Struct Bond (2005) 116: 43–117

DOI 10.1007/430_004
© Springer-Verlag Berlin Heidelberg 2005
Published online: 18 October 2005

Intermolecular Interactions via Perturbation Theory:

From Diatoms to Biomolecules
Krzysztof Szalewicz1 (u) · Konrad Patkowski1,2 · Bogumil Jeziorski2
1 Department of Physics and Astronomy, University of Delaware, Newark, DE 19716, USA
szalewic@udel.edu, patkowsk@physics.udel.edu
2 Department of Chemistry, University of Warsaw, Pasteura 1, 02-093 Warsaw, Poland

jeziorsk@tiger.chem.uw.edu.pl

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2 Convergence Properties of Conventional SAPT . . . . . . . . . . . . . . . . 47

2.1 Polarization Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.2 Symmetry-forcing Technique . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.3 Conjugate Formulation of SAPT and the HS Theory . . . . . . . . . . . . . 58

3 Convergence Properties of Regularized SAPT . . . . . . . . . . . . . . . . 63

3.1 Methods of Regularizing the Coulomb Potential . . . . . . . . . . . . . . . 63
3.2 Regularized SRS Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.3 A Posteriori Inclusion of V t . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.4 Double Perturbation Approach . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5 The “All-in-one” R-SRS+ELHAV Theory . . . . . . . . . . . . . . . . . . . 69
3.6 The R-SRS+SAM Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.7 Zero-order Induction Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 72

4 Numerical Studies of Convergence Behaviour . . . . . . . . . . . . . . . . 73

4.1 H· · ·H interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2 Li· · ·H Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

5 Extension of the Theory to Many-electron Systems . . . . . . . . . . . . . 87

5.1 Outline of Many-electron SAPT . . . . . . . . . . . . . . . . . . . . . . . . 87
5.2 SAPT Based on DFT Description of Monomers . . . . . . . . . . . . . . . . 90
5.3 SAPT/SAPT(DFT) Computer Codes . . . . . . . . . . . . . . . . . . . . . . 92

6 Helium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.1 Helium as a Thermodynamic Standard . . . . . . . . . . . . . . . . . . . . 94
6.2 Towards 0.01% Accuracy for the Dimer Potential . . . . . . . . . . . . . . . 96

7 Some Recent Applications of Wave-function-based Methods . . . . . . . . 98

7.1 Argon Dimer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.2 He – HCl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.3 He – N2 O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.4 He – HCCCN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.5 H2 – CO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.6 Methane-water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.7 Water Dimer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

8 Performance of the SAPT(DFT) Method . . . . . . . . . . . . . . . . . . . 103

44 K. Szalewicz et al.

9 Applications of SAPT(DFT) to Molecular Crystals . . . . . . . . . . . . . . 106

9.1 Benzene Dimer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
9.2 Dimethylnitramine Dimer . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

10 Transferable Potentials for Biomolecules . . . . . . . . . . . . . . . . . . . 109

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

Abstract This article is devoted to the most recent, i.e. taking place within the last few
years, theoretical developments in the field of intermolecular interactions. The most
important advancement during this time period was the creation of a new version of
symmetry-adapted perturbation theory (SAPT) which is based on the density-functional
theory (DFT) description of monomers. This method, which will be described in Sect. 5.2,
allows SAPT calculations to be performed for much larger molecules than before. In fact,
many molecules of biological importance can now be investigated. Another important
theoretical advancement was made in understanding the convergence properties of SAPT.
It has been possible to investigate such properties on a realistic example of a Li atom in-
teraction with an H atom. This is the simplest system for which the coupling of physical
states to the unphysical, Pauli forbidden continuum causes the divergence of the conven-
tional polarization expansion and of several variants of SAPT. This development will be
described in some detail in Sects. 2–4, where, in addition to a review of published work,
we shall present several original results on this subject. In an unrelated way, one of the
most interesting recent applications of ab initio methods concerns the helium dimer and
allows first-principle predictions for helium that are in many cases more accurate than
experimental results. Therefore, theoretical input can be used to create new measurement
standards. This broad range of systems that were the subject of theoretical investigations
in recent years made us choose the title of the current review. With a few exceptions,
the investigations of individual systems discussed here utilized SAPT. The calculations
for helium are described in Sect. 6, recent wave-function based applications in Sect. 7, the
performance of SAPT(DFT) on model systems in Sect. 8, and applications of SAPT(DFT)
in Sect. 9. Section 10 summarizes work on biosystems.

1
Introduction

It is remarkable how broad a range of physical, chemical, and even biologi-

cal phenomena originates from weak intermolecular interactions (also called
van der Waals interactions), i.e. the interactions that do not involve forming
a chemical bond between the interacting species. Intermolecular interactions
(or forces) determine bulk properties of gases and liquids and are responsi-
ble for the very existence of molecular liquids and crystals. The knowledge of
accurate intermolecular potential energy surfaces (PESs) is necessary to in-
terpret high-resolution spectroscopic [1, 2] and scattering [3] data, including
the spectroscopic data coming from planetary atmospheres and the inter-
stellar gas [4], as well as to construct and tune empirical potentials used in
Monte Carlo (MC) or molecular dynamics (MD) bulk simulations [5]. In re-
cent years, interactions of various molecules with helium became particularly
Intermolecular Interactions 45

important due to the development of superfluid helium nanodroplet spec-

troscopy [6, 7]. Weak intermolecular forces are responsible for the biomolec-
ular recognition patterns and the catalytic activity of enzymes, and thus
insights into intermolecular PESs are important for drug design [8, 9]. An in-
teresting example of a macroscopic effect of van der Waals forces is given by
the recent experimental evidence that Tokay geckos (Gekko gecko) owe their
exceptional ability to climb smooth vertical surfaces to the van der Waals
attractions between the surface and gecko toe-hairs [10].
For small monomers, the intermolecular potentials can be computed by
standard electronic structure methods that account for electron correlation.
This is done by using the supermolecular approach, i.e. for each configuration
of fixed nuclei (the Born-Oppenheimer approximation), the interaction en-
ergy Eint is obtained by subtracting the total energies of monomers from the
total energy of the cluster [11, 12]
Eint ≡ E = EAB – EA – EB . (1)
The main advantages of the supermolecular approach are its universal-
ity, conceptual simplicity, and availability of many sophisticated ab initio
methods and highly optimized computer codes that can be used to calcu-
late the quantities on the r.h.s. of Eq. 1. However, this approach is not free
from serious problems, mainly originating from the fact that the subtraction
in Eq. 1 involves components that are several orders of magnitude larger than
the interaction energy E; in fact, the errors with which these components are
calculated exceed the value of E for nearly all computations performed so
far. Under these circumstances, the supermolecular approach may give ac-
curate results only if a cancellation of errors occurs in Eq. 1. A necessary
condition for this cancellation to take place is size-consistency of the method
employed to calculate EAB , EA , and EB [13]. However, this condition is not
sufficient, as has been demonstrated by the failure of the supermolecular
density functional theory (DFT) applied to several rare-gas dimers [14]. We
now know that the cancellation of errors in Eq. 1 does take place if one uses
Møller-Plesset perturbation theory (MP) or the coupled-cluster (CC) method
to calculate the quantities on the r.h.s.; it is worth mentioning that this fact
has been proved by a decomposition of the supermolecular interaction en-
ergy in terms of SAPT corrections [15]. However, even employing a highly
correlated method like the supermolecular coupled-cluster with single, dou-
ble and noniterative triple excitations (CCSD(T)) and large basis sets does
not guarantee that an accurate PES will be obtained [16]. Another disadvan-
tage of the supermolecular approach is that one must take care of eliminating
the basis set superposition error (BSSE) [17, 18]. This requires extra computa-
tion time and, most importantly, it is not entirely clear how to get rid of BSSE
when the partitioning of the dimer into interacting monomers is ambiguous,
as in the case of PESs associated with chemical reactions, or when there exist
multiple PESs resulting from the presence of an open-shell monomer [12].
46 K. Szalewicz et al.

Last but not least, all one can get from a supermolecular calculation at a given
dimer geometry is a single number which tells nothing about the physics un-
derlying the interaction phenomenon.
Another difficulty arising in computational investigations of intermolecu-
lar interactions is that in virtually all cases one has to include effects of
electron correlation. The computer resource requirements of all methods in-
volving electron correlation scale as a high power (5th or higher) of system
size (as measured by the number of electrons), which severely limits the size
of molecules that can be handled. A much faster approach is provided by
the DFT method, which scales as the third power of the system size. DFT is
widely used in solid state physics and in chemistry. Unfortunately, with the
currently available functionals, DFT fails to describe an important part of
intermolecular forces, the dispersion interaction. Consequently, predictions
are poor except for very strong intermolecular interactions, as in the case of
hydrogen-bonded clusters.
An alternative to the supermolecular approach is symmetry-adapted per-
turbation theory [19–21]. In SAPT, the interaction energy is computed di-
rectly rather than by subtraction. SAPT provides both the conceptual frame-
work and the computational techniques for describing intermolecular inter-
actions, including the dispersion energy. However, the computer resources
required by SAPT, similar to those of the methods with high-level treat-
ment of correlation used in the supermolecular approach, make applications
to monomers with more than about ten atoms impractical at the present
time. For ten-atom or smaller molecules, SAPT has been very successful;
see [20, 21], and Sect. 7 for a review of applications.
The concept of calculating the interaction energy of two chemical systems
A and B perturbatively is not at all a new idea. The first intermolecular per-
turbation expansion was proposed [22] just a few years after the foundations
of quantum mechanics had been laid. Since then, numerous other expan-
sions, now known under a common name of symmetry-adapted perturbation
theory, have been introduced and the perturbation theory of intermolecu-
lar forces is now a fully mature approach. Thanks to the development of
the many-body SAPT [23] and of a general-utility closed-shell SAPT com-
puter code [24], the perturbative approach to intermolecular interactions has
been successfully applied to construct PESs for numerous interacting dimers
of theoretical and experimental interest [19–21, 25–27]. One of the notable
achievements of SAPT is an accurate description of the interactions between
water molecules [21, 28–32]. A recent paper by Keutsch et al. [33] compares
the complete spectra of the water dimer with theoretical predictions obtained
using an empirical potential fitted to extensive spectroscopic data, and with
the predictions from a SAPT potential. These comparisons show that the lat-
ter potential probably provides the best current characterization of the water
dimer force field. In another recent application, an SAPT PES for helium in-
Intermolecular Interactions 47

teracting with water has been used to calculate scattering parameters that
agreed well with the high-quality experimental data [34, 35].
The SAPT interaction energy is expressed as a sum of well-defined
and physically meaningful contributions, corresponding to the electrostatic,
induction, dispersion, and exchange components of the interaction phe-
nomenon. Thus, the SAPT theory provides the basic conceptual framework
for our understanding of the nature of the intermolecular forces [19, 36]. Fi-
nally, one may note that SAPT is much more flexible computationally than
the standard supermolecular approach. Different interaction energy com-
ponents can be computed using different levels of the electron correlation
treatment and/or employing different basis sets. One can also use custom
designed basis sets, like the so-called monomer-centered “plus” basis sets
(MC+ BS), to speed up the basis set convergence of some or all perturbation
corrections [37].

2
Convergence Properties of Conventional SAPT

The question of the convergence of the SAPT expansions has been studied ex-
tensively in the past, but numerical investigations were possible only for very
simple few-electron dimers like H2 + [38–42], H2 [43–47], HeH [48], He2 and
HeH2 [49]. Such systems do not exhibit all the complications arising in in-
teractions of general many-electron monomers. For example, on the basis of
early studies, it was believed for some time that the Hirschfelder-Silbey (HS)
perturbation expansion [50, 51] should converge quickly for many-electron
systems, just like it did for all the small systems investigated. Since low-
order energy corrections of the extremely complicated HS method were al-
most identical to the ones calculated from the much simpler symmetrized
Rayleigh-Schrödinger (SRS) theory [40], these results justified the use of the
low-order SRS method in practical applications. However, neither of the sys-
tems mentioned above can be viewed as a legitimate model for studying the
convergence of SAPT for many-electron systems, since significant complica-
tions appear when one of the interacting monomers has more than two elec-
trons. These complications, first pointed out by Adams [52–55], arise from
the fact that in such a case the physical ground state of the interacting dimer is
buried in a continuous spectrum of unphysical states (discovered by Morgan
and Simon [56]), which violate the Pauli exclusion principle. This situation is
graphically displayed in Fig. 1 of [57] for a lithium atom interacting with a hy-
drogen atom. In the presence of the Pauli-forbidden continuum, the Rayleigh-
Schrödinger (RS) perturbation theory, and thus also the SRS method which
employs the same expansion for the wave function, must diverge. Moreover,
under these circumstances the Hirschfelder-Silbey theory cannot be expected
to converge either, since in this theory one performs a perturbation expan-
48 K. Szalewicz et al.

sion of the so-called primitive [58] or localized [48] function, defined as

a sum of all asymptotically degenerate eigenfunctions of the Hamiltonian,
both physical and Pauli-forbidden. When the Pauli-forbidden eigenfunctions
belong to the continuum, the primitive function is not square integrable and
cannot be a limit of a convergent series of square integrable functions (all
finite-order perturbed wave functions in the HS theory are well defined and
square integrable).
There exists another class of SAPT expansions that is free from the Pauli-
forbidden continuum problem and can be expected to converge for many-
electron systems. This class includes the ELHAV theory – the one intro-
duced in 1930 by Eisenschitz and London [22], and rediscovered later by
Hirschfelder [59], van der Avoird [60], and Peierls [61]. Other SAPT expan-
sions of this kind are the Amos-Musher (AM) [62] and Polymeropoulos-
Adams [63] theories. These methods, however, suffer from a different prob-
lem, discovered in an early numerical investigation [43] of the H2 molecule
and confirmed in analytical studies [38, 64] of the H2 + ion: such methods
fail to recover in the second order the important induction and dispersion
components of the interaction energy, leading to wrong values of the con-
stants Cn in the Cn /Rn asymptotic expansion of the interaction energy [65]
(R denotes the intermonomer distance), starting from the C6 /R6 term for
electrically neutral monomers and C4 /R4 term when at least one monomer
is electrically charged. This failure should be contrasted with the behav-
ior of the SRS and HS methods which are asymptotically compatible with
the RS theory, i.e. have the property that each term in the 1/R expansion
of the interaction energy is recovered in finite order [66]. On the basis of
these findings, Jeziorski and Kołos proposed [67] a new SAPT expansion, the
JK theory, which, while still being free from the Pauli-forbidden continuum
problem, has the correct large-R asymptotics of the second-order energy (but
not of the third- and higher-order corrections), thus reproducing correctly
the conventional (second-order) induction and dispersion components of the
interaction energy. However, none of the SAPT expansions discussed so far
is simultaneously convergent for many-electron systems and asymptotically
correct in every order of the perturbation theory. This fact has been elabo-
rated by Adams [68], who concluded that the existing SAPT formulations are
inadequate for the study of many-electron systems, and one must search for
a new theory.
The understanding of the convergence issues described above has been
greatly improved in the past few years. In particular, studies of the high-order
convergence properties of the existing SAPT expansions for a system exhibit-
ing the Pauli-forbidden continuum have been performed [57, 69]. Although
the prediction of the divergence of several SAPT theories for many-electron
systems was confirmed, it has been shown that these theories provide useful
and accurate information despite their divergence. Other SAPT expansions
have been found to converge in the presence of a Pauli-forbidden continuum,
Intermolecular Interactions 49

and their low-order convergence properties were greatly influenced by the

asymptotic correctness/incorrectness of low-order energies. In particular,
a qualitative relation was found between modifications in the symmetry-
forcing procedure such as in the JK theory and improved asymptotics and
convergence properties of SAPT [69]. Very recently, several new SAPT ex-
pansions were developed [70–72] that are simultaneously convergent in the
presence of the Pauli-forbidden continuum and asymptotically correct to any
order. These expansions are also accurate in low order for a wide range of
intermolecular distances, and are therefore suitable for employing in practi-
cal calculations for many-electron systems. This development has been made
possible by separating the singular, short-range part of the nuclear attraction
terms in the interaction operator and treating it differently from the regular
long-range part [70, 71]. The convergence properties of the resulting so-called
“regularized expansions” will be discussed in detail in Sect. 3. In this sec-
tion we shall review results obtained for the conventional SAPT formulations,
treating the interaction operator as a whole.
All the methods discussed above have the property of being formulated
in a completely basis set independent way. Another possible approach is to
consider the Schrödinger equation in the matrix form using some specific ba-
sis set. If this matrix is decomposed appropriately, one can obtain a family
of so-called “symmetric perturbation” treatments [19, 67]. The best known
of these is the variant called intermolecular perturbation theory (IMPT) de-
veloped by Hayes and Stone, which has led to many successful applications in
the many-electron context [73–76].

2.1
Polarization Approximation

Suppose one wants to calculate the interaction energy of two systems A and
B which are, in the absence of interaction, described by the clamped-nuclei
Hamiltonians HA and HB , respectively. Let φA and φB denote the ground-state
eigenfunctions of HA and HB , respectively, and let EA and EB be the corres-
ponding eigenvalues,
HA φA = EA φA , HB φB = EB φB . (2)
When the interaction between A and B is switched on, the dimer AB is de-
scribed by the Hamiltonian H = HA + HB + V, where the operator V collects
all Coulombic interactions between particles (electrons and nuclei) belonging
to A and those belonging to B. For an eigenfunction ψ of the Hamiltonian H,
the corresponding eigenvalue is equal to EA + EB + E. Our aim is to calculate
approximations to the interaction energy E and the dimer wave function ψ by
means of the perturbation theory. The simplest way of doing so is provided
by the conventional Rayleigh-Schrödinger (RS) perturbation theory with the
zero-order Hamiltonian H0 = HA + HB and the perturbation equal to V [77]
50 K. Szalewicz et al.

(in this context, the RS method is often called, after Hirschfelder [78], the
polarization approximation).
To derive the equations for the RS perturbation corrections, it is conve-
nient to rewrite the Schrödinger equation Hψ = Eψ in the so-called Bloch
form [79],


ψ = φ0 + R0 ( φ0 |Vψ – V)ψ (3)
where φ0 = φA φB is a (normalized) eigenfunction of the unperturbed Hamil-
tonian H0 , corresponding to the energy E0 = EA + EB , and R0 is the reduced
resolvent of H0 which can be defined by the formula
R0 = (1 – P0 )(H0 – E0 + P0 )–1 , (4)


with P0 = |φ0 φ0 | being the projection onto the unperturbed function φ0 .
One can easily verify that any ψ
satisfying
 Eq. 3 fulfills the so-called interme-
diate normalization condition φ0 |ψ = 1.
Equation 3 can be solved iteratively [36, 40, 67],
ψn = φ0 + R0 (En – V)ψn–1 , (5)
where


En = φ0 |Vψn–1 (6)
can be viewed as the nth approximation to the interaction energy E. When
the iterative process is initiated with ψ0 = φ0 , the approximate energy En con-
tains all the RS energy corrections E(k)RS up to and including the nth order, as
well as some terms of the order higher than n [67]. To extract the individual
corrections E(n)
RS , one needs to substitute H → H0 + ζV, where ζ is a complex
variable, and insert the expansions
∞
 ∞

(n) (n)
ψ(ζ) = ψRS ζ n , E(ζ) = ERS ζ n (7)
n=0 n=1
(n)
into Eqs. 5 and 6. The resulting expressions for ψRS and E(n)
RS are


(n) (n–1)
ERS = φ0 |VψRS , (8)

(n) (n–1)

n–1
(k) (n–k)
ψRS = – R0 VψRS + ERS R0 ψRS , (9)
k=1
(0)
where ψRS ≡ φ0 .
Unlike the full Hamiltonian H, its zero-order part H0 , as well as the per-
turbation V, does not possess full symmetry with respect to the permutations
of electrons. As a consequence, the zero-order wave function φ0 is not com-
pletely antisymmetric – some electrons are assigned to one monomer, the
Intermolecular Interactions 51

remaining ones to the other. Therefore, the equations for the perturbed wave
(n)
functions ψRS (or the ψn functions of Eq. 5) cannot be solved in the Hilbert
space HAB of the Pauli-allowed (antisymmetric with respect to interchanges
of any two electrons) functions of the dimer, but rather in a larger, product
space HA ⊗ HB , where HA and HB are the Hilbert spaces of Pauli-allowed
functions for monomers A and B, respectively. At first glance, one could think
that V amounts to a small perturbation of H0 , at least when the intermonomer
distance R is large. However, this is not the case: the difference ||φ0 – ψ||,
where || · || is the L2 norm in the Hilbert space, is always large and does not
tend to zero when R goes to infinity [67]. The finding that V cannot be treated
as a small perturbation has a dramatic manifestation: the spectral proper-
ties of the operators H0 and H are completely different when one (or both)
of the interacting monomers has more than two electrons. Furthermore, the
lowest physical eigenstate of H lies in such cases within a continuum of Pauli-
forbidden states (for a more detailed discussion of this issue we refer the
reader to [52, 53, 57], and [72]). Under these circumstances, one must expect
that the RS perturbation theory will have serious problems converging to the
interaction energy E.
Even if each of the monomers has two electrons or fewer, the polarization
approximation, although convergent, is far from being suitable for practi-
cal applications. Large-order numerical studies for H2 + [39, 42] and H2 [46]
revealed that in low orders the sum of the polarization approximation ap-
proaches the so-called Coulomb energy Q, defined as an arithmetic mean of
the energies of the lowest gerade and ungerade states (for H2 + ) or the low-
est singlet and triplet states (for H2 ). After the value of Q is reproduced to
a good accuracy, convergence of the polarization series deteriorates dramat-
ically and the remaining part of the interaction energy – the exchange energy
– is not reproduced to any reasonable extent in finite order. The patholog-
ically slow high-order convergence of the polarization expansion manifests
itself in the values of the convergence radii ρ of the perturbation series be-
ing only marginally greater than unity (for instance, ρ = 1.0000000031 for
H2 at the van der Waals minimum distance of 8 bohr) [46]. For the helium
dimer, the situation is even worse. Not only the sum of the polarization ex-
pansion converges extremely slowly after reaching the value of Q (see [80] for
the definition of Q for He2 ), but its limit is not the physical ground-state en-
ergy, as it was for H2 + and H2 , but rather the energy of the unphysical, fully
symmetric 1(σg )4 state [49, 81].
When one of the monomers has three or more electrons, the polarization
approximation diverges, as first proved by Kutzelnigg [58]. He argued that an
avoided crossing must take place when the ground-state interaction energy
E(ζ) is analytically continued from the physical value at ζ = 0 to the low-
lying unphysical value at ζ = 1. The existence of the unphysical continuum
only makes matters worse: E(ζ) has to be continued analytically through
infinitely many avoided crossings before reaching the physically significant
52 K. Szalewicz et al.

value of ζ = 1 [55]. It is worth noting that it is the one-electron, attractive

part of the perturbation V that is responsible for the existence of the avoided
crossings and the divergence of the RS theory for many-electron systems. The
computationally much more complicated two-electron part of V plays only
a minor role in determining the convergence properties. An interesting model
for studying the convergence of intermolecular perturbation series, based on
the above observation, has recently been proposed by Adams [82]. This model
neglects the electron-electron interaction completely; however, it can pro-
vide qualitative predictions of the convergence/divergence of the polarization
approximation for various dimers. Adams’ model fails in the case of a ground-
state helium atom interacting with a ground-state hydrogen atom, for which
it predicts divergence of the RS expansion, whereas large-order numerical
calculations indicate that RS converges for this system.1 Nevertheless, con-
vergence properties of various intermolecular perturbation expansions are
closely related to the way in which the one-electron part of the interaction
is taken into account; we will return to this observation in Sect. 3 while dis-
cussing the regularized SAPT theory.

