1.

Thomas-Fermi method
We consider a system of N electrons in a stationary state, that would obey the stationary
Schrödinger equation:
 
2
h̄ X 1 X
− ∇2i + v(ri , rj ) Ψ(r1 , . . . , rN ) = Ei Ψ(r1 , . . . , rN ) . (1.1)
2m i 2 i6=j

We would try to avoid the complication of searching for the many-electron wavefunction
Ψ(r1 , . . . , rN ), concentrating on an electron density ρ(r) instead. The electron density
is a physical observable, it can be measured, calculated and easily visualized. For just
one particle with its corresponding wavefunction ϕ(r), the density is simply a probability
amplitude to find a particle near a certain position in space; it reads
ρ(r) = ϕ∗ (r) ϕ(r) . (1.2)
For a system with N particles, the definition of the density is the following,
Z
ρ(r) = N Ψ∗ (r, r2 , . . . , rN )Ψ(r, r2 , . . . , rN )dr2 . . . drN , (1.3)

i.e., the probability amplitude to find any particle near the position r in space. In prin-
ciple, one can attempt to solve the equation (1.1) iteratively (subject to certain approx-
imations), i.e. the electron density will be determined by Ψ({r}) and affect, in its turn,
the Coulomb potential in Eq.(1.1), and so on till self-consistency. But the essence of the
Tomas-Fermi method is to circumvent the discussion of the wavefunction Ψ completely,
concentrating on the density ρ(r) as a basic variable to search for.

1.1 Deriving the equation...

The approach has been proposed independently by L. Thomas and E. Fermi in 1927.1 The
Thomas-Fermi approach is semi-classical, i.e., certain ideas will be borrowed from quan-
tum mechanics, but otherwise one operates with normal functions instead of quantum-
mechanical operators. The condition for the semi-classical approach to be applicable
is that spatial variations of the de Broglie wavelength in a system in question must be
small. Specifically, we’ll consider momentum p as a spatial function (instead of quantum-
mechanical operator) and introduce wave number k(x) = 1/h̄p(x) (as a function of
any spatial coordinate x), then de Broglie wavelength λ(x) = 1/k(x), everything being
position-dependent, and the condition for the semi-classical approach to be valid reads
dλ(x)
ξ= 1. (1.4)
dx
1
L. H. Thomas, Proc. Camb. Phil. Soc. 23, 542 (1927); E. Fermi, Rend. Acad. Maz. Lancei 6, 602
(1927). The method has been since then decribed in many textbooks. For a relatively recent useful
review article which covers the history, improvements and applications, see: Larry Spruch, Pedagogic
notes on Thomas-Fermi theory (and on some improvements): atoms, stars, and the stability of bulk
matter, Rev. Mod. Phys. 63, No.1 (Jan. 1991), pp. 151–209.

1
From quantum mechanics, one retains two elements:
• the Fermi statistics; all the states up to those with some maximum energy and hence
momentum pF – that may vary over space – are occupied;
• the uncertainty principle; every cell in the phase space of volume h3 may host up to 2
electrons with opposite spin directions.
In the ground state, assuming first that pF over a volume V , one counts the number
of electrons N:
4π 3 2
pF · V · 3 =N , (1.5)
|3{z } h
|{z}

coordinate
momentum space
space

hence (uniform) electron density is

N 8π
ρ= = 3 p3F , (1.6)
V 3h
or, inverting it:
" #1/3
3h3
pF = ρ . (1.7)
8π
In the following, we assume that both ρ and pF are dependent on r, but so that the
quasi-classsical condition (1.4) holds.
We assume that all electrons move as classical particles in a common potential field
V (r). The classical energy for the fastest electron will be
p2F (r)
Emax. = + V (r) , (1.8)
2m
Whereas both kinetic and potential parts may independently depend on r, their sum –
at least in equilibrium – must remain constant. Otherwise the electrons will flow from
regions with higher maximal energy till Emax. becomes the same everywhere. Substituting
Eq. (1.7) into (1.8) yields:
" #2/3
1 3h3
Emax. = ρ2/3 + V (r) , (1.9)
2m 8π
Now, we consider the total energy of the whole electron distribution and in it – kinetic
and potential energyRparts. With the kinetic energy density t, the kinetic energy T can
be recovered as T = t(r) dr.
p2 1 pF p2 4πp2 · V · 2
Z Z
T 1
t = = dN = · dp
V V 2m V 0 2m h3
p5F
Z p
8π F 4 8π
= p dp = · . (1.10)
2mh3 0 2mh3 5
Substituting (1.7) into (1.10) yields:
" #5/3 2/3
1 3h3 3h2

8π 3
t= · ρ(r) = [ρ(r)]5/3 . (1.11)
2mh3 5 8π |10m {z8π }
CK

2
The potential energy is due to the interaction with external field Vext. (r) and electrostatic
interaction of the electron density with itself:
ρ(r)ρ(r0 )
Z Z
1
U =e ρ(r) Vext. (r) dr + e2 drdr0 . (1.12)
2 |r − r0 |
Then the total energy of the electron system is:
ρ(r)ρ(r0 )
Z Z Z
5/3 1
Etot = CK [ρ(r)] dr + e ρ(r)Vext. (r)dr + e2 drdr0 . (1.13)
2 |r − r0 |
We search in the following for such distribution of the electron density that would minimize
the total energy, subject to the condition of normalization
Z
ρ(r)dr = N : (1.14)

δ (Etot. − µN) = 0 , (1.15)

introducing a Lagrange multiplier µ. A variation with respect to ρ(r) yields:
Z ( )
ρ(r0 )
Z
5
δ (Etot. −µN) = CK [ρ(r)]2/3 + eVext. (r) + e2 dr0 − µ δρ(r)dr = 0
3 |r − r0 |
ρ(r0 )
Z
5 2/3 2
→ CK [ρ(r)] + eVext. (r) + e dr0 = µ , (1.16)
3 |r − r0 |
that is the Thomas-Fermi equation that determins the equilibrium distribution of the
electron density.
Comparing with Eq.(1.9) and taking into account that two last terms on the left side
represent the potential energy, we see that µ is the classical energy of the fastest electron,
that could be identified with the Fermi energy. On the other side, it follows from Eq.(1.15)
that µ = ∂Etot. /∂N, i.e. µ has the meaning of the chemical potential.
Eq. (1.16) is the Thomas-Fermi equation for ρ(r) in integral form. Due to the relation
between ρ and pF , it can be formulated as that on pF , or kF :
 1/3
8π 1 pF 1
ρ1/3 (r) = pF = · = kF .
3h3 (3π 2 )1/3 h̄ (3π 2 )1/3
With this, Eq. (1.16) transforms into:
e2 kF3 (r0 ) 0
Z
5 1 2
CK k (r) + eVext. (r) + 2 dr = µ , (1.17)
3 (3π 2 )2/3 F 3π |r − r0 |
or
e2 kF3 (r0 ) 0
Z
pF (r)
+ eVext. (r) + 2 dr = µ .
2m 3π |r − r0 |
One can transform the Tomas-Fermi equation in the integral form (1.16) into a differential
form. The function to search for will be the electrostatic potential generated by ρ(r). We
introduce this potential as
Z
ρ(r0 )
u(r) = e dr0 ,
|r − r0 |

3
it must satisfy the Poisson equation:

∆u(r) = −4πeρ(r) . (1.18)

From Eq.(1.16) one extracts the electron density,

 3/2
3
ρ(r) = [µ − eVext. (r) − eu(r)]3/2 , (1.19)
5CK

and with (1.18) it follows:

3 3/2
 
∆u(r) = −4πe [µ − eVext. (r) − eu(r)]3/2
5CK
32π 2 e
= − (2m)3/2 [µ − eVext. (r) − eu(r)]3/2 . (1.20)
3h3
This is the Thomas-Fermi equation in differential form.

1.2 Thomas-Fermi atom

Let us consider a neutral Thomas-Fermi atom as an example. Then we have

eZ
N = Z; Vext. (r) = − ⇒ ∆Vext. (r) = 4πeZδ(r) .
r
The energy-minimizing solution of the Thomas-Fermi equation is unique (without proof);
we’ll seqrch for a spherically symmetric solution that will be the solution. Combining
extrenal field (of nucleus) and Coulomb field created by the electron density into Veff (r);

Veff (r) = Vext. (r) + u(r); ∆Veff (r) = 4πeZδ(r) + ∆u(r) .

For r → ∞, ρ(r) → 0, then the Coulomb potential created by spherically symmetric

and neutral charge distribution falls down to zero. In Eq.(1.19) all terms approach 0 as
r → ∞, then µ = 0 at infinity, hence µ = 0 everywhere in the equilibrium. Eq. (1.20)
becomes:
32π 2 e
∆Veff (r) − 4πeZδ(r) = − 3
(2m)3/2 [−eVeff (r)]3/2
3h
32π 2 e
−∆Veff (r) = (2m)3/2 [−eVeff (r)]3/2 − 4πeZδ(r) . (1.21)
3h3
The δ-function takes care of boundary conditions, Veff (r) → −Z/r for r → 0, the other
boundary condition is Veff (r) → 0 for r → ∞, with this the solution of Veff (r) is unique.
For r 6= 0, the equation to solve is

1 d2 32π 2 e
− 2
rV eff (r) = 3
(2m)3/2 e3/2 [−Veff (r)]3/2 . (1.22)
r dr 3h
4
With the ansatz Veff (r) = −Z/rχ(αr) Eq.(1.22) transforms into:

Zα3 d2 χ(x) Zα 3/2 32π 2 e

 
2
= 3
(2me)3/2 [χ(x)]3/2 ;
x dx x 3h
2  1/2 2
d χ(x) Z −3/2 32π e
2
= α 3
(2me)3/2 [χ(x)]3/2
dx x " 3h #
2 2
d χ(x) 1 1/2 32π e 3/2 −3/2
2
= 1/2 Z 3
(2me) α [χ(x)]3/2 . (1.23)
dx x 3h
| {z }
=1
The square bracket can be set to 1 if we choose
!2/3
32π 2 e
α= · 2me Z 1/3 , (1.24)
3h3

then one arrives so at the universal form of the Thomas-Fermi equation that allows to
scale the solution for arbitrary Z and defines χ(x), the universal Thomas-Fermi function.
Let us discuss the asymptotic of the solution. If we assume Veff to behave as

1 d2
Veff ∼ r −ν for r → ∞, then − 2
rVeff ∼ r −(ν+2) ;
r dr
on the other side, [−Veff (r)]3/2 ∼ r −3ν/2 , hence ν + 2 = 3ν/2 ⇒ ν = 4 .

Then ρ(r) ∼ ∆Veff (r) ∼ [−Veff (r)]3/2 ⇒ ∼ r −6 .

Z
For r → 0, Veff (r) ∼ − and ρ(r) ∼ r −3/2 .
r
We summarize below essential properties of a Thomas-Fermi atom:

• There is a solution for neutral or positively charged ion (no bounded solution for
negative ions).

• Asymptotic is wrong: ρ(r) ∼ r −6 for r → ∞, in reality ρ(r) ∼ exp(−r).

ρ(r) ∼ r −3/2 (diverges) for r → 0, in reality: ρ(0) is finite.

• Total energy is too low, especially for light atoms2

• No shell structure of an atom appears.

• There is no chemical binding, the total energy of two close (interacting) atoms is
never lower than that of two distant atoms (the Taylor theorem3 ).

• But: the theory is asymptotically exact for N → ∞.

2
by 54% for hydrogen, by 35% for helium, by ≈15% for heavy elements (Z ≈ 100).
3
Edward Teller, On the Stability of Molecules in the Thomas-Fermi Theory, Rev. Mod. Phys. 34,
No. 4, 627–631 (Oct. 1962)

5
1.3 Corrections to the Thomas-Fermi equation
1.3.1 Exchange (Thomas-Fermi-Dirac)
So far, we considered only electrostatic interaction of each electron with the whole back-
ground charge. What is wrong with this?
• it includes self-interaction. i.e. the charge distribution related to the electron itself is
not excluded:
ρ(r)ρ(r0 )
Z
X 1
6= 0
dr dr0 ;
i6=j
|ri − rj | |r − r |
• it neglects the Pauli principle in a sense that there is no preference in electron-electron
interaction with respect to spin. The second problem can be corrected for. For an electron
with spin, say, “up”, the charge density of spin “down” is reduced in its neighbourhood.
Since this “down” density is not due to the electron in question, it can be handled as a
smooth distribution of charge, with no problem of self-interaction to care about. As will
be discussed larer in relation with the Hartree-Fock formalism, this additional interaction
due to the Pauli principle is the exchange interaction; it lowers the total energy (since it
puts electrons more apart) and can be casted, at least approximatively, in a form explicitly
dependent on density. The incorporation of exchange in the Thomas-Fermi formalism has
been done by Dirac.4 In the following, we discuss only qualitatively what functional
dependence on density could be expected, based on dimensions arguments.
Similarly to the kinetic energy density which scales ∼ ρ5/3 , one can assume a power
low for the exchange energy density. The exchange term arises from electron-electron
Coulomb interaction ∼ e2 /r, so we assume the exchange energy density to scale as ∼ e2 ρν .
The energy density in general must have the dimension [ML−1 T−2 ] (M stands for the units
of mass, L – length, T – time). In our combination ∼ e2 /r considered for the exchange
energy density, e2 has dimension [ML3 T−2 ] since e2 /r is energy. ρ is number per unit
volume, i.e. its dimensionality is L−3 .
ML−1 T−2 = ML3 T−2 L−3ν ⇒ ν = 4/3.
Dirac (1930) suggested a correction term
Z
ETFD = ETF − CX [ρ(r)]4/3 dr (1.25)

to be added to the Thomas-Fermi total energy of Eq.(1.13). The Thomas-Fermi-Dirac

equation reads:
5 4
CK [ρ(r)]2/3 − CX [ρ(r)]1/3 = µ − eVext (r) − u(r) , (1.26)
3 3
as a substitution to Eq. (1.16). We do not specify the magnitude of the exchange constant
CX here.
Considering the performance of the Thomas-fermi theory with Dirac’s exchange term,
one can note that the total energy (already erroneously too low) is lowered even fur-
ther. There are still no bound negative ions. However, neutral atoms get a finite spatial
extension [ρ(r) = 0 from certain radius on].
4
P. A. M. Dirac, Proc. Cambridge Philos. Soc. 26, 376 (1930).

6
1.3.2 Gradient expansions
A more serious problem of the Thomas-Fermi approach is an unsufficiently accurate treat-
ment of kinetic energy. Von Weizsäcker5 considered modified plane waves (1 + ar) exp(ikr)
in order to have an inhomogeneous situation and found a gradient correction term
Z h
1 i2
KW [ρ(r)] = ∇ρ1/2 (r) dr .
2
Later on Kirshnitz used systematic expansion and has shown that the 1st order term is
1/9 of KW . From different points of view, one can achieve coefficients λ, 1/9 ≤ λ ≤ 1, in
front of KW [ρ(r)].
Results: the electronic density at atomic nuclei becomes finite, and at infinity decays
exponentially. Negative ions are formed and molecules may bind. However, it is a problem
in attempting to improve the Thomas-fermi theory systematically, based on gradient
expansions. It was shown that the 6th order in the gradient exopansion already diverges.

5
C. von Weizsäcker, Z. Phys. 96, 431 (1935).

7
 2. Hartree-Fock formalism

2.1 Slater determinants

One speaks of the Hartree-Fock (HF) approximation and of the HF method. The former
is also called the self-consistent field approximation, or mean field approximation. Its
meaning is that every electron moves due to the action of an electrostatic field created by
the presence of all other electrons (plus external field sources). In this sence it is close to
what we applied while deriving the Thomas-Fermi formalism. The difference is that we
won’t deal with the charge density from the very beginning. Instead, we’ll try to trace
what happens with the wave functions. We’ll come back to the concept of density at the
end of the route.
The HF approximation, or the mean-field approximation, is an essential part of the
apparat of theory, often used to solve different models. It introduces an approximation
that sometimes allows to solve a certain physical problem analytically.
The HF method is more about solving complex many-body problems numerically. In
doing so, one essentially applies the self-consistent field approximation. But there are
many approaches that pursue this way – start from a trial wavefunction, calculate the
potential generated by corresponding charge distribution, etc. Essentially, the starting
point is again the stationary Schrödinger equation (1.1). What is specific for the HF
method is the Ansatz used to represent a many-body wavefunction. It is, in the HF
method the wave function is searched for as a Slater determinant constructed from one-
particle wave functions. Lets discuss what this is. If we consider the simplest case of
particles which do not interact, the probability to find each particle near certain place
in space is the same and independent on that of other particles. One would expect the
density ρ(r) to be simply N times the one-particle density,
ρ(r) = Nϕ(r) . (2.1)
On the other side, the density is related with many-body wavefunction via Eq.(1.3). A
possible consistent Ansatz is to construct the many-body wavefunction as a product of
(normalized) individual wave functions, then (2.1) will obviously hold:
Ψ(r1 , . . . , rN ) = ϕ1 (r1 )ϕ2 (r2 ) . . . ϕN (rN ) . (2.2)
This Ansatz is known; it leads to the Hartree method, not in use any more, because
it neglects an important property of electrons – that they are Fermi particles, and the
wavefunction of a system of fermions must be antisymmetric with respect to interchanging
any pair of particles:
Ψ(r1 , . . . rk , . . . rl , . . . rN ) = −Ψ(r1 , . . . rl , . . . rk , . . . rN ) . (2.3)
We emphasize here that the position of an argument in Ψ fixes the number of particular
electron, and the argument value – the actual coordinate of that electron. Obviously

8
enough, the wave function constructed according to (2.2) does not possess this property.
Actually, it does not have any particular symmetry property with respect to interchanging
particles. But we can easily force the wave function to obey Eq.(2.3); for this, we simply
take products like (2.2), interchange there positions of all particles pairwise and construct
the combinations which are a priori antisymmetric. This is easy to do for two particles:
1
Ψ(r1 , r2 ) = √ [ϕ1 (r1 )ϕ2 (r2 ) − ϕ1 (r2 )ϕ2 (r1 )] . (2.4)
2
√
The factor 1/ 2 is introduced in order to keep the normalization. If our one-particle
wave functions are normalized to 1, i.e., the probability to find one electron somewhere
in space Z
|ϕi (r)|2 dr = 1 , (2.5)
then the probability to find two particles anywhere in space must be 1, and indeed

Z 

1 Z Z
2
|Ψ(r1 , r2 )| dr1 dr2 = |ϕ1 (r1 )|2 dr1 |ϕ2 (r1 )|2 dr2 −
2
| {z } | {z }
=1 =1
Z  Z 
− ϕ∗2 (r1 )ϕ1 (r1 )dr1 ϕ∗1 (r2 )ϕ2 (r1 )dr2 −
| {z } | {z }
=0 =0
Z  Z 
− ϕ∗1 (r1 )ϕ2 (r1 )dr1 ϕ∗2 (r2 )ϕ1 (r2 )dr2 +
| {z } | {z }
=0 =0

Z Z 


+ |ϕ2 (r1 )|2 dr1 |ϕ1 (r2 )|2 dr2 =1.

| {z } | {z }

=1 =1

In doing so, we assumed that individual one-particle wavefunctions are all orthogonal.
This is indeed the case if they are solutions of the same Hamilton operator. If they are
not – for some reason – but still form the complete basis of solutions to the corresponding
one-particle problem, they can be one by one orthogonalized before proceeding further
with the construction of the many-body wavefunction.
How do we proceed if there are more than 2 electrons? There are N! possibilities to
interchange them. We sum up over all of them, try all possibilities to interchange every
two electrons and write down antisymmetric terms like in Eq.(2.4). This results in a fully
antisymmetric wave function:
1 X
Ψ(r1 , . . . , rN ) = √ sign(P ) · ϕP 1 (r1 ) ϕP 2(r2 ) . . . ϕP N (rN ) . (2.6)
N! P

P runs over all permutations of electrons, sign(P )=1 for even permutations and sign(P )=-
1 for odd permutations.
It is common and convenient to write down the wavefunctions as determinants, because
(2.6) is similar to how the determinant of a matrix is determined, and that’s why the wave

9
function of this form is called Slater determinant:
ϕ1 (r1 ) · · · ϕ1 (rN )
1 .. ..
Ψ(r1 , . . . , rN ) = √ . . . (2.7)
N!
ϕN (r1 ) · · · ϕN (rN )
The determinant form helps to illustrate important properties of a many-particle wave-
function. Each line corresponds to a certain one-electron state, and each column – to
a certain position in space, among N positions r1 , . . . , rN of particles we consider. The
interchange of either two rows or two columns means that we interchanged two particles;
the wave function then changes sign by construction. Moreover, if there happen two iden-
tical lines or two identical columns, it means that two particles share the same spatial
coordinates; the determinant then equals zero, meaning that such situation is physically
impossible.
Actually, the latter is only true if two electrons in question have the same spin direction;
otherwise, they can well share the same cell in the same phase space, i.e. have the same
one-particle wave function and the same spatial coordinate. In order to allow for that, we
introduce a generalized coordinate incorporating position and spin, x = {r, σ}, and we’ll
R P R
write dx . . . for σ dr . . .. then (2.7) must be written as
ϕ1 (x1 ) · · · ϕ1 (xN )
1 .. ..
Φ(x1 , . . . , xN ) = √ . . . (2.8)
N!
ϕN (x1 ) · · · ϕN (xN )
We shall keep for Slater determinants a special notation Φ, to distinguish them from
general-form Ψ(x1 , . . . , xN ).
So far, we concentrated on symmetry properties of a many-body wavefunction, and
we did not specify the shape of one-particle wavefunctions which constitute a Slater
determinant. They could be taken from the solution of a corresponding one-electron
problem (i.e., neglecting the interaction between the electrons), but it will be hardly a
good approximation to Ψ with interaction. What we’ll do next is to search for the “best”
one-electron wavefunctions ϕ(x) which will allow to construct the “best” approximation
to the “true” many-body wavefunction, keeping the determinantal form (2.8) for our
approximations to Ψ. As a criterion for the “best” functions, we’ll rely on the variational
principle. As in the course of deriving the Thomas-Fermi equation, we’ll search for those
individual ϕ’s which minimize the total energy.
How is all this justified? How do we know that the wave function of the form (2.8)
is a reasonable one? So far, we only used Slater determinants as an “Ansatz” in order
to achieve the antisymmetry of the many-body wave function, without any reference to
its other properties. The importance of Slater determinants lies in the fact that, for a
given number of electrons N, the Slater determinants (constructed from all possible one-
electron functions) form a complete basis set. In other words, if we have the Hilbert space
of one-particle wavefunctions H(1), then the Hilbert space of many-body wavefunctions
is obviously a direct product of one-particle Hilbert spaces:
H(N) = H(1) ⊗ H(1) ⊗ . . . H(1) (2.9)
| {z }
N

10
It means e.g. that any many-body wavefunction can be expanded into a sum of products
of appropriate one-particle functions:
X
Ψ(x1 , . . . , xN ) = aν1 ν2 ...νM ϕν1 (x1 ) ϕν2 (x2 ) . . . ϕνN (xN ); . (2.10)
ν1 ν2 ...νN

It can be shown that the basis of N-particle Slater determinants (constructed from all
possible one-electron wavefunctions in H(1) is a complete basis in H(N), i.e. any anti-
symmetric N-particle wavefunction can be expanded over it:
X
Ψ(x1 . . . xN ) = Cµ Φ(x1 . . . xN ) . (2.11)
µ

This is the so-called theorem of completeness. It emphasizes the importance of Slater

determinants in the treatment of many-electron systems.
Now, the HF approach assumes that from the expansion (2.11) one keeps only one
term, but the possible best one. In the search for the “best” Slater determinant, we vary
the constituent ϕi (xi ) functions.

2.2 Derivation of the HF equations

We start from the Hamiltonian of a many-body system, as in (1.1). We only write it down
in a form that emphasizes the difference between one-particle and two-particle operators:
" #
N
X h̄2 2 1X N
H= − ∇i + u(xi ) + v(xi , xj ) . (2.12)
i=1 2m 2 i6=j

∇i differentiates with respect to coordinates of the ith electron; u(xi ) is one-electron

operator that can typically be the field of a nucleus (or nuclei) in an atom (or a molecule),
situated at Rα :
X Zα e
u(xi ) =
α |ri − Rα |

independently on spin, but it can be spin-dependent as well (e.g., an external magnetic

field). v(xi , xj ) incorporates all interactions which may depend on the coordinates (and
spins) within each particle pair; in the following, it will be everywhere the Coulomb
interaction between two electrons:
e2
v(xi , xj ) = .
|ri − rj |

We construct the total energy as the expectation value of this Hamiltonian, hΦ|H|Φi,
and minimize it in a variational approach – similarly to how we have it done in the
Thomas-Fermi part – under considering the normalization condition:
" N Z #
δ X
hΦ|H|Φi − εi dy ϕi (y) ϕ∗i (y) =0. (2.13)
δϕ∗α (x) i=1

11
The Lagrange multipliers εi take care of the normalization of each one-particle function
ϕi (x) separately. Using (2.6), hΦ|H|Φi becomes:
( " #
N
h̄2 2
Z
1 X X
hΦ|H|Φi = sign (ν)sign(µ) dx1 . . . xN ϕ∗ν1 (x1 ) . . . ϕ∗νN (xN ) − ∇ + u(xi )
N! νµ i=1 2m i

N
1X 
+ v(xi , xj ) ϕν1 (x1 ) . . . ϕνN (xN ) .
2 i6=j 

Let us discuss the effect of one-particle and two-particle contributions in the Hamiltonian
separately. The one-particle operator selects only functions with R
the argument xi ; the
rest is integrated over variables different from xi . The integrals dxk ϕ∗νk (xk )ϕµk (xk ) give
1 if permuted indices νk = µk and zero otherwise; in other words, this condition works
as δνµ under the double sum over permutations. Since only identical permutations ν = µ
contribute, sign(ν) · sign(µ)=1 in all terms, independently on whether the permutation
in question is odd or even; one sum over permutations is lifted, and the remaining one
contains N! terms, all permuting the index i of the one-particle wave function. Taken
together with the summation in i, the sum over permutations gives simply a sum over
orbital numbers 1 to N, each number appearing N! times. The result is:
 
" #
N
h̄2 2
Z
 one-particle  X
∗
 to hΦ|H|Φi =
 dx ϕi (x) − ∇ + u(x) ϕi (x) . (2.14)

contribution i=1 2m

Similarly in the two-particle part, the integration over variables not appearing in the
two-particle interaction term gives 1 whenever the indices of ϕ∗ and ϕ, each obtained by
its corresponding permutation, turn out to be equal. This reduces the sum to:
 
N X
 two-particle  1 1X
 to hΦ|H|Φi = sign (ν)sign(µ) ×

contribution  N! 2 i6=j νµ
Z
× dxi dxj ϕ∗νi (xi )ϕ∗νj (xj ) v(xi , xj ) ϕµi (xi )ϕµj (xj ) .

Since all functions but two have been already used in non-zero terms, there are only two
possibilities how the indices νi , µj may relate:
1) νi = µi and νj = µj ⇒ ν and µ are identical, sign(ν)·sign(µ)=1;
2) νi = µj and νj = µi ⇒ µ is identical to ν, with subsequent interchanging of functions
i and j; sign(ν)·sign(µ)=−1.
In both cases, one can introduce δνµ and lift one summation over permutations, and the
second summartion produces N! identical terms, resulting in:
 
N Z
 two-particle  1X h
 to hΦ|H|Φi = dx dy ϕ∗i (x)ϕ∗j (y)ϕi(x)ϕj (y) −
 contribution  2 i6=j
i
− ϕ∗i (x)ϕ∗j (y)ϕi (y)ϕj (x) v(x, y) . (2.15)

12
Summarizing, the function to be varied in (2.13) is:
" # " #
N N Z
h̄2 2
X Z X
hΦ|H|Φi − εi dyϕi(y)ϕ∗i (y) = dxϕ∗i (x) − ∇ + u(x) ϕi (x) +
i=1 i=1 2m
N Z
1 X h i
+ dx dy ϕ∗i (x)ϕ∗j (y)ϕi(x)ϕj (y) − ϕ∗i (x)ϕ∗j (y)ϕi(y)ϕj (x) v(x, y) −
2 i6=j
N
X Z
− εi ϕ∗i (x) ϕi (x) dx . (2.16)
i=1

Its variation in δϕ∗α (x) gives:

(" #
h̄2 2
Z
δ[. . .] = dx δϕ∗α (x) − ∇ + u(x) ϕα (x)+
2m

XZ h i 
+ dy ϕ∗j (y) ϕα(x) ϕj (y) − ϕ∗j (y) ϕα(y) ϕj (x) v(x, y) − εα ϕα (x) = 0 .
j6=α

The factor 12 in front of the two-particle term disappears because the one-particle function
being varyed, ϕ∗α , may coincide with both ϕ∗i and ϕ∗j . The condition δ [. . .] = 0 for arbitrary
δϕ∗α (x) leads to the HF equation(s):
 
h̄2 2 X Z
− ∇ + u(x) + dy v(x, y) ϕ∗j (y) ϕj (y) ϕα (x) −
2m j6=α
XZ
− dy v(x, y) ϕ∗j (y) ϕj (x) ϕα (y) = εα ϕα (x) . (2.17)
j6=α

The term
XZ
dy v(x, y)ϕ∗j (y) ϕj (y)
j6=α

has a clear meaning of a Coulomb potential which acts on the electron in r due to the
presence of all other electrons. The next sum which can be formally represented as

XZ ϕ∗j (y) ϕ∗α(x) ϕα (y) ϕj (x)

− dy v(x, y) ϕα (x) (2.18)
j6=α ϕ∗α (x) ϕα (x)

introduces the exchange potential acting effectively on the one-particle function in ques-
tion, ϕα (x); it is a correction to the Coulomb potential and is due to the antisymmetry
of the many-particle wave function. This term depends on the unknown function ϕα (x)
itself.
The HF equations is a system of integro-differential equations, which couple N func-
tions. The solution is typically done by iterations. With all functions ϕα (x) found, the
one-determinant many-body wavefunction can be constructed.

