JChemPhys 99 4597

Relativistic regular two-component Hamiltonians
E. van Lenthe, E. J. Baerends, and J. G. Snijders

Afdeling Theoretische Chemie Vrije Universiteit, De Boelelaan 1083, 1081 HV Amsterdam,
The Netherlands
( Recei ved 22 February 1992; accepted 1 June 1993)
In this paper, potential-dependent transformations are used to transform the four-component
Dirac Hamiltonian to effective two-component regular Hamiltonians. To zeroth order, the
expansions give second order differential equations (just like the Schrodinger equation), which
already contain the most important relativistic effects, including spin-orbit coupling. One of the
zero order Hamiltonians is identical to the one obtained earlier by Chang, Pelissier, and Durand
[Phys. Scr. 34, 394 (1986)]. Self-consistent all-electron and frozen-core calculations are
performed as well as first order perturbation calculations for the case of the uranium atom using
these Hamiltonians. They give very accurate results, especially for the one-electron energies and
densities of the valence orbitals.
I. INTRODUCTION
Relativistic effects are important in the study of heavy
elements. Instead of the Schrodinger equation, one now
has to solve the Dirac equation, which involves a four-
component Hamiltonian. Fully relativistic calculations are
no more complicated than nonrelativistic ones, as far as
integral evaluation is concerned, but they are very time
consuming. The dimension of the secular problem will be
very large since there are four components and many basis
functions are required. If one wants to use a variational
technique, then one has to make sure that no spurious
solutions appear. This can be done using the so-called ki-
netically balanced basis sets,I-3 but then one needs in prin-
cipal different basis functions for the small component than
for the large component of the Dirac spin or, which in-
creases integral evaluation time and storage requirements.
An attractive alternative is to transform the Dirac
Hamiltonian to a two-component form. Standard ap-
proaches are the elimination-of-small-components (esc)
and the Foldy-Wouthuysen (FW) transformation. We re-
fer to the recent discussion by Kutzelnigg
4
for a detailed
exposition of the various approaches and the difficulties
that arise in the form of divergent terms and singularities
at r-O. The difficulties connected with the Foldy-
Wouthuysen transformation have for instance been inves-
tigated by Moss et al.
5
-
8
and Farazdel and Smith
9
have
criticized the mass-velocity term, pointing out that for
large momenta (p> a-I), the mass-velocity term - (a
2
/
8 )p4 is not obtained unless one uses the expansion of the
square-root operator for the kinetic energy outside its ra-
dius of convergence. The source of the difficulties is that
the expansions that are being used implicitly or explicitly
rely on an expansion in (E- V)/2c
2
, which is invalid for
particles in a Coulomb potential, where there will always
be a region of space (close to the nucleus) where (E- V)/
2c
2
> 1 and the expansion is not valid. This was already
noticed by Gollisch and Fritsche
lO
in their work on a scalar
relativistic approximation to the atomic Dirac equation,
essentially equivalent to the one of Koelling and Har-
mon.
11
In this paper, we consider expansions which are appro-
priate in the sense that the expansion "parameter" is < 1
over all space. In Sec. II, we first discuss expansions in the
case of classical relativistic mechanics in a Coulomb poten-
tial. This reveals the essential error in the standard expan-
sions as well as providing a simple remedy. In Sec. III, we
demonstrate that exactly the same error is made in the
standard derivations (both esc and FW) of the Pauli
Hamiltonian. It is shown that one can avoid this error,
leading in both the esc and FW methods to a two-
component relativistic Hamiltonian that is, at least in ze-
roth order, variationally stable, and contains similar rela-
tivistic corrections as are present in the Pauli Hamiltonian,
but in a regularized form. One of the two slightly different
Hamiltonians we obtain (depending on approximations
made) turns out to be identical to the zeroth order Hamil-
tonian derived earlier by Chang, Pelissier, and Durand us-
ing the theory of effective Hamiltonians.
12
We therefore
call it the (zero order) relativistic Chang-Pelissier-
Durand Hamiltonian EfoPD. The performance of the de-
rived Hamiltonians is investigated in self-consistent atomic
calculations in Sec. IV. They perform exceedingly well;
valence orbital energies especially are in much better agree-
ment than the energies obtained with the standard Pauli
Hamiltonian. For deep core levels, the error in the energy
is of the order E2/2c
2
and is still sizable in an absolute
sense. The first order Hamiltonian corrects this deficiency
in a first order perturbation calculation. The advantages of
the present formulation are that, in contrast to the Pauli
Hamiltonian, a variationally stable two-component Hamil-
tonian has been obtained that includes relativistic effects to
a high degree of accuracy. It is important that the relativ-
istic effects can be treated self-consistently, since this has
been shown to be important in very heavy elements such as
actinides, cf. Refs. 13 and 14. This advantage is shared
with the most successful two-component relativistic
method to date, the no-pair formalism with external field
projectors. This method has been developed for atomic and
molecular calculations by Hess
15
on the basis of theoretical
work by Sucher
16
and Douglas and Kroll.
17
A density-
functional implementation has been provided by Knappe
and Rosch.18 These schemes rely on momentum space
J. Chem. Phys. 99 (6),15 September 1993 0021-9606/93/99(6)/4597/14/$6.00 1993 American Institute of Physics 4597
Downloaded 28 Mar 2011 to 130.37.129.78. Redistribution subject to AIP license or copyright; see http://jcp.aip.org/about/rights_and_permissions
4598 van Lenthe, Baerends, and Snijders: Relativistic regular two-component Hamiltonians
evaluation of integrals and require the assumption of com-
pleteness of the finite basis sets employed in practical cal-
culations. It is an advantage of the present simpler ap-
proach that the required matrix elements can easily be
evaluated without further approximations in schemes that
rely on 3D numerical integration (see e.g. Refs. 19 and
20), making this method very straightforwardly applicable
to molecules.
II. CLASSICAL RELATIVISTIC MECHANICS AND THE
COULOMB POTENTIAL
In this section, the classical expression for the relativ-
istic energy of a particle in a potential is expanded in sev-
eral ways to prepare for similar expansions in relativistic
quantum mechanics. We will see that if one expands the
energy expression in c-
1
, as is usually done, this will give
rise to some problems if the momentum of the particle is
too large. For Coulomb-like potentials, there are always
regions where this is the case. A potential dependent ex-
pansion can be found, which is well behaved even if the
momentum of the particle is large. This expansion will be
considered after we have explained the shortcomings of the
c-
1
expansion.
Consider a particle that is moving in a potential V. In
the special theory of relativity, the expression for the total
energy W of the particle is
W= V. (1)
In this equation, mo is the rest mass of the particle, p is its
momentum, and c is the velocity of light. It is convenient
to define the energy of a particle as
E= W-moe2. (2)
Equation (1) can be rewritten as
E=m02{ + -l}+ V. (3)
This equation can be expanded in pl(moe>' giving
p2 p4

