Intruder States in Multireference Perturbation Theory:

The Ground State of Manganese Dimer

CRISTOPHER CAMACHO,1 HENRYK A. WITEK,1 SHIGEYOSHI YAMAMOTO2

1
Institute of Molecular Science and Department of Applied Chemistry, National Chiao Tung
University, 1001 Ta-Hsueh Road, Hsinchu 30010, Taiwan
2
Faculty of Liberal Arts, Chukyo University, 101-2 Yagoto-Honmachi, Showa-ku, Nagoya, Aichi
466-8666, Japan

Received 10 May 2008; Accepted 13 June 2008

DOI 10.1002/jcc.21074
Published online 5 August 2008 in Wiley InterScience (www.interscience.wiley.com).

Abstract: A detailed analysis of a severe intruder state problem in the multistate multireference perturbation theory
(MS-MRPT) calculations on the ground state of manganese dimer is presented. An enormous number of detected intruder
states (>5000) do not permit finding even an approximate shape of the X 1
g+ potential energy curve. The intruder states
are explicitly demonstrated to originate from quasidegeneracies in the zeroth-order Hamiltonian spectrum. The electronic
configurations responsible for appearance of the quasidegeneracies are identified as single and double excitations from the
active orbitals to the external orbitals. It is shown that the quasidegeneracy problem can be completely eliminated using shift
techniques despite of its severity. The resultant curves are smooth and continuous. Unfortunately, strong dependence of
the spectroscopic parameters of the X 1
g+ state on the shift parameter is observed. This finding rises serious controversies
regarding validity of employing shift techniques for solving the intruder state problem in MS-MRPT. Various alternative
approaches of removing intruder states (e.g., modification of the basis set or changing the active space) are tested. None
of these conventional techniques is able to fully avoid the quasidegeneracies. We believe that the MS-MRPT calculations
on the three lowest Ag states of manganese dimer constitute a perfect benchmark case for studying the behavior of MRPT
in extreme situations.
© 2008 Wiley Periodicals, Inc. J Comput Chem 30: 468–478, 2009

Key words: perturbation theory; intruder states; MCQDPT; CASPT2

Introduction for dynamical correlation using the multireference configuration

interaction (MRCI) method is prohibitively expensive; preliminary
Theoretical determination of potential energy curves for metal MRCI calculations yield repulsive curves.∗ The potential energy
dimers is very challenging because of a large number of low- curves determined1, 2 using second-order multireference perturba-
lying electronic states and highly multiconfigurational character of tion theory (MRPT) predict that manganese dimer is a weakly bound
the corresponding wave functions. This statement is particularly van der Waals molecule and that the ground state of Mn2 is 1
g+ .
true for transition metals, lanthanides, and actinides because of These theoretical predictions agree well with experimental findings.
the existence of many atomic terms arising from not fully occu-
pied atomic d and f orbitals. A molecule that possess remarkable
computational difficulties is the manganese dimer. Until recently,1 Additional Supporting Information may be found in the online version of
not only the shape of the potential energy curve was not known, this article.
but also the multiplicity or angular momentum of the ground Correspondence to: H. Witek; e-mail: [email protected]
state was a subject of controversy.1–17 This situation was mainly Contract/grant sponsor: National Science Council; contract/grant number:
caused by a truly multideterminantal character of the ground state NSC 96-2113-M-009-022-MY3
wave function and the dispersion being the primary force binding Contract/grant sponsor: Ministry of Education of Taiwan (MOE-ATU
the two manganese atoms. Many DFT calculations were reported Project)
on Mn2 ,6–16 but they led to diametrically different conclusions ∗ Our preliminary MRCI/CASSCF calculations used the (10e,10o) active
about the equilibrium bond length and the electronic character of space. For an incomplete MRCI/MCSCF model space, 1500 leading CSFs
the ground state, depending on the functional and the approach have been selected from the (12o,14e) active space. Both sets of calculations
used. The complete active space self-consistent field (CASSCF) yield repulsive curves. Larger MRCI calculations have not been feasible in
method predicts that the ground state PES is unbound.4 Accounting our computing environment (8CPUs machine with 32 GB RAM).

© 2008 Wiley Periodicals, Inc.

 MS-MRPT Calculations on the Ground State of Manganese Dimer 469

At the moment, MRPT is the only way of obtaining physically Intruder states are observed in second-order MRPT mainly in cal-
meaningful estimates of the ground state potential energy surface. culations for high-lying electronic states.18, 19, 40 It is rather unusual
In the context of the discussion earlier, it is particularly interest- to observe zeroth-order spectrum quasidegeneracies already for the
ing that the reported1 PES calculations for the low-lying states of ground state.19, 21 Therefore, we have decided to perform a detailed
manganese dimer using second-order multistate multireference per- study of the previously reported1 multistate second-order multi-
turbation theory (MS-MRPT) suffer from the intruder state problem, configurational quasidegenerate perturbation theory (MCQDPT)
and it was necessary to employ the intruder state avoidance (ISA) potential energy curve for the ground state of the manganese dimer
technique18 to ensure smooth character of the calculated curves. To to elucidate the character and the origin of the intruder states and to
make the reader more familiar with the intruder state problem, let us test various standard techniques of their removal (change of active
recall the second-order energy correction in Rayleigh-Schrödinger space, change of basis set, different number of coupled states in the
perturbation theory given by MCQDPT formalism, and various schemes of orbital canonicaliza-
tion). Another important aspect studied here is the dependence of