13
2.3 Formulation in terms of density
and density matrix
Note that the requirement j6=α can be dropped in (2.17), because for j = α the Coulomb
and exchange terms exactly cancel. The summation in j is over occupied orbitals. With
P
the condition j6=α lifted, the sum in the Coulomb term j ϕ∗j (x) ϕj (x) gives the particle
density defined by Eq. (1.3). Let us show it. In doing so, we’ll generalize (1.3) as depending
on x, i.e., our density will be r-dependent and have a certain spin index. Using the
determinant form of the many-body wave function,
Z
Ψ∗ (x, x2 , . . . , xN )Ψ(x, x2 , . . . , xN ) dx2 . . . dxN =
Z  Z 
1 X
= sign(ν) sign(µ) ϕ∗ν1 (x) ϕµ1 (x) ϕ∗ν2 (x2 )ϕµ2 (x2 )dx2 . . . ϕ∗νN (xN )ϕµN (xN )dxN .
N! νµ

The integrals over x2 . . . xN are non-zero only if all νi = µi (and are =1 in this case),
due to orthonormality of one-electron functions. This demands in non-zero terms to be
ν1 = µ1 as well and hence the permutations ν and µ to be identical, that lifts one sum
in permutations. Then
Z
1 X ∗
Ψ∗ (x, x2 , . . . , xN )Ψ(x, x2 , . . . , xN ) dx2 . . . xN = ϕ (x) ϕν (x) .
N! ν ν
The sum over permutations includes N! terms, among them N that permute the first
index and, for each of them, (N − 1)! possibilities to permute other indices, which are
however not anymore explicitly present. Hence
X X
ϕ∗ν1 (x) ϕν1 (x) = (N −1)! ϕ∗j (x) ϕj (x) .
ν j

Considering the factor N in the definition of ρ(r), Eq. (1.3), one arrives at
X
ρ(x) = ϕ∗j (x) ϕj (x) (2.19)
j

in the HF formalism. So, the Coulomb term in the HF equation (2.17) can be transformed
to explicitly include the (spin)-density. The spin component is implicitly present in x =
{r, σ}. If we wish to express the density irrespectively of spin, it will suffice to sum up
over spin components in x, i.e.,

ρ(r) = ρ(r, ↑) + ρ(r, ↓) .

Now, if we do a similar trick with the exchange term, it will be reduced to the form
containing the density matrix γ(x; y). The definition of the latter is
Z
γ(x; y) = N Ψ∗ (y, x2, . . . , xN ) Ψ(x, x2 , . . . , xN ) dx2 . . . dxN , (2.20)

so that
ρ(x) = γ(x; x) . (2.21)

14
We follow exactly the same arguments as for ρ(x), just keeping y different from x, and
arrive at X
γ(x; y) = ϕ∗j (y) ϕj (x) . (2.22)
j

The HF equation (2.17) then transforms into:

h̄2 2
Z Z
− ∇ + u(x) + dy v(x, y) ρ(y) · ϕα (x) − dy v(x, y) γ(x; y) ϕα(y) = εα ϕα (x) . (2.23)
2m
This can be looked at as an (integro-differential) operator acrting on each one-particle
function ϕα (x):
ĥHF ϕα (x) = εα ϕα (x) ; (2.24)
h̄2 2
Z Z
ĥHF = − ∇ + u(x) + v(x, y) ρ(y) dy − v(x, y) γ(x; y) | {z } dy (2.25)
2m
is called the Fock operator. It is the same for all orbitals and is hermitian. In order to
show that, we construct matrix elements of it between any functions f and g from the
same Hilbert space as one-electron functions:

h̄2
Z Z
hf |ĥHF |gi = − dx f ∗ (x)∇2 g(x) + dx f ∗ (x)u(x)g(x) +
Z 2m Z Z Z
∗ ∗
+ dxf (x)g(x) dy v(x, y)ρ(y) − dxf (x) dy v(x, y)γ(x; y) g(y) ;

on the other hand, using ρ∗ (x) = ρ(x), γ ∗ (x; y) = γ(y; x), v ∗ (x, y) = v(x, y) = v(y, x),
 ∗ h̄2 Z 2 ∗
Z
hg|ĥHF|f i = − dx g(x)∇ f (x) + dx g(x)u(x)f ∗(x) +
Z 2m Z Z Z
+ dxg(x)f ∗ (x) dy v(x, y)ρ(y) − dxg(x) dy v(x, y)γ(x; y) f ∗(y) .

Last terms are identical after renaming intergration variables x ↔ y in one of them, and
the first terms (those with ∇2 ) become identical via integration by parts for the functions
f , g which are zero at infinity. Hence the Fock operator is hermitian and ist eigenvalues
real.

2.4 Meaning of HF eigenvalues; Koopmans theorem

We discuss now the meaning of εα , which have been initially introduced as Lagrange
multipliers taking care of the normalization condition in the variational approach. We
construct the expectation value of the total energy of the system with all N electrons and
with an electron in the state α removed. EHF (N) = hΦ|H|Φi, that is (2.16) without the
last term.

" #
N Z
X h̄2 2
EHF (N − 1α ) = ϕ∗i (x) − ∇ + u(x) ϕi (x) dx +
i6=α 2m

15
 N Z h
1 X i
+ ϕ∗i (x)ϕ∗j (y)ϕi(x)ϕj (y) − ϕ∗i (x)ϕ∗j (y)ϕi(y)ϕj (x) v(x, y) dx dy =
2 i 6= α
j 6= α
" #
h̄2 2
Z
= EHF (N) − ϕ∗α (x) − ∇ + u(x) ϕα (x) dx −
2m
N Z h
1X i
− ϕ∗α (x)ϕ∗j (y)ϕα(x)ϕj (y) − ϕ∗α (x)ϕ∗j (y)ϕα(y)ϕj (x) v(x, y) dx dy −
2 j6=α
N Z
1X
− [ϕ∗i (x)ϕ∗α (y)ϕi(x)ϕα (y) − ϕ∗i (x)ϕ∗α (y)ϕi(y)ϕα (x)] v(x, y) dx dy .
2 i6=α

With i↔j and x↔y two last sums become identical, hence

EHF (N − 1α ) = EHF (N) −

 
h̄2 2 N
Z X Z
− dx ϕ∗α (x)− ∇ + u(x) + dy v(x, y) ϕ∗j (y) ϕj (y) ϕα (x) −
2m j6=α
Z N Z
X
− dx ϕ∗α (x) dy v(x, y) ϕ∗j (y) ϕα(y) ϕj (x)
j6=α

Comparing this with (2.17) we see that the underlined terms together give exactly εα ϕα (x),
then Z
EHF (N − 1α ) = EHF (N) − ϕ∗α (x)εα ϕα (x) dx ;

εα = EHF (N) − EHF (N − 1α ) . (2.26)

This is the theorem of Koopmans. It is important to mention that in the course of
deriving it, we assumed that one-electron wavefunctions remained unchanged, as one
electron was removed. In reality this would not be the case. In a practical calculation,
one would estimate the excitation energy EHF (N) − EHF (N − 1) (e.g., in photoemission)
as difference of total energies in initial and final states with relaxation of one-electron
wavefunctions taken into account, i.e. from two self-consistent calculations correponding
to different electron configurations of the system.
The Hamiltonian operator of a non-relativistic atom commutes with L̂2 , L̂z , Ŝ 2 , Ŝz ;
the exact solutions of many-body Schröderger equation are supposed to do the same. The
Fock operator by its construction does not have any special reason to commute with any
of these operators; hence the most general form of the HF solutions does not have any
particular symmetry properties. A HF orbital of general form looks like

ϕα (x) ≡ ϕ(+) (−)

α (r) χ+ (σ) + ϕα (r) χ− (σ) (2.27)

where χ(σ) are Pauli spinors

( (
1 for σ = 1/2 0 for σ = 1/2
χ+ (σ) = ; χ− (σ) = . (2.28)
0 for σ = −1/2 1 for σ = −1/2

16
The general form of the solution (2.27) allows to treat systems with non-collinear magnetic
density.
From the general form (2.27) one can demand that the solution commutes with Ŝz , i.e.
the spin direction will be fixed for each orbital, and no mixing of σ = 1/2 and σ = −1/2
is possible. This is known as unrestricted Hartree-Fock (UHF), although this is actually a
restriction within a more general formalism. The one-particle wavefunctions in UHF may
be numbered as
!
1
ϕUHF
α (x) = ϕ(+)
α (r) for α = 1, . . . , N (+) ,
0
!
0
ϕUHF
α (x) = ϕ(−)
α (r) for α = N (+) +1, . . . , N ,
1
if α runs through all orbitals. In principle, one could number orbitals corresponding to
both spin directions separately, because they are now completely decoupled. Only the
total number of particles must be conserved:
!
1
ϕUHF
α (x) = ϕ(+)
α (r) for α = 1, . . . , N (+) ,
0
!
(−) 0
ϕUHF
β (x) = ϕβ (r) for β = 1, . . . , N (−) ,
1
N (+) + N (−) = N . (2.29)
These functions are now eigenfunctions of the spin projection operator,
 n o
 + h̄ ϕUHF
α (x) for α ∈ 1, . . . , N (+) ,
2
Ŝz ϕUHF
α (x) = n o
 − h̄ ϕUHF (x) for α ∈ 1, . . . , N (−) ,
2 α

and the action of Ŝz on the Slater determinant gives:

h̄ h (+) i
Ŝz ΦUHF (x1 , . . . , xN ) =
N − N (−) ΦUHF (x1 , . . . , xN ) .
2
So, the essence of UHF is that it allows to treat magnetic systems. Not only the orbitals,
but the density matrices are now labeled by spin index:
N (±)
(±)
X ∗
γ 0
(r; r ) = ϕ(±)
α (r0 ) ϕ(±)
α (r) ; (2.30)
α

ρ(±) (r) = γ (±) (r; r) , (2.31)

and the HF system splits into two parts, labeled by spin:
" #
h̄2 2
Z Z
− ∇ + u(r) + dr ρ(r ) v(r, r ) ϕ(+)
0 0
α (r) −
0
dr0 γ (+) (r; r0 ) v(r, r0) ϕ(+) 0
α (r ) =
2m
= ε(+) (+)
α ϕα (r) ;
" #
h̄2 2
Z Z
− ∇ + u(r) + dr0 ρ(r0 ) v(r, r0) ϕ(−)
α (r) − dr0 γ (−) (r; r0 ) v(r, r0) ϕα(−) (r0 ) =
2m
= ε(−) (−)
α ϕα (r) , (2.32)

17
where the external field u(r) may be also spin-dependent (e.g., in the presence of a mag-
netic field). The coupling between two parts is achieved by the fact that the total density

ρ(r) = ρ(+) (r) + ρ(−) (r)

enters both equations; moreover, the total number of particles is constant, N = N (+) +
N (−) . The numbers N (+) and N (−) however are not fixed and may vary (from iteration to
iteration). In a practical calculation, one would expand the orbitals over a basis set which
is larger than N, and as in the course of solution on each iteration one gets eigenvalues,
separately for (+) and (−), the N lowest in energy among them all will be occupied, that
determines N (+) and N (−) , ρ(+) , ρ(−) etc. for this iteration.6
It is noteworthy that the density matrices γ (±) (r; r0 ) are labeled by spin. This is so
because the exchange interaction only involves the one-electron states with the same spin
state. The effect of exchange can be looked at as the Coulomb interaction of an orbital
searched for, say, ϕ(+) (+) 0
α (r) with a corresponding “exchange density” ρrα (r ):
"Z #
(+)
(r; r0)v(r, r0)ϕ(+) 0
Z
0 (+) 0γ α (r )
dr γ 0
(r; r )v(r, r 0
)ϕ(+) 0
α (r ) ⇒ dr (+)
ϕ(+)
α (r) ;
ϕα (r)

γ (+) (r; r0 )ϕ(+) 0

α (r )
ρ(+) 0
rα (r ) = (+)
. (2.33)
ϕα (r)
The exchange density is non-local (it depends on both r and r0 ), and for every orbital
index α it integrates over r0 exactly to 1:
Z N (+) (+) N (+)
XZ (+) ∗ ϕi (r) X
dr0 ρ(+) 0
r,α (r ) = dr0 ϕi (r0 )ϕ(+) 0
α (r ) (+)
= δiα = 1 ,
i=1 | {z } ϕα (r) i=1
δiα

that means that each electron is surrounded by its corresponding “exchange hole”, from
where the charge density if exactly one electron of the same spin is excluded.
We proceed discussing symmetry aspects. If one demands that individual orbitals do
possess a good angular momentum value, for instance can be casted as

ϕν (x) = Rnν (r)Ylν mν (θ, φ)χσ (σ)

(in an atom), then one arrives at the so-called restricted Hartree-Fock (RHF) formalism.
For a multiatomic system, one can generalize this demand by taking into account ap-
propriate (by symmetry) combinations of atom-centered functions with the same angular
momentum value. Then for each orbital
ˆlz ϕν (x) = h̄mν ϕν (x) ,

and for the Slater determinant

!
X
L̂z Φ(x1 , . . . , xN ) = h̄ mν ΦRHF (x1 , . . . , xN ) .
ν

6
this scheme is refered to as the aufbau principle.

18
The more severe the symmetry constraint is, the more restricted the variational space
for one-electron wavefunctions and the higher the calculated ground-state energies. The
ultimately best energies (the HF limit) can be obtained only with freely variable orbitals
which do not possess any particular symmetry properties. This situation is known as the
“symmetry dilemma” in the HF formalism.
The exchange density ρ(+) 0
r,α (r ) introduced in Eq.(2.33) is non local, as it should be from
physical considerations, but the fact that it depends on the orbital index is not physically
motivated. Slater (1951) proposed7 to weight it over occupied orbitals, according to their
corresponding partial densities:
P (+) P
(+) (+) i ρri (r0 ) ϕ∗i (r)ϕi (r) ij ϕ∗j (r0 )ϕj (r)ϕ∗i (r)ϕi (r0 )
ρr (r0 ) = ρX (r, r0) = P ∗ = P ∗ . (2.34)
j ϕj (r)ϕj (r) j ϕj (r)ϕj (r)

(+)
After summation, the eXchange density ρX (r, r0) does not depend on the orbital index
anymore.
We perform now this summation analitically for the easiest case of free particles, with
one-electron eigenvalues
1
ϕk (r) = √ eikr ,
V
where V is the volume of the“box” including the N electrons, and periodical boundary
conditions are assumed. The summation will be substituted by the integration in the
momentum space up to kF = pF /h̄, so that the lowest states with energies up to EF =
h̄2 kF2 /2m are occupied. kF is related to density as defined by Eq. (1.7), kF = (3π 2 ρ)1/3 ,
and Z k
X F 2V 3 V
→ dN ; dN = 3
d k = 3 k 2 dk sin θk dθk dφk .
i 0 (2π) 4π
When the electron gas is not spin polarized, N (+) = N (−) = N/2, the summation over
orbitals of each spin component runs over
V 2
dN (±) = k dk sin θk dθk dφk .
8π 3
The denominator of Eq.(2.34) gives:
X V 1 4π 3 kF3 ρ
ϕ∗j (r)ϕj (r) → 3 kF = 2 = . (2.35)
j 8π V 3 6π 2

The numerator of Eq.(2.34)

 2 Z
X V kF 1 ik(r0 −r)
ϕ∗j (r0 )ϕj (r)ϕ∗i (r)ϕi (r0 ) → k 0 2 dk 0 sin θk0 dθk0 dφk0 e ×
ij 8π 3 |0 {z V }
P
i
Z kF 1 0
× k 2 dk sin θk dθk dφk eik(r−r ) .
|0 {z V }
P
j

7
Phys. Rev. 81, 385 (1951)

19
Each of the integrals in k yields:
Z kF Z kF Z π
2 ikR 2
k dk sin θk dθk dφk e = 2π k dk sin θk dθk eikR cos θk =
0 0 0
Z
4π kF 4π
= k dk sin(kR) = [sin(kF R) − kF R cos(kF R)] .
R 0 R3
Finally
X
ϕ∗j (r0 )ϕj (r)ϕ∗i (r)ϕi(r0 ) →
ij
" #2
0 0 0
1 3 2 sin (kF |r − r |) − kF |r − r | cos (kF |r − r |)
→ (4πk F ) . (2.36)
(8π 3 )2 (kF |r − r0 |)3
Substituting (2.35) and (2.36) into (2.34), we get
" #2
(+) 3kF3 sin (kF |r − r0 |) − kF |r − r0 | cos (kF |r − r0 |)
ρX (r, r0) = (2.37)
2π 2 (kF |r − r0 |)3
" #2
9 j1 (kF |r − r0 |)
= ρ .
2 kF |r − r0 |
The plot of the function involved is shown in Fig. 2.1.
0.15 0.001
((sin(x)-x*cos(x))/x**3)**2 ((sin(x)-x*cos(x))/x**3)**2

0.1

0.0005

0.05

0 0
0 2 4 6 8 0 5 10 15 20
 2
j1 (x)
Figure 2.1: Plot of x

We now come back to the HF equations (2.32) where the exchange term was trans-
formed according to Eq.(2.33)
Z Z
− dr0 γ (+) (r; r0 ) v(r, r0) ϕ(+) 0 0 (+) 0 0 (+)
α (r ) → − dr ρrα (r )v(r, r ) ϕα (r)

so it took a form of a potential acting on the orbital we search for,

(+)
→ vX (r)ϕ(+)
α (r) ,

20
and now that after statistical averaging ρ(+)
rα does not depend on the orbital index α
anymore, the exchange potential is the same for all orbitals. Using the result (2.37) for
the exchange density,
(+)
ρX (r, r0 ) 0
Z
(+)
vX (r) = −e dr =
|r − r0 |
(centering the coordinate system at r)
" #2
1 3k 3 sin(kF r 0 ) − (kF r 0 ) cos(kF r 0 )
Z ∞
= −4πe r dr 0 F2
02 0
0 r 2π (kFr 0 )3
sin x − x cos x 2
 1/3 Z    1/3
3 1/3
∞ 3 3
= −6e ρ x dx 3
= −e ρ1/3 . (2.38)
π | 0
{z x } 2 π
1/4

We compare this result with (1.26) from the Thomas-Fermi section and see that the
exchange potential is indeed proportional to ρ1/3 , as was argued there based on dimen-
sionality considerations. The prefactor must be dependent on the spatial distribution of
density; it is constant in this case because we assumed homogeneous density distribution
(+)
in the derivation. Otherwise, vX (r) would have got an explicite dependence on r. We
(+)
note that, although the exchange density ρX (r, r0) integrates over r0 exactly to 1 for every
R 0 (+)
r, the value of the integral dr ρX (r, r0)/|r − r0 | would of course depend on the shape
(+)
of ρX (r, r0 ). However, since the Coulomb interaction 1/|r − r0 | has spherical symmetry,
(+)
only spherically averaged part of ρX (r, r0 ), i.e. dependent only on |r − r0 |, will contribute
to the value of the exchange potential at any given position r.
Finally, we note that one can cast the exchange energy in the HF scheme, after a
statistical averaging, in the form of a spatial integral obver exhange energy density,
Z
E X = dr X (r) ;

obviously X ∼ ρ4/3 for the homogeneous electron gas. This result will be later on used
and generalized for the case of a slowly varying density.

2.5 Pair correlation function

The exhange density ρX (r) has the meaning of the particle density expelled from the
vicinity of each electron as a consequence of the Pauli principle. The same effect can be
characterized by a dimensionless property, a pair correlation function, which shows how
any two particles tend to avoid each other. We start from some general definitions. In
addition to the 1st order density matrix previously defined in Eq. (2.20), we introduce
now the two-particle density matrix,

N(N −1)Z ∗
γ2 (x1 , x2 ; y1 , y2) = Ψ (y1 , y2 , x3 , . . . , xN )Ψ(x1 , x2 , x3 , . . . , xN ) dx3 . . . dxN .
2!
(2.39)

21
2! in the denominator stands for generality, showing how to introduce higher-order density
matrices. Obviously γ2 is related to the 1st order matrix γ by
Z
2
γ(x; y) = γ2 (x, x2 ; y, x2)dx2 . (2.40)
N −1
Taking a diagonal of γ2 we obtain another property depending on the coordinates of two
particles, that is the pair density ρ2 :
ρ2 (x, y) = 2γ2 (x, y; x, y) . (2.41)
From this, two different pair correlation functions are derived:
ρ2 (x, y)
g(x, y) = ; (2.42)
ρ(x)ρ(y)
h(x, y) = ρ2 (x, y) − ρ(x)ρ(y) . (2.43)
For large spatial distances between particles, g → 1 and h → 0.
These definitions are quite general. In order to conclude the HF part, we calculate
pair density in the HF approximation for the same model case, a homogeneous electron
gas. First we construct the two-particle density matrix for Slater determinants. Similarly
to how we proceeded for the one-particle density resulting in Eq. (2.19),
N(N − 1) X
γ2 (x1 , x2 ; y1 , y2) = sign(ν) sign(µ) ×
2N! νµ
N Z
Y 
× ϕ∗ν1 (y1 ) ϕ∗ν2 (y2 )ϕµ1 (x1 ) ϕµ2 (x2 ) dxi ϕ∗νi (xi ) ϕµi (xi ) .
i=3

In analogy with considering two-particle contributions to hΨ|H|Ψi, Eq. (2.15), we conclude

that νi = µi for i ≥ 3, that leaves the combinations
ν1 = µ1, ν2 = µ2; sign(ν) sign(µ) = 1;
ν1 = µ2, ν2 = µ1; sign(ν) sign(µ) = −1
and (N −2)1 possibilities to permute other indices. Finally
N h
1X
γ2 (x1 , x2 ; y1 , y2) = ϕi (x1 )ϕj (x2 )ϕ∗i (y1 )ϕ∗j (y2 ) −
2 i,j
i
− ϕi (x1 )ϕj (x2 )ϕ∗j (y1 )ϕ∗i (y2 ) . (2.44)
This can be further on written down as
 # #
N N
1 hX hX
γ2 (x1 , x2 ; y1, y2 ) = ϕ∗i (y1 ) ϕi (x1 ) ϕ∗j (y2 ) ϕj (x2 ) −
2 i j
"N #" N #
X X 
− ϕ∗i (y2 ) ϕi (x1 ) ϕ∗j (y1 ) ϕj (x2 )

i j

1 γ(x1 ; y1 ) γ(x1 ; y2)

= . (2.45)
2 γ(x2 ; y1 ) γ(x2 ; y2)

22
Note that this does not hold not in the general case anymore, but only for Slater deter-
minants. Using (2.41), we obtain for ρ2 (x, y):
1 γ(x; x) γ(x; y)
ρ2 (x, y) = 2 · = ρ(x) ρ(y) − |γ(x; y)|2 . (2.46)
2 γ(y; x) γ(y; y)
Now, in order to specify spatial dependencies in the density matrices and two-particle
density, we must decide how are occupied one-particle states distributed over two spin
components. Assuming paramagnetic electron gas, N (+) = N (−) = N/2,
N/2 " !
0
X
0 0
X (+) ∗
  1 (+)
γ(r, r ) = γ(rσ; r σ ) = ϕi (r0 ) 10 ϕi (r)+
σσ0 i=1
0
! !
(−) ∗
  1 (+) (+) ∗
  0 (−)
+ϕi (r0 ) 01 ϕi (r) + ϕi (r0 ) 10 ϕi (r) +
0 1
| {z } | {z }
=0 =0
! #
(−) ∗
  0 (−)
+ ϕi (r0 ) 01 ϕi (r) =
1
N/2 h i
X (+) ∗ (+) (−) ∗ (−)
= ϕi (r0 )ϕi (r) + ϕi (r0 )ϕi (r) =
i=1
(+)
= γ (r; r0 ) + γ (−) (r; r0) . (2.47)
From this it follows
ρ(r) = ρ(+) (r) + ρ(−) (r) ,
as it should be.
For γ2 (r1 , r2 ; r01 , r02 ) we have to consider 24 =16 terms with different spin attribution.
(+) (−) (+) (−)
We keep only those which do not contain ϕi i (...)ϕi (...) or ϕj i (...)ϕj (...) and hence
are not orthogonal in spin space. What is explicitly left in the paramagnetic (PM) case:
X
γ2PM (r1 , r2 ; r01 , r02 ) = γ2 (r1 σ1 , r2 σ2 ; r01 σ10 , r02 σ20 ) =
σ1 σ2
σ0 σ0
1 2
N/2 "
1 X (+) ∗ (+) ∗ (+) (+)
= ϕi (r01 ) ϕj (r02 ) ϕi (r1 ) ϕj (r2 ) +
2 i,j=1
(+) ∗ (−) ∗ (+) (−)
+ ϕi (r01 ) ϕj (r02 ) ϕi (r1 ) ϕj (r2 ) +
(−) ∗ (+) ∗ (−) (+)
+ ϕi (r01 ) ϕj (r02 ) ϕi (r1 ) ϕj (r2 ) +
(−) ∗ (−) ∗ (−) (−)
+ ϕi (r01 ) ϕj (r02 ) ϕi (r1 ) ϕj (r2 ) −
(+) ∗ (+) ∗ (+) (+)
− ϕi (r01 ) ϕj (r02 ) ϕj (r1 ) ϕi (r2 ) −
#
(−) ∗ (−) ∗ (−) (−)
− ϕi (r01 ) ϕj (r02 ) ϕj (r1 ) ϕi (r2 ) =
N/2 "
1 X
= 4 ϕi (r1 ) ϕj (r2 ) ϕ∗i (r01 ) ϕ∗j (r02 ) −
2 i,j=1
#
− 2 ϕi(r1 ) ϕj (r2 ) ϕ∗i (r02 ) ϕ∗j (r01 ) . (2.48)

23
From this,
N/2 "
X
ρPM
2 (r1 , r2 ) = 4 ϕi(r1 ) ϕj (r2 ) ϕ∗i (r01 ) ϕ∗j (r02 ) − 2 ϕi(r1 ) ϕj (r2 ) ϕ∗i (r02 ) ϕ∗j (r01 ) =
i,j=1
     
N/2 N/2 N/2 N/2
X X X X
= 4 |ϕi (r1 )|2  |ϕj (r2 )|2  − 2 ϕi (r1 )ϕ∗i (r2 ) ϕj (r2 )ϕ∗j (r1 ) =
i j i j
1
= ρ(r1 ) ρ(r2 ) − |γ(r1 ; r2 )|2 . (2.49)
2
The pair correlation function g(r1 , r2) is then

ρ2 (r1 , r2 ) 1 |γ(r1 ; r2 )|2

g PM (r1 , r2 ) = =1− . (2.50)
ρ(r1 )ρ(r2 ) 2 ρ(r1 )ρ(r2 )

For the gas of non-interacting electrons with constant density,

N/2
X V Z 2 eik(r1 −r2 )
γ(r1 ; r2 ) = 2 ϕ∗i (r2 ) ϕi(r1 ) ⇒ 3 k dk sin θk dθk dφk
i 4π V

and calculated as has been already done above, leading to Eq.(2.36). Finally
" #2
09 sin(kF r) − (kF r) cos(kF r)
g(r ≡ |r − r |) = 1 − . (2.51)
2 (kF r)3

For fully spin-polarized (“ferromagnetic”) electron gas, one would assume

N (+) = N; N (−) = 0; ρ(r) = ρ(+) (r);

N/2 " ! #
FM
X X (+) ∗
  1 (+)
γ 0
(r; r ) = 0 0
γ(rσ; r σ ) = ϕi (r0 ) 10 ϕi (r) = γ (+) (r, r0 ) ;
σσ0 i=1
0

N/2 "
1 X (+) ∗ (+) ∗ (+) (+)
γ2FM (r1 , r2 ; r01 , r02 ) = ϕi (r01 ) ϕj (r02 ) ϕi (r1 ) ϕj (r2 ) −
2 i,j=1
#
(+) ∗ (+) ∗ (+) (+)
− ϕj (r01 ) ϕi (r02 ) ϕi (r1 ) ϕj (r2 ) ;

2
ρFM
2 (r1 , r2 ) = ρ(r1 ) ρ(r2 ) − |γ(r1 ; r2 )| ; (2.52)
ρFM
2 (r1 , r2 ) |γ(r1 ; r2 )|2
g FM (r1 , r2 ) = =1− . (2.53)
ρ(r1 )ρ(r2 ) ρ(r1 )ρ(r2 )
Comparing with (2.50), one can see that the correlation function is zero at the origin,
since all electrons have the same spin now. k becomes larger and hence the fluctuations
in the correlation function have smaller period than in the case of non-magnetic electron
gas of the same density.

