Analytical Hartree-Fock gradients for periodic systems

K. Doll1 , V. R. Saunders1 and N. M. Harrison1,2

arXiv:cond-mat/0011285v1 [cond-mat.mtrl-sci] 16 Nov 2000

1 CLRC, Daresbury Laboratory, Daresbury, Warrington, WA4 4AD, UK

2 Department of Chemistry, Imperial College, London, SW7 2AY, UK

Abstract

We present the theory of analytical Hartree-Fock gradients for periodic sys-

tems as implemented in the code CRYSTAL. We demonstrate how derivatives

of the integrals can be computed with the McMurchie-Davidson algorithm.

Highly accurate gradients with respect to nuclear coordinates are obtained

for systems periodic in 0,1,2 or 3 dimensions.

Typeset using REVTEX

1
 I. INTRODUCTION

The determination of equilibrium structure is one of the most important targets in elec-

tronic structure calculations. Analytical gradients provide an important tool to facilitate

this and therefore the implementation of analytical gradients has become an important part
of modern codes. Although most solid state calculations are nowadays performed within the
framework of density functional theory, Hartree-Fock theory can serve as a useful starting

point for a correlation treatment. In the field of quantum chemistry, a Hartree-Fock so-
lution is necessary to make a wavefunction based correlation scheme such as, for example,
the coupled-cluster approach, applicable. Therefore, the determination of a Hartree-Fock
solution is often an important target.
The calculation of Hartree-Fock gradients was pioneered by Pulay who performed the

first implementation for multicenter basis sets1 . It should be mentioned, that the theory had
already been derived earlier independently2 . Analytical gradients have become an important
area in quantum chemistry and several review articles have been published3–6 .
Significant work has already been performed for one-dimensional systems: formulas for

analytic gradients, with respect to nuclear coordinates as well as with respect to the lattice
vector, have been derived and implemented in a periodic code7 ; and the theory has been
extended to metallic systems8 . Further progress has been the derivation and implementation
of formulas for MP2 energy9,10 and gradients11 , as well as gradients on the density functional

level12 . Even formulas for second derivatives have meanwhile been coded13 . Recently, a
scheme for an accurate treatment of long-range Coulombic effects in Hartree-Fock gradients
has been presented14 , and a new implementation of density functional energy and gradients
for periodic systems has been demonstated to be highly efficient and accurate15 .
In this article, we report on an implementation of Hartree-Fock gradients with respect

to nuclear coordinates in a general periodic code (the CRYSTAL16–19 package), which is to

the best of our knowledge the first implementation for the case of 2- and 3- dimensional
periodicity.

2
 The article is structured as follows: in section II, the basis functions and Hartree-Fock
equations are given. The calculation of integrals which relies on the McMurchie-Davidson
algorithm and the calculation of gradients of the integrals is explained in sections III and

IV. The total energy as calculated by the CRYSTAL code is given in section V. In section
VI, we explain the calculation of forces and possible sources of error. Finally, in section VII
we illustrate the accuracy of the gradients with some examples.

II. BASIS FUNCTION AND HARTREE-FOCK EQUATIONS

In this section, we summarize the basis functions used in the CRYSTAL code and give
the structure of the Hartree-Fock equations.

A. Basis functions

Unnormalized spherical Gaussian type functions (SGTF) in a polar coordinate system

~ are defined as
characterized by the set of variables (|~r|, ϑ, ϕ), and centered at A,

~ 2n+l P|m| (cos ϑ) exp(imϕ) exp(−α|~r − A|

~ n, l, m) = |~r − A|
S(α, ~r − A, ~ 2) (1)
l

|m|
with Pl being the associated Legendre function. In the context of the McMurchie-

Davidson algorithm, Hermite Gaussian type functions are necessary which are defined as:
t  u  v
∂ ∂ ∂

~ t, u, v) =
Λ(γ, ~r − A, ~ 2)
exp(−γ|~r − A| (2)
∂Ax ∂Ay ∂Az

Real spherical Gaussian type functions are defined as

~ n, l, 0) = S(α, ~r − A,
R(α, ~r − A, ~ n, l, 0)

~ n, l, |m|) = Re S(α, ~r − A,
R(α, ~r − A, ~ n, l, |m|)

~ n, l, −|m|) = Im S(α, ~r − A,
R(α, ~r − A, ~ n, l, |m|)

CRYSTAL uses real spherical Gaussian type functions, which are in the following denoted
~ µ ) = Nµ R(α, A
as φµ (~r − A ~ µ , n, l, m), with the normalization Nµ . µ is an index enumerating

3
the basis functions in the reference cell (e.g. the primitive unit cell). Although the code
allows only the use of SGTFs with n = 0, in the process of the evaluation of molecular
integrals, SGTFs with n 6= 0 are used20 .

B. Hartree-Fock equations

The Hartree-Fock treatment of periodic systems21 is briefly repeated in this section.

We assume that orbitals are doubly occupied and work within the restricted Hartree-Fock

formalism. The crystalline orbitals are linear combinations of Bloch functions

Ψi (~r, ~k) = aµi (~k)ψµ (~r, ~k)

X
(3)
µ

which are expanded in terms of real spherical Gaussian type functions with fixed con-
traction coefficients dj :

ψµ (~r, ~k) = Nµ ~ µ − ~g , n, l, m)ei~k~g

XX
dj R(αj , ~r − A (4)
~g j

The sum over ~g is over all direct lattice vectors. The Hartree-Fock-Roothaan equations
have a structure similar to the molecular case and have to be solved on a set of points in

the reciprocal lattice:

Fµν (~k)aνi (~k) = Sµν (~k)aνi (~k)ǫi (~k)

X X
(5)
ν ν

The Fock matrix F is given in detail in section V of this article, the overlap matrix S in

section III. The spin-free density matrix in reciprocal space is defined as

Pµν (~k) = 2 aµi (~k)a∗νi (~k)Θ(ǫF − ǫi (~k))

X
(6)
i

with the Fermi energy ǫF and the Heaviside function Θ; i is an index enumerating the

eigenvalues. The density matrix in real space Pµ~0ν~g is obtained by Fourier transformation.

4
 III. CALCULATION OF INTEGRALS

A. Types of integrals

In this section we summarize the appearing types of integrals. This is done with the
assumption that the basis functions, φµ , are real.

1. Overlap integral

The basic integral is the overlap integral:

Z
Sµg~1 ν g~2 = ~ µ − ~g1 )φν (~r − A
φµ (~r − A ~ ν − ~g2 )d3 r (7)

The integration is over the whole space, i.e. x from −∞ to +∞ and similarly for y and
z. Exploiting translational invariance we can rewrite this as
Z
Sµ~0ν~g = ~ µ )φν (~r − A
φµ (~r − A ~ ν − ~g )d3 r (8)

with ~g = g~2 − g~1 .

