\title{Weight-dependent local density-functional approximations for ensembles}
\author{Pierre-Fran\c{c}ois Loos}
\email{loos@irsamc.ups-tlse.fr}
\affiliation{\LCPQ}
\author{Emmanuel Fromager}
\email{fromagere@unistra.fr}
\affiliation{\LCQ}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{abstract}
We report a first generation of local, weight-dependent correlation density-functional approximations (DFAs) that incorporate information about both ground and excited states in the context of density-functional theory for ensembles (eDFT).
These density-functional approximations for ensembles (eDFAs) are specially designed for the computation of single and double excitations within eDFT, and can be seen as a natural extension of the ubiquitous local-density approximation for ensemble (eLDA).
The resulting eDFAs, based on both finite and infinite uniform electron gas models, automatically incorporate the infamous derivative discontinuity contributions to the excitation energies through their explicit ensemble weight dependence.
Their accuracy is illustrated by computing single and double excitations in one-dimensional many-electron systems in the weak, intermediate and strong correlation regimes.
Over the last two decades, density-functional theory (DFT) \cite{Hohenberg_1964, Kohn_1965} has become the method of choice for modeling the electronic structure of large molecular systems and materials. \cite{ParrBook}
The main reason is that, within DFT, the quantum contributions to the electronic repulsion energy --- the so-called exchange-correlation (xc) energy --- is rewritten as a functional of the electron density $\n{}{}(\br)$, the latter being a much simpler quantity than the many-electron wave function.
Despite its success, the standard Kohn-Sham (KS) formulation of DFT \cite{Kohn_1965} (KS-DFT) suffers, in practice, from various deficiencies. \cite{Woodcock_2002, Tozer_2003, Tozer_1999, Dreuw_2003, Sobolewski_2003, Dreuw_2004, Tozer_1998, Tozer_2000, Casida_1998, Casida_2000, Tapavicza_2008, Levine_2006}
The simplest and most widespread approximation in state-of-the-art electronic structure programs where TDDFT is implemented consists in neglecting memory effects. \cite{Casida}
As a result, double electronic excitations are completely absent from the TDDFT spectrum, thus reducing further the applicability of TDDFT. \cite{Maitra_2004,Cave_2004,Mazur_2009,Mazur_2011,Huix-Rotllant_2011,Elliott_2011,Maitra_2012,Sundstrom_2014}
When affordable (i.e., for relatively small molecules), time-independent state-averaged wave function methods \cite{Roos,Andersson_1990,Angeli_2001a,Angeli_2001b,Angeli_2002} can be employed to fix the various issues mentioned above.
The basic idea is to describe a finite ensemble of states (ground and excited) altogether, i.e., with the same set of orbitals.
Ensemble DFT (eDFT) was proposed at the end of the 80's by Gross, Oliveira and Kohn (GOK), \cite{Gross_1988, Gross_1988a, Oliveira_1988} and is a generalization of Theophilou's variational principle for equi-ensembles. \cite{Theophilou_1979}
Despite its formal beauty and the fact that eDFT can in principle tackle near-degenerate situations and multiple excitations, it has not been given much attention until recently. \cite{Franck_2014,Borgoo_2015,Kazaryan_2008,Gould_2013,Gould_2014,Filatov_2015,Filatov_2015b,Filatov_2015c,Gould_2017,Deur_2017,Gould_2018,Gould_2019,Sagredo_2018,Ayers_2018,Deur_2018a,Deur_2018b,Kraisler_2013, Kraisler_2014,Alam_2016,Alam_2017,Nagy_1998,Nagy_2001,Nagy_2005,Pastorczak_2013,Pastorczak_2014,Pribram-Jones_2014,Yang_2013a,Yang_2014,Yang_2017,Senjean_2015,Senjean_2016,Senjean_2018,Smith_2016}
Recent works on this topic are still fundamental and exploratory, as they rely either on simple (but nontrivial) models like the Hubbard dimer \cite{Carrascal_2015,Deur_2017,Deur_2018a,Deur_2018b,Senjean_2015,Senjean_2016,Senjean_2018,Sagredo_2018} or on atoms for which highly accurate or exact-exchange-only calculations have been performed. \cite{Yang_2014,Yang_2017}
A first step towards this goal is presented in this article with the ambition to turn, in the near future, eDFT into a practical computational method for modeling excited states in molecules and extended systems.
