Density Functional
Density Functional
Density Functional
Nicolas Ferré
Michael Filatov
Miquel Huix-Rotllant Editors
Density-
Functional
Methods for
Excited States
368
Topics in Current Chemistry
Editorial Board
H. Bayley, Oxford, UK
K.N. Houk, Los Angeles, CA, USA
G. Hughes, CA, USA
C.A. Hunter, Sheffield, UK
K. Ishihara, Chikusa, Japan
M.J. Krische, Austin, TX, USA
J.-M. Lehn, Strasbourg Cedex, France
R. Luque, C ordoba, Spain
M. Olivucci, Siena, Italy
J.S. Siegel, Tianjin, China
J. Thiem, Hamburg, Germany
M. Venturi, Bologna, Italy
C.-H. Wong, Taipei, Taiwan
H.N.C. Wong, Shatin, Hong Kong
V.W.-W. Yam, Hong Kong, China
S.-L. You, Shanghai, China
Aims and Scope
The series Topics in Current Chemistry presents critical reviews of the present and
future trends in modern chemical research. The scope of coverage includes all areas of
chemical science including the interfaces with related disciplines such as biology,
medicine and materials science.
The goal of each thematic volume is to give the non-specialist reader, whether at the
university or in industry, a comprehensive overview of an area where new insights are
emerging that are of interest to larger scientific audience.
Thus each review within the volume critically surveys one aspect of that topic and
places it within the context of the volume as a whole. The most significant
developments of the last 5 to 10 years should be presented. A description of the
laboratory procedures involved is often useful to the reader. The coverage should not
be exhaustive in data, but should rather be conceptual, concentrating on the
methodological thinking that will allow the non-specialist reader to understand the
information presented.
Discussion of possible future research directions in the area is welcome.
Review articles for the individual volumes are invited by the volume editors.
Density-Functional Methods
for Excited States
With contributions by
C. Adamo M. Barbatti K. Bennett M.E. Casida
R. Crespo-Otero C. Daniel M. Filatov K.J.H. Giesbertz
U. De Giovannini W. Hua M. Huix-Rotllant D. Jacquemin
M. Krykunov A.H. Larsen S. Mukamel A. Nikiforov
Y.C. Park K. Pernal A. Rubio I. Seidu W. Thiel
C.A. Ullrich Z.-h. Yang Y. Zhang T. Ziegler
Editors
Nicolas Ferré Michael Filatov
Institut de Chimie Radicalaire Inst. für Physikalische und Theoretische Chemie
Université d’Aix-Marseille Universität Bonn
Marseille, France Bonn, Germany
Miquel Huix-Rotllant
Institut für Physikalische und Theoretische Chemie
Goethe-Universität Frankfurt am Main
Frankfurt, Germany
It is with great sadness and sorrow that we have learnt about the sudden and
untimely death of our colleague, Prof. Dr. Tom Ziegler. Tom was a bright scientist
and one of the founders and pioneers of density functional theory. Tom has always
supported and encouraged new developments and always took an active part in
advancing and promoting them. Tom’s recent research on constricted variational
density functional theory holds considerable promise for the field of quantum
theoretical modeling of excited electronic states, and one of the latest accounts of
these developments is published in this book. Tom will be sorely missed by his
colleagues and the entire community of quantum theoretical chemists.
ThiS is a FM Blank Page
Preface
The rapidly expanding use of ultrafast laser spectroscopic methods to study the
photodynamics of chemical bonds underlines the importance of accurate theoretical
interpretation and modeling of experimental observations. To match this need there
was recent progress in the development and application of computational methods
based on density functional theory (DFT) designed to describe the excited elec-
tronic states and the related potential energy surfaces (PESs). These developments
are especially valuable as DFT methods enable one to study the properties of
excited states and to obtain on-the-fly the relevant dynamical parameters of large
molecular and condensed phase species occurring in natural as well as artificial
photoactive systems. As Kohn–Sham DFT is strictly formulated for ground elec-
tronic states only [1, 2], excited electronic states are typically accessed through the
use of linear response formalism implemented in time-dependent DFT (TD-DFT)
methodology [3, 4], which currently enjoys a wide popularity among theoretical
photochemists and photophysicists. The widespread use of TD-DFT, however,
revealed certain drawbacks and limitations of the methodology, which are being
constantly addressed through the development of improved computational pro-
tocols, new exchange-correlation density functionals, and conceptually new com-
putational approaches. Besides methodological developments in the domain of
linear response TD-DFT, there is growing activity in the field of development of
alternative time-independent DFT methods as well as the methods going beyond
the paradigm of electronic density and exploring the world of (one-electron) density
matrix functionals. Although a number of excellent reviews of TD-DFT formalism
can be found in the literature [5–9], the rate of new developments seems to outpace
the rate of review publishing. This book attempts to fill the gap by providing a
collection of chapters addressing the most recent developments in the realm of DFT
methodology for the excited electronic states written by leading experts in the field.
The opening chapter (p. 1) of the book gives a broad perspective on linear
response TD-DFT and its formal connection to many-body theory. By exploring
the latter, the authors expand on the possibilities to ameliorate some well-known
deficiencies of currently available TD-DFT methodology, especially with regard to
treatment of double (and, in general, multiple) excitations and proper description of
viii Preface
chemical bond dissociation. The dressed TD-DFT approach and its current and
potential capabilities are discussed in detail. An entirely different approach to
describing excited states in the context of DFT is taken in the following two
chapters, which expand on the use of time-independent methodologies. The
constricted-variational DFT method, presented in the second chapter (p. 61), has
the potential to outperform the currently available linear response TD-DFT when
describing excitations in large conjugated systems or charge transfer excitations,
both of which are notoriously difficult for the standard linear response TD-DFT
formalism. A practically accessible implementation of ensemble DFT formalism,
presented in the next chapter (p. 97), holds great promise for theoretical modeling
of non-adiabatic relaxation processes of excited electronic states, relevant to pho-
tochemistry and photovoltaics, and provides proper description of real and avoided
crossings between the ground and excited electronic states of large molecular
species. To complement these new developments, the fourth chapter (p. 125) gives
a wide perspective on the general background and practical aspects of a novel
quantum theoretical approach to the ground and excited states of electronic systems
– density matrix functional theory (DMFT). Although a younger methodology than
DFT, DMFT has the potential to overtake its counterpart once practically affordable
functionals of the one-body density matrix become available.
Methodological aspects of the theoretical description of excited electronic states
in the condensed phase and open quantum systems in the framework of TD-DFT are
amply discussed in the fifth and sixth chapters of the book. On p. 185 Ullrich and
Yang give a comprehensive survey of currently available exchange-correlation
kernels of TD-DFT, analyze their shortcomings, and outline possible remedies for
the description of excitonic states in condensed phase systems. A comprehensive
and pedagogical review of theoretical approaches, such as complex scaling and
open boundary conditions, for the description of time-dependent phenomena in
open quantum systems, especially with regard to resonance states photoemission
spectroscopy, is given in the chapter by Rubio et al. on p. 219. A contemporary and
encyclopedic presentation of various approaches for theoretical modeling of
nonlinear core and valence X-ray spectra is presented by Mukamel et al. on
p. 273 in the seventh chapter of the book. In this chapter the use of DFT/TD-DFT
methods to address the demands of nonlinear X-ray spectroscopy measurements are
analyzed in depth and the prospect of their use are outlined.
Practical aspects of using TD-DFT for computational description of molecular
electronic spectroscopy are reviewed in the chapter by Jacquemin and Adamo on
p. 347. Special emphasis was put on going beyond the vertical excitation approx-
imation in TD-DFT and including vibronic effects for realistic description of 0–0
transition energies in real-life molecular systems. The use of TD-DFT for compu-
tational modeling of absorption spectroscopy, emission properties, and ultrafast
intersystem crossing processes in transition metal complexes is surveyed by Daniel
(p. 377) in Chap. 8, where special attention is paid to the inclusion of spin-orbit and
vibronic effects in TD-DFT computations. The ability of the DFT/TD-DFT frame-
work to provide a proper description of dynamical effects on the spectroscopic and
photochemical properties of molecular species is analyzed in the chapters by
Preface ix
Nicolas Ferré
Michael Filatov
Miquel Huix-Rotllant
References
xi
xii Contents
Abstract In their famous paper, Kohn and Sham formulated a formally exact
density-functional theory (DFT) for the ground-state energy and density of a system
of N interacting electrons, albeit limited at the time by certain troubling
representability questions. As no practical exact form of the exchange-correlation
(xc) energy functional was known, the xc-functional had to be approximated,
ideally by a local or semilocal functional. Nowadays, however, the realization
that Nature is not always so nearsighted has driven us up Perdew’s Jacob’s ladder
to find increasingly nonlocal density/wavefunction hybrid functionals. Time-
dependent (TD-) DFT is a younger development which allows DFT concepts to
be used to describe the temporal evolution of the density in the presence of a
perturbing field. Linear response (LR) theory then allows spectra and other infor-
mation about excited states to be extracted from TD-DFT. Once again the exact
TD-DFT xc-functional must be approximated in practical calculations and this has
historically been done using the TD-DFT adiabatic approximation (AA) which is to
TD-DFT very similar to what the local density approximation (LDA) is to conven-
tional ground-state DFT. Although some of the recent advances in TD-DFT focus
on what can be done within the AA, others explore ways around the AA. After giving
an overview of DFT, TD-DFT, and LR-TD-DFT, this chapter focuses on many-body
corrections to LR-TD-DFT as one way to build hybrid density-functional/
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Brief Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 Density-Functional Theory (DFT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Time-Dependent (TD-) DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Linear Response (LR-) TD-DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 Many-Body Perturbation Theory (MBPT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.1 Green’s Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Diagram Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3 Dyson’s Equation and the Bethe–Salpeter Equation (BSE) . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.4 Superoperator Equation-of-Motion (EOM) Polarization Propagator (PP) Approach 24
4 Dressed LR-TD-DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.1 Basic Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.2 Practical Details and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3 Brillouin Corrections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5 Effective Exchange-Correlation (xc) Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1 Localizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6 Conclusion and Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Appendix: Order Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
1 Introduction
I have not included chemistry in my list [of the physical sciences] because, though
Dynamical Science is continually reclaiming large tracts of good ground from one side
of Chemistry, Chemistry is extending with still greater rapidity on the other side, into
regions where the dynamics of the present day must put her hand on her mouth. But
Chemistry is a Physical Science. . .
— James Clerk Maxwell, Encyclopaedia Britannica, ca. 1873 [1]
Much has changed since Maxwell first defended chemistry as a physical science.
The physics applied to chemical systems now involves as much, if not more,
quantum mechanics than classical dynamics. However, some things have not
changed. Chemistry still seems to extend too rapidly for first principles modeling
to keep up. Fortunately, density-functional theory (DFT) has established itself as a
computationally simple way to extend ab initio1 accuracy to larger systems than
1
The term ab initio is used here as it is typically used in quantum chemistry. That is, ab initio refers
to first-principles Hartree–Fock-based theory, excluding DFT. In contrast, the term ab initio used
in the solid state physics literature usually encompasses DFT.
MBPT Insights About and Corrections to TD-DFT 3
2 Brief Review
This section reviews a few concepts which in some sense are very old: DFT is about
50 years old, TD-DFT is about 30 years old, and LR-TD-DFT (in the form of the
Casida equations) is about 20 years old. Thus many of the basic concepts are now
well known. However, this section is necessary to define some notation and because
some aspects of these subjects have continued to evolve and so need to be updated.
MBPT Insights About and Corrections to TD-DFT 5
Hohenberg and Kohn [9] and Kohn and Sham [10] defined DFT in the mid-1960s
when they gave formal rigor to earlier work by Thomas, Fermi, Dirac, Slater, and
others. This initial work has been nicely reviewed in well-known texts [11–13] and
so we do not dwell on details here but rather concentrate on what is essential in the
present context. Hartree atomic units (h ¼ me ¼ e ¼ 1) are used throughout unless
otherwise specified.
Kohn and Sham introduced orthonormal auxiliary functions (Kohn–Sham
orbitals) ψi(1) and corresponding occupation numbers ni which allow the density
to be expressed as
X
ρð 1Þ ¼ ni ψ i ð 1Þ 2 ; ð1Þ
i
Here we use a notation where i ¼ ðri ; σ i Þ stands for the space ri and spin σ i
coordinates of electron i, ^t s ¼ ð1=2Þ∇2 is the noninteracting kinetic energy
operator, v is the external potential which represents the attraction
ðð of the electron
to the nuclei as well as any applied electric fields, EH ½ρ ¼ ρð1Þρð2Þ=r 12 d1d2 is
the Hartree (or Coulomb) energy, and Exc[ρ] is the xc-energy which includes
everything not included in the other terms (i.e., exchange, correlation, and the
difference between the interacting and noninteracting kinetic energies). Minimizing
the energy (2) subject to the constraint of orthonormal orbitals gives the Kohn–
Sham orbital equation:
where the Kohn–Sham Hamiltonian, ĥs[ρ](1),ð is the sum of ^t s ð1Þ þ vð1Þ, the
Hartree (or Coulomb) potential vH ½ρð1Þ ¼ ρð2Þ=r 12 d2, and the xc-potential
vxc ½ρð1Þ ¼ δExc ½ρ=δρð1Þ.
An important but subtle point is that the Kohn–Sham equation should be solved
self-consistently with lower energy orbitals filled before higher energy orbitals
(Aufbau principle) as befits a system of noninteracting electrons. If this can be
done with integer occupancy, then the system is said to be noninteracting v-
representable (NVR). Most programs try to enforce NVR, but it now seems likely
that NVR fails for many systems, even in exact Kohn–Sham DFT. The alternative is
to consider fractional occupation within an ensemble formalism. An important
6 M.E. Casida and M. Huix-Rotllant
theorem then states that only the last occupied degenerate orbitals may be fraction-
ally occupied (see, e.g., [12] pp. 55–56). Suitable algorithms are rare, as
maintaining this condition can lead to degenerate orbitals having different occupa-
tion numbers which, in turn, may require minimizing the energy with respect to
unitary transformations within the space spanned by the degenerate occupied
orbitals with different occupation numbers. These points have been previously
discussed in somewhat greater detail in [8]. Most programs show at least an
effective failure of NVR when using approximate functionals, in particular around
regions of strong electron correlation, such as where bonds are being made or
broken (e.g., avoided crossing of the S0 surfaces in Fig. 1) which often shows up as
self-consistent field (SCF) convergence failures.
As no practical exact form of Exc is known, it must be approximated in practice.
In the original papers, Exc should depend only upon the charge density. However
our notation already reflects the modern tendency to allow a spin-dependence in Exc
(spin-DFT). This additional degree of freedom makes it easier to develop improved
density-functional approximations (DFAs). In recent years, this tendency to add
additional functional dependencies into Exc has led to generalized Kohn–Sham
theories corresponding to different levels of what Perdew has referred to as Jacob’s
ladder2 for functionals (Table 1). The LDA and GGA are pure DFAs. Higher levels
no longer fall within the pure DFT formalism [17] and, in particular, are subject to a
different interpretation of orbital energies.
2
“Jacob set out from Beersheba and went on his way towards Harran. He came to a certain place
and stopped there for the night, because the sun had set; and, taking one of the stones there, he
made it a pillow for his head and lay down to sleep. He dreamt that he saw a ladder, which rested
on the ground with its top reaching to heaven, and angels of God were going up and down it.” –
The Bible, Genesis 28:10–13
MBPT Insights About and Corrections to TD-DFT 7
3
This is formalized in mathematical logic theory by G€ odel’s incompleteness theorem which
basically says that there are always more things that are true than can be proven to be true.
8 M.E. Casida and M. Huix-Rotllant
theorems are true for specific cases where they can be proven, but we believe them
to hold more generally and efforts continue to find more general proofs.
This theorem states, with two caveats, that the time-dependent external potential
v(1) is determined up to an arbitrary function of time by the initial wavefunction
Ψ0 ¼ Ψðt0 Þ at some time t0 and by the time-dependent charge density ρ(1). Here we
have enriched our notation to include time, i ¼ ði; ti Þ ¼ ðri ; σ i ; ti Þ. The statement
that the external potential is only determined up to an arbitrary function of time
simply means that the phase of the associated wave function is only determined up
to a spatially-constant time-dependent constant. This is because two external
potentials differing by an additive function of time e v ð1Þ ¼ vð1Þ þ cðt1 Þ lead to
e
associated wave functions Ψ ðtÞ ¼ e iαðtÞ
ΨðtÞ where dαðtÞ=dt ¼ cðtÞ. A conse-
quence of the Runge–Gross theorem is that expectation values of observables Â(t)
are functionals of the initial wavefunction and of the time-dependent charge
density,
A½ρ; Ψ0 ðtÞ ¼ Ψ½ρ; Ψ0 ðtÞA^ ðtÞΨ½ρ; Ψ0 ðtÞ : ð5Þ
The proof of the theorem assumes (caveat 1) that the external potential is
expandable in a Taylor series in time in order to show that the time-dependent
current density determines the time-dependent external potential up to an additive
function of time. The proof then goes on to make a second assumption (caveat 2)
that the external potential goes to zero at large r at least as fast as 1/r in order to
prove that the time-dependent charge density determines the time-dependent cur-
rent density.
h i
e 0 ð1Þψi ð1Þ ¼ i ∂ ψi ð1Þ;
^h ρ; Ψ0 ; Ψ ð6Þ
∂t
where
h i h i
e 0 ð1Þ ¼ ^t s þ vð1Þ þ vH ½ρð1Þ þ vxc ρ; Ψ0 ; Ψ
^h ρ; Ψ0 ; Ψ e 0 ð1Þ: ð7Þ
The proof of the theorem assumes (caveat 1) that the external potential is expand-
able in a Taylor series in time and (caveat 2) that the charge density is expandable in
a Taylor series in time. Work on removing these caveats is ongoing [27–30] ([24]
provides a brief, but dated, summary).
stationary subject to the conditions that δΨðt0 Þ ¼ δΨðt1 Þ ¼ 0 leads to the time-
dependent Schr€ odinger equation H^ ðtÞΨðtÞ ¼ i∂ΨðtÞ=∂t. Runge and Gross initially
suggested that A ¼ A½ρ; Ψ0 and used this to derive a more explicit formula for the
TD-DFT xc-potential as a functional derivative of an xc-action, but this led to
causality problems. A simple explanation and way around these contradictions was
presented by Vignale [31] who noted that, as the time-dependent Schr€odinger
equation is a first-order partial differential equation in time, Ψ(t1) is determined
by Ψ(t0) so that, whereas δΨ(t0) may be imposed, δΨ(t1) may not be imposed.
The proper Frenkel–Dirac–Vignale action principle is then
δA ¼ i Ψðt1 ÞδΨðt1 Þ : ð9Þ
In many cases, the original Frenkel–Dirac action principle gives the same results
as the more sophisticated Frenkel–Dirac–Vignale action principle. Messud
et al. [32] gives one example of where this action principle has been used to derive
an xc-potential within a TD-DFA. Other solutions to the Dirac–Frenkel causality
problem in TD-DFT may also be found in the literature [33–37].
The notation is a bit subtle here: ρt1 ð1Þ is ρð1Þ ¼ ρð1; t1 Þ at a fixed value of time,
meaning that ρt1 ð1Þ is uniquely a function of the space and spin coordinates, albeit at
fixed time t1. The AA approximation has been remarkably successful and effec-
tively defines conventional TD-DFT.
Going beyond the TD-DFT AA is the subject of ongoing work. Defining new
Jacob’s ladders for TD-DFT may be helpful here. The first attempt to do so was the
definition by one of us (MEC) of a “Jacob’s jungle gym” consisting of parallel
Jacob’s ladders for Exc, vxc(1), f xc ð1; 2Þ ¼ δvxc ð1Þ=δρð2Þ, etc. [3]. This permitted
the simultaneous use of different functionals on different ladders on the grounds
that accurate lower derivatives did not necessarily mean accurate higher deri-
vatives. Of course, being able to use a consistent level of approximation across
all ladders could be important for some types of applications (e.g., those involving
analytical derivatives). With this in mind, the authors recently suggested a new
Jacob’s ladder for TD-DFT (Table 2).
MBPT Insights About and Corrections to TD-DFT 11
If the applied field is sufficiently small then we are in the LR regime where we
may neglect the HOT and calculate the dipole polarizability as
αi, j ðωÞ ¼ Δμi ðωÞ=ε j ðωÞ. Electrical absorption spectra may be calculated from
this because of the sum-over-states theorem in optical physics,
X fI
αðωÞ ¼ ; ð13Þ
ω
I6¼0 I
2 ω2
where α ¼ ð1=3Þ αxx þ α yy þ αzz . Here
ωI ¼ E I E 0 ; ð14Þ
2
f I ¼ ωI 0 r I 2 ; ð15Þ
3
4
Remember that h ¼ 1 in the atomic units used here.
12 M.E. Casida and M. Huix-Rotllant
2ω
SðωÞ ¼ ℑαðω þ iηÞ ; ð16Þ
π
πN A e2
εðvÞ ¼ Sð2πvÞ ð17Þ
me cð4πε0 Þlnð10Þ
in SI units.
So far this is fine for calculating spectra but not for assigning and studying
individual states. For that, it is better to take another approach using the
susceptibility
δρð1Þ
χ ð1; 2Þ ¼ ; ð18Þ
δvappl ð2Þ
which describes the response of the density to the applied perturbation vappl,
ð
δρð1Þ ¼ χ ð1; 2Þδvappl ð2Þd2 : ð19Þ
δρð1Þ
χ s ð1; 2Þ ¼ ; ð21Þ
δvs ð2Þ
is known exactly from MBPT. Of course the effective potential is the sum of the
applied potential and the potential produced by the response of the self-consistent
field, vHxc:
ð
δvs ð1Þ ¼ δvappl ð1Þ þ f Hxc ð1; 2Þδρð2Þ d2; ð22Þ
where f Hxc ð1; 2Þ ¼ δvHxc ð1Þ=δρð2Þ is the functional derivative of the Hartree plus
exchange-correlation self-consistent field. Manipulating these equations is
MBPT Insights About and Corrections to TD-DFT 13
χ ¼ χ s þ χ s fHxc χ ; ð24Þ
Equation (23) may be solved iteratively for δρ. Alternatively δρ may be obtained
by solving
1
χ s fHxc δρ ¼ δvappl ; ð26Þ
which typically involves iterative Krylov space techniques because of the large size
of the matrices involved.
This last equation may be manipulated to make the most common form of LR-
TD-DFT used in quantum chemistry [38].5 This is a pseudoeigenvalue problem,
AðωÞ BðωÞ X 1 0 X
¼ω ; ð27Þ
B * ð ωÞ A* ðωÞ Y 0 1 Y
where
Aia, jb ðωÞ ¼ δi, j δa, b εa, i þ ia f Hxc ðωÞ jb
Bia, b j ðωÞ ¼ ia f Hxc ðωÞb j : ð28Þ
Here,
ðð
pq f rs ¼ ψ*p ð1Þψq ð1Þ f ð1; 2Þψ*r ð2Þψs ð2Þ d1d2 ; ð29Þ
5
This equation is not infrequently called the “Casida equation” in the TD-DFT literature (e.g., as
in [24], pp. 145–153.)
14 M.E. Casida and M. Huix-Rotllant
Equation (28) has paired excitation and de-excitation solutions. Its eigenvalues
are (de-)excitation energies, the vectors X and Y providing information about
transition moments. In particular, the oscillator strength, of the transition with
excitation energy ωI may be calculated from XI and YI [38]. When the adiabatic
approximation (AA) to the xc-kernel is made, the A and B matrices become
independent of frequency. As a consequence, the number of solutions is equal to
the number of one-electron excitations, albeit dressed to include electron correla-
tion effects. Allowing the A and B matrices to have a frequency dependence allows
the explicit inclusion of two-electron (and higher) excited states.
The easiest way to understand what is missing in the AA is within the so-called
Tamm–Dancoff approximation (TDA). The usual AA TDA equation,
AX ¼ ωX ; ð31Þ
ðH E0 1ÞC ¼ ωC ; ð32Þ
which includes all excitations of the system, can be put into the form of (31), but
with a frequency-dependent A(ω) matrix. This can be simply done by partitioning
the full CI Hamiltonian into a singles excitations part (A1,1) and multiple-
excitations part (A2þ, 2þ ) as
" #
ACI ACI C1 C1
1, 1 1, 2þ
¼ω ; ð33Þ
ACI
2þ, 1 ACI
2þ, 2þ
C2þ C2þ
provided we can ignore any coupling between the ground state and excited states.
Applying the standard L€owdin–Feshbach partitioning technique to (33) [40],
we obtain
h 1 CI i
ACI
1, 1 þ ACI
1, 2þ ω1 2þ , 2þ A CI
2þ, 2þ A2þ, 1 C1 ¼ ωC1 ; ð34Þ
6
Sometimes we call this the FORTRAN index convention in reference to the default variable
names for integers in that computer language.
MBPT Insights About and Corrections to TD-DFT 15
In the remainder of this chapter we first show how MBPT may be used to derive
expressions for the ACI CI CI
1, 2þ , A2þ, 1 , and A2þ, 2þ blocks and show how this may be used
in the form of dressed TD-DFT to correct the AA. Then we discuss localization of
the terms beyond the AA in order to obtain some insight into the analytic behavior
of the xc-kernel.
This section elaborates on the polarization propagator (PP) approach. As the PP was
originally inspired by the Bethe–Salpeter equation (BSE) and as the BSE often
crops up in articles from the solid-state physics community which are concerned
with both TD-DFT and MBPT [41–47], we try to make the connection between the
PP and BSE approaches as clear as possible. Although the two MBPT approaches
are formally equivalent, differences emerge because the BSE approach emphasizes
the time representation whereas the PP approach emphasizes the frequency repre-
sentation. This can and typically does lead to different approximations. In parti-
cular, it seems to be easier to derive pole structure-conserving approximations
needed for treating two-electron and higher excitations in the frequency represent-
ation than in the time representation. This and prior experience with the PP
approach in the quantum chemistry community [48–53] have led us to favor the
PP approach. We make extensive use of diagrams in order to give an overview of
our manipulations. Whenever possible, more elaborate mathematical manipulations
are relegated to the appendix.
Perhaps the most common and arguably the most basic quantity in MBPT is the
one-electron Green’s function defined by
D n o E
iGð1; 2Þ ¼ 0T ψ ψ{H ð2Þ 0 :
^ H ð1Þ^ ð35Þ
Here, the subscript H indicates that the field operators are understood to be in the
Heisenberg representation. Also T is the usual time-ordering operator, which
includes anticommutation in our case (i.e., for fermions),
n o
T ψ ψ{H ð2Þ ¼ θðt1 t2 Þ^
^ H ð1Þ^ ψ{H ð2Þ θðt2 t1 Þ^
ψH ð1Þ^ ψ{H ð2Þ^
ψH ð1Þ: ð36Þ
D n o E
Gð1; 2; 3; 4Þ ¼ ðiÞ2 0T ψ^ H ð1Þ^
ψH ð2Þ^ ψ{H ð3Þ 0 :
ψ{H ð4Þ^ ð37Þ
The usual MBPT approach to evaluating the susceptibility, χ, uses the fact that it
is the retarded form,
iχ ð1; 2Þ ¼ θðt1 t2 Þ 0½e ρ H ð 2Þ 0 ;
ρ H ð1Þ, e ð38Þ
where
D E
e
ρ H ð 1Þ ¼ ψ ψH ð1Þ 0ψ
^ {H ð1Þ^ ψH ð1Þ0
^ {H ð1Þ^ ð40Þ
is the density fluctuation operator. (See for example [54] pp. 151, 172–175.)
We will also need several generalizations of the susceptibility and the density
fluctuation operator. The first is the particle-hole (ph) propagator [52], which we
chose to write as
iLð1; 2; 3; 4Þ ¼ 0T fe γ ð4; 3Þg0 ;
γ ð1; 2Þe ð41Þ
where
D n o E
e
γ ð1; 2Þ ¼ ψ ψH ð1Þ 0T ψ
^ {H ð2Þ^ ψH ð1Þ 0
^ {H ð2Þ^ ð42Þ
We also need the polarization propagator (PP) which is the two-time quantity,
iΠðD1, 2; 3, 4; t t 0 Þ
n { þ 0 o E
¼ 0 T ψ ^ H ð2t Þ^ ψH ð1tÞ^ ^ H ð4t 0 Þ 0
ψ{H 3t þ ψ ð46Þ
D n o ED n 0 o E
0 T ψ ψH ð1tÞ 0 0T ψ
^ {H ð2tþ Þ^ ^ {H 3t þ ψ ^ H ð4t 0 Þ 0 :
The second term is often dropped in the definition of the PP. It is there to remove
ω ¼ 0 excitations in the Lehmann representation. (See for example pp. 559–560 of
[54].) The retarded version of the PP is the susceptibility describing the response of
the one-electron density matrix,
{
γ ð1; 2; tÞ ¼ 0ψ ψð1tÞ0 ;
^ ð2tÞ^ ð47Þ
δγ ð1; 2; tÞ
Πð1, 2; 3, 4; t t 0 Þ ¼ ; ð48Þ
δwappl ð3, 4; t 0 Þ
or
in matrix form.
where
D E
Πsr, q p ðt t 0 Þ ¼ iθðt t 0 Þ 0r^ {H ðtÞ^
s H ðtÞ^ q {H ðt 0 Þ p^ H ðt 0 Þ0
D E
iθðt 0 tÞ 0q^ {H ðt 0 Þ p^ H ðt 0 Þ^
r {H ðtÞ^s H ðtÞ0 ð52Þ
This makes it clear that the PP is a two time particle-hole propagator which
either propagates forward in time or backward in time. To represent it we introduce
the following rules:
1. Time increases vertically from bottom to top. This is in contrast to a common
convention in the solid-state literature where time increases horizontally from
right to left.
2. A PP is a two time quantity. Each of these twice is indicated by a horizontal
dotted line. This is one type of “event” (representing the creation/destruction of
an excitation).
3. Time-ordered diagrams use directed lines (arrows). Down-going arrows corre-
spond to holes running backward in time, i.e., to occupied orbitals. Up-going
arrows correspond to particles running forward in time, i.e., to unoccupied
orbitals.
At this point, the PP diagrams resemble Fig. 2. Fourier transforming leads us
to the representation shown in Fig. 3. An additional rule has been introduced:
4. A downward ω arrow on the left indicates forward ph-propagation. An upward ω
arrow on the right indicates backward ph-propagation.
MBPT Insights About and Corrections to TD-DFT 19
p q r s
p q r s
4 3 1 2
Fig. 5 Time-unordered
representation PP diagram
Π(ω) =
Fig. 6 Zero-order PP
diagrams
Πsr,qp (ω) = i ω a + i ω a
1
Πai, ai ðωÞ ¼ : ð53Þ
ω þ εi εa
1 1
Πia, ia ðωÞ ¼ ¼ : ð54Þ
ω þ εi εa ω þ εa εi
Fig. 7 First-order time-ordered diagrams Hugenholtz for Π ðωÞ Π s ðωÞ. a–i involve coupling
between the particle-hole space; g, h, m, and n involve coupling between particle-hole space and
particle-particle; i–l couple the particle-hole space with the hole-hole space
ð
1
pqrs ¼ ψ*p ð1Þψ*r ð2Þ ð1 P 12 Þψq ð1Þψs ð2Þd1d2: ð57Þ
r 12
9. To determine the number of loops and hence the overall sign of a diagram in
which electron repulsion integrals are expanded as dots, write each dot as a
dotted line (it does not matter which one of thetwo in Fig. 8 is chosen) and
apply rule 1. The order of indices in each integral rs pq should correspond to
the expanded diagrams. (When Goldstone diagrams are interpreted in this way,
we call them Brandow diagrams.)
22 M.E. Casida and M. Huix-Rotllant
10. An additional factor of 1/2 must be added for each pair of equivalent lines.
These are directed lines whose interchange, in the absence of further labeling,
leaves the Hugenholtz diagram unchanged.
For example, the term corresponding to Fig. 7a is
kaic
Πck, ai ðωÞ ¼
ðω þ εk εc Þðω þ εi εa Þ
akic
¼ : ð58Þ
ðω þ εk εc Þðω þ εi εa Þ
Two of the most basic equations of diagrammatic MBPT are Dyson’s equation for
the one-electron Green’s function and the BSE for the ph-propagator. Both require
the choice of a zero-order picture which we take here to be the exact or approximate
Kohn–Sham system of noninteracting electrons. We denote the zero-order quanti-
ties by the subscript s (for single particle).
Dyson’s equation relates the true one-electron Green’s function G to the zero-
order Green’s function Gs via the (proper) self-energy Σ,
ð
Gð1; 2Þ ¼ Gs ð1; 2Þ þ Gs ð1; 3ÞΣ ð3; 4ÞGð4; 2Þ d3d4; ð59Þ
G ¼ Gs þ Gs ΣG : ð60Þ
Fig. 9 Time-unordered (Feynman and Abrikosov) one-electron Green’s function diagrams: (a)
Dyson’s equation; (b) second-order self-energy quantum chemistry approximation; (c) GW self-
energy solid-state physics approximation
Lð1; 2; 7; 8Þ ¼ Ls ð1; 2; 7; 8Þ
ð
þ Ls ð1; 2; 3; 4ÞΞ Hxc ð3; 4; 5; 6ÞLð5; 6; 7; 8Þd3d4d5d6; ð61Þ
or
L ¼ Ls þ Ls ΞHxc L ; ð62Þ
is the ph-propagator for the zero-order picture (in our case, the exact or approximate
Kohn–Sham fictitious system of noninteracting electrons), and the four-point quan-
tity, ΞHxc, may be deduced from a Feynman diagram expansion as the proper part of
the ph-response function “self-energy”. This is shown diagrammatically in Fig. 10.
Again, the quantum chemical approximations emphasize antisymmetrization of the
electron repulsion integrals which is needed for proper inclusion of double
24 M.E. Casida and M. Huix-Rotllant
Fig. 10 Time-unordered (Feynman and Abrikosov) ph-propagator diagrams: (a) BSE; (b)
second-order self-energy quantum chemistry approximation; (c) GW self-energy solid-state phys-
ics approximation. Note in part (c) that the solid-state physics literature often turns the v and
w wiggly lines at right angles to each other to indicate the same thing that we have indicated here
by adding tab lines
We now concentrate on the PP and show how to obtain a “Casida-like” equation for
excitation energies and transition moments. This does not as yet give us correction
terms to AA LR-TD-DFT but it does give us some important tools to help us build
correction terms. The basic idea in this section is to take the exact or approximate
Kohn–Sham system of independent electrons as the zero-order picture,
^ ð0Þ ¼ ^hKS ;
H ð64Þ
^ ð1Þ ¼ V
H ^ þM
^ xc : ð65Þ
1X { { X
V^ ¼ pq rs p^ r^ s^q^ pr rq p^ { q^ ; ð66Þ
4 pqrs pqr
X
M^ xc ¼ pΣ^ HF
x v ^ xc q p^ { q^ ; ð67Þ
pq
and Σ^ HF
x is the HF exchange operator defined in terms of the occupied Kohn–Sham
orbitals. Heuristically this gives us a series of diagrams which we must resum
to have the proper analytic structure of the exact PP so we can take advantage
of this analytic structure to produce the desired “Casida-like” equation. Rigorously
we actually first begin with some exact equations in the superoperator equation-
of-motion (EOM) formalism to deduce the analytic structure of the PP. This
exact structure is then developed in a perturbation expansion so that we can
perform an order analysis of each of the terms entering into a basic “Casida-like”
equation. As we can see, not every diagram is generated by this procedure, either
because they are not needed or because of approximations which we have chosen
to make.
Our MBPT expansions are in terms of the bare electron repulsion (or more
exactly the “fluctuation potential” – see (66)), rather than the screened interaction
used in solid-state physics [41, 47]. The main advantage of working with the bare
interaction is a balanced treatment of direct and exchange diagrams, which is
especially important for treating two- and higher-electron excitations. Although
we automatically include what the solid state community refers to as vertex effects,
the disadvantage of our approach is that it is likely to break down in solids when
screening becomes important. The specific approach we take is the now well-
established second-order polarization propagator approximation (SOPPA) of Niel-
sen, Jørgensen, and Oddershede [48–51]. The usual presentation of the SOPPA
approach is based upon the superoperator equation-of-motion (EOM) approach
previously used by one of us [58]. However, the SOPPA approach is very similar
in many ways to the second-order algebraic diagrammatic construction [ADC(2)]
approach of Schirmer [52, 53] and we do not hesitate to refer to this approach as
needed (particularly with regard to the inclusion of various diagrammatic contri-
butions). The only thing really new here is the change from a Hartree–Fock to a
Kohn–Sham zero-order picture and the concomitant inclusion of (many) additional
terms. Nevertheless, it is seen that the final working expressions are fairly compact.
Before going into the details of the superoperator EOM approach, let us antici-
pate some of the results by looking at some of the diagrams which emerge from this
analysis. We have seen in (45) that the PP is just the restriction of the ph-propagator
to twice rather than four times. Thus, heuristically, it suffices to take the
ph-propagator diagrams, fix twice, and then take all possible time orderings.
26 M.E. Casida and M. Huix-Rotllant
Fig. 11 Topologically
different first-order time-
unordered Abrikosov
M M
diagrams for the PP
Defining order as the order in the number of times V^ and/or M ^ xc appear, all of the
time-unordered first-order terms are shown in Fig. 11. Fixing twice and restricting
ourselves to an exchange-only theory gives the 14 time-ordered diagrams shown in
Fig. 7. As we can see below in a very precise mathematical way, dangling parts
below or above the horizontal dotted lines correspond respectively to Hugenholtz
diagrams for initial-time and final-time perturbed wavefunctions. (Two other first-
order Goldstone diagrams are found in [52] with the electron repulsion dot above or
below the two dotted lines; however a more detailed analysis shows that these terms
neatly cancel out in the final analysis.) The area between the dotted lines corre-
sponds to time propagation. In this case, there are only one-hole/one-particle
excitations between the two horizontal dotted lines. Our final results are in perfect
agreement with diagrams appearing in the exact exchange (EXX) theory as
obtained by Hirata et al. [59] which are equivalent to the more condensed form
given by G€ orling [60].
Figure 12 shows all 13 second-order time-unordered diagrams. Although this
may not seem to be very many, our procedure generates about 140 time-ordered
Hugenholtz diagrams (and even more Feynman diagrams). A typical time-ordered
Hugenholtz diagram is shown in Fig. 13. The corresponding equation,
MBPT Insights About and Corrections to TD-DFT 27
Fig. 13 An example of a
second-order time-ordered
Hugenholtz PP diagram
X pqba klrs
Π srdiag
, q p ðωÞ ¼ ; ð68Þ
ε
a, b, c, i, k, l ik, bc
ðω εik, ca Þεil, ac
shows that this diagrams has poles at the double excitations εik,ca. Thus we see that
the polarization propagator does have poles at double excitations, but we are not
really ready to do calculations yet. There are two main reasons: (1) we need a more
sophisticated formalism which allows the single and double excitations to mix with
each other and (2) we would prefer a (pseudo)eigenvalue equation to solve. Thus
we still have to do quite a bit more work to arrive at a “Casida-like” equation with
explicit double excitations, but the basic idea is already present in what we have
done so far.
To do so, it is first convenient to express the PP in a molecular orbital basis as
X
Πð1, 2, 3, 4; t t 0 Þ ¼ pqrs
Πsr, qp ðt t 0 Þψ*r ð2Þψs ð1Þψ*q ð3Þψ p ð4Þ; ð69Þ
where
D E
Πsr, q p ðt t 0 Þ ¼ iθðt t 0 Þ 0r^ {H ðtÞ^
s H ðtÞ^q {H ðt 0 Þ p^ H ðt 0 Þ0
D E
þ iθðt 0 tÞ 0q^ {H ðt 0 Þ p^ H ðt 0 Þ^
r {H ðtÞ^ s H ðt Þ0 : ð70Þ
As explained in [54], this change of convention with respect to that of (46) turns out
to be more convenient. It should also be noted that, because the PP depends only
28 M.E. Casida and M. Huix-Rotllant
upon the time difference, t t 0 , we can shift the origin of the time scale so that t 0 ¼ 0
without loss of generality.
Equation (70) can be more easily manipulated by making use of the
^
X A^ ¼ X; ^ ¼ X^ A^ A^ X^ :
^ A ð71Þ
^ ^
1 A^ ¼ A^ : ð72Þ
Then
D h i
^
E
Πsr, q p ðtÞ ¼ iθðtÞ 0 eiHt r^ { s^ q^ { p^ 0
D h ^
i E
þ iθðtÞ 0q^ { p^ eiHt r^ { s^ 0 : ð74Þ
Πsr, q p ðωÞ ¼
{
p^ q^ ω1 þ H {
r^ s^ ; ð75Þ
A^ X B^ ¼ 0 A
^ { ; X; ^ 0 :
^ B ð76Þ
7
Technically this is not a metric, because the overlap matrix is symplectic rather than positive
definite. However, we will call it a metric as it can be used in much the same way as a true metric.
MBPT Insights About and Corrections to TD-DFT 29
We have defined the operator space differently from the previous work of one of
us [38] to be more consistent with the literature on the field of PP calculations. The
difference is actually the commutation of two operators which introduces one sign
change. Insertion into (75) and use of the relation
^ ^1 { ^ ^ 1 { {
T{ ω1 þ H T ¼ T{ T{ T{ ω1 þ H T{ T T ð81Þ
then gives
^ ^ 1 { {
Πsr, q p ðωÞ ¼ p^ { q^ T{ T{ ω1 þ H T{ T r^ s^ : ð82Þ
^ X { {
Γ ðωÞ ¼ ω T{ T{ þ T{ H T{ ¼ T T ZI ðω þ ωI ÞZ{I T{ T{ ; ð85Þ
I
and
h ^ i1 X
Γ 1 ðωÞ ¼ ω T{ T{ þ T{ H T{ ¼ ZI ðω þ ωI Þ1 Z{I : ð86Þ
I
So (82) reads
X
Πsr, q p ðωÞ ¼ p^ { q^ T{ ZI ðω þ ωI Þ1 Z{I T{ r^ { s^ : ð87Þ
I
This means that the PP has poles given at the pseudoeigenvalues of (83) and that
the eigenvectors may be used to calculate oscillator strengths via (87).
As the “Casida-like” (83) is so important, let us rewrite it as
A B X SA, A SA, B X
¼ω ; ð88Þ
B* A* Y SB, A SB, B Y
which is roughly
A B X 1 0 X
¼ω : ð89Þ
B* A* Y 0 1 Y
þ aikc
ð0þ1þ2Þ ð0þ1þ2Þ
A 1, 1 ¼ δi, k Fða0þ1þ2
,c
Þ
δa, c Fi, k
kc, ia
¼ δi, k bca j þ δ j, k bcai
ð1Þ
A 2, 1
kc, jbia ð91Þ
δ aik j þ δk, j bik j
b, c
ð0Þ
A 2, 2 ¼ δi, k δc, a δd, b εab, i j ;
ldkc, jbia
where Fðr0þ1
,s
Þ
¼ δr, s εr þ Mxc r, s is the matrix of the Hartree–Fock operator constructed
with Kohn–Sham orbitals and
X Ml, a Ml, c 1 X ld mc dlam
ð0þ1þ2Þ ð0þ1Þ
Fa, c ¼ Fa , c þ
εl, a 2l , m , d εlm, ad
l
ð92Þ
ð0þ1þ2Þ ð0þ1Þ
X M k, d M d , i 1 X le kd dlei
Fi , k ¼ Fi , k þ ;
d
εi , d 2l , d , e εim, de
include second-order corrections. (Note that extra factors of 1/2 occur in these
expressions when spin is taken explicitly into account.) In practice, a zero-order
approximation to A2,2 is insufficient and we must use an expression correct through
first order:
ð0þ1Þ ð0þ1Þ ð0þ1Þ ð0þ1Þ
A2, 2 ¼ δi, k δ j, l δa, c Fb, d þ δb, d Fða0þ1
,c
Þ
δa, c δb, d δ j, l Fi, k δi, k Fd, l
aibj, ckdl
δa, c f i, j, k, l ðb;d Þ δb, d f i, j, k, l ða;c Þþδ f ðb;cÞ þ δb, c f i, j, k, l ða;d Þ
a, d i, j, k, l
δa, c δb, d k jli δ j, l δk, i adbc ;
ð93Þ
where
f i, j, k, l ð p; qÞ ¼ δi, k l j pq þ δ j, l ki pq δk, j li pq δi, l k j pq : ð94Þ
4 Dressed LR-TD-DFT
We now give one answer to the problem raised in the introduction – how to include
explicit double excitations in LR-TD-DFT. This answer goes by the name “dressed
LR-TD-DFT” and consists of a hybrid MBPT/AA LR-TD-DFT method. We first
give the basic idea and comment on some of the early developments. We then go
32 M.E. Casida and M. Huix-Rotllant
into the practical details which are needed to make a useful implementation of
dressed LR-TD-DFT. Finally, we introduce the notion of Brillouin corrections
which are undoubtedly important for photochemistry.
As emphasized in Sect. 2, simple counting arguments show that the AA limits LR-
TD-DFT to single excitations, albeit dressed to include some electron correlation.
However, explicit double excitations are sometimes needed when describing
excited states. This was discussed in the introduction in the context of photo-
chemistry (Fig. 1). It is well known in ab initio quantum chemistry that double
excitations can be important when describing vertical excitations and the best
known example is briefly discussed in the caption of Fig. 14.
At first this may seem a little perplexing because the fact that the oscillator
strength is the transition matrix element of a one-electron operator – see (15) –
means that the oscillator strength of a double excitation relative to a single-
determinantal ground-state wavefunction should be zero – that is, the doubly
excited state should be spectroscopically dark. What happens is easily explained
by the two-level model shown in Fig. 15, which is sufficient to give a first
explanation of the butadiene case, for example. (In the butadiene case, the singly-
excited state to be used is already a mixture of two different one-hole/one-particle
Fig. 14 Doubles contribution to the 1Ag excited state of butadiene. Beecause the obvious two
lowest singly-excited singlets 1(1bg, 2bg) and 1(1au, 2au) are quasidegenerate in energy, they mix
qffiffiffi 1
to form new singly-excited singlets 1= 2 1bg , 2bg 1 ð1au , 2au Þ . One of these is
quasidegenerate with the doubly-excited singlet dark state 1(1b2g , 2a2u ). The resultant mixing
modifies the energy and intensity of the observed 1Ag excited state
MBPT Insights About and Corrections to TD-DFT 33
ωD
ωS f
ωa fa
fb
ωa ωb ω
states.) Figure 15 shows a bright singly-excited state with excitation energy ωS and
oscillator strength f S ¼ 1 interacting with a dark doubly-excited state with exci-
tation energy ωD and oscillator strength f D ¼ 0 via a coupling matrix element x.
The CI problem is simply
ωS x CS CS
¼ω ; ð95Þ
x ωD CD CD
ωS ¼ ωa cos 2 θ þ ωb sin 2 θ
ωD ¼ ωa sin 2 θ þ ωb cos 2 θ; ð96Þ
for some value of θ. It should be noted that the average excitation energy is
conserved in the coupled problem (ωa þ ωb ¼ ωS þ ωD) and that something similar
occurs with the oscillator strengths. This leads to the common interpretation that the
coupling “shatters the singly-excited peaks into two satellite peaks.”
Now let us see how this wavefunction theory compares with LR-TD-DFT and
how Maitra et al. [61] decided to combine the two into a hybrid method. Of course,
the proper comparison with CI is LR-TD-DFT within the TDA. Applying the
partitioning technique to (95), we obtain
x2
ωS þ CS ¼ ωCS : ð97Þ
ω ωD
Comparing this with the diagonal TDA LR-TD-DFT within the two-orbital
model,
34 M.E. Casida and M. Huix-Rotllant
ω ¼ εa, i þ ia f Hxc ðωÞia ; ð98Þ
shows that
x2
ia f Hxc ðωÞia ¼ ðωS εa, i Þ þ : ð99Þ
ω ωD
Maitra et al. [61] interpreted the first term as the adiabatic part,
AA
f Hxc ¼ ωS ε a , i ; ð100Þ
x2
Hxc ðωÞ ¼
f NA : ð101Þ
ω ωD
x 2 ¼ ωS ω D ω a ωb : ð102Þ
which is the form of the numerator used by Maitra et al. [61]. The suggestion of
Maitra et al., which defines dressed LR-TD-DFT, is to calculate the nonadiabatic
correction terms – see (101) – from MBPT [61]. Thus x and ωD in (95) are to be
calculated using MBPT rather than using DFT.
" #
ðAAÞ ð 1Þ
A1, 1 A1, 2 C1 C1
ð1Þ ð0þ1Þ ¼ω : ð103Þ
A2, 1 A2, 2 C2 C2
Thus the calculation of the A1,1 block, which is one of the most difficult to
calculate in the extended SOPPA/ADC(2) theory, is very much simplified by using
AA LR-TD-DFT. The A2,2 block must, however, be calculated through first order in
practice. It was confirmed that adding only a few (e.g., 100) double excitations led
to little difference in calculated eigenvalues unless the double excitations were
quasidegenerate with a single excitation. There is thus no significant problem in
practice with double counting electron correlation effects when using this hybrid
MBPT/LR-TD-DFT method. Tests were carried out on the test set of Schreiber
et al. consisting of 28 organic chromophores with 116 well-characterized singlet
excitation energies [66].
Note that the form of (103) was chosen instead of the form
ðAAÞ
A1, 1 þ KNA 1, 1 ðωÞ C1 ¼ ωC1
1 ð104Þ
ð1Þ ð0þ1Þ ð1Þ
KNA
1, 1 ðωÞ ¼ A1, 2 ω1 A2, 2 A2, 1 ;
the same way that we include doubles contributions to excited states so that we may
describe, for example, the photochemical funnel between S1 and S0 in Fig. 1. It is
not clear how to do this in LR-TD-DFT where the excited-state potential energy
surfaces are just obtained by adding the excitation energies at each geometry to the
ground-state DFT energies. Not only does such a procedure lead to the excited
states inheriting the convergence difficulties of the ground state surface coming
from places with noninteracting v-representability difficulties, but also there is no
coupling between the ground state and singly excited states. This is similar to what
happens with Brillouin’s theorem in CIS calculations and leads to problems
describing conical intersections. However, adding in the missing nonzero terms
(which we call Brillouin corrections) to dressed LR-TD-DFT is easy in the TDA.
It is good to emphasize at this point that we are making an ad hoc correction,
albeit one which is eminently reasonable from a wavefunction point of view.
Formally correct approaches might include: (1) acknowledging that part of the
problem may lie in the fact that noninteracting v-representability in Kohn–Sham
DFT often breaks down at key places on ground-state potential energy surfaces
when bonds are formed or broken, so that conventional Kohn–Sham DFT may no
longer be a good starting point; (2) examining nonadiabatic xc-kernels which seem
to include some degree of multideterminantal ground-state character in their
response such as that of Maitra and Tempel [68]; (3) introducing explicit
multideterminantal character into the description of the Kohn–Sham DFT ground
state. We return to this in our final section, but for now we just try the ad hoc
approach of adding Brillouin corrections to TDA dressed LR-TD-DFT. Note that
this also has an indirect effect on interactions between excited states, though the
primary effect is between excited states and the ground state.
It is sufficient to add an extra column and row to the TDA problem to take into
account the ground-state determinant in hybrid DFT. This gives
2 3
0 A0, 1 A0, 2 0 C0 1 0 1
C0
6A ðAAÞ
A1, 1
ð1Þ 7
A1, 2 5@ C1 A ¼ ω@ C1 A :
4 1, 0 ð105Þ
ð1Þ ð0þ1Þ
A2, 0 A2, 1 A2, 2 C2 C2
and
ðAAÞ
A1, 1 þ KNA 1, 1 ð ω Þ C1 ¼ ωC1
ω1 A0, 2
1
A0, 1
ð108Þ
ð 1Þ
KNA ð ω Þ ¼ A A ð0þ1Þ ð1Þ :
1, 1 1, 0 1, 2 A2, 0 ω1 A2, 2 A2, 1
Fig. 16 Potential energy surfaces of the ground and two lowest excited states of Σ þ g symmetry.
Comparison of CISD (solid lines) with adiabatic, dressed, and hybrid LR-TD-BH&HLYP/TDA
(dashed lines). All calculations have been performed with a cc-pVTZ basis set. All axes are in
Hartree atomic units (bohr for the x-axis and Hartree for the y-axis). Unlike the ethylene potential
energy curves (Fig. 17), no shift has been made in the potential energy curves
38 M.E. Casida and M. Huix-Rotllant
In Fig. 17 we show the potential energy surfaces of S0, S1, and S2 of ethylene along
the torsional coordinate. The static correlation of these three states can be essen-
tially represented by three configurations, namely the ground-state configuration
(π2π*,0), the singly-excited configuration (π1π*,1), and the doubly-excited configu-
ration (π0π*,2).
From the CASSCF(2,2)/MCQDPT2, we observe that the ground- and doubly-
excited configurations are heavily mixed at 90 , forming an avoided crossing. At
this angle, the S1 and S2 states are degenerate. These features are not captured by
adiabatic TD-DFT (Fig. 17a). Indeed, the doubly-excited configuration is missing,
Fig. 17 Potential energy cuts of the S0, S1, and S2 states of ethylene along the twisting coordinate:
x-axis in degrees, y-axis in eV. All the curves have been shifted so that the ground-state curve at
0 corresponds to 0 eV. The solid lines correspond to a CASSCF(2,2)/MCQDPT2 calculation, and
the dashed lines to the different models using the BH&HLYP functional and the Tamm–Dancoff
approximation. The 6-31++G(d,p) basis set have been employed in all calculations. (Note that
these curves are in good agreement with similar calculations previously reported in Fig. 7.3 of
Chap. 7 of [69], albeit with a different functional)
MBPT Insights About and Corrections to TD-DFT 39
and so the ground state features a cusp at the perpendicular conformation. The S1,
which is essentially represented by a single excitation, is virtually superimposed
with the CASSCF(2,2)/MCQDPT2 result. The dressed TD-DFT (Fig. 17b) includes
the double excitation, but the surfaces of S0 and S2 appear as diabatic states because
the ground- to excited-state coupling term is missing. This is largely fixed by
introducing the Brillouin corrections (Fig. 17c). The ground state is now in very
good agreement with the CASSCF(2,2)/MCQDPT2 S0 state, although the degen-
eracy of S1 and S2 at 90 is still not fully captured. Thus the picture given by
Brillouin-corrected LR-TD-DFT is qualitatively correct with respect to the multi-
reference results.
We now have the tools to deduce an MBPT expression for the TD-DFT xc-kernel. It
should be emphasized that this is not a new exercise but that we seem to be the only
ones to do so within the PP formalism. We think this may have the advantage of
making a rather complicated subject more accessible to Quantum Chemists already
familiar with the PP formalism.
The problem of constructing xc-correlation objects such as the xc-potential vxc
and the xc-kernel fxc(ω) from MBPT for use in DFT has been termed “ab initio
DFT” by Bartlet [70, 71]. At the exchange-only level, the terms optimized effective
potential (OEP) [72, 73] or exact exchange [74, 75] are also used and OEP is also
used to include the correlated case [76, 77]. At first glance, nothing much is gained.
For example, the calculated excitation energies and oscillator strengths in ab initio
TD-DFT must be, by construction, exactly the same as those from MBPT. This
approach does not give explicit functionals of the density (though it may be thought
of as giving implicit functionals). However it does allow us to formulate expres-
sions for and to calculate purely (TD-) DFT objects and hence it can provide insight
into, and computational checks of, the behavior of illusive objects such as vxc and
fxc(ω).
Here we concentrate on the latter, namely the xc-kernel. Previous work along
these lines has been carried out for the kernel by directly taking the derivative of
the OEP energy expression with the constraint that the orbitals come from a
local potential. This was first done by G€orling in 1998 [60] for the full time-
dependent exchange-only problem. In 2002, Hirata et al. redid the derivation for
the static case [78]. Later, in 2006, a diagrammatic derivation of the static result was
given by Bokhan and Bartlett [71], and the functional derivative of the kernel gx
has been treated by Bokhan and Bartlett in the static exchange-only case [79].
In this section, we take a somewhat different and arguably more direct approach
than that used in the previously mentioned articles, in that we make direct use of the
fundamental relation
40 M.E. Casida and M. Huix-Rotllant
where iþ is infinitesimally later than i. This approach has been used by Totkatly,
Stubner, and Pankaratov to develop a diagrammatic expression for fxc(ω) [80, 81]. It
also leads to the “Nanoquanta approximation,” so named by Lucia Reining because
it was simultaneously derived by several different people [41–43, 46, 44] involved
in the so-called Nanoquanta group. (See also pp. 318–329 of [24].)
The work presented here differs from previous work in two respects, namely
(1) we make a direct connection with the PP formalism which is more common in
quantum chemistry than is the full BSE approach (they are formally equivalent but
differ in practice through the approximations used) and (2) we introduce a matrix
formulation based upon Harriman’s contraction Υ^ and expansion operators Υ^ { . This
allows us to introduce the concept of the localizer Λ(ω) which shows explicitly how
localization in space results requires the introduction of additional frequency
dependence. Finally, we recover the formulae of G€orling and Hirata et al. and
produce a rather trivial proof of the Gonze and Scheffler result [82] that this
additional frequency dependence “undoes” the spatial localization procedure in
particular cases.
We first seek a compact notation for (109). Harriman considered the relation
between the space of kernels of operators and the space of functions [83, 84].
In order to main consistency with the rest of this chapter, we generalize Harriman’s
notion from space-only to space and spin coordinates. Then the collapse operator is
defined by
for an arbitrary operator kernel. The adjoint of the collapse operator is the so-called
expansion operator
for an arbitrary function f(1). Clearly Υ^ { Υ^ Að1; 2Þ ¼ Að1; 1Þδð1 2Þ 6¼ A ð1; 2Þ.
The ability to express these operators as matrices (Υ and Υ {) facilitates finite basis
set applications.
We may now rewrite (109) as
{
χ ðt1 t2 Þ ¼ ΥL t1 ; tþ þ
1 ; t2 ; t2 Υ ¼ ΥΠ ðt1 t2 ÞΥ
{
ð112Þ
Comparing
ð
χ ðt1 t2 Þ ¼ χ s ðt1 t2 Þ þ χ s ðt1 t3 Þ fHxc ðt3 t4 Þχ ðt4 t2 Þ dt3 dt4 ; ð113Þ
Lðt1 ; t2 ; t3 ; t4 Þ ¼ Ls ðt1 ; t2 ; t3 ; t4 Þ
ð
þ Ls ðt1 ; t2 ; t5 ; t6 ÞΞHxc ðt5 ; t6 ; t7 ; t8 ÞLðt7 ; t8 ; t3 ; t4 Þdt5 dt6 dt7 dt8 ;
ð114Þ
KHxc ¼ Π 1 1
s Π : ð119Þ
The price we have to pay is that the coupling matrix cannot be easily expanded in
Feynman diagrams, but that in no way prevents us from determining appropriate
algebraic expressions for it. We may then write
42 M.E. Casida and M. Huix-Rotllant
ð
ΥΠ s ðt1 t3 ÞΥ { fxc ðt3 t4 ÞΥΠ ðt4 t2 ÞΥ { dt3 dt4 ¼
ð ð120Þ
ΥΠ s ðt1 t3 ÞΥ { Kxc ðt3 t4 ÞΥΠðt4 t2 Þdt3 dt4 ;
5.1 Localizer
Evidently,
where we have introduced the notion of noninteracting (Λs) and interacting (Λ)
localizers,
1
Λs ðωÞ ¼ ΥΠ s ðωÞΥ { ΥΠ s ðωÞΥ {
1 ð123Þ
ΛðωÞ ¼ ΥΠ ðωÞΥ { ΥΠ ðωÞΥ { :
The localizer arises quite naturally in the context of the time-dependent OEP
problem. According to the Runge–Gross theory [25], the exact time-dependent
xc-potential vxc(t) is not only a functional of the density ρ(t) but also of an
initial condition which can be taken as the wavefunction Ψ(t0) at some prior time
t0. On the other hand, linear response theory begins with the static ground state case
where the first Hohenberg–Kohn theorem
tells us that the wavefunction is a
functional of the density Ψðt0 Þ ¼ Ψ ρt0 . G€orling has pointed out that this greatly
simplifies the problem [60] because we can then show that
ð ð
Πs ð1; 1; 2; 2; ωÞvx ð2; ωÞd2 ¼ Πs ð1; 1; 2; 3; ωÞΣ x ð2; 3Þd2d3 ; ð124Þ
or Σ x,
occ X
X virt
ψi ð1Þψ*a ð2Þψ*i ð3Þψa ð4Þ
Πs ð1; 2; 3; 4; ωÞ ¼
ω εa, i
i a
ð127Þ
XX ψ ð1Þψ* ð2Þψ* ð3Þψ ð4Þ
occ virt
a i a i
;
i a
ω þ ε a , i
Xocc Xvirt
ψi ð1Þψ*a ð1Þψ*i ð2Þψa ð3Þ
ðΥ Πs Þð1; 2; 3; ωÞ ¼
ω εa , i
i a
ð128Þ
XX ψ ð1Þψ ð1Þψ* ð2Þψ ð3Þ
occ virt *
a i a i
:
i a
ω þ ε a , i
Xocc Xvirt
ψi ð1Þψ*a ð1Þψ*i ð2Þψa ð2Þ
Υ Πs Υ { ð1; 2; ωÞ ¼
ω εa , i
i a
ð129Þ
X
occ X
virt
ψa ð1Þψi ð1Þψ*a ð2Þψi ð2Þ
*
:
i a
ω þ εa , i
As with the susceptibility, the two operators have poles at the independent
particle excitation energies ω ¼ εa, i ¼ ðεa εi Þ.
In order to construct the dynamic localizer, the kernel (125) has to be inverted. It
is not generally possible to do this analytically, though it can be done in a finite-
basis representation with great care. However, Gonze and Scheffler have noted that
exact inversion is possible in the special case of a frequency, ω ¼ εb, j , of a pole
well separated from the other poles [82]. Near this pole, the kernels, ΥΠs(ω) and
ΥΠs(ω)Υ {, are each dominated by single terms
where a localizer is used which is half way between the noninteracting and fully
interacting form,
1
Λ1=2 ðωÞ ¼ ΥΠ s ðωÞΥ { ΥΠðωÞΥ { : ð136Þ
1 1
f xc ðωÞ ¼ ΥΠ s ðωÞΥ { ðΠ ðωÞ Π s ðωÞÞ ΥΠ s ðωÞΥ { : ð137Þ
Thus once again we see that the frequency dependence of the localizer has
transformed the matrix element of a spatially-local frequency-dependent operator
into the matrix element of a spatially-nonlocal operator. Had the localizer been
neglected, then we would have found, incorrectly, that
ω ¼ εa, i þ iaΠ1 1
s ðεai Þ Π ðεa, i Þ ai : ð139Þ
Although the latter reduces to just εai for charge transfer excitations at a distance
(because ψi ψa ¼ 0), the former does not [85]. However, for most excitations the
overlap is non-zero. In such cases, and around a well-separated pole, the localizer
can be completely neglected.
In order to apply (137) we need only the previously derived terms represented by
the diagrams in Fig. 7. The resultant expressions agree perfectly with the expanded
expressions of the TD-EXX kernel obtained by Hirata et al. [59], which are
equivalent to the more condensed form given by G€orling [60].
Use of the GS relation then leads to
KS
a, i þ f xc
ω ¼ εKS ε
a, i
¼ εa , i þ a M
KS
^ xc a iM
^ xc i þ aiia ð140Þ
¼ εHF
a, i þ ai ia ;
which is exactly the configuration interaction singles (CIS, i.e., TDHF Tamm–
Dancoff approximation) expression evaluated using Kohn–Sham orbitals.
46 M.E. Casida and M. Huix-Rotllant
This agrees with a previous exact result obtained using G€orling–Levy perturbation
theory [82, 86, 87].
is useful because of its potential for preserving as much as possible of the basic
algebraic structure of the exact equation at (122) although still remaining compu-
tationally tractable. This is our second approximation,
{
fHxc ðωÞ ¼ Λs ðωÞ Π 1 1
s ðωÞ Π ðωÞ Λs ðωÞ: ð142Þ
Equation (142) simply reads that fHxc(ω) is a spatially localized form of KHxc(ω).
This is nothing but the PP analogue of the basic approximation (117) used in the
BSE approach on the way to the Nanoquanta approximation [41–46].
where T{2þ corresponds to the operator space of two-electron and higher excitations
and
1
1 Γ 1, 1 ðωÞ Γ 1, 2þ
Γ ðωÞ ¼ ; ð144Þ
Γ 2þ, 1 Γ 2þ, 2þ ðωÞ
^
{
^
{
Γ i, j ðωÞ ¼ Ti ω1 þ H T j : ð145Þ
Using the well-known expression for the inverse of a two-by-two block matrix
allows us to transform (143) into
h i
Πsr, q p ðωÞ ¼ p^ { q^ T{1 p^ { q^ T{2þ Γ 1 ð ω ÞΓ 2þ , 1
h 2þ, 2þ i
P ðωÞ T1 r^ s^ Γ 1, 2þ Γ 2þ, 2þ ðωÞ T{2þ r^ { s^
1 { { 1
ð146Þ
þ p^ { q^ T{2þ Γ 1 { {
2þ, 2þ ðωÞ T2þ r^ s ^ ;
where
PðωÞ ¼ Γ 1, 1 ðωÞ Γ 1, 2þ Γ 1
2þ, 2þ ðωÞΓ 2þ, 1 : ð147Þ
Although (146) is somewhat complicated, it turns out that P(ω) plays much the
same role in the smaller T{1 space that Γ(ω) plays in the full T{ space. To see how
this comes about, it is necessary to introduce the concept of order in the fluctuation
operator – see (67) – and in Mxc – see (69). We can now perform an order-by-order
expansion of (146). Through second order only the T{2 part of T{2þ contributes, so we
need not consider higher than double excitation operators. However, we make some
additional approximations. In particular, we follow the usual practice and drop the
last term in (146) because it contributes only at second order and appears to be small
when calculating excitation energies and transitions moments using the Hartree–
Fock approximation as zero-order [52, 106–109]. For response functions such as
dynamic polarizabilities, their inclusion is more critical, improving the agreement
with experiments [49]. We also have no need to consider the second term in
p^ { q^ T{1 p^ { q^ T{2þ Γ 1
2þ, 2þ ðωÞΓ 2þ, 1 : ð148Þ
This means that for the purposes of this chapter we can treat the PP in the present
work as given by
Πsr, q p ðωÞ ¼ p^ { q^ T{1 P1 ðωÞ T{1 r^ { s^ : ð149Þ
Comparing with (82) substantiates our earlier claim that P(ω) plays the
same role in the T{1 space that Γ(ω) plays over the full T{ space.
MBPT Insights About and Corrections to TD-DFT 49
where
ð0Þ { ^ ^ { ð0Þ, 1 { { ð0Þ
Πsrs , q p ðωÞ ¼ p^ { q^ T{1 T1 ω1 þ hKS T1 T1 r^ s^ : ð151Þ
(It should be noted that Pkc,ia is part of the A block, whereas Pck,ia is part of the B
block.) The sum of Pð0Þ þ Pð1Þ gives the exact pole structure up to first-order in the
SOPPA approach.
The zero-order contribution,
50 M.E. Casida and M. Huix-Rotllant
ð0Þ { {
p^ { q^ T{1 ¼ T1 T1 ; ð156Þ
1
{ 1
^ ^ ð0Þ
T{1 ω1 { ^
^
{
þ H T1 T1 ω1 þ H T1
1
1
^ ^ ð0Þ ^ ð1Þ ^ ^ ð 0Þ
ð161Þ
which is exactly the configuration interaction singles (CIS, i.e., TDHF Tamm–
Dancoff approximation) expression evaluated using Kohn–Sham orbitals. This
agrees with a previous exact result obtained using G€orling–Levy perturbation
theory [82, 86, 87].
X
n X ki
k X ðiÞ ðki jÞ
½Π s ðωÞ Π ðωÞðrs0þ1þ...þn
,q p
Þ
¼ p^ { q^ T{1 Pð jÞ, 1 ðωÞ T{1 r^ { s^ ;
k¼0 i¼0 j¼0
when the Born approximation is applied to the P(ω) in the same way as in Sect. 5.
The number of diagrams contributing to this expansion is large and, for the sake of
simplicity, we only give the resumed expressions for each block. Evidently, after
the calculation of each block there is an additional step matrix inversion in order to
apply the second approximation to the xc-kernel.
It should be emphasized that although the treatment below may seem simple,
application of Wick’s theorem is complicated and has been carried out using an
in-house FORTRAN program written specifically for the purpose. The result before
resummation is roughly 200 diagrams, which have been included as supplementary
material.
It can be shown that the operator space may be truncated without loss of
generality in a second-order treatment to only one- and two-electron excitation
operators [52]. The wavefunction may also be truncated at second order. This
truncation breaks the orthonormality of the T{1 space:
ð0Þ ð2Þ
1 0
T{1 T{1 { {
T1 T1 { {
þ T1 T1 6¼ : ð163Þ
0 1
This complication is dealt with by orthonormalizing our operator space. The new
operator set expressed in terms of the original set contains only second-order
corrections:
52 M.E. Casida and M. Huix-Rotllant
!
{
ð2Þ X 1X kdlb dkal X Mkb Mka
a^ ^i ¼ þ b^ {^i
4 ε ε
kl, bd kl, da ε ε
k , b k, a
b
kld k ! ð164Þ
X 1X md jc cidm X M jd Mdi
þ þ a^ { ^j :
j
4 mcd
ε m j , cd ε im , cd d
ε j , d ε i , d
(It should be noted that we have used the linked-cluster theorem to eliminate
contributions from disconnected diagrams. For a proof for the EOM of the one- and
two-particle the Green’s function, see [55].)
We may now proceed to calculate
ð1Þ ð1Þ, 1 { { ð0Þ
Πðsr2,Þq p ðωÞ ¼ p^ { q^ T{1 P ðωÞ T1 r^ s^
ð0Þ ð1Þ
þ p^ { q^ T{1 Pð1Þ, 1 ðωÞ T{1 r^ { s^
ð1Þ ð0Þ, 1 ð1Þ ð165Þ
þ p^ { q^ T{1 P ðωÞ T{1 r^ { s^
ð0Þ ð2Þ, 1 ð0Þ
þ p^ { q^ T{1 P ðωÞ T{1 r^ { s^ :
The only new contributions which arise at this level are from the block P(2),
which is given by
ð2Þ ð1Þ ð0Þ, 1 ð1Þ
Pð2Þ ¼ Γ 1, 1 Γ 1, 2 Γ 2, 2 ðωÞΓ 2, 1 : ð166Þ
(We are anticipating the ω-dependence of the various Γ-blocks which are
ð2Þ
derived below.) Because the block Γ 1;1 is affected by the orthonormalization
procedure, it may be useful to provide a few more details. Expanding order-by-
order,
MBPT Insights About and Corrections to TD-DFT 53
ð1Þ
^ ð0Þ ð1Þ
^
ð2Þ
Γ 1, 1 ¼ 0 T{1 ;
ω1 þ H , T1 0 {
{
ð0Þ
^ ð0Þ ð2Þ
^
{
þ 0 T1 ; ω1 þ H , T1 0
^ ð0Þ
þ 0ð2Þ T{1 ; ω1 þ H , T{1 0ð0Þ
^
{ð2Þ ^ ð0Þ ð167Þ
þ 0ð0Þ T1 ; ω1 þ H , T{1 0ð0Þ
^
^ ð0Þ ð0Þ
þ 0ð0Þ T{1 ; ω1 þ H , T1 0
^
{ð2Þ
D h i E
þ 0ð1Þ T{1 , H ^ ð1Þ ; T{ 0ð0Þ
1
D h i E
ð0Þ
þ 0 T1 , H ; T{1 0ð1Þ ;
{ ^ ð1Þ
{ ð2Þ
where T1 is the vector of second-order operators defined in (164). It is easily
shown that the first term cancels with the contributions coming from the second-
order operators, and that the contributions from second-order wave function are
exactly zero. Hence, that block is simply
D h i E D h i E
Γ 1, 1 ¼ 0ð1Þ T{1 , H
^ ð1Þ ; T{ 0ð0Þ þ 0ð0Þ T{ , H
^ ð1Þ ; T{ 0ð1Þ ;
ð2Þ
1 1 1 ð168Þ
The block Γ 1,2 and its adjoint is of at least first order because the space is
orthonormal. For that reason, it is not affected by the orthonormalization at this
level of approximation. Its calculation gives
54 M.E. Casida and M. Huix-Rotllant
h i
¼ δik bca j þ δ jk bcai δbc aik j þ δac bik j
ð2Þ
Γ 2, 1
h ikc, jbia ð171Þ
ð2Þ
Γ 2, 1 ¼ 0:
ck, jbia
Π ð0þ1þ2
0
Þ, 1
ðωÞ10 10 1
1 0 Pð0þ1þ2Þ ðωÞ Pð0þ1þ2Þ ðωÞ 1 0
¼ @ 0 1 A@ Pð0þ1þ2Þ ðωÞ Pð0þ1þ2Þ ðωÞ A@ 0 1 A
0 1 ð174Þ
ð0þ1þ2Þ ð0þ1þ2Þ
P ð ωÞ P ðωÞ
¼ @ Pð0þ1þ2Þ ðωÞ Pð0þ1þ2Þ ðωÞ A :
It follows that
MBPT Insights About and Corrections to TD-DFT 55
0 1
ð1þ2Þ
Pð1þ2Þ ðωÞ Γ 1, 1
B C
Π 1
s ð ωÞ Π
ð0þ1þ2Þ, 1
ðωÞ ¼ @ Γ ð1þ2Þ Pð1þ2Þ ðωÞ A : ð175Þ
1, 1
Localization – see (142) –complicates these formulae by mixing the Pð1þ2Þ ðωÞ
ð1þ2Þ
and Γ 1, 1 terms,
References
1. Rowlinson JS (2009) The border between physics and chemistry. Bull Hist Chem 34:1
2. Casida ME, Jamorski C, Casida KC, Salahub DR (1998) Molecular excitation energies to
high-lying bound states from time-dependent density-functional response theory: character-
ization and correction of the time-dependent local density approximation ionization thresh-
old. J Chem Phys 108:4439
3. Casida ME (2002) Jacob’s ladder for time-dependent density-functional theory: some rungs
on the way to photochemical heaven. In: Hoffmann MRH, Dyall KG (eds) Accurate descrip-
tion of low-lying molecular states and potential energy surfaces. ACS, Washington, p 199
4. Doltsinis NL, Marx D (2002) First principles molecular dynamics involving excited states
and nonadiabatic transitions. J Theo Comput Chem 1:319
5. Cordova F, Doriol LJ, Ipatov A, Casida ME, Filippi C, Vela A (2007) Troubleshooting time-
dependent density-functional theory for photochemical applications: oxirane. J Chem Phys
127:164111
6. Tapavicza E, Tavernelli I, Rothlisberger U, Filippi C, Casida ME (2008) Mixed time-
dependent density-functional theory/classical trajectory surface hopping study of oxirane
photochemistry. J Chem Phys 129(12):124108
7. Casida ME, Natarajan B, Deutsch T (2011) Non-Born-Oppenheimer dynamics and conical
intersections. In: Marques M, Maitra N, Noguiera F, Gross EKU, Rubio A (eds) Fundamen-
tals of time-dependent density-functional theory, Lecture Notes in Physics, vol 837. Springer,
Berlin, p 279
8. Casida ME, Huix-Rotllant M (2012) Progress in time-dependent density-functional theory.
Annu Rev Phys Chem 63:287
9. Hohenberg P, Kohn W (1964) Inhomogenous electron gas. Phys Rev 136:B864
10. Kohn W, Sham LJ (1965) Self-consistent equations including exchange and correlation
effects. Phys Rev 140:A1133
11. Parr RG, Yang W (1989) Density-functional theory of atoms and molecules. Oxford Uni-
versity Press, New York
12. Dreizler DM, Gross EKU (1990) Density functional theory, an approach to the quantum
many-body problem. Springer, New York
13. Koch W, Holthausen MC (2000) A chemist’s guide to density functional theory. Wiley-VCH,
New York
14. Perdew JP, Schmidt K (2001) Jacob’s ladder of density functional approximations for the
exchange-correlation energy. In: Doren VEV, Alseoy KV, Geerlings P (eds) Density func-
tional theory and its applications to materials. American Institute of Physics, Melville,
New York, p 1
15. Perdew JP, Ruzsinsky A, Constantin LA, Sun J, Csonka GI (2009) Some fundamental issues
in ground-state density functional theory: a guide for the perplexed. J Chem Theor Comput
5:902
16. Perdew JP, Constantin LA (2007) Laplacian-level density functionals for the kinetic energy
density and exchange-correlation energy. Phys Rev B 75:155109
17. Gill PM (2001) Obituary: density-functional theory (1927–1993). Aust J Chem 54:661
18. Becke A (1993) A new mixing of HartreeFock and local density functional theories. J Chem
Phys 98:1372
19. Perdew JP, Ernzerhof M, Burke K (1996) Rationale for mixing exact exchange with
density functional approximations. J Chem Phys 105:9982
20. Savin A (1995) Beyond the Kohn–Sham determinant. In: Chong DP (ed) Recent advances in
density functional theory. World Scientific, Singapore, p 129
21. Baer R, Livshits E, Salzner U (2010) Tuned range-separated hybrids in density functional theory.
Annu Rev Phys Chem 61:85
22. Marques MAL, Ullrich C, Nogueira F, Rubio A, Gross EKU (eds) (2006) Time-dependent
density-functional theory, Lecture Notes in Physics, vol 706. Springer, Berlin
MBPT Insights About and Corrections to TD-DFT 57
23. Marques M, Maitra N, Noguiera F, Gross EKU, Rubio A (2011) Fundamentals of time-
dependent density-functional theory, Lecture Notes in Physics, vol 837. Springer, Berlin
24. Ullrich CA (2012) Time-dependent density-functional theory: concepts and applications.
Oxford University Press, Oxford
25. Runge E, Gross EKU (1984) Density functional theory for time-dependent systems. Phys Rev
Lett 52:997
26. van Leeuwen R (1999) Mapping from densities to potentials in time-dependent density-
functional theory. Phys Rev Lett 82:3863
27. Maitra NT, Todorov TN, Woodward C, Burke K (2010) Density-potential mapping in time-
dependent density-functional theory. Phys Rev A 81:042525
28. Ruggenthaler M, van Leeuwen R (2011) Global fixed-point proof of time-dependent density-
functional theory. Europhys Lett 95:13001
29. Ruggenthaler M, Glesbertz KJH, Penz M, van Leeuwen R (2012) Density-potential mappings
in quantum dynamics. Phys Rev A 85:052504
30. Ruggenthaler M, Nielsen SEB, van Leeuwen R (2013) Analytic density functionals with
initial-state dependence. Phys Rev A 88:022512
31. Vignale G (2008) Real-time resolution of the causality paradox of time-dependent density-
functional theory. Phys Rev A 77(6):1. doi:10.1103/PhysRevA.77.062511
32. Messud J, Dinh PM, Reinhard P, Suraud E (2011) The generalized SIC-OEP formalism and
the generalized SIC-Slater approximation (stationary and time-dependent cases). Ann Phys
(Berlin) 523:270
33. Rajagopal AK (1996) Time-dependent variational principle and the effective action in
density-functional theory and Berrys phase. Phys Rev A 54:3916
34. van Leeuwen R (1998) Causality and symmetry in time-dependent density-functional theory.
Phys Rev Lett 80:1280
35. van Leeuwen R (2001) Key concepts in time-dependent density-functional theory. Int J Mod
Phys 15:1969
36. Mukamel S (2005) Generalized time-dependent density-functional-theory response functions
for spontaneous density fluctuations and nonlinear response: resolving the causality paradox.
Phys Rev A 024503
37. Mosquera MA (2013) Action formalism in time-dependent density-functional theory.
Phys Rev B 88:022515
38. Casida ME (1995) Time-dependent density-functional response theory for molecules. In:
Chong DP (ed) Recent advances in density functional methods, Part I. World Scientific,
Singapore, p 155
39. Casida ME (1996) Time-dependent density functional response theory of molecular systems:
theory, computational methods, and functionals. In: Seminario J (ed) Recent developments
and applications of modern density functional theory. Elsevier, Amsterdam, p 391
40. L€owdin PO (1964) Studies in perturbation theory. Part VI. Contraction of secular equations.
J Mol Spectr 14:112
41. Onida G, Reining L, Rubio A (2002) Electronic excitations: density-functional versus many-
body Greens-function approaches. Rev Mod Phys 74:601
42. Reining L, Olevano V, Rubio A, Onida G (2002) Excitonic effects in solids described by
time-dependent density-functional theory. Phys Rev Lett 88:066404
43. Sottile F, Olevano V, Reining L (2003) Parameter-free calculation of response functions in
time-dependent density-functional theory. Phys Rev Lett 91:056402
44. Marini A, Sole RD, Rubio A (2003) Bound excitons in time-dependent density-functional
theory: optical and energy-loss spectra. Phys Rev Lett 91:256402
45. Stubner R, Tokatly IV, Pankratov O (2004) Excitonic effects in time-dependent density-
functional theory: an analytically solvable model. Phys Rev B 70:245119
46. von Barth U, Dahlen NE, van Leeuwen R, Stefanucci G (2005) Conserving approximations in
time-dependent density functional theory. Phys Rev B 72:235109
58 M.E. Casida and M. Huix-Rotllant
47. Romaniello P, Sangalli D, Berger JA, Sottile F, Molinari LG, Reining L, Onida G (2009)
Double excitations in finite systems. J Chem Phys 130:044108
48. Oddershede J, Jørgensen P (1977) An order analysis of the particle-hole propagator. J Chem
Phys 66:1541
49. Nielsen ES, Jørgensen P, Oddershede J (1980) Transition moments and dynamic polarizabili-
ties in a second order polarization propagator approach. J Chem Phys 73:6238
50. Nielsen ES, Jørgensen P, Oddershede J (1980) J Chem Phys 75:499; Erratum (1980): J Chem
Phys 73:6238
51. Jørgensen P, Simons J (1981) Second quantization-based methods in quantum chemistry.
Academic, New York
52. Schirmer J (1982) Beyond the random phase approximation: a new approximation scheme for
the polarization propagator. Phys Rev A 26:2395
53. Trofimov AB, Stelter G, Schirmer J (1999) A consistent third-order propagator method for
electronic excitation. J Chem Phys 111:9982
54. Fetter AL, Walecka JD (1971) Quantum theory of many-particle systems. McGraw-Hill,
New York
55. Kobe DH (1966) Linked cluster theorem and the Green’s function equations of motion for a
many-fermion system. J Math Phys 7(10):1806
56. Wilson S (1984) Electron correlation in molecules. Clarendon, Oxford
57. Sangalli D, Romaniello P, Colò G, Marini A, Onida G (2011) Double excitation in correlated
systems: a many-body approach. J Chem Phys 134:034115
58. Casida ME (2005) Propagator corrections to adiabatic time-dependent density-functional
theory linear response theory. J Chem Phys 122:054111
59. Hirata S, Ivanov S, Bartlett RJ, Grabowski I (2005) Exact-exchange time-dependent density-
functional theory for static and dynamic polarizabilities. Phys Rev A 71:032507
60. G€orling A (1998) Exact exchange kernel for time-dependent density-functional theory. Int J
Quant Chem 69:265
61. Maitra NT, Zhang F, Cave RJ, Burke K (2004) Double excitations within time-dependent
density functional theory linear response theory. J Chem Phys 120:5932
62. Cave RJ, Zhang F, Maitra NT, Burke K (2004) A dressed TDDFT treatment of the 1Ag states
of butadiene and hexatriene. Chem Phys Lett 389:39
63. Mazur G, Włodarczyk R (2009) Application of the dressed time-dependent density functional
theory for the excited states of linear polyenes. J Comput Chem 30:811
64. Gritsenko OV, Baerends EJ (2009) Double excitation effect in non-adiabatic time-dependent
density functional theory with an analytic construction of the exchange-correlation kernel in
the common energy denominator approximation. Phys Chem Chem Phys 11:4640
65. Huix-Rotllant M, Ipatov A, Rubio A, Casida ME (2011) Assessment of dressed time-
dependent density-functional theory for the low-lying valence states of 28 organic chromo-
phores. Chem Phys 391:120
66. Schreiber M, Silva-Junior MR, Sauer SPA, Thiel W (2008) Benchmarks for electronically
excited states: CASPT2, CC2, CCSD, and CC3. J Chem Phys 128:134110
67. Hsu CP, Hirata S, Head-Gordon M (2001) Excitation energies from time-dependent density
functional theory for linear polyene oligomers: butadiene to decapentaene. J Phys Chem A
105:451
68. Maitra NT, Tempel DG (2006) Long-range excitations in time-dependent density functional
theory. J Chem Phys 125:184111
69. Huix-Rotllant M (2011) Improved correlation kernels for linear-response time-dependent
density-functional theory. Ph.D. thesis, Université de Grenoble
70. Bokhan D, Schweigert IG, Bartlett RJ (2005) Interconnection between functional derivative
and effective operator approaches in ab initio density functional theory. Mol Phys 103:2299
71. Bokhan D, Bartlett RJ (2006) Adiabatic ab initio time-dependent density-functional theory
employing optimized-effective-potential many-body perturbation theory potentials.
