called
canonical structure constants which are independent of energy, lattice constant, and sphere
radii. They have a very simple form in terms of the Slater–Koster notation [32]:
√
0
Sssσ = −2w/d 0
Sspσ = 2 3(w/d)2
0
Spp (σ, π) = 6(w/d)3 (2, −1)
√
0
Ssdσ = −2 5(w/d)3 (A.9)
√ √
0
Spd (σ, π) = 6 5(w/d)4 (− 3, 1)
0
Sdd (σ, π, δ) = 10(w/d)5 (−6, 4, −1).
The constants for arbitrary REI 0 − REI required in the expansion (A.7) can be trivially found
from Slater–Koster tables [32]. The same expansion (A.7) holds true if one uses the envelope
solutions K 0 for κν2 6= 0, with a more complicated functional form for the S(κ 2 ) matrices [103].
The construction of the LMTOs proceeds as follows. Consider one atomic sphere centred
at site I . The solution of Schrödinger’s equation inside the sphere is approximated by a linear
combination of ϕν and ϕ̇ν by matching the function and its first radial derivative to the envelope
function K 0 at the atomic sphere surface:
KI0l (rI ) ≡ (w/rI )l+1 → AK K
I l (ε)ϕl (ε, rI ) + BI l (ε)ϕ̇l (ε, rI ) (A.10)
where AK and B K are the expansion coefficients found from the matching conditions. Note
that since hϕ(ε)|ϕ̇(ε)i ≡ 0 for any ε we can perform this matching procedure for arbitrary
energy and then generate a set of energy-dependent basis functions. As we will see shortly,
this reduces the Schrödinger equation to a non-linear eigenvalue problem, which is difficult to
solve. Instead, we shall use our freedom in constructing these orbitals to go over directly to
an energy-independent basis set, which can be used effectively in a variational solution of the
Schrödinger equation, since it results in the linear eigenvalue problem.
The tail of the envelope function (A.7) is also matched in the same way inside neighbouring
spheres to the orbitals and their first energy derivatives:
Jl0 (rI 0 ) ≡ (rI /w)l /2(2l + 1) → J˜I00 l (ε, rI 0 ) ≡ AJIl (ε)ϕl (ε, rI ) + BIJl (ε)ϕ̇l (ε, rI ) (A.11)
where the expansion coefficients AJIl and BIJl are found from this matching condition, as AK Il
K
and BI l have been before them. By using these substitutions we arrive at some atom-centred,
energy-dependent orbitals which are called the MTOs χ, and we write them in the following
form for any point in the crystal:
X
χINL (ε, rE) = |ϕL (ε, rEI )iAK
I l (ε) + |ϕ̇L (ε, r
EI )iBIKl (ε) − |J˜I00 L0 (ErI 0 )iSI00 L0 I L + |Kii (A.12)
I 0 L0
 Ab initio tight binding R15

where all | is are non-zero only in their respective atomic spheres, | ii is non-zero only in the
interstitial region, and Nν is the normalization coefficient. It is more convenient to write MTOs
in a slightly different, though equivalent, fashion. Namely, instead of ϕ and ϕ̇ one can use ϕ
and J˜0 to arrive at a more frequently used representation of the energy-dependent MTOs:
X
χINL (ε, rE) = |ϕL (ε, rEI )iNI0l (ε) + |J˜L0 (ε, rEI )iPI0l (ε) − |J˜I00 L0 (ε, rEI 0 )iSI00 L0 I L + |Kii (A.13)
I 0 L0

where PI0l (ε) = BIKl (ε)/BIJl (ε), NI0l (ε) = AK

I l (ε) − PI l (ε)AI l (ε). Omitting the site and orbital
0 K

indices and summation over repeating indices, one can rewrite MTOs in the compact form
χ N (ε) = |ϕ(ε)iN(ε) + |J˜0 (ε)i[P 0 (ε) − S 0 ] + |Kii (A.14)
where N(ε) and P 0 (ε) are the so-called potential parameters which are diagonal with respect
to combined site and orbital index I L. Since χI L (ε, rE) = χI L (ε, rE − REI ), we can use the
standard construction for the Bloch wave in a perfect crystal. Since the normalization factor
is a constant we can work with
χ(ε) ≡ χINL (ε)/NI L (εν ). (A.15)
We can write a general electron wavefunction as
X (i)
ψi (ε) = CI L χI L (ε). (A.16)
IL

One is free to select the coefficients CI(i)L such that the second term in equation (A.14) vanishes:
X (i)
CI 0 L0 (ε)[PI00 l 0 (ε)δI 0 L0 I L − SI00 L0 I L ] = 0. (A.17)
I 0 L0