2.2
Symmetry-forcing Technique

As we have already stated, the zero-order wave function φ0 and the exact wave
function ψ exhibit different symmetry with respect to interchanges of elec-
trons. Whereas ψ is fully antisymmetric for all electron permutations, φ0 is
only antisymmetric with respect to interchanges of electrons belonging to
the same monomer (A or B). Therefore, one can improve convergence of the
polarization expansion by forcing the full antisymmetry, i.e. by inserting ap-
propriate projection operators into the perturbation equations, Eqs. 8 and 9.
The specific form of these projectors depends on the specific choice of the
spaces HA and HB . In the following discussion we shall use the so-called
spin-free approach [83, 84] in which HX (X = A, B) is spanned by spatial-
only functions corresponding to a specific irreducible representation (irrep)
[λ] of the symmetric group SNX , where NX is the number of electrons in
the monomer X. The choice of [λ] determines the spin multiplicity of the
monomer X since a spatial wave function of symmetry [λ] can only be com-
bined with a spin wave function of the conjugate symmetry [λ]† (the Young
diagrams of [λ] and [λ]† are a transpose of each other) to form an antisym-
metric total wave function [84]. In this approach, the symmetry operators to
be inserted into Eqs. 8 and 9 are the Young projectors A[λ] onto the subspace
of appropriate symmetry [λ] with respect to the SN group, N = NA + NB . [λ]
must be the symmetry of one of the subspaces into which the whole space
HA ⊗ HB decomposes under the action of SN , cf. Eq. 9 of [57]. If the conven-

1 Korona T, unpublished
Intermolecular Interactions 53

tional, spin approach were used, and the spaces HA and HB were spanned
by Slater determinants, the symmetry operators would be products of the
antisymmetrizer and a suitable spin projection.
Most of the existing SAPT formulations can be obtained using the general
symmetry-forcing technique developed in [67] and [40]. The iterative scheme
of Eqs. 5 and 6 is generalized as follows,
ψn = φ0 + R0 (En – V)F ψn–1 , (10)


φ0 |VGψn–1
En =
 , (11)
φ0 |Gψn–1
where F and G are symmetry-forcing operators. The denominator in the
energy is necessary since G does not have to conserve the intermediate nor-
malization of ψn–1 . Different choices of the operators F and G, as well as
of the function ψ0 used to initiate the iterations, lead to different SAPT ex-
pansions listed in Table 1. In this table, as well as throughout the rest of this
review, the symmetry index [λ] will be omitted as long as it does not lead to
any ambiguities.
Obviously, the RS method is obtained from the iterative process expression
defined by Eq. 11 with no symmetrization performed. The simplest SAPT
theory with symmetrization, taking into account the exchange part of the in-
teraction energy in finite order, is the SRS expansion formulated in [40]. In
the SRS method, the wave function corrections are taken directly from the
(n) (n)
RS theory, ψSRS ≡ ψRS , and the perturbation energies are calculated from the
formula
 

  n–1
 
ESRS = N0 φ0 VAψRS ESRS φ0 AψRS
(n) (n–1) (k) (n–k)
– , (12)
k=1

–1
where N0 = φ0 |Aφ0 . It is the SRS theory that has been implemented in the
general-utility closed-shell SAPT program [24] and widely applied in practi-
cal calculations [20].
Similar to SRS, but a little more complicated, is the MSMA theory in-
troduced in [85] and [86]. The iterative process resulting from the original
formulation of the MSMA theory starts from the symmetrized and intermedi-
ately normalized zero-order function N0 Aφ0 , and no further symmetrization
is applied (F = G = 1). It has been shown in [67] that the choice F = 1, G = A,
(n)
and ψ0 = φ0 leads to the same energy corrections EMSMA , although the wave
function corrections are different. The MSMA energies E(n)MSMA and wave func-
(n)
tions ψMSMA can therefore be calculated from Eqs. 12 and 9, respectively,
with all the RS and SRS quantities replaced by their MSMA counterparts. The
symmetry forcing employed in the SRS and MSMA methods is commonly re-
ferred to as weak [67], since no symmetrization is applied in Eq. 10 defining
54 K. Szalewicz et al.

Table 1 Symmetry-adapted perturbation theories obtained using the symmetry-forcing

technique. The last two columns give equation numbers from which the energy and wave
function corrections of a given method can be calculated

Method F G ψ0 Eq. for E(n) Eq. for ψ (n)

RS 1 1 φ0 8 9
SRS 1 1(A) a φ0 12 9
MSMA(a) 1 1 N0 Aφ0
MSMA(b) 1 A φ0 12 9
ELHAV A A N0 Aφ0 15 13–14
JK A 1 N0 Aφ0 8 13–14
JK-1 A 1 N1 A(φ0 – R0 Vφ0 ) 8 17
aIn the SRS theory the wave function corrections are taken from the RS method, and the
energy corrections are calculated from Eq. 12.

corrections to the wave function. The theories with weak symmetry forcing
cannot be expected to converge in the presence of a Pauli-forbidden con-
tinuum.
Among the theories employing the so-called strong symmetry forcing, for
which the symmetry projector A appears in Eq. 10, are the ELHAV [22, 59, 60]
and JK [67] methods, as well as the AM expansion [62] which does not fit
into the general scheme of Eqs. 10 and 11 and will be discussed separately.
The ELHAV theory has quite a few apparently different but fully equivalent
formulations. Here we have chosen the one introduced in [40]. Within this
approach, which can be referred to as involving the strong symmetry-forcing
procedure, Eqs. 10 and 11 are iterated with F = G = A and the starting func-
(n)
tion ψ0 = N0 Aφ0 . The corrections ψELHAV are then defined by the formulae
(1)
ψELHAV = N0 R0 A(E(1)
ELHAV – V)φ0 , (13)
(n) (n–1)

n
ψELHAV = – R0 VAψELHAV + E(k) (n–k)
ELHAV R0 AψELHAV (14)
k=1
(0) (n)
for n ≥ 2. In these equations, ψELHAV = N0 Aφ0 , and the energies EELHAV are
calculated from the formula

 n–1

(n) (n–1) (k) (n–k)
EELHAV = φ0 |VAψELHAV – EELHAV φ0 |AψELHAV . (15)
k=1
(n)
Since the unphysical components of ψELHAV are explicitly projected out
on the r.h.s. of Eqs. 13 and 14, the ELHAV method may converge despite
the presence of a continuum. However, unlike the methods employing weak
symmetry forcing, its second-order energy E(2)
ELHAV does not recover correctly
the leading Cn /Rn term in the asymptotic expansion of the interaction en-
Intermolecular Interactions 55

ergy [65]. This is a serious drawback, both from the theoretical (the phenom-
ena of induction and dispersion are not correctly described) and practical
(low-order energies are highly inaccurate for large R) point of view. The ori-
gin of the wrong asymptotics of the second-order ELHAV energy is explained
in detail in [67].
To correct the asymptotics of the second-order ELHAV energy, Jeziorski and
Kołos suggested a new approach [67], referred to as the JK method. This ap-
proach differs from ELHAV by the absence of the antisymmetrizer in the energy
expression (Eq. 11) i.e. F = A, G = 1, and ψ0 = N0 Aφ0 . As discussed in detail
below, the second-order JK energy exhibits the correct asymptotic behavior,
i.e. recovers the exact value of C6 and a few higher van der Waals constants.
The third- and higher-order JK energies, however, have incorrect asymptotic
behavior, and the term C11 /R11 for spherically symmetric atoms, or C9 /R9 for
polar molecules, is not fully recovered by a finite-order JK expansion.
A method to further refine the low-order asymptotics of the JK theory
has been proposed in [69]. In this approach, one retains the form of F
and G, but improves the function used to start the iterative process, using
suitable functions from the RS theory. If the iterations are started from
(1)
ψ0 = N1 A(φ0 + ψRS ) = N1 A(φ0 – R0 Vφ0 ), where the constant
1
N1 =

 (16)
φ0 |Aφ0 – φ0 |AR0 Vφ0

enforces the intermediate normalization of ψ0 , one obtains the so-called JK-1

(n)
expansion. The individual corrections ψJK–1 are calculated from the formula

n
ψ (n) = – R0 VAψ (n–1) + E(k) R0 Aψ (n–k) – N1(n–1) AR0 Vφ0 (17)
k=1


+ N1(n) Aφ0 (n–1)
+ N1 R0 (V – φ0 |Vφ0 )Aφ0 + N1(n–2) R0 AVR0 Vφ0
(n–2)
– N1 R0 VAR0 Vφ0 ,
for n ≥ 1, where
⎧
k
⎨ k+1
(k) N0 φ0 |AR0 Vφ0 k≥0
N1 = (18)
⎩0 k<0

and ψ (0) = N0 Aφ0 . The energy corrections are obtained from Eq. 8, with the
RS functions replaced by the corresponding ones calculated from Eq. 17. The
derivation of Eq. 17 involves the commutation relation
[En – V, A] = [H0 – E0 , A] (19)
and is presented in detail in [69]. One can show that the improvement of ψ0
resulting from the use of the first-order RS wave function makes the resulting
56 K. Szalewicz et al.

theory asymptotically correct to one order further than JK, i.e. JK-1 exhibits
the correct asymptotics in the second and third order. The asymptotic behav-
ior of E(4)
JK–1 and higher corrections remains incorrect.
It is quite instructive to look in more detail into the lowest-order energy
corrections and their asymptotics for the SAPT theories discussed so far. For
this purpose, we introduce a simplified notation X ≡ φ0 |Xφ0  for any op-
erator X. Using the formulae referenced in Table 1, one can show that the
low-order RS, SRS, ELHAV, JK, and JK-1 corrections are equal to


E(1)
RS = V (20)


(2)
ERS = – VR0 V (21)




(3)
ERS = VR0 VR0 V – V VR20 V (22)


(1)
ESRS = N0 VA (23)




E(2) 2
SRS = – N0 VAR0 V + N0 VA AR0 V (24)






E(3)
SRS =N0 VAR0 VR 0 V – N 0 V VAR2
0 V – N 0 VA AR0 VR0 V (25)






+ N0 VA V AR20 V – N0 VAR0 V AR0 V


2
+ N02 VA AR0 V


(1)
EELHAV = N0 VA (26)




2

E(2) 2
ELHAV = 2N0 VAR0 A VA – N0 VAR0 AV – N0 VA AR0 A
3
(27)


E(1)
JK = N 0 VA (28)




E(2)
JK = – N 0 VAR 0 V + N 2
0 VA AR 0 V (29)




(3)
EJK = – N02 VA VR0 VAR0 A + N0 VR0 VAR0 AV (30)

2



+ N03 VA VR0 AR0 A – N02 VA VR0 AR0 AV


2


+ N03 VA AR0 V – N02 AR0 V VAR0 V


E(1)
JK–1 = N0 VA (31)






(2)
EJK–1 = – N0 VAR0 V + 2N02 VA AR0 V – N0 V AR0 V (32)
Intermolecular Interactions 57




(3)
EJK–1 = N0 VR0 VAR0 V – N02 VA VR0 AR0 V (33)





– N02 VA VR0 VAR0 A + N0 V VR0 VAR0 A

2




+ N03 VA VR0 AR0 A – N02 V VA VR0 AR0 A


2

2
+ 4N03 VA AR0 V – 2N02 V AR0 V



– 2N02 AR0 V VAR0 V

It has been shown [66] that the RS expansion recovers the exact R–n asymp-
totics of the interaction energy,
 
  N 
 (n) 
E – ERS  = O(R–3N–3 ) . (34)
 
n=1

Therefore, we may analyze the asymptotic behavior of various SAPT for-

malisms by comparing the energy corrections to the RS ones. It may be shown
that the insertion of a single symmetry projector A into an expectation value
expression · · · , containing any multiple product of the operators R0 and V,
and a simultaneous multiplication of the whole expression by N0 , does not
influence the R–n asymptotic behavior. In the simplest example, the expres-
sions V and N0 VA exhibit the same asymptotics, i.e. the first-order SRS
(1) (1)
exchange energy ESRS – ERS = N0 VA – V vanishes exponentially with R.
However, no similar asymptotic equality exists for expressions · · ·  involv-
ing more than one symmetry projector. For example, the asymptotics of the
expression VAR0 AV, entering the second-order ELHAV energy, is deter-
mined not only by VR0 V, but contains also the so-called “double exchange”
terms of the form VPij R0 Pij V, where Pij transposes the coordinates of elec-
trons i and j, which do not vanish exponentially with R [67].
By removing the single occurrences of A, as explained in the previous
paragraph, one sees that the second- and third-order SRS energies are asymp-
totically equivalent to the corresponding RS ones. To derive this result we
used the fact that R0 φ0 = 0, therefore, e.g. R0 V = 0, and thus the quantity
AR0 V vanishes exponentially with R. In fact, any SRS energy correction
E(n) (n)
SRS exhibits the same correct asymptotic behavior as ERS [66]. On the other
hand, the formula for the second-order ELHAV energy contains expressions
· · ·  involving double antisymmetrizers, and the asymptotic behavior of
(2)
such expressions is by no means related to that of ERS . For the JK theory,
the second-order energy is the same as in the SRS approach, so the asymp-
(3)
totic behavior is correct. However, EJK contains expressions · · ·  with double
antisymmetrizers, and it does not behave correctly in the asymptotic limit.
In the JK-1 theory developed in [69], the second-order energy differs from
the corresponding SRS value. However, by removing the single occurrences
58 K. Szalewicz et al.

of A one may easily show that this difference vanishes exponentially, so the
(2)
asymptotic behavior of EJK–1 is also correct. The same result holds for the
(3)
third order. It is worth noting that EJK–1 is asymptotically correct despite the
fact that the 3rd through 6th terms in Eq. 33 individually exhibit unphysical
long-range (1/R9 ) behavior. It turns out that the long-range parts of these
terms mutually cancel out and the sum of these terms becomes proportional
to N0 VA – V, an exponentially vanishing quantity. This cancellation does
(4)
not take place in EJK–1 and in higher corrections, and no non-regularized
theory is both convergent and asymptotically correct to any order of pertur-
bation theory.
Obviously, one can improve the function ψ0 used to start the iterative
process, Eqs. 10–11, further, using higher wave functions of the polariza-
(1) (2)
tion theory, e.g. setting ψ0 = N2 A(φ0 + ψRS + ψRS ), where the constant N2
is such that ψ0 fulfills the intermediate normalization condition. The JK-2
method obtained with this starting point should be asymptotically correct to
the fourth order of perturbation theory. However, this approach is not suit-
able for the construction of a theory that can converge in the presence of
the Pauli-forbidden continuum and is at the same time asymptotically cor-
rect to any order. For this purpose, one must use the regularization technique
described in Sect. 3.
High-order convergence behavior of several other SAPT expansions, in-
cluding the Murrell-Shaw-Musher-Amos (MSMA) [85, 86], ELHAV, and JK
theories, was studied in [69] on the same example of the LiH system. It was
shown that expansions such as ELHAV and JK converge despite the pres-
ence of the Pauli-forbidden continuum, and that the asymptotic correctness
of the second-order JK energy leads to much better low-order convergence
compared to ELHAV. It turned out, however, that in practical applications the
asymptotic behavior of third- and higher-order corrections is far less signifi-
cant than the asymptotics of the second-order energy, and JK-1 provides no
improvement over JK except for very large intermonomer distances.

2.3
Conjugate Formulation of SAPT and the HS Theory

In view of the fact that the operators V and A do not commute, the order of
operators chosen in the definition of the symmetry-forcing technique (Eqs. 10
and 11) is not the only possible one and a different theory is obtained if one
replaces the operator VA by its Hermitian conjugate, the operator AV. This
conjugate formulation of SAPT allows one to define the Amos-Musher pertur-
bation theory [62]. It was also employed in one of the original formulations of
the ELHAV method [59].
The Amos-Musher theory in its pure, original form [62] is simply the RS
perturbation theory with the zero-order Hamiltonian H0 and the perturba-
Intermolecular Interactions 59

(n)
tion equal to AV. Thus, the energy and wave function corrections EAM and
(n)
ψAM are obtained from Eqs. 8 and 9, respectively, with all the occurrences
of V replaced by AV. As in the definition of the symmetrized Rayleigh-
(n)
Schrödinger approach, one may employ the functions ψAM in the SRS energy
expression, Eq. 12, to define another set of energy corrections, which will be
referred to as the symmetrized AM (SAM) energies. It turns out that the low-
order SAM corrections are much more accurate than the corresponding pure
AM values [72, 87].
Another modification of the original Amos-Musher theory has been pro-
posed by Adams [55]. The modified AM Hamiltonian takes the form

HAM = H0 + A(V – D) (35)

where D is a (constant) offset parameter chosen such that the perturbation

expansion in powers of A(V – D) converges faster than the original one.
Common choices for D are the first-order polarization energy φ0 |Vφ0  and
the Heitler-London energy N0 φ0 |VAφ0 . The interaction energy E can be
obtained by adding D to the sum of the AM corrections or by summing up
the SAM energies (no addition of D is needed in this case). The choice of
D has some impact on the convergence properties of the AM method but
is practically inconsequential when the SAM expansion is used; the relevant
numerical data for the lithium hydride and three choices of D are given in
Table 2 and Fig. 1 for the triplet and singlet states, respectively. A similar re-
sult has been obtained for a triplet He atom interacting with a ground-state
H atom [87]. In all cases, the low-order SAM results are much more accurate
than the ones obtained with the nonsymmetrized AM approach.
To obtain the corrections of the ELHAV theory [59] in the conjugate ap-
proach, one starts from the equation

(H0 – E0 )ψ + A(V – E)ψ = 0 , (36)

substitutes V → ζV, and expands ψ and E in powers of ζ. The relationship

between the approach of Eq. 36, the formulation by van der Avoird [60], and
the symmetry-forcing derivation of the ELHAV theory shown in Sect. 2.2 has
been discussed in detail in [40]. All three approaches lead to identical energy
corrections; however, the wave function corrections are different.
It is worth mentioning that the development of perturbation theories from
the equation

(H0 – E0 )ψ̃ + (V – E)Aψ̃ = 0 (37)

(this equation is conjugate to Eq. 36 used by Hirschfelder [59]) requires some

extra caution. Suppose that one substitutes V → ζV and performs the expan-
sion in powers of ζ directly in Eq. 37. The first-order wave function is then
60 K. Szalewicz et al.

Table 2 Convergence of the conventional AM and SAM expansions for the triplet state of
LiH, for R = 11.5 bohr and three different values of the offset parameter: D = 0 (columns
marked “0”), D = φ0 |Vφ0  (columns marked “pol”), and D = N0 φ0 |VAφ0  (columns
marked “HL”). The numbers listed are percent errors with respect to the FCI interaction
energy

AM SAM
n 0 pol HL 0 pol HL

2 –118.0469 –109.5340 –160.1655 –80.7499 –80.7501 –80.7488

3 –85.8639 –80.5432 –112.1885 –32.2093 –32.2100 –32.2062
4 –58.5816 –55.2563 –75.0332 –13.0982 –13.0989 –13.0945
5 –38.6454 –36.5674 –48.9257 –5.4104 –5.4110 –5.4071
6 –25.0044 –23.7060 –31.4279 –2.2656 –2.2661 –2.2630
7 –15.9887 –15.1774 –20.0018 –0.9609 –0.9612 –0.9590
8 –10.1477 –9.6409 –12.6548 –0.4126 –0.4129 –0.4113
9 –6.4095 –6.0929 –7.9755 –0.1794 –0.1796 –0.1785
10 –4.0354 –3.8376 –5.0135 –0.0790 –0.0792 –0.0785
15 –0.3905 –0.3718 –0.4834 –0.0016 –0.0016 –0.0016
20 –0.0373 –0.0356 –0.0462 0.0000 0.0000 0.0000
25 –0.0036 –0.0034 –0.0044 0.0000 0.0000 0.0000
30 –0.0003 –0.0004 –0.0005 0.0000 0.0000 0.0000

Fig. 1 Convergence of the conventional AM and SAM expansions for the singlet state of
LiH, for R = 3.015 bohr and three different values of the offset parameter: D = 0 (curves
marked “0”), D = φ0 |Vφ0  (curves marked “pol”), and D = N0 φ0 |VAφ0  (curves marked
“HL”). The numbers displayed are percentages of the FCI interaction energy recovered in
the nth order perturbation treatment
Intermolecular Interactions 61

given by
(1)
ψ̃ELHAV = R0 (E(1) – V)Aφ0 , (38)

where E(1) is the first-order energy. This function does not vanish when
R → ∞. On the other hand, in all the perturbation theories defined before,
including the ELHAV and JK methods, ψ (1) , and all subsequent ψ (n) , vanish
identically at infinite intermonomer separations. One can easily show that the
(1) (1)
difference ψELHAV – N0 ψ̃ELHAV is equal to

N0 R0 [V – E(1) , A]φ0 = – N0 R0 [H0 – E0 , A]φ0 = φ0 – N0 Aφ0 (39)

(1)
and, consequently, the whole correction ψ̃ELHAV does not vanish when R →
(n)
∞. One can prove by induction that the same holds for all ψ̃ELHAV , n > 1. The
same problem is encountered if one starts from Eq. 36 and does not expand it
directly, but substitutes ψ = N0 Aφ0 + χ and applies the commutation relation

[E – V, A] = [H0 – E0 , A] (40)

to obtain an expansion for χ (exactly as was done in [69] to derive the JK-1
expansion).
The nature of the problem described above can be understood clearly
when one considers the fact that a successful SAPT expansion either con-
verges to the exact wave function of the dimer, or to a primitive function that
is localized in the same way as φ0 and gives the exact wave function when
antisymmetrized. For these two cases, one should use something like N0 Aφ0
and φ0 , respectively, as a zero-order function. The “wrong” theories described
above try to do this the other way round – they either start from φ0 and con-
verge to N0 Aφ0 , or vice versa. Since φ0 and N0 Aφ0 are always completely
different, even for R → ∞, convergence of such a theory is bound to be very
slow.
Whether the perturbation functions vanish asymptotically or not depends
on the application of the commutation relation, Eq. 40, in deriving the per-
turbation series. This relation, as already noted by Hirschfelder [78], is
somewhat paradoxical since a first-order quantity on the l.h.s. is equated to
a (nonvanishing) zero-order quantity on the r.h.s. In other words, the equality
[E(ζ) – ζV, A] = [H0 – E0 , A] is not valid for arbitrary ζ, but only for ζ = 1.
Thus, each application of Eq. 40 changes the way the order of perturbation
theory is defined. This paradox is analyzed in more detail in a recent publi-
cation by Adams [88].
All the SAPT methods considered so far have used a single operator A[λ]
projecting onto a specific subspace H [λ] ⊂ HA ⊗ HB . The HS perturbation
theory, introduced by Hirschfelder and Silbey [50], follows a different, mul-
tistate philosophy. It performs a perturbation expansion for a primitive func-
62 K. Szalewicz et al.

tion

Φ= c[λ] ψ [λ] , (41)
[λ]

where the sum goes over all asymptotically degenerate eigenstates ψ [λ] of H
which dissociate into the specific states of the monomers, i.e. over all per-
mutation symmetries [λ] into which the Hilbert space HA ⊗ HB decomposes
under the action of SN , including the symmetries that lead to unphysical,
Pauli-forbidden states. The primitive function in the HS method is defined
uniquely by the localization conditions [89]


A[λ] φ0 |(H0 – E0 )Φ = 0 (42)

for all [λ]. Once Φ is known, the eigenstates of H can be extracted by taking
projections A[λ] Φ. One can easily verify that Φ satisfies the equation

(H0 – E0 )Φ = – VΦ + E [λ] A[λ] Φ . (43)
[λ]

Substitution V → ζV in Eqs. 42–43 and expansion of Φ and E [λ] in powers of

ζ gives the following equations for the HS perturbation corrections [49],

–1


[λ] (n)
EHS = φ0 |A[λ] φ0 φ0 |VA[λ] Φ (n–1) (44)


n–1

[λ] (k) [λ] (n–k)
– EHS φ0 |A Φ ,
k=1

 
n 
[λ] (k) [λ] (n–k)
H0 – E0 Φ (n) = – VΦ (n–1) + EHS A Φ , (45)
k=1 [λ]

where Φ (0)≡ φ0 . The last term in Eq. 45 couples all asymptotically degener-
ate states (all symmetries [λ]). Thus, the HS theory is far more complicated in
practical applications than the other approaches presented so far, especially
for many-electron systems where the large number of different permuta-
tional symmetries [λ] leads to unphysical states asymptotically degenerate
with each physically allowed one.
High-order convergence studies of the HS perturbation expansion (as well
as of the RS and SRS expansions) for Li – H, the simplest system includ-
ing a more-than-two-electron monomer, have been presented in [57]. These
studies show that the HS expansion, as the RS and SRS expansions dis-
cussed above, is indeed divergent for a system exhibiting the Pauli-forbidden
continuum. However, low-order SRS and HS results turned out to be quite
accurate, and it was possible to obtain extremely accurate results by sum-
ming up the perturbation corrections until these corrections start to grow
Intermolecular Interactions 63

in absolute value (the standard method of summing an asymptotically con-

vergent series). Surprisingly, the HS expansion behaved no better than the
much simpler SRS series, unlike in the case of interactions between one- and
two-electron monomers.