Further integrals appearing are:

2. Kinetic energy integrals

~ µ )(− 1 ∆~r )φν (~r − A

Z
Tµ~0ν~g = φµ (~r − A ~ ν − ~g )d3 r (9)
2

3. Nuclear attraction integrals

Z
~ µ )A(~r − A
~ a )φν (~r − A
~ ν − ~g )d3 r
X
Nµ~0ν~g = − Za φµ (~r − A (10)
a

where A is defined as the Coulomb potential in the molecular case, as the Euler-
MacLaurin potential for systems periodic in one dimension22,23 , as Parry’s potential24 for
systems periodic in two dimensions, and as the Ewald potential for systems periodic in three
dimensions25,26 . The a summation runs over all nuclei of the primitive unit cell.

5
 4. Electron-electron Coulomb interaction integrals

A bielectronic integral can be defined as

Z ~ µ )φν (~r − A
φµ (~r − A ~ ν − ~g )φτ (~r ′ − A ~ σ − ~n − ~h)
~ τ − ~n)φσ (~r ′ − A
Bµ~0ν~gτ~nσ~n+~h = d3 r d3 r′
|~r − ~r ′ |
(11)

In the context of periodic systems, it is necessary to perform summation over all lattice
vectors ~g , ~h, ~n. We define a Coulomb integral as follows
pen Z ~ µ )φν (~r − A
φµ (~r − A ~ ν − ~g )φτ (~r ′ − A ~ σ − ~n − ~h)
~ τ − ~n)φσ (~r ′ − A
d3 r d3 r′ =
X
Cµ~0ν~gτ~0σ~h = ′
~
n
|~r − ~r |
pen
X
Bµ~0ν~gτ~nσ~n+~h (12)
~
n

The penetration depth pen is defined as those terms for which

Z
~ µ )g(αmin , ~r − A
[g(αµmin , ~r − A ~ ν − ~g )]1/2 g(αmin , ~r − A
~ τ − ~n)d3 r > 10−IT OL2 (13)
ν τ

min
holds, with g(αamin , ~r) = ( 2απa )3/4 exp(−αamin |~r|2 ).
~ a . For these integrals,
g(αamin , ~r) means the lowest exponent of all Gaussians centered at A
the Coulomb interaction is evaluated without approximation. All the other integrals are
evaluated with a multipolar expansion. ”IT OL2” is a tolerance which can be chosen by the

user of the code. This criterion introduces an asymmetry in the energy expression: a given
bielectronic integral might be evaluated in different ways for Bµ~0ν~gτ~nσ~n+~h and Bτ~nσ~n+~hµ~0ν~g .
Avoiding this would however be very inefficient: To keep the symmetry, the lattice sum
would have to be further broken down into pieces which are evaluated exactly and other

pieces which are approximated — this would require a much higher computational effort
and more disk storage. The simpler criterion in equation 13 minimizes the effort, and, when
IT OL2 is chosen sufficiently large, the violation of the symmetry is negligible. We will
illustrate this with some examples in section VII.

6
 5. Electron-electron exchange interaction integrals

XZ ~ µ )φτ (~r − A
φµ (~r − A ~ τ − ~n)φν (~r ′ − A ~ σ − ~n − ~h)
~ ν − ~g )φσ (~r ′ − A
Xµ~0ν~gτ~0σ~h = ′
d3 rd3 r′ =
~
n
|~r − ~r |
X
Bµ~0τ~nν~gσ~n+~h (14)
~
n

For an individual exchange integral,

Bµ~0τ~nν~gσ~n+~h = Bτ~nµ~0σ~n+~hν~g (15)

should hold. However, for efficiency reasons, two different thresholds have been
introduced16 which leads to another possible asymmetry: an exchange integral is discarded
~ µ ) and φν (~r − A
when the pseudooverlap associated with φµ (~r − A ~ ν −~g ) or the pseudooverlap
~ σ − ~n − ~h) is smaller than certain thresh-
~ τ − ~n) and φσ (~r − A
associated with φτ (~r − A

olds, 10−IT OL4 and 10−IT OL5. It is recommended that the threshold IT OL5 associated with
~ σ − ~n − ~h) should be higher than IT OL4. This, however, will
~ τ − ~n) and φσ (~r − A
φτ (~r − A
lead to a violation of equation 15 and therefore another asymmetry in the energy expression.
A further cutoff parameter, IT OL3, selects the exchange integrals symmetrically: exchange
~ µ ) with φτ (~r − A
integrals are also neglected if the overlap of φµ (~r − A ~ τ − ~n) or the overlap

~ σ − ~n − ~h) is lower as 10−IT OL3. This is a symmetric cutoff,

~ ν − ~g ) with φσ (~r − A
of φν (~r − A
and therefore should not lead to an inaccuracy in the forces.

6. Multipolar integrals

The charge distribution is approximated with the help of a multipolar expansion up to

order L.
Z
~ c) =
ηlm (ρ; A ~ c )d3 r
ρ(~r)Xlm (~r − A (16)

with Xlm being regular solid harmonics26 and the charge ρ(~r) defined as

~ µ )φν (~r − A
~ ν − ~g )
X
ρ(~r) = − Pν~gµ~0 φµ (~r − A (17)
~g ,µ,ν

7
 7. Field integrals

The electrostatic potential is approximated with an expansion up to the maximum quan-

tum number L.
pen
~ˆc ) 1
Z  
m
= Zlm (A ~ µ )φν (~r − A
~ ν − ~g ) A(~r − A
~ c) − d3 r
X
Mlµ~0ν~ φµ (~r − A (18)
gc
~
n
~
|~r − Ac − ~n|

~ˆc ) being the spherical gradient operator, Z m (A

with Zlm (A ~ˆc ) = Nlm ~ˆc )
X m (A and a
l (2l−1)!! l

normalization Nlm (Ref. 26).

8. Spheropole

This term arises because the charge distribution is approximated by a model charge
distribution in the long range. However, the use of the Ewald potential instead of the
Coulomb potential requires a correction in the three-dimensional case26 .

2π
Z
(ρc (~r) − ρmodel (~r))|~r|2 d3 r
X X
Q= Qc = c (19)
c c 3V

with

~ µ )φν (~r − A
~ ν − ~g )
XX
ρc (~r) = − Pν~gµ~0 φµ (~r − A (20)
µ∈c ~
g ,ν

and
L X
l
ρmodel ~ c )δ m (A
ηlm (ρc ; A ~ c , ~r)
X
c (~r) = l (21)
l=0 m=−l

and

δlm (A ~ˆc )Λ(α, ~r − A

~c , ~r) = lim Z m (A ~ c , 0, 0, 0) (22)
lα→∞

~ c.
c is chosen as the set of basis functions sited at center A

8
 B. McMurchie-Davidson algorithm

In this section, we indicate how the integrals are evaluated. This is done with the

McMurchie-Davidson27 algorithm. In this formalism, the product of two Gaussian type

functions is expanded at an intermediate center in terms of Hermite Gaussian type functions.