Then the remaining part, which is a linear combination of |ϕ(ε)i and |Kii , satisfies the
Schrödinger equation for the muffin-tin potential exactly. Thus, we have recovered the famous
KKR tail-cancellation condition. This equation, which defines the electron energy bands,
is very inconvenient indeed, since it is a non-linear equation for the eigenvalues ε. Instead
of solving this equation, we would like to construct from MTOs the most compact energy-
independent basis to use for the variational solution of the Schrödinger equation. Moreover,
we can use our flexibility in choosing energy-independent linear MTOs χ (LMTOs) to make
them accurate to first order in (ε − εν ) inclusive, so that the errors will be proportional to
(ε − εν )2 inside the atomic spheres. As we mentioned earlier, our approximations mean that in
the interstitial region the error of the LMTO is proportional to (ε − εν )1 , but this contribution
is neglected in the atomic sphere approximation, where one selects space-filling overlapping
MT spheres. We need to analyse MTOs in detail to see all this.
The matching conditions at the atomic sphere boundaries give us a two-by-two linear
system of equations, and it is convenient to write their solutions in the following form:
 2l+1
W [ϕ(ε), K 0 ] w D(ε) + l + 1
P (ε) =
0
= 2(2l + 1) (A.18)
W [ϕ(ε), J ] 0 s D(ε) − 1
W [K 0 , J 0 ] w/2 hw i1/2
N 0 (ε) = = = Ṗ 0
(ε) (A.19)
W [ϕ(ε), J 0 ] W [ϕ(ε), J 0 ] 2
where D(ε) ≡ sϕ 0 (ε, s)/ϕ(ε, s) is the logarithmic derivative of the wave function at the
atomic sphere surface, which is simply related to the usual scattering phase shift δl [104], and
W [f1 , f2 ] ≡ s 2 [f10 (s)f2 (s) − f1 (s)f20 (s)] is the Wronskian. In both cases the prime stands
for differentiation with respect to the radial variable. To derive (A.19) for N via Ṗ one should
differentiate the Wronskian relation (A.18) and use
Z s
W [ϕ̇, ϕ] = s 2 (ϕ̇ϕ 0 − ϕ̇ 0 ϕ)r=s = dr r 2 ϕ 2 = 1. (A.20)
0
R16 A P Horsfield and A M Bratkovsky

The latter relation follows from the Schrödinger equation (A.2) and its first derivative with
respect to energy
(−∇ 2 + v − ε)ϕ̇ = ϕ.
It is important to note that when the tail-cancellation condition (A.17) is obeyed, the
contributions of J˜0 (ε) cancel out in the atomic sphere at the origin exactly, at least for low
orbital angular momenta l. However, if one were to use the MTOs χ(ε) with a variational
method, these unwanted contributions, generally varying linearly with the energy ε, would
remain and reduce the precision of the band calculations. The energy dependence of J˜0 (ε) is
obviously weak, since it matches continuously to energy-independent envelope function K 0 ,
and can be eliminated completely to first order when using fixed energy orbitals. At any given
energy ε = ενI l in a region of interest we select the head
N(ε) P 0 (ε)
χ̃ head ≡ |ϕ(ε)i + |J˜0 (ε)i
N(εν ) N (εν )
(the part of the MTO in its own atomic sphere) and choose J˜0 (εν ) from the condition
χ̇ head (εν ) = 0 which gives us a condition for J˜0 (εν ):
Ṅ(εν ) Ṗ 0 (εν )
|ϕ̇(εν )i + |ϕ(εν )i + |J˜0 (εν )i = 0. (A.21)
N(εν ) N (εν )
This relation has a pure L-character since the matrices of potential parameters are diagonal
matrices with respect to site and orbital indices. Thus, we arrive at the following substitution:
 
˜ ϕ̇(εν ) + ϕ(εν )Ṅ 0 (εν )/N 0 (εν ) 2 0 −1/2
J (εν , r) → −
0
≡ −ϕ̇ (r)
0
Ṗ (A.22)
Ṗ 0 (εν )/N 0 (εν ) w
where we have used (A.19) and introduced the definitions
ϕ̇ 0 (r) ≡ ϕ̇(r) + ϕ(r)σ 0 (A.23)
σ 0 ≡ Ṅ 0 /N 0 = P̈ 0 /(2Ṗ 0 ) (A.24)
where we imply that all quantities without arguments are evaluated at ε = εν . After we perform
the substitution (A.22) in the head and the tails of the MTO we obtain the energy-independent
basis χ̃ (εν ), which is accurate to terms O(ε − εν )2 .
Substituting (A.22) in the equation (A.15), and using (A.19) we finally arrive at the working
expression for the standard linear muffin-tin orbital (we mark this representation by zero as
the superscript):
χ 0 ≡ |ϕi + |J˜0 i(P 0 − S 0 )/N 0 + |K 0 ii /N 0
= |ϕi − |ϕ̇ 0 i(Ṗ 0 )−1/2 (P 0 − S 0 )(Ṗ 0 )−1/2 + |K 0 ii /N 0
 