2mo
(4)
where in second order, the so-called mass-velocity term
p4/(8m6c2) appears. The use of this expansion is not jus-
tified if plmoe> 1, i.e., if the momentum of the particle is
too large, as has been stressed by Farazdel and Smith.
9
If
the potential is Coulomb-like (V -1/r), then there is
always a region where the potential is so negative that the
momentum of the particle p is larger than moc, even if the
energy E is small.
However, another expansion can be found, which is
valid for Coulomb-like potentials over all space, even if the
momentum of the particle is at times larger than moe. The
only restriction is that the energy (a constant of the mo-
tion) is not too large lEI < (2mo?- V), which in chemi-
cal applications is always the case. At energies for which
this inequality would not apply, other effects should be
taken into account such as pair creation. The expansion
can be found by first rewriting Eq. (3)
E=
p2?
m02+ V
p2?
=2mQC2+E- V+ V
p2
2mo{l + [(E- V)/( 2moe
2
) n + V
p
2
C
2
(2mQC2- V){l + [EI(2mQC2- V) n+ v. (5)
The last two terms have been written down in order to
exhibit more clearly which expansions one can make. The
price to be paid to get rid of the root is that the equation is
now quadratic in the energy. The equation therefore has an
extraneous negative total energy solution. By choosing a
particular expansion, the spurious solution will be thrown
away. Expanding in (E- V)!(2mo?) will give in zeroth
and first order
p2
Eo=V+-
2
'
mo
(6)
Up to first order, this gives the same expansion as Eq. (4).
It is obvious that this expansion is not valid for r-+O, where
E- V> 2moe2. A correct expansion can be found (for en-
ergies smaller than 2moe2) by expanding in EI (2mo?
- V), or in 1/ (2moe2 - V) if one expresses all energies in
atomic units. In zeroth order, this expansion gives
p2?
Eo 2mQC2- V+ V.
(7)
The EI(2mo?- V) expansion is valid for Coulomb-like
potentials everywhere, whereas this is not true for the (E
- V)/(2mo?) expansion. Up to first order, the expansion
in EI(2mo?- V) gives
(2mQC2- V)2
(2moe2-2V) 2
= V+ (2mQC2- V)2P?
(2mQC2- V)3 .
(8)
Yet another expansion can be found, which is also valid for
Coulomb-like potentials, by rewriting Eq. (5). This expan-
sion uses the fact that only one of the terms in Eq. (5) was
expanded. It is possible to combine the two terms to get
another expansion:
J. Chem. Phys., Vol. 99, No.6, 15 September 1993
van Lenthe, Baerends, and Snijders: Relativistic regular two-component Hamiltonians 4599
p22
E= V+
2m
()?+E_ V'
p
2
c2 Ep
2
c
2
E= V+2moe2_ V (2m(JC2- V)(2m(JC2+E- V) , (9)
(
E- V ) p
2
C
2
E 1 + 2m()? _ V = V + 2m()? _ V'
1 (2m
0e2
- V)
E 1 + [E/(2m(JC2-2V)] 2moe2-2V
X ( V + 2mrJ- V),
If E is small compared to 2moe2 - 2 V, one can expand in
E/(2mo<!--2V). In zeroth order, this gives
(
2m0e2- V) ( p2c2)
Eo= 2m(JC2-2V V+
2m
()C2- V .
(10)
This expansion may be expected to be superior to the pre-
vious one for negative potentials because the expansion is
now in E/(2mo<!--2V) instead of in E/(2moe2- V).
III. EXPANSIONS IN RELATIVISTIC QUANTUM
MECHANICS
From now on, atomic units are used. In this section,
we will demonstrate that the traditional approaches in rel-
ativistic quantum mechanics for generating effective two-
component Hamiltonians (esc and FW) rely on the expan-
sion in (E- V)/2c
2
which is defective for Coulomb
potentials. The result of application of the regular expan-
sion in E/(2c
2
- V) is also derived.
A. Elimination of the small component
In the relativistic quantum theory, the Dirac equation
can be used as a starting point for relativistic calculations.
In this section, the transformation of the four-component
Dirac equation into an effective two-component form using
the method of eliminating the small component is consid-
ered.
The Dirac Hamiltonian works on a four-component
wave function IIJ,

(11 )
where is called the large component and X the small
component. These are both two-component wave func-
tions. The Dirac equation is
V+cuPX=E, cup+ (V-2c
2
)X=EX. (12)
Eliminating the small component gives
I 1( E_V)-l _
X 2?+E_V
cu
P
=2c 1+2C2 up=.X,
( 13)
(14)
The Hamiltonian Usc is energy dependent and works
solely on the large component , which is not normalized,
whereas IIJ is. In the standard approach, a normalized two-
component wave function <l> = O is generated by a nor-
malization operator 0 (which in fact effects a picture
change)
J <l>t<l>d
3
r= J tOtOd
3
r
= J IIJt'l1d
3
r= J (t+xtx)d
3
r= 1. (15)
Elimination of the small component gives
J (t+xtx)d
3
r= J t(1+XtX)d
3
r. (16)
Note that there is still considerable freedom in the choice
of O. One of the possible solutions for 0 is
0=
(17)
The Hamiltonian for <l> will now become
H=OUSCO-
1
=
(18)
The standard textbook approach (cf. Berestetski'i, Lifshitz,
and Pitaevski'i,21 McWeeny,22 Sakurai
23
) now proceeds
with an expansion in (E-V)/(2c
2
) of the factor [1+(E
- V)/2c
2
]-1 in X, in both USc [Eq. (14)1 and 0 [Eq.
(17)]. In addition, the square root in Eq. (17) is expanded
0= +p2/4c
2
+ ... ;:::d +p2/8c
2
, (19)
which is only justified if p2 < 42 [classically equivalent to
(E - V)/2c
2
< 1]. In quantum mechanics, there is an extra
problem if one wants to expand this operator in p2/ c
2
any
further because already the p4 term is not a well-defined
operator on the appropriate Hilbert space (it can produce
nonsquare integrable functions). If one nevertheless ex-
pands in this way, one obtains in zeroth order the nonrel-
ativistic Hamiltonian and, after some manipulation, in first
order the Pauli Hamiltonian
. p2 p4 il.V 1
!:F'auh= V +"2-8?+scr+4? u (VVXp). (20)
As pointed out before, the expansions used would only be
valid for regular potentials where the classical velocity of
the particles is everywhere small compared to the velocity
of light. As a matter of fact, for a Coulomb potential where
these conditions are not satisfied, the Pauli Hamiltonian
obtained in this way has well-known problematic features.
The Darwin term (il. V) / (8c
2
) has a c5-function singularity
at the origin, while the Dirac Hamiltonian does not pose
this problem. In a Coulomb field, the nonrelativistic eigen-
states have components of high momentum, for which the
use of the mass-velocity term as a first order perturbation
is questionable [cf. Farazdel and Smith (Ref. 9)]. This is
particularly troublesome since it has been demonstrated
that relativistic effects originate almost entirely from the
region close to the nucleus where classically the momen-
tum would be too large.
24
Finally, the relativistic effects are
so large in heavy element compounds that a perturbation
theoretical treatment is insufficient. As demonstrated in
Refs. 13 and 14, a self-consistent treatment of relativistic
effects is desirable, but the mass-velocity and spin-orbit
operators cause if one wants to solve the eigen-
value equation .nPauh=E. For one thing, due to the
mass-velocity operator, this is a fourth order differential
equation that does not lead to quantization of the energy
under the usual boundary conditions. The spin-orbit oper-
ator [11(42)]0' (VVXp) also causes difficulties since
close to the nucleus it behaves as an attractive -11? po-
tential that leads to arbitrarily large negative energies and
not to a discrete eigenvalue spectrum.
To solve most of the problems noticed above, one can,
as in the classical relativistic case, expand USC in 1I(2c
2
- V) [or V), which is equivalent as long as E is
in the order of unity]. For a Coulomb potential, the expan-
sion in 1/ - V) is justified even near the singularity of
the potential at the nucleus, in which region it is very
important to have a good expansion in view of the results
of Schwarz et al. 24 We obtain for /rsc,
USC = V +O'.p V (1 + V) -IO'.p, (21)
( ) E
::::: V +O'.p 2?- VO'p-O'p 2c2- V 2?- VO'p
+... . (22)
Since the expansion of 0 will have no effect in lowest order
(see below), we observe that the first two terms constitute
the zeroth order Hamiltonian [cf. Eq. (7)]

