An all-electron numerical method for

solving the local density functional for

polyatomic molecules
Cite as: J. Chem. Phys. 92, 508 (1990); https://doi.org/10.1063/1.458452
Submitted: 26 June 1989 . Accepted: 05 September 1989 . Published Online: 31 August 1998

B. Delley

ARTICLES YOU MAY BE INTERESTED IN

Density-functional thermochemistry. III. The role of exact exchange

The Journal of Chemical Physics 98, 5648 (1993); https://doi.org/10.1063/1.464913

A consistent and accurate ab initio parametrization of density functional dispersion

correction (DFT-D) for the 94 elements H-Pu
The Journal of Chemical Physics 132, 154104 (2010); https://doi.org/10.1063/1.3382344

A multicenter numerical integration scheme for polyatomic molecules

The Journal of Chemical Physics 88, 2547 (1988); https://doi.org/10.1063/1.454033

J. Chem. Phys. 92, 508 (1990); https://doi.org/10.1063/1.458452 92, 508

© 1990 American Institute of Physics.

An all-electron numerical method for solving the local density functional
for polyatomic molecules
B. Delley
Paul Scherrer Institut clo Laboratories RCA, Badenerstrasse 569, CH-8048 Zurich, Switzerland
(Received 26 June 1989; accepted 5 September 1989)
A method for accurate and efficient local density functional calculations (LDF) on molecules
is described and presented with results. The method, Dmol for short, uses fast convergent
three-dimensional numerical integrations to calculate the matrix elements occurring in the
Ritz variation method. The flexibility of the integration technique opens the way to use the
most efficient variational basis sets. A practical choice of numerical basis sets is shown with a
built-in capability to reach the LDF dissociation limit exactly. Dmol includes also an efficient,
exact approach for calculating the electrostatic potential. Results on small molecules illustrate
present accuracy and error properties of the method. Computational effort for this method
grows to leading order with the cube of the molecule size. Except for the solution of an
algebraic eigenvalue problem the method can be refined to quadratic growth for large
molecules.

I. INTRODUCTION that all ground state properties are functionals of the density.
Whereas quantum mechanical calculations on mole- In particular, the ground state energy of an electron system
cules started in the late 1920's right after the formulation of in an exterual potential, i.e., the potential due to a set of
quantum mechanics, practical calculations on polyatomic nuclei in a given arrangement, can be found by minimizing
molecules with approximate wave functions became feasible the functional with respect to variations in the density. This
with the introduction of variational Gaussian basis sets by leads to a set of equations to be solved self-consistently, in-
volving a local potential f..txc (r) .4,20,21
Roothaan 1 and Hall. 2 Much later the density functional ap-
proach 3 was introduced. This approach deemphasizes the
{ ; : Il + Vs(r) +f..txc,u(r)} <l>k,a(r) Ej<l>k,,, (r),
importance of the wave function and leads to a way to ap-
proximate calculations of electronic properties. This ap- (la)
proach, at its most simple approximation level, the local den-
sity functional (LDF) approach,'*-7 has been much used in
electronic structure calculations on molecules and solids.
pu(r) +
occ
2
l<I>k,a(r) 1 , P=Pu +P-u' (lb)

LDF is becoming an increasingly important method in
quantum chemistry, as it leads to much more reliable predic-
tions for large molecules. The power of LDF methods, and
in particular of the method presented here, has been demon-
strated in numerous applications. 8-12
Here, a methOd, called Dmol for short, for local density
functional calculations on molecules is described. The suc-
cess of the method and its predecessors has been shown in
applications to the ground state of highly correlated transi-
tion metal clusters such as Cr2 ,13 to energy surfaces of a
catalyst model Cu or Ag on SiC 111),14,15 to electron density
and related static properties of molecules l6• 17 and to low ly-
ing states of organo-metallic complexes. 18 In Sec. II, a brief
introduction into Kohn-Sham theory and linear variational
solutions are given. In Sec, III, the Dmol method is de-
scribed. Section III A describes the variational basis sets that
are used. Section III B deals with the numerical integration
method, Sec. III C with the calculation of the electrostatic
potential using a model charge obtained by a projection
method, Sec. III D with the errors due to charge truncation.
In Sec. IV, a discussion of computational effort is given. A
summary of the present status is given in Sec. V.
II. BRIEF DESCRIPTION OF THE LOCAL DENSITY
FUNCTIONAL APPROACH
Density functional theory starts from a theorem due to
Hohenberg and Kohn/ later generalized by Levy,19 stating
where Vs is the electrostatic potential due to the nuclei and
the electron charge distribution, u is a spin label, and f..t xc is a
potential which describes exchange and correlation effects.
An exact derivation of the local potential can be given,21 it
depends, however, on all excited states of the electron sys-
tem. In practice, an approximation is used which depends on
the electron density locally f..tXC,q [pu(r),p -a (r)] and is ex-
act in the limits of slowly varying density and in the high
density limit. It also fulfills the exchange hole sum rule ex-
actly. The total energy can be written as
(2a)
Ve is the electrostatic potential due to the electron charge
distribution and Excis a universal potential fundamentally
related to f..txc by d(pExc )/dp = f..txc'
The principal source of error in accurate LDF methods
is clearly the local approximation to the full exchange corre-
lation functional. This error is common to all LDF methods
and is not a topic of this article. For an overview of density
functional theories see, e.g., Ref. 22.
Most quantum chemical methods use a linear variation-
al expansion for the single particle orbitals