Phys Rev A 73:022502
MBPT Insights About and Corrections to TD-DFT 59
72. Talman JD, Shadwick WF (1976) Optimized effective atomic central potential. Phys Rev A
14:36
73. Talman JD (1989) A program to compute variationally optimized effective atomic potentials.
Comp Phys Commun 54:85
74. G€orling A (1999) New KS method for molecules based on an exchange charge density
generating the exact local KS exchange potential. Phys Rev Lett 83:5459
75. Ivanov S, Hirata S, Bartlett RJ (1999) Exact exchange treatment for molecules in finite-basis-
set Kohn–Sham theory. Phys Rev Lett 83:5455
76. Casida ME (1995) Generalization of the optimized effective potential model to include
electron correlation: a variational derivation of the Sham–Schlüter equation for the
exact exchange-correlation potential. Phys Rev A 51:2505
77. Casida ME (1999) Correlated optimized effective potential treatment of the derivative
discontinuity and of the highest occupied Kohn–Sham eigenvalue: a Janak-type theorem
for the optimized effective potential method. Phys Rev B 59:4694
78. Hirata S, Ivanov S, Grabowski I, Bartlett RJ (2002) Time-dependent density functional theory
employing optimized effective potentials. J Chem Phys 116:6468
79. Bokhan D, Barlett RJ (2007) Exact-exchange density functional theory for hyperpolariza-
bilities. J Chem Phys 127:174102
80. Tokatly IV, Pankratov O (2001) Many-body diagrammatic expansion in a Kohn–Sham basis:
implications for time-dependent density functional theory of excited states. Phys Rev Lett 86:
2078
81. Tokatly IV, Stubner R, Pankratov O (2002) Many-body diagrammatic expansion of the
exchange-correlation kernel in time-dependent density-functional theory. Phys Rev B 65:
113107
82. Gonze X, Scheffler M (1999) Exchange and correlation kernels at the resonance frequency:
implications for excitation energies in density-functional theory. Phys Rev Lett 82:4416
83. Harriman JE (1983) Geometry of density-matrices. 4. The relationship between density-
matrices and densities. Phys Rev A 27:632
84. Harriman JE (1986) Densities, operators, and basis sets. Phys Rev A 34:29
85. Heßelmann A, Ipatov A, G€orling A (2009) Charge-transfer excitation energies with a time-
dependent density-functional method suitable for orbital-dependent exchange-correlation
functionals. Phys Rev A 80:012507
86. Filippi C, Umrigar CJ, Gonze X (1997) Excitation energies from density functional pertur-
bation theory. J Chem Phys 107(23):9994
87. G€orling A (1996) Density-functional theory for excited states. Phys Rev A 54(5):3912
88. Li SL, Marenich AV, Xu X, Truhlar DG (2014) Configuration interaction-corrected Tamm-
Dancoff approximation: a time-dependent density functional method with the correct dimen-
sionality of conical intersections. J Chem Phys Lett 5:322
89. Fromager E, Knecht S, Jensen HJA (2013) Multi-configuration time-dependent density-
functional theory based upon range separation. J Chem Phys 138:084101
90. Seidu I, Krykunov M, Ziegler T (2014) The formulation of a constricted variational density
functional theory for double excitations. Mol Phys 112:661
91. B€ohm M, Tatchen J, Krügler D, Kleinermanns K, Nix MGD, LaGreve TA, Zwier TS,
Schmitt M (2009) High-resolution and dispersed fluorescence examination of vibronic bands
of tryptamine: spectroscopic signatures for La/Lb mixing near a conical intersection. J Phys
Chem A 113:2456
92. Minezawa N, Gordon MS (2009) Optimizing conical intersections by spin-flip density-
functional theory: application to ethylene. J Phys Chem A 113:12749
93. Huix-Rotllant M, Natarajan B, Ipatov A, Wawire CM, Deutsch T, Casida ME (2010)
Assessment of noncollinear spin-flip Tamm-Dancoff approximation time-dependent den-
sity-functional theory for the photochemical ring-opening of oxirane. Phys Chem Chem
Phys 12:12811
60 M.E. Casida and M. Huix-Rotllant
94. Rinkevicius Z, Vahtras O, Ågren H (2010) Spin-flip time dependent density functional theory
applied to excited states with single, double, or mixed electron excitation character. J Chem
Phys 133:114104
95. Minezawa N, Gordon MS (2011) Photoisomerization of stilbene: a spin-flip density func-
tional theory approach. J Phys Chem A 115:7901
96. Casanova D (2012) Avoided crossings, conical intersections, and low-lying excited states
with a single reference method: the restricted active space spin-flip configuration interaction
approach. J Chem Phys 137:084105
97. Huix-Rotllant M, Filatov F, Gozem S, Schapiro I, Olivucci M, Ferré N (2013) Assessment of
density functional theory for describing the correlation effects on the ground and excited state
potential energy surfaces of a retinal chromophore model. J Chem Theory Comput 9:3917
98. Minezawa N (2014) Optimizing minimum free-energy crossing points in solution: linear-
response free energy/spin-flip density functional theory approach. J Chem Phys 141:164118
99. Harabuchi Y, Keipert K, Zahariev F, Taketsugu T, Gordon MS (2014) Dynamics simulations
with spin-flip time-dependent density functional theory: photoisomerization and photo-
cyclization mechanisms of cis-stilbene in (π, π*) states. J Phys Chem A 118:11987
100. Nikiforov A, Gamez JA, Thiel W, Huix-Rotllant M, Filatov M (2014) Assessment of
approximate computational methods for conical intersections and branching plane vectors
in organic molecules. J Chem Phys 141:124122
101. Gozem S, Melaccio F, Valentini A, Filatov M, Huix-Rotllant M, Ferré N, Frutos LM,
Angeli C, Krylov AI, Granovsky AA, Lindh R, Olivucci M (2014) Shape of multireference,
equation-of-motion coupled-cluster, and density functional theory potential energy surfaces
at a conical intersection. J Chem Theory Comput 10:3074
102. Zhang X, Herbert JM (2014) Analytic derivative couplings for spin-flip configuration inter-
action singles and spin-flip time-dependent density functional theory. J Chem Phys 141:
064104
103. Frank I, Damianos K (2007) Restricted open-shell Kohn–Sham theory: simulation. J Chem Phys
126:125105
104. Friedrichs J, Darnianos K, Frank I (2008) Solving restricted open-shell equations in
excited state molecular dynamics simulations. J Chem Phys 347:17
105. Filatov M (2015) Spin-restricted ensemble-referenced Kohn–Sham method: basic principles
and application to strongly correlated ground and excited states of molecules. Comput Mol
Sci 5:146
106. Shibuya T, Rose J, McKoy V (1973) Equations-of-motion method including renormalization
and double-excitation mixing. J Chem Phys 58:500
107. Jørgensen P, Oddershede J, Ratner MA (1975) Two-particle, two-hole corrections to a self-
consistent time-dependent Hartree-Fock scheme. Chem Phys Lett 32:111
108. Oddershede J, Sabin JR (1983) The use of modified virtual orbitals in perturbative polar-
ization propagator calculations. J Chem Phys 79:2295
109. Oddershede J, Jørgensen P, Yeager DL (1984) Polarization propagator methods in atomic and
molecular calculations. Comp Phys Rep 2:33
110. Oddershede J, Jørgensen P, Beebe NHF (1978) Analysis of excitation energies and
transition moments. J Phys B Atom Mol Phys 11:1
111. Trofimov AB, Schirmer J (1995) An efficient polarization propagator approach to
valence electron excitation spectra. J Phys B At Mol Opt Phys 28:2299
Top Curr Chem (2016) 368: 61–96
DOI: 10.1007/128_2014_611
# Springer International Publishing Switzerland 2014
Published online: 20 March 2015
Tom Ziegler, Mykhaylo Krykunov, Issaka Seidu, and Young Choon Park
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2 Constricted Variational Density Functional Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.1 Second Order Constricted Variational Density Functional Theory . . . . . . . . . . . . . . . . . . . 64
2.2 Equivalence Between Adiabatic TDDFT and Second Order Constricted
Variational Density Functional Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2.3 Perturbative All Order Constricted Variational Density Functional Theory . . . . . . . . . 67
2.4 Self-Consistent All Order Constricted Variational Density Functional Theory . . . . . . 75
2.5 Self-Consistent All Order Constricted Variational Density Functional Theory
with Orbital Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
1 Introduction
Excited states have been studied in wave function theory by both excited state
variational theories and ground state response methods [1, 2]. Either approach has
been used extensively and is considered complementary and, in principle, able to
afford estimates to any desired degree of accuracy. Given the status of Kohn–Sham
density functional theory (KS-DFT) as a ground state theory, the natural path to
excited states in KS-DFT would seem to be the ground state response approach. In
fact, Runge and Gross [3] have formulated a time-dependent density functional
ground state response theory (TDDFT) which in principle should be able to
describe excited state properties without approximations. TDDFT in its exact
form requires knowledge of the “true” ground state functional and of the frequency
dependence of the energy response kernel corresponding to this functional. In
practical calculations, use is made of approximate ground state functionals and
the frequency dependence of the kernel is neglected in what has now become
known as the adiabatic TDDFT approach (ATDDFT) [4–9]. For more than two
decades the ATDDFT approach has remained the method of choice in DFT-based
studies of excited states and both its merits and limitations have been studied in
great detail [10–33]. Progress beyond the adiabatic approximation has, on the other
hand, been slow, although work in this direction is ongoing [34–36].
Long before TDDFT, Slater introduced a variational DFT approach to excited
states called ΔSCF [37, 38]. Excited states are reached in this scheme by promoting
electrons from occupied to virtual ground state levels followed by a KS calculation
on the new electron configuration. The ΔSCF approach has met with considerable
success for those lower excited states which can be represented by a single orbital
replacement (SOR) [39–49]. However, it is plagued by SCF-convergence problems.
Further, as it applies a ground state functional in a variational excited state
calculation, it is considered somewhat ad hoc [50] and without any theoretical
foundation [51–53]. Nevertheless, Van Voorhis et al. [39] have recently put for-
ward some theoretical justifications for ΔSCF and Besley et al. [42, 43] and Park
et al. [44] have addressed the SCF-convergence issue. Apart from ΔSCF, there are a
number of interesting variational DFT approaches to the study of excited states.
They include ensemble DFT [54–59], variation of bifunctionals [60], and excited
state perturbation theory [61]. They are discussed elsewhere in this volume.
The use of a ground state functional in variational excited state calculations
seems intuitively appealing from the point of view that electron correlation should
be quite similar in the ground and excited states, at least in the lower valence region.
In fact, based on this notion we introduced in 2009 the constricted variational DFT
method (CV-DFT) for excited states [29]. In this theory we allow for an admixture
of virtual ground
state orbitals fψ a ; a ¼ 1, virg into each of the occupied ground
state orbitals ψ j ; j ¼ 1, occ , according to
X
vir
ψi ¼ Uai ψ a ð1Þ
a
The ansatz in (1) makes it possible to construct occupied excited state orbitals
and evaluate the corresponding excited state energies based on the ground state
functional to any desired order n in the variational mixing matrix U. Such a
Constricted Variational Density Functional Theory Approach to the. . . 63
procedure gives rise to the nth order CV-DFT scheme designated as CV(n)-DFT
[26–28].
We start this review by an outline of the CV(n)-DFT theory. This framework
enables us to identify ATDDFT and ΔSCF as special cases of the CV(n)-DFT
scheme with ATDDFT being equivalent to CV(2)-DFT [29, 62] whereas ΔSCF
corresponds to CV(1)-DFT under the simplifying assumption that the excited state
under investigation can be described by a single orbital replacement ði ! aÞ [27,
44]. The theoretical exposition is followed by first assessing the general perfor-
mance of CV(n)-DFT in connection with applications to n ! π* [26–67] and
π ! π* [27] transitions in organic molecules. After that we demonstrate that CV
(n)-DFT is able to deal with a number of transitions where the performance of
ATDDFT based on local and hybrid functionals is problematic. These transitions
involve π ! π* excitations in organic dyes [64, 65] as well as π ! π* transitions in
charge transfer adducts [30] and Rydberg excitations for atoms and small molecules
[66]. We finally discuss future directions for the development and application of the
CV(n)-DFT scheme.
Here ϕocc and ϕvir are concatenated column vectors containing the sets
0 0
fϕi ; i ¼ 1, occg and fϕa ; a ¼ 1, virg whereas ϕocc and ϕvir are concatenated column
0 0
vectors containing the resulting sets ϕi ; i ¼ 1, occ and ϕa ; a ¼ 1, vir of
occupied and virtual excited state orbitals, respectively. The unitary transformation
matrix Y in (2a) is expressed in terms of a skew symmetric matrix U as
2 m 2 m
U2 X1
Um X1
U X1
U
Y ¼e ¼IþUþ
U
þ ¼ ¼ þU ð2bÞ
2 m¼0
m! m¼0 2m! m¼0
ð2m þ 1Þ!
If the summation in (2a) and (2b) over m is carried out to m ¼ n we talk about nth
order CV-DFT or CV(n)-DFT. Above U ij ¼ Uab ¼ 0 where “i,j” refer to the
occupied set fϕi ; i ¼ 1, occg whereas “a,b” refer to fϕa ; a ¼ 1, virg. Further, Uai
are the variational mixing matrix elements of (1) which combines virtual and
occupied ground state orbitals in the excited state with Uai ¼ U ia . Thus the entire
64 T. Ziegler et al.
matrix U is made up of occ vir independent elements Uai which can also be
! !
organized in the column vector U. For a given U we can, by means of (2a) and (2b),
generate a set of “occupied” excited state orbitals:
0
X
occþvir X
occ X
vir
ϕi ¼ Y pi ϕp ¼ Y ji ϕj þ Y ai ϕa ð3Þ
p j a
In the simple CV(2)-DFT theory [29] the unitary transformation of (2a) and (2b) is
carried out to second order in U. We thus obtain the occupied excited state orbitals
to second order as
0
X
vir occ X
1X vir
ϕi ¼ ϕi þ U ai ϕa Uai U aj ð4Þ
a
2 j a
from which we can generate the excited state Kohn–Sham density matrix to second
order as
The expression for ρ 0 (1, 1 ’) now makes it possible to write down the
corresponding excited state Kohn–Sham energy to second order as
X XX
EKS ½ρ0 ð1, 10 Þ ¼ EKS ½ρ0 þ Uai U *ai ε0a ε0i þ U ai U *bj K ai, bj
ai ai bj
1X X 1X X * * h i ð6Þ
þ U ai U bj K ai, jb þ U ai U bj K ai, jb þ O U ð3Þ :
2 ai bj 2 ai bj
Constricted Variational Density Functional Theory Approach to the. . . 65
Here EKS[ρ0] is the ground state energy and “a,b” run over virtual ground state
canonical orbitals whereas “i,j” run over occupied ground state canonical orbitals.
Further,
K ru, tq ¼ K ru
C
, tq þ K ru, tq
XC
ð7Þ
where
ðð
1
, tq ¼ ϕ*r ð1Þϕu ð1Þ ϕ ð2Þϕ*q ð2Þdv1 dv2 ð8Þ
C
K ru
r 12 t
whereas
ðð
XCðHFÞ 1
K ru, tq ¼ ϕ*r ð1Þϕt ð2Þ ϕu ð2Þϕ*q ð1Þdv1 dv2 ð9aÞ
r 12
ð9bÞ
for DFT exchange correlation. In (9a) msr ¼ 1=2 for a spin orbital ϕr(1) of α-spin
whereas msr ¼ 1=2 for a spin orbital ϕr(1) of β-spin. In addition, the kernel
f (τ,υ)(ρ0) is the second functional derivative of EXC with respect to ρα and ρβ:
δ2 E
τ, υ XC
f ρα ; ρβ ¼
0 0
τ ¼ α, β ; υ ¼ α, β: ð10Þ
δρτ δρυ 0
! !
Finally for KS exchange we have the case where ϕu r 1 , ϕq r 1 have the same
! !
(α) spin whereas ϕr r 1 , ϕt r 1 are of the other (β) spin. In this case we have,
according to Wang and Ziegler [67–69],
2 !! 3
ð ! ! ! !
1 1 δE δE
KSðXCÞ
K ru, tq ¼ ϕr r 1 ϕu r 1 ϕt r 1 ϕq r 1 4 0
* * XC
XC 5d !
r1
2 s δρα δρβ
ðρ0 ;s0 Þ
ð11Þ
! !
In (11) the integration is over space and ϕr r 1 , ϕt r 1 are the spatial parts of
KSðXCÞ
orbitals with β-spin. The evaluation of K ru, tq by numerical integration might lead
α α
to numerical instabilities if s ¼ ρ ρ 0. We can, in that case, carry out a Taylor
0
α β
expansion of ∂EKSXC =∂ρα , ∂EXC =∂ρβ from ρ ¼ ρ þ ρ and s ¼ 0. Thus
KS 0
66 T. Ziegler et al.
2 ! 3
ðh i1 δ2 E
KSðXCÞ 1 δ 2
E δ 2
E
ϕr ðr 1 Þϕu ðr 1 Þϕt ðr 1 Þϕq ðr 1 Þ 4 5dr 1
* XC XC XC
K ru, tq ¼ *
þ 2 2
2 2 δ 2 ρα δ ρβ δρα δρα
ðρ0 ;s0 Þ
KS, XC
¼ K KS
ru, tq K rt, uq
, XC
ð12Þ
, XC KS, XC
where K KS
ru, tq ,K rt, uq are defined in (9a) and (9b). The expression in (12) is
numerically stable and has no singularities for s0 ¼ 0. Finally, ε0i ,ε0a in (6) are the
ground state orbital energies of ϕi(1) and ϕa(1), respectively.
! ðI Þ ! ðI Þ
AKS þ BKS U ¼ λðIÞ U ð13aÞ
where
0
ai, bj ¼ δab δij εa εi þ K ai, bj ;
AKS ai, bj ¼ K ai, jb : ð13bÞ
0 KS
BKS KS
! ðI Þ
substituting U into (6) and making use of the constraint and normalization
! ðI Þþ ! ðI Þ ! ðI Þþ
condition U U ¼ 1 after multiplying on both sides with U .
Within the Tamm–Dancoff approximation [70], (13a) reduces to
! ðI Þ ! ðI Þ
AKS U ¼ λðIÞ U ð14Þ
which is identical in form to the equation obtained from ATDDFT in its adiabatic
formulation [3–9] after applying the same Tamm–Dancoff [70] approximation
(ATDDFT-TD). We have recently shown [62] that CV(2)-DFT without the
Tamm–Dancoff approximation is equivalent to the full adiabatic TDDFT scheme
developed by Gross [3], Casida [4], and others [5–9].
! ðI Þ
Having determined U from (14) allows us [28] to turn to a discussion of how we
construct the proper energy expression for excited singlet states originating from a
closed shell ground state. We first consider a spin-conserving transition from a close
shell ground state and assume without loss of generality that the transition takes
place in the α-manifold. In this case we can write the occupied excited state
KS-orbitals generated from the unitary transformation of (2a) and (2b) as [28, 71]
0
ϕi ¼ cos ½ηγ i ϕoi α þ sin ½ηγ i ϕvi α ; i ¼ 1, occ=2 ð15Þ
and
0
ϕi ¼ ϕi ; i ¼ occ=2 þ 1, vir ð16Þ
Here ΨM represents a mixed spin-state [40] which is half singlet and half triplet.
Further, γ i ði ¼ 1, occ=2Þ is a set of eigenvalues to
where γ is a diagonal matrix of dimension occ/2 whereas Uαα is the part of the
U matrix which runs over the occupied fϕi ; i ¼ 1, occg and virtual fϕa ; a ¼ 1, virg
ground state orbitals of α-spin [28, 71]. Further,
68 T. Ziegler et al.
X
occ=2
αα
ϕoi α ¼ W ji ϕj ð19Þ
j
and
X
occ=2
ϕvi α ¼ Vαα ai ϕa ð20Þ
a
Xocc=2
and
00
ϕi ¼ ϕi ; i ¼ occ=2 þ 1, vir ð22Þ
where Uβα is the part of U which runs over the virtual ground state orbitals of β-spin
and occupied ground state orbitals of α-spin. Finally
Constricted Variational Density Functional Theory Approach to the. . . 69
X
occ=2
ϕoi α ¼ Wβα ji ϕj ð25Þ
j
and
X
a¼vir
Vβα
v
ϕi β ¼ ai ϕa ð26Þ
a¼vir=2þ1
With the excitation energy of ΨT given by ΔET and that of ΨM by ΔEM, we can
write the singlet transition energy as [40]
provided that Uαα ¼ Uβα . This implies that Vαα ¼ Vβα , Wαα ¼ Wβα , and γ ¼ γ 0 . As
a result, ϕoi α of (19) and (25) become identical as do the spatial parts of ϕvi α in (20)
v
and ϕi β in (26). Straightforward manipulations [27, 28] allow us finally to write
down the mixed state transition energy to all orders in U in a compact and closed
form as
X
occ=2
þ2 sin ½ηγ i sin ½ηγ i sin ηγ j cos ηγ j K ivα ivα joα jvα
i¼1 j¼1
occ=2 X
X occ=2
2 sin ½ηγ i sin ½ηγ i sin ηγ j cos ηγ j K ioα ioα joα jvα
i¼1 j¼1
ð28Þ
for spin-conserving transition from a close shell ground state. More details are
given in Sect. 2.4. Here the indices ioα , joα , ivα , and jvα refer to α-spin orbitals with the
spatial parts ϕoi α , ϕoj α , ϕvi α , and ϕvj α , respectively. The expression for the
corresponding triplet transition energy reads
70 T. Ziegler et al.
X
occ=2
0
v v
Here the indices ivβ and jvβ refer to β-spin orbitals with the spatial parts ϕi β and ϕi β ,
respectively. More details are given in Sect. 2.4. From (28) and (29) we can readily
express ΔES using (27).
In perturbative all order constricted variational DFT (P-CV(1)-DFT) [27, 64]
we make use of the U matrix optimized to second order according to (14), which is
also the U obtained by ATDDFT-TD (CV(2)-DFT-TD). With this U we are able to
generate ΨM of (17) and ΨT of (23) by means of (2a) and (2b). From that we can
calculate ΔET and ΔES by means of (27)–(29).
Table 1 Results from TDDFT-TD and CV(1)-TD calculations on excitation energiesa for π to
π* transitions in polyenes using LDA
Group State Bestb TDDFT-TDc CV(1)-TD γ maxd
Ethene A B1u 7.80 8.44 8.39 1.177
Butadiene A Bu 6.18 6.16 6.10 1.174
B Ag 6.55 6.24 6.70 0.841
Hexatriene B Ag 5.09 5.03 5.36 0.787
A Bu 5.10 5.05 4.93 1.200
Octatetraene B Ag 4.47 4.17 4.42 0.799
A Bu 4.66 4.34 4.16 1.212
Cyclopropene A B2 7.06 6.30 7.55 1.253
Cyclopentadiene A B2 5.55 5.39 5.87 1.254
B A1 6.31 6.05 6.45 0.809
Norbornadiene A A2 5.34 4.52 5.10 1.158
B B2 6.11 4.95 5.36 0.942
Naphthalene B B3u 4.24 4.20 4.39 0.788
A B2u 4.77 4.25 4.71 1.103
B B1g 5.99 4.97 5.24 0.850
B Ag 5.87 5.80 6.02 0.854
A B2u 6.33 6.12 6.09 1.043
B Ag 6.67 6.21 6.71 0.904
B B3u 6.06 6.22 6.14 0.730
B B1g 6.47 6.47 6.36 0.791
RMSD Ae 0.48 0.35
RMSD B 0.51 0.34
RMSD A + B 0.50 0.35
a
Energies in eV
b
Ab initio benchmark calculations [73]
c
Identical to CV(2)-TD
d
Maximum γ eigenvalue for this transition; see (18)
e
Root mean square deviation in eV
mean square deviation (RMSD) of 0.34 eV compared to the best [73] wave function
results constitute an improvement over the TDDFT-TD excitation energies with an
RMSD of 0.51 eV. For the group A transitions we also note an improvement in
going from TDDFT-TD with RMSD ¼ 0.48 eV to CV(1)-TD with
RMSD ¼ 0.35 eV. From experience so far [27, 64] it seems that CV(1) can be
used with advantage in studies of π ! π* transitions involving dyes employing
simple local functionals.
The singlet 1La and 1Lb π ! π* excitations in linear polyacenes represent a set of
benchmark excitations which have been studied extensively both experimentally
72 T. Ziegler et al.
Fig. 2 Linear acenes with up to six fused rings (nr ¼ 6) considered in this study
Fig. 3 Excitation energies for the 1La and 1Lb transitions in linear acenes as a function of the
number of rings according to (a) experiment, (b) CV(2)-TD, (c) TDDFT, or (d) CV(1)-TD
[74] and theoretically [75–78]. Here the acenes consist of a number (nr) of fused
benzene rings; see Fig. 2.
The distinct properties [74] of the 1La and 1Lb states for the linear acenes have
already been described in the literature [75–78]. Essentially, the 1La (or 1B2u when
the x-axis corresponds to the long molecular axis state) is dominated by a single
electron transition HOMO ! LUMO, while the 1Lb (or 1B3u) state results from a
combination of HOMO 1 ! LUMO and HOMO ! LUMO + 1 transitions. Fur-
ther, excitations to 1La (1B2u) are short axis polarized with high intensity whereas
the transitions to 1Lb (1B3u) are long axis polarized with low intensity.
It follows from Fig. 3a that the experimental energy gap ΔE ¼ ΔE 1 B2u 2u Δ
1
E B3u 3u starts out positive at naphthalene (nr ¼ 2) with ΔE ¼ 0.53 eV before
turning negative at anthracene (nr ¼ 3) where ΔE ¼ 0.04 eV: For larger linear acenes
ΔE becomes increasingly negative, reaching ΔE ¼ 0.85 eV at hexacene. Thus,
experimentally, ΔE(1B2u) is seen to drop faster in energy than ΔE(1B3u).
Constricted Variational Density Functional Theory Approach to the. . . 73
Table 2 Root mean square deviations (dRMSD) from experiment for 1La and 1Lb π ! π*
excitations calculated by TDDFT, TDDFT-TD and CV(1)-TD for linear and nonlinear acenes
using LDA
RMSD 1B2u (1La) RMSD 1B3u (1Lb)
Systems TDDFT TDDFT-TDa CV(1)-TD TDDFT TDDFT-TDa CV(1)-TD
b
Linear 0.71 0.49 0.06 0.14 0.13 0.13
Nonlinearc 0.52 0.40 0.24 0.16 0.15 0.19
a
Identical to CV(2)-TD
b
Linear acenes of Fig. 2
c
Nonlinear acenes of Fig. 4
d
Ev
Table 3 CV(1)-TD singlet excitation energies (in eV) for linear acenes with LDA
Experimenta CV(1)-TD
Number of rings 1
B2u 1
B3u ΔEb 1
B2u 1
B3u ΔEb
2 4.66 4.13 0.53 4.73 4.39 0.34
3 3.60 3.64 0.04 3.68 3.73 0.05
4 2.88 3.39 0.51 2.91 3.32 0.41
5 2.37 3.12 0.75 2.35 3.03 0.68
6 2.02 2.87 0.85 1.93 2.82 0.89
a
Platt [74]
ΔE ¼ ΔE(1B2u) ΔE(1B3u)
b
1 2 3 4 5
6 7 8 9 10
11 12 13 14 15
Fig. 4 Nonlinear acenes with up to six fused rings (nr ¼ 6) considered in this study
and 1Lb are given in Table 2 for TDDFT, CV(2)-TD, and CV(1)-TD based on
LDA. 1La CV(1)-TD with an RMSD of 0.24 eV is seen to perform better than CV
(2)-TD (RMSD ¼ 0.40 eV) and especially TDDFT (0.52 eV). In fact our results are
of a similar quality to the best results obtained by long-range corrected (LRC)
functionals [79]. We note again that the 1La transitions involves a single
HOMO ! LUMO orbital displacement with γ i ¼ π=2. For 1Lb all three methods
perform equally well with RMSDs of 0.16 eV (TDDFT), 0.15 eV(CV(2)-TD), and
0.19 eV (P-CV(1)-DFT), respectively. It is interesting to note that the LRC-
functionals [79] in this case perform much more poorly with RMSDs around
0.4 eV. Thus, CV(1)-TD at the simple LDA level is the only scheme of the
methods discussed here which gives a balanced description of π ! π* transitions
involving a single orbital displacement (1La) and π ! π * transitions with more than
one displacement.
Constricted Variational Density Functional Theory Approach to the. . . 75
For the CV(1)-scheme we note that the singlet excitation energy for a transition
involving a single promotion i ! a such as 1 A1g 1g ! 1 B2u 2u according to (27)–
(29) has the simple form [27]:
CV ð1Þ 1 1
ΔES ð i ! aÞ ¼ ε a ε i þ K aaaa þ K iiii þ 2K aaii K iiaa ð30Þ
2 2
Here i is the HOMO π and a the LUMO π* in the current study. Further, a bar “–”
indicates an orbital of β-spin. For CV(2)-TD we obtain for the same transition
according to (14) and (27)
CV ð2Þ
ΔES ði ! aÞ ¼ εa εi þ 2K aiai K aiai ð31Þ
For HF these two expressions are identical [27] because K aaaa ¼ K iiii ¼ 0 and
K aiai ¼ K aaii . However, for any of the popular functionals this is not the case.
Thus, the two expressions give rise to different excitation energies for the same
functional. In the study at hand [27] on the 1 A1g 1g ! 1 B2u 2u transition the sum of
the K-integrals in (30) is larger than the sum of the K-integrals in (31) by 0.5 eV,
giving rise to the better performance of CV(1)-TD compared to CV(2)-TD for the
1
A1g 1g ! 1 B2u 2u transition. We note that acenes have also been well described by
the variational DFT-based spin-restricted ensemble referenced Kohn–Sham
(REKS) method [57].1
Using the U matrix from TDDFT-TD or CV(2)-TD to calculate ΔEM of (28) or ΔET
of (29), as is done in P-CV(1)-TD, might be a good approximation. However,
ultimately, one would want to use a U matrix which actually minimizes ΔEM of (28)
or ΔET of (29). Such a procedure leads us to self-consistent CV(1)-DFT (SCF-CV
(1)-DFT) [27] which we discuss next. We note that the U matrix can also be
!
organized as a vector U with pairs “ai” of occupied and virtual orbitals as running
numbers. The two formulations are used interchangeably in the following.
From the occupied excited state orbitals of (17) and (23) we can express the
electron density and spin matrices.2,3 Starting with a spin-conserving transition
from a close shell ground state, we can, without loss of generality, assume that it
1
See the chapter “Ensemble DFT approach to excited states of strongly correlated molecular
systems” by M. Filatov.
2
See Sect. 3.1 from part S1 of supporting information in Ziegler et al. [27].
3
See Sect. 3.3 from part S1 of supporting information in Ziegler et al. [27].
76 T. Ziegler et al.
takes place between orbitals of α-spin. Thus, for such a spin-conserving transition,
0
the excited state orbitals ϕi ¼ cos ½ηγ i ϕoi α þ sin ½ηγ i ϕvi α (15) are obtained by the
unitary transformation at (2a) and (2b) to all orders involving the part of the
U matrix (Uαα) which, according to (1), mixes occupied ground state orbitals of
α-spin with virtual ground state orbitals of α-spin. We can now write (see footnote
3) the change in density within the α-manifold caused by the excitation as
0 X
occ=2 h ih i
ΔρM 1; 1 ¼ sin 2 ηα γ jα φvj α ð10 Þφvj α ð1Þ φoj α ð10 Þφoj α ð1Þ
j
X
occ=2 h i h ih i
þ sin ηα γ jα cos ηα γ jα φvj α ð1Þφoj α ð10 Þ þ φvj α ð10 Þφoj α ð1Þ
j
ð32aÞ
In (32a) the scaling factor ηα is introduced to ensure that Δραð1Þ ð1, 10 Þ represents
the transfer of a single electron from the occupied orbital space density
Xocc=2 h i
j sin 2 ηα γ jα φoj α ð10 Þφoj α ð1Þ to the virtual orbital space density
Xocc=2 h i
j
sin 2 ηα γ jα φvj α ð10 Þφvj α ð1Þ or
Xocc=2 h i
j
sin 2 ηα γ jα ¼ 1 ð32bÞ
ΔEM E
ð KS ½ρ =2 þ ΔρM , ρ =2 EKS ½ρ =2, ρ =2
0 0 0 0
0
ð33Þ
¼ FKS ρ =2 þ 1=2ΔρM , ρ0 =2 ΔρM
Here the right hand side of (33) is derived by Taylor expanding [80]
EKS ½ρ0 =2 þ ΔρM , ρ0 =2 and EKS[ρ0/2, ρ0/2] from the intermediate point
ðρ =2 þ ΔρM =2, ρ =2Þ. Further, FKS ðρ0 =2 þ ΔρM =2, ρ0 =2Þ is a Kohn-Sham Fock
0 0
operator defined with respect to the intermediate point. The expression in (33) is
exact to third order in ΔρM which is usually accurate enough [80]. However, its
accuracy can be extended to any desired order [80]. Taylor expanding
FKS ðρ0 =2 þ ΔρM =2, ρ0 =2Þ to second order in ΔρM finally affords ΔEM of (28)
(see footnotes 2 and 3). The expression for ΔET of (29) can be derived along
similar routes.4,5
4
See Sect. 3.2 from part S1 of supporting information in Ziegler et al. [27].
5
See Sect. 3.4 from part S1 of supporting information in Ziegler et al. [27].
Constricted Variational Density Functional Theory Approach to the. . . 77
! ðI Þ
We now find vectors U which optimize ΔEM and ΔET. In either case there are
several which we order in terms of increasing energy with I ¼ 1, 2. . .. To this end,
!
we need the energy gradient with respect to variations in U . Considering first a spin-
conserving transition6 between orbitals of α-spin, we take as a starting point Uαα
0
6
See Sect. 4.1 from part S1 of supporting information in Ziegler et al. [27].
78 T. Ziegler et al.
!α, e 0, αα
δΔEM dΔρα
g ai U ¼ dΔEM
dΔUαα ¼ δρα 0 dΔUαα
ð
∂Δρ α
ai 0 ai 0
ð35Þ
α
¼ FKS ρ0 =2 þ 1=2ΔρM , ρ =2 dUαα dv1 ;
0 M
ai 0
Here ΔUαα αα
ai is the change in Uai in going from Uai
0;αα
and Δρα is the corresponding
α
change in ρ . The subscript “0” in (35) indicates that the derivatives are evaluated at
0, αα !α, e 0, αα
U αα
ai ¼ U ai . The calculation of g ai U in (35) requires closed form expres-
sions for dΔρα/dΔUαα ai (see footnote 6). 7
!α, e
With the evaluation of g ai U0, αα we can now begin an iterative process from
! ðI Þ
U0,αα generated by Uð2Þ to the optimal Uαα matrix where ΔU αα ¼ 0. A differenti-
ation of (35) by ΔUαα affords
!e, α
g U0, αα þ Hαα U0, αα ΔUαα ¼ 0 ð36Þ
αα
!
from which we can find the next Uαα. In the initial steps where ΔU
δtresh1 the
Hessian is calculated approximately by assuming that Hαα U0, αα ¼ εD with
ðεD Þai, bj ¼ δij δab ðεa εi Þ. Here εi, εa are the energies of the occupied and virtual
ground state orbitals, respectively. We thus get for each new step
! αα 1 !e, σ 0, αα
ΔU ¼ εD g U ð37Þ
Xocc=2 h i
If sin 2 α α
η γ e 0, αα ¼ U0, αα þ ΔUαα does not satisfy (32b),
resulting from U
j
j
Xocc=2 h i
we introduce a new ηα scaling so that sin 2 α α
η γ b 0, αα
constructed from U
j j
X
I1
b 0, αα
U0, αα ¼ U b 0, αα =Tr UK, ααþ UK, αα
U K, αα Tr UK, ααþ U ð38Þ
k¼1
7
See Sect. 3.0 from part S2 of supporting information in Ziegler et al. [27].
Constricted Variational Density Functional Theory Approach to the. . . 79
0,αα
After that, we
go back to (37) for a new step with U defined in (38). When
!αα
δtresh1 ΔU
δtresh2 , the iterative procedure is resumed by the help of the
conjugated gradient technique described by Pople et al. [81]. It is not required in
this procedure explicitly to know the Hessian. Instead, use is made of the fact that
!e, α 0, αα !e, α 0, αα
Hαα U0, αα ΔUαα ¼ g U þ ΔU g U þ O½3 ðΔUÞ ð39Þ
The value for δtresh1 is typically 102 whereas δtresh2 ¼ 104 . Convergence is
obtained when the threshold δtresh2 is reached. Typically 20–30 iterations are
required to reach δtresh1 and 5–10 to reach δtresh2. We have also attempted more
advanced Hessians for the first part of the optimization, such as the one suggested
by Fletcher [82] and implemented by Fischer and Almlöf [83]. However, it was
found to be less robust than the simple procedure in (37). The optimization
procedure outlined here for spin-conserving transitions can readily be formulated
for spin-flip transitions [27].
In the description of the excited state by the SCF-CV(1)-DFT scheme all occupied
β-orbitals are unchanged (frozen) from the ground state and the same is the case for
a number of α-orbitals which do not directly participate in the transition. Thus in the
case of the nσ ! π* transitions, all α-orbitals other than nσ , π * are frozen. To
remedy this, we allow in the RSCF-CV(1)-DFT [26] scheme for a relaxation to
second order in the mixing matrix Rσσ ðσ ¼ α, βÞ of all occupied orbitals in the
excited state [26]. Thus,
X
vir=2
1X X σσ σσ σ
vir=2 occ=2
ψ iσ ð1Þ ! ϕiσ ð1Þ þ Rσσ σ
ci ϕc ð1Þ R R ϕ ð1Þ þ Oð3Þ ½Rσ ð40aÞ
c
2 c k ci ck k
X
vir=2
1X X σσ σσ σ
vir=2 occ=2
ψ aσ ð1Þ ! ϕaσ ð1Þ Rσσ σ
ak ϕk ð1Þ R R ϕ ð1Þ þ Oð3Þ ½Rσσ ð40bÞ
k
2 c k ak ck c
Replacing in (2a)
the matrix Ũ whichcombines occupied and virtual orbitals of
the unrelaxed set ϕq ; q ¼ 1, occ þ vir with the corresponding matrix U which
mixes the occupied and virtual orbitals of the relaxed basis ψ q ; q ¼ 1, occ þ vir
leads to the unitary transformation
Constricted Variational Density Functional Theory Approach to the. . . 81
!
0
ψ occ ψ occ X1
Um ψ occ ψ occ
Y ¼e U
¼ ¼ 0 ð41Þ
ψ vir ψ vir m¼0
m! ψ vir ψ vir
where
X ðαÞocc
X ðαÞ
vir
ΔρM ðUαα Þ ¼ ΔPai ðUαα Þ ϕaα ð10 Þϕiα ð1Þ þ ϕaα ð10 Þϕiα ð1Þ
a i
ð44Þ
X
vir ðαÞ X
occ ð αÞ
αα
þ ΔPab ðU Þϕaα ð10 Þϕbα 1 þ ΔPij ðUαα Þϕiα ð10 Þϕjα 1
ab ij
is the change in density caused by Uαα alone and equivalent to (32a) but expressed
in terms of unrelaxed ground state orbitals. On the other hand
X X
α, β vir X
ðσ Þocc ðσ Þ
ð1Þσσ α 0
ð1Þαα
X
occ ðαÞ
ð1Þββ
T ai ¼ Rαα
ai þ ΔPαα αα
ij Raj ; T ai ¼ Rββ
ai ð46aÞ
j
Constricted Variational Density Functional Theory Approach to the. . . 83
ð2Þαα
X
occ ð αÞ X
occ X
ðαÞocc ðαÞ
T ab ¼ Rαα αα
ai Rbi þ ΔPαα αα αα αα
ij ðU ÞRai Rbj
i i j
ð46bÞ
ð2Þββ
X
virð βÞ
T ab ¼ Rββ ββ
ai Rbi
a
ð2Þαα
X
vir ðαÞ X
occ X
ðαÞvirð αÞ
T ij ¼ Rαα αα
ai Raj ΔPαα αα αα αα
il ðU ÞRal Raj
a l a
ð46cÞ
ð2Þββ
X
virðβ Þ
T ij ¼ Rββ ββ
ai Raj
i
where
X
virðαÞ X
occ ðαÞ
ΔEM ðUÞ ¼ εaα ΔPαα αα 2
aa ðU Þ εiα ΔPαα αα 2
ii ðU Þ
a i
X
vir X
ðαÞocc ð αÞ
þ ΔPαα αα αα αα
ai ðU ÞΔPbj ðU Þ K aα iα bα jα þ K aα iα bα jα
ab i
1X
occðαÞ
þ ΔPαα αα αα αα
ij ðU ÞΔPkl ðU ÞK iα jα kα lα
2 ijkl
1Xvir
þ ΔPαα ðUαα ÞΔPαα αα
cd ðU ÞK aα bα cα d α ð48Þ
2abcd ab
X
vir X
ðαÞocc ð αÞ
þ ΔPαα αα αα αα
ab ðU ÞΔPij ðU ÞK aα bα iα jα
ab ij
X
vir X
ðαÞocc ð αÞ
þ2 ΔPαα αα αα αα
ab ðU ÞΔPck ðU ÞK aα bα cα kα
abc k
X
vir ðαÞocc
X ð αÞ
þ2 ΔPαα αα αα αα
ij ðU ÞΔPck ðU ÞK iα jα cα kα
ijk c
is the excitation caused by Uαα alone and equivalent to ΔEM for the SCF-CV(1)-
DFT scheme of (28) but expressed in terms of canonical and unrelaxed ground state
orbitals. Further
84 T. Ziegler et al.
!
X X
α, β virðσ Þ X
vir ðσ Þ
ð2Þσσ
ΔEMR ¼ T ðaa2Þσσ εaσ T ii εiσ
σ a i
X
α, β X X
α, β vir X
ðσ Þocc X
ðσ Þvir ðμÞ h i
ð1Þσσ ð1Þμμ
þ T ai T bj K aσ iσ bμ jμ þ K aσ bμ iσ jμ
σ μ a c
i
"virðαÞ #
X X
α, β occ X
ðσ Þvirðσ Þ
ð1Þσσ
X X
occ ðαÞ
þ T ck ΔPαα αα
ab ðU ÞK aα bα cσ kσ þ ΔPαα αα
ij ðU ÞK iα jα cσ kσ
σ k c ab ij
ð49Þ
is the relaxation contribution to the excitation energy. The total energy for ΨM is
given as EM ¼ E0 ðρ0 Þ þ ΔEM ðUÞ þ ΔEMR where E0(ρ0) is the ground state energy
expressed in terms of unrelaxed orbitals. The expression for ΔERM is derived after
orthogonalization of ΨM to the ground state to second order in R.
We optimize ΔEM of (49) by first performing a Taylor expansion from the
starting point reference (U0,αα, R0,αα, R0,ββ) to Uαα ; Rαα ; Rββ ¼ U0, αα þ ΔUαα ,
R0, αα þ ΔRαα , R0, ββ þ ΔRαα Þ:
EM Uαα ;Rαα ;Rββ ¼ EM U0, αα ;R0, αα ;R0, ββ
!
X
dEM Xα, β X
dEM
XX d 2 EM
σσ 1
þ αα
αα
ΔU ai þ σσ ΔRai þ αα αα ΔUαα αα
ai ΔU bj
dU ai σ
dR ai 2 dU ai dU bj
ai 0 ai
! 0 ai bj
! 0
1X X X X XXX
α, β α, β α, β
d 2 EM d 2
E
ΔRσσ ττ
ΔU αα σσ
M
þ ai ΔRbj þ ai ΔRbj
2 ai bj σ τ dRσσ ai dRbj
ττ
ai bj σ
dU αα
ai dR σσ
bj
0 0
þO½3
ð50Þ
Here the subscript “0” indicates that the derivative is evaluated at the reference
(U0,αα,R0,αα,R0,ββ). We can alternatively write the expansion in terms of energy
gradients and energy Hessians as
0 αα 1
!U
g
!αα !αα
!ββ B Rαα C
EM U αα ; Rαα ; Rββ ¼ EM U 0, αα ; R0, αα ; R0, ββ þ ΔU B! C
ΔR ΔR @g A
ββ
!R
g
0 αα αα αα αα αα 10 Rββ , Rββ 1
ββ
1 !αα HU , U HU , R HU , R H αα
Rαα , Rββ AB C
! αα ! αα !
þ @ HRαα , Uαα HRαα , Rαα @ ΔR A þ O½3
2 ΔU ΔU ΔU H ββ
ββ αα ββ αα ββ ββ !
HR , U HR , R HR , R ΔR
ð51Þ
αα αα ββ
!U !R !R αα αα
where the expressions for the gradients g , g , g and Hessians HU , U ,
Rαα , Uαα
H , etc. can be obtained by a comparison between (50) and (51). Specific
Constricted Variational Density Functional Theory Approach to the. . . 85
αα αα ββ
!U !R !R
formula for g , g , and g are also given in [26] for the spin-flip transition.
For the spin conserving transition, a differentiation of (51) with respect to the
! αα !αα !ββ
individual components of ΔU , ΔR , and ΔR affords, after rearrangement,
0 αα 1
0 αα αα 10 !αα 1
!U αα αα αα ββ
ð 0Þ
C B H αα αα ð0Þ HU , R ð0Þ HU , R ð0Þ ΔU
g U ,U
B !Rαα CB !αα C
B g ð 0Þ C þ @ H R , U ð 0Þ Rαα , Rαα
ð 0Þ HRαα , Rββ
ð0Þ A@ ΔR A ¼ 0 ð52Þ
@ A H
ββ ββ αα
Rββ , Rαα Rββ , Rββ !ββ
!R
g ð 0Þ H R , U ð 0Þ H ð 0Þ H ð 0Þ ΔR
! αα !αα !ββ
from which we can find ΔU , ΔR , ΔR iteratively. More details can be
found in [26] which also covers the case of spin-flip transitions.
It follows from Table 4 that the RSCF-CV(1)-DFT scheme with full orbital
relaxation gives nσ ! π* transition energies which on average are within 0.15 eV
of the ΔSCF results. This is acceptable given the fact that the RSCF-CV(1)-DFT
scheme is only second order in relaxation and that it satisfies constraints not
fulfilled by ΔSCF. In comparison to the “Best” ab initio results [73], RSCF-CV
(1)-DFT fares as well as ΔSCF and CV(2)-TD (TDDFT-TD) with an RMSD of
0.32 eV. Thus, although RSCF-CV(1)-DFT is somewhat more costly (~twice) for
each transition, it does not fare much better than CV(2)-TD in those cases where the
latter is reliable and fares well. However, what we show shortly is that RSCF-CV
(1)-DFT has a similar accuracy (RMSD ~0.3–0.2 eV) where CV(2)-TD fails such
as Rydberg and charge transfer transitions.
Table 5 Root mean square deviations of Rydberg excitation energiesa calculated with RSCF-CV
(1)-DFT using five functionals with the extended basis set [66]
Functionals
Species Nr. of States LDA BP86 B3LYP LCBP86b LCBP86*c
N2 5 0.27 0.34 0.05 0.23 0.62
CO 7 0.22 0.43 0.13 0.12 0.37
CH2O 8 0.21 0.28 0.12 0.20 0.34
C2H2 8 0.31 0.50 0.52 0.25 0.24
H2O 10 0.27 0.17 0.14 0.21 0.24
C2H4 13 0.15 0.20 0.28d 0.28 0.29
Be 6 0.45 0.60 0.47 0.31 0.23
Mg 6 0.18 0.35 0.19 0.13 0.12
Zn 8 0.18 0.25 0.27 0.34 0.46
Average root mean square deviation 0.24 0.32 0.24 0.23 0.32
a
Energies in eV
b
Refers to LC functional combined with BP86 and ω ¼ 0.4
c
Represents LC functional combined with BP86 and ω ¼ 0.75
d
Comprised of 12 states
It has been demonstrated that regular adiabatic TDDFT employing the general
gradient approximation (GGA) as well as hybrid functionals with a fraction (α) of
exact Hartree–Fock exchange included ð0:0 α 0:5Þ underestimate charge
transfer excitation energies by as much as 2–4 eV [24, 25, 33]. This failure has
been discussed and analyzed extensively [24, 25, 32, 33]. By contrast, ATDDFT in
conjunction with long range corrected (LC) functionals affords charge transfer
excitation energies in good agreement with experiment [33]. In these functionals,
Hartree–Fock exchange is given a growing weight towards longer inter-electronic
distances.
We have recently [30] applied the RSCF-CV(1)-DFT scheme to a series of
charge transfer molecular complexes (CTMC) of the type X-TCNE where an
aromatic molecule (X ¼ benzene, toluene, o-xylene, naphthalene, anthracene) is
bound to tetracyanoethylene (TCNE) [33]. All of these complexes have one or more
distinct charge transfer transitions involving the excitation of an electron from an
Constricted Variational Density Functional Theory Approach to the. . . 87
(1)-DFT where the energy is minimized with respect to both R and U. After full
optimization in RSCF-CV(1)-DFT, the calculated excitation energies are lowered
from CV(1)-TD to values in reasonable agreement with experiment. The best fit is
provided by B3LYP (3.56 eV) and the largest deviation is observed for LC-BP86
(3.10 eV). We must conclude that the RSCF-CV(1)-DFT method in general gives
reasonably good agreement with experiment for the different DFT schemes. Thus
the RSCF-CV(1)-DFT energy expressions of (28) and (29) seem to be relatively
robust with respect to the choice of functional; see Table 6. The relaxation brings
the calculated excitation energy to 2.85 eV for RSCF-CV(1)-HF; see Table 6 [30].
We present in Table 7 [30] RSCF-CV(1) results for calculations on X-TCNE
adducts I-IV of Fig. 7 using both local functionals and hybrids together with long
range corrected (LC) functionals. We notice again that the standard functionals
LDA, BP86, B3LYP, and BHLYP all are close to experiment. The LC-BP86
functional fares somewhat worse here. However, we have not optimized the LC
parameter which usually improves the results [33]. It should be noted that the
corresponding ATDDFT results are off by 2 eV for LDA, BP86, 1 eV for B3LYP
and BHLYP [30, 33]. For optimized LC functionals the ATDDFT results are in
excellent agreement with experiment [33].
Constricted Variational Density Functional Theory Approach to the. . . 89
Table 8 Singlet excitation energies (in eV) for π (donor) to π* (TCNE) transitions in
X-anthracene complexes based on a TZP-basis and the RP-CV(4) scheme with different
functionals
Substituents(X) LDA BP86 BLYP BPErev SAOP GRAC SKB Exp.
None 1.69 1.71 1.66 1.73 1.60 1.71 1.82 1.73
9,10-Dimethyl 1.43 1.46 1.41 1.47 1.34 1.45 1.77 1.44
9-Carbo-methoxy 1.74 1.78 1.70 1.80 1.71 1.77 1.84 1.84
9-Chloro 1.74 1.78 1.71 1.80 1.66 1.78 1.82 1.74
9-Cyano 2.00 2.03 1.96 2.04 1.97 2.00 2.03 2.01
9-Formyl 10-chloro 2.02 2.06 1.99 2.08 1.80 2.06 1.96 1.96
9-Formyl 1.99 2.03 1.97 2.05 1.97 2.04 1.95 1.90
9-Methyl 1.48 1.50 1.45 1.51 1.44 1.49 1.71 1.55
9-Nitro 1.94 1.97 1.92 1.99 1.96 1.98 2.12 2.03
RMSD 0.06 0.07 0.08 0.07 0.10 0.07 0.10 1.73
VWN [90], BP86 [91, 92], BLYP [91, 93], revPBE [94–96], SAOP [97], SKB [33], GRAC [98]
calculated by the GGA functionals BP86, BLYP, and revBEP differ on average by
less than 0.05 eV from LDA estimates and introducing SAOP or GRAC with the
right asymptotic 1/r behavior does not lead to any significant change. Charge
transfer transitions can also be described well by the variational DFT-based spin-
restricted ensemble referenced Kohn–Sham (REKS) method [57] (see footnote 1).
It is at this point important to note that the experimental excitation energies for
the anthracene systems were all obtained in solution with CH3Cl as the solvent. We
do not expect the solvent effect to be significant. In fact, theoretical calculations
[33] using a continuum model revealed that the excitation energies were lower by
only 0.05 eV. We have, as a consequence, decided to compare our gas-phase results
directly with the experimental solvent data. We obtain from such a comparison that
the RMSD is 0.06 for LDA followed by 0.07 for BP86, revPBE, and 0.08 for BLYP.
The two 1/r asymptotically corrected functionals afford 0.10 for SAOP and 0.07 for
CRAC. Stein et al. [33] “SKB” introduced in their DZP gas-phase study a uniform
correction of 0.32 eV to simulate solvation effects; see Table 8. The magnitude
and sign of this correction was given without much explanation [33]. After applying
their correction the authors obtained an RMSD of 0.1 which is quite similar to the
one found here for our gas-phase results without any solvent correction. It should be
noted that ATDDFT with the same functionals carries errors of up to 1 eV.
3 Concluding Remarks
Acknowledgement T.Z. would like to thank the Canadian government for a Canada research
chair in theoretical inorganic chemistry and NSERC for financial support.
References
theory) from an excited state variational approach based on the ground state functional. J
Chem Phys 140:18A502
31. Ziegler T, Krykunov M (2010) On the calculation of charge transfer transitions with standard
density functionals using constrained variational density functional theory. J Chem Phys
133:074104
32. Ziegler T, Seth M, Krykunov M, Autschbach J, Wang F (2008) A revised electronic Hessian
for approximate time-dependent density functional theory. J Chem Phys 129:184114
33. Stein T, Kronik L, Baer R (2009) Reliable prediction of charge transfer excitations in
molecular complexes using time-dependent density functional theory. J Am Chem Soc
131:2818
34. Cave RJ, Zhang F, Maitra NT, Burke K (2004) A dressed time-dependent density functional
treatment of the 21A states of butadiene and hexatriene. Chem Phys Lett 389:39
35. Mazur G, Wlodarczyk R (2009) Application of the dressed time-dependent density functional
theory for the excited states of linear polyenes. J Comp Chem 30:811
36. Elliott P, Goldson S, Canahui C, Maitra NT (2011) Perspectives on double-excitations in
TDDFT. Chem Phys 391:110
37. Slater JC, Wood JH (1971) Statistical exchange and the total energy of a crystal. Int J Quant
Chem Suppl 4:3
38. Slater JC (1972) Statistical exchange-correlation in the self-consistent field. Adv Quant Chem
6:1
39. Kowalczyk T, Yost SR, Van Voorhis T (2011) Assessment of the ΔSCF density functional
theory approach for electronic excitations in organic dyes. J Chem Phys 134:054128
40. Ziegler T, Rauk R, Baerends EJ (1977) On the calculation of multiplet energies by the
Hartree-Fock-Slater method. Theor Chim Acta 43:261
41. Ziegler T, Rauk A, Baerends EJ (1976) The electronic structure of tetrahedral oxo-
complexes. The nature of the “charge transfer” transitions. J Chem Phys 16:209
42. Gilbert A, Besley N, Gill P (2008) J Phys Chem A 122:13164
43. Besley N, Gilbert A, Gill P (2009) Self-consistent-field calculations of core excited states. J
Chem Phys 130:124308-1
44. Park YC, Krykunov M, Seidu I, Ziegler T (2014) On the relation between adiabatic time
dependent density functional theory (TDDFT) and the ΔSCF-DFT method. Introducing a
numerically stable ΔSCF-DFT scheme for local functionals based on constricted variational
DFT. Mol Phys 112:661
45. Gavnholt J, Olsen T, Engelund M, Schiøtz J (2008) Δ self-consistent field method to obtain
potential energy surfaces of excited molecules on surfaces. J Phys Rev B 78:075441/1
46. Liu TQ, Han WG, Himo FG, Ullmann M, Bashford D, Toutchkine A, Hahn KM, Noodleman
L (2004) Density functional vertical self-consistent reaction field theory for solvatochromism
studies of solvent-sensitive dyes. Phys Chem A 108:3545
47. Ceresoli D, Tosatti E, Scandolo S, Santoro G, Serra S (2004) Trapping of excitons at
chemical defects in polyethylene. J Chem Phys 121:6478
48. Zhekova H, Seth M, Ziegler T (2014) Application of time dependent and time independent
density functional theory to the first π to π* transition in cyanine dyes. Int J Quant Chem
114:1019
49. Ziegler T (2011) A chronical about the development of electronic structure theories for
transition metal complexes. Struct Bond 47:1
50. von Barth U (1979) Local-density theory of multiplet structure. Phys Rev A 20:1693
51. Gunnarsson O, Lundqvist BI (1976) Exchange and correlation in atoms, molecules, and
solids by the spin-density-functional formalism. Phys Rev B 13:4274
52. Levy M, Perdew JP (1985) Hellmann-Feynman, virial, and scaling requisites for the exact
universal density functionals. Shape of the correlation potential and diamagnetic suscepti-
bility for atoms. Phys Rev A 32:2010
53. Gaudoin R, Burke K (2005) Lack of Hohenberg-Kohn theorem for excited states. Phys Rev
Lett 93:173001
94 T. Ziegler et al.
54. Oliveira LN, Gross EKU, Kohn W (1988) Density-functional theory for ensembles of
fractionally occupied states. II. Application to the He atom. Phys Rev A 37:2821
55. Filatov M, Shaik S (1998) Spin-restricted density functional approach to the open-shell
problem. Chem Phys Lett 288:689
56. Filatov M, Shaik S (1999) Spin-restricted density functional approach to the open-shell
problem. Chem Phys Lett 304:429
57. Filatov M, Huix-Rotllant M (2014) Description of electron transfer in the ground and excited
states of organic donor–acceptor systems by single-reference and multi-reference density
functional methods. J Chem Phys 141:024112
58. Gidopoulos NI, Papaconstantinou PG, Gross EKU (2002) Density-functional theory for
ensembles of fractionally occupied states. II. Application to the He atom. Phys Rev Lett
88:03300
59. Gross EKU, Oliveira LN, Kohn W (1988) Spurious interactions, and their correction, in the
ensemble-Kohn-Sham scheme for excited states. Phys Rev A 37:2809
60. Levy M, Nagy A (1999) Variational density-functional theory for an individual excited state.
Phys Rev Lett 83:4361
61. Görling A, Levy M (1993) Correlation-energy functional and its high-density limit obtained
from a coupling-constant perturbation expansion. Phys Rev B 47:13105
62. Ziegler T, Krykunov M, Autschbach J (2014) Derivation of the RPA (random phase approx-
imation) equation of ATDDFT (adiabatic time dependent density functional ground state
response theory) from an excited state variational approach based on the ground state
functional. J Chem Theory Comput 10:3980
63. Ziegler T, Krykunov M, Cullen J (2011) The application of constricted variational density
functional theory to excitations involving electron transitions from occupied lone-pair
orbitals to virtual π* orbitals. J Chem Theory Comput 7:2485
64. Krykunov M, Grimme S, Ziegler T (2012) Accurate theoretical description of the 1La and 1Lb
excited states in acenes using the all order constricted variational density functional theory
method and the local density approximation. J Chem Theory Comput 8:4434
65. Zhekova H, Krykunov M, Autschbach J, Ziegler T (2014) Applications of time dependent
and time independent density functional theory to the first π to π* transition in cyanine dyes. J
Chem Theory Comput 10:3299
66. Seidu I, Krykunov M, Ziegler T (2014) Applications of time--dependent and time-- indepen-
dent density functional theory to Rydberg transitions. J Phys Chem A ASAP. doi:10.1021/
jp5082802
67. Wang F, Ziegler T (2004) Time-dependent density functional theory based on a noncollinear
formulation of the exchange-correlation potential. J Chem Phys 121:12191-1
68. Wang F, Ziegler T (2005) The performance of time-dependent density functional theory
based on a noncollinear exchange-correlation potential in the calculations of excitation
energies. J Chem Phys 122:074109-1
69. Wang F, Ziegler T (2006) Use of noncollinear exchange-correlation potentials in multiplet
resolutions by time-dependent density functional theory. Int J Quant Chem 106:2545–2550
70. Hirata S, Head-Gordon M (1999) Time-dependent density functional theory within the
Tamm-Dancoff approximation. Chem Phys Lett 291:314
71. Amos AT, Hall GG (1961) Single determinant wave functions. Proc R Soc A 263:483
72. Martin RLJ (2003) Natural transition orbitals. J Chem Phys 118:4775
73. Schreiber M, Silva-Junior M, Sauer S, Thiel W (2008) Benchmarks for electronically excited
states: CASPT2, CC2, CCSD, and CC3. J Chem Phys 128:134110
74. Platt JR (1949) Classification of spectra of Cata condensed hydrocarbons. J Chem Phys
17:484
75. Grimme S, Parac M (2003) Substantial errors from time-dependent density functional theory
for the calculation of excited states of large π systems. Chemphyschem 4:292
76. Parac M, Grimme S (2003) TDDFT of the lowest excitation energies of polycyclic aromatic
hydrocarbons. Chem Phys 292:11
Constricted Variational Density Functional Theory Approach to the. . . 95
77. Jacquemin D, Wathelet V, Perpete EA, Adamo C (2009) Assessment of functionals for
TD-DFT calculations of singlet-triplet calculations. J Chem Theory 5:2420
78. Goerigk L, Grimme S (2010) Assessment of TD-DFT methods and of various spin scaled CIS
(D) and CC2 versions for the treatment of low-lying valence excitations of large organic dyes.
J Chem Phys 132:184103
79. Richard RM, Herbert JM (2011) Time-dependent density-functional description of the 1La
state in polycyclic aromatic hydrocarbons: charge-transfer character in disguise? J Chem
Theory Comput 7:1296
80. Ziegler T, Rauk A (1977) On the calculation of bonding energies by the Hartree-Fock-Slater
method. Theor Chim Acta (Berl) 46:1
81. Pople JA, Krishnan R, Schlegel HB, Binkley JS (1979) Derivative studies in configuration
interaction theory. Int J Quant Chem S13:225
82. Fletcher R (1980) Practical methods of optimization, vol 1. Wiley, New York
83. Fischer H, Almlöf J (1992) General methods for geometry and wave function optimization. J
Phys Chem 96:9768
84. Prochorow J, Tramer AJ (1967) Photoselection study of charge transfer complexes. J Chem
Phys 47:775
85. Frey JE, Andrews AM, Ankoviac DG et al (1990) Charge-transfer complexes of tetracya-
noethylene with cycloalkanes, alkenes, and alkynes and some of their aryl derivatives. J Org
Chem 55:606
86. Merrifield RE, Phillips WD (1958) Cyanocarbon chemistry. II.1 Spectroscopic studies of the
molecular complexes of tetracyanoethylene. J Am Chem Soc 80:2778
87. Masnovi JM, Seddon EA, Kochi JK (1984) Electron transfer from anthracenes. Comparison
of photoionization, charge-transfer excitation and electrochemical oxidation. Can J Chem
62:2552
88. Hanazaki I (1972) Vapor-phase electron donor–acceptor complexes of tetracyanoethylene
and of sulfur dioxide. J Phys Chem 76:1982
89. Garcia-Cuesta I, Sanchez de Meras AMJ, Koch H (2003) Coupled cluster calculations of the
vertical excitation energies of tetracyanoethylene. J Chem Phys 118:8216
90. Vosko SH, Wilk L, Nusair M (1980) Accurate spin-dependent electron liquid correlation
energies for local spin density calculations: a critical analysis. Can J Phys 58:1200
91. Becke AD (1988) Density-functional exchange-energy approximation with correct asymp-
totic behavior. Phys Rev A 38:3098
92. Perdew JP, Wang Y (1986) Density-functional approximation for the correlation energy of
the inhomogeneous electron gas. Phys Rev B 33:8822
93. Lee C, Yang W, Parr RG (1988) Development of the Colle-Salvetti correlation-energy
formula into a functional of the electron density. Phys Rev B 37:785
94. Perdew JP, Burke K, Ernzerhof M (1996) Generalized gradient approximation made simple.
Phys Rev Lett 77:3865
95. Hammer B, Hansen LB, Norskøv JK (1999) Improved adsorption energetics within density-
functional theory using revised Perdew-Burke-Ernzerhof functionals. Phys Rev B 59:7413
96. Zhang Y, Yang W (1998) Comment on “generalized gradient approximation made simple”.
Phys Rev Lett 80:890
97. Gritsenko OV, Schipper PRT, Baerends EJ (1999) Approximation of the exchange-
correlation Kohn–Sham potential with a statistical average of different orbital model poten-
tials. Chem Phys Lett 302:199
98. Grüning M, Gritsenko OV, van Gisbergen SJA, Baerends EJ (2001) Shape corrections to
exchange-correlation Kohn-Sham potentials by gradient-regulated seamless connection of
model potentials for inner and outer region. J Chem Phys 114:652
99. Filatov M (2013) Assessment of density functional methods for obtaining geometries at
conical intersections in organic molecules. J Chem Theory Comput 9:4526
100. Ziegler T (1983) Extension of the statistical energy expression to multi-determinantal wave
functions. In: Dahl JP, Avery J (eds) Density functional theory of atoms, molecules and
solids. Plenum, New York
Top Curr Chem (2016) 368: 97–124
DOI: 10.1007/128_2015_630
# Springer International Publishing Switzerland 2015
Published online: 24 April 2015
Michael Filatov
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
2 Ensemble DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3 REKS Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.1 REKS Method for Ground States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.2 REKS Method for Excited States: SA-REKS and SI-SA-REKS . . . . . . . . . . . . . . . . . . . . 111
4 Applications of the REKS Method to Excited States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
M. Filatov (*)
Institut für Physikalische und Theoretische Chemie, Universität Bonn, Beringstr. 4, 53115
Bonn, Germany
e-mail: mike.filatov@gmail.com
98 M. Filatov
1 Introduction
The founding principles of density functional theory (DFT) were initially formu-
lated only for the ground states of fermionic many-body systems [1, 2]. It is
therefore commonly accepted that the excited states in the context of DFT can be
accessed by the use of some form of response formalism implemented, for instance,
in the time-dependent DFT (TD-DFT) methods [3, 4]. In principle, TD-DFT is a
rigorous formulation of the ground state DFT for time-dependent phenomena [5].
However the excitation energies of many-body systems are typically accessed with
the use of the linear response (LR) formalism, which assumes that the time
dependence stems from a weak (usually oscillatory) perturbing potential [3–5]. In
practice, LR-TD-DFT yields a very reasonable description of optical absorption
spectra with the use of the commonly available ground-state approximate density
functionals [3, 4]; however, some spectacular failures of the formalism are also
known. In particular, standard implementation of LR-TD-DFT relies on the adia-
batic approximation (i.e., locality of the exchange-correlation (XC) kernel in the
time domain) and consequently cannot take proper account of multiple excitations
[6, 7], which become important, e.g., for excited states of conjugated molecular
systems [8]. Yet another failure of the standard LR-TD-DFT to describe the excited
states of strongly correlated systems, e.g., H2 at stretched bondlength [9], can be
traced back to the use of the standard ground-state Kohn–Sham (KS) formalism [2]
which fails to take proper account of the non-dynamic electron correlation.
In the domain of wavefunction theory (WFT), the excited states of molecules can
be obtained from the ground-state response formalism as well as the variational
excited state formalism [10]. An appealing idea is to employ the (time-independent)
variational formalism to obtaining excitation energies in the context of DFT.
Indeed, the first attempts to calculate the excitation energies by taking the energy
differences between the variationally obtained ground state energy and the energy
of a state obtained by promoting an electron to unoccupied energy level, the so-
called ΔSCF approach,1 date back to the early 1970s [11, 12]. However, despite
some attempts to justify the ΔSCF approach for computing the energies of one-
electron transitions between the states of different spatial symmetry [13, 14], the
idea of variationally obtaining the energy of an individual excited state in the
context of DFT lacks firm theoretical background [15–17].
A rigorous way of developing time-independent formalism for obtaining exci-
tation energies in the context of DFT is offered by ensemble DFT [18, 19] which
1
For more details on ΔSCF, see the chapter “A Constricted Variational Density Functional Theory
Approach to the Description of Excited States” by T. Ziegler, M. Krykunov, I. Seidu, and Y. C.
Park.
Ensemble DFT Approach to Excited States of Strongly Correlated Molecular Systems 99
2 Ensemble DFT
The basic tenet of KS DFT is that any physical fermionic ground state density ρ(r)
can be uniquely mapped onto the ground state density ρs(r) of a fictitious system of
non-interacting particles moving in a suitably modified external potential vs(r). If
such a vs(r), which is also known as the KS potential, can be found, the respective
KS Hamiltonian Ĥs is minimized by a single Slater determinant (KS determinant)
constructed from the lowest-energy one-electron functions (KS orbitals) φs,i(r) and
the non-interacting density ρs(r) is
X
ρs ð r Þ ¼ 2φs, i ðrÞ2 ; 8εi μ ; ð1Þ
i
where εi are the respective eigenvalues, μ is the Fermi level, and a closed electronic
100 M. Filatov
shell is assumed [1, 2]. The physical density ρ(r) which can be mapped onto such a
non-interacting density ρs(r) is said to be non-interacting pure state v-representable,
or PS-VR for brevity [21]. Naturally, in the case of non-interacting particles, such a
pure state wavefunction is represented by a single KS determinant and one may
speak of a determinantal v-representability (D-VR) as well [21].
A general proof of the existence of such a KS potential vs(r) and of the PS-VR
property has never been achieved for an arbitrary physical density. By contrast,
rigorous theoretical arguments have been given in favor of an alternative represent-
ation of an arbitrary fermionic density ρ(r) by an ensemble (weighted sum) of a
finite number (M ) of the densities ρK(r) originating from the same physical external
potential vext(r) [19–21]:
X
M X
M
ρð r Þ ¼ λK ρK ðrÞ, λK 0, λK ¼ 1 : ð2Þ
K¼1 K¼1
X
ρs ðrÞ ¼ ni φs, i ðrÞ2 ; ð3Þ
i
n j ¼ 2, εj < μ
0 nk 2, X εk ¼ μ ; ð4Þ
ni ¼ N
i
that is, only a few KS orbitals which degenerate at the Fermi level μ are allowed to
have fractional occupations [26]. Alternatively, the ensemble density (3) can be
written down as in (2) as a weighted sum of the densities ρs,K (r) of several KS
determinants constructed from a common set of KS orbitals; the ensemble
weighting factors λK are then connected to the FONs in (3) via λK ¼ nk/m where
m is the number of electrons in the KS orbitals degenerate at the Fermi level and all
the KS orbitals in the determinant yielding the ρs,K (r) density are set doubly
occupied. Recently, the degeneracy of the fractionally occupied KS orbitals at the
Fermi level was rigorously proved [64].
For the ensemble density (2), Lieb proved [20] that the ground state energy is
given by a weighted sum
X
M
E½ ρ ¼ λK E½ρK ; ð5Þ
K¼1
of the energies E[ρK] of the ensemble components taken with the same weighting
factors as in (2). Englisch and Englisch proved the differentiability of the ensemble
energy E[ρ] with respect to the ensemble density, thus demonstrating the existence
of vs(r) and the ensemble KS reference system [21].
The energies of the non-interacting KS reference states constructed in [26] for
the C2 molecule and for the H2 + H2 system satisfy (5), provided that the ensemble
densities are allowed. If, however, one insisted on having PS-VR (or D-VR) KS
reference states for these molecular systems, holes below the Fermi level were
observed which implied the breakdown of the basic assumption behind the KS
method, namely that the density ρs(r) is constructed from the lowest one-particle
eigenstates of the non-interacting KS Hamiltonian. Besides that, the single deter-
minant KS states found in [26] (and in [28]) had somewhat higher energies than the
respective ensemble KS solutions. Thus, the ensemble KS solutions had to be
preferred on the grounds of the variational principle. These conclusions have
been fully confirmed by Morrison [28] in the study of Be isoelectronic series of
atomic ions, for which mapping of the exact densities onto the KS reference could
only be achieved with the use of ensemble densities, i.e., densities with the
fractional occupation numbers of the valence 2s and 2p atomic orbitals. An attempt
to formalize these observations and to develop ensemble variants of the KS theory
102 M. Filatov
was undertaken in [30]; however this did not materialize in the form of a practically
accessible computational scheme.
The ground-state ensemble DFT formalism was extended into the domain of
excited state calculations in the works of Theophilou [18] and Gross et al. [23] who
demonstrated that the Hohenberg–Kohn theorem is satisfied not only by the ground-
state density and the energy but also by the density and the energy of an ensemble of
several lowest energy states (i.e., the ground and excited states) of a many-body
fermionic system. In particular, Gross et al. [23] proved that the ensemble energy
(5) constructed from several lowest eigenstates of a many-body Hamiltonian Ĥ
satisfies the variational principle
X
M X M X
M
^ ΦK
λK ΦK H λ K EK ; 0 λK 1; λK ¼ 1 ; ð6Þ
K¼1 K¼1 K¼1
where Φk are the trial wavefunctions and EK are the exact eigenvalues of the
Hamiltonian Ĥ [23].
The variational character of the ensemble energy enables one to calculate
excitation energies rigorously using (formally ground-state) density functionals.
Considering only two state ensembles (the ground state E0 and the lowest excited
state E1), for which the energy and the density are given by (7) and (8),
the excitation energy ΔE ¼ E1–E0 can be obtained in two ways [5]. The first obtains
ΔE for some fixed weighting factor ω, which trivially leads to
E ω E0
ΔE ¼ E1 E0 ¼ ; ð9Þ
ω
and the second employs derivatives of Eω with respect to the weighting factor [5,
29]
dEω
ΔE ¼ ; ð10Þ
dω
A practical exploration of (10) was attempted by Gross et al. [29] who used the
quasilocal density approximation (qLDA) [65] with fractional occupation numbers
of the KS orbitals, although the excitation energies obtained for the He atom were
unsatisfactory. Similarly poor results (with the errors on the order of a few eV) were
obtained in several other works by employing various approximations for the
exchange-correlation functional to calculate the excitation energies of atoms and
small molecules [66–68].
Ensemble DFT Approach to Excited States of Strongly Correlated Molecular Systems 103
where the Hartree electron–electron repulsion energy and Exc are calculated for the
total ensemble density,
ð ð 0
1 3 0 3 ρω ðrÞρω r
U H ½ ρω ¼ d r d r ; ð12Þ
2 jr r0 j
and a suitable approximate functional is employed for Exc [29]. As the UH energy
depends nonlinearly on the density, the dependence of (12) on the ensemble
weighting factors becomes nonlinear, which leads to the emergence of unphysical
“ghost” contributions, i.e., cross-terms between the ensemble components. These
terms are supposed to be eliminated by the XC functional, which should also
become nonlinear in the ensemble weighting factors [31, 33]. The commonly
available approximations for the XC functional were incapable of accurately
compensating for the “ghost” contributions and, consequently, the results obtained
with the use of these functionals were quite poor [29, 66–68].
Considerably better excitation energies from the ensemble DFT calculations
were obtained by Pernal et al. [31], who employed a “ghost”-free formulation for
the ensemble energy functional. The “ghost”-free Hartree electron–electron repul-
sion in [31] was calculated:
104 M. Filatov
ð ð 0
1X 3 0 3 ρI ðrÞρI r
UH ½ρω ¼ ωI d r d r ; ð13Þ
2 I jr r0 j
where ρI are the densities of the individual components of the ensemble. In their
work, Pernal et al. separated the ensemble XC energy into the long-range (lr)
component which was treated at the multi-reference WFT level and the short-
range (sr) XC energy approximated by a density functional:
X
ð
Eω ¼ ωI T s, I þ U H ½ρI þ d r vext ðrÞρI ðrÞ þ
3
Elrxc, I þ Exc
sr
, DFT ½ρω : ð14Þ
I
Although, with the use of this approach, the excitation energies of Be atom and LiH
and BH molecules were considerably improved, there still remained substantial
residual errors on the order of 0.6–0.8 eV. Furthermore, the sr-XC energy in (14)
still remained nonlinear in the ensemble weighting factors and inseparable into
the individual contributions of ensemble components. These shortcomings of the
currently available implementations of ensemble DFT are not present in the REKS
method which is described in the following section.
3 REKS Methodology
In this section the basic aspects of the REKS method are explained. The REKS
method was initially developed to deal with the non-dynamic correlation in the
ground electronic states of molecules [35, 38] and was later extended to treat the
excited states [39–41]. The latter method is known as the state-averaged REKS
(SA-REKS) [39] and the state interaction SA-REKS (SI-SA-REKS or SSR, for
brevity) [40, 41].
The REKS method for ground states is a practical implementation of ensemble DFT
formalism that depends upon (2) and (5) [20]. Let us consider a situation that
requires the use of the ensemble formalism at the DFT level and the multi-reference
description at the WFT level. For instance, let us take two H2 molecules in a
rectangular arrangement as shown in Fig. 1. The H2 + H2 system was studied in
[26] with the use of both the multi-reference configuration interaction (MRCI)
method of WFT and the ensemble DFT formalism. In the latter case, the non-
interacting KS reference state and the KS potential vs(r) were constructed from the
MRCI density using the reverse engineering approach of Zhao, Morrison, and Parr
[62]. It was found that one has to use the ensemble representation and the fractional
Ensemble DFT Approach to Excited States of Strongly Correlated Molecular Systems 105
Fig. 1 Definition of
geometry and frontier
orbitals for the H2 + H2
system
The ensemble KS solution obtained which comprises the densities and energies
of two electronic configurations, where each is represented by a single KS deter-
minant, corresponds to a system of non-interacting particles. To derive an energy
expression that would conform with the ensemble representation and would yield
the energy of a system of interacting electrons, let us make use of the adiabatic
connection formalism [69], i.e., let us gradually switch the electron–electron
interaction on and simultaneously modify the external potential in such a way
that the total density remains unchanged [70]; see the Hamiltonian
X 1 X X α
^α ¼
H ∇2i þ vext, α ðri Þ þ ; ð15Þ
i
2 i
r
i> j i j
naα
nα
Eα ¼ E α . . . ϕ a ϕ a þ b Eα . . . ϕ b ϕ b
2 2
1 α α 1=2
þ na nb E α ½ . . . ϕ a ϕ b Eα . . . ϕ a ϕ b þ Eα . . . ϕ a ϕ b Eα . . . ϕ a ϕ b ;
2
ð16Þ
where the energies of the electronic configurations are calculated using the Hamil-
tonian (15) and the barred orbitals and the unbarred orbitals are occupied with the
beta-spin and the alpha-spin electrons, respectively. The energy term in parentheses
in the second line of (16) represents the negative of the exchange integral
(ϕaϕb|ϕbϕa) expressed via the energy differences between the singlet and triplet
configurations.2
Using the coupling strength integration [69] and making an assumption that the
α-dependent occupation numbers nαa and nαb can be replaced by the respective
median values, one arrives at the formula
2
Note that the kinetic energy is independent of the spin and the total densities of the electronic
configurations in the second line of (16) are identical.
Ensemble DFT Approach to Excited States of Strongly Correlated Molecular Systems 107
na
nb
1
Eens ¼ EDFT . . . ϕa ϕa þ EDFT . . . ϕb ϕb þ ðna nb Þ1=2
2 2 2
EDFT ½. . . ϕa ϕb EDFT . . . ϕa ϕb þ EDFT . . . ϕa ϕb EDFT . . . ϕa ϕb ;
ð17Þ
where EDFT denotes the total energy calculated for a single-determinant configu-
ration using the conventional KS DFT formalism. It is noteworthy that the paren-
thesized term in the second line of (17) does not contribute to the total density, as
the densities of these configurations cancel each other identically. Hence, the total
density of a strongly correlated state can be calculated using
na
nb
ρens ¼ ρ . . . ϕa ϕa þ ρ . . . ϕb ϕb
2 2
1 1=2
þ ð na nb Þ ρ½ . . . ϕ a ϕb ρ . . . ϕa ϕb þ ρ . . . ϕa ϕ b . . . ϕa ϕb ;
2
nb
na
¼ ρ . . . ϕa ϕa þ ρ . . . ϕb ϕb
2 2
ð18Þ
which is the weighted sum of the densities of the configurations in (17) taken with
the same weighting factors.
To illustrate the derivation of (17), let us expand the ensemble energy (16)
obtained from quasi-degenerate perturbation theory near α ¼ 0. Equation (16)
is obtained as the most negative eigenvalue of the secular matrix
α
Eα . . . ϕa ϕa K ab
α
; ð19Þ
K ab Eα . . . ϕ b ϕ b
α 1
where K ab ¼ Eα ½...ϕa ϕb Eα ...ϕa ϕb þ Eα ...ϕa ϕb Eα ...ϕa ϕb
α 2
K ab ¼ α ϕa ϕb ϕb ϕa , for α!0 is the coupling element between the configu-
rations ...ϕa ϕa and ...ϕb ϕb (for a small α, the exchange integral between
the orbitals ϕa and ϕb). Expanding this matrix with respect to α and keeping
only the first term in the expansion, one obtains
(continued)
108 M. Filatov
α
Eα . . . ϕ a ϕ a K ab E0 . . . ϕ a ϕ a 0
α
¼
K ab Eα . . . ϕ b ϕ b 0 E 0 . . . ϕb ϕb
0
1
dEα . . . ϕa ϕa α
dK ab
B C
þ αB
@
dα dα
C
dK abα dEα . . . ϕb ϕb A
dα dα
2
þO α ;
ð20Þ
naα
nα
Eα ¼ E0 . . . ϕ a ϕ a þ b E0 . . . ϕ b ϕ b
2 2
!
α dE . . . ϕ ϕ
na α a a nbα dEα . . . ϕb ϕb α α 1=2 dK ab
α
þα þ na nb : ð21Þ
2 dα 2 dα dα
where the EHxc terms comprise the Hartree and the XC energy of the given
configuration. Equation (17) is obtained from (22) using the density ρens in
(18) and noting that the sum of the kinetic energy and the interaction with the
external potential vext,1 is the same for the four terms in parentheses in the
third line of (22). When deriving (17), it was also assumed that no further
degeneracies (except the point α ¼ 0) occur along the adiabatic connection
path.
The formulae obtained for the density and the energy are valid for the case of
strong non-dynamic correlation, where the occupation numbers of the fractionally
occupied orbitals are close to unity, na nb 1. When the multi-reference
Ensemble DFT Approach to Excited States of Strongly Correlated Molecular Systems 109
1 1na nb þδ
f ðna ; nb Þ ¼ ðna nb Þ12 1þδ : ð24Þ
2
The damping factor in (24) is set to a value δ ¼ 0.4 to provide for a stable
convergence of the REKS self-consistent field (SCF) iterations near the regime
when E-VR solution collapses to the PS-VR solution [73]. In the described version
of REKS, the FONs of the two frontier orbitals are restricted to sum up to two
electrons; hence the name REKS(2,2), which is similar to the notation adopted for
the complete active space SCF (CASSCF) method in multi-reference WFT.
In the strict implementation of KS theory, the derived REKS total energy should
be minimized with respect to the REKS density (naturally, the FONs too). As the
REKS energy is not an explicit functional of the density, such a minimization
should inevitably rely on a variant of the optimized effective potential (OEP)
approach [74], which is known to suffer from steep computation time scaling and
certain stability issues when used in connection with the localized basis sets for
expanding the KS orbitals [75]. Therefore, the REKS total energy is minimized
with respect to the orbitals, as is being commonly done in connection with the
hybrid and meta GGA density functionals, thus avoiding the need to tackle the
density–density response function3 used in the OEP formalism. The FONs are
obtained variationally by minimizing the energy (23) under the constraint
na + nb ¼ 2. The latter constraint is imposed explicitly, without using the method
of Lagrange multipliers. The REKS orbitals are optimized using the coupling
operator technique of the open-shell SCF theory [77]. For brevity, the REKS one-
electron equations are not presented here and the reader is referred to the original
publications [34, 35, 43]; see also a review article [73].
3
See [76] for the derivation of density–density response function for ensemble densities.
110 M. Filatov
between the configurations, although the BS-UKS curve deviates stronger from the
target MRCI PES and underestimates the reaction barrier height. The FONs of the
frontier orbitals (only the b2u FON shown in Fig. 2) obtained by the REKS method
are in a good agreement with the exact ensemble KS values, whereas the BS-UKS
occupations (the natural orbital’s occupation numbers are shown in lieu of FONs)
deviate strongly from the exact ones, suggesting that BS-UKS overestimates the
effect of the non-dynamic correlation. Furthermore, BS-UKS displays an abrupt
onset of the non-dynamic correlation (after ca. R ¼ 2.75 bohr), whereas the REKS
method yields a smooth transition between the PS-VR and E-VR regimes and a
more accurate description of the reaction PES profile.
The comparison vis-a-vis the exact ensemble KS results demonstrates the
validity of the approximations made in the REKS working equations. Besides the
H2 + H2 system, the REKS method was applied to study bond-breaking/bond-
formation reactions in several chemical systems as well as the electronic structure
of biradicals, magnetic coupling in metal complexes and organic charge transfer
crystals. The reader is advised to inspect the original publications [37, 38, 42–53,
56, 57] for more examples of the method performance.
Let us consider a model system with two strongly correlated electrons in two
orbitals, such as the H2 molecule with the bond stretched beyond the Coulson–
Fischer point [84]. Near the equilibriumbondlength, the electronic structure of H2
is dominated by a single configuration 1σ g 1σ g and the doubly excited configu-
ration j1σ u 1σ u i lies high in energy (1σ g is the bonding MO and 1σ u the anti-bonding
MO). When the bond is stretched beyond the Coulson–Fischer point, the energy gap
between the two electronic configurations narrows to a limit that allows for an
efficient mixing of the configurations and the strong non-dynamic electron corre-
lation ensues. In the minimal basis of the two orbitals (the bonding 1σ g MO denoted
to ϕa and the anti-bonding 1σ u to ϕb), the ground-state wavefunction of stretched H2
can be represented by a two-configurational wavefunction:
rffiffiffiffiffi rffiffiffiffiffi
na nb
Φ0 ¼ ϕa ϕa ϕb ϕb ; ð25Þ
2 2
where na and nb are the FONs of the orbitals ϕa and ϕb. Promoting a single electron
from ϕa to ϕb orbital leads to a singlet excited state Φ1 which can be represented by
the wavefunction
112 M. Filatov
1 1
Φ1 ¼ pffiffiffiϕa ϕb þ pffiffiffiϕb ϕa ; ð26Þ
2 2
For a homosymmetric molecule, such as H2, the two states belong in different
symmetry species and therefore do not interact with one another.
Using the ensemble DFT for excited states, described in Sect. 3.1, the excitation
energy can be obtained from the variational optimization of the energy of an
ensemble of the two states [5]. The ground state (25) can be described by the
REKS(2,2) method and the excited state (26) by the spin-restricted open-shell KS
(ROKS) method for an open-shell singlet (OSS) state [12, 34]. Within the latter
approach, the energy of the OSS state is given by [12, 34]
EDFT . . . ϕa ϕb : ð27Þ
2
The use of the REKS and ROKS energies in (7) leads to the SA-REKS energy
expression [39]:
ESA-REKS
ω ¼ ð1 ωÞEREKSð2;2Þ þ ωEROKS ; ð28Þ
gradually to zero. As seen in the upper panel of Fig. 3, the SA-REKS method
correctly describes the H–H bond dissociation, whereas the single-reference RKS
approach fails to yield the correct dissociation limit for the H2 molecule. Thus, it is
the failure of the conventional KS DFT approach to describe the non-dynamic
electron correlation for a dissociating covalent bond that is responsible for the
failure of TD-DFT to describe correctly the excitation energy of a dissociating
molecule.
The described SA-REKS method is capable of describing the ground and excited
states of a homosymmetric molecule when the mixing of the two states is prevented
by symmetry. In the case of a heterosymmetric molecule, e.g., dissociating LiH, the
two states in (25) and (26) are allowed to mix and therefore their representation as a
purely covalent state and a purely ionic state is no longer accurate. To correct for
this deficiency of the SA-REKS description and to construct an ensemble of two
decoupled states, one can obtain a pair of new states by solving a 2 2 secular
problem with the Hamiltonian matrix that spans the EREKS(2,2) and the EROKS
energies as the diagonal elements and the off-diagonal (coupling) element given
in (29):
pffiffiffiffiffi ^ pffiffiffiffiffi ^ pffiffiffiffiffi pffiffiffiffiffi
H 01 ¼ na ϕb na F a ϕa nb ϕa nb F b ϕb ¼ ð na nb Þεab ð29Þ
which was obtained in [40, 41] by applying the Slater–Condon rules in the space of
the two CSFs Φ0 and Φ1 and the variational condition for the open-shell orbitals ϕa
114 M. Filatov
and ϕb [34, 35, 77]. In (29), F^ a and F^ b are the Fock operators for the open-shell
orbitals and εab is the off-diagonal Lagrange multiplier4 in the open-shell Lagrang-
ian [73]. As the two states, Φ0 and Φ1, are mutually orthogonal, the average of the
new energies E0 and E1 obtained from the above secular problem remains the same
as the average of the REKS(2,2) and ROKS energies. This implies that the orbitals
for the new approach, dubbed SI-SA-REKS or SSR, can still be obtained from the
SA-REKS orbital optimization, provided that ω ¼ 1/2 was employed in the latter. In
practical applications of the SI-SA-REKS method [40, 41, 59–61], it was found that
the described state-interaction scheme is important for obtaining the correct shape
of the ground and excited state PESs in the vicinity of conical intersections and near
avoided crossings. For other situations, when the energy gap between the ground
and excited states is sufficiently wide the SI-SA-REKS method yields nearly the
same excitation energies as the SA-REKS method [59].
4
The matrix of Lagrange multipliers in open-shell SCF becomes Hermitian (but not diagonal)
upon convergence to the variational minimum [77].
Ensemble DFT Approach to Excited States of Strongly Correlated Molecular Systems 115
ωPBE functional and the aug-cc-pVTZ basis set are compared with the results of the
conventional TD-DFT calculations (with the same basis set and functional) and with
the literature data. The x1Σ + a1Σ + excitation energy of LiH was recently studied
using an ab initio restricted active space CI (RASCI) method with the aug-cc-pVTZ
basis set [86]. For the individual states, the x1Σ + and the a1Σ + states, the potential
energy curves along the dissociation path were obtained in [31] using the CCSD
method (presumably the EOM-CCSD was used to obtain the excited state curve).
As seen in Fig. 4, the SI-SA-REKS potential energy curves follow closely the ab
initio results, whereas the conventional KS DFT curves fail to reproduce the correct
dependence on distance. Near ca. RLiH ¼ 7 bohr, the two states undergo an avoided
crossing as seen in the curves obtained by the ab initio WFT calculations and the SI-
SA-REKS calculations. The RKS ground-state curve does not converge to the
correct dissociation limit and the ground state remains ionic along the whole
dissociation path. The excitation energy from the SI-SA-REKS calculations closely
follows the RASCI excitation energy curve and correctly yields the avoided
crossing. The TD-DFT excitation energy, although close to the ab initio value
near the equilibrium distance, fails to display the correct distance dependence and
vanishes at the dissociation limit. This example illustrates yet another failure of the
conventional KS DFT/TD-DFT approach to describe the ground and excited state
potential energy surfaces of molecules with dissociating bonds (or, more generally,
strongly correlated molecular systems). By contrast, the SI-SA-REKS method
describes these situations with high accuracy and can be applied with confidence
to study the excited states of strongly correlated molecules.
116 M. Filatov
REKS method can be used as a general purpose computational scheme for describ-
ing the valence excitation energies.
Obviously, the ability of the SI-SA-REKS method to describe electronic transi-
tions in strongly correlated molecules enables one to apply this method beyond the
realm of applicability of the conventional adiabatic linear-response TD-DFT. Thus,
the method was employed to study the optically bright 1La electronic transitions
(1B1u symmetry) in a series of linear n-acenes (1) [59].
Polyacenes are known to have a strongly correlated ground state and this is
illustrated by a sketch of the valence Lewis structures in the diagram above [94].