no= V+O'p O'.p. (23)
2c -V
So our zeroth order Hamiltonian is not the nonrelativistic
Hamiltonian. In fact, it is expected to incorporate relativ-
istic effects that are traditionally only introduced at the
level of the Pauli Hamiltonian. The great advantage of our
no is that it may be used variationally and that it does not
suffer from the singularities for r-+O that plague the Pauli
Hamiltonian. It is the primary objective of this paper to
study the performance of this no in self-consistent calcu-
lations.
In order to obtain the contributions to first order in
V), one needs, apart from the last term in Eq.
(22), contributions from the transformation with O. The
square root in the operator 0 will be expanded as
!vt-
O=l+iA%+'"
= 1 O'.p V)2 ( 1+ V) -20'.p+.... (24)
In classical relativistic mechanics, it would be allowed to
expand 0 in this way because the classical analog of xtx is
smaller than one
E-V
(2?+E- V)2 2c
2
+E- V< 1.
(25)
In quantum mechanics, the situation is slightly more com-
plicated, a point to which we shall return below. Expansion
of the factor [1+E/(2c
2
-V)]-2 in Eq. (24) yields
1 1
O=l+-O'p
2 2c--V 2c-V
(
) 1 E
-O'.p
2c-- V 2c-- V2c-- V
(26)
The second term in the above expansion of 0 is of first
order, and in zeroth order, 0 is just the unit operator. We
already used this result to obtain no ofEq. (23). Using Eq.
(22) for !Fe and the first two terms ofEq. (26) for 0, the
Hamiltonian up to first order HI will now be
(27)
This Hamiltonian can be compared with Eq. (8). HI
cannot be used in variational calculations without extra
assumptions. It posesses fourth order derivatives and has
no lower bound for the expectation value of the energy, just
like the Pauli Hamiltonian.
The zeroth and first order Hamiltonians obtained here
are identical to Hamiltonians of these orders derived by
Chang, Pelissier, and Durand 12 using the theory of effec-
tive Hamiltonians. Chang et al. obtain these Hamiltonians
when imposing the condition of hermiticity following des
Cloizeaux.
25
Our approach here is energy dependent and is
not convenient to use for higher orders, but it gives direct
physical insight in the origin of the trouble with the Pauli
Hamiltonian and its cure [expansion in E/(2c
2
- V) in-
stead of (E- V)/(2c
2
)].
We finally return to the expansion of the square root in
the operator 0 [see Eq. (24)]. In practice, we will be look-
ing for solutions in a finite model space of two-component
functions in which at least the bound states of no, or equiv-
alently the large components of the bound states of the
Dirac Hamiltonian, can be well represented. The expan-
sion of 0 would be justified if for any normalized vector f
in the model space the norm of the image vector xt X f
would be smaller than one. Some justification for this as-
sumption may be given. Taking for f the large component
? of some eig.enstate and the norm of
IS seen to be Just the norm of XX. Since
1 (cO' VV)
=2?+E- VcO'.p-i (2?+E- V)2' (28)
we have, using Eq. (14),
(29)
Since the factors in front of q, and X are much smaller than
1 everywhere except possibly very close to the nucleus, the
norm of gt X is expected to be small.
It is interesting to note that the quantum mechanical
effects, arising from the noncommuting of p and V and
leading to the second term in Eq. (28), are actually quite
small. If we neglect this term with respect to the first,
which seems justified because of the (2c
2
+ E - V) -
2
fac-
tor, then we may proceed as in the classical case to obtain
an imprOVed zeroth order Hamiltonian. Simplifying gtg
with the help of Eq. (14), we obtain
q,totOq,=q,t( I +XtX)q,
=q,t[ 1 +O'.p V)2 O'.p]q,
1 VO'p 22::'- vO'p)q,
(
E-V)
=q,t 1+22+E_ V q,.
(30)
We recognize in (E- V)/(2c
2
+E- V) the "classical"
form of gtg [cf. Eq. (25)]. In zeroth order of the expan-
sion in E/(2c
2
_ V), a solution for 0 is
P?=2V
O=V-2T-V'
(31)
This would be the true zeroth order expansion in E/(2c
2
- V) of 0 if one is justified in neglecting the last term in
Eq. (28). After some calculation, the zeroth order of ex-
pansion for the Hamiltonian is
f2c2"=V ( c?) f2c2"=V
Ifrt<t> = V2?-=-iV V + 0'. p 22 _ V 0'. p V2c2-=-iV <t>
=E. (32)
The similarity between this "extended" zeroth order
Hamiltonian and Eq. (10) is obvious. We will also inves-
tigate the performance of this Hamiltonian, which we ex-
pect to be superior to IfJ [Eq. (23)] except possibly very
close to the nucleus.
B. The Foldy-Wouthuysen transformation
The most straightforward way to generate an effective
two-component Hamiltonian would consist of finding a
unitary transformation

1
1 r)
Jl+X+X Jl+X+X
(33)
1 1 '

Jl+XX+

r 1 )
-
(34)
X
Jl+XX+
that brings the Dirac Hamiltonian If D,
(
V
If-
D-, CO'.p
cu p )
V-2c
2
(35)
to block diagonal form. Foldy and Wouthuysen
26
intro-
duced a systematic procedure for decoupling the large and
small components to successively higher orders of c-
2
In
this section, we follow closely the approach and notation of
Kutzelnigg,4 who has stressed the problems connected
with this and other procedures to obtain the Foldy-
Wouthuysen transformation, which are already apparent
from the lowest order (c-
2
) approximate Hamiltonian
(the Pauli Hamiltonian) obtained in the FW method. We
pause briefly to demonstrate that part of the trouble is
again caused by the neglect of E - V with respect to 2c
2
rather than E with respect to 2c
2
- V. Suppose the trans-
formation U [Eq. (33)] generates the desired two-
component FW wave function

(36)
so that
1 1

(37)
The last line is an identity if X satisfies X=Xq,. If we ap-
proximate X by expanding the energy dependent X for
which X=Xq,= (2C+E- V)-ICO'pq, in (E- V)/2c
2
, we
obtain in lowest order X ( l/2c
2
) CO' P and U becomes to
order c-
2
,
(

8c
U-
cO'p
---V
cO'p )
--V
2
P
1-g?
(38)
This is precisely the traditional FW transformation to or-
der c-
2
that leads to the much criticized Pauli Hamil-
tonian. It is interesting to see what happens if we follow, in
order to arrive at an improved effective Hamiltonian, the
same strategy as before and avoid the erroneous expansion
of (2c
2
+ E - V) -I by expanding in E/ (2c
2
- V). This
leads to inserting X (2c
2
- V) - 1 CO' P in the expression
for U. We also have to expand the square root operator
+xtx, just as in the esc method. Following exactly the
same procedure of ordering the terms according to the
number of (2c
2
- V) factors in the denominator, the trans-
formation matrix U will now be
c
uop 22- V
u=
c