24
 3. Total energy as functional
of electron density

3.1 Expectation values

of one- and two-particle operators
We remember the separation of a Hamiltonian into one-particle and two-particle operators
in Eq. (2.12). It can be shown that the expectation value of any one-particle operator
can be expressed in terms of one-particle density matrix, that of two-particle operator –
in terms of two-particle matrix etc. We consider an arbitrary single-particle operator
N
X
F̂ = ˆ i) ,
f(x (3.1)
i=1

then its expectation value in a many-body state Ψ is

Z N
X
hF̂ i = ∗
dx1 . . . dxN Ψ (x1 . . . xN ) ˆ i ) Ψ(x1 . . . xn ) =
f(x
i=1
Z N
X Z 
= dx fˆ(x) ∗
dx2 . . . dxn Ψ (y, x2 . . . xN )Ψ(x, x2 . . . xN ) .
i=1 y=x

In doing so, we used the fact that simultaneous interchange of any two arguments in both
Ψ∗ and Ψ does not change their product,
Ψ∗ (x1 . . . xi . . . xN )Ψ(x1 . . . xi . . . xN ) = Ψ∗ (xi . . . x1 . . . , xN )Ψ(xi . . . x1 . . . xN ) ,
and the integration parameter y different from x was introduced in Ψ∗ in order to ensure
that the operator fˆ acts on Ψ only, not on Ψ∗ . For a simple multiplication (e.g., fˆ
being an external potential) this is not important so that one can simplify in the last
formula [· · ·]y=x ⇒ dx2 . . . dxn Ψ∗ (x, x2 . . . xN )Ψ(x, x2 . . . xN ), but if, e.g., fˆ = ∇2 , then
R

it is essential to keep first arguments in Ψ∗ and Ψ formally different and set y = x after
the integration. With the definition of the density matrix (2.20),
Z h i
hF̂ i = fˆ(x) γ(x; y) dx , (3.2)
y=x

or, considering both fˆ and γ as operators in the x-space,

hF̂ i = tr (fˆ γ) , (3.3)
For an operator fˆ depending only on the coordinates of particles (like external electrostatic
field, or kinetic energy) the summation over spin components can be done in Eq. (3.2),
resulting in Z h i
hF̂ i = dr fˆ(r) γ(r; r0) 0 . (3.4)
r =r

25
 If the spin degrees of freedom matter for the one-particle operator in question, then its
expectation values can be expressed via spin-dependent single-particle density matrices.
P
The vector spin operator σ̂ = α eα σ̂α is defined via cartesian unit vectors eα and Pauli
matrices
! ! !
1 1 0 1 0 1 1 0 −i
σ̂z = , σ̂x = , σ̂y = , (3.5)
2 0 −1 2 1 0 2 i 0

then the vector spin density will be

N
X
Ŝ(r) = σ̂i δ(r − ri ) , (3.6)
i=1

and the spin magnetization density

X
m(r) = 2 hŜ(r)i = 2 σ̂s0 s ρss0 (r) = 2 trs [σ̂ γ(r)] (3.7)
ss0

with !
ρ++ (r) ρ+− (r)
ρss0 (r) = ,
ρ−+ (r) ρ−− (r)
then 
2 Re ρ−+ (r)


m(r) = 2 Im ρ−+ (r) . (3.8)

 ++ −−
ρ (r) − ρ (r)
1 P
Similarly, for any two-particle operator V̂ = 2
v̂(xi , xj )
i6=j

Z
1X
h V̂ i = dx1 . . . dxN Ψ∗ (x1 . . . xN )v̂(xi , xj )Ψ(x1 . . . xN ) =
2 i6=j
Z Z
1X
= dx dy Ψ∗ (x, y, x3 . . . xN ) v̂(x, y) Ψ(x, y, x3 . . . xN ) dx3 . . . dxN =
2 i6=j
(using the same argument as above, interchanging xi ↔ x, xj ↔ y in both Ψ∗ and Ψ)
Z
= dx dy [v̂(x, y) γ2(x, y; x0, y 0)] x0 =x . (3.9)
y 0 =y

Again, we introduced x0 independent from x and y 0 independent from y in order to prevent

that operator v̂(x, y) acts on the arguments in Ψ∗ . However, if v̂(x, y) is simple multipli-
cation (as is the case with Coulombic two-particle interaction), this is not necessary, so
one can write down using Eq. (2.41)
Z
1
h V̂ i = dx dy v(x, y) ρ2 (x, y) . (3.10)
2
Therefore the expectation value of any operator in a certain state of a many-electron
system can be obtained through the knowledge of two functions, the single-particle density
matrix γ(x, y) defined in Eq. (2.20) and the pair density ρ2 (x, y) defined in Eq. (2.41).

26
3.2 Hohenberg–Kohn theorem
Of course the last statement applies to the total energy as the expectation value of the
Hamilton operator in a particular state. However, if we consider specifically the ground
state, the total energy must be extremal, that allows to formulate a more strong condition,
as was done by Pierre Hohenberg and Walter Kohn in 19648:
In the ground state, the total energy is a functional of the electron density ρ(r).
That means actually that not only one doesn’t need to know density matrix and the
pair density to determine the ground-state total energy, but the dependence of the latter
on spin variables is not relevant as well. We’ll prove this ad absurdum, following the
original derivation of Hohenberg and Kohn. Underway, we’ll distinguish between what is
sometimes referred to as the first Hohenberg–Kohn theorem and the second one.
We start by specifying once again the many-electron Hamiltonian,
h̄2 X
N XN
1X N
H[u] = − ∇2i + u(xi ) + v(xi , vj ) .
2m i=1 i=1 2 i6=j
| {z } | {z } | {z }
≡ T̂ ≡ Û ≡ V̂

u(xi ) is an external potential, and in the following we assume we are free to vary it.
Each external potential defines a corresponding ground-state wavefunction via solution
of the Schrödinger equation. Such wavefunctions are of course different for u1 and u2 , if
u1 (r) 6= u2 (r) + const.. What about the particle density? It results from the integration
over N −1 variables in the wavefunction. It will turn out that still, there is no way to
get ρ[u1 ] the same as ρ[u2 ] if the potentials differ by more than a constant. But first we
assume that this is possible. Let Ψ1 be the ground-state wavefunction corresponding to
the external potential u1 , then the corresponding total energy is:
Z
E1 = hΨ1 |H[u1]|Ψ1 i = hΨ1 |T̂ + V̂ |Ψ1 i + u1 (r) ρ(r) dr .
This must be lower than the expectation value obtained with any other wavefunction Ψ2 :
E1 < hΨ2 | H[u1] |Ψ2 i
E1 < hΨ2 | H[u2] − u2 + u1 |Ψ2 i
Z Z
E1 < hΨ2 | H[u2] |Ψ2 i + dr [u1(r) − u2 (r)] N Ψ∗2 (r2 . . . rN )Ψ2 (r2 . . . rN ) dr2 . . . drN .
| {z }
E2
| {z }
= ρ(r)

On the other side, the ground-state energy E2 calculated with its “true” wavefunction Ψ2
must be lower than the expectation value of H[u2 ] calculated with any other wavefunction,
including Ψ1 :
E2 < hΨ1 | H[u2] |Ψ1 i
E2 < hΨ1 | H[u1] − u1 + u2 |Ψ1 i
Z Z
E2 < hΨ1 | H[u1] |Ψ1 i + dr [u2(r) − u1 (r)] N Ψ∗1 (r2 . . . rN )Ψ1 (r2 . . . rN ) dr2 . . . drN .
| {z }
E1
| {z }
= ρ(r)

8
Inhomogeneous Electron Gas, Phys. Rev. 136, B864 (1964)

27
Summing up two inequalities, we arrive at E1 +E2 < E2 +E1 , that is a contradiction. The
origin of this contradiction is that we assumed ρ(r) to be the same for two wavefunctions
generated by two different potentials u1 and u2 . Hence this assumption was wrong, and the
external potential uniquely determines the density. Since ρ[u1 ] 6= ρ[u2 ] for u1 6= u2 +const,
for any given ρ(r) there is at most one potential u(r) for which ρ(r) is the ground-state
density.
At this point, one sometimes singles out as the first Hohenberg–Kohn theorem:
For isolated many-electron system, its ground-state one-electron
density ρ(r) determines uniquely the external potential u(r)
and as the second Hohenberg–Kohn theorem:
The exact ground state energy E = E[Ψ0 ] of many-electron system with external
potential u(r) is a functional of the associated ground-state electron density ρ0 (r).
This second statement follows from the fact that since u(r) uniquely fixes H and hence
the many-particle ground state, the latter must be a unique functional of ρ(r):
Z
E [ρ0 (r)] = T [Ψ0 ] + V [Ψ0 ] + dr u(r) ρ(r) . (3.11)

As a brief discussion, it is important to emphasize the following:

– this is simply an existence theorem, which tells absolutely nothing about the actual
dependence of the kinetic part T [Ψ0 [u[ρ(r)]]] and the electron-electron interaction part
V [Ψ0 [u[ρ(r)]]] in the total energy on particle density. This is a serious challenge for
practical applications.
– there is a problem of u-representability, i.e. a question of whether an arbitrary function
ρ(r) can be mapped onto corresponding u(r). There has been a history of works aimed
at more strict mathematical formulation.

3.3 Kohn–Sham equation

There are two essential problems of applying Eq. (3.11) to practical calculations: the
concretization of T [ρ] and V [ρ]. The Thomas–Fermi theory provides approximations for
both, but they are not accurate enough, particularly in what regards the kinetic energy
part. Walter Kohn and Lu Sham in 19659 proposed a trick allowing to represent the
kinetic energy better.
In a many-electron system without interaction (V switched off), the ground-state
wavefunction (when non-degenerate) is a single Slater determinant. Then the kinetic
energy can be easily evaluated:
1
H0 = T̂0 + u(r); Ψ0 = √ det kϕi (xk )k ;
N!
h̄2 X
N
T0 [Ψ0 ] = hΨ0 | T̂ |Ψ0 i = − hϕi | ∇2i |ϕi i . (3.12)
2m i=1
9
Phys. Rev. 140, A1133 (1965)

28
 P
The particle density is ρ(r) = ϕ∗i (x)ϕi (x), as in the HF formalism, Eq. (2.19). With
is
the interaction actually present, the representation for the density can still be applied,
understanding that ϕi (x) are not true one-particle wavefunctions anymore, but just any
support functions, or pseudofunctions, of our convenience to represent the density. This
is fully general. Further on, we approximate the kinetic energy of a true system by T̂0
constructed from {ϕi } according to Eq. (3.12). This introduces an error, which is further
on attributed to the yet undefined part, which describes electron-electron interaction. the
functions ϕi (x) are searched for using variational approach (instead of searching for ρ(x)
directly, as in the Thomas-Fermi formalism). Once {ϕi } are found, we can immediately
reconstruct the density. Specifically, the total energy from Eq. (3.11)
e2 Z ρ(x)ρ(x0 ) Z
Etot = T [ρ] + 0
dx dx0 + EXC [ρ] + e u(x)ρ(x) dx (3.13)
2 |r − r |
| {z } | {z }
Coulomb interaction, corrections:
including exchange +
self-interaction correlation,
so good we can
| {z }
V [ρ]
will be substituted by
e2 ρ(x)ρ(x0 )
Z Z
Etot = T0 [ρ] + 0
dx dx0 + EXC
0
[ρ] + e u(x)ρ(x) dx . (3.14)
2 |r − r |
| {z } | {z }
kinetic energy EXC +T [ρ]−T0 [ρ]
of non-interacting
system of N particles
with the same density

We apply the variational approach, searching for those pseudofunctions ϕi (x) which would
minimize the total energy. From Eq. (3.12)
δT0 h̄2 2 δρ(x0 )
= − ∇ ϕ α (x) ; = ϕα (x) δ(x − x0 ) ;
δϕ∗α (x) 2m δϕ∗α (x)
" Z # "Z #
δ 1 ρ(x)ρ(x0 ) 0 ρ(x)
dx dx = dx ϕα (x) ;
δϕ∗α (x) 2 |r − r0 | |r − r0 |
hence
(
e2 ρ(x)ρ(x0 )
Z Z
δ
T0 [ρ] + dx dx0 + EXC
0
[ρ] + e u(x)ρ(x) dx−
δϕ∗α (x) 2 |r − r0 |
Z 
− εα ρ(x) dr − N =0. (3.15)

where the Lagrange multipliers εα , as before, take care of maintaining the normalization
of the density in the course of variation. The variation of (yet unknown) XC-functional

29
in ϕ∗α can be reformulated via its variation in ρ:
0 Z 0 0 0
δEXC 0 δEXC δρ(x ) δEXC
= dx = ϕα (x) .
δϕ∗α (x) δρ(x0 ) δϕ∗α (x) δρ(x)
| {z } | {z }
0
δ(x−x ) ϕα (x) XC-potential,
VXC (x)

Eq. (3.13) then reduces to

" #
h̄2 2 Z
ρ(x0 )
− ∇ + e2 dx0 + u(x) + VXC (x) − εα ϕα (x) = 0 . (3.16)
2m |r − r0 |
This is the system of equations resembling that of the HF method after introducing
the Slater averaging for exchange – the equations are linear in ϕα , which mix via the
contribution of all of them to the total density. The Slater averaged exchange potential
VX can be considered as an approximation to a more general VXC . VXC is a simple function
of r (and spin), but it depends on the distribution of density in the whole space.
However, the present derivation is more general; it does not impose any particular
restrictions on the wave function (we assumed for it a single-determinant form only in the
absence of electron-electron interaction, in order to construct T0 ).
• This rearrangement puts everything yet unknown into VXC (x); the rest is explicitly
defined.
• The equations give ϕα and then total density, but provide absolutely no information
about the ground-state wavefunction (not even about the density matrix).
• Everyting applies to the ground state only – that is nevertheless quite important, because
problems like equilibrium structure, magnetization, phonons, etc. fall into this category.
• This approach essentially fixes the problems of the Thomas-Fermi method related to
poor description of the kinetic energy.
The closed form for the ground-state total energy in the Kohn–Sham approach can be
obtained from Eq. (3.14), using T0 as defined by Eq. (3.12):
!Z
N
X h̄2
T0 [ρ] = − ϕ∗i (x)∇2 ϕi (x) dx
i=1 2m
from Eq. (3.16)
" #
h̄2 2 ρ(x0 )
Z
− ∇ ϕα (x) = εα ϕα (x) − u(x) + VXC (x) + e2 dx0 ϕα (x) ,
2m |r − r0 |
that results in
N Z
X N Z
X
Etot = ϕ∗i (x) εi ϕi (x) dx − ϕ∗i (x) VXC (x) ϕi (x)dx −
i=1 i=1
N Z
ρ(x0 ) e2 ρ(x)ρ(x0 )
X Z Z
− e2 dx ϕ∗i (x) ϕi (x) dx0
+ dx dx0 + EXC
0
[ρ] =
i=1 |r − r0 | 2 |r − r0 |
N
e2 ρ(x)ρ(x0 )
X Z Z
= εi − 0
dx dx0 − 0
VXC (x) ρ(x) dx + EXC [ρ] . (3.17)
i=1 2 |r − r |

30
{εi } is a set of N lowest eigenvalues of the Kohn–Sham equations, similarly to the aufbau
pronciple in the HF method.

3.4 Relation between exchange-correlation energy and

electron pair correlation function
We already discussed earlier that the exchange and correlation are essentially due to a
Coulomb interaction of each electron with electron density missing in its environmet,
so no wonder that EXC can be directly related to the pair density ρ2 (x, y) definned by
0
Eq. (2.41). However, one must consider that EXC incorporates as well the difference
between the true kinetic energy T and that of non-interacting electron gas of the same
density, T0 . We model this situation by an adiabatic approximation: scale the electron-
electron interaction in the Hamiltonian with λ, 0 ≤ λ ≤ 1, and assure that the density
ρ(x) remains the same for all values of λ by adding a corresponding λ-dependent external
potential uλ :
X (λ) X
H = H0 + λW = ĥi + λ ŵij . (3.18)
i ij

For all λ along the part 0→1, there is a groud-state function Ψλ and total energy Eλ .
Taking into account

∂ ∂Ψλ ∂Ψλ
hΨλ |Ψλi = 0 = h |Ψλ i + hΨλ | i,
∂λ ∂λ ∂λ
we get:

∂ D ∂Ψ
λ
E D ∂H E D ∂Ψλ E
Eλ = H Ψλ + Ψλ Ψλ + Ψλ H =
∂λ ∂λ
" ∂λ # ∂λ
D ∂Ψ E D ∂Ψλ E D ∂H E
λ
= Eλ Ψλ + Ψλ + Ψλ Ψλ =
∂λ ∂λ ∂λ
D E
= Ψλ W Ψλ . (3.19)

This is the Hellmann–Feynman theorem. Then the total energy can be recovered as

Z1
E = E0 + dλ hΨλ| W |Ψλ i , (3.20)
0

Assuming that we switch on the interaction adiabatically, i.e. dλ/dt → 0 and the system
is everywere in the ground state, we specify the Hamiltonian as

X e2 X λ
H = H0 + uλ (xi ) + ;
i 2 i6=j |ri − rj |
dH X duλ (xi ) e2 X 1
= + .
dλ i dλ 2 i6=j |ri − rj |

31
To calculate its expectation value, we use Eq. (3.2) and (3.10). Since both operators here
are simply numbers, one does not need to care about y → x in (3.2) and writes simply in
terms of single and pair density:
Z
D X duλ(xi ) E duλ(x)
Ψλ Ψλ = ρ(x) dx ; (3.21)
i dλ dλ
D X 1 E Z
ρ(x, x0 )
Ψλ Ψλ = 0|
dx dx0 , (3.22)
i6=j |r i − r j | |r − r

We express the pair density in terms of pair correlation function g (2.42), which gets the
dependennce on λ (the single-particle density is the same for all λ, but the correlation
function and hence the pair density certainly not):

ρ(x, x0 ) = gλ (x, x0 ) ρ(x) ρ(x0 ) . (3.23)

e2 ρ(x)ρ(x0 )
Z Z
dEλ ∂uλ (x)
= ρ(x) dx + gλ (x, x0 ) dx dx0 .
dλ ∂λ 2 |r − r0 |
Then
Z1 Z
dEλ
E − E0 = dλ = ρ(x) [uλ=1 (x) − uλ=0 (x)] dx +
dλ
0
1
e2 Z ρ(x) ρ(x0 ) 0
Z
+ 0
dx dx gλ (x, x0 ) dλ . (3.24)
2 |r − r |
0

Without interaction, Z
E0 = T0 [ρ] + ρ(x) uλ=0 (x) dx ,
with full interaction (3.14)
e2 ρ(x)ρ(x0 )
Z Z
E = T0 [ρ] + dx dx0 + 0
ρ(x) uλ=1 (x) dx + EXC [ρ] ;
2 |r − r0 |
hence
e2 Z ρ(x)ρ(x0 ) 0
Z
0
E − E0 = dx dx + ρ(x) [uλ=1 (x) − uλ=0 (x)] dx + EXC [ρ] ,
2 |r − r0 |
and comparing with (3.24) one gets:
e2 ρ(x)ρ(x0 )
Z
0
EXC [ρ] = 0
[g̃(x, x0 ) − 1] dx dx0 . (3.25)
2 |r − r |
with
Z1
0
g̃(x, x ) = gλ (x, x0 ) dλ . (3.26)
0
Hennce the exact exchange-correlation energy is fully determined by the shape of the
exchange-correlation hole (but, for the full range of interaction strengths).

32
3.5 EXC for homogeneous electron liquid
A large amount of useful information about general properties of EXC [ρ] has been obtained
in the course of simulating a simple benchmark system, homogeneous electron liquid. This
is a system of interacting electrons with homogeneous density, compensated by equally
homogeneous background positive charge.
We begin with some general considerations, in order to incorporate the effect of spin
polarization in the following treatment. We start from the exchange energy and then
generalize the results over the correlation energy. Introducing explicitly spin variables,
we get
(+) (−)
e2 Z Z ρX (r, r0 ) ρ(+) (r0 ) 0 e2 Z Z ρX (r, r0) ρ(−) (r0 )
EX = − dr dr − dr dr0 , (3.27)
2 |r − r0 | 2 |r − r0 |
since exchange interaction does not “mix” two spin components. Therefore
EX [ρ(+) , ρ(−) ] = EX [ρ(+) ] + EX [ρ(−) ] . (3.28)
Further on, we introduce exchange energy density X (r) that satisfies
Z
EX = dr X (r) ρ(r) (3.29)

without loss of generality.10 Comparing it with Eq. (3.27)

(+) (−)
e2 ρX (r, r0 ) ρ(+) (r0 ) 0 e2 ρX (r, r0 ) ρ(−) (r0 ) 0
Z Z
ρ(r) X (r) = − dr − dr . (3.30)
2 |r − r0 | 2 |r − r0 |
For uniform electron gas of constant density ρ(+) and ρ(−) , we may perform the integration
in Eq. (3.30) in analogy with (2.38). We emphasize that it doesn’t matter how the total
density is split into ρ(+) and ρ(−) ; the integration limit in k-space is fixed by the ρ value
in each spin channel separately. Therefore we use, instead of (2.37) where it was assumed
ρ(+) = ρ/2, explicitly
" #2
0
(+) 0 (+) j1 (kF |r − r |)
ρX (r, r ) = 9ρ ;
kF |r − r0 |
instead of (2.35) we have
kF3 (+)
 
2 (+) 1/3
= ρ ; k F = 6π ρ .
6π 2
The integration over r00 = r0 − r, with fixed r, yields:
" #2 Z∞ " #2
j1 (kF |r − r0 |) ρ(+) j1 (kF r 00 )
Z
(+) 0
h i
(+) 2 00 2 00 1
9ρ dr = 9 ρ · 4π r dr =
kF |r − r0 | |r − r0 | kF r 00 r 00
0

36π h i4/3 Z∞(sin x − x cos x)2 

3
1/3 h i4/3
= ρ(+) dx = 3 ρ(+) . (3.31)
(6π 2 )2/3 x5 4π
0
| {z }
1/4

10
this is not yet local density approximation as long as X (r) depends on the density in all space, not
just at r.

33
Further on, we introduce spin polarization ζ(r)

ρ(+) − ρ(−)
ζ= with ρ = ρ(+) + ρ(−) , (3.32)
ρ
whence
ρ ρ
ρ(+) = (1 + ζ), ρ(−) = (1 − ζ); , (3.33)
2 2
and exchange energy density, now dependent on spin polarization, becomes:
1/3
3e2 ρ1/3 h

3 4/3 4/3
i
X (r, ζ) = − (1 + ζ) + (1 − ζ) . (3.34)
2 4π 24/3
In the following, we’ll use instead of ρ the density parameter rs , that is the radius of a
sphere inclusing unit charge,
!1/3
4π 3 3
r ρ=1 → rs = . (3.35)
3 s 4πρ

Note that rs is r-dependent for inhomogeneous density distribution. With this definition,
1/3
3e2 (1 + ζ)4/3 + (1 − ζ)4/3

9π
X (r, ζ) = − . (3.36)
4π 4 2rs
In what regards the dependence on the spin polarization, the exchange energy density
can be looked at as an interpolation between limiting values of “paramagnetic” (ζ=0)
and “ferromagnetic” (ζ=1) cases,
1/3 1/3
3e2 3e2 24/3
 
9π 2 9π
X (r, 0) = − · , X (r, 1) = − · ,
4π 4 2rs 4π 4 2rs
so that
X (r, ζ) = X (r, 0) + [X (r, 1) − X (r, 0)] f (ζ), (3.37)
1
with
f (ζ)
(1 + ζ)4/3 + (1 − ζ)4/3 − 2
f (ζ) = (3.38)
24/3 − 2
known as the von Barth – Hedin interpolation function
0
−1 0 1 (see plot).
Von Barth and Hedin calculated11 the correlation energy of spin-polarized electron
liquid in the lowest order in the random phase approximation for different values of ζ
and rs . It follows that the same interpolation function is accurate enough for correlation
energy density, so that one can use:

XC (r, ζ) = XC (r, 0) + [XC (r, 1) − XC (r, 0)] f (ζ) . (3.39)
11
U. von Barth and L. Hedin, A local exchange-correlation potential for the spin polarized case: I,
J. Phys. C: 5, 1629 (1972).

34
The exact results analysis of XC (rs ) are not so far known even for the simplest model
case of free electron liquid, or jellium – a homogeneous distribution of interacting electrons
at the background of uniformly smeared compensating positive charge. The analytic
results have been otained (up to certain acccuracy) only for limiting cases of low and high
densities. The high-density limit (rs → 0) in the paramagnetic case is given by

 1/3
3 9π 1
XC (rs , 0) = − already described by Eq. (3.36)
4π 4 rs
1 − ln 2 W. Macke,
+ ln rs
π2 Z. Naturforsch. 5a, 192 (1950)

+B 
 Carr + Maradudin,
+ Crs ln rs
 Phys. Rev. 133, A371 (1964)
+ Drs 

+...

The solution for the ferromagnetic case is known in the random phase approximation:12

1 RPA 4
RPA
XC (ρ, 1) =  (2 ρ, 0) . (3.40)
2 XC

This scaling is consistent with Eq. (3.36) for the exchange energy density.
The low-density limit is essentially that of electron crystallization, i.e. their localiza-
tion near well separated (and ordered) positions is space. The possibility of such behaviour
at sufficiently low densities was first recognized by E. Wigner.13 Wigner compared dif-
ferent crytsal lattices by their contributions to the total energy – essentially, Madelung
terms evaluated by Ewald summation. The results obtained for several lattice types are:
−1.79186 rs−1 Ry (body centered cubic)
−1.79172 (face centered cubic)
−1.79168 (hexagonal close-packed)
−1.760... (simple cubic)
As is seen, the bcc lattice seems to be the most probable candidate for the Wigner
crystallization of jellium, but it is not strictly proven, and at least fcc and hcp lattices
become competitive. The excitations from the “crystallized” state are electron lattice
vibrations, or “phonons”. With zero-point energy of such “phonons” taken into account,
the perturbation expansion in the low-density limit is in powers of rs−1/2 and yields (in

12
L. Hedin, Phys, Rev. 139, A796 (1965)
13
E. Wigner, Trans. Faraday Soc. 34, 678 (1938)

35
 14
the paramagnetic case) for the energy in Ry per electron – see, e.g., Carr (1961):

1.792 Madelung energy of Wigner crystal

E=−
rs
2.66 zero-point energy
+ 3/2
rs of electron lattice vibrations

+ b2 
rs  perturbations to
!
+O 1 

 oscillator functions; b < 1
rs5/2


+ terms ∼ exp(−rs1/2 )

Fig. 1 of Carr (1961). Energy of the

electron gas plotted against rs . The
open circles are points given by the
above formula with anharmonic terms
neglected. The filled circles were calcu-
lated from the Gell-Mann and Brueck-
ner equation [Phys. Rev. 106, 364
(1957)]. In constructing the dashed
line connecting the two end regions,
consideration was given to the fact that
it must fall below the sum of Fermi
plus exchange energy for plane waves.
The open circle points below rs =6 have
lower accuracy, since for these points
the exponential terms are appreciable.

The above figure shows the estimations of exchange-correlation energy of homogeneous

electron liquid obtained by perturbation expansion in two limiting cases. The solid curve
is a free interpolation connecting the data is small rs and large rs regions. The region
of “relevant” rs values is 2 – 7 a.u., where none of limiting approximations works well.
A number of calculations has been done by Hedin and Lundqvist, von Barth and Hedin,
Gunnarsson and Lundqvist15 to bridge the region of intermediate rs for both paramag-
netic and ferromagnetic cases. For example, the parametrization of the correlation energy
density by Hedin and Lundqvist16 (ζ=0 case), adopted also by von Barth and Hedin for
the ζ=1 case, uses the function
 
1 z 1
F (z) = (1 + z 3 ) ln 1 + + − z2 −
z 2 3
14
W. J. Carr, Jr., Energy, Specific Heat, and Magnetic Properties of the Low-Density Electron Gas,
Phys. Rev. 122, 1437 (1961).
15
O. Gunnarsson and B. I. Lundqvist, Exchange and correlation in atoms, molecules, and solids by the
spin-density-functional formalism, Phys. Rev. B 13, 4274 (1976).
16
L. Hedin and B. I. Lundqvist, Explicit local exchange-correlation potentials, J. Phys. C 4, 2064 (1971).

36
and    
rs rs
(rs , 0) = −C PM F PM
, (rs , 1) = −C FM F FM ,
r r
with parameter values (rs in a.u., energy in Ry):

C PM = 0.0504; C FM = 0.0254; r PM = 30; r FM = 75.

This is consistent (within 1%) with the scaling relation (3.40) for the random phase
approximation:
1
C FM = C PM ; r FM = 24/3 r PM .
2
For more accurate parametrizations in the intermediate range of rs , it was extremely im-
portant that (ultimately exact) results of Monte Carlo simulation became available due to
Ceperley and Alder.17 Their scheme was a search for “optimized” many-body wavefunc-
tion and its corresponding energy in a diffusion process, starting from a trial wavefunction
and an arbitrary reference energy. The trial function was a product of two-body correla-
tion factors (that remove singularities in the local energy as electrons approach each other)
times a Slater determinant, constructed either from plane-wave states (for Fermi liquid),
or from bcc-centered Gaussians (for the Wigner crystal). Paramagnetic or ferromagnetic
case was fixed in advance.
With the Hamiltonian of usual form,

h̄2 X
N
e2 X 1
H= ∇2i − ,
2m i=1 2 i6=j |ri − rj |

the parameter space R includes 3N spatial coordinates of the system of N electrons. A

random walk in dR3N at time t is given by f (R, t) · dR3N , f being the probability density.
The value of f at t=0 is |ΨT (R)|, e.g., starting trial function. For large times,

f (R, t → ∞) = ΨT (R) ϕ0 exp[−t(Eref − E0 )],

with ϕ0 as asymptotic (“true”) wavefunction and E0 – “true” eigenvalue. The diffusion

equation for f (R, t) is
" #
∂f h̄2 X
N   h HΨ
T
i
= ∇2i f − ∇i f ∇i ln |ΨT |2 − −Eref f
∂t 2m i=1 ΨT
| {z }
trial
energy

and implies:
– random diffusion;
– drift induced by the trial “quantum force” ∇ ln |ΨT |2
(since ln |ΨT |2 has the meaning of energy);
– branching with probability = (time step)×(Etrial − Eref ).
17
D. M. Ceperley and B. J. Alder, Ground State of the Electron Gas by a Stochastic Method,
Phys. Rev. Lett. 45, 566 (1980).