Therefore the use of multi-reference approaches is mandatory for proper description
of their ground state. The single-reference KS DFT is incapable of taking accurate
account of the non-dynamic correlation in the ground state of longer polyacenes and
the TD-DFT excitation energies become unrealistically low for these molecules. The
SI-SA-REKS method describes accurately the ground state of polyacenes and yields
excitation energies in good agreement with the experimental figures.
Another situation where the description of the non-dynamic correlation in the
ground state becomes important is the real crossing between the ground and lowest
excited states of the same spin and space symmetry, the so-called conical inter-
sections. The SI-SA-REKS method was successfully applied to study conical inter-
sections in a series of organic molecules and models of biological chromophores
[40, 41, 60, 95], molecular switches [55], and molecular motors [54, 58, 61]. In
these applications and benchmarks, the SI-SA-REKS method was capable of
describing the geometry at the minimum of the conical intersection seam (the so-
called minimum energy conical intersection, MECI) with an accuracy matching
high level ab initio multi-reference methods such as MRCI and CASPT2. The
results of the application of SI-SA-REKS to conical intersections are described in
another chapter of this book;5 here it is only mentioned that the root mean square
deviation of the SI-SA-REKS MECI geometries from the ab initio reference
geometries is less than 0.1 Å on average (0.0609 Å was obtained in [95] for a set
of 12 MECIs).
Besides being capable of describing excitations of strongly correlated molecular
species, the SI-SA-REKS method displays an outstanding performance in other
situations which proved to be difficult for standard linear response methods.
5
See the chapter “Description of conical intersections with density functional methods” by M.
Huix-Rotllant, A. Nikiforov, W. Thiel, and M. Filatov.
Ensemble DFT Approach to Excited States of Strongly Correlated Molecular Systems 119
Table 3 Lowest electronic excitation energy (eV) of cyanine dyes. The aug-cc-pVTZ basis set is
employed in DFT calculations
BH&HLYPa CAM-B3LYPa
Molecule TD SSR TD SSR CASPT2b DMCb
CN5 5.35 4.87 5.19 4.71 4.69 5.03
CN7 4.19 3.72 4.07 3.65 3.52 3.83
CN9 3.49 3.06 3.39 3.03 2.81 3.09
CN11 3.02 2.62 2.93 2.62 2.46 2.62
a
Geometries are taken from [96]
b
CASPT2 and diffusion Monte–Carlo (DMC) data from [101]
In cyanine dyes (2), the lowest singlet 1B1 (in C2v symmetry) excited state
(π ! π* transition) is notoriously difficult for linear response methods [96–98].
TD-DFT with the commonly available density functionals overestimates the exci-
tation energies by ca. 0.4–0.5 eV; this deviates from the trend typical for TD-DFT
which has a tendency to underestimate the valence excitation energies by ca. 0.3–
0.4 eV. The cyanine dyes do not have a strongly correlated ground state and it was
the incorrect description of the differential correlation effects between the ground
and excited states that was blamed for the poor performance of TD-DFT [99].
However, this conjecture was challenged by Ziegler et al. [100] who showed that
going beyond the linear response approximation leads to considerable improvement
of the calculated excitation energies.6
The 1B1 excitation energies in a series of cyanine dyes were studied in [59] with
the use of the SI-SA-REKS method in connection with a few commonly available
density functionals and the aug-cc-pVTZ basis set. The results of the SI-SA-REKS
calculations are compared in Table 3 with TD-DFT and with a number of high level
ab initio calculations, the second-order complete active space perturbation theory
(CASPT2), and the diffusion Monte–Carlo (DMC) calculations from [96]. The
results in Table 3 show that SI-SA-REKS noticeably outperforms TD-DFT in the
accuracy of description of the target excitation energies, thus demonstrating the
advantage of the ensemble formalism. Indeed, the KS orbitals in the SI-SA-REKS
method are variationally optimized for both states, the ground and the excited state,
and the good performance of SI-SA-REKS seems to agree with the conclusions of
Ziegler et al. [100, 102] drawn from the results of the application of the relaxed
constricted variational DFT (RSCF-CV(1)-DFT) method, a method that goes
beyond the linear response and affords a variational optimization of the orbitals
6
See the chapter “A Constricted Variational Density Functional Theory Approach to the Descrip-
tion of Excited States” by T. Ziegler, M. Krykunov, I. Seidu, and Y. C. Park.
120 M. Filatov
Table 4 Excitation energies (eV) of the lowest CT transitions of the Ar-TCNE adducts. The cc-
pVDZ basis set is employed in all DFT calculations
BLYPa BH&HLYPa LC-ωPBEa
Arene TD SSR TD SSR TD SSR Lit.b Exp.c
Benzene 1.54 3.53 2.96 3.52 4.00 3.69 3.80 3.59
Naphthalene 0.34 2.28 1.84 2.46 3.01 2.74 2.70 2.60
Toluene 1.37 2.72 2.67 3.26 3.65 3.30 3.40 3.36
o-Xylene 1.47 2.61 2.42 2.85 3.40 3.01 3.00 3.15
MADd 2.00 0.39 0.70 0.15 0.34 0.11 0.13
a
Geometries are taken from [101]
b
Literature data: results of TD-DFT calculations using the tuned range separated BNL functional
from [101]
c
Gas phase excitation energies of CT transitions from [106]
d
Mean absolute deviations from the experimental data
partaking in the electronic transition [100, 102]. The SI-SA-REKS method achieves
the same effect by using the ensemble formalism.
The use of ensemble formalism based on the variational principle also turns out
to be beneficial for the description of charge transfer transitions. Linear response
methods, such as TD-DFT, experience considerable difficulties when describing
this type of electronic excitation, especially when used in connection with the
commonly available approximate density functionals [103, 104]. Although it was
not designed with these particular excitations in mind, the SI-SA-REKS method
was found to be surprisingly accurate for charge transfer excitations, even when
used in connection with the stock parameterization of the commonly available
GGA and hybrid density functionals [105].
Table 4 reports excitation energies of the lowest charge transfer transitions of a
series of arene–TCNE (tetracyanoethylene) adducts, for which the gas phase optical
absorption spectra are available [106]. For these electronic transitions, the TD-DFT
excitation energies obtained with the use of the usual density functionals deviate
from the experimental figures by a wide margin and only the use of individually
tuned range-separated density functionals brings these errors down to an acceptable
level [101]. However, the accuracy achieved with the fine-tuned density functionals
is easily surpassed by the SI-SA-REKS method employed in connection with the
standard parameterizations of commonly available density functionals. Even when
used in connection with the GGA functional, such as BLYP, the SI-SA-REKS
method yields more accurate charge transfer excitation energies than does TD-DFT
with the use of range-separated hybrid functional (see Table 4). The observed
excellent performance of SI-SA-REKS is consistent with the analysis of the
description of various types of excitations undertaken by Ziegler et al. [100, 104]
who showed that it is the use of approximate density functionals in connection with
the adiabatic linear response approximation that is to blame for ludicrous perfor-
mance of the adiabatic TD-DFT and not the density functional alone.
To conclude this section, ensemble DFT for excited states as implemented in the
SI-SA-REKS method is a versatile and accurate approach to the calculation of
Ensemble DFT Approach to Excited States of Strongly Correlated Molecular Systems 121
Ensemble DFT [18, 20, 21, 23, 29] holds considerable promise for theoretical
description of the excited states of strongly correlated molecular systems. Although
it was conceived more than three decades ago, ensemble DFT still did not find its
way to the repertoire of the methods used by computational chemists on a daily
basis. Perhaps it is the perceived lack of practical implementations of ensemble
DFT that holds down its adoption by a wider computational chemistry community.
Although there is a renewed interest in developing ensemble DFT further [31–33]
and in implementing it in the form of practically affordable computational schemes,
these approaches are largely unknown to ordinary computational chemists.
The REKS computational method, reviewed in this chapter, makes ensemble
DFT affordable. The method has already been used to study various types of
electronic transitions occurring in usual as well as strongly correlated molecular
systems and its ability to describe excitation energies in these systems with a
remarkable accuracy has been demonstrated. Although the currently available
implementation of the REKS formalism is not free of certain limitations, in
particular the size of the active space and the number of excited states are restricted,
these limitations will be removed in the near future and this should considerably
improve the prospects for practical use of the method. Especially promising for
obtaining multiple excited states and for simulating the entire excitation spectra of
strongly correlated molecules appears to be a merger of the REKS methodology
with the variational constricted DFT formalism proposed by Ziegler et al. [102,
104] (see Footnote 6). The work in these directions is currently in progress and will
continue in the future.
References
Symposium on nonlinear optical materials – theory and modeling, at the 208th National
Meeting of the American Chemical Society, Washington, DC
4. Casida ME, Huix-Rotllant M (2012) Annu Rev Phys Chem 63:287
5. Marques MAL, Gross EKU (2003) In: Fiolhais C, Nogueira F, Marques MAL (eds) A primer
in density-functional theory, lecture notes in physics, vol 620. Springer, Berlin, pp 144–184
6. Maitra NT, Zhang F, Cave R, Burke K (2004) J Chem Phys 120:5932
7. Cave RJ, Zhang F, Maitra NT, Burke K (2004) Chem Phys Lett 389:39
8. Huix-Rotllant M, Ipatov A, Rubio A, Casida ME (2011) Chem Phys 391:120
9. Aryasetiawan F, Gunnarsson O, Rubio A (2002) Europhys Lett 57:683
10. McWeeny R (1992) Methods of molecular quantum mechanics. Academic, San Diego
11. Slater JC, Wood JH (1970) Int J Quantum Chem S4:3
12. Ziegler T, Rauk A, Baerends EJ (1977) Theor Chim Acta 43:261
13. Gunnarsson O, Lundqvist BI (1976) Phys Rev B 13:4274
14. Kowalczyk T, Yost SR, Van Voorhis T (2011) J Chem Phys 134:054128
15. Gaudoin R, Burke K (2004) Phys Rev Lett 93:173001
16. Gaudoin R, Burke K (2005) Phys Rev Lett 94:029901
17. Li YQ, Pan XY, Li B, Sahni V (2012) Phys Rev A 85:032517
18. Theophilou AK (1979) J Phys C Solid State Phys 12:5419
19. Valone SM (1980) J Chem Phys 73:4653
20. Lieb EH (1983) Int J Quantum Chem 24:243
21. Englisch H, Englisch R (1984) Phys Stat Sol (b) 123:711
22. Englisch H, Englisch R (1984) Phys Stat Sol (b) 124:373
23. Gross EKU, Oliveira LN, Kohn W (1988) Phys Rev A 37:2805
24. Gross EKU, Oliveira LN, Kohn W (1988) Phys Rev A 37:2809
25. Oliveira LN, Gross EKU, Kohn W (1988) Phys Rev A 37:2821
26. Schipper PRT, Gritsenko OV, Baerends EJ (1998) Theor Chem Acc 99:329
27. Schipper PRT, Gritsenko OV, Baerends EJ (1999) J Chem Phys 111:4056
28. Morrison RC (2002) J Chem Phys 117:10506
29. Oliveira LN, Gross EKU, Kohn W (1990) Int J Quantum Chem Quantum Chem Symp 24:707
30. Ullrich CA, Kohn W (2001) Phys Rev Lett 87:093001
31. Pastorczak E, Gidopoulos NI, Pernal K (2013) Phys Rev A 87:062501
32. Franck O, Fromager E (2014) Mol Phys 112:1684
33. Pribram-Jones A, Yang ZH, Trail JR, Burke K, Needs RJ, Ullrich CA (2014) J Chem Phys
140:18A541
34. Filatov M, Shaik S (1998) Chem Phys Lett 288:689
35. Filatov M, Shaik S (1999) Chem Phys Lett 304:429
36. Filatov M, Shaik S (2000) Chem Phys Lett 332:409
37. Filatov M, Shaik S (2000) J Phys Chem A 104:6628
38. Moreira IDPR, Costa R, Filatov M, Illas F (2007) J Chem Theory Comput 3:764
39. Kazaryan A, Heuver J, Filatov M (2008) J Phys Chem A 112:12980
40. Filatov M (2013) J Chem Theory Comput 9:4526
41. Huix-Rotllant M, Filatov M, Gozem S, Schapiro I, Olivucci M, Ferré N (2013) J Chem
Theory Comput 9:3917
42. Filatov M, Shaik S (1999) J Phys Chem A 103:8885
43. Filatov M, Shaik S (1999) J Chem Phys 110:116
44. De Visser SP, Filatov M, Shaik S (2000) Phys Chem Chem Phys 2:5046
45. Filatov M, Shaik S, Woeller M, Grimme S, Peyerimhoff SD (2000) Chem Phys Lett 316:135
46. De Visser SP, Filatov M, Shaik S (2001) Phys Chem Chem Phys 3:1242
47. Kraka E, Anglada J, Hjerpe A, Filatov M, Cremer D (2001) Chem Phys Lett 348:115
48. Cremer D, Filatov M, Polo V, Kraka E, Shaik S (2002) Int J Mol Sci 3:604
49. Gräfenstein J, Kraka E, Filatov M, Cremer D (2002) Int J Mol Sci 3:360
50. Filatov M, Cremer D (2003) Phys Chem Chem Phys 5:2320
51. De Visser SP, Filatov M, Schreiner PR, Shaik S (2003) Eur J Org Chem 21:4199
Ensemble DFT Approach to Excited States of Strongly Correlated Molecular Systems 123
52. Illas F, Moreira IDPR, Bofill JM, Filatov M (2004) Phys Rev B 70:132414
53. Illas F, Moreira IDPR, Bofill JM, Filatov M (2006) Theor Chem Acc 116:587
54. Kazaryan A, Kistemaker JCM, Schäfer LV, Browne WR, Feringa BL, Filatov M (2010)
J Phys Chem A 114:5058
55. Filatov M (2011) ChemPhysChem 12:3348
56. Filatov M (2011) Phys Chem Chem Phys 13:12328
57. Filatov M (2011) Phys Chem Chem Phys 13:144
58. Kazaryan A, Lan Z, Schäfer LV, Thiel W, Filatov M (2011) J Chem Theory Comput 7:2189
59. Filatov M, Huix-Rotllant M (2014) J Chem Phys 141:024112
60. Gozem S, Melaccio F, Valentini A, Filatov M, Huix-Rotllant M, Ferré N, Frutos LM, Angeli C,
Krylov AI, Granovsky AA, Lindh R, Olivucci M (2014) J Chem Theory Comput 10:3074
61. Filatov M, Olivucci M (2014) J Org Chem 79:3587
62. Zhao Q, Morrison RC, Parr RG (1994) Phys Rev A 50:2138
63. Sinanoglǔ O (1964) Adv Chem Phys 6:315
64. Giesbertz KJH, Baerends EJ (2010) J Chem Phys 132:194108
65. Kohn W (1986) Phys Rev A 34:737
66. Andrejkovics I, Nagy Á (1998) Chem Phys Lett 296:489
67. Paragi G, Gyémánt IK, Van Doren VE (2000) Chem Phys Lett 324:440
68. Paragi G, Gyémánt IK, Van Doren VE (2001) J Mol Struct Theochem 571:153
69. Noziéres P, Pines D (1966) The theory of quantum liquids. Perseus Books Publishing LLC,
Cambridge
70. Becke AD (1988) J Chem Phys 88:1053
71. Wang SG, Schwarz WHE (1996) J Chem Phys 105:4641
72. Csányi G, Arias T (2000) Phys Rev B 61:7348
73. Filatov M (2015) WIREs Comput Mol Sci 5:146
74. Lathiotakis NN, Helbig N, Rubio A, Gidopoulos NI (2014) Phys Rev A 90:032511
75. Fernandez JJ, Kollmar C, Filatov M (2010) Phys Rev A 82:022508
76. Farid B (1998) J Phys Condens Matter 10:L1
77. Hirao K, Nakatsuji H (1973) J Chem Phys 59:1457
78. Hoffmann R, Woodward RB (1965) J Am Chem Soc 87:395
79. Hoffmann R, Woodward RB (1965) J Am Chem Soc 87:2046
80. Woodward RB, Hoffmann R (1969) Angew Chem Int Ed 8:781
81. Vydrov OA, Heyd J, Krukau A, Scuseria GE (2006) J Chem Phys 125:074106
82. Vydrov OA, Scuseria GE (2006) J Chem Phys 125:234109
83. Vydrov OA, Scuseria GE, Perdew JP (2007) J Chem Phys 126:154109
84. Coulson CA, Fischer I (1949) Philos Mag 40:386
85. Kolos W, Wolniewicz L (1965) J Chem Phys 43:2429
86. Van Meer R, Gritsenko OV, Baerends EJ (2014) J Chem Phys 140:024101
87. Schirmer J (1982) Phys Rev A 26:2395
88. Schirmer J, Trofimov AB (1995) J Phys B At Mol Opt Phys 28:2299
89. Trofimov AB, Stelter G, Schirmer J (1999) J Chem Phys 111:9982
90. Trofimov AB, Stelter G, Schirmer J (2002) J Chem Phys 117:6402
91. Schirmer J, Trofimov AB (2004) J Chem Phys 120:11449
92. Silva-Junior MR, Schreiber M, Sauer SPA, Thiel W (2008) J Chem Phys 129:104103
93. T€onshoff C, Bettinger HF (2010) Angew Chem Int Ed 49:4125
94. Plasser F, Pašalić H, Gerzabek MH, Libisch F, Reiter R, Burgd€
orfer J, Müller T, Shepard R,
Lischka H (2013) Angew Chem Int Ed 52:2581
95. Nikiforov A, Gamez JA, Thiel W, Huix-Rotllant M, Filatov M (2014) J Chem Phys 141:
124122
96. Send R, Valsson O, Filippi C (2011) J Chem Theory Comput 7:444
97. Jacquemin D, Zhao Y, Valero R, Adamo C, Ciofini I, Truhlar DG (2012) J Chem Theory
Comput 8:1255
98. Moore B II, Autschbach J (2013) J Chem Theory Comput 9:4991
124 M. Filatov
K. Pernal (*)
Institute of Physics, ul. Wolczanska 219, 90-924 Lodz, Poland
e-mail: pernalk@gmail.com
K.J.H. Giesbertz
Theoretical Chemistry, VU University, De Boelelaan 1083, 1081 HV Amsterdam,
The Netherlands
e-mail: k.j.h.giesbertz@vu.nl
126 K. Pernal and K.J.H. Giesbertz
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
2 Construction of Density Matrix Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
2.1 Functionals Based on a Paradigm Two-Electron Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
2.2 Functionals Based on Reconstruction of 2-RDM in Terms of 1-RDM . . . . . . . . . . . . . 137
2.3 Going Beyond Explicit Density Matrix Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
2.4 Empirical Density Matrix Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
3 Predicting Properties of Electronic Systems with Density Matrix Functionals . . . . . . . . . . . 151
3.1 Response Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
3.2 Ionization Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
3.3 Fundamental Gap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4 Optimization of Density Matrix Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
5 Time-Dependent RDMFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
5.1 Equation of Motion of the 1-RDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
5.2 Time-Dependent Response Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.3 Standard Adiabatic Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5.4 Alternative Adiabatic Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.5 Phase Including Natural Orbitals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
6 Summary and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
1 Introduction
The most widely used methods in quantum chemistry are either wavefunction-
based or they exploit density functional theory (DFT). The former class of methods
offers high accuracy, but unfavorable scaling of the computational cost with system
size limits their scope of applicability to relatively small systems. Density func-
tional approximations are known to offer a good balance between computational
efficiency and accuracy. Nevertheless, most approximations to DFT are plagued by
generic problems related to the fact that DFT employs a simple local object – the
electron density. Accurate description of the electronic structure of multireference
systems or predicting multiple electronic excitations are examples of problems that
still await satisfactory solutions in DFT. There has recently been growing interest in
approaches embracing simplicity (in the sense that a wavefunction is not involved),
computational efficiency, and versatility of DFT, while lacking the drawbacks.
Functionals of one-electron reduced density matrix (1-RDM) γ, defined for an N-
electron wavefunction Ψ as
ð ð
γ ðx; x Þ ¼ N Ψðx; x2 ; . . . ; xN ÞΨ* ðx0 ; x2 ; . . . ; xN Þdx2 dxN ;
0
ð1Þ
where x ¼ (r, s) is a combined spatial and spin coordinate, should, in principle, lead
to formulating methods superior to existing density functional approximations,
especially when static electron correlation effects cannot be neglected. An imme-
diate advantage of using 1-RDM as a main variable instead of the electron density,
ρ, is that the kinetic energy is an explicit functional of γ but not of ρ. Thus, in
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 127
Eigenvalues of 1-RDM, {np}, are called natural occupation numbers, whereas its
eigenfunctions, {φp}, are known as natural spinorbitals. Throughout the chapter we
assume a convention that the indices p, q, r, s pertain to natural spinorbitals and a, b,
c, d to arbitrary one-electron functions. Self-adjointness of γ implies orthonormality
of the natural spinorbitals. Additionally, because γ is assumed to be normalized to a
number of electrons N, cf. definition given in (1), the natural occupancies sum up to
N. Taking into account that each np is nonnegative and not greater than 1 [2, 3], the
overall properties of the natural spinorbitals and occupation numbers read
ð
8 p, q φ*p ðxÞφq ðxÞdx ¼ δ pq ; ð3Þ
8p 0 n p 1; ð4Þ
X
n p ¼ N: ð5Þ
p
Coleman [2] has proved that if a given Hermitian 1-RDM satisfies the conditions
(3)–(5) there exists an ensemble of N-electron antisymmetric wavefunctions that
yield γ. The conditions are called N-representability conditions. It should be noted
that similar sufficient and necessary conditions that would ensure pure-state N-
representability are not known, though some significant progress has been reported
by Klyachko [A.A. Klyachko, J. Phys. Conf. Ser. 36, 72–86 (2006), doi: 10.1088/
1742-6596/36/1/014].
A one-to-one mapping between pure-state v-representable 1-RDMs and
non-degenerate ground state wavefunctions has been demonstrated by Gilbert
who extended the Hohenberg–Kohn theorem to nonlocal potentials [1, 13]. This
establishes existence of a 1-RDM functional [1, 11]
128 K. Pernal and K.J.H. Giesbertz
^ ^
v ½γ ¼ Tr h γ^ þ Ψ½γ V ee Ψ½γ ;
EHK ð6Þ
h^ ¼ ^t þ v^ ext ; ð7Þ
XN 1
V^ ee ¼ i> j r
is an electron interaction operator (note that atomic units are
ij
employed throughout the chapter), and Ψ[γ] denotes a ground state wavefunction
pertinent to a v-representable γ. A variational principle for the functional defined in
(6) exists and reads
8γ2v-rep Ev ½ γ E0 ; ð8Þ
where the minimization is carried out with respect to N-electron density matrices
Γ(N ) that yield γ. Because of the linearity of the map Γ(N ) ! γ and the fact that the
set of N-representable γ is convex, a functional Eee[γ] is also convex [6]. For a given
external potential v^ ext , minima of the Hohenberg–Kohn functional given in (6), the
ð13Þ
ee ½γ ¼ EH ½γ þ Ex ½γ :
EHF ð15Þ
Another paradigm case for which an exact density matrix functional is known, is a
two-electron closed-shell system. We discuss this case extensively in Sect. 2.1.
130 K. Pernal and K.J.H. Giesbertz
Ec ½γ ¼ Exc ½γ Ex ½γ : ð19Þ
1X pffiffiffiffiffiffiffiffiffiffi
xc ½γ ¼
EBB n p nq pq q p ; ð20Þ
2 pq
where the natural occupation numbers {np} and the spinorbitals {φp} are eigen-
values and eigenfunctions of γ and the exchange integrals {hpq|qpi} are written in
the representation of the natural spinorbitals. The following notation is adopted in
this chapter for two-electron integrals
ðð
pqrs ¼ φ*p ðx1 Þφ*q ðx2 Þjr1 r2 j1 φr ðx1 Þφs ðx2 Þ dx1 dx2 : ð21Þ
The BB functional is convex [21] and reduces to the exchange functional (17) for an
idempotent γ. It is not exact for two-electron systems, though. It has been shown
that this functional severely overestimates correlation energy of atoms and mole-
cules [22–27]. However, the BB functional has been successfully used as a base for
developing more sophisticated functionals, as discussed in Sect. 2.1.
Goedecker and Umrigar (GU) have modified the BB functional by removing
diagonal ( p ¼ q) terms, called electron self-interaction, from the Hartree and the
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 131
1X pffiffiffiffiffiffiffiffiffiffi 1X
xc ½γ ¼
EGU n p nq pq q p þ n p n2p pppp : ð22Þ
2 pq 2 p
Even though the correlation part of the CHF functional satisfies the particle-hole
symmetry condition, cf. (30), which is also satisfied by the exact functional [15], it
provides little or no improvement over the HF method for molecules around the
equilibrium distances, and it breaks down in the dissociation limit [22, 32].
Although the aforementioned first generation of density matrix functionals has
not turned out to be overall competitive with DFT approximations, understanding
the origins of their failures has provided insight to developing more advanced and
successful functionals described in the next section.
1X
Eee ¼ Γabcd cd ab ; ð24Þ
2abcd
where, for a given basis set {χ a} and the pertinent sets of the creation and
a{}, {^
annihilation operators {^ a}, the elements of the 2-RDM are defined as
Γabcd ¼ 0c^ { d^ { b^ a^ 0 : ð25Þ
1X
Eee ½γ ¼ Γpqrs ½fnt g rs pq : ð26Þ
2 pqrs
1X
Ec ½ γ ¼ λpqrs ½γ rs pq : ð29Þ
2 pqrs
It has been shown that the exact correlation 1-RDM functional possesses a particle-
hole symmetry [15]
Ec ½ γ ¼ Ec ½ 1 γ ð30Þ
Ec γ η ¼ ηEc ½γ ; ð31Þ
where coordinates in γ η are scaled with a real number η and the normalization is
preserved, i.e.,
Some density matrix functionals rely on the reconstruction scheme given in (28). In
other cases, the exchange-correlation functional (18) is not partitioned any further
and is modeled as a whole. Different approaches to approximating electron–elec-
tron density matrix functionals proposed in recent years are discussed in the
remaining part of this section.
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 133
where p and p are spinorbitals of the opposite spin and φ p φ p denotes a normalized
Slater determinant. The normalization of the wavefunction imposes the following
condition of the expansion coefficients {cp}
X
c2p ¼ 1: ð34Þ
p
Employing the LS wavefunction given in (33) in (1) defining 1-RDM, one imme-
diately obtains γ in its spectral representation, which indicates that squares of the
expansion coefficients are simply the natural occupation numbers, i.e.,
8p n p ¼ c2p : ð35Þ
Taking the expectation value of the Hamiltonian with the LS wavefunction (33)
leads to a simple expression for the energy
X 1X
E¼ c2p hpp þ c p cq ppqq ; ð36Þ
p
2 pq
The exact ground state energy for a two-electron system follows by minimizing the
energy with respect to the expansion coefficients {cp} and the natural spinorbitals
{φp} under the orthonormality condition for the orbitals and a normalization
condition given in (34). An exact electron interaction density matrix functional
can be immediately written as
1 X pffiffiffiffiffiffiffiffiffiffi
ee ½γ ¼
ELS min f f n p nq pqq p ; ð37Þ
2 f f q g pq p q
8p f p ¼ 1 ; ð38Þ
where it has been taken into account that the orbitals are real so the integrals hpp|qqi
are equal to the exchange integrals hpq|qpi, and the relation (35) between expansion
coefficients in the LS wavefunction and the occupation numbers have been
exploited. It is known that for two-electron atoms and molecules at equilibrium
geometry the sign of the factor f1 corresponding to the highest occupation n1 is
predominantly
opposite
to signs {fp} of all other factors corresponding to weakly
occupied n p < 12 orbitals [29]. It should be noted that cases when this rule is
violated are known and they include, for example, a hydrogen molecule far from
equilibrium bond distance [36–38] or a strongly correlated Hooke’s atom [39]. In
such cases, natural orbitals that violate the phase rule, i.e., those orbitals whose
phase factor coincides with f1, are very weakly occupied and their contribution to
the energy is small. Consequently, a two-electron functional explicitly depending
on the occupation numbers defined as
X
e LS ½γ ¼ 1
E ee pq pq q p ;
GLS ð39Þ
2 pq
8
<np p¼q
pffiffiffiffiffiffiffiffiffiffi
GLS ¼ n p nq p ¼ 1, q > 1 or p > 1, q ¼ 1 ; ð40Þ
pq
: pffiffiffiffiffiffiffiffiffi
n p nq
ffi
otherwise
is not always fully equivalent to the exact LS functional (37) but it provides a very
good approximation to it. Inspecting the structure of the functional (39), it is evident
that terms corresponding to two weakly occupied orbitals ( p, q > 1) are treated
differently (enter the functional with a different signs) from products of strongly-
weakly occupied orbitals ( p ¼ 1, q > 1 or p > 1, q ¼ 1).
Evidently the form of the BB functional given in (20) does not reflect the orbital
structure of the functional for two electrons given in (39). Restoring this structure
and correcting for the overcorrelation by the BB functional have been motivations
behind proposing a number of corrections to it [27]. This has resulted in the
development of BB-corrected (BBC) functionals consisting of the Hartree part
(16) and the exchange-correlation functional comprising products of exchange
integrals and occupation number depending factors GBBC pq , namely
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 135
1X 1X BBC
ee ½γ ¼
EBBC n p nq ppqq þ G pqq p : ð41Þ
2 pq 2 pq pq
where p, q 2 weak denotes that both spinorbitals are weakly occupied. Self-
consistent calculations revealed that, similar to BB, BBC1 overbinds diatomic
molecules, which indicates a need for further repulsive corrections [27]. The
BBC2 functional emerged after correcting interaction between two different
strongly occupied orbitals and replacing square roots of products of the pertinent
occupancies with products npnq. The third functional, BBC3, is a result of adding
another two corrections to the BBC2 functional. In the BBC3 functional a pair of
bonding and antibonding (both called frontier) spinorbitals which form a breaking
bond is distinguished. It is assumed that a bonding orbital belongs only to a set of
frontier orbitals, and that, at the same time, antibonding orbitals belong to sets of
frontier and weakly occupied orbitals. The first BBC3 correction, added on top of
the BBC1 and BBC2 corrections, replaces square roots of products of occupation
numbers with pertinent products if one occupancy corresponds to an antibonding
orbital and the other to a strongly occupied one. The replacement does not affect
pairs of antibonding-bonding orbitals. The second correction removes self-
interaction from all orbitals except the frontier (i.e., bonding and antibonding)
orbitals. The form of the Gpq elements in the BBC3 functional therefore read
8
> p 6¼ q, p, q 2 weak,
>
> pffiffiffiffiffiffiffiffiffiffi
>
> p 2 weak, q 2 frontier ðweakÞ
>
> n p nq
>
> p 2 frontier ðweakÞ, q 2 weak
>
>
>
>
>
< p 6¼ q, p, q 2 strong,
GBBC3 ¼ p 2 strong, q 2 frontier, ð43Þ
pq
>
> n p nq
>
> p 2 frontier, q 2 strong,
>
>
>
>
>
>
> n2p
> p ¼ q, p=
2frontier,
>
>
: pffiffiffiffiffiffiffiffiffiffi
n p nq otherwise;
136 K. Pernal and K.J.H. Giesbertz
where the frontier (weak) spinorbitals are those that belong to a set of frontier
orbitals and their occupancy number is smaller than 1/2 (antibonding orbitals).
BBC3 functional possesses a correct orbital structure of the two-electron functional
(39) if applied to hydrogen molecule in the dissociating limit and it is able to
reproduce very accurately the whole dissociation curve [27]. Moreover, based on
examples of small atoms and diatomic molecules at equilibrium geometries and in
their dissociation limits, it has been shown that BBC3 accounts for both dynamic
and static correlation yielding correct potential energy and recovering most of the
correlation energy.
A difficulty with practical usage of the BBC3 functional is a need to select
bonding and antibonding orbitals. In a computer implementation a strongly occu-
pied orbital of the lowest occupancy is taken as bonding although a weakly
occupied orbital of the highest occupancy is taken as an antibonding with a
straightforward extension for this selection rule for molecules with more than one
bond. This, however, leads to numerical problems because in the optimization
procedure occupation numbers are varied and the antibonding character of orbitals
may change, which may cause problems with convergence or may result in
obtaining discontinuities in potential energy curves. To avoid the previously
described problems with selecting frontier orbitals and to improve the overall
accuracy of the BBC3 functional, it has been proposed to replace the elements
Gpq present in the BBC functional, cf. (41), with a function G(np,nq). The function
mimics the behavior of the GBBC3pq elements but does it automatically, based on the
values of its arguments. The optimal function has been found by introducing two
parameters and fitting the resulting AC3 functional to accurate energies of two
molecules at equilibrium and stretched-bond geometries [40]. The AC3 functional
yields decent quality potential energy curves for ten-electron molecules, although
for some molecules a small hump in the curve is visible.
A two-electron wavefunction (33) is a special case of the more general closed-
shell N-electron ansatz (N assumed to be even) involving, apart from a reference
determinant Φ0, all determinants arising from diagonal double, diagonal quadruple,
etc., excitations; cf. (67). Taking the expectation value of the Hamiltonian with
such a wavefunction yields an energy expression involving only Coulomb,
exchange, and integrals of the hpp|qqi type [41]. In [42] the previously mentioned
ansatz for the wavefunction has been used in development of the extended L€owdin–
Shull (ELS) functional. The functional is applicable to systems for which a set of
the natural spinorbitals can be partitioned into “inner” orbitals localized on atoms
and the occupancies close to 1 and “outer” orbitals including a bonding orbital and
all weakly occupied orbitals, i.e., orbitals localized on a single bond. For N-electron
(N being even) systems (N/2 1) strongly occupied orbitals (in a spin-restricted
formulation each orbital gives rise to two spinorbitals with opposite spins) would be
considered “inner” and the remaining strongly occupied orbital of the lowest
occupancy would belong to the “outer” orbital set. By analyzing a structure of the
energy expression resulting from the assumed ansatz, the following form of the ELS
functional has been proposed
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 137
1 X pffiffiffiffiffiffiffiffiffiffi 1 X
ee ½γ ¼
EELS f p f q n p nq ppqq þ n p pppp
2 p, q2outer 2
X X
p2inner
þ n p nq pq pq pqq p
p2inner q> p
X
X q2inner
ð44Þ
where the phase factors {fp} for the outer orbitals are set according to the rule valid
for two-electron systems, namely
8
< 1 if n p > 1
8 p2outer f p ¼ 2 : ð45Þ
: 1 if n p < 1
2
Note that even though the exchange integrals hpq|qpi are identical to hpp|qqi if the
spinorbitals are real, the two types of integrals make different contributions to time-
dependent linear response equations so they are kept separately in the ELS func-
tional. It is evident that for a two-electron system the set of inner spinorbitals is
empty and, unlike the BBC3 or AC3 functionals, the ELS reduces to the accurate
functional given in (39). A few models have been tried for the function FL, which is
responsible for correlating inner and outer orbitals. The most successful ones
include one or two empirical parameters fitted to reproduce potential energy curves
of LiH, Li2, and BH+ molecules. Very accurate potential energy curves have been
obtained for these molecules [42]. Unfortunately, applications to other systems
have not been presented, because the functional has been designed to treat only
molecules with one single bond and no lone electron pairs. Nevertheless, ELS is a
promising step towards extending the L€owdin–Shull functional to more than two
electrons, aiming at providing a balanced description of the dynamic and static
correlation.
Γpqrs ¼ Γ∗
rspq ; ð46Þ
antisymmetry
have been imposed. To narrow down the possible form of the 2-RDM as a function
of the occupation numbers, it has been required that the final correlation energy
functional includes only Coulomb integrals hpq|pqi, exchange integrals hpq|qpi,
and integrals of the type hpp|qqi. It should be noted that the last two sets are
identical if the orbitals are real, which is the case in practical calculations, but
they enter the time-dependent density matrix functional equations in different terms
as discussed in Sect. 5. Piris and Ugalde [49, 50] proposed the following structure of
the spin-blocks of the cumulant matrix in a spin-restricted formalism
λσσ σσ
pqrs ¼ Δ pq δ pr δqs δps δqr ; ð49Þ
λαβ
pqrs ¼ Δαβ
pq δ pr δqs þ Πrp δ pq δrs ; ð50Þ
where σ ¼ α, β, the Δ matrices are symmetric, and the Π matrix is Hermitian. Such
an ansatz for the cumulant results in the 2-RDM given in (28) satisfying the
symmetry and antisymmetry conditions; cf. (46) and (47). For Systems in a singlet
state, for which n pα ¼ n pβ ¼ n p , and λαα ββ
pqrs ¼ λpqrs , PNOF functionals, resulting
from employing a reconstruction of Γ given in (28) with the ansatz (49) and (50),
are of the following spin-summed form:
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 139
X
PNOF
Eee ½γ ¼ n p nq 2 pq pq pqq p
X h i
pq
Δαα αβ
pq þ Δ pq pq pq Δαα
pq pq q p ð51Þ
X pq
þ Π pq ppqq ;
pq
where the indices p, q pertain to spatial parts of the natural spinorbitals. Diagonal
elements of the Δαβ and Π matrices have been fixed by imposing conservation of
spin [51] which for high-spin states amounts to requiring that the expectation value
of the Sˆ 2 operator computed with the assumed form of the 2-RDM is equal to
Ms(Ms + 1), with Ms ¼ (Nα Nβ)/2, Nα Nβ. The pertinent diagonal elements read
[50]
Δαβ α β
pp ¼ n p n p ; ð52Þ
Πpp ¼ n βp : ð53Þ
The final forms of the off-diagonal elements of the Δ and Π matrices have been
proposed by further imposing a sum rule given in (48) and exploiting the
so-called D, G, Q-conditions that state that 2-RDM, the electron–hole density
matrix G, and two-hole density matrix Q must be positive [50]. The first PNO
functional, PNOF1 [43], has been proposed for singlets after setting Δαα ¼ Δαβ,
assuming dependence of the cumulant matrix on two occupation numbers with
relevant indices, i.e.,
λ pq ¼ λ pq n p ; nq ; ð54Þ
8 p, q2P n p þ nq ¼ 1: ð55Þ
It should be noted that imposing the condition (55) immediately implies that the
normalization condition for 1-RDM, namely
N=2 X
X
2 np ¼ N ð56Þ
P¼1 p2P
is satisfied. In (56) the first summation runs over pairs of electrons and the condition
(55) has been employed. Analogously to other PNOF functionals, the diagonal
elements of the Δ and Π matrices employed in PNOF5 are given by (52) and (53),
whereas the off-diagonal elements for the coupled indices p and q have been
proposed to take the form
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 141
8 p, q2P Δ pq ¼ n p nq ; ð57aÞ
pffiffiffiffiffiffiffiffiffiffi
8 p, q2P Π pq ¼ n p nq : ð57bÞ
The resulting spin-summed expression of the PNOF5 functional reads [47, 49]
N=2 X X
X
EPNOF5
ee ½γ ¼ n p nq 2 pq pq pqq p
P6¼Q p2P q2Q
XX X ð58Þ
pffiffiffiffiffiffiffiffiffiffi X
N=2
n p nq pp qq þ n p pp pp ;
P p2P q2P p
q6¼ p
where P and Q stand for indices of pairs of coupled orbitals. PNOF5 has
outperformed all its PNOF predecessors in describing multireference systems. In
particular it has been shown that it describes qualitatively correctly dissociation
curves yielding accurate dissociation energies [54–56] and products of dissociation
are of integer numbers of electrons [47, 56]. Dissociating of molecules with
multiple bonds, e.g., N2 or CO, leads to products of a correct high-spin symmetry
[56]. The ability of the PNOF5 functional to treat homolytic bond cleavage has been
exploited in its application to radical formation reactions [54]. Unfortunately, good
performance of PNOF5 in recovering static correlation in nearly degenerate sys-
tems is paralleled by its insufficient inclusion of the dynamic correlation [49, 57,
58]. Application of the PNOF5 functional for such challenging systems as Cr2, Mo2,
and W2 dimers revealed that, although it yields energies of an accuracy between
that of the CASSCF and CASPT2 methods, the lack of an important portion of
dynamic correlation energy spoils the results [55]. In order to add the missing
interpair dynamic correlation to PNOF5 Piris has considered a second-order
multiconfiguration perturbation theory [59] and has adopted it for a wavefunction
which leads to the PNOF5 energy expression [57]. The method has been named
PNOF5-SC2-MCPT. Quite unexpectedly, its application to description of the
helium dimer has led to a curve with multiple minima. Moreover, homolytic
dissociation of diatomic molecules with the perturbation method resulted in break-
down of the dissociation curves because of singularities in the second-order energy
appearing for quasi-degenerate systems. The former problem has been avoided by
excluding from the perturbative expansion determinants corresponding to double
excitations from spinorbitals of the same spatial parts, whereas singularities have
been eliminated by removing second-order terms corresponding to quasi-
degenerate orbitals [57]. Such a modified perturbation method has been called
PNOF5-PT2. Application of PNOF5-SC2-MCPT and PNOF5-PT2 to the G2/97
test set of molecules has shown that, on average, the methods recover, respectively,
around 80% and 70% of the correlation energy (with respect to Hartree–Fock
energies) [58].
Good performance of the uncorrected PNOF5 for chemical reactions is a con-
sequence of the observation that the functional can also be derived within the
142 K. Pernal and K.J.H. Giesbertz
Y
N=2
ΨAPSG ðx1 ; . . . ; xN Þ ¼ A^ ψ P ðx2P1 ; x2P Þ; ð59Þ
P¼1
ð
which are strongly orthogonal, i.e., 8P6¼Q ψ P ðx1 ; x2 Þψ Q x01 ; x2 dx2 ¼ 0 [64, 65]. It
can be shown that if geminals are expanded in the natural orbitals {φp(r)}
corresponding to the 1-RDM obtained from the ansatz (59), then the expansion
for each geminal P is diagonal, i.e.,
X
ψ P ðx1 ; x2 Þ ¼ 21=2 c p φ p ðr1 Þφ p ðr2 Þ½αð1Þβð2Þ αð2Þβð1Þ; ð60Þ
p2P
the coefficients {cp} are simply square roots of the corresponding occupation
numbers taken with “+” or “” sign
8p n p ¼ c2p ð61Þ
and the strong orthogonality of geminals implies that the sets of orbitals belonging
to individual geminals are disjointed, i.e., each natural orbital belongs to only one
geminal [66]. It should be noted that for a closed-shell two-electron system the
APSG wavefunction is exact and identical with the L€owdin and Shull function
given in (33). The expectation value of the Hamiltonian with the APSG
wavefunction yields the following spin-summed electron–electron repulsion energy
expression
X
N=2 X
pffiffiffiffiffiffiffiffiffiffi
APSG
Eee f p ; np ; φp ¼ f p f q n p nq ppqq
P p, q2P
ð62Þ
X
N=2 X
þ n p nq 2 pq pq pqq p ;
P6¼Q p2P, q2Q
where {fp} are phase factors of the value +1 or 1. The APSG ground state energy
follows from optimization of the total energy with respect to phase factors, occu-
pation numbers, and the orbitals. Actually, it turns out that typically each geminal is
composed of one strongly occupied orbital of the occupation number greater than
1/2 and a pertinent phase factor f1 ¼ +1, and all other orbitals, which are weakly
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 143
occupied (np < 1/2), of the corresponding phase factors fp ¼ 1 [61, 67]. As already
mentioned in Sect. 2.1 for two-electron singlet systems (described within the APSG
theory by one geminal), exceptions to this rule have been observed [36, 39], but
they occur for very weakly occupied orbitals. Practically, fixing the phases in the
APSG functional given in (62) according to the aforementioned rule, instead of
finding them variationally, has only a small effect on the APSG energy. The APSG
functional with the phase factors fixed can be seen as a density matrix functional.
Comparison of the PNOF5 functional defined by (58) and (55) with (62) immedi-
ately reveals that they are identical if the dimension of the expansion space for each
geminal in the APSG approach is limited to 2 and the phase factors of the two
orbitals which form a given geminal are opposite, i.e., f1 ¼ f2 [60]. Because
PNOF5 is equivalent to such constrained APSG approximation, it inherits its
features from the latter, which explains the good performance of the PNOF5
functional for predicting dissociation energy curves of molecules [47] and the
localized character of its optimal orbitals [68].
Lifting the restriction on the dimensionality of expansion spaces for the geminals
in PNOF5 functional should allow one to recover a part of the correlation energy
missing in this functional. This procedure has been proposed in [69] but clearly
such extended PNOF5 functional (PNOF5e) is identical to the APSG functional
(62) with fixed phases. For PNOF5 and PNOF5e functionals a systematic recon-
struction of the 2-RDM in terms of the 1-RDM has merged with a theory based on
the ansatz for the wavefunction [49]. On one hand this may seem to be a desirable
result – the functionals are N-representable and bound by the exact ground state
energy, but the drawback is that the functionals suffer from the same deficiencies as
the APSG approximation.
An interesting idea that leads to incorporating the dynamic correlation that
PNOF5 lacks has been proposed in [48]. The intrapair correlation is included in
PNOF5 by proposing the elements Δpq (57a) and Πpq (57b) corresponding to
uncoupled orbitals p and q (belonging to different pairs) to be nonzero and
expressing them as functions of the occupation numbers. The new functional,
PNOF6, employs, similarly to PNOF5, a paired-orbitals picture. Compared to
PNOF5, the PNOF6 functional underestimates the dissociation energies to a lesser
degree. Unlike its predecessor, PNOF6 yields delocalized orbitals and it avoids
spatial symmetry breaking of the benzene equilibrium geometry [48].
An ongoing development of natural orbital functionals, PNOF, originating from
reconstruction of 2-RDM in terms of 1-RDM, has already resulted in functionals
competing in accuracy with MP2 method for single-reference systems. Unlike the
MP2 method, the PNOF4, PNOF5, and PNOF6 functionals are useful in describing
potential energy surface also when bonds are stretched and dissociation potential
energy curves are often of the quality of the much more expensive CASSCF
approach.
144 K. Pernal and K.J.H. Giesbertz
The resulting Yasuda correlation functional, cf. Eq.(29), Ec[γ] ¼ ∑pεp[γ] is there-
fore implicitly dependent on spectral components of γ. Parameters {εp} are found
from a set of auxiliary equations. Despite the fact that the Yasuda correlation
functional possesses a number of desirable features, i.e., it satisfies the exact
conditions given in (30) and (31), it gives rise to dispersion interaction [70] and
recovers a logarithmic divergence of the correlation energy of the homogeneous
electron gas in high-density limit [71], its usefulness in practical electronic struc-
ture calculations has been undermined by showing that it does not seem to be bound
from below even for two-electron systems [72].
Quite a different approach has been assumed in [41, 73–76] where explicit
density matrix functionals have been derived by assuming a configuration interac-
tion (CI) ansatz for a wavefunction and parameterizing CI coefficient. In all cases
the CI wavefunctions were such that the resulting expression for the energy
involved only Coulomb and exchange two-electron integrals. Because the former
integrals are often denoted with the letter J and the latter with K, the functionals
involving only these two types of integrals are sometimes called “JK-only” func-
tionals. The idea of constructing functionals by parameterizing the CI ansatz is
evidently directly related to the Levy constrained search functional (9) which for
the CI wavefunction
X
Ψ¼ CI ΦI ; ð64Þ
I
C stands for a vector of all CI coefficients and the minimization in the functional
(65) is carried out with respect to all vectors C corresponding to the assumed ansatz
for a wavefunction (64) which yield a given density matrix γ. Were the expansion in
(64) complete, the functional (65) would be exact, i.e., it would be equivalent to the
Levy functional (9). Otherwise, for truncated CI expansion, the functional is only
approximate. The strategy adopted in [41, 73–76] was to use a CI ansatz leading to a
“JK-only” approximation and to replace the whole set of the CI parameters by
auxiliary variational matrices, say A and B, such that the functional (65) turns into
( )
X X
Eee n p ; φ p ¼ min A pq pq pq þ B pq pqq p ; ð66Þ
A, B!n
pq pq
where the two-electron integrals are computed with the natural spinorbitals {φp}.
The minimization is done with respect to the matrices A, B constrained to yield a
given vector of occupation numbers n and to satisfy some conditions, for example
conditions imposing size-consistency on the functional. If the conditions are such
that there is a mapping from A and B to the CI coefficients C, the functional given
in (66) is variational, i.e., it constitutes an upper bound to the functional (65) and the
exact Levy functional (9). If, however, the matrices A, B are constructed to ensure
that the underlying 2-RDM only satisfies some of the necessary N-representability
conditions, the functional (66) is not necessarily variational. The main advantage of
replacing functionals (65) with approximations (66) is to obtain a more efficient
method than CI, because the complex objects (CI coefficients) are replaced by
matrices of much smaller dimensionalities. Moreover, if the starting CI ansatz (64)
is not size-consistent, the proposed reparameterization in terms of A, B could
restore this property (but then variationality is lost).
In [74] Kollmar and Hess considered a CI wavefunction being a combination of
a closed-shell reference Slater determinant Φ0 and determinants arising from Φ0 by
doubly exciting electrons from spinorbitals of the same spatial parts to virtual
a a
orbitals also sharing spatial functions, i.e., Φiααiβ β , where i and a stand for, respec-
tively, occupied and unoccupied orbitals in the reference state. Such an ansatz leads
to an energy expression involving only Coulomb and exchange integrals but it lacks
size-consistency. To recover this property a normalization condition has been
replaced by a new condition on the CI coefficients. The resulting functional of
the form of (66) has been applied to the description of symmetric dissociation of
water molecule which has led to a potential energy curve of a reasonable shape. At
the same time, it became evident that the functional misses dynamic correlation.
In [41] the most general form of the closed-shell CI wavefunction which leads
only to Coulomb and exchange integrals in the energy expression has been consid-
ered. The wavefunction can be called pair-excited CI because it includes all
possible Slater determinants, each built of N/2 spatial orbitals entering a determi-
nant with the α and β spin component, i.e.,
146 K. Pernal and K.J.H. Giesbertz
X
1
Ψ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi AQN=2 ϕq1 ϕq1 ϕqN=2 ϕqN=2 ; ð67Þ
ðN=2Þ!QN=2
where QN/2 ¼ q1,. . .,qN/2 is a string of indices, and ϕq1 ϕq1 ϕqN=2 ϕqN=2 is a Slater
determinant built of N/2nelements
o from the set of one-electron spatial functions
{ϕp}. The coefficients AQN=2 are symmetric with respect to exchanging two
indices in a string QN/2 and they satisfy a normalization condition. The
wavefunction (67) is size-consistent and it yields a diagonal 1-RDM, i.e.,
X 2
γ pq ¼ 2δ pq N A pQN=21 ¼ 2δ pq n p : ð68Þ
QN=21
Therefore one immediately recognizes that the orbitals {ϕp} are simply the natural
orbitals, i.e., φ p ðrÞ ϕ p ðrÞ. As already mentioned, another property of the ansatz
(67) is that the spin-summed electron interaction part of the corresponding energy
reading
X
X
Eee ¼ F pq 2 pq pq pqq p þ G pq ppqq ; ð69Þ
pq pq
where
X 2
N N
F pq ¼ 1 A pqQN=21 ; ð70Þ
2 2 Q N=22
NX
G pq ¼ A pQN=21 AqQN=21 ; ð71Þ
2Q
N=21
is a strict upper bound to the exact energy functional. A large number of possible
phase factors makes minimization of the functional practically impossible. Intro-
ducing some fixed pattern for signs of those factors could destroy variationality of
the functional. This has been called a “phase dilemma” in [41] and identified as a
serious bottleneck in constructing density matrix functionals. A functional free of
the phase factors has been obtained by Kollmar and Hess by reconstructing 2-RDM
in terms of 1-RDM by imposing N-representability condition (a strategy similar to
the one adopted in construction of PNOF functionals presented in Sect. 2.2
[75]). The Kollmar–Hess functional is identical to (72) if a simple approximation
for phase factor products is assumed [41]. The functional is therefore not variational
in general (except for two-electron systems for which the functional is exact).
Numerical applications showed that it is very accurate for four-electron systems
[75]. The results for water molecule undergoing symmetric dissociation are much
less satisfactory. They are very close, however, to those corresponding to the
closed-shell MC-SCF approach with the CI ansatz given in (67). One can therefore
conclude that the phase dilemma is not such a serious limitation if a proper model is
assumed for the phase factors. Another confirmation of this conclusion comes from
considering the exact functional for a two-electron closed-shell system which is a
special case of the functional given in (72) for N ¼ 2. As has already been men-
tioned, fixing the signs of the phase factors corresponding to the weakly occupied
orbitals to be opposite to the sign of the phase factor associated with the strongly
occupied orbitals leads to only a small change in the energy.
In [76] it has been shown that a computationally hard MC-SCF problem can be
replaced by the optimization of a simple 1-RDM functional which parallels the
MC-SCF method in accuracy. However, it has also been pointed out that the ansatz
(67), i.e., the best possible wavefunction leading to a “JK-only” expression for the
energy (67), recovers only a small fraction of the correlation energy for systems as
small as a ten-electron molecule. Any variational (or based on an N-representable
2-RDM) “JK-only” functional suffers from the same deficiency. The density matrix
functionals tested in [76] developed by imposing basic necessary N-representability
conditions on the underlying 2-RDM do not recover more correlation than the
wavefunction-based approach, even though they are not variational. In other words,
results of parallel accuracy are obtained by minimizing the CI energy given by
(69)–(71) with respect to the CI coefficients and the orbitals {ϕp} and by optimizing
“JK-only” functionals proposed as approximations to (69) [76]. This poses a
question as to whether the pair-excited CI ansatz (67) is a good starting point for
developing functionals. This question is addressed in Sect. 6.
Apart from the implicit density matrix functionals discussed earlier which
involve some auxiliary parameters, cf. (66), a promising class of functionals
depending on γ and electron density ρ has been proposed by combining density-
functional and density matrix functional theory [77, 78]. The method is based on the
range-separation of electron–electron interaction operator, V^ ee , into short- and
long-range parts, V^ eesr
and V^ lree , respectively [79, 80]. Dynamic correlation energy
should mostly be described by the short-range density functional, and static
148 K. Pernal and K.J.H. Giesbertz
ργ is a diagonal part of the density matrix γ and the definition of the short-range
universal density functional Fsr also employs Levy constrained search construction.
Partitioning the long- and short-range functionals into Hartree and exchange-
correlation parts results in obtaining the following srDFT-lrRDMFT energy
functional
Ground state energy follows from minimizing the range-separated functional with
respect to N-representable γ. A short-range PBE exchange-correlation functional
[81] has been used for Esr
xc [ρ]. This is a short-range version of the PBE functional
derived for the error function employed in range-separation of electronic interac-
tion. The long-range density matrix functional, Elrxc , has been approximated by the
long-range BB functional ElrBB
xc obtained by simply replacing two-electron integrals
in the full-range BB functional, cf. (20), by their long-range counterparts, namely
the spin-summed expression which reads
X pffiffiffiffiffiffiffiffiffiffi
lrBB
EXC ½γ ¼ n p nq h pqjq pilr ; ð76Þ
pq
ðð
erf ðμr12 Þ
h pqjq pilr ¼ φ*p ðr1 Þφ*q ðr2 Þ φq ðr1 Þφ p ðr2 Þdr1 dr2 ; ð77Þ
r 12
where r12 ¼ |r1 r2| and erf stands for the error function. Both short- (sr) and long-
range (lr) functionals involve a range-separation parameter μ, the value of which
has been empirically chosen to be 0.4 bohr1. Such an obtained srPBE-lrBB
functional has been applied to a few diatomic molecules and, in contrast to full-
range BB or PBE functional, the range-separated density matrix functional turned
out to be capable of reproducing correct dissociation energy curves [78]. Another
direct advantage of range-separated functionals over full-range density matrix
functionals is that the former, unlike the latter, are weakly basis set dependent
which adds to their favorable computational efficiency.
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 149
2X 0
γ ðr; r0 Þ ¼ nðkÞeikðrr Þ ; ð78Þ
V k
where k ¼ |k|, n(k) is called momentum distribution, and V is the volume of the
system (V ! 1). BBC functionals, cf. (41)–(43), developed for molecules involve
in their definition partitioning natural orbitals into strongly and weakly occupied,
which is based on the value of the pertaining occupation number. A straightforward
generalization of the BBC functionals to extended systems would assume
establishing a type of the natural orbital (a plane wave) on the basis of the k-number
i.e., whether it is smaller or greater than some reference value kc [82]. The most
obvious choice would be kc ¼ kF, where kF is the Fermi wave vector. As mentioned
before, this choice implemented in the BBC1 or BBC2 functionals does not lead to
accurate correlation energy of HEG. Lathiotakis et al. proposed two variants of the
BBC1 modifications [82]. In the first, kc was treated as a parameter, whereas the
second variant assumes keeping kc ¼ kF, multiplying the exchange-correlation
terms of the BBC1 functional corresponding to two weakly occupied orbitals by a
parameter s (s-functional). In both cases, values of parameters were chosen to
reproduce the exact correlation energy of HEG. Unfortunately, momentum distri-
butions resulting from such proposed functionals obtained for metallic densities,
even though they show discontinuity, quantitatively still deviate strongly from the
150 K. Pernal and K.J.H. Giesbertz
accurate references. Adaptation of the s-functional developed for the electron gas to
finite systems has led to surprisingly accurate values of energy for molecules at
their equilibrium geometry but it has been also revealed that the functional is not
size-consistent and it fails in the description of potential energy curves [85].
Motivated by the fact that the exchange-correlation functional (18) in many
density matrix functionals is approximated by an expression involving only
exchange integrals multiplied by factors depending on two pertinent occupation
numbers, i.e.,
1X
Exc ½γ ¼ G n p ; nq pqq p ; ð79Þ
2 pq
Marques and Lathiotakis (ML) proposed to find the function G fully empirically by
using a Padé approximant depending on a variable x ¼ npnq [87]. Coefficients in the
Padé approximant were found by minimizing the error of the correlation energies of
selected test-molecules. Computing the correlation energies of molecules in a G2
test with different methods has revealed that the empirical ML functional is on
average the most accurate of all functionals tested, competing with or being
superior to the MP2 method [87]. However, because the exchange-correlation
part depends only on products of two occupation numbers, it inevitably lacks the
structure needed to describe the breaking of a two-electron bond. The ML func-
tional is not appropriate for describing molecules at geometries far from their
equilibrium.
In the quest to develop a computationally efficient 1-RDM functional which is
useful for solids, a very simple idea has been proposed and leads to remarkable
results. The first and simplest approximate density matrix functional proposed is the
BB functional (also known as the Müller functional) [18, 19], cf. (20). Müller has
arrived at the particular form for the exchange-correlation functional given in (18)
by considering a generalization of the Hartree–Fock exchange functional (17),
which assumes replacing |γ(x, x0 )|2 present in the HF two-particle density matrix
and, consequently, in the functional (17), by a product γ 1α(x, x0 )γ α(x, x0 )*. The
power α was constrained to belong to the interval h0, 1i, to assure convexity of the
functional and integrating of the corresponding exchange-correlation hole to 1
[18]. The BB functional results from choosing α ¼ 1/2. Sharma et al. proposed to
consider an approximate exchange-correlation functional of the form [84]
1X α
α
Exc ½γ ¼ n p nq pqq p ; ð80Þ
2 pq
that for α ¼ 1 is just an exchange Hartree–Fock functional (17) whereas for α ¼ 1/2
it turns into a BB form (20). It should be mentioned that a 2-RDM
α
Γpqrs ¼ n p nq δ pr δqs n p nq δps δqr ; ð81Þ
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 151
giving rise (via (26)) to the exchange-correlation power functional (80) satisfies the
sum rule (48) only for α ¼ 1/2. Application of the power functional to the homo-
geneous electron gas revealed that momentum distribution resulting from the
functional optimization not only lacks a step structure but is very different from
the exact distributions in general [83, 88]. Even though the power functional does
not recover the exact high-density limit of the correlation energy of the HEG [83,
89], it has been shown that with a carefully chosen value of α it is possible to obtain
rather accurate values of the correlation energy for this system in the broad range of
densities [86]. Moreover, the power functional performs remarkably well in
predicting accurate band gaps of semiconductors and insulators [84]. The test set
included materials of covalent or ionic character with band gaps ranging from 1 to
14.2 eV. It is striking that all these systems are incorrectly predicted to be metallic if
described with the α ¼ 1/2 power functional, whereas choosing α ¼ 0.65 or α ¼ 0.7
results in obtaining nonzero gaps deviating on average from experimental values by
less than 10%. Reducing α below the value 0.65 leads to zero energy gap for some
materials, so it seems the range of admissible values of α is quite narrow.
Performance of the simple power functional (80) with α 2 (0.65, 0.7) when
applied to transition metal oxides (TMO) is even more impressive. TMOs can be
regarded as prototypes of strongly correlated Mott insulators, the nonzero band gap
of which is a result of strong Mott–Hubbard correlations. Most approximate density
functionals incorrectly predict TMO to be metallic. The power density matrix
functional, however, yields finite values for band gaps of nonmagnetic TMOs,
although deviations of the computed gaps from their experimental counterparts
are larger than in the case of conventional insulators [84]. Sharma et al. also showed
that apart from band gaps the power functional is capable of accurately predicting
other properties of solids such as equilibrium lattice constants. Another successful
application of the power functional includes predicting the photoelectron spectra of
strongly correlated Mott insulators within a density matrix functional method
proposed in [90]. Despite its very simplistic form, the power functional has been
shown to be a useful tool for studying solids, including those for which most density
functionals provide unreliable results.
X
N
^0 ¼
H w^ ðλ; xi Þ; ð82Þ
i
where the occupancies {np} and the natural spinorbitals {φp} correspond to
unperturbed γ. Second-order properties are given by the expression (valid for real
orbitals)
* +
∂ E½γ X
2 ∂2 w Xh ð1Þ i ^
∂ w
^
¼ n p φp 2 φp þ n p δ pq þ nq n p U pq φp φp ; ð84Þ
ð1Þ
∂λ 2
p
∂λ pq
∂λ
ð1Þ
where {np } and U(1) are components of the first-order response of γ, namely
1Þ
γ ðpq
1Þ
¼ nðp1Þ δ pq þ nq n p U ðpq : ð85Þ
They can be found by solving a set of linear coupled-perturbed equations [91]. If the
perturbation is particle-number-conserving then a condition must be imposed that a
ð1Þ
sum of perturbations {np } vanishes. Applying the response equations to compute
the static polarizabilities has revealed that even functionals which perform well in
predicting energies of atoms and diatomic molecules, e.g., BBC3, do not provide
satisfactory results for the second-order response properties [91]. The values for
polarizabilities are of comparable or even worse quality than those obtained within
the coupled-perturbed Hartree–Fock method [91]. Much more encouraging results
have been obtained for hyperpolarizabilities of the H2 molecule using the PNOF5
functional within a finite field approach [92]. Good accuracy could have been
expected though, because the PNOF5 functional, cf. (58), is equivalent to the
two-electron functional (39) if the number of orbitals with nonzero occupancy is
restricted to two [60]. Despite this constraint, the PNOF5 functional captures the
right physics of two-electron systems.
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 153
A number of methods for predicting vertical ionization potentials (IPs) have been
proposed within RDMFT. The most straightforward approach involves performing
calculations of energies for a neutral and ionized species [93–95]. Apart from the
fact that such multiple calculations are time consuming, it has been pointed out in
[95] that inaccuracy may arise because of different treatment of closed- and open-
shell systems in RDMFT.
Another way of computing IPs is provided by the Extended Koopmans’ Theo-
rem (EKT) which connects 1- and 2-RDM of a Coulombic system with its ioniza-
tion potentials [96–98]. It has been shown in [99] that the Lagrangian matrix λ
related to imposing orthonormality of the natural orbitals in optimizing a density
matrix functional, reading
ð
δEee ½γ *
λ pq ¼ n p hqp þ φ ðxÞ dx; ð86Þ
δφ*p ðxÞ q
is equivalent to the generalized Fock matrix of the EKT equations. This implies that
diagonalization of a Hermitian matrix Λ defined as
λ pq
Λ pq ¼ pffiffiffiffiffiffiffiffiffiffi ð87Þ
n p nq
yields IPs as eigenvalues. For small molecules the BBC and GU functionals
employed in the EKT formalism yield ionization energies with errors with respect
to experimental references of the order of 4–6%. Similar accuracy has been
obtained with the PNOF functionals [93, 100]. On average the accuracy is higher
than that of the standard Koopmans’ theorem.
The EKT method is not practical for solids as it would require diagonalization of
a very large matrix. Sharma et al. proposed an alternative method [90]. This consists
of assigning to each natural spinorbital an orbital energy εp obtained as a derivative
of the total energy with respect to the pertinent occupation np. The derivative is
taken at np ¼ 1/2 with the rest of occupation numbers set equal to their ground state
optimal values, i.e.,
∂E½γ
εp ¼ : ð88Þ
∂n p n p ¼1=2
Employing orbital energies obtained in this way for predicting densities of states of
transition metal oxides has led to excellent agreement with experimental data. The
orbital energies defined in (88) have also been used as approximations to ionization
energies and electron affinities of molecules. Performance of density matrix func-
tionals within such an approach is satisfactory and for IPs the errors are of the same
order as those obtained from the much more theoretically grounded EKT method [95].
154 K. Pernal and K.J.H. Giesbertz
Another quantity of key importance for solids accessible in RDMFT is the band gap
or more generally the fundamental gap, which is defined as the difference between
the ionization potential I and the electron affinity A as
Δ ¼ I A: ð89Þ
Helbig et al. proved that within exact formulation of RDMFT a Lagrange multiplier
μ, used to impose the normalization condition (5) on the occupation numbers in
variational equations, possesses a discontinuity at integer particle numbers. This
discontinuity amounts to the fundamental gap [106, 107], i.e.,
φ ¼ eX χ ð92Þ
which assures orthonormality of the orbitals φ, cf. (3), [26, 27]. To satisfy the N-
representability condition given in (4) the natural occupation numbers may be
parameterized by cosine functions, namely 8p np ¼ cos2(xp) where parameters
{xp} are unconstrained. The normalization condition (5) is taken into account by
means of a Lagrange multiplier. A bottleneck of a two-step procedure is optimiza-
tion of the orbitals. It takes many iterations to meet tight convergence criteria,
because energy is almost completely insensitive to variations of very weakly
occupied orbitals.
Because of unsatisfactory efficiency of the gradient orbital optimization algo-
rithms, efforts have been made to turn the optimization problem for orbitals into an
eigenproblem for an effective Hamiltonian [1, 103, 110–113]. For a given
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 157
5 Time-Dependent RDMFT
The extension of ground state RDMFT to the time domain was recently considered
[117–121]. The main motivation to develop time-dependent RDMFT
(TD-RDMFT) is the poor performance of time-dependent DFT (TDDFT) in the
adiabatic approximation in combination with the approximate ground state density
functionals. The best known failure is the inability of approximate TDDFT to
capture charge transfer excitations [122, 123], though this deficiency has been
remedied with some success using range-separated hybrid functionals [123, 124]
and by an explicit reconstruction of the spatial divergence in the kernel [125,
126]. Other failures of adiabatic TDDFT are bond-breaking excitations which are
predicted to be too low in energy upon dissociation (they can even go to zero) [127,
128] and a lack of double excitations [128–130]. All these failures are connected to
the inability of approximate adiabatic density functionals to deal with static
158 K. Pernal and K.J.H. Giesbertz
The equation of motion for the 1-RDM is readily obtained from the time-dependent
Schr€
odinger equation
i∂t γ ðx; x0 ; tÞ ¼ h^ ð
x; tÞ h^ ðx0 ; tÞ γ ðx; x0 ; tÞþ
ð
1 1 ð94Þ
Γðxx2 , x0 x2 ; tÞdx2 ;
jr r2 j jr0 r2 j
So we find that we need the 2-RDM to determine the evolution of the 1-RDM. It
turns out that the evolution of the 2-RDM is coupled to the 3-RDM and so on, till we
hit the full N-RDM. This chain of p-RDMs coupled to each other is known as the
Bogoliubov–Born–Green–Kirkwood–Yvon (BBGKY) hierarchy [131–137]. To be
of any practical use, the BBGKY hierarchy needs to be truncated at some level. In
TD-RDMFT the hierarchy is truncated already at the level of the 1-RDM and it is
assumed that the time-dependent 2-RDM is a functional of the 1-RDM, Γ[γ]. For
Hamiltonians with only local potentials, we know from the Runge–Gross theorem
[138] and its extension [139, 140] that this is indeed true, because all observables
are already functionals of the density, so they are also functionals of the 1-RDM.
The use of density for the formal foundations of TD-RDMFT is not satisfactory,
however, because we would have to limit ourselves to local-potential representable
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 159
1-RDMs. The formulation of a solid foundation for TD-RDMFT which allows for
non-local potentials is still an open challenge.
The time-dependent 1-RDM can be diagonalized at each time t, allowing one to
consider the time-dependent natural spinorbitals, φp(xt), and time-dependent occu-
pation numbers, np(t). The equation of motion for the 1-RDM (94) can be trans-
formed to the time-dependent natural spinorbital basis, which gives the equations of
motion for the natural spinorbitals and occupation numbers separately [118, 119]
i nq ðtÞ n p ðtÞ φ p ðtÞjφ_ q ðtÞ ¼ nq ðtÞ n p ðtÞ h pq ðtÞ þ W {pq ðtÞ W pq ðtÞ 8 p6¼q ,
ð96aÞ
in_ p ðtÞ ¼ W {pp ðtÞ W pp ðtÞ ; ð96bÞ
It should be noted that the two-electron integrals are also time-dependent here,
because they are evaluated with the time-dependent natural spinorbitals.
The time-dependent response equations can be derived from the equation of motion
of the 1-RDM (96) by considering a small time-dependent perturbation to a
stationary system, with the stationary 1-RDM γ 0. The first-order perturbation in
the 1-RDM is directly related to perturbation in the natural spinorbitals and
occupation numbers as [compare with (85)]
δγ pq ðtÞ ¼ δn p ðtÞδ pq þ nq n p δUpq ðtÞ; ð98Þ
The response matrix K[γ 0](τ) is the coupling matrix defined as [117, 119, 121,
141, 142]
δ {
ð Þ ð Þ
0
W pq t W pq t
0
K pq, rs γ ðt t Þ ¼ ð100Þ
0
δγ rs ðt Þ
γ0
and plays the same role as the Hartree-exchange-correlation kernel, fHxc (r, r0 ,
t t0 ), in TDDFT [138, 143]. To obtain the frequency-dependent response equa-
tions, we simply need to take the Fourier transform of the time-dependent response
equations. Because the time-integral over the coupling matrix K(t t0 ) and the
perturbation in the 1-RDM has the form of a convolution, the Fourier transform
turns this integral into a simple product. If we further assume that the natural
spinorbitals of the unperturbed 1-RDM can be chosen to be real (no magnetic
fields), the frequency-dependent response equations can be cast into a particular
simple matrix form
0 10 R 1 0 1
ω1M Aþ
MM ðωÞ 0 δγ ðωÞ 0
@ N1 A ðωÞN1 ω1M N1 CðωÞ A @ iδUI ðωÞ A ¼ @ δvR ðωÞ A ;
0 AþmM ðωÞ ω1m δnðωÞ 0
ð101Þ
Positive and negative combinations of the response matrix A(ω) enter the
frequency-dependent RDMFT response equations (101) as
K γ 0 ðt t0 Þ K γ 0 δðt t0 Þ: ð104Þ
If the initial state of the system was the ground state, a ground state functional
would probably provide a reasonable approximation and the full standard adiabatic
approximation becomes
Using this approximate 2-RDM in the definition for W(t) (97), and inserting the
result into the equation of motion of the natural occupation numbers (96b), we find
that they are time-independent, in_ p ðtÞ ¼ 0.
More work is needed to demonstrate that the use of a ground state functional for
the 2-RDM always leads to stationary occupation numbers in the standard adiabatic
approximation [117, 144, 145]. First we note that, because the natural orbitals are
the eigenfunctions of the self-adjoint kernel, γ(x,x0 ;t), their phases are undetermined
by the 1-RDM. Therefore, a 1-RDM functional formulated in terms of the natural
orbitals and occupation numbers is not allowed to depend on the phase of the natural
orbitals. Making the phase of the natural orbital explicit φ p ðxtÞ ¼ eiα p ðtÞ ϕ p ðxtÞ, we
have the following condition on the derivative of any 1-RDM functional, F
162 K. Pernal and K.J.H. Giesbertz
ð !
dF ∂F ∂F
0¼ ¼ i dx φ p ðxtÞ *
*
φ ðxtÞ : ð107Þ
dα p ðtÞ ∂φ p ðxtÞ ∂φ p ðxtÞ p
To connect these derivatives with the contractions W(t) (97) in the adiabatic
approximation, we express the exact ground state functional as [99]
h n o i 1 X
W φ p ; φ*p ; n p ¼ min Γpqrs fξi g; n p rs pq ; ð108Þ
2 fξi g pqrs
where the set of variables {ξi} indicates the additional degrees of freedom over
which we can vary the 2-RDM, keeping the 2-RDM ensemble N-representable and
such that it yields the prescribed 1-RDM (48). This expression assumes that only
the occupation numbers are part of the N-representability conditions for the 2-RDM
and not the natural orbitals. This is reasonable, because N-representability should
not depend on the particular orthonormal basis we are working in. An advantage of
expressing the exact interaction-energy function in this manner is that the func-
tional W is even defined for non-orthogonal natural spinorbitals. Although the value
of the functional W does not make any physical sense for non-orthonormal orbitals,
it allows us to define derivatives with respect to ϕp(x) and ϕ
q (x) separately in an
unambiguous manner and impose the orthonormality conditions afterwards with the
help of Lagrange multipliers or in other ways, e.g., (92).
The optimal 2-RDM parameters which attain the minimum are functionals of the
natural orbitals and
h occupation numbers, and we write these optimal value for the
n ∗ o i
parameters as ξi φ p ; φ p ; n p . Using ξi the exact interaction-energy
functional can also be written as
h n o i 1X
W φ p ; φ*p ; n p ¼ Γpqrs ξi ; n p rs pq ; ð109Þ
2 pqrs
ð X
δW
dx φ q ð xÞ ¼ Γprst stqr ¼ W pq : ð111Þ
δφ p ðxÞ rst
Using this result together with the phase invariance condition (107) in the equation
of motion for the occupation numbers (96b) in the standard adiabatic approxima-
tion, we find the claimed result
ð !
∂W ∂W
in_ p ðtÞ ¼ dx φ∗p ðxtÞ ∗ φ ðxtÞ ¼ 0: ð112Þ
∂φ p ðxtÞ ∂φ p ðxtÞ p
Because no δn(ω) term is present, we find that even in the static limit ω ! 0 the
occupation numbers are not perturbed, in contrast to the time-independent response
equations presented in Sect. 3.1 [120, 121, 142]. This discrepancy has been
demonstrated to be sizable by calculating the polarizability of HeH+ [117, 120,
144, 145]. The L€owdin–Shull functional has been exclusively used for these
calculations. There are two possible variants of this functional: one with the
exchange integrals (39) and one where we restore the original integrals hpp|qqi of
the singlet two-electron system (36) and replace products cpcq with GLS pq given in
(40). For real natural spinorbitals there is no difference, but in the time-dependent
case the natural orbitals are complex and hence the two different integrals give rise
to different coupling matrices. The advantage of using exchange integrals is that the
functional is phase invariant, which is a requirement for a proper 1-RDM func-
tional. Therefore, this variant is called the density matrix LS (DMLS). The variant
with the original hpp|qqi integrals is not phase invariant, so not a proper 1-RDM
functional. Because of its phase dependence it is called the phase including LS
(PILS). Though the PILS is not a proper 1-RDM functional, its use is appealing,
because the breaking of phase invariance implies that the natural occupation
numbers do change.
One would expect that the DMLS functional should give superior results. This is
indeed the case for the polarizability of HeH+ if only a limited number of transitions
between the natural orbitals are taken into account [117, 144, 147], typically only
the transitions from the two highest occupied NOs to all the others. If all transitions
between the natural orbitals are taken into account, the DMLS functional has
spurious divergences in the polarizability at low frequencies [117, 145, 147],
severely deteriorating the DMLS result. Though the polarizability from the PILS
164 K. Pernal and K.J.H. Giesbertz
functional does not follow the exact polarizability as closely as the DMLS, it has the
main advantage that no spurious divergences emerge, even when all natural orbital
transitions are taken into account in the response calculation [117, 121, 147].
Another disadvantage of the standard adiabatic approximation is that we effec-
tively lose excitation energies. Because the original frequency-dependent response
matrix has M + M + m ¼ m2 entries, the standard adiabatic approximation yields m2
excitation energies, because the response matrices A and C are frequency indepen-
dent. However, we only obtain 2M ¼ m(m 1) sensible excitation energies and
m roots ω ¼ 0, which are physically meaningless, so, effectively, these m excitation
energies are lost in the standard adiabatic approximation [142].
Test calculations on the excitation spectrum of the hydrogen molecule as a
function of the bond length have been carried out to test the performance of both
LS functionals. Because divergences in the polarizability correspond to excitations,
the spurious divergences of the DMLS functional already indicate that the perfor-
mance of the DMLS functional for the calculation of excitation energies is not very
good. Indeed, the test calculations on H2 have shown that the DMLS functional
predicts many spurious low lying excitations which completely clutter the excita-
tion spectrum when all natural orbital transitions are taken into account [117, 145,
148]. Reducing the number of transitions to only transitions from the two heaviest
occupied natural orbitals is very effective in cleaning up the DMLS excitation
spectrum [117]. Such an approach would not be desirable in practice, because it is
orthogonal to the idea that expanding a basis brings one closer to the desired result.
Because the PILS functional is dependent on the phase of the natural orbitals, the
occupation numbers are not necessarily stationary any more in the standard adia-
batic approximation. However, it can be demonstrated that there are still only
2M ¼ m(m 1) non-trivial roots of the response equations (113) and m zero exci-
tations [117]. Nevertheless, the PILS functional gives a huge improvement over the
DMLS functional for excitation energies. Most notably, no spurious low lying
excitations appear when we exhaust the response basis by including more natural
orbital transitions. Furthermore, one can show that the 1 Σ þ
u , Πg and Πu excitations
1 1
become equal to the full CI result when all natural orbital transitions are taken into
account [117, 141, 142]. This is caused by the fact that these excitations do not need
any perturbation in the natural occupation numbers to be described exactly, which
is related to symmetry. This also holds for excitations in other irreducible repre-
sentations (irreps) that do not couple to the completely symmetric irreducible part
of the response matrix, such as the xy component of the Δg excitations. The x2 y2
component does couple to the occupation numbers, however, so the Δg excitations
of the H2 molecule is symmetry broken when using the PILS functional: the xy
components are equal to the full CI result and the x2 y2 components are not
[117]. Symmetry breaking does not occur for the DMLS functional, because the
occupation numbers are never involved in the standard adiabatic response.
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 165
Because the standard adiabatic approximation used in TD-RDMFT has the unde-
sirable features of stationary occupation numbers and a mismatch with the static
response equations (Sect. 3.1) in the ω ! 0 limit, an alternative adiabatic approx-
imation has been proposed. The idea is to replace the dynamic equation for the
occupation numbers (96b) by its static counterpart and to make the perturbed
quantities frequency-dependent, which leads to the following equation
X X
0¼2 C Tp, rs δγ rsR ðωÞ þ 2 W p, r δnr ðωÞ þ δvpp ðωÞ; ð114Þ
r>s r
where we use
2
1 ∂ W
C Tp, rs ¼ Crs, p and W p, q ¼ : ð115Þ
2 ∂n p ∂nq
Though the occupation numbers are not determined by an equation of motion, but
follow instantaneously from δγ R(ω) and the diagonal elements of the potential
δvD(ω), there is at least a response of the occupation numbers. The fact that this
alternative adiabatic approximation is an instantaneous relaxation of the natural
occupation numbers at each time t has been stressed in [149] where the more
descriptive name “instantaneous occupation number relaxation” was introduced.
The frequency-dependent response equations in this alternative adiabatic approxi-
mation become
0 10 R 1 0 1
ω1M AþMM 0 δγ ðωÞ 0
@ N1 A N1 ω1M N1 C A @ iδUI ðωÞ A ¼ @ δvR ðωÞ A: ð116Þ
CT N1 0 W δnðωÞ δvD ðωÞ=2
The correction for the ω ! 0 limit to the standard adiabatic approximation proves to
be quite effective and improves the description of the polarizability for small
frequencies [117, 120, 121]. Additionally, because the frequency-dependent
response equations now reduce correctly to the static response equations in the
ω ! 0 limit, both the DMLS and PILS functionals coincide at ω ¼ 0. The general
trend from the standard adiabatic approximation remains: the DMLS is closer to the
exact polarizability, though has some spurious divergences which are absent in the
PILS calculations [117].
Because ω is only present in the upper two M M blocks, the determinant of the
response matrix is only a 2M ¼ m(m 1) order polynomial in ω. We therefore find
that the alternative adiabatic approximation does not restore the lost roots in the
standard adiabatic approximation. Calculations on the H2 and HeH+ have demon-
strated that the excitation spectrum does not change much compared to the standard
adiabatic approximation for both the DMLS and PILS functionals [117]. The lowest
166 K. Pernal and K.J.H. Giesbertz
It is unlikely that the missing m diagonal double excitations can be restored with
any decent adiabatic approximation to the TD-RDMFT equations. The problem is
that the excitation energies should come out of the response equations in pairs +ω
and ω. This pairing of the frequencies is dictated by an important symmetry of the
response function χ ðωÞ ¼ χ*ðωÞ, which follows directly from the Lehmann [150]
(sum-over-states) representation. We therefore need to increase the number of roots
by m in some manner, because m is not necessarily even. Increasing the number of
roots to 2(M + m) results in an even number of roots, which in turn ensures that all
excitations are present in both the positive and negative parts of the spectrum.
This partially explains why we had m zero excitations in the standard adiabatic
approximation, because ω ¼ 0 is the only number which is its own negative, so it
does not destroy the χ ðωÞ ¼ χ*ðωÞ symmetry even if an odd number of these roots
is present. This does not explain why we could not have bm/2c excitation energies
occurring both at +ω and ω in an adiabatic approximation. To explain this, we
observe that for a proper quantum evolution a quantity needs to be able to have a
complex phase. All the off-diagonal elements of the 1-RDM are able to obtain a
complex phase-factor, but because the diagonal is necessarily real, the occupation
numbers do not have a quantum phase [117]. This lack of a corresponding quantum
phase for the natural occupation numbers is not limited to the 1-RDM, but exists for
the diagonal of any p-RDM if the BBGKY hierarchy is truncated at the pth order [151].
The way to solve all these problems together is to include an additional set of
m complex phase factors which can act as the conjugate variables for the natural
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 167
where we assume that the action can be considered as a functional of the PINOs and
occupation numbers. The equations of motion for the PINOs and their occupation
numbers follow by making the action stationary with respect to all variations.
However, we have to keep in mind that the action is now only a functional of the
PINOs and occupation numbers and not of the full many-body wavefunction.
Therefore, we cannot set the variation at the end-point t ¼ T to zero and need to
take this term into account explicitly [153], so the variational principle becomes
δA ¼ i ΨðT ÞδΨðT Þ : ð118Þ
A 6 π p ; n p ¼ A0 6 π p ; n p AHxc 6 π p ; n p : ð119Þ
A0 6 π p; n p ¼ dt ^ 0 ðtÞjΦP ðtÞi;
dP ðtÞhΦP ðtÞji∂t H ð120Þ
0 P
where p 2 P means that the sum runs over all PINOs present in the determinant ΦP.
Because the 1-RDM of the ensemble should have the prescribed occupation num-
bers, all the weights of the states which contain a particular 6 π p need to sum to the
corresponding occupation number, np(t). The non-interacting action therefore sim-
plifies even further to
ðT X
A0 6 π p; n p ¼ dt n p ðtÞ 6 π p ðtÞi∂t h^ ðtÞ6 π p ðtÞ ; ð122Þ
0 p
Considering variations in δA0 separately, we find the expected result that the PINOs
are solutions of one-electron Schr€odinger equations i∂t 6 π p ðxtÞ ¼ h^ ðtÞ 6 π p ðxtÞ and
that the occupation numbers (weights) are time-independent. We are not interested
in the solutions of the non-interacting system, however, but we want the solutions of
the interacting system. Therefore, we should add a “bath” term which takes into
account that the electrons do not behave independently but move in the “bath” of
other electrons. Hence, we subtract the following term
δW 6 π p ; n p ¼ δAHxc 6 π p ; n p X
þ i Ψ 6 π p ; n p ðT ÞjδΨ 6 π p ; n p ðT Þ i n p ðT Þ 6 π p ðT Þjδ6 π p ðT Þ
p
ð124Þ
from the left-hand side, to make the variational principle equal to the interacting
one (118)
X
δA0 6 π p ; n p δW 6 π p ; n p ¼ i n p ðT Þ 6 π p ðT Þjδ6 π p ðT Þ : ð125Þ
p
Enforcing the orthonormality of the PINOs with the standard Lagrange multiplier
technique, we can work out the variations in the action produced by perturbations in
the PINOs [117, 145], which recovers the equation of motion for the 1-RDM in the
natural orbital basis (96)
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 169
h D Ei
i n_ p ðtÞ þ nq ðtÞ n p ðtÞ 6 π p ðtÞj6 π_q ðtÞ
{ ð126Þ
¼ nq ðtÞ n p ðtÞ h pq ðtÞ þ W 6 πpq ðtÞ W 6 πpq ðtÞ ;
where the effects of the electron–electron interaction are now expressed as varia-
tions δW produced by perturbations in the PINOs
ð
6π δW
W pq ðtÞ ¼ dx 6 π ðxtÞ: ð127Þ
δ6 π p ðxtÞ q
An equation of motion for the phase factors of the PINOs is obtained by considering
variations produced by perturbations in the occupation numbers, which give [117,
144, 145]
D E δW
i 6 π p ðtÞj6 π_ p ðtÞ ¼ hpp ðtÞ þ : ð128Þ
δn p ðtÞ
One can combine the equation of motion for the PINO phase factors with the
off-diagonal terms of the equation of motion for the 1-RDM (126) to write a
Schr€ o dinger
equation
for the PINOs with an effective potential,
v^ PINO
6 π p ; n p ðtÞ, [117, 144, 145]
i∂t 6 π p ðxtÞ ¼ h^ ðtÞ þ v^ PINO ½f6 πr ; nr gðtÞ 6 π p ðxtÞ; ð129Þ
where v^ PINO 6 π p ; n p ðtÞ is an effective potential which takes the two-body effects
into account and is defined via its matrix elements which can be read off from (126)
and (128)
8 {
>
> W 6 πpq ðtÞ W 6 πpq ðtÞ
>
< for p 6¼ q
nq ðtÞ n p ðtÞ
pq ½f6 π r ; nr gðtÞ ¼
v PINO : ð130Þ
>
> δW
>
: for p ¼ q
δn p ðtÞ
The degeneracy of the natural spinorbitals [103] mentioned in Sect. 4 can therefore
also be regarded as the complete in-phase time evolution of the PINOs. This makes
170 K. Pernal and K.J.H. Giesbertz
complete sense, especially for the two-electron system, because the phase factors of
the PINOs can be used to reconstruct completely the two-electron wavefunction as
[117, 155]
m qffiffiffiffiffiffiffiffiffiffiffi
1 X
Ψðx1 ; x2 ; tÞ ¼ pffiffiffi n p ðtÞ 6 π p ðx1 tÞ6 π p ðx2 tÞ 6 π p ðx1 tÞ6 π p ðx2 tÞ ; ð132Þ
2 p¼1
so the PINOs are coupled in pairs in the two-electron wavefunction. This expression
is valid for an arbitrary spin state. In the case of a singlet state, the spin-up and spin-
down components of the same spatial part form the PINO pairs and, in the case of
triplet states, two PINO which are spatially different are coupled together [117]. It
is clear from this expression that for all np 6¼ 0, all the PINOs need to have the same
time-dependent phase factor in order for the full two-electron wavefunction to be a
stationary state, Ψðx1 ; x2 ; tÞ ¼ eiEt Ψðx1 ; x2 Þ.