α|Ĥ|k2 the calculated potential energy curves and the spectroscopic param-
Eα(2) = , (1) eters of Mn2 on the value of the shift parameter used to remove
k Eα(0) − Ek(0) intruder states. To show that the reported problems are not only spe-
cific to the MCQDPT method, additional calculations are performed
where |α is a reference wave function, |k denotes all singly with different variants of MRPT: fully- and partially-contracted
and doubly excited intermediate states, and Eα(0) and Ek(0) are second-order complete active space perturbation theory (CASPT2).
the zeroth-order energies of the reference and intermediate states, The results of these investigations—that by far have exceeded our
respectively. Intruder states are defined as those of the intermedi- expectations—are presented here.
ate states that have their zeroth-order energies quasidegenerate with
the zeroth-order energy of the reference state, Eα(0) ≈ Ek(0) . Pres-
ence of such states results in small energy denominators in eq. (1)
and, consequently, leads to unphysically large—either positive or Computational Details
negative—value of Eα(2) . In PES calculations, intruder states usually
manifest themselves as discontinuities. Because the character of the The 1
g+ ground state potential energy curve of Mn2 has
discontinuities is usually very local, they are frequently overlooked been studied in details using MCQDPT41 implemented in the
owing to large separations between the points defining PES. When gamess22 quantum chemistry package. More than 4000 single-
the number of such singularities is considerable, or when a large point MCQDPT energy calculations have been performed to pro-
number of points is requested, the possibility of their accidental vide a consistent description of the studied PES. All-electron
encountering can be large. Consequently the calculated PES is not (18s15p8d4f 2g)/[7s6p4d4f 2g] basis set, (14e, 12o) complete
smooth and displays peculiar local features. We believe that this is active space (14 active electrons in 12 active orbitals originating
a serious problem of MRPT because in practice it leads to reduc- from the 3d and 4s AOs), and other computational details have
tion of its popularity. Many researchers—especially those not very been adopted from the previous study by Yamamoto et al.1 The
familiar with mathematical details of perturbation theory—after three lowest roots of symmetry Ag (corresponding to the two lowest
1 +
applying MRPT to their problems and obtaining nonsmooth PES,
g states and a single component of the 1 g state) are used to
feel discouraged and decide to switch to a different computational construct the effective MCQDPT Hamiltonian. Two orbital canon-
method. Various techniques have been developed18–21 to overcome icalization schemes are tested: state-averaged (for the three studied
the intruder state problem. Most of them employ a real or imaginary states) and state-specific (only for the ground state). The 1s, 2s,
shift operator that modifies the zeroth-order energies of the refer- and 2p orbitals of each manganese atom are not correlated. Anal-
ence or intermediate states and leads to a modified second-order ogous calculations are performed with two alternative versions of
energy correction given by MRPT, the CASPT2 programs implemented in the MOLCAS24 and
MOLPRO23 quantum chemistry programs. In the MOLPRO version

α|Ĥ|k2 of CASPT2 serious computational problems occur if the number
Eα(2) = , (2) of doubly-occupied and active electrons approaches 30 (as it is in
k Eα(0) − Ek(0) +  our case). Therefore, the CASPT2/MOLPRO calculations have not
included any of the doubly occupied orbitals in the perturbation
where  is a shift parameter that may depend on k. Intruder state treatment. The following shift techniques have been tested for their
removal shift techniques are available in most of quantum chemistry ability of removing intruder states: the intruder state avoidance (ISA)
packages22–24 and are conventionally used to ensure smooth char- technique,18 the real shift technique of Roos and Andersson,21 and
acter of the calculated potential energy surfaces.25–39 It is important the imaginary shift technique of Forsberg and Malmqvist.19 Simul-
to stress here that the shift techniques are not a panacea for all prob- taneously, various traditional approaches of avoiding the intruder
lems originating from a selection of poor partitioning in MRPT. state problem are tested; more detailed information is given in the
It may happen that using these techniques may lead to neglect next Section. Exemplary GAMESS, MOLPRO, and MOLCAS input
of substantial portion of dynamical correlation or skipping some files for the CASSCF and MRPT calculations are given in Auxiliary
potentially important intermediate states in the perturbation expan- Materials to facilitate reproducing our results. Numerical data used
sion that otherwise would interact strongly with the reference wave to prepare the presented plots are available in electronic form in
function. Auxiliary Materials as GRACE input files.