508 J. Chem. Phys. 92 (1),1 January 1990 0021-9606/90/010508-10$03.00 © 1990 American Institute of Physics
 B. Delley: Density functional for polyatomic molecules
This converts the differential eigenvalue problem ( la) into a
generalized matrix eigenvalue equation
2: hiJ.uCj,k.u =
j
Ek 2: SiJCj,k,u,
j
(lc)
where
_,,2
h iJ•u =
f { tPi(r) ~ a + Vs(r)
+/-lxc,u[pu(r),p_u(r)]} tPj(r)d
3
r rors which are difficult to discuss. Naturally, Omol relies on
and
SiJ = f tPi (r)tPj (r)d r.
3
LOF methods are typically characterized by the type of orbi-
tal expansion functions tPi (r) used, e.g., Gaussian linear
combination of atomic orbitals (LCAO), Slater type orbital
LCAO, numerical LCAO, muffin tin orbitals, multiple scat-
tering, linearized augmented plane waves, plane waves with
pseudopotential, etc.
It is instructive to compare briefly this method with
standard ab initio quantum chemistry methods. The use of
Gaussian basis sets in Hartree-Fock (HF) and configura-
tion interaction (CI) calculations for general polyatomic
molecules is apparently the only practical way to calculate
the four center matrix elements arising from the Fock term.
Principal limitations ofthese methods are thus basis set limi-
tations at the orbital level and limitations at the level of the
multiconfiguration expansion of the wave function. The
principal merit of the LDF approach results from the possi-
bility to use better basis sets, which is fully exploited in the
present method. There are two basic reasons for the possibil-
ity to use better basis sets in a LOF method. First, the Har-
tree-like matrix elements to be calculated leave more free-
dom of choice for the function classes to be used as basis sets
than do the Fock matrix elements. Second, the growth of
computational effort with basis set size N is proportional to
N 4 and higher in HF and CI methods but only proportional
to N 3 in LOF, making the use of larger basis sets feasible.
III. METHODOLOGY
Omol uses numerically tabulated functions for the Ritz
variation procedure. Details about the basis sets currently
used are given in Sec. IlIA. Omol requires no further ana-
lytic properties for its orbital expansion functions than the
ones guaranteed by single electron theory of bound states,
namely square integrability and cusp singularities only at the
nuclei. Admitting such a large class of expansion functions
forces us to use three-dimensional numerical integrations to
calculate the matrix elements of the eigenvalue equation.
The density, which is a sum of squares of eigenfunctions
according to Eq. (1 b), does not lend itself automatically to
an efficient calculation of the electrostatic potential. Most
molecular LOF methods rely on some fitting procedure to
calculate the electrostatic potential from a model density,
e.g., Refs. 23 and 24. Such a procedure can be viewed as
J. Chern. Phys., Vol. 92, No.1, 1 January 1990
509
integrand smoothing. It introduces a density fitting basis
with possible limitations and consequent inaccuracies. The
numerical integration used here, however, facilitates the use
of density partitioning and a basis set free expansion into
atomic density multipoles which allow calculation of the
static potential.
If no three-dimensional integration is used to calculate
the exchange and correlation matrix elements directly, an-
other fitting is necessary to expand the exchange correlation
potential. The fitting necessarily relies on a three dimension-
al integration, and the fitting of the potential introduces er-
direct three dimensional integration to calculate exchange
and correlation matrix elements.
A. Variational basis sets
In principle, Omol could use any type offunction which
satisfies the regularity and normalizability conditions for
molecular orbitals. 25 For convenience of generating the basis
functions, which may also be symmetrized optionally, nu-
merically tabulated mUltipolar functions calculated in a
LOF spherical atom program are used. Thus, the present
approach can be considered a generalized LCAO type ap-
proach.
Dmol uses exact LOF spherical-atom orbitals in the ba-
sis set. This has the consequence that the molecule can be
dissociated exactly within LOF into its constituent atoms.
Necessarily, there is always net binding energy ofthe calcu-
lated molecular ground state with respect to the sum of exact
atomic total energies, even in a weakly bonded molecule.
This is as it should be in LOF, however, in other methods
there may be no molecular ground state bindirig energy with
respect to the sum of exact atomic energies because of basis
set limitations in the molecular calculation. If the molecular
basis sets fall short of attaining exact atomic energies for the
chosen model; LOF, HF, ... , there is room for basis set super-
position errors improving in fact on atomic energies. In the
present method, with its perfect basis set for the separated
atoms, generally, excellent description of weak bonds has to
be expected. The lack of inclusion of long range correlation
in the LOF currently prevents the accurate calculation of
van der Waals bonds, however. In this case an exponential
decrease of binding energy at long distances results as an
artifact ofLDF.
It is standard in any LCAO method to provide two basis
functions for each occupied atomic valence orbital. If fewer
trial functions are provided, variational flexibility beyond
basic promotion can only come from basis set superposition
effects which is generally not desired. If the present method
is used in the spirit of an LCAO method (with basis func-
tions kept fixed through the entire calculation), it is reasona-
ble to use a double set of numerically tabulated basis func-
tions. In previous publications this kind of basis set has been
referred to as "double basis." A more precise term would be
"double numerical basis" to be contrasted with double zeta
basis, where the radial functions are defined as Slater zeta
functions. In many respects the double basis has more in
common with linearized augmentation methods, which all
use a double set of numerical radial functions inside the muf-
510 B. Delley: Density functional for polyatomic molecules
fin tins. For B, C, N, 0, F atoms it turns out that augmenting
the functions from neutral atom by taking atomic valence
functions from the corresponding 2 + ions is generally close
to a fully optimized double basis sets. It is gratifying to see
that valence functions from 1 + or 1.5 + ions are almost as
good. This implies that the optimum does not come from a
sharply peaked function.