The equations of motion can be used again to formulate time-dependent
response equations. Because the zeroth-order time-dependent PINOs already have
a (time-dependent) phase factor, we expand the perturbation in the PINO in the
order of the perturbation as
6 π p ðxtÞ ¼ eiεp t 6 π p ðxÞ þ δ6 π p ðxtÞ þ ð133Þ
The first order of the perturbation in the PINOs, δ6 π p ðxtÞ is expanded in the time-
independent PINO basis as
X
δ6 π p ðxtÞ ¼ 6 πr ðxÞδUrp ðtÞ: ð134Þ
r
The advantage of expressing the first-order perturbation in this manner is that the
connection between and δγ(t) used in the TD-RDMFT response equations at (98) is
still valid. Following the same procedure as before, collecting all perturbations up
to first order and taking the Fourier transform, the frequency-dependent PINO
response equations in the standard adiabatic approximation, W W can be cast
in the following form [117, 144, 145, 156]
0 10 1 0 1
ω1M 0 AþMM AþMm δγ R ðωÞ 0
B 0 ω1m Aþ Aþ C B δnðωÞ C B 0 C
B mM mm C B C¼B C: ð135Þ
@ N1 A N1 N1 C ω1M 0 A @ iδUI ðωÞ A @ δvR ðωÞ A
CT N1 W 0 ω1m iδUD ðωÞ=2 δvD ðωÞ=2
These PINO response equations in the adiabatic approximation have all the desired
properties:
• The ω ! 0 exactly coincides with the linear response equations of static RDMFT
(see Sect. 3.1)
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 171
• The diagonal double excitations are properly accounted for without destroying
the important symmetry of the response function, χ ðωÞ ¼ χ*ðωÞ
• The PILS functional is a proper PINO functional, so the exact PINO functional is
known for the two-electron system
The dimensionality of the response equations can be reduced by half by elim-
inating the imaginary components from the response equations, giving
N1 A N1 N1 C δγ R ðωÞ δvR ðωÞ
ω2 A þ ¼ Aþ ; ð136Þ
CT N1 W δnðωÞ δvD ðωÞ=2
which immediately demonstrates that indeed all the roots occur both at +ω and ω.
The dimensionality of the TD-PINO response equations (m(m + 1)/2) is signifi-
cantly larger than in TDDFT, where only the transitions between the occupied and
unoccupied KS orbitals need to be taken into account (so the dimension would be
only m 1 for two electrons). Though the results from the adiabatic TD-PINO
equations are far superior to those from adiabatic TDDFT, the computational cost is
equivalent to a full CI calculation. However, one would expect that the transitions
between all the low occupied PINOs are not important for the description of low
lying excited states. Test calculations have been performed where only transitions
from the k highest occupied PINOs to all other PINOs are taken into account. No
reduction was made in δn(ω), because its full treatment turned out to be important
for particle number conservation.
Indeed, calculations with low values of k demonstrated that the polarizabilities
[147], excitations [156], and oscillator strengths [157] are in excellent agreement
with the exact results. Taking transitions from only the highest occupied PINO into
account (k ¼ 1) gives reasonable results for the low lying excitations of the hydro-
gen molecule at its equilibrium. To take properly into account the static correlation
effects on the excitation spectrum, one also needs transition from the 1σ u PINO,
because that PINO also obtains a significant occupation when the bond is stretched.
The truncation to k ¼ 2 already gives results very close to the exact ones along the
complete bond-breaking coordinate. Going to k ¼ 3 only provides a small additional
improvement over k ¼ 2.
The same idea has also been tested in the time-domain [155, 158]. The same
effect as in the frequency-domain has been observed: only a small number of the
highest occupied PINOs need explicitly to be taken into account to give a reliable
description of the physical processes. This is particularly interesting for the calcu-
lation of the double ionization yield of He in strong laser fields, which needs an
accurate description of non-sequential double ionization, a highly correlated pro-
cess [159, 160]. An accurate account of the non-sequential double ionization
process has only been given in one dimension by solving the full many-body
Schr€ odinger equation for a one-dimensional He model [161]. A full three-
dimensional treatment is still out of reach, because the grid (number of basis
functions) needs to be very large to describe the electrons moving very far away
from the nucleus and coming back. In a one-dimensional pilot study it has been
172 K. Pernal and K.J.H. Giesbertz
demonstrated that only ten PINOs are sufficient to give a quantitative prediction of
the double ionization yield [162], so an accurate three-dimensional calculation
should now come within reach.
The main challenge is to make the PINO approach applicable to systems with
more than two electrons. For the two-electron system it is clear how to define the
PINO phase factors which gives a very simple expression for the two-body effects,
W, and is even exact. For systems with more electrons, it is not so clear what a
suitable and convenient definition for the PINO phase would be and which prefer-
ably reduces to the exact functional for two electrons. For multi-electron systems
which only have one electron pair constituting a chemical bond (Li2 and LiH for
example), one can try to use a Hartree–Fock (HF) functional for the core electrons
and the PILS functional for the “HONO” which is the highest strongly occupied
natural orbital and some encouraging results have already been obtained for
diatomic molecules with a single chemical bond [163].
A different route is also explored by combining features from the PINO response
equations to the extended RPA equations [164, 165] obtained from Rowe’s equa-
tion of motion framework [166]. The advantage is that the response matrices are
now formulated as partial contractions of the 1-RDM and 2-RDM instead of
functional derivatives with respect to PINOs and occupation numbers. This
makes it easier to use other sources for approximate 2-RDMs such as the APSG
wavefunction or other correlated methods. However, the APSG wavefunction can
also be used to construct a PINO functional (62). The adiabatic PINO response
equations with the APSG functional are actually identical to those obtained by
applying time-dependent response theory to the APSG wavefunction directly
[165]. Calculations on small molecular systems have demonstrated that the lowest
excitation energies for the APSG functional (62) are in very good agreement with
more sophisticated approaches. Higher excitation energies seem to be less reliable.
Experiments using a range-separated version of the APSG functional indicate a
shortcoming of the APSG functional rather than an inherent limitation of the
adiabatic TD-PINO linear response equations [165]. However, more evidence
needs to be gathered before we can make any conclusive statement.
1-RDM, it is evident that retaining solely Coulomb and exchange integrals in the
functional and reconstructing N-representable 2-RDM would lead to a variational
functional deficient in accounting for dynamic electron correlation. However, N-
representability conditions for 2-RDM employed in developing novel functionals
are of necessary but insufficient character. Consequently, functionals resulting from
employing a reconstructional approach are not necessarily variational. As a result,
they can yield lower energy than the best “JK-only” functional if the limited “JK-
only” form is properly compensated by the lack of N-representability of the
underlying 2-RDM.
BBC, ELS, or PNOF functionals are orbital-dependent and, by incorporating a
subtle orbital structure of the exact two-electron functional, they are more appro-
priate for molecular systems than for solids. A functional proposed to work mainly
for extended systems is the power functional (80) with the value of the power α
found empirically. Because of a simple form of the power functional its optimiza-
tion is highly efficient. Taking into account that its form has been proposed rather
ad hoc without imposing any exact conditions, it is remarkable how well it works
for solids. The most spectacular application of the power functional is for Mott
insulators which are properly predicted to be nonmetallic [84]. In general, power
functional cannot compete with the BBC3 or the recent PNOF functionals in
describing the electronic structure of molecular systems.
It should be mentioned that most of the functionals have been proposed in spin-
restricted formulation but extensions to open-shell systems are also available
[167]. So far, RDMFT for high-spin systems has been tested for only a limited
set of systems.
Size-consistency is another property that a useful functional should possess.
Apart from the BB or power functionals, most of the other available approximations
are, in principle, not size-consistent. However, in [168] it has been shown that
violation of size-consistency is negligible for BBC, AC3, and ML functionals.
Undoubtedly there has been significant progress in the last 10 years in the
development of methods in RDMFT. More accurate and versatile functionals
have been proposed. Surprisingly, the “JK-only” form has not yet been fully
exploited and the most recent functionals, ELS [42] and PNOF6 [48], still stay
within this form. As has been discussed, future functionals can either include other
integrals than Coulomb and exchange or stay within the “JK-only” form at the price
of abandoning variationality from the start. Unfortunately, development of
RDMFT-based methods is hindered by slow advances in improving computational
efficiency of optimization algorithms for density matrix functionals. The lack of
sufficiently fast methods has not allowed for application of the existing functionals
to systems consisting of more than a few tens of electrons. Only the very recently
proposed local-RDMFT approach [104] holds any promise of extending limits of
the size of systems that can be treated with RDMFT by at least one order of
magnitude.
The practical use of a time-dependent version of RDMFT has recently been
explored to calculate excitation energies and other frequency-dependent response
properties. A rigorous mathematical foundation for TD-RDMFT is still lacking,
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 175
Acknowledgements The authors wish to acknowledge Dr. Ewa Pastorczak for critically reading
the manuscript. KJHG gratefully acknowledges a VENI grant by the Netherlands Foundation for
Research NWO (722.012.013). KP acknowledges support by the National Science Centre of
Poland under grant DEC-2012/07/E/ST4/03023.
References
1. Gilbert TL (1975) Hohenberg–Kohn theorem for nonlocal external potentials. Phys Rev B 12
(6):2111–2120. doi:10.1103/PhysRevB.12.2111
2. Coleman AJ (1963) Structure of fermion density matrices. Rev Mod Phys 35(3):668–687.
doi:10.1103/RevModPhys.35.668
3. Smith DW (1966) N-representability problem for fermion density matrices. II. The first-order
density matrix with N even. Phys Rev 147(4):896–898. doi:10.1103/PhysRev.147.896
4. Levy M (1979) Universal variational functionals of electron-densities, first-order density-
matrices, and natural spin-orbitals and solution of the v-representability problem. Proc Natl
Acad Sci U S A 76(12):6062–6065. doi:10.1073/pnas.76.12.6062
5. Levy M (1987) Correlation energy functionals of one-matrices and Hartree–Fock densities.
In: Erdahl R, Smith VHJ (eds) Density matrices and density functionals. Reidel, Dordrecht,
pp 479–498
176 K. Pernal and K.J.H. Giesbertz
6. Zumbach G, Maschke K (1985) Density-matrix functional theory for the N-particle ground-
state. J Chem Phys 82(12):5604–5607. doi:10.1063/1.448595
7. Valone SM (1980) Consequences of extending 1 matrix energy functionals from pure-state
representable to all ensemble representable 1 matrices. J Chem Phys 73(3):1344–1349.
doi:10.1063/1.440249
8. Valone SM (1980) A one-to-one mapping between one-particle densities and some normal-
particle ensembles. J Chem Phys 73(9):4653–4655. doi:10.1063/1.440656
9. Nguyen-Dang TT, Lude~na EV, Tal Y (1985) Variation of the energy functional of the
reduced first-order density operator. J Mol Struct (THEOCHEM) 120:247–264.
doi:10.1016/0166-1280(85)85114-9
10. Donnelly RA (1979) Fundamental difference between energy functionals based on 1st-order
and on 2nd-order density matrices. J Chem Phys 71(7):2874–2879. doi:10.1063/1.438678
11. Donnelly RA, Parr RG (1978) Elementary properties of an energy functional of first-order
reduced density matrix. J Chem Phys 69(10):4431–4439. doi:10.1063/1.436433
12. L€owdin PO (1955) Quantum theory of many-particle systems. I. Physical interpretations by
means of density matrices, natural spin-orbitals, and convergence problems in the method of
configurational interaction. Phys Rev 97(6):1474–1489. doi:10.1103/PhysRev.97.1474
13. Hohenberg P, Kohn W (1964) Inhomogeneous electron gas. Phys Rev B 136(3B):B864–
B871. doi:10.1103/PhysRev.136.B864
14. Lieb EH (1981) Variational principle for many-fermion systems. Phys Rev Lett 46
(7):457–459. doi:10.1103/PhysRevLett.46.457
15. Yasuda K (2001) Correlation energy functional in the density-matrix functional theory. Phys
Rev A 63:032517. doi:10.1103/PhysRevA.63.032517
16. Cioslowski J (2005) New constraints upon the electron–electron repulsion energy functional
of the one-electron reduced density matrix. J Chem Phys 123:164106
17. Cioslowski J, Pernal K, Ziesche P (2002) Systematic construction of approximate one-matrix
functionals for the electron–electron repulsion energy. J Chem Phys 117(21):9560–9566.
doi:10.1063/1.1516804
18. Müller AMK (1984) Explicit approximate relation between reduced two-particle and
one-particle density-matrices. Phys Lett A 105(9):446–452. doi:10.1016/0375-9601(84)
91034-X
19. Buijse MA, Baerends EJ (2002) An approximate exchange-correlation hole density as a
functional of the natural orbitals. Mol Phys 100(4):401–421. doi:10.1080/
00268970110070243
20. Buijse MA (1991) Ph.D. thesis, Electron correlation. Fermi and Coulomb holes, dynamical
and nondynamical correlation. Vrije Universiteit, Amsterdam
21. Frank RL, Lieb EH, Seiringer R, Siedentop H (2007) Muller’s exchange-correlation energy in
density-matrix-functional theory. Phys Rev A 76(5):052517. doi:10.1103/PhysRevA.76.
052517
22. Staroverov VN, Scuseria GE (2002) Assessment of simple exchange-correlation energy
functionals of the one-particle density matrix. J Chem Phys 117(6):2489–2495.
doi:10.1063/1.1491395
23. Herbert JM, Harriman JE (2003) Self-interaction in natural orbital functional theory. Chem
Phys Lett 382(1–2):142–149. doi:10.1016/j.cplett.2003.10.057
24. Herbert JM, Harriman JE (2003) N-representability and variational stability in natural orbital
functional theory. J Chem Phys 118(24):10835–10846. doi:10.1063/1.1574787
25. Cohen AJ, Baerends EJ (2002) Variational density matrix functional calculations for the
corrected Hartree and corrected Hartree–Fock functionals. Chem Phys Lett 364
(3–4):409–419. doi:10.1016/S0009-2614(02)01345-3
26. Cioslowski J, Pernal K (2001) Response properties and stability conditions in density matrix
functional theory. J Chem Phys 115(13):5784–5790. doi:10.1063/1.1383292
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 177
49. Piris M, Ugalde JM (2014) Perspective on natural orbital functional theory. Int J Quant Chem
114(18, SI):1169–1175. doi:10.1002/qua.24663
50. Piris M (2013) A natural orbital functional based on an explicit approach of the two-electron
cumulant. Int J Quant Chem 113(5, SI):620–630. doi:10.1002/qua.24020
51. Piris M, Matxain JM, Lopez X, Ugalde JM (2009) Spin conserving natural orbital functional
theory. J Chem Phys 131(2):021102. doi:10.1063/1.3180958
52. Leiva P, Piris M (2006) Calculation of vertical ionization potentials with the Piris natural
orbital functional. J Mol Struct (THEOCHEM) 770(1–3):45–49. doi:10.1016/j.theochem.
2006.05.001
53. Lopez X, Piris M, Matxain JM, Ugalde JM (2010) Performance of PNOF3 for reactivity
studies: X[BO] and X[CN] isomerization reactions (X ¼ H, Li) as a case study. Phys Chem
Chem Phys 12(40):12931–12934. doi:10.1039/c003379k
54. Lopez X, Ruiperez F, Piris M, Matxain JM, Matito E, Ugalde JM (2012) Performance of
PNOF5 natural orbital functional for radical formation reactions: hydrogen atom abstraction
and C-C and O-O homolytic bond cleavage in selected molecules. J Chem Theory Comput 8
(8):2646–2652. doi:10.1021/ct300414t
55. Ruiperez F, Piris M, Ugalde JM, Matxain JM (2013) The natural orbital functional theory of
the bonding in Cr2, Mo2 and W2. Phys Chem Chem Phys 15(6):2055–2062. doi:10.1039/
c2cp43559d
56. Matxain JM, Piris M, Ruiperez F, Lopez X, Ugalde JM (2011) Homolytic molecular
dissociation in natural orbital functional theory. Phys Chem Chem Phys 13
(45):20129–20135. doi:10.1039/c1cp21696a
57. Piris M (2013) Interpair electron correlation by second-order perturbative corrections to
PNOF5. J Chem Phys 139(6):064111. doi:10.1063/1.4817946
58. Piris M, Ruiperez F, Matxain JM (2014) Assessment of the second-order perturbative
corrections to PNOF5. Mol Phys 112(5–6, SI):1–8. doi:10.1080/00268976.2013.854933
59. Szabados Á, Rolik Z, Toth G, Surján PR (2005) Multiconfiguration perturbation theory: size
consistency at second order. J Chem Phys 122(11):114104. doi:10.1063/1.1862235
60. Pernal K (2013) The equivalence of the Piris Natural Orbital Functional 5 (PNOF5) and the
antisymmetrized product of strongly orthogonal geminal theory. Comput Theor Chem 1003
(SI):127–129. doi:10.1016/j.comptc.2012.08.022
61. Surján PR (1999) An introduction to the theory of geminals. In: Surján PR (ed) Correlation
and localization, vol 203, Topics in current chemistry. Springer, Berlin/Heidelberg, pp 63–88
62. Rassolov V (2002) A geminal model chemistry. J Chem Phys 117(13):5978–5987.
doi:10.1063/1.1503773
63. Rassolov VA, Xu F (2007) Geminal model chemistry. IV. Variational and size consistent
pure spin states. J Chem Phys 127(4):044104. doi:10.1063/1.2755738
64. Hurley AC, Lennard-Jones J, Pople JA (1953) The molecular orbital theory of chemical
valency. A theory of paired-electrons in polyatomic molecules. Proc R Soc Lond A Math
Phys Sci 220(1143):446–455. doi:10.1098/rspa.1953.0198
65. Kutzelnigg W (1964) Direct determination of natural orbitals and natural expansion coeffi-
cients of many-electron wavefunctions. I. Natural orbitals in geminal product approximation.
J Chem Phys 40(12):3640–3647. doi:10.1063/1.1725065
66. Arai T (1960) Theorem on separability of electron pairs. J Chem Phys 33(1):95–98.
doi:10.1063/1.1731142
67. Mehler EL, Reudenberg K, Silver DM (1970) Electron correlation and separated pair
approximation in diatomic molecules. II. Lithium hydride and boron hydride. J Chem Phys
52(3):1181–1205. doi:10.1063/1.1673115
68. Matxain JM, Piris M, Uranga J, Lopez X, Merino G, Ugalde JM (2012) The nature of
chemical bonds from PNOF5 calculations. Chemphyschem 13(9):2297–2303. doi:10.1002/
cphc.201200205
69. Piris M, Matxain JM, Lopez X (2013) The intrapair electron correlation in natural orbital
functional theory. J Chem Phys 139(23):234109. doi:10.1063/1.4844075
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 179
70. Cioslowski J, Pernal K (2002) Density matrix functional theory of weak intermolecular
interactions. J Chem Phys 116(12):4802–4807. doi:10.1063/1.1446028
71. Cioslowski J, Ziesche P, Pernal K (2001) Description of a high-density homogeneous
electron gas with the Yasuda density matrix functional. J Chem Phys 115(19):8725–8730.
doi:10.1063/1.1412604
72. Cioslowski J, Pernal K (2002) Variational density matrix functional theory calculations with
the lowest-order Yasuda functional. J Chem Phys 117(1):67–71. doi:10.1063/1.1481384
73. Cioslowski J, Buchowiecki M, Ziesche P (2003) Density matrix functional theory of four-
electron systems. J Chem Phys 119(22):11570–11573. doi:10.1063/1.1623741
74. Kollmar C, Hess BA (2003) A new approach to density matrix functional theory. J Chem
Phys 119(9):4655–4661. doi:10.1063/1.1590635
75. Kollmar C, Hess BA (2004) The structure of the second-order reduced density matrix in
density matrix functional theory and its construction from formal criteria. J Chem Phys 120
(7):3158–3171. doi:10.1063/1.1634554
76. Kollmar C (2004) The “JK-only” approximation in density matrix functional and wave
function theory. J Chem Phys 121(23):11581–11586. doi:10.1063/1.1819319
77. Pernal K (2010) Long-range density-matrix-functional theory: application to a modified
homogeneous electron gas. Phys Rev A 81(5):052511. doi:10.1103/PhysRevA.81.052511
78. Rohr DR, Toulouse J, Pernal K (2010) Combining density-functional theory and density-
matrix-functional theory. Phys Rev A 82(5):052502. doi:10.1103/PhysRevA.82.052502
79. Stoll H, Savin A (1985) Density functionals for correlation energies of atoms and molecules.
In: Dreizler R, da Providencia J (eds) Density functional methods in physics. Plenum,
New York, pp 177–207
80. Toulouse J, Colonna F, Savin A (2004) Long-range-short-range separation of the electron–
electron interaction in density-functional theory. Phys Rev A 70(6):062505. doi:10.1103/
PhysRevA.70.062505
81. Goll E, Werner HJ, Stoll H (2005) A short-range gradient-corrected density functional in
long-range coupled-cluster calculations for rare gas dimers. Phys Chem Chem Phys 7
(23):3917–3923. doi:10.1039/b509242f
82. Lathiotakis NN, Helbig N, Gross EKU (2007) Performance of one-body reduced density-
matrix functionals for the homogeneous electron gas. Phys Rev B 75(19):195120.
doi:10.1103/PhysRevB.75.195120
83. Cioslowski J, Pernal K (2000) Description of a homogeneous electron gas with simple
functionals of the one-particle density matrix. Phys Rev A 61(3):034503
84. Sharma S, Dewhurst JK, Lathiotakis NN, Gross EKU (2008) Reduced density matrix
functional for many-electron systems. Phys Rev B 78(20):201103. doi:10.1103/PhysRevB.
78.201103
85. Lathiotakis NN, Helbig N, Zacarias A, Gross EKU (2009) A functional of the one-body-
reduced density matrix derived from the homogeneous electron gas: performance for finite
systems. J Chem Phys 130(6):064109. doi:10.1063/1.3073053
86. Lathiotakis NN, Sharma S, Dewhurst JK, Eich FG, Marques MAL, Gross EKU (2009)
Density-matrix-power functional: performance for finite systems and the homogeneous
electron gas. Phys Rev A 79(4):040501. doi:10.1103/PhysRevA.79.040501
87. Marques MAL, Lathiotakis NN (2008) Empirical functionals for reduced-density-matrix-
functional theory. Phys Rev A 77(3):032509. doi:10.1103/PhysRevA.77.032509
88. Cioslowski J, Ziesche P, Pernal K (2001) On the exactness of simple natural spin-orbital
functionals for a high-density homogeneous electron gas. Phys Rev B 63(20):205105
89. Cioslowski J, Pernal K (1999) Constraints upon natural spin orbital functionals imposed by
properties of a homogeneous electron gas. J Chem Phys 111(8):3396–3400. doi:10.1063/1.
479623
90. Sharma S, Dewhurst JK, Shallcross S, Gross EKU (2013) Spectral density and metal-
insulator phase transition in Mott insulators within reduced density matrix functional theory.
Phys Rev Lett 110(11):116403. doi:10.1103/PhysRevLett.110.116403
180 K. Pernal and K.J.H. Giesbertz
110. Giesbertz KJH, Baerends EJ (2010) Aufbau derived from a unified treatment of occupation
numbers in Hartree-Fock, Kohn-Sham, and natural orbital theories with the Karush-Kuhn-
Tucker conditions for the inequality constraints ni 1 and ni 0. J Comput Chem 132
(19):194108. doi:10.1063/1.3426319
111. Piris M, Ugalde JM (2009) Iterative diagonalization for orbital optimization in natural orbital
functional theory. J Comput Chem 30(13):2078–2086. doi:10.1002/jcc.21225
112. Requist R, Pankratov O (2008) Generalized Kohn-Sham system in one-matrix functional
theory. Phys Rev B 77(23):235121. doi:10.1103/PhysRevB.77.235121
113. Baldsiefen T, Gross EKU (2013) Minimization procedure in reduced density matrix func-
tional theory by means of an effective noninteracting system. Comput Theor Chem 1003
(SI):114–122. doi:10.1016/j.comptc.2012.09.001
114. Talman JD, Shadwick WF (1976) Optimized effective atomic central potential. Phys Rev A
14(1):36–40. doi:10.1103/PhysRevA.14.36
115. Kümmel S, Kronik L (2008) Orbital-dependent density functionals: theory and applications.
Rev Mod Phys 80(1):3–60. doi:10.1103/RevModPhys.80.3
116. Gidopoulos NI, Lathiotakis NN (2012) Nonanalyticity of the optimized effective potential
with finite basis sets. Phys Rev A 85(5):046502. doi:10.1103/PhysRevA.85.052508
117. Giesbertz KJH (2010) Time-dependent one-body reduced density matrix functional theory;
adiabatic approximations and beyond. Ph.D. thesis, Vrije Universiteit, Amsterdam
118. Appel H (2007) Time-dependent quantum many-body systems: linear response, electronic
transport and reduced density matrices. Ph.D. thesis, Freie Universität, Berlin
119. Pernal K, Gritsenko O, Baerends EJ (2007) Time-dependent density-matrix-functional the-
ory. Phys Rev A 75(1):012506. doi:10.1103/PhysRevA.75.012506
120. Pernal K, Cioslowski J (2007) Frequency-dependent response properties and excitation
energies from one-electron density matrix functionals. Phys Chem Chem Phys 9(45):5956.
doi:10.1039/b704797e
121. Pernal K, Giesbertz K, Gritsenko O, Baerends EJ (2007) Adiabatic approximation of time-
dependent density matrix functional response theory. J Chem Phys 127:214101. doi:10.1063/
1.2800016
122. Tozer DJ, Amos RD, Handy NC, Roor BO, Serrano-Andrés L (1999) Does density functional
theory contribute to the understanding of excited states of unsaturated organic compounds?
Mol Phys 97(7):859–868
123. Dreuw A, Weisman JL, Head-Gordon M (2003) Long-range charge-transfer excited states in
time-dependent density functional theory require non-local exchange. J Chem Phys 119
(6):2943. doi:10.1063/1.1590951
124. Yanai T, Tew DP, Handy NC (2004) A new hybrid exchange-correlation functional using the
Coulomb-attenuating method (CAM-B3LYP). Chem Phys Lett 393(1–3):51
125. Gritsenko O, Baerends EJ (2004) Asymptotic correction of the exchange-correlation kernel
of time-dependent density functional theory for long-range charge-transfer excitations.
J Chem Phys 121(2):655
126. Neugebauer J, Gritsenko O, Baerends EJ (2006) Assessment of a simple correction for the
long-range charge-transfer problem in time-dependent density-functional theory. J Chem
Phys 124(21):214102
127. Gritsenko O, van Gisbergen SJA, G€orling A, Baerends EJ (2000) Excitation energies of
dissociating H2: a problematic case for the adiabatic approximation of time-dependent
density functional theory. J Chem Phys 113(19):8478
128. Giesbertz KJH, Baerends EJ (2008) Failure of time-dependent density functional theory for
excited state surfaces in case of homolytic bond dissociation. Chem Phys Lett 461:338.
doi:10.1016/j.cplett.2008.07.018
129. Maitra NT, Zhang F, Cave RJ, Burke K (2004) Double excitations within time-dependent
density functional theory linear response. J Chem Phys 120(13):5932
182 K. Pernal and K.J.H. Giesbertz
130. Neugebauer J, Baerends E (2004) Vibronic coupling and double excitations in linear response
time-dependent density functional calculations: dipole allowed states of N2. J Chem Phys 121
(13):6155. doi:10.1063/1.1785775
131. Yvon J (1935) Theorie statistique des fluides et l’equation et l’equation d’etat (French). In:
Actes Sientifique et Industrie, vol. 203. Hermann, Paris
132. Bogoliubov NN (1946) Kinetic equations. J Phys USSR 10(3):265 (in English)
133. Bogoliubov NN (1946) Kinetic equations. J Exp Theor Phys 16(8):691 (in Russian)
134. Kirkwood JG (1946) The statistical mechanical theory of transport processes I. General
theory. J Chem Phys 14(3):180
135. Born M, Green HS (1946) A general kinetic theory of liquids. I. The molecular distribution
functions. Proc R Soc Med 188(1012):10–18
136. Kirkwood JG (1947) The statistical mechanical theory of transport processes II. Transport in
gases. J Chem Phys 15(1):72
137. Bogoliubov NN, Gurov KP (1947) Kinetic equations in quantum mechanics (rus.). J Exp
Theor Phys 17(7):614
138. Runge E, Gross EKU (1984) Density-functional theory for time-dependent systems. Phys
Rev Lett 52(12):997. doi:10.1103/PhysRevLett.52.997
139. Ruggenthaler M, van Leeuwen R (2011) Global fixed-point proof of time-dependent density-
functional theory. Europhys Lett 95(1):13001. doi:10.1209/0295-5075/95/13001
140. Ruggenthaler M, Giesbertz KJH, Penz M, van Leeuwen R (2012) Density-potential mappings
in quantum dynamics. Phys Rev A 85(5):052504. doi:10.1103/PhysRevA.85.052504
141. Giesbertz KJH, Baerends EJ, Gritsenko OV (2008) Charge transfer, double and bond-
breaking excitations with time-dependent density matrix functional theory. Phys Rev Lett
101:033004. doi:10.1103/PhysRevLett.101.033004
142. Giesbertz KJH, Pernal K, Gritsenko OV, Baerends EJ (2009) Excitation energies with time-
dependent density matrix functional theory: singlet two-electron systems. J Chem Phys 130
(11):114104. doi:10.1063/1.3079821
143. Peuckert V (1978) A new approximation method for electron systems. J Phys C 11(24):4945
144. Giesbertz KJH, Gritsenko OV, Baerends EJ (2010) Response calculations with an indepen-
dent particle system with an exact one-particle density matrix. Phys Rev Lett 105(1):013002.
doi:10.1103/PhysRevLett.105.013002
145. Giesbertz KJH, Gritsenko OV, Baerends EJ (2010) The adiabatic approximation in time-
dependent density matrix functional theory: response properties from dynamics of phase-
including natural orbitals. J Chem Phys 133(17):174119. doi:10.1063/1.3499601
146. Requist R, Pankratov O (2011) Time-dependent occupation numbers in reduced-density-
matrix-functional theory: application to an interacting Landau–Zener model. Phys Rev A 83
(5):052510. doi:10.1103/PhysRevA.83.052510
147. Giesbertz KJH, Gritsenko OV, Baerends EJ (2014) Response calculations based on an
independent particle system with the exact one-particle density matrix: polarizabilities.
J Chem Phys 140(18):18A517. doi:10.1063/1.4867000
148. Giesbertz KJH, Gritsenko OV, Baerends EJ (2012) Time-dependent reduced density matrix
functional theory. In: Marques MAL, Maitra N, Nogueira F, Gross EKU, Rubio A (eds)
Fundamentals of time-dependent density functional theory, vol 837, Lecture notes in physics.
Springer, Berlin, pp 485–498. doi:10.1007/978-3-642-23518-4
149. Requist R, Pankratov O (2010) Adiabatic approximation in time-dependent reduced-density-
matrix functional theory. Phys Rev A 81(4):042519. doi:10.1103/PhysRevA.81.042519
150. Lehmann H (1954) Über Eigenschaften von Ausbreitungsfunktionen und Renormierungs-
konstanten quantisierter Felder. Nuovo Cimento 11(4):342
151. Requist R (2012) Hamiltonian formulation of nonequilibrium quantum dynamics: geometric
structure of the Bogoliubov–Born–Green–Kirkwood–Yvon hierarchy. Phys Rev A 86
(2):022117. doi:10.1103/PhysRevA.86.02211
Reduced Density Matrix Functional Theory (RDMFT) and Linear Response Time. . . 183
Abstract This chapter gives an overview of the description of the optical and
dielectric properties of bulk insulators and semiconductors in time-dependent den-
sity-functional theory (TDDFT), with an emphasis on excitons. We review the linear-
response formalism for periodic solids, discuss excitonic exchange-correlation ker-
nels, calculate exciton binding energies for various materials, and compare the
treatment of excitons with TDDFT and with the Bethe–Salpeter equation.
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
2 What Is an Exciton? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
3 A Tale of Three Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
4 Linear Response and Optical Properties in Periodic Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
4.1 Microscopic and Macroscopic Dielectric Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
4.2 Linear-Response Theory and TDDFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
5 TDDFT for Excitons in Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
5.1 Why Are Excitons a Difficult Problem? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
5.2 Formalism: Direct Calculation of Exciton Binding Energies . . . . . . . . . . . . . . . . . . . . . . . 197
5.3 Why Does ALDA Fail? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
5.4 Excitonic xc Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
1 Introduction
2 What Is an Exciton?
The optical properties of materials are determined by the way in which the electrons
and the ions respond to light. In this chapter we focus exclusively on the electronic
response and ignore the lattice dynamics or any effects related to the coupling of
electronic and lattice excitations (such as polarons; for details see, e.g., Yu and
Cardona [20]).
The response of a system of N electrons is often characterized as having either
“single-particle” or “collective” character. What do we mean by this? If a system
consists of noninteracting particles, the response is always purely single-particle, or
can be viewed as the sum of many individual, uncorrelated single-particle excita-
tions. In the presence of interactions this simple picture is no longer valid, because
any change of state of one electron has an immediate influence on all other electrons
Excitons in Time-Dependent Density-Functional Theory 187
in the system; the response is collective and involves, in principle, all electrons. The
question is how dominant these effects are.
There are many situations where the interactions give rise to rather straightfor-
ward behavior. Imagine an experiment where we can turn off the interactions
between the electrons in a molecule, and measure the resulting excitation spectrum.
If we now gradually turn the interactions back on, all the while keeping an eye on
the excitation spectrum, we find that the peaks in the spectrum shift, but we can
keep track of each of them and hence can in principle interpret them as single
particle excitations of an effective system such as Kohn–Sham.
On the other hand, interacting systems have certain excitations without a coun-
terpart in any corresponding noninteracting system. A most drastic example is
plasmons in a metal, where all electrons respond collectively and with a fixed
phase relationship. A plasmon requires dynamical electron–electron interactions at
least at the level of the random-phase approximation (RPA).
An exciton is another example of a collective excitation, occurring in non-
metallic systems. The ideal exciton can be described [21] as an electrically neutral
quantum of electronic excitation energy travelling in the periodic structure of a
crystal. It can be viewed as a bound electron–hole pair and can hence be associated
with the transportation of energy, but not of net charge. Excitons are a crucial stage
in the photovoltaic process, where free carriers are generated after separation of the
electron–hole pairs.
Excitons come in different types [22]:
• Frenkel excitons [23, 24] are excitations localized at the atomic sites of wide-gap
insulators such as solid rare gases (neon, argon) or certain ionic solids (e.g., LiF).
• Davydov excitons [25] are found in molecular crystals with ring units, such as
benzene and anthracene. Because the excitations remain localized on the indi-
vidual molecules, Davydov excitons can be viewed as a subclass of Frenkel
excitons.
• Mott–Wannier excitons [26, 27] typically occur in semiconductors such as
GaAs, CdSe, or Cu2O. They tend to be delocalized over several atomic unit cells.
The concept of excitons was originally introduced in bulk crystals, but they also
exist in many lower-dimensional systems such as surfaces, quantum wells, quantum
wires, nanotubes, polymers, nanocrystals, and quantum dots [28–33]. In this chap-
ter we limit ourselves to three-dimensional periodic crystals.
Excitons are usually described as bound electron–hole pairs, i.e., as an effective
two-particle system. Within the effective-mass approximation [34, 35], where con-
duction band electrons have effective mass m
e and valence band holes have effective
mass m
h , we can define a reduced effective mass mr ¼ m*e m*h = m*e þ m*h . Next, we
separate center-of-mass and relative degrees of freedom. The former describes how
the exciton travels through the crystal, and the latter determines the exciton binding
energy according to the following hydrogen-like Schr€odinger equation:
ℏ 2 ∇2 e2
ψ j ðrÞ ¼ E j ψ j ðrÞ: ð1Þ
2mr 4πε0 εr
188 C.A. Ullrich and Z.-h. Yang
a b
Fig. 1 Optical transitions in a two-band insulator, where the lower (valence) band is filled and the
upper (conduction) band is empty. The vertical direction is energy, the horizontal direction is
wavevector. (a) A single-particle transition, in which only one single-particle state gets excited
and all other states do not participate. The associated excitation energy is the difference of the
initial and final single-particle states. (b) Excitonic transition, which is a collective excitation in
which many states participate. The thickness of the arrows indicates that transitions close to the
band gap are dominant. The associated excitation energy can be lower than the band gap, because
the collective nature of the response is energetically favorable
Here, ε0 is the vacuum permittivity, e is the free electron charge, and ε is the
dielectric constant of the material (we talk about the dielectric constant in greater
detail in the following section). Equation (1) is also known as the Wannier equation.
It yields a Rydberg series of bound states as well as a continuum of unbound states
[36]. The lowest (1s) excitonic state determines the exciton binding energy Eex 0 and
the exciton Bohr radius a
0 . In GaAs, a material in which the Wannier equation
works particularly well, one obtains E0ex ¼ 4:6 meV and a*0 ¼ 118 Å, which clearly
shows that Wannier excitons are weakly bound and extend over many lattice
constants which, a posteriori, justifies the simplified treatment via (1).
The Wannier picture of excitons as bound electron–hole pairs, described by (1),
is generally not quantitatively accurate, and breaks down completely if the exciton
radius becomes comparable to a lattice constant. In this chapter we present an ab
initio approach, based on TDDFT and/or other many-body techniques, which is
universally valid and in principle exact. This approach reveals an alternative point
of view, in which excitons are described as collective excitations of the many-
electron system. This picture is schematically illustrated in Fig. 1.
The left panel of Fig. 1 shows a single-particle transition in a simple model of an
insulator, going vertically from the filled valence band to the empty conduction
band. The energy associated with this transition is just the difference between the
levels in the two bands. By contrast, an exciton arises from a superposition of many
single-particle transitions, as illustrated in the right panel of Fig. 1. Not all transi-
tions contribute with equal weight, as indicated by the different thicknesses of the
arrows, but all of them have a fixed phase relationship; hence, the exciton is a
collective excitation. The energy of the exciton (i.e., the energy of this collective
excitation) is lower than the lowest single-particle transition. This happens because
the collective behavior induced by the dynamical many-body effects is energeti-
cally favorable compared to any single-particle transition. We see later how the two
viewpoints of the nature of an exciton can be reconciled with each other [18].
Excitons in Time-Dependent Density-Functional Theory 189
Eg ðN Þ ¼ I ðN Þ AðN Þ; ð2Þ
where I(N ) and A(N ) are the ionization potential and the electron affinity of the
system, respectively. These two quantities can be obtained in a straightforward
manner from ground-state DFT: the ionization potential is formally exactly given
by the highest occupied Kohn–Sham eigenvalue of the N-electron system, εN (N ),
and the electron affinity is the corresponding quantity of the N + 1-electron system.
Hence, we obtain
Eg ðN Þ ¼ εNþ1 ðN þ 1Þ εN ðN Þ: ð3Þ
It is important to note that the right-hand side of (3) contains the highest occupied
Kohn–Sham eigenvalues of two different systems, namely with N and with N + 1
electrons. In a macroscopic solid with 1023 electrons, it would be impossible (or at
least highly impractical) to calculate the band gap according to this definition.
The band gap in the noninteracting Kohn–Sham system, also known as the
Kohn–Sham gap, is defined as
In contrast with the interacting gap Eg, the Kohn–Sham gap Eg,s is simply the
difference between the highest occupied and lowest unoccupied single-particle
levels in the same N-particle system. This quantity is what is usually taken as the
band gap in standard DFT band-structure calculations. We can relate the two gaps
by
E0ex
xc
N 1
Energy
Fig. 2 Schematic illustration of the different types of gaps in DFT and TDDFT. The Kohn–Sham
gap is defined as the difference of the highest occupied and lowest unoccupied Kohn–Sham
eigenvalues of the N-electron system; see (4). The fundamental band gap [or quasiparticle
(QP) gap] is the Kohn–Sham gap plus the derivative discontinuity; see (5). The optical gap is
the band gap minus the lowest exciton binding energy Eex 0 . The Kohn–Sham gap can be viewed as
an approximation for the optical gap
The so-called band-gap problem of DFT reflects the fact that in practice Eg,s is
often a poor approximation to Eg, typically underestimating the exact band gap by
as much as 50%. The reason for this is twofold: commonly used approximate xc
functionals (such as LDA and GGA) tend to underestimate the exact Kohn–Sham
gap Eg,s, and they do not yield any discontinuity correction Δxc. An extreme
example for the second failure is Mott insulators, which are typically predicted to
be metallic by DFT. This is no accident: in Mott insulators, the exact Kohn–Sham
system is metallic (i.e., Eg,s ¼ 0) so that Eg ¼ Δxc. Clearly, standard xc functionals
(where Δxc vanishes) are unfit to describe Mott insulators.
It is important to distinguish between the fundamental band gap and the optical
gap [43]. The band gap describes the energy which an electron must have so that,
when it is added to an N-electron system, the result is an N + 1 electron system in its
ground state. The total charge of the system changes by 1 in this process. By
contrast, the optical gap describes the lowest neutral excitation of an N-electron
system: here, the number of electrons remains unchanged. The two gaps are
schematically illustrated in Fig. 2 together with the Kohn–Sham gap.
The band gap of insulators can be accurately obtained from the so-called
quasiparticle energies, which are defined as the single-particle energies of a
noninteracting system whose one-particle Green’s function is the same as that of
the real interacting system (it should be noted that this effective noninteracting
system is very different from the Kohn–Sham system, which is defined as that
noninteracting system which reproduces the exact density). In practice, quasiparti-
cle calculations are often done using the GW method [4, 5, 14]. GW calculations are
more demanding than DFT, but they produce band structures of solids which agree
very well with experiment.
Excitons in Time-Dependent Density-Functional Theory 191
In recent years, generalized Kohn–Sham schemes [37, 44, 45] have become
quite popular for calculating band gaps of solids. Generalized Kohn–Sham theory
means, in essence, using exchange-correlation (xc) functionals which contain an
admixture of Hartree–Fock (HF) nonlocal exchange. These functionals (for
instance B3LYP [46] or PBE0 [47]) have been crucial for the enormous success
of DFT in theoretical chemistry. Hybrid xc functionals for solids are not
unproblematic: HF exchange is difficult to implement in periodic systems, and
B3LYP fails for metals [48]. However, for insulators, hybrid functionals generally
produce excellent results, comparable to what can be achieved by more sophisti-
cated many-body calculations [49–59]. Another promising approach is meta-GGA
functionals such as the Tran–Blaha exchange potential [60–64], which gives good
band structures and band gaps and is much less costly than hybrid functionals.
However, it has the drawback of not being derivable from an energy functional.
It should be emphasized that current implementations of hybrid xc functionals
do not actually calculate the Kohn–Sham gap (this would require a self-consistent
calculation with a local potential); instead, hybrid functionals yield an approxima-
tion to the quasiparticle gap. Local xc potentials can, in principle, be constructed
from hybrid functionals using the OEP (optimized effective potential) method,
which was successfully done for the case of exact exchange (see Betzinger
et al. [65] and references therein).
While the band gap can be measured using techniques in which electrons are
added or removed from the system (such as photoemission spectroscopy), the
optical gap refers to the lowest neutral excitation. The difference between quasi-
particle band gap and optical gap is the lowest exciton binding energy, Eex 0 . In the
previous section we have seen that excitons can be viewed as bound electron–hole
pairs, whose bound states form a Rydberg series, analogous to the hydrogen atom.
The band gap is given by the asymptotic limit of the excitonic Rydberg series [66]
(at least for direct-gap insulators and semiconductors).
∇ D ¼ n f; ð6Þ
∂B
∇E¼ ; ð7Þ
∂t
192 C.A. Ullrich and Z.-h. Yang
∇ B ¼ 0; ð8Þ
∂D
∇ H ¼ jf þ ; ð9Þ
∂t
where all fields (D, E, B, and H) and all source terms (the density of free charges nf
and the free current density jf) are functions of position r and time t. We consider
situations where all time dependence is periodic, and we Fourier transform from
time t to frequency ω. Of main interest to us (because we are concerned with
nonmagnetic materials) is the relationship between the electric displacement D and
the total electric field E:
ð
Dðr; ωÞ ¼ d3 r 0 ¼ε ðr, r0 , ωÞEðr0 , ωÞ; ð10Þ
where V is the crystal volume, k is a wave vector in the first Brillouin zone (BZ),
and G and G0 are reciprocal lattice vectors. In the following we use the notation
ε
¼GG0
ðk; ωÞ ¼ ¼ε ðk þ G, k þ G0 , ωÞ: ð12Þ
where εGG0 ðk; ωÞ, the longitudinal component of the dielectric tensor for the cubic
system, is often called the dielectric matrix. Calculating the macroscopic dielectric
tensor without imposing cubic symmetry is technically more involved [70].
In the optical spectroscopy of materials, a central quantity is the complex
refractive index n~, defined as [20]
The real and imaginary parts of n~ determine two key optical properties of materials:
the refractive index n and the extinction coefficient κ, where
ℜεmac ¼ n2 þ κ2 ; ð17Þ
Jεmac ¼ 2nκ: ð18Þ
We now make a connection between the dielectric function and the linear-response
formalism. The linear density response n1(r, ω) caused by a frequency-dependent
scalar perturbation v1 (r, ω) is given by
ð
n1 ðr; ωÞ ¼ d 3 r 0 χðr, r0 , ωÞv1 ðr0 , ωÞ; ð19Þ
ð ð
3 0 0 n1 ðr00 , ωÞ
0 3 00
v1 ðr; ωÞ ¼ d r εðr, r , ωÞ v1 ðr , ωÞ þ d r : ð20Þ
jr0 r00 j
ε1
GG0 ðk; ωÞ ¼ δGG0 þ vG ðkÞχ GG0 ðk; ωÞ; ð22Þ
4π
vG ðkÞ ¼ : ð23Þ
k þ G 2
Thus, the inverse dielectric function follows directly from the response function.
In linear-response TDDFT [72], the interacting response function χ can be
expressed in terms of the response function of the Kohn–Sham system χ s and the
Hartree and xc kernels:
χ ðr, r0 , ωÞ ¼ χ s ðr, r0 , ωÞ
ð ð
1
þ d3 x d 3 x0 χ s ðr; x; ωÞ þ f ð x, x 0
, ω Þ χ ðx0 , r0 , ωÞ: ð24Þ
j x x0 j xc
X
1 φ j ðrÞφ*k ðrÞφ*j ðr0 Þφk ðr0 Þ
χ s ðr, r0 , ωÞ ¼ ð f k f jÞ ; ð25Þ
j, k¼1
ω ω jk þ iη
where fj and fk are occupation numbers referring to the configuration of the Kohn–
Sham ground state (1 for occupied and 0 for empty Kohn–Sham orbitals), φj (r) are
the Kohn–Sham orbitals, and the ωjk are defined as the differences of the Kohn–
Sham eigenvalues,
ω jk ¼ ε j εk : ð26Þ
0 0 δvxc ½nðr; tÞ
f xc ðr, t, r , t Þ ¼ : ð27Þ
δn ðr0 ; t0 Þ n0 ðrÞ
ð28Þ
featuring the Kohn–Sham band structure εjk and Bloch functions φjk(r). The
so-called head (G ¼ G0 ¼ 0) of the xc kernel f xcGG0 ðk; ωÞ gives the largest contri-
bution to the change from χ s to χ; the contributions from bigger Gs decay rapidly.
Thus the sums in (28) can usually be restricted to a small number of reciprocal
lattice vectors, which reduces the computational effort significantly.
Let us now come back to the macroscopic dielectric constant. It can be shown
[1, 14, 15] that εmac(ω) takes on the following form:
Here, χ GG0 ðk; ωÞ differs from the full response function χ GG0 ðk; ωÞ, as defined in
(28), in the following way: instead of using the full Coulomb interaction vG(k) [see
(23)], it uses the modified Coulomb interaction
8
< 0 for G ¼ 0,
vG ðkÞ ¼ 4π ð31Þ
: k þ G2 for G 6¼ 0;
in which the long-range part v0(k) ¼ 4π/k2 has been left out. This seemingly small
modification turns out to be quite important.
196 C.A. Ullrich and Z.-h. Yang
Until a few years ago, common wisdom held that TDDFT may be good for
molecular excitations but fails for excitons in solids. We now know better, as we
see in this section. However, let us first discuss why excitons are such a hard
problem for TDDFT.
Figure 3 shows the imaginary part of the macroscopic dielectric function of bulk
silicon [73], comparing experimental data with calculations using the RPA (where
fxc ¼ 0) and the adiabatic local-density approximation (ALDA), which is the sim-
plest and most commonly used approximation of TDDFT. There are drastic devi-
ations between theory and experiment. First, the onset of absorption is red-shifted in
RPA and ALDA by about half an eV; this is not very surprising, and reflects the
“band gap problem” of ground-state DFT: as discussed in Sect. 3, the Kohn–Sham
gap of standard local and semilocal xc functionals is smaller than the quasiparticle
gap. One can correct for this error and shift the empty bands via a scissors operator
[74, 75] or one can use other methods to obtain band structures with a better gap,
such as GW or hybrid functionals (see Sect. 3).
The second deviation is more problematic: both RPA and ALDA lack the first
excitonic peak (labeled E1 in the experimental data), and instead only have a weak
shoulder. This discrepancy persists even if a better band structure (such as GW) is
used as input to calculate the noninteracting response function χ s [14]. This failure
of the ALDA as well as the GGA xc functionals is by no means unique to silicon,
The indices i, j and a, b run over occupied and unoccupied Kohn–Sham orbitals,
respectively. fHxc denotes the sum of the Hartree and xc kernels.
198 C.A. Ullrich and Z.-h. Yang
The real space representation of the Hxc kernel is related to the momentum space
representation as
1 X X iðqþGÞr 0 0
f Hxcσσ 0 ðr, r0 , ωÞ ¼ e f Hxcσσ0 ðq, G, G0 , ωÞeiðqþG Þr : ð37Þ
V q2FBZ G, G0
1 X X
K Hxc
iaσ , jbσ 0
¼ hiki σ jeiðqþGÞr jaka σ i f Hxcσσ0 ðq, G, G0 Þ
V q2FBZ G, G0
0 0
hbkb σ 0 jeiðqþG Þr j jk j σ 0 iδka ki þq, G0 δkb k j þq, G0 ; ð38Þ
0
where the ks are the Bloch wavevectors of the corresponding wavefunctions, and
G0, G00 can be any reciprocal lattice vector. The Kronecker-δs in (38) are a
consequence of Bloch’s theorem.
In the following, we do not consider any spin-dependent excitations (see Yang
and Ullrich [19] for a discussion of triplet excitons within TDDFT). Because we are
interested in optical absorption, only vertical single-particle transitions need to be
considered, so that ki ¼ ka and kj ¼ kb, which implies q ¼ 0 in (38). Equation (36)
then becomes, in reciprocal space,
Xh i
δik, jk0 δak, bk0 ðεak εik Þ þ K Hxc
iak, jbk0 X jbk0 ¼ ΩX iak : ð40Þ
jbk0
Excitons in Time-Dependent Density-Functional Theory 199
Here, i, j and a, b denote occupied and unoccupied band indices, respectively. This
includes, in principle, all empty bands including continuum states; in practice,
however, only a limited number of valence and conduction bands in the vicinity
of the band gap need to be included if (40) is to be solved numerically.
The Hartree part of the coupling matrix is given by
2X 4π
H
K iak , jbk0 ¼
hikjeiGr jakihbk0 jeiGr j jk0 i: ð41Þ
V G6¼0 G2
2 X
iðqþGÞr
iak, jbk0 ¼
K xc lim f 0 ðqÞhikje jakihbk0 jeiðqþGÞr j jk0 i: ð42Þ
V q!0 GG0 xc, GG
The solutions of (40) can be used to calculate the macroscopic dielectric function,
using the following expression [l labels the lth eigenvalue of (40)]:
X 2
iqr ðlÞ
4π X iakh ik je j akiX iak
εmac ðωÞ ¼ 1 lim : ð43Þ
q!0 q2 l
ω Ωl þ iη
Now let us discuss the behavior of the head, wings, and body of the coupling matrix
iðqþGÞr
(42). For G ¼ 0, the matrix element ik e ak vanishes as OðqÞ when q ! 0,
0 iðqþG0 Þr 0
and similarly for the other matrix element, bk e jk . This means that the
head (G ¼ G0 ¼ 0) of K xc iak, jbk0 vanishes unless the head of f xc, GG0 ðqÞ diverges at
2 0
least as q . Likewise, the wings (G ¼ 0 and G finite, or vice versa) vanish unless
the wings of the xc kernel diverge at least as q1.
All local and semilocal xc kernels (ALDA and adiabatic GGAs) remain finite for
all G, G0 , and q. This is easy to see for the ALDA, whose real-space form is
d 2 exc ðnÞ
ALDA
f xc ðr, r0 Þ ¼ δðr r0 Þ; ð44Þ
dn2 nn0 ðrÞ
where exc (n) is the xc energy density of a homogeneous electron gas of uniform
density n and n0(r) is the ground-state density of the material. Carrying out the
Fourier transform we find that the q-dependence simply drops out [79]:
200 C.A. Ullrich and Z.-h. Yang
ð
1 iðGG0 Þr d2 exc ðnÞ
, GG0 ðqÞ ¼ ; ð45Þ
ALDA 3
f xc d re
V cell dn2 n¼n0 ðrÞ
where the integral runs over one unit cell with volume Vcell. The adiabatic GGA xc
0 ðqÞ is substituted into (42), then the
ALDA
kernels exhibit similar behavior. If f xc , GG
contribution from the head and wings of fxc to K xc iak, jbk0 vanishes. For ALDA and
GGA kernels, all changes to the Kohn–Sham spectrum can thus only come from the
0
iak, jbk0 (where both G 6¼ 0 and G 6¼ 0), but these are not sufficiently strong
body of K xc
to produce excitons.
The case q ! 0 in reciprocal space corresponds to r ! 1 in real space. The
long-range behavior of the xc kernel is relatively unimportant for low-lying exci-
tations in finite systems such as atoms and molecules, which means that local and
semilocal xc kernels work reasonably well (an exception to this statement are
charge-transfer excitations [80–86]). However, for extended and periodic systems
it is crucial to have xc kernels with the proper long-range behavior to obtain correct
optical spectra [14, 73]. Gonze et al. [87, 88] pointed out that the head of fxc has to
diverge as q2 for q ! 0 to describe correctly the polarization of periodic insulators.
With the q2 divergence, the head of fxc contributes in the sum of (42), dominating
the other parts of fxc (the wings and the body). Local and semilocal xc kernels do not
have this long-range behavior, and there is no obvious and consistent way of
modifying them to include the long-rangedness. Hence, a different class of approx-
imate xc kernels – excitonic xc kernels – is needed.
f xc ¼ f qp
xc þ f xc :
ex
ð46Þ
The two parts are responsible for the opening of the gap and the excitonic effects,
respectively. Further justification of (46) is given in Sect. 5.4.6. Let us now focus on
the excitonic part and give some examples of how it can be approximated.
Excitons in Time-Dependent Density-Functional Theory 201
Let us begin with an apparent paradox, namely, the so-called contact exciton. Even
though we have stressed that the proper long-range behavior of the xc kernel is
crucial, an ultra-short-range xc kernel of the general form
cont
f xc ðr, r0 Þ ¼ Acont δðr r0 Þ ð47Þ
can produce excitonic features if the constant Acont is properly chosen [15, 18,
92]. By the same token, an ad hoc scaled ALDA, α f xc
ALDA
, GG0 , can, in principle, produce
excitons, although the scaling factor α would have to be rather absurdly large
(typically of the order of ~103).
The resolution of the contact-exciton paradox is that the contact kernel and the
scaled ALDA work via the body of the coupling matrix K xc iak, jbk0 . In other words, the
missing long-range behavior is, somewhat unphysically, compensated by an
ultrastrong short-range electron–hole interaction. It is found [15, 90] that the
contact xc kernel can be tuned to reproduce certain features of the optical spectrum
(for instance, a bound-exciton peak) but at the cost of a poor description of other
parts of the spectrum.
Because we know from Sect. 5.3 that the long-range 1/q2 behavior of the head of the
xc kernel is the key to excitonic effects, it is straightforward to construct a simple ad
hoc approximation which captures the right physics for the right reason. The
resulting so-called long-range corrected (LRC) kernel has the following form:
ALRC
xc, GG0 ðqÞ ¼
f LRC δGG0 ; ð48Þ
q þ G2
where ALRC is a system-dependent fitting parameter. Despite its simple form, LRC
spectra (with properly chosen ALRC) can be in good agreement with experiment
[73, 92] because the head contribution of the kernel tends to dominate over the
local-field effects contained in the contributions of the body of K xc
iak, jbk0 . A simple
connection with the high-frequency dielectric constant ε1 has been suggested [73]:
ALRC ¼ 4:651ε1
1 0:213: ð49Þ
The purpose of this empirical formula was to reproduce the continuum spectrum;
hence, it cannot be expected to (and, in fact, does not) perform well for bound
excitons (see Sect. 5.4.9).
202 C.A. Ullrich and Z.-h. Yang
Rv is the first-order vertex diagram (third on the right-hand side in Fig. 4a),
X ð f i f j Þð f k f l Þ
RV ðr, r0 , ωÞ ¼ 2 φi ðrÞφ*j ðrÞφ*k ðr0 Þφl ðr0 Þhiljwj jki ; ð51Þ
ijkl
z ωi j ðz ωlk Þ
where z ¼ ω + i0+, and the fj are the usual occupation factors. RΣ denotes the sum of
all the remaining four diagrams (the self-energy diagrams):
( )
X φi ðrÞφ*i ðr0 Þφ j ðrÞφ*k ðr0 Þh jjΔjki ð f f Þωik ð f j f i Þωi j
0
RΣ ðr,r ,ωÞ ¼ 4 k i
2 ;
i jk
ωk j z2 ω2ik z ω2i j
ð52Þ
where Δ(r1, r2)¼Σx(r1, r2)νx(r1)δ(r1 r2). Here, Σx is the exchange part of the
self-energy, and νx is the exact exchange potential of DFT, defined as an orbital
functional via the optimized effective potential (OEP) method [45, 97]. It is also
fx
vx
a
vx
b
f xcex
Fig. 4 Diagrammatic representations of (a) the exact exchange kernel fx (50) and (b) the excitonic
xc kernel fex
xc (61), the so-called nanoquanta kernel. Full lines represent noninteracting Kohn–Sham
Green’s functions and dashed lines represent quasiparticle Green’s functions. Thin wavy lines are
bare Coulomb interactions and the thick wavy line is a screened interaction
Excitons in Time-Dependent Density-Functional Theory 203
possible to derive the exact exchange kernel directly as the functional derivative
fx(ω)¼δvx(ω)/δn(ω) [98, 99].
In periodic insulators, the exact-exchange kernel has the long-range behavior
necessary for the formation of excitons [100, 101]. However, the resulting
unscreened electron–hole interaction tends to lead to a dramatic overbinding of
the excitons; in extreme cases this causes a collapse of the optical spectra (i.e., the
exciton would be so strongly bound that it falls below the valence band edge). This
collapse can be prevented by a cutoff of the Coulomb singularity [100, 101]; this is
equivalent to an evaluation of the xc kernel with a screened interaction [15, 102].
however, the implementation of meta-GGAs for the linear response in solids turns
out to be technically rather difficult.
The exact exchange kernel can be approximated in various ways. The simplest
approximation is known as the PGG kernel (after Petersilka, Gossmann, and Gross
[117, 118]). In real space, it is defined as
Xocc 2
2 i φ*i ðrÞφi ðr0 Þ
f PGG ðr, r0 Þ ¼ ; ð53Þ
x
jr r0 jnðrÞnðr0 Þ
where n is the ground-state electronic density. In this form, the PGG kernel has been
successfully applied to calculate atomic and molecular excitation energies as well
as plasmons in nanostructures [119]. Thus, one might be optimistic regarding its
performance for excitons.
We convert the PGG kernel into reciprocal space, assuming that the Kohn–Sham
1=2
orbitals have the form φik ðrÞ ¼ N cell uik ðrÞeikr , where Ncell is the number of unit
cells in the crystal and uik(r) are Bloch functions. fPGG
x can then be written as
Xocc X
occ 0 0
2eiðkk Þðrr Þ
0
f PGG ð r, r Þ ¼ H ikmk0 ðr, r0 Þ; ð54Þ
x
ik mk 0 j r r 0j
where Hikmk0 (r, r0 ) is periodic within one unit cell and defined as
8π X
occ X
occ X e
H ikmk0 ðG G0 , G0 G0 Þ
f PGG
x ðq, G, G0 Þ ¼ ; ð56Þ
V ik mk0 G
0
jq ðk0 kÞ þ G0 j2
8π Xocc e
H ikmk ð0; 0Þ
f PGG
x ðq; 0; 0Þ ¼ : ð57Þ
V i, m, k q2
Excitons in Time-Dependent Density-Functional Theory 205
Unfortunately, explicit calculations show that the performance of the PGG kernel is
disappointing for solids: it does not produce any bound excitons at all, despite
having a nonzero head contribution with the correct 1/q2 behavior [19]. How can
this be reconciled with the fact that the PGG kernel seems to work well in finite
systems such as atoms and molecules? Periodic systems are dominated by the head
of the xc kernel in reciprocal space; however, the situation is very different in finite
systems, where the electron dynamics can be viewed as coming entirely from local-
field effects. Thus, the strongly attractive nature of the PGG kernel in finite systems
would at most translate into a strong body of the xc matrix in periodic systems
(which, however, is irrelevant for excitons), but does not necessarily guarantee a
strong head. Indeed, if one fits the head of the PGG kernel to the LRC kernel (48),
one finds that the resulting constant APGG LRC is orders of magnitude too weak [19].
The underlying reason for the failure of the PGG kernel for periodic insulators
can be inferred
from its real-space definition (53), which can be written in the form
f PGG ¼ ρðr, r0 Þ2 = 2r r0 nðrÞnðr0 Þ , where ρ(r, r0 ) is the Kohn–Sham density
x
matrix. For periodic solids, the long-range behavior of fPGG x is determined by both
the Coulomb singularity and the density matrix. It is a well-known fact [120–124]
that the one-particle density matrix in insulators decays exponentially as ρ(r, r0 )
~ exp(γ|r–r0 |). This effectively cuts off the required long-range behavior and
explains why the head of the PGG kernel is so weak.
χ ¼ χ s þ χ s f xc e
e χ ð58Þ
(for simplicity, we drop all arguments and integrals). At the beginning of this
section we defined the quasiparticle and excitonic parts of the xc kernel; see (46).
It is then easy to write down the following relations for the two parts of fxc [91]:
χ qp ¼ χ s þ χ s f qp
xc χ qp ; ð59Þ
e
χ ¼ χ qp þ χ qp f xc
ex
e
χ: ð60Þ
χ qp f xc
ex
e
χ χ qp f xc
ex
χ qp . Second, find a diagrammatic representation of χe χ qp ; the
details of this representation are quite technical and are not given here (in essence, it
involves two-particle Green’s functions and four-point vertex functions [1, 14, 15,
125]). The key point is that this diagrammatic representation can be very easily
approximated, and one ends up with the following expression:
ð ð ð ð
d3 d4χ qp ð1; 3Þ f xc
ex
ð3; 4Þχ qp ð4; 2Þ ¼ d3 d4Gqp ð1; 3ÞGqp ð4; 1ÞW ð3; 4ÞGqp ð3; 2ÞGqp ð2; 4Þ:
ð61Þ
This is the xc kernel of Reining et al. [89] and many others [also known as the
“nanoquanta” kernel (http://www.cmt.york.ac.uk/nanoquanta/)]. Figure 4b shows
its diagrammatic representation. The numbers in (61) represent space-time argu-
ments, e.g., 1 ¼ (r1, t1). Gqp is a quasiparticle Green’s function, and W is a screened
interaction, formally defined as
ð ð
W ð1; 2Þ ¼ wð1; 2Þ þ d3 d4wð1; 3Þe χ ð3; 4ÞW ð4; 2Þ; ð62Þ
Compared to the exact exchange and nanoquanta kernels, the simplicity of the LRC
kernel is desirable for practical use. The adjustable parameter ALRC requires prior
knowledge to the system, however, and therefore the LRC kernel cannot be used as
a black-box method. The bootstrap kernel proposed by Sharma et al. [132, 133] can
be seen as an attempt to determine the ALRC parameter (which now depends on q,
G, and G0 ) self-consistently while retaining the simplicity of the LRC kernel. The
original definition [132] is written in terms of symmetrized quantities to avoid
0 1=2 1=2
xc ðq, G, G Þ ¼ vG
singularities: for example, f sym ðqÞ f xc ðq, G, G0 ÞvG0 ðqÞ and
χ sym ðq, G, G0 Þ ¼ vG ðqÞχðq, G, G0 ÞvG0 ðqÞ. In terms of regular, non-symmetrized
1=2 1=2
h i
jqj2 δGG0 þ vG ðqÞvG0 ðqÞχðq, G, G0 Þ
1=2 1=2
boot
f xc ðq, G, G0 Þ ¼ ; ð63Þ
jq þ Gjjq þ G0 jχ s ðq; 0; 0Þ
where vG(q) ¼ 4π/|q + G|2 is the Coulomb potential, χ s is the Kohn–Sham linear
response function of (29), and χ is the TDDFT linear response function obtained via
(28). Equations (28) and (63) are solved self-consistently for the xc kernel. The
q ! 0 behavior of χ s(q, 0, 0) is O(q2), and it is canceled by the |q|2 in the numerator.
Thus the bootstrap fxc has the correct q ! 0 behavior because of the presence of
|q + G| |q + G0 | in the denominator.
The bootstrap kernel has been reported to yield good continuum spectra (includ-
ing the enhancement of the band-edge spectra by continuum excitons) for a wide
range of materials [132]. Unfortunately, numerical applications of this xc kernel are
plagued by its slow convergence with respect to the total number of bands included
in χ s. When convergence is finally achieved (which may require including dozens
of unoccupied bands), the results for bound excitons tend to be disappointing, with
exciton binding energies typically orders of magnitude smaller than the experimen-
tal values [19]. In fact, contrary to Sharma et al. [132], the bootstrap kernel does not
yield bound excitons for wide-gap insulators such as LiF and solid Ar (we present
numerical results in Sect. 5.4.9). Improving the performance of the bootstrap kernel
by suitable modification is a subject of ongoing research.
where [135]
π d ½ r s εc ð r s Þ
CðnÞ ¼ 1=3
; ð65Þ
2ð3π nÞ
2 dr s
208 C.A. Ullrich and Z.-h. Yang
αnβ E2g 1 þ Eg
k n; Eg ¼ þ
; ð66Þ
BðnÞ 4πq2 n BðnÞ þ Eg
and [136]
1=2 3=2
1 þ a1 r s þ a2 r s
Bð n Þ ¼ 1=2 3=2
: ð67Þ
3 þ b1 r s þ b2 r s
Here, rs ¼ [3/(4πn)]1/3, εc is the LDA correlation energy per particle [137], and
α ¼ 0.02552, β ¼ 0.6916, a1 ¼ 2.15, a2 ¼ 0.435, b1 ¼ 1.57, and b2 ¼ 0.409. The
performance of the JGM kernel is similar to that of the bootstrap kernel for continuum
spectra; unfortunately, it also severely underestimates exciton binding energies.
We now present some results for exciton binding energies, obtained using (40), to
demonstrate the performance of several of the xc kernels we have discussed. All
TDDFT calculations are done on top of scissor-corrected [74, 75] LDA band
structures, so that only the excitonic part of the exact xc kernel is approximated
by the functionals that we consider. All calculations include only the head of the xc
coupling matrix K xc
iak, jkk0 . Ignoring the wings and the body of the xc coupling matrix
gives an estimated error of less than 5%. Other details specific to our numerical
implementation are described in Yang and Ullrich [19].
Experimental and calculated exciton binding energies for several materials are
collected in Table 1. We compare the performance of three different xc kernels: the
LRC kernel (48), evaluated using (49), the bootstrap kernel (63),1 and the JGM
Table 1 Exciton binding energies calculated with different TDDFT xc kernels, compared with
experimental values (all energies in meV)
GaAs β-GaN α-GaN CdS CdSe Ar Ne LiF AlN ZnO MgO
Exp. 3.27 26.0 20.4 28.0 15.0 1,900 4,080 1,600 75 60 80
LRC 0.858 0.514 0 0.513 1.40 0.304 0.127 1.14 0 0.810 0.076
Boot 0.332 0.199 0 0.461 0.895 1.70 852 32.2 0 1.09 0.051
JGM 0.833 0.382 0 0.741 1.42 41.0 0.593 993 0 4.45 1.79
1
We find that the convergence of the bootstrap kernel strongly depends on the number of bands
used in the iterative calculation of the kernel. The results for solid Ar, solid Ne, and LiF are
obtained by calculating the bootstrap kernel with 30 bands. It turns out that some of our previous
results reported in Yang and Ullrich [19], where the bootstrap kernel seemed to work very well for
Ar, Ne, and LiF, were in fact not fully converged.
Excitons in Time-Dependent Density-Functional Theory 209
Table 2 Top row: fitted ALRC parameter which reproduces the experimental exciton binding
energies with the LRC kernel. Second and third rows: heads of the xc coupling matrix of the
bootstrap and JGM kernels
GaAs β-GaN α-GaN CdS CdSe Ar Ne LiF AlN ZnO MgO
Fit 0.595 2.409 3.6285 4.244 2.144 21.45 96.5 9.5 3.0006 1.6285 4.0405
Boot 0.0884 0.3048 0.2147 0.5895 0.3183 6.448 46.34 4.236 0.3412 0.3620 1.230
JGM 0.2056 0.5245 0.5782 0.7829 0.4631 9.685 10.76 7.78 1.568 0.8008 2.357
kernel (64). Except for α-GaN and AlN, all three TDDFT methods produce a bound
exciton; however, the results are quantitatively not very accurate. By and large, the
excitons are significantly underbound, sometimes by up to two to three orders of
magnitude.
The LRC kernel contains the adjustable parameter ALRC, which can be fitted
to reproduce the experimental exciton binding energy. The fitted values of ALRC
are given in the first row of Table 2. From the head of the bootstrap and the JGM
xc coupling matrices, we can extract the corresponding ALRC parameters; the
results are given in the second and third rows of Table 2. Clearly, the bootstrap
and JGM kernels produce heads which are significantly too weak in comparison
with what would be needed to reproduce the experimental data. Clearly, the
exciton binding energy depends crucially on the strength of the head of the xc
coupling matrix.
Figure 5 shows optical absorption spectra of solid Ar. The spectra in the top
panel were calculated from the solutions of (40) using (43). The bottom panel
(adapted from Sottile et al. [128]) was calculated via (30). The top panel compares
RPA and LRC, both of them using a scissor-corrected LDA band structure as input.
The LRC kernel clearly produces a very strong bound exciton, but the continuum
part is too weak compared to the excitonic peak. The bottom panel compares
experimental results with calculations using GW-BSE and GW-TDDFT, using
the nanoquanta kernel. Clearly, BSE and the nanoquanta kernel are in excellent
agreement, and both agree well with the experimental data (the splitting of the
peaks is caused by spin-orbit coupling, which is not included in the calculations). It
should be noted that a second bound exciton is present, which is missing in the LRC
calculation. For comparison, the bottom panel of Fig. 5 also shows ALDA results,
based on an uncorrected LDA band structure. Clearly, the ALDA bears no resem-
blance whatsoever to experiment.
210 C.A. Ullrich and Z.-h. Yang
25
LRC fit
20 RPA
15
Im[εM]
10
0
12 14 16 18
ω (eV)
15 n=1
n’=1 BSE
fxcex
ALDA
Experiment
10
Abs(ω)
n=2
n’=2
0
10 11 12 13 14
ω (eV)
Fig. 5 Absorption spectra of solid Ar obtained from (43) (top panel) and from (30) (bottom panel,
adapted from Sottile et al. [128]). Reproduced with permission from APS from Sottile
et al. [128]. ©2007
Exciton binding energies in TDDFT are determined via (40). The Hxc coupling
matrix is given by
where the Hartree and xc matrices are defined in (41) and (42). It turns out that the
BSE leads to an equation for exciton binding energies which is formally identical to
(40), except that it features the coupling matrix
Excitons in Time-Dependent Density-Functional Theory 211
where
1X 0
W
K iak , jbk0 ¼ W GG0 ðqÞhikjeiðqþGÞr j jkihbk0 jeiðqþG Þr jak0 i: ð70Þ
V GG0
The Hartree part is the same as in TDDFT, but the xc part is replaced by a coupling
matrix featuring the screened interaction (62). In practice, one ignores the fre-
quency dependence of W; explicitly, one finds
4πε1
GG0 ðq, ω ¼ 0Þ
W GG0 ðq, ω ¼ 0Þ ¼ : ð71Þ
jq þ G0 j2
Let us now compare the two coupling matrices Kxc and KW. Two main differences
become apparent: first, the order of the band indices i, j, a, b is different; second, the
xc matrix only depends on the long-range (q ¼ 0) behavior, while the W matrix also
depends on other values of q.
Figure 6 shows contour plots of the xc and W coupling matrices, calculated for a
one-dimensional model insulator with a soft-Coulomb interaction [18]. The xc
kernel here is the long-range corrected kernel f LRC
xc with a fitting parameter chosen
such that the lowest exciton binding energy in TDDFT and BSE is the same.
The two coupling matrices shown in Fig. 6 are strikingly different. This is not
surprising, because the screened interaction WGG0 (q) has an extra degree of
freedom over fxcGG0 (q ¼ 0); hence, it cannot be expected that an adiabatic xc
kernel can be found that reproduces the full BSE coupling matrix. One can only
hope to reproduce a portion of the BSE coupling matrix, unless the xc kernel is
made frequency-dependent so that at least some of the information from the
Fig. 6 Contour plots of the coupling matrices Kxc (left panel) and KW (right panel). Reproduced
with permission from AIP from Yang et al. [18]. ©2012
212 C.A. Ullrich and Z.-h. Yang
where R, R0 are direct lattice vectors. Because Wannier exciton radii extend over
many lattice constants, one may replace R by a continuous spatial variable r.
Assuming, furthermore, parabolic valence and conduction bands, and using the
effective-mass approximation, (40) becomes, after Fourier transformation,
ð
∇2
XðrÞ þ d 3 r 0 V xc 0 0
eh ðr, r ÞX ðr Þ ¼ EX ðrÞ; ð73Þ
2mr
where E is the exciton binding energy, and the integration goes over all space. This
shows that the TDDFT analog of the Wannier equation (1) is a nonlocal
Schr€odinger equation. With a proper choice of the xc kernel, the nonlocal effective
electron–hole interaction potential V xc
eh supports bound excitonic states.
Fig. 7 Contour plots of the effective nonlocal electron–hole interaction potentials in TDDFT (left
panel) and BSE (right panel). Reproduced with permission from AIP from Yang et al. [18]. ©2012
Excitons in Time-Dependent Density-Functional Theory 213
Because of the formal similarity between BSE and TDDFT in the transition
space representation, we can also define a BSE effective electron–hole interaction
W
potential V BSE
eh in analogy to (72), using the BSE kernel K . Figure 7 shows a
xc BSE
comparison of V eh and V eh for our one-dimensional model insulator. It can clearly
be seen that the nonlocal potentials in both cases are dominated by the diagonal
part; however, under further examination it turns out that V xceh is shallower than
BSE
V eh , and hence is only able to sustain a single bound exciton.
7 Conclusions
Acknowledgments This work was supported by NSF grant DMR-1408904. We thank Lucia
Reining and Francesco Sottile for many helpful discussions. C.U. thanks the ETSF-Palaiseau
group for its hospitality and the Ecole Polytechnique for its support during an extended visit
in 2014.
References
the form of complex absorbing potentials and mask functions. The last two are
regularly used together with time-dependent density-functional theory to describe
the electron emission dynamics of atoms and molecules. Finally, we discuss
approaches to the calculation of energy and angle-resolved time-dependent
pump–probe photoelectron spectroscopy of molecular systems.
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
2 Resonances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
2.1 Definition and Properties of Resonant States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
3 Calculation of Resonances from Complex Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
3.1 Formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
3.2 Bound States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
3.3 Continuum States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
3.4 Resonant States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
3.5 Exterior Complex Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
3.6 Example: Resonance in One Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
4 Density Functional Resonance Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
4.1 Complex Scaling and DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
4.2 Complex Scaling of Exchange and Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
4.3 Resonance Lifetimes in DFRT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
4.4 Time-Dependence in Complex Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
5 Open Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
5.1 Transparent Boundary Conditions Using Green Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 244
5.2 Time-Dependent Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
5.3 Absorbing Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
5.4 Complex Absorbing Potentials (CAPs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
5.5 Mask Function Absorbers (MFAs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
5.6 Time-Dependent Exterior Complex Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
6 Electron Photoemission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
6.1 Sampling Point Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
6.2 Surface Flux Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
6.3 Mask Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
1 Introduction
All natural phenomena occur away from equilibrium. Non-equilibrium systems can
range in scale from microscopic (such as nanostructures and bacteria) to geological
phenomena, and away-from-equilibrium processes occur on timescales ranging
from nanoseconds to millennia. Despite the ubiquitous non-equilibrium systems
and processes, most of the current understanding of physical and biological systems
is based on equilibrium concepts. In fact, in interacting many-body systems, more
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 221
often than not we face the fact that the electronic states have finite lifetimes because
of the coupling to the environment or to a continuum of states (resonance pro-
cesses). Even if we were able to prepare a perfectly isolated quantum system, we
would need to regard a measurement of the system as bringing the system into
contact with an environment. Already a single atom in vacuum cannot be regarded
as completely isolated, because the atom is embedded in the surrounding photon
field (spontaneous emission). Other examples where the coupling to the surround-
ing plays a prominent role include hot electron relaxation in bulk systems and
surfaces after laser irradiation, thermalization caused by electron–phonon coupling,
decoherence in pump–probe experiments, exciton propagation and relaxation in
biological chromophores, and vibrational relaxation in nanomaterials and molecu-
lar systems. Understanding these decay mechanisms provides important informa-
tion about electron correlations, quantum coherence, dissipative and decoherence
processes, and control of these processes has important implications. For instance,
this would make it possible to enhance the performance of molecular/solid-based
optoelectronic devices.
In this context, density-functional theory (DFT) provides an exact theoretical
framework which could yield observable quantities directly, by-passing the need to
calculate the many-body wavefunction Ψ. Hohenberg and Kohn [1] proved that all
observable properties of a static many-electron system can be extracted exactly
from the one-body ground-state density alone (density–potential mapping). Later,
Runge and Gross extended this theorem to time-dependent systems [2]. Time-
dependent density-functional theory (TDDFT) is a rigorous reformulation of the
non-relativistic time-dependent quantum mechanics of many-body systems. The
central theorem of TDDFT is the Runge–Gross theorem which proves a one-to-one
correspondence between the time-dependent external potential νext(r, t) and the
electronic one-body density n(r, t) for many-body systems evolving from a fixed
initial state Ψ0. This implies that the time-dependent electronic density determines
all properties of the interacting many-electron system: all observable properties of a
many-electron system can be extracted from the one-body time-dependent density
alone [2]. What has made both DFT and TDDFT so successful is the Kohn–Sham
scheme [3]: the density of the interacting many-electron system is obtained as the
density of an auxiliary system of non-interacting fermions, living in a one-body
potential. Because of the excellent balance between the computational load it
requires and the accuracy it provides, TDDFT is now a tool of choice for quite
accurate and reliable predictions for excited-state properties in solid state physics,
chemistry, and biophysics, in both the linear and nonlinear regimes. However, there
exist many situations where the electronic degrees of freedom are not isolated but
must be treated as a subsystem embedded in an environment, which influences it in
a non-negligible way. Those situations go beyond the realm of the original formu-
lation of TDDFT which is meant to tackle the isolated dynamics of electronic
systems. It is therefore clear that there is a need to extend density-functional
approaches to the realm of open quantum systems to allow us to treat the processes
described above.
Burke and co-workers recently introduced a TDDFT approach based on a Kohn–
Sham master equation [4], and in recent work this has been pursued by the group of
222 A.H. Larsen et al.
2 Resonances
decays as governed by the damping force, and the rate of decay is proportional to
the width of the resonance peak. Second, if we consider the phase of the oscillation
of the system with respect to that of the external force, we see that it shifts quickly
by up to π as the energy passes that of the resonance. The rate with which it shifts is
inversely proportional to the decay rate.
We mention here a few commonly studied types of resonance in atomic,
molecular, and condensed-matter physics:
• Plasmon resonances where the whole electron charge density in a material
resonates with incoming light. Surface plasmon resonances are central to the
field of plasmonics.
• Scattering resonances where incident electrons interact with an atom or mole-
cule. Near a resonance energy, the electrons couple strongly and the scattering
cross-section shows a peak. The process may be understood as the incoming
electron becoming temporarily trapped in a metastable state before escaping.
• Asymmetric Fano resonances [10]. These occur when two coupled excitation
pathways interfere with each other.
• Autoionizing resonances, wherein a system such as an atom or molecule is
unstable with respect to the ejection of one or more electrons. These are similar
to those that would be observed in time-resolved spectroscopies and electron
scattering experiments as mentioned above.
• Electron transport processes with molecular junctions, where a bias voltage
causes electrons to jump from one metallic lead across a metastable state at a
molecule, then escapes through another lead. Such processes have, for example,
been studied using DFT plus non-equilibrium Green functions represented with
atomic basis sets [11–14].
• Adsorption of an atom onto a surface where the continuum states of the surface
couple with the discrete atomic states which then become unstable, broadening
into resonances. The Newns–Anderson model [15] describes this process for a
one-electron adsorbate.
There are many further classes of resonance which we do not mention here.
Below we consider only a small class of resonances, namely scattering or
autoionizing ones. In this context, a resonance is a metastable quantum mechanical
state that the system possesses, and which can be associated with a wavefunction.
Below we describe some mathematical properties of such resonances, with the
objective of eventually calculating them from static or time-dependent DFT.
asymptotic region, taking in one dimension the form cos(kx + δ). See for instance
the simple demonstration by Gellene [16] which we consider again later. A reso-
nance energy and width can be estimated by locating the energy where the phase
shift δ changes most rapidly, and the width can be estimated from the maximum rate
of change. This intuitively relates the resonance to a strong coupling of the system
with continuum states in a narrow energy interval, as we noted in the beginning.
A more mathematically precise way of identifying a resonance is, following the
work of Siegert [17], to search for complex energies corresponding to singularities
of the scattering cross section. A pole close to the real energy axis would produce a
peak in the scattering cross-section for real energies, consistent with a resonance.
As noted by Siegert, the corresponding condition on a wavefunction1 ψ(r) is that far
away from the scattering region:
dψ ðr Þ
¼ ikψ ðr Þ; ð1Þ
dr
k2 =2 ¼ ε iΓ=2; ð2Þ
where ε > 0 is the real resonance energy, and Γ > 0 its width. This yields a discrete
set of resonant states characterized by being purely outgoing waves. States obeying
the boundary condition (1) are frequently called Siegert or Gamow–Siegert states,
and they diverge as r ! 1. See, for example, Hatano et al. [18] for a detailed
description of resonant states.
Most computational methods in quantum mechanics work in terms of square
integrable states, and thus cannot straightforwardly represent a resonance
wavefunction. An elegant solution to this problem is the complex scaling method,
where one uses complex spatial coordinates to suppress the exponential divergence.
One thus solves for functions that obey the usual boundary conditions, ψ(r) ! 0 for
r ! 1. This also has the convenient advantage that the boundary conditions no
longer depend on k. The method relies on the properties of analytic functions to
transform the Hamiltonian into a non-Hermitian operator whose point spectrum
consists exactly of that of the bound states along the negative real axis plus the
complex resonance energies which have positive real part and negative imaginary
part. The wavefunctions of bound as well as resonance states are square integrable
analytic continuations of the original ones. These properties make the complex
scaling method a powerful computational tool as it can make use of many existing
methods which do not otherwise apply to unbounded scattering states.
1
We mention for completeness that Siegert worked in a spherical system where the represented
quantity is really r times the wavefunction; this however happens to yield the same equation as in
the one-dimensional case.
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 225
Although the complex scaling method clearly works with any kind of particles in
a finite system, here we explicitly assume that we are dealing with electrons
temporarily trapped by simple potentials representing atoms or molecules. The
electrons eventually tunnel out to a far-away region which we do not wish to
represent explicitly in the calculation. We are thus dealing with the specific case
of an open quantum system where we only have particles leaving the system.
The complex scaling method was initially developed by Aguilar, Balslev, Combes,
and Simon [19–21], and is based on a scaling r ! reiθ of the position variable in the
Schr€odinger equation. This is referred to as uniform complex scaling. Here we
review uniform complex scaling in the simple case of independent particles. Most
recent work is based on a later generalization called exterior complex scaling [22],
which we consider later. The following is a rather informal description of complex
scaling, focusing on a few important cases. More information can be found in any of
the many existing reviews.[23–27]
3.1 Formalism
^ ψ ðrÞ ¼ εψ ðrÞ:
H ð3Þ
where N is the number of spatial coordinates on which the scaling is applied (thrice
the number of particles in the 3D many-body case). θ, the scaling angle, is a fixed
226 A.H. Larsen et al.
Resonances
0
Bound states
[a.u.]
θ
−1
Co
Im
ntin
−2
uum
θ = 0.1
sta
θ = 0.3
tes
θ = 0.5
−3
−4 −2 0 2 4
Re θ [a.u.]
Fig. 1 Effect of complex scaling on the spectrum for the 1D potential ν(x) ¼ 3(x2 2)exp(x2/4).
Bound-state eigenvalues (bold circles) are independent of θ while the continuous spectrum rotates
by 2θ around the threshold 0. Because of the finite size of the simulation box, the numerically
calculated unbound states (uncircled) do not fall exactly on the line arg z ¼ 2θ. Resonances (thin
circles) are resolved when θ is sufficiently large for them to segregate from the continuum states.
Calculated using a uniform real-space grid from 18 to 18 a.u. with 250 points and fourth-order
Laplacian finite-difference stencil
number formally supposed to lie within 0 θ π/4, although this depends on the
analyticity of the potential. The scaling operation transforms the position and
momentum operators as x ! xeiθ and d/dx ! eiθd/dx, wherefore the Hamiltonian
transforms to
^ θ ψ θ ðrÞ ¼ εθ ψ θ ðrÞ
H ð5Þ
with
H ^ R^ 1 ¼ 1ei2θ ∇2 þ v reiθ :
^ θ ¼ R^ θ H ð6Þ
θ
2
The transformation maps the potential to its analytic continuation on reiθ in the
complex plane. The interesting property of Ĥθ is how its spectrum and eigenstates
are related to that of Ĥ. First of all, Ĥθ is non-Hermitian and therefore admits
complex eigenvalues. The continuous spectrum “swings down” by an angle of 2θ as
shown in Fig. 1. Meanwhile, the energies of any bound states remain unaffected.
Finally, for sufficiently large θ, new eigenvalues materialize which are independent
of further increase of θ and which are taken to represent resonances. Let us have a
closer look at each of these three effects separately.
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 227
Im(xeiθ )
0
L 0 L
Re(xeiθ )
Fig. 2 Complex integration path with directions indicated by arrows. If the integrand is suitably
localized and analytic on the integration path, the indicated path becomes equivalent to that over
the real axis from 1 to 1 as L ! 1. This ensures that the unphysical complex scaling angle
does not affect matrix elements or expectation values
Suppose ϕ(r) and ψ(r) are square-integrable and reasonably well-behaved states.
We consider first the scaling operation U^η ψ(r) ¼ eNη/2 ψ(reη) where η is a real
number. This operation is easily seen to be unitary; for example it preserves scalar
products:
D E ð ð D E
^{^
ϕ U U
η η ψ ¼ ϕ *
ð re η
Þψ ð re η
Þdre Nη
¼ ϕ* ðr0 Þψ ðr0 Þdr0 ¼ ϕψ ; ð7Þ
where we used the substitution r0 ¼ reη. A real scaling therefore preserves matrix
elements and eigenvalues.
The derivation of the complex scaling method starts with the unitarity of the real
scaling, then considers the extension to the complex plane of the scaling parameter
η. However, as we see, the situation becomes radically different when the scaling is
complex. In order for the method to be correct, the scaling operation must retain
some property resembling unitarity to make sure that observables do not arbitrarily
change with the scaling parameter. The analytic continuations of functions defined
originally on the real axis are not always within the Hilbert space (hence breaking
unitarity). However, for suitable states and operators, as we see later, the complex
scaling operation corresponds simply to a change of integration path which pre-
serves scalar products. Let us consider a matrix element of some local operator
D E ð
^ ^ ðrÞψ ðrÞdr:
ϕ O ψ ¼ ϕ* ðrÞO ð8Þ
This integral is taken for each coordinate axis over all real numbers 1 to 1.
