What can lattice DFT teach us about real-space DFT?

Nahual Sobrino,1 David Jacob,1, 2 and Stefan Kurth∗, 1, 2, 3

1) Nano-Bio Spectroscopy Group and European Theoretical Spectroscopy Facility (ETSF),
Departamento de Polímeros y Materiales Avanzados: Física, Química y Tecnología, Universidad del País Vasco UPV/EHU,
Avenida Tolosa 72, E-20018 San Sebastián, Spain
2) IKERBASQUE, Basque Foundation for Science, Plaza Euskadi 5, E-48009 Bilbao, Spain
3) Donostia International Physics Center (DIPC), Paseo Manuel de Lardizabal 4, E-20018 San Sebastián,
Spain
(* Corresponding author: [email protected])
(Dated: 23 October 2023)
arXiv:2308.01185v2 [cond-mat.str-el] 20 Oct 2023

In this paper we establish a connection between density functional theory (DFT) for lattice models and common real-
space DFT. We consider the lattice DFT description of a two-level model subject to generic interactions in Mermin’s
DFT formulation in the grand canonical ensemble at finite temperature. The case of only density-density and Hund’s
rule interaction studied in earlier work is shown to be equivalent to an exact-exchange description of DFT in the real-
space picture. In addition, we also include the so-called pair-hopping interaction which can be treated analytically and,
crucially, leads to non-integer occupations of the Kohn-Sham levels even in the limit of zero temperature. Treating the
hydrogen molecule in a minimal basis is shown to be equivalent to our two-level lattice DFT model. By means of the
fractional occupations of the KS orbitals (which, in this case, are identical to the many-body ones) we reproduce the
results of full configuration interaction, even in the dissociation limit and without breaking the spin symmetry. Beyond
the minimal basis, we embed our HOMO-LUMO model into a standard DFT calculation and, again, obtain results in
overall good agreement with exact ones without the need of breaking the spin symmetry.

I. INTRODUCTION late a DFT for lattice models, such as the Hubbard and Ander-
son impurity models.22–26 This allows to incorporate electron-
electron interactions in a controlled way, and study their effect
Density Functional Theory (DFT) is an in principle exact
on the Hartree-exchange-correlation (Hxc) functional. Lattice
framework for solving the quantum many-body problem of
DFT (LDFT) studies have for example revealed the ubiquitous
interacting electrons.1 The Kohn-Sham (KS) formulation of
presence of steps in the exact functionals.27 It turns out that
DFT allows to cast the problem into the particularly simple
these steps are crucial for DFT to capture important aspects
form of a system of effectively non-interacting electrons mov-
of strongly correlated systems by DFT, such as the Kondo
ing in a mean-field potential.2 Owing to its computational ef-
plateau in the zero-bias conductance28–30 or the opening of
ficiency, conceptual simplicity and in many cases high accu-
a Mott gap.24,31 Furthermore, an extension of DFT which in-
racy, KS-DFT has become one of the cornerstones of elec-
corporates the current through an interacting system in addi-
tronic structure theory and is now the standard tool for the
tion to its density, called steady-state DFT or i-DFT, is ca-
description of electronic matter in computational condensed
pable of capturing strongly correlated phenonomena both in
matter physics, material science and chemistry.3–5
the conductance and the many-body spectral function, such as
One of the challenges for KS-DFT is the accurate de- Coulomb blockade, Kondo effect and the Mott metal-insulator
scription of so-called strongly correlated systems, for which transition.32–36 Again the proper description of these phenom-
the physics is dominated by the effects of strong electron- ena is typically linked to the presence of steps in the corre-
electron interactions. Important examples of these systems are sponding Hxc potentials.
transition-metal oxides, rare-earth compounds and transition- An interesting question then is whether the insights ob-
metal complexes. In these systems strong electronic cor- tained from LDFT can somehow be exploited to improve the
relations lead to very diverse phenomena such as the Mott performance of standard functionals of real-space DFT for
metal-insulator transition, high-temperature superconductiv- molecules and materials in the presence of strong electronic
ity and spin-crossover behavior.6–9 One of the problems stan- correlations. The challenge lies in finding features of strong
dard approximations to DFT struggle with is the proper correlations that are generic enough to incorporate in approx-
description of multi-determinant, open-shell ground states imate Hxc functionals in order to comply with the univer-
within the KS framework which by construction is single- sal character of the true functional. A good testbed is the
determinant.10–12 To address this problem, various extensions dissociation of the hydrogen (H2 ) molecule, which despite
to DFT have been proposed, such as the DFT+U method13,14 its apparent simplicity presents a challenge for many elec-
or the combination of DFT with dynamical mean-field theory tronic structure methods including DFT. Most approximate
(DMFT).15–19 However, these methods introduce additional functionals encounter difficulties in accurately stretching H2
parameters which can be hard to determine in an ab initio without breaking the spin symmetry.10 At its heart H2 disso-
fashion and may still fail to properly describe a system in spe- ciation is a strongly correlated problem: at large bond dis-
cific circumstances.20,21 tances the many-body ground state acquires a pronounced
While DFT is usually formulated in real space and applied multi-determinant character which poses a challenge to nor-
to real materials and molecules, it is also possible to formu- mal KS-DFT, as explained above.
 2