In the following, the present methodology is illustrated on \emph{strict} one-dimensional (1D), spin-polarized electronic systems. \cite{Loos_2012, Loos_2013a, Loos_2014a, Loos_2014b}
In other words, the Coulomb interaction used in this work describes particles which are \emph{strictly} restricted to move within a 1D sub-space of three-dimensional space.
Despite their simplicity, 1D models are scrutinized as paradigms for quasi-1D materials \cite{Schulz_1993, Fogler_2005a} such as carbon nanotubes \cite{Bockrath_1999, Ishii_2003, Deshpande_2008} or nanowires. \cite{Meyer_2009, Deshpande_2010}
%Early models of 1D atoms using this interaction have been used to study the effects of external fields upon Rydberg atoms \cite{Burnett_1993, Mayle_2007} and the dynamics of surface-state electrons in liquid helium. \cite{Nieto_2000, Patil_2001}
This description of 1D systems also has interesting connections with the exotic chemistry of ultra-high magnetic fields (such as those in white dwarf stars), where the electronic cloud is dramatically compressed perpendicular to the magnetic field. \cite{Schmelcher_1990, Lange_2012, Schmelcher_2012}
In these extreme conditions, where magnetic effects compete with Coulombic forces, entirely new bonding paradigms emerge. \cite{Schmelcher_1990, Schmelcher_1997, Tellgren_2008, Tellgren_2009, Lange_2012, Schmelcher_2012, Boblest_2014, Stopkowicz_2015}
where the $K$th energy level $E^{(K)}$ [$K=0$ refers to the ground state] is the eigenvalue of the electronic Hamiltonian $\hat{H}=\hat{h}
%\sum^N_{i=1}\hat{h}(i)
+\hat{W}_{\rm ee}$, where $\hat{h}\equiv\sum^N_{i=1}\left(-\frac{1}{2}\nabla_{\br_i}^2+v_{\rm ne}(\br_i)\right)$ is the one-electron operator describing kinetic and nuclear attraction energies, and $\hat{W}_{\rm ee}$ is the electron repulsion operator.
The (positive) ensemble weights $w_K$ decrease with increasing index $K$. They are normalized, i.e.
\beq
w_0=1-\sum_{K>0}w_K,
\eeq
so that only the weights $\bw\equiv\left(w_1,w_2,\ldots w_K,\ldots\right)$ assigned to the excited states can vary independently.
For simplicity we will assume in the following that the energies are not degenerate. Note that the theory can be extended to multiplets simply by assigning the same ensemble weight to all degenerate states~\cite{}. In GOK-DFT, the ensemble energy is determined variationally as follows:
\beq\label{eq:var_ener_gokdft}
E^{{\bw}}=
\underset{\opGamma{{\bw}}}{\rm min}\left\{
{\rm
Tr}\left[\opGamma{{\bw}}\hat{h}\right]
+
{E}^{{\bw}}_{\rm
Hx}\left[n_{\opGamma{\bw}}\right]
+
{E}^{{\bw}}_{\rm
c}\left[n_{\opGamma{\bw}}\right]
\right\},
\eeq
where ${\rm
Tr}$ denotes the trace and the trial ensemble density matrix operator reads
The determinants (or configuration state functions) $\Phi^{(K)}$ are all constructed from the same set of (ensemble Kohn--Sham) orbitals that is optimized variationally and the trial ensemble density is simply the weighted sum of the individual densities:
As readily seen from Eq.~(\ref{eq:var_ener_gokdft}), both Hartree-exchange and
correlation energies are described with density functionals that are {\it weight-dependent}. We focus here on the (exact) Hx part which is defined as follows:
where the KS wavefunctions fulfill the ensemble density constraint
\beq
\sum_{K\geq 0}w_Kn_{\Phi^{(K)}[n]}(\br)=n(\br).