2c -V
1 c 2 c
1-
2
2c
2
- v
P
2c
2
- V
(39)
This transformation will give in zeroth and first order
c
2
Jt>= V+uop (22- V) uop,
(40)
IJt>1 ? Jt>
H = -2
UoP
(2c2_V)2UoP
1 ?
-2 Jt>u
o
p (2c2_V)2
Uo
P.
(41 )
These are exactly the same zero and first order Hamilto-
nians as obtained before in the esc method [cf. Eqs. (23)
and (27)].
Co The CPO Hamiltonian
We have derived a regularized two-component Hamil-
tonian by simply modifying the traditional esc and FW
approaches so as to take care of the radius of convergence
of the employed expansions. The present simple approach
becomes more complicated for higher orders. It has been
presented in some detail in order to stress the important
point concerning the validity of the expansions to be used.
The Hamiltonian obtained has actually been derived earlier
by Chang, Pelissier, and Ourand
l2
and will therefore be
denoted as the CPD Hamiltonian lfo
PD
(we leave aside the
further regularization of the kinetic energy applied by
Chang et al. ) . We will not study higher order terms of the
Hamiltonian, but we note that Chang et al. have derived
and used high order terms in the expansion of Ire. Al-
though our zero order Hamiltonian is identical to the one
obtained by Chang et al. this does not hold for the higher
orders since they did not use the renormalization operator
o and obtained non-Hermitian higher orders. Our HI is
actually identical to the Hermitian Des Cloizeaux HI.12
The zeroth order CPD effective Hamiltonian lfo
PD
may be further developed
2
m
PD
= V+uoP(2C2C_ v)u
op
c
2
c
2
= V+p 2C2"=VP+ (22- V)2 u
o
(VVXp).
(42)
One can now see that the spin-orbit splitting is already
present in the zeroth order Hamiltonian. This spin-orbit
term is regular because of the (2c
2
- V) -2 factor in it. It
poses no problems in variational calculations. The eigen-
value equation
I4
PD
o= [V V )u
op
(43)
is only a second order differential equation. The two-
component wave function <1>0 will now be referred to as the
(zero order) CPD wave function. In first order perturba-
tion calculations (FOPT), the Hamiltonian of Eq. (41) is
used. Defining
(
V)-I
/= 1-2(02 ,
(44)
the FOPT energy E?D can be defined as
Ef
PD
= < <1>0 I HII <1>0)
....cPD 2
=J!,i) (uop<l>olf luop<l>o) (45)
Because the Hamiltonian .a<gPD is energy independent
and Hermitian, the eigenfunctions belonging to different
eigenvalues Eo are orthogonal. An important question is
whether the eigenvalue spectrum is bounded from below.
We briefly investigate this problem along the lines of the
analysis of Landau and Lifshitz
27
for the nonrelativistic
case. We only consider potentials V for which 2c
2
- V> 0
everywhere, so that the CPD kinetic energy operator
rc
PD
,
?

2c -V
(46)
is a positive operator. If the potential is bounded from
below by V min' then all the CPD eigenvalues If;,PD
;> V min because the mean value of rcPD;>O. Now suppose
the potential V is of the form
Z
V=-;r
(47)
near the origin. Consider a wave function localized in some
small region (of radius 70) around the origin. The uncer-
tainty in the momentum of the particle is then of order 1/70
(uncertainty principle). The sum of the mean values of the
potential and kinetic energy then is of the order
Z ?
-lo+''6[2c
2
+ (Z/l(';)] .
(48)
For s < 1, this energy expression cannot take arbitrarily
large negative values. However, whereas in the nonrelativ-
istic case the energy is bounded from below for s < 2 and
conditionally so for s=2 (depending on the strength of the
potential, cf. Chap. 35 of Ref. 27), the case s= 1 is already
a special one for the CPD eigenvalue spectrum. Equation
( 48) suggests that for s = 1, i.e., a Coulomb potential, the
CPD Hamiltonian will only be a bounded operator for
Z < c. Note that pathological behavior also occurs for the
Dirac equation for Z> c. Analogous conditions for the po-
tential in the Dirac and the CPD equations may perhaps be
expected from the identical asymptotic behavior for r-+O
for the regular solutions, which is rr-
1
in both cases. Here

(49)
and K is the usual relativistic quantum number. However,
these arguments are qualitative. A more detailed treatment
of this important question, including a rigorous proof that
the CPD eigenvalue spectrum is bounded from below if
Z < C and an investigation of the precise relation between
the Dirac and CPD eigenvalues and eigenfunctions in a
pure Coulomb potential, is deferred to a subsequent paper
(see Ref. 28).
IV. RESULTS
A. Methodological details
1. Self-consistent calculations
We discuss first the construction of the potential VCr)
in the Hamiltonian m
PD
from the one-electron solutions
to Eq. (43) during the iterations of a self-consistent calcu-
lation. In the present work, the simple Xa version
29
of the
density functional theory is used. The electron-electron
potential Vee is thus split in the classical Coulomb interac-
tion V c and the exchange-correlation potential V
xc
' Mag-
netic effects and retardation were not taken into account
for the electron-electron potential. The potentials are cal-
culated from the electron density p in the following way:
V= V
N
+ Vee= V
N
+ V
c
+ V
xc
'
where
z
VN(r)=- L _I -'-I'
i r-ri
(50)
(51)
(52)
Vxc(r) = Vxc(p) = -3a[ per) r
3
, with a=0.7.
(53)
The Dirac equation with this approximation for the
exchange-correlation potential is called the Dirac-Slater
equation
[
VN+ V
c
+ Vxc(p)
cup
(
<Pi)
=Ei Xi'
(54)
Our two-component calculations are to be considered as
approximations to Dirac-Slater calculations. [In such cal-
culations, no attempt is made to include relativistic effects
to the exchange-correlation potential (see e.g., Ref. 30).
When carrying the calculations through to self-
consistency, the density and the potential are derived from
the solutions <1>0 to Eq. (43). This involves approxima-
tions, as can be seen as follows:
The potential VCr) in Eq. (43) should be derived from
the Dirac density pD(r). Denoting the components of the
Dirac one-electron spinor \11 by \I1i' i= 1..4, those of the
transformed wave function <1>= U\JI by <l>i' i= 1..4 (<1>3 and
<1>4 are therefore zero), and the eigenstates of the r operator
in the direct product space of spatial and spinor coordi-
nates by 1 r,i), we have
The components of the transformed wave function <1>= []\II
along 1 r,;) are not identical to those of \11, but only the
components of along the transformed basis states UI r,i)
are-(r,i 1 at 1 <1 = (r,iI \11). This simply reflects the well-
known fact that a picture change effected by U requires
that not only the wave function is transformed, but also the
observables, in this case, r to UrUt =q, in order that the
physics remains unaltered. The inverse transformation of
the operator r, utrU, is called R. It is the operator that
describes, in the Dirac picture, the famous average position
or mass position rmass of the electron (see Foldy and
Wouthuysen
26
and MosS
31
). A clear example of the differ-
ence we introduce by using (r) instead of (q) is pro-
vided by the nodes that are present in the solutions <1>0 to
Eq. (43) and therefore in the (orbital) density
Such nodes do not occur in pD(r) since the nodes of the
large component <p do not coincide with those of the small
component X. The effect of (neglect of) the picture change
is probably small, but such effects are visible for core states,
as discussed extensively by Baerends et al. 32 Although the
Dirac density is not 1 (r) 1
2
, it could in principle, be cal-
culated from (r) for instance by using
xt I )
-