Recalling that this basis set is ideally suited to treat also
completely dissociated atoms, it may be assumed that it is
good on the entire energy hypersurface. An s function calcu-
lated in the potential of a nucleus with charge 1.3 is well
suited to complement the LDF hydrogen atomic s function.
The basis set can be significantly improved by adding higher
angular momentum valence polarization functions and also
by core polarization functions. Tables I(A) and I(B) illus-
trate the lowering of total energy obtained from adding more
functions 'to the basis set for N2 and H 2, respectively. Two
types of additional basis orbitals are considered here. First,
numerical orbitals of positive ions. Second, orbitals of hy-
drogenic atoms with a specified nuclear charge. The positive
ion charge is generated by depopulating the highest valence
orbitals. Other possibilities can readily be realized but are
not considered here. If more than one orbital is added, the
total lowering of energy is equal to the sum of the corre-
sponding table values. It should be noted that interchanging
the order of adding orbitals with different angular momen-
tum changes little on the values of the table. Interchanging
the order of adding orbitals of the same angular momentum
increases the value for lower ranked (more to the right)
TABLE I. (A) Total energy improvement A E"" for extended basis sets for
N 2 • The left-hand side of the table characterizes the generation of additional
orbitals in the basis set ordered according to angular momentum s,p,d, ... ; an
entry p+ denotes a valence orbital belonging to a NP+ ion, an entry q de-
notes a hydrogenic orbital belonging to nuclear charge q. The right-hand
side of the table gives total energy lowering in 10- s Ha for each additional
orbital set with respect to the total energy of the double basis consisting of
neutral atom orbitals and the valence orbitals from the 2 + ion. Missing en-
tries imply that no function with minimal improvement oflO- s Ha or more
has been found. If more than one orbital is added, the total lowering of
energy is equal to the sum of the corresponding table values. It should be
noted that interchanging the order of adding orbitals with different angular
momentum changes little on the values in the table. Interchanging the order
of adding orbitals of the same angular momentum increases the value for
lower ranked (more to the right) orbitals significantly while the sum re-
mains constant. The order was chosen to maximize the energy lowering for
the high ranked subsets. (1 Ha = 27.21165 eV = 627.53 Kcal/mol). (B)
Total energy improvement AE"" for extended basis sets for H 2 • Energy im-
provement beyond minimal basis consisting of Is LDF orbital of neutral
atom. Otherwise as 1 (A).
Type
AE,M
(A) s 1+ 5 3+ 4 179 97 13
P 1+ 5 3+ 4 121 17 14
d 7 5 4 5967 200 2
/ 11 9 14 226 134 2
g 17 18
h 24 6 the core eigenvalues, and has little to do with basis set im-
(B) s 1.3 0.3 2 7730 3
4 311
P
d 11 8 set size has slowed down on going to the two largest basis
J. Chern. Phys., Vol. 92, No.1, 1 January 1990
orbitals significantly while the sum remains constant. The
order in Table I was chosen to maximize the energy lowering
for the high ranked subsets. In order to make the overlap
matrix better conditioned and also to make the components
of the orbital expansion smaller, orbitals of the second and
further basis subsets are Schmitt orthogonalized against all
previous basis functions.
Even the double basis set may be considered as a large
basis especially for the larger molecules. Since core polariza-
tion is oflesser importance for energy properties and since in
the double basis essentially no core polarization is possible, a
frozen core approximation may not cause a significant loss in
accuracy but offer considerable savings in computer time,
especially for molecules containing many heavy atoms. In
the frozen core approximation, valence molecular orbitals
are Schmitt orthogonalized against core molecular orbitals.
Core eigenvalues are calculated by first order perturbation
theory when needed. The frozen core potential may be
viewed as a nonlocal angular momentum (I) and energy
dependent pseudopotential. Since the method has no prob-
lem in handling this most complicated pseudopotential, it
should readily be possible to incorporate standard pseudo-
potentials together with coreless basis sets. This would spare
the Schmitt orthogonalization step at the usual price in accu-
racy. This possibility has not yet been explored, however.
Table II shows eigenvalues and total energies of the N2
molecule for some basis sets. The bond length was set to 2.07
A. The column labeled "Exact" contains basis set free results
obtained with specialized methods for diatomic mole-
cules. 26,27 For comparison with the published results of Ref.
26 a local density functional with J.txc = - 0.73/2
{[3p(r) ]/1r}I/3 [a.u.] is used for all calculations on N 2. The
present results for N2 were calculated as if this were a general
polyatomic molecule. The label froz4 refers to a double basis
set with frozen core approximation where 4 means that four
functions participate in the trial function for the U g orbitals,
doubl5 refers to standard double basis containing a double
set of valence functions and a single set of core functions; d 6
denotes a basis augmented by a single set of d-polarization
functions; /12 contains two /-polarization functions and
two d-polarization function sets and additional sand p func-
tions, in fact all the functions in Table I(A) with a value of
more than 50X 10- 5 Ha (1 Ha = 27.21165 eV = 627.53
Kcallmol); gl9 includes 1=4 polarization functions as
well; and h 37 includes I = 5 polarization functions. Clearly,
convergence towards the exact result is demonstrated, the
best basis set produces errors in self-consistent eigenvalues
and total energy of ::::::0.000 03 Ha. The good accuracy of the
self-consistent eigenvalues testifies to the accuracy of the nu-
merical evaluation of the potential to be discussed in Sec.
III C. It should be mentioned here, however, that the model
charge density for h 37 was obtained by formula (16) and for
the other basis sets by formula ( 15) (a discussion follows in
Sec. IV). The improvement is especially clearly visible for
provement as the eigenvalues for the fixed potential testify.
It should be noted that the rate of convergence versus basis
sets. This is not too surprising, however, since the two large
 B. Delley: Density functional for polyatomic molecules 511