Imagine now that we liberate each position coordinate and allow it to take complex
values. Then, instead, we take the integral over some complex path, such as the one
in Fig. 2 with three segments. If the diagonal segment is long enough (L ! 1 in the
228 A.H. Larsen et al.
figure), and the integrand is analytic and sufficiently localized, then the integral
along the vertical segments is zero. Thus the integral over the diagonal z ¼ xeiθ is
independent of θ and equal to that along the real line:
ð ð aeiθ
dx ¼ lim dx: ð9Þ
a!1 aeiθ
The substitution r0 ¼ reiθ then transforms the integral back so the integration
variable is (unlike the integrand) real:
D E ð
^ ^ θ ðrÞψ ðrÞdr;
ϕ O ψ ¼ ϕθ ðrÞO θ ð10Þ
with
ψ θ ðrÞ ¼ eiNθ=2 ψ reiθ ¼ R^ θ ψ ðrÞ; ð11Þ
*
ϕθ ðrÞ ¼ eiNθ=2 ϕ* reiθ ¼ R^ θ ϕðrÞ ; ð12Þ
O ^ R^ 1 :
^ reiθ ¼ R^ θ O
^ θ ðr Þ ¼ O ð13Þ
θ
Note how (1) the complex prefactors of eiNθ/2 from (4) serve to “absorb” exactly the
volume element eiNθ produced by the variable substitution, and (2) the left states or
bras are effectively rotated by θ. Furthermore, if the unscaled state ϕ(r) is real, the
cumbersome notation for ϕθ ðrÞ of (12) can be avoided:
The previous discussion does not apply to states that are not localized, such as
continuum states. Let us consider the complex-scaled Schr€odinger equation for a
free particle in one dimension:
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 229
1 d2 ψ θ ðxÞ i2θ
e ¼ εθ ψ θ ðxÞ: ð15Þ
2 dx2
We immediately see that this is the same differential equation as the unscaled one,
and thus has the usual set of solutions:
As per the standard procedure, let us say that the particle is confined to some finite
box. We then require that ψ θ(x) be 0 on the boundaries, which quantizes k to a set of
real positive numbers. Taking the limit of large boxes, we see that solutions exist for
all k > 0. It follows that the energy εθ in (15) must become complex according to
1
εθ ¼ k2 ei2θ , k > 0: ð17Þ
2
Evidently the spectrum has been rotated by an angle of 2θ into the fourth quadrant
of the complex plane. Meanwhile, the solution wavefunctions for the free particle
have the same form as without the complex scaling operation.
What, then, is so interesting about the complex-scaled solutions ψ θ(x)? Because
they are not normalizable, and because their energy depends on the scaling angle θ,
they are not of much use computationally. However, we can gain some insight by
scaling them back to θ ¼ 0 to obtain
R^ θ ψ θ ðxÞ ¼ eiθ=2 Aeikx cos θ ekx sin θ þ Beikx cos θ ekx sin θ : ð18Þ
For x ! 1 the right-going term diverges whereas the left-going one dies out. For
x ! 1 it is the left-going one which survives. The solution ψ θ(x) to the complex-
scaled problem therefore resembles an outgoing, exponentially diverging state. We
see intuitively that the complex scaling operation may have something to say about
the outgoing character of states. However, as mentioned, the states ψ θ(x) are not
normalizable and their energies depend on θ. The main effect of the complex
scaling operation was to move the continuous spectrum of the Hamiltonian away
from the real axis, close to which we find the resonance eigenvalues as we see later.
If the system consists of a central, (almost) localized potential surrounded by
vacuum, an unbound state still has the form (16) almost everywhere in space.
Importantly and non-trivially, this also works with the Coulomb potential in spite
of its long range. The complex scaling transformation still causes the continuous
spectrum to rotate by exactly 2θ. In numerical representations this is only
approximately true because of incompleteness of the basis and in particular finite
simulation boxes as in Fig. 1.
We note here that the method cannot in general be combined with extended
(periodic) systems, because complex scaling fundamentally works in terms of the
asymptotic form of decaying functions. For example, a metal would possess
occupied continuum states which do not decay at the end of the cell. This makes
230 A.H. Larsen et al.
their properties depend on the scaling angle θ as we saw for free particles. However,
from what we have seen so far, one could well imagine using complex scaling in
some directions and not others – for example, to describe electrons escaping in the
z direction from a surface which is periodic along x and y, or radially from a
one-dimensional nanowire.
From standard scattering theory we know that resonances are associated with
wavefunctions that diverge exponentially at increasing distances. If the resonance
is generated by a short-range potential, the resonance wavefunction must far away
equal or approach that of a free particle.
In one dimension the resonance wavefunction must therefore have the form
where we have used the complex wavenumber k ¼ p iq with positive p (so the
wave is outgoing) and q (so it diverges exponentially). Now apply the complex
scaling transformation to this function:
R^ θ ψ ðxÞ ¼ eiθ=2 eið piqÞxe ¼ eiθ=2 eið p cos θþq sin θÞx eð p sin θþq cos θÞx :
iθ
ð20Þ
2
The above discussion is, of course, very informal. Scrinzi and Piraux have presented a more
complete argument on the link between outgoing wavefunctions and square integrability after
complex scaling; see Scrinzi and Piraux [31], Appendix A.
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 231
“complex virial theorem,” would best approximate the fully converged complex
energy [32, 33].
The uniform and exterior scaling integration contours are shown in Fig. 3. The
important condition for exterior scaling is that the scaling retains the asymptotic
form x ! xeiθ which ensures outgoing-wave boundary conditions. Because the
integration contour is not differentiable, neither is an exterior complex-scaled
function that corresponds to a smooth original function. Recall from uniform
scaling that we needed to multiply by eiθ/2 to “absorb” the now complex volume
element when integrating. We have not done anything to the volume element in
(21), and therefore we need to apply a factor of eiθ when calculating integrals over
the complex segments. Alternatively, most authors define the exterior scaling
operation so the wavefunction in the exterior segments includes the complex
prefactor; the functions then become discontinuous [38], but we do not need to
consider the volume element when integrating. Here we have followed the original
convention of Simon [22] where the function is always continuous. As long as the
discontinuities of the complex-scaled functions or their derivatives are well
incorporated into the numerical basis set used to represent them, they are harmless.
232 A.H. Larsen et al.
Uniform
Exterior
Smooth exterior
Im Fθ (x)
θ
−R0 0 R0
Re Fθ (x)
Fig. 3 Possible complex integration contours for uniform, exterior and smooth exterior complex
scaling. θ ¼ 0.6. The contours must be continuously deformable (without crossing any poles) back
to the real axis in order for them to be equivalent to a real integration. Note that, as per basic
complex analysis, the contours themselves do not have to be differentiable – it is sufficient that the
integrand be analytic
where
1 3 5 4
V 0F ðxÞ ¼ ½F0 ðxÞ F000 ðxÞ ½F0 ðxÞ ½F00 ðxÞ ;
2
ð24Þ
4 8
3
V 1F ðxÞ ¼ ½F0 ðxÞ F00 ðxÞ: ð25Þ
3
This would be less of an advantage in Cartesian 3D calculations where a smooth scaling could be
applied spherically, whereas the sharp scaling would need a cube to align its boundary with
the grid.
234 A.H. Larsen et al.
Fig. 4 Rectangular
potential barrier supporting V0
resonances. Exterior
V (x < 0) = ∞
V (x) [a.u.]
complex scaling ensures exterior
that resonance complex
wavefunctions localize and scaling
appear as eigenstates of the
scaled Hamiltonian
0
0 a b c
x [a.u.]
8
< 0, 0 x < a,
V ðxÞ ¼ V 0 , a x < b, ð26Þ
:
0, b x;
seen in Fig. 4. Both Gellene [16] and Simons [43] have considered this problem
previously. As the rectangular barrier is not an analytic function, we cannot use
uniform complex scaling. However, nothing stops us from using exterior scaling,
with the scaling transformation starting somewhere outside the barrier at x ¼ c > b.
We thus use the contour
x, 0 x < c,
Fθc ðxÞ ¼ ð27Þ
c þ eiθ ðx cÞ, c x:
each of which the wavefunction must be a solution to the Schr€odinger equation for a
free particle but with different local momenta k1, k2, and kθ3 which may be complex:
ψ 1 ðxÞ ¼ iA eik1 x eik1 x ¼ 2A sin ðk1 xÞ, 0 x < a; ð28Þ
ik2 x
ψ 2 ðxÞ ¼ Ce ik2 x
þ De , a x < b; ð29Þ
ψ 3 ðxÞ ¼ Feik1 x þ Geik1 x , b x < c; ð30Þ
ik3θ x ik3θ x
ψ 4θ ðxÞ ¼ Ie þ Je , c x: ð31Þ
The expression for ψ 1(x) has been chosen to fulfill the boundary condition ψ 1(0) ¼
0, and A eventually determines the normalization of the state. To relate the three
wavenumbers k1, k2, and kθ3 , we note that applying the Hamiltonian to the
wavefunction must yield the same energy eigenvalue εθ ¼ k21 =2 ¼ k22 =2 þ V 0 ¼
iθ θ 2
e k3 =2 within each segment. From this may take k3θ ¼ k1 eiθ .
The segments must be joined continuously and differentiably, i.e., ψ 1(a) ¼ ψ 2(a)
and ψ 01ðaÞ ¼ ψ 02ðaÞ at x ¼ a. Likewise ψ 2(b) ¼ ψ 3(b) and ψ 02ðbÞ ¼ ψ 03ðbÞ. At x ¼ c,
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 235
the onset of the scaled exterior region, the derivative ψ 0 3(c) must match the
scaled derivative ψ θ4 0 (c), so the derivative becomes discontinuous [22]:
(We have here, for esthetic reasons, chosen not to include the square root of the
volume element or Jacobian in the definition of ψ θ4 (x); if we had, the function itself
would have been discontinuous as discussed in Sect. 3.5.)
We thus have two equations at each of the points a, b, and c, for a total of six
equations. A seventh equation follows from the requirement that the function be
square integrable. These seven equations determine the six unknown coefficients C,
D, F, G, I, and J, and further quantize the energy so that we get solutions only for
specific wavenumbers k1, k2, and kθ3 .
Gellene [16] provides expressions for the coefficients C, D, F, and G in terms of
A so that ψ 1(x), ψ 2(x), and ψ 3(x) match at the points a and b. The resonances are
then found by considering the phase shift between the incoming and outgoing
coefficients F and G of ψ 3(x). However, this is very different in our case using
complex scaling; here, the coefficients I and J of ψ θ4 (x) must ensure square
integrability.
Physically, we would expect of a resonance that its energy is much greater than
its width. The wavenumber k1 then has real and imaginary parts k1 ¼ p iq such
that p
q. The wavefunction can thus be written as
ψ 4θ ðxÞ ¼ Ieð p sin θþq cos θÞx eið p cos θþq sin θÞx
ð34Þ
þJeð p sin θq cos θÞx eið p cos θþq sin θÞx :
For scaling angles θ not too close to zero, the first term converges whereas the
second diverges as x ! 1, and so we conclude that J ¼ 0. Relating the right and left
values and derivatives of ψ 3(x) and ψ θ4 (x) at x ¼ c we get
−2
0 2 4 6 8 10 12
x [a.u.]
For any energy ε iΓ/2, the wavenumbers k1 and k2 are uniquely determined. The
solutions can then be determined numerically. The three complex resonance ener-
gies closest to 0 are given by the real parts ε ¼ 0.421, 1.65, 3.57, and half-widths
Γ/2 ¼ 0.00138, 0.0189, 0.138. Figure 5 shows the corresponding resonance
wavefunctions. The eigenvalues slightly disagree with those by Gellene who
works effectively on the real axis. This is because the two methods are different:
With complex scaling we find an eigenvalue in the complex plane which corre-
sponds exactly to an outgoing wave. Working on the real axis, we would find the
real energy which responds most strongly to that eigenvalue. However, as the
complex eigenvalue gets further away from the real axis, location and width soon
begin to differ.
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 237
DFT is based on the minimization of a functional of the real electron density. The
minimum of the functional and the corresponding electron density are the ground-
state energy and electron density [1, 3]. For practical calculations one uses a set of
single-particle states or Kohn–Sham states to facilitate evaluation of the kinetic part
of the functional. The Kohn–Sham energy functional contains the following contri-
butions: the kinetic energy, the Hartree energy, the exchange–correlation
(XC) energy, and the energy from a system-dependent external potential. The
kinetic energy functional depends explicitly on the Kohn–Sham wavefunctions
whereas the others depend on them only through the density. Either way, all the
terms can be understood as sums of matrix elements of operators. We know from
Sect. 3.2 how the complex scaling operation conserves matrix elements of states
that are spatially localized, provided that the operators are analytic. We can
therefore reasonably expect complex scaling to be made to work within DFT,
once we know how each term in the energy functional scales. The combination
has been dubbed density functional resonance theory (DFRT) [50].
One would thus propose a complex-valued energy functional
1 X ð 1
ðð
ρθ ðrÞρθ ðr0 Þ
Eθ ¼ ei2θ f n ψ θn ðrÞ∇2 ψ θn ðrÞdr þ eiθ 0k
drdr0
2 n ð
2 k r r
ð38Þ
θ
þ Exc ½nθ þ vext reiθ nθ ðrÞdr
X
nθ ð r Þ ¼ f n ψ θn ðrÞψ θn ðrÞdr ¼ eiNθ n reiθ ; ð39Þ
n
where fn are occupation numbers, and N the number of dimensions in which the
coordinates are complex-scaled. In (38) the kinetic and external contributions are
complex-scaled as normal. In the Hartree energy, ρθ(r) denotes the complex-scaled
charge density which is the electron density nθ(r) plus any other contributions such
as pseudopotential charges (whose complex-scaled form is uniquely determined by
requiring that their Hartree potential scales as normal). The Hartree energy itself
scales as EHθ ½nθ ¼ eiθ EH ½nθ , i.e., the standard Hartree functional is applied to the
complex density, with the factor eiθ appearing because of the 1/r kernel. We
discuss the complex XC energy functional Eθxc [nθ] later.
Being complex, “minimizing” the energy functional (38) does not strictly make
sense. Nevertheless, the lowest-energy resonance is obtainable as a stationary point
of the complex energy functional [51]. An equation for the stationary point can,
as normal, be obtained by taking the derivative with respect to the wavefunctions
plus a set of Lagrange multipliers which ensure normalization. This yields the
complex scaled Kohn–Sham equations
θ 1
HKS ψ θn ðrÞ ¼ ei2θ ∇2 þ vθ ðrÞ ψ θn ðrÞ ¼ εθn ψ θn ðrÞ ð40Þ
2
for ψ θn(r) and εθn, where we have taken the derivative with respect to the left states
ψ θn ðrÞ. If the unscaled Hamiltonian is real, the states can be chosen to be real so that
ψ θn ðrÞ ¼ ψ θn ðrÞ. In general, however, we could equally well have derived a
Hamiltonian for the left states ψ θn ðrÞ.
In the Kohn–Sham equations (40) we have introduced the effective potential
vθ ðrÞ ¼ vHθ ðrÞ þ vxc
θ
ðrÞ þ vext reiθ ð41Þ
which allows the potential to be determined from the charge density by solving a
complex Poisson problem using standard techniques. What remains to be discussed
now is the XC functional.
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 239
The first DFRT calculations were carried out in a one-dimensional model potential
with two electrons in the same (singlet) state [50]. The method was demonstrated
using the exact KS potential, which in this case is
θ ∇2 ψ θ ð x Þ
vexact ðxÞ ¼ ei2θ þ εθ ; ð43Þ
2ψ θ ðxÞ
along with exact exchange (EXX) which, in this case, simply cancels out half the
Coulomb energy. However, for systems with more particles, and indeed for realistic
numerical calculations in the style of modern DFT software, the XC functional
would have to be one of the many commonly used approximations. For simplicity
we ignore any notion of spin below. The simplest functional is the local density
approximation (LDA), the complex scaling of which was studied by Larsen
et al. [49]. The first question is whether the functional is analytic. The exchange
energy is given by
ð
3 3 1=3 4=3
Ex ½n ¼ n ðrÞdr; ð44Þ
4 π
where the fractional power n4/3 is three-valued on the complex numbers and we
must mind the branch cuts. Following the arguments of Sect. 3.2 for handling the
change in complex contour, the integral scales as follows as long as we do not run
into a branch cut:
ð ð ð
4=3 iNθ
n4=3 ðrÞdr ¼ n4=3 reiθ dreiNθ ¼ eiNθ nθ ðrÞ dre
ð ð45Þ
¼ eiNθ=3 nθ ðrÞdr:
4=3
Taking the derivative with respect to nθ(r) we get the exchange potential
1=3
3 iθ
eiNθ=3 nθ ðrÞ ¼ vLDA
1=3
xθ ðrÞ
vLDA ¼ re ; ð47Þ
π x
i.e., the expression is consistent with analytically continuing the expression for the
unscaled potential.
240 A.H. Larsen et al.
Fig. 6 “Stitching” branches of the cube root for the LDA exchange potential. The procedure starts
at x ¼ 0 where we know that the potential must be real. When the density takes the value of a
branch cut of the cube root (indicated by arrows), the function must switch to a different branch to
retain analyticity. The stitched function, indicated by the shaded gray band, is analytic everywhere
and always follows one of the three branches of the cube root. In this example the density is a
Gaussian function. From Larsen et al. [49]
1 dεc ðr s Þ
v c ð r s Þ ¼ εc ð r s Þ rs ; ð48Þ
3 dr s
where
and rs is the Wigner–Seitz radius, i.e., rs(r) ¼ [3/(4πn(r))]1/3. The complex log-
arithm can be stitched quite analogously to the cube root. Other XC functionals can
be stitched similarly, provided that they do not contain poles that get in the way of
the integration contour. With exterior complex scaling we avoid scaling the regions
of space where most of the action happens, potentially avoiding these problems. We
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 241
Fig. 7 Ionization rates of the helium atom in static electric fields from different methods. The
accuracy at low field strengths is determined by how well the XC functional predicts the energy of
the highest occupied orbital, which LDA is known to greatly overestimate. From Larsen et al. [49]
mention a recent time-dependent study [53] which uses smooth exterior complex
scaling with the LB94 [54] XC model potential for spin σ:
with
k∇nσ ðrÞk
xσ ðrÞ ¼ 4=3
: ð52Þ
nσ ðrÞ
This expression also has several issues with analyticity as it involves both division
and fractional powers. In Telnov et al. [53] the exterior scaling contour was
probably chosen so as to avoid these, but unfortunately the issue was not mentioned.
In this section we present a few results from DFRT on physical systems. Figure 7
shows the ionization rate of a helium atom in an electric field as a function of field
strength calculated with different methods: LDA, EXX (Hartree–Fock), ADK [55],
and an accurate correlated-electron calculation by Scrinzi [45].
ADK is a simple approximation which is correct in the limit of weak fields. The
ionization potential of the atom entirely determines the form of the curve in this
limit. Precisely because low-field asymptotics are determined by the value of the
ionization potential, the utility of a functional in this limit is directly linked to the
242 A.H. Larsen et al.
precision with which it estimates the ionization potential, i.e., which energy it
assigns to the highest occupied state.
LDA is well known to overestimate this energy, and therefore calculates too high
ionization rates for low fields. This problem is attributed to the wrong asymptotic
decay of the LDA potential [54]. Meanwhile, Hartree–Fock is known to produce
accurate orbital energies, and the decay of the exact exchange potential has the
correct asymptotic form. EXX also yields results that are close to the reference by
Scrinzi. This all suggests that a good XC functional for DFRT resonance lifetime
calculations is one retaining the correct asymptotic form of the potential, such as the
previously mentioned LB94 functional.
∂ψ ðrtÞ ^ ψ ðrtÞ:
i ¼H ð53Þ
∂t
∂ψ θ ðrtÞ ∂ψ ðrtÞ ^ ^ ^ 1 ^
i ¼ iR^ θ ^ θ ψ θ ðrtÞ:
¼ R θ H R θ R θ ψ ðrtÞ ¼ H ð54Þ
∂t ∂t
^
ϕθ ðrtÞ ¼ eiH θ t ϕ0θ ðrÞ ¼ eiεθ t ϕ0θ ðrÞ: ð55Þ
If, further, the eigenstate represents a resonance, so that its energy has a negative
imaginary part, ϕθ(rt) decays exponentially while everywhere maintaining its
shape. To calculate a general expectation value after a certain time, we would,
according to (10), need to apply the left state ψ θ ðrtÞ ¼ ½ψ θ ðrtÞ* as per (12). The
left state can be time evolved using (54) with θ. The Hamiltonian Ĥθ is the
conjugate of Ĥθ so all eigenvalues are likewise conjugated. If ψ θ(rt) contains
exponentially decaying components, the corresponding components of ψ θ(rt)
exponentially increase at the same rate (one could equivalently say that they
propagate backward in time [59]). In principle the increase of the left state would
be cancelled by the decay of the right so that the norm, calculated using both left
and right states, is time independent, but any numerical error accumulates over the
course of the time evolution and eventually causes the procedure to break down.
Although Bengtsson and co-workers have demonstrated that a complex time
propagation path can be used to stabilize the time evolution [58], most applications
of complex scaling with time evolution have been handled differently. The typical
approach is to use exterior complex scaling and time evolve only the right states,
then calculate all physical quantities using only the right states although this in
general is not formally justified. This approach is discussed further in Sect. 5.6.
Fig. 8 A system localized in a bounded region A exchanges electrons with the environment B. We
look for the correct boundary conditions for the TDSE in A such that the bounded wavefunction
ΨA(t) matches the complete wavefunction Ψ(t) at all times t
In general, the problem of finding the Green function for an arbitrary system is hard
to solve. However, including in A most of the atomic and molecular structure leaves
us in B with a problem which, in many cases, can be easily solved.
The simplest case consists of choosing B to represent the empty space, and the
method lends itself to the description of scattering or ionization [62, 63]. On a more
advanced level, one may choose B to represent a bulk system and, in conjunction
with a time-dependent potential, create a base model for electron transport [64,
65]. Alternatively, by mixing both bulk and empty space Green functions, the
frameworks can adapt to the description of ionization from surfaces [66, 67].
The approach is adaptable to a large variety of situations. This versatility has,
however, to face the fact that discretizing the otherwise exact equations often leads
to computationally demanding implementations with limited application. On the
practical level, either one introduces an approximation which affects the quality of
the results, or one just uses a simple time propagation of a full-dimensional system,
which represents a challenging task [68].
In spite of the technical limitations, the approach provides a fundamental and
illustrative description of the open boundary problem. Below we discuss two of the
most notable derivations present in the literature.
The original Green function embedding was developed in the context of surface and
solid state physics for the static Schr€odinger equation by Inglesfield [69]. It was
subsequently extended to the time-dependent case in Inglesfield [67, 70] by the
same author, but similar derivations have been proposed earlier in different fields,
for instance, to describe the interaction of a strong laser with atoms in Boucke
et al. [62] and Ermolaev et al. [63], and for electron transport in Hellums and
Frensley [64].
Below we introduce the theory following an approach similar to the one used to
describe molecular transport with TDDFT by Kurth [65].4 We first restrict our-
selves to the single-electron case and then discuss the extension to the many-
electron one with TDDFT.
Let us consider the case of a system in contact with a reservoir as shown in
Fig. 9. We want to find a closed set of conditions that have to be imposed on the
equations for a wavefunction in A such that it correctly matches its outer part in
B for all times. Following the division in the figure, we can write the time-
dependent Schr€ odinger equation for the system A coupled with a reservoir/environ-
ment B using a block matrix representation:
4
An analogous approach was first presented by Hellums [64] in a single-particle picture.
246 A.H. Larsen et al.
∂ ψ A ðr; tÞ ^ ðtÞ H
H ^ ðtÞ ψ A ðr; tÞ
i ¼ ^ A, A ^ A, B ; ð56Þ
∂t ψ B ðr; tÞ H B, A ðtÞ H B, B ðtÞ ψ B ðr; tÞ
where ψ A(r,t) and ψ B(r,t) are the wavefunctions projected onto each separate
region. Here we consider the general case where the Hamiltonian is time-
dependent, and its components include two diagonal terms ĤA,A(t) and ĤB,B(t)
operating within each separate region and two coupling terms ĤA,B(t) and ĤB,A(t)
connecting the environment to the system.
To derive the embedded time-dependent equations we introduce the retarded
Green function G0 for the reservoir, defined as
∂ ^ B, B ðtÞ G0 ðr, r0 , t, t0 Þ ¼ δðr r0 Þδðt t0 Þ;
i H ð57Þ
∂t
with boundary conditions G0(r,r0 ,t+,t) ¼ i, G0(r,r0 ,t,t+) ¼ 0, and where t+ repre-
sents a time approaching t from above. Because of the explicit time dependence of
ĤB,B(t), it generally depends on both the time variables t and t0 . We note however
that the solution greatly simplifies if we consider B to represent empty space. In this
case, G0(r,r0 ,t,t0 ) is the free propagator, which depends only on the time difference
t t0 and is known analytically.
Using G0(r,r0 ,t,t0 ) we can directly build the solution of the differential equations
in B. This corresponds to considering only the second row in (56), and results in5
5
To simplify notation we avoid explicitly writing out all the coordinates. We also use the same
convention used in Kurth et al. [65] where operators are thought of as matrices with continuous
indices along the spatial coordinates. We thus omit explicit reference to r and r0 and interpret
operator products as integrals.
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 247
ðt
ψ B ðtÞ ¼ iG0 ðt; 0Þψ B ð0Þ þ ^ 0 ðt, t0 ÞH
G ^ B, A ðtÞψ A ðt0 Þdt0 : ð58Þ
0
The final equation governing the time evolution for ψ A(t) can be written in a closed
form simply by plugging (58) into the first row of (56). After that we obtain
with
ðt
^ Σ ½ψ A ðtÞ ¼
H ^ A, B ðtÞG
Σ^ ðt, t0 Þψ A ðt0 Þdt0 þ iH ^ 0 ðt; 0Þψ B ð0Þ: ð60Þ
0
^ A, B ðtÞG
In this equation, Σ^ ðt, t0 Þ ¼ H ^ 0 ðt, t0 ÞH
^ B, A ðt0 Þ can be identified with the self-
energy responsible for the hopping in and out of the system, whereas the last term is
responsible for imposing the initial conditions in the reservoir. It is zero if the
wavefunction is completely localized in A at t ¼ 0. The time evolution of ψ A(t) is
thus governed by a modified Hamiltonian containing an additional time-dependent
embedding operator ĤΣ[ψ A](t). The dependence on the wavefunction is written
in square brackets to stress the fact that ĤΣ[ψ A](t) is not just a simple local potential
but involves a more general non-local action.
The kernel Σ^ ðt, t0 Þ of the time integral in (60) is, in the most general case, an
explicit function of t and t0 . This is the case, for instance, when one wants to apply
this method to model molecular transport and B represents an electrode with a time-
dependent voltage bias. Evaluating (60) thus requires one to keep track of ψ A(t) for
all times up to t. This is one of the biggest drawbacks of the approach as it restricts
the propagation to short times because of storage limitations. Direct approximations
of the kernel intended to mitigate this problem have to face the fact that the kernel is
often non-analytical and highly oscillating, especially for t ! t0 [65]. However, we
note that when the Hamiltonian in B is not explicitly time-dependent, Σ^ ðt; t0 Þ
depends only on the time difference t t0 and we are left with a much easier
convolution integral.
In this last case, i.e., when the Hamiltonian in B is time-independent, an
alternative but equivalent form for the embedding operator can be obtained follow-
ing the derivation of Inglesfield [67]. In this approach we are given two
wavefunctions ψ A(r,t) and ψ B(r,t) which have equal amplitude on the surface
S separating A and B, but arbitrary derivative as illustrated in Fig. 9. Assuming
that ψ B(r,t) is a solution of the time-dependent Schr€odinger equation in B, we need
to find a closed set of equations for ψ A(r,t) to connect perfectly to ψ B(r,t) on S for
all t.
The problem is solved with the use of what in the field of partial differential
equations goes under the name of Dirichlet-to-Neumann and its inverse Neumann-
to-Dirichlet maps [68, 71, 72]. These maps allow one to transform Dirichlet
248 A.H. Larsen et al.
where ∂/∂ns denotes the directional derivative out of A and perpendicular to S, and
ð1
1 1
G0 ðrS ; r0 S ; tÞ ¼ eiεt G1 0
0 ðrS ; r S ; εÞdε: ð62Þ
2π 1
Here G1 0
0 ðrS ; r S ; εÞ is the inverse of the Green function defined by (57) evaluated on
the boundary surface S with rS , r0 S 2 S. Because G0(r,r0 ,t t0 ) depends only on time
differences it is conveniently expressed in the energy domain ε with a Fourier
transform over the time domain. Because of the presence of the δ(r rS), the
embedding operator (61) is non-zero only on the boundary surface and involves
normal and time derivatives of ψ A(r,t) over that surface.
Because of the equivalence of Ĥℰ and ĤΣ defined in (60) and (61), we refer in the
following to an embedding operator with the symbol ℰ^ ½ψ ðtÞ for simplicity. We are
A
now in the position to comment on the most characteristic features of ℰ^ ½ψ A ðtÞ. In
general, it involves complex quantities which make it an explicitly non-Hermitian
operator. This fact implies that the total number of electrons is no longer conserved
during the propagation. Furthermore, it contains a memory term in the form of a
time integral. In Frensley [73] it was postulated that transparent boundary condi-
tions should break time reversal symmetry. The presence of a memory term in (59)
turns the time propagation into a non-Markovian process and precisely breaks this
symmetry.
The extension to the many-electron case is straightforward using the same 2 2
block structure of (56) with the difference that the entries must be interpreted as
operators acting on the N-body Hilbert space. The previous steps of the derivation
hold in a completely equivalent way up to (59) and (60) provided the interacting
many-body Green function G is used in place of G0.
Formulating this in the language of TDDFT, the OQS-TDDFT theory establishes
a one-to-one connection between potential and density for non-unitary dynamics
[5–7]. The evolution from an initial state is uniquely defined if we find a way to
write the coupling with the environment as a functional νB[n] of the total density n.
Once again the equations retain the block structure of (56) with entries interpreted
as multi-index tensors, each index being associated with a Kohn–Sham orbital. The
result is a set of equations equivalent to (59) for each orbital, where the exact
embedding operator ℰ^ ½n depends on the total density of the system (i.e., in A[B)
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 249
through each orbital and the full many-body Green function G[n]. The total
embedding operator can thus be interpreted as the coupling functional νB[n] with
the environment. Obviously, this connection involving the full many-body Green
function is of little use in practical situations, but it provides a clear starting point
for further approximations.
Describing charge transfer between a system and its environment implies a modifi-
cation of the isolated Hamiltonian. In the previous section we showed how the exact
condition requires the addition of an embedding operator ℰ^ ½ψ ðtÞ that turns the
A
Hamiltonian non-Hermitian. The evaluation of such an operator can, however, be
very demanding and one needs to resort to simpler strategies.
Absorbing boundaries (ABs) or boundary absorbers are cheaper options. They
can be defined as any approximation of the form
^ AB ½ψ A ðtÞðtÞ ℰ^ ½ψ A ðtÞ
H ð63Þ
to an embedding operator such as the one given by (60) or (61). This approximation
is specific to the case where B represents the empty space and we only have to
absorb outgoing electrons. We know that ℰ^ ½ψ ðtÞ can be spatially localized on the
A
boundary surface. The absorbing boundary operator is instead generally allowed to
act on the wavefunctions over a larger region close to the boundaries, as illustrated
in Fig. 10. In the large majority of approximations, this operator is taken to be a
local potential:
Its purpose is to absorb completely any outgoing wave packet entering the region
(striped in the figure) of its support. The main goal here is to apply the absorber that
best simulates the exact embedding operator with the minimum computational cost.
From a TDDFT perspective, when we apply ĤAB to each Kohn–Sham orbital,
on top of all the approximations which might be involved in the description of the
embedding operator, we are also approximating the interaction between the system
and the environment by setting it to zero.
The absorbing properties of a boundary depend strongly on the numerical
implementation. We do not enter any specific implementation here but just point
out the fact that none of the absorbers presented in the literature are completely free
from reflections. We refer to De Giovannini et al. [74] for a recent review on the
reflection properties of members of each boundary family.
We discuss below two of the most popular families of absorbing boundaries: the
complex absorbing potentials (CAPs) and the mask function absorbers (MFAs).
250 A.H. Larsen et al.
Fig. 10 Absorbing boundaries. An absorbing boundary Hamiltonian ĤAB(t) acting on the striped
region is added to the original one Ĥ(t) to prevent reflections from the boundaries during time
propagation. The perfect absorber is the one that matches the full solution ψ(t) with ψ A(t) in the
inner (non-striped) region for all times t
We already noted above that the exact embedding potential has to be a complex
quantity to turn the Hamiltonian non-Hermitian, and the fundamental mechanism of
CAPs is precisely based on this observation. The idea was originally introduced
from a different standpoint by Neuhauser and Baer [75, 76] with the use of negative
imaginary potentials for the Schr€odinger equation. This was in connection with the
so-called optical potentials or perfectly matched layers developed for electro-
magnetic waves [77].
The effect of a CAP can be easily understood by observing the action of the
infinitesimal time evolution operator on a wavefunction
when V^ CAP is a negative imaginary potential with support on a region close to the
boundaries of A. In this case, the effect simply results in an exponential suppression
of the wavefunction in the absorbing region. In other words, the time evolution
operator associated with the non-Hermitian Hamiltonian modified with V^ CAP is
non-unitary and no longer conserves the wavefunction norm. The norm decreases if
V^ CAP is negative and increases if it is positive. In the latter case it becomes possible
to simulate charge injection, and this fact has been used to mimic reservoirs acting
as sinks or sources in the attempt to simulate electron transport [78–80].
CAPs are by no means restricted to purely imaginary potentials and there is a
huge body of literature describing their different forms and declinations [81]. We
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 251
Fig. 11 Neon and argon atom absorption cross-sections above the first ionization threshold
calculated with TDDFT and different exchange and correlation functionals: LDA, CXD-LDA
[82], PBE [83], and LB94 [54]. A CAP is introduced to reduce reflections in an energy window
centered around E ¼ 93 eV (Ne) and E ¼ 105 eV (Ar). Adapted from Crawford-Uranga et al. [84]
stress the fact that their properties strongly depend both on their mathematical form
and the specific implementation, and, without exception, they all reflect in some
energy range [74]. For practical purposes it is thus very important to ensure that the
CAP we choose for our calculations has good absorption properties in the range of
interest.
As an example, in Fig. 11 we show the absorption cross-sections for argon and
neon in the continuum, above the first ionization threshold, calculated in linear
response with TDDFT and a CAP. The CAP is chosen to minimize reflections
around E ¼ 93 eV for neon and E ¼ 105 eV for argon. The spectra are in good
agreement with the experimental ones in a fairly large range around those energies
and reflections appear as oscillations.
What is interesting about this result is that we are able to calculate a quantity
involving transitions to infinitely extended continuum states just performing a time
propagation in a bounded volume. Although at first it might seem counterintuitive,
the explanation is actually quite intuitive. In fact, we are calculating here a quantity
involving the dipole matrix element between an initial state, the ground state of our
system Ψ0, to a final state, a continuum state ΨE>0 : hΨ0 jd^ jΨE>0 i. The main
contribution to this matrix element comes from an integration over the overlap
region between the two wavefunctions and, because the ground state is bounded,
this region is safely included in A. The extent to which we manage to remove
reflection thus directly relates to the quality with which we calculate this integral
and, eventually, the quality of the absorption cross-section.
252 A.H. Larsen et al.
MFAs are an alternative formulation of CAPs. They have been employed to study a
variety of phenomena including high harmonic generation [85], electron and proton
emission [86], and above-threshold ionization [87].
They are defined by directly modifying the infinitesimal time evolution operator
with a mask function M(r) as follows:
^ M ðt þ dt, tÞ ¼ MðrÞU
U ^ ðt þ dt, tÞ: ð66Þ
ð1Þ
MCAP ðrÞ ¼ eiV CAP ðrÞdt : ð67Þ
The mask function can thus be a complex function, and becomes real when V^ CAP is
purely imaginary. The inverse relation can be obtained in a similar way, and to first
order it reduces to
ð1Þ i
V MFA ðrÞ ¼ ln½MðrÞ: ð68Þ
dt
In De Giovannini et al. [74] it was shown that the first-order relations above, for a
given pair of CAP and MFA, yield reflection properties in excellent agreement with
each other.
One important feature of the MFA approach is that by multiplying M(r) and
1 M(r) by a wavefunction it is possible to split its propagation in two different
components moving in separate regions. This property is fundamental for split-
domain propagation schemes initially derived in Chelkowski et al. [88] and Grobe
et al. [89] and later extended to the study of electron photoemission with TDDFT in
De Giovannini et al. [90]. We return to this point in Sect. 6.3.
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 253
for left ψ θ ðr; tÞ and right states ψ θ(r,t), where ĤECS θ (t) represents the scaled
Hamiltonian. Extrapolating from the discussion in Sect. 4.4 we can interpret (69)
as imposing purely outgoing boundary conditions and (70) as the incoming
counterpart.
In the theory of complex scaling, the calculation of the expectation value of an
observable Ôθ on the scaled path as of (10) involves left and right states on an
equal footing. This extends to the time-dependent case with the requirement of
having both left and right states at the same time to calculate Ôθ. Hence, we need,
in principle, to solve (69) and (70) simultaneously.
The fact that the scaling path lies exactly on the real axis in a certain region
simplifies the equations. In fact, on the real axis, left and right states are complex
conjugates: ψ θ ðr; tÞ ½ψ θ ðr; tÞ* ¼ ψ θ ðr; tÞ* for r in the interior region. This is
particularly true when the system contains only a local potential and the propa-
gation is initialized with a state localized in the unscaled region at t ¼ 0 and
propagating outward. If we restrict ourselves to observables in the unscaled region
and we want to describe a purely outgoing process, we resolve to use the right state
ψ θ(r,t) only. This state can be obtained by propagating (69) which involves only
right states [38]. Following this, most applications of exterior complex scaling are
limited to a use with the decaying right states and observables evaluated in the
unscaled region.
Equation (69) perfectly describes problems where imposing purely outgoing
boundary conditions represents an exact condition similar to that, for example, in
ionization processes. In those cases it can be regarded as equivalent to a transparent
boundary condition described with a Green function. Here, because we are dealing
with purely outgoing conditions, we should note that the title of perfect absorber is
254 A.H. Larsen et al.
more appropriate than that of transparent boundary, because electrons can flow in
only one direction.
However, we note an important difference between the two approaches.
Whereas the Green function embedding defines the exact matching conditions at
the boundary of a finite volume A, the scaled (69) acts on a wavefunction defined in
the full space A[B. This makes the size of the simulation box a weakness in
numerical simulations if a wave is capable of reaching the end of the box. The
scaling transformation imposes an asymptotic form which can be efficiently cap-
tured by exponential functions eαr. By employing a finite element approach with
an element at infinity which captures the exponential tail, it was numerically shown
by Scrinzi [38] that exterior complex scaling indeed provides perfectly absorbing
conditions for numerical precision.
Restricting (69) to A otherwise implies a truncation which irrevocably breaks its
perfect properties. In this case the scaling transformation reduces to an absorbing
boundary which can be regarded as a simple CAP and, as such, presents reflections
[74, 91]. We should mention that the use of (69) restricted to A in combination with
a smooth exterior complex scaling in the literature has been going under the
misleading name of reflection-free CAP, in spite of presenting a certain degree of
reflection [39, 91–93].
In the context of TDDFT, exterior complex scaling has been applied purely as an
absorbing boundary [53, 94].
6 Electron Photoemission
We focus here on the approaches that can be employed in the description of multi-
electron ionization initiated by external electromagnetic fields within TDDFT. As
in previous sections, we are interested only in electronic processes, neglecting any
ionic motion, and we restrict ourselves to the class of methods that requires
knowledge of the wavefunctions only on a bounded region of space A much as in
Fig. 8.
We are interested in the family of problems characterized by time-dependent
electronic Hamiltonians with the structure
^ 1 A ðtÞ 2
H ðtÞ ¼ i∇ þ vext þ vee ; ð71Þ
2 c
where νee is the electron–electron Coulomb interaction, νext is the external potential
which generally consists of a static potential produced by the nuclei, A ðtÞ is the
vector potential of the external field, and c is the speed of light. In writing (71) we
implied the choice of the velocity gauge to describe the action of the field. The
associated electric field can easily be obtained as a time derivative:
ℰðtÞ ¼ ∂t A ðtÞ. Typically, one would want to perform a simulation by choosing
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 255
a vector potential representing one or more laser pulses, then investigate the
induced dynamic.
Ionization takes place whenever the field is capable of inducing a bound-to-
continuum transition, resulting in electrons escaping with a given kinetic energy.
Calculation of observables characterizing these ionized electrons is at the center of
our interest here.
To some extent we already approached this problem in Sect. 5. In fact, total
ionization can be naturally described using only information contained in a
bounded volume A surrounding our system. The total number of electrons
contained in A can be simply calculated from the knowledge of the time-dependent
density as
ð
N ðtÞ ¼ nðr; tÞdr: ð72Þ
A
Combined with the use of one of the boundary conditions described above, (72)
implements a practical strategy for the calculation of N(t). The total ionization
probability, i.e., the probability of ejecting an electron in the long-time limit, is thus
naturally obtained using only quantities defined in A as
N N ðtÞ
P ¼ lim ; ð73Þ
t!1 N
where N represents the total number of electrons in the system before ionization.
Being a direct functional of the density, P is an exact quantity within TDDFT and
does not present any further approximation besides the one involved with the use of
the boundary conditions.
In many situations the quantities containing relevant physical information are
more complex objects than the simple total ionization probability, and one would
wish to access differential probabilities with respect to energy or momentum:
3
∂P ∂ P
Pð EÞ ¼ , PðkÞ ¼ : ð74Þ
∂E ∂kx ∂k y ∂kz
The calculation of these observables within TDDFT is, however, not as straight-
forward as the evaluation of P.
The first reason is the intrinsic complexity of the ionization process already with
only one electron. There are situations, especially when strong laser fields are
involved, where the electron dynamics are so complex that one has to propagate
explicitly the wavefunction in time to account for the process. In principle, the
differential probabilities can then be obtained by projecting the scattering
wavefunction Ψs(t) onto the appropriate set of scattering wavefunctions ϕE as
256 A.H. Larsen et al.
2
PðEÞ ¼ lim ϕE Ψs ðtÞ with E > 0: ð75Þ
t!1
Besides the issues related to the correct evaluation of the projecting set, owing to
their large spatial extension, the propagation of the total electronic wavefunction at
long times can be practically performed only for highly symmetrical systems, such
as atoms and small molecules, or for short times. Alternative approaches, such as
those we describe below, involve the knowledge of the wavefunction only in a
bounded region of space.
The second reason has to do with the multi-electron nature of ionization at the
TDDFT level. In fact, whereas the connection between P and the total density is
explicitly known, the differential quantities (74) cannot be easily expressed in terms
of the density. The derivation and the use of appropriate density functionals to
describe P(E) and P(k) from (74) are thus important and have to take into account
in our model.
In the following we discuss the methods that have been developed to tackle these
problems numerically.
1 X N
PrS ðEÞ / pffiffiffi jψe i ðrS ; EÞj2 : ð77Þ
E i¼1
Fig. 12 Electron
photoemission with the
sampling method. The
energy-resolved
photoelectron probability is
calculated by recording the
time evolution of the
wavefunction at the points
marked in red
ð
Pð EÞ / PrS ðEÞdrS : ð78Þ
S
In practical calculations, the integral over the S is of course still discretized, and
open boundary conditions, often in the form of absorbing boundaries, may be
employed. The choice of the absorber must be such that it efficiently removes
reflected wave packets in the energy range where photoelectrons are simulated.
pffiffiffi
In the original paper, (78) was introduced without the energy density factor 1= E
and the surface integral [95]. The results where thus, in general, strongly dependent
on the choice of the sample point rs and applicable only in the situations where the
electrons are emitted as s-waves, hence not presenting any angular fluctuations. Even
taking into account the integral over S, the method is not free from problems. It
requires S to be placed at a distance from the parent system such that two conditions
are fulfilled: (i) the electronic wave packets can be considered to be composed of
outgoing waves only, and (ii) photoelectrons must hit the surface at a time for which
the external field is turned off. It was later realized that a time- and energy-dependent
phase eiΦ(E,t) must be included in the integral (76) to account for the wrong kinetic
energy reference when the external field is still active [96].
Although this method is straightforward and easy to implement in existing
TDDFT codes, the above drawbacks render it of limited use in many interesting
physical situations, especially when strong laser fields are employed.
This method is based on the idea that differential photoemission probabilities (74)
can be calculated by recording the electron flux through a surface. It was originally
introduced by Scrinzi and co-workers in Caillat et al. [97] in the context of multi-
configuration Hartree–Fock and then further developed in Scrinzi [98] and Tao and
Scrinzi [99] for one- and two-electron problems. Although no applications in the
context of TDDFT have been attempted so far, it presents an interesting approach
for the calculation of photoelectron differential probabilities in bounded volumes.
Let us consider the case of a one-electron system governed by a Hamiltonian
Ĥ(t) (71) such that at large distances it matches an exactly solvable one Ĥv(t),
258 A.H. Larsen et al.
^ ðt Þ ¼ H
H ^ v ðt Þ for jrj r s and all t; ð79Þ
as shown in Fig. 13. We are assuming here for simplicity that the surface S sepa-
rating the Hamiltonians is spherical, but what follows can be easily extended to a
generic surface. If we consider the case of a short range external potential νext(r) ¼
0 for |r| > rs, Ĥv is the Volkov Hamiltonian, i.e., the Hamiltonian governing the
motion of free electrons in an external field:
2
^ v ðtÞ ¼ 1 i∇ A ðtÞ :
H ð80Þ
2 c
Provided the external field has no spatial dependence, i.e., A ðtÞ is constant in space,
the associated TDSE can be solved exactly. The solutions can then be expressed as
plane waves with an additional time and momentum-dependent phase:
ðt
1 ikr iΦðk;tÞ 1 A ðt Þ 2 0
ϕk ðr; tÞ ¼ 3 e e , Φðk; tÞ ¼ k dt : ð81Þ
ð2π Þ2 2 1 c
Let us imagine the situation where a laser pulse ionizes our system. In the long time
limit t > T, some time after the field as been turned off, A ðt > T Þ ¼ 0, the
electronic configuration is described by a scattering wavefunction which can be
partitioned into bound and scattering components,
which are approximately localized in the bound and unbound regions A and B of
Fig. 8. The quality of this approximation is ultimately connected to rs and T, and the
time that it takes the slowest components of the scattering wave packet ΨB(r,t) to
cross S.
In order to calculate the emission amplitude, we just need to evaluate the
projection of Ψ(r,t) over the asymptotic wavefunctions ϕk(r) as in (75). The
information about the scattering process is contained only in ΨB(r,t). Because
ΨB(r,t) is exponentially vanishing in A for t T, we can write the emission
amplitude as
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 259
Sðk; T Þ ¼ hϕk ðT Þjθðr; r s ÞjΨðT Þi ¼ ϕk ðT ÞΨB ðT Þ ; ð83Þ
Equation (83) can be written as a time integral of the derivative of S(k,t). Combined
with the Ehrenfest theorem and the fact that both states in (83) evolve with the same
Hamiltonian Ĥv within the support of θ(r,rs) we obtain
ðT
Sðk; T Þ ¼ i ϕ k ðt Þ H^ v , θðr; r s Þ ΨðtÞ dt: ð85Þ
0
a b
Fig. 14 The main traits of the mask method. In this method, photoelectrons are time propagated
with a mixed real and momentum-space representation. A red striped area identifies the region
where the matching between the two representations is performed. (a) Spatial and Hamiltonian
partitioning. (b) The mask function. (c) A wavefunction Ψ is split into two parts, ΨA and ΨB, using
the mask function
This mask method is based on the idea that the photoelectron emission probability
can be calculated by explicitly propagating the ionized electron wave packets as a
superposition of plane waves. The problem of matching inner and outer solutions is
solved here with the aid of a mask function [90]. This approach as been successfully
employed within TDDFT in situations involving atoms and molecules under the
influence of a variety of external fields ranging from strong and weak laser fields
[90] to pump and probe configurations [100, 101].
We begin here by introducing the equations governing time propagation for the
single-electron case and then turn to the many-electron one. Let us consider the case
where the Hamiltonian Ĥ(t) is of short range and satisfies the asymptotic condition
(79), i.e., it coincides with Ĥv(t) for |r| rs as illustrated in Fig. 14a.
As discussed in the previous section, in the long-time limit of an ionization
process, we can assume the electronic wavefunction splits into two spatially
separated parts, namely the bound and the scattering parts (82). A practical way
to implement this splitting for a generic time t is to use a mask function M(r) similar
to what was discussed in Sect. 5.5:
We consider here the case where M(r) is a continuous function equal to 1 in an inner
part of A, where |r| rs, equal to 0 in B, and smoothly decays over the intermediate
region as shown in Fig. 14b. The splitting defined with this procedure is smooth and
the wavefunctions ΨA(r,t) and ΨB(r,t) are not sharply separated but are allowed to
overlap in the region where the mask decays to zero, as illustrated in Fig. 14c. The
mask function M(r) is such that this overlap region is entirely contained in A.
The solution of the TDSE associated with the full Hamiltonian Ĥ(t) in the whole
space A [ B can be formally written as a set of coupled equations:
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 261
j ΨA ð t 0 Þ i ¼ M^ ^ 0
U ðt ;
tÞ½jΨ0A ðtÞiþjΨB ðtÞi
0
j ΨB ð t Þ i ¼ 1 M ^ U ^ ðt ; tÞ½jΨA ðtÞi þ jΨB ðtÞi ; ð89Þ
and imposing the boundary condition |ΨB(t ¼ 0)i ¼ 0. Here the mask operator is
given by hrjPM ^ jr0 i ¼ MðrÞδðr r0 Þ.
Owing to the asymptotic condition (79) on the Hamiltonian, |ΨB(t)i evolves
under the action of Ĥv defined in (80). In what follows we indicate with Uv(t0 ,t) the
evolution operator associated with Ĥv. Because Ĥv is diagonal in momentum, the
action of Uv(t0 ,t) is easily described in this space. It is thus convenient to expand the
equations for |ΨB(t)i using plane waves: hr|ki ¼ (2π)3/2exp{ikr}. On the other
hand, owing to the presence in Ĥ(t) of V(r), which has an explicit dependence on r,
the equations for |ΨA(t)i are better solved in real space. The use of a mixed real and
momentum space representation seems the more natural one for the problem.
Using a mixed representation we can integrate (89) by recursively applying the
discrete time evolution operator U(Δt)^ ^ + Δt,t) as
U(t
(
hrjΨA ðt þ ΔtÞi ¼ rM^U^ ðΔtÞΨA ðtÞ þ rM ^ ðΔtÞΨB ðtÞ
^U
v
hkjΨB ðt þ ΔtÞi ¼ k 1 M^ U ^ ðΔtÞΨA ðtÞ þ k 1 M
^ U ^ v ðΔtÞΨB ðtÞ ;
ð91Þ
with initial condition hk|ΨB(t ¼ 0)i ¼ 0. These equations can be written in a closed
form for hr|ΨA(t)i and hk|ΨB(t)i by including the additional set
8
>
> rM^U ^ ðΔtÞΨA ðtÞ ¼ MðrÞ rU
ð
^ ðΔtÞΨA ðtÞ
>
>
>
> M
^U ^ v ðΔtÞΨB ðtÞ ¼ MðrÞ rk kU ^ v ðΔtÞΨB ðtÞ dk
>
> r
< ð
:
> 1M ^ ^ ð Δt Þ ΨA ðtÞ ¼ kr ½1 MðrÞ rU ^ ðΔtÞΨA ðtÞ dr
>
>
k U
>
> ð
>
>
>
: k 1 M ^ U ^ v ðΔtÞΨB ðtÞ ¼ kU ^ v ðΔtÞΨB ðtÞ kr rM ^U ^ v ðΔtÞΨB ðtÞ dr
ð92Þ
The equations in (92) have an intuitive interpretation in terms of electron flow. The
first and second equations account, respectively, for electrons leaving and returning
to A. The third equation is responsible for introducing charge in B whereas the
fourth is composed of a term of pure time evolution minus a term balancing
the backward flow of the second equation. In the limit of infinitesimal steps Δt¼dt,
the complete set defined by (91) and (92) is equivalent to (89), and it fully accounts
262 A.H. Larsen et al.
for the description of outgoing and incoming particle flows. In this respect it
directly relates to the exact boundary conditions of Sect. 5.1.
Once (91) and (92) are propagated up to a time T such that the external field has
vanished and the bound and scattering components of Ψ(t) are well separated, the
momentum-resolved probability can be obtained just by taking the square modulus
of the wavefunction in B: P(k) ¼ |hk|ΨB(T )i|2. This definition is consistent with that
of the surface flux method, noting that, at time T, Volkov and plane waves differ
only by a phase |ϕk(T )i ¼ |ki exp{iΦ(k,T )}. Extending the mask method to the case
of infinite-range potentials incurs the same approximation errors as in the flux
method with Volkov states.
The extension of the method to the many-electron case, on the other hand, is less
trivial. It can be derived from a phase-space standpoint given the interpretation of
the Wigner transform of the one-body density matrix ρ(r,r0 ,t),
ð
eiks s s R ¼ ðr þ r0 Þ=2
W ðR; k; tÞ ¼ ρ R þ , R , t ds with ; ð93Þ
ð2π Þ2
3
2 2 s ¼ r r0
The connection with TDDFT can be established using the Kohn–Sham one-body
density matrix
X
N=2
ρKS ðr; r0 ; tÞ ¼ 2 ψ *i ðr; tÞψ i ðr0 ; tÞ ð95Þ
i¼1
in (93) to calculate the Wigner distribution. For simplicity we assume here a closed-
shell system where each orbital ψ i(r,t) is doubly occupied. There is no fundamental
restriction in extending to the more general case where spin polarization is taken
into account.
We now assume that it is possible to establish an approximate asymptotic
connection, similar to (79), between the Kohn–Sham Hamiltonian ĤKS(t) and
Ĥv(t) after a certain radius |r| > rs (see De Giovannini et al. [90]). Under this
assumption we can partition each orbital according to (88) and use (91) and (92)
to propagate them in time. By plugging the Wigner distribution obtained from
ρKS(r,r0 ,t) into (94) we then obtain that the momentum-resolved probability distri-
bution can be expressed as an sum of orbital densities
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 263
a b
Fig. 15 Ionization of randomly oriented N2 molecules by a strong infrared laser field. Angle and
energy-resolved photoelectron probability P(E,θ) (log scale) obtained from the experiment [102]
(a) and with the theory (b) using the mask method of (91) and (92). The laser is a six-cycle pulse
with wavelength λ ¼ 750 nm and intensity I ¼ 4.3 1013 W/cm2. Adapted from De Giovannini
et al. [90]
X
N=2
PðkÞ ¼ 2 kψ i, B ðT Þ 2 : ð96Þ
i¼1
P(E,τ)
2.9 60
π π *
a z y b
z -3
2.7 40
10
2.6 30
1
σ 2.5 20
π -4
y 10 2.4 10
0
τp 5 9 τm 13 τp 5 9 τm 13
τ (fs) τ (fs)
8 ^ ^
ΨA ðtÞ ¼ MðrÞ rU
^ ðΔtÞΨA ðtÞ
>
> r M U ð Δt Þ
>
>
< rM ^U^ v ðΔtÞΨB ðtÞ ¼ 0
ð
>
> k 1M ^ U^ ð Δt Þ ΨA ðtÞ ¼ kr ½1 MðrÞ rU ^ ðΔtÞΨA ðtÞ dr : ð97Þ
>
>
:
k 1M ^ U ^ v ðΔtÞΨB ðtÞ ¼ kU ^ v ðΔtÞΨB ðtÞ
7 Summary
change from real space to momentum space representation. This scheme allows for
the description of a wide range of processes and time resolved pump–probe
spectroscopies.
References
15. Newns DM (1969) Self-consistent model of hydrogen chemisorption. Phys Rev 178:
1123–1135. doi:10.1103/PhysRev.178.1123, http://link.aps.org/doi/10.1103/PhysRev.178.1123
16. Gellene GI (1995) Resonant states of a one-dimensional piecewise constant potential. J Chem
Educ 72(11):1015. doi:10.1021/ed072p1015, http://dx.doi.org/10.1021/ed072p1015
17. Siegert AJF (1939) On the derivation of the dispersion formula for nuclear reactions.
Phys Rev 56:750–752. doi:10.1103/PhysRev.56.750, http://link.aps.org/doi/10.1103/
PhysRev.56.750
18. Hatano N, Sasada K, Nakamura H, Petrosky T (2008) Some properties of the resonant state in
quantum mechanics and its computation. Prog Theor Phys 119(2):187–222. doi:10.1143/
PTP.119.187, http://ptp.oxfordjournals.org/content/119/2/187.abstract
19. Aguilar J, Combes J (1971) A class of analytic perturbations for one-body Schr€ odinger
Hamiltonians. Commun Math Phys 22:269–279. doi:10.1007/BF01877510, http://dx.doi.
org/10.1007/BF01877510
20. Balslev E, Combes JM (1971) Spectral properties of many-body Schr€ odinger operators with
dilatation-analytic interactions. Commun Math Phys 22(4):280–294
21. Simon B (1973) Resonances in n-body quantum systems with dilatation analytic potentials
and the foundations of time-dependent perturbation theory. Ann Math 97:247–274
22. Simon B (1979) The definition of molecular resonance curves by the method of
exterior complex scaling. Phys Lett A 71(2):211–214
23. Ho Y (1983) The method of complex coordinate rotation and its applications to atomic
collision processes. Phys Rep 99(1):1–68. doi:10.1016/0370-1573(83)90112-6, http://www.
sciencedirect.com/science/article/pii/0370157383901126
24. McCurdy CW, Baertschy M, Rescigno TN (2004) Solving the three-body Coulomb breakup
problem using exterior complex scaling. J Phys B At Mol Opt 37(17):R137, http://stacks.iop.
org/0953-4075/37/i ¼ 17/a ¼ R01
25. Moiseyev N (1998) Quantum theory of resonances: calculating energies, widths and cross-
sections by complex scaling. Phys Rep 302(5–6):212–293. doi:10.1016/S0370-1573(98)
00002-7, http://www.sciencedirect.com/science/article/pii/S0370157398000027
26. Reinhardt WP (1982) Complex coordinates in the theory of atomic and molecular structure
and dynamics. Annu Rev Phys Chem 33(1):223–255. doi:10.1146/annurev.pc.33.100182.
001255, http://www.annualreviews.org/doi/abs/10.1146/annurev.pc.33.100182.001255
27. Simon B (1978) Resonances and complex scaling: a rigorous overview. Int J Quantum Chem
14(4):529–542. doi:10.1002/qua.560140415, http://dx.doi.org/10.1002/qua.560140415
28. Cerjan C, Hedges R, Holt C, Reinhardt WP, Scheibner K, Wendoloski JJ (1978) Complex
coordinates and the Stark effect. Int J Quantum Chem 14(4):393–418. doi:10.1002/qua.
560140408, http://dx.doi.org/10.1002/qua.560140408
29. Herbst IW (1979) Dilation analyticity in constant electric field. Commun Math Phys 64(3):
279–298. doi:10.1007/BF01221735
30. Herbst IW, Simon B (1978) Stark effect revisited. Phys Rev Lett 41:67–69. doi:10.1103/
PhysRevLett.41.67, http://link.aps.org/doi/10.1103/PhysRevLett.41.67
31. Scrinzi A, Piraux B (1998) Two-electron atoms in short intense laser pulses. Phys Rev A 58:
1310–1321. doi:10.1103/PhysRevA.58.1310, http://link.aps.org/doi/10.1103/PhysRevA.58.1310
32. Junker BR (1983) Complex virial theorem and complex scaling. Phys Rev A 27:2785–2789.
doi:10.1103/PhysRevA.27.2785, http://link.aps.org/doi/10.1103/PhysRevA.27.2785
33. Moiseyev N, Friedland S, Certain PR (1981) Cusps, θ trajectories, and the complex virial
theorem. J Chem Phys 74(8):4739–4740. doi:10.1063/1.441624, http://scitation.aip.org/con-
tent/aip/journal/jcp/74/8/10.1063/1.441624
34. McCurdy CW (1980) Complex-coordinate calculation of matrix elements of the resolvent of
the Born–Oppenheimer Hamiltonian. Phys Rev A 21:464–470. doi:10.1103/PhysRevA.21.
464, http://link.aps.org/doi/10.1103/PhysRevA.21.464
35. McCurdy CW, Rescigno TN (1978) Extension of the method of complex basis functions to
molecular resonances. Phys Rev Lett 41:1364–1368. doi:10.1103/PhysRevLett.41.1364,
http://link.aps.org/doi/10.1103/PhysRevLett.41.1364
268 A.H. Larsen et al.
36. Moiseyev N, Corcoran C (1979) Autoionizing states of H2 and H2 using the complex-
scaling method. Phys Rev A 20:814–817. doi:10.1103/PhysRevA.20.814, http://link.aps.org/
doi/10.1103/PhysRevA.20.814
37. Morgan JD, Simon B (1981) The calculation of molecular resonances by complex scaling.
J Phys B At Mol Opt 14(5):L167
38. Scrinzi A (2010) Infinite-range exterior complex scaling as a perfect absorber in time-
dependent problems. Phys Rev A 81(5):053845
39. Moiseyev N (1999) Derivations of universal exact complex absorption potentials by the
generalized complex coordinate method. J Phys B At Mol Opt 31(7):1431–1441
40. Krylstedt P, Carlsund C, Elander N (1989) On the calculation of electron–atom collision
properties using exterior complex dilatated s-matrix expansions. J Phys B At Mol Opt 22(7):
1051, http://stacks.iop.org/0953-4075/22/i ¼ 7/a ¼ 014
41. Rescigno TN, Baertschy M, Byrum D, McCurdy CW (1997) Making complex scaling work
for long-range potentials. Phys Rev A 55:4253–4262. doi:10.1103/PhysRevA.55.4253,
http://link.aps.org/doi/10.1103/PhysRevA.55.4253
42. Scrinzi A, Elander N (1993) A finite element implementation of exterior complex scaling for
the accurate determination of resonance energies. J Chem Phys 98(5):3866–3875. doi:10.
1063/1.464014, http://scitation.aip.org/content/aip/journal/jcp/98/5/10.1063/1.464014
43. Simons J (1980) The complex coordinate rotation method and exterior scaling: a simple
example. Int J Quantum Chem 18(S14):113–121. doi:10.1002/qua.560180814, http://dx.doi.
org/10.1002/qua.560180814
44. Kar S, Ho YK (2009) Isotope shift for the 1De autodetaching resonance in H and D. J Phys
B At Mol Opt 42(5):055001, http://stacks.iop.org/0953-4075/42/i ¼ 5/a ¼ 055001
45. Scrinzi A, Geissler M, Brabec T (1999) Ionization above the Coulomb barrier. Phys Rev Lett
83:706–709. doi:10.1103/PhysRevLett.83.706, http://link.aps.org/doi/10.1103/
PhysRevLett.83.706
46. McCurdy CW, Rescigno TN, Davidson ER, Lauderdale JG (1980) Applicability of self-
consistent field techniques based on the complex coordinate method to metastable electronic
states. J Chem Phys 73(7):3268–3273. doi:10.1063/1.440522, http://scitation.aip.org/content/
aip/journal/jcp/73/7/10.1063/1.440522
47. Samanta K, Yeager DL (2008) Investigation of 2P Be shape resonances using a quadrati-
cally convergent complex multiconfigurational self-consistent field method. J Phys Chem B
112(50):16214–16219. doi:10.1021/jp806998n, http://dx.doi.org/10.1021/jp806998n
48. Zdánská PR, Moiseyev N (2005) Hartree–Fock orbitals for complex-scaled configuration
interaction calculation of highly excited Feshbach resonances. J Chem Phys 123(19):194105.
doi:10.1063/1.2110169, http://scitation.aip.org/content/aip/journal/jcp/123/19/10.1063/
1.2110169
49. Larsen AH, Whitenack DL, De Giovannini U, Wasserman A, Rubio A (2013) Stark ion-
ization of atoms and molecules within density functional resonance theory. J Phys Chem Lett
4:2734–2738
50. Whitenack DL, Wasserman A (2011) Density functional resonance theory of unbound
electronic systems. Phys Rev Lett 107(16):163002
51. Wasserman A, Moiseyev N (2007) Hohenberg-Kohn theorem for the lowest-energy reso-
nance of unbound systems. Phys Rev Lett 98:093003. doi:10.1103/PhysRevLett.98.093003,
http://link.aps.org/doi/10.1103/PhysRevLett.98.093003
52. Perdew JP, Wang Y (1992) Accurate and simple analytic representation of the electron-gas
correlation energy. Phys Rev B 45:13244–13249. doi:10.1103/PhysRevB.45.13244,
http://link.aps.org/doi/10.1103/PhysRevB.45.13244
53. Telnov DA, Sosnova KE, Rozenbaum E, Chu SI (2013) Exterior complex scaling method in
time-dependent density-functional theory: multiphoton ionization and high-order-harmonic
generation of Ar atoms. Phys Rev A 87(5):053406
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 269
54. van Leeuwen R, Baerends EJ (1994) Exchange–correlation potential with correct asymptotic
behavior. Phys Rev A 49:2421–2431. doi:10.1103/PhysRevA.49.2421, http://link.aps.org/
doi/10.1103/PhysRevA.49.2421
55. Ammosov MV, Delone NB, Krainov VP (1986) Tunnel ionization of complex atoms and
atomic ions in a varying electromagnetic-field. Zh Éksp Teor Fiz 91:2008–2013
56. Parker SD, McCurdy C (1989) Propagation of wave packets using the complex basis
function method. Chem Phys Lett 156(5):483–488. doi:10.1016/S0009-2614(89)87316-6,
http://www.sciencedirect.com/science/article/pii/S0009261489873166
57. Bengtsson J, Lindroth E, Selstø S (2008) Solution of the time-dependent Schr€ odinger
equation using uniform complex scaling. Phys Rev A 78:032502. doi:10.1103/PhysRevA.
78.032502, http://link.aps.org/doi/10.1103/PhysRevA.78.032502
58. Bengtsson J, Lindroth E, Selstø S (2012) Wave functions associated with time-dependent,
complex-scaled Hamiltonians evaluated on a complex time grid. Phys Rev A 85:013419.
doi:10.1103/PhysRevA.85.013419, http://link.aps.org/doi/10.1103/PhysRevA.85.013419
59. Gilary I, Fleischer A, Moiseyev N (2005) Calculations of time-dependent observables in
non-Hermitian quantum mechanics: the problem and a possible solution. Phys Rev A 72:
012,117. doi:10.1103/PhysRevA.72.012117, http://link.aps.org/doi/10.1103/PhysRevA.72.012117
60. Garcı́a-Moliner F, Flores F (2009) Introduction to the theory of solid surfaces.
Cambridge University Press, Cambridge
61. Kudrnovský J, Drchal V, Turek I, Weinberger P (1994) Magnetic coupling of interfaces: a
surface-Green’s-function approach. Phys Rev B 50:16105–16108. doi:10.1103/PhysRevB.
50.16105, http://link.aps.org/doi/10.1103/PhysRevB.50.16105
62. Boucke K, Schmitz H, Kull HJ (1997) Radiation conditions for the time-dependent
Schr€odinger equation: application to strong-field photoionization. Phys Rev A 56(1):763–771
63. Ermolaev A, Puzynin I, Selin A, Vinitsky S (1999) Integral boundary conditions for the time-
dependent Schr€odinger equation: atom in a laser field. Phys Rev A 60(6):4831–4845
64. Hellums J, Frensley W (1994) Non-Markovian open-system boundary conditions for the
time-dependent Schr€odinger equation. Phys Rev B 49(4):2904–2906
65. Kurth S, Stefanucci G, Almbladh CO, Rubio A, Gross EKU (2005) Time-dependent quantum
transport: a practical scheme using density functional theory. Phys Rev B 72(3):035308
66. Inglesfield JE (2001) Embedding at surfaces. Comput Phys Commun 137(1):89–107
67. Inglesfield JE (2011) A time-dependent embedding calculation of surface electron emission.
J Phys Condens Matter 23(30):305004
68. Antoine X, Arnold A, Besse C, Ehrhardt M, Schädle A (2008) A review of transparent and
artificial boundary conditions techniques for linear and nonlinear Schr€ odinger equations.
Commun Comput Phys 4:729–796
69. Inglesfield JE (1981) A method of embedding. J Phys C Solid State 14(26):3795–3806
70. Inglesfield J (2008) Time-dependent embedding. J Phys Condens Matter 20:095215
71. Ehrhardt M (1999) Discrete transparent boundary conditions for general Schr€ odinger-type
equations. VLSI Des 9(4):325–338
72. Szmytkowski R, Bielski S (2004) Dirichlet-to-Neumann and Neumann-to-Dirichlet embed-
ding methods for bound states of the Schr€odinger equation. Phys Rev A 70(4):042103
73. Frensley W (1990) Boundary conditions for open quantum systems driven far from equili-
brium. Rev Mod Phys 62(3):745–791
74. De Giovannini U, Larsen AH, Rubio A (2015) Modeling electron dynamics coupled to
continuum states in finite volumes. Eur Phys J B 88(3):56. doi:10.1140/epjb/e2015-50808-0
75. Neuhauser D, Baer M (1989) The application of wave packets to reactive atom–diatom
systems: a new approach. J Chem Phys 91(8):4651–4657
76. Neuhauser D, Baer M (1989) The time-dependent Schr€ odinger equation: application of
absorbing boundary conditions. J Chem Phys 90(8):4351
77. Berenger JP (1994) A perfectly matched layer for the absorption of electromagnetic waves.
J Comput Phys 114(2):185–200
270 A.H. Larsen et al.
78. Elenewski JE, Chen H (2014) Real-time transport in open quantum systems from
PT-symmetric quantum mechanics. Phys Rev B 90(8):085104
79. Varga K, Pantelides S (2007) Quantum transport in molecules and nanotube devices.
Phys Rev Lett 98(7):076804
80. Wibking BD, Varga K (2012) Quantum mechanics with complex injecting potentials.
Phys Lett A 376(4):365–369
81. Muga J, Palao JP, Navarro B, Egusquiza IL (2004) Complex absorbing potentials. Phys Rep
395(6):357–426
82. Andrade X, Aspuru-Guzik A (2011) Prediction of the derivative discontinuity in density
functional theory from an electrostatic description of the exchange and correlation potential.
Phys Rev Lett 107(18):183002
83. Perdew JP, Burke K, Ernzerhof M (1996) Generalized gradient approximation made simple.
Phys Rev Lett 77(18):3865–3868
84. Crawford-Uranga A, De Giovannini U, Räsänen E, Oliveira MJT, Mowbray DJ,
Nikolopoulos GM, Karamatskos ET, Markellos D, Lambropoulos P, Kurth S, Rubio A
(2014) Time-dependent density-functional theory of strong-field ionization of atoms by
soft X-rays. Phys Rev A 90(3):033412
85. Krause J, Schafer K, Kulander K (1992) Calculation of photoemission from atoms subject to
intense laser fields. Phys Rev A 45(7):4998–5010
86. Kulander K, Mies F, Schafer K (1996) Model for studies of laser-induced nonlinear processes
in molecules. Phys Rev A 53(4):2562–2570
87. Lein M, Marangos J, Knight P (2002) Electron diffraction in above-threshold ionization of
molecules. Phys Rev A 66(5):051404R
88. Chelkowski S, Foisy C, Bandrauk AD (1998) Electron–nuclear dynamics of multiphoton H2+
dissociative ionization in intense laser fields. Phys Rev A 57(2):1176–1185
89. Grobe R, Haan S, Eberly J (1999) A split-domain algorithm for time-dependent multi-
electron wave functions. Comput Phys Commun 117(3):200–210
90. De Giovannini U, Varsano D, Marques MAL, Appel H, Gross EKU, Rubio A (2012) Ab initio
angle- and energy-resolved photoelectron spectroscopy with time-dependent density-func-
tional theory. Phys Rev A 85:062515
91. Shemer O, Brisker D, Moiseyev N (2005) Optimal reflection-free complex absorbing poten-
tials for quantum propagation of wave packets. Phys Rev A 71(3):032716
92. McCurdy CW, Stroud C, Wisinski M (1991) Solving the time-dependent Schr€ odinger
equation using complex-coordinate contours. Phys Rev A 43(11):5980–5990
93. Riss UV, Meyer HD (1995) Reflection-free complex absorbing potentials. J Phys B At Mol
Opt 28(8):1475–1493
94. Sosnova KE, Telnov DA, Rozenbaum EB, Chu SI (2014) Exterior complex scaling method in
TDDFT: HHG of Ar atoms in intense laser fields. J Phys Conf Ser 488(1):012022
95. Pohl A, Reinhard PG, Suraud E (2000) Towards single-particle spectroscopy of small metal
clusters. Phys Rev Lett 84(22):5090–5093
96. Dinh PM, Romaniello P, Reinhard PG, Suraud E (2013) Calculation of photoelectron spectra:
a mean-field-based scheme. Phys Rev A 87(3):032514
97. Caillat J, Zanghellini J, Kitzler M, Koch O, Kreuzer W, Scrinzi A (2005) Correlated
multielectron systems in strong laser fields: a multiconfiguration time-dependent Hartree–
Fock approach. Phys Rev A 71(1):012712
98. Scrinzi A (2012) t-SURFF: fully differential two-electron photo-emission spectra. New J
Phys 14(8):085008
99. Tao L, Scrinzi A (2012) Photo-electron momentum spectra from minimal volumes: the time-
dependent surface flux method. New J Phys 14(1):013021
100. Crawford-Uranga A, De Giovannini U, Mowbray DJ, Kurth S, Rubio A (2014) Modelling the
effect of nuclear motion on the attosecond time-resolved photoelectron spectra of ethylene.
J Phys B At Mol Phys 47(12):124018
Dynamical Processes in Open Quantum Systems from a TDDFT Perspective:. . . 271
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
2 Nonlinear X-Ray Spectroscopies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
2.1 Time-Resolved Four-Wave Mixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
2.2 Double-Quantum-Coherence Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
2.3 Stimulated X-Ray Raman Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
2.4 Correlation Function Expressions for SXRS Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
2.5 Discussion of Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
3 Quantum Chemistry Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
3.1 ΔSCF-Based Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
3.2 TDDFT Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
3.3 MCSCF Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
3.4 Other Core Hole State Simulation Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
Abbreviations
TS Transition state
TSH Trajectory surface hopping
TXAS Transient X-ray absorption spectroscopy
XANES X-Ray absorption nearedge structure
XCH Excited core hole
XDQC X-Ray double-quantum-coherence
XES X-Ray emission spectroscopy
XFEL X-Ray free electron laser
XPS X-Ray photoelectron spectroscopy
1 Introduction
In linear spectroscopy experiments, the incident light field interacts with the studied
system only once and is relatively weak compared to the intrinsic interaction
potential of the system. The signal can be considered as the linear response of the
system to the perturbation of the light field. Linear spectroscopy provides useful
information about the atomic and electronic structure of the system. Nonlinear
spectroscopy techniques provide more detailed information. Nonlinear spectro-
scopy [1] employs multiple light fields to probe the correlations between different
spectral features. Many controlling factors, such as the frequencies, wavevectors,
and polarizations of the light fields and the time delays between them, can be varied
so that detection of the correlation between certain spectroscopy features and their
dynamics is possible.
Nonlinear optical spectroscopy became feasible soon after the invention of the
laser in the 1960s. With the development of laser technology, pulse durations were
reduced from picoseconds (1970s) to femtoseconds (1980s) [2], and now to
attoseconds [3]. Nonlinear infrared and optical spectroscopy techniques have
proved to be very successful for studying various excited state couplings and
dynamics in molecules and materials [4, 5]. Nonlinear spectroscopy techniques in
the X-ray regime made possible by new X-ray free electron lasers (XFEL) and high
harmonic generation (HHG) sources provide a unique window into the motions of
electrons, holes, and excitons in molecules and materials. Because of their broad
bandwidth (about 10 eV for a 100-attosecond pulse), X-ray pulses can create
coherent superpositions of many excited states localized at the target atoms. In
analogy to how optical pulses manipulate molecular vibrations, attosecond X-ray
pulses triggering and probing valence excited state dynamics have been considered
recently [6] and explored experimentally [7–10]. Sequences of coherent broadband
X-ray pulses can reveal the dynamics of nuclei and electrons in molecules with
attosecond temporal, and nanometer spatial resolution.
X-Ray pulses can be used in various ways:
• Off-resonant diffraction detects the charge density. This technique can be
extended to multiple dimensions to provide multipoint correlations of the charge
density [11, 12].
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 277
• Core resonant spectroscopy offers a fast and versatile way to trigger valence
excitations at selected positions and times via a stimulated Raman process [6]
and to study their dynamics.
• Multiple cores can be excited at various delays, allowing the study of nonlinear
response of valence electrons.
The complex nature of excited state correlations and dynamics leads to charac-
teristic patterns in nonlinear spectroscopy signals, whose interpretation calls for
state-of-the-art theoretical simulation methods. The simulation of time-domain
nonlinear X-ray spectroscopy signals poses numerous challenges to quantum
chemists. First, resonant X-ray spectroscopy involves core excited state. Most
computational molecular electronic structure activity had focused on the ground
state and valence excitations. Core excitations received much less attention because
they do not participate in typical chemical processes. Core excited states lie well
above many valence excited states, and calculating them directly with bottom-up
algorithms is not practical. Core electrons also have special asymptotic behavior
(cusp condition) close to the nuclei and large relativistic effects. Second, signals
obtained by broadband X-ray pulses require many excited states. A state-by-state
calculation scheme is tedious and it is better to obtain all excited states with an
energy range in one shot. Third, resonant X-ray signals require not only the energies
of excited states but also the transition dipoles between them. It is usually necessary
to calculate high order excited state energy gradients to determine these quantities,
which complicates the simulation. Furthermore, multiple X-ray pulses can easily
create excited states with multiple core holes, which are not well described by
single-reference-based excited state quantum chemistry methods such as adiabatic
time-dependent density functional theory (TDDFT). Most of the discussions in this
chapter are based on adiabatic TDDFT. Non-adiabatic frequency-dependent kernels
are discussed in Sect. 4.3. Finally, many electrons may respond to the core hole
created by the X-ray pulses (e.g., shake-up and shake-off processes) [13], so that
many-body effects are very important in these signals. The single-particle picture
may break down and high level methods such as multireference configuration
interaction (MRCI) or multireference perturbation theory (MRPT) are often neces-
sary to account for electron correlation. These challenges are addressed in the
following sections. We focus on the theoretical methods (mainly DFT/TDDFT)
which have been extensively used in X-ray spectroscopy simulation. There are
excellent reviews on using TDDFT to simulate linear X-ray spectroscopy signals
[13–15]. Here we emphasize the specific issues associated with nonlinear X-ray
spectroscopy simulations and mainly discuss the methods applied to molecules.
This chapter is organized as follows. We first briefly describe the calculation of
various nonlinear X-ray spectroscopy signals, and then review existing quantum
chemistry simulation methods. We then discuss several key issues in nonlinear
X-ray spectroscopy simulation. Finally, conclusions and future directions are
outlined.
278 Y. Zhang et al.
X p^ 2 1X
^ ¼
H i
þ V^ ri r j ; ð1Þ
i
2mi 2 i j
where p^i is the momentum of the ith electron and V^ is the Coulomb potential. In the
minimal-coupling Hamiltonian, the effects of an external electromagnetic field are
^ where qi is the charge and  is the
included by the substitution p^i ! p^i qci A
electronic magnetic vector potential [1, 16]. The minimal coupling is well-suited to
discuss X-ray diffraction, which arises from the A2 term, but it is often more
convenient to work with the electric and magnetic fields (which are gauge invari-
ant) rather than the vector potential. This is accomplished by the Power–Zienau
canonical transformation [1, 16]. The Hamiltonian of the system then becomes
H ^ þH
^ S ðt Þ ¼ H ^int ðtÞ; ð2Þ
where H is the material Hamiltonian and, in the dipole approximation, the inter-
action Hamiltonian is
ð
H^int ðtÞ ¼ dr ℰ^ ðr; tÞ þ ℰ^ { ðr; tÞ μ^ ; ð3Þ
with μ^ the dipole operator and ℰ^ þ ℰ^ Eis
{ ^ the electric field which is separated
into positive and negative Fourier components. Within the rotating wave approxi-
mation, the dipole moment is also separated into positive and negative Fourier
components μ^ ¼ V^ þ V^ { and only the terms ℰ^ V^ { þ ℰ^ V^ are retained [1]. Through-
{
out, we work in the interaction picture with respect to this Hamiltonian and in the
Hartree units, which simplifies the coefficients in the resulting expressions. The
detected quantity in the signals coincided here is the integrated photon number
ð D E ð h i
^_
SðΛÞ ¼ dt N ðtÞ ¼ dtdrℑ ℰ^ ðr; tÞ hμ^ ðtÞi ;
{
ð4Þ
where the last equality follows from the Heisenberg equation of motion for the
photon number operator and the signal is a function of the parameters defining the
pulse envelope (collectively denoted Λ). In the following, we take the field to be
polarized along the dipole and avoid the tensor notation (this restriction is easily
relaxed). Note that this form for the signal does not include any frequency- or time-
resolved detection. This could be done by adding gating functions [11, 17, 18] in
nonlinear spectroscopic applications; the electric field is a superposition of more
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 279
than one externally applied pulses or continuous wave (CW) laser field. To simulate
such experiments, we calculate the propagated wavefunction of the driven system
^ ðtÞjψ 0 i;
jψ ðtÞi ¼ U ð5Þ
Linear signals are determined by the first order Ĥint. In the X-ray regime, such
signals include X-ray absorption near edge structure (XANES) and extended X-ray
absorption fine structure (EXAFS) [19–21]. The third order techniques (four-wave
mixing) provide more detailed information [6, 22]. In this section, we describe a
class of techniques that utilize four pulses well-separated in time. The pulses
interact with the molecule sequentially and the signal is defined as the change in
transmission of the final pulse. In the limit of ultrashort pulses, the signal is
parameterized by the time delays between successive pulses. In the semiclassical
approximation (where the electric field is treated classically and the molecule is
quantum), we have
280 Y. Zhang et al.
X
Eð t Þ ¼ ℰ~ t τp eik p riω p ðtτp Þþiϕ p þ c:c: ð9Þ
p¼1, 2, 3, 4
where ϕp is the phase of the pth pulse, τp , ωp the central times and frequencies of the
temporal and spectral pulse envelopes and ℰ~ ðtÞ the temporal pulse envelopes
p
centered at t ¼ 0. The system interacts once with each pulse and the signal can then
be plotted as a function of the pulse parameters. The terms in the perturbative
expansions are conveniently depicted diagrammatically. Besides facilitating
enumeration of all terms, this procedure allows one to write quickly the signal
corresponding to a particular diagram and to discern in which time periods parti-
cular coherences appear. For macroscopic samples longer than the relevant radi-
ation wavelength a delta function δ(k4 k3 k2 k1) results. This is known as
phase matching for our level and dipole scheme. The ground state, singly excited
state and doubly excited state manifolds involved in these four-wave mixing
experiment are shown in Fig. 1. The three possible signals are denoted
kI k1 + k2 + k3, kII k1 k2 + k3, and kIII k1 + k2 k3. Below we focus on
two techniques: the double quantum coherence four wave mixing, and the stimu-
lated Raman simulations, and analysis of these signals are given later. The diagrams
of other two four-wave mixing techniques are also provided in Fig. 2 for reference.
Fig. 2 Loop diagrams for the kI (top row) and kII (bottom row) signals. The system begins in the
ground state (or more generally, a distribution of states as in (30)) and then interacts once with each
of the four sequentially applied pulses. For diagram rules see [23]
Fig. 3 Schematic depiction of the double quantum coherence technique. Four time-ordered pulses
are shown impinging on the sample from different directions. The transmission of the final pulse is
recorded as a function of the delay times or their conjugate frequencies. Figure taken from [24]
282 Y. Zhang et al.
Fig. 4 The two loop diagrams contributing to the kIII (DQC) signal. The system begins in
the ground state (or more generally, a distribution of states as in (30)) and then interacts
once with each of the four sequentially applied pulses. Note that the phase choice for this signal
(kIII ¼ + k1 + k2 k3) guarantees that the second excitation manifold is reached and provides
resonances between the ground state and this manifold
From the diagrams in Fig. 4, one can immediately obtain the time-domain
signals:
ð ð1
SkIII ðΛÞ ¼ ℑ dt dt3 dt2 dt1 ℰ~4 ðt τ4 Þeiω4 ðtτ4 Þ ℰ~3 ðt t3 τ3 Þeiω3 ðtt3 τ3 Þ
0
whereas the t3 resonances contain single excitations from the ground state or
double excitations from the single-excitation manifold depending on the dia-
gram. We can therefore fix one of t1 or t3 and transform with respect to the other
two time arguments to obtain a two-dimensional frequency plot which reveals
correlations between the double excitations and single excitations (either from
the ground state or the single-excitation manifold). We denote this impulsive
signal by a 0 superscript:
S0kIII ðt3 ; t2 ; t1 Þ ¼ ℰ4 ℰ3 ℰ2 ℰ1 eiω1 t1 iðω1 þω2 Þt2 iðω1 þω2 þω3 Þt3 δðω1 þ ω2 ω3 ω4 Þ
{
ψ 0 U ^ ðt þ t2 þ t3 ÞV^ U ^ ðt3 ÞV^ U ^ ðt1 ÞV^ { ψ 0
^ ðt2 ÞV^ { U
{ 1
in order to reveal resonances better. Finite pulse envelopes may now be incorpo-
rated and, when the correlation functions are expanded in material eigenstates,
we obtain
Xℰ e ∗ ω ω 0 V 0 ℰ e∗ e
4 eg ge 3 ω3 ω f e0 V e0 f ℰ2 ω3 ω f e V f e
SkIII , a ðΩ3 ; Ω2 ; Ω1 Þ ¼ 4
f e0 e Ω3 ωe0 g þ iγ e0 g Ω2 ωfg iγ fg
e ω ω V
ℰ1 1 eg eg
;
Ω1 ωeg iγ eg
ð13Þ
e∗ ω
Xℰ
e ∗ ω
ω f e0 V e0 f ℰ
e ω ω V∗
ωe0 g V ge0 ℰ
4 3 2 3 fe
4 3 fe
SkIII , b ðΩ3 ; Ω2 ; Ω1 Þ ¼
f e0 e Ω3 ω f e0 þ iγ f e0 Ω2 ωfg iγ fg
e ω ω V ∗
ℰ
1 1 eg
;
eg
Ω1 ωeg iγ eg
ð14Þ
where ωi j εi ε j and γ ij are the frequency and the dephasing rate of the i ! j
transition, respectively. The contributions from diagrams a and b may be read
directly from Fig. 4. The numerator contains all transition dipoles as well as the
field-envelope factors which determine the material transitions permitted by the
bandwidths. The denominators contain the resonance factors for these material
transitions.
284 Y. Zhang et al.
At the level of Hartree theory, which assumes independent electrons, the DQC
signal vanishes because of interference. TDHF (or TDDFT) goes one step further
and provides a picture of independent transitions (quasiparticles). Here the signal
no longer vanishes, but shows a limited number of peaks. When correlation effects
are fully incorporated, the many-electron wavefunctions become superpositions of
states with different numbers and types of e-h pairs. The Ω2 and Ω3 axes then
contain many more peaks corresponding to all many-body states (in the frequency
range spanned by the pulse bandwidths), which project into the doubly-excited
states. Thus, along Ω2 the peaks are shifted, reflecting the level of theory used to
describe electron correlations. Along Ω3, the effect is even more dramatic and new
peaks show up corresponding to splittings between various levels. We show the
X-ray DQC signals of formamide as an example in Sect. 4.3. This highly-resolved
two-dimensional spectrum provides an invaluable direct dynamical probe of elec-
tron correlations (both energies and wavefunctions) [25, 26].
Linear techniques contain the single-excitation spectrum whereas we have just seen
that the DQC (kIII) signal gives access to the double-excitation spectrum. Both of
these spectra thus characterize the intermanifold structure of the material (the
transitions between manifolds). We may obtain a window into the intramanifold
structure (transitions within the same manifold) by using the stimulated Raman
signal (SXRS in the X-ray regime) [2, 27, 28]. As with the DQC signal, this technique
is third-order (involving four interactions with the electromagnetic field). However,
rather than four sequential pulses, 1D-SXRS employs only two pulses, each of which
interacts twice with the material. This process is shown diagramatically in Fig. 5.
Note that, because the pair of interactions with each pulse are of opposite
Hermiticity, the overall absolute phase is ϕ1 ϕ1 þ ϕ4 ϕ4 ¼ 0 and this technique
therefore does not require phase control to obtain a finite signal.
The first pulse in the SXRS process creates a superposition of excited states in
the ground state manifold. After a controlled delay period, the sample interacts with
the second pulse which returns the system to the original state.