37
For Etrial > Eref , the configuration is removed from the ensemble, for Etrial < Eref –
duplicated in the ensemble.
The results of Ceperley and Alder are organized in a table of ground-state energies
for several simulation objects (paramagnetic Fermi liquid, ferromagnetic Fermi liquid,
Wigner crystal, and Bose liquid) over a range of rs . These results can be also represented
as the following “phase diagram”:

Fig. 2 of Ceperley and Alder. The energy of the four phases studied relative to that of the lowest
boson state times rs2 in rydbergs vs rs in Bohr radii. Below rs =160 the Bose fluid is the most
stable phase, while above, the Wigner crystal is most stable. The energies of the polarized and
unpolarized Fermi fluid are seen to intersect at rs =75. The polarized (Ferromagnetic) Fermi
fluid is stable between rs =75 and rs =100, the Fermi Wigner crystal above rs =100, and the
normal paramagnetic Fermi fluid below rs =75.

A number of analytical expressions interpolating the results by Ceperley and Alder has
been proposed, which are now widely in use (Perdew and Zunger 1981 and others).

3.6 Local density approximation

The expression used for the exchange-correlation energy, in analogy with Eq. (3.29),
Z
EXC = dr XC (r) ρ(r) ,

is quite general, as long as we assume XC to be a function of r, dependent on the whole

distribution of density, i.e., XC [ρ(r)]. However, the benchmark studies we just discussed
refer to the case of homogeneous density and are not, strictly speaking, applicable to vary-
ing ρ(r). The local density approximation (LDA) represents an ill-justified but reasonable

38
generalization, that is expected to work at least for the case of slowly varying density:
Z  
LDA
EXC [ρ] = dr ρ(r) XC ρ(r), ζ(r) . (3.41)

We’ll see that the accuracy of this approximation is better than could be a priori ex-
pected. The exchange-correlation potential VXC (r) as introduced in by Eq. (3.16) can be
generalized for the spin-polarited case as follows:

δEXC δρ(+) (x0 )

Z
δEXC δEXC
= (+)
dx0 = (+) ϕα (x)
δϕ∗α (x) 0 ∗
δρ (x ) δϕα (x) δρ
| {z }
δ(x0 −x)ϕα (x)

is ϕα contributes to, say, spin (+), so that

(+) ∂ n  o
VXC (r) = ρ(r)  ρ(r), ζ(r) .
∂ρ(+) (r)

There is no more integration in spatial coordinates since

Z Z
∂ ∂...
. . . dr0 = δ(r − r0 ) dr0 .
∂ρ(r) ∂ρ(r)

The differentiation in spin-resolved components of density, taking into account relations

(3.32), reduces to

∂ ∂ ∂ρ ∂ ∂ζ ∂ 1−ζ ∂
(+)
= (+)
+ (+)
= + ;
∂ρ ∂ρ ∂ρ ∂ζ ∂ρ ∂ρ ρ ∂ζ
∂ ∂ 1+ζ ∂
(−)
= − .
∂ρ ∂ρ ρ ∂ζ

Then
   
(±)
  ∂XC ρ(r), ζ(r) ∂XC ρ(r), ζ(r)
VXC (r) = XC ρ(r), ζ(r) + ρ(r) ± [1 ∓ ζ(r)]
∂ρ ∂ζ
 
LDA (±)
≡ VXC ρ(r), ζ(r) . (3.42)
 
Since the energy won’t change in case of simultaneous reversal of all spins, XC ρ(r), ζ(r) =
 
XC ρ(r), −ζ(r) , and

∂XC (ρ, −ζ) ∂XC (ρ, ζ) (+) (−)

=− , hence VXC (ρ, −ζ) = VXC (ρ, ζ) .
∂ζ ∂ζ

Although seemingly poorly justified, the LDA works in practical calculations surprizingly
well. The reason is that the XC energy is the integral over the XC hole of density times
Coulomb interaction, see Eq. (3.25). The Coulomb interaction is spherically symmetric
and leaves only spherically averaged, around r, part of ρXC (r, r0) to contribute to EXC .

39
The discussion to this effect can be found in the paper by Gunnarsson et al.18 and in the
review by Jones and Gunnarsson19 and can be illustrated by the following figures, which
show the cuts of the exact exchange hole in the Ne atom in comparison with its local
density approximation. The actual shape of the exchange hole is drastically different
from the approximated one (the true exchange hole in an atom has maximum at the
nucleus whereas the averaged one is centered at the electron). However, the results of
averaging over angles (with the center taken at the electron) are much closer for exact
and LDA cases; moreover, only the radial integral of the angular-averaged function (times
density and the Coulomb factor) actually affects the exchange energy.

Fig. 5 of the Gunnarsson et al. (1979) paper. The exchange hole around an electron at r in
the neon atom shown as a function of distance from the electron along a line connecting the
electron and the nucleus. Full curves: the exact hole, dashed curves: the LDA result. Two
panels correspond to the electron at two different distances from the nucleus, 0.09 and 0.4 a.u.

Fig. 7 of the Gunnarsson et al. (1979) paper. Spherical average of the neon exchange hole
shown above.

18
O. Gunnarsson, M. Jonson and B. I. Lundqvist, Description of exchange and correlation effects in
inhomogeneous electron systems, Phys. Rev. B 20, 3136 (1979).
19
R. O. Jones and O. Gunnarsson, The density functional formalism, its applications and prospects,
Rev. Mod. Phys. 61, 689 (1989).

40
 Qualitatively, the success of the LDA even for such unhomogeneous densities as that in
atoms can be traced to the fact that an important sum rule is by construction implemented
in the LDA: the exchange-correlation hole is integrated to −1 everywhere in space, i.e.,
for EXC [ρ] generally defined as
e2 ρ(r) ρXC (r, r0 − r)
Z
EXC [ρ] = dr dr0 , (3.43)
2 |r−r0|
Z
ρXC (r, r0 −r) dr0 = −1 for all r . (3.44)
If one separates exchange-correlation hole into exact exchange part (as would follow from
the HF treatment) and the correlation part as the rest, the corresponding sum rules would
be Z Z
0 0
ρX (r, r −r) dr = −1 ; ρC (r, r0 −r) dr0 = 0 .
The LDA substitutes
e2 ρXC (r, r0 −r) 0
Z  
dr ⇒ XC ρ(r)
2 |r − r0 |
where the XC-energy density is some function of ρ(r) only. Because of this, density may
appear only as a prefactor to the integral over r0 and not in the integral over r0 :

ρLDA 0 0
XC (r, r − r) = ρ(r) [g̃(r, r ) − 1] ,

where g̃ is the correlation function integrated over interaction strength, see Eq. (3.26).
The performance of LDA in real systems with chemical bonding (solids, molecules) has
been addressed in large number of publications. When discussing ground-state properties,
one can note that the equilibrium volume is usually within −5 − 0% of the experiment,
i.e. almost always underestimated, that reflects the fact that the LDA somehow over-
estimates the chemical bonding. Results for the bulk modulus, or generally on elastic
constants, vary from very good to hardly satisfactory; actually calculated elastic proper-
ties are quite sensitive to a practical realization of DFT calculation (i.e., on the accuracy
of representing a general-shape charge density). The same applies to the calculation of
phonon frequencies. For cohesive energies, the following table gives some idea about the
accuracy of LDA results. MJW (Moruzzi – Janak – Willliams) and VW (Vosko – Wilk)
are two different “flavours”, i.e. different parametrizations of the LDA energy density.

Cohesive energy (in eV) calculated using the LDA

Li Na K Be Sr Al
LDA-MJW 1.65 1.12 0.90 3.97 1.89 3.84
LDA-VW 1.74 1.21 0.96
Exp. 1.66 1.26 0.94 3.33 1.70 3.34

Still, the error of LDA in reproducing ground-state properties is sometimes larger than
acceptable; there are some known errors in the equation of state (wrong crystal structure
in the ground state of Fe etc.)

41
3.7 Self-interaction correction
One of shortcomings of LDA is the presence of spurious self-interaction. It leads to no-
ticeable errors in light atoms (especially hydrogen) and in systems with strongly localized
states. Let us consider the hydrogen atom,

ρ(+) = |φ(r)|2 ; ρ(−) = 0 ; ζ = 1.

The exact expression for the XC energy would be:

e2 Z |φ(r)|2 |φ(r0 )|2

EXC = − drdr0 ,
2 |r − r0 |

that exactly cancels Hartree energy (there is no electron density to interact with). Com-
paring to the general expression (3.25), we see that this corresponds to vanishing cor-
relation function, g̃(r; r0 ) = 0. In LDA, the XC energy is constructed implying some
non-zero correlation function. Explicitly for hydrogen atom, the electrostatic self-energy
(e.g., Hartree term) of 8.5 eV is canceled by the XC energy of −8.1 eV to ≈95%.
A possible (not quite satisfactory) cure is to introduce explicit dependency on the
index of one-particle functions (Kohn–Sham orbitals) into the functional to be optimized.
This procedure is called self-interaction correction (SIC):
X
E SIC = E LDA [ρ(r), ζ(r)] − Θiσ , (3.45)
iσ
e2 ρiσ (r) ρiσ (r0 )
Z  
0 LDA
with Θiσ = dr dr + E ρiσ (r), 1 . (3.46)
2 |r − r0 |
Each correction term Θiσ involves the contribution to the density from the Kohn–Sham or-
bital in question, which, if occupied, carries one electron and is hence fully spin-polarized:
(
|ϕiσ (r)|2 , ϕiσ occupied,
ρiσ (r) =
0, otherwise

Similarly to the derivation of Kohn–Sham equations, Eq. (3.16), the variation of individual
ϕiσ yields:
" #
h̄2 2 ρ(x0 )
Z
− ∇ + e2 SIC
dx + u(x) + VXC,α (x) − εα ϕα (x) = 0 , (3.47)
2m |r − r0 |
SIC
where VXC,α (x), now taking place of VXC (x), is:

ρα (r0 ) 0
Z
SIC
VXC,α (x) = VXC [ρ, ζ](x) − VXC [ρα , 1](x) − dr . (3.48)
|r − r0 |

In a sense, some similarity to the HF treatment is thus recovered (since there is no spurious
self-interaction in the HF), but VXC (x) may well include correlation effects beyond the
exchange as well. The problems in the SIC scheme are the following:
• The Hamiltonian is formally different for different Kohn–Sham functions, that results

42
in non-orthogonality of the orbitals;
• The scheme is not invariant under a unitary transformation of the orbitals (i.e., different
results may be obtained when using the basis of spherical harmonics or cubic harmonics).
Other useful extensions beyond the LDA can be sorted out into two groups. One group
of approximations deals with certain assumptions about the shape of the XC hole, using,
for instance, appropriate model functions. The other group of approximations proceeds
with treating inhomogeneities in a systematic way – either by Taylor expansion, or Fourier
expansion, of patially varying density. Let us consider several examples from both groups.

3.8 Averaged density approximation and

Weighted density approximation
These two approximations have been proposed and discussed in the above cited paper by
Gunnarsson – Jonson – Lundqvist (1979). The Averaged density approximation (ADA)
uses, instead of density ρ at r only, an average of density over some vicinity of r. The
XC-density is then represented by
Z1 h   i
ρADA 0 0
XC (r, r −r ) = ρ(r) g r−r0, λ, ρ(r) − 1 dλ , (3.49)
0

where g is some reasonable model for the pair correlation function. The averaging pro-
cedure for ρ is not uniquely predetermined, and the freedom existing in its choice can be
exploited to satisfy certain important criteria, like the sum rule (3.44) and consistency
with LDA results in the limit of constant density. Using such criteria, Gunnarsson emphet
al. (1979) constructed the differential equation for the averaging function w that yields
Z  
ρ(r) = w r−r0 , ρ(r) ρ(r0 ) dr0 ,

and solved this equation numerically. This approximation has not been widely used so
far.
The Weighted density approximation (WDA) uses another prescription for the XC-
density,  
ρWDA
XC (r, r 0
−r 0
) = ρ(r 0
) G |r−r 0
|, ρ̃(r) (3.50)
that has the advantage of keeping a correct prefactor in the XC-density, compare Eqs. (3.25)
and (3.43). ρ̃(r) is again an in some way averaged density, and G(r, ρ) is a model for the
pair correlation function that doesn’t have to be exact; one is free to choose a convenient
analytical form which would guarantee the sum rule and correct asymptotics. A choice
by Gunnarsson and Jones20
( " #)
λ(ρ)
G(r, ρ) = C(ρ) 1 − exp − 5 . (3.51)
|r|
20
O. Gunnarsson and R. O. Jones, Phys. Scr. 21, 394 (1980); R. O. Jones and O. Gunnarsson, The
density functional formalism, its applications and prospects, Rev. Mod. Phys. 61, 689 (1989).

43
Two conditions, Z
ρ G (|r|, ρ) dr = −1
and
e2
Z
G (|r|, ρ)
ρ dr = XC (ρ) ,
2 |r|
allow to determine C(ρ) and λ(ρ). The asymptotic of (3.51) is such that several important
limiting cases are described correctly. The WDA can be constructed to be exact in the
following cases:
(1) for homogeneous system;
(2) for one-electron system, such as hydrogen atom (exact cancellation of the electron
self-interaction);
(3) for an atom, it provides correct asymptotic of the XC-energy far from nucleus,

e2
XC (r) = −
;
2r
(4) far outside the crystal surface, the correct asymptotic of image potential is recovered:

e2
XC (z) = − .
4z
For comparison, the LDA holds only for (1), for (2) the cancellation is only partial, and
the criteria (3) and (4) are not satisfied.
Singh21 tested several flavours of WDA for ground-state properties of solids. According
to present-day experience, the WDA provides considerable improvement over LDA in the
description of ground-state properties for essentially all systems. A selection of some
results by Singh is given in the following table.

Ground-state properties calculated within WDA by Singh (1993, 1998).

a is lattice constant (in Å), B - bulk modulus (in GPa)

C GaAs V KNbO3 SrO

a B a B a B a a

LDA 3.53 483 5.62 85 2.93 215 3.96 5.06

WDA 3.57 404 5.74 62 3.00 145 4.02 5.16
Exp. 3.57 442 5.65 75 3.02 157 4.02 5.16

21
David J. Singh, Weighted-density-approximation ground-state studies of solids, Phys. Rev. B 48,
14099 (1993); I. I. Mazin and D. J. Singh, Weighted Density Functionals for ferroelectric Materials,
http://xxx...cond-mat/9801301

44
3.9 Inhomogeneities of density: k-space analysis and
gradient expansions
Another approach to inhomogeneities of particle density involves the Fourier transforma-
tion and analysis of terms dependent on wavevector (of charge fluctuations). Below, we
follow the analysis of “inhomogeneous electron gas” in the original paper by Hohenberg
and Kohn.
Consider the total energy functional as

e2 ρ(r)ρ(r0 )
Z Z
E[ρ] = v(r) ρ(r) dr + dr dr0 + G[ρ] , (3.52)
2 |r − r0 |

where G[ρ] incorporates kinetic and exchange-correlation parts (the expression for the
rest is exact). We assume that spatial deviations of density from its mean value are small,

ρ̃(r)
ρ(r) = ρ0 + ρ̃(r) , with  1 everywhere.
ρ0

Moreover, the normalization (number of particles) remains constant, that leads to

Z
ρ̃(r) dr = 0 .

One can imagine the following formal expansion of G[ρ] in ρ̃ as small parameter:
Z
G[ρ] = G[ρ0 ] + I(r) ρ̃(r) dr +
Z
+ K(r, r0) ρ̃(r) ρ̃(r0 ) dr dr0 +
Z
+ L(r, r0, r00 ) ρ̃(r) ρ̃(r0 ) ρ̃(r00 ) dr dr0 dr00 + . . . (3.53)

The linear expansion coefficient must be zero, because the result cannot depend on the
uniform shift ∆ of the coordinate system:
Z Z Z
I(r) ρ̃(r) dr = I(r) ρ̃(r−∆)dr , hence [I(r)−I(r+∆)] ρ̃(r) dr = 0 for any ∆;
R
therefore I(r) = const, and I ρ̃(r) dr = 0, due to normalization of ρ̃(r).
Similarly, the 2nd-order kernel K(r, r0 ) can only depend on |r − r0 |, because the result
must be independent on the displacement of the ρ̃(r) distribution, as well as on rotations
about the (r−r0 ) line. We skip the following terms in the expansion (3.53) for the moment,
so Z
G[ρ] = G[ρ0 ] + K(|r−r0|) ρ̃(r) ρ̃(r0 ) dr dr0 . (3.54)

We introduce the Fourier transformation of the quadratic kernel,

1X 0
K(r − r0 ) = K(k)e−ik(r−r ) (3.55)
Ω k

45
and analyze the properties of K(k). For this, we introduce a small parameter λ in the
Fourier expansion of density:
λX
ρ(r) = ρ0 + b(k) e−ikr . (3.56)
Ω r
Hohenberg and Kohn argue that K(k) can be related (comparing terms with the same
powers of λ in the perturbation theory series for the interaction energy) to the electronic
polarizability α(k), or to dielectric constant ε(k) of the electronic liquid:
" #
2π 1 2π 1
K(k) = 2 −1 = 2 (3.57)
k α(k) k ε(k) − 1
E (primary field) 1
with ε(k) = = .
E− E · α}
| {z 1 − α(k)
induced field
The polarizability α(k) has the following
general properties:

k→0: a(k) = 1 + c2 k 2 + c4 k 4 + ...

dα
q → 2kF : → −∞;
dk
1
q→∞ : a(q) ∼ 4 .
q

Fig. 1 of Hohenberg and Kohn (1964).

Electronic polarizability α(q), for electron
density = 4 × 1022 cm−3 .

For free electron gas, ε(k) is given by the Lindhard function, that leads for K(k):
" ! #−1
4π k 2 1 kF k2 k + 2kF
K(k) = 2 · + 1− ln . (3.58)
k 8kF 2 2q 4kF k − 2kF

The prefactor 4π/k 2 is the Fourier transform of

(∼ q 2 ) the Coulomb interaction, and the rest – Fourier
transform of the correlation function. In order
to discuss exchange-correlation hole separately,
we concentrate in the following on the exchange-
correlation local field factor

k2
IXC (k) = KXC (k) .
4π

(const1 + const2 ·q 4 ) Fig. 2 of Hohenberg and Kohn (1964). Behavior of

the kernel K(q), for the same electron density.

46
Here, KXC (k) refers to the XC part of the kernel only, with the contribution due to kinetic
energy of non-interacting particles subtracted.
For IXC (k), Monte Carlo results are available for several values of rc , as well as differ-
ent analytical approximations. Actually, any model concretizing the exchange-correlation
energy density, be it LDA or different flavours of WDA, can be mapped onto a correspond-
ing local-field XC factor. Examples have been given, e.g., by Mazin and Singh (1998).22
Gunnarsson-Jones
1.4 Perdew-Wang
Gritsenko et al
LDA
MC interpolation
1.2

1
Fig. 1 of Mazin and Singh (1998). Exchange-
0.8
correlation local field factor in the WDA by
Ixc(q)

0.6 Gunnarsson and Jones 1980 (Phys. Scr. 21, 394),

0.4 Gritsenko et al. 1993 (Chem. Phys. Lett. 205,
0.2 348), and derived from the homogeneous electron
gas pair correlation function, Perdew and Wang
2

1.8
Gunnarsson-Jones
Perdew-Wang 1993 (Phys. Rev. B 46, 12947), as compared with
Gritsenko et al

1.6
LDA
Monte Carlo
MC interpolation
the Monte Carlo results and the interpolating
1.4 formula thereof, as given by Moroni, Ceperley and
1.2 Senatore 1995 (Phys. Rev. Lett. 75, 689). Den-
Ixc(q)

1
sities, from top to bottom, correspond to rs =1, 2, 5.
0.8

0.6

0.4

0.2
One can see that the LDA prescription for
2
Gunnarsson-Jones
Perdew-Wang
the local field factor proceeds as KXC (0)/(4π) ·
1.8 Gritsenko et al

1.6
LDA
Monte Carlo
MC interpolation
k 2 ,and it is a quite good approximation if fluc-
1.4
tuations of density are of long wavelength, k <
1.2 2kF. At k = 2kF , the local-field factor, as the
kernel K(k)) in general, experiences a kink, that
Ixc(q)

0.8
gives rise to Friedel oscillations of electron den-
0.6
sity in the direct space. Beyond 2kF , IXC (k)
0.4

0.2
increases ∼ k 2 rather than tending to a con-
0
stant; this is an indication that EXC includes
0 1 2 3 4
q/kF
some kinetic-energy part.
A different method of taking into account inhomogeneities of density is via gradient
expansions. One can construct the following expression for the energy density:
3
X 3 h
X i
(1,1) (2)
g[ρ] = g0 (ρ) + gi (ρ)∇i ρ + gij (ρ)∇i ρ ∇j ρ + gij ∇i ∇j ρ + . . . (3.59)
i=1 ij

If the density is scaled as

r
ρ(r) = ρ0 ( ),
r0
22
I. I. Mazin and D. J. Singh, Nonlocal density functionals and the linear response of the homogeneous
electron gas, Phys. Rev. B 57, 6879 (1998)

47
the above series is an expansion in powers of (r0−1 ). It does not strictly converge. However,
we’ll use it to sort out leading terms, inclusing gradients of density. The validity of the
gradient expansion requires
|∇ρ| |∇i ∇j ρ|
 qF ;  qF .
ρ |∇ρ|
Making use of invariance relations under uniform displacements and rotations, one arrives
at the following form:
(2)
g[ρ] = g0 (ρ) + g2 (ρ) ∇ρ · ∇ρ +
h i
(2) (3) (4)
+ g4 (ρ)(∇2 ρ)(∇2 ρ) + g4 (ρ)(∇2 ρ)(∇ρ · ∇ρ) + g4 (ρ)(∇ρ · ∇ρ)2 + O(∇6i ) .
The expansion coefficients can be expressed via Taylor expansion coefficients of electronic
polarizability α(q).
A formally correct gradient expansion
Z   Z  
EXC [ρ] = ρ(r)LDA
XC ρ(r) dr + BXC ρ(r) |∇ρ(r)|2dr + . . . (3.60)

meets certain problems in its practical implementation:

– the terms of the 4th order and higher in the expansion of the exchange energy diverge;
– the exchange potential diverges already in the 2d order;
– in the 2d order, gradient corrections to exchange energy bring an improvement,
but gradient corrections to corellation energy yield worse results that the LDA.
The last fault can be traced to the fact that the normalization of the XC-density (3.44),
valid in the LDA, does not necessarily hold in a truncated gradient expansion (3.60).
Particularly, it is violated in what regards the correlation hole (that must integrate to
zero around any electron). Therefore, one has to impose necessary constrains, probably
using additional gradient terms (nonlinear and not necessarily quadratic) beyond those
included in the formal gradient expansion (3.60). These working schemes with explicite
nonlinear dependence of the integrand on density and its gradients are refered to as
different flavours of generalized gradient approximation (GGA)
Z  
GGA
EXC [ρ] = ρ(r)GGA
XC ρ(r), |∇ρ(r)|, ∇2 ρ(r) dr . (3.61)

The parametrization of GGA

XC is chosen so as to satisfy certain limiting cases, like the LDA
limit, scaling property EX [ρλ] = λEX [ρ] etc.
An example of such arguing was given by Becke23 who introduced a now broadly
used (especially in quantum chemistry applications) form of a GGA approximation to the
exchange energy. The correlation was explicitly left out in the Becke analysis, in order to
use HF results as a benchmark for tuning this density functional scheme. The asymptotic
properties explicitly satisfied were the following: the Coulomb potential of the exchange
charge, i.e. of Fermi hole density far from the electron distribution
ρXσ (r, r0 ) 0
Z
1
lim 0
dr = − , (3.62)
r→∞ |r − r | r
23
A. D. Becke, Density-functional exchange-energy approximation with correct asymptotic behavior,
Phys. Rev. A 38, 3098 (1988)

48
and the exchange energy defined by Eq. (3.27) must behave according to this asymptotic.
Morever, the asymptitic of the spin density is

lim ρσ (r) = e−aσ r ,

r→∞
(3.63)

where aσ is a constant related to the ionization potential of the system. The functional
form of the exchange energy proposed by Becke guarantees that the density satisfying
Eq. (3.63) generates asymptotic potential as in Eq. (3.62):

(Becke) XZ x2σ
EX = EXLDA −β ρσ4/3 dr (3.64)
σ 1 + 6βxσ sinh−1 xσ

where xσ is the dimensionless ratio

|∇ρσ |
xσ = 4/3
ρσ

and β is an adjustable parameter, for which Becke proposed, based on least-square fit to
Hartree-Fock results over a variety of systems, a value of β=0.0042 a.u.
A more sophisticated form incorporating both exchange and correlation, that satis-
fied more essential conditions, has been proposed in 1991 by Perdew and Wang24 for
GGA
XC in Eq. (3.61), in terms of ρ and ζ of Eq. (3.32); rs as related to ρ by Eq. (3.35);
g = [(1+ζ)2/3 +(1−ζ)
q
2/3
]/2; two differently scaled density gradients s = |∇ρ|/(2ρkF ) and
t = |∇ρ|/(2gρ 4kF/π), where kF = (3π 2 /ρ)1/3 is identical to definition of Eqs. (1.7) and
(2.35), and many tabulated numerical parameters:
(PW91)
XC = X + C ;
" 2 #
1 + a1 s sinh−1 (a2 s) + (a3 + a4 e−100s )s2
X = LDA
X (ρ, ζ) ;
1 + a1 s sinh−1 (a2 s) + a5 s4
C = LDA
C (ρ, ζ) + H(ρ, t) ,
" #
3 2
g β 2α t2 + At4 4 4
3 2 −100g πkF t
2
H = log 1 + + C c0 [C c (ρ) − C c1 ] g t e ,
2α β 1 + At2 + A2 t4
( " # )−1
2α 2αLDA
C (ρ, ζ)
A = exp − −1 ,
β g3β 2
C2 + C3 rs + C4 rs2
Cc (ρ) = C1 + . (3.65)
1 + C5 rs + C6 rs2 + C7 rs3

The inconvenience of having many adjustable parameters in a practical GGA scheme mo-
tivated Perdew, Burke and Ernzerhof25 to drop several least important constraints in the
formulation of GGA and present a somehow more simplified, as compared to Eq. (3.65),
24
J. P. Perdew, in Electronic Structure of Solids ’91, edited by P. Ziesche and H. Eschrig (Academie
Verlag, Berlin, 1991), p. 11, reproduced in Perdew et al., Phys. Rev. B 46, 6671 (1992).
25
John P. Perdew, Kieron Burke and Matthias Ernzerhof, Generalized Gradient Approximation Made
Simple, Phys. Rev. Lett. 77, 3865 (1996)

49
and simultaneously more physically transparent scheme. The effect of inhomogeneity of
ρ is isolated in the “enhancement factor” FXC , modulating the local exchange:
h i Z
(PBE)
EXC ρ(+), ρ(−) = ρ(r) unif.
X (r)FXC (rs , ζ, s) dr . (3.66)

Whereas the correlation part of FXC (rs , ζ, s) is ex-

pressable via the function H resembling (but ana-
lytically more simple than) that of Eq. (3.65), the
exchange part could be mapped onto
κ
FX (s) = 1 + κ − ,
1 + µs2 /κ

where µ ' 0.21951 and κ = 0.804 are numerical

constants. The figure on the right shows the
behaviour of the “enhancement factor” for rele-
vant values of rs and s in non-polarized and fully
polarized cases.

Fig. 1 of Perdew, Burke and Ernzerhof (1996).

Encancement factors FXC shownig GGA non-locality.
Solid curves denote the PBE results, open circles –
those of PW91.

An extensive literature exists on the performance of GGA in different systems, see,

e.g., Dal Corso et al. (1996).26 Generally, GGA tends to correct to some extent the
overbonding that is typical in LDA results. The equilibrium volumes are typically larger
according to calculations using GGA, and in many cases agreement with experiment
improves. In other cases, the GGA slightly overestimates the volume. Quite important
is the inclusion of GGA in the treatment of magnetic systems; one of the most known
examples of LDA failures is the hierarchy of energies of different (magnetic and structural)
phases of iron, where the LDA wrongly predicts fcc phase to be favourable over the bcc.
The inclusion of gradient corrections restores the correct relation between these phases.