In this work we first introduce the lattice DFT approach as εis = εi + vHxc,i with
for a two-level system with different interactions (Sec. II) for
which we present the Hxc energy and potential at N = 2 and ∂ ΩHxc (n1 , n2 )
vHxc,i (n1 , n2 ) = (4)
functionals at low temperature. In Sec. III we propose a for- ∂ ni
mal connection between lattice and real-space DFT and we
identify the interactions in terms of KS orbitals. In Sec. IV, where ΩHxc (n1 , n2 ) is the Hxc contribution to the grand-
we apply this connection to the binding energy curve of the canonical potential. Of course, the exact form of ΩHxc (n1 , n2 )
hydrogen molecule in minimal basis and beyond. Finally, we depends on the interaction Ŵ and, in the zero-temperature
present our conclusions in Sec. V. limit, it reduces to the Hxc energy contribution EHxc (n1 , n2 )
to the total ground state energy. For given, fixed single-
particle energies εi of the interacting Hamiltonian (1), the self-
II. LATTICE DENSITY FUNCTIONAL THEORY FOR A consistent KS equations for our model then read
TWO-LEVEL SYSTEM
ni = 2 f (εis ) = 2 f (εi + vHxc,i (n1 , n2 )) (5)
First we review some results for the Hxc potentials for the where f (x) = (1+exp(−β (x− µ )))−1 is the Fermi function at
particular interactions studied in Ref. 27. Then we will in- inverse temperature β and chemical potential µ and the pref-
troduce an additional term to the interaction, the pair-hopping actor 2 is due to spin degeneracy. Since the r.h.s. of Eq. (5)
interaction, which at zero temperature, can be treated analyti- depends on both densities, the two equations (5) for i ∈ {1, 2}
cally within a density functional framework. are coupled and have to be solved simultaneously.
We consider a two-level system described by the Hamilto- In Ref. 27 we studied a two-level system with interactions
nian of the form
2
Ĥ = ∑ εi n̂i + Ŵ (1) Ŵ1 = ∑ Ui n̂i↑n̂i↓ + U12 n̂1n̂2
i=1 i
!
where εi is the single-particle energy of level i and n̂i = −J ∑ n̂1σ n̂2σ + ĉ†1↑ĉ1↓ ĉ†2↓ ĉ2↑ + ĉ†1↓ĉ1↑ ĉ†2↑ ĉ2↓ (6)
† †
∑σ =↑,↓ n̂iσ with n̂iσ = ĉiσ ĉiσ and the ĉiσ (ĉiσ ) are the creation σ
(annihilation) operators for an electron with spin σ in level
i. Finally, Ŵ is the electron-electron interaction whose exact and found from reverse-engineering of exact results that the
form will be specified below. We emphasize that the non- corresponding Hxc potentials in the zero-temperature limit
interacting part of the Hamiltonian of Eq. (1) is diagonal and have a structure dominated by step features which are in-
that spin symmetry is not broken. In the following we work timately related to the famous derivative discontinuity of
in the grand-canonical ensemble at finite temperature T = β1 DFT.38 Even in the zero-temperature limit, the exact form of
the Hxc potential depends on the relative magnitude of the dif-
and chemical potential µ . As usual, expectation values of any
ferent interaction parameters. For parameters U1 ,U2 ≥ U12 ≥
operator Ô are given by O = Tr{ρ̂ Ô} where the trace is over
J we found that the Hxc potential for level i can be decom-
a complete set of states spanning the Fock space and the sta-
posed as
tistical operator is defined as
1
vHxc,i (n1 , n2 ) = vCIM
Hxc (U12 − J)[N]
ρ̂ = exp −β (Ĥ − µ N̂) . (2) +vSSM Inter
Z Hxc (Ui − U12 + J)[ni ] + vHxc (J/2)[N] (7)