3
Convergence Properties of Regularized SAPT

3.1
Methods of Regularizing the Coulomb Potential

The existence of an unphysical continuum surrounding the physical states

of interacting many-electron systems is caused by the fact that, when the
wave function does not obey the Pauli principle, electrons that were ini-
tially assigned to one monomer can fall into the Coulomb wells of the other
monomer, ejecting some other electrons into the continuum (as in the Auger
process). This is possible since the negative Coulomb wells in V are of exactly
the same magnitude as the ones in H0 . The main idea behind the regulariza-
tion of the Coulomb potential, a concept first employed by Herring [90] in his
studies of the asymptotics of the exchange energy, is to remove all negative
singularities from V so that the electrons belonging to one monomer cannot
fall into the wells around the nuclei of the other monomer. Such a modified
Coulomb potential should be as similar as possible to the original one. In
particular, it must exhibit the same large-R asymptotics, i.e. the difference
between regularized and non-regularized Coulomb potentials must vanish
exponentially with the distance from the nucleus. Moreover, the one-electron
integrals involving the regularized potential should be easy to evaluate.
A simple choice for a regularized Coulomb potential would be


1/c r≤c
vp (r) = (46)
1/r r>c

where c > 0 is a parameter. This is the potential used by Adams in his re-
cent work on the regularized SAPT [71]. However, it is preferable that vp (r)
is smooth for r > 0. Moreover, the one-electron integrals involving vp (r)
of Eq. 46 and Gaussian basis functions are not so easy to compute. There-
fore, two slightly more complicated analytic forms of vp (r) were employed in
recent work [70]:

1 2
vp (r) = (1 – e–ηr ) , (47)
r
64 K. Szalewicz et al.

and
1 √
vp (r) = erf( ωr) , (48)
r
where η > 0 and ω > 0 are parameters defining the strength of the regulariza-
tion, and erf(z) is the standard error function
z
2 2
erf(z) = √ e–t dt . (49)
π
0

The choice defined by Eq. 47, which will be referred to as the Gaussian
regularization, has the advantage that the matrix elements of vp between
Gaussian basis functions can be evaluated in exactly the same way as the
ordinary one-electron potential energy integrals. The regularized potential
of Eq. 48 was first employed by Ewald [91] in his calculations of the Madelung
constants in crystals. More recently, it was used in the linear scaling electronic
structure theory [92, 93] and in the description of the electron correlation
cusp [94]. This potential is analytic for any r – there is no cusp at r = 0. It is
worth noting that, since
 3/2  –ωr2 √
ω e  erf( ωr)
dr = , (50)
π |r – r  | r
R3

the potential of Eq. 48 can be interpreted as the electrostatic potential of

a smeared unit charge, with the charge distribution defined by a Gaussian
2
function e–ωr . Thus, the regularization defined by Eq. 48 corresponds to re-
placing a point nuclear charge by a smeared charge of the same total value,
and it will be referred to as the smeared nuclear charge (SNC) regulariza-
tion.
The difference between the original and regularized Coulomb potentials,
1
vt (r) = – vp (r) , (51)
r
which will be referred to as the singular or residual part of the Coulomb po-
tential, is a short-range function with a singularity at r = 0. We have been
using the subscripts p and t to remind the reader that the potentials vp and
vt are responsible for the polarization and tunneling aspects, respectively, of
the interaction phenomenon.
The interaction operator V of two atoms A and B can now be split into its
regular part Vp and singular part Vt as follows,

ZA ZB   
A 
A Z B Z Z BZ
1
Vp = – ZB vp (rBi ) – ZA vp (rAj ) + , (52)
rAB rij
i=1 j=1 i=1 j=1
Intermolecular Interactions 65

and

ZA 
ZB
Vt = – ZB vt (rBi ) – ZA vt (rAj ) , (53)
i=1 j=1

where rpq = |r p – r q | denotes the distance between particles p and q. The par-
ticles in Eqs. 52–53 are the nuclei A and B (with atomic numbers ZA and
ZB , respectively), and the electrons initially assigned to A (enumerated by i)
and to B (enumerated by j). In Eqs. 52–53, as well as throughout the whole
text, atomic units are used. Note that only the one-electron, attractive part
of the Coulomb potential has been regularized in Eq. 52, therefore, the ap-
proach of Eqs. 52–53 will be referred to as the one-electron regularization. The
full regularization, corresponding to the partitioning of V into Vpfull and Vtfull
operators given by,

ZA ZB   
A 
A Z B Z B Z Z
Vpfull = – ZB vp (rBi ) – ZA vp (rAj ) + vp (rij ) , (54)
rAB
i=1 j=1 i=1 j=1

ZA 
ZB 
ZA 
ZB
Vtfull = – ZB vt (rBi ) – ZA vt (rAj ) + vt (rij ) (55)
i=1 j=1 i=1 j=1

has also been tested but only in the case of two interacting hydrogen
atoms [70]. This approach turned out to perform very similarly to the one-
electron regularization [70]. It is, however, significantly more complicated
computationally, since the regularized two-electron integrals are required in
this case. The one-electron regularization is preferable also on the theoretical
grounds since it does not affect the dispersion interaction at all. This results
from the fact that the dispersion part of the interaction energy E does not
depend on the one-electron part of V.

3.2
Regularized SRS Expansion

When one neglects the operator Vt completely, the Schrödinger equation

takes the form
 
H0 + Vp (η) – E0 ψp (η) = Ep (η)ψp (η) , (56)

where it has been explicitly stated that the eigenfunction ψp and the eigen-
value Ep depend on the value of the regularization parameter η (or ω, when
the SNC regularization is employed). Equation 56 can be solved by means of
the standard RS perturbation theory; the resulting expansion in powers of ζ,
which will be referred to as the regularized RS (R-RS) expansion, takes the
66 K. Szalewicz et al.

form
∞
 (n)
Ep (η) = ER-RS (η)ζ n , (57)
n=1
∞
(n)
ψp (η) = ψR-RS (η)ζ n , (58)
n=0

with the individual terms given by Eqs. 8 and 9 with V replaced by Vp ,

and the RS functions and energies replaced by their regularized counterparts
(0)
(obviously, ψR-RS ≡ φ0 ). Note that in Eq. 56, unlike in the non-regularized
Schrödinger equation, the permutational symmetry is broken. Thus, one may
expect that ψp (η) will be localized in the same way as φ0 . In fact, for some
range of values of the regularization parameter, the function ψp (η) should be
close to the exact primitive function, i.e. the function Aψp (η) should provide
a good approximation to the exact eigenfunction ψ for any permutational
symmetry forced by the projector A.
Knowing ψp (η), one can obtain an approximation to the exact interaction
energy E by an SRS-like energy formula,


φ0 |VAψp (η)
E ≈ ER-SRS (η) =
 . (59)
φ0 |Aψp (η)

Substituting V → ζV and expanding Eq. 59 in powers of ζ leads to the expan-

sion
∞

ER-SRS = E(n) n
R-SRS ζ , (60)
n=1

where the coefficients E(n)R-SRS , which will be referred to as the regularized SRS
(k) (k)
(R-SRS) corrections, are given by Eq. 12 with ψRS replaced by ψR-RS and E(k)
SRS
(k)
replaced by ER-SRS . All the energies in Eq. 60 depend on the value of the reg-
ularization parameter η (or ω). This dependence will not be explicitly shown
as long as it does not lead to ambiguities. Note that Eq. 59, as well as the ex-
pression for the corrections E(n) R-SRS , contains the full interaction operator V,
not just Vp . In the limit η → ∞ (for the Gaussian regularization) or ω → ∞
(for the SNC regularization) the R-RS and R-SRS expansions defined above
are identical to the ordinary RS and SRS theories, respectively.
For a suitable range of values of the regularization parameter, the R-SRS
expansion, unlike the non-regularized SRS theory, can be expected to con-
verge since the neglect of Vt shifts the unphysical continuum upwards in the
energy, possibly above the physical states. Moreover, as vt (r) is a short-range
potential, the R-RS and R-SRS energy corrections exhibit the same correct
asymptotic behavior as the standard RS and SRS theories. The main draw-
Intermolecular Interactions 67

back of the R-SRS series is that its sum ER-SRS (η) differs somewhat from the
exact interaction energy E. To account for this difference, one has to construct
a theory that takes into account not only Vp , but also Vt . There are several
possible choices of such a theory, as discussed in the next subsections.

3.3
A Posteriori Inclusion of V t

As a first step towards construction of a regularized SAPT theory that in-

cludes both Vp and Vt , let us note that if ψp and Ep are known, the remaining
part of the interaction energy can be recovered by means of a perturba-
tion expansion in powers of Vt . Since this expansion does not influence the,
already correct, asymptotics of Ep , we can now employ a method that is
convergent despite the presence of the Pauli-forbidden continuum, i.e. the EL-
HAV, AM, or JK theory. Unlike the case of non-regularized expansions, there
is no asymptotics-related reason to expect that the JK method will perform
better than the other two. If we choose the ELHAV theory, the successive cor-
rections to the energy and the wave function, referred to as the regularized
ELHAV (R-ELHAV) corrections, are obtained from the formulae [40]

 n–1

E(n)
R-ELHAV
(n–1)
= ψp |Vt AψR-ELHAV – E(k)
R-ELHAV ψp |Aψ (n–k)
R-ELHAV , (61)
k=1
(0)
ψR-ELHAV ≡ Np Aψp , (62)
 
(1) (1)
ψR-ELHAV = Np Rp A ER-ELHAV – Vt ψp , (63)

and

(n) (n–1)

n
(k) (n–k)
ψR-ELHAV =– Rp Vt AψR-ELHAV + ER-ELHAV Rp AψR-ELHAV (64)
k=1

for n ≥ 2. In these equations, Rp is the ground-state reduced resolvent of the

operator H0 + Vp , and Np = ψp |Aψp –1 . If one chose to use the JK method
(n)
instead of ELHAV, the wave function corrections ψR-JK would be calculated
(n)
from Eqs. 62–64 as well, only the energy corrections ER-JK would be defined
differently,


E(n)
R-JK = ψ |V ψ (n–1)
p t R-JK . (65)

(k)
Once the wave function corrections ψR-ELHAV have been calculated, one
can use the SRS-like formula, Eq. 59, and define an alternative expansion for
the interaction energy, called in [70] the R2-ELHAV expansion. The expres-
68 K. Szalewicz et al.

sion for the R2-ELHAV energy corrections is


E(n)
R2-ELHAV
(n–1)
=N0p φ0 |VAψR-ELHAV (66)


n–1

– E(k) φ
R2-ELHAV 0 |Aψ (n–k)
R-ELHAV ,
k=1

where


ψp |Aψp
N0p =
 . (67)
φ0 |Aψp

The R2-ELHAV approach has the slight advantage that E(1) R2-ELHAV = ER-SRS , so
the second- and higher-order corrections can be viewed as small contribu-
tions improving the, already quite accurate, infinite-order R-SRS energy.

3.4
Double Perturbation Approach

If the R-ELHAV expansion is able to effectively reproduce the part of the in-
teraction energy missing in ER-SRS in a low-order treatment (as will be seen
in Sect. 4, this is the case), it is desirable to extend this theory to obtain
a perturbation expansion that starts from φ0 and takes both Vp and Vt into
account. For this purpose, the most straightforward idea is to develop some
double perturbation expansion in Vp and Vt which treats these two perturba-
tions in an SRS-like and ELHAV-like way, respectively. The formulae defining
the wave function corrections in this double perturbation theory can be ob-
tained by expanding the equation

(H0 – E0 + µVp )ψ + A(νVt – E)ψ = 0 (68)

in powers of µ and ν. Note that if Vp = 0 and Vt = V, Eq. 68 defines the ELHAV

theory in Hirschfelder’s formulation [59]. Assuming the convention that the
first index refers to the order in the perturbation Vp , and setting ψ (0,0) ≡ φ0 ,
one obtains


i 
j
(i,j) (i–1,j) (i,j–1) 
ψ = – R0 Vp ψ – R0 AVt ψ + R0 A E(k,l) ψ (i–k,j–l) ,
k=0 l=0
(69)
Intermolecular Interactions 69

where the energies E(i,j) are given by



E(i,j) =N0 φ0 |Vp ψ (i–1,j) + φ0 |AVt ψ (i,j–1) (70)

 j
i 

 (k,l) (i–k,j–l)
– E φ0 |Aψ ,
k=0 l=0

the prime in the summation over (k, l) denotes omission of the term k = l =
0, and the double prime – omission of the terms k = l = 0 and k = i, l = j. To
keep Eqs. 69 and 70 compact, we defined here ψ (i,j) ≡ 0 whenever i < 0 or j <
0. An alternative formula for the energy corrections, corresponding to the R2-
ELHAV approach, can be obtained by expanding the equation


φ0 |(µVp + νVt )Aψ
E=
 (71)
φ0 |Aψ

in powers of µ and ν. The result is



E (i,j) =N0 φ0 |Vp Aψ (i–1,j) + φ0 |Vt Aψ (i,j–1) (72)


i 
j

 (k,l) (i–k,j–l)
– E φ0 |Aψ .
k=0 l=0

Double perturbation theory calculations are very time-consuming if one

wants to go to high orders. Therefore, it would be highly advantageous to
combine Vp and Vt in a single perturbation theory, related to Eqs. 69–72
as closely as possible. We will present such a theory – the R-SRS+ELHAV
method [72] – in the next subsection.

3.5
The “All-in-one” R-SRS+ELHAV Theory

To derive perturbation equations for a theory that uses φ0 as the zero-order

function, takes into account both Vp and Vt , and avoids the complications of
a double perturbation theory framework, we start from the following equa-
tion [72],
 
H0 – E0 + Vp – Ep + A(Vt – Et ) ψ = 0 , (73)

where Et = E – Ep . Performing the substitution Vp → ζVp , Vt → ζVt , and

using the already known R-RS perturbation expansion for Ep , Eq. 57, one
70 K. Szalewicz et al.

finds that the coefficients in the expansions

∞
 (n)
Et (ζ) = Et ζ n , (74)
n=1
∞

ψ(ζ) = ψt(n) ζ n (75)
n=0

can be calculated from the equations

 

 n–1

(n) (n–1) (k) (n–k) (n)
Et = N0 φ0 |(Vp + AVt )ψt – Et φ0 |Aψt – ER-RS (76)
k=1

and
 
(n) (n–1)

n–1
(k) (n–k)

n
(k) (n–k)
ψt =– R0 (Vp + AVt )ψt – ER-RS ψt – Et Aψt , (77)
k=1 k=1
(0) (n)
where ψt = φ0 . Once the wave function corrections ψt are known, the
energy corrections are calculated from the SRS-like formula, Eq. 15, with
(k) (k)
ψELHAV replaced by ψt . The SAPT expansion defined in this way may be
regarded as a single-step combination of the R-SRS method and the ELHAV
theory and will be referred to as the R-SRS+ELHAV expansion. It has been
found that the low-order R-SRS+ELHAV energies are significantly more accu-
rate than the corresponding sums of the coefficients E(n) (n)
R-RS and Et (as is also
the case for the non-regularized AM and SAM energies, cf. Sect. 2.3).
(n)
One should note that in order to calculate ψt one has to obtain the R-
RS energy corrections up to nth order from a separate expansion. This is not
a significant computational complication, and it should be contrasted with
(n)
the R-ELHAV theory, where to calculate ψR-ELHAV for any n one must know
the energy Ep to infinite order in Vp . As in the preceding subsection, we em-
ployed the conjugate formulation of SAPT (cf. Eq. 36 of Sect. 2.3) to obtain
the starting point for the development of the R-SRS+ELHAV method. One
can also try to develop a theory by an extension of the symmetry-forcing
formalism, i.e. starting from the equation
 
H0 – E0 + Vp – Ep + (Vt – Et )A ψ = 0 . (78)

The perturbation equations for such a conjugate R-SRS+ELHAV formal-

ism [95] turned out to be significantly more complicated than Eqs. 76–
77, both formally and computationally. However, numerical results that
we have obtained using this conjugate approach (for the LiH system) dif-
fered insignificantly from the results of the R-SRS+ELHAV theory discussed
here.
Intermolecular Interactions 71

3.6
The R-SRS+SAM Approach

The R-SRS+ELHAV theory outlined above is not the only possible way of in-
cluding both Vp and Vt in a single perturbation treatment. Another method
of doing so has been introduced by Adams in a recent contribution [71].
Adams called his method the corrected SRS (cSRS), however, we will use the
name R-SRS+SAM to emphasize the relation of his theory to the symmetrized
Amos-Musher approach.
In the R-SRS+SAM theory, the Schrödinger equation takes the form

H0 + Vp + A(Vt – D) ψR-SRS+AM = (E0 + ER-SRS+AM )ψR-SRS+AM . (79)
The choice of a particular offset D does not influence the results significantly
(cf. Sect. 2.3). Note that if Vp = 0 and Vt = V, Eq. 79 would be identical to the
one appearing in the AM perturbation theory [62]; in other words, the Hamil-
tonian H0 + Vp + A(Vt – D) includes Vp and Vt in an RS-like and AM-like
way, respectively. The eigenproblem 79 can be solved by means of the stan-
dard RS perturbation theory. In the resulting expansion of the R-SRS+AM
wave function
∞
 (n)
ψR-SRS+AM = φ0 + ψR-SRS+AM , (80)
n=1
(n)
the coefficients ψR-SRS+AM can be obtained from Eqs. 8 and 9 with V replaced
by Vp + A(Vt – D) and all the RS corrections replaced by their R-SRS+AM
counterparts. The interaction energy can be calculated as
∞

E =D+ E(n)
R-SRS+AM . (81)
n=1
However, significantly more accurate results are obtained when one fol-
lows the SAM (or SRS) algorithm and notes that AψR-SRS+AM satisfies the
Schrödinger equation with the full Hamiltonian H. Thus,


φ0 |VAψR-SRS+AM

 =E (82)
φ0 |AψR-SRS+AM

and the energy corrections can be calculated, as in the SRS method,

(k) (k) (n)
from Eq. 12 with ψRS replaced by ψR-SRS+AM . The corrections ER-SRS+SAM ob-
tained in this way will be named R-SRS+SAM energies, as opposed to the
(n)
nonsymmetrized R-SRS+AM energies ER-SRS+AM calculated along with the
R-SRS+AM wave functions using an analog of Eq. 8.
One may note that the eigenproblems 73 and 79 differ, apart from the
(insignificant) presence of the offset D in the latter, only by the term AEt
72 K. Szalewicz et al.

in Eq. 73 versus Et in Eq. 79. Speaking more generally, these two approaches
both start from φ0 and apply weak symmetry forcing to Vp and strong sym-
metry forcing to Vt , so they both can be viewed as refinements of the regu-
larized SRS theory. In view of these similarities, these approaches were given
similar names.

3.7
Zero-order Induction Theory

Apart from the R-SRS+SAM theory itself, Adams also proposed [71] a very
interesting extension to this method in which the induction effects are in-
cluded already in the zeroth-order energy and wave function – the so-called
zero-order induction (ZI) theory. The ZI scheme can be employed to refine
the R-SRS+ELHAV approach in the same manner as Adams used it on top of
the R-SRS+SAM theory; we will refer to these methods as R-SRS+ELHAV+ZI
and R-SRS+SAM+ZI, respectively. Performing third-order (first-order in the
wave function) calculations for the singlet state of LiH, Adams found [71] that
R-SRS+SAM+ZI provides a significant improvement over SRS for distances
around the chemical minimum, i.e. around 3 bohr. For larger interatomic sep-
arations, the accuracy of the third-order R-SRS+SAM+ZI energy decreased
significantly, although the results were still better than the SRS ones.
The zero-order induction approach differs from its parent R–SRS+ELHAV
(Eq. 73) or R–SRS+SAM (Eq. 79) theory in the specific choice of the zero-order
Hamiltonian and the regular part of the perturbation operator. These operators
are replaced by new operators H 0 = HA + HB and 
Vp defined such that the ef-
fects of the (regularized) induction interaction are included in the zeroth order.
Specifically, one can set H A = HA + ΩB and H B = HB + ΩA , where
   1 (0)
ΩB = – ZB vp (rBi ) + ρ (r j )drj (83)
i∈A i∈A
rij B

is the operator of the electrostatic potential of atom B resulting from the

regularized Coulomb attraction of the nucleus and the repulsion of the un-
perturbed electronic charge distribution ρB(0) (r j ) of atom B. The definition of
ΩA is obtained by interchanging A and B in Eq. 83. In accordance with the
changes in H0 , the long-range part of the perturbation now takes the form
Vp = Vp – ΩA – ΩB , and the short-range part remains unchanged, 
 Vt = Vt .
The zero-order wave function has the form  φ0 =  φAφB and the zero-order
energy is 
E0 = 
EA +  X
EB , where H EX
φX =  φX for X = A, B. Note that this defin-
0 differs from the original Adams’ formulation [71] by the absence
ition of H
of a small combinatorial factor multiplying Vt . This difference does not ap-
pear to be significant in practice.
The X+ZI energy corrections E(n) X+ZI , where X = R–SRS+ELHAV or X =
R–SRS+SAM, are now calculated as in its parent approach, except that all
Intermolecular Interactions 73

functions and operators are replaced by their tilded counterparts. The inter-
action energy differs from the sum of the X+ZI corrections by the induction
contribution E0 – E0 contained in the zero-order energy 
E0 .
It is worth noting that the R-SRS+ELHAV+ZI theory inherits all the advan-
tages of the R-SRS+ELHAV method. The Vt perturbation is treated as in the
ELHAV theory, so the R-SRS+ELHAV+ZI expansion may converge despite the
presence of the Pauli-forbidden continuum. Simultaneously, the long-range
perturbation Vp is treated such that the correct asymptotics of the interac-
tion energy is ensured. The same is true for the R-SRS+SAM+ZI approach. It
should also be emphasized that the ZI procedure makes sense only when the
electron-nucleus attraction is regularized. Otherwise, the singular part of ΩX
would generate unphysical electron transfer between monomers (polarization
catastrophe [96]) and the infinite-order induction energy  E0 – E0 would not
vanish at large R.

4
Numerical Studies of Convergence Behaviour

4.1
H· · ·H interaction

The regularized approach was first tested on a simple example of two in-
teracting hydrogen atoms [70]. Such a system obviously does not possess
any Pauli-forbidden states. However, even for H2 serious pathologies in the
SAPT convergence were observed [45], and these pathologies were success-
fully eliminated by the regularization technique.
The computations reported in [70] were carried out for the lowest singlet and
triplet states of the H· · · H system at the interatomic distance of 8.0 bohr (cor-
responding to the minimum of the van der Waals well in the triplet state) and
employed the basis set formed by 180 explicitly correlated Gaussian geminals

exp(– α1 |r1 – RA |2 – α2 |r2 – RA |2 – β1 |r1 – RB |2 – β2 |r 2 – RB |2 – γ |r 1 – r2 |2 )

(84)

with the nonlinear parameters α1 , α2 , β1 , β2 , and γ optimized variation-

ally for the total energy of the hydrogen molecule. This basis was supple-
mented by two functions representing the orbital products 1sA (r 1 )1sB (r 2 ) and
1sB (r 1 )1sA (r 2 ) with the hydrogenic 1s orbital expanded in terms of 60 prim-
itive Gaussian orbitals with even-tempered exponents, so that the hydrogen
atom energy in this basis differed from – 0.5 by only 2 × 10–14 . The perturba-
tion corrections were computed by expanding the perturbed functions in the
basis that diagonalizes the zero-order Hamiltonian H0 (or Hp in case of the
74 K. Szalewicz et al.