26
Andrea Dal Corso, Alfredo Pasquarello, Alfonso Baldereschi and Robeto Car, Generalized-gradient
approximations to density-functional theory: A comparative study for atoms and solids, Phys. Rev. B 53,
1180 (1996)

50
 4. Practical implementation of DFT
and HF calculation schemes

4.1 Basis sets: an overview

Historically, the invention of the HF method made possible numerical solution, albeit
approximative, of many-electron problems too difficult to be treated analytically. During
1930s and 40s, HF equations were looked at as a system of (integro-)differential equations,
and corresponding methods of solution were applied. In many cases it turned out too
difficult to solve these equations for HF-functions ϕ(r) of general form. A further progress,
at least in what regards numerical accuracy, was related with the use of basis sets in
which the unknown functions can be expanded. Certain shape approximations on the
shape of one-electron wavefunctions and/or potential helped further on to simplify the
problem. An example of this development is e.g. the muffin-tin approximation for the
potential proposed by Slater which gave rise to powerful computational methods like
APW, KKR and their derivatives – LAPW and LMTO – to be discussed in some detail
below. Finally, with the increase of computer power the trend of solving Kohn-Sham
equations still numerically but “directly”, i.e. without relation to any particular basis
system, becomes noticeable and promising.
The idea to expand one-electron functions over a limited basis set and thus to reduce
integro-differential equation system to a matrix eigenvalue problem is probably due to
Roothaan.27 While originally giving rise to what is referred to as the Hartree-Fock-
Roothaan method, the same approach will work for the Kohn-Sham equations. Let us
look at it in more detail. Eq. (2.17), with j 6= α dropped in both sums, reads:
" #
h̄2 2 ρ(x0 ) dx0 ρ(x0 , x) ϕα (x0 ) 0
Z Z
− ∇ + u(x) + ϕ α (x) − dx = εα ϕα (x) (4.1)
2m |r − r0 | |r − r0 |
in the HF formalism, and similarly, substituting
ρ(x0 , x) ϕα (x0 ) 0
Z
− dx ⇒ VXC (x)ϕα (x)
|r − r0 |
in the Kohn–Sham equations. We expand ϕα (x) over a basis set χ:
Q
X
ϕα (x) = Cαp χp (x) , (4.2)
p=1

where the dimension of basis Q is reasonably larger than the number of occupied states
N, and χp (x) may retain the dependence on both spatial and spin variables. Eq. (4.1)
then reduces to X X
Cαp Hχp (x) = εα Cαp χp (x) ,
p p

27
C. C. J. Roothaan, New Developments in Molecular Orbital Theory, Rev. Mod. Phys. 23, 69 (1951).

51
with H being the Fock operator defined by Eq. (2.25) in the HF method, or its counterpart
in the Kohn–Sham approach. We shall refer to it as “Hamiltonian”, since it is formally
indeed a Hamiltonian of the reference system of independent particles, which produce the
one-electron density equal to the correct one. Multiplying the above equation on the left
by χ∗q (x) and integrating in x, we get

X Z Z 
Cαp χ∗q (x)Hχp (x)dx − εα χ∗q (x) χp (x) dx =0 (4.3)
p

That is a system of algebraic equations on the expansion coefficients Cαp , or a generalized

(i.e., with overlap) diagonalization problem. The overlap matrix elements
Z
Sqp ≡ χ∗q (x) χp (x) dx (4.4)

need to be calculated only once (for basis functions fixed in advance). The matrix elements
of the Hamiltonian
" #
Z
h̄2 2 Z
ρ(x0 )dx0
Hqp ≡ χ∗q (x) − ∇ + u(x) + + VXC (x) χp (x)dx
2m |r − r0 |

are actually dependent (via ρ and VXC ) on yet unknown coefficients Cαp . Specifically,
only the eigenvectors corresponding to N lowest eigenvalues εα must be included in the
construction of the ground-state density (similarly to the aufbau pronciple in the HF
formalism), so the density is

N
" N
#
X X X
ρ(x) = ϕ∗α (x) ϕα (x) = ∗
Cαq Cαp χ∗q (x) χp (x) . (4.5)
α=1 pq α=1
| {z }
≡Dpq , density matrix

The density can be either diagonal in spin variables, if basis functions for both spin com-
ponents remain completely decoupled, or it may retain a general form, if spin components
interact due to spin-orbit interaction and/or spin noin-collinearity effects being included.
The internal dependency of Hpq on Cαq via the density matrix prevents the solution of
Eq. (4.3) in a single matrix manipulation. However, an iterative solution is possible: one
fixes a trial set of Cαq in the density matrix and constructs the Hamiltonian, then (4.3)
becomes a system of linear equations solvable by a single diagonalization. Among the
Cαq (q = 1, . . . Q > N) found by diagonalization one selects those corresponding to the N
lowest eigenvalues, uses them to update the density matrix and proceeds till everything
converges.
In practical calculations, one would typically use a dampfing parameter β < 1,

C next = βC new + (1 − β)C old

or a more spphisticated mixing scheme, in order to prevent instability of expansion coef-

ficients C in the course of iterations, as is illustrated by the following figure.

52
 out out
C C
diverges...

converges...

in in

start
C start
C

A schematically shown dependency of the output solution vector on an input one. The
self-consistency condition is shown by a dashed line. Increased dempfing (smaller β)
scales the dependency towards the type of behaviour shown on the left.

In so doing, one has to solve in each iteration a system of linear equations. While
Sqp remains unchanged if the basis functions do not change (that is the case, at least, in
solving a static electronic-structure problem with fixed positions of atoms), Hqp has to be
re-calculated and includes the following terms:
" #
h̄2 2
Z
Hqp = χ∗q (x) − ∇ χp (x)dx (≤2-center integrals)
2m
Z
+ χ∗q (x) u(x) χp (x)dx (≤3-center integrals)
" #Z
X N
X ∗ χ∗q (x) χ∗q0 (x0 ) χp0 (x0 ) χp (x)
+ Cαq 0 Cαp0 dx dx0 (≤4-center integrals)
q 0 p0 α=1 |r − r0 |
| {z }
density matrix
Z
+ χ∗q (x) VXC (x) χq (x)dx . (4.6)

The term “n-center integral” refers to the situation when the basis functions are centered
somewhere in space, usually at atom sites. This is often the case – but now always, as the
site-independent basis functions (e.g., plane waves) may be also used. For site-centered
bases, one usually takes into account the fact that the external potential u(x) contains,
along with other possible contributions, the Coulomb potential from atom nuclei, either
screened by core electrons or not:
X eZµ
u(x) = + ...
µ |r − Rµ |
Then, depending on whether both basis functions and the corresponding atom-centered
component relate to the same center or not, the contributions of the form
Z
1
χαq ∗ (r − Rα ) χβp (r − Rβ ) dr
|r − Rµ |
may be one- , two- or three-center integral. Similarly, the Hartree contribution to the
Hamiltonian (the 3d line of Eq. (4.6) may contain one- to four-center terms. The exchange-
correlation part depends on ρ(x) and may also give rise to up to four-center integrals, as

53
its particular form in the HF-formalism, the exchange term, would obviously do. The
calculation of one-center integrals is usually trivial; two-center integrals may be simplified
by an appropriate coordinate transformation (e.g., using elliptic coordinates); the three-
and four-center integrals however may present a problem, and their evaluation deviates
between different calculation methods.
After eigenvectors Cαq are finally converged and the density matrix Dpq thus deter-
mined, the total energy can be recovered as follows from Eq. (3.14):

h̄2 X X Z
Etot = − Dpq χ∗q (x)∇2i χp (x)dx +
2m pq i
e2 X dx dx0
Z
+ Dpq Drs χ∗q (x) χp (x) χ∗r (x0 ) χs (x0 ) +
2 pqrs |r − r0 |
X Z
+ e Dpq χ∗q (x) χp (x) u(x) dx + EXC [ρ] , (4.7)
pq

or, similarly to Eq. (3.17):

N
e2 X dx dx0
X Z
Etot = εi − Dps Drs χ∗q (x) χp (x) χ∗r (x0 ) χs (x0 ) −
i=1 2 pqrs |r − r0 |
X Z
− e Drq χ∗q (x) χp (x) VXC (x) dx + EXC [ρ] . (4.8)
pq

So far, we haven’t yet specified the form of basis functions χ. As already mentioned,
a quite common choise is some atom-centered functions localized in real space, but the
functions localized in reciprocal space (plane waves) could be a convenient option, or some
hybride bases. The basis of choice is often related to the nature of boundary conditions
used. Thus, for a simulation of a periodic solid with Bloch – von Kármán boundary
conditions the use of plane wave basis may seem a natural choice, whereas non-periodic
systems like molecules or clusters may seem to be more naturally treated with localized
atom-centered bases. However, in both these cases an alternative choice of basis is equally
possible and in effect broadly used: molucules are simulated in a repeated “simulation
box” with periodic boundary conditions thus imposed and plane waves used as basis set;
for periodic solids, the (k-dependent Bloch sums can be constructed by lattice summation
of localized functions, as:
1 X α
χ̃αpk (r) = √ χp (r − uα − Rν ) exp[ik(Rν + uα )] , (4.9)
N ν
where Rν runs over lattice vectors and uα is the basis vector of atom α in the unit cell.
Between the two extremities of plane waves and atom-centered functions as candidates
for a basis set definition in a workable method, large variety of combined methods has
been developed and is in use. The need for combined methods is due to the fact that none
of both “pure” choices is fully satisfactory. The practical difficulty is that the electron
density in real materials behave very differently near atomic cores (large fluctuations
with short spatial period) and far from the cores (smooth distribution of density with
possibly elevated values on the chemical bonds, i.e. away from atoms). In discussing a

54
compromise between computational efficiency and accuracy, let us sort out advantages
and disadvantages of plane waves and atom-centered basis functions.
Plane waves
(+) Ultimately complete basis; systematic augmentation of accuracy is controlled by a
single parameter (planewave cutoff);
(+) Easy analytical manipulation, e.g., when calculating matrix elements of different ob-
servables, derivatives of the total energy etc.
(−) Boundary conditions can be only periodic, therefore in the course of simulating finite
fragments “in the box” the spurious interactions across the box boundary are built in,
and their suppression may demand for a large box size.
(−) The number of plane waves necessary to describe fluctuations of all-electron charge
density is usually beyond the reasonable computational resources. The use of pseudopo-
tentials is a typical solution, but this is a certain approximation.
(−) For a given cutoff (i.e., the largest wavevector used in the planewave expansion), the
size of basis grows very rapidly with the size of simulation cell, irrespectively on whether
it contains extra atoms or not. This is a serious handicap in a simulation of large systems,
but also of open systems, like surfaces or molecules.
Atom-centered functions
(+) Since all charge is physically delivered by one-electron functions centered on atoms,
the basis size scales linearly with the number of atoms, irrespectively of empty space in
the system. This is especially important for simulations of open systems.
(+) Boundary conditions can be either periodic or strictly “vacuum-like”, with no spuri-
ous interaction between repeated fragments.
(−) The lack of systematics in gradually enhancing the completeness of basis; additional
basis functions are added ad hoc, and no asymptotic completeness of the basis is guaran-
teed.
(−) Difficulties in calculating matrix elements. This is probably the most serious draw-
back that can be overcome by the following tricks:
– if possible, calculate in advance and store in tables for subsequent fast interpolation
(good, e.g., for dynamical simulations);
– use efficient statistical scheme rather than straightforward integration;
– use localized basis functions which allow analytical (or otherwise easy) integration
(Gaussian basis sets).
Among ”analytical” localized basis functions, important are so-called Slater-type and
Gaussian orbitals. The Slater-type orbitals (STO) imitate the asymptotics of one-electron
wavefunctions, known for the hydrogen atom:

χSTO (r) ∼ r n−1 e−ζr Ylm (θ, φ) . (4.10)

The STOs form a complete basis of solutions for the central-field problem, while the
textbook solutions for the hydrogen atom are only complete if taken together with the
eigenfunctions corresponding to the continuous energy spectrum. n in STOs has the
meaning of principal quantum numberm and ζ is “effective nuclear charge”, but essentially
it is an adjustable (free to vary) parameter. A peculiar feature of STOs is that, differently

55
from the hydrogen-atom solutions, they have no node structure, but a combination of
several STOs makes it possible to imitate the nodes – and hence fluctuations of charge
density near the atomic core. One speaks of “minimal basis”, “double-zeta”, “triple-zeta”
etc., the latter combining several exponents with different ζ values. This terminology is
also transferred to the classification of Gaussian bases discussed below. The number of
Slater-type basis functions per atoms of different periods is shown in the following table:

Minimal basis Double-zeta

H, He 1s [1] 1s1s0 [2]
Li → Ne 1s2s2p [5] 1s1s0 2s2s0 2p2p0 [10]
Na → Ar 1s, · · · , 3p3d [14] ··· [28]
K → Kr 1s, · · · , 4s4p [18] ··· [36]
Rb → Xe 1s, · · · , 5s5p [27] 1s1s0 , · · · , 5p5p0 [54]

STO bases are known to be quite efficient in a sense that convergency of results is
achieved fast with the augmentation of basis. However, the evaluation of integrals is as
difficult as with any other (numerical) functions.

4.2 Gaussian basis sets

Gaussian-type orbitals (GTO) have been proposed by Boys in 195028 :
2
χGTO ∼ r 2(n−1) e−αr Ylm (θ, φ) . (4.11)

Their big advantage is the possibility to calculate the integrals (2- to 4-centered in the
construction of Hamiltonian, as well as in many cases matrix elements of other operators
of interest) analytically. In GTO, only functions with l = n − 1 are explicitly used, i.e.
1s, 2p, 3d etc. but not 2s, 3p. Thus, the node structure of one-electron wavefunctions is
not included in any single GTO, but can be imitated by a linear combination of several
basis functions. The drawback of gaussian
bases is their slower convergency as com-
pared to STO. This follows from the fact
that individual GTOs are relatively poor rep-
resentatives of true one-electron wavefunc-
tions: GTOs have wrong asymptotics in the
infinity (fall down too fast) and wrong be-
haviour near the nucleus. The following fig-
ure shows the “optimum” GTO obtained for
the H1s state by least-square fit, preserving
the normalization.
28
S. F. Boys, Electronic wave functions. I. A general method of calculation for the stationary states of
any molecular system, Proc. Roy. Soc. (London) A200, 542 (1950).

56
The performance can be improved by using
several GTO (3 – 4) to imitate a single STO.
If keeping all them independent, the bottle-
neck is the storage of many integrals. A possi-
ble solution is to use fixed linear combinations
of GTO in a calculation rather than individ-
ual GTOs; such combinations are referred to
as contracted GTOs. One can fit contracted
GTOs either to STOs, or to numerical solu-
tions for single atoms or ions. The following
figure shows again the H1s wavefunction, now
approximated by a contracted 4-GTO set.
Let us discuss now the evaluation of integrals. The GTOs may be rewritten in Carte-
sian form,
Gijk (r, α, A) ≡ xiA yAj zAk exp(−α rA
2
) (4.12)
where A is the origin and rA ≡ r − A; the Cartesian Gaussians can be factorized:
Gijk (r, α, A) = Gi (x, α, Ax ) Gj (y, α, Ay ) Gk (z, α, Az ) (4.13)
with Gi (x, α, Ax ) = xiA exp(−αx2A ) etc. The differentiation of Gaussians (necessary e.g.
for calculating forces) yields
∂Gn ∂Gn
=− = 2αGn+1 − n Gn−1 . (4.14)
∂Ax ∂x
For higher-order differentiation, one can use the recursion
!q
∂ q+1 Gn ∂ ∂ q Gn+1 ∂ q Gn−1
= [ 2αGn+1 − n Gn−1 ] = 2α − n ,
∂Aq+1
x ∂Ax ∂Aqx ∂Aqx
or more compact:
(q) (q)
G(q+1)
n = 2α Gn+1 − n Gn−1 . (4.15)
Of central importance for analytical evaluation of integrals is the Gaussian product
rule, according to which a product of two Gaussians is again a Gaussian, centered at some
point on the line connecting original centers. For s-Gaussian (without polynomial in r)
in one dimension the relation is straightforward:
exp(−αx2A ) exp(−βx2B ) = exp(−q Q2x ) × exp(−p x2P ) (4.16)
| {z } | {z }
constant a new
≡ KAB Gaussian
αβ
with xP ≡x−Px , p=α+β, q= , pPx = αAx + βBx , Qx = Ax − Bx .
α+β
Along y (or z)-direction, the product of two original exponents would yield exp[−(α+β)y 2 ]
with the same exponent prefactor p as in the x-dependence, therefore the above product
formula is valid in 3-dimensional case:
exp(−αr2A ) exp(−βr2B ) = KAB exp(−pr2P ) .

57
In the product of two non-s-Gaussians that contain powers of (r − A) and (r − B) as well,
the polynomial terms can be rearranged in another polynomial of (r − P). So a product
of two arbitrary Gaussians yields
X
Gijk (r, α, A) Gi0j 0 k0 (r, β, B) = C (α) G(α) (r, p, P)
(α)

where (α) incorporates different possible combinations of powers of x, y and z, up to the

maximal power i+j+k+i0 +j 0 +k 0 . After such decomposition, the overlap matrix elements
Sqp (4.4) reduce to the sum of integrals
X Z Z∞
(α) 2
Sqp = C d ΩYlm (θ, φ) r n(α) e−γr r 2 dr
(α) 0

that are taken using (with γ > 0, n = 0, 1, . . .):

Z∞ s Z∞
2n −γx2 (2n − 1)!! π 2 n!
x e dx = ; x2n+1 e−γx dx = .
2(2γ)n γ 2γ n+1
0 0

Analyzing different terms in Eq. (4.6), one can sort them out into
• multipole moments, hGA | xeC yCf zCg |GB i with GA = xiA yAj zAk exp(−αrA 2
) etc. and inte-
gration done over coordinates of one electron. Such integrals describe, for example, the
effect of the external field;
• integrals hGA | eikC xC |GB i that appear, e.g., in the calculation of exchange-correlation
term, when the Fourier transformation of a general-form density is used. These two types
of integrals can be analytically taken by parts and are available in closed form;
• kinetic energy. Since differential operators acting on Cartesian Gaussians are separable,
!q !r !s
∂ ∂ ∂ ∂q ∂q ∂q
hGA | |GB i = hGi | q |Gj ihGk | q |Gl ihGm | q |Gn i .
∂x ∂y ∂z ∂x ∂x ∂x
By the virtue of recursion (4.15), these matrix elements are related to overlap integrals
between GTOs of reduced and increased order, up to ±2.
More involved is calculation of (up to) four-center Coulomb integrals,
dr dr0
Z
χ∗q (r) χp (r) χ∗q0 (r0 ) χp0 (r0 )
.
|r − r0 |
For simplicity, we discuss in the following the case of all four GTOs being of s-type, i.e.
without polynomial prefactor. For general-form Gaussians analytical integration is also
possible but much more lengthy.
q
r’ (q)+(p) −> (A) r’
p

p’
(q’)+(p’) −> (B)
r q’ r
The figure above explains the geometry of our problem: the origins of four basis
functions, if all different, are at q, p, q 0 and p0 ; two spatial variables r and r0 are involved.

58
The Gaussian product rule allows to reduce this to, at most, 2-center problem, with new
centers A and B that can be straightforwardly found. Now, the integral to be evaluated is
formally identical to a Coulomb interaction between two Gaussian distributions of charge:
Z Z
ρA (rA ) ρB (rB )
VAB = dr1 dr2 . (4.17)
|rA − rB |

In order to evaluate this, we specify first the electrostatic potential (at arbitrary point
C) of such distribution centered at A, then the electrostatic energy of the second charge
distribution (centered at B) in this potential. For the following, we assume the charge
distribution around A to be normalized to unity,
 3/2 2d distribution
α 2
ρA (rA ) = e−αrA ;
π
Z
ρA (r − A) 1st distribution
rA B
VA (C) = dr .
|r − C|
A
For 1/|r − C| ≡ 1/rC we make a formal sub- rC rB
stitution using
RAB
Z∞ √
2 t2 π r C
e−R dt = ;
R
−∞

Z∞
1 1 2 2 global origin
=√ e−rC t dt .
|r − C| π
−∞

2
Applying the Gaussian product rule (4.16) to exp(−αRA ) exp(−rC2 t2 ), we transform
 ∞

α3/2 Z 2
−αrA
Z
2 2
VA (C) = 2
dr e  e−rC t dt =
π
−∞
Z∞
α3/2
Z 
αt2
− R2 −(α+t2 )(r−S)2
= dt e α+t2 AC dr e .
π2
−∞

Note that t is a dimensionless integration

rA variable, and S marks the position of the re-
rC sulting (overlapped) Gaussian along the A–
A
C C line (different for each t), as shown on the
left figure:
t2 RAC
S S= ,
α + t2
Thus the origin of integration in r is simply
center of displaced to S, with the result unchanged:
overlapped Gaussians
Z
2 ) r2 π 3/2
dr e−(α+t = ,
(α + t2 )3/2

59
hence
Z∞ exp(− α t2 2
α3/2 α+t2
RAC )
VA (C) = √ dt .
π (α + t2 )3/2
−∞

A variable substitution
t2 (α + t2 )3/2
2
→ u2 ; dt = du
α+t α
results in s
Z1
4α −αR2 u2
VA (C) = e AC du . (4.18)
π
0

This integral can be expressed via error function:

Zx
2 2
erf(x) = √ e−t dt ,
π
0

hence √
Z1 ZA r √
−Ax2 1 −t2 π
e dx = √ e dt = erf( A)
A 4A
0 0

and
1 √
VA (C) = erf(RAC α) .
RAC
Now we are ready to calculate the Coulomb integral (4.17). Assuming that the charge
distribution centered at B has an exponent β and is normalized to unity, ρB (rB ) =
2
(β/π)3/2 e(−βrB ) , similarly to how it was for ρA (rA ), (4.17) becomes:
s
Z !3/2 !3/2 Z Z1
β −β(r−B) 4α β 2 2 2
VAB = dr VA(r) e = dr du e−αrAu e−βrB .
π π π
0

Once more applying the product rule yields:

!
2 2 2
h
2
i αu2 β
exp(−αrA u ) exp(−βrB ) = exp −(αu + β)rS2 exp − 2 R2
αu + β AB

where S, from which rS is measured, is the center position of overlapped Gaussians some-
where on the A–B line.
s !3/2 Z1 ! Z
4α β αu2 β 2
VAB = du exp − 2 RAB dr exp[−(αu2 + β)rS2 ]
π π αu + β
0 | {z }
 3/2
π
αu2 +β
s 1 !
4αβ Z β αu2 β 2
= exp − 2 R du .
π (αu2 + β)3/2 αu + β AB
0

60
Another variable substitution
α+β 2 βdu
u → t2 ; (α + β)1/2 → dt
αu2 + β (αu2 + β)3/2

leads to an expression completely symmetric in (A,α) and (B,β), as it of course should

be, with γ ≡ αβ/(α + β):
s s
Z1 ! 1
4αβ 1 αβ 2 2 4γ Z −γR2 t2
VAB = √ exp − R t dt = e AB dt . (4.19)
π α+β α + β AB π
0 0

The above derivation helps to calculate the values of (up to) 4-center Coulomb integrals,
provided they do not contain polynomials (i.e., only s-Gaussians enter). For GTO of
general form, closed analytical expressions, albeit of much more complicated form, have
also been obtained. For a review, one can consult, e.g., an article by Helgaker and Taylor29
Whereas in HF calculations all terms in the Hamiltonian reduce to 2-, 3- or 4-center
integrals, in a DFT formalism the exchange-correlation potential VXC is present that is a
nonlinear function (or, beyond the LDA, a functional) of density. Moreover, the Hartree
potential, although reducible to Coulomb interaction between individual Gaussians, is
expensive (in terms of computation time and storage) due to the necessity to re-calculate
and store multicenter integrals. It is more efficient to fit the total resulting density to
a superposition of support functions, probably the same Gaussians (typically taken to
a higher order than for the representation of basis functions). Therefore, it may be
both advantageous and unavoidable to deal with potential and charge density of spatially
unconstrained, i.e., general form. Practically used solutions are:
• expand everything in Gaussians centered at the atoms, or in the interstitial (i.e. at
chemical bonds).
• single out slowly varying interstitial density (by subtracting atom-like contributions)
and represent it on a discrete mesh (e.g., by splines, or by Fourier transformation) with
the aim to solve the Poisson equation easily.
A typical structure of the latter decomposition of density could be:

ρ(r) = ρloc (r) + ρnon-loc (r) + ρnucl (r)

= [ρloc (r) + ρnucl (r) + ξ(r)] + [ρnon-loc (r) − ξ(r)] ,

where ρnucl (r) is a well localized near nuclei contribution of deep core states, ρloc (r)
is a valence charge density near atoms, that can be typically expanded over spherical
harmonics inside a sphere of appropriate radius around each site. ρnon-loc (r) is a smoothly
varying charge density in the interstitial region. The separation density ξ(r) is chosen in
such way that each square bracket remains neutral and as smooth as possible. For ξ(r), a
dual representation is used: by spherical harmonic expansion around centers in the first
bracket and e.g. by plane wave expansion in the second bracket. Then the contribution
to the potential follows by multipole expansion from the first bracket and by Fourier
transformation from the second bracket.
29
T. Helgaker and P. R. Taylor, “Caussian Basis Sets and Molecular Integrals”, in: “Modern Electronic
Structure Theory”, Ed. D. R. Yarkony, World Scientific, Singapore 1995.

61
4.3 Planewaves and pseudopotentials
Earlier, we already dealt with planewave basis functions in order to calculate different
integrals related to a system of non-interacting electrons. When dealing with planewave
bases in crystals, we’ll write them down as
1 i(k+G)r
χG
k = √ e .
Ω
Here, k stands for a point in the Brillouin zone labeling a certain irreducible representation
of the translation group. In other words, this is an “artefact” of the Born – von Kármán
periodic boundary conditions. Infinite number of cells in a crystal maps onto infinite
number k points, effectively a continuous distribution, over the Brillouin zone. As we
scan the Brilluoin zone, we study the energy dispersion. As the repeated unit cell gets
larger (in one dimension, as for slabs, or in all dimensions, as for a “molecule in a box”),
the Brilluoin zone shrinks correspondingly, and the dispersion effectively disappears – one
gets instead a sequence of discrete energy levels. In this case, it may be sufficient to use
just one k-point, k = 0.
For each k (the results for different k do not mix other than via the construction of
density), G runs over plane waves actually used as basis functions. G labels the sites of
the reciprocal lattice, usually within some cutoff, |k + G| ≤ Gmax . (The actual set of G
can be therefore slightly different for different k). Further on, G will number the matrix
elements in the planewave representation, e.g.,
1Z 0
SGG0 = dr ei(G −G) r = δGG0 .
Ω
Ω

The big advantage of the planewave basis is the easiness with which some matrix elements
can be calculated. The momentum operator and hence the kinetic energy are diagonal in
the planewave basis, yielding

h̄2
k
TGG 0 = (G + k)2 δGG0 .
2m
This solves analytically an im-
G =2b G =−b G =0
portant (for numerical control
of accuracy of a computational
scheme in question) case of
empty lattice. The figure on EF
the right shows several lowest
bands (labeled by G) for one- G =−b G =b
dimensional empty lattice. The G =2b
occupied eigenstates (for sev- k
eral k within the first Brillouin −b b 2b
zone) are indicated by dots.
_ __
b __
b
2 2

62
 The presence of non-constant potential disturbs this band structure, and eigenfunc-
tions are not “pure” planewaves anymore, but their linear combinations. In order to
calculate them by diagonalization, one needs the matrix elements of the Hamiltonian that
read

h̄2
Z
1
(k + G)2 δGG0 +
0
k
HGG 0 = ei(G −G)r u(r)dr +
2m Ω
2 X
exp[i(G0 − G)r] exp[i(G2 − G1 )r0 ]
Z Z
e
+ D k
dr dr0 +
Ω2 G1 G2 G1 G2 |r − r0 |
Z
1 0
+ ei(G −G)r VXC (r)dr , (4.20)
Ω

k
where DG 1 G2
is the density matrix as introduced in Eq. (4.5).
The Hartree term in the second line reduces to the Fourier transform of the Coulomb
interaction,
Z
0
Z
exp[i(G2 − G1 )(r0 − r)]
drei(G −G+G2 −G1 )r d(r0 − r) ,
| {z } |r0 − r|
Ω δ(G0 −G+G2 −G1 )

whereas external potential and VXC (r) contribute via their Fourier transforms:
X
hG|VXC |G0 i = VXC (Q) δ(Q + G0 − G) .
Q

Only these last terms are actually problematic in a planewave calculation, because they
may be strongly localized in real space, so that their Fourier expansions do not converge
up to a prohibitively large G-cutoff. This can be understood as a consequence of the fact
that these terms provide a strong perturbation to the empty lattice case and hence mix
the basis states far separated in energy. Or, a deep potential in the vicinity of atomic cores
results in rapidly fluctuating density and hence high G cutoff for its proper description.
Such problems did not arise in tight-binding-type schemes, where the basis functions could
be tailored to account for such fluctuations.
The only possibility to save a purely planewave basis for practical calculations is to
discard core states, considering them as “frozen”. In other words, their spatial form can
be obtained from a single-atom calculation and further on to be believed to remain the
same in a molecule or a solid. The justification for such treatment could be, that the
core states do not participate in chemical bonding. The exclusion of core states from
the direct consideration means that we have to cope in the following not with the effect
of bare nuclei potentials, but rather with the potential screened by core electrons. This
potential is more smooth and shallow than the “true” one. This simplification gives rise
to a family of pseudopotential methods, which are contrasted to “all-electron methods”
(i.e., those where all electrons, valence and core, are treated on equal footing). gives rise
to a family of pseudopotential methods, which are contrasted to “all-electron methods”
(i.e., those where all electrons, valence and core, are treated on equal footing).