2 0 −1/2
= |ϕi + |ϕ̇ 0 ih0 + |K 0 ii Ṗ . (A.25)
w
Here we have introduced the two-centre first-order Hamiltonian
h0 = −(Ṗ 0 )−1/2 (P 0 − S 0 )(Ṗ 0 )−1/2
and have used (A.19). One can now use the constructed LMTOs to solve the Schrödinger
equation by minimizing hψi |Ĥ − E (i) |ψi i with ψi taken from equation (A.16). This leads to
the standard linear problem
X (i)
CI 0 L0 (HI 0 L0 I L − E (i) OI 0 L0 I L ) = 0
I 0 L0
 Ab initio tight binding R17

where the matrix elements are

H ≡ hχ 0 |−∇ 2 + v|χ 0 i
= h0 + h0 σ 0 h0 + (1 + h0 σ 0 )εν (σ 0 h0 + 1) + h0 εν ph0 + i hκ|−∇ 2 + v|κii
(A.26)
O ≡ hχ |χ i = (1 + h σ )(σ h + 1) + h ph + hκ|κi
0 0 0 0 0 0 0 0 i i
(A.27)
where we have introduced an extra potential parameter
pI l = h(ϕ̇I l )2 i = −ϕ̈l (sI )/3ϕl (sI )
and
 −1/2
2 0
|κii ≡ |K 0 ii Ṗ .
w
The last terms in (A.26) and (A.27) are the so-called combined corrections. Recalling that |κii
is a solution of the Schrödinger equation at ε = κ 2 we see that i hκ|−∇ 2 + v|κii = κ 2 (i hκ|κii ).
With the use of Green’s theorem and the definition of |κii we can express these integrals in terms
of the canonical structure constants S 0 and their first energy derivatives Ṡ ≡ ∂S/∂κ 2 [105,106].
Therefore, all one needs to know in order to solve the Schrödinger equation for a crystal is
the values of the partial wavefunctions and their derivatives with respect to the coordinate and
energy at the atomic sphere surface at the energy εν in the window of interest. These values are
easily found from radial solutions of the Schrödinger equation for the muffin-tin potentials.
Let us now discuss in more detail the first-order Hamiltonian h0 :
√ √
h0 = −|ϕ̇ 0 i(Ṗ 0 )−1/2 (P 0 − S 0 )(Ṗ 0 )−1/2 = c0 − εν + d 0 S 0 d 0 (A.28)
where
√  1/2  1/2  l+1/2
2 s s l−D
d 0 ≡ (Ṗ 0 )−1/2 = W [ϕ, J 0 ] = ϕ(s)
w 2 w 2l + 1
2 (D + l + 1)(l − D)
c0 − εν ≡ −P 0 /Ṗ 0 = − W [ϕ, K 0 ]W [ϕ, J 0 ] = sϕ 2 (s)
w 2l + 1
where D ≡ D[ϕν (r)] ≡ ∂ ln ϕν (r)/∂ ln r at r = s. Potential parameters c0 ≡ cI0l δI 0 L0 I L and
d 0 ≡ dI0l δI 0 L0 I L are the diagonal matrices easily found from the preceding equations.
It is instructive to relate the potential parameters c0 and d 0 to the scattering characteristics
of the muffin-tin potential for which they are calculated. First we recall the relation between the
potential function and the scattering phase shift, namely (P 0 )−1 ∝ − tan δl . Since tan δl has a
resonant character [104], it is customary to parametrize [P 0 (ε)]−1 in the following functional
form:
1
[P 0 (ε)]−1 = +γ (A.29)
ε−C
where the canonical potential parameters Cl (centre), 1l (width), and γl of the band l are
readily found from ϕ 0 (ε, s), ϕ(ε, s), and their derivatives with respect to energy. Cl is the
energy at which D(Cl ) = −l − 1. Substituting this expression into equation (A.28) we arrive
at the following relations for the parameters of the first-order Hamiltonian h0 :
√  
√ εν − C
d0 = 1 1 + γ
1
  (A.30)
εν − C
c0 − εν = (C − εν ) 1 + γ .
1
R18 A P Horsfield and A M Bratkovsky

If one selects εν = C, then d 0 = 1, c0 = C, and the first-order Hamiltonian becomes

√ √ √ √
h0 = 1S 0 1 H ≈ H (1) ≡ εν + 1S 0 1.
Thus we have arrived at the simplest two-centre first-order Hamiltonian in real space, useful
in qualitative discussions of the band problem. For an ideal crystal one transfers H and O
into k-space by substituting for the only site-non-diagonal matrix S with its Fourier transform,
calculated by the standard Ewald procedure. Since the canonical potential parameters Cl , 1l ,
and γl are somewhat less sensitive to a choice of εν , one can substitute equation (A.3) into
equation (A.18) and compare with the resonant expression for P 0 , equation (A.29), with the
following results suitable for practical calculations:
W [ϕ, K 0 ] ϕ(s) D[ϕ(s)] + l + 1
C − εν = − =−
W [ϕ̇, K 0 ] ϕ̇(s) D[ϕ̇(s)] + l + 1
√
√ w/2 1 (s/w)l+1/2 1
1= =√ √ (A.31)
|W [ϕ̇, K ]|
0
2 s|ϕ̇(s)| |D[ϕ̇(s)] + l + 1|
 2l+1
1 W [ϕ̇, J 0 ] 1 s D[ϕ̇(s)] − l
γ ≡ = 0
= .
P [ϕ̇(s)] W [ϕ̇, K ] 2(2l + 1) w D[ϕ̇(s)] + l + 1
In order to see that the width 1 has the correct dimensionality of Ryd we recall that we are
working in atomic units, where s has the dimensionality (Ryd)−1/2 and ϕ has a dimensionality
of s −3/2 = (Ryd)3/4 . This completes our discussion of the standard LMTO method.
In order to treat large systems (meaning a few hundred atoms or more) one would like to
have a localized, accurate, and minimal basis set which one can use in a fully ab initio method.
The standard LMTO method operates with long-range orbitals and is not suitable for this
purpose. However, the LMTOs can be made localized with the use of the information about
the environment of each atom [83]. To screen the long-ranged tail of K 0 (ErI ) one introduces,
instead of J 0 , a function J α in the expansion (A.7) which includes the screening multipoles
with charges αI l (=0 for l > lα , where usually lα 6 2):
JIαl (rI ) ≡ JI0l (rI ) − αI l KI0l (rI ). (A.32)
Introducing such screening charges will inevitably change the multipole field everywhere,
including the vicinity of the origin, where K 0 was centred. Thus, one has to introduce new
envelope functions K α which are defined in all space and are linearly related to K 0 everywhere.
The tail expansion now reads
X
KIαL (ErI ) = − JIα0 L (ErI 0 )SIα0 L0 I L (A.33)
L0
where now S has non-zero on-site elements, since K α 6= K 0 at the origin. Let us use
α