Here, Z = Tr{ρ̂ } is the partition function and N̂ = ∑2i=1 n̂i is where

the operator for the total electron number. 3
Of course, since our model is so simple (and the corre- vCIM
sponding Fock space small), it is straightforward to (numeri-
Hxc (U)[N] = ∑ U θ (N − k) (8)
k=1
cally) exactly diagonalize the Hamiltonian Ĥ for any conceiv-
able interaction Ŵ . However, a conceptually alternative ap- is the Hxc potential of the Constant Interaction Model
proach for the calculation of the “densities” ni is to use DFT (CIM)26,39 with N = n1 + n2
in its incarnation of LDFT22,24 . Moreover, since we work
in the grand-canonical ensemble, the proper DFT framework vSSM
Hxc (U)[ni ] = U θ (ni − 1) (9)
is given by Mermin’s finite temperature DFT37 . Therefore, is the Hxc potential of the Single Site Model (SSM)28 and
we are looking for a non-interacting Hamiltonian, the Kohn-
Sham (KS) Hamiltonian, of the form vInter
Hxc (U)[N] = U θ (N − 2) (10)
2
is the interorbital term. Here we would also like to emphasize
Ĥs = ∑ εis n̂i (3)
i=1
that in Eqs. (7) - (10) the function θ (x) should always be read
as a function which in the zero-temperature limit approaches
with KS orbital energies εis which yields the same densities ni the Heaviside step function but in contrast to the Heaviside
as the interacting one. The KS orbital energies can be written function remains continuous for any finite temperature T . By
 3

integrating Eq. (7) one can then also derive the Hxc energy We note that out of all eigenstates, only states |1̃i and |6̃i can
which reads lead to a non-vansihing density difference ∆n = n1 − n2 . With-
3 out loss of generality (since ϑ hasn’t been specified yet), we
(1) assume that out of the two eigenstates |1̃i and |6̃i, |1̃i is the
EHxc (n1 , n2 ) = (U12 − J) ∑ (N − k)θ (N − k)
k=1 one with lower eigenvalue. The densities of this state can eas-
2 ily be computed as
+ ∑ (Ui − U12 + J)(ni − 1)θ (ni − 1)
i=1 n1 = h1̃|n̂1 |1̃i = 2 cos2 ϑ
J
+ (N − 2)θ (N − 2) (11) n2 = h1̃|n̂2 |1̃i = 2 sin2 ϑ (20)
2
Finally we would like to emphasize a peculiarity of the Hamil- or
tonian (1) with interaction Ŵ = Ŵ1 of Eq. (6): for this par-
ticular interaction, all eigenstates of Ĥ are at the same time ∆n = n1 − n2 = 2(cos2 ϑ − sin2 ϑ ) . (21)
eigenstates of the operators n̂i with integer eigenvalues.
Eqs. (20) can then be used to express
A. Including the pair-hopping interaction  
1 ∆n
cos2 ϑ = 1+
2 2
Of course, the interaction of Eq. (6) is not the most general  
interaction possible in a two-level system. One generaliza- 1 ∆n
sin2 ϑ = 1− (22)
tion additionally includes the so-called pair hopping interac- 2 2
tion (with strength P) which is written as
  For our model, the Hohenberg-Kohn (HK) functional (which
Ŵ2 = Ŵ1 − P ĉ†1↑ ĉ†1↓ ĉ2↑ ĉ2↓ + ĉ†2↑ĉ†2↓ ĉ1↑ ĉ1↓ . (12) only depends on ∆n) can be found from the constrained search
approach as
We here restrict ourselves to the two-particle sector for which
we choose the following 2-electron basis F(∆n) = min hΨ|Ŵ2 |Ψi . (23)
Ψ→∆n
|1i = ĉ†1↑ ĉ†1↓ |0i |4i = ĉ†1↓ ĉ†2↑ |0i
Since for given ∆n 6= 0, the minimizing state |Ψ0 i must have
|2i = ĉ†1↑ ĉ†2↑ |0i |5i = ĉ†1↓ ĉ†2↓ |0i the form (17), the HK functional can easily be evaluated as
|3i = ĉ†1↑ ĉ†2↓ |0i |6i = ĉ†2↑ ĉ†2↓ |0i (13)
F(∆n) = h1̃|Ŵ2 |1̃i
where |0i is the vacuum state. It is easy to check that states r
(∆n)2
   