~ ñ, ˜l, m̃) =

~ n, l, m)S(β, ~r − B,
S(α, ~r − A,

E(n, l, m, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t, u, v)

X
(23)
t,u,v

~ ~
with γ = α + β and P~ = αA+β
α+β
B
. ~ − A;
E also depends on α, β and the distance B ~

however, the dependence on these parameters is suppressed in the notation.

This makes a very efficient evaluation of integrals feasible. The starting point
αβ ~ ~ 2 ) can be derived from the Gaussian product
E(0, 0, 0, 0, 0, 0, 0, 0, 0) = exp(− α+β |B − A|
rule28,29 :

αβ ~
  ~ + βB
αA ~ 2
~ 2 ) exp(−β|~r − B|
exp(−α|~r − A| ~ 2 ) = exp − ~ 2 exp
|B − A| − (α + β) ~r −
α+β α+β
(24)

The coefficients E can be generated by recursion relations27,20 . They are zero for the
case t + u + v > 2n + 2ñ + l + ˜l and for all negative values of t, u or v.

IV. CALCULATION OF DERIVATIVES

A. Gradients within the McMurchie-Davidson algorithm

The evaluation of gradients of the integrals is closely related to the evaluation of

the integrals themselves. All the integrals can be expressed with the help of the E-
coefficients27,20,26,30 . In the following we show how derivatives of the integrals can be ex-

pressed in a similar way with Hermite Gaussian type functions. Starting from equation 23,
we obtain

9
 ∂  ~ n, l, m)S(β, ~r − B, ~ ñ, ˜l, m̃) =

S(α, ~r − A,
∂Ax
∂ X
E(n, l, m, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t, u, v) =
∂Ax t,u,v
X  ∂
E(n, l, m, ñ, ˜l, m̃, t, u, v) Λ(γ, ~r − P~ , t, u, v) +


t,u,v ∂Ax
α

E(n, l, m, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t + 1, u, v) =
α+β
X ∂ α
E(n, l, m, ñ, ˜l, m̃, t, u, v) + E(n, l, m, ñ, ˜l, m̃, t − 1, u, v) Λ(γ, ~r − P~ , t, u, v)) =


t,u,v ∂Ax α+β

GA ˜ r − P~ , t, u, v)
X
x (n, l, m, ñ, l, m̃, t, u, v)Λ(γ, ~ (25)
t,u,v

Therefore, the gradients can be obtained in a quite similar way as the integrals. Instead
of the E-coefficients, the coefficients GA A A B B
x (and, for the other derivatives, Gy , Gz , Gx , Gy

and GB
z ) have to be used, as defined in equation 25. We obtain the following relation from

equation 25:

GA ˜
x (n, l, m, ñ, l, m̃, t, u, v) =

∂
E(n, l, m, ñ, ˜l, m̃, t, u, v) +
∂Ax
α
E(n, l, m, ñ, ˜l, m̃, t − 1, u, v) (26)
α+β

We could thus derive the G-coefficients from the E-coefficients. However, a more conve-
nient way would be to have a recursion relation similar to the E-coefficients. Indeed, these
relations can be obtained in an analogous way19 . We give the relations here for the case

of complex spherical Gaussian type functions; a transformation to real spherical Gaussian

type functions is possible along the lines given in ref. 18. The starting point can be obtained
from equation 26 and the definition of E(0, 0, 0, 0, 0, 0, 0, 0, 0) given in section III B:

GA
x (0, 0, 0, 0, 0, 0, 0, 0, 0) =
∂
E(0, 0, 0, 0, 0, 0, 0, 0, 0) +
∂Ax
α
E(0, 0, 0, 0, 0, 0, −1, 0, 0) =
α+β
αβ
2 (Bx − Ax )E(0, 0, 0, 0, 0, 0, 0, 0, 0)
α+β

10
 and

GA
x (0, 0, 0, 0, 0, 0, 1, 0, 0) =
∂
E(0, 0, 0, 0, 0, 0, 1, 0, 0) +
∂Ax
α
E(0, 0, 0, 0, 0, 0, 0, 0, 0) =
α+β
α
E(0, 0, 0, 0, 0, 0, 0, 0, 0) (27)
α+β

All the other GA

x (0, 0, 0, 0, 0, 0, t, u, v) are zero. Similarly, we obtain

GA
y (0, 0, 0, 0, 0, 0, 0, 1, 0) =

GA
z (0, 0, 0, 0, 0, 0, 0, 0, 1) =

GA
x (0, 0, 0, 0, 0, 0, 1, 0, 0)

and

GA
y (0, 0, 0, 0, 0, 0, 0, 0, 0) =
αβ
2 (By − Ay )E(0, 0, 0, 0, 0, 0, 0, 0, 0)
α+β
and

GA
z (0, 0, 0, 0, 0, 0, 0, 0, 0) =
αβ
2 (Bz − Az )E(0, 0, 0, 0, 0, 0, 0, 0, 0) (28)
α+β

Recursion relations for the G-coefficients can be derived using similar arguments as for

the E-coefficients19 . There exist recursion relations to generate E(n + 1, l, m, ñ, ˜l, m̃, t, u, v),
E(n, l + 1, m, ñ, ˜l, m̃, t, u, v), E(n, l + 1, l + 1, ñ, ˜l, m̃, t, u, v) and E(n, −l − 1, −l −
1, ñ, ˜l, m̃, t, u, v). Recursions are now necessary for GA A A
x , Gy and Gz .

1. Recursion in l and m

~ n, l + 1, l + 1) = (2l + 1)((x − Ax ) + i(y − Ay ))S(α, ~r − A,

With S(α, ~r − A, ~ n, l, l), we

obtain:

~ ñ, ˜l, m̃) =

~ n, l + 1, l + 1)S(β, ~r − B,
S(α, ~r − A,

11
 E(n, l, l, ñ, ˜l, m̃, t, u, v)((x − Ax ) + i(y − Ay ))Λ(γ, ~r − P~ , t, u, v) =
X
(2l + 1)
t,u,v

E(n, l + 1, l + 1, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t, u, v)

X
(29)
t,u,v

In the case of gradients, we obtain:

∂ ~ n, l + 1, l + 1)S(β, ~r − B, ~ ñ, ˜l, m̃) =

S(α, ~r − A,
∂Ax
∂
X 
(2l + 1) ˜ ~
E(n, l, l, ñ, l, m̃, t, u, v)((x − Ax ) + i(y − Ay ))Λ(γ, ~r − P , t, u, v) =
∂Ax t,u,v

−E(n, l, l, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t, u, v) +

X
(2l + 1)
t,u,v
!
∂

((x − Ax ) + i(y − Ay )) E(n, l, l, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t, u, v) =
∂Ax