For calculating this signal, we find it more convenient to work with the actual
interaction times τ rather than the time delays t j ð j ¼ 1, 2, 3Þ. It is straightforward
to write down a time-domain expression for the 1D-SXRS signal directly from the
diagrams in Fig. 5. Its form is similar to (10):
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 285
Fig. 5 Two contributing loop diagrams (labeled as a, b in the figure) for the 1D-SXRS technique.
As before, the system begins in the ground state but this time interacts twice with each of the two
sequentially applied pulses. Note that the phase for this signal (ϕ1 ϕ1 + ϕ2 ϕ2) automatically
vanishes, making the signal incoherent. The first pulse prepares a wavepacket of valence excita-
tions that evolves for the interpulse delay period before being probed with the second pulse
ð ð τ4 ð τ3 ð τ2
*
SSXRS ðΛÞ ¼ ℑ dτ4 dτ3 dτ2 dτ1 ðiÞ3 ℰ2 ðτ4 Þℰðτ3 Þ
h 1 1 1
ℰ1 ðτ2 Þℰ1 ðτ1 Þ ψ 0 V^ ðτ4 ÞV^ { ðτ3 ÞV^ ðτ2 ÞV^ { ðτ1 Þψ 0 : ð15Þ
*
i
þ ℰ1 ðτ2 Þℰ ðτ1 Þ ψ 0 V^ { ðτ1 ÞV^ ðτ2 ÞV^ ðτ4 ÞV^ { ðτ3 Þψ 0
*
1
Because the interactions are paired within a given pulse and the pulses are tempo-
rally well-separated, we may extend the upper limit for the τ2 integration to infinity.
This permits us to define formally the polarizability α^ p induced by the pth pulse:
ð1 ðτ
dτ0 V^ ðτÞV^ { ðτ0 Þℰ p ðτÞℰ p ðτ0 Þ;
*
α^ p Λ p i dτ ð16Þ
1 1
which is both a material operator and a function of Λp, the parameters of the pth
pulse. In the limit of ultrashort pulses, the primary Λp parameter is the central pulse
time τp and the principal control variable for the 1D-SXRS signal is the interpulse
delay T ¼ τ2 τ1 and the signal is recast as
where we have set τ1 ¼ 0 as the origin of time. Taking matrix elements in the
Hamiltonian eigenbasis gives the sum-over-states expression
286 Y. Zhang et al.
X ð1 ð τ2
dτ1 ℰ p ðτ2 Þℰ p ðτ1 Þe ð
* i ω p ωeg0 þiγ e Þτ2 i ω p ωeg00 þiγ e τ1
α p, g0 g00 i V g0 e V eg00 dτ2 :
e 1 1
ð18Þ
Here, ωev is the frequency for the v ! e transition, and γ e is the inverse of the
excitation lifetime. In some applications, it may be more convenient to work in the
frequency domain where the spectral (rather than temporal) field envelopes are
used. This can be accomplished by explicitly writing the propagators in (16) and
replacing the temporal field envelopes by their Fourier transforms yielding
ð
dωdω p dω0p
ℰ p ω0p ℰ p ω p
*
α^ p ¼
ð2π Þ3
ð19Þ
1 1 1
V^ V^ { :
ω þ ω p ω0p H^0 þ iη ω þ ω p H
^0 þ iη ω H
^0 þ iη
Starting from (20), we may now write the frequency-domain 1D-SXRS signal as
X ℜ α2;gg0 α1;g0 g γ g0 iΩ þ ℑ α2;gg0 α1;g0 g ωg0 g
SSXRS ðΩÞ ¼
g0 γ 2g0 2iγ g0 Ω Ω2 þ ω2g0 g
; ð21Þ
ℜ α*1;gg0 α2;g0 g γ g0 iΩ þ ℑ α*1;gg0 α2;g0 g ωg0 g
þ
γ 2g0 2iγ g0 Ω Ω2 þ ω2g0 g
which is the Fourier transform of (15) with respect to the interpulse delay T. The
first term in (17) and (21) can be viewed as a valence wavepacket α1|ψ 0i, created by
pulse 1, which propagates forward in time T and overlaps with a wavepacket hψ 0|α2
created by pulse 2. The second term can be viewed as a wavepacket α2|ψ 0i created
by pulse 2 propagating backward in time T to overlap with the wavepacket hψ 0|α{1
created by pulse 1. The SXRS technique creates a wavepacket of valence exci-
tations and, after a specified delay period T, probes this wavepacket so as to track its
evolution. A 2D extension of this 1D-SXRS in which three successive pulses are
employed is shown in Fig. 6. The resulting signal SSXRS(Ω1, Ω2) requires expansion
to fifth order in the field and carries information about correlations between
dynamics during the two delay periods which would not be available in
1D-SXRS [29]. This technique can also be applied following a pump pulse which
prepares the system by exciting a core hole. The subsequent SXRS process then
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 287
Fig. 6 Four contributing loop diagrams (labeled as a, b, c, d in the figure) for the 2D-SXRS
technique. The system begins in the ground state then interacts twice with each of three sequen-
tially applied pulses. As with 1D-SXRS, the phase vanishes and the signal is incoherent. The
additional delay period allows information about couplings and correlations of valence excitations
that are not available in 1D-SXRS to be extracted
creates a valence excitation and tracks its motion along the potential energy surface
produced by the core hole [30].
τ~pi sup t : E p ðt0 Þ ¼ 0 8 t0 < t
; ð22Þ
τ~p f inf t : E p ðt0 Þ ¼ 0 8 t0 > t
where the properties of the pulse and the choice of the τ~ guarantee the appropriate
sign of the propagators. This is an operator in the valence excitation space. It may
be recast in the frequency domain as before:
ð
dωdω p dω0p
ℰ p ω0p ℰ p ω p eiωðτ~p f ~τpi Þ eiðω p ω p Þτ~p f
* 0
^p τ~p f ; τ~pi
α ¼ 3
ð2π Þ
;
1 1 1
V^ V^ {
^0 þ iη ω þ ω p H
ω þ ω p ω0p H ^0 þ iη
^0 þ iη ω H
ð26Þ
and may differ from (19) in the appearance of τ~p -dependent phase-factors. Note
that, because the choice of the τ~p is set by the pulse shape, their appearance on the
left hand side is not necessary and we merely include them for clarity. With this
definition, the contribution to the signal from diagram a in Fig. 5 is
hD ^ ^ Ei
eiH0 τ~2 f α^2 ~τ2 f ; ~τ2i eiH0 ðτ~2i ~τ1 f Þ α^1 ~τ1 f ; ~τ1i eiH0 τ~1i :
^
Sa ¼ ℜ ð27Þ
Defining the duration of the pth pulse t p τ~p f τ~pi and the interpulse delay T τ~2i
~ τ1 f we have, in the limit of well-separated pulses, T >> t p p 2 f1; 2g. Assuming
that, having accounted for the finite pulse duration in the definition of α, we may take
t p ! 0 for the purposes of the free evolution in (27), the signal becomes
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 289
hD Ei
^ ^
Sa ð T Þ ¼ ℜ eiH0 T α^2 eiH0 T α^1 ¼ ℜ½hα^2 ðT Þα^1 ð0Þi; ð28Þ
where we have set τ~1i ¼ τ1 f as the zero point of time and the last equality defines the
interaction picture polarizability αp(t). Note that this expression for Sa(T ) matches
the first term in (17) (with the second term standing for Sb(T )). Besides being
necessary for certain applications, the separation of interpulse and intrapulse
propagations prominently features the dependence on the key time parameter, the
interpulse delay T. All other parameters defining the pulses are encoded in the
definition of the α^ p . We pause to recall that the only assumption necessary in
reaching (28) (just as for (17)) is that the interpulse delay be much larger than the
temporal pulse widths.
It is important to note that separating interpulse and intrapulse propagation
periods yields a formally identical expression and may seem an unnecessary arti-
fice, as indeed it is within an eigenstate representation. The utility then is manifest
when the eigenstates are prohibitively expensive to calculate. For example, in the
configuration interaction representation, states are expanded in a basis consisting of
the many-body ground state (the orbitals being filled up to some maximum energy
level) and excitations on this ground state obtained by successively higher orders of
electron creation-annihilation operator pairs:
X
j ψ i ¼ j gi þ Ci j c^ {i c^j jgi þ . . . ð29Þ
ij
Because the material may generally be taken to begin an experiment in the many-
body ground state, perturbative treatment of nonlinear spectroscopies naturally
produces such states. At low order, there are many fewer states in this treatment
than in the full eigenbasis and a significant numerical speedup can be achieved.
In order to exploit this form requires a similar recasting of the α^ p and this is
explored in Sect. 4.2. Corresponding expressions for the 2D signal SSXRS(T2, T1)
(Fig. 6) are given in [29, 31].
In the above sections we provided two different types of expressions for the DQC
and SXRS signals. The first ((10) and (15)) are given in terms of time correlation
functions of the dipole operator. This form is convenient for direct ab initio
dynamic simulations of electrons and nuclei [31, 32]. It can take into account,
e.g., in nonadiabatic dynamics, conical intersections, etc. Real-time time-dependent
density functional theory can then be applied to calculate the signal. Alternatively,
the second procedure ((13), and (17)–(20)) expands the correlation functions in
molecular eigenstates. This is convenient for simpler models when only a few
electronic states participate and for relatively small systems where the many-
290 Y. Zhang et al.
body states can be computed. The sum-over-states expansions also facilitate inter-
pretation of the resulting signal as the meaning of the various resonances is
transparent in this form. Both the correlation function and sum-over-states forms
may be displayed either in the time-domain (as a function of t3, t2, t1) or the frequency
domain (as a function of the conjugate variables Ω3, Ω2, Ω1). It is often useful to
employ a mixed representation, e.g., S(Ω1, t2, Ω3), which is 2D in frequency and 1D
in time so that correlations between resonances observed at the two frequencies can
be observed and monitored as the time argument is allowed to vary. These tech-
niques can therefore provide a high degree of selectivity and carry a rich abundance
of information on the electronic and nuclear structure and dynamics.
Our correlation function expressions (10) and (15) are given by the expectation
values with respect to |ψ 0i. Alternatively, we may describe the system using the
density matrix
X
ρ^ ¼ Pi jψ i ihψ i j; ð30Þ
i
ρ^_ ¼ i H;
^ ρ^ i H
^int ; ρ^ : ð31Þ
Here Pi is the probability that the system is found in state |ψ ii. When all degrees of
freedom are treated at the Hamiltonian level, it is more convenient to remain in
Hilbert space rather than recasting in Liouville space (as is done in [1]). This
facilitates computations because Hilbert space has far fewer dimensions than the
associated Liouville space. In these cases, the above equations may still be utilized
formally with appropriate choice of the Pi. In terms of the density matrix, the
expectation value of the dipole is given by
One can then include the effects of coupling to a bath by introducing further terms
to the equation of motion – see (30) – which represent the dissipation of system
excitations into the bath. One numerically inexpensive strategy to implement this
idea is the stochastic Liouville equation (SLE)
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 291
ρ^_ ¼ i H;
^ ρ^ þ L^b ρ^ ; ð34Þ
where L^b represents the stochastic Markovian dynamics of the bath. The SLE is an
equation for the field-free evolution of the joint system-bath density matrix and can
be used to write a reduced equation of motion for the system density matrix which
incorporates (perturbatively) the effects of the bath. This can, for example, be done
at the level of the Lindblad equation [33, 34]. The model for the bath and system-
bath coupling determines the form of L^b . Examples are the n-state jump and
Brownian oscillator models [1, 35].
Equations (32) and (33) give one procedure for obtaining the nth order signal and
generates 2n terms when the commutators with the initial density matrix are fully
expanded. Equations (8) and (7) offer an alternative procedure which, upon
expanding the time-ordered exponentials in the U^({)(t), generates n þ 1 terms at
nth order. The latter procedure obviously involves less terms and it is often
numerically preferable to propagate the wavefunction rather than the density
matrix. On the other hand, only a density matrix based procedure can properly
account for system-bath interactions and the dephasing-effects these cause. More-
over, the real-time interpulse delays appear more naturally in a density-matrix
formulation. Equations (32) and (33) are therefore more expensive to implement
but provide a more intuitive picture and are necessary when a proper account of
system-bath dynamics is crucial [1].
As previously mentioned, the nth-order expansion of the density matrix as per
(33) generates 2n terms. Interpreting the resulting signal requires expanding
the interaction Hamiltonian into its constituent terms (which are, in the rotating
wave, (ℰ^ V^ and ℰ^ V^ { )). There are thus a total of 4n terms which may be depicted
{
Signal expressions in Sect. 2 require the calculation of core excited states and
transition dipole moments. Here we review the quantum chemistry methods that
can be used in their simulation and present a few examples.
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 293
Fig. 7 Approximation schemes for core hole excitations. ECH equivalent core hole (Zþ1)
approximation, FCH full core hole approximation, XCH excited core hole approximation, TS
transition state method, TP(HCH) transition potential method (half core hole approximation). Full
discs represent electrons and half discs represent half electrons. Numbers in circles at the bottom
represent nuclear charges, where Z is the number of electrons of the system
An X-ray photon usually excites a core electron leaving a core hole in the system.
Describing core holes properly is the primary task of resonant X-ray spectroscopy
simulation. In Fig. 7 we show the most widely used approximation schemes for core
holes [45]. The simplest is to represent a core hole as an additional nuclear charge.
This equivalent core hole (ECH) also known as (Zþ1) approximation [46, 47] is
adequate for deep core holes because for electrons in the exterior shells a deep core
hole behaves as does a positive nuclear charge. It is simple to apply (no additional
coding in standard quantum chemistry packages is necessary) and multiple core
holes can also be easily represented by additional nuclear charges. However, it is a
crude approximation which does not apply to shallow core holes. It further arti-
ficially changes the spin state of the system.
The ECH approximation was used in our early X-ray nonlinear spectroscopy
simulations [48–52]. The photon echo signal kI ¼ k1 þ k2 þ k3 of the para and
ortho isomers of aminophenol was calculated in [48] (see Fig. 8). The second time
delay t2 is set to zero. The signals reveal the correlation between the O1s core
excitations (Ω1) and the N1s core excitations (Ω3). The equivalent-core molecular
orbitals corresponding to the three strong O1s XANES peaks (marked A, B, and C)
are also shown. In a simple single orbital picture, orbital A is populated by the
excited O1s electron in the lowest O1s excitation. The XANES signals are not
sensitive to the corresponding core excited states, as can be seen from the top of
Fig. 8. Although the orbitals corresponding to peak Bs of the two isomers look very
294 Y. Zhang et al.
Fig. 8 Simulated O1s XANES and O1s/N1s photon echo crosspeak (at t2 ¼ 0) of para- (left) and
ortho-aminophenol (right) with the ECH approximation. All single core excitation energies are
shown as ω ωj, where ωj is the lowest O1s or N1s excitation energy. Molecular orbitals
populated by the promoted O1s electron for each of core-excited states contributing to the signal
are show on the top. The positions of N and O atoms are labeled. Figure adapted from [48]
different, the two peak Bs are similar in energy and lineshape. This is because
XANES only detects the local electronic structure of atoms. However, the photon
echo signals show many differences. In the para isomer, the orbital corresponding
to peak B is delocalized and extends from the O to N atom. So, the N1s and O1s
core excitations corresponding to this orbital affect each other and generate a
crosspeak (Fig. 8, left). However, in the ortho isomer, because the orbital
corresponding to peak B vanishes in the surroundings of the N atom, the N1s and
O1s core excitations corresponding to this orbital do not affect each other and thus
we cannot see a crosspeak (Fig. 8, right). Although it is much stronger in O1s
XANES of the para isomer, peak A contributes to a much weaker crosspeak than
peak B. This is because the orbital corresponding to peak A in the para isomer is
highly localized to the O atom and far away from the N atom. It is also understand-
able that peaks A and C produce much stronger crosspeaks of the ortho isomer than
those of the para isomer because O and N atoms are closer in the ortho isomer.
In all, photon echo signals carry detailed information about the wavefunctions of
the core excited states involved in the experiment.
More than four decades ago Slater had proposed the transition state (TS) method
for calculating core excitation energies. In this method the two orbitals involved in
the transition are occupied by a half electron and solved self-consistently. The
excitation energy is given by the difference between the two orbital energies
[53, 54]. The excitation energy obtained in this way is accurate up to second
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 295
order in the occupation number changes of the transition orbitals. The transition
state method is convenient and sufficiently accurate in many cases, but it is not
suitable for calculating many excited states because of the nonorthogonality and
self-consistent field (SCF) collapse of the excited states. An alternative is the
transition potential method (TP) [55, 56], in which the potential corresponding to
the transition hole state (a half electron on the hole orbital; see Fig. 7) is used to
produce a set of orthogonal excited states. The excitation energy is determined by
the differences between transition potential orbital energies. TP is widely used in
X-ray absorption spectroscopy simulation [57].
Similar core hole approximations have been proposed in solid state physics. The
half core hole approximation (HCH; see Fig. 7) is similar to the transition potential
method, and the full core hole approximation (FCH; see Fig. 7) is similar to the
direct exchange method (STEX) [58–60] in quantum chemistry (this is explained in
detail in the next subsection). In FCH, the impact of the excited electron on the core
hole is neglected. If it is included, we obtain the excited core hole approximation
(XCH) [61]. We have used XCH combined with TDDFT to simulate X-ray double-
quantum-coherence spectroscopy [24].
N E
Ψ i ¼ p1ffiffiffi a^ { ΨN1 þ a^ { ΨN1 iÞ; ð35Þ
j, l lα jα lβ jβ
2
N
where ΨN1
jσ i a ^ jσ Ψref i and σ ¼ α, β are spin states, j is the core orbital index, l is
N
the virtual orbital index, Ψref i is the N-electron neutral reference state, and a^ and a^{
are annihilation and creation operators, respectively. The excited orbitals within the
N1 approximation satisfy the eigenvalue equations:
F^ STEX
j
ψ l j ¼ εl j ψ l j ; ð36Þ
where ψ jl is the exited orbital and εjl is the corresponding orbital energy. The STEX
Fock operator
X
occ
F^ STEX
j
¼ h^ þ 2J^i K^i þ J^j þ K^ j ; ð37Þ
i6¼ j
X
occ
^j ¼
P ψ j ihψ j ; ð39Þ
k k
k6¼ j
to project out all occupied orbitals of the (N1)-electron ionic system and solve the
projected STEX equation
j
^1 P ^
^j F ^ ^j j j j
STEX 1 P ψ l ¼ εl ψ l : ð40Þ
The solutions of this equation should serve as a good approximation to the excited
orbitals. The core excitation energy is finally given by
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 297
ω j, l ¼ IP j þ ε j, l ; ð41Þ
where IPj is the ionization potential of the core electron j (determined from the
energy differences of the (N1)-electron ionic system and the N-electron neutral
system).
STEX is a single excitation theory in which particle and hole are independent.
Channel interaction can be included by diagonalizing the configuration interaction
matrix in the space of linear combinations of different single excitation channels.
This extension is known as the STEX-CIS (configuration interaction singles)
method. Double excitations can also be treated with a STEX Hamiltonian based
on the (N2)-electron ionic system, but the SCF convergence problem is more
serious and the spin coupling schemes are complicated [66].
State-to-state transition dipoles are necessary in order to simulate the nonlinear
X-ray spectroscopy signals. Because the STEX orbitals are not orthogonal to the
MOs of the N-electron system, L€owdin’s rule [67] may be used for calculating the
transition dipoles between two states with nonorthogonal single particle orbitals.
The transition dipoles for single excitations are
NX config: X
ΨA d^ΨB ¼ am bn ð1Þiþ j d mn
i j MinorðS Þi j ;
mn
ð42Þ
m, n i, j
where ΨA,B are two excited states with nonorthogonal single particle orbitals, d^ is
the transition dipole operator, and am and bn are configuration interaction (CI)
coefficients for different single excitation configurations (m and n) of state A
and B, respectively.
X ð
* ^
d mn
ij ¼ c ip, m, A jq, n, B ϕ p d ϕq dτ;
*
c ð43Þ
p, q
is the transition dipole matrix between the single excitation configurations m and n,
cip,m,A and cj,q,n,B are MO coefficients for the configurations m and n of state A
and B, respectively, and
X ð
Si j ¼
mn
cik, m, A c jl, n, B ϕ*i ϕ j dτ;
*
ð44Þ
k, l
is the overlap matrix between the MOs of the configurations m and n of state A
and B, respectively. ϕi,j in (43) and (44) are basis functions and i, j, p, q, k, l are
indices for these basis functions. Minor (Smn)ij denotes the (i, j) minor of the matrix
Smn.
We next present 1D and 2D SXRS signals calculated using STEX [29, 68]. Fig-
ure 9 shows the 2D-SXRS spectrum of N-methylacetamide (NMA) with the OOO
(O1s pump with two O1s probes) pulse sequence, together with 1D projections
298 Y. Zhang et al.
Fig. 9 The 2D-SXRS signal SSXRS(Ω1, Ω2) and its 1D traces of NMA (right) from STEX
calculations. Left: the OOO spectrum. All these pulses are resonant with the O K-edge. Middle:
horizontal and diagonal slices of the 2D spectrum on the left (in red) plotted together with the
corresponding traces from the corresponding OON (dashed, blue) to highlight the effect of
changing the probe pulse in the three-pulse sequence. Figure adapted from [29]
along several horizontal and diagonal traces. We also show the corresponding
traces from the OON 2D-SXRS signal in blue dashed lines for comparison. The
only difference between the two types of 1D signals are the third pulse. Peaks along
the diagonal line (Ω2 ¼ Ω1, i in Fig. 9) resemble those from the 1D-SXRS spectrum
of the same molecule [29]. Peaks along the horizontal lines drawn at the represen-
tative valence excitation energies (Ω2 ¼ 8.95, 8.14 eV, ii and iii in Fig. 9) reveal the
interference of the two Liouville space quantum pathways represented by diagrams
a and d in Fig. 6, and peaks along the diagonal lines shifted with representative
valence excitation energies (Ω2 ¼ Ω1 6.91, 8.95, 12.68 eV, vi, v, and vi in Fig. 9)
reveal the interference of the other two quantum pathways represented by diagrams
b and c in Fig. 6. Figure 9 illustrates that multidimensional SXRS signals reveal
couplings of different valence excitations, and interferences of quantum pathways.
Comparison of STEX with another method for calculating core excited states,
the restricted excitation window time-dependent density functional theory
(REW-TDDFT), was given in [69]. Core excitation energies from both methods
must be shifted to match experiment. Because of the inclusion of core orbital
relaxation, the shifts of STEX core excitation energies (<10 eV) are usually smaller
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 299
than those of TDDFT (>10 eV). However, being an independent particle and hole
theory, STEX cannot account for core hole mixing in X-ray spectroscopy, whereas
REW-TDDFT can. The SCF calculation of a core ionized state is often tricky.
Convergence is not guaranteed. In addition, because the STEX equation (40) is not
solved self-consistently, the occupied and virtual STEX orbitals are not variational
for the total energy. Evaluating core excited state properties, e.g., electron density,
thus become complicated [70]. State-to-state transition dipole calculations are
expensive.
where Cn1 and Cn are the molecular orbital coefficient matrices in the last and
current SCF iteration, respectively, S is the overlap matrix of basis functions, and
the matrix element Oij represents the overlap between the ith old orbital and the jth
new orbital. The projection of the jth new orbital onto the old occupied orbital space
may be defined as
" ! #
X
occ X
occ Xocc X
occ
Pj ¼ Oij ¼ n1
Cik Skl Cljn : ð46Þ
i l k i
The orbitals with the largest Pjs are chosen as the new occupied orbitals. In some
cases (46) is not robust when selecting new occupied orbitals. Alternative pro-
jections such as
X
occ
Pj ¼ Oij ; ð47Þ
i
and
300 Y. Zhang et al.
X
occ
Pj ¼ O i j 2 ; ð48Þ
i
have been implemented in the quantum chemistry packages Q-CHEM [75] and
GAMESS [76]. The same projection scheme as in (48) was also proposed by other
authors very recently [77]. When all Oijs do not have the same sign, or a number of
orbitals are nearly degenerate during the SCF iteration, (48) is believed to perform
better than (47) [77].
Unlike STEX, a spin-unrestricted scheme is employed in ΔSCF-DFT. Thus for a
system with a closed shell ground state, ΔSCF-DFT usually gives a broken-
symmetry spin state which is a mixture of a singlet and a triplet state. The singlet
excitation energy can be obtained through the spin-purification formula [72]
ES ¼ 2EBS ET ; ð49Þ
where ES is the energy of the open shell singlet, EBS is the spin broken-symmetry
state energy, and ET is the triplet energy from a separated ΔSCF-DFT calculation.
Spin-purification is necessary in valence excitation calculations [78, 79] but is less
important in core excitations, where EBS ET so that ES EBS.
ΔSCF-DFT can be easily extended to calculate various excited state properties
other than the excitation energy [79]. It also includes orbital relaxation upon
excitation, which is neglected in TDDFT. Note that the (N1) approximation in
STEX may not be necessary for ΔSCF-DFT because the DFT virtual orbitals
experience the same potential as do the occupied orbitals. However, it has some
drawbacks. First, ΔSCF-DFT is a state-specific approach; one should calculate the
excited states one by one. This makes it unsuitable for broadband spectroscopy
simulations, where many excited states are needed. Second, it only gives excited
states which can be well described by a single determinant. Excited states with
strong configuration interactions are missed. Third, excited states from separated
ΔSCF-DFT calculations are not orthogonal, and there is no unique way to enforce
the orthogonality requirement. Finally, it is an open question how to run variational
DFT calculations of excited states because there is no Hohenberg–Kohn theorem
for a generic excited state [80]. Despite its drawbacks, ΔSCF-DFT has been revived
recently in charge-transfer excitation [81], Rydberg excitation [82], and excited
state potential energy surface calculations [83], and looks very promising in the
X-ray regime.
As explained in the introduction section, the discussions in this section are based on
adiabatic TDDFT. Although similar formulation for time-dependent Hartree-Fock
(TDHF) theory had existed for more than two decades [84], the time-domain
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 301
extension of DFT was not possible until Runge and Gross established the one-to-
one mapping of electron density and time-varying external potential [85]. There are
two categories for applying TDDFT to calculate excited state properties: the
response theory based on perturbation in the frequency domain and the direct
real-time propagation methods in the time domain. The frequency-domain formal-
ism of TDDFT, which is based on linear response theory, became popular after
Casida proposed a density matrix response equation which is very similar to the
renowned random phase approximation (RPA) equation [86]. The Casida equation
can be derived by solving the equation of motion (EOM) of the single electron
reduced density matrix to the first order of external perturbation in the frequency
domain. Similar expressions have been obtained for TDHF in the collective elec-
tronic oscillator (CEO) method [87–89]. Nonlinear response functions of the
system can be calculated in CEO by applying high order perturbation theory [89].
Another way to obtain excited state properties is to solve the EOM of
single-electron reduced density matrix by direct propagation in the time domain.
Time-dependent properties of the system induced by the time-dependent external
perturbation can be calculated directly and Fourier transform can recover the
excited state information in the frequency domain.
In this section, we start with the linear-response formalism of Casida, and then
present a specific variant of linear-response TDDFT applied to core excited state
(restricted excitation window time-dependent density functional theory,
REW-TDDFT). Moreover, high order perturbation theory methods for nonlinear
response properties of the system are introduced and finally the real-time propa-
gation methods are discussed. It should be noted that the response and real-time
propagation methods are very general and not restricted to DFT/TDDFT, but we
focus on the DFT/TDDFT formalisms of these methods in this chapter.
The Casida equation can be derived by calculating the linear response of the
density matrix or through an EOM approach. One may start with the EOM of the
one-particle transition density matrix PI ¼ Iih0:
^ PI ¼ i ∂PI ¼ ωPI ;
H; ð50Þ
∂t
where Ii and 0i are the Ith excited state and ground state, respectively, H^ is the
Hamiltonian of the system, and ω is the excitation energy. Considering the idem-
potency property of density matrix, the transition density matrix can be expanded as
302 Y. Zhang et al.
X
PI ¼ Xia a^ {^i þ Y ia^i { a^ ; ð51Þ
i, a
where Xia and Yia are one-particle/one-hole (1p1h) excitation and de-excitation
amplitudes, respectively, and i, a represent an occupied and virtual orbital, respec-
tively. Substituting (51) into (50), after some algebraic manipulations, one can
obtain the Casida equation:
A B X 1 0 X
¼ω ; ð52Þ
B* A* Y 0 1 Y
where
and
ðð
1
i σ aσ j τ bτ ψ jτ ðr0 Þ ψ bτ ðr0 Þdrdr0 ,
*
¼ ψ iσ ðrÞ* ψ aσ ðrÞ 0
jr r j
ðð ð54Þ
δ2 EXC
iσ aσ f xc jτ bτ ψ ðr0 Þ ψ bτ ðr0 Þdrdr0 :
* *
¼ ψ iσ ðrÞ ψ aσ ðrÞ
δρσ ðrÞδρτ ðr0 Þ jτ
Here i, j and a, b represent occupied and virtual orbitals, respectively, σ, τ are spin
indices, ε is the orbital energy, and fxc is the exchange-correlation kernel which is
expressed as the second-order functional derivative of the exchange-correlation
energy with respect to electron density (54). In (52), X and Y should be considered
as column vectors. Alternatively, it is possible to derive these equations for
the reduced single electron density matrix. This has been done in the CEO method
[87–89] for both TDHF [89] rather than TDDFT [90, 91].
Because of its balance of accuracy and computational cost, and its robustness
and black-box character, linear-response TDDFT has become the method of choice
for computing excited states, including core excited states. We have also based our
nonlinear X-ray spectroscopy simulation work on TDDFT [24, 69, 92, 93]. Unlike
ΔSCF-DFT, linear-response TDDFT does not target a single excited state. Only
ground state orbitals are necessary in the calculation, so that a manifold of excited
states is obtained in one shot. However, linear-response TDDFT also has its
limitations. Usually based on a single-referenced Kohn–Sham state, linear-response
TDDFT cannot handle excited states calculations for a ground state with a heavy
multiconfigurational character. Orbital relaxation for different excited states is
missed. Approximate energy functionals do not have proper long-range asymptotic
behavior, and thus linear-response TDDFT has difficulties in handling charge-
transfer excited states [94] and Rydberg states. The same limitation applies to
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 303
core excited states [14]. The long-range corrected density functionals are discussed
in Sect. 4.1. In addition, within the adiabatic (frequency-independent) exchange-
correlation kernels, linear-response TDDFT cannot properly describe double exci-
tations [95]. This is a major obstacle for simulating nonlinear spectroscopy experi-
ments, which directly access double or multiple excited states. Double excitations
and frequency-dependent exchange-correlation kernels are discussed in Sect. 4.3.
Fig. 10 Four-wave mixing and I2P-SXRS signals of cysteine (structure shown on the bottom
right) from REW-TDDFT calculations. Top: constatnt-Ω2 slices of the 3D kII signal SkII
(Ω1, Ω2 ¼ 6.6, 8.9, 11.4 eV, Ω3) using an OOSS pulse sequence with xxxx polarization. Bottom
left: the integrated two-pulse SXRS signal using an OS pulse sequence with xx polarization.
Figure adapted from [108]
Correlation between the motion of the doorway wavepacket and the fluctuations of
the SXRS signal profiles can be established. The time-domain signals provide a
real-time image of EET in the system, which is not possible for time-resolved
fluorescence anisotropy decay studies. SXRS could become a powerful tool in
revealing EET mechanisms in molecular systems. Further geometrical and struc-
tural factors that control EET in a series of porphyrin heterodimer systems were
studied in [93].
In another SXRS simulation study we investigated long-range electron transfer
(ET) in the small redox protein azurin [109]. Borrowing the ET kinetic parameters
from time-resolved infrared (IR) and optical measurements, time-resolved SXRS
signals at the electron donor, hopping intermediate and electron acceptor were
simulated with REW-TDDFT. We found that the SXRS signals depend sensitively
on the local electronic structure changes around the excited atoms, and could serve
as an excellent indicator for detecting electron transfer dynamics. The atomic
306 Y. Zhang et al.
Fig. 11 Time-domain 1D SXRS signals that reveal excitation energy transfer in porphyrin
heterodimers from REW-TDDFT calculations. Top: the molecular structure of the Zn-Ni porphyrin
heterodimer studied. Bottom left: (a, b) Spatially integrated hole and electron densities on the Ni
(red) and Zn monomer (blue). (c) The time-resolved integrated two-pulse SXRS signals of the
porphyrin dimer between 0 and 120 fs. The single color Zn2p/Zn2p signal is in blue and the
two-color Zn2p/Ni2p signal is in red. Bottom right: electron and hole densities of the Zn2p valence
superposition state prepared by SXRS for various times after excitation. The isosurfaces are colored
according to which monomer they reside on, red for Ni and blue for Zn. Figure adapted from [92]
pinpoint spatial accuracy also makes SXRS a convenient tool in studying different
ET pathways. Such measurements are difficult for conventional IR or optical
techniques. SXRS should complement linear transient X-ray absorption [110] in
studying ultrafast ET molecular processes.
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 307
Nonlinear optical response properties can be obtained by going beyond the first
order of perturbation in equations similar to (50). Response equation (Sternheimer
equation) of wave functions instead of density matrix can also been considered
[111]. Following the standard time-dependent perturbation theory, Orr and Ward
derived the sum-over-state (SOS) expressions of the nonlinear optical polarizations
four decades ago [112]. The SOS expressions are general and excited states from
any level of theory could be used. Much TDDFT work has been done along these
lines [91, 111, 113–121]. A weak time-dependent external electric field is intro-
duced as a perturbation to the original Kohn–Sham system. Then the coupled-
perturbed TDDFT equations, which are similar to the coupled-perturbed Kohn-
Sham equations in DFT geometry optimization calculations, are solved at different
orders of the perturbation and nonlinear response properties are evaluated by
perturbed wave function or density matrix. Using the 2n þ 1 rule [122], third
order response properties can be obtained through the first perturbed wave function.
These approaches can also be used to calculate the nonlinear response to X-ray
pulses.
The complex polarization propagator (CPP) method for XANES simulation was
proposed by Norman and coworkers [123–125]. They first parameterized the orbital
rotation during time evolution, then started from an EOM of the state-transfer
operators with external perturbations and a phenomenological damping term.
With this damping term, decay of excited states can be considered in the
now-complex response functions and resonance divergences are eliminated. With
perturbation techniques, the EOM can be solved at different orders of perturbation
and the corresponding response properties can be calculated. For the linear polar-
izability, one has
h i1
½1{ ½1
αi j ðωÞ ¼ μi E½2 ðω þ iγ ÞS½2 μj ; ð55Þ
½1
where i, j ¼ x, y, z are coordinate axis indices, μi;j is the electric-dipole property
gradient along the coordinate axis i, j, respectively, E[2] is the electronic Hessian, γ
is the phenomenological damping parameter, and S[2] is the overlap matrix
[125]. The CPP method is general and when DFT orbitals are used, it gives
excellent XANES spectra of large molecules such as copper phthalocyanine
[126]. A constant shift (usually it is a blue shift) is still needed to match the
calculated XANES to experiment. The constant shift depends on the system and
functional used, e.g., for the water molecule the shifts are 15.15 and 4.0 eV for the
CAM-B3LYP and LB94 functional, respectively [125]. Nevertheless, third order
electronic Hessian is needed to calculate second-order response, making the calcu-
lation quite involved.
Coupled-perturbed TDDFT results in SOS expressions of nonlinear response
functions [127], which become increasingly more complex for higher order
308 Y. Zhang et al.
In the frequency domain, each relevant excited state must be explicitly calculated
when the SOS expressions of nonlinear X-ray spectroscopy signals are employed.
The calculation becomes very expensive when many excited states contribute to the
signals. In recent attosecond laser spectroscopy experiments [128–140], the orbital
relaxation as well as the nonadiabatic dynamics with significant geometry changes
involve many excited states. Real-time methods are then preferable.
In real-time time-dependent density functional theory (RT-TDDFT), rather than
solving for eigenstates, the wave function or the one-electron reduced density
matrix1 is directly propagated in the time domain. Spectroscopic signals can be
extracted from Fourier transform of time-dependent system properties such as the
polarization of the molecule driven by the external electric field. The entire
spectrum can be obtained at once and direct calculation of specific excited states
is avoided.
The Liouville–von Neumann equation of motion of the reduced single electron
density matrix σ(t) is [141]
∂σðtÞ
i ¼ ½FðtÞ, σðtÞ; ð56Þ
∂t
where F(t) is the Fock matrix in DFT. The time-dependent electric dipole moment
μ(t), can be calculated by
^ ðt2 ; t1 Þψ ðt1 Þ:
ψ ðt2 Þ ¼ U ð58Þ
1
Throughout this chapter we mean one-electron reduced density matrix for density matrix unless
explicitly explained with another meaning.
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 309
where t is a time point and Δt is a time delay. Formally, the time evolution operator
can be expanded in time-ordered products:
ð tþΔt
^
Uðt þ Δt, tÞ ¼ T exp i ^
F ðτÞdτ
t
X
1 ð ð tþΔt ð tþΔt
ðiÞn tþΔt
¼ dτ1 dτ2 dτn T^ F^ ðτ1 ÞF^ ðτ2 Þ F^ ðτn Þ ; ð60Þ
n¼0
n! t t t
numerical difficulties is by using the Magnus (cumulant) expansion [1, 141, 155,
156]. The Magnus expansion of U(t þ Δt, t) is
ð tþΔt
Uðt þ Δt, tÞ ¼ T^ exp i FðτÞdτ ¼ eΩ1 þΩ2 þ ; ð61Þ
t
and
ð tþΔt
Ω1 ðt þ Δt, tÞ ¼ i FðτÞdτ; ð62Þ
t
ð tþΔt ð τ1
1
Ω2 ðt þ Δt, tÞ ¼ ½Fðτ1 Þ, Fðτ2 Þdτ2 dτ1 : ð63Þ
2 t t
One problem with using (66) in direct propagation is that F(t þ Δt/2) is unknown at
time t. F(t þ Δt/2) should be estimated by linearly extrapolating the F values at
previous times, or through some predictor-corrector technique [157]. However, the
latter breaks the time evolution symmetry. Alternatively, we can say backward time
propagation with the predictor-corrector cannot reproduce the original state of the
system at early time. The modified midpoint unitary transform method (MMUT)
[158, 159] maintains this time-reversibility. In this method, U(t þ Δt, t Δt) is
constructed from the eigenvectors C(t) and the eigenvalues ε(t) of F(t) at the time
midpoint:
Uðt þ Δt, t ΔtÞ ¼ exp½i 2Δt FðtÞ ¼ CðtÞexp½i 2Δt εðtÞC{ ðtÞ: ð67Þ
For other popular integrators such as the one combined split-operator with the
enforced time-reversal symmetry method, see [143]. Implementations of
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 311
where D is the dipole matrix and E(t) is the time-dependent external electric field.
An impulsive external electric field is used in the calculation:
μ~i j ðωÞ
α i j ð ωÞ ¼ ; ð71Þ
k
4πω Tr½αðωÞ
Sð ω Þ ¼ Im ; ð72Þ
c 3
where i ¼ x, y, z is the coordinate axis index and the summation runs over negative
and positive frequency domains to keep the external electric field real. Expansion of
the time-dependent dipole under this external field gives
312 Y. Zhang et al.
X
μi ðtÞ ¼ μ0i þ αi j ðω; ωÞE ωj eiωt þ
ω
1X
β ðωs ; ω1 , ω2 ÞEωj 1 Eωk 2 eiωs t þ ð74Þ
2ω1 , ω2 i jk
1 X
γ ðωs ; ω1 , ω2 , ω3 ÞEωj 1 Eωk 2 Eωl 3 eiωs t þ ;
6ω1 , ω2 , ω3 ijkl
be represented by loop diagrams where the ket moves first forward and then
backward to account for the bra [168]. The second protocol uses the density matrix
and can be represented by ladder diagrams [1]. The first protocol does not maintain
the bookkeeping of relative time ordering of bra and ket interactions and results in
n þ 1 basic terms for the nth order response. The second protocol fully keeps track
of time ordering. Both ket and bra move forward, yielding more 2n terms [1, 23]. 2
Even though the Casida (or CEO) equations of motion represent the reduced single
electron density matrix, this density matrix is simply used for parameterizing the
many electron wavefunction given by a single Slater determinant. The response
predicted by the equations of motion for the density matrix turns out to correspond
to the many-electron wavefunction rather than the density matrix [86, 89].
RT-TDDFT should have many advantages in nonlinear X-ray spectroscopy
simulation. Because it does not calculate individual states, it saves computing
time when many excited states are involved, which is the case for ultrashort
broadband X-ray pulse excitation. Direct propagation of the density matrix, only
involves occupied orbitals, so the computational scaling of RT-TDDFT is much
better than any SOS method [112], which involves a large number of virtual
orbitals. There are already many linear scaling algorithms both in both time [169]
and frequency domains [170] for RT-TDDFT in excited state calculations. Both
methods rely on the diagonal dominance of single-electron density matrices or
transition density matrices. The key issue is to calculate the highly nonlocal
exchange components in the popular hybrid density functionals efficiently, which
has only recently been addressed [171]. Unlike the perturbation method (to be
discussed in the next section), high order functional derivatives are not necessary in
RT-TDDFT calculations, so the well-behaved but complicated energy functionals,
such as the orbital-dependent functionals or the optimized effective potential (OEP)
functionals, can be readily used. In addition, RT-TDDFT has advantages for
nonlinear response because the calculations are no more difficult than for linear
response, whereas for the frequency domain methods such as SOS, they become
increasingly more complex for higher order response. Nuclear motions can also be
accounted for by Ehrenfest dynamics [159, 172]. RT-TDDFT offers a direct
simulation of nonlinear spectroscopy experiments with short pulses. However, we
still have some tradeoffs in using RT-TDDFT. Because individual excited states are
not available in RT-TDDFT, it is hard to interpret the spectral features. So far,
RT-TDDFT applications have been restricted to calculating standard dynamical
hyperpolarizabilities. Simulations of the signals presented in Sect. 2 constitute
challenges.
2
Each electronic oscillator which parameterizes the evolution of a single-electron density matrix
corresponds to a single electronic excited state appearing in the linear response regime. Nonlinear
SOS response calculations are then reformulated as sum-over-oscillator expressions, where mul-
tiply excited oscillators appear in the higher order responses leading to 2n terms. It is interesting
that the expressions for the TDHF CEO [89] (or equivalently the Casida TDDFT [86]) response
obtained from the equations of motion correspond to the wavefunction, not the density matrix.
314 Y. Zhang et al.
MCSCF core state calculations were first performed in the 1980s by Ågren
et al. [193–196], who studied the state-specific low-lying single and double core
hole (DCH) states of a series of small molecules, systematically investigated the
influence of correlation and relaxation effects on energies, and computed the
effective transition dipole moments (TDMs) between separately optimized
MCSCF states. The method was rediscovered about 20 years later [197]. In recent
years it was employed for work on DCH states and spectra of various small
molecules by Tashiro et al. [198–201]. Odelius et al. [202] first employed the
state-averaged RASSCF (SA-RASSCF) method [192] in conjunction with the
state-interaction treatment of SO effects [203] for L-edge XANES and RIXS
spectra of transition-metal-based complexes, and soon calculations were performed
for various other similar systems [204–207]. Hua et al. applied it to study the O
K-edge ASRS signals of furan conical intersections in the photo-induced ring-
opening reaction [208].
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 315
Table 1 SCF and MCSCF core hole orbital relaxation energies (eV) for single and double core
ionized states of formamide. Singlet energies are used although triplet energies are included in
parenthesis if different. Rebuilt based on [196]
Core hole SCF MCSCF
1
Single O1s 0.90 0.87
N1s1 0.88 0.86
C1s1 0.85 0.83
One-site double O1s2 0.28 0.38
N1s2 0.23 0.35
C1s2 0.17 0.34
Two-site double O1s1N1s1 1.79 1.72 (1.73)
O1s1C1s1 1.78 (1.76) 1.69
N1s1C1s1 1.74 1.68 (1.67)
Core hole calculations require a RAS1 space for the core hole orbital with fixed
occupation number. The RAS2 space (and RAS3 space, if necessary) is used for
valence correlation. Optimization of the MCSCF wavefunction with core holes
must avoid the variational collapse. A rigorous treatment of orbital relaxation can
be realized by a two-step procedure [193, 196]. One first freezes the orbital with
core hole (RAS1) and relaxes the rest; the frozen core orbital is relaxed in the
second step. Different optimization algorithms may be used for the two steps, for
example, combining the second-order norm-extended optimization (NEO) algo-
rithm [209, 210] and a straight Newton–Raphson (NR) algorithm. The effect of the
second step can be illustrated by the resulting energy change (i.e., the core orbital
relaxation energy). Table 1 gives an example for single and double core hole
ionization of formamide [196]. The core orbital relaxation energy is a few electron-
volts. For a single core hole, it is almost 1 eV; for a one-site double core hole, it is
even less; whereas for two-site double core hole, it reaches about 2 eV. This gives
an estimate of the effect of freezing the core orbital.
The optimization in step two in practice is more difficult to implement, espe-
cially when a state-averaged MCSCF is used, because in this case the object
function to optimize is more complicated. It is usually sufficient to skip step two.
The underlying physics is that the core orbital is well separated in energy from the
valence orbitals, so core orbital relaxation hardly influences the nature of the
valence orbitals, but mainly leads to a few electronvolts red shift of transition
energies. The calibration can be obtained by aligning the main peak in the calcu-
lated XANES spectrum to experiment (this also covers relativistic effects and basis
set incompleteness). Because the relaxation energies are similar at the SCF and
MCSCF levels (Table 1), one can also estimate the shift value at the simpler SCF
level.
316 Y. Zhang et al.
A practical way to perform the core hole calculations is to run a valence CASSCF
calculation first. One chooses the most important orbitals (e.g., localized or natural
orbitals around a breaking bond, 3d orbitals of an excited transition metal) which
have flexible occupations in the active space. The converged wavefunction serves
as the initial guess for the RASSCF calculation. The original active space is
included in the RAS2 space, and the excited core orbital is placed in the RAS1
space. Overall, one extra orbital (the core orbital) and two additional electrons (two
core electrons) are added to the active space for core hole calculations to achieve
more consistency in valence and core excited states calculations. Including addi-
tional virtual orbitals in the RAS3 space can help get a broader energy range for
easier comparison with experimental XANES spectra, especially when the ionic
potential is relatively high. For example, while studying the metal L2,3-edge spectra
from the np shell, the RAS2 space should be the 3d orbitals of the metal, and some
ligand orbitals can be included in the RAS3 space to represent the ligand-to-metal
or metal-to-ligand charge transfer effects [202].
Even though there are general guidelines for choosing the active space [189,
190], it is still necessary to verify by checking the convergence of energy and/or
spectra with the active space size. Figure 12 gives an example for the “UV
absorption” and “O1s XANES” spectra of a furan CoIn in the photo-induced
ring-opening reaction, obtained from a non-adiabatic molecular dynamics trajec-
tory [208]. Such spectra may not be directly observed (because the system is in a
superposition of valence states instead of the ground electronic state), but serve as
good tests for the active space. They also give an estimate of the accuracy of the
calculated transition dipole moments (see next section) which are essential for
simulation of the time-domain nonlinear X-ray signals. In Fig. 12a, 10-state-
averaged CASSCF was used, and, as expected, a smaller active space expands a
broader energy range. Increase of the active space introduces more states with
smaller oscillator strengths, and 10 electrons in 10 orbitals (10, 10) is sufficient to
obtain essentially converged UV spectra. The optimized valence state wavefunction
serves as the initial guess for core hole calculations. In Fig. 12b, 50 states were
calculated using SA-RASSCF. Because core states have a higher density of states
than valence states, additional states are needed to get an energy range of several to
10 eV for spectral usage. The 10 valence orbitals (RAS2) plus 1 core orbital (RAS1)
[labeled as “(12, 1/10/0)”. Twelve electrons include the 10 valence electrons and
2 electrons originally in the O1s orbital; numbers separated by slashes refer to the
sizes of RAS1, RAS2, and RAS3] can give converged O1s XANES. Such active
space settings are enough to obtain accurate and consistent electronic structure for
both the valence and core-excited state manifolds. Note that here the convergence is
much faster than the valence level. A test that includes 20 more orbitals in the RAS3
space “(6, 1/4/20)” shows that it can generate more states in the higher-energy
region (534–540 eV) and modify the fine structure of spectra.
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 317
a b
Fig. 12 Simulated (a) UV absorption and (b) O1s XANES spectra of a furan conical intersection
(snapshot at T ¼ 62.5 fs in trajectory 3 of the nonadiabatic MD simulation) at different active
spaces by using the state-averaged CASSCF or RASSCF method. Active spaces are labeled by
(n, m) in panel a or (n, m1/m2/m3) in panel b, where n, m are the number of electrons and orbitals in
the active space, and m1, m2, and m3 are the numbers of orbitals in the RAS1, RAS2, RAS3 spaces,
respectively. All core hole energies have been uniformly shifted by 3.05 eV. Inset in a: geometry
of the snapshot with C–O distances labeled in Å. Rebuilt based on [208]
MCSCF X m m E
Φ ¼ λ ϕ ; ð75Þ
m j j
j
MCSCF X e e
Θ ¼ κ θ ; ð76Þ
e i i
i
X X m * e D m E
ΦmMCSCF jμ^ jΘeMCSCF ¼ λ j κ i ϕ j jμ^ jθie : ð77Þ
i j
D E
Because the MOs of the two manifolds are non-orthogonal, terms of ϕ mj jμ^ jθie are
calculated using the L€owdin rules [67, 211]. Because this method is based on the
determinants, it has a lengthy expansion over configurations. The advantage is that
it is easy for programming and parallelization. This algorithm was employed to
compute the TDMs between the O1s core excited states and valence excited states
of furan [208].
A more efficient approach is the CAS or RAS state interaction (CASSI/RASSI)
method based on configuration state functions (CSFs) developed by Malmqvist and
Roos [212, 213]. Because a many-electron wavefunction can be equivalently
described in different sets of molecular orbitals, the orbitals are rotated (and the
coefficients are changed correspondingly) to be biorthonormal. We can then simply
use the Slater–Condon rule. Methods for including the spin-orbit (SO) coupling
have been developed [203]. This algorithm was employed in early effective TDMs
calculations between state-specific valence and core MCSCF states [195] and
widely used in recent L-edge XANES and RIXS calculations [202, 204–207].
Fig. 13 (a) Simulated ASRS signals of for furan during the passage of V1/V0 conical intersection
(T ¼ 27.5–33.0 fs in MD trajectory 4) by using the MCSCF method. From left to right, total signals
and contributions from ρ00 and ρ11 populations, and from ρ01 + ρ10 coherence (0 and 1 refer to
states V0 and V1). (b) Absolute amplitudes of the two constituting states. (c) NTOs [215–217] for
the two main peaks in the ASRS signals (denoted by “o” and “+”). Dominant hole and particle
orbitals are plotted with contour isovalue ¼ 0.08. Rebuilt based on [208]
320 Y. Zhang et al.
resulting signals are sensitive to the change in chemical and electronic structure
(Fig. 13a, left). Moreover, the contributions from ρ00 and ρ11 populations and ρ01
and ρ10 coherences may be separated. Below 29.0 fs the signals mainly come from
the V1 state and after 29.0 fs from the V0 state. At the transition point T ¼ 29.0 fs
both V0 and V1 populations have comparable contributions to the signals, and
coherence terms have larger contributions than at any other time. We had further
analyzed two major peaks in the total signals (denoted as “o” and “+”, the former is
always lower in energy) and found that they can be tracked by varying time-
dependent strength: peak “o” is stronger before T ¼ 29.0 fs, when peak “+”
becomes stronger. Both peaks correspond to transitions from the oxygen p orbital
to π* transitions, as indicated by the dominant natural transition orbital (NTO)
[215–217] pairs (Fig. 13c). As time increases, for peak “o” the hole orbital changes
from localized on the C–O bond to delocalized. For peak “+”, the particle orbital
changes from delocalized to localized on the C–O bond.
[219]. The GW/BSE approach has been successfully applied to XANES calcula-
tions of solids [225–231]. Until now, GW/BSE is still a theory of heavy solid state
flavor. Adapting it to a molecular theory is currently in progress [232–234].
The many-body Green’s function techniques with algebraic diagrammatic con-
struction (ADC) [235] was also used to study core excited states [236] and the
dynamics of core holes and particles [237, 238]. So far, most applications of the
above many-body methods have been made to XANES, and their use to calculate
core excited states is limited to small systems because of computational cost.
Nonlinear X-ray spectroscopy simulation of large systems is an important
future goal.
Other theoretical methods were designed primarily for metal L-edge calcu-
lations. These are more challenging than the ligand or metal K-edge calculations
because of multiplet effects, spin-orbital coupling, and metal-to-ligand or ligand-to-
metal charge transfer. Such transition metal-based systems have attracted broad
interest because of the numerous applications in biology (e.g., metallic enzyme
centers) and artificial light harvesting (e.g., dye-sensitized solar cells). Because the
2p ! 3d transitions are dipole-allowed, the metal L2,3-edge spectra can better
reflect the valence electron structure of the metal 3d orbitals which are more
essential to the chemistry. For metal K-edge spectra, the 1s ! 3d transitions are
dipole forbidden. These methods have mainly been applied to XANES and RIXS
spectra. Nonlinear X-ray spectra require accurate transition dipole moments. To
obtain these, both the valence and core-excited states must be treated with consis-
tent accuracy.
Early theoretical efforts on transition metal L-edge X-ray spectra were based on
semi-empirical methods developed by de Groot and coworkers [239, 240], namely
the crystal field multiplet (CFM) and the charge transfer multiplet (CTM) models.
These methods start with the SO-coupled multiplets of the excited metal atom and
include the effect of ligands using ligand field theory (LFT). Adjustable parameters
include the crystal field splitting and the charge transfer energy. In recent years,
there are developments in ab initio theory including the ab initio CTM based on
DFT-CI [241], ab initio multiplet ligand-field theory (MLFT) method with Wannier
orbitals [242], and method employing the Russel–Saunders coupling [243]. These
methods are usually computationally expensive and were employed for relatively
small systems with high symmetry. Neese and co-workers [244, 245] proposed an
efficient approach by combining DFT and the restricted-open-shell configuration
interaction singles (DFT/ROCIS). It introduces global empirical parameters for the
periodic table to scale the CI matrix and includes dynamic correlation and the SO
coupling effects. Excellent agreement with experiment was obtained for most
systems. Besides RASSCF, Odelius et al. [202] further tested the influence of
dynamic correlations by using the multiconfigurational second-order perturbation
theory (RASPT2) for a [NiII(H2O)6]2+ complex. TDDFT has also been tested for
this topic. Although it is widely believed that TDDFT is only valid for transition
metal-based systems with closed-shell (e.g., the low spin form of FeII, S ¼ O) but
not to those open-shells (e.g., the high spin form of FeII, S ¼ 2), it is still necessary
to examine its performance because of its high efficiency. It was found [246] that,
322 Y. Zhang et al.
for a SCO complex [FeII(tren(py)3)]2+, the TDDFT approach can predict the Fe
L3-edge XANES spectra of both the low-spin and high-spin complexes which agree
well with experiment [247–249]. However, such agreement depends on the system.
When the same procedure was applied to a variety of FeII and FeIII complexes with
different spin states, the accuracy did not always persist (Hua et al., submitted).
Other post-HF methods have been employed for simulating the K-edge XES,
XAS, or RIXS spectra of very small molecules. The intentions are to examine the
effect of electron correlation, to include the effect of shake-up/shake-off processes,
and/or to consider the influence of bond breaking. Ågren and coworkers had
reported early CI studies of XAS [250, 251] and XES [252, 253] spectra of CO
and N2. Recently, Neese et al. [254] studied the vibrationally-resolved RIXS
spectra of CO2 using the MRCI method. Coupled cluster (CC)-based methods
have been developed for core state calculations, including single-reference
equation-of-motion CC (EOM-CC) [255, 256], state-specific multireference CC
(SS-MRCC) [257], and open-shell symmetry-adapted cluster configuration inter-
action (SAC-CI) methods [258, 259].
excitation energies [262]. Core excitations share many similarities with long-range
charge transfer excitations, because the transition orbitals involved in these exci-
tations have negligible overlaps. This suggests that a similar strategy may be used
for designing energy functionals for core excitation as is done for long-range charge
transfer excitations. The range-separated functionals [263–277] are good choices.
In these functionals, the long-range part of the exchange energy is evaluated using
Hartree–Fock theory, and in the short-range DFT exchange is used. The 1/r12
operator is partitioned to two parts:
where erf(r) is the error function and μ is a parameter to control the separation of the
long- and short-ranges. We had employed this type of long-range corrected func-
tionals in SXRS simulations [92, 109].
To improve the performance of these long-range corrected functionals, Hartree–
Fock exchange should be introduced because core orbitals are very localized. This
can be done by adding a Gaussian correction term in the 1/r12 operator partition
scheme:
where the CSHF and CLHF parameters control the Hartree–Fock exchange contri-
bution in the long- and short-range. In the above equation the terms in the first box
are evaluated with Hartree–Fock exchange and the second with DFT exchange. The
resulting SRC1 functional is
324 Y. Zhang et al.
ESRC1
xc ¼ CSHF ESR-HF
x ðμSR Þ þ CLHF ELR-HF
x ðμLR Þ
ð81Þ
CSHF ESR-DFT
x ðμSR Þ þ CLHF ELR-DFT
x ðμLR Þ þ EDFT
xc :
ESRC1
xc ¼ CSHF ESR-HF
x ðμSR Þ þ CLHF ELR-HF
x ðμLR Þ
ð82Þ
ð1 CSHF ÞESR-DFT
x ðμSR Þ þ ð1 CLHF ÞELR-DFT
x ðμLR Þ þ EDFT
c :
The two functionals coincide when μSR ¼ μLR. Both functionals can predict light
atom core excitation energies with sub-1 eV accuracy [279].
Instead of partitioning the Coulombic operator in real space, Nakai and
coworkers had divided the electron density (orbitals) into the core and valence
groups. They proposed to use hybrid functionals with large Hartree–Fock exchange
components for core electrons and common hybrid functionals for valence elec-
trons. In the total energy expression, the hybrid scheme varies in the core–core,
core–valence, and valence–valence interaction terms. After numerical fitting of the
hybrid parameters, the resulting core–valence-(CV) B3LYP functional [280] gives
very good core excitation energies for light atoms (error less than 1 eV). In addition,
this scheme can be extended to Rydberg states [281]. Despite their success in core
excitation calculations, such orbital-specific functionals not only lead to a compli-
cated TDDFT implementation, but also bring some conceptual difficulties such as
the lack of a unique Fock operator.
In summary, exchange-correlation functionals specific for core excitations can
be designed along the same lines for long-range charge transfer excitations. The key
issue is that core excitation functionals cannot be too specific, because, in many
nonlinear X-ray spectroscopy experiments, both core and valence excited states are
involved and should be treated on the same footing. The core excitation functionals
discussed above should be tested in future nonlinear X-ray spectroscopy
simulations.
X ∂α^v X ∂2 α^v
α^v ¼ α^v0 þ ^i þ
Q ^Q
Q ^ þ ; ð83Þ
^
i ∂Qi
^ ^ i j
i j ∂Qi ∂Q j
where
rffiffiffiffiffiffiffiffiffiffiffiffi
^i ¼ h
Q a^ {i þ a^i : ð84Þ
2mi ωi
Here αv0 is the polarizability at the equilibrium geometry, and a^{i (^ ai) is the ith
phonon creation (annihilation) operator.
In analogy to this, the effective polarizability in electronic X-ray Raman scat-
tering ((18) or (20)) can also be expanded as
ð pÞ
X ð pÞ
X ð pÞ
α^ p ¼ α0 þ K i j c^ {i c^j þ Lijkl c^ {i c^j c^ {k c^l þ ; ð85Þ
i, j i, j, k , l
where c^{i (^
ci) is the creation (annihilation) operator for an electron in the ith valence
orbital and i, j, k, l are valence orbital indices. The super- and subscript ps indicate
the pulse inducing the polarizability and appear on the right hand side as a super-
ð pÞ
script for typographical convenience. α^ 0 is the effective polarizability responsible
for Rayleigh scattering. The electron–hole pairs created by c^{i c^j play the same role in
X-ray Raman scattering as do the vibrational normal modes Q ^i in optical Raman
scattering.
The direct use of the sum-over-state expression of the effective polarizability
((18) or (20)) would require calculation of large number of many-body states and
the corresponding state-to-state transition dipoles, which is tedious for large sys-
tems. X-Ray Raman signals may be calculated alternatively by solving equations of
motion for the reduced, single-electron density matrix in the valence space [87–89,
158]. The polarizability should then be expanded in valence electron creation and
annihilation operators (see (85)) avoiding the eigenstate expansion. Core excita-
tions can be included approximately in the calculation of the expansion coefficients,
but then the X-ray response is calculated in the valence space. We can view the
valence (occupied and unoccupied) orbitals as an open system that exchanges
electrons with the core space. This is formally analogous to molecular junctions
and the same methods can be applied to the X-ray signals [282–288].
ðpÞ
In the following we show how to calculate the expansion coefficients of α^ p, Kij
ðpÞ
and Lijkls , starting with a model Hamiltonian. In solid state applications it is
common to construct an electron-boson model Hamiltonian to represent all core
and valence excitations [289–293]. Model Hamiltonians can be obtained semi-
empirically by numerical fitting to experimental results, or from high level quantum
chemistry calculations of model systems. The model Hamiltonian is written in
ð{Þ
terms of the creation and annihilation operators c^ iðκÞ for single-particle valence
326 Y. Zhang et al.
n o
(core) orbitals ϕi(κ) (where the anticommutation relations c^{i ; ^cj ¼ δi j are satis-
fied). The following model Hamiltonian describes the valence and core orbitals and
their interaction:
X { X X X
H¼ εi c^ i c^ þ U ijkl c^ {i c^ {j c^k c^l þ εκ c^ {κ c^κ þ U i jκ c^ {i c^j c^κ c^ {κ ; ð86Þ
i ijkl κ i jκ
We thus write the Hamiltonian of the valence electrons in the absence of core holes
X { X
H^0 ¼ εi c^ i c^i þ U ijkl c^ {i c^ {j c^k c^l : ð88Þ
i ijkl
where κ labels core orbitals. We may now treat the core-valence interaction term in
(86) perturbatively to obtain an effective Hamiltonian for the valence electrons in
the presence of a single core hole or by recalculating the orbitals of the model in the
presence of the core hole potential. In the latter case we must start on the basis of
1-hole orbitals (denoted by ϕa, etc.), in which the valence Hamiltonian is
X X
~^ 0 ¼
H εa c^ {a c^a þ U abcd c^ {a c^ {b c^c c^d ; ð90Þ
a abcd
and then transform to the original orbitals with the overlap matrix T:
X ð*Þ ð{Þ
c^ ða{Þ ¼ tai c^ i : ð91Þ
i
X X
~^ 0 ¼
H ε~i j c^ {i c^j þ ~ijkl c^ { c^ { c^k c^l ;
U ð92Þ
i j
ij ijkl
which encode all time dependence and depend on the pulse parameters (collectively
denoted as Λ). This auxiliary function enters proportional to terms in (85) which
result from lth order expansion in the first propagator, mth order in the second, and
nth order in the third. Note that, because of the properties of the pulse and the
definition of the τ~ , the lower limits of integration may be truncated at τ~pi and the
upper limit of the dτ integration may be truncated at τ~p f . These auxiliary functions
vanish in the limit of t p ! 0 and, moreover, higher order auxiliary functions
(resulting from higher-order terms in the exponential expansion) vanish progres-
sively faster so that the ratio of successive functions also vanishes and the series
converges for sufficiently short pulses (pulses shorter than the inverse of any
relevant material energy scales). Further insight is gained by considering flat pulses,
ðlmnÞ
in which case fp is roughly proportional to a power of the pulse duration (Tp)N
where N ¼ l þ m þ n (neglecting factors of i and factorials). Because this procedure
naturally separates the parametric field dependence from the material operators, we
can write the expansion coefficients (Kij, etc.) generically without specifying a
328 Y. Zhang et al.
X
pulse envelope. To simplify these expressions, we define μi j μ* μ . A
κ iκ jκ
straightforward but lengthy calculation then yields the first few terms in (85):
ð pÞ
X X
α0 ¼ μii þ f ðp010Þ ðΛÞ μii ε~ii ; ð95Þ
i i
! !
X X
ð pÞ
Ki j ¼ μi j þ f ðp100Þ þ f ðp001Þ μll εi δi j þ U illj
l l !; ð96Þ
X X
f ðp010Þ μ jk ε~ik þ μki ε~k j þ ~kijl
μkl U
k kl
ð pÞ
Xn o
Lijkl ¼ ~iklm þ μmk U
f ðp001Þ μi j U kmml þ f ðp010Þ μ jm U ~mijl þ f ð100Þ μlk Uimmj ;
p
m
ð97Þ
where, in the first line, we have explicitly notated the dependence on the pulse
parameters via the auxiliary function f (omitting this dependence in further expres-
sions for brevity). Note that the above is an expansion up to first order roughly in the
product Tpε where ε is the material energy scale and higher order expansions
generate higher order terms in this quantity which contribute to every coefficient
in (85). Because the order of the expansion is determined by the sum l þ m þ n,
ð101Þ ð110Þ ð011Þ ð200Þ
expanding to second-order adds terms proportional to f p , f p , f p , f p etc.
Double core excitations or double excitations with a core hole and a valence hole
are directly probed by nonlinear X-ray four-wave mixing spectroscopy (e.g., photo
echo [48] and double-quantum-coherence [24]) and multidimensional SXRS [108]
experiments. Calculating double excitations is of essential importance in nonlinear
X-ray spectroscopy simulation. Double core states studied in X-ray two-photon
photoelectron spectroscopy can be considered as doubly core excited states of the
system with two electrons removed. These double core states are more sensitive to
the chemical environment and carry more pronounced electronic many-body
effects compared to singly core excited states [201, 294–297]. Strong double
excitations (shake-up) features can be found even in linear XANES signals [298].
A doubly excited state can be phenomenologically defined as an excited state
with energy close to the sum of two single excitation energies. This definition may be
misleading. First, near-degeneracy in energy does not mean the excited states share
the same character, and, second, there exists double excitations lower in energy than
any single excitation [299]. A proper definition for doubly excited states in quantum
chemistry relies on a single Slater determinant reference state. If an excited state can
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 329
Fig. 14 Comparison of the absolute values of the ONNO (left) and NONO (right) SXDQC (t3 ¼ 5 fs,
Ω2, Ω1) signals of formamide (structure shown on the top) from calculations combined ΔSCF with
REW-TDDFT. XANES spectra are shown in the marginals. Figure adapted from [24]
correlated systems [328]. XDQC signals exist only when single core excitations
interact with each other and doubly core-excited states are not simply an outer
product of two singly core-excited states. XDQC signals vanish for noninteracting
single core excitations.
We have simulated the XDQC signals of formamide at the N and O K-edges
[24]. In Fig. 14 we show the 2-D projections (we set the last time delay to 5 fs) of
the XDQC signals for different pulse orders (ONNO and NONO). The Ω1 axis
represents the N1s or O1s single excitation energies and the Ω2 axis represents the
N and O1s double excitation energies. The intensity of the 2-D XDQC signal at the
position (Ω1, Ω2) represents the correlation between the single excitation at Ω1 and
the double excitation at Ω2. The (shifted) XANES spectra are also placed in the
marginals to help the signal analysis. In the plot of the ONNO (O excitation first,
and then N excitation; see the left column of Fig. 14) signal we can find features
scattered along several diagonal straight lines, which indicates almost uniform
interaction between the O1s and N1s single excitations, so that the double excita-
tions are almost sums of the single excitations with constant shifts. This can also be
confirmed by comparing the energy differences of spectroscopic features in the
XANES and XDQC spectra. Take peak A and A0 in the N1s XANES spectrum
(right marginal of the ONNO signal) for an example. The energy difference
between these two peaks is 1.29 eV in XANES and 1.34 eV in the ONNO spectrum,
which are very close to each other. This means the first O1s excitation shifts all N1s
excitations in almost the same way, so the energy differences between peaks would
not change too much. The same linear pattern is not observed in the NONO signal
(the right column of Fig. 14). This tells us the second O1s excitations are affected
very differently from each other by the first N1s excitations. From the ONNO signal
we can also find that peak B in the N1s XANES only correlates weakly with peak C
in the O1s XANES when the O1s electron is excited first, whereas if the N1s
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 331
electron is excited first, the N1s excitation which corresponds to peak B heavily
affects the following O1s excitation which corresponds to peak C (a strong feature
connecting peak B and C in the NONO spectrum). By comparing XDQC signals of
different pulse orders, we can understand the subtle many-body correlations
between core excitations. Generally, X-ray four-wave mixing signals including
the XDQC signals can measure double core excited states directly, thus providing
new experimental tests for the accuracy of the electronic structure methods and
offering a way to visualize projections of the complicated many-electron wave
functions [25, 26].
X-Ray photons with energies higher than the ionization threshold can excite the
system to a metastable resonance state, and then this resonance decays to a cationic
species and a leaving electron. In spectroscopy, either the cationic species or the
leaving electron can be detected, both providing us with windows into the elec-
tronic structure of the parent neutral system.
In addition to excitation, molecular ionization is another way to trigger impul-
sively rich electronic and nuclear dynamics such as geometry relaxation, molecular
dissociation [329], charge migration [330–334], and radiation emission [335]. Ion-
ization has advantages over excitation in pumping the system because there is no
selection rule to restrict which ionization is allowed, and there is usually a lower
number of energetically accessible cationic states compared to the large number of
high energy exited neutral states in excitation experiments. Moreover, after ioni-
zation, we have the freedom to detect the cationic states or the ejected electron,
which can provide more information about the parent neutral species. X-Ray
ionization and fragmentation have been used to probe transient molecular structures
during a photoinduced chemical reaction process [336]. Recently we studied the
cationic states of the amino acid glycine prepared by a sudden N1s core ionization
produced by an attosecond X-ray pulse [337]. The created superposition of cationic
states is probed by 2D transient X-ray absorption (TXAS) and 3D ASRS. Our
simulated ASRS results reveal the complex coupling of the valence and core
excited states of the cation.
X-Ray photoelectron spectroscopy (XPS) [338] is a powerful technique for
probing the chemical compositions and electronic states of molecular systems
and materials. By measuring the kinetic energy of the ejected photoelectrons, the
electronic energy levels of the ionized species can be determined. The photo-
electrons also carry momenta. Additional information about the initial and final
electronic wavefunctions can be obtained by measuring photoelectron angular
distributions (PAD) [339–341]. TRPES [342, 343] is also used to monitor the
electronic structure changes during a chemical reaction.
332 Y. Zhang et al.
The total TRPES signal STRPES (Ek, t) can be expressed as a sum of the proba-
bilities of transitions from the neutral
electronic
state I to the cationic electronic
state F and free electron state η PI!Fη ðEk ; tÞ :
X
STRPES ðEk ; tÞ ¼ PI!Fη ðtÞ; ð98Þ
I , F, ηðEk Þ
where Ek is the kinetic energy of the ejected electron, and t is the observation time.
For the probability of a single transition, we have
2
PI!Fη ðtÞ / Eprobe ψ F ψ η μ^ ψ I δðhω Ek ðηÞ ΔEIF Þ; ð99Þ
where Eprobe and ω are the polarization vector and frequency of the probe light field,
respectively, ψI, ψF, and ψη are the wavefunction of the neutral state I, the cationic
state F, and the free electron state η, respectively, and ΔEIF ¼ EF EI is the
ionization energy corresponding to the neutral state I and cationic state F. Here
we employ the sudden ionization approximation [344, 345], in which the ionization
process is very fast and the leaving electron does not interact with the cation, so the
final state function can be written as a product of ψ F and ψ η. We can also assume
that ψ η is orthogonal with ψ I, so (99) can be recast as
D 2
PI!Fη ðtÞ / ψ η Eprobe μ^ ψ IF δðhω Ek ðηÞ ΔEIF Þ: ð100Þ
where N is the number of electrons in the system. If the polarization of the probe
field and the angular distribution of the photoelectron is ignored, (100) can be
simplified as [346, 347]
D 2 2
PI!Fη ðtÞ / ψ IF μ
η δðhω Ek ðηÞ ΔEIF Þ; ð102Þ
where μη is an appropriate average value for the transition dipole matrix element
between the Dyson orbital ψ DIF and free electron state ψ η. If all free electron states η
with a kinetic energy Ek are summed over in (100), we have
D 2
where μη ðEk Þ represents an average transition dipole matrix element corresponding
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 333
to a kinetic energy Ek, and ρ(Ek) is the density of the free electron state at the kinetic
energy Ek. Photoelectron experiments tell us that usually in the low kinetic energy
2
region, μ η ðEk Þ ρðEk Þ const: [348], so (103) can be further simplified as
D 2
PI!F ðEk ; tÞ / ψ IF δðhω Ek ΔEIF Þ; ð104Þ
where wIF is some weighting factor because different PI!F ðEk ; tÞ has different
prefactors in (104). In a simplified treatment we can set all wIFs to be equal [349], or
treat them as adjustable parameters [347]. Accurate determination of those
weighting factors requires complicated electron-molecule scattering calculations
[350], which is beyond the scope of this chapter. Recently, we extended 2D TRPES
technique (see (105)) to multidimension by adding more pump pulses before the
probe ionization pulse [351]. The TRPES of thioflavin T in its photoisomerization
has also been studied with ab initio molecular dynamics (AIMD) and TDDFT
simulations [349].
The Dyson orbital defined in (101) can be considered as the diagonal element of
the one-electron reduced transition density matrix between the neutral and the
cationic states. Dyson orbitals are generally not normalized. Their norms reflect
the one-electron character of the ionization process. In the simplest case when the
neutral and cationic states are well described by Hartree–Fock orbitals and the
Koopmans theorem applies, the Dyson orbitals reduce to the canonical Hartree–
Fock orbitals and their norms are one. Dyson orbitals are solutions of an effective
single-particle equation with ionization energies as their eigenvalues [352–
354]. Dyson orbitals are widely used in calculating Compton profiles [355, 356]
and electron momentum spectra [357], and interpreting orbital imaging experi-
ments [358–360]. Krylov and coworkers [346] describe an implementation of
Dyson orbital calculation at the coupled cluster singles and doubles (CCSD) or
EOM-CCSD level of theory.
X-Ray photons often bring the molecule into a superexcited state (excitation
above the ionization threshold, or resonance). These resonances are usually short-
lived with strong coupling with the continuum leading to the final ionization or
dissociation of the system. Generally, resonances can be divided into shape reso-
nance and Feshbach type [361, 362]. Resonances are ubiquitous in radiation
damage studies of biomolecules [363], molecular electronic device design [364],
attosecond pulse generation [365], and X-ray ionization [13].
Resonances may not be captured by conventional quantum chemistry methods
because of their unbound nature and lack of variational principle. To compute these
unbound states with finite lifetimes, one must use a non-Hermitian Hamiltonian
[366]. One approach for calculating resonance is the complex absorbing potential
334 Y. Zhang et al.
(CAP) method [367, 368]. In the method, a complex potential is added to the
exterior region of the metastable system to absorb the scattering electron and
makes the wave function square-integrable, i.e.,
^η ¼ H
H ^ iηW; ð106Þ
r0 ! reiθ : ð107Þ
Here θ is the rotation angle. As in the CAP method, trajectory calculations are
necessary to find an optimal θ value. However, because this is the only parameter to
be determined, the degree of uncertainty is greatly reduced compared to the CAP
calculations. A DFT combined with the complexed scaling method has been
developed for resonances [376–379] and has been used to study Stark ionization
of atoms and molecules [380]. The trajectory calculation becomes tedious when the
system is large. Moreover, there are still some fundamental questions needs to be
answered in extending DFT to resonance. For example, the complex version of the
v-representability problem.
Quantum chemistry method development for resonance is still in its infancy.
Most applications so far are for resonance energy and lifetime calculations of model
or very small atomic and molecular systems. Recently, a non-Hermitian
RT-TDDFT study of near and above ionization excitations of small molecules
was reported [381]. In this study an absorption boundary condition was used to
emulate the continuum. This scheme has a potential to be used in the future for
X-ray ionization and photoelectron spectroscopy simulations.
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 335
In the Frank–Condon (FC) region, nuclear motion acts as a bath for electronic
transitions. Including this effect introduces high-resolution fine structure to the
recorded spectra. Theoretical simulations are usually under the Born–Oppenheimer
(BO) and the harmonic oscillator approximations, which are usually good in the FC
region. The potential energy surfaces (PESs) of the ground and excited states are
well separated. The simplest approach to account for vibronic coupling is the linear
coupling model (LCM). It assumes the mode-k PES of excited state has the same
curvature as that of the ground state but only shifted by a displacement. This
approach is efficient and can be applied to medium and large molecules. It has
been well illustrated in various linear and nonlinear X-ray spectroscopy calcu-
lations including the XPS, XANES, RIXS, and SXRS spectra (see, e.g., [246,
382–388]). A more rigorous method is to include the Duschinsky rotation and/or
the non-Condon effects. For time-domain nonlinear spectroscopy, the nuclear and
electronic coordinates are mixed together. The response function can be evaluated
via the cumulant expansion [1] till truncated order. The vibrationally resolved
SXRS spectra were studied by Hua et al. [387] combining the LCM and cumulant
expansion till the second-order. With the inclusion of vibronic coupling, a faster
decay in the time domain signals, and new splitting and shoulder structures in the
frequency-domain were observed.
In this chapter we have surveyed some typical nonlinear X-ray spectroscopy signals
and the quantum chemistry methods used for their simulation. Because of their
balance in accuracy and computational cost, DFT/TDDFT methods are commonly
used in excited state calculations. With the fast development of new exchange-
correlation functionals and linear scaling algorithms, these methods provide a most
valuable quantum chemistry tool for nonlinear X-ray spectroscopy simulation.
DFT/TDDFT often provides an adequate zero order electronic structure at reason-
able cost, which paves the way for the application of high level methods. DFT/
TDDFT-based semiempirical methods such as density functional tight binding
(DFTB) [389–392] or time-dependent density functional tight binding (TDDFTB)
[393, 394], and their linear scaling forms [395] have been shown useful in spectro-
scopy simulations of large systems.
DFT/TDDFT works well in many cases but fails for double excitations, long-
range charge transfer excitations, and conical intersections. Much effort has been
made to address these difficulties by designing more elaborate functionals and
schemes. We believe that rather than putting the burden on the functionals, it
makes more sense to use DFT/TDDFT results as fast zero-order inputs to high
level wave-function approaches and many-body techniques. The recent
336 Y. Zhang et al.
Acknowledgements This work was supported by the U.S. Department of Energy, Office of
Science, Basic Energy Sciences under Award #DE-FG02-04ER15571. Yu Zhang and Weijie
Hua acknowledge a postdoctoral fellowship by the Department of Energy grant. We also gratefully
acknowledge the support of the National Science Foundation (Grant CHE-1361516), and the
National Institutes of Health (Grant GM-59230). Helpful comments from Dr. Sergei Tretiak,
Dr. Niri Govind and Prof. Kieron Burke are greatly appreciated.
References
22. Abramavicius D, Palmieri B, Voronine DV, Sanda F, Mukamel S (2009) Chem Rev 109(6):
2350
23. Mukamel S, Rahav S (2010) In: Berman P, Arimondo E, Lin C (eds) Advances in atomic,
molecular, and optical physics, vol 59. Elsevier, Amsterdam, pp 223–263
24. Zhang Y, Healion D, Biggs JD, Mukamel S (2013) J Chem Phys 138(14):144301
25. Mukamel S, Oszwaldowski R, Yang L (2007) J Chem Phys 127(22):221105
26. Li Z, Abramavicius D, Mukamel S (2008) J Am Chem Soc 130(11):3509
27. Weiner AM, Silvestri SD, Ippen EP (1985) J Opt Soc Am B 2(4):654
28. Tanaka S, Mukamel S (2002) Phys Rev Lett 89(4):043001
29. Biggs JD, Zhang Y, Healion D, Mukamel S (2012) J Chem Phys 136(17):174117
30. Dorfman KE, Fingerhut BP, Mukamel S (2013) J Chem Phys 139(12):124113
31. Dorfman KE, Fingerhut BP, Mukamel S (2013) Phys Chem Chem Phys 15(29):12348
32. Fingerhut BP, Dorfman KE, Mukamel S (2014) J Chem Theory Comput 10(3):1172
33. Kossakowski A (1972) Rep Math Phys 3(4):247
34. Lindblad G (1976) Commun Math Phys 48(2):119
35. Ando H, Fingerhut BP, Dorfman KE, Biggs JD, Mukamel S (2014) J Am Chem Soc 136(42):
14801
36. Brixner T, Stiopkin IV, Fleming GR (2004) Opt Lett 29(8):884
37. Scheurer C, Mukamel S (2001) J Chem Phys 115(11):4989
38. Tian P, Keusters D, Suzaki Y, Warren WS (2003) Science 300(5625):1553
39. Turner DL (2007) Encyclopedia of magnetic resonance. Wiley, New York
40. Ernst RR, Bodenhausen G, Wokaun A (1990) Principles of nuclear magnetic resonance in
one and two dimensions. Oxford University Press, New York
41. Stolow A, Bragg AE, Neumark DM (2004) Chem Rev 104(4):1719
42. Seel M, Domcke W (1991) J Chem Phys 95(11):7806
43. McFarland B, Farrell J, Miyabe S, Tarantelli F, Aguilar A, Berrah N, Bostedt C, Bozek J,
Bucksbaum P, Castagna J, Coffee RN, Cryan JP, Fang L, Feifel R, Gaffney KJ, Glownia JM,
Martinez TJ, Mucke M, Murphy B, Natan A, Osipov T, Petrović VS, Schorb S, Schultz T,
Spector LS, Swiggers M, Tenney I, Wang S, White JL, White W, Gühr M (2014) Nat
Commun 5:4235
44. Bennett K, Kowalewski M, Mukamel S (2014) Faraday Discuss. doi:10.1039/C4FD00178H
45. Hua W, Gao B, Luo Y (2012) Prog Chem 24(06):964
46. Nozières P, De Dominicis CT (1969) Phys Rev 178:1097
47. Schwarz WHE, Buenker RJ (1976) Chem Phys 13:153
48. Schweigert IV, Mukamel S (2007) Phys Rev Lett 99:163001
49. Schweigert IV, Mukamel S (2007) Phys Rev A 76:012504
50. Schweigert IV, Mukamel S (2008) J Chem Phys 128:184307
51. Schweigert IV, Mukamel S (2008) Phys Rev A 78:052509
52. Healion DM, Schweigert IV, Mukamel S (2008) J Phys Chem A 112:11449
53. Slater JC (1972) Adv Quantum Chem 6:1
54. Slater JC, Johnsson KH (1972) Phys Rev B 5:844
55. Stener A, Lisini M, Decleva P (1995) Chem Phys 191:141
56. Triguero L, Pettersson LGM, Ågren H (1998) Phys Rev B 58:8097
57. Brena B, Siegbahn PE, Ågren H (2012) J Am Chem Soc 134(41):17157
58. Hunt WJ, Goddard WA III (1969) Chem Phys Lett 3:414
59. Ågren H, Carravetta V, Vahtras O, Pettersson LGM (1994) Chem Phys Lett 222:75
60. Ågren H, Carravetta V, Vahtras O, Pettersson LGM (1997) Theor Chem Acc 97:14
61. Prendergast D, Galli G (2006) Phys Rev Lett 96:215502
62. Gilbert ATB, Besley NA, Gill PMW (2008) J Phys Chem A 112:13164
63. Sleeman DH (1968) Theor Chim Acta 11(2):135
64. Saunders V, Hillier I (1973) Int J Quantum Chem 7(4):699
65. Guest M, Saunders V (1974) Mol Phys 28(3):819
66. Ågren H, Carravetta V, Pettersson LG, Vahtras O (1996) Phys Rev B 53(23):16074
338 Y. Zhang et al.
Cossi M, Rega N, Millam JM, Klene M, Knox JE, Cross JB, Bakken V, Adamo C, Jaramillo J,
Gomperts R, Stratmann RE, Yazyev O, Austin AJ, Cammi R, Pomelli C, Ochterski JW,
Martin RL, Morokuma K, Zakrzewski VG, Voth GA, Salvador P, Dannenberg JJ,
Dapprich S, Daniels AD, Farkas Ö, Foresman JB, Ortiz JV, Cioslowski J, Fox DJ (2009)
Gaussian 09. Gaussian Inc, Wallingford CT
108. Biggs JD, Zhang Y, Healion D, Mukamel S (2013) J Chem Phys 138(14):144303
109. Zhang Y, Biggs JD, Govind N, Mukamel S (2014) J Phys Chem Lett 5:3656
110. Chen LX, Zhang X (2013) J Phys Chem Lett 4(22):4000
111. Andrade X, Botti S, Marques MAL, Rubio A (2007) J Chem Phys 126(18):184106
112. Orr B, Ward J (1971) Mol Phys 20(3):513
113. van Gisbergen SJA, Snijders JG, Baerends EJ (1998) J Chem Phys 109(24):10644
114. Larsen H, Jørgensen P, Olsen J, Helgaker T (2000) J Chem Phys 113(20):8908
115. Iwata JI, Yabana K, Bertsch GF (2001) J Chem Phys 115(19):8773
116. Furche F (2001) J Chem Phys 114(14):5982
117. Hait Heinze H, Della Sala F, Gorling A (2002) J Chem Phys 116(22):9624
118. Salek P, Vahtras O, Helgaker T, Ågren H (2002) J Chem Phys 117(21):9630
119. Ye A, Autschbach J (2006) J Chem Phys 125(23):234101
120. Ye A, Patchkovskii S, Autschbach J (2007) J Chem Phys 127(7):074104
121. Wang F, Yam CY, Chen G (2007) J Chem Phys 126(24):244102
122. Shavitt I, Bartlett RJ (2009) Many-body methods in chemistry and physics: MBPT and
coupled-cluster theory. Cambridge University Press, Cambridge
123. Norman P, Bishop DM, Jensen HJA, Oddershede J (2001) J Chem Phys 115(22):10323
124. Norman P, Bishop DM, Jensen HJA, Oddershede J (2005) J Chem Phys 123(19):194103
om U, Norman P, Carravetta V, Ågren H (2006) Phys Rev Lett 97(14)
125. Ekstr€
126. Linares M, Stafstr€om S, Rinkevicius Z, Ågren H, Norman P (2011) J Phys Chem B 115(18):5096
127. Norman P, Ruud K (2006) In: Leszczynski J (ed) Challenges and advances in computational
chemistry and physics, vol 1. Springer, Dordrecht, pp 1–49
128. Bertrand JB, W€orner HJ, Bandulet HC, Bisson É, Spanner M, Kieffer JC, Villeneuve DM,
Corkum PB (2011) Phys Rev Lett 106(2):023001
129. Niikura H, W€ orner HJ, Villeneuve DM, Corkum PB (2011) Phys Rev Lett 107(9):093004
130. W€orner HJ, Bertrand JB, Fabre B, Higuet J, Ruf H, Dubrouil A, Patchkovskii S, Spanner M,
Mairesse Y, Blanchet V, Mével E, Constant E, Corkum PB, Villeneuve DM (2011) Science
334(6053):208
131. Ranitovic P, Hogle CW, Riviere P, Palacios A, Tong XM, Toshima N, Gonzalez-Castrillo A,
Martin L, Martin F, Murnane MM et al (2014) Proc Natl Acad Sci 111(3):912
132. Bergues B, Kübel M, Johnson NG, Fischer B, Camus N, Betsch KJ, Herrwerth O,
Senftleben A, Sayler AM, Rathje T et al (2012) Nat Commun 3:813
133. Neppl S, Ernstorfer R, Bothschafter EM, Cavalieri AL, Menzel D, Barth JV, Krausz F,
Kienberger R, Feulner P (2012) Phys Rev Lett 109(8):087401
134. Krausz F, Stockman MI (2014) Nat Photonics 8(3):205
135. Mikosch J, Boguslavskiy A, Wilkinson I, Spanner M, Patchkovskii S, Stolow A (2013)
Phys Rev Lett 110(2):023004
136. Leone SR, McCurdy CW, Burgd€orfer J, Cederbaum LS, Chang Z, Dudovich N, Feist J,
Greene CH, Ivanov M, Kienberger R et al (2014) Nat Photonics 8(3):162
137. Frumker E, Kajumba N, Bertrand JB, W€orner HJ, Hebeisen CT, Hockett P, Spanner M,
Patchkovskii S, Paulus GG, Villeneuve DM et al (2012) Phys Rev Lett 109(23):233904
138. Ivanov M, Smirnova O (2011) Phys Rev Lett 107(21):213605
139. Shafir D, Soifer H, Bruner BD, Dagan M, Mairesse Y, Patchkovskii S, Ivanov MY,
Smirnova O, Dudovich N (2012) Nature 485(7398):343
140. Mairesse Y, Higuet J, Dudovich N, Shafir D, Fabre B, Mével E, Constant E, Patchkovskii S,
Walters Z, Ivanov MY et al (2010) Phys Rev Lett 104(21):213601
141. Jakowski J, Morokuma K (2009) J Chem Phys 130(22):224106
142. Kosloff R (1988) J Phys Chem 92(8):2087
340 Y. Zhang et al.
324. Knippenberg S, Starcke JH, Wormit M, Dreuw A (2010) Mol Phys 108:2801
325. Nakayama K, Nakano H, Hirao K (1998) Int J Quantum Chem 66:157
326. Kurashige Y, Nakano H, Nakao Y, Hirao K (2004) Chem Phys Lett 400:425
327. Zhuang W, Abramavicius D, Mukamel S (2005) Proc Natl Acad Sci 102(21):7443
328. Kim J, Mukamel S, Scholes GD (2009) Acc Chem Res 42(9):1375
329. Fausti D, Tobey RI, Dean N, Kaiser S, Dienst A, Hoffmann MC, Pyon S, Takayama T,
Takagi H, Cavalleri A (2011) Science 331:189
330. Schinke R (1993) Photodissociation dynamics: spectroscopy and fragmentation of small
polyatomic molecules. Cambridge University Press, Cambridge
331. Weinkauf R, Schanen P, Metsala A, Schlag EW, Bürgle M, Kessler H (1996) J Phys Chem
100(47):18567
332. Cederbaum L, Zobeley J (1999) Chem Phys Lett 307(3–4):205
333. Breidbach J, Cederbaum LS (2005) Phys Rev Lett 94(3):033901
334. Remacle F, Levine RD (2006) Proc Natl Acad Sci U S A 103(18):6793
335. Breidbach J, Cederbaum LS (2007) J Chem Phys 126(3):034101
336. Kuleff AI, Cederbaum LS (2011) Phys Rev Lett 106(5):053001
337. Petrović VS, Siano M, White JL, Berrah N, Bostedt C, Bozek JD, Broege D, Chalfin M,
Coffee RN, Cryan J, Fang L, Farrell JP, Frasinski LJ, Glownia JM, Guhr M, Hoener M,
Holland DMP, Kim J, Marangos JP, Martinez T, McFarland BK, Minns RS, Miyabe S,
Schorb S, Sension RJ, Spector LS, Squibb R, Tao H, Underwood JG, Bucks-baum PH (2012)
Phys Rev Lett 108(25):253006
338. Zhang Y, Biggs JD, Hua W, Mukamel S, Dorfman KE (2014) Phys Chem Chem Phys 16:
24323
339. van der Heide P (2012) X-Ray photoelectron spectroscopy. Wiley, Hoboken, NJ
340. Wang K, McKoy V (1995) Annu Rev Phys Chem 46(1):275
341. Reid KL (2003) Annu Rev Phys Chem 54(1):397
342. Sanov A (2014) Annu Rev Phys Chem 65(1):341
343. Wu G, Hockett P, Stolow A (2011) Phys Chem Chem Phys 13(41):18447
344. Suzuki T (2012) Int Rev Phys Chem 31(2):265
345. Pickup B, Goscinski O (1973) Mol Phys 26(4):1013
346. Pickup BT (1977) Chem Phys 19(2):193
347. Melania Oana C, Krylov AI (2007) J Chem Phys 127(23):234106
348. Thompson AL, Martı́nez TJ (2011) Faraday Discuss 150:293
349. Rabalais JW (1977) Principles of ultraviolet photoelectron spectroscopy. Wiley, New York
350. Ren H, Fingerhut BP, Mukamel S (2013) J Phys Chem A 117(29):6096
351. Lucchese R, Raseev G, McKoy V (1982) Phys Rev A 25(5):2572
352. Rahav S, Mukamel S (2010) Phys Rev A 81(6):063810
353. Cederbaum LS, Domcke W (1977) Adv Chem Phys 36:205
354. Ortiz J (1999) Adv Quantum Chem 35:33
355. Linderberg J, Öhrn Y (2004) Propagators in quantum chemistry, 2nd edn. Wiley, Hoboken
356. Kaplan I, Barbiellini B, Bansil A (2003) Phys Rev B 68(23):235104
357. Barbiellini B, Bansil A (2004) J Phys Chem Solids 65(12):2031
358. Miao YR, Ning CG, Liu K, Deng JK (2011) J Chem Phys 134(20):204304
359. Murray R, Spanner M, Patchkovskii S, Ivanov MY (2011) Phys Rev Lett 106(17):173001
360. Saliéres P, Maquet A, Haessler S, Caillat J, Taı̈eb R (2012) Rep Prog Phys 75(6):062401
361. Mignolet B, Kus T, Remacle F (2013) In: Grill L, Joachim C (eds) Imaging and manipulating
molecular orbitals. Springer, Berlin, pp 41–54
362. Schulz G (1973) Rev Mod Phys 45(3):378
363. Simons J (1984) In: Truhlar DG (ed) Resonances in electron-molecule scattering van der
Waals complexes, and reactive chemical dynamics. American Chemical Society,
Washington, DC, pp 3–16
364. Sanche L (2009) Nature 461(7262):358
365. Goyer F, Ernzerhof M, Zhuang M (2007) J Chem Phys 126(14):144104
Nonlinear Spectroscopy of Core and Valence Excitations Using Short X-Ray. . . 345
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
2 Protocol to Determine the 0–0 Energies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
2.1 Gas Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
2.2 Condensed Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
2.3 Further Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
3 Benchmarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
D. Jacquemin (*)
CEISAM, UMR CNRS 6230, Université de Nantes, 2, rue de la Houssinière, 44322 Nantes,
France
Institut Universitaire de France, 103, blvd Saint-Michel, 75005 Paris, France
e-mail: Denis.Jacquemin@univ-nantes.fr
C. Adamo
IRCP UMR CNRS 8247, ENSCP-Chimie ParisTech, 11, rue Pierre et Marie Curie, 75005
Paris, France
Institut Universitaire de France, 103, blvd Saint-Michel, 75005 Paris, France
348 D. Jacquemin and C. Adamo
1 Introduction
In this section we present a popular approach to compute the 0–0 energies with
TD-DFT. This also allows us to define a series of different energies which are
subsequently used, and to propose an easy-to-follow protocol to obtain all the
relevant parameters which are represented in Fig. 1, in which RGS and RES stand
for the optimal geometries of the ground- and excited-states, respectively, whereas
EGS and EES are the total energies of these two states. Following [15], we first
explain the more straightforward gas-phase situation before extending the protocol
to the condensed phase.