compressed matrix indices a ≡ I L which imply a summation over repeated indices, and mark
by the superscript ∞ the envelope functions defined in the whole space, while the functions
without this sign are assumed to be truncated beyond their atomic sphere of origin. Then we
have
∞ α
Ka ≡ Ka00 δa 0 a − Jaα0 Saα0 a = Ka00 (δa 0 a + αa 0 Saα0 a ) − Ja00 Saα0 a
∞ 0
(A.34)
Ka ≡ Ka00 δa‘a − Ja00 Sa00 a .
From our earlier discussion we expect these envelope functions to be related by a unitary
transformation Ua 0 a such that ∞ Kaα = ∞ K 0a 0 Ua 0 a . Comparing this with equation (A.34) we
obtain a relation between S α and S 0 by comparing the coefficients before K 0 and J 0 in two of
these expansions:
Ua 0 a = δa 0 a + αa 0 Sa00 a ≡ 1 + αS 0
(A.35)
Saα0 a = Sa00 a 00 Ua 00 a = S 0 (1 + αS α ) = S 0 + S 0 αS α .
 Ab initio tight binding R19

Thus we have obtained a Dyson-like equation for the screened structure constants. A solution
of the Dyson equation (A.35) can be given in the following convenient form:
S α = S 0 + S 0 αS 0 + · · · ≡ α −1 (1 + αS 0 + αS 0 αS 0 + · · · − 1)
 
= α −1 (1 − αS 0 )−1 − 1 = α −1 (α −1 − S 0 )−1 α −1 − α −1 . (A.36)
Since in a regular lattice with lattice vectors RE and TE and atomic position τE within the unit
cell,
h i−1 Xh i−1
E E E
αR−1 E TE τE0 L0 − SRE τEL,R+
E τEl δRE τEL,R+
0
E TE τE0 L0 = (1/N ) αR−1
E τEl − SτEL,E
0
τ 0 L0 (k) e−ik·T (A.37)
kE

where by definition
X E E
E =
Sτ0L,τ 0 L0 (k) SR0EτEL,R+
E TE τE0 L0 e
−ik ·T
(A.38)
TE

the long-range behaviour of S α is defined by the analytical properties of the Fourier-transformed

E which is customarily evaluated by means of a Ewald summation. It follows
function S 0 (k),
from the form of the canonical structure constants (A.9) that S 0 (k) E is bounded from above;
E 6= 0.
therefore it is possible to find sufficiently small and positive αs for which det[α −1 −S 0 (k)]
Since there are no poles for real values of k in equation (A.37), its Fourier transform will be
falling off exponentially with distance |RE − RE 0 |/w. This conclusion does not, of course, depend
on whether the lattice is ordered or not. As mentioned above, it should just be sufficiently
close packed or, for open structures, packed with empty spheres so that this recipe for screening
might work.
We now give a simple example illustrating this procedure for screening LMTOs. Consider
a system with only one s orbital per site [106]. The relevant structure constant is Sssσ ,
equation (A.9), whose lattice Fourier transform is
X −2w E E Z 3
E = d T 2w −ikE·TE 6
S(k) e−ik·T ≈ − e + constant = − 2
+ constant (A.39)
E
T  T (kw)
T

where the constant is determined from the

Pcondition that the on-site element of the S-matrix in
real space vanishes, SR=0
E = 0 = (1/N) kE S( E
k),  = (4π/3)w 3
is the unit-cell volume, and
the Brillouin zone was approximated by a sphere with radius k0 such that (4π/3)k03 = (2π )3 /.
After performing the integration we obtain