Journal of Computational Chemistry DOI 10.1002/jcc

470 Camacho, Witek, and Yamamoto • Vol. 30, No. 3 • Journal of Computational Chemistry

Results and Discussion ground state PES and the corresponding spectroscopic parameters
depend strongly on the employed value of the shift parameter. These
MCQDPT Potential Energy Curves findings rise serious controversy regarding mathematical validity of
applying the shift techniques for solving the intruder state problem
Previously reported1 MCQDPT potential energy curve for the low-
in MRPT. We discuss this phenomenon further in Section “Failure
est 1
g+ state of Mn2 was apparently smooth and continuous. of Shift Techniques”.
However, the regular and uniform character of this curve could only
be recovered after employing the ISA technique18 in the MCQDPT CASPT2 Potential Energy Curves
calculations. The original MCQDPT PES calculated without the
ISA technique shows numerous discontinuities related to quaside- To verify that the reported intruder state problem is not a specific
generacies in the zeroth-order MCQDPT Hamiltonian. The extent feature of the MCQDPT method, we have performed analogous
and magnitude of distortions introduced in the MCQDPT PES by calculations using two alternative versions of multistate MRPT: the
the intruder states have motivated us to present here a detailed, partially-contracted CASPT2 method implemented in the MOL-
spatially-resolved study of this curve. PRO23 package and the fully-contracted CASPT2 method imple-
Diagonal elements of the effective MCQDPT Hamiltonian cor- mented in the MOLCAS24 package. The calculations have been
responding to the two lowest 1
g+ states of manganese dimer are performed using a standard canonical definition of the Fock oper-
shown in Figure 1 as a function of internuclear separation. Analo- ator without additional shift techniques; analogous calculations
gous data for the distance-dependent coupling between these states, employing shift techniques are discussed in Section “Failure of
which corresponds to the only nonvanishing off-diagonal elements Shift Techniques”. No solution of the CASPT2 equations could
of the 3 × 3 MCQDPT effective Hamiltonian, is shown in Figure be obtained with the CASPT2/MOLPRO program at any of the
2. The upper panels of both plots give an overview of the computed selected 27 test geometries between 1.9 and 5.4 Å. The norm of
curves between 1.9 and 4.0 Å. Each curve contains an enormous the CASPT2 wave function grows unlimitedly clearly indicating
number of discontinuities—estimated to be larger than 5000—at presence of intruder states. The CASPT2 potential energy curves
distances shorter than 3.2 Å. The abundance of singularities and obtained with the CASPT2/MOLCAS program are shown in the
the magnitude of changes introduced in their vicinity do not permit lower panel of Figure 4. Numerous intruder states are detected—
finding even an approximate shape of the computed curves. This is especially for the higher two states—but it is clear that the intruder
best visible in the middle and bottom panels of Figures 1 and 2 that state problem in CASPT2 is far less serious than in MCQDPT (a
show two subsequent magnifications of a small segment of the upper direct comparison of both sets of curves is somewhat impeded by
panels. The curves in Figures 1 and 2 may seem erratic at the first the fact that the MCQDPT curves are plotted using more than
sight, but the performed magnifications show that in fact the curves 4000 single point energies, while for CASPT2, only about 450
energies are computed). Less severe problems with intruder states
are locally smooth and show a typical up-down branching18 around
in CASPT2 are easily understood if one takes into account much
every singular geometry. It must be stressed that even at nonsingular
smaller number of intermediate states used to construct the CASPT2
geometries (e.g., 2.3104 Å) the error introduced by intruder states
first-order wave function. The minimum of the CASPT2/MOLCAS
can be quite substantial, which effectively prevents guessing even
ground state PES is located approximately at 3.2 Å; a number of
an approximate shape of the curves.
intruders present around this geometry prohibit determining other
The MCQDPT potential energy curves are obtained by diag-
spectroscopic parameters. Similarly to MCQDPT, employing shift
onalizing the 3 × 3 MCQDPT effective Hamiltonian for every
techniques allows for obtaining smooth potential energy curves with
considered internuclear separation. It is clear that this procedure
both versions of CASPT2. (See Figs. 4 and 6.) Unfortunately, again
cannot be a remedy for the intruder states present in the diag-
the shape of the shifted CASPT2 potential energy curves and the
onal and off-diagonal elements of the effective Hamiltonian. As
computed spectroscopic parameters depend strongly on the value
expected, the calculated MCQDPT curves display an abundance of
of the shift parameter; the character of this dependence is discussed
singularities that completely prevent one from finding the actual
in details in Section “Failure of Shift Techniques”.
shape of the potential energy curves. A fragment of the calculated
MCQDPT potential energy curves between 2.3071 and 2.3284 Å is Analysis of Intruder States in the MCQDPT Calculations
shown in Figure 3. It is amazing that only the presented small seg-
ment (approximately 0.02 Å) of the total PESs contains more than The first-order interacting space (FOCI) used to account for dynami-
35 singularities! This example shows that the multistate (12o,14e) cal correlation in the MCQDPT method is spanned by all singly and
MCQDPT calculations without the shift corrections are completely doubly excited CSFs with respect to the CASSCF configuration
meaningless for the lowest 1
g+ state of Mn2 . In a practical compu- state functions. These excited CSFs—often referred to as interme-
tational scheme, usually only a small number of probe geometries diate states—can be classified into eight different groups, depending
(say 30) are used to determine the shape of the whole PES. In the on the number of excited electrons and their original and final loca-
case studied by us such an approach inevitably would lead to an tions.42, 43 Dividing the molecular orbitals into the inactive (I), active
almost random shape of the calculated curve. After employing the (A), and external (E) subspaces allows for symbolic representation
ISA shift technique with the shift parameter equal to 0.02, it is pos- of these eight classes as: I→A, II→AA (internal excitations), I→E,
sible to completely remove all the discontinuities from all the three A→E, II→AE (semiinternal excitations), II→EE, IA→EE, and
studied curves (see the upper panel of Fig. 4). Unfortunately, a closer AA→EE (external excitation). Here, the notation II→AE should
inspection of the ISA-MCQDPT curves (or more general: shifted be understood as an excitation of two inactive electrons into sin-
MRPT curves) shows a very undesirable feature. The shape of the gle active and single external orbital. Number of excited electronic

Journal of Computational Chemistry DOI 10.1002/jcc

 MS-MRPT Calculations on the Ground State of Manganese Dimer 471

Figure 1. Distance-dependent diagonal elements of the MCQDPT Figure 2. Distance-dependent off-diagonal elements of the MCQDPT
(2) (2)
effective Hamiltonian Ĥeff for the ground and the first excited 1
g+ effective Hamiltonian Ĥeff between the ground and the first excited 1
g+
states of Mn2 . states of Mn2 .

Figure 3. A fragment of the MCQDPT potential energy curves for the three lowest Ag roots.