R-ELHAV method) and using an appropriate spectral representation for the

reduced resolvent R0 (Rp ).
The results of [70] confirm that the conventional, non-regularized polar-
ization series converges to the ground-state interaction energy 1 E [46]; how-
ever, after approaching quickly the Coulomb energy Q, which differs from
1 E by 31.9841%, the convergence becomes pathologically slow, and the ex-

change part of the interaction energy is not reproduced to any reasonable

extent in a finite-order treatment. These results, including the convergence ra-
dius ρ equal to 1.0000000031, are in perfect agreement with those obtained
earlier [46] using the explicitly correlated basis of Kołos-Wolniewicz [97].
The conventional SRS series, as for the polarization one, also converges
quickly in low orders; however, after reproducing the value of the interaction
energy to better than 0.01%, the convergence deteriorates dramatically. This
fact is understandable since the RS and SRS expansions possess the same con-
vergence radius. In case of the singlet state, the SRS series converges to the exact
interaction energy 1 E. For the triplet state this is not possible since the RS expan-
sion for the wave function, from which the SRS energy corrections are calculated
(Eq. 12), converges to the fully symmetric singlet function which is annihilated
by the antisymmetrizer. As a result, the infinite-order SRS treatment for the low-
est triplet state of H2 yields only the so-called apparent interaction energy which
differs from the exact value of the triplet energy 3 E by 0.012% [70].
The study of [70] shows also that the regularization of the Coulomb po-
tential removes all the pathologies in the convergence behavior of the RS and
SRS theories. The stronger the regularization (the smaller the value of the pa-
rameter η), the faster the R-RS expansion approaches its limit. For any finite
value of η, the R-RS series converges smoothly (unlike in the case of LiH, as
we will see in the next subsection). Comparison of the results obtained with
the one-electron regularization and with the full regularization [70] demon-
strates that it is the one-electron, attractive part of the perturbation that
is responsible for the convergence problems in SAPT. Regularization of the
electron-electron repulsion does not change the results significantly. In fact,
the calculated values of the convergence radius ρ are the same for both regu-
larization algorithms [70].
Obviously, the limit Ep of the regularized polarization expansion depends
on the value of the regularization parameter η. Fortunately, this dependence
is rather weak and for a wide range of η the value of Ep is very close to the
Coulomb energy Q. Similarly, the limit ER-SRS of the regularized SRS expan-
sion exhibits only weak dependence on the value of η, both for the singlet and
triplet state (cf. Table 4 of [70]). Even for quite a strong regularization corres-
ponding to η = 5, one can recover the exact interaction energy to better than
one percent from the R-RS wave function that does not include the effects of
Vt . Moreover, the value of ER-SRS (η) can be successfully approximated by a fi-
nite sum of the R-SRS energy corrections: the regularized SRS series, unlike
the non-regularized one, converges quickly and smoothly.
Intermolecular Interactions 75

When one knows the function ψp , the small part of the interaction energy
that is missing in the infinite-order R-SRS energy ER-SRS can be easily recov-
ered by means of the R-ELHAV (or R2-ELHAV) expansion of Sect. 3.3. This
expansion converges really fast and, unlike the non-regularized ELHAV se-
ries suffering from the wrong asymptotic behavior, gives very accurate results
already in a low-order treatment. The rapid high-order convergence of the R-
ELHAV expansion results from large values of the convergence radius ρ. In
the whole range of η studied in [70], ρ is greater than two, and it increases
rapidly when η increases, i.e. when a larger part of the interaction is already
included in ψp . Switching from the one-electron regularization to the full one
does not significantly affect the convergence radii of the R-ELHAV series.
The low-order R-ELHAV and R2-ELHAV energies are similar, although the
R-ELHAV results are consistently somewhat more accurate [70]. The success
of a low-order R-ELHAV approach is, however, a little paradoxical since the
expression for the first-order R-ELHAV energy


E(1)
R-ELHAV = N ψ
p p t|V Aψ p (85)

involves a one-electron operator only and is completely different from the

well-established Heitler-London formula for the leading contribution to the
exchange energy [19]. On the other hand, the good accuracy of low-order R2-
ELHAV results is well understood since E(1)
R2-ELHAV = ER-SRS , and higher R2-
ELHAV corrections provide an improvement to the already accurate infinite-
order R-SRS energy.

4.2
Li· · ·H Interaction

Interacting lithium and hydrogen atoms are the simplest system for which the
convergence of the polarization or SRS expansions is destroyed by the Pauli-
forbidden continuum in which the physical ground state is submerged. The
numerical studies for this system were performed in [57, 69, 71], and [72].
These studies have shown that the regularization of the Coulomb potential
leads to expansions that are both convergent and asymptotically correct to
any order, not only for the distances around the van der Waals minimum for
the triplet state, but also in the region of the chemical minimum for molecu-
lar, singlet LiH.

Conventional SAPT Expansions

To make a high-order perturbation treatment computationally feasible, all

the numerical calculations of [57, 69], and [72] have been carried out using
a rather moderate basis set of 32 Gaussian orbitals. The orbital exponents
have been taken from [98] (Li) and [48] (H) and augmented by a set of dif-
76 K. Szalewicz et al.

fuse functions optimized for the dispersion interaction to obtain a realistic

description of the interaction energy in the van der Waals minimum region.
To provide reference values for the interaction energies obtained with
SAPT, the full configuration interaction (FCI) calculations for the lowest
singlet ([λ] = [22]) and triplet ([λ] = [211]) states of LiH, as well as for the
unphysical resonance state ([λ] = [31]) asymptotically degenerate with the
former two, were performed [57] for 10 ≤ R ≤ 20 bohr. The singlet and res-
onance potential curves are negative for this range of R (the singlet state
exhibits a chemical minimum at R = 3.015 bohr) whereas for the triplet state
there is a shallow van der Waals minimum at R = 11.5 bohr, and the curve
passes through zero at about 10.3 bohr. Interestingly enough, it has been
found that the position of the resonance state is approximated extremely well
[22]
by a weighted average of the physical energies, [31] E ≈ 23 E + 13 [211] E, cf. the
last two columns of Table I in [57]. The theoretical basis of this approximate
equality is not clear at the moment.
The results of [57] show that, as predicted by Adams [52–55], the RS,
SRS, and HS expansions diverge. In low order, however, the SRS results are
quite accurate, although the accuracy of a second-order treatment is some-
what worse than that obtained for the interactions of typical closed-shell
systems [20]. In fact, the conventional SRS theory is capable of providing re-
ally accurate results only when one goes to somewhat higher orders. As shown
in Fig. 4 of [57], the 20th-order SRS treatment for the triplet LiH is in perfect
agreement with the FCI values for the whole range of distances considered. In
even higher orders, the divergence starts to show up, and the 30th-order SRS
results are significantly less accurate for small R. The best way to obtain really
accurate results using the SRS approach is to sum the corrections E(n) SRS until
they start to grow in absolute value. This is the standard method of obtaining
a sum for a series that converges asymptotically.
Reference [57] also reported results obtained with the 1s2 core of Li frozen.
Freezing the core of the lithium atom has a very little effect on low-order
perturbation corrections. In high orders, however, a significant difference is
observed between the frozen core results obtained with and without the in-
clusion of the configuration state functions of 1s3 occupancy. When these
functions are included, the high-order behavior of the SRS series mimics very
well that observed without freezing the core. When the 1s3 functions are re-
moved from the basis set, the high-order perturbation corrections change
their behavior and the SRS series appears to converge, although extremely
slowly. As in the case of two hydrogen atoms (Sect. 4.1), the limit of the SRS
series for the triplet state is very close but slightly different from the super-
molecular interaction energy. This situation could be expected because the
frozen-core RS series for the wave function converges to a state of different
permutational symmetry so the SRS energy expression (Eq. 59 in the limit
η → ∞) takes the form 0/0 at ζ = 1. The results of [57] clearly show that it is
Intermolecular Interactions 77

the bosonic 1s3 state of the lithium atom and the associated Pauli-forbidden
continuum of the perturbed Hamiltonian that are responsible for the diver-
gence of the SRS perturbation series for LiH. Removing this continuum makes
the SRS expansion convergent; however, it does not improve the rather mod-
erate low-order convergence rate of the perturbation series.
The formal similarity of the SRS and MSMA theories (Sect. 2.2) leads to
almost the same convergence behavior of these two methods [69]. As in the
case of the SRS expansion, the divergence of the fully correlated theory turns
into a moderately fast convergence when the frozen-core approximation with
the neglect of 1s3 configurations is applied. The close similarity of the SRS
and MSMA results is further evidenced by the convergence radii of these se-
ries [69]. Thus, the MSMA theory does not provide any improvement over
the simple SRS approach. As a matter of fact, the same is true for the much
more complicated HS theory [57] – the SRS and HS results are practically
the same in low orders, and in higher orders the HS expansion diverges even
more rapidly than SRS. Freezing the core of the lithium atom has a similar
effect on the SRS and HS corrections; however, when the 1s3 configurations
are absent in the basis set, the HS series, unlike SRS, converges to the ex-
act supermolecular interaction energy for the triplet state [57]. Divergence of
the HS theory, predicted by Adams [55, 68], may be viewed as disappointing
since this method was exhibiting the best performance among all investigated
methods for systems involving one- and two-electron monomers [47–49].
However, even if the HS theory were convergent, its complex multistate struc-
ture would make it extremely difficult to use for larger systems.
The results of [69] show that the conventional, non-regularized ELHAV,
JK, and JK-1 expansions converge despite the presence of a Pauli forbidden
continuum. However, the wrong asymptotics of the second-order ELHAV en-
ergy spoils dramatically its low-order convergence rate, especially for larger
intermonomer distances. The JK expansion, exhibiting correct asymptotics in
second order, is far better. One may raise a question whether improving further
the asymptotic properties while maintaining the strong symmetry forcing, i.e.
going from JK to JK-1, results in further improvement of the low-order SAPT
convergence rate. The answer is obviously yes for very large intermonomer dis-
tances where the potential curve is perfectly described by a truncated Cn /Rn
expansion. However, even for R as large as 20 bohr, the difference between the
low-order JK and JK-1 energies is not really significant, and a simple SRS the-
ory is almost identical to JK-1 in low orders [69]. For smaller distances, the JK-1
results are not much better than the JK ones. Surprisingly, sometimes they are
even slightly worse, as it is the case for the triplet LiH at R = 11.5 bohr. This
means that there is no point in trying to introduce further improvement to JK by
improving the function used to initiate the iterative process defined by Eqs. 10
and 11, as, for example, in the JK-2 approach outlined in [69].
The results of [69] also show that the ELHAV and JK expansions converge
uniformly better for the triplet than for the singlet state. This observation is
78 K. Szalewicz et al.

further supported by the convergence radii of ELHAV/JK/JK-1 series given

in Table II of this reference. These convergence radii were obtained by ex-
trapolating the d’Alembert ratios or by fitting the large-order energies to the
formula [99]
J0 ((n – 32 )θ) – J0 ((n + 12 )θ)
E(n) ≈ Cρ –n , (86)
2n – 1
where θ is an argument of the branch points ρe±iθ (θ  1) determining the
convergence radius, J0 is the standard Bessel function, and C is some con-
stant. It has been found that in the considered range of distances R, the
convergence radii for the ELHAV, JK, and JK-1 methods were identical to the
number of digits computed. At R = 11.5 bohr (i.e. at the bottom of the van der
Waals well), ρ is equal to 1.214 for the singlet and 1.600 for the triplet state.
In fact, these radii are not exactly identical but the differences between them
are visible only at smaller interatomic separations [69]. It may also be noted
here that in the case of these three convergent expansions, both algorithms for
freezing the core were found to give results very similar to the fully correlated
calculations.
The significant difference between the convergence rate of ELHAV/JK/JK-
1 approaches for the singlet and triplet state of LiH can be understood by
looking at the behavior for R → ∞. One can show, using somewhat heuristic
arguments, that at this limit the convergence radius can be expressed as
N0
ρ= , (87)
N0 – 1

–1
where N0 = φ0 |Aφ0 . The limit of N0 for R → ∞ is determined by the
weight of the unit operator in the symmetry projector A. This weight is equal
to 1/4 for the singlet state and 3/8 for the triplet state, cf. Eq. (36) of [57],
which gives ρ = 1.333 for the singlet and 1.600 for the triplet state. It is not
clear why the triplet limit is reached already for the smallest distance con-
sidered, whereas even for R = 20 bohr the singlet convergence radius differs
significantly from its asymptotic value. Equation 87 appears to hold also for
the H· · · H interaction and the AM theory, as recently found by Adams [100],
as well as for the HeH system.2
Some information about the convergence of the non-regularized AM and
SAM expansions (Sect. 2.3) has been given in [72] for the case of the triplet
state at the van der Waals minimum (R = 11.5 bohr). The results reported
in this reference were obtained with the offset parameter D equal to zero,
however, as shown earlier in Table 2, different choices of D lead to very
similar results. The low-order convergence of the pure AM theory appears
to be the worst of all the SAPT methods considered. Significant improve-
ment is achieved by the symmetrization of the energy expression, and the
2 Przybytek M et al., to be published
Intermolecular Interactions 79

Fig. 2 Convergence of the non-regularized SRS, SAM, ELHAV, JK, and JK-1 expansions for
the singlet state of LiH at R = 3.015 bohr. The numbers displayed are percentages of the
FCI interaction energy recovered in the nth order perturbation treatment

SAM energies are much more accurate, although still not superior (and in
fact, almost equal) to those resulting from the ELHAV expansion, let alone
the JK one. Analogous results were obtained in the case of the singlet state.
The same close similarity between the SAM and ELHAV energies, which
could have been expected on the grounds of the formal similarity of these
methods (cf. Sect. 2.3), was obtained for the interaction of a triplet helium
atom with a hydrogen atom [87]. This similarity disappears when one goes to
shorter intermonomer distances, as illustrated in Fig. 2 by the results of non-
regularized ELHAV and SAM expansions for the singlet state at the chemical
minimum distance of 3.015 bohr. In this region the superiority of ELHAV is
clear; in fact, the SAM expansion was found to diverge in this case. The EL-
HAV, JK, and JK-1 series remain convergent for such a short intermonomer
distance, and in low orders JK and JK-1 give slightly more accurate results
than ELHAV.

Regularized SAPT Expansions

The lowest eigenvalue Ep of the regularized Hamiltonian H0 + Vp for the hy-

drogen dimer was found to be very close to the Coulomb energy Q defined
as the weighted average of the energies of all asymptotically degenerate states
(including the Pauli forbidden ones) with weights proportional to the ex-
80 K. Szalewicz et al.

change degeneracy of these states. Comparison of Ep (η) and Q for LiH at a few
intermonomer distances is presented in Fig. 3. An analogous comparison of
the infinite-order R-SRS energy ER-SRS (η) and the FCI interaction energy for
the triplet state of LiH has been made in Fig. 3 of [72]. These figures show
that, unlike for the hydrogen dimer, the regularization must be sufficiently
strong to make Ep resemble Q. There exists a critical value ηc (or ωc for the
SNC regularization) at which the curve Ep (η), and the corresponding eigen-
function ψp (η), undergo a dramatic change of character. Above ηc , the value
of Ep (η) exhibits a steep fall towards the energy of the mathematical ground
state of H in the space HA ⊗ HB , belonging to the Pauli-forbidden [31] sym-
metry. The wave function ψp (η), resembling φ0 for η < ηc , for η > ηc changes
its character and starts to resemble the mathematical ground-state function.
This result shows that to have a chance of obtaining convergent regularized
SAPT expansions one must regularize the potential strongly enough, i.e. one
must set η < ηc . It is interesting to note that the critical value ηc , equal to 5.43,
is independent of R for the range of distances considered. The R → ∞ limit of
ηc can be calculated from monomer properties [87]. This limit amounts also
to 5.43 in the basis set used for the calculations in [72]. Figure 3 also contains
the curve Ep (ω) obtained using the SNC regularization at R = 11.5 bohr. The
critical value ωc , equal to about 1.6, is significantly smaller than ηc . The rea-

Fig. 3 Dependence of the infinite-order R-RS energy Ep on the value of the regulariza-
tion parameter η (ω) for a few intermonomer distances R. The curve marked “SNC” has
been obtained using the SNC regularization, for all the other curves the Gaussian regu-
larization was used. The numbers displayed are percentages of the FCI Coulomb energy
Q calculated for the same distance R
Intermolecular Interactions 81

son for ωc  ηc is that if η = ω, the SNC regularization is significantly milder

at short distances from the nucleus, cf. Fig. 1 of [70], so the Gaussian regular-
ization leads to a much smaller volume of the regularized Coulomb well. The
results of Fig. 3, and of Fig. 3 in [72], show that it is not really the removal of
the negative singularities from V that is crucial for the success of the regular-
ized SAPT. It is rather the breaking of the permutational symmetry of the full
Hamiltonian H0 + V and the weakening of the Coulomb attraction between
electrons of one monomer and the nucleus of the other one that is really rel-
evant. Therefore, the procedure of [70–72] and [87] could also be referred to
as a “short-range attenuation” rather than a “regularization” of the Coulomb
potential.
Even when the regularization parameter is smaller than its critical value,
the results of Fig. 3 show that the infinite-order energies Ep (η) and ER-SRS (η)
can still be quite different from the reference FCI values. Only for larger R
do both Ep and ER-SRS become very accurate. This result might be expected
since both these quantities exhibit the same asymptotic behavior as the exact
interaction energy. However, as demonstrated in [72], the description of ex-
change interactions by the infinite-order R-SRS energy also improves when R
increases. As a measure of the accuracy of the exchange energy (in the triplet
state), one can choose the parameter
 
 [211] E [211] E 
 R-SRS – FCI 
∆ = 100% ×  [211] , (88)
 EFCI – [22] EFCI 

i.e. the percentage error of ER-SRS relative to the exponentially vanishing

singlet-triplet splitting. The value of ∆ is found to decrease with increasing
R, however, this decrease is slow and it is not clear from the results obtained
thus far (also for the interaction of a triplet helium atom with a hydrogen
atom [87]) whether ∆ goes to zero for R → ∞. It is worth noting that the SNC
regularization, despite having a smaller critical value of the regularization pa-
rameter than the Gaussian one, performs similarly to the latter as far as the
accuracy of Ep (ER-SRS ) for ω < ωc is concerned. This fact is demonstrated by
the results in Fig. 3 for R = 11.5 bohr.
Convergence of the regularized RS and SRS expansions is presented in
Tables 3 and 4. As expected, the condition η < ηc is necessary for the conver-
gence of the R-SRS series. As η increases beyond its critical value, the R-SRS
expansion diverges faster and faster, and in the limit η → ∞ it becomes iden-
tical with the non-regularized SRS theory (having the convergence radius of
0.7167 at R = 11.5 bohr [69]). As η decreases below ηc , the SRS series starts
to converge smoothly to its limit ER-SRS , and for a wide range of values of the
regularization parameter the R-RS convergence radius is significantly greater
than unity. However, the condition η < ηc does not guarantee the convergence
of perturbation series. If η is too small, the R-SRS series starts to diverge
in an oscillatory way. This happens when η < 0.7, or ω < 0.16 in case of the
82 K. Szalewicz et al.

Table 3 Convergence of the regularized polarization expansion for R = 11.5 bohr and both
regularizations (Gaussian and SNC). The numbers listed are percent errors of the sum of
the first n R-RS corrections with respect to Ep (η) (Ep (ω)). The values of Ep (and of the
Coulomb energy Q for η = ∞), given in microhartrees, are displayed in the row marked
Ep , whereas the last row lists convergence radii of the regularized polarization series

n η = 0.5 η=1 η=5 ω = 0.09 ω=1 ω = 2.25 η=∞

2 3.0127 –0.3668 –5.4387 –21.1765 –2.5056 –5.2368 –16.6158

3 –2.1651 –1.1629 –3.8041 13.4992 –1.8990 –3.6771 –13.8633
4 0.9617 –0.0591 –1.9792 –9.8130 –0.6486 –1.8921 –10.9174
5 –0.6032 –0.1114 –1.1939 7.5266 –0.3341 –1.1376 –8.9167
6 0.3537 –0.0032 –0.7081 –5.8669 –0.1505 –0.6710 –7.2920
7 –0.2163 –0.0144 –0.4295 4.6026 –0.0745 –0.4050 –5.9860
8 0.1311 0.0000 –0.2623 –3.6206 –0.0361 –0.2459 –4.9206
9 –0.0803 –0.0021 –0.1625 2.8522 –0.0182 –0.1512 –4.0489
10 0.0493 0.0000 –0.1022 –2.2492 –0.0093 –0.0942 –3.3330
15 –0.0064 0.0000 –0.0154 0.7042 –0.0005 –0.0127 –1.2249
20 0.0042 –0.0067 –0.2597 –0.0001 –0.0046 –0.2657
25 –0.0061 –0.0053 0.1672 0.0000 –0.0033 0.5214
30 0.0097 –0.0048 –0.2182 –0.0028 1.9714
40 0.0244 –0.0041 –0.7169 –0.0023 17.1733
50 0.0618 –0.0037 –2.5729 –0.0019 175.8250
60 0.1571 –0.0032 –9.2761 –0.0016
70 0.4005 –0.0028 –33.4818 –0.0013
Ep –34.4311 –39.1431 –44.3471 27.7881 –41.6989 –44.1406 –51.4609
ρ 0.910 1.270 1.013 0.879 1.164 1.019 0.717

SNC regularization. This divergence can be viewed as an artifact of the one-

electron regularization. It is probably caused by a mechanism similar to the
one that makes the Møller-Plesset perturbation theory divergent for many
systems – the so-called backdoor intruder state [101], i.e. the complex func-
tion E(z) has a singularity with a negative real part inside the unit circle.
The appearance of such a singularity has been confirmed by recent model
studies of Adams [100]; the singularity is likely to show up because, when
a strong one-electron regularization is applied, the weak nucleus-electron
repulsion present in the operator – Vp cannot compensate for the strong, non-
regularized electron-electron attraction from the two-electron part of – V, so
the electrons from one monomer fall into Coulomb wells around the other
monomer’s electrons. This divergence could be avoided if the two-electron
part of V were regularized as well. However, the range of values of the reg-
ularization parameter η (or ω) for which the R-SRS expansion converges is
sufficiently broad for the divergence for small η to bear no practical impor-
tance. For the SNC regularization, the convergence patterns are similar to
those for the Gaussian one; R-RS diverges for ω > ωc as well as for very small
ω. For the values of ω between the divergence regions, the regularized RS and
Intermolecular Interactions 83

Table 4 Convergence of the regularized SRS expansion for the lowest singlet and triplet
states of LiH, for R = 11.5 bohr and both regularizations. The numbers listed are percent
errors defined with respect to the sum ER–SRS of the R-SRS series. The last two rows dis-
play values of this sum (in microhartrees) and percent differences δ between ER–SRS and
the FCI interaction energy

singlet state triplet state

n η=1 η=4 ω=1 ω = 2.25 η=1 η=4 ω=1 ω = 2.25

2 –2.4559 –6.4502 –4.3457 –7.0680 2.6812 –0.2710 0.9801 –1.0023

3 –0.9980 –3.8285 –2.3196 –4.3374 –1.3357 –2.4001 –1.7090 –2.9252
4 –0.3313 –1.9832 –0.9734 –2.3729 0.2771 –0.6874 –0.1690 –1.0579
5 –0.0936 –1.1253 –0.4729 –1.4091 –0.0682 –0.4175 –0.1350 –0.6598
6 –0.0443 –0.6368 –0.2263 –0.8355 0.0561 –0.1848 –0.0194 –0.3373
7 –0.0122 –0.3657 –0.1119 –0.5008 –0.0003 –0.0873 –0.0022 –0.1791
8 –0.0069 –0.2111 –0.0558 –0.3005 0.0129 –0.0341 0.0080 –0.0867
9 –0.0022 –0.1226 –0.0285 –0.1803 0.0024 –0.0091 0.0084 –0.0368
10 –0.0013 –0.0715 –0.0148 –0.1076 0.0034 0.0027 0.0073 –0.0094
15 0.0000 –0.0041 –0.0007 –0.0028 0.0002 0.0085 0.0018 0.0182
20 0.0011 0.0000 0.0072 0.0000 0.0055 0.0005 0.0163
25 0.0014 0.0079 0.0040 0.0002 0.0143
30 0.0012 0.0075 0.0030 0.0001 0.0128
40 0.0008 0.0064 0.0019 0.0000 0.0105
50 0.0005 0.0054 0.0012 0.0087
60 0.0003 0.0045 0.0007 0.0072
70 0.0002 0.0038 0.0005 0.0059
sum –68.7700 –73.7824 –71.4893 –74.3672 –15.6509 –15.9859 –15.8418 –16.1331
δ –18.83 –12.92 –15.62 –12.23 –13.99 –12.15 –12.94 –11.34

SRS series converge, and the accuracy of the low-order corrections, as well
as the convergence radii, are similar to the ones obtained with the Gaussian
regularization.
The regularized SRS expansion, however quickly convergent, cannot pro-
vide a very accurate description of the potential energy curve for distances
around the van der Waals minimum. To obtain accurate results one must re-
sort to a theory that takes into account the singular operator Vt . As shown in
Table 5, this can be achieved using the R-ELHAV and R2-ELHAV approaches.
The R-JK theory gives similar results in this case since the better asymp-
totic behaviour of its non-regularized version does not matter here as vt (r)
is a short-range potential. Once ψp and Ep are known, the remaining part of
the interaction energy is recovered by the regularized ELHAV theory quite ef-
fectively, especially for the triplet state. For the singlet state the convergence
is somewhat slower; in fact, the regularized theory appears to have the same
convergence radius as the non-regularized one. As for H – H, the R-ELHAV
and R2-ELHAV results are of similar quality.
84 K. Szalewicz et al.