E =− 6 1
S(k) 2
+ (A.40)
(kw) αc
where αc = 2−5/3 3−2/3 π 2/3 = 0.32. Substituting this into (A.36) we obtain for RE 6= 0
X 11 1 6
−1
1 −ikE·RE
S0EαRE = (1/N) − + 2
e . (A.41)
α α αc (kw) α
Note that this expression has no poles only if α < αc = 0.32, i.e. the screening charges cannot
exceed a certain limit. In this case, if we were to extend the integration over k to infinity, we
would get
αc2 2w −√6ξ (R/w)
S0EαRE = − e (A.42)
(αc − α)2 R
where we have introduced ξ ≡ ααc /(αc − α). Indeed, this matrix decays exponentially in
real space. This seemingly obvious statement [83] does not actually apply to real lattices.
Obviously, the decay will be strictly exponential only when all multipole moments of the
R20 A P Horsfield and A M Bratkovsky

given charge distribution are exactly zero. Note that an actual integration is cut off at the
Brillouin-zone boundary, and the actual functional dependence on the distance R/w is given
in terms of a damped oscillatory asymptotic expansion. The complicated shape of the actual
Brillouin zone, however, produces strong compensation of different oscillatory terms, so the
actual behaviour is close to fast exponential decay. The optimal numerical choice for close-
packed structures is αs = 0.3485, αp = 0.053 03, and αd = 0.0107, and the orbitals are
essentially limited to first and second neighbours [83].
A screened LMTO χ α can be constructed in exactly the same way as the standard LMTO
(A.25), with the replacements of all superscripts 0 → α, and by substituting for the potential
parameter γ with γ − α in (A.29), (A.30). Indeed, the matching conditions now read
 l+1
(rI /w)l w
JI l (r) ≡
α
− αl → J˜Iαl (rI ) (A.43)
2(2l + 1) rI
 l+1
w
KI0l (r) ≡ → ϕI l (ε, rI )NIαl (ε) + J˜Iαl (rI )PIαl (ε). (A.44)
rI
We immediately have from these matching conditions
W [ϕ(ε), K 0 ] P 0 (ε)
P α (ε) = = (A.45)
W [ϕ(ε), J α ] 1 − αP 0 (ε)
0
W [K , J ]α h w α i1/2
N α (ε) = = Ṗ (ε) (A.46)
W [ϕ(ε), J α ] 2
where P 0 is from (A.18). Rewriting (A.45) in the form
 α −1  0 −1
P (ε) = P (ε) −α
and substituting the resonant expansion (A.29), we arrive at
 α −1 1
P (ε) = + γ − α. (A.47)
ε−C
In the expression for the Hamiltonian in the screened LMTO representation below (A.54)
we shall see that all potential parameters are defined by P α (ε) and its energy derivatives.
Comparing (A.47) with (A.29) we see that indeed the only difference is the replacement
γ → γ − α. Repeating our reasoning for the choice of J˜α , we arrive at the following
substitution, which makes the screened LMTO independent of energy everywhere:
 
˜ ϕ̇(εν ) + ϕ(εν )Ṅ α (εν )/N α (εν ) 2 α −1/2
J (εν , r) → −
α
≡ −ϕ̇ (r)
α
Ṗ (A.48)
Ṗ α (εν )/N α (εν ) w
where
ϕ̇ α (r) ≡ ϕ̇(εν ) + ϕ(εν )σ α (A.49)
Ṅ α (εν )
σα ≡ α = P̈ α /(2Ṗ α ) (A.50)
N (εν )
and obtain a screened LMTO (again, we want it to start from just the atomic orbital |ϕi, so we
divide the initial MTO (A.44) by N α (εν )):
χ α ≡ |ϕi + |J˜α i(P α − S α )/N α + |K α ii /N α
= |ϕi − |ϕ̇ α i(Ṗ α )−1/2 (P α − S α )(Ṗ α )−1/2 + |K α ii /N α
 
2 α −1/2
= |ϕi + |ϕ̇ α ihα + |K α ii Ṗ . (A.51)
w
 Ab initio tight binding R21

Here we have introduced the two-centre first-order screened Hamiltonian

hα = −(Ṗ α )−1/2 (P α − S α )(Ṗ α )−1/2 .
One can now use the screened LMTOs to solve the Schrödinger equation, which leads to the
standard linear problem
X α(i)
CI 0 L0 (HIα0 L0 I L − E (i) OIα0 L0 I L ) = 0
I 0 L0
where the matrix elements are
H α ≡ hχ α |−∇ 2 + v|χ α i
= hα (1 + σ α hα ) + (1 + hα σ α )εν (σ α hα + 1) + hα εν phα + i hκ α |−∇ 2 + v|κ α ii
(A.52)
O ≡ hχ |χ i = (1 + h σ )(σ h + 1) + h ph + hκ |κ i
α α α α α α α α α i α α i
(A.53)
where
 
2 α −1/2
|κ i ≡ |K i
α i α i
Ṗ .
w
One can gain more insight into the screened LMTOs by looking at the parametrization of
the Hamiltonian
√ √
hα = −(Ṗ α )−1/2 (P α − S α )(Ṗ α )−1/2 = cα − εν + d α S α d α (A.54)
where
 1/2  
√ 2 √ εν − C
d α ≡ (Ṗ α )−1/2 = W [ϕ, J α ] = 1 1 + (γ − α)
w 1
 