Journal of Computational Chemistry DOI 10.1002/jcc

472 Camacho, Witek, and Yamamoto • Vol. 30, No. 3 • Journal of Computational Chemistry

configurations in each class is given in Table 1. It is natural to ask Table 1. Total Number of Electronic Configurations with the Ag Symmetry
which of those excitation classes are responsible for the occur- for the Eight Possible Classes of FOCI Excitations. (For Details, See Text.)
rence of intruder states in our MCQDPT calculations. To answer The Original CASSCF Space Corresponds to 10 Core, 8 Inactive, 12
this question, we plot in Figure 5 the zeroth-order energies Ek(0) Active, and 152 External Orbitals and 14 Active Electrons.
of all the intermediate states together with the distance-dependent
Subspace Number of configurations
zeroth-order energy Eα(0) of the reference state as a function of inter-
nuclear separation. The zeroth-order energies are given separately A→E 1, 327, 256
for each of the eight classes of excitations. For six out of the eight AA→EE 107, 367, 376
classes, Ek(0) are well separated from Eα(0) . However, for two classes I→A 43, 760
of excitations, A→E and AA→EE, Eα(0) is completely dipped in the I→E 8, 877, 904
FOCI section of the zeroth-order Hamiltonian spectrum. For A→E, IA→EE 809, 420, 160
the number of detected quasidegeneracies at various geometries is II→AA 129, 028
equal to 3486, which accounts for approximately 0.3% of the total II→AE 29, 633, 136
II→EE 3, 050, 580, 208
number of the A→E excited electron configurations. For AA→EE,
the number of detected quasidegeneracies at various geometries is
equal to 1500, which accounts for approximately 0.001% of the total
number of the AA→EE excited electron configurations. Note that
the actual number of intermediate CSFs responsible for appearance configuration may give rise to a larger number of CSFs via differ-
of intruder states is larger, because a single open-shell electronic ent spin coupling schemes. However, all such CSFs would manifest
themselves as a single intruder state in PES because they share the
same value of the zeroth-order energy.
A closer inspection of both excitation classes shows that the
external orbitals responsible for the quasidegeneracies always cor-
respond to the lowest virtual orbitals in each irrep. At short distances,
the number of such low-lying external orbitals can be considerable;
the detected intruder states are caused by excitations to MOs origi-
nating from the atomic 4p, 5p, and 5s orbitals. For longer distances,
the observed quasidegeneracies are caused only by the atomic 4p
orbitals. This analysis shows that the standard method of avoiding
intruder states by shifting the low-lying external orbitals to the active
space is not applicable here owing to prohibitively large dimension
of the resulting CAS problem.
It is important to mention that even if the zeroth-order energies of
some of the intermediate states do not cross with that of the reference
state, Eα(0) − Ek(0) can still be very small in some regions of PES. In
such situations, the potential energy curve can exhibit strange local
features (bubbles).† For Mn2 , even if such peculiarities are present
in the potential energy curve, they are effectively masked by the
enormous number of singularities.

Failure of Shift Techniques

As can be seen from Figure 4, using shift techniques implies

systematic shift of the calculated MRPT energies upward, which
can be explained by partial cancellation of the dynamical corre-
lation energy by larger value of denominators in the perturbation
series. Even though this effect is rather small,‡ it may have
strong impact on calculated potential energy curves because the
amount of the dynamical energy cancellation is different at various
geometries. This behavior seems to have serious practical con-
sequences. The calculated spectroscopic parameters (equilibrium

Figure 4. Comparison of the MCQDPT and ISA-MCQDPT (upper † For an example of such a feature, see the curve for the 23  state of AgH at

panel) and CASPT2 and IS-CASPT2 (lower panel) potential energy 3.9 Å given in ref. 18.
curves for the X 1
g+ (), 11
g+ (
), and 11 g () states of Mn2 . The ‡ At 4.0 Å, i.e., at a distance not affected by intruder states, the ISA-MCQDPT

value of the ISA shift parameter is 0.02. The value of imaginary shift in method accounts for 99.4, 99.3, and 99.3% of the MCQDPT dynamical
the IS-CASPT2 calculations is 0.3. correlation energy for X 1
g+ , 11
g+ , and 11 g , respectively.

Journal of Computational Chemistry DOI 10.1002/jcc

 MS-MRPT Calculations on the Ground State of Manganese Dimer 473

Figure 5. Zeroth-order MCQDPT energies of the reference state (dark red line) and the intermediate states
(light blue line) for all the possible classes of single and double excitations spanning the FOCI space.

bond lengths, dissociation energies, harmonic vibrational frequen- applicable here. Later we discuss other alternative methods of avoid-
cies, and excitation energies) depend strongly on the value of the ing intruder states in the MRPT calculations. Probably, the simplest
shift parameter used to remove intruder states. In Table 2, we have way is to choose a different set of geometries defining PES in hope
collected various spectroscopic parameters for the ground state of to avoid the singular geometries. Unfortunately, in the three-state
Mn2 computed using the MCQDPT and CASPT2 method with MRPT calculations on Mn2 virtually all geometries are affected by
various shift techniques: the intruder state avoidance (ISA) tech- some—more or less local—intruder states. Another way is to mod-
nique,18 the real shift (RS) technique of Roos and Andersson,21 and ify the one-electron basis set used in the calculations, preferably by
the imaginary shift (IS) technique of Forsberg and Malmqvist.19 removing some of the diffuse functions. Sometimes it is also pos-
Five different shift values are employed for each technique to sible to change the set of MOs used in the MCQDPT calculations
visualize the scale of changes induced by varying the shift parame- by performing appropriate state-averaging in the CASSCF calcula-
ter. The corresponding ISA-MCQDPT, IS-CASPT2/MOLCAS, and tions. Another possibility can be to enlarge/reduce the number of
RS-CASPT2/MOLPRO potential energy curves are shown in Figure states in the effective MCQDPT Hamiltonian, which can modify
6. The position of the minimum on each of the curves—marked both, the definition of the Fock operator and the set of canonical
with a symbol—is systematically shifted toward longer distances MOs. Finally, we mention here probably the most popular solution,
for larger value of the shift parameter. This behavior can be read- i.e., employing other active space to define the CASSCF reference
ily understood if one recollects that the CASSCF potential energy wave function. Later we discuss our results obtained when applying
curve for the ground state of Mn2 is repulsive and that the larger these techniques to the multistate MRPT calculations for the ground
value of shift effectively diminishes the amount of dynamical corre- state of Mn2 .
lation recovered by MRPT; it is clear that for the infinite value of the
shift parameter, the CASSCF curve should be reproduced. Analo- Modification of Atomic Basis Set
gous pattern can be observed for the other spectroscopic parameters
presented in Table 2. Note that in the studied case it is impossible to Our main objective here is to investigate the possibility of modifying
predict unbiased values of spectroscopic parameters for the ground the basis set for the manganese atom in a way that enables avoiding
state of manganese dimer using MRPT with shift techniques. the intruder state problem in the multistate MCQDPT calculations
for the ground state of Mn2 . We request that the basis set modi-
Alternative Methods of Removing Intruder States fications should not change the energy of the CASSCF reference
wave function (or do it in a minimal degree). To this end, we have
It is a common belief in the computational chemistry community gradually kept removing some portion of the atomic basis set and
that the intruder state problem in second-order MRPT can be solved inspected the effect of this operation on the CASSCF and MCQDPT
in a number of ways. The most popular solution is including the curves for the 1
g+ of Mn2 . We have been able to remove the g, f ,
low-lying external orbitals that are responsible for quasidegenera- and the two most diffuse p atomic orbitals without significant alter-
cies in the active space. Another solution is using shift techniques. ation of the CASSCF curve. It has not been possible to modify the s
Unfortunately, as we have shown earlier, these solutions are not and d segments of the basis set in the same manner. The MCQDPT

