See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/215830919

On the use of the Laplace transform in local correlation methods

Article in Physical Chemistry Chemical Physics · January 2008

CITATIONS READS

10 85

3 authors, including:

Daniel Kats Denis Usvyat

Max Planck Institute for Solid State Research Universität Regensburg
58 PUBLICATIONS 1,977 CITATIONS 89 PUBLICATIONS 2,859 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Transcorrelated methods on solids View project

Laplace Transform in Local Methods View project

All content following this page was uploaded by Daniel Kats on 05 September 2016.

The user has requested enhancement of the downloaded file.

 View Article Online / Journal Homepage / Table of Contents for this issue

This paper is published as part of a PCCP Themed Issue on:

Explicit-r12 Correlation Methods and Local Correlation Methods
Guest Editors: Hans-Joachim Werner and Peter Gill
Editorial Slater-type geminals in explicitly-correlated perturbation
theory: application to n-alkanols and analysis of errors and
basis-set requirements
Explicit-r12 correlation methods and local correlation Sebastian Höfener, Florian A. Bischoff, Andreas Glöß and Wim
methods Klopper, Phys. Chem. Chem. Phys., 2008
Phys. Chem. Chem. Phys., 2008 DOI: 10.1039/b803575j
DOI: 10.1039/b808067b
Accurate calculations of intermolecular interaction energies
Papers using explicitly correlated wave functions
Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.

Oliver Marchetti and Hans-Joachim Werner, Phys. Chem. Chem.

Implementation of the CCSD(T)-F12 method using cusp
Phys., 2008
conditions
DOI: 10.1039/b804334e
Denis Bokhan, Seiichiro Ten-no and Jozef Noga, Phys. Chem.
Chem. Phys., 2008
Variational formulation of perturbative explicitly-correlated
DOI: 10.1039/b803426p
coupled-cluster methods
Martin Torheyden and Edward F. Valeev, Phys. Chem. Chem.
Analysis of non-covalent interactions in (bio)organic
Phys., 2008
molecules using orbital-partitioned localized MP2
DOI: 10.1039/b803620a
Stefan Grimme, Christian Mück-Lichtenfeld and Jens Antony,
Phys. Chem. Chem. Phys., 2008
Resolution of the identity atomic orbital Laplace transformed
DOI: 10.1039/b803508c
second order Møller–Plesset theory for nonconducting
periodic systems
Tighter multipole-based integral estimates and parallel
Artur F. Izmaylov and Gustavo E. Scuseria, Phys. Chem. Chem.
implementation of linear-scaling AO–MP2 theory
Phys., 2008
Bernd Doser, Daniel S. Lambrecht and Christian Ochsenfeld,
DOI: 10.1039/b803274m
Phys. Chem. Chem. Phys., 2008
DOI: 10.1039/b804110e
On the use of the Laplace transform in local correlation
methods
Local correlation domains for coupled cluster theory: optical
Danylo Kats, Denis Usvyat and Martin Schütz, Phys. Chem. Chem.
rotation and magnetic-field perturbations
Phys., 2008
Nicholas J. Russ and T. Daniel Crawford, Phys. Chem. Chem.
DOI: 10.1039/b802993h
Phys., 2008
DOI: 10.1039/b804119a
Intracule densities in the strong-interaction limit of density
functional theory
Local and density fitting approximations within the short-
Paola Gori-Giorgi, Michael Seidl and Andreas Savin, Phys. Chem.
range/long-range hybrid scheme: application to large non-
Chem. Phys., 2008
bonded complexes
DOI: 10.1039/b803709b
Erich Goll, Thierry Leininger, Frederick R. Manby, Alexander
Mitrushchenkov, Hans-Joachim Werner and Hermann Stoll,
Intracule functional models
Phys. Chem. Chem. Phys., 2008
Part III. The dot intracule and its Fourier transform
DOI: 10.1039/b804672g
Yves A. Bernard, Deborah L. Crittenden and Peter M. W. Gill,
Phys. Chem. Chem. Phys., 2008
Equations of explicitly-correlated coupled-cluster methods
DOI: 10.1039/b803919d
Toru Shiozaki, Muneaki Kamiya, So Hirata and Edward F.
Valeev, Phys. Chem. Chem. Phys., 2008
Density matrix renormalisation group Lagrangians
DOI: 10.1039/b803704n
Garnet Kin-Lic Chan, Phys. Chem. Chem. Phys., 2008
DOI: 10.1039/b805292c
Vanadium oxide compounds with quantum Monte Carlo
Annika Bande and Arne Lüchow, Phys. Chem. Chem. Phys.,
The interaction of carbohydrates and amino acids with
2008
aromatic systems studied by density functional and semi-
DOI: 10.1039/b803571g
empirical molecular orbital calculations with dispersion
corrections
Second-order Møller–Plesset calculations on the water
Raman Sharma, Jonathan P. McNamara, Rajesh K. Raju, Mark A.
molecule using Gaussian-type orbital and Gaussian-type
Vincent, Ian H. Hillier and Claudio A. Morgado, Phys. Chem.
geminal theory
Chem. Phys., 2008, 10, 2767
Pål Dahle, Trygve Helgaker, Dan Jonsson and Peter R. Taylor,
Phys. Chem. Chem. Phys., 2008
The principle-quantum-number (and the radial-quantum-
DOI: 10.1039/b803577f
number) expansion of the correlation energy of two-electron
atoms
Werner Kutzelnigg, Phys. Chem. Chem. Phys., 2008
The - states of the molecular hydrogen
DOI: 10.1039/b805284k
Jacek Komasa, Phys. Chem. Chem. Phys., 2008
DOI: 10.1039/b803548b
 PAPER
On the use of the Laplace transform in local correlation methods
Danylo Kats, Denis Usvyat and Martin Schütz*
Received 20th February 2008, Accepted 27th March 2008
First published as an Advance Article on the web 20th May 2008
DOI: 10.1039/b802993h
Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.
The applicability of the Laplace transform ansatz of Almlöf in the context of local correlation
methods with a priori restricted sets of wavefunction parameters is explored. A new local MP2
method based on the Laplace transform ansatz is described, its relation to the local MP2 method
based on the Pulay ansatz is elucidated, and its accuracy and efficiency are compared to the latter.
1. Introduction
During the last decade much progress has been made in
developing high-level electron correlation methods applicable
to extended molecular systems. Among the most successful
ones in terms of efficiency are the local correlation methods
based on the ansatz of Pulay.1 In order to exploit the short-
range nature of dynamic correlation effects in nonmetallic
systems the de-localized canonical orbitals spanning the occu-
pied and virtual space are substituted by (i) localized mole-
cular orbitals (LMOs) obtained from the occupied canonical
MOs by unitary transformation, and (ii) by projected atomic
orbitals (PAOs) generated by projecting the AOs onto the
virtual space. Local approximations on the relevant excited
configuration functions, i.e., the related wavefunction para-
meters (amplitudes) in this local basis are then imposed a priori
based on spatial locality arguments. This is achieved by
specifying, e.g., restricted LMO pair lists and LMO/pair
specific excitation subspaces of PAOs (excitation domains).
Based on these ingredients O(N) methods were developed for
local second-order Møller–Plesset perturbation theory
(LMP2),2 local coupled cluster singles and doubles (LCCSD),3
and LCCSD with triples corrections.4,5 Particularly efficient
algorithms were devised by combining local correlation with
density fitting (DF).6–8 Recently, also local methods based on
time-dependent coupled cluster (CC) response theory, capable
of treating excited state properties of extended molecular
systems,9,10 as well as local MP2 methods for infinite periodic
systems11,12 were presented.
The local correlation methods are all based on an orbital
invariant formalism of the related underlying electron correla-
tion theory. Such an orbital invariant formalism, e.g., for
MP2, is straightforwardly devised by perturbative analysis of
the (orbital invariant) CC equations. Yet, in comparison to the
canonical formulation some complication arises. Due to the
non-diagonality of the Fock operator in a general (local)
orbital basis the first-order amplitudes no longer decouple
and a linear equation system has to be solved (usually by an
iterative procedure) to get the amplitudes and from those the
second-order energy. In a canonical formulation, on the other
Institute of Physical and Theoretical Chemistry, University of
Regensburg, Universitätsstraße 31, D-93040 Regensburg, Germany
3430 | Phys. Chem. Chem. Phys., 2008, 10, 3430–3439
www.rsc.org/pccp | Physical Chemistry Chemical Physics
hand, the second-order energy can be calculated directly.
Similar complications with regard to the canonical formula-
tion occur also for the CC2 model and the related response
theory, and for perturbative triples corrections.
An alternative orbital invariant formalism for MP2, pio-
neered by Almlöf and Häser, is based on the Laplace trans-
form of the denominators in the expression for the canonical
MP2 amplitudes.13,14 Here, instead of solving a linear equa-
tion system a numerical quadrature of the Laplace integral
must be performed, the latter requiring only a few quadrature
points. So far, the Laplace transform approach has been
primarily utilized to devise formulations of MP2 theory in
AO basis representation.15–18 Some studies on the use of the
Laplace transform technique in the context of other localized
orbitals have also been reported.19,20 Recently, Nakajima and
Hirao presented an implementation of a Laplace MP2 method
based on localized orbitals and DF.21
A major disadvantage of the Laplace transform ansatz,
present in all these methods, is the requirement of an indivi-
dual ERI evaluation/transformation for each quadrature
point, even though the related transformation matrices, re-
sembling density matrices, are sparse. In the present work we
develop a local correlation method for MP2 theory along the
lines of the Pulay ansatz based on the Laplace transform of
canonical MP2 theory. In contrast to other methods the
required ERI sets for the individual quadrature points are
generated from an initial compact integral set in local basis by
transforming the PAOs within the related excitation domains.
These transformations are rather inexpensive and no new ERI
evaluation in AO basis is needed for the individual quadrature
points. DF is also utilized to make the method more econom-
ical. Overall, the new Laplace transformed DF-LMP2
(LT-DF-LMP2) method is competitive to DF-LMP2, in par-
ticular for extended domains, where the linear equation system
in the latter becomes more expensive to solve. A local Laplace
transform method for scaled opposite-spin MP2 has recently
been presented by Jung et al.22,23 In contrast to our method,
orthogonalized localized orbitals rather than non-orthogonal
PAOs were used to span the virtual space, and (ii) locality was
exploited entirely based on screening, not on a priori specified
domains. The main motivation of the present work was not to
propose yet another MP2 method for extended systems, but to
work out a general methodology on how to use the Laplace
This journal is