−E(n, l, l, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t, u, v) +

X
(2l + 1)
t,u,v
!
((x − Ax ) + i(y − Ay ))GA ˜ − P~ , t, u, v) =
x (n, l, l, ñ, l, m̃, t, u, v)Λ(γ, ~
r

GA ˜ r − P~ , t, u, v)
X
x (n, l + 1, l + 1, ñ, l, m̃, t, u, v)Λ(γ, ~ (30)
t,u,v

By substituting the recursion relation for Hermite polynomials

1
(x − Px )Λ(γ, ~r − P~ , t, u, v) = Λ(γ, ~r − P~ , t + 1, u, v) + tΛ(γ, ~r − P~ , t − 1, u, v) (31)
2γ

in the penultimate expression of equation 30, we obtain:

−E(n, l, l, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t, u, v) +

X
(2l + 1)
t,u,v
1

Λ(γ, ~r − P~ , t + 1, u, v) +
2γ
tΛ(γ, ~r − P~ , t − 1, u, v) +

(Px − Ax )Λ(γ, ~r − P~ , t, u, v) +
1
Λ(γ, ~r − P~ , t, u + 1, v) +

i
2γ
uΛ(γ, ~r − P~ , t, u − 1, v) +

(Py − Ay )Λ(γ, ~r − P~ , t, u, v) GA ˜
x (n, l, l, ñ, l, m̃, t, u, v) (32)

From this, we deduce the following relation:

12
 GA ˜
x (n, l + 1, l + 1, ñ, l, m̃, t, u, v) =

(2l + 1) − E(n, l, l, ñ, ˜l, m̃, t, u, v) +
1 A
G (n, l, l, ñ, ˜l, m̃, t − 1, u, v) +
2γ x
(t + 1)GA (n, l, l, ñ, ˜l, m̃, t + 1, u, v) +
x

(Px − Ax )GA ˜
x (n, l, l, ñ, l, m̃, t, u, v) +
1 A
G (n, l, l, ñ, ˜l, m̃, t, u − 1, v) +

i
2γ x
(u + 1)GA (n, l, l, ñ, ˜l, m̃, t, u + 1, v) +
x

(Py − Ay )GA ˜
x (n, l, l, ñ, l, m̃, t, u, v) (33)

In an analogous way, we obtain the following recursion relation:

GA ˜
x (n, l + 1, −l − 1, ñ, l, m̃, t, u, v) =

(2l + 1) − E(n, l, −l, ñ, ˜l, m̃, t, u, v) +
1 A
G (n, l, −l, ñ, ˜l, m̃, t − 1, u, v) +
2γ x
(t + 1)GA (n, l, −l, ñ, ˜l, m̃, t + 1, u, v) +
x

(Px − Ax )GA ˜
x (n, l, −l, ñ, l, m̃, t, u, v) −
1 A
G (n, l, −l, ñ, ˜l, m̃, t, u − 1, v) +

i
2γ x
(u + 1)GA (n, l, −l, ñ, ˜l, m̃, t, u + 1, v) +
x

(Py − Ay )GA ˜
x (n, l, −l, ñ, l, m̃, t, u, v) (34)

2. Recursion in n

~ n + 1, l, m) = |~r − A|
With S(α, ~r − A, ~ 2 S(α, ~r − A,
~ n, l, m) the following relation can be

derived19 :

GA ˜
x (n + 1, l, m, ñ, l, m̃, t, u, v) =
1  A
G (n, l, m, ñ, ˜l, m̃, t − 2, u, v) +
4γ 2 x

13
 GA ˜
x (n, l, m, ñ, l, m̃, t, u − 2, v) +

˜

GA
x (n, l, m, ñ, l, m̃, t, u, v − 2) +
1 ˜
(Px − Ax )GA
x (n, l, m, ñ, l, m̃, t − 1, u, v) −
γ
E(n, l, m, ñ, ˜l, m̃, t − 1, u, v) +

(Py − Ay )GA ˜
x (n, l, m, ñ, l, m̃, t, u − 1, v) +

˜

(Pz − Az )GA
x (n, l, m, ñ, l, m̃, t, u, v − 1) +
3 
~ 2 + (t + u + v + 2 ) GA (n, l, m, ñ, ˜l, m̃, t, u, v) −
|P~ − A|

x
γ
2(Px − Ax )E(n, l, m, ñ, ˜l, m̃, t, u, v) +
˜

2 (Px − Ax )(t + 1)GA
x (n, l, m, ñ, l, m̃, t + 1, u, v) −

(t + 1)E(n, l, m, ñ, ˜l, m̃, t + 1, u, v) +

(Py − Ay )(u + 1)GA ˜

x (n, l, m, ñ, l, m̃, t, u + 1, v) +

˜

(Pz − Az )(v + 1)GA
x (n, l, m, ñ, l, m̃, t, u, v + 1) +

(t + 2)(t + 1)GA ˜
x (n, l, m, ñ, l, m̃, t + 2, u, v) +

(u + 2)(u + 1)GA ˜
x (n, l, m, ñ, l, m̃, t, u + 2, v) +

(v + 2)(v + 1)GA ˜
x (n, l, m, ñ, l, m̃, t, u, v + 2) (35)

3. Recursion in l

~ n, l + 1, m) = ((2l + 1)(z − Az )S(α, ~r − A,

Using that S(α, ~r − A, ~ n, l, m) − |~r − A|
~ 2 (l +

~ n, l − 1, m))/(l − |m| + 1) the following relation can be derived19 :

|m|)S(α, ~r − A,

GA ˜
x (n, l + 1, m, ñ, l, m̃, t, u, v) =
G (n, l, m, ñ, ˜l, m̃, t, u, v − 1)
 A
(2l + 1) x +
2γ
(Pz − Az )GA (n, l, m, ñ, ˜l, m̃, t, u, v) +
x