Journal of Computational Chemistry DOI 10.1002/jcc

474 Camacho, Witek, and Yamamoto • Vol. 30, No. 3 • Journal of Computational Chemistry

Table 2. Spectroscopic Parameters for the Ground State of Mn2 Computed Using MRPT with Different Shift
Techniques (For Details, See Text). Note Strong Dependence of the Presented Results on the Value of the Shift
Parameter.

Evert (X 1
g+ → 11
g+ ) (eV)

Shift parameter re (Å) ωe (cm−1 ) De (eV) At minimum At 4.02 Å

ISA-MCQDPT/GAMESS
0.007 3.11 47.3 0.17 1.05 1.43
0.010 3.18 50.6 0.16 1.11 1.44
0.020a 3.29 53.4 0.14 1.22 1.46
0.050 3.49 48.5 0.10 1.40 1.53
0.100 3.60 41.1 0.07 1.52 1.60
IS-CASPT2/MOLCAS
0.100 3.21 67.6 0.16 0.92 1.26
0.200 3.28 65.8 0.15 1.26 1.47
0.300 3.38 59.9 0.12 1.41 1.57
0.400 3.45 52.2 0.10 1.55 1.67
0.500 3.55 46.0 0.08 1.69 1.77
RS-CASPT2/MOLPRO
0.100 . . .b . . .b . . .b . . .b 2.69
0.200 3.15 60.1 0.22 3.17 3.45
0.300 3.28 71.1 0.15 3.15 3.35
0.400 3.39 61.6 0.11 3.11 3.27
0.500 3.49 53.6 0.09 3.09 3.20
Exp. 3.4c 68,d 76,e 59,f 68f 0.13 ± 0.13g ... ...

a Values reported previously in ref. 1.

b Intruder states do not permit determination of spectroscopic values.
c Ref. 44.
d In xenon matrix, ref. 45.
e In krypton matrix, Ref. 46.
f At two different sites of argon matrix, ref. 46.
g Ref. 47.

potential energy curves for the Mn2 ground state calculated using the number of intruder states present in multistate MCQDPT calcula-
resultant reduced atomic basis sets are shown in Figure 7 together tions. The departure point for constructing these schemes is a set of
with the original (18s15p8d4f 2g)/[7s6p4d4f 2g] MCQDPT curve. state-averaged CASSCF orbitals. The canonical Fock matrix, given
It is clear from the curves presented in Figure 7 that the reduction by
of the atomic basis set is not a general remedy for the intruder state

 
problem. The reduction of the one-particle basis allows for dimin- 1
ishing the number of intruder states, but at the same time leads Fpq = hpq + Drs (pq|rs) − (pr|qs) , (3)
rs
2
to considerable reduction of the portion of dynamical correlation
recovered by MCQDPT. At 4.0 Å, the reduced basis set calculations
account only for the following fraction of the original MCQDPT where D is a one-particle density matrix, is constructed and sub-
correlation energy: 91.6% for the (18s15p8d4f )/[7s6p4d4f ] basis sequently diagonalized in each of the three orbital subspaces
set, 54.7% for the (18s15p8d)/[7s6p4d] basis set, and 53.3% for (doubly-occupied, active, and virtual) yielding a set of canoni-
the (18s13p8d)/[7s4p4d] basis set. For the smallest reduced basis, cal MOs and a set of orbital energies. The first canonicalization
the MCQDPT curve is free from intruders around its minimum and scheme uses the standard definition of the density matrix D com-
therefore it is possible to determine the spectroscopic parameters puted as an average over the three lowest CASSCF states with the
for the X 1
g+ state of Mn2 . They are: re = 3.47 Å, ωe = 52 cm−1 , Ag symmetry. In the second canonicalization scheme D is computed
and De = 0.07 eV. These values are comparable to those obtained only from the ground state CASSCF wave function. The multistate
previously1 with ISA-MCQDPT (re = 3.29 Å, ωe = 54 cm−1 , and MCQDPT potential energy curves obtained with the first canoni-
De = 0.14 eV) and from experiment44–46, 48 (re = 3.4 Å, ωe = 59, calization scheme are shown in the upper panel of Figure 4 and the
68, and 76 cm−1 , and De = 0.02 − 0.15 eV). PESs obtained with the second canonicalization scheme, in Figure
8. Because both sets of curves are heavily deformed by numerous
Modification of Orbital Canonicalization Scheme intruder states, the corresponding ISA-MCQDPT curves with the
shift parameter of 0.02 are given for comparison. It is clear that
Two different orbital canonicalization schemes have been employed the ground state PES obtained with the orbitals canonicalized only
to test how the choice of the set of molecular orbitals influences the for the ground state (Scheme 2) is of better quality, because the

Journal of Computational Chemistry DOI 10.1002/jcc

 MS-MRPT Calculations on the Ground State of Manganese Dimer 475

intruders are present only at short distances. This allows for find-
ing spectroscopic parameters for the ground state PES computed
with MCQDPT. They are: re = 3.31 Å, ωe = 56 cm−1 , and De =
0.13 eV.