c the Owner Societies 2008
 transform in the context of local correlation methods based on whereas the transformation matrix Q = CvwSAO8 transforms
the Pulay ansatz, i.e., with a priori specified domains, which canonical virtuals to PAOs,
can then be easily adapted to other local methods like CC2,
time-dependent CC2, local triples, or periodic local MP2. The X
theory of Laplace transformed local correlation methods fPAO
a
¼ jfCAN
a iQa
a : ð5Þ
based on pre-defined excitation domains with lower rank than a

the space of the canonical virtuals is less straightforward than In the definition of Q the matrix Cv is the virtual block of the
for methods where the rank of the space spanned by the local
Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.

(canonical) MO-coefficient matrix and SAO is the AO overlap

functions is either larger than or equal to that of the canonical matrix. Here and in the following we use barred indices i,
j,.

..
virtuals like in the previous methods. This issue is outlined in
. . for LMOs and PAOs, respectively. The matrix
and ā, b,.
the following section. product
QwQ = SPAO (6)
2. Theory yields the metric of the PAOs. Using these transformations the
relations
Formalism of Laplace transformed DF-LMP2 X y y
ði

¼

ajj
bÞ Wii
Qa
a ðiajjbÞWjj
Qbb
ð7Þ
As shown by Almlöf13 the energy denominator
ijab
Dab
ij = ea  ei + eb  ej appearing in the energy expression
of canonical MP2, X


X ij ij
ij
Tab ¼ Wii
Qa
a Ta
ijb
Wjj
y Qybb

ð8Þ
ij
E ð2Þ ¼ K~ T ; K~ ¼ 2ðiajjbÞ  ðibjjaÞ;
ab ab ab
i
j
a
b

ijab
straightforwardly follow. Eqn (8) is obtained from the expres-
sion for the local correlation energy,
ij ðiajjbÞ
Tab ¼ ð1Þ X i
j
i
j

Dab
ij
ð2Þ
ELOC ¼ K~ a
b
Ta
b
; ð9Þ
R1 i
j
a
b

is eliminated via the identity 1=x ¼  0 ext dt,
Z 1 and eqn (7). Integrals and amplitudes always transform con-
ab
ij
Tab ¼  ðiajjbÞeDij t dt: ð2Þ versely, as covariant and contravariant objects. Relation (7)
0
also holds for the case when the PAO indices ā, b
are confined
The indices i,j,. . . and a,b,. . . refer to canonical occupied and to excitation domains, i.e., when the number of PAOs ā, b
is
virtual orbitals, respectively. The ei, ea, etc. in Dab ij denote smaller than the number of canonical virtuals. Yet the con-
related orbital energies. The (ia|jb), finally, represent the usual versely transforming ‘canonical’ amplitudes of (8) are not
electron repulsion integrals (ERIs). The Laplace integration in identical to the canonical amplitudes of eqn (2), but represent
eqn (2) can be replaced by a numerical quadrature, a fit to those within the specified local basis (vide infra), and the
nq
X local MP2 energy defined in (9) deviates from the canonical
ij
Tab ¼ wq eðei eF Þtq eðea þeF Þtq ðiajjbÞ MP2 energy of eqn (1).
q The transformation of eqn (3) to the local basis, i.e., the
ð3Þ inversion of eqn (8), is not straightforward due to two reasons:
nq
X
e ðej eF Þtq ðeb þeF Þtq
e ¼ wq Tijab ðqÞ: in contrast to AOs the PAOs form a non-orthogonal redun-
q dant set of functions. Furthermore, the PAOs are usually
restricted to excitation domains and are therefore spanning a
eF = (eHOMO + eLUMO)/2 cancels in eqn (3) and ensures that
space with lower dimensions than that of the canonical
the individual exponential factors are always smaller than one.
virtuals. In order to investigate this transformation we intro-
Tijab(q) is implicitly defined in (3) and has the meaning of an
= QV with property
duce the matrix Q
incremental amplitude with regard to the Laplace quadrature, X y
rather than that of an ERI. The number of quadrature points
a
a hfPAO j ¼ hfPAO j;
Qc
a Q ð10Þ
a
c

nq required to achieve mH accuracy is rather small (5–7). A a

a

short discussion about how to obtain points tq and weights wq with fPAO
ā and fPAO
c
belonging to the same domain. Multi-
for the quadrature is given below. plying (10) with the ket PAO fPAO
d
(again of the same domain)
In order to develop a formalism for local correlation and integrating yields
methods we introduce the usual transformation matrices con-
necting LMOs and PAOs with canonical occupied and virtual QwQS

PAO = SPAOVSPAO = SPAO, (11)
orbitals. The unitary transformation matrix W specified by the
chosen localization criterion (Boys,24 Pipek-Mezey,25 etc.) defining the unknown matrix V as a pseudoinverse S+ PAO of

transforms canonical occupied orbitals to LMOs, SPAO. V is obtained by (i) diagonalizing SPAO, (ii) inverting
elements of the diagonal matrix so obtained, which are above
a given threshold defining redundancies, and zeroing the
X remaining ones, and (iii) back-transforming of the diagonal
fLMO
i
¼ jfCAN
i iWii
; ð4Þ matrix. We emphasize at this point, that if PAO domains are
i

This journal is

c the Owner Societies 2008 Phys. Chem. Chem. Phys., 2008, 10, 3430–3439 | 3431
 introduced, V is domain specific (even though we omit the formula
explicit dependence on domains at the present stage), while Q X X
j
b

ði

¼

ajj
bÞ ði
1

ajPÞJPQ
¼
ðQjj
bÞ Bi

a
is not. For a transformation involving Q the columns related P BP ; with
PQ P
to the relevant domain are just extracted from Q. Now
utilizing eqn (10) the relations

X
X Bi

a
P ¼

ajQÞðQjPÞ1=2
ði
ð18Þ

¼

ajj
bÞ
ði

ði

ðQy QÞ



y
a
c
cjj dÞðQ QÞd
b
ð12Þ Q
c
d

can be used, while otherwise, e.g. in case of local fitting,26,27
Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.