TABLE II. Energy eigenvalues and total energies for the nitrogen molecule with the Xa = 0.7 functional for dilferent basis sets. 11 ... indicates average
deviation of valence eigenvalues from the exact ones.I1ET indicates overestimation of total energy. All values are in hartree units.

Exact h37 g19 /12 d6 doubl5 froz4

For fixed potential

ug - 13.98107 - 13.98107 - 13.98107 - 13.9810 - 13.981
Uu - 13.97966 - 13.97966 - 13.97966 - 13.9797 - 13.980
ug - 1.00721 - 1.00721 - 1.007 14 - 1.0053 - 1.002 - 1.002
Uu - 0.460 71 - 0.460 71 - 0.46071 -0.4606 -0.460 -0.460
1fu - 0.404 21 -0.40420 -0.40413 - 0.4035 -0.395 - 0.395
ug - 0.35005 -0.35003 -0.34992 -0.3486 -0.346 -0.346
!l. av8 0.00000 0.00000 0.00007 0.0010 0.006 0.006

Self-consistent
ug -13.981068 - 13.98104 - 13.98149 - 13.981 89 - 13.9821 - 14.000
Uu - 13.979659 -13.97963 -13.98008 - 13.98048 -13.9807 - 13.999
ug - 1.007215 - 1.007 19 - 1.007 16 - 1.00685 - 1.0080 - 1.069 - 1.070
Uu - 0.460725 - 0.46070 - 0.46064 - 0.460 14 - 0.4587 -0.452 - 0.452
1fu -0.404235 -0.40420 -0.40412 - 0.403 62 -0.4027 - 0.421 - 0.421
ug -0.350059 - 0.35003 - 0.35000 -0.34945 - 0.3481 -0.344 -0.344
b. avg 0.000000 0.00003 0.00009 0.000 56 0.0013 - 0.016 - 0.016
ET - 108.346 609 - 108.34658 - 108.34650 - 108.34584 - 1083363 - 108.277 - 108.276
I1ET 0.000000 0.00003 0.000 11 0.000 77 0.0103 0.070 0.070

basis sets were constructed as tempered basis sets, while the TABLE III. Energy eigenvalues, total energies (E T) and C-C bond length
/12 and d 6 were partially optimized by condensation and (dee) for ethane, ethylene and acetylene with the Hedin-Lundqvist func-
nonlinear variations. tional (Ref. 5) for dilferent basis sets. Energies are in hartree units, distance
in angstroms.
Table III shows a comparison of eigenvalues, total ener-
gy, and C-C bond length for ethane, ethylene, and acetylene doubl5 d6 fl2
obtained from double basis sets d 6 and /12. For the hydro-
gen atom the double basis set contains two s functions, the d 6 Ethane
includes another s function and p-polarization functions, a. - 9.76150 - 9.760 54 - 9.763 23
a2 - 9.76138 -9.76041 - 9.76311
and the /12 basis set includes yet another s function and d- a. - 0.687 90 -0.68204 -0.68244
polarization functions, in fact, all functions represented in a2 -0.55458 -0.54928 -0.54982
Table I(B). Ethane was calculated in a configuration with a e -0.40160 -0.39658 - 0.39679
a. -0.32888 -0.32443 - 0.32422
torsion angle halfway between the staggered and the eclipsed
e - 0.306 87 - 0.303 25 - 0.303 58
configuration. Tetrahedral angles were used for the carbon-
hydrogen bonds. In ethylene, the LHCH was set to 116.8°. ET -79.07858 -79.10413 -79.10760
dee 1.54846 1.53905 1.53789
The CH bond length was fixed at 1.09 A. Total energy im-
provement on going from d 6 to /12 is more pronounced for Ethylene
the shorter multiple bonds. Total energy for the double basis ag - 9.785 21 - 9.783 41 - 9.78434
is quite uniformly higher than the ones for the extended basis btu -9.78448 - 9.782 69 - 9.783 61
sets. This testifies to the good quality of the double basis set. ag -0.69814 - 0.691 06 -0.69006
Bond lengths are quite consistently slightly too long for the b. u - 0.528 62 -0.52471 -0.52360
double basis by ::::;0.01 A. This must be expected since this b3u -0.43468 -0.42807 -0.42678
ag -0.38816 -0.38148 -0.37989
basis set is perfectly adapted for large separation and less
b2g -0.32042 - 0.31695 -0.31556
well so for short separations. The uniformity of the deviation b2u - 0.259 96 -0.25781 - 0.25802
is contrary to bond length errors in a HF approach, where
understanding is complicated by the admixture of ionic ET -77.86404 -77.88607 -77.89042
dee 1.33804 1.33146 1.32678
states. The too short C-C bond lengths calculated here for
ethylene and acetylene are likely to be consequence of the too Acetylene
long CH distance, which was fixed beforehand, artificially ug - 9.80145 - 9.799 38 - 9.796 80
strengthening the C-C bond. Uu - 9.799 06 - 9.797 05 -9.79444
ug -0.69462 -0.68619 - 0.68209
Uu - 0.51493 -0.513 45 - 0.510 39
ug -0.45725 - 0.45102 -0.44710
B. Numerical integrations 1fu -0.27847 - 0.27519 - 0.27359
Three-dimensional numerical integration of the matrix ET -76.62733 -76.64330 -76.65200
elements occurring in the eigenvalue equation must suitably dec 1.21532 1.20861 1.205 18
treat the fast oscillations of the molecular orbitals near the