Modification of Effective MCQDPT Hamiltonian Dimension

Previously reported MCQDPT calculations1 used the three lowest

Ag roots—X 1
g+ , 11
g+ , and a single component of 11 g —to define
the effective MCQDPT Hamiltonian. Performing state-averaged
CASSCF optimization of orbitals and averaging density matrices
of three states in the canonicalization scheme may affect the num-
ber of intruder states present in the MCQDPT calculations. To
test this dependence, we have performed two alternative sets of
MCQDPT calculations: (a) with the lowest Ag state (X 1
g+ ) and
(b) with the two lowest Ag states (X 1
g+ and 11
g+ ). Figure 9
shows the potential energy curves obtained from these calculations.
Modifying the dimension of the effective MCQDPT Hamiltonian
could not prevent the zeroth-order spectrum quasidegeneracies com-
pletely, but the number of intruder states could be reduced, especially
in case (a). Unfortunately, this MCQDPT curve displays another
problem; the calculated PES is discontinuous at approximately
4.23 Å. The discontinuity is related to a switch of two orbitals
between the active and inactive MO subspaces in the CASSCF cal-
culations. No discontinuity is visible in the CASSCF curve, only
accounting for dynamical correlation allows for noticing this defi-
ciency. A more detailed analysis of this phenomenon is given in
ref. 49. Additional discussion on the multistate MCQDPT calcula-
Figure 6. Potential energy curves for the X 1
g+ state of Mn2 computed tions performed for the lowest ten Ag states is given in Auxiliary
using various variants of MRPT with additional shift techniques. The Materials.
presented curves correspond to different values of the shift parameter.
Minimum of each curve is marked with a symbol. Note strong depen-
Modification of Active Space
dence of the position of the minima on the value of the shift parameter.
[Color figure can be viewed in the online issue, which is available at Probably the most standard way of avoiding the intruder state prob-
www.interscience.wiley.com.] lem in MRPT is an appropriate modification of the active space. Note

Figure 7. MCQDPT potential energy curves for the X 1
g+ state of Mn2 computed with four
4 different basis sets. [Color figure can be viewed in the online issue, which is available at
www.interscience.wiley.com.]

Journal of Computational Chemistry DOI 10.1002/jcc

476 Camacho, Witek, and Yamamoto • Vol. 30, No. 3 • Journal of Computational Chemistry

Figure 8. Comparison of the MCQDPT and ISA-MCQDPT potential energy curves for the X 1
g+ (),
11
g+ (
), and 11 g () states of Mn2 computed with a set of orbitals canonicalized only for the ground
state. Analogous curves computed with orbitals canonicalized over all studied states are shown in the upper
panel of Figure 4. The value of the ISA shift parameter is 0.02.

that in the studied case, including in the active space all low-lying these values can be significantly altered in multistate MCQDPT
virtual orbitals responsible for appearing of intruder states is infea- calculations.
sible since the dimension of the resulting CASSCF problem would
be too large. Here, we test two alternative active spaces applicable Modification of the Zeroth-Order Hamiltonian
to the multistate MRPT calculations for the ground state of man-
ganese dimer. The first of the active spaces, denoted as (12o,12e), is Recently, a new definition of the zeroth-order Hamiltonian has been
obtained by moving the σg4s active orbital to the inactive subspace proposed for the CASPT2 method.50 In principle, it was introduced
4p to diminish systematic errors observed for open shell systems, but
and by moving the σg virtual orbital to the active subspace. The
it was also found to reduce significantly the intruder state prob-
second active space, denoted as (13o,14e), is obtained by moving the
4p lem. In the new zeroth-order Hamiltonian, a real shift—referred to
σg virtual orbital to the active subspace. The MCQDPT potential as the IPEA shift—weighted by a function of the diagonal density
energy curve for the first active space is calculated using two-state matrix elements is used to correct the energies of active orbitals. The
MCQDPT formalism (X 1
g+ and 11
g+ ) and for the second active default value of the IPEA shift in the CASPT2/MOLCAS program
space, using three-state MCQDPT formalism (X 1
g+ , 11
g+ , and a is 0.25 hartree; this value was determined empirically for dissocia-
single component of 11 g ). Both calculated curves, shown in Figure tion energies of 49 diatomic molecules. Applying the IPEA shift to
10, are affected by intruder states. The number of detected singu- the multistate CASPT2/MOLCAS calculations on Mn2 allows for
larities is noticeably smaller than for the original (12o,14e) active significant improvement of the computed potential energy curves.
space, but on the other hand the singularities are present in a larger Intruder states can be observed only at short distances. The ground
segment of the calculated PESs. state spectroscopic parameters determined from these calculations
Several other choices of active space—including the (12o,12e) are: re = 4.23 Å, ωe = 29 cm−1 , and De = 0.06 eV. Unfortunately,
and (13o,14e) CAS studied here—are tested in ref. 49 in con- the computed spectroscopic parameters of Mn2 depend strongly on
junction with the single-state MCQDPT formalism. None of the the value of the IPEA shift suggesting certain arbitrariness of the
employed active spaces allows for avoiding the intruder state prob- values presented earlier. Additional details of the IPEA-CASPT2
lem entirely, however, the observed intruders are present only calculations are presented in Auxiliary Materials.
at short interatomic distances. The computed curves are smooth Summarizing the results obtained in this Section, we state that no
around their minima allowing for determination of spectroscopic standard technique of avoiding intruder states could be fully success-
parameters for the ground state of Mn2 . The equilibrium bond fully employed to solve the intruder state problem in our multistate
lengths are between 3.24 and 3.50 Å, the harmonic vibrational MRPT calculations on the ground state of Mn2 . We have found
frequencies, between 44 and 72 cm−1 , and the dissociation ener- that removing diffused atomic orbitals from the basis set, reducing
gies, between 0.05 and 0.09 eV. Because the coupling between the number of interacting states in the effective MCQDPT Hamil-
the X 1
g+ and 11
g+ states is relatively strong, it is possible that tonian, using the set of orbitals canonicalized only for the ground