(56)
and writing the Dirac electron density as
per) =<pt(r)<p(r) +xt(r)x(r)
= [ (r) r[ (r) 1
+[x
(57)
Note that one cannot use the turnover rule here. We will,
however, approximate the electron density that is used for
calculating the potential in the self-consistent field (SCF)
calculations by
J. Chern. Phys., Vol. 99, No.6, 15 September 1993
(58)
where <1>0 are the solutions of Eq. (43). So apart from using
the approximate Hamiltonian H?D of Eq. (43), we also
make two presumably small errors for the potential V(r)
in that equation. In the first place, the effect of the picture
change is ignored, meaning we use <l>t(r)(r) instead of
Eq. (57). In the second place, we neglect-consistent with
the order in which we work-the fact that the transforma-
tion U that we effectively use is only correct to order
E/(2r?- V), so the small components are not completely
annihilated by U, as they are assumed to be in Eqs. (56)
and (57). Maybe this approximation is also most serious
when the small component is relatively large. Note that the
second approximation, neglect of residual small compo-
nents, would disappear when more accurate transforma-
tions U would be used, but the first approximation, exem-
plified by the problem of the nodes, would not improve in
that case. It can only be remedied by making the correct
picture change for the position variable. The effect of these
approximations on the results of actual self-consistent cal-
culations will be discussed in the next section.
We note that Chang et al
12
did not carry through self-
consistent calculations on many-electron atoms. From one-
electron calculations, they obtained relativistic correction
potentials that were used in many-electron systems (noble
gases). The potentials they use thus do not take screening
effects on the relativistic correction potential into account,
which will be treated in the present self-consistent ap-
proach.
2. Separation of the radial variable from angular and
spin variables
To solve the equations for an atom, it is useful to sep-
arate variables just as in the nonrelativistic case. Because
the potential of an atom is spherically symmetric, the total
angular momentum j = 1 + s of a particle is conserved. j
commutes with the CPO Hamiltonian, so we may con-
struct simultaneous eigenfunctions of H, j2, and jz. The
eigenfunctions can also be classified according to parity,
which depends on the I value. It is convenient to introduce
the operator K,
K=ul+ 1 =u (rXp) + 1. (59)
Eigenfunctions of this operator are written as 77:' with ei-
genvalue - K,
(60)
These are functions of angular and spin variables with a
definite parity. The relativistic quantum number K is given
by
K= 1-(/+ 1) = -