63
 p −pseudofunction
The most straightforward procedure of screening the
“true” potential with fixed core density is not practi-
cally useful. For one thing, the Coulomb field of a not
4p function
fully compensated bare charge remains singular at the
nuclei. Moreover, a true valence wavefunction must have
nodes in the intraatomic region for ensuring its orthogo-
nality to the core states. The description of these nodes
by plane waves needs high cutoffs. In reality, one works
p −pseudopotential with smooth node-free pseudofunctions generated in a
shallow pseudopotential. In historical perspective, pseu-
dopotentials have often been “model” ones, constructed
ad hoc and tunable with just several parameters. Nowa-
rc days, pseudopotentials in use are usually of ab initio
type. They are “cooked” (with the use of certain ap-
proximations and criteria, to be discussed further on)
true potential from “true” (all-electron; in a sense that all electrons
participate in the solution of the Kohn–Sham equations)
solutions for free atoms or ions.
The construction of a pseudopotential typically starts with the choice of an appropriate
reference configuration (e.g., Fe3d7 4s1 ) and “pseudoization radii” rC (see the figure) that
can be different for different l-channels. As a rule, the following conditions are imposed:
• the pseudofunction must have no nodes (in order to avoid wiggles that would demand
for higher cutoff);
• the pseudofunction matches the all-electron one beyond the cutoff radius;
• norm conservation, meaning that the charge contained within the pseudoization radius
is the same for the pseudofunction and the all-electron one. (This condition however can
be relaxed; deviations from this rule give rise to ultrasoft pseudopotentials, to be specially
discussed below);
• the eigenvalues corresponding to pseudofunctions must be equal to those of the all-
electron solution – at least for the reference configuration, and in the ideal case over a
range of reasonable configurations, meaning the pseudopotential is transferable.
The pseudopotential approach is related to the orthogonalized planewave method
(OPW), in which the basis set consists of plane waves, orthogonalized to lower-lying
core states χαcore :
XX
χk+G (r) = ei(k+G)r − hχαcore |ei(k+G)r i χαcore (r) . (4.21)
α core

We now consider the Phillips–Kleinmann construction of pseudopotential which, if

formally appearing in the Kohn–Sham equations, leads, as a solution, to the pseudofunc-
tion ϕPS
c . Let H be the Hamiltonian whose core and valence wavefunctions are ϕc and
ϕv , correspondingly:

Hϕαc = εαc ϕαc ;

Hϕkv = εkv ϕkv .

64
In analogy with Eq. (4.21), we construct pseudofunctions from valence and core states
(we remove wiggles in valence states by cancelling them by those in core states):
X
ϕPS
v = ϕv + aαvc ϕαc . (4.22)
αc

By multiplying with ϕc’ and integrating, one gets the expansion coefficients:
X
hϕβc0 |ϕPS β
v i = hϕc0 |ϕv i + aαvc hϕβc0 |ϕαc i
| {z } αc | {z }
= 0, since core and = δαβ δcc0 ;
valence states core states at different ;
correspond to the centers do not overlap
same Hamiltonian

whence aαvc = hϕαc0 |ϕPS

v i. (4.23)

Now we act on ϕPS

v of Eq. (4.22) by H:
X
HϕPS
v = Hϕv + aαvc (Hϕαc )
| {z } αc
ε v ϕv
" #
X X
= εv ϕPS
v − aαvc ϕαc + aαvc εαc ϕαc
αc αc
X
= εv ϕPS
v + aαvc (εαc − εv )ϕαc .
αc
h X i
H+ (εαc − εv ) |ϕαc ihϕαc | |ϕPS PS
v i = εv |ϕv i . (4.24)
αc | {z }
projection
operator

It should be emphasized here that H in Eq. (4.24) is the conventional Kohn–Sham Hamil-
tonian, that includes contributions from ion cores, all electrons in the Hartree term and
exchange-correlation potential depending on the total density. The transformation leading
to (4.24) is merely a technical regrouping of terms. The additional term in the Hamilto-
nian yielding the pseudofunction, as compared to the Hamiltonian H yielding the “true”
wavefunction, can be added to the conventional potential V (r):

e2 Z ρ(r) ρ(r0)
V (r) ≡ u(r) + dr dr0 + VXC (r)
2 |r − r0 |

and gives rise to the Phillips – Kleinman pseudopotential:

X
V PK = V (r) + (εαv − εc )|ϕαc ihϕαc | . (4.25)
αc

It is energy-dependent (via εv ); outside the core region it coincides with V (r). Since
εv > εv , the additional term is repulsive and hence the pseudopotential is softer than
the original potential. But the valence eigenvalues are by construction identical with those
of the corresponding all-electron problem.

65
 Actually, the construction of pseudopotential in reality is seldom done this way –
because the main idea is to eliminate core wavefunctions from calculation, rather than
calculate projections of pseudofunctions on them. In practice, one typically starts from a
smooth pseudofunction – there is a certain freedom in choosing it – and then inverts the
radial Schrödinger equation to get the corresponding pseudopotential:
pseudo l(l + 1) 1 d2 h PS i
Vl screened (r) = εl − + r Rl (r) . (4.26)
r2 2rRlPS(r) dr 2
This gives a screened pseudopotential, which incorporates the Hartree and exchange-
correlation terms related to the pseudocharge only, i.e. to the charge distribution in a
pseudo-wavefunction. When using the pseudopotential in the following, we would need
to re-calculate these terms according to how the pseudo-WF evolves; so we subtract the
corresponding terms and get an unscreened pseudopotential,
pseudo pseudo
Vl unscreened (r) = Vl screened (r) − Vl pseudo-H (r) − Vl pseudo-XC (r) .

A further progress in the use of pseudopotential methods was achieved along the line
of implementing the following priorities:
• the PP should be as soft as possible;
• it should be as transferable as possible (i.e., the PP generated for a given atomic con-
figuration should reproduce others accurately), that is essential for applications dealing
with simulations of chemical bonding;
• The pseudo-charge density should reproduce the “true” valence charge density as accu-
rately as possible.
These goals are essentially conflicting, for the reasons to be discussed below. In the
search for a compromise, the concept of norm conservation remained for a long time
generally adopted. The condition of norm-conservation reads
Zrc Zrc
2 2 2
r dr ϕPS
v (r) = r 2 dr ϕall-electron
v (r) . (4.27)
0 0

that must hold for all valence pseudofunctions. Since all-electron (single-atom) wavefunc-
tions and eigenvalues εv are different for different angular momentum values l, so would be
cutoff radii and pseudopotentials. Such (l-dependent) pseudopotentials are called semi-
local.
It is good for the transferability of pseudopotential if the slope of the pseudofunction
equals that of the true wavefunction at the pseudoization radius. Taken together with
the norm-conservation, this formulates the condition on the logarithmic derivatives:
1 d ϕPS (rc , E) 1 d ϕall-el.(rc , E)
= . (4.28)
ϕPS (rc , E) dr r=rc
ϕall-el.(rc , E) dr r=rc

The larger the energy “window” within which the logarithmic derivatives match approx-
imately, the better the transferability.
In 1982, Bachelet, Hamann and Schluter proposed a scheme to construct transferable
semi-local pseudopotentials. The scheme consists of several steps:

66
(1) Choose reference configuration, rc small rc large
calculate all-electron eigenvalues and
eigenfunctions of a free atom.
(2) Choose pseudoization radii for each
l-channel. The value of choice is
slightly beyond, or inside, the outmost
maxima of the all-electron wavefunc-
tion. Larger rc results in softer, but less
transferable, pseudopotential, smaller
rc in hard but more transferable one.
(3) Construct “first-step pseudopotential” Vl (r by cutting off the singularity in the all-
atomic potential:     
r r
Vl (r) = V (r) 1 − f + Cl f .
rc rc
Here, f (x) is essentially an arbitrary monotonic function such that f (0) = 1, f (∞) = 0,
cutting off rapidly about x = 1.
h  i
r
V (r) 1−f rc
Vl (r)
rc rc
1 
r
 r
f rc

Cl
1  
r
1−f rc
V (r)
1

(4) Norm conservation is imposed by adding a correction to the core region, e.g.
h i
ϕPS (r) = γl ϕ(1) (r) + δl gl (r) ,
where δl is a variable parameter that enforces norm conservation, γl is a scaling fac-
tor γl = ϕall-el. /ϕ(1) (in the outer region r > rc these functions may differ at most by a
constant); gl (r) is an arbitrary, in principle, function, that must however have correct
behaviour (∼ r l+1) near the origin and fall down at rc ; gl (r) = r l+1 f (r/rc ) is a possible
candidate.
(5) The screened pseudopotential (the one that would result in ϕPS (r)) is obtained by
numerical inversion of the Schrödinger equation.
(6) This pseudopiotential is then unscreened by removing Hartree and exchange-correlation
contributions related to valence pseudocharge, as discussed above.
One should note that in the course of constructing the pseudopotential, a particular
charge distribution (dependent on a reference atomic configuration) was used, that gave
rise to Hartree and exchange-correlation parts. Later on, the pseudo-ϕ would vary in the
course of iterations, and one should take into account how Hartree and XC contributions

67
will change. The Hartree term presents no problem since it is linear in charge. The
exchange-correlation potential is not linear in ρ, and the subtraction of just valence con-
tribution for its “unscreening” is often not satisfactory. What is instead done in reality –
a fixed smooth function imitating the core density (just for the sake of calculating VXC is
removed at the unscreening step and then added on each iteration to the actual pseudo-
density, before calculating VXC . 30 (One could of course remove a true core density, but
it is strongly fluctuating, and we would like to avoid the wiggles in the pseudopotential.
Usually a more smooth analytical function does the job well.)
As a result of the above scheme, a semi-local pseudopotential is generated that has
the following form:
lX
max

V PS (r) = V loc (r) + VlPS (r)P̂l , (4.29)

l=0

where V loc is a local pseudopotential, P̂l is the angular momentum projection operator;
the corresponding expansion coefficients of the pseudopotential VlPS converge fast with l,
so that lmax = 1 − 2 usually suffices.
The matrix elements in the planewave basis reduce to
Z
1 0
dr e−i(k+G)r VlPS (r) P̂l ei(k+G )r =
Ω
Z ∞
2l + 1
= 4π Pl (cos γ) VlPS (r) jl (|k + G| r) jl (|k + G0 | r) r 2 dr , (4.30)
Ω
0
(k + G)(k + G0 )
with cos γ = .
|k + G| |k + G0 |

Here, Pl are Legendre polynomials. An obvious disagvantage is that matrix elements de-
pend not on (k+G) − (k−G0 ) = G − G0 , but on (k + G) and (k + G0 ) separately. Thus,
for n planewaves in basis, the number of matrix elements of a semi-local pseudopotential
is ∼ n(n + 1)/2, that may rapidly become prohibitively large for their storage or calcu-
lation. Note that the matrix elements of local (Phillips – Kleinmann) pseudopotential
of Eq. (4.25) obviously depend on (G − G0 ) only and hence their number scales ∼ 8n
(number of reciprocal latttice sites within the sphere of the radius 2Gmax , where G is the
planewave cutoff and 2Gmax – the maximal value of G−G0 ).
This problem can be solved if semi-local pseudopotential VlPS (r) is transformed to a
fully non-local but separable form, as proposed by Kleinmann and Bylander.31
ED
VlPS (r) ϕPl ϕPl VlPS (r)
VlPN (r) = . (4.31)
hϕPl |VlPS (r)|ϕPl i

One can see that E E

VlPN (r) ϕPl = VlPS (r) ϕPl . (4.32)
30
Steven G. Louie, Sverre Froyen and Marvin L. Cohen, Nonlinear ionic pseudopotentials in spin-
density-functional calculations, Phys. Rev. B 26, 1738–1742 (1982).
31
Kleinmann and Bylander, Phys. Rev. Lett. 48, 1425 (1982).

68
The advantage of such separation is that the pseudopotential aquires a form similar to
X
V PN = |V i hVj | ,
j

and its matrix elements in the planewave basis

X
hG| V PN |G0 i = hG|Vj ihVj |G0 i
j

become convenient for calculation.

Let us emphasize the origin of the problem with semi-local pseudopotential once more.
0
The P̂l projection in Eq. (4.29) acts on ei(k+G ) and produces results depending on the
direction of (k + G). If then (k + G0 ) has another direction, the angle between them
needs to be treated specially, as in Eq. (4.30). In the Kleinmann–Bylander construction,
a pseudo-wavefunction is built in that “absorbs” the effect of projection angle – the result
is merely a function of r. So one can decouple the effects of (k + G) and (k + G0 ). The
price we pay is the acccuracy: (4.32) holds exactly only for the reference function, i.e. the
one used for the construction of pseudopotential. But in reality, the separation is used
over a certain range of energies and eigenstates in the valence band, that is not exactly
justified – but may acceptable for well transferable pseudopotentials.

4.4 Ultrasoft pseudopotentials

Vanderbilt32 suggested to abandon the norm-conservation condition (4.27), that would
allow to make the pseudoization radius rc large – essentially, limited only by the con-
dition that the spheres of this radius centered on different atoms must not overlap in a
simulation. A big advantage would be that pseudopotentials generated with larger rc are
much softer and hence much lower planewave cutoff would be needed. There are also
disadvantages:
• the formalism becomes less transparent and
more mathematically involved, because the norm−conserving
functions get a non-trivial overlap; charge den- pseudofunction
P PS
sity is not |ϕ (r)|2 anymore but includes
contributions from “local functions”;
• a pseudopotential becomes less transferable.
The latter disadvantage is not very serious be-
cause one can typically afford generating the
pseudopotential anew several times, as the cal-
culation converges and the final charge config- pseudofunction
uration becomes apparent. The mathematical of ultrasoft PP
complications are briefly discussed below, fol-
lowing the original paper by Vanderbilt.
32
David Vanderbilt, Soft self-consistent pseudopotentials in a generalized eigenvalue formalism,
Phys. Rev. B 41, 7892–7895 (1990)

69
 We label all-electron wavefunctions ψi (r), where i is composite index, i = {εilm},
whereby one allows in principle more than one reference energy εi per l-channel. ψi (r) is
hence a radial solution of the Schrödinger equation at such fixed energy:
h i
T̂ + V AE (r) ψi (r) = εi ψi (r) . (4.33)

One introduces the following cutoff radii:

rcl for wavefunction,
rcloc for local pseudopotential,
R – diagnostic radius, beyond which all all-electron and pseudo properties (wavefunctions,
potential) agree for all l channels. In the following the bra-ket notation stands for integrals
up to R:
Z ZR
∗
hψi |ψj i ≡ Ylm (Ω)Yl0 m0 (Ω)dΩ r 2 ψi (r) ψj (r) .
0
Next we introduce pseudofunctions φi that match corresponding ψi smoothly at rcl and
is norm-conserving:
hφi |φii = hψi |ψi i .
It will be a solution of the Schrödinger equation with an, in principle, nonlocal pseudopo-
tential: h i
T̂ + V loc (r) + V NL (r) φi (r) = εi φi (r) . (4.34)
that is exact for the reference energy εi . This allows to construct the “difference” function
χi , h i
|χi i = εi − T̂ − V loc (r) |φi i . (4.35)
From (4.34) and (4.35),

V NL |φi i = |χi i , hence

|χi i hχi |
V NL = . (4.36)
hχi |φii
This bears some similarity to the Phillips – Kleinmann pseudopotential (4.5). Indeed,
|χi i has similarity to core states |φαc i in (4.25) in a sense that it incorporates the wiggles
to be suppressed in a pseudofunction.
An important extension in the Vanderbilt scheme is the possibility to allow two or
more reference energies εi at which the scattering properties of the pseudopotential will
be correct. The reference functions φi taken at different εi within the same l-channel are
of course not orthogonal. these non-orthogonalities will be accounted for by augmentation
coefficients Qij :
Qij = hψi |ψj i − hφi |φj i (4.37)
which are Qij = 0, if norm-conservation condition is imposed – but this is not necessary.
Further on, we define
Bij = hφi|χj i (4.38)
and X  X  
|βi i = B −1 |χj i hβi | = hχj | B −1 . (4.39)
ji ij
j j

70
The meaning of |βi i is that they are localized (since they disappear outside R, as |χi i)
projectors which are dual to pseudofunctions |φi i, i.e.,

hβi |φi i = δij . (4.40)

Indeed, X 
hβi |φi i = B −1 hχl |φj i = δij .
il | {z }
l
Bjl

With this, we arrive at X

V NL = Bmn |βm i hβn | , (4.41)
mn

because from (4.36)

|χiihχi |φi i
V NL |φi i = = |χi i .
hχi |φi i
and from (4.41),
X
V NL |φi i = Bmn |βm i hβn |φi i =
mn | {z }
=δni
X X X 
= Bmi |βm i = Bmi B −1 |χl i =
lm
m m l
XX 
= B −1 Bmi |χl i .
lm
l m
| {z }
=δli

The norm of all-electron wavefunction is hψi |ψi i. It can be recovered from pseudo-
wavefunction if one introduces a non-trivial overlap,
X
S=1+ Qmn |βm ihβn | , (4.42)
mn

so that
X
hφi |Ŝ|φj i = hφi |φj i = hφi |φj i + Qmn hφi |βm i hβn |φj i =
mn | {z } | {z }
δim δjn

= hφi |φj i + Qij = [using (4.37)] = hψi |ψj i .

Then (non norm-conserving) functions |φii satisfy the equation:

" #
X
T̂ + Vloc + Dmn |βm ihβn | |φii = εi Ŝ|φi i (4.43)
mn

with
Dmn = Bmn + εn Qmn . (4.44)
Indeed, the left side of Eq. (4.43) reduces to
  X
T̂ + Vloc |φi i + |χi i + εi Qni |βm i ,
| {z } m
εi |φi i

71
that equals the right side.
As was discussed before for norm-conserving pseudopotentials, for the final construic-
tion of pseudopotential one has to “unscreen” its ion and non-local parts:

ρ(r0 )
Z
ion
Vlov (r) = Vloc (r) − dr0 − VXC (r) ;
|r − r0 |
Z
(0)
Dmn = Dmn − dr0 Vloc (r) ρ(r0 ) , (4.45)

where ρ(r) is the valence density – a “pseudo” density, but corresponding to norm con-
servation.
Exactly as the integral over this density over the R-sphere can be recovered from
Eq. (4.37), the density at arbitrary r < R can be reconstructed in terms of augmentation
functions Qij (r):
Qij (r) = ψi∗ ψj (r) − φ∗i φj (r) . (4.46)
the indices i, j numbered so far reference energies. The equations like (4.43) hold exactly
for such reference states. Our aim is, however, to solve the bandstructure problem in
a whole interval of valence band energies, using plane waves as basis. Therefore the
generalized gradient problem
(H − εS) Φαk = 0 ,
of which Eq. (4.43) is an example, must be solved for each k point, yielding a sequence of
bands numbered by α. The explicit form of Hamiltonian and overlap includes summations
over reference states i, n etc. The (corrected) density includes summation over occupied
band states, like in all other methods. Explicitly,
X X
ρ(r) = φ∗αk (r)φαk (r) + ρij Qji (r) . (4.47)
α, k ij
(occupied)

with ρij constructed from projections of occupied band states onto |βi i:
X
ρij = hβi |φαk ihφαk |βj i . (4.48)
α, k
(occupied)

These terms must be recalculated on each iteration.

While the pseudo-component of the density (4.47) is now smooth, the fluctuating
part of it is shifted into augmentation functions. They can be represented by spherical
harmonics expansion inside R,
X ij
Qij (r) = Clm Ylm (r̂) Qij (r) ,
lm

but still they have to be transformed to reciprocal space for integration (e.g., for solving
the Poisson equation). Therefore, the fact that they are localized may result in high
cutoff in their planewave expansion. But this is the price paid for keeping the dimension
of of planewave basis compact. Moreover, it is not necessary to keep true augmentation

72
functions; in practice it suffices to imitate them by more smooth l-dependent functions
which faithfully represent moments of true augmentation functions, up to a certain lmax .
Let us summarize some important features of ultrasoft potentials scheme.
• The cutoff radius R is only limited by next-neighbor distance (pseudoization spheres
must not overlap).
• Within this R, individual rc can be adjusted in order to keep pseudofunction as smooth
one can manage.
• The necessary planewave cutoff for the planewave basis is drastically reduced.
• The amount of computational work (in the generation of pseudopotential etc.) is in-
creased. But most of these additional efforts need not to be repeated in the course of
iterations.
• As augmentation charge varies in the course of iterations, so does the local potential
in the sphere. It can be considered as part of pseudopotential, so the pseudopotential
develops itself as the calculation proceeds (similar to all-electron methods).
• The transferability of pseudopotentials is worse than of conventional pseudopotentials
generated with smaller rc . But this is compensated by an option to use several reference
energies instead of just one, and by the possibility to generate pseudopotential anew in the
course of iterations, as final electron configuration emerges. The main area of application
of ultrasoft pseudopotentials is large systems, where the relative cost of the pseudopoten-
tial generation is low, as compared to solving the electronic structure problem. For the
latter, a low planewave cutoff is a major advantage.

4.5 All-electron methods using muffin-tin geometry

We discuss now methods tat allow to treat both valence and core states on the same
footing, as it was already the case with the use of Gaussian orbitals. Differently from the
latter, we consider now the methods where the basis is not fixed (i.e., it develops from it-
eration to iteration), or where there is no basis, in a sense that the solutions come out not
from a single diagonalization. Historically, this branch of methods’ development starts
with the muffin-tin approximation and augmented planewave method as proposed by
Slater in 1937 and 1964. The idea behind it was S
to use arbitrarily deep – but spherically symmet-
ric – potential near nuclei and a flat one in the
interstitial (I) region, with the following dual rep-
resentation for the wavefunction: Rα
 P
 √1 CG e i(k+G)r
, r∈I
 Ω
G
ψk (r) = P αk α (4.49)

 Alm ul (r)Ylm (r̂) , r<S
lm

ul (r) is a regular (at the origin) solution of radial

Schrödinger equation in spherical symmetric po-
tential:
" #
d2 l(l + 1)
− 2+ + V (r) − εl rul (r) = 0 . (4.50)
dr r2

73
εl is a parameter, variable in order to achieve the continuity of the wavefunction at the
sphere boundary. The continuity of the dual representation (4.49) is necessary in order to
have well defined kinetic energy. The matching condition, relating CG and Alm , follows
from the planewave expansion
X
eiKr = 4πeiKRα il jl (K|r − Rα |) Ylm ∗ (K̂)Ylm (r −ˆRα ) (4.51)
lm

around an arbitrary cite Rα .

1 X X
√ CG 4πei(k+G)Rα il jl (|k+G|r) Ylm ∗ (k+G)Ylm (r̂)
Ω G lm

and X
Alm ul (r) Ylm (r̂)
lm

must match at r = S for all {l, m}, hence

4πil X
Aαk
lm = √ ei(k+G)Rα jl (|k+G|S) Ylm ∗ (k+G) (4.52)
α
Ωu (S) G
l

with εl fixed as external parameter. For a given composition of wavefunctions from plane
waves in the interstitial, one can thus recover Alm which would guarantee the matching
at S.
Since the matching can be achieved for every single G-wave, one could exploit the
idea of using a global basis set, labeled by G, and inside each sphere there will be an
P
augmentation to exp[i(k + G)r] by corresponding lm Aαlm ul (|r − Rα |). There will be
Hamiltonian matrix HGG0 and overlap SGG0 ,and one could search for the solutions of the
secular equation
khG1 |H − εS|G2 ik = 0 .
Note that ul are only solutions inside sphere for a certain energy. If ε deviates from it,
ul is not a solution anymore. One cannot simply fix the energy parameter εl to any ar-
bitrary value. Instead, one has to scan the values of this parameter that fulfil the condition

khG1 , ε|H − εS|G2 , εik ≡ D(ε) = 0 ,

D (ε )
that is not a single diagonalization
problem. In practice, one calculates the
values of the determinant D(ε) on a suf-
ficiently fine energy mesh and tries to
localize its roots. This has to be done
ε
for each k point. The determinant is an
ever-decreasing function of energy con-
sisting of many branches; the number
of roots within a given energy interval
is not known in advance.
74
An additional complication comes from the fact
that for some energy values one or another of
ul ( r ) S
atomic solutions ul (r) may have a node at the
sphere boundary, then (4.52) would diverge and "bonding"
ε1 r
augmentation of this particular radial solution
to planewaves breaks down – an additional nu-

<
merical problems, if there is a root nearby. This
situation is almost unavoidable, because radial
ε2
solutions of the Schrödinger equation for each

<
given l show a systematic behavior as the in-
tegration energy grows: the function gradually "antibonding" ε3
becomes more compact, and at certain energies
a next node squeezes into the sphere of radius S.

<
This is shown schematically in the following fig-
ure. Two special cases u0l (S)=0 and ul (S)=0 are ε4
labeled as “bonding” and “antibonding”, corre-
spondingly, for further reference.
A related calculation scheme, that is however derived from quite different starting
point, is the Korrings–Kohn–Rostoker (KKR) method. It uses the fact that in the
interstitial, i.e. in the region with constant potential V0 , the Schrödinger equation degen-
erates into the Helmholz equation
!
h̄2 2
− ∇ + V0 −  ψ = 0
2m

with a solution s 
2m
ψ = exp i (ε − V0 ) r  ≡ eiκr
h̄2

that can be expanded in spherical harmonics with jl (κ2 , r) – regular at the origin and
nl (κ2 , r) – singular at the origin radial functions. Such functions to be used in the following
are, up to a normalization, spherical Bessel and Neumann functions:

(2l − 1)!! (κS)l+1

jl (κ2 , r) ≡ jl (κr) ; nl (κ2 , r) ≡ −nl (κr) . (4.53)
2(κS)l (2l − 1)!!

These plane waves are scattered on individual atomic potentials, that gives rise to phase
shifts η. The phase shift shows in which relation regular and singular solutions are present
in the scattered wave. In the most direct way the relation between Jl and nl is accounted
for by the potential function of scattering problem P( , κ2 ). In terms of it, the wavefunction
can be expanded over spherical harmonics inside and outside the sphere with scattering
potential in the following way:

X  ul (ε, r) , r<S
ψ(ε, r) = Ylm (r̂) × (4.54)
lm
 nl (κ2 , r) − Pl (ε, κ2 ) jl (κ2 , r) r > S

75
The potential function can be obtained from the condition that both the radial wavefunc-
tion ul (ε, r) and its first derivative match a combination of nl (κ2 , r) and jl (κ2 , r):

{ul (ε), nl (κ2 )}

P (ε, κ2) = , (4.55)
{ul (ε), jl (κ2 )}

where the Wronskian of two radial functions, taken at the sphere boundary, is defined as

{f, g} ≡ S 2 [f (S)g 0(S) − f 0 (S)g(S)]

that gives specifically for the functions (4.53)

S
{nl , jl } = .
2
The so calculated potential function relates to the phase shift ηl in the following way:

2 2(Sκ)2l+1
Pl (ε, κ ) = − cot ηl (ε, κ) . (4.56)
[(2l − 1)!!]2

Indeed, for the case of no scattering (the potential equal to V0 everywhere inside the
sphere) ul = jl , P (ε, κ2 ) = ∞ and η = 0. It follows from Eq. (4.56) that Pl (ε, κ2 ) is an al-
ways increasing function of ε. Adding Pl (ε, κ2 )jl (κ2 , r) to the wavefunction representation
(4.52) results in muffin-tin orbital (MTO) χl :

 ul (ε, r) + Pl (ε, κ2 )jl (κ2 , r) , r<S
χlm (ε, κ2 , r) = Ylm (r̂) × 2
(4.57)
n
l (κ , r) . r>S

One could in principle fix ε and use muffin-tin orbitals as localized basis functions, but the
quality of this basis would be quite bad for any general ε, for the same reasons as discussed
above for augmanted plane waves. Here also, instead of performing a diagonalization one
can, as a more practical scheme, scan the integral of energies of interest and search for
those values of ε at which the matching conditions are satisfied at all sphere boundaries
simultaneously. This is achieved by taking into account the expansion of non-regular
solution, taken at any center Rα , around any other center Rβ , that has the following
form:
X
nlm (κ2 , |r − Rα |) = jl0 m0 (κ2 , |r − Rβ |) ×
l0 m0
0 00
X 0 m00 −l+l0 −l00 2(κRα )l +l−l (2l00 −1)!! ∗
× 4π Clm m
l0 l00 i nl00 (m0 −m) (κ2 , |Rα − Rβ |)
l00 (2l0 −1)!! (2l−1)!!
X
≡ jl0 m0 (κ2 , |r − Rβ |) SRβ l0 m0 , Rα lm (κ2 ) . (4.58)
l0 m0

The condition that a linear combination of MTOs is the solution in all space reads
Xh i
PR0 l0 (ε, κ2 ) δRR0 δll0 δmm0 − SR0 l0 m0 , Rlm (κ2 ) CRlm = 0 . (4.59)
Rlm

76
The zeros of the determinant of this equation

||P(ε, κ2) − S(κ2 )|| = 0

determine the solutions of the scattering problem.