\eeq
The (approximate) description of the correlation part is discussed in Sec.~\ref{sec:eDFA}.\\
In practice, one is not much interested in ensemble energies but rather in excitation energies or individual energy levels (for geometry optimizations, for example). The latter can be extracted exactly as follows~\cite{}:
where, according to the {\it variational} ensemble energy expression of Eq.~(\ref{eq:var_ener_gokdft}), the derivative in $w_K$ can be evaluated from the minimizing KS wavefunctions $\Phi^{(K)}=\Phi^{(K),\bw}$ as follows:
Since, for given ensemble weight $\bw$ and $\bxi$ values, the ensemble densities $n^{\bxi,\bxi}$ and $n^{\bw,\bxi}$ are generated from the {\it same} KS potential (which is unique up to a constant), it comes
from the exact expression in Eq.~(\ref{eq:exact_ens_Hx}) that
The coefficients $\cMO{\mu p}{\bw}$ used to construct the density matrix $\bGamma{\bw}$ in Eq.~\eqref{eq:Gamma} are obtained by diagonalizing the following Fock matrix
In Eq.~\eqref{eq:F}, $\hHc$ is the core Hamiltonian (including kinetic and electron-nucleus attraction terms), $\eG{\mu\nu\la\si}=(\mu\nu|\la\si)-(\mu\si|\la\nu)$,
$\bE{Hxc}{\bw}[\n{}{}(\br)]=\n{}{}(\br)\be{Hxc}{\bw}[\n{}{}(\br)]$ and $\be{Hxc}{\bw}[\n{}{}(\br)]$ is the weight-dependent Hartree-exchange-correlation functional to be built in the present study.
The self-consistent process described above is carried on until $\max\abs{\bF{\bw}\,\bGamma{\bw}\,\bS-\bS\,\bGamma{\bw}\,\bF{\bw}} < \tau$, where $\tau$ is a user-defined threshold and $\eS{\mu\nu}=\braket{\AO{\mu}}{\AO{\nu}}$ are elements of the overlap matrix $\bS$.
Note that because the second term of the RHS of Eq.~\eqref{eq:Ew} is quadratic in $\bGamma{\bw}$, the weight-dependent energy contains the so-called ghost interaction which makes the ensemble energy non linear. \cite{Gidopoulos_2002, Pastorczak_2014, Alam_2016, Alam_2017, Gould_2017}
Following Deur and Fromager, \cite{Deur_2018b} it is possible to extract individual energies, $\E{}{(I)}$, from the ensemble energy [see Eq.~\eqref{eq:Ew}] as follows:
The only remaining piece of information to define at this stage is the weight-dependent Hartree-exchange-correlation functional $\be{Hxc}{\bw}(\n{}{})$.
where $\be{Hx}{\bw}(\n{}{})$ is a weight-dependent Hartree-exchange functional designed to correct the ghost interaction \cite{Gidopoulos_2002, Pastorczak_2014, Alam_2016, Alam_2017, Gould_2017} [see Subsec.~\ref{sec:GIC}] and $\be{c}{\bw}(\n{}{})$ is a weight-dependent correlation functional [see Subsec.~\ref{sec:Ec}].
The construction of these two functionals is described below.
Note that, because we consider strict 1D systems, one cannot decompose further the Hartree-exchange contribution as each component diverges independently but their sum is finite. \cite{Astrakharchik_2011, Lee_2011a, Loos_2012, Loos_2013, Loos_2013a}
Most of the standard local and semi-local DFAs rely on the infinite uniform electron gas (UEG) model (also known as jellium). \cite{ParrBook, Loos_2016}
One major drawback of the jellium paradigm, when it comes to develop eDFAs, is that the ground and excited states cannot be easily identified like in a molecule. \cite{Gill_2012, Loos_2012, Loos_2014a, Loos_2014b, Agboola_2015, Loos_2017a}
Moreover, because the infinite UEG model is a metal, it is gapless, which means that both the fundamental and optical gaps are zero.