X y the three term formula has to be utilized. In eqn (18) indices

Ta
ijb
¼
QÞ T i
j

ðQy QÞ
ðQ


ð13Þ
a
c
c
d db
P,Q,. . . denote functions of the auxiliary basis set with Cou-
c
d

lomb metric JPQ = (P|Q). Inserting (18) in (16) yields
straightforwardly follow. Exploiting (13) and unitarity of W XX y



j
d

eqn (8) can now be transformed with Q
and W and yielding Tai
jb
ðqÞ ¼ Va
c
Bi

c
P ðqÞBP ðqÞVd
b
; with ð19Þ
X X c
d
P



Ta
ijb
¼ Wii
y Va
yc
Qyc
a Tab
ij
Wj j
Qbd
Vd
b
: ð14Þ
X
ijab c
d




Bi

a
P ðqÞ ¼ Xi
ok
ðqÞXa
vc
ðqÞVc
d
BPkd : ð20Þ

cd

k

Eqn (14) is valid for both the true and the fitted canonical
amplitudes. If instead of the MP2 the scaled opposite-spin MP2 energy is
The related ERIs again transform conversely, requested, then the energy of the related quadrature point can
XX
ðiajjbÞ ¼ Wii
Qa
c Vc
a
ði


y V y

Qy
:

ajj
bÞW ð15Þ be computed with a nominal scaling of the cost of O(N4)
jj bd db
i
j
a
b
c
d
instead of O(N5) (cf. ref. 22 for details).
The whole formalism developed so far is by construction
Combining eqns (3), (14), and (15) leads to also valid for the case of PAO domains, i.e., if the number of
nq
X PAOs involved in the transformations is smaller than the





Ta
ijb
¼  wq Tia
jb
ðqÞ; with ð16Þ number of canonical virtuals. At this point we now want to
q specify our local approximation: as in ordinary local MP2 we
X impose a restricted pair list and pair domains to confine the







Tia
jb
ðqÞ ¼
cjl
dÞX
Xi
ok
ðqÞYa
vc
ðqÞðk


o
ðqÞY v

ðqÞ;
lj db
amplitudes Ta
ijb
, and accordingly Tai
jb
. Then the matrices V
k
l


cd
appearing in eqn (19) are actually pseudoinverses of the PAO
overlap matrices restricted to the ½i
j
pair domains, which we
where we have introduced the quadrature point dependent


denote as Vdi
jb
. Eqn (19) explicitely written for the specified
matrices
X y domains takes the form
Xi
oj
ðqÞ ¼ Wii
eðei eF Þtq Wij
;

X X i
jy

c j
d
i
j

i Tai
jb
ðqÞ ¼ Va
c
Bi

P ðqÞBP ðqÞVd
b
: ð21Þ

i
j
c
d2½
P
X
Xa
vb
ðqÞ ¼ Qya
a eðea þeF Þtq Qab
;
c|P) (and hence for the Bk

c
a
For the ERIs (k
P ), on the other
hand, we do not explicitly introduce any restriction on the set
X of indices. Consequently, the matrix V of eqn (20) is the
Ya
vb
ðqÞ ¼ Va
yc
Xc
vd
Vd
b
: ð17Þ
c
d

pseudoinverse of the non-truncated PAO overlap matrix.

a
In order to achieve higher efficiency in computing the Bi
P
Eqns (16) and (17) show that the compact ERI set (k

c|l
d)

in quantities the sparsity of various matrices can be utilized. For
local basis can be re-used at each quadrature point by applying example, since LMOs and PAOs k
and c
are localized func-

c
a sequence of rather inexpensive transformations within the tions the ERIs (k

c|P) and therefore the Bk


P for a given k have
PAO
local basis. No new AO integral evaluation and (AO to MO or negligibly small values when fc
is spatially far from fLMO
k
.
o v
AO to AO) transformation is required for each quadrature The matrices X (q) and X (q) are also sparse. This is due to the
point. fact that these quantities are analogous to the exponentially
In contrast to ordinary local MP2 the local approximation decaying Hartree–Fock density matrices for the occupied and
in the amplitudes does not carry over to the ERIs, as is clearly virtual spaces, but with Laplace-weighted occupation num-
evident from eqn (16), i.e., introducing a priori a restricted pair bers.15,16 Yet this sparsity might depend on the quadrature
list fi
jg
for T i
j

does not imply a

and pair domains ā, b
A ½i
j point tq. This will be explored further for a test case in
a
b
priori restrictions on the pairs {kl} and the PAO ranges c
, d
of


section 3.
the ERIs. This has some important consequences distinguish- The computational cost of calculating the pseudoinverse V
ing the local Laplace approach from ordinary local correlation of the non-truncated PAO overlaps the matrix scales as
methods (vide infra). O(N3). Another O(N3) step (with a larger prefactor) is the
We now proceed by invoking the density fitting approxima- inversion of the Coulomb metric for density fitting (18). In
tion to the ERIs. We assume that the auxiliary basis set is order to reduce the scaling of the latter either attenuated
always identical for all orbital pair densities in the ERIs. If the Coulomb fitting (O(N2), cf. ref. 23) or local fit domains
Coulomb metric is employed for such a case the one-term (O(N), cf. refs. 26 and 27) could be introduced. In our present

3432 | Phys. Chem. Chem. Phys., 2008, 10, 3430–3439 This journal is

c the Owner Societies 2008
 program we have not yet implemented any of these schemes. Discrete quadrature of Laplace integral
The O(N3) scaling of calculating the pseudoinverse V can be
Following Häser and Almlöf14 the points tq and weights wq for
avoided by confining the PAO index c
of the ERIs

¼ S ½k
l,
½UðkÞ
the union of all pair domains [k
l]
with the discrete quadrature of the Laplace integral are obtained
8l from a least-squares fit,
common k.
These, dictated by the truncated pair list, reflect
the r6 decay behavior of the pair energies and safely cover the Z xmax nq
!2
1 X xtq
relevant range of the exponentially decaying ERIs. Then the f ðxÞ  wq e ¼ min ð23Þ
xmin x q
dimensions of the pseudoinverses are confined to the related
Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.

orbital specific united pair domains and the computational

xmin and xmax are the minimal and maximal energy denomi-
cost of their construction decreases to O(N). In our test
nators of all Dabij , respectively. f(x) is a distribution function
calculations such a restriction of the integrals to the corre-
counting the Dabij in the vicinity of a certain x. It is constructed
sponding united domains had only a negligible effect on the
by generating a histogram of all Dai = ea  ei based on 600
energy.
equidistant bins subdividing the interval between the min-max
The MP2 energy is finally obtained as
values of Dai. The two identical histograms reflecting the
nq
X X X

X distributions of the Dai and the Dbj are then added to a new
ð2Þ ij



jb
ELOC ¼  wq L~a
b
Bi

a
P ðqÞBP ðqÞ; with histogram based on 1200 equidistant bins subdividing the
q i
j2f

i
jg
i
j

a
b2½
P interval between xmax and xmin. The final histogram so ob-
tained then corresponds very closely to the desired distribution



function f(x), which would be obtained by generating a
ij

i
j

L~a
b
¼ 2Lia
jb
 Lb


a ; and histogram of the Dab ij directly. The computational cost of
forming f(x) in that way however scales as O(N2) rather than
! O(N4) and in practice takes only a negligible amount of time.