Table 5 Convergence of the R-ELHAV (columns marked “R”) and R2-ELHAV (columns
marked “R2”) perturbation expansions (in powers of Vt ). The numbers displayed are per-
cent errors of the sum of the first n corrections with respect to the FCI interaction energy.
Results for R = 11.5 bohr and the Gaussian regularization are shown

singlet state triplet state

η=1 η=4 η=1 η=4
n R R2 R R2 R R2 R R2

1 –17.7309 –18.8325 –8.9794 –12.9165 –13.4716 –13.9868 –12.6418 –12.1456

2 –14.4166 –15.2865 –6.4253 –9.4414 –5.6683 –5.7610 –4.4571 –4.3651
3 –11.7393 –12.4256 –4.6347 –6.8710 –2.4193 –2.4217 –1.6586 –1.6224
4 –9.5667 –10.1100 –3.3516 –4.9893 –1.0424 –1.0302 –0.6245 –0.6080
5 –7.8001 –8.2320 –2.4264 –3.6193 –0.4524 –0.4417 –0.2436 –0.2319
6 –6.3619 –6.7068 –1.7575 –2.6243 –0.1977 –0.1904 –0.0944 –0.0882
7 –5.1900 –5.4666 –1.2733 –1.9024 –0.0869 –0.0825 –0.0381 –0.0341
8 –4.2347 –4.4574 –0.9226 –1.3789 –0.0385 –0.0359 –0.0151 –0.0131
9 –3.4557 –3.6355 –0.6686 –0.9995 –0.0172 –0.0156 –0.0063 –0.0051
10 –2.8203 –2.9657 –0.4845 –0.7244 –0.0077 –0.0068 –0.0026 –0.0020
15 –1.0218 –1.0734 –0.0969 –0.1449 –0.0002 –0.0001 –0.0001 0.0000
20 –0.3703 –0.3889 –0.0194 –0.0290 0.0000 0.0000 0.0000
25 –0.1342 –0.1409 –0.0039 –0.0058
30 –0.0486 –0.0511 –0.0008 –0.0012

In view of the success of the regularized ELHAV method in recovering

the contribution to the interaction energy missing in Ep , one may hope
that the “all-in-one” R-SRS+ELHAV theory presented in Sect. 3.5 will con-
verge well for the LiH system despite the presence of the Pauli-forbidden
continuum. The similarity between the non-regularized ELHAV and SAM
theories suggests also that the corresponding R-SRS+SAM approach will
perform equally well, at least for larger intermonomer distances where the
ordinary AM expansion converges. The results of Table 6, as well as the
results of [72], show that both R-SRS+ELHAV and R-SRS+SAM indeed con-
verge rapidly for the van der Waals minimum of the triplet state. In low
orders, the R-SRS+ELHAV and R-SRS+SAM energies for the same η are prac-
tically identical, in higher orders the R-SRS+ELHAV approach is slightly
superior, which is indicated by the estimated convergence radii. For the re-
sults given in this table, the smaller the value of η, the faster the series
converges. However, as for the regularized SRS theory, if η is too small,
the R-SRS+ELHAV and R-SRS+SAM series start to diverge in an oscilla-
tory way. This fact is illustrated by the low-η R-SRS+SAM results displayed
in Fig. 4.
The success of the R-SRS+ELHAV theory in reproducing the potential
energy curve around the van der Waals minimum of triplet LiH is docu-
mented in some detail in [72] where the R-SRS+ELHAV results are compared
Intermolecular Interactions 85

Table 6 Convergence of the R-SRS+SAM and R-SRS+SAM+ZI expansions for the triplet
state of LiH, for R = 11.5 bohr and different values of η. The numbers displayed are per-
cent errors of the sum of the first n corrections with respect to the FCI interaction energy.
The row marked “ρ” lists estimated convergence radii of the perturbation series

R-SRS+SAM R-SRS+SAM+ZI
n η=1 η=2 η=4 η=1 η=2 η=4

2 –5.7950 –8.4520 –9.8868 –7.0785 –7.5480 –7.2844

3 –6.5376 –7.9306 –9.3955 –7.8929 –9.1141 –9.5884
4 –2.4007 –4.1470 –5.7491 –3.6992 –4.1591 –4.2570
5 –1.5688 –2.5528 –3.9076 –2.4317 –2.8883 –3.1613
6 –0.6520 –1.4407 –2.5471 –1.3379 –1.5658 –1.7093
7 –0.3755 –0.8270 –1.6662 –0.8018 –0.9724 –1.1220
8 –0.1608 –0.4645 –1.0822 –0.4563 –0.5539 –0.6506
9 –0.0872 –0.2614 –0.7025 –0.2668 –0.3326 –0.4105
10 –0.0383 –0.1464 –0.4559 –0.1538 –0.1941 –0.2474
15 –0.0012 –0.0090 –0.0556 –0.0107 –0.0160 –0.0265
20 –0.0001 –0.0012 –0.0094 –0.0010 –0.0022 –0.0053
25 –0.0001 –0.0005 –0.0032 –0.0002 –0.0007 –0.0022
30 0.0000 –0.0003 –0.0019 –0.0001 –0.0003 –0.0013
40 –0.0001 –0.0011 0.0000 –0.0001 –0.0006
50 0.0000 –0.0007 0.0000 –0.0003
60 –0.0004 –0.0002
70 –0.0003 –0.0001
ρ 1.16 1.12 1.05 1.20 1.13 1.07

with the non-regularized SRS and ELHAV ones for distances ranging from
8 to 16 bohr. The R-SRS+ELHAV method performs similarly to ELHAV for
R = 8 bohr, and clearly better for larger distances where the non-regularized
theory starts to suffer from its incorrect asymptotics. The R-SRS+ELHAV en-
ergies are also far more accurate than the SRS ones.
One may ask if the zero-order induction technique of Adams [71] brings
about some improvement to the convergence rate of the R-SRS+ELHAV and
R-SRS+SAM theories. The results presented in [72] suggest that this is not
the case, at least for the triplet state around the van der Waals minimum. The
R-SRS+ELHAV+ZI (R-SRS+SAM+ZI) energies are sometimes even slightly
worse than the ones of the corresponding R-SRS+ELHAV (R-SRS+SAM) the-
ory, and the convergence radii remain practically unchanged. Thus, the ZI
trick does not improve the (already very good) description of the van der
Waals well of the triplet LiH.
Except for the non-regularized ELHAV method, all the methods consid-
ered here, even the simple SRS theory, provide quite an accurate potential en-
ergy curve for the triplet LiH already in the second order (see Fig. 7 in [72]).
The most accurate results are obtained from the R-SRS+ELHAV approach; the
R-SRS+ELHAV+ZI energies are slightly worse, but both these theories per-
86 K. Szalewicz et al.

Fig. 4 Percentage of the full CI interaction energy for the triplet state of LiH recovered
by the Gaussian-regularized R-SRS+SAM and R-SRS+SAM+ZI expansions through nth
order. The interatomic distance is 11.5 bohr, and the numbers displayed in the legend
show the corresponding values of η

form far better than SRS, and in the fourth order their accuracy is excellent
in the range of distances studied.
There does exist a case when the ZI technique gives a significant im-
provement – when η is very small. As illustrated by the results in Fig. 4, the
R-SRS+SAM+ZI expansions converge for small η and even in the limit η → 0,
whereas the ordinary R-SRS+SAM series oscillate and are divergent when η
is too small. Unfortunately, the accuracy of low-order low-η R-SRS+SAM+ZI
energies is really bad, and the advantage of the R-SRS+SAM+ZI expansion
over the R-SRS+SAM one (or of R-SRS+ELHAV+ZI over R-SRS+ELHAV) for
small η bears no practical importance.
Since the regularization of the Coulomb potential, together with an ap-
propriate symmetry forcing, provided a quickly convergent and fully asymp-
totically correct description of the van der Waals interaction for triplet LiH,
it is interesting to see how the regularized SAPT expansions perform in de-
scribing the chemical bond in the singlet LiH. This question was considered
by Adams [71], who found that the third-order R-SRS+SAM+ZI approach
provides a significant improvement over R-SRS+SAM in this region. Since at
small interatomic distances the non-regularized ELHAV expansion is known
to exhibit a much better convergence than the SAM expansion (the latter
even diverges at the chemical minimum of singlet LiH), one may expect
Intermolecular Interactions 87

that the R-SRS+ELHAV and R-SRS+ELHAV+ZI approaches will perform bet-

ter than the corresponding SAM-based methods for this system. The results
reported in [72] show that this is indeed the case: the R-SRS+ELHAV expan-
sion remains convergent whereas R-SRS+SAM diverges like its parent SAM
method. In fact, the R-SRS+ELHAV approach remains clearly superior to the
R-SRS+SAM method for all values of η. The ZI technique of Adams improves
the low-order energies of both R-SRS+ELHAV and R-SRS+SAM theories and
makes the R-SRS+SAM expansion convergent, at least for some range of η.
However, the low-order R-SRS+ELHAV+ZI corrections remain much more
accurate than the corresponding R-SRS+SAM+ZI ones [72].
It is very gratifying that the R-SRS+ELHAV approach, both with and with-
out the ZI extension, can successfully describe such different aspects of the
interaction phenomenon as the chemical bonding in the singlet state, and
the van der Waals attraction in the triplet state of LiH. One may ask, of
course, if this conclusion can be transferred to larger systems. We believe
that the LiH molecule, unlike H2 or the ground state of He2 , is a system
for which all essential complications plaguing the SAPT treatment of many-
electron systems, including the coupling with the Pauli-forbidden continuum,
are present. Therefore, we think that the main conclusions reached while in-
vestigating the model LiH system will remain valid for larger systems as well,
so that the newly developed regularized SAPT may be regarded as a univer-
sal theory that is able to accurately describe both chemical and van der Waals
interactions – the achievement of a goal set by Eisenschitz and London in the
first years of quantum chemistry [22].

5
Extension of the Theory to Many-electron Systems

5.1
Outline of Many-electron SAPT

Sections 2–4 presented the SAPT approach assuming that the Schrödinger
equation for the monomers can be solved exactly or nearly exactly. This is
not the case for many-electron systems and therefore an extension of SAPT to
these systems requires further theoretical developments. In two-body, many-
electron SAPT, one uses the following partitioning of the total Hamiltonian:
H = F + V + W, (89)
where V is the intermolecular interaction operator collecting all Coulomb re-
pulsion and attraction terms between all particles of monomer A and those
of monomer B (the same as used in Sects. 2–4), F = FA + FB is the sum of
the Fock operators for monomers A and B, and W = WA + WB is the in-
tramonomer correlation operator with WX = HX – FX , X = A or B. The WX
88 K. Szalewicz et al.

operators are the same as the Møller-Plesset (MP) fluctuation potentials used
in many-body perturbation theories (MBPT) of the electron correlation.
The simplest zero-order wave function in a perturbational approach based
on Eq. 89 is the product of monomer Hartree-Fock determinants
Φ0HF = ΦAHF ΦBHF . (90)
The function Φ0HF is an eigenfunction of the operator F. The simplest per-
turbation expansion that one may employ is the RS (polarization) method
(extended to the case of two perturbation operators) discussed in Sect. 2.
However, as explained in that section, the RS method is not adequate, ex-
cept for large intermonomer separations. The underlying reason is that the
wave functions in this approach do not completely fulfill the Pauli exclusion
principle, i.e. the wave functions are not fully antisymmetric with respect to
exchanges of electrons [the antisymmetry is satisfied for exchanges within
monomers but not between them]. As described in Sect. 2, the antisymme-
try requirement can be imposed by acting on the wave functions with the
N-electron antisymmetrization operator. This (anti)symmetrization can be
performed in many ways and leads to various versions of SAPT. The sim-
plest of them, the SRS method, has been implemented in the many-electron
context [24].
The SRS energy corrections are obtained by expanding the expression for
the interaction energy in powers of the perturbation operators
∞  ∞ ∞ 
 ∞  
Eint = E(ni)
SRS = E (ni)
pol + E(ni)
exch , (91)
n=1 i=0 n=1 i=0
where n denotes the order with respect to V and i denotes the order with
(ni)
respect to W. The last expression splits each ESRS component into the polar-
ization part obtained from the RS expansion and the remainder due to the
electron exchanges. General expressions for each correction E(ni) (ni)
pol and Eexch
were given in [80]. In order to solve the resulting equations for the correc-
tions to the wave function, one usually assumes a finite orbital basis (so-called
algebraic approximation), although other approaches are also possible [80].
Derivation of explicit forms of the SRS expressions in terms of two-electron
integrals and orbital energies is quite involved. The formulae for several of
those components have been given in [23, 25, 102–107].
Each polarization component appearing in Eq. 91 can be related to the
physical picture of intermolecular interactions resulting from the long-range
(1i)
expansion of the interaction energy. The first-order corrections Epol repre-
sent the electrostatic interaction of unperturbed charge distributions and are
(1i)
usually denoted by Eelst . In the second order in V, the contributions are
naturally separated into the induction and dispersion components. Also the
exchange energies can be related to the corresponding polarization terms. In
this way one can, for example, distinguish exchange-dispersion interactions.
Intermolecular Interactions 89

The currently implemented level of SAPT represents the four fundamental

interaction energy components by the following expansions:
(10) (12) (13)
Eelst = Eelst + Eelst,resp + Eelst,resp (92)
(20) (22)
Eind = Eind,resp + t Eind (93)
Edisp = E(20) (21) (22)
disp + Edisp + Edisp (94)
Eexch = E(10) (1) (20)
exch + exch (CCSD) + Eexch-ind,resp + t E(22) (20)
exch-ind + Eexch-disp (95)
The terms with the subscript “resp” are calculated using monomer orbitals
(1)
distorted in the field of the interacting partner. The quantity exch (CCSD) is
the intramonomer correlation contribution to the first-order exchange energy
calculated using monomers correlated at the coupled-cluster single and dou-
ble excitation (CCSD) level, and t E(22) t (22)
ind and Eexch-ind denote those portions of
the second-order intramonomer correlation correction to the induction and
exchange-induction energies, respectively, which are not included in E(20)
ind,resp
and E(20)
exch-ind,resp . In most applications one adds to the set of SAPT corrections
the term:
(10) (10) (20) (20)
δEHF HF
int,resp = Eint – Eelst – Eexch – Eind,resp – Eexch-ind,resp , (96)

where EHF int is the supermolecular Hartree-Fock interaction energy. This pro-
cedure involves an approximation since the relation between SAPT and the
supermolecular methods can be established rigorously only in the asymptotic
limit of large intermolecular separations.
Recently, the complete set of E(30) corrections has been developed [108].
This work allowed us in particular to understand better the role of the term
δEHF
int,resp . It has been found that this term should be used only for interactions
of polar and polarizable systems, where the induction effects are very large.
For systems not fulfilling this criterion, like rare gas atoms or H2 – CO, one
should not use δEHF int,resp .
For large intermonomer separations, the exchange and overlap effects van-
ish and SAPT interaction energies become equal to those obtained from the
multipole expansion [36]. This expansion, often multiplied by some damping
factors [109], is applied as the analytic form of functions used to fit the in-
teraction energies computed by SAPT. For small monomers, one can use the
expansion in terms of spherical multipole moments, polarizabilities, and dy-
namic polarizabilities located at the centers of masses of the monomers [36,
110–114]. The coefficients in this expansion are calculated ab initio using
methods and programs developed by Wormer and Hettema [115, 116]. For
larger systems and for the purpose of producing inexpensive potentials for
molecular simulations, the center-of-mass expansion is replaced by site-site
multipole expansions. The parameters of these expansions can be fitted to the
interaction energies of the center-of-mass expansion, but an alternative and
90 K. Szalewicz et al.

a much better approach is to use distributed multipoles [117–121] and po-

larizabilities [121–126] for the electrostatic and induction energies. Recently,
effective methods have been developed for distributing the dynamic polariz-
abilities, leading to distributed asymptotic dispersion energies [127]. Earlier
distributed dispersion approaches [128, 129] are less practical.

5.2
SAPT Based on DFT Description of Monomers

Williams and Chabalowski (WC) [130] have proposed a perturbational ap-

proach where the interaction energies are obtained using only the lowest-
order, computationally least demanding SAPT expressions, but replacing the
Hartree-Fock (HF) self-consistent field (SCF) orbitals and orbital energies by
their DFT counterparts, called Kohn-Sham (KS) orbitals and orbital energies.
The corrections included were
(10) (10) (20) (20) (20) (20)
Eint ≈ Eelst + Eexch + Eind + Eexch-ind + Edisp + Eexch-disp . (97)

Such an approach is significantly less time consuming than the regular SAPT
with high-order treatment of electron correlation. However, the accuracy of
the predictions was found to be disappointing [130] even for the electro-
static energy, which is potentially exact in this approach. It has been demon-
strated [131] that some deficiencies of the original method stem from an
incorrect asymptotic behavior of exchange-correlation potentials. Upon ap-
plying an asymptotic correction in monomer DFT calculations, the revised
approach was not only able to accurately recover the electrostatic energy, but
also the exchange and induction energies. The dispersion energies in the ori-
ginal method of WC were computed from an expression that asymptotically
corresponds to the use of uncoupled KS dynamic polarizabilities. Such dis-
persion energies remained inaccurate even with the asymptotic correction.
Misquitta and two of the present authors have proposed [132] a method of
computing the dispersion energy that asymptotically corresponds to the use
of coupled KS (CKS) polarizabilities. The new method gives extremely ac-
curate dispersion energies, probably more accurate than those predicted by
the regular SAPT at the currently programmed level. We will refer to the
KS-based SAPT method with the dispersion energy computed in this way as
the SAPT(DFT) approach. This approach turned out to be much more ac-
curate than could have been expected. In fact, there are indications that at
least for some systems it may be more accurate than SAPT at the currently
programmed level [131–134]. A similar approach has been independently de-
veloped by Hesselmann and Jansen [135–137].
In the SAPT(DFT) method, the partitioning of Eq. 89 is replaced by

H = K + W KS + V (98)
Intermolecular Interactions 91

where K = KA + KB is the sum of the KS operators (counterparts of FX ), and

W KS may be formally defined as W KS = HA + HB – K. In SAPT, several powers
of the W operator, as seen in Eqs. 92–95, have to be included in the expansion
to achieve a reasonably accurate description of intermolecular interactions.
Since SAPT(DFT) includes only the components listed in Eq. 97 [with the
induction and dispersion energies in the CKS versions], SAPT(DFT) effec-
tively neglects the operator W KS and uses the Hamiltonian H KS = K + V. Only
in the CKS dispersion and induction energies are the effects of W KS partly
included (those corresponding to response effects in Hartree-Fock based ap-
proaches). The first hint that this approach can work is the fact that if KS
orbitals are used in the expression for the electrostatic energy of zero order
(10)
in W, Eelst , one gets the exact value of the fully correlated electrostatic en-
(1)
ergy, Eelst , provided that DFT gives the exact electron density for a given
system. The DFT-based SAPT method in the uncoupled KS version is equally
time-consuming as regular SAPT in zero order with respect to W [this level
of SAPT will be denoted by SAPT(0)]. The coupled version is somewhat
more time-consuming, but still the requirements are much closer to those of
SAPT(0) than to those of SAPT at the level of Eqs. 92–95. Thus, the major
criterion for evaluation of the effectiveness of SAPT(DFT) is the increase in
accuracy of predictions compared with SAPT(0).
Although the original formulation of the SAPT(DFT) method [130] pro-
duced interaction energies which were not accurate enough to be competitive
with SAPT(0), further modifications of this method not only reversed this re-
lation but made SAPT(DFT) results competitive with the regular SAPT results
obtained with high-level treatment of the electron correlation. As mentioned
above, the first reason for the initial poor performance was the wrong asymp-
totic behavior of exchange-correlation potentials in KS equations and this
problem can be fixed by applying the asymptotic correction [131]. Then
the electrostatic, induction, and exchange energies were reproduced with as-
tounding accuracy and only the (uncoupled KS) dispersion energy was not
sufficiently accurate. We will now describe a solution to this problem. The
exact (all orders in W) second-order dispersion energy is given by


  | Φ0A Φ0B |VΦkA ΦlB |2
(2)
Edisp = , (99)
EA B A
0 + E0 – Ek – El
B
k =0 l =0

where ΦiX and EXi are the exact eigenfunctions and eigenvalues of HX . This
energy can be represented by an alternative expression [77, 138–140], the so-
called generalized Casimir-Polder formula:
∞    
1 dr1 dr 2 dr 1 dr 2
E(2) =– αA (r 1 , r1 |iu)αB (r 2 , r2 |iu) du ,
disp 2π |r 1 – r2 | |r 1 – r2 |
0
(100)
92 K. Szalewicz et al.

where
 EXm – EX0


αX (r, r |ω) = 2 Φ X
|ρ̂(r)|Φ X
Φ X
|ρ̂(r 
)|Φ X
(101)
m =0
(EXm – EX0 )2 – ω2 0 m m 0

is the frequency-dependent density susceptibility (FDDS) of monomer X

computed at frequency  ω. The symbol ρ̂(r) stands here for the electronic
density operator ρ̂(r) = i δ(r – ri ), the summation extending over all elec-
trons of the considered molecule. For real ω, α(r, r |ω) describes the linear
change of electronic density at r under the influence of a one-electron per-
turbation localized at r  and oscillating with frequency ω. FDDSs are closely
related to the dynamic polarizabilities. It should be stressed that the disper-
sion energy expressions, Eqs. 99 and 100, account not only for the asymptotic
dipole-dipole (1/R6 ) term but also include the effects of all higher instanta-
neous multipoles as well as the short-range contributions resulting from the
overlap of monomer charge distributions [19, 140].
If the wave functions Φ0X in Eq. 101 are replaced by appropriate HF/KS de-
terminants, Φm X by singly-excited HF/KS determinants, and the differences
X X
Em – E0 by the corresponding HF/KS orbital excitation energies, one obtains
the uncoupled HF/KS (UCHF/UCKS) FDDSs and the corresponding expres-
sion for dispersion energy reduces to the expression for E(20)
disp . An obviously
better option is to use FDDSs computed using time-dependent DFT (TD-
DFT). This choice defines the CKS dispersion energy.