From this point of view, using finite UEGs \cite{Loos_2011b, Gill_2012} (which have, like an atom, discrete energy levels) to construct eDFAs can be seen as more relevant. \cite{Loos_2014a, Loos_2014b, Loos_2017a}
As a finite uniform electron gas, we consider the ringium model in which electrons move on a perfect ring (i.e., a circle). \cite{Loos_2012, Loos_2013a, Loos_2014b}
The most appealing feature of ringium (regarding the development of functionals in the context of eDFT) is the fact that both ground- and excited-state densities are uniform.
As a result, the ensemble density will remain constant (and uniform) as the ensemble weights vary.
This is a necessary condition for being able to model derivative discontinuities.
The present weight-dependent eDFA is specifically designed for the calculation of excitation energies within eDFT.
As mentioned previously, we consider a three-state ensemble including the ground state ($I=0$), the first singly-excited state ($I=1$), and the first doubly-excited state ($I=2$) of the (spin-polarized) two-electron ringium system.
All these states have the same (uniform) density $\n{}{}=2/(2\pi R)$ where $R$ is the radius of the ring where the electrons are confined.
We refer the interested reader to Refs.~\onlinecite{Loos_2012, Loos_2013a, Loos_2014b} for more details about this paradigm.
Based on highly-accurate calculations (see {\SI} for additional details), one can write down, for each state, an accurate analytical expression of the reduced (i.e., per electron) correlation energy \cite{Loos_2013a, Loos_2014a} via the following Pad\'e approximant
In order to make the two-electron-based eDFA defined in Eq.~\eqref{eq:ecw} more universal and to ``center'' it on the jellium reference (as commonly done in DFT), we propose to \emph{shift} it as follows:
specifically designed for 1D systems in Ref.~\onlinecite{Loos_2013} as been used, where $F(a,b,c,x)$ is the Gauss hypergeometric function, \cite{NISTbook} and
This procedure can be theoretically justified by the generalized adiabatic connection formalism for ensembles (GACE) which was originally derived by Franck and Fromager. \cite{Franck_2014}
Within this in-principle-exact formalism, the (weight-dependent) correlation energy of the ensemble is constructed from the (weight-independent) ground-state functional (such as the LDA), yielding Eq.~\eqref{eq:eLDA}.
This is a crucial point as we intend to incorporate into standard functionals (which are ``universal'' in the sense that they do not depend on the number of electrons) information about excited states that will be extracted from finite systems (whose properties may depend on the number of electrons).
Having defined the eLDA functional in the previous section [see Eq.~\eqref{eq:eLDA}], we now turn to its validation.
Our testing playground for the validation of the eLDA functional is the ubiquitous ``electrons in a box'' model where $\Nel$ electrons are confined in a 1D box of length $L$, a family of systems that we call $\Nel$-boxium.
In particular, we investigate systems where $L$ ranges from $\pi/8$ to $8\pi$ and $2\le\Nel\le7$.
\alert{Comment on the quality of these density: density- and functional-driven errors?}
These inhomogeneous systems have non-trivial electronic structure properties which can be tuned by varying the box length.
We use as basis functions the (orthonormal) orbitals of the one-electron system, i.e.
\begin{equation}
\AO{\mu}(x) =
\begin{cases}
\sqrt{2/L}\cos(\mu\pi x/L), &\mu\text{ is odd,}
\\
\sqrt{2/L}\sin(\mu\pi x/L), &\mu\text{ is even,}
\end{cases}
\end{equation}
with $\mu=1,\ldots,\Nbas$ and $\Nbas=30$ for all calculations.
For the self-consistent calculations (such as HF, KS or eKS), the convergence threshold has been set to $\tau=10^{-7}$.
For KS and eKS calculations, a Gauss-Legendre quadrature is employed to compute numerical integrals.
In order to test the present eLDA functional we have performed various sets of calculations.
To get reference excitation energies for both the single and double excitations, we have performed full configuration interaction (FCI) calculations with the Knowles-Handy FCI program described in Ref.~\onlinecite{Knowles_1989}.