2 εν − C
c − εν ≡ −P /Ṗ = − W [ϕ, K ]W [ϕ, J ] = (C − εν ) 1 + (γ − α)
α α α α α
w 1
where we again expressed everything in terms of our primary potential parameters C, 1, and
γ (equation (A.31)). Note that it differs from (A.30) merely by substitution γ → γ − α, as
we discussed earlier; see equation (A.47).
This finalizes our construction of the tight-binding LMTO. We see that the first-order
Hamiltonian (A.54) is short ranged because it contains the screened S α -matrices. The range
of H α (equation (A.52)) is larger, since it contains two-hop terms, like hσ h. These terms may
be treated exactly or perturbatively.
Finally we offer some comments on the accuracy of the LMTO method and recent
improvements. We have used the atomic spheres approximation (ASA) by taking the energy
of our basis wave functions to be κν2 = 0. Thus we make an error that is linear in E − Eν in our
basis in the interstitial region, whereas inside the MT spheres the corrections are proportional to
(ε − εν )2 . The LMTO-ASA is a powerful tool for close-packed solids, but these inaccuracies
make it unsuitable for calculation of forces and dynamics. Full potential LMTO has been
designed to allow the inclusion of any non-ASA corrections and the accurate treatment of any
interstitial region (see [107] and references therein). It is somewhat involved, making use of
multiple-κν2 basis sets to describe accurately the interstitial region.
It is worth mentioning that it is in principle possible to reformulate the standard LMTO
method so that (i) the error of the basis in the interstitial region is not ε − εν but (ε − εν )2 ,
(ii) the basis set is localized in space, and (iii) the full (non-spherical) charge density and
potential are expanded via the same functions as were used to construct the basis [108]. The
main idea of using solutions of the Schrödinger equation in the MT spheres as well as the
interstitial region (A.5) with matching at the MT spheres does not change. However, in order
R22 A P Horsfield and A M Bratkovsky

to localize the basis one has to solve the wave equation (A.5) such that the solution has zero
YL0 (rI )-projections on other non-touching spheres with radii aI 0 l 0 (these spheres are analogous
to hard impenetrable spheres). Such solutions always exist, they produce a complete set, and
they may be localized. One needs these functions to obey the matching conditions of the
wave functions at the MT spheres, which is done by introducing so-called kinky partial waves.
The matching conditions at the MT spheres are then reduced to algebraic form involving the
kink (KKR) matrix K. The LMTO overlap matrix and the Hamiltonian matrix can then be
expressed solely in terms of the kink KKR matrix at the selected energy Eν and its first three
···
energy derivatives K̇, K̈, and K. This, as we have seen, extends the spatial range of the matrix
elements.

References

[1] Ashcroft N W and Mermin D 1976 Solid State Physics (Tokyo: Holt-Saunders)
[2] Tan K E, Bratkovsky A M, Harris R M, Horsfield A P, Nguyen-Manh D, Pettifor D G and Sutton A P 1997
Modell. Simul. Mater. Sci. Eng. 5 187
[3] Kohn W and Sham L J 1965 Phys. Rev. 140 A1133
[4] Sutton A P, Finnis M W, Pettifor D G and Ohta Y 1988 J. Phys. C: Solid State Phys. 21 35
[5] Goodwin L, Skinner A J and Pettifor D G 1989 Europhys. Lett. 9 701
[6] Xu C H, Wang C Z, Chan C T and Ho K M 1992 J. Phys.: Condens. Matter 4 6047
[7] Kwon I, Biswas R, Wang C Z, Ho K M and Soukoulis C M 1994 Phys. Rev. B 49 7242
[8] Tang M S, Wang C Z, Chan C T and Ho K M 1996 Phys. Rev. B 53 979
[9] Allen P B, Broughton J Q and McMahan A K 1986 Phys. Rev. B 34 859
[10] Khan F S and Broughton J Q 1989 Phys. Rev. B 39 3688
[11] Sawada S 1990 Vacuum 41 612
[12] Mercer J L and Chou M Y 1993 Phys. Rev. B 47 9366
[13] Mercer J L and Chou M Y 1994 Phys. Rev. B 49 8506
[14] Sigalas M M and Papaconstantopoulos D A 1994 Phys. Rev. B 49 1574
[15] Cohen R E, Mehl M J and Papaconstantopoulos D A 1994 Phys. Rev. B 50 14 694
[16] Menon M and Subbaswamy K R 1997 Phys. Rev. B 55 9231
[17] Liu F 1995 Phys. Rev. B 52 10 677
[18] Pople J A, Santry D P and Segal G A 1965 J. Chem. Phys. 43 S129
[19] Pople J A and Segal G A 1965 J. Chem. Phys. 43 S136
[20] Pople J A and Segal G A 1966 J. Chem. Phys. 44 3289
[21] Santry D P and Segal G A 1967 J. Chem. Phys. 47 158
[22] Pople J A, Beveridge D L and Dobosh P A 1967 J. Chem. Phys. 47 2026
[23] Harris J 1985 Phys. Rev. B 31 1770
[24] Foulkes W M C 1987 PhD Thesis University of Cambridge
[25] Foulkes W M C and Haydock R 1989 Phys. Rev. B 39 12 520
[26] Finnis M W 1990 J. Phys.: Condens. Matter 2 331
[27] Polatoglou H M and Methfessel M 1988 Phys. Rev. B 37 10 403
[28] Chan C T, Vanderbilt D and Louie S G 1986 Phys. Rev. B 33 2455
[29] Sankey O F and Niklewski D J 1989 Phys. Rev. B 40 3979
[30] Horsfield A P 1997 Phys. Rev. B 56 6594
[31] Porezag D, Frauenheim Th, Köhler Th, Seifert G and Kaschner R 1995 Phys. Rev. B 51 12 947
[32] Slater J C and Koster G F 1954 Phys. Rev. 94 1498
[33] Drabold D A, Fedders P A, Sankey O F and Dow J D 1990 Phys. Rev. B 42 5135
[34] Drabold D A, Fedders P A, Klemm S and Sankey O F 1991 Phys. Rev. Lett. 67 2179
[35] Fedders P A, Drabold D A and Klemm S 1992 Phys. Rev. B 45 4048
[36] Fedders P A, Fu Y and Drabold D A 1992 Phys. Rev. Lett. 68 1888
[37] Fedders P A and Drabold D A 1993 Phys. Rev. B 47 13 277
[38] Kilian K A, Drabold D A and Adams J B 1993 Phys. Rev. B 48 17 393
[39] Fedders P A 1995 Phys. Rev. B 52 1729
[40] Sankey O F, Niklewski D J, Drabold D A and Dow J D 1990 Phys. Rev. B 41 12 750
[41] Adams G B and Sankey O F 1991 Phys. Rev. Lett. 67 867
[42] Feil H, Zandvliet J W, Tsai M-H, Dow J D and Tsong I S T 1992 Phys. Rev. Lett. 69 3076
 Ab initio tight binding R23