77:' is given explicitly by

. I \'
J= -2"
(61)
(62)
In this equation, m=m' +ms holds and (}sm is the eigen-
s
function of the spin operator. One can now separate vari-
ables because K commutes with the Hamiltonian, as does j.
Writing
=R(r)77:' , (63)
it is possible to calculate u p<l>,
(64)
u p<l> = a;.u p<l>
ur (u.r )
=-r- --r- u p 
[
1 i ] (a i )
=a -rp+-u(rXp) =a -i-+-ul
r r r r ar r
(
aR K+ 1) (aR K+ I ) m
=a -i--i--R "I1
m
=i -+--R 77- .
r ar r "K ar r K
(65)
Now it is possible to separate the spin and angular vari-
ables from the radial variable in the CPO-Slater equations
because the u p operator appears twice. The first one will
give something proportional to 77'::.K' the second one will
give back something proportional to 77:'.
Due to the separation of radial from angular and spin
variables, the Dirac-Slater, the CPO-Slater, and the ex-
tended CPO-Slater equations can be solved numerically.
This will be done for the uranium atom (see the next sec-
tion).
3. Basis set calculations
To prepare for future applications to molecules, the
CPD-Slater equation is also solved in a basis set expansion.
To solve the CPO-Slater equation (43) for atoms in a
basis, one needs to calculate the following matrix elements:
where
(67)
is a basis function. The first two are the same as in the
nonrelativistic case. The last one can be written as [with
the aid of Eq. (65)]
TABLE I. Optimized Slater exponents for all-electron CPO-Slater calculations on uranium.
Sill PI/2 P3/2
d
3
/
2 dS/2
n
~
n
~
n
~
n
~
n
~
27700.0 2 3587.50 2 150.26 3 64.151 3 47.917
5409.50 2 958.970 2 58.433 3 30.866 3 28.208
1440.000 2 337.190 2 41.219 3 21.393 3 20.346
454.960 2 138.288 3 25.240 4 15.350 4 14.630
180.480 2 66.387 3 19.950 4 10.650 4 10.260
1 106.185 2 45.709 4 16.135 5 7.330 5 7.054
2 223.615 3 28.467 4 11.726 5 4.703 5 4.526
2 47.215 3 22.424 5 8.350 6 2.946 6 2.772
3 60.110 4 17.715 5 5.775 6 1.644 6 1.520
3 25.833 4 12.915 6 3.817 6 0.915 6 0.829
4 15.769 5 9.259 6 2.346
4 11.075 5 6.464
5 10.457 6 4.365
5 7.268 6 2.741
6 5.008
6 3.274
7 2.035
7 1.199
1 1
2(<1>;10" pfO' P 1 <1>j) = -2(0" p<1>ilfl 0" p<1>j)
(68)
. Because f = ( 1 - V /2c
2
) - 1 depends on the potential, it
IS convenient to do a numerical integration. To solve the
CPO-Slater equations, one can now use the same tech-
niques as in the nonrelativistic case. The difference is that
in the nonrelativistic case, the eigenfunctions are labeled by
the quantum numbers I and lz' whereas now they are la-
beled by the quantum numbers K and jz=m. The basis set
consists of normalized Slater-type functions
(69)
The different n's and ~ ' s used in the analytical calculation
are listed in Table I. The exponents were fitted to the nu-
merical orbitals. Note that six Is functions are used, with
very high exponents. The reason one needs such high ex-
ponents for the s1l2 orbitals is that for r ..... O, the wave func-
tion has the weak singularity noted in Eq. (49), just as in
the fully relativistic case.
B. All-electron calculations on U
Results of all-electron calculations on the U atom are
given in Table II. The nonrelativistic (NR) results are
from a numerical solution of the one-electron Hartree-
Fock-Slater equations. The results with the Pauli Hamil-
tonian (Pauli pert) are from a perturbation theory ap-
proach from Snijders and Baerends.
33
The difference between the nonrelativistic and relativ-
istic (Dirac) results is quite large, the nonrelativistic one-
electron energy of a valence level like 5f deviating by sev-
/SI2 /7/2
n
~
n
~
4 22.727 4 21.562
4 13.541 4 12.901
4 8.544 4 8.241
4 4.517 4 3.792
5 3.024 5 4.553
5 1.734 5 2.052
5 0.983 5 1.056
eral tenths of an atomic unit from the relativistic one. The
nonrelativistic levels are even in the wrong order. The
CPO energies reproduce the Dirac ones to an accuracy of
better than 0.001 a.u. in the valence region. The Pauli pert
results do give the right qualitative picture, but the 6s
1l2
and 6P1I2 orbital energies still have errors of about 0.1 a.u.
compared with the Dirac values. One reason for this dif-
ference between CPO and Pauli pert is that if one uses the
zero order CPO Hamiltonian, which includes much of the
relativistic effects, one can also fully incorporate indirect
effects (from the relativistic orbital contractions and ex-
pansions) through the self-consistent calculations. The
Pauli pert results include such effects only as a first order
perturbation to the potential VCr). It has been noted be-
fore
l3
that considerable improvement over the Pauli pert
results can be obtained by performing so-called quasirela-
tivistic self-consistent calculations with the Pauli Hamil-
tonian projected onto the space of nonrelativistic orbitals
to avoid collapse to the nucleus and other problems men-
tioned before (cf. the last column of Table IV). Although
this approach has proven quite successful in applications to
heavy element compounds (cf. Refs. 13, 14 and 34), it is
not devoid of theoretical flaws. The method of the present
paper is, however, well founded and achieves an even bet-
ter result.
In order to examine the results of Table II more
closely, it is interesting to compare the zeroth order CPO
effective Hamiltonian with the energy-dependent Hamil-
tonian for the large component if> of the Dirac wave func-
tion obtained by eliminating the small component X in the
Dirac equation. This Hamiltonian n= is [Eq. (14)]
c
2
}Fc= V+O"p 2c2+E- VO'p.
(70)
Comparing the }Fc Hamiltonian with the zeroth order
CPD effective Hamiltonian
TABLE II. Results of all-electron calculations on the U atom.
CPD
basis CPD CPD Pauli
Orbital NR Dirac CPD set FOPT ext pert
ISI/2
-3690.78 -4255.55 -4872.99 -4872.78 -4158.88 -4043.07 -4114.71
2SI/2
-640.21 -795.00 -818.96 -818.91 -793.51 -781.76 -753.91
2Pl/2
-619.59 -766.70 -789.87 -789.67 -765.24 -749.63 -728.13
2Pl/2
-625.96 -641.99 -641.97 -625.15 -616.36 -626.95
3SI/2
-161.26 -200.69 -202.63 -202.62 -200.51 -199.62 -190.53
3Pl/2
-151.14 -187.82 -189.62 -189.58 -187.65 -186.60 -178.43
3Pl/2
-155.38 -156.70 -156.70 -155.29 -154.68 -155.59
3dl/2 -132.14 -134.98 -136.14 -136.13 -134.93 -134.44 -135.29
3d
s
/2 -128.40 -129.46 -129.46 -128.36 -127.94 -128.93
4s
1
/2 -40.57 -51.09 -51.24 -51.22 -51.05 -50.98 -48.28
4Pl/2
-35.89 -45.29 -45.42 -45.41 -45.26 -45.19 -42.78
4Pl/2
-36.82 -36.92 -36.92 -36.80 -36.77 -36.90
4dl /
2
-27.16 -27.58 -27.66 -27.66 -27.58 -27.56 -27.68
4d
s
/2 -26.03 -26.09 -26.10 -26.02 -26.00 -26.19
5s
1
/2 -8.813 -11.32 -11.33 -11.32 -11.31 -11.31 -10.62
5pI/2 -7.006 -9.073 -9.076 -9.075 -9.066 -9.062 -8.490
5p3/2
-7.057 -7.061 -7.062 -7.055 -7.052 -7.087
4/S/2
-15.06 -13.87 -13.91 -13.91 -13.87 -13.87 -14.00
4/1/2
-13.47 -13.50 -13.50 -13.47 -13.47 -13.57
5dl/2 -3.850 -3.764 -3.767 -3.767 -3.764 -3.764 -3.793
5d
s
/2 -3.465 -3.468 -3.469 -3.466 -3.465 -3.500
6s1/2
-1.298 -1.720 -1.719 -1.716 -1.718 -1.718 -1.580
6pI/2
-0.7945 -1.069 -1.069 -1.068 -1.068 -1.068 -0.9725
6Pl/2
-0.7410 -0.7409 -0.7415 -0.7407 -0.7406 -0.7476
5fs/2 -0.3419 -0.1033 -0.1040 -0.1045 -0.1040 -0.1043 -0.1047
5h/2
-0.0728 -0.0735 -0.0743 -0.0735 -0.0738 -0.0647
6d
l
/
2
-0.1157 -0.0710 -0.0711 -0.0716 -0.0711 -0.0712 -0.0728
6d
s
/2 -0.0537 -0.0538 -0.0543 -0.0538 -0.0539 -0.0512
7S
1
/
2
-0.1071 -0.1340 -0.1339 -0.1339 -0.1339 -0.1339 -0.1231
(71)
one notes there are two differences between JrSc and the
present use of m
PD
JrSc is energy dependent, but if E is
small compared to 2c
2
, this is a difference of the order
E2/22 (here one assumes that the kinetic energy is about
equal to minus the energy). The percentage error is thus
expected to increase linearly with E. The second difference
arises from our use of 1 (r) 12 to generate the density,
which does not yield exactly the Dirac density and poten-
tial V (cf. the discussion in Sec. IV A). This may affect the
energies of the higher lying orbitals since those depend
strongly on the screening of the nucleus by the densities of
all lower lying orbitals. Near the nucleus, however, the
nuclear attraction dominates and the Hamiltonians are
practically the same. So one expects the CPD wave func-
tions to be approximately the same there (apart from nor-
malization) as the large component of the Dirac wave
function. The CPD wave function indeed exhibits the same
mild singularity near the nucleus as does the large compo-
nent of the Dirac spinor.
If one looks at the CPD-orbital energies in Table II,
one can see that for the innermost orbitals, which see the
nuclear potential almost unscreened, the expectation that
the error is about E2/2c
2
is borne out (14.5% error for
1S1/2, 11.1 % expected). If one looks at higher lying orbit-
als, the expected error of E2/2c
2
is very small in an abso-
lute sense. The actual error is in fact larger than E2/22 due
to the error in the self-consistent potential used, though
still very small (in the upper valence region typically
0.0001 a.u., Le., 0.1% error, 10-
4
% expected). Comparing