(v + 1)GA ˜
x (n, l, m, ñ, l, m̃, t, u, v + 1) −

(l + |m|) GA ˜
x (n, l − 1, m, ñ, l, m̃, t − 2, u, v) +

14
 GA ˜
x (n, l − 1, m, ñ, l, m̃, t, u − 2, v) +

˜

GA 2
x (n, l − 1, m, ñ, l, m̃, t, u, v − 2) /(2γ) +

˜

(Px − Ax )GA
x (n, l − 1, m, ñ, l, m̃, t − 1, u, v) −

E(n, l − 1, m, ñ, ˜l, m̃, t − 1, u, v) +

(Py − Ay )GA ˜
x (n, l − 1, m, ñ, l, m̃, t, u − 1, v) +

˜

(Pz − Az )GA
x (n, l − 1, m, ñ, l, m̃, t, u, v − 1) /γ +
3 
~ 2 + (t + u + v + 2 ) GA (n, l − 1, m, ñ, ˜l, m̃, t, u, v) −
|P~ − A|

x
γ
2(Px − Ax )E(n, l − 1, m, ñ, ˜l, m̃, t, u, v) +
˜

2 (Px − Ax )(t + 1)GA
x (n, l − 1, m, ñ, l, m̃, t + 1, u, v) −

(t + 1)E(n, l − 1, m, ñ, ˜l, m̃, t + 1, u, v) +

(Py − Ay )(u + 1)GA ˜

x (n, l − 1, m, ñ, l, m̃, t, u + 1, v) +

˜

(Pz − Az )(v + 1)GA
x (n, l − 1, m, ñ, l, m̃, t, u, v + 1) +

(t + 2)(t + 1)GA ˜
x (n, l − 1, m, ñ, l, m̃, t + 2, u, v) +

(u + 2)(u + 1)GA ˜
x (n, l − 1, m, ñ, l, m̃, t, u + 2, v) +
!
(v + 2)(v + 1)GA ˜
x (n, l − 1, m, ñ, l, m̃, t, u, v + 2) /(l − |m| + 1) (36)

The recursion relations for the GA 20

x -coefficients are similar to those for the E-coefficients ,

∂
with the extra terms arising from the derivative ∂Ax
(x − Ax ) for the recursion in l and m,
and from the derivative ∂
|~r ~ 2 for the recursion in l and the recursion in n.
− A|
∂Ax

This can similarly be done for GA A

y and Gz . Finally, all G-coefficients are zero for the

case t + u + v > 2n + 2ñ + l + ˜l + 1 and for all negative values of t, u or v.

B. Gradients with respect to other centers

To obtain the derivatives with respect to center B, the following relation is used31 (note
~ =A
R ~ − B):
~

∂ ∂ ∂Px ∂ ∂Rx
= + =
∂Ax ∂Px ∂Ax ∂Rx ∂Ax

15
 α ∂ ∂
+
γ ∂Px ∂Rx
(37)

and

∂ ∂ ∂Px ∂ ∂Rx
= + =
∂Bx ∂Px ∂Bx ∂Rx ∂Bx
β ∂ ∂
−
γ ∂Px ∂Rx
(38)

Therefore,

∂ ∂ ∂
+ = (39)
∂Ax ∂Bx ∂Px

This means that

∂ X
E(n, l, m, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t, u, v) =
∂Bx t,u,v
!
∂ ∂
E(n, l, m, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t, u, v)
X
− (40)
∂Px ∂Ax t,u,v

Applying equation 25, we obtain:

GB ˜ r − P~ , t, u, v) =
X
x (n, l, m, ñ, l, m̃, t, u, v)Λ(γ, ~
t,u,v

E(n, l, m, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t + 1, u, v)

X

t,u,v

GA ˜ r − P~ , t, u, v)
X
− x (n, l, m, ñ, l, m̃, t, u, v)Λ(γ, ~ (41)
t,u,v

Finally, we conclude:

GB ˜ ˜ A ˜
x (n, l, m, ñ, l, m̃, t, u, v) = E(n, l, m, ñ, l, m̃, t − 1, u, v) − Gx (n, l, m, ñ, l, m̃, t, u, v) (42)

and similar for derivatives with respect to y and z direction. All the integrals can be
expressed with the help of the E-coefficients. Taking the derivative therefore reduces to

replacing E with the corresponding GA A

x , Gy , etc coefficients.

16
 As an example, we show how overlap integral and gradient are obtained. First, the
overlap is computed as follows:
Z
~ ñ, ˜l, m̃)d3 r =
~ n, l, m)S(β, ~r − B,
S(α, ~r − A,
Z X
E(n, l, m, ñ, ˜l, m̃, t, u, v)Λ(γ, ~r − P~ , t, u, v)d3r =
t,u,v
!3
π 2
E(n, l, m, ñ, ˜l, m̃, 0, 0, 0) (43)
γ

We have used that

!3
π 2
Z
3
Λ(γ, ~r, t, u, v)d r = δt0 δu0 δv0 (44)
γ

because of the orthogonality of the Hermite Gaussian type functions (δt0 is the Kronecker
delta).
The gradient is computed similarly:

∂
Z
~ ñ, ˜l, m̃)d3 r =
~ n, l, m)S(β, ~r − B,
S(α, ~r − A,
∂A
Z x
GA ˜ r − P~ , t, u, v)d3r =
X
x (n, l, m, ñ, l, m̃, t, u, v)Λ(γ, ~
t,u,v
!3
˜ π 2
GA
x (n, l, m, ñ, l, m̃, 0, 0, 0) (45)
γ

In our implementation, we therefore compute the gradient of the two Gaussians which
are associated with the integrals, by replacing E-coefficients with G-coefficients. As a con-
sequence, if an operator, which might appear in the integral, has a nonvanishing derivative
(such as, for example, the nuclear attraction), this must be taken into account additionally.

This derivative with respect to the third center can be obtained by applying translational
invariance with respect to a simultaneous uniform translation of the three centers. In the
case of bielectronic integrals, products with two E-coefficients appear. Obviously, when dif-
ferentiating, the corresponding rules of differentiating a product must be applied and two

derivative terms appear, each of them consisting of a product of one set of E− and one set
of G−coefficients. Finally, the nuclear-nuclear term must be differentiated which is trivial.

17
 It is interesting to compare this implementation with that of the Namur group14 where
also gradients within the McMurchie-Davidson algorithm are computed. Whereas our
scheme computes the derivatives of the two Gaussians appearing in the integral and a possi-

bly necessary derivative of an operator is obtained by applying translational invariance, the

alternative implementation14 relies on explicitly computing derivatives of E-coefficients and
of the auxiliary function27 Rt,u,v appearing in the integrals.

V. TOTAL ENERGY

The total energy consists of kinetic energy, Coulomb energy (nuclear-nuclear repulsion,
nuclear-electron attraction and electron-electron repulsion), and exchange energy. We as-

sume that all the orbitals are either empty or doubly occupied.