In the gas phase, the vertical absorption can simply be defined as the difference
between the ES and GS energies at the optimal ground-state geometry,
Everta ¼ EES RGS EGS RGS ; ð1Þ
whereas the vertical fluorescence is the corresponding data estimated at the optimal
geometry of the relevant excited-state,
350 D. Jacquemin and C. Adamo
E vert-a
E vert-f
E adia
E 0-0
et al. [15]. Copyright 2012,
American Chemical Society
GS
E GS(R ES)
Ereorg-GS
EZPVE(RGS)
E GS(R GS)
RGS R ES Geometry
Evert f ¼ EES RES EGS RES : ð2Þ
1 vert f
1
In this latter equation the first term tends to be dominant, and, in a first crude
approximation the second term can be neglected. Indeed, the second term is the
difference of reorganization energies between the two considered states, which is
significant only when there is a strong difference between RGS and RES. Next, one
needs to determine the difference of zero-point vibrational energy (ZPVE) between
the ES and GS,
ΔEZPVE ¼ EZPVE RES EZPVE RGS ; ð5Þ
We note that ΔEZPVE is almost systematically negative, as the PES of the ES tends
to be flatter than its GS counterpart and, consequently, E00 is generally smaller
than Eadia. As stated above, E00 can be directly compared to the absorption-
fluorescence crossing point for solvated molecules, and it subsequently offers a
much more solid basis for theory–experiment comparisons than Everta, which often
has no straightforward experimental counterpart.
for emission. For the former phenomenon, one starts from an eq GS and goes to a
neq ES, whereas for the latter phenomenon, the ES is in equilibrium whereas the GS
is in non-equilibrium, and a proper modeling of the latter process requires quite
advanced computational approaches [30–32]. Differences between eq and neq
vertical transition energies can be significant in polar solvents [26]. By definition,
both the adiabatic and 0–0 energies are equilibrium properties as they correspond to
a transition between two states at their respective minima:
352 D. Jacquemin and C. Adamo
1 verta
The rationale for this correction can be obtained by examining (4). Indeed, in (10),
the only approximations are the neglect of the difference between non-equilibrium
and equilibrium environmental effects on the difference between the reorganization
energies of the two states, a very small contribution, and the consideration of
equilibrium limit during the computation of ΔEZPVE, but the eq–neq variations
for this average term are generally trifling.
The most popular approach for modeling solvent effects is the Polarizable Contin-
uum Model (PCM) which treats the environment as a structureless material
presenting the macroscopic properties of the actual solvent. The solute is embedded
in a cavity inside this solvent, and charges located on the surface of this cavity are
determined self-consistently to account for the electrostatic interactions between
the solute and the solvent [26]. We briefly describe here the different variations of
the PCM model which have been developed for ES. In the (simplest) linear
response (LR) model [28, 29], the GS-to-ES transition densities are used to deter-
mine the variations of the charges localized on the cavity when the solute changes
its electronic configuration. In the corrected linear response (cLR) [30], the
one-particle TD-DFT density matrix (the actual density of the ES within the
selected approximation) is used in a perturbative approach, to evaluate the changes
of the charges of the cavity when the solute changes electronic state [30]. The use of
the one-particle TD-DFT density, rather than the transition density, advantageously
allows one to account for orbital relaxation, and this density is also used in the two
Computational Molecular Electronic Spectroscopy with TD-DFT 353
self-consistent approaches, namely the state specific (SS) [31] and the vertical
excitation model (VEM) [32] approaches. One of the principal differences between
the two approaches is that the former implies a modification of the GS reference
during the self-consistent process, whereas the latter does not. When the change of
polarity of the chromophore between the GS and ES is large, e.g., for charge-
transfer (CT) transitions, going beyond the LR-PCM approximation is
recommended, though the most adequate model in that case remains a matter of
debate [34, 35].
We underline that, although all these approaches can be used to determine EES
analytically, analytic gradients (and hence efficient access to RES) are only avail-
able with the LR approach [20]. Subsequently, a popular approach is to determine
Everta , Evert f , and Eadia with one of the three refined PCM approaches (cLR, SS,
and VEM) on geometries computed within the LR-PCM model. Likewise, ΔEZPVE
is often calculated at the LR-PCM level, so that the results of (10) are generally
obtained with mixed environmental models, the energy (geometry and vibrations)
being obtained with a refined (simpler) PCM level of theory [33].
Besides TD-DFT, there is a wide panel of alternative and (very) accurate ab initio
methods with, on the one hand, multi-reference approaches, e.g., Complete Active
Space second-order Perturbation Theory (CAS-PT2) [36] and Multi-Reference
Configuration Interaction (MR-CI) [37], and, on the other hand, single-reference
(highly-)correlated schemes, e.g., Equation-of-Motion Coupled Cluster (EOM-CC)
[38–41], Symmetry Adapted Cluster CI (SAC-CI) [42], Algebraic Diagrammatic
Construction (ADC) [43], and Configuration Interaction singles with a perturbative
correction for double excitations [CIS(D)] [44, 45]. Despite the rapid developments
of these approaches and the implementations of efficient protocols (e.g., the reso-
lution of identity scheme), their less favorable scalings with system size than
TD-DFT generally limit their applications to vertical calculations but for rather
small molecules. Therefore, it has been proposed to combine TD-DFT’s ES geom-
etries and vibrations to Everta and Evert f obtained with these more advanced
approaches. In the protocol proposed by Goerigk and Grimme [13], the experimen-
tal 0–0 energies are first transformed into “experimental” vertical energies by
applying successive corrections for solvation, vibration, and geometrical reorgani-
zation effects determined with TD-DFT. Alternatively, one can determine AFCP
energies through (10) and next correct them through wavefunction (Ψ) vertical
calculations performed on the DFT GS and TD-DFT ES geometries [46, 47]. For
approaches that can only be used for gas-phase vertical transition energies, the
corrected AFCP energy simply becomes
354 D. Jacquemin and C. Adamo
AFCP
EBE ðneqÞ ¼ ETDDFT
AFCP
ðneqÞ þ EΨadia ðgasÞ ETDDFT
adia
ðgasÞ ; ð13Þ
where BE stands for best estimates. Compared to (10), (13) only requires, for the
TD-DFT part, two additional vertical gas-phase calculations (one for each optimal
geometry) and the time-limiting step generally remains the wavefunction compu-
tation. The accuracy of the results obtained with (13) of course depends not only on
the quality of the wavefunction model but also partly on the “starting” accuracy
obtained with TD-DFT. When TD-DFT strongly underestimates the transition
energies, using (13) could be less efficient.
Once the GS and ES vibrational signatures have been determined, for instance in
the course of computing the ΔEZPVE contribution to E00 , it is possible to obtain
vibronic couplings and hence to estimate absorption and emission band shapes.
This requires the calculation of the coupling factors between the different vibra-
tional states of the GS and the ES, a task often achieved by the Franck–Condon (for
strongly dipole allowed transitions) and/or Herzberg–Teller (for forbidden or
weakly allowed transitions) approaches [7, 9, 12, 48–51]. Such schemes are now
implemented in several codes, and can also be used to gain access to absolute
intensities, i.e., the molar absorptivity (generally noted ε in the well-known Beer–
Lambert’s law). This offers additional direct comparisons with experimental data.
Though this topic is discussed in more detail in Sect. 3, is it probably worth giving
some general comments regarding the selection of an appropriate XCF. First, one
can select a hybrid functional, incorporating a fraction of the so-called exact
exchange: they generally yield much more accurate results than the typical LDA
or GGA approaches which tend to provide much too low transition energies in most
compounds. If valence ES are investigated, one should distinguish the localized ES,
typically resulting for n ! π ⋆ and π ! π ⋆ transitions, for which standard global
hybrids such as B3LYP [52] or PBE0 [53] are well suited from charge-transfer
excited-states, for which the selection of range-separated hybrids which present an
amount of exact exchange increasing with the interelectronic distance, e.g.,
CAM-B3LYP [54] of ωB97X-D [55], generally provide more accurate transition
energies. Eventually, range-separated hybrids are also often a better choice for
Rydberg ES [56].
Computational Molecular Electronic Spectroscopy with TD-DFT 355
Similar to DFT, TD-DFT is relatively less sensitive to the size of the atomic basis
set than the corresponding highly-correlated wavefunction theories, though excep-
tions have been reported [57]. Irrespective of the agreement with experimental data,
reaching ES data which are converged with respect to the extension of the basis set
generally requires the selection of larger atomic basis sets than for GS properties.
For electronic transitions to low-lying excited-states in conjugated molecules, a
double-ζ (or, better, triple-ζ) polarized atomic basis set augmented with diffuse
orbitals appears to be a judicious choice. In other words, 6-31+G(d) or aug-cc-
pVDZ could be advised as reasonable compromises between computational cost
and accuracy for both Everta and Evert f . Of course, for Rydberg ES, a much larger
basis set may be necessary, e.g., aug-cc-pVTZ. When optimizing the geometry of a
given ES, one should also be cautious as the PES are often quite flat and
diffusionless basis sets could yield rather poor results but in strongly constrained
fluorophores. The interested reader can find elsewhere longer discussions regarding
basis set effects for both small [58] and large [15] molecules in the context of
TD-DFT spectroscopic investigations.
3 Benchmarks
In this section we focus on investigations treating the EAFCP of large molecules [7,
13, 15, 65], though there are several works dealing with small gas-phase com-
pounds for which the 0–0 band can be accurately measured [17, 19, 66–69]. First, as
ΔEZPVE is the most computationally expensive term, let us discuss its magnitude
356 D. Jacquemin and C. Adamo
O O O
N N N
B NH2 N
MeO O O N
F F
O
I II III IV Me2N V VI
O NHEt H
N O
O N O
N N O
N
O NHEt
NMe2
VII VIII S IX X XI XII NH2
Ph N
Cl S O
N S N N
Me2N N S N N
N S
2 HN NO2
Cl Th S Th
S Ph
XIII XIV XV XVI XVII
R(exp)=C8H17 R(exp)=C8H17
O R R(theo)=CH3 R(theo)=CH3
N NHMe H2N O
S R R
O O
S Th Th
2 N 2
R O MeHN
XVIII XIX XX XXI XXII
CF3 MeO
O
N
N
N N
N N
N O O B
N Ph F F Ph H2N O NH2
XXIII XXIV XXV XXVI XXVII OMe
R 2
O N O R(exp)=C8H17
O O
R(theo)=CH3 CO2H
Br S
R N N R
CN Ph Ph
R(exp)=C4H9 N
O O R(theo)=CH3 O
XXVIII R XXIX XXX
O O O N
H N Et2N O O
F N Th2
O N
N F N
H NMe2
O O
XXXI XXXII XXXIII XXXIV XXXV
N
Ph O
C6F5 CO2Et N N N
N N
N B O Al N Zn
N N
C6F5 OC O
O N N
Os N
OC Cl
NMe2 N
CO
XXXVI XXXVI XXXVIII XXXIX XL
Fig. 2 Set of 40 test molecules used in several benchmarks of AFCP energies [15, 70,
71]. Reproduced with permissions from Jacquemin et al. [70]. Copyright 2014, American Chem-
ical Society
and XCF dependence. For the 40 molecules displayed in Fig. 2, it has been found
that the variations when changing the XCF are weak (ca. 0.02 eV) [15], a
conclusion also reached in other studies on smaller systems [17, 66], indicating
that ΔEZPVE can, in general, be evaluated with any XCF. In addition, this term was
Computational Molecular Electronic Spectroscopy with TD-DFT 357
Table 1 MSE and MAE obtained during benchmarks of EAFCP of large structures. All data in
eV. LC-PBE* and LC-PBE0* are optimally-tuned range-separated hybrid functionals
XCF Molecular set MSE MAE References
BP86 41 conjugated molecules 0.56 0.57 [7]
BLYP 12 large dyes 0.49 0.51 [13]
B3LYP 41 conjugated molecules 0.33 0.34 [7]
12 large dyes 0.22 0.31 [13]
40 dyes (Fig. 2) 0.14 0.27 [15]
APF-D 40 dyes (Fig. 2) 0.06 0.27 [71]
PBE0 40 dyes (Fig. 2) 0.03 0.22 [15]
M06 40 dyes (Fig. 2) 0.05 0.23 [15]
PBE0-1/3 40 dyes (Fig. 2) 0.14 0.22 [71]
BMK 12 large dyes 0.07 0.19 [13]
SOGGA11-X 40 dyes (Fig. 2) 0.21 0.24 [70]
M06-2X 40 dyes (Fig. 2) 0.25 0.26 [15]
BHHLYP 41 conjugated molecules 0.01 0.18 [7]
CAM-B3LYP 12 large dyes 0.11 0.18 [13]
40 dyes (Fig. 2) 0.24 0.25 [15]
ωB97X-D 40 dyes (Fig. 2) 0.30 0.30 [70]
LC-PBE 40 dyes (Fig. 2) 0.56 0.57 [15]
LC-PBE* 40 dyes (Fig. 2) 0.12 0.20 [70]
LC-PBE0* 40 dyes (Fig. 2) 0.25 0.26 [71]
B2PLYP 12 large dyes 0.11 0.20 [13]
B2GPPLYP 12 large dyes 0.01 0.16 [13]
[63, 64] and CAM-B3LYP [54] provide more consistent values, i.e., larger corre-
lation coefficients with respect to experimental results (than B3LYP [52] or PBE0
[53, 62]) and can therefore be valued if design is sought: they overshoot the
transition energies in a rather systematic way. The best results for the set of
molecules of Fig. 2 are obtained with the optimally-tuned LC-PBE*, but at the
cost of a systematic (non-empirical) optimization of the attenuation parameter.
Grimme and coworkers also performed a series of benchmarks [7, 13, 65] with a
similar focus on “real-life” structures, and the results are collected in Table 1. In
their first contribution, they evaluated 3 XCF (BP86 [59, 73], B3LYP [52], and
BHLYP [74]) on 30 singlet–singlet and 13 doublet–doublet transitions in aromatic
and radical dyes, respectively. Solvent effects were empirically accounted for by
applying a standard correction to the experimental 0–0 energies. These authors
concluded that global hybrids with 30–40% exact exchange emerged the best
compromises [7]. More recently, the same group treated 12 molecules,
transforming the measured energies in reference vertical values thanks to
TD-DFT calculations. With this model, they could obtain deviations smaller than
0.2 eV with a recent global hybrid (BMK [75]), a range-separated hybrid
(CAM-B3LYP [54]), and their double hybrid (B2GPPLYP [65]).
In short, the typical TD-DFT errors for EAFCP are of the order of 0.2–0.3 eV,
when hybrid XCF are used. It should also be noted that XCF including a large share
of exact exchange (ca. 50%) deliver too large transition energies but tend to yield a
good consistency (large correlation coefficient) with experiment. The most accurate
results are obtained with double-hybrids or optimally-tuned range-separated XCF
but for an increased computational effort.
The accuracy of the band topologies obtained with several XCF has been evaluated
by several groups [7, 16, 27, 71]. For the sake of consistency with the EAFCP works
presented above, we discuss here the two latter investigations which relied on a set
of 20 conjugated molecules belonging to the same families as the one shown in
Fig. 1. The selected protocol also relied on the 6-31+G(d) atomic basis set and
included environmental effects thanks to the PCM approach. Selected key statisti-
cal data are given in Table 2. As all vibronic calculations have been performed on
the basis of GS and ES vibrations obtained in the harmonic approximation, the clear
trend is to overestimate the separation between the different vibronic peaks,
irrespective of the selected XCF, an error which could be reduced by including
anharmonic effects [16, 76, 77]. It is also obvious that the average absolute errors
are smaller for absorption (ca. 100 cm1) than for emission (ca. 250 cm1). All
XCF, apart from LC-PBE, provide rather similar deviations, and it is therefore
difficult to select an unambiguously more accurate hybrid functional. Nevertheless,
it should be noted that the obtained accuracy is significantly system dependent, e.g.,
most XCF are able to reproduce accurately the characteristic multi-peak structure of
Computational Molecular Electronic Spectroscopy with TD-DFT 359
fused aromatics but they fail to provide the correct height of the shoulder in
cyanines (see below for a discussion on the latter derivatives) [16]. For the relative
intensities (setting the intensity of the most intense peak to 1), the typical TD-DFT
error attains 10–15% for both absorption and emission, an average discrepancy
which is again rather independent of the selected XCF. Eventually, as for EAFCP,
optimally-tuned approaches vastly improve the original LC-PBE results, though
they do not outperform other XCF for band shapes. In other words, optimal tuning
improves the transition energies without deteriorating the accuracy of the computed
band shapes [71].
In this last part of this section, we consider a limited number of known TD-DFT
problems for low-lying singlet ES. In these cases, the accuracy of TD-DFT is either
worse than expected (cyanines) or can only be maintained with the selection of a
specific XCF (charge-transfer). It should also be noted that triplet ES and, conse-
quently, singlet-triplet splittings may be challenging for conventional TD-DFT
[78–81] but this particular error is beyond our scope here.
modeled with adiabatic TD-DFT [33, 82–93]. Indeed, the TD-DFT transition
energies are too large (by ca. 0.3–1.0 eV) in cyanines, and this conclusion has
been reached through comparisons of both TD-DFT’s Everta with their highly-
correlated wavefunction counterparts [82, 87] and TD-DFT’s EAFCP with experi-
mental references for fluoroborate emitters [88, 89]. More puzzling is the fact that
the errors seem to be almost independent of the selected XCF and that this error is
not related to a multi-determinant nature. The fundamental reasons explaining this
failure of TD-DFT have been given in [90–92, 94, 95] and summarized in a recent
account [93]. A pragmatic approach to obtain accurate EAFCP is to apply (13)
selecting an appropriate variant of the CIS(D), ADC(2) or CC2 approaches as the
wavefunction method [47, 96]. Examples of applications of such mixed approach
are given in Sect. 4.
4 Illustrations
Fig. 4 Comparison between theoretical ( full lines) and measured (dashed lines) absorption (red)
and emission (black) band shapes of an inorganic complex. No shifting of the AFCP energies was
applied. For the theoretical absorption and emission spectra, both the convoluted and stick spectra
are displayed with numbering for the most contributing modes. Reproduced with permission from,
Steffen et al. [115]. Copyright 2014, American Chemical Society
Theory (eV)
been performed at the 3.5
PCM-TD-M06-2X/6-311
+G(2d,p)//PCM-M06-2X/6- 3.0
31G(d) level, using either
the cLR or the SS PCM 2.5
approach for the transition
energies. The central line 2.0
indicates a perfect match
between theory and 1.5
experiment
1.0
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
Experiment (eV)
BODIPY and other similar derivatives relying on a fluoroborate group to ensure the
chemical stability of the dyes constitute one of the most important classes of
organic emitters [116–118]. Indeed, they present sharp fluorescence emission
bands and large quantum yields. A large panel of chemical groups can be added
around the central chromogens so as to modify the absorption and emission
energies. These fluorophores present ES of cyanine nature, which is known to be
challenging for TD-DFT (see above). Figure 5 displays the EAFCP obtained with
TD-DFT for a set of 83 fluoroborates using (10). This large set was obtained by
putting together the panel of molecules considered in [47, 89, 96, 119] and was
modeled using the M06-2X XCF. It is obvious that TD-DFT overestimates the
EAFCP in an almost systematic way (TD-DFT underestimates this energy in only
1 out of 83 cases), and this error is significant, as the MAE attains 0.354 eV.
However, the variations of EAFCP with the chemical structures is well reproduced
by TD-DFT, and this can be seen by computing the linear determination coefficient,
R2, which attains 0.965 eV. This indicates that this protocol misses only 3.5% of the
total variability of the experimental energies. To obtain values in better absolute
agreement with experiment, it has been shown that applying a scaled opposite spin
(SOS) variant of the CIS(D) model [45], that is using (13) with Ψ ¼ SOS-CIS(D), is
a very effective approach. Indeed, it allows the MAE to decrease by a factor of
3 (0.115 eV), at the same time inducing only a slight decrease of the R2 (0.949).
This is well illustrated in Fig. 5
Despite the systematic overestimation of the transition energy, it has been shown
that TD-DFT allows reproduction of the band shapes of both the absorption and
emission of fluoroborates with good to excellent accuracy [34, 35, 47, 88, 89,
120]. In other words, the PES provided by TD-DFT are reasonably accurate for
364 D. Jacquemin and C. Adamo
Fig. 6 Comparison between theoretical and experimental band topologies for a typical BODIPY
derivative. The impact of the SOS-CIS(D) correction which shifts the EAFCP is shown. Reproduced
with permissions from Chibani et al. [47]. Copyright 2014, American Chemical Society
this class of dyes. This is illustrated in Fig. 6 for a strongly conjugated BODIPY
designed to redshift the optical spectra. In Fig. 6 the band shape – which of course
remains unchanged when applying the SOS-CIS(D) correction to the energy –
clearly fits the experimental reference, with a marked shoulder displaced by
ca. 1,500 cm1 from the 0–0 band. The accuracy of TD-DFT’s vibronic coupling
has also been confirmed by computing the Huang–Rhys factors which were used to
provide an estimation of the non-radiative deactivation vibrational pathways in
selected BODIPY [89]. These factors correlated well with the measured quantum
yields of emission: the larger the Huang–Rhys factors, the more efficient the
non-radiative pathways, and the smaller the emission quantum yields.
Fig. 7 Schematic
representation of an ESIPT
*
O R1
system containing an enol
and a keto isomer. Adapted N ESIPT *
with permissions from R2 O R1
HO
Benelhadj
N
et al. [124]. Copyright 2014, S1 Enol E* R2 H O
Wiley
S1 Keto K*
Fluorescence
Fluorescence
ESIPT
Enol
h
O R1
Tautomerism
O N
R1
R2 H O
N
R2 HO
SO Keto K
SO Enol E
Fig. 8 Potential energy surfaces obtained for two HBO dyes. Left: alkyl substituted structure
presenting only ESIPT emission experimentally. Right: amino-substituted structure displaying
dual fluorescence in several solvents. For both dyes, the PES go from the enol (small O–H
distance) to the keto (large O–H distance). Adapted from Houari et al. [139] with permission
from the Royal Society of Chemistry
emission for a solvent in which the measured fluorescence bands for these two
tautomers overlap.
In a subsequent investigation [124], TD-DFT was used to rationalize the prop-
erties of seven large hydroxybenzofuranbenzoxazole (HBBO) derivatives differing
by their substitution patterns. A comparison between the experimental ratio of
ESIPT and normal emissions (Iketo/Ienol) with the theoretical relative stabilities of
the two tautomers determined for the ES (ΔGES) is given in Fig. 9. When the ΔGES
are smaller than 0.1 eV, the driving force is sufficient to yield a quantitative
proton transfer and only ESIPT emission is observed. Between 0.1 and 0.0 eV,
there is an equilibrium between the two forms which emit and dual emission can
only be obtained in this narrow energetic window. The correlation between the
measured relative fluorescence intensities and the computed driving force for
ESIPT is obvious in Fig. 9. This study led to the development of single-molecule
white organic light emitting diodes [124].
As stated above, TD-DFT can be coupled with several models to reproduce the
impact of the environment on the spectral properties of a chromophore. Besides the
Computational Molecular Electronic Spectroscopy with TD-DFT 367
Fig. 9 Comparison
between the theoretical
relative free energies of the 100
enol and keto isomers
determined at the ES and
20
ESIPT
0
-0.20 -0.15 -0.10 -0.05 0.00
Relative free energies (eV)
most widely treated case of organic solvents, such an environment can involve a
biomolecule [145–148], a cage [149, 150] a metal [151–153], an inorganic solid
[154–156], or a molecular crystal [157] to cite a few examples. Depending on the
exact nature of the environment, one needs to set up a specific computational
protocol, but the general idea is to split the total system into two parts: the
chromophore where the electronic excitation takes place and which is treated
with TD-DFT whereas the surroundings are modeled with a simpler theoretical
model, typically Molecular Mechanics (MM). We illustrate here such a procedure
for an organic cage and redirect interested readers to a previous review on the topic
for other examples and references [27]. The selected system consists of a squaraine
dye encapsulated in a tetralactam macrocycle (see Fig. 10). Such an assembly was
experimentally investigated by Smith and coworkers [158] and later modeled
[149]. The macrocyclic cage aims to protect the dye from (bio-)chemical degrada-
tions and was not designed to tune the observed color. Indeed, the hallmark
absorption band of the dye is shifted after complexation by 0.06 eV only [158],
a bathochromic effect which can be almost perfectly reproduced by TD-DFT
calculations considering the full system quantum mechanically (0.07 eV). How-
ever, such a brute force approach implies a large computational cost. As the
excitation is clearly localized on the squaraine, using a hybrid TD-DFT/MM is
justified. The first approach proposed in [149] was to account self-consistently for
the ground-state polarization by determining atomic point charges of the cage
equilibrated with the density of the dye. Such a procedure yields a qualitatively
incorrect hypsochromic shift of +0.10 eV. In a second approach, the response of the
cage density to the change of electronic state of the dye was modeled through a
368 D. Jacquemin and C. Adamo
N N
O O
N
NH HN
NH HN
N
O O
Fig. 10 Representation of the squaraine dye (top left) and cage (bottom left) used by Smith and
coworkers [158]. On the right hand side, a side view of the DFT (PBE0) optimized complex is
given [149]
polarizable continuum model inspired from PCM. This second scheme, denoted
Electronic Response of the Surroundings, yields, for a negligible computational
cost, a shift of 0.09 eV, in good agreement with both experiment and full TD-DFT
calculation.
Subsequently, one divides Δρ(r) into two parts according to the increase/
decrease of the density resulting from the electronic transition. For the former,
this reads
8
< ΔρðrÞ if ΔρðrÞ > 0
ρþ ðrÞ ¼ 0 if ΔρðrÞ < 0 ; ð15Þ
:
and similarly for ρ ðrÞ. The amount of charge transferred is obtained by integration
ð
qCT ¼ ρþ ðrÞdr; ð16Þ
and an equivalent result is obtained by integrating ρ ðrÞ. One next computes the
barycenters corresponding to the ρþ ðrÞ and ρ ðrÞ functions
ð
1
r ¼ ðx ; y ; z Þ ¼ CT rρþ ðrÞdr;
þ þ þ þ
ð17Þ
q
ð
1
r ¼ ðx ; y ; z Þ ¼ CT rρ ðrÞdr: ð18Þ
q
μCT is also equal to the difference of dipoles computed from the total GS and ES
densities. This procedure was applied to design rod-like dyes with a maximal CT
distance, using densities obtained with TD-DFT and more precisely with the
CAM-B3LYP functional [169]. The compounds considered in [169] consist of an
electron-donor group and an electron-acceptor moiety separated by a π-conjugated
linker. All parameters were investigated (nature of the donor, size and nature of the
linker, strength of the acceptor. . .). An illustration of the results obtained is given in
Fig. 11 for three typical push–pull systems. For the shortest system, one indeed
notices a typical CT state, the nitro (amino) group gaining (losing) density upon
electronic excitation and dCT is large. When the π-conjugated chain gets longer, one
observes, contrary to expectations, that dCT decreases. This can be qualitatively
understood from Fig. 11: as the chain gets longer the excited-state starts to be
370 D. Jacquemin and C. Adamo
Fig. 11 Representation of Δρ(r) for three oligomers (trimer, hexamer, and nonamer). The green
vector indicates the CT distance. The blue (red) regions indicate decrease (increase) of density
after photon absorption. Adapted with permission from Ciofini et al. [169]. Copyright 2012,
American Chemical Society
localized on the central part of the dye, with a minimal involvement of the terminal
groups and the CT character is lost, because the excited-state eventually corresponds
to a delocalized but symmetric π ! π ⋆ transition. This means that, to maximize
CT, there is an optimal linker length. For α,ω-NMe2,NO2 oligomers, this maximal
CT is obtained for an oligomeric length of ca. 3–5 connecting rings, smaller (larger)
systems being limited by the lack of efficient delocalization (the ineffective com-
munication between the end groups). In [169] it was therefore concluded that there
is a systematic fine balance between the three elements of the rod-like compounds,
and simply increasing the strength of the terminal electro-active groups or improv-
ing the delocalizability by adding more π-electrons in the bridge does not neces-
sarily mean improvement of the CT properties.
5 Conclusions
Acknowledgements D.J. acknowledges the European Research Council (ERC) and the Région
des Pays de la Loire for financial support in the framework of a Starting Grant (Marches – 278845)
and a recrutement sur poste stratégique, respectively. The COST-CMTS Action CM1002: COn-
vergent Distributed Environment for Computational Spectroscopy (CODECS) and its members
are acknowledged for many fruitful discussions.
References
23. Barbatti M, Crespo-Otero R (2015) Density-functional methods for excited states. In:
Ferré N, Filatov M, Huix-Rotllant M (eds) Topics in current chemistry. Springer, Berlin/
Heidelberg, pp 1–30. doi:10.1007/128_2014_605
24. Laurent AD, Jacquemin D (2013) Int J Quantum Chem 113:2019
25. Ziegler T, Krykunov M, Cullen J (2012) J Chem Phys 136:124107
26. Tomasi J, Mennucci B, Cammi R (2005) Chem Rev 105:2999
27. Jacquemin D, Mennucci B, Adamo C (2011) Phys Chem Chem Phys 13:16987
28. Cammi R, Mennucci B (1999) J Chem Phys 110:9877
29. Cossi M, Barone V (2001) J Chem Phys 115:4708
30. Caricato M, Mennucci B, Tomasi J, Ingrosso F, Cammi R, Corni S, Scalmani G (2006) J
Chem Phys 124:124520
31. Improta R, Scalmani G, Frisch MJ, Barone V (2007) J Chem Phys 127:074504
32. Marenich AV, Cramer CJ, Truhlar DG, Guido CG, Mennucci B, Scalmani G, Frisch MJ
(2011) Chem Sci 2:2143
33. Jacquemin D, Zhao Y, Valero R, Adamo C, Ciofini I, Truhlar DG (2012) J Chem Theory
Comput 8:1255
34. Chibani S, Charaf-Eddin A, Le Guennic B, Jacquemin D (2013) J Chem Theory Comput
9:3127
35. Chibani S, Charaf-Eddin A, Mennucci B, Le Guennic B, Jacquemin D (2014) J Chem Theory
Comput 10(2):805
36. Andersson K, Malmqvist P, Roos BO (1992) J Chem Phys 96(2):1218 http://scitation.aip.org/
content/aip/journal/jcp/96/2/10.1063/1.462209
37. Buenker RJ, Peyerimhoff SD (1968) Theor Chim Acta 12(3):183
38. Stanton JF, Bartlett RJ (1993) J Chem Phys 98(9):7029
39. Christiansen O, Koch H, Jørgensen P (1995) Chem Phys Lett 243:409
40. Kállay M, Gauss J (2004) J Chem Phys 121(19):9257
41. Hättig C, Weigend F (2000) J Chem Phys 113:5154
42. Nakatsuji H, Ehara M (1993) J Chem Phys 98:7179
43. Schirmer J, Trofimov AB (2004) J Chem Phys 120:11449
44. Head-Gordon M, Maurice D, Oumi M (1995) Chem Phys Lett 246:114
45. Rhee YM, Head-Gordon M (2007) J Phys Chem A 111(24):5314
46. Boulanger P, Chibani S, Le Guennic B, Duchemin I, Blase X, Jacquemin D (2014) J Chem
Theory Comput 10(10):4548
47. Chibani S, Laurent AD, Le Guennic B, Jacquemin D (2014) J Chem Theory Comput 10:4574
48. Guthmuller J, Zutterman F, Champagne B (2009) J Chem Phys 131:154302
49. Improta R, Barone V (2009) J Mol Struct (THEOCHEM) 914(1–3):87
50. Avila Ferrer FJ, Improta R, Santoro F, Barone V (2011) Phys Chem Chem Phys 13(38):17007
51. Avila Ferrer FJ, Santoro F (2012) Phys Chem Chem Phys 14(39):13549
52. Becke AD (1993) J Chem Phys 98:5648
53. Adamo C, Barone V (1999) J Chem Phys 110:6158
54. Yanai T, Tew DP, Handy NC (2004) Chem Phys Lett 393:51
55. Chai JD, Head-Gordon M (2008) Phys Chem Chem Phys 10:6615
56. Peach MJG, Benfield P, Helgaker T, Tozer DJ (2008) J Chem Phys 128:044118
57. Picconi D, Avila Ferrer FJ, Improta R, Lami A, Santoro F (2013) Faraday Discuss 163:223
58. Ciofini I, Adamo C (2007) J Phys Chem A 111:5549
59. Becke AD (1988) Phys Rev A 38:3098
60. Lee C, Yang W, Parr RG (1988) Phys Rev B 37:785
61. Perdew JP, Burke K, Ernzerhof M (1996) Phys Rev Lett 77:3865
62. Ernzerhof M, Scuseria GE (1999) J Chem Phys 110:5029
63. Zhao Y, Truhlar DG (2008) Acc Chem Res 41:157
64. Zhao Y, Truhlar DG (2008) Theor Chem Accounts 120:215
65. Goerigk L, Moellmann J, Grimme S (2009) Phys Chem Chem Phys 11:4611
66. Send R, Kühn M, Furche F (2011) J Chem Theory Comput 7(8):2376
Computational Molecular Electronic Spectroscopy with TD-DFT 373
109. Pl€
otner J, Tozer DJ, Dreuw A (2010) J Chem Theory Comput 6(8):2315
110. Guido CA, Mennucci B, Jacquemin D, Adamo C (2010) Phys Chem Chem Phys 12:8016
111. Charaf-Eddin A, Cauchy T, Felpin FX, Jacquemin D (2014) RSC Adv 4:55466
112. Wetzel C, Mishra A, Mena-Osteritz E, Liess A, Stolte M, Würthner F, Bäuerle P (2014) Org
Lett 16(2):362. doi:10.1021/ol403153z. http://pubs.acs.org/doi/abs/10.1021/ol403153z
113. Lanthier E, Reber C, Carrington T Jr (2006) Chem Phys 329(1–3):90, Electron correlation
and multimode dynamics in molecules (in honour of Lorenz S. Cederbaum)
114. Latouche C, Baiardi A, Barone V (2015) J Phys Chem B (in press)
115. Steffen A, Costuas K, Boucekkine A, Thibault MH, Beeby A, Batsanov AS, Charaf-Eddin A,
Jacquemin D, Halet JF, Marder TB (2014) Inorg Chem 53(13):7055
116. Loudet A, Burgess K (2007) Chem Rev 107:4891
117. Ulrich G, Ziessel R, Harriman A (2008) Angew Chem Int Ed 47:1184
118. Nepomnyashchii AB, Bard AJ (2012) Acc Chem Res 45(11):1844
119. Chibani S, Laurent AD, Le Guennic B, Jacquemin D (2015) J Phys Chem A. PMID:25522826
120. Zakrzewska A, Zalesny R, Kolehmainen E, Osmialowski B, Jedrzejewska B, Agren H,
Pietrzak M (2013) Dyes Pigm 99(3):957
121. Henary MM, Wu Y, Fahrni CJ (2004) Chem Eur J 10(12):3015
122. Wu Y, Peng X, Fan J, Gao S, Tian M, Zhao J, Sun S (2007) J Org Chem 72(1):62
123. Massue J, Ulrich G, Ziessel R (2013) Eur J Org Chem 2013(25):5701
124. Benelhadj K, Muzuzu W, Massue J, Retailleau P, Charaf-Eddin A, Laurent AD,
Jacquemin D, Ulrich G, Ziessel R (2014) Chem Eur J 20:12843
125. Sobolewski AL, Domcke W (1999) Phys Chem Chem Phys 1:3065
126. Aquino AJA, Lischka H, Hattig C (2005) J Phys Chem A 109:3201
127. Aquino AJA, Plasser F, Barbatti M, Lischka H (2009) Croat Chem Acta 82:105
128. Barbatti M, Aquino AJA, Lischka H, Schriever C, Lochbrunner S, Riedle E (2009) Phys
Chem Chem Phys 11:1406
129. Plasser F, Barbatti M, Aquino AJA, Lischka H (2009) J Phys Chem A 113(30):8490
130. Randino C, Ziolek M, Gelabert R, Organero JA, Gil M, Moreno M, Lluch JM, Douhal A
(2011) Phys Chem Chem Phys 13:14960
131. Cui G, Lan Z, Thiel W (2012) J Am Chem Soc 134(3):1662
132. Xie L, Chen Y, Wu W, Guo H, Zhao J, Yu X (2012) Dyes Pigm 92(3):1361
133. Hayaki S, Kimura Y, Sato H (2013) J Phys Chem B 117(22):6759
134. Moreno M, Ortiz-Sanchez JM, Gelabert R, Lluch JM (2013) Phys Chem Chem Phys
15:20236
135. Padalkar VS, Ramasami P, Sekar N (2013) J Fluoresc 23(5):839
136. Phatangare KR, Gupta VD, Tathe AB, Padalkar VS, Patil VS, Ramasami P, Sekar N (2013)
Tetrahedron 69(6):1767
137. Savarese M, Netti PA, Adamo C, Rega N, Ciofini I (2013) J Phys Chem B 117(50):16165
138. Savarese M, Netti PA, Rega N, Adamo C, Ciofini I (2014) Phys Chem Chem Phys 16:8661
139. Houari Y, Charaf-Eddin A, Laurent AD, Massue J, Ziessel R, Ulrich G, Jacquemin D (2014)
Phys Chem Chem Phys 16:1319
140. Laurent AD, Houari Y, Carvalho PHPR, Neto BAD, Jacquemin D (2014) RSC Adv 4:14189
141. Hubin PO, Laurent AD, Vercauteren DP, Jacquemin D (2014) Phys Chem Chem Phys
16:25288
142. Wilbraham L, Savarese M, Rega N, Adamo C, Ciofini I (2015) J Phys Chem B 119:2459
143. Raucci U, Savarese M, Adamo C, Ciofini I, Rega N (2015) J Phys Chem B 119:2650
144. Houari Y, Chibani S, Jacquemin D, Laurent AD (2015) J Phys Chem B 119:2180
145. Riccardi D, Schaefer P, Yang Y, Yu H, Ghosh N, Prat-Resina X, Konig P, Li G, Xu D, Guo H,
Elstner M, Cui Q (2006) J Phys Chem B 110(13):6458
146. Wanko M, Hoffmann M, Frähmcke J, Frauenheim T, Elstner M (2008) J Phys Chem B 112
(37):11468
147. Konig C, Neugebauer J (2011) Phys Chem Chem Phys 13:10475
Computational Molecular Electronic Spectroscopy with TD-DFT 375
Chantal Daniel
C. Daniel (*)
Laboratoire de Chimie Quantique, Institut de Chimie UMR 7177 CNRS/Université de
Strasbourg, 1 Rue Blaise Pascal BP 296/R8, F-67008 Strasbourg Cedex, France
e-mail: c.daniel@unistra.fr
378 C. Daniel
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
2 Methodological Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
2.1 Spin-Orbit Coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
2.2 Vibronic Couplings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
2.3 Intersystem Crossings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
3 Absorption Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
3.1 Electronic Spectroscopy of Ir(III) Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
3.2 Electronic Spectroscopy of Re(I) Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
4 Emission Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
4.1 Phosphorescence of Ir(III) Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
4.2 Emissive Properties of Square Planar Pt(II) Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
4.3 Luminescence of Re(I) Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
5 Ultra-Fast Intersystem Crossings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
5.1 Excited States Dynamics in Spin Crossover Fe (bpy)3]2+ Complex . . . . . . . . . . . . . . . . 401
5.2 Ultra-Fast Relaxation Processes in [Ru (bpy)3]2+ Complex . . . . . . . . . . . . . . . . . . . . . . . . . 404
5.3 Ultra-Fast Luminescence Decay in [Re (Br)(CO)3(bpy)] Complexes . . . . . . . . . . . . . . 405
6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
Abbreviations
MC Metal-centered
MCQDPT Multiconfiguration quasi-degenerate perturbation theory
MCTDH Multiconfiguration time-dependent Hartree
MLCT Metal-to-ligand-charge-transfer
MS-CASPT2 Multi-state CASPT2
PBE0 Perdew–Burke–Ernzerhof
PCM Polarized continuum model
PES Potential energy surfaces
PW91 Perdew–Wang 1991
RASSI Restricted active space state interaction
SOC Spin-orbit coupling
XANES X-Ray absorption near-edge structure
ZFS Zero field splitting
ZORA Zeroth order regular approximation
1 Introduction
Triplet electronic excited states play a central role in the spectroscopy, photochem-
istry, and photophysics of transition metal complexes. They perturb the fine struc-
ture of absorption spectra and are responsible for long-lived emission via low-lying
metal-to-ligand-charge-transfer (MLCT) states over a wide range of energy
domains [1–12]. They can quench emission by triggering electron transfer pro-
cesses via charge-separated (CS) states [13–15] or by inducing competitive disso-
ciation via metal-centered (MC) states [16–20]. Triplet sigma-bond charge transfer
states induce metal–alkyl bond homolysis [21, 22] sigma-bond sigma-bond excited
states are precursors of metal–metal bond homolysis [23] whereas intra-ligand
(IL) localized triplet states conduct isomerization pathways under visible irradia-
tion [24–26].
The kinetics of intersystem crossing (ISC) processes entirely control the popu-
lation of the low-lying triplet states after UV/visible absorption and strongly
influence the branching ratio between radiative and non-radiative decays. The
development of time-resolved spectroscopy, within femtosecond (fs)/picosecond
(ps) time scales, has opened the route to new experimental investigations in the field
of first-, second-, and third-row transition metal complexes photophysics
supporting evidence of ultra-fast ISC [27–32]. In order to understand the role of
the high spin states and to interpret these experimental findings, quantum chemistry
needs powerful methods able to describe correctly the excited states properties:
(1) electronic and geometrical structures; (2) transition energies; (3) spin-orbit
interaction between states of different multiplicities; and (4) multiplet spin-orbit
splitting. The electronic structure calculations should also figure out the shape of
the potential energy surfaces (PES) underlying the non-adiabatic excited states
dynamics. A direct correlation between the experimental data and the outcome of
380 C. Daniel
theory, namely branching ratio, time scales, or luminescence quantum yields, needs
simulation of the dynamics in real time. Obviously, whereas a complete quantum
treatment is possible for small molecules involving 5 or 6 atoms [33, 34] other
strategies have to be employed for large transition metal complexes with ~50–100
atoms and a metal center.
The purpose of this chapter is to review density functional theory (DFT)-based
methods for computing with reasonable accuracy excited states properties, includ-
ing spin-orbit coupling (SOC) in transition metal complexes, and to point to
pioneering strategies to simulate ultra-fast ISC processes in this class of molecules.
Whenever possible, the time-dependent DFT (TD-DFT) excited states properties
are compared either to the results obtained by more accurate ab initio methods or to
experimental data. The first section is devoted to methodological highlights ori-
ented to the computation of SOC and its interplay with vibronic coupling, and to the
simulation of ISC. The two next sections are dedicated to SOC effects on the
absorption and emission spectroscopies exemplified by recent theoretical studies
performed on third-row Ir(III), Re(I), and Pt(II) complexes. The last section reports
on recent pioneering simulations of ultra-fast ISC processes in various complexes
from first to third row on the basis of different approaches.
2 Methodological Highlights
Molecules that contain heavy elements (in particular 5d transition metals) play an
important role in the photochemistry and photophysics of coordination compounds
with regard to their luminescent properties and their implication in catalysis and
energy/electron transfer processes. Whereas molecular properties and electronic
spectroscopy of light molecules can be studied in a non-relativistic quantum
chemical framework, one has to consider the theory of relativity when dealing
with elements that belong to the lower region of the periodic table. As far as
transition metal complexes are concerned, one has to distinguish between different
manifestations of relativity. Important but not directly observable manifestations of
relativity are the mass velocity correction and the Darwin correction. These terms
lead to the so-called relativistic contraction of the s- and p-shells and to the
relativistic expansion of the d- and f-shells. A chemical consequence of this is,
for instance, a destabilization of the 5d shells with respect to the 3d shells in
transition metals.
Other important evidence of relativity in electronic spectroscopy, photophysics,
and photochemistry is the spin-orbit coupling between states of different multiplic-
ities. Indeed, most light-induced processes in transition metal complexes involve a
change of spin state and are not allowed in non-relativistic quantum formalism. A
fully relativistic treatment based on the Dirac equation [55] and a four-component
Hamiltonian including scalar and spin-orbit contributions for many electrons sys-
tems is unrealistic and beyond the scope of the systems and problems of interest in
this review. The reduction of the Dirac equation and its extension to many-electron
problems has opened the route to several relativistic approaches based on approx-
imate Hamiltonians applied with success to chemistry [56]. The spin-orbit coupling
terms arise from one- and two-electron operators developed within the
two-component formalism of the relativistic theory obtained by transformation of
the four-component equation. The Douglas–Kroll (DK) [57, 58] and Breit–Pauli
(BP) [59–62] forms are the most popular relativistic two-component operators.
In most of today’s applications the SOC effects in large transition metal com-
plexes are included by means of two approaches: (1) the restricted active space state
interaction (RASSI) including SOC [63] developed on the basis of a one-electron
Fock-type spin-orbit Hamiltonian [64] within the atomic mean field approximation
(AMFI) [65]; and (2) the zeroth-order regular approximation (ZORA) to the full
relativistic Hamiltonian based on a one effective two-component regular Hamilto-
nian developed at the zeroth-order [66–68]. Both methods are derived from the BP
spin-orbit Hamiltonian and are a good approximation to the BP theory. Whereas
RASSI-SOC formalism has been developed for correlated wave functions, the
ZORA operator is better adapted to perturbation and Kohn–Sham (KS) theories.
Most ZORA applications are performed within the framework of DFT despite some
limitations [35, 69]. SOC has recently been evaluated by means of a full BP
Hamiltonian applied to multi-reference CI wave functions or combined with
382 C. Daniel
^ ¼H
H ^0þW
^ ð0Þ þ W
^ ð1Þ þ W
^ ð2Þ þ ð1Þ
where the first term includes the kinetic energy operator and a harmonic term
representing the ground state Hamiltonian and W ^ ð0Þ is the zero-order diagonal
coupling matrix which contains the vertical excited state energies calculated at
the FC geometry. The first-order term W ^ ð1Þ contains the linear coupling elements
and the second-order non-adiabatic coupling term W ^ ð2Þ is included to take into
account the change of frequency in the electronic excited states. The truncation of
the Taylor expansion has to be adapted to the problematic and to the size of the
molecule.
The multi-state spin-vibronic interactions within a set of n electronic excited
states are deduced from the diabatic electronic representation including all pertinent
coupling terms. Explicitly, the intrastate κ n and interstate λn,m vibronic coupling
constants between the n and m electronic states are derived from the gradient and
Hessian of the potential energy with respect to the nuclear coordinates. The δn,m
SOC and the necessary ingredients, potential energy and its derivatives, are
extracted from the electronic structure data obtained by means of wave function
or DFT approaches. Not all coupling elements survive to the integration because of
group symmetry constraints. For instance, within the linear vibronic coupling
(LVC) approximation the non-vanishing intrastate κ n and interstate, λn,m coupling
constants are those for which the product of the irreducible representations of states
n and m and of the nuclear normal mode coordinate Qi contain the totally symmet-
ric representation. Recent applications to vibronic spectra of first-row transition
384 C. Daniel
restricted number of active electronic excited states in the dynamical process. The
lack of coherence and phase of the nuclei and the total time per trajectory are the
limiting factors of semi-classical trajectory surface hopping. The advantages are the
inclusion of all nuclear degrees of freedom, the use of on-the-fly calculated poten-
tials, and the ease of improving accuracy by including more trajectories.
The applicability of the full quantum approach is limited by the nuclear dimen-
sionality and requires very robust electronic structure methods for excited states
[70, 110]. The drawbacks of wavepacket dynamics are the complexity of setting up
an appropriate effective Hamiltonian, the use of approximate fitted potentials, and
the choice of selected vibrational modes.
The pioneering dynamical simulations performed on transition metal complexes
are far from being routine and need specific developments to be applicable to a wide
range of systems and ultra-fast phenomena circumscribed by spin-vibronic cou-
pling [111]. Recent applications to first-, second-, and third-row transition metal
complexes based on various approaches are developed in section 5 dedicated to
ultra-fast ISC processes.
3 Absorption Spectroscopy
The purpose of the next sections devoted to the absorption spectroscopy of third-
row transition metal complexes is to illustrate by recent examples the importance of
SOC in the computation of TD-DFT vertical electronic spectra using ZORA
approach. Whenever possible, the SOC-TD-DFT approach is compared to the
SOC-CASSCF/MS-CASPT2 method based on RASSI. In these examples the
electronic SOC effect is treated independent of any nuclear relaxation that could
influence its contribution to the absorption spectra.
In this application the absorption spectra of [Ir (ppy)3] 1 and [Ir (ppy)2(CO)Cl)] 2
(ppy ¼ tris(2-phenylpyridine) (Scheme 1) have been calculated by means of
TD-DFT methods based on optimized structures in vacuum and including spin-
orbit coupling [80].
Both TD-DFT/B3LYP and TD-DFT/PW91 “spin-free” absorption spectra have
been computed. The TD-DFT/B3LYP results shift the theoretical spectrum of
[Ir (ppy)3] 1 to the blue, as compared to the PW91 results, by 0.75 eV, with a
first transition calculated at 25,080 cm1 of significant oscillator strength
( f ¼ 0.023). Moreover, this overestimated transition is characterized by an impor-
tant unrealistic ligand-to-ligand-charge-transfer (LLCT) character which is only
minor in the TD-DFT/PW91.
386 C. Daniel
Scheme 1 Optimized structures of fac-[Ir (ppy)3] 1 and cis-[Ir (ppy)2(CO)(Cl)] 2 (data from [80])
Wavelength [nm]
250 300 350 400 500 600 700
E25 A22
50
A23 A24
45
ε [103 I - moI-1 cm-1]
40
35
Intensity
A28 A29 30
E12 E13 25
E30
Intensity in arbitrary unit
A11 A12 20
15
10
5 lr(ppy)3
0
45000 40000 35000 30000 25000 20000 15000
Wavenumber [cm-1]
A5 A6
E3 E5 E6 E7
A2
Table 1 TD-DFT/PW91 “spin-free” states (in cm1) of fac-[Ir (ppy)3] 1 and associated oscillator
strengths (adapted from Brahim and Daniel [80])
Transition Transition
State Label Character energy in cm1 energy in eV fa
1
A S2 MLCT 18,560 2.32 7 103
1
E S5 MLCT 20,720 2.59 0.015
1
E S7 MLCT 21,760 2.72 0.041
3
E T5 MLCT 22,080 2.76
1
E S8 MLCT 22,720 2.84 0.013
1
A S9 MLCT 23,280 2.91 0.033
3
E T7 LC 27,440 3.43
3
E T12 LC/MLCT 29,920 3.74
1
A S12 LC 30,000 3.75 0.042
1
A S17 LC 32,240 4.03 2 103
1
A S18 LC/MLCT 32,560 4.07 0.05
1
E S21 LC/MLCT 33,280 4.16 0.040
1
A S22 MLCT 33,440 4.18 0.046
1
E S23 LC/MLCT 33,520 4.19 0.015
1
A S24 MLCT 33,600 4.20 0.106
1
E S33 MLCT 36,720 4.59 0.075
1
A S34 MLCT 37,040 4.63 0.083
1
A S35 MLCT 37,280 4.66 5 103
3
A T34 LC/MLCT 37,360 4.67
3
A T36 LC 37,520 4.69
1
A S36 LC/MLCT 37,600 4.70 0.012
1
A S44 LC 39,520 4.94 0.109
3
E T43 MLCT 39,920 4.99
1
E S45 LC 40,000 5.00 0.023
a
Only the “spin-free” states of interest entering in the composition of the “spin-orbit” states
described in Table 2 are reported
388 C. Daniel
Table 2 TD-DFT/PW91 “spin-orbit” states (in cm1) of fac-[Ir (ppy)3] and associated oscillator
strengths ( f > 0.005) (adapted from Brahim and Daniel [80])
Transition Transition
State Compositiona energy in cm1 energy in eV f
E3 23% S5 12% S2 12% T5 20,080 2.51 0.011
E5 54% S5 22% S7 21,280 2.66 0.017
E6 41% S7 32% T7 22,720 2.84 0.017
E7 68% S8 23,120 2.89 0.014
A2 90% S9 23,440 2.93 0.030
A5 25% S17 20% T12 17% S12 29,680 3.71 0.010
A6 66% S17 16% S12 29,840 3.73 0.009
A11 38% S22 13% S24 10% T28 32,400 4.05 0.030
A12 74% S18 32,560 4.07 0.040
E12 27% S21 8% S23 33,280 4.16 0.015
E13 38% S24 20% S22 33,280 4.16 0.053
E25 25% S33 12% T36 37,040 4.63 0.019
A22 16% S36 14% T34 10% T36 37,120 4.64 0.017
A23 18% T36 15% S36 11% S35 37,200 4.65 0.027
A24 65% S34 37,440 4.68 0.054
A28 73% S44 39,520 4.94 0.080
E30 82% S45 40,000 5.00 0.019
A29 31% T43 10% S44 40,640 5.08 0.012
a
The label of the singlet (Sn) and triplet (Tn) states refer to Table 1
characterized by a series of low-lying MLCT states between 18,480 cm1 (S1) and
23,280 cm1 (S9). Above 30,000 cm1 the ligand-centered LC state’s contributions
become more important with two intense peaks calculated at 33,600 cm1 (S24) and
39,520 cm1 (S44). These LC states and the peripheral transitions calculated above
33,360 cm1 contribute to the two intense experimental UV bands centered at
34,840 and 40,820 cm1. Whereas the maximum observed at 34,840 cm1 is well
reproduced by the theoretical “spin-free” spectrum, the maximum at 40,820 cm1 is
red shifted by 0.5 eV by the calculation. Moreover, the theoretical maximum at
36,160 cm1, not observed in the experimental spectrum, corresponds to a metal-
centered transition corresponding mainly to a dIr ! 6 s excitation with a diffuse
Rydberg character. Knowing that TD-DFT is not the method of choice for describ-
ing the Rydberg excited states [113], this assignment has to be taken with care. The
quality of the upper part of TD-DFT spectrum could certainly be improved by
requesting more roots (actually 200).
The experimental and theoretical maxima of the first band observed between
20,000 and 25,000 cm1 do not coincide exactly, the theoretical band being slightly
shifted to the red. However, several MLCT states with rather large oscillator
strengths are calculated in this region (Table 1). The accuracy of the calculations
performed in vacuum does not allow further comparison. The data reported in
Table 1 illustrate the high density of singlet and triplet excited states within
Absorption Spectroscopy, Emissive Properties, and Ultrafast Intersystem. . . 389
X
OC N
Re
OC N
CO
(Fig. 3), they give a very different interpretation of the absorption bands and only
the SO treatment can account for all the observed spectral features, namely the
low-energy shoulders.
SO-TD-DFT transitions are spread over a broader energy range and have more
similar relative intensities than the SO-MS-CASPT2 ones. The SO-TD-DFT theo-
retical spectrum thus accounts better for the large widths and shoulders observed
experimentally. The SO-MS-CASPT2 and solvent corrected SO-TD-DFT transi-
tion energies calculated for the lowest states of [Re(I)(CO)3(bpy)] are reported in
Table 3. In contrast to the experiment, SO-MS-CASPT2 predicts an increase of the
lowest absorption band intensity on going from Cl to Br and I, with increasing
oscillator strengths of the strongest contributing transition in the order Cl
( f ¼ 0.038) < Br (0.068) < I (0.082). On the other hand, SO-TD-DFT predicts
decreasing molar absorptivity of the lowest band Cl ( f ¼ 0.047) > Br (0.036) > I
(0.013), in qualitative agreement with the experimental trend (Fig. 3). This differ-
ence between the two computational techniques is probably caused by a limited
active space, smaller MLCT-XLCT delocalization, and the neglect of solvent
effects in SO-MS-CASPT2.
392 C. Daniel
Fig. 3 Comparison between experimental UV–vis spectra in different solvents (left) and ”spin-
free” TD-DFT/PBE0/COSMO-CH2Cl2 absorption spectra (right) of [ReX(CO)3(bpy)] (X ¼ Cl,
Br, I) (reprinted with permission from Heydova et al. [79] Copyright 2012 American Chemical
Society)
Within the TD-DFT spin-free model, the lowest absorption band originates from
a single strong transition and two satellites occurring at higher and lower energy,
respectively. The spin-orbit analysis interprets the lowest band as resulting from a
whole series of weaker transitions and assigns the two lowest lying shoulders as
transitions to spin-mixed states. Notably, even the strongest transition contributing
to the lowest band possesses only partial singlet character which decreases in the
order Cl (88%) > Br(81%) >> I(58%) at the SO-TD-DFT level (Table 3).
A correlation diagram between “spin-free” and “spin-orbit” states of [ReI
(CO)3(bpy)] is shown in Fig. 4. Left and right columns show “spin-free” singlet
and triplet states, respectively, and the SO states are presented in the middle, with
dashed lines indicating the principal contributions. Figure 4 illustrates the spin-orbit
interactions of “spin free” singlet and triplet states in the formation of two sets of
“spin orbit” states. Transition to the fourth and sixth “spin orbit” excited states cA0
and dA0 have an oscillator strength of 0.0061 and 0.0018, respectively, explaining
the occurrence and relative intensities of the two low-energy bands in the experi-
mental spectrum of [ReI(CO)3(bpy)] and the presence of shoulders for the other two
complexes. These features cannot be accounted for by the “spin-free” calculations,
where the lowest transitions to 1,3A00 and 3A0 states are forbidden.
Including SOC explicitly not only improves the quantitative correspondence
with the experimental spectra but also provides a physically more correct insight
into the nature of the excited-states involved and their deactivation pathways.
Absorption Spectroscopy, Emissive Properties, and Ultrafast Intersystem. . . 393
Table 3 “Spin orbit” MS-CASPT2 and solvent corrected (CH2Cl2) SO-TD-DFT transition
energies to the lowest excited states of [Re(I)(CO)3(bpy)], associated wavelengths, in nm and
oscillator strengths f
SO Composition of the SO-states in terms Transition energies Wavelength
state of spin-free states in cm1 in nm f
MS-CASPT2
2A0 a3A00 (87%) + b1A0 (9%) 21,085 474 0.0078
00
1A a3A00 (86%) + a3A0 (12%) 21,100 474
0
3A a3A00 (87%) + a3A0 (9%) 21,110 474 0.0031
00
2A a1A00 (84%) + a3A0 (14%) 21,400 467 0.0048
00
3A a3A0 (86%) + a3A00 (9%) 24,460 409 0.0002
0
4A a3A0 (86%) + a3A00 (12%) 24,500 408 0.0009
00
4A a3A0 (85%) + a1A00 (11%) 24,570 407 0.0008
5A0 b1A0 (86%) + a3A00 (12%) 24,930 401 0.082
6A0 b3A00 (90%) + c3A0 (7%) 27,800 360 0.0001
5A00 b3A00 (90%) + c3A0 (8%) 27,800 360
7A0 b3A00 (93%) 27,825 359 0.0011
TD-DFT
1A00 a3A00 (51%) + a3A0 (47%) 19,160 522
0
2A a3A00 (52%) + a3A0 (47%) 19,170 521 0.0001
00
2A a3A0 (55%) + a1A00 (44%) 19,410 515 0.0003
0
3A a3A00 (65%) + a1A0 (31%) 19,560 511 0.0073
00
3A a3A0 (48%) + a3A00 (44%) 21,920 456
0
4A a3A0 (47%) + a3A00 (45%) 22,010 454 0.0019
00
4A a1A00 (51%) + a3A0 (44%) 22,200 450 0.0003
0
5A b1A0 (58%) + a3A00 (30%) 22,535 444 0.013
00
5A b3A0 (84%) 24,960 400
0
6A b3A00 (84%) 24,970 400
0
7A b3A00 (94%) 25,120 398 0.0015
00
6A b3A00 (49%) 26,040 384 0.0001
00
7A b3A00 (95%) 26,205 382
0
8A b3A0 (30%) + c1A0 (26%) + c3A0 (14%) 26,320 380 0.0006
0
9A c3A0 (40%) + b3A0 (36%) + d1A0 (6%) 26,630 375 0.0036
Only the contributions of the “spin-free” states 5% are given (the “spin-free” and “spin orbit”
electronic ground state are labeled a1A0 and 1A0 , respectively) (adapted with permission from
Heydova et al. [79] Copyright 2012 American Chemical Society)
Similar SOC effects have been put in evidence on the electronic spectroscopy of
[Re(imidazole)(CO)3(phen)]+ [76] for which correlation diagrams between “spin-
free” and “spin-orbit” states become very complicated as illustrated in Fig. 5.
The lowest parts of the two correlation diagrams (Fig. 5) are qualitatively similar
with little differences because of solvent corrections not being included in the
MS-CASPT2 calculations. The lowest MLCT states (a3A00 , a3A0 , b3A00 ) remain
predominantly triplets but contribute to the “spin-orbit” spectrum by admixture of
singlets.
394 C. Daniel
Fig. 4 Correlation of SO-TD-DFT/PBE0 lowest singlet (left) and triplet (right) “spin-free” states
with “spin-orbit” states (middle) of [Re(I)(CO)3(bpy)] in solvent. Red, blue, and black arrows
indicate transitions with oscillator strengths larger than 0.01, 0.001–0.01, and 0.0005–0.001,
respectively (reprinted with permission from Heydova et al. [79]. Copyright 2012 American
Chemical Society)
Fig. 5 Correlation of (a) SO-MS-CASPT2 and (b) SO-TD-DFT between singlet (left) and triplet
(right) “spin-free” states and “spin-orbit” states (middle) of [Re(imidazole)(CO)3(phen)]+. Red,
blue, and black arrows indicate transitions with oscillator strengths larger than 0.01, 0.001–0.01,
and 0.0005–0.001, respectively (adapted from Bakova et al. [76])
Absorption Spectroscopy, Emissive Properties, and Ultrafast Intersystem. . . 395
4 Emission Spectroscopy
1 4 3 X k 2
¼ α30 ΔEk M ; ð2Þ
α
τk 3t0 α2fx;y;zg
moment between the ground state and the k-spin sub-levels of the triplet T1. Both
the S0 electronic ground state and triplet T1 states were fully optimized at the
DFT/B3LYP level.
In the pioneering study performed on the reference complex [Ir (ppy)3], and the
piq-substituted complexes depicted in Scheme 3, neither solvent nor vibronic
coupling effects have been included. The electric dipole moments associated with
the transitions between T1 and the ten lowest Tn triplets and between S0 and the ten
lowest Sn singlets, as well as SOC matrix elements between T1, S0, and the ten
lowest Sn singlet states have been computed [123].
From this detailed investigation it has been shown that at vertical S0 ! T1
transition the T1 state is highly delocalized over the three ligands with a mixed
LC/MLCT character. When relaxed to its minimum potential energy, T1 becomes
Absorption Spectroscopy, Emissive Properties, and Ultrafast Intersystem. . . 397
Table 4 ZFS sub-levels (in cm1) of T1 state of [Ir (ppy)3] at the DFT/B3LYP/6-311G*/SDD
level (reprinted with permission Jansson et al. [123] Copyright 2007 Elsevier)
T1 spin-orbit sub-levels S0-geom T1-geom
Tz1 19,989.06 16,609.87
Ty1 20,021.79 16,663.47
Tx1 20,092.98 16,678.48
localized on a single ligand. The zero-field splitting (ZFS) spin sub-levels and their
spontaneous emission characteristics have been obtained at some intermediate
geometry between S0 and T1 structures because of the anharmonicity of the
potential associated with T1. The large SOC between T1 and the S5 singlet MLCT
state together with large dipole interactions between S5 and S0 are responsible for
the intense phosphorescence of the reference complex [Ir (ppy)3]. The ZFS
sub-levels of T1 calculated at S0 and T1 geometries are reported in Table 4
The values reported in Table 4 and the lowest position of the Tz1 sub-level, as
well as the calculated lifetimes originated from these spin-orbit sub-levels agree
rather well with the data obtained by temperature-dependent refined spectroscopic
experiments [124]. Moreover, the theoretical model, despite the neglect of vibronic
coupling effects, recovers the overall phosphorescence experimental lifetime of
~2 μs in the high temperature limit as soon as the S0 geometry is chosen in the
phosphorescence rate calculation.
On the basis of the same computational strategy, the photophysical properties of
[Ir (piq(ppy)2] 5, [Ir (piq)2(ppy)] 4, and [Ir(piq)3] 3 have been elucidated and
successfully compared to the accurate experimental data available for this class
of molecules [81]. It has been shown that increasing the number of piq ligands shifts
the emission maximum to the red by about 10 nm and enhances radiative rate
constants by 60% within the range of the experimental trends. Interestingly, the
SOC strength and the radiative rate constant are diminished by the presence of
fluorine atoms in [Ir(Fnppy)3] complexes [42] resulting from the inverse heavy-
atom effect also observed in the [Re (X)(CO)3(bpy)] complexes discussed in
Heydova et al. [79], Cannizzo et al. [120], and Gourlaouen et al. [121].
A recent systematic study of the optical properties of a series of five Pt(II) planar
complexes with bidentate ligands, namely [Pt (bpy)Cl2] (bpy ¼ 2,20 -bipyridine) 6
and [Pt (ppy)Cl2] (ppy ¼ 2-phenylpyridine) 7 and terdentate ligands, namely
[Pt (tpy)Cl]+ (tpy ¼ 2,20 :60 ,200 -terpyridine) 8, [Pt (phbpyR)Cl] (phbpy ¼ 6-phenyl-
2,20 -bipyridine; R ¼ H) 9, and [Pt (dpybR)Cl] (dpyb ¼ 2,6-di(2-pyridyl)benzene;
R ¼ CH3) 10 (Scheme 4) by means of TD-DFT including solvent correction has
allowed us to rationalize the puzzling emissive behavior of this class of
molecules [75].
398 C. Daniel
Scheme 4 Structures of [Pt (bpy)Cl2] 6, [Pt (ppy)Cl2] 7, [Pt (tpy)Cl]+ 8, [Pt (phbpyR)Cl] 9, and
[Pt (dpybR)Cl] 10
In this example, SOC has been included only for the computation of the
absorption spectra and has been neglected for the determination of the emission
wavelengths reported in Table 5. The nearly pure triplet character of the “spin-
orbit” states and the negligible singlet/triplet mixing in these molecules justify this
choice [75].
On the basis of fully optimized structures and energetics, we have shown that the
structures remain nearly planar in the low-lying singlet and triplet excited states of
charge transfer character, namely MLCT and XLCT, whereas a significant distor-
tion corresponding to the out-of-plane-bending of the Pt–Cl bond characterizes the
geometry of the metal-centered (MC) states. The presence of these strongly
distorted non-radiative MC states minima, situated well below the charge transfer
states minima (ΔE ¼ 0.3 to 0.8 eV) and easily accessible upon irradiation in the
visible, explains the poor luminescence of the bipyridine and terpyridine
non-cyclometalated complexes 6 and 8 at room temperature.
In contrast, the minima of the emissive states of mixed MLCT/XLCT/LC
character are efficiently populated in 7, 9, and 10, especially in the terdentate
complexes. The luminescence of complex 10, cyclometalated in axial position, is
particularly efficient because the minimum of the lowest emissive state is well
separated from those of the MC states (ΔE ¼ +0.23 eV) in contrast to its analog,
complex 9, cyclometalated in lateral position where the emissive MLCT/LC state
minimum is nearly degenerate with the lowest MC state minimum (Δ ¼ +0.01 eV).
Whereas SOC correction probably overstabilizes the MC states with respect to the
charge transfer states, it should not overturn the relative order of the minima
Absorption Spectroscopy, Emissive Properties, and Ultrafast Intersystem. . . 399
Table 5 TD-DFT/B3LYP vertical Tn ! S0 transition energies (in eV) and corresponding emis-
sion wavelengths (in nm) of complexes 6 to 10 depicted in Scheme 4 (adapted with permission
from Gourlaouen and Daniel [75] Copyright 2014 Royal Society of Chemistry)
Vertical Tn ! S0 Emission
State transition energy (eV) wavelength (nm)
[Pt (bpy)Cl2] 6 T1 MC 0.36 –
T2 MC/XLCT 0.92 1,355
[Pt (tpy)Cl]+ 8 T1 MC 0.41 3,012
T2 MLCT/LC 1.92 645
[Pt (ppy)Cl2] 7 T1 MLCT/XLCT/LC 2.34 530
T2 MC/LC 2.28 544
T3 MC 2.33 531
T4 MC/LMCT/MLCT 2.41 515
[Pt (phbpyH)Cl] 9 T1 MLCT/LC 2.02 613
T2 MC 0.43 2,852
S1 MLCT/LC 2.35 527
T3 MLCT/XLCT/LC 2.71 458
[Pt (dpybR)Cl] 10 T1 MLCT/LC/XLCT 2.42 513
(R ¼ CH3) T2 MLCT/LC/XLCT 2.48 500
T3 MC 0.98 1,271
T4 MC 1.47 846
Table 6 SOC-TD-DFT vertical Sn, Tn ! S0 transition energies (in eV) and corresponding
emission wavelengths (in nm) of the lowest relaxed singlet and triplet states of [Re(Cl)
(CO)3(bpy)] 11, [Re(Br)(CO)3(bpy)] 12, and [Re(I)(CO)3(bpy)] 13 calculated in CH3CN (adapted
from Gourlaouen et al. [121])
Vertical Sn, Tn ! S0 Emission wavelength (in nm)
State transition energy (in eV) with SOC
X ¼ Cl 11
a3A00 2.048 610 (A00 )
2.049 610 (A0 )
2.060 607 (A0 )
a3A0 2.171 575 (A00 )
2.172 576 (A0 )
2.191 570 (A00 )
a1A00 2.173 575 (A00 )
b1A0 2.521 496 (A0 )
b3A00 2.593 482 (A0 )
2.648 472 (A00 )
2.681 466 (A00 )
X ¼ Bra 12
T1 a3A00 2.054 609 (A00 )
2.055 608 (A0 )
2.069 604 (A0 )
T2 a3A0 2.130 587 (A00 )
2.130 587 (A0 )
2.145 583 (A00 )
S1 a1A00 2.169 576 (A00 )
S2 b1A0 2.475 505 (A0 )
T3 b3A00 2.565 487 (A0 )
2.675 467 (A00 )
2.675 467 (A0 )
X ¼ I 13
a3A00 2.017 620 (A00 )
2.018 619 (A0 )
2.043 612 (A0 )
a3A0 2.016 620 (A00 )
2.017 620 (A0 )
2.027 617 (A00 )
a1A00 2.102 595 (A00 )
b1A0 2.164 577 (A0 )
b3A00 2.443 512 (A0 )
2.589 483 (A0 )
2.609 479 (A00 )
a
The labels Sn and Tn are used in Sect. 5.3 dedicated to the ultra-fast luminescence decay of
[Re (Br)(CO)3bpy]
Absorption Spectroscopy, Emissive Properties, and Ultrafast Intersystem. . . 401
400 nm. We may notice a very small spin-orbit splitting of the triplets. The
structural deformations when going from the electronic ground state to the relaxed
low-lying excited states are quite small. After nuclear relaxation into the potential
wells of the excited states, the Re–X shortening does not exceed 0.097 Å in the
singlet states and 0.043 Å in the triplet states. This deformation is accompanied by
an elongation of the Re–Cax bonds (<3%) together with a minor shortening of the
Re–N bonds in all excited states.
According to the calculations, the luminescence should start from the b1A0 state
calculated at 505 nm (or 496 nm with SOC) and 522 nm (or 505 nm with SOC) for
the chloride 11 and bromide 12 complexes, respectively. These values lead to
theoretical Stokes shifts of 5,640 (5,200) cm1 and 5,880 (5,640) cm1, respec-
tively, in agreement with the experimental data (~6,000 cm1). Accordingly, the
two complexes behave similarly and the small calculated red shift is also observed
on the experimental luminescence spectra when going from Cl to Br [120].
The emission wavelengths reported in Table 6 correlate nicely with the three
domains of luminescence detected by ultra-fast resolved spectroscopy [120],
namely in the ranges 500–550, 550–600, and 600–620 nm for the three molecules.
It is noteworthy that both b3A00 and b1A0 states participate in the early short-lived
emission whereas both a1A00 and a3A0 contribute to the intermediate band. The only
purely phosphorescent process is the long-lived emission generated by a3A00 in the
chloride and bromide complexes and by both a3A0 and a3A00 in the iodide complex,
these states keeping a nearly pure triplet character.
A recent promising approach, applied to the spin crossover complex [Fe (bpy)3]2+
[90] has been developed by C. M. Marian et al. [101, 102]. This method is based on
time-dependent calculations of ISC rates in the multi-mode harmonic oscillator and
Condon approximations and beyond, where the electronic spin-orbit matrix ele-
ments depend linearly on the nuclear coordinates within a spin-vibronic coupling
scheme. The ISC rate can be decomposed into three contributions, namely direct,
mixed direct vibronic, and vibronic.
402 C. Daniel
Scheme 5 Schematic
diagram of the
1
MLCT ! 5T2g
deactivation channel in
[Fe (bpy)3]2+ (adapted from
Bressler et al. [95, 96])
where ϕi and ϕf are the initial and final electronic states, respectively, Ωi is a matrix
that contains vibrational frequencies of the initial state, and G(t) is a time-dependent
correlation function which contains information about the vibrational frequencies
and normal coordinates of the initial and final states.
The SOC matrix of 144 144 dimension generated by the 16 lowest electronic
states in each multiplicity contains nearly 4,500 non-zero elements. This illustrates
the complexity of the spin-orbit interactions in transition metal complexes charac-
terized by a high density of various electronic states. The 1MLCT/3MLCT and
1
MLCT/3MC SOC do not exceed 200 cm1 and the 3MLCT/5T2g is very small
(~6 cm1). In contrast, the 3MC/5T2g and 3MC/1A1g are characterized by large
values (~500 cm1). The calculated 1MLCT ! 3MLCT and 1MLCT ! 3MC ISC
rates are reported in Table 7.
The 3MLCT ! 5T2g ISC rate has been estimated to be greater than 10 ps,
whereas the 3MC ! 5T2g ISC process is fast with a rate estimated to be 62 fs.
This gives an overall time scale qualitatively comparable to the experimental
one [95].
This theoretical study has pointed to a step-by-step mechanism showing that the
3
MC states play a key role in the deactivation process from the initially populated
1
MLCT state.
The interest of the method is that spin-orbit and vibrational contributions to the
ISC rates can be easily extracted. However, some drawbacks should be pointed out.
In this approach the vibrational relaxation at Franck–Condon and the variation of
SOC as a function of the normal modes are not considered. The validity of the
Fermi golden rule approximation is questionable because we are facing ultra-fast
non-BO processes which need to be described by spin-vibronic coupling model
Hamiltonians [84, 87, 88]. An alternative is to run semi-classical trajectories on
PES computed on the fly, coupled non-adiabatically and by SOC. Such a tentative
proposal, under some approximation, is presented in the next example dedicated to
the 1MLCT ! 3MLCT ISC in [Ru (bpy)3]2+ [126]. This type of simulation is based
on TD-DFT calculations of the energies valid for second- and third-row transition
metal complexes but hardly applicable to first-row transition metal complexes such
as [Fe (bpy)3]2+ because of electron correlation and multireference electronic
configurations.
Table 7 Calculated 1
MLCT ! 3MLCT and MLCT ! 3MC intersystem-crossing rates in
1
[Fe (bpy)3]2+
ϕi ϕf kISC (s1) τ (fs)
1
MLCT 3
MLCT 3.31 108 28
1
MLCT 3
MC 2.28 108 23
1
MLCT 3
MCa 2.26 107 718
Reprinted with permission from Sousa et al. [90] Copyright 2013 Wiley
a
This 3MC state is the lowest one in Scheme 5
404 C. Daniel
The first simulation based on TD-DFT energies and forces (gradient and Hessian)
computed on-the-fly and introducing both vibronic and SO coupling effects has
been able to reproduce semi-quantitatively the ultra-fast relaxation of the photo
excited 1MLCT state of [Ru (bpy)3]2+ (bpy ¼ 2,20 -bipyridine) followed by ISC to
the lowest 3MLCT state [126].
This study focuses on the ultra-fast excited states dynamics of [Ru (bpy]2+ in
water with emphasis on the ISC processes through the seven low-lying singlet and
triplet states. Time-resolved spectroscopy [127, 128] has shown a fast decay within
less than 100 fs after absorption leading to a long-lived 3MLCT state during a few
hundreds of ps.
The non-adiabatic dynamics of the early events (50 fs) is simulated on the fly by
means of linear response TD-DFT-based trajectories surface hopping according to
the Tully algorithm [129]. SOC between qualitatively selected singlet and triplet
states is obtained from a perturbative approach [130].
In order to take into account the solvent effects, the initial system constituted of
the Ru(II) complex surrounded by 3,298 water molecules and the counter ions Cl
is heated at 300 K. The non-BO dynamics represented by the panels in Fig. 6 show
that the internal conversion processes through the individual singlet states manifold
occur within a few to 10 fs, showing strong singlet triplet SOC occurring within the
Fig. 6 Non-adiabatic molecular dynamics of [Ru (bpy)3]2+ in water represented by the population
of the low-lying seven singlet (in gray) and seven triplet states (in red) as function of time. The
SOC at states crossing are represented by white (weak SOC), gray (medium SOC), and black
(strong SOC) filled circles (reprinted with permission from Tavernelli et al. [126] Copyright 2011
Elsevier)
Absorption Spectroscopy, Emissive Properties, and Ultrafast Intersystem. . . 405
first 40–50 fs. The largest SOC is predicted between the lowest 1MLCT and its
associated triplet state. The importance of solvation dynamics is also emphasized.
The drawback of this pioneering promising approach is the limited number of
trajectories and the qualitative estimation of the SOC, the value of which should be
determined quantitatively on the fly as well. However, this strategy opens the route
to new applications in large transition metal complexes as soon as the electronic
structure can be described correctly by means of TD-DFT.
In this very recent application we focus on the excited states decay of [Re (Br)
(CO)3(bpy)] within the first 500 fs by means of non-adiabatic quantum dynamics
including spin vibronic couplings within the linear approach and based on TD-DFT
energies and frequencies [122]. For this purpose we have constructed a model
Hamiltonian including five electronic states, namely S2 which absorbs at 400 nm,
S1 and the three lowest triplet states T3, T2, and T1 (Scheme 6) and up to six
vibrational normal modes. The five states are coupled by SOC and vibronically. The
SOC values are assumed to be constant as functions of the nuclear displacements.
This is justified in a first approximation, for the bromide-substituted complex in
which the calculated SOC remain nearly constant as function of the Re–Br
stretching bond [121]. The calculated emission wavelengths of these five excited
states, including SOC, are reported in Table 6 (Sect. 4.3). To validate the PES
generated within the LVC harmonic model starting from Franck–Condon geome-
try, we have computed the TD-DFT PES as function of the mass and frequency
weighted Re–Br stretching normal mode [121].
Scheme 6 proposes a qualitative correlation between the experimental data
obtained from time-resolved luminescence spectroscopy and the state diagram
built on the basis of the calculated emission wavelengths reported in Table 6
(Sect. 4.3).
The purpose of the simulation based on wave packet propagations with multi-
configuration time-dependent Hartree MCTDH [131, 132] is to recover the
populations of the involved electronic states as function of time within the first
500 fs and to interpret the ultra-fast luminescent decay observed
experimentally [120].
The interplay between SOC and vibronic coupling, based on symmetry rules,
controlled entirely the ISC process. Indeed, when including SOC matrix elements
together with four normal modes of symmetry a0 , namely two modes associated
with the Re–Br stretching and two modes associated with the Re–CO stretching, we
406 C. Daniel
Scheme 6 State diagram representing the low-lying potentially emissive singlet and triplet states
of [Re (Br)(CO)3(bpy]. Experimental data are reported on the left side (adapted from Cannizzo
et al. [120])
observe a fast and efficient population of T1 and T3 (Fig. 7, top) with a rapid decay
of the S2 population to less than 40% within 20 fs. The nearly instantaneous
population of T1 is because of a large value of SOC between this state and S2
(600 cm1), whereas the population of T3 is controlled by the significant degree of
spin-orbit mixing and the small energy gap between this state and S2.
When two additional normal modes of symmetry a00 corresponding to out-of-
plane bending of the bpy ligand are taken into account (Scheme 7), the population
of T1 and T3, controlled mainly by SOC, is not modified but we observe a modest
population of S1 and T2 starting after a few tens of fs and increasing until a
maximum of 20% at 350 fs (Fig. 7, bottom, black dashed and red). Interestingly,
the population of T3 optimum within 25 fs decays rapidly to less than 20%. This
confirms the participation of T3 together with S2 to the early signal observed in the
500–550 nm domain of energy (Scheme 6).
The increase in population of S1 and T2 coupled by spin-orbit (>600 cm1) and
activated by vibronic coupling follows the decay of S2 until about 300 fs, but most
of the triplet population is trapped into T1 which remains stable around 20%. As
expected from the luminescent static properties analyzed in Sect. 4.3, both S1 and
T2 should contribute to the intermediate signal observed within a few hundred of fs
Absorption Spectroscopy, Emissive Properties, and Ultrafast Intersystem. . . 407
Fig. 7 Non-adiabatic quantum dynamics of [Re (Br)(CO)3(bpy)] including SOC with four normal
modes associated to Re-Br and Re-CO a0 stretching modes (top) and SOC with six modes
including two normal modes associated to bpy a00 out-of-plane-bending modes (bottom). The
population of S2 (in black), S1 (dashed line), T1 (in blue), T2 (in red), and T3 (in sienna) are
represented as function of time
408 C. Daniel
Scheme 7 Out-of-plane bending normal modes displacements of the bpy ligand of symmetry a00
calculated at 557 cm1 (left) and 820 cm1 (right)
between 550 and 600 nm (Scheme 6). The states population analysis at longer time
scale is meaningful because of the restricted number of active normal modes
included in the simulation. After 300 fs we observe an exchange of population,
mainly between S2 and T1 strongly coupled with SOC.