[43] Adams G B, Page J B, Sankey O F, Sinha K, Menendez J and Huffman D R 1991 Phys. Rev. B 44 4052
[44] O’Keeffe M, Adams G B and Sankey O F 1992 Phys. Rev. Lett. 68 2325
[45] Drabold D A, Wang R, Klemm S, Sankey O F and Dow J D 1991 Phys. Rev. B 43 5132
[46] Yang S H, Drabold D A and Adams J B 1993 Phys. Rev. B 48 5261
[47] Alfonso D R, Yang S H and Drabold D A 1994 Phys. Rev. B 50 15 369
[48] Drabold D A, Fedders P A and Stumm P 1994 Phys. Rev. B 49 16 415
[49] Adams G B, Page J B, Sankey O F and O’Keeffe M 1994 Phys. Rev. B 50 17 471
[50] Alfonso D R, Drabold D A and Ulloa S E 1995 Phys. Rev. B 51 1989
[51] Alfonso D R, Drabold D A and Ulloa S E 1995 Phys. Rev. B 51 14 669
[52] Cappelletti R L, Cobb M, Drabold D A and Kamitakahara W A 1995 Phys. Rev. B 52 9133
[53] Demkov A A and Sankey O F 1993 Phys. Rev. B 48 2207
[54] Phillips R, Drabold D A, Lenosky T, Adams G B and Sankey O F 1992 Phys. Rev. B 46 1941
[55] Tsai M-H, Sankey O F and Dow J D 1992 Phys. Rev. B 46 10 464
[56] Yang S H, Drabold D A, Adams J B and Sachdev A 1993 Phys. Rev. B 47 1567
[57] Adams G B, O’Keeffe M, Demkov A A, Sankey O F and Huang Y-M 1994 Phys. Rev. B 49 8048
[58] Caro A, Drabold D A and Sankey O F 1994 Phys. Rev. B 49 6647
[59] Ortega J, Lewis J P and Sankey O F 1994 Phys. Rev. B 50 10 516
[60] Demkov A A, Sankey O F, Schmidt K E, Adams G B and O’Keeffe M 1994 Phys. Rev. B 50 17 001
[61] Demkov A A, Ortega J, Sankey O F and Grumbeth M P 1995 Phys. Rev. B 52 1618
[62] Huang Y M, Spence J C H and Sankey O F 1995 Phys. Rev. Lett. 74 3392
[63] Landman J I, Morgan C G and Schick J T 1995 Phys. Rev. Lett. 74 4007
[64] Park Y K, Estreicher S K, Myles C W and Fedders P A 1995 Phys. Rev. B 52 1718
[65] Demkov A A, Windl W and Sankey O F 1996 Phys. Rev. B 53 11 288
[66] Demkov A A and Sankey O F 1997 Phys. Rev. B 56 10 497
[67] Lewis J P, Ordejon P and Sankey O F 1997 Phys. Rev. B 55 6880
[68] Demkov A A, Sankey O F, Gryko J and McMillan P F 1997 Phys. Rev. B 55 6904
[69] Fritsch J, Sankey O F and Schmidt K E 1998 Phys. Rev. B 57 15 360
[70] Windl W, Sankey O F and Menendez J 1998 Phys. Rev. B 57 2431
[71] Delley B 1990 J. Chem. Phys. 92 508
[72] Seifert G, Porezag D and Frauenheim T 1996 Int. J. Quantum Chem. 58 185
[73] Eschrig H 1979 Phys. Status Solidi b 96 329
[74] Seifert G, Eschrig H and Bieger W 1986 Z. Phys. Chem. 267 529
[75] Frauenheim Th, Weich F, Köhler Th, Uhlmann S, Porezag D and Seifert G 1995 Phys. Rev. B 52 11 492
[76] Widany J, Frauenheim Th, Porezag D, Köhler Th and Seifert G 1996 Phys. Rev. B 53 4443
[77] Sitch P and Frauenheim Th 1996 J. Phys.: Condens. Matter 8 6873
[78] Haugk M, Elsner J and Frauenheim Th 1997 J. Phys.: Condens. Matter 9 7305
[79] Jackson K A, Jungnickel G and Frauenheim Th 1998 Chem. Phys. Lett. 292 235
[80] Elsner J, Jones R, Heggie M I, Sitch P K, Haugk M, Frauenheim Th, Öberg S and Briddon P R 1998 Phys.
Rev. B 58 12 571
[81] Ziman J 1965 Principles in the Theory of Solids (Cambridge: Cambridge University Press)
[82] Andersen O K 1975 Phys. Rev. B 12 3060
[83] Andersen O K, Pawlowska Z and Jepsen O 1986 Phys. Rev. B 34 5253
Andersen O K and O Jepsen 1984 Phys. Rev. Lett. 53 2571
[84] Haydock R, Heine V and Kelly M J 1975 J. Phys. C: Solid State Phys. 8 2591
[85] Heine V, Bullett D W, Haydock R and Kelly M J 1980 Solid State Physics vol 35, ed F Seitz and D Turnbull
(New York: Academic)
[86] Nowak H J, Andersen O K, Fujiwara T and Jepsen O 1991 Phys. Rev. B 44 3577
[87] Bratkovsky A M and Smirnov A V 1993 Phys. Rev. B 48 9606
[88] Bratkovsky A M, Smirnov A V, Nguyen Manh D and Pasturel A 1995 Phys. Rev. B 52 3056
[89] Smirnov A V and Bratkovsky A M 1996 Phys. Rev. B 53 8515
[90] Smirnov A V and Bratkovsky A M 1996 Phys. Rev. B 54 R17 371
[91] Smirnov A V and Bratkovsky A M 1997 Phys. Rev. B 55 14 434
[92] Stokbro K, Chetty N, Jacobsen K W and Norskov J K 1994 Phys. Rev. B 50 10 727
[93] Hansen L B, Stokbro K, Lundqvist B I, Jacobsen K W and Deaven D M 1995 Phys. Rev. Lett. 75 4444
[94] Stokbro K, Jacobsen K W, Norskov J K, Deaven D M, Wang C Z and Ho K M 1996 Surf. Sci. 360 221
[95] Hansen L B, Stokbro K, Lundqvist B I and Jacobsen K W 1996 Mater. Sci. Eng. B 37 185
[96] Elstner M, Porezag D, Jungnickel G, Elsner J, Haugk M, Frauenheim Th, Suhai S and Seifert G 1998 Phys.
Rev. B 58 7260
R24 A P Horsfield and A M Bratkovsky