J. Chern. Phys., Vol. 92, No.1, 1 January 1990

512 B. Delley: Density functional for polyatomic molecules
nuclei having a cusp behavior at the point nucleus. In general
polyatomic molecules, it is difficult to devise a global inte-
gration scheme that avoids integrating over the cusps al-
though the location of the cusps is perfectly well known.
Since a multipolar expansion of the molecular orbitals is rel-
atively rapidly convergent near the nucleus where the expan-
sion is centered, efficient integration methods can be con-
structed for the portion of the integrals near the nuclei. If
multipolar expansions are done inside Wigner-Seitz cells,
treatment of the boundaries for the integration is difficult. If
the expanded function is forced to vanish outside the cell, to
transform to integrals running over all space, very slowly
convergent expansions result. The concept of partition func-
tions leads to faster converging expansions much better suit-
ed for numerical integrations of multipolar functions. Parti-
tion functions have been used previously to divide molecular
charge densities into atomic charges that do not depend on
orbital basis sets. 28 In the present context the partition func-
tions are constructed by
where a is an atom index and ga (r - Ra) is a function
which typically is large for small arguments and small for
large arguments. Simple examples are Pa (Irl), Vlrl,
etc., wherepa is a spherical atomic density. Currently, pre-
ferred choices are the functionsga = p~, ga = Pair, andga
= Pa (e(ralr) - 1 - rolr). These partition functions auto-
matically depend on atom sizes viapa and have given consis-
tently good results for all sorts of compounds including he-
teronuclear compounds containing heavy transition metals
and light elements including hydrogen. The last function
leads to continuous vanishing of all partition functions and
all their derivatives on all nuclei where they are not centered.
Different partition functions, in which atomic radii need to
be inserted as parameters, have been presented recently by
Becke. 29 The partition function is used to rewrite integrals
over all space for functions 1= "pi (r) h (r)"pj (r), etc. as
~ J
l(r)Pa(r)d 3r= J l(r)d 3r, (4)
where the partitioned integrals have now a rapidly conver-
gent multipolar expansion around sites. The site integrals
can be decomposed into radial integrals and angular inte-
grals to be evaluated numerically. Obviously, there are many
methods to achieve efficiently sufficient accuracy for the ra-
dial integrals. The angular integrals can be done by Gaussian
quadratures on a sphere given in Stroud. 30 Higher order
methods have been described by Lebedev 31 •32 and Kon-
yaev. 33 Arbitrarily high order can be obtained readily by
spherical product Gaussian integration. An important point
is, however, that not very high accuracy needs to be obtained
for the matrix elements, as will be seen from an error analy-
sis. To be specific it should be mentioned that typically one
part in a thousand for the least accurate overlap matrix ele-
ments is presently considered sufficiently accurate for most
applications of this method. This order of accuracy is ob-
tained with about a thousand integration points per atom.
If the eigenvalue equation is expressed in the basis of the
J. Chern. Phys., Vol. 92, No.1, 1 January 1990
Ritz eigenfunctions calculated without numerical integra-
tion errors, one has
L (Ei~ij + N [ihj] )(~j,k + Cj,k)
j
= (E k +Ei) L (~ij +N[ij])(~j,k +Cj,k)' (5)
j
where N[ihj],N[ij] are the numerical truncation errors oc-
curring in the approximate evaluation of f "pi(r)h(r)
X"pj (r)d 3r and f "pi (r)"pj (r)d3 r. The Ci,k are the expansion
coefficients of the error part for a molecular orbital belong-
ing to eigenvalue E k' and El is the error for the eigenvalue E k'
To lowest order, the set ofEq. (5) is
El =N[khk] -EkN[kk] (6)
and
N [ihk ] - EkN [ik ] N [kk ]
Ci,k = , Ck,k = - . (7)
Ek - Ei 2
The first order equation for Ck,k results from the normaliza-
tion of the single particle functions. The first order errors Ci,k
can be brought down to a second order error, except for the
normalization error Ck,k' by using the same numerical inte-
gration for the f "pi"pjd 3r and the f "pih"pjd3r as has long
been done in the discrete variational method. 34 Since
(8)
where ifJk is a smooth residual function nonvanishing be-
cause of basis set truncation
El = EkN [kifJk] (9)
and
N [iifJk]
Ci,k = , i#k (10)
Ek - Ei
should vanish to first order with vanishing basis set trunca-
tion error. This mixed second order error term is a character-
istic of the discrete variational methods. On the basis ofEq.
(9) one should expect the eigenvalue errors to scale with the
square root of the total energy error due to basis set trunca-
tion. Table IV gives an illustration of such errors relative to
the values of Table II due to reduced integration accuracy on
the matrix elements of the eigenvalue (lc). In particular the
ten times greater sensitivity of eigenvalues with doubl5 basis
as compared to the ones from the 112 basis is in line with
these expectations. While accurate integrations of Table II
used 1024 points total, integrations for Table IV total to 722
points with reduced accuracy in the angular part used. The
relatively high number of integration points required in
these examples appeared necessary to make high order con-
vergence set in also for the very large basis sets.
C; Model charge density and Coulomb potential
Since in the LDF equations exchange and correlation
are treated by a local effective potential Ilxc (r) there is no
need to calculate four center terms. However, the accurate
calculation of the electrostatic potential requires an expan-
sion of the density per) to be performed efficiently. In
LCAO-like methods, the charge density is fitted typically to
multipolar densities attached to the atoms. In the present
 B. Delley: Density functional for polyatomic molecules 513

TABLE IV. Energy eigenvalue errors A(E,) for the nitrogen molecule due and
to reduced accuracy integrations. (Accurate integrations of Table II used
1024 points total, integrations for this table with reduced accuracy in the V(r) = 41T.,f4ii ') 1
angular part used 722 points total.) Values are in 10- 5 hartree units. a7'm ~2/ + 1
/12 d6 doublS X Y/,m (r - Ra) Va,/,m (Ir - Ra I)· (12c)

ug 0 -1 0 Values of Va,/,m on tabulation points are obtained analytical-

U. 0 0 0 ly from the spline functions Pa,/,m' Values of Va,/,m at inter-
ug 0 -1 mediate points are obtained by spline interpolation. Figure 1
u. 0 -1 shows the radial functions Pa,/,m and Va,/,m for the nitrogen
1T. 0 -1 -1 molecule calculated with basis h 37 (see Table II) calculated
ug 0 0 -1 according to Eq. (16) below. Clearly, the 1= 1 component
1:, n,A(~,) 2 -4 -10 shows a nodal structure not easily anticipated from S-P hy-
bridization. It is also notable that the / = 3 component is still
of comparable magnitude as the dipolar component. These
almost exact deformation density functions suggest that
method Poisson's equation is solved in a completely numeri- flexible basis sets would be needed for accurate density fit-
cal (nonbasis set) approach. ting.
The point of the method is to first partition the molecu- For a brief discussion of the !!.leaning of self-consistency
lar charge density into atomic densities with a rapidly con- it is useful to define an operator P with the property
vergent multipolar expansion. In a second step, the atomic p(r) = Pp(r) (13)
densities are decomposed into multipolar components by
projection methods. In a third step, the static potential is
calculated for each component and reassembled later on. 35
The partitioning approach, Eqs. (3) and (4), used for
the calculation of three dimensional integrals can be used to [au] 1-----------------1
obtain densities attached to the atoms. The best partition
function for atomic densities with rapid convergent multipo-
lar expansions is not necessarily the same as the one used for
the decomposition of the three dimensional integrations.
These atomic densities can then be decomposed into their
multipolar components by projection rather than by fitting
1
Pa,/,m (s) = ~ ~
-y41T-y2/ + 1

X f Y/,m (r - Ra )Pa (r)p(r)d!l.

)s=lr-Ral
(11 )
Because ofthe necessary angular truncation the model den-
sity
p(r) =.,f4ii

X a~m ~2/ + I Y/,m (r - Ra )Pa,/,m (Ir - Ra I)

(12a)
is not equal to the orbital density p. Sampling the projected
density components at many radii gives a flexibility to the
charge density representation which goes much beyond the
one obtained with least-squares fitting procedures. The den-
sity components Pa,/,m at a radius not equal to one of the
sampling radii are defined to be the cubic spline interpola-
0.0 1.0 2.0 3.0 r lao]
tion function. The potential and its first radial derivative for
each component are calculated using the Green's function of FIG. 1. Radial charge and radial potential for N 2• Charge represented as
the Laplacian rpt, .. _o(r) full line [Eq. (11)]. electrostatic potential dashed [Eq.
( l2b) ]. The lowest pair of lines represent the monopole component. fol-
Va,/,m (r) = {r-/- I
Sa' Pa,/,m (X)x/+ 2 dx lowed by the higher angular components in ascending order. The curves
have been displaced by an arbitrary shift for clarity. the zero for each pair of

+,.t 1 00

Pa,/,m (x) 1 - / dX} (12b)

curves is defined by the value of rp~m _ 0 (r) at r = O. The charge and poten-
tial functions for the spherical atom are subtracted out in the monopole
component.