This study, which does not take into account the anharmonicity of the excited
states potentials and keeps constant the SOC values as function of the nuclear
displacements, points to the importance of spin vibronic effects in describing ISC
processes. The interplay between geometry distortion and SOC has also been
shown in the phosphorescence decay of [Ir (ppy)3] complexes described in
Sect. 4.1 and in the excited states dynamics of spin crossover complex
[Fe (bpy)3]2+ in Sect. 5.1 of the present contribution, as well as in the pioneering
work by Domcke et al. [87].
6 Concluding Remarks
time attributed to the lowest triplet T1 state. In this case, a detailed study of spin-
induced ZFS and vibronic effects may result in a direct comparison with accurate
experimental findings. The second category of molecules represented by Re
(I) complexes is characterized by shorter lived signals following a cascade of
ultra-fast luminescence decays attributed to Sn and Tn states. The interpretation of
the processes underlying emissive properties has to be based on non-adiabatic
excited states dynamics, either within a full quantum method from first principle
or following surface hopping trajectories within a semi-classical approach. Never-
theless, a TD-DFT static approach based on the optimized structures of the key
emissive states is very useful for deciphering the mechanism as preamble to the
dynamical study. The in-between luminescent behaviors exemplified by flexible Pt
(II) square planar complexes have to be treated with care, some basic approxima-
tion such as Kasha rules being meaningless in that case as illustrated by the
applications reported in the present contribution.
The simulation of ISC processes by means of wave packet propagations on spin-
vibronic coupled multi-dimensional PES is out of reach for large transition metal
complexes. Pioneering studies based on recent methodological developments have
been able to reproduce with reasonable accuracy ultrafast luminescent time scales
observed in first-, second- and third-row transition metal complexes. Counterintu-
itive heavy atom effects on ISC kinetics observed experimentally have found
explanations in detailed theoretical analysis.
Whereas TD-DFT absorption spectra have been validated by a number of joined
experimental/theoretical studies or by comparison with accurate ab initio methods
in the past decade, we do not have the benefit of hindsight as far as the emissive
properties are concerned.
The interpretation of ultra-fast time-resolved spectroscopy outcomes is espe-
cially challenging because both spin-orbit and vibronic coupling effects have to be
considered. In the systems investigated so far involving mostly singlet and triplet
states, the standard level of approximation used for SOC assessment seems realistic.
For higher multiplicities or situations with large mixing between electronic states of
different multiplicities, a more refined treatment of spin-orbit interactions could be
mandatory. The newly developed perturbational treatment for generally applicable
high-level multireference methods is one useful approach [133].
The simulation and computation of ISC rates in pioneering applications are
based on different strategies: (1) non-adiabatic molecular dynamics where
TD-DFT is coupled to classical trajectory-based methods; (2) quantum dynamics
where both electronic and nuclear wave functions are treated exactly within a given
level of approximation; and (3) time-dependent non-radiative rate theory in the
multi-mode harmonic oscillator and Condon approximations, where the electronic
spin-orbit matrix elements depend linearly on the nuclear coordinates within a spin-
vibronic coupling scheme. As illustrated by the examples developed above for first-
, second- and third-row transition metal complexes, each approach has its strengths
and weaknesses. Several important aspects have to be considered in this expanding
field, namely the solvation dynamics, the variation of spin-vibronic couplings with
nuclear relaxation, the construction of realistic model Hamiltonians, and the
410 C. Daniel
References
32. Renske M, van der Veen RM, Cannizzo A, van Mourik F, Vlček A Jr, Chergui M (2011) J Am
Chem Soc 133:305
33. Schinke R (1995) Photodissociation dynamics spectroscopy and fragmentation of small
polyatomic molecules, Cambridge monograph on atomic, molecular and chemical physics.
Cambridge University Press, Cambridge
34. Hochlaf M (2013) Spectroscopy and dynamics of medium sized molecules and clusters. Phys
Chem Chem Phys 15:9967
35. Daniel C (2015) Coord Chem Rev 282–283:19
36. Neese F (2009) Coord Chem Rev 253:526
37. González L, Escudero D, Serrano-Andres L (2012) Chem Phys Chem 13:28
38. Vancoillie S, Zhao H, van Tran T, Hendrickx MFA, Pierloot K (2011) J Chem Theor Comp
7:3961
39. Dreuw A, Head-Gordon M (2005) Chem Rev 105:4009
40. Adamo C, Jacquemin D (2013) Chem Soc Rev 42:845
41. Labat F, Le Bahers J, Ciofini I, Adamo C (2012) Acc Chem Res 45(8):1268
42. Li X, Minaev B, Ågren H, Tian H (2011) Eur J Inorg Chem 2517
43. Zhekova HR, Seth M, Ziegler T (2014) Int J Quantum Chem 114:1019
44. Casida ME, Huix-Rotlant M (2012) Annu Rev Phys Chem 63:287, Johnson MA, Martinez TJ
(eds)
45. Rosa A, Baerends EJ, Gisbergen SJA, van Lenthe E, Groeneveld JA, Snijders JG (1999) J Am
Chem Soc 121:10356
46. Wakamatsu K, Nishimito K, Shibahara T (2000) Inorg Chem Commun 3:677
47. Van Gisbergen SJA, Groeneveld JA, Rosa A, Snijders JG, Baerends EJ (1999) J Phys Chem
A 103:6835
48. Ricciardi G, Rosa A, van Gisbergen SJA, Baerends EJ (2000) J Phys Chem A 104:635
49. Rosa A, Ricciardi G, Baerends EJ, van Gisbergen SJA (2001) J Phys Chem A 105:3311
50. Ricciardi G, Rosa A, Baerends EJ (2001) J Phys Chem A 105:5242
51. Turki M, Daniel C, Záliš S, Vlček A Jr, van Slageren J, Stufkens DJ (2001) J Am Chem Soc
123:11431
52. Full J, González L, Daniel C (2001) J Phys Chem A 105:184
53. Daniel C (2006) Photochemistry of transition metal complexes. In: Crabtree RH
(ed) Encyclopedia of inorganic chemistry. Wiley, Chichester. doi:10.1002/9781119951438.
eibc0175
54. Daniel C (2005) In: Yersin H (ed) Topics in current chemistry, vol 241. Springer, Heidelberg,
p 119
55. Dirac PAM (1928) Proc Roy Soc Lond A 117:610
56. Marian CM (2012) WIREs Comput Mol Sci 2:187
57. Douglas N, Korll NM (1974) Ann Phys 82:89
58. Hess BA (1986) Phys Rev A 33:3742
59. Breit G (1929) Phys Rev 34:553
60. Bethe H, Salpeter E (1957) Quantum mechanics of one- and two-electron atoms. Springer,
Berlin
61. Breit G (1930) Phys Rev 36:363
62. Breit G (1932) Phys Rev 39:616
63. Malmqvist PÅ, Roos BO, Schimmelpfennig B (2002) Chem Phys Lett 357:230
64. Hess BA, Marian C, Wahlgren U, Gropen O (1996) Chem Phys Lett 251:365
65. Christiansen O, Gauss J, Schimmelpfennig B (2000) Phys Chem Chem Phys 2:965
66. van Lenthe E, Baerends EJ, Snijders JG (1993) J Chem Phys 99:4597
67. van Wüllen C (1998) J Chem Phys 109(2):392
68. van Lenthe E, Baerends EJ, Snijders JG (1994) J Chem Phys 101:9783
69. Neese F (2003) J Chem Phys 118:3939
70. Ando H, Iuchi S, Sato H (2012) Chem Phys Lett 535:177
71. Zeng T, Fedorov DG, Schmidt MW, Klobukowski MJ (2012) Chem Theor Comput 8:3061
412 C. Daniel
72. Herzberg G (1966) Molecular spectra and molecular structure III. Electronic spectra and
electronic structure of polyatomic molecules. van Nostrand, New York
73. Koseki S, Schmidt MW, Gordon MS (1998) J Phys Chem A 102(50):10430
74. Koseki S, Fedorov DG, Schmidt MW, Gordon MS (2001) J Phys Chem A 105(35):8262
75. Gourlaouen C, Daniel C (2014) Dalton Trans 43(47):17806
76. Bakova R, Chergui M, Daniel C, Vlček A Jr, Záliš S (2011) Coord Chem Rev 255:975
77. Brahim H, Daniel C, Rahmouni A (2012) Int J Quantum Chem 112:2085
78. Vallet V, Strich A, Daniel C (2005) Chem Phys 311:13
79. Heydova R, Gindensperger E, Romano R, Sykora J, Vlček A Jr, Záliš S, Daniel C (2012) J
Phys Chem A 116:11319
80. Brahim H, Daniel C (2014) Comp Theor Chem 1040/104:219
81. Li X, Minaev B, Ågren H, Tian H (2011) J Phys Chem C 115:20724
82. Ronca E, De Anglis F, Fantucci S (2014) J Phys Chem C 118:17067
83. Fantacci S, Ronca E, De Angelis F (2014) J Phys Chem Lett 5:375
84. K€oppel H, Domcke W, Cederbaum LS (1984) Adv Chem Phys 57:59
85. Hazra A, Nooijen M (2005) Phys Chem Chem Phys 7:1759
86. Neugebauer J, Baerends EJ, Nooijen M (2005) J Phys Chem A 109:1168
87. Mondal P, Opalka D, Poluyanov LV, Domcke W (2012) J Chem Phys 136:084308
88. Mondal P, Opalka D, Poluyanov LV, Domcke W (2011) Chem Phys 387:56
89. Capano G, Penfold TJ, R€othlisberger U, Tavernelli I (2014) CHIMIA 68(4):227
90. Sousa C, de Graaf C, Rudavskyi A, Broer R, Tatchen J, Etinski M, Marian CM (2013) Chem
Eur J 19:17541
91. El Nahhas A, Cannizzo A, van Mourik F, Blanco-Rodrı́guez AB, Zalis S, Vlček A Jr, Chergui
M (2010) J Phys Chem A 114:6361
92. Kuimova MK, Alsindi WZ, Blake AJ, Davies ES, Lampus DJ, Matousek P, McMaster J,
Parker AW, Towrie M, Sun X-Z, Wilson C, George MW (2008) Inorg Chem 47:9857
93. Huse N, Kim TK, Jamula L, McCusker JK, de Groot FMF, Schoenlein RW (2010) J Am
Chem Soc 132:6809
94. Huse N, Cho H, Hong K, Jamula L, de Groot FMF, Kim TK, McCusker JK, Schoenlein RW
(2011) J Phys Chem Lett 2:880
95. Bressler CH, Milne C, Pham V-T, El Nahhas A, van der Veen RM, Gawelda W, Johnson S,
Beaud P, Grolimund D, Kaiser M, Borca CN, Ingold G, Abela R, Chergui M (2009) Science
323:489
96. Zhang W, Alonso-Mori R, Bergmann U, Bressler C, Chollet M, Galler A, Gawelda W, Hadt
RC, Hartsock RW, Kroll T, Kjær KS, Kubiček K, Lemke HT, Liang HW, Meyer DA, Nielsen
MM, Purser C, Robinson JS, Solomon EI, Sun Z, Sokaras D, van Driel TB, Vankó G, Weng
T-C, Zhu D, Gaffney KJ (2014) Nat Lett 509:345
97. Blanco-Rodrı́guez AM, Kvapilová H, Sýkora J, Towrie M, Nervi C, Volpi G, Záliš S, Vlček
A Jr (2014) J Am Chem Soc 136:5963
98. El-Sayed MA (1963) J Chem Phys 38:2834
99. Englman R, Jortner J (1970) Mol Phys 18:145
100. Freed KF, Jortner J (1970) J Chem Phys 52:6272
101. Etinski M, Tatchen J, Marian CM (2011) J Chem Phys 134:154105
102. Etinski M, Rai-Constapel V, Marian CM (2014) J Chem Phys 140:114104
103. Tatchen J, Gilka N, Marian CM (2007) Phys Chem Chem Phys 9:5209
104. Curchod BFE, Rothlisberger U, Tavernelli I (2013) Chem Phys Chem 14:1314
105. K€oppel H, Domcke W (1998) In: von Ragué Schleyer P (ed) Encyclopedia in computational
chemistry. Wiley, New York, p 3166
106. Domcke W, Yarkony DR, K€oppel H (eds) (2004) Conical intersections – electronic structure,
dynamics and spectroscopy, vol 15, Advanced series in physical chemistry. World Scientific,
Singapore
107. Baer M (2006) Beyond Born-Oppenheimer electronic nonadiabatic coupling terms and
conical intersection. Wiley-Interscience, Hoboken
Absorption Spectroscopy, Emissive Properties, and Ultrafast Intersystem. . . 413
108. Curchod BFE, Penfold TJ, Rothlisberger U, Tavernelli I (2013) Cent Eur J Phys 11:1059
109. de Carvalho FF, Bouduban MEF, Curchod BFE, Tavernelli I (2014) Entropy 16:62
110. Worth GA, Welch G, Paterson MJ (2006) Mol Phys 104:1095
111. McKinlay RG, Zurek JM, Paterson MJ (2010) Advances in inorganic chemistry, vol 62.
Academic, p 351
112. Holzer W, Penzkofer A, Tsuboi T (2005) 308:93
113. Ciofini I, Adamo C (2007) J Phys Chem A 111:5549
114. Vlček A Jr, Záliš S (2007) Coord Chem Rev 25:258
115. Hay PJ (2002) Phys Chem A 106:1634
116. Nozaki KJ (2006) Chin Chem Soc 53:101
117. Jansson E, Minaev B, Schrader S, Ågren H (2007) Chem Phys 110:1849
118. Fine J, Diri K, Krylov AI, Nemirow C, Lu Z, Wittig C (2012) Mol Phys 110:1849
119. Nemirow C, Fine J, Lu Z, Diri K, Krylov AI, Wittig C (2012) Mol Phys 110:1893
120. Cannizzo A, Blanco-Rodrı́guez AM, El Nahhas A, Sebera J, Záliš S, Vlček A Jr, Chergui M
(2008) J Am Chem Soc 130:8967
121. Gourlaouen C, Otsuka M, Gindensperger E, Daniel C (2015) J Chem Theory Comp 11(1):99
122. Eng J, Gourlaouen C, Gindensperger E, Daniel C (2015) Acc Chem Res 48(3):809
123. Jansson E, Minaev B, Schrader S, Ågren H (2007) Chem Phys 333:157
124. Finkenzeller WJ, Yersin H (2003) Chem Phys Lett 377:299
125. de Graaf C, Sousa C (2011) Int J Quantum Chem 111:3385
126. Tavernelli I, Curchod BFE, Rothlisberger U (2011) Chem Phys 391:101
127. Yeh AT, Shank CV, McCusker JK (2000) Science 289(5481):935
128. Cannizzo A, van Mourik F, Gawelda W, Zgrablic G, Bressler C, Chergui M (2006) Angew
Chem Int Ed 3174
129. Tully JC (1990) J Chem Phys 93:1061
130. Wang F, Ziegler T (2005) J Chem Phys 123(15):154102
131. Manthe U, Meyer H-D, Cederbaum LS (1992) J Chem Phys 97:3199
132. Meyer H-D, Manthe U, Cederbaum LS (1990) Chem Phys Lett 165:73
133. Mai S, Müller T, Plasser F, Marquetand P, Lischka H, González L (2014) J Chem Phys
141:074105
Top Curr Chem (2016) 368: 415–444
DOI: 10.1007/128_2014_605
# Springer International Publishing Switzerland 2014
Published online: 24 February 2015
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
2 Surface Hopping Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
3 Excited States in DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
3.1 Excitation Energies in DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
3.2 Nonadiabatic Couplings in DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
3.3 Critical Appraisal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
M. Barbatti (*)
Max-Planck-Institut für Kohlenforschung, Kaiser-Wilhelm-Platz 1, 45470 Mülheim an der
Ruhr, Germany
e-mail: barbatti@kofo.mpg.de
R. Crespo-Otero
School of Biological and Chemical Sciences, Queen Mary University of London, Mile End
Road, London E1 4NS, UK
416 M. Barbatti and R. Crespo-Otero
Abbreviations
1 Introduction
d2 Rm 1 ∂EK
¼ ; ð1Þ
dt2 Mm ∂Rm
where EK is the adiabatic potential energy of the current state K. Given a set of
initial positions and velocities, (1) is numerically integrated.
Along with the Newton’s equations, the probability for the system to hop to
another state L is evaluated. Diverse schemes for the evaluation of such probabil-
ities have been developed [19, 23–30]. The most successful and popular approach
has been the fewest switches proposed by Tully in the early 1990s [28].
In the fewest switches, the number of hopping events within one time step Δt is
minimized. Under this condition, the hopping probability between states K and L is
The coefficients cJ are obtained by solving a local approximation for the time-
dependent electronic Schrödinger equation, given in the adiabatic representation by
[28]
420 M. Barbatti and R. Crespo-Otero
dcL i X
þ EL c L þ σ LJ cJ ¼ 0: ð4Þ
dt h J
In this equation, the coupling terms between any pair of states L and M are
∂
σ LM ΨL ΨM ¼ τLM v; ð5Þ
∂t
where SLM ðtÞ hΨL ðt ΔtÞjΨM ðtÞi are wavefunction overlaps between different
time steps. This method can be generally used for any electronic-structure method,
provided that a configuration interaction representation of the electronic
wavefunction can be worked out [32–35]. In the last part of (7), the coupling is
conveniently written in terms of full time steps (t, t Δt, t 2Δt) rather than in
terms of midpoints (t + Δt/2, t Δt/2) as in the original model. This shift is
explained in [33]. Comparisons between couplings computed with the finite-
difference approach and with analytical derivatives are made in [33, 35, 36].
Alternatively, cL can still be obtained by the local diabatization approach [37]. In
this case, instead of integrating (4), the array of coefficients after one time step is
given by
1
1 1 EðtÞ þ TEðt þ ΔtÞT
cðt þ ΔtÞ ¼ T exp ih Δt cðtÞ; ð8Þ
2
XD E
SLM ðtÞ ¼ ηL ðt ΔtÞηJ ðtÞ T JM ðtÞ: ð9Þ
J
In this equation, {jηi} represents the diabatic basis, which is obtained along the
trajectory as explained in [37]. It has been shown that this method is more stable in
the presence of weak nonadiabatic couplings than conventional algorithms [38]. An
alternative surface-hopping diabatization method is discussed in [39].
Either via (4) or (8), as soon as the coefficients cL are obtained, the hopping
probability can be computed and within the fewest switches approach in the
adiabatic representation it is given by
" #
2Δt
PK!L ¼ max 0, Re cK c*L σ LK : ð10Þ
jcK j2
In the most recent implementations of the fewest switches, the coefficients cL are
corrected for decoherence effects [8, 40–42] before probabilities are computed [43,
44].
Jaeger, Fischer, and Prezhdo recently proposed the decoherence-induced surface
hopping (DISH) method, a hopping algorithm which relies entirely on the
decoherence times of each adiabatic state to determine the state branching
[30]. Another recently proposed alternative to the fewest switches is global-flux
surface hopping (GFSH) [23], which computes the hopping probability between
groups of states with reduced or increased population. In this way, hops can occur
even between indirectly coupled states (super-exchange).
With hopping probabilities at a time t, a stochastic algorithm is invoked to decide
in which state the dynamics continue to be in the next time step. A hopping from
state K to state L occurs if a uniformly selected random number rt in the [0, 1]
interval is such that
X
L1 X
L
PK!J ðtÞ < r t PK!J ðtÞ; ð11Þ
J¼1 J¼1
and the energy gap between the final and initial states satisfies [22]
Equation (12) ensures that, if the nuclear kinetic energy (Ekin) cannot compensate
the variation of potential energy, the hop is rejected (“frustrated hop”). If the state
changes, the momentum is changed accordingly to ensure conservation of total
energy. Normally, the momentum adjustment is carried out in the direction of the
nonadiabatic coupling vector. When the vector direction is not available, as in the
case of computation of the coupling terms via (7), then the adjustment may be
carried out in the linear momentum direction.
422 M. Barbatti and R. Crespo-Otero
where HSOLM are the spin-orbit couplings [46] and the last term accounts for the
dipole interaction of the molecule with the field (μML is the transition dipole
moment between the L and M and A0 is the vector potential) [49]. In Sect. 3.2,
we discuss how these couplings can be computed within the DFT framework.
In this section, different approximations for the computation of excited states based
on DFT are analyzed. We start from a general description of the popular linear-
response (LR) TDDFT. A number of other methods for computing excitation
energies based on DFT have been used for surface hopping as well. They are also
described here, highlighting the hierarchic relations between them. Methods other
than linear response – such as real-time TDDFT [51, 52] – or beyond linear
response [53] can also be used to study excited states, but discussion of these is
beyond of the scope of this chapter as these methods have not yet been generally
applied for surface hopping. Multiconfigurational DFT is briefly addressed in
Sect. 3.3.
The excitation energies ωL in the LR-TDDFT are given by [54]
! !
ΩF L ¼ ω2L F L ; ð14Þ
where
In these equations, i and j denote occupied orbitals and a and b denote virtual
orbitals. A and B are defined here for a restricted approach, single excitation, and a
functional without any fraction of Hartree–Fock exchange. For more general
expressions, see Eq. (4.33) of [54]. For hybrid functionals, see Eqs. (95) and
(96) of [55].
In (16) and (17),
Δεjb ¼ εb εj ; ð18Þ
for real KS orbitals ψ k and within the adiabatic local density approximation
(ALDA) [54]:
δ2 EXC
f XC ¼ ; ð22Þ
δρðrÞδρðr0 Þ
GS p→ r p→ s
Fig. 1 Schematic illustration of a reduced system with one occupied and two virtual orbitals. The
ground state (GS) determinant can give an origin to two single excitations
Ωpr, pr Ωpr, ps F1 F1
¼ω 2
: ð23Þ
Ωps, pr Ωps, ps F2 F2
s
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
Ωpr, pr þ Ωps, ps Ωpr, pr Ωps, ps 2
ω22, 1 ¼ þ Ω2pr, ps : ð24Þ
2 2
If the nondiagonal terms of Ω are null (which occurs when K pr, ps ¼ K ps, pr ¼ 0,
see (26)), the excitations energies are simply
ω1 ¼ Ω1=2
pr , pr ω2 ¼ Ω1=2
ps, ps ; ð25Þ
implying that the energy of each state is associated with a unique singly-excited
determinant (for instance, p ! r) and independent from the remaining
determinants.
The matrix elements of Ω can be explicitly written as (see Eq. (4.33) of [54])
pffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffi
Ωia, jb ¼ δij δab Δε2jb þ 2 Δεia K ia, jb Δεjb : ð26Þ
This result corresponds to the excitation energy computed with density functional
perturbation theory to the first order [58] and it is also directly obtained with TDA
(B ¼ 0).
If Kpr,pr is completely neglected, the excitation energy is given simply by the
bare energy gap between orbitals
Surface Hopping Dynamics with DFT Excited States 425
ω1 ’ Δεpr ¼ εr εp : ð29Þ
Nonadiabatic couplings between different electronic states are needed for propa-
gation of surface hopping dynamics. While analytical energy gradients for excited
states computed with TDDFT are well established [72] and implemented in diverse
computational-chemistry programs, analytical nonadiabatic couplings are still
mostly unavailable.
426 M. Barbatti and R. Crespo-Otero
Ψ0 ¼ jKSi; ð30Þ
rffiffiffiffiffiffiffiffiffi
X Δεjb jb
ΨL ¼ F j jbi; ð31Þ
j, b
ωL L
where |KSi is the ground-state KS determinant and jjbi is the determinant with a
single excitation from j to b. With these wavefunctions, which are analogous to a
CIS expansion, the coupling terms σLM can be evaluated according to (7) in terms of
atomic orbital overlap integrals [33]. Although the validity of Casida’s ansatz for
computation of couplings approach still needs to be extensively tested, it has been
shown that TDDFT dynamics computed with these couplings compares well with
dynamics based on CASSCF [79] and MRCI [80].
Equation (31) forms a non-orthogonal set, which can have consequences for the
evaluation of the couplings. Werner et al. [81] recommended the use the linear-
response coefficients without the square-root term in (31) as the CIS coefficients,
which forms an orthogonal set within TDA. An alternative solution is to orthogo-
nalize the approximate wavefunctions given by (31) before computing the cou-
plings. In practical terms, far from conical intersections with the ground state, either
of these approximations provides similar dynamical pictures. In some cases, neg-
ative excitation energies are obtained (see Sect. 3.3), and consequently the
wavefunctions are complex and imaginary couplings may be obtained. In such
situations, as we discuss later, it is not recommended to continue the SH/TDDFT
propagation.
The Casida wavefunctions given in (30) and (31) have been employed by
Tavernelli and co-workers to derive analytical nonadiabatic coupling vectors
Surface Hopping Dynamics with DFT Excited States 427
between the ground and the first excited states [82] and also between excited states
[83]. They are given by
rffiffiffiffiffiffiffiffiffi
X Δεia ia
τ0L ¼ F hψ j∇R H KS jψ a i; ð32Þ
ia
ωL L i
rffiffiffiffiffiffiffiffiffirffiffiffiffiffiffiffiffiffi
X Δεia Δεib ia{
τLM ¼ F Fib hψ j∇R H KS jψ b i
ωL
iab rffiffiffiffiffiffiffiffi
ωM L M a
ffirffiffiffiffiffiffiffiffiffi ð33Þ
X Δεia Δεja ia{ ja
FL FM ψ i j∇R HKS jψ j :
ija
ωL ωM
(As in the previous section, ψ k corresponds to the molecular orbitals, i, j run over
occupied orbitals, and a, b over virtual orbitals.)
In [84] it is shown that τ0L in (32) is equivalent to the analytical coupling vector
derived by Hu et al. in [74], but, also as the Hu et al.’s result, it does not fully agree
with Chernyak–Mukamel [73]. Ou et al. [36] have pointed out that in the formalism
leading to (32) and (33), only the KS Fock matrix is differentiated, omitting the full
coupling induced by the second derivative of the exchange-correlation functionals.
Still working with a CIS expansion based on TDA amplitudes, these authors have
derived nonadiabatic coupling vectors (including excited-excited terms) fully con-
sistent with Chernyak–Mukamel [36].
Apart from their limitations, (32) and (33) are general results, which are still
valid by replacing the operator ∇R H KS by any single-body operator Ô [83], such as,
for example, the electronic dipole operator, to obtain the transition dipole moment
μLM.
Recently, spin-orbit coupling elements [85] were also derived based on the
Casida wavefunctions, opening the possibility of performing surface hopping
between surfaces with different multiplicities within the TDDFT approach. The
coupling between singlet and triplet states, for instance, is given by
rffiffiffiffiffiffiffiffiffiffiffi rffiffiffiffiffiffiffiffiffiffiffi
X Δεiaρ iaρ Δεjbρ0 jbρ0
SO
H ST ¼ FS FT hiaS jH SO jjbT i; ð34Þ
iaρ, jbρ0 ω S ω T
where ρ and ρ0 are spin indexes. In [85], HSO is approximated by the one-electron
Breit–Pauli operator [86]. The computation of the two-electron operator is
discussed by Chiodo and Russo in [87, 88].
In [89], the computation of Dyson orbitals based on Casida wavefunctions is
discussed. These Dyson orbitals were used to evaluate photoionization cross-
sections during the dynamics and to simulate time-dependent photoelectron imag-
ing spectra.
In SDKS methods, the computation of the couplings is largely simplified by the
singe-determinant approximation. As shown in (21) of [61], the couplings in such
428 M. Barbatti and R. Crespo-Otero
Fig. 2 Illustration of the main problems with DFT and TDDFT occurring in different regions of
the ground- and excited-state surfaces
Surface Hopping Dynamics with DFT Excited States 429
and, consequently, of the excited-states based upon. In particular, the methods for
computing excitation energies discussed in Sect. 3.1 have in common the assump-
tion that the ground state can be described by a single KS determinant. (For a
conceptual discussion of nondynamic electron correlation in DFT, see [95]. For a
discussion about conical intersection in DFT, see [96]).
The development of a density functional theory including nondynamic
(or “strong”) electron correlation – which Becke designated as the “last frontier”
in DFT [97] – has been pursued by many groups following different approaches.
Among these approaches, we may cite the use of restricted open-shell ground-state
representations [98], configuration ensembles with fractional occupations [91, 99–
101], configuration interaction [102, 103], multiconfigurational DFT [104], hybrid
multiconfiguration/(TD)DFT [105, 106], and spin-unrestricted broken-symmetry
(UBS) [107] approaches. Unfortunately, analytical energy gradients are not avail-
able for most of these methods, which rules out their use in surface hopping
dynamics.
All the troubles caused by nondynamic electron correlation are rather restricted
to the crossing between the ground and the first excited state. In the case of crossing
between excited states, however, if each of these states is well represented by single
excitations from a well-behaved ground state, the description of the crossing does
not present further problems.
SH/DFT has other potential problems (Fig. 2) besides nondynamical electron
correlation. First, the results are deeply dependent on the functional. Second,
double and higher excitations are not properly described by LR-TDDFT. Third,
diffuse and charge-transfer states may be poorly described by conventional func-
tionals. All these problems, though, are not exclusive of dynamics simulations and
are part of routine investigations of excited states based on DFT. As such, they are
addressed in the same way, by systematic test of functionals, methodological
comparisons, and use of range-separated functionals.
To illustrate the current situation of excited state description for SH/DFT, we
show in Fig. 3 the potential-energy surfaces for the S1 state and for the S1/S0 gap of
ethylene along two important reaction coordinates, H2CCH2-torsion (θ) and CCH2-
pyramidalization (φ). All other coordinates are kept at their ground-state values
optimized at the same level as used for energy calculations. These surfaces were
computed with several different DFT-based methods and with ab initio MR-CISD,
which is taken as the reference method. With the exception of the DFT-MRCI
[102], computed with the SV(P) basis set [108], all other DFT-based surfaces were
computed with the 6-31G* basis set [109]. LC-BLYP [110–112] was computed
with μ ¼ 0.2 a01 [113]. TD-DFTB was based on analytic expressions for the matrix
elements [114]. MR-CISD was based on a small (two electron, two orbital)-space
[115] with the aug-cc-pVTZ basis set [116] assigned to C and cc-pVDZ assigned
to H.
Ethylene S1 surface is a specially challenging problem, starting from (i) the
adequate computation of the excitation into the V state [6], then (ii) the description
of the right topography of the state, which includes a crossing with the Z state near
the twisted structure and a conical intersection at twisted-pyramidalized geometries
430 M. Barbatti and R. Crespo-Otero
Fig. 3 S1 and S1/S0 potential-energy surfaces of ethylene computed with diverse methods. Green
regions indicate the crossing seam. Red/orange regions are negative energy gaps. In the illustra-
tion, ethylene has θ ¼ 90 and φ ¼ 45
[115], to finally (iii) the dynamics evolution itself, which has motivated a long
debate between theorists and experimentalists [117–119]. In fact, Levine
et al. [120] have used maps similar to those in Fig. 3 (but computed for stretched
CC distances) to discuss the qualitative deficiencies in the excited-state description
provided by TDDFT.
The MR-CISD result in Fig. 3 has two main features which dominate the
topography and the dynamics in the S1 state: (a) there is an S1 minimum near
θ ¼ 90 and φ ¼ 60 and (b) there is a conical intersection at θ ¼ 90 and φ ¼ 110 .
Among all tested methods, the only ones able to reproduce both features are the
DFT-MRCI [102] and REKS [91, 99, 101], which makes clear the importance of
having a multiconfigurational description of the ground state. TDDFT-UBS, which
is usually considered a good alternative for qualitatively recovering nondynamic
electron correlation near a degeneracy [71], can describe reasonably well the
Surface Hopping Dynamics with DFT Excited States 431
conical intersection, but the topography along the θ ¼ 90 line shows diverse
spurious minima.
Spurious minima were also observed in TDDFT and TDA with B3LYP [121,
122]. The minimum is at a qualitatively wrong position at TDDFT-B3LYP with
restricted open shell (TDDFT-RO). The same happens for TDDFT-BLYP. TDA
can qualitatively predict the conical intersection (θ ¼ 90 , φ > 70 ). Somewhat
surprisingly, the simple HOMO–LUMO gap is the only other method predicting a
twisted-pyramidalized conical intersection (at θ ¼ 90 , φ ¼ 70 ). All other methods
wrongly predict an extended S1/S0 seam along the θ ¼ 90 line, starting from
φ ¼ 0 .
A disturbing feature which can be observed in TDDFT-B3LYP, TDDFT-UBS-
B3LYP, TDA-B3LYP, and TDDFT-LC-BLYP is that the excitation may become
negative near the crossing seam. This is not an exclusive problem with DFT-based
methods; it can be observed in other single-reference methods as well, such as
coupled cluster with either equation-of-motion or linear-response approaches.
As discussed in [123], near the degeneracy, self-interaction errors may cause the
HOMO to be less stable than the LUMO, leading to imaginary excitation energies
in TDDFT – see (27). This does not happen at TDA because the excitation energy is
given by terms such as those in (28). This feature has been said to represent an
improved stability of the TDA-based dynamics in comparison to that based on
TDDFT [123]. Nevertheless, as we can see in the negative gap regions in Fig. 3,
both methods are still unstable near the degeneracy. These negative excitations are
clearly caused by the mixing with the other singly-excited determinants, which
cannot be avoided either in TDDFT or in TDA.
In SDKS methods, the excitation energies are given by the bare KS gaps [61]. In
the case of the ethylene, we can see in Fig. 3 that the bare KS energy (given by the
HOMO–LUMO gap and neglecting double excitations near the twisted structure)
gives an adequate representation of the S1 state and of the S1/S0 gap. This good
behavior, however, should not be taken for granted. Maitra [60], in a critical
discussion of these approximation in the context of surface hopping, showed that
the bare KS energies may be far from adiabatic and closer to diabatic energies.
Ethylene is admittedly too harsh a case for DFT-based methods, as its dynamics
is deeply controlled by coordinates involved in the nonadiabatic events. SH/DFT
methods have been developed to deal with large molecular systems and in these
cases the dynamics may evolve in the configurational space spanning regions
relatively far from any muticonfigurational ground state. Under such situation, the
excited-state dynamics involving a large manifold of excited states can be well
simulated with DFT-based methods.
Take, for instance, the schematic dynamics illustrated in Fig. 4 (top). The
dynamics starts at a high excited state (here, the second state to simplify the
picture). Using a multireference method such as MRCI, we would observe a
relaxation to the first excited state (a), then oscillation around the minimum of
this state (b). From this minimum, the molecule can eventually fluoresce to the
ground state (c) or cross a barrier (d) to reach a conical intersection, from where it
relaxes to the ground state minimum (e).
432 M. Barbatti and R. Crespo-Otero
Fig. 4 Schematic
comparison of a trajectory
computed with surface
hopping based on
multireference method (top)
and on TDDFT (bottom)
4 Surface-Hopping/DFT Implementations
One of the reasons for the popularity of the surface hopping method is that its
implementation is straightforward. This has given rise to several home-made
private codes to simulate SH/DFT [62, 63, 81, 129]. There are also a few general
programs of public access (either commercial or non-commercial) with SH/DFT
capability, including Newton-X [130, 131], PYXAID [61, 132], Turbomole [78],
and CPMD (www.cpmd.org).
The main difference between the several implementations of SH/DFT is exactly
which DFT method is being used for computing the electronic-structure quantities.
The main features of the most common implementations are summarized in
Table 1.
The most computationally efficient implementation of SH/DFT available is
based on SDKS approximations [134], such as that implemented in the PYXAID
program using plane waves. Variants of this method have also been developed by
Fischer, Handt, and Schmidt [50, 62, 135] using Gaussian functions, by Gao and
co-workers using DFTB KS orbitals [63], and by Shenvi, Roy, and Tully based on a
model Hamiltonian parameterized by DFT data [64, 136]. SDKS has been used to
investigate diverse problems, especially in condensed matter (see Sect. 5). Besides
the computational efficiency granted by the single-determinant approximation,
Shenvi and co-workers [64] have pointed out that, while in TDDFT the electronic
Hilbert space must be truncated to include only a relatively small number of states,
this restriction does not apply to SDKS, being a big advantage for the treatment of
systems with large density of states. Moreover, SDKS also allows the inclusion of
double and higher excitations [61]. On the down side, the bare KS energy gaps may
Table 1 Survey of diverse implementations of SH/DFT. Excited states can be computed with
linear response time-dependent (LR-TD) theory, single determinant KS (SDKS), or restricted open
shell KS (ROKS); single (SS) or multiple (MS) excited states can be included; states can be
restricted to single (SE) or multiple (ME) excitations; propagation can be done in terms of
Gaussian functions (GF) or plane waves (PW)
DFT Number of Excitation Basis Public
excitations states level type Refs. implementation
LR-TDDFT MS SE GF [32] Newton-X
LR-TDDFT SS SE GF [77] Turbomole
LR-TD-DFTB MS SE GF [133]
TDA MS SE GF [38] Newton-X
TDA MS SE PW [34, 83] CPMD
ROKS SS SE PW [68] CPMD
SDKS MS ME GF [63]
SDKS MS ME GF [62]
SDKS MS ME PW [61, PYXAID
134]
434 M. Barbatti and R. Crespo-Otero
not only deviate strongly from the exact energies, but also provide a poor approx-
imation for adiabatic surfaces [60].
Still aiming at maximum computational optimization, Prezhdo and co-workers
have also introduced the so-called “classical path approximation” (CPA) in
PYXAID [61]. The CPA implies that the nuclear dynamics is supposed to evolve
independently of the electrons, driven by kinetic effects. In practical terms, it means
that the nuclear dynamics is propagated in the ground state and then is used to
compute the nonadiabatic electronic events in the excited states using either the
fewest switches surface hopping or one of the other surface hopping algorithms
developed by the Prezhdo group [132]. The CPA may be especially inadequate for
systems undergoing significant chemical changes such as photofragmentation and
chemical reactions.
Another efficient implementation of SH/DFT is based on ROKS and was
developed by Doltsinis and Marx [137]. The spin-adapted wavefunctions and the
KS orbitals optimized for the excited-state density should in principle be an
improvement over the SDKS. SH/ROKS, however, is limited to a single excited
state, which is a very strong limitation for most realistic problems. Such an
approach is implemented in the CPMD program.
SH/LR-TDDFT has been pioneered by Tavernelli and Rothlisberger [34] and by
Mitrić and Bonačić-Koutecký [81]. Linear response should provide a better descrip-
tion of the excited-state surfaces than single determinant and ROKS approaches, at
higher computational costs naturally (see Sect. 3.1). TDDFT is also not limited to a
single surface as ROKS, but it cannot describe multiple excitations. In the frame-
work of linear response, surface hopping dynamics has been investigated with
TDDFT, TDA, and TD-DFTB (see Sect. 5). In the CPMD program, this approach
is implemented based on plane waves, while in Newton-X, it is implemented on a
localized basis. It is also implemented in Turbomole, but limited to couplings
between the ground state and the first excited state. Ehrenfest dynamics [138],
another related semiclassical nonadiabatic dynamics method, can be performed
with TDDFT using the Octopus program [139].
In the case of Newton-X, the coefficients cL to compute the hopping probabilities
(10) can be obtained either by integrating (4) or by using the local diabatization
method (8). The program also allows the computation of surface hopping through
interfaces with different programs (Turbomole, Gaussian [140], Columbus [141,
142], Gamess [143]) and using different methods (TDA, TDDFT, ADC(2), CC2,
CASSCF, MRCI), making it particularly convenient for comparative analysis.
5 Case Studies
There are a large number of molecular systems which have been investigated with
SH/DFT. This section does not aim at providing a comprehensive review of them,
but instead at pointing out the main classes of problems in different fields, from
where the reader can search for more information.
Surface Hopping Dynamics with DFT Excited States 435
SH/DFT has been developed mainly for the treatment of large systems, for
which wavefunction-based methods are computationally too expensive. It is possi-
ble, however, to find a series of investigations for photoreactions of small molecules
(see Table 2), which has been useful to gauge the quality of these simulations.
CH2NH2+, for instance, has been used as a test case of SH/TDA [34, 79] and
SH/SDKS [149]. In both cases, qualitative agreement with surface SH/CASSCF
[174] is observed. We should note, however, that this agreement may be accidental,
as CH2NH2+ is a very small system with only two relevant excited states, a single
dominant reaction path, and a lifetime shorter than 100 fs. More impressive is the
semi-quantitative agreement observed between SH/TDDFT [80] and SH/MRCI
[175] for pyrrole. In this case, TDDFT dynamics with ten excited states has
successfully predicted the excited-state lifetime and the split of population among
several reaction paths.
A more applied class of systems investigated with SH/DFT excitations involves
photoinduced proton transfer and photoinduced isomerization in medium-sized
molecules. A typical example of such a class is azobenzene, which has been studied
with SH/ROKS [164] and SH/SDKS [165]. Dynamics based on both provides a
good description of the cis-trans isomerization of azobenzene in the gas phase in
comparison to other semiempirical and ab initio wavefunction-based surface hop-
ping simulations [176–180]. Azobenzene has been chosen by a number of groups,
including ours, as a standard test case for methods. We should consider, however,
that this molecule may not be challenging enough to be a good test case. After S1
excitation, azobenzene evolves adiabatically until finding the intersection to the
ground state, approaching the crossing seam along torsional coordinates. These
features imply that neither nonadiabatic effects between excited states nor the
dimensionality of the crossing seam can be really tested with this system. On the
other hand, azobenzene is an excellent system to probe the topography of S1 and the
S1/S0 coupling strength, which can be done through simulations of excited-state
lifetime and isomerization yield.
One of the main problems with SH/DFT is the deep dependence on the chosen
functional. This can be illustrated with an example we have recently investigated,
the excited-state dynamics of N-methylformamide dimer (NMF) [127]. These
simulations showed that NMF dimers are protected against photodissociation by a
proton-transfer mechanism. The excited-state proton transfer occurs according the
Sobolewski–Domcke mechanism [181], where an electron is transferred first,
followed by the proton (see Fig. 5). For properly describing the charge-transfer
state, SH/TDDFT was done with the range-separated LC-BLYP functional [110–
112]. The range-separation parameter was fixed at μ ¼ 0.2 a01, a value based on a
non-empirical parameterization [113]. Our tests with diverse other values of μ
showed that the ratio between dissociation and proton transfer was deeply depen-
dent on this parameter. Not surprisingly, larger values favored dissociation by
under-stabilizing the charge-transfer state. (In the Gaussian program [140], for
instance, the default value of μ is 0.47 a01.) Lower values favored proton transfer
for the opposite reason.
Diverse systems of biological interest have also been investigated with SH/DFT
(see Table 2). In this class, a challenging case has been the description of purine
nucleobases. An indication of the problem was already in the earliest simulations of
9Me-keto guanine with SH/ROKS [151, 152], whose trajectories did not reveal any
conical intersection with the ground state. At that point, the internal conversion of
Surface Hopping Dynamics with DFT Excited States 437
8 fs 15 fs
Fig. 6 Ground-state 80
Ground State Population at 1 ps (%)
SH/SDKS has allowed the limits of the simulations to be stretched much beyond
what can currently be done with TDDFT or wavefunction-based methods. In
particular, it has allowed one to investigate large organic chromophores [63],
adsorbance of molecules on metal [64, 136] and semiconductor [134, 172] surfaces,
Auger dynamics in quantum dots [23], carbon nanostructures [160, 162], and
organic crystals [61, 166].
An interesting example in this class of systems related to condensed matter and
materials science is the recently published simulations for a P3HT/carbon-nanotube
heterojunction [161]. Organic heterojunctions have been intensively explored for
the development of organic photovoltaics based on photoinduced electron transfer
[183, 184]. P3HT (poly(3-hexylthiophene)) is a standard organic polymer used as
chromophore and electron donor [185], while carbon nanotubes (CNT) are
electron-acceptor materials with enhanced charge-transport properties [186]. Not-
withstanding the limitations of DFT to approach this type of system [187], dynam-
ics with SH/SDKS predicts that there is a strong asymmetry between the electron
and hole transports in P3HT/CNT interface (Fig. 7). While photoexcitation of P3HT
leads to an electron transfer within 100 fs, the hole transfer takes much longer,
occurring on the few picoseconds scale.
6 Conclusions
In the last decade, surface hopping dynamics has become an essential tool for the
investigation of nonadiabatic processes in diverse fields, providing fundamental
information to interpret data from time-resolved spectroscopy, to explain photo-
chemical processes, and to predict new properties with potential technological
applications. Motivated by the advances in computational capabilities and algo-
rithms, such simulations are under constant pressure to address ever larger and more
complex systems. The development of SH/DFT has opened possibilities to go far
beyond where wavefunction ab initio methods could achieve.
Surface Hopping Dynamics with DFT Excited States 439
Acknowledgments We would like to thank Dr. Michael Filatov for kindly providing the REKS
data for Fig. 3 and also Dr. W. Arbelo-González, Dr. Fazzi, Dr. G. Rodrigues, and Dr. T. Very for
helpful discussions.
References
1. Winter NOC, Graf NK, Leutwyler S, Hattig C (2013) Phys Chem Chem Phys 15:6623
2. Silva-Junior MR, Thiel W (2010) J Chem Theory Comput 6:1546
3. Silva-Junior MR, Schreiber M, Sauer SPA, Thiel W (2010) J Chem Phys 133:174318
4. Schreiber M, Silva-Junior MR, Sauer SPA, Thiel W (2008) J Chem Phys 128:134110
5. Silva-Junior MR, Schreiber M, Sauer SPA, Thiel W (2008) J Chem Phys 129:104103
6. Wu W, Zhang H, Braı̈da B, Shaik S, Hiberty P (2014) Theor Chem Acc 133:1
7. Dreuw A, Weisman JL, Head-Gordon M (2003) J Chem Phys 119:2943
8. Yonehara T, Hanasaki K, Takatsuka K (2011) Chem Rev 112:499
9. Yarkony DR (2011) Chem Rev 112:481
10. Worth GA, Cederbaum LS (2004) Annu Rev Phys Chem 55:127
11. Barbatti M (2011) WIREs Comput Mol Sci 1:620
12. Nelson T, Fernandez-Alberti S, Roitberg AE, Tretiak S (2013) J Chem Phys 138:224111
13. Zheng J, Meana-Pañeda R, Truhlar DG (2014) J Phys Chem Lett 5:2039
14. Nelson T, Fernandez-Alberti S, Chernyak V, Roitberg AE, Tretiak S (2012) J Chem Phys
136:054108
15. Akimov AV, Prezhdo OV (2014) Phys Rev Lett 113:153003
16. Gorshkov VN, Tretiak S, Mozyrsky D (2013) Nat Commun 4:2144
17. Kelly A, van Zon R, Schofield J, Kapral R (2012) J Chem Phys 136:084101
18. Nikitin EE (1968) In: Hartmann H, Heidberg J, Heydtmann H, Kohlmaier GH (eds)
Chemische Elementarprozesse. Springer, Berlin
19. Tully JC, Preston RK (1971) J Chem Phys 55:562
20. Curchod BFE, Rothlisberger U, Tavernelli I (2013) ChemPhysChem 14:1314
21. Akimov AV, Neukirch AJ, Prezhdo OV (2013) Chem Rev 113:4496
22. Fabiano E, Keal TW, Thiel W (2008) Chem Phys 349:334
440 M. Barbatti and R. Crespo-Otero
Barbatti M, Matsika S, Schuurmann M, Yarkony DR, Brozell SR, Beck EV, Blaudeau J-P,
Ruckenbauer M, Sellner B, Plasser F, Szymczak JJ (2012) COLUMBUS, an ab initio
electronic structure program, release 7.0: www.univie.ac.at/columbus
143. Gordon MS, Schmidt MW (2005) In: Dykstra CE, Frenking G, Kim KS, Scuseria GE (eds)
Theory and applications of computational chemistry the first forty years. Elsevier, Amster-
dam, p 1167
144. Crespo-Otero R, Barbatti M, Yu H, Evans NL, Ullrich S (2011) ChemPhysChem 12:3365
145. Tomasello G, Humeniuk A, Mitrić R (2014) J Phys Chem A 118:8437
146. Klaumunzer B, Kroner D, Lischka H, Saalfrank P (2012) Phys Chem Chem Phys 14:8693
147. Frank I, Damianos K (2007) J Chem Phys 126:125105
148. Barbatti M, Lan Z, Crespo-Otero R, Szymczak JJ, Lischka H, Thiel W (2012) J Chem Phys
137:22A503
149. Fischer SA, Habenicht BF, Madrid AB, Duncan WR, Prezhdo OV (2011) J Chem Phys
134:024102
150. Rodrigues GP, Ventura E, do Monte SA, Barbatti M (2014) J Phys Chem A 118:12041
151. Langer H, Doltsinis NL (2004) Phys Chem Chem Phys 6:2742
152. Langer H, Doltsinis NL, Marx D (2005) ChemPhysChem 6:1734
153. Nieber H, Doltsinis NL (2008) Chem Phys 347:405
154. Markwick PRL, Doltsinis NL, Schlitter J (2007) J Chem Phys 126:045104
155. Wohlgemuth M, Bonačić-Koutecký V, Mitrić R (2011) J Chem Phys 135:054105
156. Mercier SR, Boyarkin OV, Kamariotis A, Guglielmi M, Tavernelli I, Cascella M,
Rothlisberger U, Rizzo TR (2006) J Am Chem Soc 128:16938
157. Hu W, Lendvay G, Maiti B, Schatz GC (2008) J Phys Chem A 112:2093
158. Mališ M, Loquais Y, Gloaguen E, Biswal HS, Piuzzi F, Tardivel B, Brenner V, Broquier M,
Jouvet C, Mons M, Došlić N, Ljubić I (2012) J Am Chem Soc 134:20340
159. Guglielmi M, Tavernelli I, Rothlisberger U (2009) Phys Chem Chem Phys 11:4549
160. Habenicht BF, Prezhdo OV (2008) Phys Rev Lett 100:197402
161. Long R, Prezhdo OV (2014) Nano Lett 14:3335
162. Habenicht BF, Prezhdo OV (2009) J Phys Chem C 113:14067
163. Böckmann M, Doltsinis NL, Marx D (2010) Angew Chem 122:3454
164. Bockmann M, Doltsinis NL, Marx D (2010) J Phys Chem A 114:745
165. Neukirch AJ, Shamberger LC, Abad E, Haycock BJ, Wang H, Ortega J, Prezhdo OV, Lewis
JP (2013) J Chem Theory Comput 10:14
166. Akimov AV, Prezhdo OV (2014) J Am Chem Soc 136:1599
167. Crespo-Otero R, Barbatti M (2011) J Chem Phys 134:164305
168. Garavelli M (2006) Theor Chem Acc 116:87
169. Tavernelli I, Curchod BFE, Rothlisberger U (2011) Chem Phys 391:101
170. Jasper AW, Zhu CY, Nangia S, Truhlar DG (2004) Faraday Discuss 127:1
171. Casida M, Natarajan B, Deutsch T (2012) In: Marques MAL, Maitra NT, Nogueira FMS,
Gross EKU, Rubio A (eds) Fundamentals of time-dependent density functional theory.
Springer, Berlin/Heidelberg, p 279
172. Akimov AV, Muckerman JT, Prezhdo OV (2013) J Am Chem Soc 135:8682
173. Casida ME, Huix-Rotllant M (2012) Annu Rev Phys Chem 63:287
174. Barbatti M, Aquino AJA, Lischka H (2006) Mol Phys 104:1053
175. Vazdar M, Eckert-Maksic M, Barbatti M, Lischka H (2009) Mol Phys 107:845
176. Weingart O, Lan Z, Koslowski A, Thiel W (2011) J Phys Chem Lett 2:1506
177. Pederzoli M, Pittner J, Barbatti M, Lischka H (2011) J Phys Chem A 115:11136
178. Carstensen O, Sielk J, Schonborn JB, Granucci G, Hartke B (2010) J Chem Phys 133:124305
179. Ootani Y, Satoh K, Nakayama A, Noro T, Taketsugu T (2009) J Chem Phys 131:194306
180. Gao A-H, Li B, Zhang P-Y, Liu J (2014) Comput Theor Chem 1031:13
181. Sobolewski AL, Domcke WG (2007) J Phys Chem A 111:11725
182. Evans NL, Ullrich S (2010) J Phys Chem A 114:11225
444 M. Barbatti and R. Crespo-Otero
Abstract Conical intersections are perhaps the most significant mechanistic fea-
tures of chemical reactions occurring through excited states. By providing funnels
for efficient non-adiabatic population transfer, conical intersections govern the
branching ratio of products of such reactions, similar to what the transition states
do for ground-state reactivity. In this regard, intersections between the ground and
the lowest excited states play a special role, and the correct description of the
potential energy surfaces in their vicinity is crucial for understanding the mecha-
nism and dynamics of excited-state reactions. The methods of density functional
theory, such as time-dependent density functional theory, are widely used to
describe the excited states of large molecules. However, are these methods suitable
for describing the conical intersections or do they lead to artifacts and, conse-
quently, to erroneous description of reaction dynamics? Here we address the first
part of this question and analyze the ability of several density functional
approaches, including the linear-response time-dependent approach as well as the
spin-flip and ensemble formalisms, to provide the correct description of conical
intersections and the potential energy surfaces in their vicinity. It is demonstrated
that the commonly used linear-response time-dependent theory does not yield a
proper description of these features and that one should instead use alternative
computational approaches.
M. Huix-Rotllant
Institute of Physical and Theoretical Chemistry, Goethe University Frankfurt, Max-von-Laue-
Str. 7, 60438 Frankfurt am Main, Germany
A. Nikiforov and W. Thiel
Max-Planck-Institut für Kohlenforschung, Kaiser-Wilhelm-Platz 1, 45470 Mülheim an der
Ruhr, Germany
M. Filatov (*)
Institut für Physikalische und Theoretische Chemie, Universität Bonn, Beringstr. 4, 53115
Bonn, Germany
e-mail: mike.filatov@gmail.com
446 M. Huix-Rotllant et al.
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
2 Conical Intersections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
2.1 Branching Plane Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
2.2 Double Bond Torsion and Conical Intersections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
3 Computational Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
3.1 REKS Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
3.2 Linear-Response Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
4 Application of DFT Methods to Conical Intersections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
4.1 Ethylene . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
4.2 Penta-2,4-Dieniminium Cation, PSB3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
4.3 Ketene . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
5 Conclusions and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
1 Introduction
The molecular potential energy surface (PES) is perhaps one of the most significant
concepts in molecular physics which enables one to model molecular structure in
terms of specific spatial arrangements of atoms (molecular geometries) [1]. Defining
(meta)stable molecular conformations (local minima) and transient species (saddle
points or transition states) in terms of the molecular electronic energy–molecular
geometry relationship [2], the PES concept is a basis for theories of molecular
structure and reactivity and serves as a starting point for even more approximate
models, such as force-field molecular mechanics [3]. Rooted in the Born–Oppen-
heimer approximation [4], the PES concept is naturally valid in situations where the
coupling between the motion of the nuclei and electrons is negligibly weak, that is,
when the separation between the electronic states is much greater than the charac-
teristic energy of nuclear motion [5]. This assumption breaks down in the vicinity of
points where two (or more) PESs corresponding to distinct electronic states become
degenerate, the surface crossing points [5–8]. Near these points, the non-adiabatic
coupling between the electronic and nuclear degrees of freedom becomes decisive
for the dynamics of transformations occurring in the excited states [9–12] as well as
in the ground state [13, 14] of molecules.
Especially interesting and at the same time especially challenging for their
accurate theoretical description are situations of accidental (that is, not
symmetry-imposed) crossing of PESs of states of the same spatial and spin sym-
metry, the conical intersections (CIs) [6, 15–19]. Such crossings occur in the
subspace of M2 internal molecular degrees of freedom (for a nonlinear molecule,
M ¼ 3 N 6, N – the number of atoms). The degeneracy between the electronic
states is lifted along the two remaining directions, which span the branching plane
Description of Conical Intersections with Density Functional Methods 447
(BP) of the CI [6]. At a CI, the non-adiabatic coupling between the nuclear and
electronic degrees of freedom becomes divergent [6, 9]; hence, the manifold of CI
points, the CI seam, plays a decisive role for the dynamics of the electronic states by
providing funnels for the efficient population transfer [9–12, 20]. As the BP defines
all possible directions of exiting the strong non-adiabatic coupling region [6], it
becomes a very important descriptor of the CI which, very similar to the transition
vector in transition state theory [21, 22], determines the branching of possible
products of the reaction occurring through the CI [23].
The correct theoretical description of CIs requires the use of computational
methods capable of describing various electron correlation effects in the ground
and the excited states with high and unbiased accuracy [10, 12, 24]. Naturally, since
the first realization of the importance of CIs for the dynamics of excited states [9,
10], the computational investigation of CIs has been the realm of multi-reference
(MR) methods of wavefunction theory (WFT), such as the complete active space
self-consistent field (CASSCF) [25–28], the CAS-based second-order perturbation
theory (CASPT2) [29], and the MR configuration interaction (MRCI) [30] methods.
Although capable of delivering highly accurate results, these methods are compu-
tationally demanding and can be used routinely to calculate small molecular
systems limited to just a handful of atoms. To go beyond this limitation and to
match the challenges presented by the rapidly expanding use of ultrafast spectros-
copy in photochemistry and photobiology [31, 32] one needs to employ computa-
tional methods capable of realistically describing large molecules. As the methods
of density functional theory (DFT) [33, 34] offer a reasonably accurate description
of electron correlation effects (which are vital for molecular bonding and structure)
at a typical mean-field computational cost, these methods seem to represent a
viable alternative to MR-WFT approaches.
However, there is a caveat. Originally [33], DFT was formulated for ground
states only, and it was deemed inappropriate to apply it to variational calculation of
individual excited states [35–37]. A way around this problem is offered by the
response formalism implemented in the currently widely used linear-response
(LR) time-dependent DFT (TD-DFT) approach [38, 39]. LR-TD-DFT (or TD-DFT
for brevity) enables one to obtain the excitation energies from the poles of the
ground-state density-density response function [38, 39], thus enabling the use of
ground-state theory for obtaining excited states. Although capable of describing
crossings between excited states correctly, TD-DFT may experience difficulties
with the S0/S1 Cls as the ground electronic state (S0) is the variationally optimized
reference state for the response calculation and is thus described on a different
footing than the excited state (S1), which is a response state. In particular, there is no
coupling between the ground and excited states in most approximate LR-TD-DFT
methods, which results in a wrong dimensionality of the S0/S1 crossing seam, M1
instead of M2 [40–43].
An interesting modification of the original LR-TD-DFT formalism is
implemented in the spin-flip (SF) TD-DFT method which employs a variationally
optimized high-spin open-shell reference state (e.g., triplet state) to access lower
spin states (e.g., singlets) by allowing one-electron transitions with simultaneous
448 M. Huix-Rotllant et al.
1
See the chapter “Surface Hopping Dynamics with DFT Excited States” by M. Barbatti and
R. Crespo-Otero.
Description of Conical Intersections with Density Functional Methods 449
(MECI) and its relative energy, we focus on the BP of the intersection and compare
the BP vectors obtained in the DFT calculations with the corresponding MRCISD-
derived vectors [59]. When benchmarking the DFT methods, we address a set of
CIs of diverse topography occurring in unsaturated molecules during double-bond
torsion; however, the conclusions drawn from these benchmarks should remain
valid for other CIs as well, such as those in cyclic molecules or in nucleobases. The
reason why we focus on relatively small molecules is that it would be extremely
difficult to carry out MRCISD calculations with good-quality basis sets for larger
molecules.
2 Conical Intersections
CIs are manifolds of points at which there is a real crossing (degeneracy) between
two (or more) adiabatic Born–Oppenheimer PESs of electronic states of the same
spatial and spin symmetry. The fundamental conditions for the occurrence of such
intersections have been known since the early days of quantum mechanics [15,
17]. For two adiabatic electronic states Ψm and Ψn, which can be represented by the
solutions of a 2 2 secular problem
H mm H mn Cmm Cmn Em 0 Cmm Cmn
¼ ; ð1Þ
H nm H nn Cnm Cnn 0 En Cnm Cnn
H mm Hnn ¼ Em En ¼ 0; ð2aÞ
H mn ¼ H nm ¼ 0; ð2bÞ
are fulfilled [6, 9, 11, 17–19, 74]. In the above equations, Hmn are the matrix
elements, Hmn ¼ hΦm| H ^ |Φni, of the Hamiltonian H^ in Born–Oppenheimer approx-
imation evaluated with respect to the diabatic states. As was pointed out in the
Introduction, these conditions can be fulfilled in the space of M2 internal molec-
ular coordinates; thus, the CIs may occur in molecules with three or more atoms.
The degeneracy of the two electronic states is lifted in the subspace of the two
coordinates which can be found expanding the Hamiltonian in (1) around the point
of degeneracy through the first order in internal nuclear coordinates Q,
450 M. Huix-Rotllant et al.
H mm Hmn
I2 ∇Q H:δQ
Hnm H nn
0 1
1
B ∇ Q ð H mm H nn Þ ∇ Q H nm C
þ @2 1 A δQ; ð3aÞ
∇Q H mn ∇Q ðH mm H nn Þ
2
g h
¼ I2 s δQ þ δQ; ð3bÞ
h g
span the branching plane of the CI which contains all the nuclear displacements
lifting the degeneracy of the electronic states Ψm and Ψn. In the vicinity of a CI,
the adiabatic PESs depend linearly on the nuclear displacements and have the
topography of a double cone [6, 17]; hence the name.
The definition of the BP vectors of a CI given in (3) and (4) is not unique [74];
the vectors spanning the same plane can be defined via the adiabatic states Ψm and
Ψn, which leads to the following definition:
0
0 g 0
x1 ¼ ; g ¼ ∇Q ðEm En Þ; ð5aÞ
kg0 k
0
h
^ Ψn :
h ¼ Ψm ∇Q H
0 0
x2 ¼ 0 ; ð5bÞ
h
0 0
Although the planes spanned by the (x1, x2) and (x1 , x2 ) vector pairs are the same,
the individual vectors are not necessarily aligned with each other. Furthermore, as
0 0
was first demonstrated by Ruedenberg et al. [6], the vectors (x1 , x2 ) are rigidly
rotated within the plane when traveling along a loop around a CI. This leads to a
0 0
certain arbitrariness in the directions of the (x1 , x2 ) vectors obtained from CI
optimization carried out with finite numerical accuracy. As the crossing point at
which the optimization stops may be infinitesimally close to the CI, but not exactly
0 0
at the CI, the BP vectors (x1 , x2 ) obtained in a series of CI optimizations utilizing
Description of Conical Intersections with Density Functional Methods 451
the same computational method but started from different initial conditions may not
coincide with each other.2
0 0
When comparing BP vectors (x1 , x2 ) obtained using different computational
schemes, the arbitrariness in their orientation within the BP can be bypassed by
aligning the vector pairs such as to maximize the projection of one of the vectors,
0 0
e.g., x1 , obtained using computational method A onto the matching vector (x1 or x1)
obtained using method B [59, 76]. As proposed recently [59], such an alignment can
be carried out by a similarity transformation SRS1 that spans a 2D orthogonal
rotation R and a shear transformation S. The latter transformation is necessary
because the BP vectors are not, in general, orthogonal with respect to one another,
0 0
i.e., the inner products x1 x2 and (x1 · x2) are not zero. The shear transformation
leaves the inner product invariant upon rotation within the plane. The same publica-
tion [59] also introduced a number of useful measures to compare the BPs obtained
using different computational methods. Thus, projections pB xkA , k ¼ 1, 2 of the
0
vectors xk (or xk ) obtained using method A onto the BP obtained by method B and
the projection rAB of a rectangle spanned by the vector pair (xA1 , xA2 ) onto the (xB1 , xB2 )
rectangle were introduced to quantify the similarity (or discrepancy) between the
BP vectors produced by different computational methods [59]. These measures are
used in the following, when comparing the computational methods addressed in
this chapter.
The use of the alignment procedure and the numerical measures described above
for comparing BP vectors become especially useful when addressing computational
schemes for which the BP vectors cannot be obtained using definitions (4) or (5),
0
such as the TD-DFT or SF-TD-DFT methods. For these methods, only the x1 vector
can be determined explicitly by differentiation of the respective electronic energies.
0
The x2 vector is obtained during optimization for a CI using the branching space
0
update method of Maeda et al. [77] as an orthogonal complement to x1 in a plane
0 0
that iteratively converges to the BP of the CI. The (x1 , x2 ) vector pairs so obtained
have been shown to approximate sufficiently accurately the true BPs of the opti-
mized CIs [77]. However, the resulting vectors are strictly orthogonal, a property
not shared by the explicitly calculated vectors. In the context of the REKS method,
the BP vectors are defined by (4), which uses diabatic states and, therefore, yields a
unique orientation of the vectors. This feature becomes especially important when
analyzing the Cls and their BPs in chemical terms.
2
As shown by Yarkony [74, 75], a rotation of the crossing states which orthogonalizes the BP
vectors brings the vectors to a unique orientation, especially when symmetry is present. Applica-
tion of such a prescription, however, modifies the BP vectors, while leaving the BP unchanged.
452 M. Huix-Rotllant et al.
In this chapter, we focus on the Cls arising during double-bond torsion in unsatu-
rated molecules; thus, let us look more closely at this situation. The origin of Cls for
double-bond torsion can be traced back to a crossing between the electronic states
caused by homolytic and heterolytic breaking of the π-component of the double
bond [78]. The homolytic bond breaking results in a diradicaloid electronic con-
figuration, whereas the heterolytic bond breaking leads to an ionic (or zwitterionic)
electronic configuration. According to the sign-change theorem of Longuet-Higgins
[19], a CI should be present inside a loop connecting the conformations that
correspond to the two bond-breaking mechanisms [12, 79–81]; see Fig. 1.
Let us now assume a nuclear movement in the direction of the x1 vector while
keeping the interstate coupling element Hmn at zero. When passing through the CI,
the S0 wavefunction experiences a sudden switch from ionic to diradical (or vice
versa) and the S1 wavefunction does precisely the opposite. It is therefore natural to
associate with the x1 vector a direction that corresponds to the transition between
the uncoupled ionic and diradical states; it is these two electronic states that are
included into the ensemble averaging in the SI-SA-REKS method; see Sect. 3.1.
Conversely, a displacement along the x2 vector while keeping the energy difference
Hmm–Hnn at zero should correspond to increasing (decreasing) the coupling
between the states; hence, this motion should contain the torsion about the double
bond axis. Indeed, when the two fragments connected by the double bond attain an
3 Computational Methods
The REKS method [54–56] employs the ideas behind ensemble DFT to describe the
non-dynamic correlation in the ground and excited states of molecules and to access
the excited states through the application of the variational principle [42, 57,
58]. Perhaps the most important realization in ensemble DFT that was rigorously
proved by Lieb [53] and computationally verified by Schipper et al. [82, 83] and by
Morrison [84] is that the density and the ground-state energy of a strongly
454 M. Huix-Rotllant et al.
X
M X M X
M
^ ΦK λK EK ; 0 λK 1;
λK ΦK H λK ¼ 1 ; ð6Þ
K¼1 K¼1 K¼1
E ω E0
ΔE ¼ E1 E0 ¼ ; ð8Þ
ω
where ϕa and ϕb are the fractionally occupied frontier orbitals, na and nb are the
respective FONs, and the unbarred and barred orbitals are occupied with α- and β-
spin electrons, respectively. When the two active orbitals belong to two different
irreducible representations of the molecular symmetry group (e.g., a
homosymmetric biradical, H2 with stretched bond, double bond torsion in C2H4,
Description of Conical Intersections with Density Functional Methods 455
etc.), the lowest singlet excited state arising from a one-electron excitation in the
space of the two orbitals ϕa and ϕb can be approximated by an open-shell singlet
(OSS) wavefunction [85],
1 1
Φ1 ¼ pffiffiffi . . . ϕa ϕb i þ pffiffiffi . . . ϕb ϕa i:: ð10Þ
2 2
3
See also the chapter “Ensemble DFT approach to excited states of strongly correlated molecular
systems” by M. Filatov.
456 M. Huix-Rotllant et al.
The BP vectors as defined in the SI-SA-REKS method are in accord with the
chemical interpretation presented in Sect. 2 [81]. Indeed, the x1 (or g) vector points
in the direction of the difference of the gradients of the energies of the non-
interacting states S0 and S1, exactly as it should be when moving along the x1
vector in Fig. 1. The same is true for the x2 vector as well. As the SI-SA-REKS
energy formula can be obtained by the adiabatic connection argument, and hence
within the domain of DFT, the BP vectors given in (13) can be regarded as
approximations to the true DFT BP vectors, had these vectors been known from
the exact theory.
Thus, when using the SI-SA-REKS method, the excited states are obtained from
a variational calculation in strict correspondence with the ensemble variational
principle of DFT [49]. The variational calculation of excited states is also
implemented in constricted variational DFT (CV-DFT) [88, 89] reviewed in
another chapter of this book. The use of the SI-SA-REKS method offers markedly
improved accuracy when describing certain types of excitations, particularly exci-
tations of extended π-conjugated molecular systems [90], charge-transfer excita-
tions in donor-acceptor systems [91], and excitations of strongly correlated ground-
state systems (biradicals, molecules with broken bonds, etc.) [90]. With regard to
CIs, the SI-SA-REKS method yields the correct dimensionality (M 2) of the CI
seam and the correct shape of the S0 and S1 PESs in its vicinity [43, 59]. This
follows not only from the numerical tests presented below in this chapter but also
from the general theoretical argument that the coupling between the S0 and S1 states
is consistently taken into account in the SI-SA-REKS method.
where χ LR(r,r0 ,ω) is the LR function, χ s(r,r0 ,ω) is the zeroth-order response func-
tion, and fHxc(r1,r2,ω) ¼ |r1 – r2|1 + fxc(r1,r2,ω) is the Hartree plus the exchange-
correlation (xc) kernel. All LR approaches extract the poles of some form of (15),
differing mainly in the choice of the zeroth-order density used to construct the
non-interacting response function and the xc kernel.
3.2.1 LR-TD-DFT
In LR-TD-DFT, the poles of the LR function (15) are cast in matrix form
AðωÞ BðωÞ X 1 0 X
¼ω ; ð16Þ
B * ð ωÞ A* ðωÞ Y 0 1 Y
In (17), i, j, . . ., a, b,. . . and p, q,. . . are the indices for occupied, virtual, and general
(occupied or virtual) spin-orbitals, and εKS p are the eigenvalues of the KS
Hamiltonian.
Adiabatic LR-TD-DFT based on a KS initial density employs the XC kernel
taken in the adiabatic approximation (i.e., implying locality of the exchange-
correlation (XC) kernel in the time domain),
0 δ2 Exc ½ρ
AA
f xc r; r ¼ : ð18Þ
δρðrÞδρðr0 Þ
The fact that the xc kernel is frequency independent makes it impossible to account
properly for doubly excited states within the adiabatic approximation [92,
93]. These reasons are believed to be the major causes of the failure of LR-
TD-DFT methods to describe correctly the low-lying excited states in
458 M. Huix-Rotllant et al.
Several schemes have been developed that attempt to ameliorate some of the
problems arising in the adiabatic LR-TD-DFT description of S0/S1 conical inter-
sections. Among them, one of the most successful is the spin-flip time-dependent
DFT, SF-TD-DFT. In SF-TD-DFT, electronic configurations arising from the
excitation operators with ΔMs ¼ 1 are coupled, unlike the usual formulation of
TD-DFT in which only spin-preserving excitation operators with ΔMs ¼ 0 are
allowed. Several formulations of SF-TD-DFT appear in the literature [44–
47]. Ziegler and Wang’s formulation of SF-TD-DFT [45] relies on the
non-collinear spin DFT framework, which operates with the KS spinors, rather
than spin-free orbitals. In this formulation, the xc kernel is given by
0 ν α ðrÞ ν β ðrÞ
f SF
xc r; r ¼ xc xc
; ð19Þ
ρα ðr0 Þ ρβ ðr0 Þ
where νσxc and ρσ (r0 ) are the xc potential and the electronic density of spin σ,
respectively.
The configurations arising from the spin-flip αβ and βα excitations are
decoupled. Starting from an open-shell triplet reference in a two-electron
4
See the chapter “Current status and recent developments in linear response time-dependent
density-functional theory” by Mark E. Casida and Miquel Huix-Rotllant.
Description of Conical Intersections with Density Functional Methods 459
Fig. 3 Superimposed geometries at the respective MECI points optimized using MRCISD/6-31
+ G** (standard colors), SSR-BH&HLYP/6-31 + G** (solid blue), SF-TD-DFT-BH&HLYP/6-
31 + G** (solid green), LR-TD-DFT-BH&HLYP/6-31 + G** (solid red). MRCISD calculations
employed the 6-31G** basis set for stilbene and anionic HBI. The MRCISD, SSR, and SF
geometries are taken from [59]
Description of Conical Intersections with Density Functional Methods 461
energies of the lowest singlet π ! π* transition (best estimates given boldface in parentheses). Energies of optimized MECI structures refer to conical
intersections between the S1 and S0 states. The MRCISD, SSR, and SF data are taken from [59]
Geometry MRCIa SSRb SFc TDd Geometry MRCI SSR SF TD
Ethylene Methyliminium
FC pointe (7.80) [99] 8.11 7.55 7.70 7.53 FC point 9.52 8.62 9.22 8.32
tw-pyr MECI 4.79 4.96 4.82 4.73 tw-BLA MECI 3.82 3.83 4.27 4.14
eth MECIf 4.69 3.96 4.52 4.26 meth MECIg 5.63 4.54 5.46 5.22
Styrene Penta-2,4-dieniminium cation, PSB3
FC point (4.88) [100] 5.76 5.10 5.29 5.25 trans-FC pointk 4.47 4.18 4.43 4.97
tw-pyr CI1h 4.68 4.29 4.01 n.a.i 0 l 0.16 0.15 0.13 0.14
ΔEScistrans
tw-pyr CI2j 5.28 4.97 4.83 4.72 cis-FC pointm (4.20) [101] 4.24 4.08 4.28 4.93
tw-BLA MECI 2.38 3.00 3.08 2.82
Butadiene Ketene
trans-FC pointn (6.18) [99] 6.89 5.76 5.94 6.43 FC point (3.84) [102] 4.01 3.93 4.01 4.06
0 l 0.07 0.14 0.11 0.14 MECI 2.43 3.00 2.82 3.27
ΔEScistrans
cis-FC pointm,o 6.68 5.54 5.96 6.05
transoid MECI 4.89 5.58 5.19 5.40
Stilbene Anionic HBI
trans-FC pointk (4.1) [103] (5.35)q 4.18 4.35 4.42 FC point (3.06) [104] (3.88)p 3.00 3.14 3.56
0 l (0.13) 0.18 0.35 0.33 tw-pyr MECIImq (2.87) 2.97 2.84 2.60
ΔEScistrans
cis-FC pointm (4.6) [103] (5.74) 4.54 4.49 4.86 tw-pyr MECIphr (3.20) 3.24 3.20 2.81
tw-pyr MECI (4.50) 4.26 4.20 n.a.i
M. Huix-Rotllant et al.
a
MRCISD/6-31+G** method
b
SI-SA-REKS-BH&HLYP/6-31+G** method
c
SF-TD-DFT-BH&HLYP/6-31+G** method
d
LR-TD-DFT-BH&NLYP/6-31+G** method
e
Franck–Condon point corresponding to the So minimum
f
Ethylidene MECI
g
Methylimine MECI
h
Twisted-pyramidalized MECI1
i
LR-TD-DFT energy is not available because of lack of convergence
j
Twisted-pyramidalized MECI2
k
trans-conformation
l
Energy difference between cis- and trans-conformations in the ground state
m
cis-conformation
n
trans-conformation, transition to the bright Bu state
o
Transition to the B1 state
p
6-31G** basis set
q
Twisted-pyramidalized MECI with twisted imidazole ring
r
Twisted-pyramidalized MECI with twisted phenyl ring
Description of Conical Intersections with Density Functional Methods
463
464
Table 2 Root-mean-square deviations (RMSDs) between the MECI geometries and branching plane projections averaged over all molecules in Fig. 3. The
MRCISD, SSR, and SF data are taken from [59]
RMSD (Å)a BP projectionb
MRCI SSR SF TD MRCI SSR SF TD
MRCI 0.0000 0.0609 0.0698 0.0658 MRCI 1.0000 0.8059 0.9027 0.6307
SSR 0.0609 0.0000 0.0712 0.0420 SSR 0.7845 1.0000 0.8089 0.6459
SF 0.0698 0.0712 0.0000 0.0304 SF 0.8996 0.8064 1.0000 0.6721
TD 0.0789 0.0504 0.0365 0.0000 TD 0.7378 0.7677 0.8111 1.0000
a
RMSDs are given as a matrix with the off-diagonal elements corresponding to the values for methods A and B
b
BP projections are given as a matrix with the off-diagonal elements rAB representing the projection of the BP obtained using method A onto the BP obtained
using method B
M. Huix-Rotllant et al.
Description of Conical Intersections with Density Functional Methods 465
MRCISD RMSDs in Table 2, which vary in the range between 0.061 Å (SSR) and
0.079 Å (TD). An even better agreement is seen between the geometries obtained
by the different DFT methods, especially for the SF and TD methods (0.030 Å). The
BP projections in Table 2 indicate that all methods produce nearly the same BP
vectors at the MECI geometries. The very good agreement between the DFT
methods and MRCISD seems to indicate that these much simpler methods can be
confidently used in lieu of MRCISD for locating CIs and analyzing the S0 and S1
PESs in their vicinity. However, there is a caveat: despite the seemingly good
agreement with the high-level MR-WFT approach, not all the DFT methods
produce the correct dimensionality of the CI seam. In the following, we focus on
three molecules and their respective CIs, ethylene and its tw-pyr MECI, cationic
PSB3 and its tw-BLA MECI, as well as ketene molecule and its n/π* MECI, as
representatives of the specific types of CIs discussed above.
4.1 Ethylene
On the S0/S1 CI seam in ethylene there are two distinct MECI points that correspond
to a tw-pyr MECI (the major photorelaxation funnel) and to a MECI point with
partial transfer of a hydrogen atom; see the tw-pyr and ethylidene MECIs of C2H4
in Fig. 3 [105–108]. The tw-pyr MECI originates from a crossing between the PESs
for homolytic and heterolytic π-bond breaking along the double bond torsion mode.
As the heterolytic π-bond breaking is strongly disfavored in ethylene, the tw-pyr
MECI features a strong pyramidalization of one of the carbon atoms to stabilize the
ionic electronic configuration.
All the DFT methods employed here predict the vertical excitation energy at the
FC point in a good agreement with the best theoretical estimate of 7.8 eV. The energy
level of the tw-pyr MECI is much lower than the FC point. The DFT methods yield
tw-pyr MECI energies varying between 4.96 eV (SSR) and 4.73 eV (TD), in good
agreement with the MRCISD value 4.79 eV; see Table 1. The BP vectors predicted
by the DFT methods are in very good agreement with the MRCISD BP vectors, as
can be seen from a visual comparison of the vectors in Fig. 4 and from the BP
projections (see Sect. 2) which lie in the range 0.99–0.97 in each case.
From these results it seems that all the DFT methods are capable of describing
the tw-pyr MECI in ethylene correctly. However, a closer look at the S0 and S1 PESs
near the MECI point reveals that the surface crossing predicted by adiabatic LR-
TD-DFT is not actually a conical intersection but is rather a linear crossing. As seen
in Fig. 5, the S1 PES produced by adiabatic LR-TD-DFT falls below the S0 PES
when moving along the x1 direction.5 Furthermore, the shape of the S0 and S1 PESs
5
The conventional KS DFT/LR-TD-DFT calculations experienced severe convergence problems
in the vicinity of the MECI point, which are reflected in the shape of the S1 PES in the upper right
panel of Fig. 5.
466 M. Huix-Rotllant et al.
Fig. 4 BP vectors of
twisted-pyramidalized
MECI of ethylene
calculated using MRCISD,
SSR-BH&HLYP, SF-TD-
BH&HLYP, and LR-TD-
BH&HLYP methods
Fig. 5 S0 (blue) and S1 (red) PES profiles near the tw-pyr MECI of ethylene. Upper panels –
profiles along the x1 BP vector, lower panels – along the x2 BP vector. Leftmost panels – profiles
obtained using the SSR-BH&HLYP/6-31 + G** method, middle panels – obtained by the SF-TD-
BH&HLYP/6-31 + G** method, rightmost panels – using the TD-BH&HLYP/6-31 + G** method.
For each computational method, the MECI energy level is taken as the origin of the energy scale
(in kcal/mol). The variable x corresponds to displacement (in Å) from the MECI geometry in the
direction of the x1 or x2 vector, respectively
Fig. 6 The S0–S1 energy difference (in kcal/mol) along a loop around the MECI point of ethylene.
The solid line corresponds to the SSR-BH&HLYP/6-31 + G** method, the dashed line to the
SF-TD-BH&HLYP/6-31 + G** method, and the dotted line to the TD-BH&HLYP/6-31 + G**
method. The loop radius is 0.002 Å. For each method, the angle θ (in deg) is taken from the
direction of the x1 vector
468 M. Huix-Rotllant et al.
θ ¼ 180 ; note that the off-diagonal element Hmn in (1) is supposed to zero out when
moving along the x1 direction. The SSR and SF methods predict the maximum S0–
S1 splitting in the x1 direction, whereas adiabatic LR-TD-DFT yields the maximum
splitting approximately in the x2 direction (θ 0 ).
As found by Gozem et al. [43, 98] for the example of PSB3, the multi-state
CASPT2 and MRCISD methods yield energy difference curves around a MECI
which are similar in shape to those predicted by SSR and SF and dissimilar to
TD-DFT. Therefore, in spite of its seemingly good predictions for the geometry and
energy of the tw-pyr MECI of ethylene, TD-DFT fails to give the correct conical
topography of the crossing and yields the wrong shape of the two PESs in the
vicinity of the crossing point. As discussed by Barbatti and Crespo-Otero (see
Footnote 1) this has severe consequences for the description of the dynamics near
the crossing points, rendering TD-DFT practically useless for modeling
non-adiabatic population transfer through PES crossings.
Fig. 8 S0 (blue) and S1 (red) PES profiles near the tw-BLA MECI of PSB3. See caption of Fig. 5
for the legend
470 M. Huix-Rotllant et al.
surface crossing has a linear rather than conical character. Furthermore, the adia-
batic LR-TD-DFT method experiences severe convergence problems in the vicinity
of the crossing point which prevented us from obtaining a reasonable S0–S1 energy
difference plot along a loop around the MECI. Nevertheless, the curves in Fig. 8
corroborate the fact that the adiabatic LR-TD-DFT method is incapable of yielding
the correct topography of the S0/S1 crossing point and that the other two DFT
methods, SSR and SF, are thus preferred for studying the non-adiabatic dynamics of
the excited state of PSB3 and related molecules.
4.3 Ketene
Ketene displays an S0/S1 CI along the C–C–O bending mode which, as first
analyzed by Yarkony [97], originates from a crossing between the 1A’ and 1A”
electronic states. This crossing is linear under the Cs symmetry constraint; however,
it becomes conical when the symmetry restriction is lifted, as there emerges another
direction which lifts the degeneracy between these states. Hence, the BP vectors of
this CI correspond to the a0 -symmetric C–C–O bending mode (x1) and the out-of-
plane H–C–C–O torsional mode (x2).
As shown in Fig. 9, the DFT methods yield BP vectors in very good agreement
with the MRCISD vectors. However, all the DFT methods predict the MECI point
Fig. 9 BP vectors of
twisted MECI of ketene
calculated using MRCISD,
SSR-BH&HLYP, SF-TD-
BH&HLYP, and LR-TD-
BH&HLYP methods
Description of Conical Intersections with Density Functional Methods 471
Fig. 10 S0 (blue) and S1 (red) PES profiles near the MECI of ketene. See caption of Fig. 5 for the
legend
at a somewhat higher, by ca. 0.4–0.8 eV, energy level than MRCISD. This differ-
ence may likely be caused by a somewhat stiffer C–C bond predicted by the DFT
calculations. This can be asserted from the x1 vectors in Fig. 9, which describes a
C–C bond stretching combined with the C–C–O bending motion: there is a notably
greater degree of bond stretching in the MRCISD vector. As can be seen from
Fig. 10, which shows the S0 and S1 PES profiles along the BP vectors, this motion
noticeably destabilizes the S0 (1A0 ) state in which the C–C π-bonding orbital is
doubly occupied. It is therefore plausible that MRCISD somewhat underestimates
the strength of the C–C bond in ketene because of the rather small basis set and the
limited active space employed in the calculations [59].
Although the DFT methods agree with one another on the strength of the C–C
bond in ketene, the TD-DFT method differs from the other two methods as to the
character of the S0/S1 crossing; whereas SSR and SF predict it to be conical, TD
yields a linear crossing. This is confirmed by the plots of the S0–S1 energy
difference along a loop around the minimal energy crossing point in Fig. 11. In
this plot, a loop with a greater radius, 0.02 Å, was chosen to bypass certain SCF
convergence problems in the DFT calculations with a smaller radius (0.002 Å was
used in Fig. 6). The adiabatic LR-TD-DFT energy difference in Fig. 11 (dotted
curve) again crosses the zero line twice which indicates a linear crossing. The same
linear character of the S0/S1 PES crossing can be seen in the rightmost plots in
Fig. 10, which show the PES profiles along the BP vectors. Similar to the cases of
ethylene and PSB3, there is no coupling between the crossing states in the adiabatic
LR-TD-DFT curves. By contrast, the SSR and SF methods correctly predict
(nearly) the symmetric splitting of the states which indicates a proper description
of the coupling and therefore a correct conical character of the PES crossing point.
472 M. Huix-Rotllant et al.
References
477
478 Index
Constricted variational density functional Equivalent core hole (ECH), (Z+1), 293
theory (CV-DFT), 63 Ethylene, 4, 46, 70, 116, 429, 461, 465, 471
Contact exciton, 201 electronic excitation energies, 116
Core excitations, 273 photochemical isomerization, 4
density functionals, 322 torsion, 38
Core hole, 314 Exact exchange (EXX) theory, 26
Crystal field multiplet (CFM), 321 Exchange–correlation (XC) energy, 237
Cyanine, 359 functionals, 347
Cyclopentadiene, 70, 116 kernel, 39, 49, 98, 185
Cyclopropene, 70, 116 Excitation energy, 11
Cytosine, 435 transfer (EET), 304
Excited electronic states, 1, 97
Excited-state intramolecular proton transfer
D (ESIPT), 364
Davydov excitons, 187 Excited states, 415, 445
Decoherence-induced surface hopping (DISH), dynamics, 377
421 REKS, 116
Density-functional approximations (DFAs), 6 Excitons, 185, 213, 276, 329
Density functional resonance theory (DFRT), Extended Koopmans’ theorem (EKT), 153
237 Extended X-ray absorption fine structure
Density-functional theory (DFT), 1, 61, 273, (EXAFS), 279
415, 447 Exterior complex scaling, 253
constricted variational, 61
ensemble, 97, 99, 445
Kohn–Sham (KS-DFT), 62 F
time-dependent (TDDFT), 1, 7, 61, 185, Fe(bpy)32+, 401
219, 347, 377 Fluoroborates, 363
Density functional theory, constricted Formaldehyde, 3
variational (CV-DFT), 63 Frank–Condon (FC) region, 335, 348, 417, 461
Density matrix functional theory, 125, 131 Frenkel–Dirac action, 9
Density–potential mapping, 221 Frenkel excitons, 187
Dielectric functions, 185 Fundamental gap, 154
Diffusion Monte–Carlo (DMC), 119 Furan, 116, 314, 318, 435
Diphenydibenzofulvene, 435
Double core hole (DCH) states, 314
Double excitation, 273 G
Double-quantum coherence, 273, 280, 328 Gas phase, 349
Douglas–Kroll (DK), 381 Generalized gradient approximation (GGA),
Dye sensitized solar cells (DSSC), 366 86, 322, 389
Dyes, inorganic, 119, 357, 362, 468 Global-flux surface hopping (GFSH), 421
organic, 63 Global hybrid (GH), 7
Dyson’s equation, 22 Graphene, 435
Green’s function, 15, 18, 22, 52, 190, 244, 320
Guanine, 435
E
Electron correlation, 125
nondynamic, 97 H
Electron–hole operators, expansion of Hartree–Fock (HF), 2, 129, 237, 242, 295
polarizability, 324 exchange correlation, 35, 42, 65, 86, 148,
Electron photoemission, 254 160, 197
Electronic excited states, 1, 97 nonlocal exchange, 191
Ensemble density functional theory, 97, 445 time-dependent, 423
Equation of motion (EOM), 24, 301, 353 Heptacene, 117
Index 479