Comparing the performance of the KKR method with the augmented plane wave
scheme one can note the following advantages:
• a much more compact basis set can be typically used. Whereas a high G-cutoff in the
augmented plane wave method may typically result in ∼ 1000 plane waves per atom, the
size of the basis set in the KKR is essentially limited by maximal l in the expansion of
structure constants. The latter converge quite fast, with lmax. ≈ 2 − 3 being typically
acceptable, resulting in (l + 1)2 ∼ 16 functions per atom (for the description of valence
states);
• radial solutions and functions in the interstitial are allowed to match along with their
derivatives.
• The structure constants do not depend on scattering potentials and they are scaled in a
simple way as the lattice constant changes; therefore they do not need to be recalculated
in the course of electronic-structure iterations and can be stored in advance (as functions
of κ2 ) for any given crystal structure.
Still the problem preventing to use a single diagonalization in order to obtain all
eigenvalues at once is unsufficient variational freedom of numerical solutions inside the
spheres, taken at certain fixed ε. This can be helped by linearization, due to an observation
by Andersen33 that ul (ε, r) as functions of energy is with good accuracy linear:

ul (ε, r) ≈ ul (εν , r) + (ε − εν )u̇l (εν , r) . (4.60)

The linearization of augmented plane waves leads to the LAPW method, the linearization
of scattering-wave formalism – to the LMTO method. The advantage of the former is
a relatively easy generalization over non-spherical potentials, the advantage of the latter
– compact and highly efficient basis; the inclusion of potentials of general form is also
possible but somehow more mathematically involved.
Differentiating Eq. (4.50) in energy yields:
" #
d2 l(l + 1)
− 2+ + V (r) − εl r u̇l (r, εl ) = r ul (r, εl ) , (4.61)
dr r2

with
d
u̇l ≡
ul (r, ε) .
dε
In the LAPW formalism, both functions ul and u̇l are used to match the plane wave on
the sphere boundary, whereby the wavefunction expansion (4.49) takes a form:
 P
 √1 CG ei(k+G)r , r∈I
 Ω
G
ψk (r) = Ph k i (4.62)

 k
Alm ul (r) + Blm u̇l (r) Ylm (r̂) , r <S.
lm

33
O. Krogh Andersen, Linear methods in band theory, Phys. Rev. B 12, 3060–3083 (1975).

77
This combination now has sufficient variational freedom: if band energy deviates from
an in advance fixed value εν , the shape of the wavefunction inside the sphere can still be
recovered using the approximation (4.60). One can then use as basis functions ψ G (r),
corresponding to different G in the dual representation (4.62), with the coefficients Alm
and Blm obtained from matching conditions there. The basis functions are then energy
independent, and the electronic-structure problem can be solved by matrix diagonaliza-
tion, as in any method with fixed basis. The error in the band energy due to linearization
is ∼ (ε − εν )4 . Typically, it is negligible within an energy window of 1–1.5 Ry.
Two parameters A and B for each (l, m) are fixed by two requirements, that the wave-
function and its radial derivative match the plane wave. Therefore the asymptote problem
of ul (S)=0 in the denominator, as in Eq. (4.52) does not arise: ul (S) and u0l (S) cannot be
equal to zero simultaneously. ul (S) = 0 corresponds to the “antibonding” solution (see
preceding figure) and u0l (S) = 0 – to the “bonding” solution; their corresponding energies
lie apart by ∼ the valence band width.
With this, the secular equation becomes:
h i
k k
HGG 0 − ε SGG0 CG0 = 0 . (4.63)

The expression for the overlap matrix

Z
1 0 X
SGG0 = drei(G−G )r Θ(r) + Sα (G, G0 ) (4.64)
Ω Rα

contains a step function, Θ(r), that is = 0 inside any muffin-tin sphere, and Sα (G, G0 ),
contributions to overlap from a sphere centered at Rα . The Hamiltonian matrix
Z
1 h i 0
k
HGG 0 = dre−i(G+k)r Θ(r) T̂ + V̂PW ei(G +k)r +
Ω h i
X
+ Hα (G, G0 ) + VαNS (G, G0 ) (4.65)
Rα

has Hα (G, G0 ) – spherical and VαNS (G, G0 ) – non-spherical (i.e., from l 6= 0) contributions
from muffin-tin sphere at Rα . The details of practical implementations of full-potential
LAPW method are well described in a book by D. Singh.34

34
David J. Singh, Planewaves, pseudopotentials and the LAPW method, Kluwer (1994), 115 pp.

78
 5. Response properties
in ab initio schemes
A number of important physical observables is expressed via derivatives of total energy
(or free energy) E. Examples are:

∂E forces on the nuclei; critical points on the potential energy surface

∂R (minima and saddle points)

∂2E force constants; vibrational frequencies;

∂Ra ∂Rb infrared and Raman spectra

∂3E
anharmonic constributions to vibrational frequencies
∂Ra ∂Rb Rc
∂E
(E – electric field): dipole moment
∂E

∂2E
infrared intensities
∂Ra ∂Ei

∂2E
polarizability; light scattering
∂Ei ∂Ej

∂3E
hyperpolarizability; second harmonic generation
∂Ei ∂Ej ∂Ek

∂3E
Raman intensity
∂Ra ∂Ei ∂Ek
etc.

If total energy is available in a calculation with {R} or {E} as parameters, the above
properties can be obtained by three methods:
1. A fit of sample points to an analytical form and subsequent differentiation. This
approach is very often used, but the results are subject to numerical instability (much
more data points are needed than there are parameters in the fitting function, for a
numerically stable fit).
2. Finite differences. This method is good if calculation algorithm is accurate, so that
numerical “noise” can be neglected. The following formulae are for the first derivative of

79
a numericall function f (x1 , . . . , xN ), defined on a mesh with a step h;:
!
∂f 1 h i 1 d3 ∂f 2
= f (. . . , x0i +h, . . .) − [f (. . . , x0i −h, . . .) + h
∂x i
2h 6 ∂3

...and for the second derivatives, with obvious notation of f (+) ≡ f (. . . , x0i + h, . . .) etc. :

∂2f 1 1 ∂4f 2
= 2 [f (+) + f (−) − 2f (0)] + h
∂x2i h 12 ∂x4i
∂2f 1
= [f (++) + f (−−) + 2f (00) − f (+0)−
∂xi ∂xj 2hi hj
−f (−0) − f (0+) − f (0−)] + O(f IV h2 ) .

3. Analytical differentiation, built in into the code. This typically means a considerable
programming load and an additional amount of calculation. However, the differentiation
is “exact”, and there exist a possibility to calculate many response properties (all compo-
nents of force constants etc.) in a single run. Also this may be advantageous in reducing
a number of different calculations to be done. For example, calculation of a second-order
dynamical matrix of a system with N atoms demands at least 9N(N −1)/2 + 1 calcu-
lations for different displacement patterns if only the total energies are available, but
[9N(N −1) + 2]/[2(3N −2)], i.e. much less, if the forces are available as well.

5.1 General considerations

In the following, we consider the problem of differentiating an expression which is varia-
tional (like, e.g., energy) and depends implicitly and explicitly on some parameters. Ex-
amples of explicit parameters are, e.g., positions of nuclei (and possibly of basis functions
pinned to them); the implicit dependence of the total energy is on variational parame-
ters – those describing the decomposition of orbitals over basis functions. Moreover, we
introduce a set of constarints that could be target pressure, or fixed spin moment. So
we deal with constrained minimization, using Lagrange multipliers. The function to be
minimized has a form
X
W = W 0 (C, λ, R) − fm (C, R)λm , (5.1)
m

where W 0 is e.g. the total energy expression, C – variational parameters, λ – Lagrange

multipliers, R – explicit parameters (e.g., positions of nuclei), f – constraints, formulated
as
fm (C, R) = 0 (5.2)
(like orthonormality of orbitals, fixation of total magnetic moment etc.) In less general
case, one can omit the constraints. Variational parameters will be determined from the
equations

80
 
 ∂W

 = 0, ∀ i
∂Ci (5.3)


 fm (C, R) = 0 , ∀ m
In the following, we use the notation:
∂W ∂W
≡ Wi ; ≡ Wa (5.4)
∂Ci ∂Ra
The optimized energy will be
 
E = W C(R), λ(R), R ,
and its gradient X X
Ea = W a + Wi Cia − fm λam = W a , (5.5)
i m
where W a can be obtained by direct differentiating in Ra , and Wi =0, fm =0 at equilibrium.
Similarly, the second derivative yields:
X X
E ab = W ab + Wia Cib − a b
fm λm , (5.6)
i m

and other terms, Wi Ciab and fm λab

m , disappear for the same reason. Since W is variational,
it must be stationary with respect to variations of C:
X X
Wi + Wij Cj − fmi λm = 0 . (5.7)
j m

This is something like Hartree-Fock, or Kohn–Sham equation equation, written down for
fixed positions of nuclei. Differentiating in Ra yields:
dWi X X
= Wia + Wij Cja − fmi λam = 0 , (5.8)
dRa j m

and since fm = 0 in a stationary state,

dfm X
= fa + fmj Cja = 0 . (5.9)
dRa j

So we arrive at a system of response equations

 X X



 Wij Cja − fmi λam = −Wia ,
 m
j
X (5.10)



 fmj Cja = −fm
a
,

j

which has to be solved in addition to “zero-order” equation (5.7). Once this is done, Wia
a
and fm can be inserted in Eq. (5.6) for Eab .
A similar analysis, but including more coupled terms, can be done for third derivatives.
This approach is described by Pulay.35
35
Peter Pulay, Analytical Derivative Techniques and the Calculation of Vibrational Spectra, in: Modern
Electronic Structure Theory, Part II, ed. by David R. Yarkony, World Scientific, Singapore (1995), pp.
1191–1240.

81
5.2 Forces
Let us discuss in more detail the evaluation of force, that must be, according to Eq. (5.5),
E a = W a . In practical calculations, evaluation of this derivative may be considerably
complicated by the necessity to take into account derivatives of the basis functions. This
is necessary if e.g. basis functions are centered on atoms and are displaced with them.
This is not necessary if a position-independent basis set (e.g., planewaves) is used. The
reason is the following. For the exact wavefunction,

E a = hψ|Ha |ψi

by the power of the Hellmann–Feynman theorem, as was shown in Eq. (3.19). We repro-
duce here this argumentation for completeness:
d
Ea ≡ hψ|H|ψi = hψ a | H|ψi +hψ|Ha |ψi + hψ|H |ψ a i =
dRa | {z } | {z }
E|ψi Ehψ|

= hψ|H |ψi + E [hψ |ψi + hψ|ψ a i] =

a a
h ia
= hψ|Ha |ψi + E hψ|ψi = hψ|Ha |ψi . (5.11)
| {z }
=const

However, if the wavefunction contains parameters pt dependent on perturbation (i.e.,

displacement of ions) either implicitly or explicitly, then (5.11) is not generally valid,
because
X ∂hψ|H|ψi
hψ a |H|ψi + hψ|H|ψ a i = 2 Re hψ a |H|ψi = pat 6= 0 . (5.12)
t ∂p t

The validity of the Hellmann–Feynman theorem can be restored if either

∀ pat = 0 ,

i.e. the basis is independent on positions of nuclei, as is for instance the case for plane
waves, or
∂ hψ|H|ψi
= 0,
∂ pt
i.e. all parameters are fully optimized, that is normally the case if the basis is complete.
Or, at least, for each basis function its derivative with respect to perturbation must also
be present in the basis. In practical terms, Hellmann–Feynman forces are seldom useful,
because even basis sets extended to include their gradients are not complete enough.
The total energy in the density functional theory is given by Eq. (3.17):
e2 ρ(r)ρ(r0 )
X Z Z
el.
Etot = εi − dr dr0 − VXC (r) ρ(r) dr + EXC [ρ] .
(i occupied)
2 |r − r0 |

When considering a dynamical problem, one must add here a contribution from atomic
cores that is not a constant anymore:

el. e2 X Zα Zβ
Etot [ρ] = Etot + . (5.13)
2 αβ |Rα −Rβ |

82
Given a fixed configuration of atoms {Rα }, the displacement of one atom Rα → Rα + ∆α
changes the total energy by
X X Z
(core)
δE = δ εi + δ β − dr ρ(r) δVKS(r) − FHF
α ·∆α . (5.14)
(i occupied) β (core)

The Hellmann–Feynman force is purely electrostatic, due to the displacement of density:

 
Z
d  XX eZβ ρ(r)
FHF
α = Zα − +e dr , (5.15)
d∆α β6=α τ |R α +∆ α −R β +τ | |r−R α −∆α |

τ are lattice translation vectors. Other contributions adding up to the total force are:36
d Etot
Fα ≡ − = FHF IBS core
α + Fα + Fα . (5.16)
d ∆α
FIBS is “incomplete basis set”, or “Pulay” force, due to the use of finite number of position-
dependent basis functions:
" #
X D δϕi E 1 X
FIBS
α = −2 Re T̂ + Veff − εi ϕi − hϕi | δ T̂ |ϕi i . (5.17)
i
δ∆α δ∆α i
(occupied) (occupied)

Fcore is a “core correction”, due to the fact that for core electrons (at least in the FLAPW
method) only spherical part of the potential is taken into account.
Z
Fcore
α = − ρα (r)∇VKS(r) . (5.18)

To give an idea of relative importance of dif-

ferent contributions to the force, the values (in
F numericali diff. = −41.45
mRy/a.u., for ∆α =0.089 a.u.) as calculated by
F by Eq. (5.16) = −41.38
the FLAPW method for Si are shown on the right.
F HF = −64.90
(after C. Ambrosch-Draxl, lecture at the Workshop
F IBS = 3.06
The Physics of the Electronic Behaviour in the Core
F core = 20.46
Region: All-Electron LAPW Electronic Structure Cal-
culations, 22.06–04.07.1998, Trieste. )

5.3 Dynamical matrix

We discuss now the calculation of second derivatives of energy in ion displacements,
according to Eq. (5.6). For this, we need to know Cib , i.e., Kohn–Sham orbitals to 1st order
in displacements. The Hellmann–Feynman theorem holds up to 3d order in perturbations:
Z " #
∂2E ∂ρR ∂VR ∂ 2 VR
= +ρ dr . (5.19)
∂Ra ∂Rb ∂Ra ∂Rb ∂Ra ∂Rb
36
B. Kohler, S. Wilke, M. Scheffler, R. Kouba and C. Ambrosch-Draxl, Force calculation and atomic-
structure optimization for the full-potential augmented plane-wave code WIEN, Comp. Phys. Commun.
94, 31 (1996).

83
The set of Kohn–Sham equations (3.16)
" #
h̄2 2
− ∇ + VSCF (r) ϕi (r) = εi ϕi (r) ;
2m
ρ(r0 )
X −Zα Z
2 2 δEXC
VSCF (r) = e +e 0
dr0 + ; (5.20)
α |r−Rα | |r − r | δρ(r)
X
ρ(r) = |ϕi (r)|2 .
i
(occupied)

can be “linearized”, introducing small parameter λ:

(0) (1)
ϕi (r) = ϕi (r) + λ ϕi (r) ;
(0) (1)
VSCF (r) = VSCF (r) + λ VSCF (r) ; (5.21)
ρ(r) = ρ(0) (r) + λ ρ(1) (r) .
(0)
The “zero-order” system for ϕ(0) , VSCF and ρ(0) is (5.20), and the “first-order” system
(where terms linear in λ are retained) can be easily written once the nature of pertur-
bation (that must be proportional to λ) is specified. As we are interested at present in
displacements of ions λwα as an example of such perturbations, the “first-order” equations
aquire the form:
" #
h̄2 2 (0) (1) (1) (0)
− ∇ + VSCF (r) − εi ϕi (r) = −VSCF (r) ϕi (r) ;
2m
" #
ρ(1) (r0 ) 0
Z
(1) X Zα wα (r−Rα) dVXC (r)
VSCF (r) = e2 + e2 dr + ρ(1) (r) ; (5.22)
α |r−Rα|3 |r − r |0 dρ ρ(0)
X (0) ∗ (1)
ρ(1) (r) = 2 Re ϕi (r) ϕi (r) .
i
(occupied)

The first of this equations is Sternheimer equation, i.e. the Schrödinger equation to the
first order. The “perturbation” λwα could be constrained to just one atom, or it could
be a phonon with given wavevector q and polarization A,

wα = AeiqRα + A∗ e−iqRα .

The system of equations (5.22) has to be solved self-consistently together with (5.20).
(1)
ρ(1) and VSCF can be directly used in Eq. (5.19), and the second term on the right side
of Eq. (5.19) can be obtained from analytical differentiation of the second equation in
(5.22), similarly to how the forces were obtained from zero-order Kohn–Sham results.

84
 6. Order-N methods

6.1 Basic considerations

The computation time and necessary resources (memory) in ab initio calculation increase
rapidly with the size of problem N (number of atoms or electrons, basis size). The scaling
with N is governed by several factors and is generally ∼ N 3 in both DFT and HF-based
realizations. Depending on technical realization of the algorithm, the scaling may be
worse than that. In the following, we’ll discuss physical origins of the scaling with N, and
the ways to improve it.
In traditional implementation of the DFT, like most of those discussed in Chapter 3,
one has to solve Kohn–Sham equations (3.16), where one-electron functions are expanded
over a fixed basis set of size, say, Q, Eq. (4.2). This results in a generalized eigenvalue
problem, Eq. (4.3). So at some point we have to perform diagonalization that is a ∼ Q3
procedure, at least for algorithms so far known. Typically, the basis size Q is much
larger than the number of electrons N.37 It is not necessary to diagonalize the secular
equation matrix in full – it suffices to know the N lowest eigenvalues and corresponding
eigenvectors, necessary to reconstruct the density along Eq. (2.19). For this, iterative
diagonalization techniques are known that scale, in the best case, as ∼ N 2 Q. Even if this
may be much better tan ∼ Q3 , the scaling with the number of atoms is still at least ∼ N 3 .
It is important to realize that the limitation ∼ N 3 is not purely technical. Even
if more efficient diagonalization scheme appear, there is an essential intrinsic constant
that actually sets this ∼ N 3 limit. namely, this is the condition that the solutions,
i.e., eigenvectors Cαp in the expansion (4.2), must be orthogonal for all α≤N. As the
system grows, each wavefunction is represented by a longer eigenvector, ∼ N. There are
more occupied states, that brings in one more ∼ N factor. Each eigenvector must be
orthogonalized to all others, that results in the total ∼ N 3 scaling, on the very best.
What are then the arguments that better scaling can exist at all? From physical point
of view, there seem to be some reasons to argue that the scaling as slow as linear is pos-
sible. Indeed, it is known – at least from chemistry – that electronic properties are to a
large extent governed by near neighborhood. There are many justifications for different
types of systems – indeed, cluster methods are used to simulate crystalline or amorphous
solids in view of solid experience that details of electronic structure, local charge density,
magnetic moments etc. converge reasonably fast to perfect bulk values as the cluster size
grows. A somehow more sensitive property is total energy per atom; the following figure
from a review paper by Goedecker38 shows how rapidly it converges to its asymptotic
(bulk) value with the cluster size.
37
As we discussed earlier, that depends very much on the basis set used. In compact tight-binding
bases, there are usually just several (few) basis functions per electron; in plane-wave schemes there can
be hundreds.
38
Stefan Goedecker, Linear scaling electronic structure methods, Rev. Mod. Phys. 71, 1085–1123
(1999); http://arXiv.org/abs/cond-mat/9806073.

85
 1
Fig. 1 of Goedecker: The de-

energy deviation
viation of the total energy per 0.1
silicon atom from its asymp-
totic bulk value as a function 0.01
of the size of the periodic vol-
ume in which it is embedded. 0.001
The calculation was done with
a tight-binding scheme using 0.0001
exact diagonalization. 0 100 200 300 400 500 600
number of atoms

Another manifestation of the dominance of short-range interactions is the behaviour

of the density matrix depending on the separation between particles. The density matrix
was introduced in Eq. (2.20). In the Hartree-Fock approximation, it is expressed via
one-electron orbitals as in Eq. (2.22):
X
γ(r, r0 ) = ϕ∗α (r) ϕα(r0 ) .
α

The summation is straightforward for fully polarized uniform electron gas:

ZkF ZkF
X V
→ dN → k 2 dk sin ϑk dϑk dφk
α 8π 3
0 0

(for non-spin-polarized case, see Sec. 2.4). Hence

ZkF Zπ
0 V 2 0
γ(r, r ) = 2π k dk sin ϑ dϑ eik|r−r | cos ϑ
8π 3
0 0
ZkF 0 0 ZkF
V eik|r−r | − e−ik|r−r | V
= kdk = 2 k dk sin(k|r − r0 |)
2π 2 0
2i|r − r | 2π |r − r0 |
0 0
" #
0 0
V sin(kF |r − r|) kF cos(kF|r − r |)
= −
2π 2 |r − r0 |3 |r − r0 |2
V kF2
= 2 0
j1 (kF |r − r0 |) (6.1)
2π |r − r |

with kF3 /(3π 2 ) = ρ, see (2.35).

This HF asymptotics is essentially preserved in “true” metallic systems, as inter-
electron correlations influence the specific short-range behaviour but not long-range decay.
One can show that the decay of γ(r, r0 ) with increasing |r − r0 | becomes even faster (i.e.,
exponential) in two cases, with the presence of an insulating band gap and when the Fermi
distribution function is smeared out (for instance, due to the introduction of electrionic
temperature). The latter trick is sometimes used in order to control, or force, the decay
of the density matrix. This option will be addressed later on.

86
 Walter Kohn39 discussed these issues of the dominance of short-range interactions and
the decay of the density matrix with distance on a somehow more systematic basis, and
he introduced a concept of “nearsightedness” of equilibrium system consisting of many
quantum mechanical particles moving in an external potential without long-range (e.g.,
electric) fields, in the following way:
Let F (r1, . . . , rν ) be a static property (for example: charge density, pair correlation
function etc.) depending on coordinates of ν particles, all within a restricted volume of
linear dimension ∼ λ, a typical de Broglie wavelength in the ground state wavefunction.
Then, at fixed chemical potential µ, a change of the external potential ∆v(r0 ), no matter
how large, has a small effect on F , provided only that ∆v(r0 ) is limited to a distant region,
i.e., for all r0 , |r − r0 |  λ. Thus, F does not “see” ∆v(r0 ).
Kohn gives the following explanations to this principle.
1. The principle is a manifestation of wave-mechanical destructive interference.
2. It is essential to have many particles (not necessarily interacting).
3. There are exceptions: noninteracting bosons below critical point, systems with trans-
lationally invariant long-range order, like Wigner crystal in a torus.
4. If ∆v includes long-range electric fields (unscreened Coulomb, as in ionic crystals),
this must be added self-consistently to the effective potential. In other words, a Coulomb
field from, say, an extra charge certainly will have effect on the many-particle system in
question. But this effect will be, primarily, a rigid shift of potential over the whole region
of dimension λ, that would shift the chemical potential there. However in the formulation
above it was assumed that the chemical potential must be fixed. The principle says that,
once the effect of bare Coulomb field is considered by a correspondingly shifted chemical
potential, as determined in a self-consistency procedure, the variations ∆v on top of bare
Coulomb field won’t affect the property F .
W. Kohn argued further on that the possibility of mλ
O(N) methods immediately follows from the principle
of “nearsightedness”. For a system enclosed in a large
volume Ω, one can introduce a system of overlapping
smaller volumes, ω ∼ (mλ)3 “where m is, say 100” –
writes W. Kohn. One can consider each subvolume in-
dependently, because of nearsightedness, and enclose it
in a hard wall boundary. The computational effort is
proportional to the number of subvolumes ∼ N, albeit
probably with a quite high prefactor.
A quite straightforward implementation of this program would be applying a cluster
method with open boundary conditions, centered on one atom after another (or selecting
small subvolumes. One can imagine calculation of, say, electronic structure of DNA base
by base (shown schematically as black boxes in the next figure), including in each step
only certain number of neighboring structure elements in such “cluster” (shaded boxes)

39
W. Kohn, Density Functional and Density Matrix Method Scaling Linearly with the Number of Atoms,
Phys. Rev. Lett. 76, 3168–3171 (1996).

87
and discarding more distant ones (white boxes):

cluster 1
     
  

     
  

  
  



     
   

     
  

  
  



     
   

     
  

  
  



     
   

     
  

  
  



     
   

     
  

  
  



     
   

     
  

  
  

     
   

  
  

  
  

  
  

  
  
  
  

  
  
  
 

  
  
  
  

  
  
  
 

  
  
  
  

  
  
  
 

  
  
  
  

  
  
  
 

  
  
  
  

  
  
  
 

  
  
  
  

  
 

  
  

cluster 2
The disadvantage is clearly that the
computational effort, although lin-
early scalable with number of “clus-
ters” considered, has so large pref-
actor that it hardly is affordable. ~N, large prefactor
Generally, when comparing O(N 3 )
with O(N) methods one is always
a “crossover” of computational load 3
~N
(CPU time, memory requirements).
The point in developing more sophis-
ticated order-N methods is to shift
the crossover point towards, say, 100 ~N, small prefactor
atoms or less, that would only make
100 200 300 atoms
these methods practically interesting.
Most of workable schemes now in use deal with density matrix in one or another
representation. The “most general” definition (2.20) won’t be used in the following,
because many-body wavefunction is not specified in density functional theory. However,
we’ll rely on the following expressions for kinetic and potential energies of electron systems,
both being expectation values of a single-particle operators:
h̄2
Z
Ekin = − ∇2r γ(r, r0) dr0 , (6.2)
Z 2m
r=r0

Epot = γ(r, r) veff(r) dr , (6.3)

where γ(r, r) = ρ(r) according to Eq. (2.21). The above formulae follow from Eq. (3.4)
under consideration that kinetic energy is a single-particle operator; the potential en-
ergy of Coulomb interaction in many-body system is two-particle operator, but effective
interaction in the Kohn-Sham formalism is a single-particle operator all right,
Z
ρ(r0 ) dr0
veff (r) = + uext (r) + VXC (r) . (6.4)
|r − r0 |
As long as we build on the Kohn-Sham approach with no two-particle operators explic-
itly appearing, we can quite consistently construct γ(r, r0) as in the HF formalism, see

88
Eq. (2.22). The obvious generalization in terms of fixed basis functions will be [compare
(4.5)]:  
N (occ.)
X X
γ(r, r0 ) =  ∗
Cαq Cαp  χ∗q (r0 ) χp (r), . (6.5)
pq α=1

The main reason behind these manipulations is to reformulate the Kohn-Sham approach
in such way as to avoid matrix diagonalization and/or the need to orthogonalize eigen-
vectors, that are essentially ∼ N 3 algorithms. Actually, in many cases we do not care so
much about individual eigenvectors (band structure, that becomes messy in a large system
anyway) as about total energy and density matrix. The first allows us to perform dynam-
ical calculations, and the second – to calculate expectation values of all single-particle
operators, according to Eq. (3.3):
hF i = Tr(F̂ γ) .
The Kohn-Sham total energy of Eq. (3.17)
N
e2 ρ(r)ρ(r0 )
X Z Z
Etot = εα − 0
dr dr0 − 0
VXC (r) ρ(r) dr + EXC [ρ]
α=1 2 |r − r |
can be casted into
Etot = EBS − EDC . (6.6)
| {z } | {z }
P (double-
≡ α εα counted
(band- terms)
structure
energy)

P
EBS = Ekin + Epot in terms of (6.2) and (6.3). Indeed, with γ(r, r0) = α ϕ∗α (r0 ) ϕα (r)
(6.2) and (6.3) become
h̄2 N
Z X
Ekin + Epot = − ∇2r0 ϕ∗α (r0 ) ϕα (r) dr0 +
2m α=1 r=r0
N Z
X
+ ϕ∗α (r0 ) veff (r) ϕα(r) dr0 =
α=1 r=r0
N Z
" 2
#
X h̄
= dr ϕ∗α (r) − ∇2 + veff (r) ϕα (r) =
α=1 2m
N
X
= [compare to Eq. (3.16)] εα .
α=1

The way it is written,

EBS = Tr (γ̂ Ĥ) .
This is obvious in the coordinate representation, or, one can transform to the preselected
basis of χq :
Q " Q Q#
N Z
X X
∗ ∗ h̄2 2 X X
EBS = dr Cαq χq (r) − ∇ + veff (r) Cαp χp (r) = Dpq Hqp ,
α=1 p=1 2m p=1 p=1
q=1

89
where Hqp incorporates underlined terms, and Dpq – those with the bars above.
Double-counting terms can be similarly casted in a form expressed via a density matrix.
Let’s add to the potential energy a yet undefined potential U(r) and find out what it
must be, in order to arrive at the correct expression for the total energy, with the double-
counting terms subtracted.
" #
N Z
X h̄2 2
Etot = dr ϕ∗α (r) − ∇ + U(r) ϕα (r) =
α=1 2m
" #
XN Z
h̄2 2
= dr ϕ∗α − ∇ + veff (r) ϕα (r) +
α=1 2m
XN Z
+ dr ϕ∗α [U(r) − veff (r)] ϕα (r)
α=1
N
X Z
= εα − dr [veff (r) − U(r)] .
α=1

Comparing to Eq. (3.17) yields

e2 ρ(r0 )
Z
δEXC [ρ]
veff (r) − U(r) = 0
dr0 + VXC (r) − ,
2 |r − r | δρ
| {z }
VXC

hence
e2 ρ(r0 )
Z
U(r) = vXC (r) − dr , and
2 |r − r0 |

h̄2 Z 2
Etot = − ∇r γ(r, r0 )|r=r0
2m " Z #
Z
e2 ρ(r0 )
+ γ(r, r) dr0 + uext (r) + VXC (r) dr ,
2 |r − r0 |

without reference to band energies. It is important for the following that this expression
doe not contain anymore the results of diagonalization, εα nor Cαq , if we address the
density matrix Dpq later on directly, without reference to its representation as a sum over
eigenvectors.
Since we want to abandon direct reference to εα whenever possible, we face a problem
how to populate the states numbered by α. Conventionally we kept trace on those α ≤ N
corresponding to the lowest εα . In other formulation, the necessary sums may run over
all states, weighted by the step function (=1 for α≤N, =0 for α > N). Now a trick will
be that we introduce instead of this Fermi distribution function,
1
F (ε) = ε−µ , (6.7)
e kT +1
corresponding to a certain chemical potential µ and, probably to a certain electronic
temperature T . It was mentioned earlier that this may be also advantageous to force

90
better spatial decay of the density matrix. (Note however that this temperature has
nothing to do with “lattice” temperature introduced in dynamical simulations). With the
Fermi function thus introduced, the density matrix and other properties (energy etc.) can
be expressed via sums over all basis functions:
Q
X
0
γ(r, r ) = F (εα ) ϕ∗α (r) ϕα (r0 ) ,
α=1
XQ
∗
Dpq = F (εα ) Cαq Cαp , (6.8)
α=1
Q
X Q
X
EBS = Dpq Hqp = F (εα ) εα .
p,q=1 α=1

After this general introduction, we review several methods that allow in principle
order-N scaling in some detail.