J. Chern. Phys., Vol. 92. No.1, 1 January 1990

514 B. Delley: Density functional for polyatomic molecules

A
and an operator S which generates the orbital density p ac-
cording to Eq. (1 b) where the orbitals have been calculated
from Eq. (la) modified to
{ ; : a+ Vs(r,jj) + ILxc." [p(r)]} <I>;.,,(r)
= Ej<l>j." (r). (1d)
S
This defines an operator which produces an orbital density
p from a density p
A
per) =Sp(r). (14)
Equations (13) and (14) have to be solved self-consistently.
This can be realized in an iterative procedure where the out-
put density of one iteration is used, in principle, to generate
the potentials to be used in the next iteration. For stability
reasons damping has to be introduced in this procedure, but
this will not be discussed here. Expressed in the model den-
sity, the self-consistency condition is
AA
0= (PS-1)p(r) (15)
which is the usual self-consistency criterion. True self-con-
sistency would imply thatp = p but this condition cannot be
fulfilled exactly if a truncated model density is introduced.
However, it is possible to investigate the condition for self-
consistency
A
in the function space spanned by Eq. (12a).
Since P is not an indempotent operator an improved self-
consistency criterion
(16)
O=P(S-l)p(r)
can be made, where residue p = p is zeroed in the function
space of p. This procedure amounts to an iteratively defined
partitioning of p which can reduce the need for high angular
components. It is noted that the rate of convergence slows
down significantly beyond the level reached by the approach
shown in Fig. 2(b), the residual for the I = 3,2 charge den-
sity model is reduced by a factor of - 5 (note change in
contour spacing) and has a completely different character.
The I = 4,3 model yields a further improvement of the resid-
ual. However, it appears that the ring-shaped positive resid-
ual will only disappear with a model incorporating rather
high angular components. These high angular components
produce a potential which falls off with a high power of the
radius r towards the expansion center and a high inverse
power of r towards infinity. The resulting local perturbation
of the potential will only weakly affect calculated integral
properties. The second iteration scheme has been used in a
calculation with I = 3,2 truncation in Fig. 2(d). Clearly this
residual is much improved as compared even to the 1=4,3
model. Approach (16) with a truncation at 1=4,3, not
shown, leads to a residual barely exceeding 0.004 e/A at its
peak. Since the second iteration scheme currently needs sig-
nificantly more iterations to obtain self-consistency, the first
scheme is more appropriate for most calculations. Table V
shows the influence of model truncation on the total energy
ofN 2 • In most applications a truncation at the 1= 2 level will
be sufficient.
D. Effect of charge truncation errors
In the following, the effect of charge truncation errors
on energy eigenvalues and total energies is discussed. The
approximate eigenvalue equation expressed in the basis of
variational eigenstates of the exact eigenvalue equation is
I (EjO jJ + VjJ ) (OJ.k + Cj,k)
j

( 15). This does not come as a surprise. As in a fitting proce-

dure, a density far from the site where a particular multipole = (Ek + El) I (OJ.k + Cj.k), (17a)
j
is attached may influence this multipole in Eq. (16) but not
where V;J is the matrix element of the error due to the differ-
in (15). Unlike a fitting procedure, Eq. (16) does not re-
ence op = PSCF - Pexact between the model density and the
quire solution of linear equations and can be solved with
exact self-consistent orbital density (SCF represents self-
minimal overhead in an iterative procedure. Equation (16)
consistent field). op originates from basis set truncation, nu-
requires more iterations than Eq. (15) because of the slower
merical integration, and model expansion truncation as well
final rate of convergence. The residual density, i.e., the dif-
as lack of full self-consistency. It should be noted that the
ference between the total orbital density and the model den-
origin of op does not matter. To first order one finds
sity, especially near midbond, can however, be much im-
proved. It should also be noted that the self-consistent radial (17b)
densities resulting from Eq. (16) are less dependent on the
choice of partition function.
At this point it is interesting to ask whether it is possible
to define functions ga beforehand which produce the density
resulting from Eq. (16) when used in Eq. (15). Empirically, Thus, first order errors are produced both on energy eigen-
it was found that thefunctionga (r) = p(r)/rleads to better values and molecular orbitals. The orbital errors translate
agreement than the functions preferred for the integral parti- into first order errors of the density. To first order in op the
tioning. error of the approximate total energy
Figures 2( a)-2 (d) demonstrate the improvement of the
residual density for the ethane molecule, at the /12 basis
level, that is possible at a given level of angular truncation of
E, = ~ Ek + ~ f p,,(Exc - [Lxc." - Ve/2)d 3 r
p. The residual density for the model density truncated at
quadrupoles for carbon and dipoles for hydrogen (/ = 2,1) +I ZaZp (2b)
still has significant magnitude. The shape of the residual IRa -Rpl
a.p
near carbon is suggestive of an 1= 3 component. Indeed, as compared to the total energy (2a) is

J. Chern. Phys., Vol. 92, No.1, 1 January 1990

 B. Delley: Density functional for polyatomic molecules 515

lao] lao] i
i
(a) (e) -
,, .. ... , i
(-', ,, \

1.5 1.5
\ . \I
\ .
i..I

_0 ____ -"
Y Y

-""'- '-"'-
.........
-1.5 -1.5 . ",.
\ ,.,.. ,
.........