For the single excitations, we have also performed time-dependent HF (TDHF), configuration interaction singles (CIS) and TDLDA calculations. \cite{Dreuw_2005}
For TDLDA, the validity of the Tamm-Dancoff approximation (TDA) has been also tested.
Concerning the eKS calculations, two sets of weight have been tested: the zero-weight limit where $\bw=(0,0)$ and the equi-ensemble (or state-averaged) limit where $\bw=(1/3,1/3)$.
In Fig.~\ref{fig:EvsL}, we report the error (in \%) in excitation energies (compared to FCI) for various methods and box sizes in the case of 5-boxium (i.e., $\Nel=5$).
Similar graphs are obtained for the other $\Nel$ values and they can be found --- alongside the numerical data associated with each method --- in the {\SI}.
For the single excitation, TDHF is extremely accurate over the whole range of $L$ values, while CIS is slightly less accurate and starts to overestimate the excitation energy by a few percent at $L=8\pi$.
TDLDA yields larger errors at large $L$ by underestimating the excitation energies.
TDA-TDLDA slightly corrects this trend thanks to error compensation.
Concerning the eLDA functional, our results clearly evidences that the equi-weights [i.e., $\bw=(1/3,1/3)$] excitation energies are much more accurate than the ones obtained in the zero-weight limit [i.e., $\bw=(0,0)$].
This is especially true for the single excitation which is significantly improved by using state-averaged weights.
The effect on the double excitation is less pronounced.
Overall, one clearly sees that, with state-averaged weights, the eLDA functional yields accurate excitation energies for both single and double excitations.
This conclusion is verified for smaller and larger number of electrons (see {\SI}).
\alert{Shall I test the one-electron system for self-interaction?}
Figure \ref{fig:EvsN} reports the error (in \%) in excitation energies, for the same methods, as a function of $\Nel$ and fixed $L$ (in this case $L=\pi$).
The graphs associated with other $L$ values are reported as {\SI}.
Again, the graph for $L=\pi$ is quite typical and we draw similar conclusions as in the previous paragraph: irrespectively of the number of electrons, the eLDA functional with state-averaged weights is able to accurately model single and double excitations.
As a rule of thumb, we see that eLDA single excitations are of the same quality as the ones obtained in the linear response formalism (such as TDHF or TDLDA), while double excitations only deviates from the FCI values by a few tenth of percent for $L=\pi$, an error of the same order as CIS or TDA-TDLDA.
Even for larger boxes, the discrepancy between FCI and eLDA for double excitations is only a few percent.
\alert{Need further discussion on DD and LZ shift. Linearity of energy wrt weights?}
\alert{For small $L$, the single and double excitations are ``pure''. In other words, the excitation is dominated by a single reference Slater determinant.
However, when the box gets larger, there is a strong mixing between different degree of excitations.
In particular, the single and double excitations strongly mix.
This is clearly evidenced if one looks at the weights of the different configurations in the FCI wave function.
In one hand, if one does construct a eDFA with a single state (either single or double), one clearly sees that the results quickly deteriorates when the box gets larger.
On the other hand, building a functional which does mix singles and doubles corrects this by allowing configuration mixing.}
\alert{It might be useful to add eHF results where one switch off the correlation part.
For both zero weight and state-averaged weights?
It would highlight the contribution of the derivative discontinuity.}
This eDFA delivers accurate excitation energies for both single and double excitations.
Generalization to a larger number of states is straightforward and will be investigated in future work.
Using similar ideas, a three-dimensional version \cite{Loos_2009,Loos_2009c,Loos_2010,Loos_2010d,Loos_2017a} of the present eDFA is currently under development to model excited states in molecules and solids.
Similar to the present excited-state methodology for ensembles, one can easily design a local eDFA for the calculations of the ionization potential, electron affinity, and fundamental gap.
This can be done by constructing DFAs for the one- and three-electron ground state systems, and combining them with the two-electron DFA in complete analogy with Eqs.~\eqref{eq:ec} and \eqref{eq:ecw}.