5.3
SAPT/SAPT(DFT) Computer Codes

The computer codes evaluating the SAPT expressions described in Sect. 5.1
have been developed over the last 20 or so years. The current version, named
SAPT2002 [24], is available for downloading from the web free of charge
(see http://www.physics.udel.edu/∼szalewic/SAPT/SAPT.html). The codes are
used by about 150 research groups around the world.
The SAPT codes have been parallelized in the period 1997–2004. This work
has generated an efficient and portable parallel implementation of a highly
correlated electronic structure code suitable for both shared- and distributed-
memory parallel architectures. The code features good parallel performance
even on Linux clusters with slow communication channels and distributed
scratch disk space. Other systems that the parallel SAPT was developed on
are IBM SP3/4 and SGI O3K. The tests performed on platforms with common
scratch disk space, shared between all processors, show that the competition
for I/O bandwidth is the main reason for performance deterioration of I/O-
bound codes on such platforms. The parallel implementation of the CCSD
method within the SAPT codes was, to our knowledge, the first one efficiently
running on more than 32 processors.
Intermolecular Interactions 93

The SAPT program scales well up to about 128 processors. The scaling de-
pends on the size of the problem and improves for larger problems. On SP4
and on Linux clusters, the speedups upon doubling the number of processors
are generally around 1.8 for the whole range of processors used, both for the
integral transformation part and the evaluation of many-body sums. For ex-
ample, the water dimer calculation in the basis of 430 functions speeds up on
SP4 1.7 times upon increasing the number of processors from 32 to 64.
The SAPT(DFT) method is algorithmically almost identical to the SAPT(0)
subset of SAPT. In fact, the same code is used for most parts of calcula-
tions. The only exceptions are the CKS-type terms whose CHF counterparts
in SAPT are computed from different types of formulae. If the uncoupled KS
dispersion energies are used, the calculation of the expressions for the cor-
rections listed in Eq. 97 requires a small fraction of the time needed for the
calculations of all higher-order corrections listed in Eqs. 92–95. For medium-
size (few-atomic) monomers and medium-size (triple-zeta) basis sets, the
former calculation is about 2 or 3 orders of magnitude faster than the lat-
ter one [130, 133]. If the time spent in SCF/DFT calculations and in the
transformation is included, the speedup is reduced to about 1 order of mag-
nitude [133]. Of course, the speedup increases with the size of the system
and/or basis set. Thus, the DFT-based approach with the uncoupled KS dis-
persion energy allows calculations for much larger systems than is possible
with the regular SAPT (a further extension of system sizes is possible due
to the lesser basis set requirements of SAPT(DFT) [133]). Calculations of the
CKS dispersion energies make the timing issues more involved. The expres-
sion for this energy scales as the sixth power of the system size, but has
a very small prefactor so that this scaling is not consequential for the range
of systems of current practical interest. However, somewhat paradoxically,
the limiting factor becomes the TD-DFT calculation for monomers. In fact,
this calculation wipes out a large part of the speedup quoted above, although
for larger systems the better scaling of SAPT(DFT) compared with regular
SAPT does result in very large performance gains. Significant work has re-
cently been performed on the optimization of codes and the development
of new algorithms for SAPT(DFT) [141]. After only some rudimentary opti-
mizations, the SAPT(DFT) codes became about 1 order of magnitude faster
than the initial version and calculations could be performed using mod-
est (workstation-level) computer resources for monomers containing about
a dozen atoms in basis sets with a couple of hundred orbitals [142, 143].
Regular-SAPT calculations for such systems are very costly. Using signifi-
cant computer resources, SAPT(DFT) codes could be used for calculations for
dimers containing about 20-atom monomers, which could not be investigated
with regular SAPT. A major further speedup has recently been achieved by
using the density fitting methods [141]. This approach was particularly im-
portant for the TD-DFT and transformation stages of calculations. Similar
developments have also been achieved by Hesselmann et al. [144]. The new
94 K. Szalewicz et al.

codes make calculations for systems of the size of benzene dimer in medium
size basis sets virtually as inexpensive as the standard supermolecular DFT
calculations. Although the SAPT(DFT) method scales worse than DFT with
system size N: as N 5 vs. N 3 , both methods can be reduced to linear scaling for
very large systems.

6
Helium

6.1
Helium as a Thermodynamic Standard

Ab initio computed interaction potentials for the helium dimer turned out
to be of significant importance in thermal physics. Thermodynamical meas-
urements in industry and science are utilizing international standards of
temperature and pressure and the value of the Boltzmann constant. Improve-
ment of the accuracy of these standards is the major goal of the international
metrology community. These quantities are interrelated and can be measured
in several ways. If the measurements are performed in helium, it is now pos-
sible to utilize ab initio values of some quantities, such as the density and
dielectric virial coefficients, that previously had to be determined experimen-
tally.
The Boltzmann constant kB (or equivalently the gas constant, R = NA kB ,
where NA is the Avogadro constant known to 0.17 ppm) was measured in
argon by Moldover et al. in 1988 with an accuracy of 1.7 ppm [145]. With
improvements in technology, the accuracy can be increased to about 1 ppm
within a few years.3 There appears to be a consensus in the metrology com-
munity that after the uncertainty of kB is reduced to 1 ppm or better, this
value should be fixed and further work should concentrate on improving
the temperature standard. Once the value of kB is fixed, e.g. kB = 1.38065 ×
10–23 J/K, one kelvin can be defined as the change of temperature leading to
a change of kB T equal to 1.38065 × 10–23 J. Then, no particular way of measur-
ing the temperature would be favored. However, the required reduction of the
uncertainty in the value of kB is not a simple task. In principle, the Boltzmann
constant can be determined with any primary thermometer by measuring the
product kB T at the triple point of water (TPW), which is assumed by the cur-
rent standard, ITS-90, to be T = 273.16 K exactly. In practice, there will always
be an uncertainty resulting from the realization of TPW (in the current value
of kB , this uncertainty amounts to 0.9 ppm [145]).
The ITS-90 defines the temperature scale using a set of fixed points and in-
terpolates between these points with platinum thermometers. The uncertain-

3 Moldover MR (2005) Personal communication

Intermolecular Interactions 95

ties of the standard compared to thermodynamics measurements approach

70 ppm in some regions. In addition to the limits of accuracy, this standard
has the disadvantage of being generally inconsistent with thermodynamic
variables. Therefore, creation of a new temperature standard is one of the
high-priority goals. The conceptually simplest method can be based directly
on the equation of state
 
p = RTρ 1 + B(T)ρ + C(T)ρ 2 + ... , (102)
where p is pressure, B and C are the second and third virial coefficients, re-
spectively, and ρ is density. If p and ρ are measured and R, B(T), and C(T) are
known, this equation determines T. Since R is known to 1.7 ppm, volume and
mass measurements giving ρ can be of similar accuracy, and pressure can be
measured to a few ppm at low pressures, T could be determined to a better ac-
curacy than given by ITS-90 provided that the virials are known. Such virials
can be computed if the interaction potential between helium atoms is known.
An accurate potential for the helium dimer [146, 147] based on SAPT was de-
veloped in 1996. This potential has been widely used in thermal physics and
in other fields. The current most accurate values of B(T) are those from first-
principles calculations and have been computed by Janzen and Aziz [148]
using the SAPT96 potential for the helium dimer [146, 147] and by Hurly
and Moldover [149] using a modification of SAPT96. By comparing with pre-
dictions of other ab initio works, which gave the depth of the potential at
the minimum about 100 mK different from SAPT96, Hurly and Moldover es-
timated the uncertainty of theoretical B(T) at 300 K to be 2200 ppm [149].
Recently, more accurate calculations of the He2 interaction energies have been
performed for a few intermolecular separations using the supermolecular ap-
proach [150, 151]. These calculations are converged to better than 10 mK at
the minimum. If such newer ab initio calculations [150–152] are taken into
account (showing that the SAPT96 potential was within about 50 mK of the
current well depth at the minimum), the uncertainty in B(T) can be reduced
to about half of the value assumed by Hurly and Moldover [149]. The cur-
rent goal is to obtain B(T) accurate to about 100 ppm. This accuracy should be
more than sufficient for the future standards as B(T)ρ is about 10–3 at 300 K.
The contribution of C(T) comes with a still lower weight but will also be re-
quired for the next generation of measurements. The three-body potential
for helium needed to predict C(T) has been computed some time ago [153],
showing that many-body effects are very small for helium.
The simplest approach based on Eq. 102 is in practice replaced by meas-
urements involving the dielectric constant of a gas [154]. This leads to the
need for the dielectric virial coefficients, discussed below in the context of the
pressure standard.
Whereas with the best conventional piston gauges one can measure pres-
sure to a few ppm at low range, the accuracy of the current standard based on
such mechanical pistons can be questioned at high pressures. Moldover [155]
96 K. Szalewicz et al.

proposed that the techniques used in thermometry can be inverted to pro-

vide a new standard of pressure. This is possible by combining measurements
of the dielectric constant of helium with ab initio values of the virials and
polarizability of helium. The dielectric constant  can be expressed as [155]:
–1  
= A ρ 1 + b(T)ρ + c(T)ρ 2 + ... (103)
+2
where A is the molar polarizability of gas and b and c are respectively, the
second and third dielectric virial coefficients. This equation can be combined
with the equation of state to give
  2 
 – 1 NA  – 1  – 1
p= kB T 1 + B(T)∗ + C(T)∗ + ... (104)
 + 2 A +2 +2
where B(T)∗ and C(T)∗ are simple algebraic expressions in terms of the dens-
ity and dielectric virial coefficients and of A . The uncertainties of kB T near
TPW and of NA are below 1.7 ppm. It is expected that  – 1 can be measured
to 5 ppm. Recent theoretical work [156, 157] has determined A to within
0.2 ppm. Before this work was performed, the uncertainty of the theoretical
value of the molar polarizability amounted to 20 ppm, as estimated by Luther
et al. [154]. This uncertainty used to be the major limitation of all methods
involving dielectric measurements. Now, the standard with accuracy of about
1 ppm can be created if sufficiently accurate values of B(T)∗ and C(T)∗ can be
predicted by theory.
Accurate ab initio calculations for the helium dimer and trimer can be
used in several other ways in thermal physics. One can use theoretical predic-
tions to calibrate instruments designed to measure the density and dielectric
virial coefficients, viscosity, thermal conductivity, speed of sound, and other
properties of gases based on comparisons of theory with experiment for he-
lium.

6.2
Towards 0.01% Accuracy for the Dimer Potential

To determine the second virial coefficients to the required accuracy of about

100 ppm, one needs to know the helium dimer potential to a few mK at
the minimum. Consider first the interaction of two helium atoms in the
nonrelativistic Born-Oppenheimer (BO) approximation. The best published
calculations using the supermolecular method have estimated error bars of
8 mK [150, 151]. More recently, improved calculations of this type reached an
accuracy of 5 mK [158]. Also, the SAPT calculations have been repeated with
increased accuracy, resulting in an agreement to 5 mK with the supermolec-
ular method. Furthermore, the upper bound from four-electron explicitly
correlated calculations is 5 mK above the new supermolecular value [159]. All
this evidence from three different theoretical models seems to show convinc-
Intermolecular Interactions 97

ingly that the BO helium dimer interaction energy at the minimum is now
known to within a few mK.
At the accuracy level of a few mK near the minimum, effects beyond the
BO approximation have to be considered. The adiabatic effects have been
computed by Komasa et al. [160] and for 4 He2 amount to – 13.2 mK at the
minimum. Recently, these results were found to contain small errors and will
have to be recomputed. The next effects to be included are the leading rela-
tivistic terms (∼ α2 ). A preliminary value of the relativistic correction at the
minimum is + 15.6 ± 0.4 mK [159].
Once the relativistic correction to the helium dimer potential is known,
one should consider the QED corrections. A comparison of the relativistic
and QED contributions to the polarizability of the helium atom [156, 157]
shows that the latter is only slightly smaller in magnitude than the former.
Thus, the QED effects on the helium dimer interaction energies may be of the
order of a few mK, i.e. non-negligible. A part of the QED effect, the retarda-
tion correction, has been calculated for He2 by several authors for asymptotic
separations. In this region, as first shown by Casimir and Polder, the re-
tardation effects change the 1/R6 behavior of the potential into 1/R7 . This
asymptotic contribution has to be included in the He2 potentials in investi-
gations of some effects sensitive to the tails of the potential. For example, the
size of the He2 molecule changes by about 4% upon the inclusion of the retar-
dation effects [147]. The expressions for the QED corrections to energy of the
order α3 are well known; see for example [161]. Some of the terms in these
expressions are analogous to those appearing in the calculations of the rela-
tivistic corrections and have already been programmed [159]. The so-called
Araki-Sucher term [161] and the Bethe logarithm term are more complicated.
A complete calculation of the latter term for He2 is probably not feasible at
this time. However, a recent study for H2 has shown [162] that the R depen-
dence of the Bethe logarithm is very weak. Therefore, it should be sufficient
to use its well-known value for the helium atom. This is similar to the finding
for the polarizability of helium that the value of the QED correction computed
with no approximations [157] agrees to within 1% with the analogous correc-
tion estimated by Pachucki and Sapirstein [163] by neglecting the dependence
of the Bethe logarithm on the electric field. When the R dependence of the
Bethe logarithm is neglected, one finds [159] that the α3 QED correction to
the dimer well depth amounts to – 1.3 ± 0.1 mK – a value somewhat smaller
than suggested by the estimate discussed above.
The second dielectric virial coefficient b(T) is 2 orders of magnitude
smaller than the density virial coefficient but the existing absolute uncertain-
ties in the values of these coefficients are expected to be similar. Calculations
of b(T) were performed by Moszynski et al. using a fully quantum mechan-
ical approach [164, 165] and the collision induced polarizability from SAPT
calculations with a moderately large orbital basis set. However, more recent
calculations from [166] and [167] disagree with this work to such an extent
98 K. Szalewicz et al.

that the resulting estimated uncertainties are too large from the experimental
point of view. It should be noted, however, that the collision-induced Ra-
man spectra predicted theoretically from the SAPT polarizabilities of [164]
and [165] agree very well with the experimental data [168–170].

7
Some Recent Applications of Wave-function-based Methods

7.1
Argon Dimer

Argon gas is widely used in physics and chemistry and is also of importance
due to its presence in the atmosphere. In particular, it is used in developing
standards for thermal physics. The popular, empirical argon dimer poten-
tial of Aziz [171] has generally been assumed to be the best representation
of this system. However, this potential was not able to fit all data to within
experimental error bars [171, 172]. More importantly, neither this potential
nor any other potential for argon was able to predict even qualitatively prop-
erties dependent on the highly repulsive region of the potential wall [173].
To resolve these problems, a new ab initio potential has recently been de-
veloped for the argon dimer [174], extending earlier accurate calculations by
Slavíček et al. [175]. This potential was based on calculations using the CCSD
method supplemented with the noniterative triple excitations contribution
[CCSD(T)] in a sequence of very large basis sets, up to augmented sextuple-
zeta quality and containing bond functions, followed by extrapolations to the
complete basis set limit. The calculations included intermolecular distances
as small as 0.25 Å, where the interaction potential is of the order of 4 keV.
The computed points were fitted by an analytic expression. The new poten-
tial has the minimum at 3.767 Å with a depth of 99.27 cm–1 , very close to the
experimental values of 3.761 ± 0.003 Å and 99.2 ± 1.0 cm–1 [176]. The poten-
tial was used to compute spectra of the argon dimer and the virial coefficients.
The latter calculations suggest a possible revision of the established experi-
mental reference results. From the agreement achieved with experimental
values and from comparisons of the fit with available piece-wise information
on specific regions of the argon-argon interaction, one can assume that the
potential of [174] provides the best overall representation of the true argon-
argon potential to date, although Aziz’s potential [171] is probably slightly
more accurate near the minimum. To improve the accuracy of ab initio pre-
dictions, one has to include electron excitations beyond triples and relativistic
effects [174].
For the argon dimer, standard-level SAPT calculations overestimate the
van der Waals well depth by about 10%. The reason for this discrepancy has
been identified recently [108]: the term δEHF int,resp (see Sect. 5.1) significantly
Intermolecular Interactions 99

overestimates the 3rd- and higher-order induction and exchange-induction

effects for this system. For example, in the aug-cc-pVQZ basis set supple-
mented by a set of spdfg midbond functions, replacing δEHFint,resp by the sum
(30) (30)
Epol + Eexch [108] reduces the well depth predicted by SAPT by 7%, resulting
in a potential that agrees well with supermolecular CCSD(T) results.

7.2
He – HCl

He – HCl dimer has been of interest for a long time since its spectra may be used
to determine the abundance of HCl in planetary atmospheres or in cold inter-
stellar clouds. Recently, a two-dimensional intermolecular PES for the He – HCl
complex has been obtained [177] from ab initio calculations utilizing SAPT and
an spdfg basis set including midbond functions. HCl was kept rigid with a bond
length equal to the expectation value r in the ground vibrational state of iso-
lated HCl. In the region of the minimum, the He – HCl interaction energy was
found to be only weakly dependent on the HCl bond length, at least as com-
pared with the case of Ar – HF. This result can be attributed to the smaller dipole
moment of HCl relative to HF and the subsequently smaller induction energy
component in the case of HCl. The calculated points were fitted using an an-
alytic function with ab initio computed asymptotic coefficients. As expected,
the complex is loosely bound, with the dispersion energy providing the ma-
jority of the attraction. The SAPT PES agrees with the semi-empirical PES of
Willey et al. [178] in finding that, atypically for rare gas–hydrogen halide com-
plexes including the lighter halide atoms, the global minimum is on the Cl side
(with an intermonomer separation 3.35 Å and a depth of 32.8 cm–1 ), rather than
on the H side, where there is only a local minimum (3.85 Å, 30.8 cm– 1). The
ordering of the minima was confirmed by single-point calculations in larger
basis sets and complete basis set extrapolations, and also using higher levels
of theory. Reference [177] has shown that the opposite findings in the recent
calculations of Zhang and Shi [179] were due to the fact that the latter work
did not use midbond functions in the basis set. Despite the closeness in depth
of the two linear minima, the existence of a relatively high barrier between
them invalidates the assumption of isotropy, a feature of some literature po-
tentials. The accuracy of the SAPT PES was tested by performing calculations
of rovibrational levels. The transition frequencies obtained were found to be in
excellent agreement (to within 0.02 cm–1 ) with the experimental measurements
of Lovejoy and Nesbitt [180]. The SAPT PES predicts a dissociation energy for
the complex of 7.74 cm–1 , which is probably more accurate than the experimen-
tal value of 10.1 ± 1.2 cm–1 . The characterizations of three low-lying resonance
states through scattering calculations can also be expected to be more accurate
than the experimentally derived predictions. The analysis of the ground-state
rovibrational wave function shows that the He – HCl configuration is favored
100 K. Szalewicz et al.

over the He–ClH configuration despite the ordering of the minima. This is due
to the greater volume of the well in the former case. The potential has been used
to calculate the spectra of the HCl dimer in helium [181]. The HCl dimer is analo-
gous to the much investigated HF dimer [182], but much more floppy, so that the
tunneling splittings in the former case are about an order of magnitude larger.
The ordering of the minima in He – HCl raises the question of general
trends in shapes of potential energy surfaces in the Rg – HX family, where
Rg denotes a rare-gas atom and X is a halogen atom. Based on the results
for He – HCl, single-point SAPT calculations for some systems, and literature
data, the authors of [177] investigated this issue. By analyzing the behavior of
individual components of the interaction energy, they were able to rational-
ize the trends within this family and relate them to monomer properties. The
changes of monomer properties are very significant across the family: the po-
larizabilities increase many times as the monomer size increases, whereas the
dipole moments of HX decrease. As an effect, the importance of the induction
energy – which strongly favors the Rg – H – X configuration – decreases by
a factor of 200 as one moves from Kr – HF to He – HI. Thus, the demarcation
line for the ordering of the minima lies between these two complexes. One
has to take account of the exchange and dispersion interactions for complexes
close to this line in order to predict the correct structure.

7.3
He – N2 O

Reference [183] describes SAPT calculations that were performed to deter-

mine a two-dimensional potential for the interaction of the helium atom
with the nitrous oxide molecule. This system has been very popular in re-
cent years, as shown by publications of two later potentials from two different
groups [184, 185]. The reason for this interest is experimental investigations
of N2 O embedded in superfluid helium nanodroplets [186]. To estimate the
accuracy of SAPT interaction energies, the authors of [183] performed su-
permolecular CCSD(T) calculations for selected geometries. The ab initio
interaction energies were fitted to an analytic function and rovibrational
energy levels of He – N2 O were computed on the resulting surface. Exten-
sive comparisons were made with a literature ab initio He – CO2 potential
and rovibrational states [240] in order to rationalize the highly counterintu-
itive observations concerning spectra of N2 O and CO2 in superfluid helium
nanodroplets [186], in particular, the degree of reduction of the rotational
constants of the molecules between the gas phase and the nanodroplet. Ref-
erence [183] argued that the large reduction of the N2 O rotational constant
compared to CO2 is related to the greater potential depth and the resulting
greater probability of attaching helium atoms in the former case. Also, the
characteristics of the lowest vibrational levels for the two systems are quite
different in a way related to the experimental findings. As a byproduct of
Intermolecular Interactions 101

this work, accurate multipole moments of N2 O have been computed. The

quadrupole, octupole, and hexadecapole moments are significantly different
from the experimental values and are probably more accurate than the latter.
Based on simple minimizations of potential energy surfaces, the authors of
reference [183] predicted the structures of Hen – N2 O clusters. Independently,
and virtually simultaneously, similar structures were seen in experiments [187].
The same structures were later found in quantum Monte Carlo calculations
by Paesani and Whaley [188] who used the potential of [183]. Although the
latter results are more reliable than the static predictions, the underlying phys-
ical mechanisms leading to the creation of given types of clusters are more
transparent from the minimization work and analysis of the features of PES.

7.4
He – HCCCN

Another molecule investigated in superfluid helium nanodroplets is cyano-

acetylene (HCCCN) [189, 190]. This molecule is of interest as a model of
elongated species. Five two-dimensional potential energy surfaces for the in-
teraction of He with HCCCN were obtained from ab initio calculations using
symmetry-adapted perturbation theory and the supermolecular method at
different levels of electron correlation [195]. HCCCN was taken to be a rigid
linear molecule with the interatomic distances fixed at the experimental “r0 ”
geometry extracted from ground-state rotational constants. The complex was
found to have the global minimum at a T-shaped configuration and a sec-
ondary minimum at the linear configuration with the He atom facing the
H atom. Two saddle points were also located. There was good agreement
between the positions of the stationary points on each of the five surfaces,
though their energies differed by up to 19%. Rovibrational bound state calcu-
lations were performed for the 4 He – HCCCN and 3 He – HCCCN complexes.
Spectra (including intensities) and wave functions of 4 He – HCCCN obtained
from these calculations were presented. No experimental spectra have been
published for He – HCCCN, so theory may guide future experiments. The
effective rotational constant of HCCCN solvated in a helium droplet was esti-
mated by minimizing the energy of Hen – HCCCN for n = 2–12, selecting the
n = 7 complex as giving the largest magnitude of the interaction energy per
He, and shifting the resulting ring of He atoms to the position corresponding
to the average geometry of the ground state of the He – HCCCN dimer. This
estimate was within 4.8% of the measured value [189].