X

X
c j
d


In our test calculations the use of an f(x) reflecting the
Lia
jb
¼ Va
ijc
Bi

P BP Vdi
jyb
: ð22Þ

i
j

P
distribution of the Dab ij as described turned out to be superior
c
d2½
to the simpler uniform distribution, i.e., f(x) = 1, certainly so


for small nq.
The L~ij matrices do not depend on the quadrature point and
The tq are obtained by using a Simplex algorithm, the
have to be computed just once at the beginning. After that also


related weights needed in each Simplex step by solving the
the V ij matrices can be discarded. Hence, just one four-
set of linear equations obtained from (23) by differentiation
indexed quantity in local basis must be kept, in comparison
with regard to wq (for details see ref. 14). The necessary
to at least three in ordinary DF-LMP2. Moreover, for a given


numerical integrations are carried out by subdividing the
quadrature point just a single pass through the list of L~ij
interval of integration defined by xmin and xmax into 1200
matrices is required, whereas in the original DF-LMP2 multi-
subintervals (logarithmic scale). The related f(x) for a certain x
ple passes through the list of amplitude matrices are needed for
was obtained by linear interpolation. For tq = 0 we have
each iteration to calculate the coupling between amplitudes via
a
a
Bi
i

P ð0Þ = BP (this immediately follows from eqns (12), (17)
the Fock matrix2 (see also eqn (26)).
and (19)), thus, the energy contribution for tq = 0 can be
It can also be seen from eqn (22) that even for the case of an

c|l
d)

in eqn (16) the PAO index ā of computed virtually for free as a by-product of the initial
unrestricted set of ERIs (k




a
i
calculation of L~ia
jb
. For that reason, the tq = 0 point is always
the quantity BP ðqÞ entering the assembly step is still confined
included and kept fixed in the Simplex minimization.
to the pair domain ½i
j.

A priori restriction of the ERI set, i.e.,
restricting index d
in eqn (20) to [U(k)]
helps reduce the cost Laplace transformed vs. original LMP2

a
for computing the Bi
P ðqÞ intermediates, and, especially, for
The Laplace transformed and the original local MP2 method
getting the related pseudoinverse. On the other hand, the
differ in some interesting aspects. A common basis for both is
scheme with a completely unrestricted set of the ERI benefits
the equation
from the possibility of pre-multiplying XvV before transform-
ing the three-index object BkPd .


X


Rijab ¼ Wii
Qa
a Ta
ijb
Qybb y

Wjj


However independent of all that, the essential point here is i
j
a
b

that all the assembly steps in our algorithm always go just over ð24Þ
the pair domains and restricted pair lists, whether the ERI 1 X
a
a ði

Q

ajj
bÞ y


W
y :
þ Wii
Q bb

c|l
d)
(k

in eqn (16) is a priori truncated to the united domain, or Dab jj
ij i
j
a
b

not, and are hence always computationally inexpensive.
The algorithm as outlined scales nominally cubically (calcu- Provided that the local ERI set ði

(but not the local

ajj
bÞ
i
j

lation of the inverse of the Coulomb metric for DF which is amplitude set Ta
b
) is non-truncated, i.e., corresponds to full
also independent of the quadrature point), yet O(N) scaling domains, then the second term of eqn (24) is identical to the
could be achieved by introducing local fit domains.26,27 More- canonical amplitudes and Rijab represents the difference be-
over, the Laplace scheme allows for an easy parallelization of tween the genuine canonical amplitudes and the local ampli-
the code by a partitioning of the quantities occurring in the tudes in canonical basis, obtained according to eqn (8). Except
energy expression (22), which is lacking any coupling between for full amplitude domains, the Rijab cannot be strictly zero,
large objects. since the number of the equations exceeds the number of the

This journal is

c the Owner Societies 2008 Phys. Chem. Chem. Phys., 2008, 10, 3430–3439 | 3433
 variables. Dab ij
ij Rab is the canonical MP2 residual (with regard to the weighted residuals, and, for most basis sets, to an energy
the fitted amplitudes obtained from the local ones) entering the slightly closer to that of canonical MP2 than the original local
canonical Hylleraas functional. method provides.
In the Laplace transformed LMP2 method presented here
P LT-LMP2 Lagrangian and the first order properties
the functional ijab(Rijab)2 is minimized with regard to the local
amplitudes, As mentioned above, application of the Laplace transforma-
P ij 2
tion in the context of the local scheme, i.e. for a truncated set
d ijab ðRab Þ
X
Wii
y Qya
a Rijab Qbb
Wjj
: of amplitudes, implies fitting of the latter to the full set of
Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.

0¼

¼2 ð25Þ
dðTa
ijb
Þ ijab canonical amplitudes. This implies that this method is no
longer solely based on perturbation theory (in a reduced local
Note that the PAO and LMO indices of the transformation space), but rather is a fit to results of perturbation theory
matrices are restricted to the space of the local amplitudes. carried out in the non-truncated first-order interacting space.
Dividing eqn (25) by two, multiplying it from the left and the The Hylleraas functional is hence not a proper Lagrangian for
right with Vw and V, respectively, inserting Rijab as defined in the LT-LMP2 energy. The LT-LMP2 Lagrangian can be
(24), making use of eqn (13), and finally Laplace transforming specified instead as
the denominator leads to the LT-LMP2 amplitude equations !
X i
j
i
j
0 ij 0 ij
ij
Kab
given in eqn (16). The LT-LMP2 therefore corresponds to a 0 0
L½T; K; T ; K  ¼ ~
K a
b
Ta
b
þ Lab Tab þ ab
least-squares minimization of the canonical MP2 residual, i
j
a
b

Dij
weighted with the inverse of the related denominator. Ampli- !
tudes corresponding to a small denominator (i.e. correspond- X



X y 0 ij
þ Lia
jb
Ta
ijb
 Wii
y


Qa
a Tab Wj j
Qbb
;
ing to the HOMO–LUMO region) hence have a larger weight
i
j
a
b
ijab
and are fitted more accurately.
The original LMP2, on the other hand, corresponds ð27Þ
to a projection onto the space of the local amplitudes of where T 0 and T represent the canonical and fitted MP2
eqn (24) after multiplying it with the denominator. amplitudes, respectively. The first term of the right-hand-side
Multiplication with Ww, Qw from the left, and W, Q from the of (27) corresponds to the LT-LMP2 energy, the second
right then yields comprises the conditions for the (canonical) MP2 amplitudes
X y y (1), while the third term refers to the fit of the canonical
Wii
Qa
a ðDab ij
þ ðfT i
j
SÞ


ajj
bÞ
ij Rab ÞQbb
Wj j
¼ ði
a
b
ijab
amplitudes by the local ones and represents the set of eqns
X ð26Þ (25). The multipliers of the second and the last term obviously