A. Kinetic energy

The kinetic energy of the electrons is obtained as:

~ µ ) − 1 ∆~r φν (~r − A
Z  
kinetic ~ ν − ~g )d3 r =
X
E = Pν~gµ~0 φµ (~r − A
~g ,µ,ν
2
X
Pν~gµ~0 Tµ~0ν~g (46)
~g ,µ,ν

B. Exchange energy

The exchange energy is obtained as:

E exch−el =
1 X X
− Pν~gµ~0 Pσ~h+~nτ~n
4 ~g,µ,ν ~ h,~
n,τ,σ
Z
φµ (~r − A ~ µ )φτ (~r − A~ τ − ~n)φν (~r ′ − A ~ σ − ~h − ~n)
~ ν − ~g )φσ (~r ′ − A
d3 r d3 r′ =
|~r − ~r ′ |
1 X X
− Pν~gµ~0 Pσ~hτ~0 Xµ~0ν~gτ~0σ~h (47)
4 ~g,µ,ν ~ h,τ,σ

18
 where we have exploited translational invariance of the density matrix with respect to
direct lattice vectors ~n: Pσ~h+~nτ~n = Pσ~hτ~0 .
We can define a Fock operator for the exchange energy which is

1 X
Fµexch−el
~0ν~
g
=− Pσ~hτ~0 Xµ~0ν~gτ~0σ~h (48)
2~
h,τ,σ

C. Coulomb energy

Both kinetic energy and exchange energy must converge independently. However, a
separation of the contributions to the Coulomb energy is not possible: for example, in a one
dimensional periodic system with lattice constant a, and n being an index enumerating the
cells, the electron-electron interaction per unit cell would have contributions like:
∞
X 1
(49)
n=1 na

This sum is divergent (similarly in two and three dimensions). Therefore, in CRYSTAL
a scheme based on the Ewald method is used to sum the interactions32,26 . We only quote
the results for the individual contributions:

1. Nuclear-nuclear repulsion:

1X ~b − A
~ a)
E NN = Za Zb A(A (50)
2 a,b

2. Nuclear-electron attraction:

1
The energy E ,
2 NE
which is the Ewald energy of the nuclei in the primitive unit cell

with the all the electrons of all cells, is the same as the energy 21 EEN , which is the Ewald
energy of the electrons of the primitive unit cell with all the nuclei in all cells, as long as
no approximations are introduced30 . CRYSTAL uses the following expression as the sum of
these interactions:

19
 E coul−nuc = Pν~gµ~0 Fµcoul−nuc
X
~0ν~
g
(51)
~g ,µ,ν

with the Fock matrix Fµcoul−nuc

~0ν~
g
containing the nuclear-electron contributions defined as
Z
Fµcoul−nuc ~ µ )φν (~r − Aν − ~g )A(~r − A
~ a )d3 r
X
~0ν~
g
=− Za φµ (~r − A (52)
a

3. Electron-electron repulsion:

1 X
E coul−el = P ~ F coul−el (53)
2 ~g,µ,ν ν~gµ0 µ~0ν~g

with the Fock matrix Fµcoul−el

~0ν~
g
containing the electron-electron contributions defined as

L X
l
Fµcoul−el ~ c )M m~
ηlm (ρc ; A
X XX
~0ν~
g
= −QSµ~0ν~g + Pσ~hτ~0 Cµ~0ν~gτ~0σ~h − lµ0ν~
gc
(54)
~h,τ,σ c l=0 m=−l

D. Total energy

Finally, the total energy can be expressed as

E total = E kinetic + E NN + E coul−nuc + E coul−el + E exch−el =

= E NN +
X
Pν~gµ~0 Tµ~0ν~g
~g,µ,ν
Z
~ µ )φν (~r − A
~ ν − ~g )A(~r − A
~ a )d3 r
X X
− Pν~gµ~0 Za φµ (~r − A
~g ,µ,ν a
L X l
1 X
 
~ c )M m~
ηlm (ρc ; A
X XX
+ Pν~gµ~0 − QSµ~0ν~g + Pσ~hτ~0 Cµ~0ν~gτ~0σ~h − lµ0ν~
gc
2 ~g,µ,ν ~ c l=0 m=−l
h,τ,σ
1 X X
− Pν~gµ~0 Pσ~hτ~0 Xµ~0ν~gτ~0σ~h (55)
4 ~g,µ,ν ~ h,τ,σ

The Fock operator used has the structure:

Fµtotal
~0ν~
g
= Tµ~0ν~g + Fµcoul−nuc
~0ν~
g
+ Fµcoul−el
~0ν~
g
+ Fµexch−el
~0ν~
g
(56)

We note that this expression for the Fock operator would be exact if we could guar-
antee that the penetration depth and screening was symmetric. This would require that

20
Cµ~0ν~gτ~0σ~h = Cτ~0σ~hµ~0ν~g should always hold. This, however, as aforementioned, cannot be
guaranteed because the truncation is not necessarily symmetric. In addition, the screening
of the exchange interaction is not necessarily symmetric. Therefore, an inaccuracy in the

Fock operator will show up which will be stronger the more asymmetric the truncation in
the energy expression is.
The total energy can be expressed as

1
 
total NN
Fµcoul−nuc + (Fµcoul−el + Fµexch−el
X
E =E + Pν~gµ~0 Tµ~0ν~g + ~0ν~ ) (57)
~g ,µ,ν
g 2 ~0ν~g ~0ν~
g

and the Hartree-Fock equations become as in equation 5. In ref. 26, it was pointed out
that the quantity QSµ~0ν~g can be removed from the Fock operator which has been done in
CRYSTAL. This leads to eigenvalues shifted by Q as we now use the modified equation

total ~
(k) + QSµν (~k))aνi (~k) = Sµν (~k)aνi (~k)(ǫi (~k) + Q)
X X
(Fµν (58)
ν ν

VI. GRADIENT OF THE TOTAL ENERGY

The force on the nuclei can be calculated similarly to the molecular case2,1 . The deriva-

tives of all the integrals are necessary, and the derivative of the density matrix is expressed
with the help of the energy-weighted density matrix. One important assumption is that

Bµ~0ν~gτ~nσ~n+~h = Bτ~nσ~n+~hµ~0ν~g = Bσ~n+~hτ~nν~gµ~0 (59)

holds. Taking the derivative leads, for example, to terms like the following

∂
 X X 
Pν~gµ~0 Pσ~hτ~0 Cµ~0ν~gτ~0σ~h =
~i
∂A ~g ,µ,ν ~h,τ,σ
X  ∂ ∂
 X X X  
P Pσ~hτ~0 Cµ~0ν~gτ~0σ~h + Pν~gµ~0 P C
~
g ,µ,ν
~ i ν~gµ~0
∂A ~h,τ,σ ~g ,µ,ν ~h,τ,σ
~ i σ~hτ~0 µ~0ν~gτ~0σ~h
∂A
∂
X X  
+ Pν~gµ~0 Pσ~hτ~0 C (60)
~g ,µ,ν ~h,τ,σ
~ i µ~0ν~gτ~0σ~h
∂A

When equation 59 holds, we rename the indices in the second addend and obtain:

21
 X  ∂ ∂
 X X X  
2 P Pσ~hτ~0 Cµ~0ν~gτ~0σ~h + Pν~gµ~0 Pσ~hτ~0 C (61)
~
g ,µ,ν
~ i ν~gµ~0
∂A ~h,τ,σ ~g ,µ,ν ~h,τ,σ
~ i µ~0ν~gτ~0σ~h
∂A