to the nonrelativistic energies, in all cases the CPD ener-
gies recover the bulk of the relativistic effect, except for
Is1/2' The extended CPD results are superior to the CPD
results, reducing the error in most cases by more than
50%. The CPD FOPT results are considerably better still,
especially for the inner shell orbitals, where the error is
reduced to about 10% of the CPD error. The largest re-
maining error, both absolute and relative, is in Is1/2 (2.2%
TABLE III. Uranium all-electron spin-orbit splitting.
Spin-orbit CPD CPD Pauli
splitting Dirac CPD FOPT ext pert
2p 140.74 147.83 140.09 133.27 101.18
3p 32.435 32.921 32.358 31.915 22.84
4p 8.469 8.503 8.452 8.419 5.88
Sp 2.0157 2.0152 2.011 4 2.009 1 1.403
6p 0.3284 0.3278 0.3277 0.3275 0.225
3d 6.584 6.681 6.577 6.511 6.36
4d 1.5579 1.564 5 1.5565 1.5527 1.49
5d 0.29843 0.29857 0.29821 0.29804 0.293
6d 0.01731 0.01731 0.01730 0.01730 0.0262
4/ 0.40637 0.40797 0.40626 0.40580 0.43
5/
0.03045 0.03047 0.03046 0.03049 0.0400
8
I DIRAC-I
6
;
!
t
:i
~ 4
c.
~
10'6 10,5 10'4 IO,J 10'2 10,1 100
r (a,u.)
FIG. 1. r times the square root of Dirac 1s
1
/2' CPD 1s
1l2
, and NR Is
orbital density of uranium.
for CPO FOPT). The CPO and particularly CPO FOPT
results are much superior to the Pauli pert ones.
If one looks at the spin-orbit splitting, the regular
Hamiltonians give much better results than the perturba-
tive Pauli result (see Table III). The CPD FOPT orbital
splittings all differ less than 0.5% from the Dirac result,
whereas the Pauli pert results differ up to 40% [Pauli pert
results are much better in frozen core calculations (see
Ref. 33 and see below)] Chang, Pelissier, and Ourand
l2
listed some spin-orbit splitting energies. Especially their
valence shells are in poorer agreement with exact relativ-
istic values than the 5p and 6p results listed here. For
example, the 6p spin-orbit splitting energy for radon (Z
4
I DIRAC-I
10,3 10'2
r (a.u.)
FIG. 2. r times the square root of Dirac 2P1/2' CPD 2P1/2' and NR 2p
0.8
0.6
i" 0.4
0.001 0.01 0.1
r (a.u.)
;
f
~ .
;
\
i
\
\
10
FIG. 3. r times the square root of Dirac 6P1I2' CPD 6P1/2' and NR 6p
= 86) is in their work 0.123 a. u. compared to 0.156 a. u. for
the exact value. The less accurate results for the valence
orbital energies can possibly be understood from the fact
that these authors did not take screening effects into ac-
count, whereas we do. On the other hand, Schwarz et 0/.
24
have demonstrated that spin-orbit splittings of p atomic
orbitals (AO) originate very close to the nucleus (mostly
from the spatial region occupied by the K shell electrons).
The spin-orbit splitting should therefore not be particu-
larly sensitive to inaccuracy of the (valence) electronic
potential.
We now tum to the electron density. In Figs. 1-4,
some of the orbital electron densities are compared for the
0.5
0.4
; 0.3
c.
t:
0.2
0.1
0.001 0.01 0.1 10
r (a.u.)
FIG. 4. r times the square root of Dirac 7s
1/2
, CPD 7S
1/2
, and NR 7s
TABLE IV. Valence-only calculations on uranium using the Dirac core density.
CPD Pauli Pauli
Orbital EL Dirac CPD FOPT Ext FOPT QR
6s
112
2 -1.71981 -1.72039 -1.71988 -1.71976 -1.580 -1.749
6PI12
2 -1.06935 -1.06963 -1.06939 -1.06931 -0.9718 -1.019
6p3/2
4 -0.740 957 -0.741090 -0.740 959 -0.740940 -0.7468 -0.7456
5/
s12
3 -0.103296 -0.103470 -0.103440 -0.103300 -0.1090 -0.0928
5/7!2
0 -0.072 841 -0.072 987 -0.072 967 -0.072 845 -0.0690 -0.0676
6d3/2 -0.071033 -0.071021 -0.071016 -0.071031 -0.0726 -0.0722
6d
s12
0 -0.053728 -0.053715 -0.053712 -0.053726 -0.0510 -0.0560
7SI/2
2 -0.133988 -0.133989 -0.133984 -0.133987 -0.1231 -0.1330
uranium atom. In these pictures, the nonrelativistic,
Dirac-Slater and CPD-Slater electron densities are pre-
sented. It is obvious that the CPD-Slater electron densities
are in much better agreement with the Dirac-Slater ones
than are the nonrelativistic densities, in particular, for the
valence levels in the outer region. In contrast to the non-
relativistic densities, the Dirac and CPD densities practi-
cally coincide over all shells except the innermost ones. A
notable difference is that the CPD orbital electron densities
have nodes, whereas the Dirac ones are always larger than
zero, because the large and small component of the Dirac
wave function do not vanish at the same place. As dis-
cussed in Sec. IV A, this is an expected effect of neglecting
to transform also the operator r in the change from Dirac
picture to (approximate) Schrodinger picture (also called
the Foldy-Wouthuysen picture). We may use the results of
Baerends et al.
32
concerning the difference of expectation
values of the Dirac position rcharge and of the average po-
sition r mass in order to infer what consequences we may
expect when the orbital density of the CPD wave function
<l> is given in terms of the untransformed r rather than the
transform UrUt. The CPD wave function <l> is the relativ-
istic wave function transformed from Dirac picture to
Schrodinger picture
(72)
The superscripts D and S denote Dirac picture and Schro-
dinger pictures, respectively. That we are dealing with the
relativistic wave function is indicated explicitly by the sub-
script reI. The expectation value (<l> I r I <l
= ('I':ell r I is just the expectation value of the average
position r mass' with operator R = ut rU in the Dirac pic-
ture, for the relativistic wave function
rl 'I':el) = rl = I
(73)
As discussed in detail in Ref. 32, the expectation value (R)
is for a 1sl/
2
orbital significantly lower than the expectation
value (r). This lowering has been calculated for the 1sI/2 of
uranium to be 1.0.10-
3
a. u., 36% of the nonrelativistic
to relativistic contraction of 2.8X 10-
3
a.u. [according to
the conventional definition (Ref. 32)]. The shift of the
pCPD(r) curve to lower r with respect to the pD(r) curve
observed in Fig. 1 is perfectly in line with this result. There
may of course be other influences, particularly for the outer
tail of lS1/2, which already extends sufficiently far out that
screening effects by (the core tails of) other orbitals take
effect. It is interesting to observe that the 2pi!2 orbital
shows opposite behavior, the CPD density being lower
than the Dirac density in the inner tail of the wave func-
tion, which would have an increasing effect on (r). This is
again consistent with the findings about picture change
effects. Since according to Ref. 32 for a hydro genic orbital
we have
(74)
(which shows that for large Z and small n the effect is in
the order of the Compton wavelength), one may expect
opposite behavior for K < 0 orbitals (the j = 1+ 1/2 upper
spin-orbit components such as 1s
l
!2) compared to the
K> 0, j = /-1/2 lower spin-orbit components such as the
2Pl/2 AO. The latter have indeed been found to show ex-
pansion when going from (r) to (R).
These picture change arguments may give some insight
in the observed deviations (presence of nodes; contraction
or expansion) of the CPD density from the Dirac density.
Keeping in mind that <l> resembles a renormalized large
component N> of the Dirac wave function and using Eq.
(57), other details of the differences in Figs. 1 and 2 be-
tween pD and pCPD, such as pCPD being higher than pD
close to the maximum, may also be understood. These ar-
guments are, however, qualitative. There may be other ef-
fects, such as the neglect of contributions of residual small
components, which may not be completely negligible after
transformation with an approximate U. Altogether, the
CPD density is so close to the Dirac density that the va-
lence orbital energies and outer tail densities of the CPD
wave functions are very accurate, well within "chemical
accuracy" of the Dirac values. We will nevertheless con-
sider in the next section possible further improvement by
using more accurate potentials in the Hamiltonian JfciPD.
C. Valence-only calculations on uranium using the
Dirac core density
In view of its great utility in molecular calculations on
large heavy-element compounds, we briefly investigate
valence-only calculations. In this approach, the electron
density is split into a core density Prore and a valence den-
sity Pval' Only the valence orbitals are optimized and the
core density is taken from an all electron Dirac-Slater cal-
:i 0
.e
.g-
.
"t.
-2
t'
:' \
: \ IEXT-DIRACI
\ ~
,
,
,
,
,
,
,
,
10'5 10' 10') 10'2 10'1 10 101
r (a,u,>
FIG. 5. ? times the difference in uranium 7s
1l2
orbital density between
CPD and Dirac calculations (CPD-Dirac) and between extended
CPD and Dirac calculations (ext-Dirac). Results are from all-electron
calculations.
culation. Although the difference with a CPD core density
would not be large, this procedure does correct for the
deviations from the Dirac density we discussed in the pre-
vious section and which should not enter the potential. The
core density is frozen during the SCF calculations for the
valence orbitals. We solved the CPD-Slater and extended
CPD-Slater equations numerically. In this particular ap-