þ ðST ij f Þa
b
 ðSðfi
k
T kj þ fk
j
T ik ÞSÞa
b
: live in the non-truncated canonical and in the local restricted
k space, respectively.
Differentiation of L[T,K,T 0 ,K 0 ] with regard to the ampli-
If the local ERIs in the second term of eqn (24) are non- tudes T yields the equations for the multipliers K simply as
truncated then the transformed ERIs correspond to the cano-





nical MO integrals and eqn (7) can be utilized to expose the Lia
jb
¼  K~a
ijb
: ð28Þ
local ERIs in eqn (26). The very same equations however are
also obtained if these ERIs are restricted to the space of the Note that the multipliers and the contravariant integrals
local amplitudes and eqn (12) can be used instead. After in (28) are confined to the restricted local space. Finally,
projection, the ERIs outside the space of the local amplitudes, differentiation with regard to the canonical amplitudes
even if sizeable, cannot influence the local amplitudes. The T 0 provides a set of equations relating the two sets of multi-
usual LMP2 amplitude equations are obtained from (26) by pliers T 0 and T,
requiring that the canonical MP2 residual Dab ij
ij Rab, projected 0 X
Labij ¼
a
a Li
j

W
y Q
Wii
Q
y
; ð29Þ
onto the local space, is zero. Residual elements outside the a
b jj bb
i
j
a
b

local space are entirely neglected even though there is no
guarantee that they are negligible (in particular for small which can be used to eliminate the canonical amplitudes and
Boughton–Pulay (BP) domains). multipliers in the Lagrangian (27). This yields (after invoking
The Laplace transformed LMP2 presented here and the the Laplace identity)
original LMP2 are therefore two non-equivalent local methods !
X i
j
i
j
nq
X
based on a priori restrictions of the wavefunction parameters.






L½T; K ¼ K~

T
þ Lij
T ij
þ
ab a
b
wq Tij
ðqÞ : ð30Þ
a
b a
b a
b
The former corresponds to a weighted least-squares fit of the i
j
a
b
q
canonical amplitudes, the latter to a projection of the canoni-
cal residual onto the specified local space (in order to reduce and the condition specifying the local amplitudes is then
the number of equations to the number of variables) with the identical to eqn (16).
requirement that the projected canonical residual is strictly In order to introduce a (one-body) perturbation into the
zero. For the Laplace transformed LMP2 none of the residual Hamiltonian the condition specifying the canonical ampli-
elements is strictly zero yet those outside the local space are tudes of L[T,K,T 0 ,K 0 ] is augmented by the two connected
also required to be as small as possible and are not left un- diagrams contracting the perturbation with the canonical
controlled. This corresponds to a better fit of the whole set of amplitudes, which are divided by the energy denominator,

3434 | Phys. Chem. Chem. Phys., 2008, 10, 3430–3439 This journal is

c the Owner Societies 2008
 i.e., [VT 0 ]ijab/Dab
ij . The derivative of this Lagrangian with respect
to the perturbation strength at zero perturbation strength then
leads to an expression for the LT-LMP2 density matrix which
contains the canonical first-order amplitudes of the un-
restricted space T 0 , i.e.,
0 1
X X i
j

Dba ¼ 2 @ y

a
a K~
W
Q
Wii
Q

AT 0 ij ;
y i
j

a
b jj bc bc
Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.
ijc i
j
a
b

0 1
X X i
j

Dji ¼ 2 @ Wii
Q

AT 0 jk :

a
a K~
W
y Q y
ð31Þ
a
b jk bb ab
kab i
j
a
b

Eqn (31) is therefore impractical for extended systems.
The occurrence of the canonical amplitudes in eqn (31)
can be avoided by substituting the genuine canonical
amplitudes in the perturbation diagrams by their fit through
the local amplitudes in the canonical basis. The structure
of the Lagrangian given in eqn (27) is then not altered,
but the influence of the perturbation on the canonical
amplitudes is approximated (note that in the case of the
LMP2 method both the perturbation and the amplitudes
refer to the restricted local space). As a result of this
approximation the genuine canonical amplitudes T0 in eqn
(31) are replaced by the amplitudes T of the local space,
transformed to the canonical basis according to eqn (8). The
resulting LT-LMP density matrix, in contrast to MP2 or
LMP2, is non-symmetric, as is generally the case for coupled
cluster theory. This is a consequence of the two different sets of
ERIs (truncated and non-truncated) which appear in the
first and second term of the Lagrangian, eqn (27), respectively,
and implies that two different sets of amplitudes have to be
determined. This is, of course, a severe disadvantage of
LT-LMP2 compared to LMP2, yet if the Laplace approach
is employed in the context of a method like CC2
then amplitudes and multipliers have to be determined sepa-
rately anyhow.
Symmetric LT-LMP2 density matrices are obtained if the
full set of ERIs appearing in the second term of eqn (27) is
replaced by its fit through a truncated local set of ERIs, again
transformed to the canonical basis according to eqn (15). The
local space of this ERI set must coincide with that of the local
amplitudes. In such a case only one set of amplitudes has to be
determined and the resulting LT-LMP2 density matrix is
symmetric. On the other hand, when density fitting is em-
ployed for the ERIs, it is not convenient to insist on an
identical truncation of amplitudes and ERIs, since a factoriza-
tion of the local ERIs, as exploited in eqn (21), is then not
possible (pair domain restrictions cannot be applied for the
relevant three-index quantities). As a consequence the quad-
rature-dependent assembly step of (21) has to be replaced by a
quadrature-dependent four-index transformation of the local
ERI set. Moreover, the results obtained for the properties are
anticipated to be of lower quality if the full set of ERIs is
replaced by a set of ERIs truncated in the same strict manner
as the amplitudes. Instead, an alternative symmetrical ap-
proach is proposed here, where the amplitude set related to
the truncated set of ERIs is replaced by that related to the full
ERI set. The expression for the density matrix in the local
This journal is

c the Owner Societies 2008
basis then takes the form
X


jk


Di
j
¼ 2 T~a
b
Sa
c
Tc
ikb
Sd
b
;

ab

k

cd

X


ij


Da
b
¼ 2 T~b
d
Sd
c
Ta
ijc
; ð32Þ

cd

which from its structure is identical to that of the original
LMP2 method.
Such a definition of the density matrix is related to multi-
pliers which do not satisfy the stationary conditions for the
Lagrangian, hence the first-order properties cannot strictly be
considered as the derivatives of the LT-LMP2 energy with
regard to the corresponding perturbation. This implies that
such a strategy cannot be used in the context of nuclear energy
gradients used for geometry optimizations. On the other hand,
(i) the deviations from the proper derivative of the LT-LMP2
energies are small (especially for extended domains), and (ii)
the resulting properties are slightly closer to the corresponding
canonical result than either those obtained from LMP2 or
from the non-symmetric LT-LMP2 approach (since the un-
restricted set of ERIs obviously provides a better fit for the
canonical amplitudes than the restricted set). A first illustra-
tion of this issue is given in the next section.
3. Test calculations
Sparsity of transformation matrices
For an efficient algorithm the sparsity of the transformation
matrices Xo(q), Xv(q) as defined in eqn (17) and of the matrix
product Xv(q)V occurring in eqn (20) is of a major concern.
The Xo(q) and Xv(q) matrices are essentially the density
matrices of the occupied and virtual space in LMO and
PAO basis, respectively, but with exponentially weighted
occupation numbers.16 In order to analyze the sparsity of
such matrices in a large molecule a crosslink to solid state
theory may be helpful.15 It has been shown by Des Cloizeaux
in ref. 28 that the density matrix in nonmetallic systems decays
exponentially and the rate of this decay depends on the width
of the band gap. The density matrix for the virtual states
should behave in a similar way due to the finite size of the basis
set. The decay rates of Xo(q) and Xv(q) for different tq points
should therefore remain the same, yet the prefactor and
therefore the actual sparsity of the matrices may differ.
Trivially, for the point tq = 0 these matrices correspond to
the identity and the PAO overlap matrix, respectively. For
non-zero, but small tq the prefactor may increase since the
Laplace exponentials are cutting out regions with very low (for
Xo(q)) or very high (for Xv(q)) energies. For larger tq the
exponential factor becomes very small and the sparsity of the
matrices increases again.
We have investigated the sparsity of the Xo(q), Xv(q) ma-
trices and the Xv(q)V matrix product in detail for the case of a
solvent cluster consisting of a phenothiazine-phenyl-isoallox-
azine dye dyad solvated by 20 acetonitrile solvent molecules
(cf. Fig. 1). This solvated dye dyad has been investigated
before in our group in the context of an application project
on artificial blue-light photoreceptors.29 In Fig. 2 the mean
Phys. Chem. Chem. Phys., 2008, 10, 3430–3439 | 3435

Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.
Fig. 1 A cluster of a phenothiazine-phenyl-isoalloxazine dye dyad
solvated by 20 acetonitrile solvent molecules (173 atoms in total).
values of the elements Xi
oj
ðqÞ and Xvāb
(q) as a function of the
spatial separation of the functions fLMO , fLMO , and fPAO ,
i
j
ā
PAO
fb
, respectively, are plotted.
For all calculations presented here the maximum of the
quadrature points max(tq)8q remains below 10 Eh1, hence, for
Fig. 2, representative tq values of 0 (0.01 for Xo(q)) and 10
Eh1 were taken. Evidently, the transformation matrices Xo(q)
and Xv(q) all decay exponentially according to (nearly) the
same rate which is governed by the HOMO–LUMO gap. All
matrices appear to be reasonably sparse. The sparsity of the
matrix Xv(q) for a moderately small tq is similar to that of the
PAO overlap matrix (tq = 0). For tq = 10 Eh1 the elements of
Xv(q) become small due to the larger exponents in the ex-
ponential factors. The sparsity of the Xo(q) matrix is decreas-
ing on going from tq = 0.01 to tq = 10 Eh1, but Xo(q)
remains more sparse than Xv(q). For the whole relevant range
of the tq the sparsity of both matrices is sufficient to allow for
an efficient pre-screening in the transformation steps. Test
calculations show that the pre-screening of transformations
Fig. 2 Mean absolute values of the Xi
oj
ðqÞ and Xvāb
(q) for the system
shown in Fig. 1 versus the spatial separation between the i
and j
atoms.
Filled markers correspond to Xo, open ones to Xv. The dependencies
are given for small (triangles) and large (squares) values of the Laplace
quadrature points tq.
3436 | Phys. Chem. Chem. Phys., 2008, 10, 3430–3439
involving the Xv(q) matrix is much more important for mole-
cules of the present size range since the transformations with
Xo(q) take only a small fraction of the overall computational
time.
In Fig. 3 an analogous plot of the decay behaviour of the
matrix product Xv(q)V is given. V is the pseudoinverse of the
PAO overlap matrix SPAO (cf. eqn (11)) and hence might
destroy the sparsity of the product Xv(q)V, regardless of the
fact that Xv(q) itself is sparse. On the other hand, if there were
no redundancies in SPAO the matrix product at point tq = 0,
Xv(0)V would be identical to the identity matrix, and one may
anticipate that Xv(q)V is nevertheless sparse, also for non-zero
tq, in spite of a subset of redundant PAOs. Fig. 3 shows that
the decay rate of Xv(q)V is somewhat lower than that of Xv(q)
itself, yet it is still exponential. The exponent of the exponen-
tial decay appears to be insensitive to a particular value of tq,
whereas the prefactor is increasing for intermediate values of
tq, and then decreasing again for larger tq due to the smallness
of the exponential factors.
An efficient pre-screening, which exploits the sparsity of the
Xo(q), Xv(q) matrices, and the Xv(q)V product can therefore
significantly reduce the scaling of the computational cost with
regard to the molecular size N. It ties spatially the LMOs and
PAOs related to the ERI of eqn (16) to those of the ampli-
tudes, thus, consequently, all functions have to be spatially
‘‘close’’ to fLMO
i
.
Accuracy of the discrete quadrature
The accuracy of the discrete quadrature was investigated for
the different distribution functions f(x) = 1 (uniform distribu-
tion), f(x) = Dab ij (genuine distribution of the denominators
a 2
Dab
ij ), and f(x) = (Di ) (distribution obtained via the addition
theorem from the distribution of the denominators Dai). As
already pointed out in the previous section the computational
a 2
costs for generating f(x) = Dab ij and f(x) = (Di ) scale as
4 2
O(N ) and O(N ), respectively. The former is thus too
expensive for its use, while for the latter only an insignificant
amount of time in comparison to the rest of the calculation has
to be spent.
Fig. 3 Mean absolute values of the matrix product (XvV)āb
(q) for the
solvent cluster displayed in Fig. 1 versus the distance between the


atoms ā and b.
This journal is

c the Owner Societies 2008
 Table 1 Convergence of the energy with respect to the number of points taken for the numerical Laplace quadrature. DEn is the deviation from a reference calculation using 10 quadrature points,

analyzed for standard Boughton–Pulay domains (denoted as BP) and Boughton–Pulay domains extended by the nearest neighbours (denoted as iext = 1). For the function f(x), involved in optimizing

11.526849
11.526613
11.430101

except explicitly stated otherwise. The convergence has been

ij energies (denoted as Dij ) were tested, as well as the distribution defined as the ‘‘sum’’

1.41712
0.04778
0.00572
0.00027
0.00019
0.00000
CH4/MgO
cc-pVDZ

(Dai)2

LT-DF-LMP2 calculation, no a priori restrictions apply for BkPd defined in eqn (18). c DF-LMP2 calculation.
BP

13.446021
13.446020
13.439857
0.45904
0.06055
0.00612
0.00090
0.00022
0.00001
iext = 1
(Dai)2
Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.

13.261057
13.261054
13.235349
0.46403
0.06177
0.00618
0.00092
0.00023
0.00001
Fig. 4 Two-layer MgO cluster, consisting of 98 atoms, with 5

(Dai)2
adsorbed methane molecules. The distance between neighbouring
Mg and O atoms was set to 2.105 Å. The methane molecules are

13.261057
13.261054
13.235349
placed on top of the Mg atoms in the dipod orientation. The distance

8.85062
0.45237
0.01707
0.00157
0.00021
0.00002
cc-pVDZ
between the carbon atoms and the surface was set to 3.25 Å and the
C–H bond length to 1.0854 Å.

DA

BP
1

13.758252
13.758252
13.751505
ab

0.21044
0.02522
0.00330
0.00011
0.00015
0.00001
 100%. The BkPd quantity defined in eqn (18) was a priori restricted to the united pair domains [U(k)],
As test systems we have chosen (i) the dye dyad solvent
cluster above (cf. Fig. 1), (ii) linear chains of polyglycine

(Dai)2
peptides [gly]n, C2nH3n+2NnOn+1 (cf. ref. 2), and (iii) a two-
layer MgO cluster with five adsorbed methane molecules (cf.

13.758256
13.758252
13.751505
Fig. 4, such a system is studied in our group in the context of

1.41899
0.15002
0.00006
0.00149
0.00004
0.00005
cc-pVTZ
iext = 1
periodic LMP2 calculations of physisorption30). These systems
represent three rather different situations for the application of
the quadrature weights and points (23) the uniform (denoted as 1) and the true distribution of the denominator Dab

1
the method. The results of these test calculations are compiled

11.594812
11.594720
11.605419
in Table 1. The relative deviations (in %) from the correspond-

0.73212
0.03753
0.00404
0.00094
0.00064
0.00015
iext = 1
ing reference value based on nq = 10 quadrature points are
given for calculations involving nq = 2. . .7 points. Ordinary,

(Dai)2
of two Dai distributions (denoted as (Dai)2). Absolute energies are given in Eh. For details see section 2

as well as augmented BP domains (by all next nearest neigh- aug-cc-pVDZ

bour atoms) were used. As a criterion for the construction of

11.386892
11.386600
11.404204
0.72008
0.03598
0.00358
0.00061
0.00037
0.00009
the BP domains the default value of 0.98 was generally used.
The cc-pVDZ, aug-cc-pVDZ, and cc-pVTZ basis sets of
(Dai)2

Dunning were employed as AO basis sets, together with the

BP

related fitting basis sets of Hättig and Weigend. A comparison

b
11.038798
11.038798
11.036429
of these relative deviations obtained at nq = 4 or nq = 5

LT-DF-LMP2 calculation, BP defined in eqn (18) is restricted to d A [U(k)].