[97] Horsfield A P, Dunham S and Fujitani H 1998 MRS Proc. (Fall 1998) (Warrendale, PA: Materials Research
Society) Symposium J
[98] Löwdin P 1950 J. Chem. Phys. 18 365
[99] Lin Z and Harris J 1992 J. Phys.: Condens. Matter 4 1055
[100] Ordejon P, Artacho E and Soler J M 1996 Phys. Rev. B 53 R10 431
[101] Bowler D, Fearn M, Goringe C, Horsfield A and Pettifor D 1998 J. Phys.: Condens. Matter 10 3719
[102] Mercer J L, Nelson J S, Wright A F and Stechel E B 1998 Modell. Simul. Mater. Sci. Eng. 6 1
[103] Varshalovich D A, Moskalev A N and Khersonskii V K 1989 Quantum Theory of Angular Momentum
(Singapore: World Scientific)
[104] Landau L D and Lifshitz E M 1977 Quantum Mechanics, Non-Relativistic Theory 3rd edn (Oxford: Pergamon)
[105] Bratkovsky A M and Savrasov S Yu 1990 J. Comput. Phys. 88 243
[106] Andersen O K, Jepsen O and Sob M 1987 Electronic Band Structure and its Applications (Springer Lecture
Notes in Physics, vol 283) ed M Yussouff (Berlin: Springer)
[107] Savrasov S Yu and Savrasov D Yu 1992 Phys. Rev. B 46 12 181
Savrasov S Yu 1992 Phys. Rev. Lett. 69 2819
Savrasov S Yu 1998 Phys. Rev. Lett. 81 2570
[108] Andersen O K, Jepsen O and Krier G 1994 Preprint
Andersen O K, Arcangeli C, Tank R W, Saha-Dasgupta T, Krier G, Jepsen O and Dasgupta I 1998 Proc. Mater.
Res. Soc. vol 491 (Warrendale, PA: Materials Research Society) p 3