plication, the only condition for the valence orbitals was
that the number of nodes was correct. This ensures orthog-
onality on the lower states that are solutions in the same
potential.
Results are given in Table IV for uranium. For com-
parison, results are shown from calculations which use the
standard Pauli Hamiltonian and the frozen core approxi-
mation. First order perturbation theory (FOPT) and qua-
sirelativistic (QR) results are shown (results taken from
Ref. 14). We observe a considerable improvement over the
results of the previous section. The CPD results have an
accuracy of about 0.0006 a.u., the CPD FOPT results
about 0.000 15 a.u. and the extended CPD (ext) results are
even better with an error of about 0.000 06 a.u. compared
with the Dirac results. The results with the Pauli Hamil-
tonian have errors more than 0.01 a.u. for the 6s
1l2
and
6P1I2 orbital energies.
In Figs. 5 and 6, we plot? times the CPD minus Dirac
orbital densities for the 7s
112
orbital, in Fig. 5 for all-
electron calculations, and in Fig. 6 for the present valence-
only calculations. The same plots are shown for extended
CPD minus Dirac densities. Both CPD and EXT are doing
very well (note the scale of the plots); the differences with
the Dirac orbital density are very small. The maximum
differences are about 0.0005 a.u. for both CPD and ext. If
one would calculate the same maximum difference between
the nonrelativistic (NR) 7s and the Dirac 7s
1l2
orbital,
then one would find 0.12 a.u., which is more than two
2
:i 0
.e
~
"t.
-2
-
" , ,
! l_IEXT-DIRACI
.
.
.
, , .
: '
. .
, '
. '
. .
, .
. ,
,/ ~
" ,
.. ,;' :
,
,
,
,
,
,
10'5 10'4 10'3 10'2 10'1 10 101
r (a.u,)
FIG. 6. ? times the difference in uranium 7S
I
/
2
orbital density between
CPD and Dirac calculations (CPD-Dirac) and between extended CPD
and Dirac calculations (ext-Dirac). Results are from valence-only calcu-
lations using the Dirac core density.
orders of magnitude larger than the present result (see also
Fig. 4). The valence-only results of Fig. 6 are significantly
improved compared to the all-electron results of Fig. 5,
particularly in the valence region. In the most important
outer lobe of the 7s, between 1 and 10 bohr (cf. Fig. 4), the
difference between CPD and Dirac is now only 3.10-
5
a.u.
and between ext and Dirac one order better still.
We make two more observations concerning the curves
of Figs. 5 and 6. The Dirac orbital density is always larger
than zero, whereas the CPD and ext orbital densities have
nodes. The difference curve is therefore expected to show
negative minima close to these nodes. When the nodes of
the CPD and ext densities (nearly) coincide with each
other and with the minima in the Dirac densities, a behav-
ior that is observed in Fig. 4 except in the deep core, the
minima in the CPD-Dirac and ext-Dirac curves will co-
incide and be equal to the Dirac density at those points.
This explains the behavior of the curves at the minima. As
for the maxima, we observe that the ext results oscillate
around the Dirac value, whereas the CPD results do not
deviate much from Dirac in the positive direction and the
deviation is even strictly negative in the innermost region.
Detailed explanation of this behavior can be given, but is
omitted here. In the all-electron case, one might argue that
CPD deviates less from Dirac, although one might also
prefer the smaller average deviation of the ext results. If
one includes other orbitals in the comparison (not shown),
there seems to be little ground for preference of one over
the other. In the valence-only calculations, the ext results
are maybe to be preferred because of the better agreement
in the outer region. However, the differences between CPD
and ext are so small that if one uses a basis set expansion,
they would be much smaller than basis set effects in any
but extremely large basis sets. Summarizing, we may there-
fore conclude that there is not much reason to use the ext
J_ Chem. Phys., Vol. 99, No.6, 15 September 1993
CPO-Slater equation instead of the somewhat simpler
CPD-Slater equation. Using valence-only calculations
with frozen Dirac core densities can, however, be recom-
mended.
v. CONCLUSIONS
We have explored in this paper the consequences of a
proper expansion (in a Coulomb potential) in E/(2c
2
- V)
instead of in (E- V)/2c
2
in the transformation of the
Dirac equation to a two-component form. The results show
that two-component relativistic calculations are possible
that give accurate results for the valence orbitals. To ze-
roth order, the potential dependent expansion gives a sec-
ond order differential equation (just like the Schrodinger
equation) that is variationally stable and incorporates the
relativistic effects, including spin-orbit splitting, to high
accuracy. This makes self-consistent field calculations pos-
sible that take indirect effects (relativistic screening effects
in the potential) fully into account. The results are very
promising. All-electron calculations with the new Hamil-
tonian yield agreement with Dirac one-electron energies to
0.000 1 a. u. in the upper valence region (0.1 % ). Also the
density agrees at the 0.1 % level. A voiding an approxima-
tion to the potential made in the all-electron calculations
by performing valence-only calculations using a Dirac core
potential improves the results in both energies and densi-
ties to the 0.01 % level. Our approach thus makes it possi-
ble to get "chemically" accurate results for the valence
orbitals without solving the much more complicated Dirac
equation. The calculations are only slightly more involved
than in the nonrelativistic case. Even higher accuracy can
be obtained by employing a slightly different Hamiltonian
which is based on expansion in E/(2c
2
-2V), but the ad-
ditional complications in the implementation seem hardly
warranted.
A significant advantage of the present approach is the
possibility to apply it in molecular calculations. The matrix
elements of the Hamiltonian can be calculated straightfor-
wardly with our recently developed accurate 3D numerical
integration method.
2o
The method seems potentially no
more time consuming than nonrelativistic calculations. Ba-
sis set choice and the treatment of frozen cores require
further attention and work along these lines is in progress.
IR. E. Stanton and S. Havriliak, J. Chern. Phys. 81, 1910 (1984).
2W. Kutzelnigg, lnt. J. Quantum Chern. 25, 107 (1984).
3p. J. C. Aerts and W. C. Nieuwpoort, Chern. Phys. Lett. 125, 83
(1986).
4W. Kutzelnigg, Z. Phys. D 15, 27 (1990).
5 J. D. Morrison and R. E. Moss, Mol. Phys. 41, 491 (1980).
61. J. Ketley and R. E. Moss, Mol. Phys. 48, 1131 (1983).
71. J. Ketley and R. E. Moss, Mol. Phys. 49, 1289 (1983).
8 R. E. Moss, Mol. Phys. 53, 269 (1984).
9 A. Farazdel and V. H. Smith, Jr., Int. J. Quantum Chern. 29, 311
( 1986).
10 H. Gollisch and L. Fritsche, Phys. Status Solidi B 86, 145 (1978).
llD. D. Koelling and B. N. Harmon, J. Phys. C 10, 3107 (1977).
12Ch. Chang, M. Pelissier, and Ph. Durand, Phys. Scr. 34, 394 (1986).
13p. M. Boerrigter, thesis, Vrije Universiteit, Amsterdam, 1987.
14T. Ziegler, V. Tschinke, E. J. Baerends, J. G. Snijders, and W. Ravenek,
J. Phys. Chern. 93, 3050 (1989).
15B. A. Hess, Phys. Rev. A 33, 3742 (1986).
16J. Sucher, Phys. Rev. A 22,348 (1980).
17M. Douglas and N. M. Kroll, Ann. Phys. 82, 89 (1974).
18p. Knappe and N. Rosch, J. Chern. Phys. 92, 1153 (1990).
19J. G. Snijders, E. J. Baerends, and P. Ros, Mol. Phys. 38,1909 (1979).
20G. te Velde and E. J. Baerends, J. Compo Phys. 99, 84 (1992).
21 V. B. Berestetski'f, E. M. Lifshitz, and L. P. Pitaevski'f, Relativistic
Quantum Theory (Pergamon, Oxford, 1971).
22R. McWeeny and B. T. SuthcIiffe, Methods of Molecular Quantum
Mechanics (Academic, London, 1976).
23J. J. Sakurai, Advanced Quantum Mechanics (Addison-Wesley, Read-
ing, Mass., 1967).
24W. H. E. Schwarz, E. M. van Wezenbeek, E. J. Baerends, and J. G.
Snijders, J. Phys. B 22, 1515 (1989).
25J. des Cloizeaux, NucI. Phys. 20, 321 (1960).
26L. L. Foldy and S. A. Wouthuysen, Phys. Rev. 78, 29 (1950).
27L. D. Landau and E. M. Lifshitz, Quantum Mechanics (Pergamon,
Oxford, 1977).
28R. van Leeuwen and E. van Lenthe (unpUblished).
29J. C. Slater, Phys. Rev. 81, 385 (1951).
30M. P. Das, M. V. Ramana, and A. K. Rajagopa1, Phys. Rev. A 22,9
(1980).
31 R. E. Moss, Advanced Molecular Quantum Mechanics (Chapman and
Hall, London, 1973).
32E. J. Baerends, W. H. E. Schwarz, P. Schwerdtfeger, and J. G. Snijders,
J. Phys. B 23,3225 (1990).
33 J. G. Snijders and E. J. Baerends, Mol. Phys. 36, 1789 (1978).
34p. M. Boerrigter, E. J. Baerends, and J. G. Snijders, Chern. Phys. 122,
357 (1988).

JChemPhys 99 4597

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

JChemPhys 99 4597

Uploaded by

Copyright:

Available Formats

Relativistic regular two-component Hamiltonians

E. van Lenthe, E. J. Baerends, and J. G. Snijders

You might also like