0.32443
0.00625
0.00019
0.00001
0.00002
0.00004
between the different variants of f(x) shows that the uniform
(Dai)2

a 2
distribution is inferior to f(x) = Dab
ij and f(x) = (Di ) , while the
latter two show, as expected, nearly the same performance
(shown for the [gly]6 case). We therefore advocate the use of
11.038798
11.038798
11.036429
1.32132
0.00632
0.00293
0.00007
0.00002
0.00000

f(x) = (Dai)2 due to its much lower cost. For nq = 6 or nq = 7

cc-pVDZ
iext = 1

all three distributions lead to well converged energies with

[gly]18

deviations of a few mEh from the nq = 10 reference value. In

1

contrast to the findings of ref. 14 we did not find in all cases a

3.756888
3.756899
3.748798
31

systematic convergence pattern for f(x) = Dab

0.36239
0.00645
0.00029
0.00002
0.00001
0.00000

ij and f(x) =
(Dai)2. Since in ref. 14 just one particular example was exa-
(Dai)2

mined we conclude that this behaviour cannot be anticipated

generally.
3.756888
3.756899
3.748798
EðNq ¼nÞEðNq ¼10Þ

0.36667
0.00570
0.00049
0.00004
0.00000
0.00000

LT-DF-LMP2 vs. DF-LMP2: energies and efficiency

k
d

jEðNq ¼10Þj

Dab
ij

The correlation energies as obtained with DF-LMP2 and LT-

DF-LMP2 for the different test systems, domains, and basis
3.756888
3.756899
3.748798
1.20185
0.01256
0.00303
0.00007
0.00001
0.00000
cc-pVDZ

sets, are included in Tables 1 and 2. For the present imple-

calculated as DEn ¼

mentation based on DF the most severely truncated set of

[gly]6

BP

ERIs which is possible includes all LMO pairs and all PAOs
1

within the related united pair domains,

Domains

DEa2 (%)
DEa3 (%)
DEa4 (%)
DEa5 (%)
DEa6 (%)
DEa7 (%)
Enq=10 a
Enq=10 b

X



ELMP2 c
System

cjl
dÞ

¼ Bk

c ld


2 ½UðkÞ;

d
2 ½UðlÞ:


Basis

ðk
P BP ; 8k; 8l; c ð33Þ

f(x)

P
a

This journal is

c the Owner Societies 2008 Phys. Chem. Chem. Phys., 2008, 10, 3430–3439 | 3437
 Table 2 DF-LMP2 and LT-DF-LMP2 correlation energies (in Eh) obtained for the individual test molecules with different basis sets and
domains. For the LT-DF-LMP2 calculations five quadrature points were used. CPU times are also given for individual steps of the calculations

System [gly]18 DA CH4/MgO

Basis cc-pVTZ cc-pVDZ aug-cc-pVDZ cc-pVTZ cc-pVDZ aug-cc-pVDZ

NAO, NAUX 2974, 7593 1733, 6474 2905, 8703 3962, 10127 1738, 7028 2304, 8072

Domains iext =1 BP iext = 1 BP iext = 1 BP iext = 1 BP iext = 1 BP iext = 1

Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.

DF-LMP2
E(2)
LOC 13.751505 13.236139 13.440706 13.813185 14.006510 16.243006 16.407678 11.430101 11.956126 12.697749 —
Min. mem.a/MW 377 70 221 298 889 318 1038 57 1166 134 2835
No. of iter. 7 8 8 8 8 8 8 9 9 9 —
CPU timeb/min
ERIs/FCs 104 67 89 336 374 431 520 130 199 289 562
ITER 13 10 20 51 118 67 172 13 109 36 —
TOTAL 189 146 232 706 1232 914 1738 232 1108 585 —

LT-DF-LMP2
E(2)
LOC 13.758237 13.261747 13.446759 13.793212 13.991243 16.279686 16.416874 11.526571 11.967917 12.693230 12.869608
CPU timeb/min
SIMPLEX 4 1 1 5 6 9 9 1 1 2 2
ERIs/FCs 200 120 127 609 636 734 784 181 232 365 534
Q-POINT 31 21 27 93 124 183 235 25 64 58 163
TOTAL (nq = 5) 325 201 234 986 1134 1474 1733 281 488 598 1189
a b
Memory needed to hold the local amplitudes (1 MW = 8 MByte). Intel Xeon 5150, 2.66 GHz.

This corresponds to a much larger restricted local space than
that of the local amplitudes, the latter being confined to the
restricted local pair list and the related pair domains. The four-
index ERIs (k

c|l
d)

are of course never constructed explicitly,
cf. eqns (20) and (22). Most of the calculations presented in
Table 1 are based on such a restricted ERI set and it can be
seen that the deviations from results of calculations based on
the full ERI set are indeed very small. Without much higher
cost though it is also possible to abandon any a priori
restrictions on the ERIs altogether (only the calculation of

c
the Bk
P quantities is affected), and all calculations reported in
Table 2 actually employed the full ERI set.
It can be seen that LT-DF-LMP2 recovers for all basis sets
(including the aug-cc-pVTZ basis, as calculations on smaller
molecules reveal) apart from the aug-cc-pVDZ case a some-
what larger fraction of the correlation energy than DF-LMP2.
This directly reflects the special feature of LT-DF-LMP2
discussed in the previous section, which, for the case of a
larger set of ERIs than amplitudes always provides a better fit
to the canonical amplitudes than DF-LMP2 does (the latter
entirely disregards non-zero residual elements outside the
amplitude domain boundaries). For ordinary BP domains
LT-DF-LMP2 usually recovers about 0.2% more of the
correlation energy than DF-LMP2. The two-layer MgO clus-
ter appears to represent a case where the BP domains are
particularly insufficient. Here, LT-DF-LMP2 recovers almost
1% more than DF-LMP2. With augmented BP domains the
discrepancy decreases by almost an order of magnitude to
0.02–0.05% (0.09% for the MgO cluster). For the aug-cc-
pVDZ basis set a rather large fraction of the functions has
diffuse character. As a consequence, numerous energetically
low-lying Rydberg-like diffuse MOs may occur, which do not
contribute significantly to the correlation energy, even though
the related denominators can be small. Since the LT-DF-
LMP2 method is based on a least-squares fit of the canonical
MP2 residual weighted by the inverse of the related denomi-
nators (vide supra) the importance of such amplitudes in the fit
is overrated, which may lead to the observed behaviour.
The CPU times of the individual relevant steps of the
DF-LMP2 and LT-DF-LMP2 calculations are compiled in
Table 2. Comparing the overall times it can be seen that the
efficiency of LT-DF-LMP2 is comparable to that of
DF-LMP2. For small compact domains DF-LMP2 is quicker,
for large extended domains LT-DF-LMP2 is faster. For the
case of the MgO cluster with extended domains even a single
quadrature point is cheaper than a single LMP2 iteration. For
the same system a calculation within the aug-cc-pVDZ basis
was not possible with the present DF-LMP2 implementation