We derived the equation for the force this way although equation 59 does not always

hold. Therefore, inaccuracies will appear when equation 59 is strongly violated. The full
force is obtained as:
total
∂E
F~Ai = − =
∂A ~i
X ∂Tµ~0ν~g ∂E NN
− Pν~gµ~0 −
~g ,µ,ν ∂ ~
A i
~i
∂A
∂ Z
 
~ µ )φν (~r − A
~ ν − ~g )A(~r − A
~ a )d3 r
X X
+ Pν~gµ~0 Za φµ (~r − A
~
g ,µ,ν a
~i
∂A
1 X 2π X X

− Pν~gµ~0 − Sµ~0ν~g Pσ~hτ~0
2 ~g,µ,ν 3V c ~
h,σ,τ ∈c
∂
Z 
~ σ − ~h)
~ τ )φσ (~r − A
− φτ (~r − A
~i
∂A
L X
l Z 
′ ~ σ − ~h)X m (~r ′ − A
~ τ )φσ (~r ′ − A ~ c )d3 r′ δ m (A
~ c , ~r) r 2 d3 r
X
+ φτ (~r − A l l
l=0 m=−l
X ∂Cµ~0ν~gτ~0σ~h
+ Pσ~hτ~0
τ,σ ∂A ~i
L X
l
∂
Z 
~ σ − ~h)X m (~r − A
~ τ )φσ (~r − A ~c )d3 r M m~
XX X
− Pσ~hτ~0 φτ (~r − A l
c l=0 m=−l ~h,τ ∈c,σ ~i
∂A lµ0ν~
gc

1 X X ∂Xµ~0ν~gτ~0σ~h
+ Pν~gµ~0 Pσ~hτ~0
4 ~g,µ,ν ~ ∂A ~i
h,τ,σ
∂Sµ~0ν~g Z
exp(i~k~g ) 2aνj (~k)a∗µj (~k)(ǫj (~k) + Q)Θ(ǫF − ǫj (~k) − Q)d3 k
X X
− (62)
~g ,µ,ν
~
∂ Ai BZ j

The last addend is the energy weighted density matrix; the integral is over the first

Brillouin zone. It is worthwhile mentioning that the factor Pν~gµ~0 Sµ~0ν~g is equal to the number
~ i vanishes. We note
of electrons in the unit cell and therefore its derivative with respect to A
three important points:

• Equation 62 is correct for the exact solution of the Hartree-Fock equations. Thus, in
practice, a well converged solution is necessary to achieve accurate forces.

22
 • The energy-weighted density matrix is ~k-dependent. Therefore, the accuracy of the

forces will become dependent on the number of ~k-points.

• The derivation of equation 62 assumes that equation 59 holds. The treatment of the
Coulomb series with finite penetration depth leads to an asymmetry associated with
IT OL2 in the CRYSTAL code. In addition, in the treatment of the exchange series

an asymmetry can be introduced if the screening parameters (IT OL4 and IT OL5 in
the CRYSTAL code) are chosen differently. Therefore, the choice of IT OL2, IT OL4
and IT OL5 will influence the accuracy of the gradients.

VII. RESULTS FROM TEST CALCULATIONS

With a few examples, we want to illustrate the accuracy of the analytically computed
gradients.
In table I, all the integrals are evaluated without approximation and the analytical
derivative agrees to five digits with the numerical derivative. As the numerical derivative is

only accurate up to five digits, this is certainly satisfying.

In table II the variation in accuracy when penetration depth and overlap criteria are
altered, is displayed. As described in the article, lowering IT OL2 to low values leads to
inaccuracies in the gradients. Lowering only one of the parameters IT OL4 or IT OL5 also
leads to inaccuracies, whereas lowering both to the same value gives an analytical gradient

which is consistent with the numerical gradient — however, as a value of 1 for IT OL4
and IT OL5 was chosen, energy and force are completely different from calculations with
reasonable values for IT OL4 and IT OL5. The parameter IT OL1, which selects the one-
electron and Coulomb integrals according to the overlap, can lead to numerical instabilities

in the energy calculation and inaccurate gradients when chosen too low, and should therefore
be reasonably high. The parameter IT OL3 does not influence the accuracy of the gradients:
although when chosen much too low with a value of e.g. 1, numerical and analytical gradient
still agree. The default values for the IT OL-parameters for the energy calculation are 6, 6,

23
6, 6 and 12. In all calculations performed so far, these default values did not lead to serious
errors for the gradients.
Another example (table III) is the CO molecule arranged as a single molecule

(”molecule”), as a molecule periodically repeated in one dimension (”polymer”), in two

dimensions (”slab”) and in three dimensions (”bulk”). The forces agree well, and it is
demonstrated that using stricter real-space truncation parameters improves the agreement.
The forces seem to be relatively insensitive to the number of sampling points and changing

their number changed the error in the forces only slightly.

Finally, in table IV we compare analytical and numerical derivatives for MgO when
moving the oxygen atom in x-direction which would correspond to a longitudinal phonon.
Again, agreement is to the order of 10−5 E
aB
h
(default IT OL parameters were used).

VIII. CONCLUSION

We presented the theory of analytic Hartree-Fock gradients for periodic systems. This
has been implemented in the code CRYSTAL which is to the best of our knowledge the first
implementation of Hartree-Fock gradients in systems periodic in 2 and 3 dimensions. The

results are in excellent agreement with numerical derivatives.

Future directions will be the improvement of the efficiency of the code (implementation
of symmetry and various technical improvements), derivatives with respect to the lattice
vector, as well as an extension to metallic systems.

IX. ACKNOWLEDGMENTS

The authors would like to acknowledge support from EPSRC grant GR/K90661.

24
 REFERENCES

1
P. Pulay, Mol. Phys. 17, 197 (1969).

2
S. Bratoz̆, in Calcul des fonctions d’onde moléculaire, Colloq. Int. C. N. R. S. 82, 287
(1958).

3
P. Pulay, Adv. Chem. Phys. 69, 241 (1987).

4
P. Pulay, in Applications of Electronic Structure Theory, edited by H. F. Schaefer III, 153

(Plenum, New York, 1977).

5
H. B. Schlegel, Adv. Chem. Phys. 67, 249 (1987).

6
T. Helgaker and P. Jørgensen, Adv. in Quantum Chem. 19, 183 (1988)

7
H. Teramae, T. Yamabe, C. Satoko and A. Imamura, Chem. Phys. Lett. 101, 149 (1983);
H. Teramae, T. Yamabe and A. Imamura, J. Chem. Phys. 81, 3564 (1984).