7.5
H2 – CO

The H2 – CO system is highly important in astrophysics and has been the

subject of investigations from several groups. The 1998 SAPT potential pre-
102 K. Szalewicz et al.

dicted spectra of this dimer [192], in excellent agreement with experimental

results of McKellar [193, 194], as well as with later experiments by McKel-
lar [195, 196]. The agreement was so good that there was hope that theoretical
results can be used to assign the measured, very dense spectrum of the
ortho-H2 – CO complex. The ab initio spectrum was actually not sufficiently
accurate, but one could expect that if the potential were tuned to the para-
H2 – CO data, the goal could be reached. This turned out not to be the case,
and a new four-dimensional energy surface was developed recently [197]. The
ab initio calculations have actually been performed on a five-dimensional
grid and the 4-D surface has been obtained by averaging over the intramolec-
ular vibration of H2 (the C – O distance was fixed in [197]). Since the goal was
to get as accurate results as possible, the supermolecular CCSD(T) method
was used. The correlation part of the interaction energy was obtained by per-
forming extrapolations from a series of basis sets. An analytical fit of the ab
initio set of points has the global minimum of – 93.049 cm–1 for an inter-
molecular separation of 7.92 bohr in the linear geometry with the C atom
pointing towards the H2 molecule. Thus, the potential is significantly shal-
lower than the SAPT potential of 1998. It has been found that the major
reason for this discrepancy is that the SAPT potential included an estimate
of the induction and exchange-induction energies of the third and higher
orders given by δEHFint,resp . This approximation turned out to work poorly for
H2 – CO. The new potential has been used to calculate rovibrational energy
levels of the para-H2 – CO complex, which agree very well with those ob-
served by McKellar [195]: their accuracy is better than 0.1 cm–1 , whereas the
1998 energies agreed to better than 1 cm–1 . The calculated dissociation en-
ergy is equal to 19.527 cm–1 and is significantly smaller than the value of
22 cm–1 estimated from the experiment. It turned out that the use of the ex-
perimental dissociation energy was the main reason for the tuning of the 1998
potential not achieving the anticipated accuracy. The predictions of rovibra-
tional energy levels for ortho-H2 – CO have been performed and will serve
to guide assignment of the recorded experimental spectra. Also, the inter-
action second virial coefficient has been calculated and compared with the
experimental data. The agreement between theory and experiment here is
unprecedented.

7.6
Methane-water

Methane-water interactions are of significant interest. Not only is this system

an important model of hydrophobic interactions, but methane-water clathrates
(also commonly called methane hydrates) are one of the largest energy re-
sources on Earth. These clathrates are nonstoichiometric mixtures of the two
molecules, with the water molecules forming a cage (clathrate) around a me-
thane molecule. Even conservative estimates predict that methane clathrates
Intermolecular Interactions 103

contain twice the energetic resources of the conventional fossil fuels in world
reserves [198]. A 6-dimensional potential energy surface has been developed
for interactions between water and methane [199]. The global minimum of the
potential, with a depth of 1.0 kcal/mol, was found in a “hydrogen bond” config-
uration with water being the proton donor and the bond going approximately
through the midpoint of a methane tetrahedral face. However, the interaction
was shown to have few of the characteristics of typical hydrogen bonds [200].
The SAPT calculations on a grid of about 1000 points were fitted by an ana-
lytic site-site potential using previously developed methods [28, 201]. Most of
the sites were placed on the atoms. The asymptotic part of the potential was
computed ab initio using the Wormer and Hettema set of codes for monomer
properties [115, 116]. The classical cross second virial coefficient was calculated
and agreed well with some experiments but not with others, allowing an evalu-
ation of the quality of experimental results. The potential is now being used in
simulations of methane interactions in aqueous solutions.

7.7
Water Dimer

Interactions between water molecules have been the subject of investigations

much more often than any other intermolecular interactions, obviously due
to the importance of water in all aspects of human life. An elaborate poten-
tial for water using IMPT was developed by Millot and Stone [202] and named
ASP. Later an improved version was published [203]. SAPT was used to de-
velop a water dimer potential called SAPT-5s [28–30]. The latter potential
and its applications in spectroscopy [29, 204, 205] and in simulations of li-
quid water [31, 32] have been recently reviewed in [20, 21]. Reference [21] also
described an improved potential obtained using SAPT(DFT).
A potential for the water dimer with flexible monomers has recently been
developed [206]. The resulting potential energy surface is 12-dimensional,
which is at the limits of complexity that one can deal with. About half a mil-
lion points have been computed (compared to about 2500 in [30] using rigid
monomers). Fitting these data required a very substantial effort. The poten-
tial will be used in calculations of the spectra of water dimer and of the second
virial coefficient for water.

8
Performance of the SAPT(DFT) Method

A recent paper [133] is an extension of earlier work [131] and describes the
implementation of the SAPT(DFT) version without the coupled Kohn-Sham
dispersion energies. Since this version is based only on Kohn-Sham orbitals
and orbital energies, it is called SAPT(KS). In addition to the He2 and (H2 O)2
104 K. Szalewicz et al.

systems investigated already in [131], the paper describes calculations for Ne2
and (CO2 )2 dimers at several geometries. In all cases, calculations were per-
formed using several basis sets and a number of DFT functionals. The role
of the asymptotic corrections to the exchange-correlation potentials has been
investigated. It was shown that the Fermi-Amaldi correction [207] provides
the most reliable results. The role of this correction ranges from dramatic in
the case of the electrostatic energy of He2 (order of magnitude improvements
in the intramonomer correlation contribution) to rather small for the carbon
dioxide dimer at the level of accuracy possible for this system.
It has been found [133] that SAPT(KS) converges much faster with basis set
size than the regular SAPT. This is due to the fact that the rate of convergence
in the regular SAPT is determined by the slow convergence of the correla-
tion cusps in products of orbitals. Such terms do not appear in SAPT(KS)
except for the dispersion energy. However, even the cusps in the dispersion
energy are not a problem for SAPT(KS). This is due to the fact that the disper-
sion energy converges fast already in the regular SAPT provided that the basis
set contains diffuse orbitals (possibly optimized for the dispersion energy)
and that the so-called midbond functions are used (orbitals placed mid-way
between monomers). Reference [133] found that both the dispersion energy
and all other terms of SAPT(KS) are very accurate if this type of basis set is
used. This result contrasts with regular SAPT, where bases optimal for dis-
persion energies lead to slow convergence of other terms. It was possible to
achieve practically converged SAPT(KS) results for all systems but the car-
bon dioxide dimer without any need for extrapolations to the complete basis
set (CBS) limit. When such extrapolations were performed for the regular
SAPT, it was possible to make comparisons free of any basis set artifacts. In-
terestingly enough, in all cases the agreement between the two approaches
was much better at the CBS limit than in finite bases, where the regular SAPT
result had significant basis set incompleteness errors.
SAPT(KS) was found [133] to be relatively independent of the DFT func-
tional used. This is in contrast to supermolecular DFT calculations of interac-
tion energies where this dependence is dramatic. All the modern functionals
applied gave reasonably close SAPT(KS) components. However, hybrid poten-
tials were generally performing better than the non-hybrid ones. In particu-
lar, the functionals PBE0 [208, 209] and B97-2 [210, 211] gave accurate results.
For He2 , accurate benchmarks exist for all the components of the inter-
action energy [147]. In some cases, the SAPT(KS) components were closer
to the benchmark than the regular SAPT components. This shows that the
truncation of the expansion in W in SAPT introduces larger errors than the
inaccuracies of the current best DFT functionals. For systems larger than
He2 , however, the only benchmarks available are those from SAPT. Thus, it
is very difficult to answer the question of whether SAPT(KS) components
may be more accurate for such systems than the regular SAPT ones. However,
in a few cases this was possible. In particular, since recently the electro-
Intermolecular Interactions 105

static energy has been available in SAPT at the CCSD level,4 see also [212].
E(1)
elst (CCSD) inlcudes a much higher level of electron correlation than
does E(12) (13)
elst,resp + Eelst,resp . Some additional estimates could also be performed
based on asymptotic comparisons [133]. In all cases, the SAPT(KS) results
were closer to the values computed at the higher level than to the regular
SAPT results. This shows that – for the electrostatic, first-order exchange,
second-order induction and exchange-induction energies – SAPT(KS) is not
only approaching but occasionally surpassing the accuracy of regular SAPT at
the currently programmed theory level.
Reference [133] provided theoretical justifications for high accuracy of
SAPT(KS) predictions for the electrostatic, first-order exchange, and second-
order induction energies. For the electrostatic energy, the argument is very
simple and has already been given here in Sect. 5.2. The CKS induction en-
ergies developed in [133] are analogous to the CKS dispersion energies. As
for the dispersion energy [cf. Eqs. 99 and 100], the exact second-order in-
duction energy can be written in terms of the FDDSs, now computed at zero
frequency, and of the electrostatic potentials of the unperturbed monomers.
In the CKS induction energy, one uses CKS FDDSs and regular DFT elec-
trostatic potentials. Since both quantities are potentially exact in DFT, i.e.
would be exact if the exact exchange-correlation potential were known, the
induction energy is potentially exact. Since modern density functionals are,
for these purposes, reasonable approximations to the exact functional, the
CKS induction energies are quite accurate. In fact, somewhat surprisingly, the
uncoupled KS induction energies are also very accurate. Reference [133] ra-
tionalized this behavior by conjecturing that there is a systematic cancellation
between uncoupled and coupled polarizability differences and the respective
differences in the overlap effects. Indeed, asymptotically – where overlap ef-
fects are small – the CKS induction energies are more accurate.
For the exchange energies, the justification of the good performance
of SAPT(KS) is more difficult. An asymptotic expression has been de-
veloped [133] for the interaction density matrices which determine the first-
order exchange energy in the case of the KS determinants and the exact wave
functions. It was shown that in the limit of exact DFT both densities decay in
the same way. Thus, at least asymptotically, the first-order exchange energies
are potentially exact.
A related manuscript has been devoted to the coupled Kohn-Sham dis-
persion energies in SAPT(DFT) [134]. The method utilizes a generalized
Casimir-Polder formula and frequency-dependent density susceptibilities of
monomers obtained from time-dependent DFT. Numerical calculations were
performed for the same systems as in [133]. It has been shown that for a wide
range of intermonomer separations, including the van der Waals and the
short-range repulsion regions, the method provides dispersion energies with
4 Wheatley RJ (2003) unpublished results
106 K. Szalewicz et al.

accuracies comparable with those that can be achieved using the current most
sophisticated wave function methods. The dependence of the CKS dispersion
energy on basis sets and on variants of the DFT method has been investigated
and the relations were found to be very similar to those for the SAPT(KS)
theory discussed above. For the carbon dioxide dimer, the dispersion energy
predicted by SAPT(DFT) turned out to be significantly different from that
given by SAPT at the level of Eq. 94. An asymptotic analysis strongly sug-
gested that it is the CKS dispersion energy which is more accurate. A further
confirmation comes from the fact that whereas SAPT predictions for the CO2
dimer are significantly different from supermolecular CCSD(T) interaction
energies, SAPT(DFT) is in very good agreement with the latter. This finding
resolves the long-standing issue [201] of the somewhat unsatisfactory per-
formance of SAPT for this particular system.
If the CKS dispersion energy is combined with the electrostatic and exchange
interaction energies from the SAPT(KS) method and the CKS induction en-
ergies, very accurate total interaction potentials are obtained. For the helium
dimer, the only system with nearly exact benchmark values, SAPT(DFT) re-
produces the interaction energy to within about 2% at the minimum and to
a similar accuracy for all other distances ranging from the strongly repulsive
to the asymptotic region. An accurate SAPT(DFT) potential for this system has
also been published by Hesselmann and Jansen [213]. For the remaining sys-
tems investigated in [134], the quality of the interaction energies produced by
SAPT(DFT) is so high that these energies may actually be more accurate than
the best available results obtained with wave function techniques. At the same
time, SAPT(DFT) is much more efficient computationally than any method
previously used for computing the dispersion and other interaction energy
components at this level of accuracy, as discussed in Sect. 5.3.

9
Applications of SAPT(DFT) to Molecular Crystals

One of the most interesting applications of SAPT(DFT) is to predicting struc-

tures of molecular crystals. Until recently, these crystals could only be inves-
tigated using empirical potentials since, due to the size of typical monomers,
SAPT calculations were not practical for such systems.
Molecular crystals are an important class of matter. The current progress
in molecular biology and medicine would not be possible if structures of
biomolecules were not known from an X-ray analysis of crystalline forms of
these compounds. Typical medical drugs are molecular crystals and so are
most energetic materials. The ability to predict polymorphic forms of medical
drugs would be extremely important in pharmacology. Significant experi-
mental efforts are directed towards creating new materials in the form of
molecular crystals with some desired properties.
Intermolecular Interactions 107

Many properties of molecular crystals can be predicted theoretically [214–

216]. The first step in theoretical investigations of such crystals has to be
a determination of interactions (force fields) between constituent molecules.
The force field for a crystal can be built from pair and pair-nonadditive con-
tributions computed for isolated dimers, trimers, etc. Such force fields can
be obtained empirically or from first-principles calculations. The next step
is to determine the low-energy crystal structures applying molecular pack-
ing programs. So far, this has been done mostly using empirical force fields.
However, the inherent inaccuracies of such force fields limit the predictability
of this approach [214–217]. A solution could be to use force fields obtained
from first-principles quantum mechanical calculations. Unfortunately, wave-
function-based methods are too time consuming for this, whereas, as dis-
cussed earlier, the supermolecular DFT approach cannot predict dispersion
energies which can be significant for these systems. For a recent demonstra-
tion of this fact, see the work of Byrd et al. [218]. Clearly, the solution to this
problem is to use SAPT(DFT).
It is hoped that the force fields computed using first-principles methods
should be able to predict properties of molecular crystals significantly better
than it is currently possible using empirical potentials. In particular, the-
ory could play a very important role in screening notional materials, i.e.
molecules that may not have been synthesized, but based on their expected
structures appear to have desired properties. For such molecules, specific em-
pirical potentials are simply unknown. Generic potentials can be used, but
these are not very reliable in predicting crystal structures.
In the subsections that follow, we will describe some pilot SAPT(DFT)
calculations for molecules that can be considered models for those forming
molecular crystals.

9.1
Benzene Dimer

Solid benzene is one of the most thoroughly investigated molecular crystals as it

represents the model system for aromatic compounds. However, there exist no
ab initio potentials for the benzene dimer. Ab initio calculations have been per-
formed only for selected geometries and it is difficult to estimate their accuracy.
Recently, SAPT(DFT) was applied to this system [143]. An augmented double-
zeta basis supplemented by a set of bond functions was used and calculations
were performed for a range of intermolecular separations in the “sandwich”
configuration. As it is well known, supermolecular DFT methods fail completely
for the benzene dimer, in most cases predicting the wrong sign of the interaction
energy [219]. SAPT(DFT) curve agreed very well with the best previous calcu-
lations performed by Tsuzuki et al. at the CCSD(T) level and including some
extrapolations [220]. Near the minimum, at R = 3.8 Å, the two methods pre-
dict interaction energies of – 1.62 and – 1.67 kcal/mol, respectively. However,
108 K. Szalewicz et al.

the residual error of either result (due to basis set truncations and other factors)
is somewhat larger than their difference, as indicated by extensive R12-MP2
plus CCSD(T) single-point calculations by Sinnokrot et al. [221] who obtained
– 1.81 kcal/mol at 3.7 Å. The latter result agrees very well with preliminary
SAPT(DFT) calculations in very large basis sets. Reference [143] also included
calculations for Ar2 and Kr2 , in both cases obtaining excellent agreement with
the best existing potentials. The accuracy for all three systems was signifi-
cantly higher than that of some recent DFT approaches created specifically for
calculations of intermolecular interactions [222, 223].

9.2
Dimethylnitramine Dimer

Dimethylnitramine (DMNA) is an important model compound for energetic

materials and was investigated by SAPT in the past [224]. Recently, interac-
tion energies were computed for the DMNA dimer containing 24 atoms [142].
In Table 7, the total interaction energies at a near minimum geometry com-
puted using SAPT, SAPT(DFT), and several supermolecular methods are
shown. The basis set used was of double-zeta quality with bond functions
in monomer-centered “plus” basis set (MC+ BS) form [37] [dimer-centered
“plus” basis set (DC+ BS) form in the (counterpoise corrected) supermolecu-
lar calculations]. The “plus” denotes here the use of bond functions (in the
MC+ BS case, also the use of the isotropic part of the basis of the interact-
ing partner). The regular SAPT results given in Table 7 employ the complete
(20)
standard set of corrections, in contrast to [224] which used EHF
int + Edisp .

Table 7 Interaction energies (in kcal/mol) for the DMNA dimer, from [142]. The geom-
etry was the near-minimum one denoted by M1 in Table 3 of [224] and the basis set was
also taken from that reference. MPn denotes many-body perturbation theory with the MP
Hamiltonian

Hartree-Fock 2.25

Frozen-core
MP2 –7.90
MP4 –7.85
CCSD –5.31
CCSD(T) –6.85
All electrons
CCSD(T) –6.86
(20)
EHF
int + Edisp –10.58
SAPT (Eqs. 92–95) –7.36
SAPT(DFT)/PBE0 –6.22
SAPT(DFT)/B97-2 –6.56
Intermolecular Interactions 109

Table 8 Individual components of the DMNA dimer interaction energy for SAPT and
SAPT(DFT) with PBE0 and B97-2 functionals. Energies are in kcal/mol. The value in
parentheses is E(20)
disp . All data are from [142]

Component SAPT PBE0 B97-2

Electrostatic –10.51 –10.25 –10.05

1st order exchange 18.28 17.43 16.85
Induction –6.07 –6.54 –6.35
Exchange-induction 4.49 5.02 4.82
Dispersion –13.50 (–12.83) –11.83 –11.75
Exchange-dispersion 1.34 1.34 1.30
δEHF
int –1.38 –1.38 –1.38
Total –7.36 –6.22 –6.56

Table 7 shows first that the higher-order terms neglected in [224] are im-
portant for the DMNA dimer and decrease the magnitude of the interaction
(20)
energy by more than 3 kcal/mol. The value of EHF int + Edisp in Table 7 differs
from the minimum energy of – 11.06 kcal/mol given in Table 3 of [224] due
to the use of the DC+ BS vs. MC+ BS scheme. SAPT(DFT) gives interaction en-
ergies within about 1 kcal/mol of the regular SAPT and about 0.5 kcal/mol
of the CCSD(T) method. This constitutes excellent agreement taking into ac-
count that both the regular SAPT and CCSD(T) methods are much more
computer resource intensive than SAPT(DFT). The SAPT(DFT) calculations
were performed with two very different functionals: PBE0 and B97-2, which
gave results within 0.3 kcal/mol of each other, showing again that SAPT(DFT)
is only weakly dependent on the choice of the functional.
The framework of SAPT provides insights into the physical structure of the
interaction energy. Table 8 shows the individual contributions. It can be seen
that, as already pointed out in [224], SAPT results do not support the conven-
tional description of interactions of large molecules, which considers only the
electrostatic component. Clearly, the first-order exchange and the dispersion
energies are actually larger in magnitude than the electrostatic interactions.
An attempt to describe the DMNA dimer at the Hartree-Fock level, as it is of-
ten done for large molecules, would lead to completely wrong conclusions as
the interaction energy at this level is positive. Note also the good agreement
between the individual SAPT and SAPT(DFT) components.

10
Transferable Potentials for Biomolecules
Weak interactions between biomolecules govern a significant part of life’s
processes. With the greatly expanded range of systems that can be investi-
110 K. Szalewicz et al.

gated ab initio as a result of the development of SAPT(DFT), some of these

processes become important subjects for computational research. The small-
est biomolecules, such as DNA bases, polypeptides, and sugars, are within
the range of systems for which the complete potential energy surfaces can
be obtained. So far, ab initio calculations for such systems have been re-
stricted to single-point calculations [225, 226]. Due to the size of these sys-
tems, often even single-point calculations could only be performed at low
levels of theory and using very small basis sets. If PESs are available, prop-
erties of biomolecules in aqueous solutions can be studied by molecular
simulations. Such studies with the use of empirical potentials have been
very popular [227]. For some systems experimental data exist also for iso-
lated dimers [228] so that direct comparisons with ab initio calculations are
possible.
Ab initio methods can also be used in several ways to investigate much
larger molecules than one can afford by direct calculations. One avenue is to
perform calculations on a fragment of a large biological molecule interacting
with another fragment or with a smaller molecule, for example with water.
Another way to extend the range of applicability is to use ab initio informa-
tion to construct universal force fields for biomolecules. Such force fields are
similar to the currently used empirical ones [229–231] in the sense that the
interaction between two arbitrary molecules depends only on predetermined
interactions between pairs of atoms belonging to different molecules. A given
atom may come in a few “varieties”, depending on its chemical surround-
ing. In the existing biomolecular force fields, this information comes from
atomic properties such as polarizabilities, van der Waals radii, and partial
charges. In many force fields, some of the parameters, in particular the par-
tial charges, have been obtained from ab initio calculations for monomers.
In fact, very intensive investigations have been performed to represent elec-
trostatic interactions using the distributed multipole analysis [117, 119, 121].
In recent years, a subset of parameters in force fields has often been tuned
by adjusting them within molecular dynamic simulations of some model sys-
tems. Such force fields are usually specific for a class of systems, for example
proteins.
An ongoing research effort is aimed at proposing an universal force field
based on SAPT calculations for a set of model complexes.5 This project had
started before SAPT(DFT) was fully developed, and therefore mostly the
regular SAPT approach was used. The future use of SAPT(DFT) will make
it possible to increase the size of the model complexes. One component of
the new force field, the electrostatic energy, is not parametrized but actu-
ally computed as the Coulomb interaction of the charge distributions of the
interacting monomers [232]. In this way, the overlap (penetration) effects
are fully taken into account and there are no problems appearing related to

5 Volkov A, Coppens P, Macchi P, Szalewics K unpublished results

Intermolecular Interactions 111

the divergence of multipole expansions. The monomer (molecular) charge

distributions are represented as linear combinations of the so-called pseu-
doatom densities [233, 234]. The pseudoatom densities have been extracted
from ab initio molecular densities of a large number of small molecules using
a least-squares projection technique in Fourier-transform space [235], and
are available in the form of a “databank”, with the first applications reported
in [236]. For pseudoatoms that are far apart (separations larger than 4–5 Å),
the electrostatic energy can be computed from distributed multipoles eval-
uated directly from pseudoatom parameters [234]. In effect, the calculation
of the electrostatic energy is reasonably fast although, of course, not as fast
as a summation of point-charge interactions. The remaining terms of the in-
teraction energy are obtained by simultaneous fits of the computed SAPT
components for all model systems. Thus, the interaction energy is approxi-
mated as5
⎡ ⎤
  An An
Eint = E(1) ⎣Ba Bb e–(Ca +Cb )rab – a b⎦
elst + n
rab
. (105)
a∈A,b∈B n=6,8,10

This formula is somewhat reminiscent of the one proposed by Spack-

man [237–239], but the method of determination of the parameters is com-
pletely different. The parameters appearing in the exponential term were
fitted to reproduce the exchange energy of Eq. 95 whereas those in front of
inverse powers of distances were fitted to the sum of the induction and dis-
persion energies of Eqs. 93 and 94 for the model set. In practice, some more
time-consuming terms in these equations were neglected. A variant of fitting
the induction and dispersion energies separately was also explored. There
were more than one hundred dimer configurations in the model set. The
monomers included amino acids such as: L-serine, L-glutamine, α-glycine,
and several other; amino-acid-like compounds such as, for example, DL-
norleucine; L-(+)-lactic acid, and benzene. The force field was tested by
predicting interaction energies for configurations not included in the model
set and for other dimers for which ab initio calculations were possible. It
was also tested by finding the lattice binding energies for the crystals built of
model monomers.

Acknowledgements This research was supported by the NSF grant CHE-0239611 and by
an ARO DEPSCoR grant. B.J. acknowledges a generous support from the Foundation for
Polish Science.

References
1. Hutson JM (1990) Annu Rev Phys Chem 41:123
2. Cohen RC, Saykally RJ (1991) Annu Rev Phys Chem 42:369
3. Buck U (1975) Adv Chem Phys 30:313
112 K. Szalewicz et al.