Table 3 DF-LMP2 and LT-DF-LMP2 correlation contributions to the dipole moment (in a.u.) for the case of [gly]6 using different basis sets. The
reference calculations were performed with extended domains by augmenting the BP domains (criterion 0.98) by all atoms within a range of 10 a0
(rext = 10). The LT-DF-LMP2 calculations were performed with 10 quadrature points and by using eqn (32) to calculate the correlated density
matrix. The X-component of the dipole moment is zero due to symmetry

cc-pVDZ aug-cc-pVDZ cc-pVTZ

ma rext = 10 LMP2 LT rext = 10 LMP2 LT rext = 10 LMP2 LT

mY 0.00998 0.00895 0.00916 0.00999 0.00854 0.00876 0.01073 0.01018 0.01040
mZ 0.03260 0.03526 0.03505 0.03093 0.03230 0.03214 0.02782 0.03064 0.02965

3438 | Phys. Chem. Chem. Phys., 2008, 10, 3430–3439 This journal is

c the Owner Societies 2008
 due to a memory bottleneck, whereas the same calculation
could be easily performed with the LT-DF-LMP2 method,
which has only very modest memory requirements.
Table 3 compares the orbital-unrelaxed LT-DF-LMP2 and
DF-LMP2 dipole moments of [gly]6 calculated with different
basis sets. This shall serve as a first illustration for our
discussion about LT-DF-LMP2 properties in the previous
section. Ordinary BP domains were used for these calculations
Published on 20 May 2008. Downloaded by Universitat Stuttgart on 24/08/2016 13:56:47.
(with extended domains there would hardly be any visible
difference), for the reference calculations the BP domains were
extended with all centres within a range of 10 a0 from the core
BP domains. The LT-DF-LMP2 density matrices were calcu-
lated according to eqn (32), i.e., using the symmetrical ap-
proach. Evidently, even for non-extended BP domains the DF-
LMP2 and LT-DF-LMP2 are very close, with the latter being
slightly closer to the reference value. This is most pronounced
for the biggest basis set, for the two other basis sets the
differences between LT-DF-LMP2 and DF-LMP2 are hardly
noticeable. From this first preliminary example it appears that
the simple symmetrical approach provides LT-DF-LMP2
properties which are at least as accurate as those provided
by the original DF-LMP2 method. Further investigation
concerning the issue about properties in the context of the
local Laplace method is deferred to a forthcoming paper.
4. Conclusions
In this paper a new local method based on the Laplace
transform ansatz of Almlöf is presented. As in the previous
local methods based on the local ansatz of Pulay the occupied
and virtual orbital spaces are spanned by mutually orthogonal
localized MOs and non-orthogonal projected AOs, respec-
tively, and the set of wavefunction parameters (amplitudes) is
a priori restricted. An implementation for local second-order
Møller–Plesset perturbation theory is described, which is
utilizing density fitting for factorizing the electron repulsion
integrals. A detailed analysis of the relation between the
original and the new Laplace-based local method reveals that
the two methods are complementary: The former corresponds
to a projection of the canonical residual onto the a priori
restricted local space with the requirement that this projection
is exactly vanishing, and the latter to a least-squares mini-
mization weighted with the inverse of the related energy
denominators. The Laplace-based local method generally re-
covers a slightly larger fraction of the correlation energy.
Properties and energy gradients can also be calculated, gra-
dients though at a somewhat higher cost compared to ordinary
local MP2 due to an asymmetry in the related density matrices.
The computational efficiency for calculating correlation en-
ergies is comparable for local MP2 and the Laplace-based
local method, the former being more efficient for calculations
involving small domains, the latter for calculations involving
large domains. The primary purpose of this work though was
This journal is

c the Owner Societies 2008
View publication stats
not to present yet another efficient local MP2 method, but to
investigate generally the usefulness of the Laplace transform
ansatz in local methods with a priori restricted sets of wave-
function parameters. We will report on a local coupled cluster
response implementation based on the Laplace transform
ansatz in a forthcoming paper.
Acknowledgements
Financial support from the Deutsche Forschungsgemeinschaft
DFG (priority program SPP1145) is gratefully acknowledged.
References
1 P. Pulay, Chem. Phys. Lett., 1983, 100, 151.
2 M. Schütz, G. Hetzer and H.-J. Werner, J. Chem. Phys., 1999, 111,
5691.
3 M. Schütz and H.-J. Werner, J. Chem. Phys., 2001, 114, 661.
4 M. Schütz, J. Chem. Phys., 2000, 113, 9986.
5 M. Schütz, J. Chem. Phys., 2002, 116, 8772.
6 H.-J. Werner, P. Knowles and F. R. Manby, J. Chem. Phys., 2003,
118, 8149.
7 M. Schütz and F. R. Manby, Phys. Chem. Chem. Phys., 2003, 5,
3349.
8 M. Schütz, H.-J. Werner, R. Lindh and F. R. Manby, J. Chem.
Phys., 2004, 121, 737.
9 D. Kats, T. Korona and M. Schütz, J. Chem. Phys., 2006, 125,
104106.
10 D. Kats, T. Korona and M. Schütz, J. Chem. Phys., 2007, 127,
064107.
11 L. Maschio, D. Usvyat, F. Manby, S. Cassassa, C. Pisani and M.
Schütz, Phys. Rev. B: Condens. Matter Mater. Phys., 2007, 76,
075101.
12 D. Usvyat, L. Maschio, F. Manby, S. Cassassa, M. Schütz and C.
Pisani, Phys. Rev. B: Condens. Matter Mater. Phys., 2007, 76,
075102.
13 J. Almlöf, Chem. Phys. Lett., 1991, 181, 319.
14 M. Häser and J. Almlöf, J. Chem. Phys., 1992, 96, 489.
15 M. Häser, Theor. Chim. Acta, 1993, 87, 147.
16 P. Y. Ayala and G. E. Scuseria, J. Chem. Phys., 1999, 110, 3660.
17 P. Y. Ayala, K. N. Kudin and G. E. Scuseria, J. Chem. Phys., 2001,
115, 9698.
18 D. Lambrecht, B. Doser and C. Ochsenfeld, J. Chem. Phys., 2005,
123, 184102.
19 G. Rauhut and P. Pulay, Chem. Phys. Lett., 1996, 248, 223.
20 A. Wilson and J. Almlöf, Theor. Chim. Acta, 1997, 95, 49.
21 T. Nakajima and K. Hirao, Chem. Phys. Lett., 2006, 427, 225.
22 Y. Jung, R. Lochan, A. Dutoi and M. Head-Gordon, J. Chem.
Phys., 2004, 121, 9793.
23 Y. Jung, Y. Shao and M. Head-Gordon, J. Comput. Chem., 2007,
28, 1953.
24 S. F. Boys, in Quantum Theory of Atoms, Molecules, and the Solid
State, ed. P. O. Löwdin, Academic Press, New York, 1966, p. 253.
25 J. Pipek and P. G. Mezey, J. Chem. Phys., 1989, 90, 4916.
26 H.-J. Werner, F. R. Manby and P. J. Knowles, J. Chem. Phys.,
2003, 118, 8149.
27 M. Schütz and F. R. Manby, Phys. Chem. Chem. Phys., 2003, 5,
3349.
28 J. D. Cloizeaux, Phys. Rev., 1964, 135, A685.
29 K. Sadeghian and M. Schütz, J. Am. Chem. Soc., 2007, 129, 4068.
30 C. Pisani, L. Maschio, S. Casassa, M. Halo, M. Schütz and D.
Usvyat, J. Comput. Chem., 2008, DOI: 10.1003/jcc.20975.
31 J. W. Boughton and P. Pulay, J. Comput. Chem., 1993, 14, 736.
Phys. Chem. Chem. Phys., 2008, 10, 3430–3439 | 3439