-
...... ...... .-' _.... / i
.\ i
\ 0_.,
\
,
, , '., ,
"i
;"-,-",,

!
-1.5 X 1.5 lao] -1.5 X 1.5 lao]

..... ,
lao] / ;
\" /
, ..... ,'-. i
,
",~- ....... .,,0_.,. i
/ j
( . . .i \
(
.\ / , i \
1.5 ._._.i
(bl
.\
\
.I
. 1.5 ..;' ,.I
/
I
'-."
I·..... i \
\
\
, ....
J (dl "
'-'
;'
....... ~.i '1 i
\'-.:.-.-.~,...-.
.",'
~.,c:~- /
,.
.,.
,. ,

i
y y "._-
.",.- ....... __ .-._.
-.
,.- .............. _.
i
i
-1.5
-1.5
",, ...... /

-1.5 X 1.5 lao] -1.5 X 1.5 lao]

FIG. 2. Residual density for ethane eSp = p - p. see the text. The model hasa torsion halfway between the staggered and eclipsed conformation to show more
details of the density. (a) An expansion including 1= 2 components for carbon and I = I for hydrogen; (b) includes Ie = 3 and IH = 2; (c) Ie = 4 and
IH = 3; (d) includes Ie = 3 and IH = 2 but is calculated usingEq. (16). Contour spacing is 0.05 elA.l in (a) andO.OI elA.lin (b)-(d). Dash-dotcurve is zero
contour. Full contours show positive density eSp.

numerical evaluation· of the integral :I q f Pq(Exc

A=~El+~f 8[pq(Exc - fixc - Ve/2)d 3 r, since this leads in principle to a first
3
order error N [pq(Exc - P,XC,q - Ve/2)]. Controlled accu-
-P,XC,q - Ve/2)] d r=O. (18)
racy integrations using partitioning as described in Sec. III B
This follows from the Seitz relation 8 (pExc ) = p,8p. From may be used to evaluate this particular integral as precisely
Ecf. (17a) and the definition of Vlc,k one has :Ire El as desired. The integration scheme proposed recently by
= :Ifc Vlc,k = f p8(P,xc + Vel. Care must be taken in the Boerrigter et al. 36 appears also well suited for this purpose.

J. Chern. Phys., Vol. 92, No.1. 1 January 1990

516 B. Delley: Density functional for polyatomic molecules

TABLE V. Energy eigenvalues and total energies for the nitrogen molecule (Xa = 0.7) for different trunca-
tions of the model charge density p. A basis oftype g35, see the text, is used. All values are in hartree units.

[max

5 4 3 2 0

For fixed potential, truncated to Imax

ug - 13.98107 - 13.98033 - 13.977 77 -13.97695 - 14.013 93 - 14.00900
u. -13.97966 -13.97892 -13.97636 - 13.97554 - 14.01250 -14.00756
ug - 1.00721 -1.00710 - 1.00653 - 1.00708 - 1.044 67 - 1.043.38
u. - 0.46071 - 0.46063 - 0.46021 -0.46002 - 0.477 38 - 0.47445
tT. - 0.40421 - 0.40418 -0.40390 - 0.403 77 - 0.415 35 - 0.412 76
ug - 0.35005 - 0.349 92 -0.34932 -0.34938 -0.37911 -0.37674

Self-consistent
u. - 13.981 51 - 13.98126 - 13.97995 - 13.97933 - 13.99290 - 13.98949
u. - 13.98010 - 13.97985 - 13.97854 -13.97792 - 13.99147 - 13.98806
ug - 1.00719 - 1.007 16 - 1.00705 - 1.00748 -1.03082 -1.03039
u. - 0.46071 -0.46071 - 0.46066 - 0.460 61 - 0.468 41 - 0.46611
tT. - 0.404 19 - 0.40421 - 0.404 27 -0.40420 -0.40534 - 0.403 22
u. -0.35005 - 0.35002 - 0.34983 - 0.349 97 -0.36919 - 0.367 63
ET - 108.34657 -108.34656 - 108.34654 -108.34649 - 108.35773 - 108.35901

IV. DISCUSSION FOR COMPUTATIONAL EFFORT
In order to discuss the approximate number (to leading
order) of operations required for each part of the algorithm
presented here, it is useful to define numbers N that grow
linearly with molecule size, and numbers M that remain con-
stant with molecule size. The algorithms have been imple-
mented in a way permitting full advantage to be taken of any
eventual symmetry of the cluster under consideration. Sym-
metry reduces the number of floating point operations in the
cubic parts approximately by N} and by N r N G' where N r is
the number of representation partners and N G is the order of
the group. Since most larger molecules and clusters cannot
be expected to be symmetric, the following discussion as-
sumes the nonsymmetric case.
To set up Ns basis functions of the integration mesh
consisting of Np points NsNpM; floating point operations
are necessary, where M; is the number of operations per in-
terpolation and symmetrization. M; is a small number of
order 10. If core orthogonalization is done, the total number
of expansion functions is split into core functions Nsc and
"valence" functions Nsv' and the number of operations re-
quired is 2Nsv N sc Np- The calculation of the overlap and the
Fock matrix requires 3Nsv (Nsv + I)Np operations. If no
core orthogonalization was done N sv = N s. The following
computational steps are repeated in each iteration in the self-
consistency procedure. The number of iterations Mit de-
pends on the kind of system studied and the requirement on
self-consistency. Many molecules converge in seven itera-
tions with the present procedures. To calculate the potential
on the integration mesh from a model density characterized
by N m multipolar components, M;tNm Np~ operations be-
come necessary, where M j is the number of operations per
interpolation for densities and the potential together. The
integration of the potential matrix requires MitNsv (Nsv
+ I )Np operations. The number of operations to diagona-
lize the generalized eigenvalue problem is the standard
MitN ;vMd where Md can be reduced by not calculating non-
occupied eigenvectors. Reconstruction of the density needs
another 2Mit N so N. vN p operations. Here Nso denotes the
number of occupied valence orbitals, always a fraction of
N sv . The total number of operations for large molecules is
thus dominated by the cubic growth NpNsv [ Mit (Nsv
+ 2Nso + I) + 3(N.v + 1)] + N;vMd' Since Np
::::: 100 N s , density reconstruction and potential matrix set-
up are responsible for the majority of floating point opera-
tions.
If it is realized that the basis functions in an LCAO
approach have finite range because ofthe exponential decay,
the algorithm can be modified such that only a limited num-
ber of integration points contribute (to within roundoff er-
ror) to each integral. In that case, Np becomes a constant
Mp ::::: 104 and only the small cubic term from the diagonali-
zation remains. The numerical effort necessary to obtain
self-consistent local density functional for very large mole-
cules becomes proportional to the one for most simple se-
miempirical MO calculations. The proportionality factor is
mit [NsvINsv(MO) ]3.
V. CONCLUSIONS
An approach to quantum chemical calculations has
been presented, preliminary versions of which have been
used for years by the author and others on small and large
molecules, organo-metallic compounds, polymers and met-
al clusters. Since the calculation of matrix elements is done
fully numerically any variational basis set of interest can be
used in principle. A practical choice of basis set with excel-
lent initial convergence properties was presented. The three-
dimensional iritegration is done in a novel way which inher-