4. Melnick GJ, Stauffer JR, Ashby MLN, Bergin EA, Chin G, Erickson NR, Goldsmith
PF, Harwit M, Howe JE, Kleiner SC, Koch DG, Neufeld DA, Patten BM, Plume R,
Schieder R, Snell RL, Tolls V, Wang Z, Winewisser G, Zhang YF (2000) Astrophys J
539:L77
5. Allen MP, Tildesley DJ (1987) Computer Simulation of Liquids. Clarendon Press,
Oxford
6. Toennies JP, Vilesov AF (1998) Annu Rev Phys Chem 49:1
7. Callegari C, Lehmann KK, Schmied R, Scoles G (2001) J Chem Phys 115:10090
8. Burger A (1983) A Guide to the Chemical Basis of Drug Design. Wiley, New York
9. Naray-Szabo G (ed) (1986) Theoretical Chemistry of Biological Systems, vol 41 of
Studies in Theoretical and Physical Chemistry. Elsevier, Amsterdam
10. Autumn K, Liang YA, Hsieh ST, Zesch W, Chang WP, Kenny TW, Fearing R, Full RJ
(2000) Nature 405:681
11. Chałasiński G, Szczȩśniak MM (1994) Chem Rev 94:1723
12. Chałasiński G, Szczȩśniak MM (2000) Chem Rev 100:4227
13. Bartlett RJ (1989) J Phys Chem 93:1697
14. Peréz-Jordá JM, Becke AD (1995) Chem Phys Lett 233:134
15. Chałasiński G, Szczȩśniak MM (1988) Mol Phys 63:205
16. Rode M, Sadlej J, Moszyński R, Wormer PES, van der Avoird A (1999) Chem Phys
Lett 314:326
17. Sadlej AJ (1991) J Chem Phys 95:6705
18. van Duijneveldt FB, van Duijneveldt-van de Rijdt JGCM, van Lenthe JH (1994) Chem
Rev 94:1873
19. Jeziorski B, Moszyński R, Szalewicz K (1994) Chem Rev 94:1887
20. Jeziorski B, Szalewicz K (2003) In: Wilson S (ed) Handbook of Molecular Physics and
Quantum Chemistry. Wiley, vol 3, Part 2, Chap. 9, p 232
21. Szalewicz K, Bukowski R, Jeziorski B (2005) In: Dykstra CE, Frenking G, Kim KS,
Scuseria GE (eds) Theory and Applications of Computational Chemistry: The First
40 Years. A Volume of Technical and Historical Perspectives, Chap 33. Elsevier, Am-
sterdam, p 919
22. Eisenschitz R, London F (1930) Z Phys 60:491
23. Rybak S, Jeziorski B, Szalewicz K (1991) J Chem Phys 95:6579
24. SAPT2002: An Ab Initio Program for Many-Body Symmetry-Adapted Perturbation
Theory Calculations of Intermolecular Interaction Energies Bukowski R, Cencek
W, Jankowski P, Jeziorska M, Jeziorski B, Kucharski SA, Lotrich VF, Misquitta AJ,
Moszyński R, Patkowski K, Rybak S, Szalewicz K, Williams HL, Wormer PES, Univer-
sity of Delaware and University of Warsaw (http://www.physics.udel.edu/∼szalewic/
SAPT/SAPT.html)
25. Szalewicz K, Jeziorski B (1997) In: Scheiner S (ed) Molecular Interactions – from van
der Waals to strongly bound complexes. Wiley, New York, p 3
26. Jeziorski B, Szalewicz K (1998) In: von Ragué Schleyer P et al. (eds) Encyclopedia of
Computational Chemistry, vol 2. Wiley, Chichester, UK, p 1376
27. Moszyński R, Wormer PES, van der Avoird A (2000) In: Bunker PR, Jensen P (eds)
Computational Molecular Spectroscopy. Wiley, New York, p 69
28. Mas EM, Szalewicz K, Bukowski R, Jeziorski B (1997) J Chem Phys 107:4207
29. Groenenboom GC, Mas EM, Bukowski R, Szalewicz K, Wormer PES, van der Avoird
A (2000) Phys Rev Lett 84:4072
30. Mas EM, Bukowski R, Szalewicz K, Groenenboom G, Wormer PES, van der Avoird A
(2000) J Chem Phys 113:6687
31. Mas EM, Bukowski R, Szalewicz K (2003) J Chem Phys 118:4386
Intermolecular Interactions 113

32. Mas EM, Bukowski R, Szalewicz K (2003) J Chem Phys 118:4404

33. Keutsch FN, Goldman N, Harker HA, Leforestier C, Saykally RJ (2003) Mol Phys
101:3477
34. Patkowski K, Korona T, Moszyński R, Jeziorski B, Szalewicz K (2002) J Mol Struct
(Theochem) 591:231
35. Brudermann J, Steinbach C, Buck U, Patkowski K, Moszyński R (2002) J Chem Phys
117:11166
36. Stone AJ (1996) The Theory of Intermolecular Forces. Clarendon Press, Oxford
37. Williams HL, Mas EM, Szalewicz K, Jeziorski B (1995) J Chem Phys 103:7374
38. Chipman DM, Hirschfelder JO (1973) J Chem Phys 59:2838
39. Chałasiński G, Jeziorski B, Szalewicz K (1977) Int J Quantum Chem 11:247
40. Jeziorski B, Szalewicz K, Chałasiński G (1978) Int J Quantum Chem 14:271
41. Chałasiński G, Szalewicz K (1980) Int J Quantum Chem 18:1071
42. Jeziorski B, Schwalm WA, Szalewicz K (1980) J Chem Phys 73:6215
43. Certain PR, Hirschfelder JO, Kołos W, Wolniewicz L (1968) J Chem Phys 49:24
44. Bowman JD (1973) PhD Thesis, University of Wisconsin
45. Ćwiok T, Jeziorski B, Kołos W, Moszyński R, Szalewicz K (1992) J Chem Phys 97:7555
46. Ćwiok T, Jeziorski B, Kołos W, Moszyński R, Rychlewski J, Szalewicz K (1992) Chem
Phys Lett 195:67
47. Ćwiok T, Jeziorski B, Kołos W, Moszyński R, Szalewicz K (1994) J Mol Struct
(Theochem) 307:135
48. Korona T, Jeziorski B, Moszyński R, Diercksen GHF (1999) Theor Chem Acc 101:282
49. Korona T, Moszyński R, Jeziorski B (1997) Adv Quantum Chem 28:171
50. Hirschfelder JO, Silbey R (1966) J Chem Phys 45:2188
51. Chipman DM, Bowman JD, Hirschfelder JO (1973) J Chem Phys 59:2830
52. Adams WH (1990) Int J Quantum Chem S24:531
53. Adams WH (1991) Int J Quantum Chem S25:165
54. Adams WH (1994) Chem Phys Lett 229:472
55. Adams WH (1996) Int J Quantum Chem 60:273
56. Morgan JD III, Simon B (1980) Int J Quantum Chem 17:1143
57. Patkowski K, Korona T, Jeziorski B (2001) J Chem Phys 115:1137
58. Kutzelnigg W (1980) J Chem Phys 73:343
59. Hirschfelder JO (1967) Chem Phys Lett 1:343
60. van der Avoird A (1967) J Chem Phys 47:3649
61. Peierls R (1973) Proc R Soc London, Ser A, 333:157
62. Amos AT, Musher JI (1969) Chem Phys Lett 3:721
63. Polymeropoulos EE, Adams WH (1978) Phys Rev A 17:18
64. Kutzelnigg W (1978) Int J Quantum Chem 14:101
65. Ahlrichs R (1976) Theor Chim Acta 41:7
66. Jeziorski B, Kołos W (1982) In: Ratajczak H, Orville-Thomas W (eds) Molecular
Interactions, vol 3. Wiley, New York, p 1
67. Jeziorski B, Kołos W (1977) Int J Quantum Chem Suppl 1 12:91
68. Adams WH (1999) Int J Quantum Chem 72:393
69. Patkowski K, Jeziorski B, Korona T, Szalewicz K (2002) J Chem Phys 117:5124
70. Patkowski K, Jeziorski B, Szalewicz K (2001) J Mol Struct (Theochem) 547:293
71. Adams WH (2002) Theor Chem Acc 108:225
72. Patkowski K, Jeziorski B, Szalewicz K (2004) J Chem Phys 120:6849
73. Stone AJ, Hayes IC (1982) Faraday Discuss 73:19
74. Hayes IC, Stone AJ (1984) Mol Phys 53:69
75. Hayes IC, Stone AJ (1984) Mol Phys 53:83
114 K. Szalewicz et al.

76. Hayes IC, Hurst GJB, Stone AJ (1984) Mol Phys 53:107
77. Longuet-Higgins HC (1965) Disc Farad Soc 40:7
78. Hirschfelder JO (1967) Chem Phys Lett 1:325
79. Bloch C (1958) Nucl Phys 5:329
80. Szalewicz K, Jeziorski B (1979) Mol Phys 38:191
81. Claverie P (1971) Int J Quantum Chem 5:273
82. Adams WH (2002) Int J Quantum Chem 90:54
83. Matsen FA (1964) Adv Quantum Chem 1:59
84. Kaplan IG (1975) Symmetry of Many-Electron Systems. Academic Press, New York
85. Murrell JN, Shaw G (1967) J Chem Phys 46:1768
86. Musher JI, Amos AT (1967) Phys Rev 164:31
87. Przybytek M, Patkowski K, Jeziorski B (2004) Collect Czech Chem Commun 69:141
88. Adams WH (2002) J Mol Struct (Theochem) 591:59
89. Korona T, Moszyński R, Jeziorski B (1996) J Chem Phys 105:8178
90. Herring C (1962) Rev Mod Phys 34:631
91. Ewald PP (1921) Annalen der Physik, Ser. 4 64:253
92. Panas I (1995) Chem Phys Lett 245:171
93. Dombroski JP, Taylor SW, Gill PMW (1996) J Phys Chem 100:6272
94. Sirbu I, King HF (2002) J Chem Phys 117:6411,
95. Patkowski K (2003) PhD Thesis, University of Warsaw
96. Gutowski M, Piela L (1988) Mol Phys 64:337
97. Kołos W, Wolniewicz L (1965) J Chem Phys 43:2429
98. Jankowski P, Jeziorski B (1999) J Chem Phys 111:1857
99. Baker JD, Freund DE, Hill RN, Morgan JD III (1990) Phys Rev A 41:1241
100. Adams WH (2005) Int J Quantum Chem (in press)
101. Larsen H, Halkier A, Olsen J, Jørgensen P (2000) J Chem Phys 112:1107
102. Jeziorski B, Moszyński R, Rybak S, Szalewicz K In Kaldor U (1989) (eds) Many-Body
Methods in Quantum Chemistry, Lecture Notes in Chemistry, vol 52. Springer, New
York, p 65
103. Moszyński R, Jeziorski B, Szalewicz K (1993) Int J Quantum Chem 45:409
104. Moszyński R, Jeziorski B, Ratkiewicz A, Rybak S (1993) J Chem Phys 99:8856
105. Moszyński R, Jeziorski B, Rybak S, Szalewicz K, Williams HL (1994) J Chem Phys
100:5080
106. Moszyński R, Cybulski SM, Chałasiński G (1994) J Chem Phys 100:4998
107. Williams HL, Szalewicz K, Moszyński R, Jeziorski B (1995) J Chem Phys 103:4586
108. Patkowski K et al. to be published
109. Tang KT, Toennies JP (1984) J Chem Phys 80:3726
110. Stone AJ (1975) Mol Phys 29:1461
111. Stone AJ (1976) J Phys A 9:485
112. Tough RJA, Stone AJ (1977) J Phys A 10:1261
113. Stone AJ (1978) Mol Phys 36:241
114. Stone AJ, Tough RJA (1984) Chem Phys Lett 110:123
115. Wormer PES, Hettema H (1992) J Chem Phys 97:5592
116. Wormer PES, Hettema H (1992) POLCOR package, University of Nijmegen
117. Stone AJ (1981) Chem Phys Lett 83:233
118. Price SL, Stone AJ, Alderton M (1984) Mol Phys 52:987
119. Stone AJ, Alderton M (1985) Mol Phys 56:1047
120. Buckingham AD, Fowler PW, Stone AJ (1986) Int Rev Phys Chem 5:107
121. Stone AJ (1991) In: Maksic ZB, editor, Theoretical Models of Chemical Bonding,
vol 4. Springer, New York, p 103
Intermolecular Interactions 115

122. Stone AJ (1985) Mol Phys 56:1065

123. Fowler PW, Stone AJ (1987) J Phys Chem 91:509
124. Stone AJ (1989) Chem Phys Lett 155:102
125. Stone AJ (1989) Chem Phys Lett 155:111
126. Le Sueur RC, Stone AJ (1993) Mol Phys 78:1267
127. Williams GJ, Stone AJ (2003) J Chem Phys 119:4620
128. Stone AJ, Tong CS (1989) Chem Phys 137:121
129. Hättig C, Jansen G, Hess BA, JG Ángyán (1997) Mol Phys 91:145
130. Williams HL, Chabalowski CF (2001) J Phys Chem A 105:646
131. Misquitta AJ, Szalewicz K (2002) Chem Phys Lett 357:301
132. Misquitta AJ, Jeziorski B, Szalewicz K (2003) Phys Rev Lett 91:033201
133. Misquitta AJ, Szalewicz K (2005) J Chem Phys 122:214109
134. Misquitta AJ, Podeszwa R, Jeziorski B, Szalewicz K, to be published
135. Hesselmann A, Jansen G (2002) Chem Phys Lett 357:464
136. Hesselmann A, Jansen G (2002) Chem Phys Lett 362:319
137. Hesselmann A, Jansen G (2003) Chem Phys Lett 367:778
138. Zaremba E, Kohn W (1976) Phys Rev B 13:2270
139. Dmitriev Y, Peinel G (1981) Int J Quantum Chem 19:763
140. McWeeny R (1984) Croat Chem Acta 57:865
141. Bukowski R, Podeszwa R, Szalewicz K (2005) Chem Phys Lett 414:111
142. Szalewicz K, Podeszwa R, Misquitta AJ, Jeziorski B (2004) In: Simos T, Maroulis G
(eds) Lecture Series on Computer and Computational Science: ICCMSE 2004, vol 1.
VSP, Utrecht, p 1033
143. Podeszwa R, Szalewicz K (2005) Chem Phys Lett 412:488
144. Hesselmann A, Jansen G, Schütz M (2005) J Chem Phys 122:014103
145. Moldover MR, Trusler JPM, Edwards TJ, Mehl JB, Davis RS (1988) J Res Natl Inst
Stand Technol 93:85
146. Williams HL, Korona T, Bukowski R, Jeziorski B, Szalewicz K (1996) Chem Phys Lett
262:431
147. Korona T, Williams HL, Bukowski R, Jeziorski B, Szalewicz K (1997) J Chem Phys
106:5109
148. Janzen AR, Aziz RA (1997) J Chem Phys 107:914
149. Hurly JJ, Moldover MR (2000) J Res Natl Inst Stand Technol 105:667
150. Jeziorska M, Bukowski R, Cencek W, Jaszuński M, Jeziorski B, Szalewicz K (2003)
Coll Czech Chem Commun 68:463
151. Cencek W, Jeziorska M, Bukowski R, Jaszuński M, Jeziorski B, Szalewicz K (2004)
J Phys Chem A 108:3211
152. Anderson JB (2004) J Chem Phys 120:9886
153. Lotrich VF, Szalewicz K (2000) J Chem Phys 112:112
154. Luther H, Grohman K, Fellmuth B (1996) Meterologia 33:341
155. Moldover MR (1998) J Res Natl Inst Stand Technol 103:167
156. Cencek W, Szalewicz K, Jeziorski B (2001) Phys Rev Lett 86:5675
157. Łach G, Jeziorski B, Szalewicz K (2004) Phys Rev Lett 92:233001
158. Patkowski K et al. to be published
159. Cencek W, Komasa J, Pachucki K, Szalewicz K (2005) Phys Rev Lett, submitted
160. Komasa J, Cencek W, Rychlewski J (1999) Chem Phys Lett 304:293
161. Pachucki K, Komasa J (2004) Phys Rev Lett 92:213001
162. Lach G et al. to be published
163. Pachucki K, Sapirstein J (2001) Phys Rev A 63:213001
164. Moszyński R, Heijmen TGA, van der Avoird A (1995) Chem Phys Lett 247:440
116 K. Szalewicz et al.

165. Moszyński R, Heijmen TGA, Wormer PES, van der Avoird A (1996) J Chem Phys
104:6997
166. Koch H, Hättig C, Larsen H, Olsen J, Jørgensen P, Fernandez B, Rizzo A (1999)
J Chem Phys 111:10106
167. Maroulis G (2000) J Phys Chem A 104:4772
168. Rachet F, Chrysos M, Guillot-Noel C, Le Duff Y (2000) Phys Rev Lett 84:2120
169. Rachet F, Le Duff Y, Guillot-Noel C, Chrysos M (2000) Phys Rev A 61:062501
170. Guillot-Noel C, LeDuff Y, Rachet F, Chrysos M (2002) Phys Rev A 66:012505
171. Aziz RA (1993) J Chem Phys 99:4518
172. Boyes SJ (1994) Chem Phys Lett 221:467
173. Phelps AV, Greene CH, Burke JP Jr (2000) J Phys B 33:2965
174. Patkowski K, Murdachaew G, Fou CM, Szalewicz K (2005) Mol Phys 103:2031
175. Slavíček P, Kalus R, Paška P, Odvárková I, Hobza P, Malijevský A (2003) J Chem Phys
119:2102
176. Herman PR, LaRocque PE, Stoicheff BP (1988) J Chem Phys 89:4535
177. Murdachaew G, Szalewicz K, Jiang H, Bacic Z (2004) J Chem Phys 121:11839
178. Willey DR, Choong VE, De Lucia FC (1992) J Chem Phys 96:898
179. Zhang Y, Shi HY (2002) J Mol Struct (Theochem) 589:89
180. Lovejoy CM, Nesbitt DJ (1990) J Chem Phys 93:5387
181. Jiang H, Sarsa A, Murdachaew G, Szalewicz K, Bacic Z (2005) J Chem Phys, submit-
ted
182. Sarsa A, Bacic Z, Moskowitz JW, Schmidt KE (2002) Phys Rev Lett 88:123401
183. Chang BT, Akin-Ojo O, Bukowski R, Szalewicz K (2003) J Chem Phys 119:11654
184. Zhu YZ Xie DQ (2004) J Chem Phys 120:8575
185. Song XG, Xu YJ, Roy PN, Jager W (2004) J Chem Phys 121:12308
186. Nauta K, Miller RE (2001) J Chem Phys 115:10254
187. Xu YJ, Jager W, Tang J, Mc Kellar ARW (2003) Phys Rev Lett 91:163401
188. Paesani F, Whaley KB (2004) J Chem Phys 121:5293
189. Callegari C, Conjusteau A, Reinhard I, Lehmann KK, Scoles G (2000) J Chem Phys
113:10535
190. Merritt JM, Douberly GE, Miller RE (2004) J Chem Phys 121:1309
191. Akin-Ojo O, Bukowski R, Szalewicz K (2003) J Chem Phys 119:8379
192. Jankowski P, Szalewicz K (1998) J Chem Phys 108:3554
193. Mc Kellar ARW (1990) J Chem Phys 93:18
194. Mc Kellar ARW (1991) Chem Phys Lett 186:58
195. Mc Kellar ARW (1998) J Chem Phys 108:1811
196. Mc Kellar ARW (2000) J Chem Phys 112:9282
197. Jankowski P, Szalewicz K (2005) J Chem Phys 123:104301
198. Sloan ED Jr (1998) Clathrate Hydrates of Natural Gases, 2nd edn. Marcel Dekker,
New York
199. Akin-Ojo O, Szalewicz K (2005) J Chem Phys, in press
200. Szalewicz K, Hydrogen bond (2002) In: Meyers R et al. (eds), Encyclopedia of Phys-
ical Science and Technology, third edition, vol 7. Academic Press, San Diego, CA,
p 505–538
201. Bukowski R, Sadlej J, Jeziorski B, Jankowski P, Szalewicz K, Kucharski SA, Williams
HL, Rice BS (1999) J Chem Phys 110:3785
202. Millot C, Stone AJ (1992) Mol Phys 77:439
203. Millot C, Soetens JC, Costa MTCM, Hodges MP, Stone AJ (1998) J Phys Chem 102:754
204. Groenenboom GC, Wormer PES, van der Avoird A, Mas EM, Bukowski R, Szalewicz
K (2000) J Chem Phys 113:6702
Intermolecular Interactions 117

205. Smit MJ, Groenenboom GC, Wormer PES, van der Avoird A, Bukowski R, Szalewicz
K (2001) J Phys Chem A 105:6212
206. Murdachaew G et al. to be published
207. Fermi E, Amaldi G (1934) Mem Accad Italia 6:117
208. Perdew JP, Burke K, Ernzerhof M (1996) Phys Rev Lett 77:3865
209. Adamo C, Barone V (1999) J Chem Phys 110:6158
210. Becke AD (1997) J Chem Phys 107:8554
211. Wilson PJ, Bradley TJ, Tozer DJ (2001) J Chem Phys 115:9233
212. Korona T, Moszyński R, Jeziorski B (2002) Mol Phys 100:1723
213. Hesselmann A, Jansen G (2003) Phys Chem Chem Phys 5:5010
214. Brunsteiner M, Price SL (2001) Cryst Growth Des 1:447
215. Gavezotti A (2002) Mod Simul Mat Sci Eng 10:R1
216. Price SL (2004) Cryst Eng Comm 6:344
217. Motherwell WDS, Ammon HL, Dunitz JD, Dzyabchenko A, Erk P, Gavezzotti A,
Hofmann DWM, Leusen FJJ, Lommerse JPM, Mooij WTM, Price SL, Scheraga H,
Schweizer B, Schmidt MU, van Eijck BP, Verwer P, Williams DE (2002) Acta Cryst
B 58:647
218. Byrd EFC, Scuseria GE, Chabalowski CF (2004) J Phys Chem B 108:13100
219. Tsuzuki S, Lüthi HP (2001) J Chem Phys 114:3949
220. Tsuzuki S, Honda K, Mikami M, Tanabe K (2002) J Am Chem Soc 124:104
221. Sinnokrot MO, Valeev EF, Sherrill CD (2002) J Am Chem Soc 124:10887
222. Dion M, Rydberg H, Schröder E, Langreth DC, Lundqvist BI (2004) Phys Rev Lett
92:246401
223. von Lilienfeld OA, Tavernelli I, Rothlisberger U, Sebastiani D (2004) Phys Rev Lett
93:153004
224. Bukowski R, Szalewicz K, Chabalowski CF (1999) J Phys Chem A 103:7322
225. Sponer J, Leszczynski J, Hobza P (2001) Biopolymers 61:3
226. Hobza P, Sponer J (1999) Chem Rev 91:3247
227. Karplus M (2003) Biopolymers 68:350
228. Desfrancois C, Carles S, Schermann JP (2000) Chem Rev 100:3943
229. Kollman P, Caldwell JW, Ross WS, Pearlman DA, Case DA, DeBolt S, Cheatham TE
III, Ferguson D, Siebel G (1998) Encyclopedia of Computational Chemistry. Wiley,
Chichester, UK, p 11
230. MacKerell AD Jr, Brooks B III, Nilsson L, Roux B, Won Y, Karplus M (1998) Encyclo-
pedia of Computational Chemistry. Wiley, Chichester, UK, p 271
231. van Gunsteren WF, Daura X, Mark AE (1998) Encyclopedia of Computational Chem-
istry. Wiley, Chichester, UK, p 1211
232. Volkov A, Koritsanszky T, Coppens P (2004) Chem Phys Lett 391:170
233. Hansen NK, Coppens P (1978) Acta Cryst A34:909
234. Coppens P (1997) X-ray Charge Densities and Chemical Bonding. Oxford University
Press, New York
235. Koritsanszky T, Volkov A, Coppens P (2004) Chem Phys Lett 391:170
236. Volkov A, Li X, Koritsanszky T, Coppens P (2004) J Phys Chem A 108:4283
237. Spackman MA (1986) J Chem Phys 85:6579
238. Spackman MA (1986) J Chem Phys 85:6587
239. Spackman MA (1987) J Phys Chem 91:3179
240. Korona T, Moszyński R, Thibault F, Launay JM, Bussery-Honvault B, Boissoles J,
Wormer PES (2001) J Chem Phys 115:3074