8
M. Kertesz, Chem. Phys. Lett. 106, 443 (1984).

9
S. Suhai, Phys. Rev. B 27, 3506 (1983); S. Suhai, Chem. Phys. Lett. 96, 619 (1983).

10
J.-Q. Sun and R. J. Bartlett, J. Chem. Phys. 104, 8553 (1996).

11
S. Hirata and S. Iwata, J. Chem. Phys. 109, 4147 (1998).

12
S. Hirara and S. Iwata, J. Chem. Phys. 107, 10075 (1997); S. Hirata, H. Torii and M.
Tasumi, Phys. Rev. B 57, 11994 (1998); S. Hirata and S. Iwata, J. Chem. Phys. 108, 7901
(1998); S. Hirata and S. Iwata, J. Phys. Chem. A 102, 8426 (1998).

13
S. Hirata and S. Iwata, J. Mol. Struct.: THEOCHEM 451, 121 (1998).

14
D. Jacquemin, J.-M. André and B. Champagne, J. Chem. Phys. 111, 5306 (1999); J.
Chem. Phys. 111, 5324 (1999).

15
K. N. Kudin and G. E. Scuseria, Phys. Rev. B 61, 16440 (2000).

25
16
V. R. Saunders, R. Dovesi, C. Roetti, M. Causà, N. M. Harrison, R. Orlando, C. M.
Zicovich-Wilson, crystal 98 User’s Manual, Theoretical Chemistry Group, University
of Torino (1998).

17
C. Pisani and R. Dovesi, Int. J. Quantum Chem. 17, 501 (1980); R. Dovesi and C. Roetti,
Int. J. Quantum Chem. 17, 517 (1980).

18
C. Pisani, R. Dovesi, and C. Roetti, Hartree-Fock Ab Initio Treatment of Crystalline
Systems, edited by G. Berthier et al, Lecture Notes in Chemistry Vol. 48 (Springer, Berlin,
1988).

19
V. R. Saunders, N. M. Harrison, R. Dovesi, C. Roetti, Electronic Structure Theory: From
Molecules to Crystals (in preparation)

20
V. R. Saunders, in Methods in Computational Molecular Physics, edited by G. H. F.
Diercksen and S. Wilson, 1 (Reidel, Dordrecht, Netherlands, 1984).

21
J. M. André, L. Gouverneur and G. Leroy, Int. J. Quant. Chem. 1, 427 (1967); Int. J.
Quant. Chem. 1, 451 (1967); J. M. André, J. Chem. Phys. 50, 1536 (1969).

22
J. Delhalle, J. M. André, Ch. Demanet and J. L. Brédas, Chem. Phys. Lett. 54, 186
(1978).

23
V. R. Saunders, C. Freyria-Fava, R. Dovesi, and C. Roetti, Comp. Phys. Comm. 84, 156
(1994).

24
D. E. Parry, Surf. Science 49, 433 (1975); 54, 195 (1976) (Erratum).

25
P. P. Ewald, Ann. Phys. (Leipzig) 64, 253 (1921).

26
V. R. Saunders, C. Freyria-Fava, R. Dovesi, L. Salasco, and C. Roetti, Mol. Phys. 77, 629
(1992).

27
L. E. McMurchie and E. R. Davidson, J. Comput. Phys. 26, 218 (1978).

26
28
S. F. Boys, Proc. Roy. Soc. A 200, 542 (1950).

29
R. McWeeny, Nature 166, 21 (1950).

30
R. Dovesi, C. Pisani, C. Roetti, and V. R. Saunders, Phys. Rev. B 28, 5781 (1983).

31
T. Helgaker and P. R. Taylor, Theor. Chim. Acta 83, 177 (1992).

32
F. E. Harris, in Theoretical Chemistry: Advances and Perspectives, Vol. 1, 147 (1975),
edited by H. Eyring and D. Henderson, Academic Press, New York

27
 TABLES
TABLE I. CO molecule, all tolerances high enough that all the integrals are done without

approximation. The carbon atom is placed at (0 Å, 0 Å, 0 Å), the oxygen atom at (0.8 Å, 0.5

Å, 0.4 Å). To calculate the numerical force in x-direction, the x-coordinate of the oxygen atom is

changed (column 1). The Hartree-Fock energy is displayed in column 2. The numerical force is

determined from 2 points at (0.8 ±δ) Å and displayed in column 3. The analytical derivative for

x=0.8 Å is displayed in column 4. Basis sets of the size [2s2p1d] were used.

x-coordinate of oxygen energy numerical derivative analytical derivative

(x-component) (x-component)

Å Eh Eh /aB Eh /aB

0.799 -1.107642201574E+02 0.376915

0.7999 -1.107648645627E+02 0.376913

0.79999 -1.107649287000E+02 0.376914

0.80000 -1.107649358230E+02 0.37691274

0.80001 -1.107649429453E+02

0.8001 -1.107650070153E+02

0.801 -1.107656446886E+02

28
 TABLE II. CO molecule, when varying the IT OL parameters. The first and second parameter

should be reasonably high, the fourth and fifth should both be high or identical to obtain accurate

forces.

IT OL values numerical derivative analytical derivative

(x-component) (x-component)

Eh /aB Eh /aB

6 6 6 6 12 0.37691 0.376913

20 20 20 20 20 0.37691 0.376913

4 20 20 20 20 0.37691 0.376912

3 20 20 20 20 0.37747 0.377389

2 20 20 20 20 0.38883 0.388260

20 2 20 20 20 0.37684 0.376821

20 1 20 20 20 -7.40676 -1.246409

20 20 1 20 20 -8.09240 -8.09240

20 20 20 1 20 -0.51905 -3.504654

20 20 20 20 1 0.40892 0.359847

20 20 20 1 1 -33.34823 -33.348229

29
 TABLE III. CO molecule, arranged periodic in 0 dimensions (”molecule”), in 1 dimension

(”polymer”), in 2 dimensions (”slab”) and in 3 dimensions (”bulk”). We display the accuracy

of the gradient as a function of the number of sampling points, and as a function of the IT OL

parameters.

dimension number of sampling points IT OL parameters numerical derivative analytical derivative

(x-component) (x-component)

Eh /aB Eh /aB

0 - 6 6 6 6 12 0.37691 0.376913

1 5 6 6 6 6 12 0.37659 0.376633

1 5 6 9 6 12 12 0.37662 0.376647

2 34 6 6 6 6 12 0.37630 0.376335

3 8 6 6 6 6 12 0.37571 0.375742

3 260 6 6 6 6 12 0.37565 0.375679

3 260 6 8 6 8 14 0.37570 0.375720

3 260 8 8 6 8 14 0.37573 0.375721

30
 TABLE IV. MgO at a lattice constant of 4.21 Å. We compare numerical and analytical deriva-

tives when moving the oxygen ion parallel to the x-direction. Basis sets of the size [3s2p] were

used.

displacement of oxygen analytical derivative numerical derivative

(x-component) (x-component)

in % of lattice constant Eh /aB Eh /aB

+1 -0.005923 -0.00593

+2 -0.012254 -0.01226

+3 -0.019385 -0.01940

31