J. Chern. Phys., Vol. 92, No.1, 1 January 1990

 B. Delley: Density functional for polyatomic molecules
ently keeps track of atomic properties and of cusp-like
behavior of the integrand at the nuclei. It was shown how the
static potential can be calculated efficiently with high accu-
racy without recourse to a charge fitting compromise. Sym-
metry has been mentioned under speed considerations only,
but is being used routinely where applicable. Error charac-
teristics and computational speed of the approach have been
discussed. Error characteristics and computational speed of
the approach have been discussed. The method can be used
to give efficiently highly accurate solutions ofthe LDF equa-
tions for small molecules as well as to yield efficiently results
of useful accuracy on rather large molecules and clusters.
ACKNOWLEDGMENTS
The author wishes to thank George Fitzgerald and Ar-
thur J. Freeman for reading the manuscript and constructive
suggestions. He is indebted to Jan Andze1m for implement-
ing a fast algorithm on Cray computers and for benchmark-
ing.
IC. C. J. Roothaan, Rev. Mod. Phys. 23, 69 (1951).
2G. G. Hall, Proc. R. Soc. LondonSer. A 205, 541 (1951).
3p. Hohenberg and W. Kohn, Phys. Rev. B 136, 864 (1964).
4W. Kohn and L. J. Sham, Phys. Rev. A 140, Al133 (1965).
5L. Hedin and B. I. Lundqvist, J. Phys. C 4, 2064 ( 1971 ).
"U. v.Barth and L. Hedin, J. Phys. C 5,1629 (1972).
70. Gunnarsson and B. I. Lundqvist, Phys. Rev. B 13, 4274 (1976).
8E. Wimmer, A. J. Freeman, C.-L. Fu, P.-L. Cao, S.-H. Chou, and ij. Del-
ley, in Supercomputer Research in Chemistry and Chemical Engineering,
ACS Symposium Series No. 353, edited by K. F. Jensen and D. G. Truhlar
(American Chemical Society, Washington, D.C., 1987), p. 49ff.
9R. O. Jones and O. Gunnarsson, Rev. Mod. Phys. 61, 689 (1989).
J. Chem. Phys., Vol. 92, No.1, 1 January 1990
517
100. Gunnarsson, J. Harris, and R. O. Jones, Phys. Rev. B 29, 703 (1984).
IIp. J. Braspenning, R. Zeller, A. Lodder, and P. H. Dederichs, Phys. Rev.
B 29, 703 (1984).
12E. J. Baerends and P. Ros, Int. J. Quant. Chern. Symp. 12,169 (1978).
13B. Delley, D. E. Ellis, andA. J. Freeman, Phys. Rev. Lett. SO,488 (1983).
14S._H. Chou, A. J. Freeman, S. Grigoras, T. M. Gentle, B. Delley, and E.
Wimmer, J. Am. Chern. Soc. 109, 1880 (1987); J. Chern. Phys. 89, 5177
(1988).
15Ye Ling, A. J. Freeman, and B. Delley, Phys. Rev. B 39, 10144 (1989).
I"B. Delley, Chern. Phys. 110, 329 (1986).
17p. Moeckly, D. Schwarzenbach, H.-B. Biirgi, J. Hauser, and B. Delley,
Acta Cryst. B 44,636 (1988).
18B. Delley (to be published).
19M. Levy, Proc. Nat!. Acad. Sci. 76, 6062 (1979).
20J. C. Slater, Phys. Rev. 81, 385 (1951); 82,538 (1951).
21L. Fritsche, Phys. Rev. B 33,3976 (1986).
22U. v Barth, in Many-Body Phenomena at Surfaces, edited by D. C. Lan-
greth and H. Suhl (Academic, New York, 1984), p. 3ff.
23B. Delley and D. E. Ellis, J. Chern. Phys. 76,1949 (1982).
24B. I. Dunlap, J. W. Connolly, and J. R. Sabin, J. Chern. Phys. 71, 3396
(1979).
25F. W. Averill and D. E. Ellis, J. Chern. Phys. 59, 6412 (1973).
2"D. Heinemann, B. Fricke, and D. Kolb, Chern. Phys. Lett. 145, 125
( 1988).
27L. Laaksonen, D. Sundholm, and P. Pyykko, Int. J. Quant. Chern. 27, 601
(1988).
28F. L. Hirshfeld, Theor. Chim. Acta B 44, 129 (1977).
29A. D. Becke, J. Chern. Phys. 88, 2547 (1988).
30A. H. Stroud, Approximate Calculation of Multiple Integrals (Prentice-
Hall, Englewood Cliffs, NJ, 1971).
31V. I. Lebedev, Zh. Vychis!. Mat. Mat. Fiz. 15,48 (1975).
32V. I. Lebedev, Zh. Vychis!. Mat. Mat. Fiz. 16, 293 (1977).
33S. I. Konyaev, Mat. Zametki 25,629 (1979).
34D. E. Ellis and G. S. Painter, Phys. Rev. B 2, 2887 (1970).
35This approach is similar to the method of A. D. Becke and R. M. Dickson,
J. Chern. Phys. 89, 2993 (1988). The present method has been used to
calculate the electrical field at the nucleus in Ref. 16 and is routinely being
used in the selfconsistent calculations since 1987.
36p. M. Boerrigter, G. te Velde, and E. J. Baerends, Int. J. Quant. Chern. 33,
87 (1988).