Journal of Computational Chemistry DOI 10.1002/jcc

 MS-MRPT Calculations on the Ground State of Manganese Dimer 477

the reported serious intruder state problem is not only specific to the
MCQDPT method, but also it is rather a common problem of MS-
MRPTs. Using standard shift techniques in the multistate MRPT
formalism allows for removing all intruders, but unfortunately, it
also introduces a large arbitrariness in the calculated results. The
shape of the potential energy curve and the values of the spectro-
scopic parameters of the X 1
g+ state of Mn2 are found to depend
strongly on the value of shift parameters. These findings rise serious
controversies regarding validity of employing shift techniques for
solving the intruder state problem in MRPT. A set of various alterna-
tive methods of removing intruder states is tested: removing diffused
atomic orbitals from the basis set, reducing the number of interacting
states in the effective MCQDPT Hamiltonian, using the set of
orbitals canonicalized only for the ground state, including low-lying
virtual orbitals in the active space, and modifying the zeroth-order
Hamiltonian. None of these tested alternative approaches is able to
completely eliminate the intruder state problem, but some of these
approaches could reduce its severity considerably. Probably the most
promising approach is using the zeroth-order Hamiltonian modified
with the IPEA shift.

Figure 9. Potential energy curves for the X 1
g+ state of Mn2 computed
with (a) one-state and (b) two-state (12o,14e) MCQDPT formalism.
[Color figure can be viewed in the online issue, which is available at
www.interscience.wiley.com.]

state, or including low-lying virtual orbitals in the active space helps

to reduce the extent of the intruder state problem, however, none of
these techniques is able to remove it entirely. Probably, the most
promising approach is to use the IPEA shift in the definition of
the modified zeroth-order Hamiltonian. Strong dependence of the
computed spectroscopic parameters on the value of the IPEA shift
suggests that this issue requires further clarification.

Conclusion

A detailed, spatially-resolved analysis of the previously reported1

multistate MCQDPT potential energy curve for the X 1
g+ state of
manganese dimer is presented. We show that the computed PES
suffers from numerous (more than 5000!) discontinuities, which
originate from quasidegeneracies in the zeroth-order Hamiltonian
spectrum. The one- and two-electron excited intermediate states that
are responsible for appearance of these intruder states are identified
and analyzed. These states belong to two classes of excitations: sin- Figure 10. Potential energy curves for the X 1
g+ state of Mn2 com-
gle and double excitations from active orbitals to external orbitals. puted with (a) two-state (12o,12e) MCQDPT and with (b) three-state
Additional calculations performed with other versions of multi- (13o,14e) MCQDPT. [Color figure can be viewed in the online issue,
state MRPT (fully- and partially-contracted CASPT2) show that which is available at www.interscience.wiley.com.]