6.2 Divide-and-Conquer
We begin with a historically important order-N method, proposed by Yang and Zhao in
1991–1995, the “divide-and-conquer” (D&C) method. It performs better than merely su-
perposition of free clusters, because D&C allows interaction between subsystems. There-
fore the “clusters” can be kept reasonably small.
The system is split into several subsystems, as already discussed earlier, and each
subsystem is surrounded by its buffer, like e.g. for the linear chain:

buffer subsys. A buffer p

buffer subsys. B buffer

For subsystem A, the eigenvalue problem is solved (e.g., by diagonalization) in terms of

H and S matrices between pairs of functions within the subsystem and its buffer:
Xh i
A
Hqp − εA A A
α Sqp Cαp = 0 . (6.9)
p

The size of these matrices is (N A×N A ), counting the basis functions within the system +
buffer, therefore the related diagonalization amounts to (hopefully moderate) ∼ (N A )3 . In
order to compute the total energy, we need density matrix for the whole system, computed
A
from individual Dpq using partition matrix dApq such that
X X
global
dA
pq = 1 , Dpq = dA A
pq Dpq . (6.10)
A A

91
The partition scheme is subject to an (arbitrary, up to the sum rule above) choice, e.g.,


 0 if p ∈ buffer and q ∈ buffer,
1
dA
pq = 2
if p ∈ A and q ∈ buffer, or vice versa


1 if p ∈ A and q ∈ A,

or more sophisticated. For the case of linear chain as shown above, the (two-dimensional)
A
density matrix would be reconstructed from the overlap of individual Dpq as follows:

q
ignored
p

overlap
ignored subsystems
+ buffers
subsystems

The global density matrix, by virtue of Eq. (6.10), is

X X Q
X
global ∗
dA
pq = 1 , Dpq = dA
pq F (εα ) C Aαq Cαp
A
. (6.11)
A A α=1

The value of the chemical potential µ appearing in F (ε) is shared by all subsystems; this
allows for the charge transfer between them. µ must be self-consistently determined from
the normalization condition,
global
QX
Dpq Sqp = N . (6.12)
p,q=1
|{z}
contains
F (ε)

that is identical to
N
XX Z N Z
X Q
X Q
X
∗
Cαq Cαp χ∗q (r) χ(r) dr = dr ∗
Cαq χ∗q (r) Cαp χp (r) = N .
pq α=1 α=1 q p
| {z } | {z }
ϕ∗α (r) ϕα (r)

The convergence of results with the buffer size may look like in the following figure,

92
reproduced from the review by Ordejón40 and presenting the results of Yang41 for a
tetrapeptide molecule (shown below), with different buffer sizes (varying from 0 to 5
neighbouring groups in the polypeptide chain) and different choice of bases: “single-zeta”
(SZ), “double-zeta” (DZ) and “double-zeta including polarization orbitals” (DZP). The
basis size increases from SZ (99) to DZ (181) to DZP (308), the number in brackets giv-
ing the total number of basis functions in the whole molecule. The accuracy typically
becomes acceptable after already just several shells of
neighbors (=buffer size) included. However, the 1

method is not variational, i.e., the error is not nec-

0
essarily reduced while using larger (or better) basis
DZ
set. The Hellmann-Feynman forces are neverthe- -1
less available.
DZP

log10(|∆E|)
O O -2
H2 H H2
H2 N C C N C C OH
C N C C N C
H2 H H2 H -3
O O

-4
Results from Table 1 of Yang (1991): absolute
SZ
value of errors in the total energies for a tetrapep- -5
tide molecule for the D&C approach, as com-
pared with the Kohn-Sham result for the whole
0 1 2 3 4 5 6
molecule with the corresponding choice of basis.
Shells of neighrours in buffer region

6.3 Fermi operator expansion

The starting point of this method is the Fermi distribution function (6.7), which, as a
function of ε, is approximated by a polynomial expansion. The temperature T enters as
an extrenal parameter; from physical considerations, the case T → 0 is of most interest,
but keeping the temperature small non-zero is advantageous for getting better polynomial
approximation, as well as for achieving better spatial localization of the density matrix.
The idea behind the method is to avoid the diagonalization and the selection of N lowest
states. Instead, one fixes the chemical potential µ in advance (and corrects it afterwards,
if the total number of electrons comes out wrongly), and the calculation of the density
matrix is done iteratively, from a “reasonable” starting guess. The Fermi distribution
function can be generalized to a “Fermi operator”
1
F (H) = H−µ ,
e kT + 1
which, for T = 0, reduces to merely H if acting on occupied states, and zero for the states
whose eigenvalues are beyond µ.
Casting the Kohn-Sham equation in the form
F (H)|ϕα i = F (εα)|ϕα i , (6.13)
40
Pablo Ordejón, Order-N tight-binding methods for electronic-structure and molecular dynamics, Com-
put. Mater. Sci. 12, 157–191 (1998).
41
Weitao Yang, Direct calculation of electron density in density-functional theory: Implementation for
benzene and a tetrapeptide, Phys. Rev. A 44, 7823 (1991).

93
we’ll now use instead of the the “true” Fermi operator F (ε) its polynomial approximation
P (ε):
P (H)|ϕα i = P (εα)|ϕα i . (6.14)
The polynomial approximation P (H) can be taken of any appropriate form; it is advan-
tageous to use Tchebyshev polynomials Tj for better numerical stability:
T0
1

T0 1 =
T1
T1 x =
0 T2 2x3 − 1
=
T3 4x3 − 3x
=
T2 ···
Tj+1 (x) = 2x Tj (x) − Tj−1 (x) .
−1
−1 0 1

Then the polynomial approximation P (H) reads:

npoly
C0 I X
P (H) = + Cj Tj (H) . (6.15)
2 j=1

This allows, instead of direct diagonalization, to construct ”Tchebyshev matrices” (which

will sum up to give the density matrix) iteratively, as follows:
|t0l i = |el i
|t1l i = H|el i
···
|tl i = 2H|tjl i − |tj−1
j+1
l i.
Here |tjl i is the l’th column of the j’th Tchebyshev matrix, |el i is a starting guess, typically
a l’th column of the unit matrix. On constructing all tjl ’s up to j = npoly by matrix
multiplication, one can recover the l’th column of the density matrix:
npoly
C0 0 X
|Dl i = |tl i + Cj |tjl i . (6.16)
2 j=1

Note that the chemical potential µ was fixed at the beginning. If this would result in
a wrong number of electrons (obtained from the trace of the density matrix), µ can be
corrected by
∆Nel
∆µ = ,
Tr [P 0 (H)]
where P 0 is the derivative of the Tchebyshev polynomial P chosen to represent the Fermi
distribution, by the virtue of
Nel = Tr [F ] → Tr [P ] ,
" #
dNel dP
→ Tr .
dµ dµ

94
Without further assumptions, the method is of the order N 2 . Indeed, for Mb basis func-
tions, the density matrix is Mb ×Mb , i.e. one has to calculate Mb full columns. For the
calculation of each column, one performs npoly matrix × vector multiplications, each costs
Mb nH , assuming H is a sparse matrix with nH non-zero elements per row (or column).
The total effort is then
∼ Mb2 npoly nH .
| {z }
independent
on the
system size

The linear scaling is recovered if one introduces a localization region for each column
of the density matrix, outside of which the elements are negligibly small. Denoting the
number of such non-vanishing elements in each column of the density matrix by Mloc , we
arrive at the scaling ∼ Mb Mloc npoly nH , that is linear.

6.4 Variational methods

We consider now two different approaches which both avoid the diagonalization and di-
rectly optimize total energy in terms of either elements of the density matrix, or individual
orbitals. The total energy, for our further purposes, will be casted as
 
Etot = Tr γ̂ Ĥ ,
N
X
or Etot = hϕα |H|ϕα i .
α=1

In order to make te function to be minimized insensitive to the rigid shift of the potential,
we substract µN from Etot and thus define the grand potential Ω,
h i
ˆ ,
Ω = Tr (γ̂H) − µN = Tr γ̂(H − µ I) (6.17)
N
X
or Ω = ˆ αi .
hϕα |H − µ I|ϕ (6.18)
α=1

Under constant shift, the potential energy increases by ∆ U·N; in order to conserve the
number of electrons, µ must also shift upwards and thus Ω remains constant. One can
achieve minimization of Ω only by applying non-uniform potential, that would lead to a
redistribution of charge density, reducing the total energy in a ”non-trivial” way. In the
following, we shall refer to
H0 ≡ H − µ Iˆ
as ”shifted Hamiltonian”. The minimization of either (6.17) in terms of density matrix,
or of (6.18) in terms of orbitals is subject to constraints that are automatically satisfied in
the course of a conventional diagonalization procedure, but we have to build in explicitly
if we envisage to abandon the diagonalization. These constraints are, on the orbitals, that
they must be orthonormal,
hϕα |ϕβ i = δαβ ,

95
 42
and on the density matrix – that it has the property of idempotency, i.e.

γ̂ 2 = γ̂ . (6.19)

Moreover, it must conserve the number of electrons:

N = Tr(γ̂) . (6.20)

In a conventional energy-minimization scheme one proceeds iteratively; in each step, the

gradient of Ω with respect to either γ, or {ϕα }, is calculated, the variables are moved
in the direction of gradients, and the constraints imposed. A disadvantage is that im-
posing directly orthonormality or idempotency constraints is an over-linear operation.
An alternative would be, instead of imposing constraints explicitly in each iteration, to
gradually influence the variables in such a way that the constraints would become ever
better satisfied in the course of iterations. Several functionals which do not require ex-
plicit imposition of constraints have been developed. Li, Nunes and Vanderbilt 43 used a
purification transformation of the density matrix via the McWeeny function:
1 γ̃ = 3γ 2 − 2γ 3 . (6.21)
If eigenvalues of γ are close to either 0, or 1, the
eigenvalues of γ̃ become even close to the same
values. The Li–Nunes–Vanderbilt (LNV) grand
potential is then, instead of Tr[γ̂(H − µI],ˆ
h  i
0
0 1 ΩLNV = Tr 3γ̂ 2 − 2γ̂ 3 H − µIˆ , (6.22)
and no constraints imposed during the minimization. On the convergency, 3γ 2 − 2γ 3 = γ
by the virtue of γ 2 = γ, hence ΩLNV = Ωtrue .
The gradient of ΩLNV can be found using the following differentiation rules for the
trace of a matrix:
X ∂M
for M = Aij Bji, = Bβα ;
ij ∂Aαβ
X ∂M X
for M = AkiAij Bjk , = (Bβk Akα + Aβk Bkα ) ;
ijk
∂Aαβ k
X ∂M X
for M = Aij Ajk Akl Bli , = (Aβk Akl Blα + Aβk Bki Aiα + Bβk Aki Aiα ) ,
ijkl ∂Aαβ k

hence
∂ΩLNV  
= 3 (γ̂H0 + H0 γ̂) − 2 γ̂ 2 H0 + γ̂H0 γ̂ + H0 γ̂ 2 , (6.23)
∂γ̂
42
for instance, in the coordinate representation it follows immediately from the definition of the
density matrix, Eq.(2.20),
Z
γ(r, r0 ) = γ(r, r00 ) γ(r00 , r) dr00 .

43
X.-P. Li, R. W. Nunes, and David Vanderbilt, Density-matrix electronic-structure method with linear
system-size scaling, Phys. Rev. B 47, 10891–10894 (1993).

96
 ˆ Since Hϕ = εϕ and γ̂ϕ = f (ε)ϕ,
with H0 ≡H − µI.

Hγ̂ϕ = εf (ε)ϕ = γ̂Hϕ and hence [γ̂, H0 ] = 0,

then
∂ΩLNV
= 6γH0 − 6γ 2 H0 = 0
∂γ̂
for true γ, i.e., γ = γ 2 .
Since the LNV functional is a cubic polynomial in all its degrees of freedom, and cubic
polynomial may have only one local minimum, there is exactly one local minimum in a
multidimensional minimization, which can be easily found. That means that the LNV
functional is a “well behaved” one.
A complication may arise that, unless µ is set to a “correct” band gap from the
beginning, the minimization may end up with a wrong number of electrons. To overcome
this, one must simply foresee on option to adjust µ in the course of minimization (see the
discussion to this point at the end of Section 6.3.
The forces on atoms are available as
dΩ ∂Ω ∂γ ∂Ω ∂H
= + .
dRα ∂γ ∂Rα ∂H ∂Rα
∂Ω
∂γ
vanishes at the solution, and the previous formula simplifies to
" #
dΩ ∂Ω ∂H   ∂H
= = Tr 3γ 2 − 2γ 3 . (6.24)
dRα ∂H ∂Rα ∂Rα
Now we come back to a possibility to minimize the functional Ω in a different way –
not via the optimization of density as in Eq. (6.17) but via optimization of wavefunctions
as in Eq. (6.18). Remember that we want to allow an unconstrained variation of |ϕα i,
in order to avoid the orthonormalization which is an order-N 3 procedure. Therefore |ϕα i
will remain not necessarily orthogonal underway and will only become orthonormalized
in the course of optimization. So there will exist an overlap

Sαβ = hϕα |ϕβ i , (6.25)

and the “band structure” energy takes the form

N
X
EBS = (S −1 )αβ hϕβ |H|ϕα i . (6.26)
α,β=1

Indeed, as has been introduced earlier,

N
X N
X
EBS = εα = hϕα |H|ϕα i
α=1 α=1

for an orthonormal basis. For an arbitrary basis, it holds anyway

H|ϕα i = εα |ϕα i ⇒ hϕβ |H|ϕβ i = εα hϕβ |ϕα i

| {z }
Sβα

97
hence
X  N 
X 
−1
εα = S hϕβ |H|ϕα i and EBS = S −1 hϕβ |H|ϕαi .
αβ αβ
β α,β=1

For the density matrix defined as

N
X  
γ̂ = |ϕα i S −1 hϕβ | ,
αβ
α,β=1

X   X
hϕi|γ̂|ϕj i = hϕi |ϕα i S −1 hϕβ |ϕj i = δiβ Sβj = Sij , as it should be.
| {z } αβ | {z }
ij β
Siα Sβj

With (6.26), the grand potential Ω (6.18) reads

N 
X 
Ω= S −1 (Hβα − µSβα ) (6.27)
αβ | {z }
α,β=1
0
≡ Hβα , shifted
Hamiltonian.

The problem with minimizing this functional is that, no matter how localized the orbitals
are, the inversion of S is not an order-N operation and results in a not sparse matrix. In
order to get rid of (S −1 ), Mauri et al.44 proposed to use a Taylor expansion,
  ∞ 
X n
S −1 = Iˆ − S . (6.28)
n=0

Indeed, expanding 1/x about x = 1 yields

1 2 6
= 1 − (x − 1) + (x − 1)2 − (x − 1)3 + . . . = 1 + (1 − x) + (1 − x)2 + . . .
x 2! 3!
Truncating this series at certain n leaves us with the n-th order approximation to the true
overlap−1 ,
m 
X 
Q(n) = Iˆ − S , (6.29)
n=0

and
N
X (m)
Ω⇒ Qαβ (Hβα − µSβα ) . (6.30)
α,β=1

Now the functional Ω can be obtained by merely matrix multiplication. In the Taylor
expansion (6.29), it suffices to take the lowest non-trivial term (m = 1), if we would show
that this results in the correct minimum value of the functional. Let’s try it.

Q(1) = 2Iˆ − S ;
44
Francesco Mauri, Giulia Galli, and Roberto Car, Orbital formulation for electronic-structure calcu-
lations with linear system-size scaling, Phys. Rev. B 47, 9973–9976 (1993).

98
the variation of |ϕα i will be, in practice, a variation of their expansion coefficient in a
fixed basis system, like in Eq. (4.2).

N
X
0
Ω = (2δαβ − Sαβ ) Hβα =
α,β

X Q
N X N
X Q
X X
∗ 0 ∗ 0 ∗
= 2 Cαq Hpq Cαq − Cαp Hpq Cβq Cαr Cβr . (6.31)
α=1 p,q=1 α,β=1 p,q=1 r
| {z }
=Sαβ

P ∗
In the orthonormal basis ( r Cαr Cβr ≡Sαβ = δαβ ) (6.31) reduces to

X Q
N X N
X
∗ 0
Ω= Cαp Hpq Cαq = εα − µN ,
α=1 p,q=1 α=1

as it should be. In the general (i.e. not necessarily orthonormal) case, the differentiation
of (6.31) yields:

Q N X Q Q N X Q
∂Ω X
∗ 0
X
∗ 0
X
∗
X
∗ 0 ∗
=2 Cγp Hps − Cαp Hps Cαq Cγq − Cαp Hpq Cγq Cαs . (6.32)
∂Cγs p=1 α=1 p=1 q=1 α=1 p,q=1

Using
X X
∗ ∗
Hpq Cγq = εγ Cγp , Cγp Hpq = εγ Cγq
q p

and hence
X X
0 ∗ 0 ∗
Hpq Cγq = (εγ − µ) Cγp , Cγp Hpq = (εγ − µ) Cγq ,
q p

eq. (6.32) reduces to:

N Q N X Q
∂Ω ∗
X
∗
X
∗
X
∗ ∗
= 2(εγ − µ)Cγs − (εα − µ) Cαs Cαq Cγq − Cαp (εγ − µ) Cγp Cαs
∂Cγs α=1 q=1 α=1 p=1
N
X Q
X
∗ ∗
= 2(εγ − µ)Cγs − Cαs Cp∗ Cγp (εα + εγ − 2µ)
α=1 p=1
| {z }
= δαγ
for exact solution,
when orthonormality
is restored

∗ ∗
= 2(εγ − µ)Cγs − Cγs (2εγ − 2µ) = 0 .

We have shown that the derivative ∂Ω/∂Cγs of the form (6.32), which is valid for any Cαp
in the course of minimization (i.e. also away from the orthonormality) really vanishes if
the orthonormality is enforced. In other words, the extremum ∂Ω/∂Cγs = 0 corresponds
to the situation when eigenvectors are orthonormal.

99
 ∂2
Differentiating (6.32), the second derivative 2
∂Cγs
is:

N X Q N X Q
∂2 X
∗ 0 ∗
X
∗ 0 ∗
2
= − C H C
αp ps αs − Cαp Hps Cαs
∂Cγs α=1 p=1 α=1 p=1
N
X Q
X
∗ 0 ∗
= −2 Cαp Hps Cαs .
α=1 p=1
| {z }
= (εα − µ)Cαs∗

at the exact solution

∂ Ω 2
εα −µ≤0 for the lowest states, hence ∂C 2 > 0 at the exact solution, corresponding to the
γs
ground state (i.e., when N lowest Kohn-Sham levels are occupied). Therefore, the above
formalism presents a variational approach.
Once we reduce our computational task to unconstrained minimization of the func-
tional Ω (6.17/6.18) in one or another representation (in terms of either density matrix
elements, or expansion coefficient over the orbital basis), the matrix multiplication re-
mains the main time-consuming step, so it is advantageous to keep it fast, using sparse
matrices. For maintaining the representation of either the density matrix, or of the or-
bitals |ϕα i in a sparse form, it is essential to use well localized basis functions. Possible
choices are:
– Gaussian-type orbitals (sparseness enforced by discarding overlaps smaller than a cer-
tain cutoff value);
– generalized Wannier functions (centered somewhere at the chemical bonds; falldown
exponential for insulators and power-low for metals);
– ad hoc strictly localized functions, as e.g. Sankey-Niklewski functions45
It is essential for the control over the accuracy that the method to use is variational. In
this case the localization constraint slightly increases the total energy, but keeping trace
on the latter helps to choose an acceptable degree of localization for the basis functions.
More details can be found in the review by Ordejón cited above.

45
Otto F. Sankey and David J. Niklewski, Ab initio multicenter tight-binding model for molecular-
dynamics simulations and other applications in covalent systems, Phys. Rev. B 40, 3979–3995 (1989).

100
 7. Lattice dynamics

7.1 Basic consideratons

In this section we’ll follow a number of well-known textbooks, and only try to keep the
notations consistent. Suppose we have a system of N atoms, m=1, . . . , N, with their
positions defined by some (arbitrary chosen) set of coordinates qi , i=1, . . . , 3N. In classical
treatment, one has to solve the system of Lagrangian equations of motion:
d ∂T ∂V
+ = 0, j = 1, . . . , 3N (7.1)
dt ∂ q̇i ∂qi
Let us first assume the coordinates qj to be Cartesian,
(
i = 1, 3
Xmi , ,
m = 1, N

and consider a more general case later on. Then

N 3
1 X X 2
T = Mm Ẋmi , (7.2)
2 m=1 i=1
3
N X
! N 3
X ∂V 1 X X ∂2V
V = V0 + Xmi + Xmi Xnj r; . (7.3)
m=1 i=1 ∂Xmi 2 m = 1 i = 1 ∂Xmi ∂Xnj
| {z }
n=1 j=1
i=0
in the
equilibrium

the neglection of higher expansion terms for V corresponds to the harmonic approximation
that we introduce here. With
∂T
= Mm Ẋmi ,
∂ Ẋmi
N X 3
!
∂V X ∂2V
= Xnj ,
∂Xmi n=1 j=1 ∂Xmi ∂Xnj

the equation of motion

3
N X
!
X ∂2V
Mm Ẍmi + Xnj = 0 (7.4)
n=1 j=1 ∂Xmi ∂Xnj

is a system of 3N coupled equations, labeled by {m, i}. For a finite cluster or a molecule,
one may try to simplify the problem by choosing appropriate symmetry coordinates,
reduce the matrix of force constants to block-diagonal form making use of the point

101
group symmetry, and project out 3 global translation and 3 global rotation modes – see
details below. For a perfect periodic crystal, a simplification is possible through applying
periodic boundary conditions. The “individual” atom displacements are then labeled by
the index of the unit cell in crystal, α, and the index of atom in each cell, l, e.g.:

Xmi → Xlαi ; Xnj → Xkβ j .

l and k run now from 1 to n, number of atoms per unit cell, and i, j remain the Cartesian
indici. Similarly we emphasize the dependency on cell index and atomic position in the
cell for the force constants:
∂2V
→ Φαli,βkj .
∂Xmi ∂Xnj

We search for solutions in lattice-periodic form,

Xliα = Ali (q) exp[−iωt + iqRα ] . (7.5)

Rα is the origin of the α’th unit cell, and the atomic displacements get the dependency
on q. There are exactly as many allowed q values as there are unit cells in crystal,
with periodic boundary conditions imposed. So, Eq. (7.5) still defines 3N individual
displacements. The trick is, due to imposing of periodic boundary conditions the equations
of motion become decoupled in q. Substituting the second derivative of (7.5) into Eq. (7.4)
yields
3
n X
XX
−Ml ω 2 Ali (q) exp[−iωt + iqRα ] + Φαli,βkj Akj (q) exp[−iωt + iqRβ ] = 0 .
β k=1 j=1

Multiplying by exp[iωt − iqRα ] yields:

 
X 3
n X X αβ
−Ml ω 2 Ali (q) +  Φ li, kj Akj (q) exp[iq(Rβ − Rα )] Akj (q) = 0 . (7.6)
k=1 j=1 β
| {z }
≡Dli, kj (q)

The summation in β must be done over all cells, and since real-space force constants
depend only on relative distances between the atoms, the origin α does not play a role
anymore (and could be set to 0). The system (7.6) becomes that of matrix equation,
decoupled in q, of the size 3n:
X 3 h
n X i
Ml ω 2δlk δij − Dli, kj (q) Akj (q) = 0 . (7.7)
k=1 j=1

The matrix to diagonalize is not symmetric, that may be unconvenient. One can redefine
eigenvectors (incorporating there square root of mass), in order to restore the symmetry:
3
n X
" # q 
X
2 Dli, kj (q)
ω δlk δij − √ Mk Akj (q) = 0 . (7.8)
k=1 j=1 Ml Mk

102
More precisely, the matrix is now Hermitian. Indeed,
X X
∗
Dli, kj (q) = Φαli,βkj e−iq(Rβ −Rα ) = Φαli,βkj eiq(Rα −Rβ ) =
β | {z } β | {z }
real ≡ Φβ α
kj, li
X
= Φβkj,αli eiq(Rα −Rβ ) = Dkj, li (q) .
α

It was used that since the lattice summation runs over relative vectors (Rα − Rβ ), instead
of summing over β for α = 0 we can as well fix β and run over α. The force constants
matzrix is related to the forces as follows:
N X 3
∂V XX
Fliα = − = − Φαli,βkj Xkj
β
. (7.9)
∂Xliα β k=1 j=1

Inversely,  
β
∂F α ∂Fkj
Φαli,βkj = − liβ = −  . (7.10)
∂Xkj ∂Xliα
This can be used for calculating force constants if forces due to displacements of all
cartesian displacements of all atoms are available. If we make use of the electronic-
structure code that makes use of periodic boundary conditions, we cannot displace just
one atom in an infinite crystal. But we can choose different supercells and try different
collective displacements. Once again,
β
X ∂Fkj
Dli, kj (q) = − exp[iq(Rβ − Rα )] ,
β ∂Xliα
if we displace the atom at the origin only (in the cell α) and sum over forces induced all
over the lattice. The relation between dynamical matrices in real and reciprocal spaces:
X
Dli, kj (q) = Φ0li,βkj eiqRβ , k
β
X l
Φ0li,βkj (Rβ ) = Dli, kj (q) e−qRβ . Rβ
q
In principle it would be nice to displace atoms one by one (running through Rβ and
k) in all three cartesian directions and look at the force on the atom of our choise (say,
l), and then make Fourier transformation to recover q-dependent force constants:
displace this one

measure the force

Unfortunately, this is not possible without breaking translation symmetry of the lattice.
What is more feasible is either creation of a phonon with a given q, or individual dis-
placements of atoms in a supercell.

103
 Let us first consider the first approach, that is the frozen phonon one. In this approach,
the atoms of type k are displaced in every cell, proportionally to exp(iqRβ ):

measure the force here... ... or here

This becomes quite straightforward if the calculation of forces is done in a supercell, and
the frozen phonon in question is commensurable with this supercell. Inversely, a choice of
supercell defines several commensurable vectors Q, for which the elements of dynamical
matrix can be sampled exactly. Consider an example for 2-dim. lattice:
reciprocal lattice cell
single cell
Q1 Q2
Γ

supercell
Q4 Q3
x2 Q3 Q4

Q1 Q2
x3 Brillouin zone

A reduced reciprocal cell maps onto four different vectors in the original Brillouin zone,
which correspond to the following displacement patterns (according to the eiQR wave).
Of course the displacement patterns need not be necessarily harmonic inside the su-
percell. Disregarding the symmetry, we can displace any atom individually and probe
different force constants. Of course this “any” refers to atom in a supercell, and in the
adjacent supercell we’ll have an identically displaced atom:

displace only this atom

in each supercell
get the force here

In the lattice summation we’ll now specify Rβ = L + ∆, where L runs over supercells in
β ∆
crystal, and ∆ numbers single cells within a supercell. Xkj is then Xkj for all L.

XX 3
n X 3 X
n X
XX
Fliα=0 = − Φ0li,βkj Xkj
β
=− β
Dli, kj (q)e−iqRβ Xkj =
β k=1 j=1 β k=1 j=1 q

104
 X 3
n X X X
∆
= − Xkj Dli, kj (q) exp[−iq∆] e−iqL =
k=1 j=1 q L
| {z }
=δqQL

X 3
n X X
∆
= − Xkj Dli, kj (QL ) exp[−iQL ∆] . (7.11)
k=1 j=1 QL

QL have been introduced here as vectors in the reciprocal space, which are “conjugated”
to the supercell vectors: L · QL = 2π×(integer). Note that Dli, kj (QL ) from different QL
are now mixed in the final force, but they can be recovered if we try all possible ∆, since
there are so many different QL (in the first Brillouin zone) as there are ∆’s.
Let us skip for brevity the (li) and (kj) indices and concentrate on the dependencies
on ∆ and QL :
X
F (∆) = − D(QL )X(∆)e−iQL ∆ ;
QL
 
F Xh i
= − e−iQL ∆ D(QL ) ;
X ∆ QL
 
Xh i
−iQL ∆ −1 F
D(QL ) = − e .
∆ X ∆

This is a matrix equation with the dimension equal to that of the supercell size (= number
of different ∆), that will be, say S. Since on top of it we’ll have mixing up of (kj) and
(li) components everywhere, the full problem of finding all elements of D amounts to the
matrix inversion of the size 3nS. The use of point symmetry may help to reduce this
number considerably.46

46
Of course matrix inversion as such is not a problem; expensive are total energy/force calculations for
large number of different displacements in the supercell. The symmetry analysis may hopefully help to
reduce the number of trial displacements below 3n + 1 (three Cartesian displacements for each atom in
the unit cell, plus equilibrium). For the forces, we assume that all 3nS components are available from
each displacement pattern.

105