Journal of Computational Chemistry DOI 10.1002/jcc

478 Camacho, Witek, and Yamamoto • Vol. 30, No. 3 • Journal of Computational Chemistry

We believe that the lowest three Ag states of manganese dimer Berning, A.; Cooper, D. L.; Deegan, M. J. O.; Dobbyn, A. J.; Eckert,
constitute a perfect benchmark case for studying the behavior of F.; Hampel, C.; Hetzer, G.; Lloyd, A. W.; McNicholas, S. J.; Meyer, W.;
MRPT in extreme situations. The problems detected and communi- Mura, M. E.; Nicklass, A.; Palmieri, P.; Pitzer, R.; Schumann, U.; Stoll,
cated in this article can provide an invaluable benchmark framework H.; Stone, A. J.; Tarroni, R.; Thorsteinsson, T. Molpro, version 2006.1,
A Package of Ab initio Programs, 2006. http://www.molpro.net.
for development of reliable and computationally stable MS-MRPT.
24. Karlstrom, G.; Lindh, R.; Malmqvist, P. Å.; Roos, B. O.; Ryde,
We also hope that the discovered strong dependence of the com-
U.; Veryazov, V.; Widmark, P.-O.; Cossi, M.; Schimmelpfennig, B.;
puted potential energy curves and the spectroscopic parameters on Neogrady, P.; Seijo L. Comput Mater Sci 2003, 28, 222.
the value of the shift parameter will issue a warning for all users of 25. Fedorov, D. G.; Nakajima, T.; Hirao, K. J Chem Phys 2003, 118,
MRPT that use the shift techniques in a completely automated fash- 4970.
ion to solve the intruder state problem.25–39 We advocate here using 26. Iuchi, S.; Morita, A.; Kato, S. J Chem Phys 2004, 121, 8446.
two different shift parameters and comparing the resulting data. If 27. Patchkovskii, S. Phys Chem Chem Phys 2006, 8, 926.
the results are similar, it is safe to assume that the employed shift 28. Sliznev, V. V.; Vogt, N.; Vogt, J. J Mol Struct 2006, 780–781, 247.
technique does not have large influence on the calculated results. 29. Gopakumar, G.; Lievens, P.; Nguyen, M. T. J Chem Phys 2006, 124,
214312.
30. Doyle, R. J.; Campo, R. D.; Taylor, P. R.; Mackenzie, S. R. J Chem Phys
References 2004, 121, 835.
31. Pierloot, K.; Vancoillie, S. J Chem Phys 2006, 125, 124303.
1. Yamamoto, S.; Tatewaki, H.; Moriyama, H.; Nakano, H. J Chem Phys 32. Oliver, T. A. A.; Taylor, P. R.; Doyle, R. J.; Mackenzie, S. R. J Chem
2006, 124, 124302. Phys 2007, 127, 024301.
2. Wang, B.; Chen, Z. Chem Phys Lett 2004, 387, 395. 33. Olaso-González, G.; Roca-Sanjuán, D.; Serrano-Andrés, L.; Merchán,
3. Wang, B.; Chen, Z. J Chem Phys 2005, 123, 134306. M. J Chem Phys 2006, 125, 231102.
4. Bauschlicher, C. W. Jr. Chem Phys Lett 1989, 156, 95. 34. Zou, W.; Liu, W. J Chem Phys 2006, 124, 154312.
5. Nesbet, R. K. Phys Rev A 1964, 135, 460. 35. Chen, H.; Li, S. J Chem Phys 2006, 124, 154315.
6. Harris, J.; Jones, R. O. J Chem Phys 1979, 70, 830. 36. Pierloot, K.; van Besien, E. J Chem Phys 2005, 123, 204309.
7. Pederson, M. R.; Reuse, F.; Khanna, S. N. Phys Rev B 1998, 58, 5632. 37. de Graaf, C.; Hozoi, L.; Broer, R. J Chem Phys 2004,
8. Desmarais, N.; Reuse, F. A.; Khanna, S. N. J Chem Phys 2000, 112, 120, 961.
5576. 38. Pierloot, K.; Delabie, A.; Groothaert, M. H.; Schoonheydt, R. A. Phys
9. Yanagisawa, S.; Tsuneda, T.; Hirao, K. J Chem Phys 2000, 112, 545. Chem Chem Phys 2001, 3, 2174.
10. Barden, C. J.; Rienstra-Kiracofe, J. C.; Schaefer, H. F. III, J Chem Phys 39. Rubio, M.; Roos, B. O.; Serrano-Andrés, L.; Merchán, M. J Chem Phys
2000, 113, 690. 1999, 110, 7202.
11. Bobadova-Parvanova, P.; Jackson, K. A.; Srinivas, S.; Horoi, M. J Chem 40. Choe, Y.-K.; Witek, H. A.; Finley, J. P.; Hirao, K. J Chem Phys 2001,
Phys 2005, 122, 014310. 114, 3913.
12. Valiev, M.; Bylaska, E. J.; Weare, J. H. J Chem Phys 2003, 119, 5955. 41. Nakano, H. J Chem Phys 1993, 99, 7983.
13. Gutsev, G.; Bauschlicher, C. J Phys Chem A 2003, 107, 4755. 42. Andersson, K.; Malmqvist, P. Å.; Roos, B. O.; Sadlej, A. J.; Wolinski,
14. Nayak, S. K.; Jena, P. Chem Phys Lett 1998, 289, 473. K. J Phys Chem 1990, 94, 5483.
15. Fujima, N.; Yamaguchi, T. J Phys Soc Jpn 1995, 64, 1251. 43. Andersson, K.; Malmqvist, P. Å.; Roos, B. O. J Chem Phys 1992, 96,
16. Yamanaka, S.; Ukai, T.; Nakata, K.; Takeda, R.; Shoji, M.; Kawakami, 1218.
T.; Takada, T.; Yamaguchi, K. Int J Quantum Chem 2007, 107, 3178. 44. Baumann, C. A.; Van Zee, R. J.; Bhat, S. V.; Weltner, W. Jr. J Chem
17. Yamaguchi, K.; Tsunekawa, T.; Toyoda, Y.; Fueno, T. Chem Phys Lett Phys 1983, 78, 190.
1988, 143, 371. 45. Bier, K. D.; Haslett, T. L.; Kirkwood, A. D.; Moskovits, M. Chem Phys
18. Witek, H. A.; Choe, Y.-K.; Finley, J. P.; Hirao, K. J Comput Chem 2002, Lett 1988, 89, 6.
23, 957. 46. Kirkwood, A. D.; Bier, K. D.; Thompson, J. K.; Haslett, T. L.; Huber,
19. Forsberg, N. Malmqvist, P. Å. Chem Phys Lett 1997, 274, 196. A. S.; Moskovits, M. J Phys Chem 1991, 95, 2644.
20. Khait, Y. G.; Song, J.; Hoffmann, M. R. J Chem Phys 2002, 117, 4133. 47. Kant, A.; Lin, S.; Strauss, B. J Chem Phys 1968, 49, 1983.
21. Roos, B. O.; Andersson, K. Chem Phys Lett 1995, 245, 215. 48. Haslett, T. L.; Moskovits, M.; Weitzman, A. J Mol Spectrosc 1989, 135,
22. Schmidt, M. W.; Baldridge, K. K.; Boatz, J. A.; Elbert, S. T.; Gordon, 259.
M. S.; Jensen, J. H.; Koseki, S.; Matsunaga, N.; Nguyen, K. A.; Su, S.; 49. Camacho, C.; Yamamoto, S.; Witek, H. A. Phys Chem Chem Phys (in
et al. J Comput Chem 1993, 14, 1347. press), DOI:10.1039/B805125A.
23. Werner, H.-J.; Knowles, P. J.; Lindh, R.; Manby, F. R.; Schütz, M.; 50. Ghigo, G.; Roos, B. O.; Malmqvist, P. Å. Chem Phys Lett 2004, 396,
Celani, P.; Korona, T.; Rauhut, G.; Amos, R. D.; Bernhardsson, A.; 142.

Journal of Computational Chemistry DOI 10.1002/jcc