U1 ∆n U2 ∆n
|2i and |5i are two degenerate eigenstates of the Hamiltonian = 1+ + 1− −P 1− (24)
(1) (with Ŵ = Ŵ2 ) with eigenvalue E2/5 = ε1 + ε2 + U12 − J. 2 2 2 2 4
From |3i and |4i we form the two eigenstates
where the minus sign has to be chosen for the last (square root)
1 term for the functional to be minimal for a given ∆n. For ∆n =
|3̃i = √ (|3i − |4i) (14)
2 0, F(∆n) of Eq. (24) coincides with the HK functional if the
1 interaction parameters are such that (U1 +U2 )/2 − P < U12 −
|4̃i = √ (|3i + |4i) (15) J. For (U1 +U2 )/2 − P = U12 − J the states |1̃i and |4̃i become
2
degenerate ground states (at ∆n = 0) while for (U1 +U2 )/2 −
with eigenvalues P > U12 − J the state |4̃i becomes the unique ground state.
E3̃ = ε1 + ε2 + U12 + J It is important to note that the states |1̃i and |4̃i have a very
different character. While the former is a spin singlet formed
E4̃ = ε1 + ε2 + U12 − J , (16) as linear combination of Slater determinants each with a single
i.e., |4̃i is degenerate with |2i and |5i (spin triplets). The re- orbital occupied by two electrons, the latter is a triplet formed
maining eigenstates |1̃i and |6̃i can be written as linear com- as linear combination of Slater determinants each with two
bination of states |1i and |6i different orbitals occupied by a single electron. It is exactly
for this reason that in the limit J = P = 0, Eq. (24) reduces
|1̃i = cos ϑ |1i + sin ϑ |6i (17) to (U1 + U2 )/2 (for ∆n = 0) while at the same densities the
|6̃i = − sin ϑ |1i + cos ϑ |6i (18) proper HK functional for (U1 + U2 )/2 > U12 (which follows
from integrating Eq. (13) of Ref. 27) gives the value U12 .
with some parameter ϑ to be specified and with eigenvalues Since in our model there is no kinetic energy contribution
1 (neither in the interacting nor the non-interacting case), the
E1̃/6̃ = ε1 + ε2 + (U1 + U2 ) functional for the Hxc energy is identical to the HK functional
2
1
q
∓ (2ε1 − 2ε2 + U1 − U2 )2 + 4P2 . (19) EHxc (∆n) = F(∆n) (25)
2
 4

2
and the Hxc potentials are
1 -3
U1 − U2 P∆n P=0.2; T=10
vHxc,1 (∆n) = + q P=0.2; T=10
-4
4

∆n
(∆n)2 0
4 1− 4 P=0.5; T=10
-3

-4
U1 − U2 P∆n -1 P=0.5; T=10
vHxc,2 (∆n) = − − q (26)
4 2
4 1 − (∆n)
4 -2

or 0.003
s -3
ε1, T=10
0

s
εi
s -3
∆vHxc (∆n) = vHxc,1 (∆n) − vHxc,2(∆n) ε2, T=10
s -4
U1 − U2 P∆n -0.003 ε1, T=10
= + q . (27) s
ε2,
-4
T=10
2 2
2 1 − (∆n)4
-0.006
-4 -2 0 2 4
∆ε
Knowing the HK functional F(∆n), we can for given values of
FIG. 1. Upper panel: Density imbalance ∆n = n1 − n2 versus sin-
the single-particle parameters (on-site energies) ε1 = ∆2ε0 and gle particle energy difference ∆ε = ε1 − ε2 for the model described
ε2 = − ∆2ε0 find the corresponding ground state density ∆n0 by by the Hamiltonian (1) with interaction Ŵ2 of Eq. (12) for differ-
minimizing the HK total energy functional ent values of the pair-hopping interaction P and different tempera-
tures T . The other parameters are U1 = U2 = U12 = 1.0 and J = P.
∆ε0 Lower panel: KS eigenvalues εis = εi + vHxc,i at self-consistency for
E∆ε0 (∆n) = ∆n + F(∆n) , (28) P = J = 0.5 and two different temperatures T . The other parameters
2
are U1 = U2 = U12 = 1.0. All energies given in units of U1 .
i.e., we solve

∂ E∆ε0 (∆n)
=0
∂ ∆n ∆n=∆n0
U1 − U2 P ∆n0 system and both solutions are identical on the scale of the fig-
= ∆ε0 + + q . (29)
4 4 2
ure (not shown). Moreover, we have computed ∆n for two dif-
1 − (∆n40 )
ferent, small temperatures T and found no appreciable differ-
This can be rearranged for ∆n0 and gives ence in the corresponding density imbalances. However, we
would like to draw attention to the particular way in which the
2(2∆ε0 + U1 − U2) KS system achieves the same density as the interacting one.
∆n0 = − p (30) We first note that the Fermi functions in the zero temperature
(2∆ε0 + U1 − U2)2 + 4P2
limit approaches a step function and naively, Eq. (31) seems
where the minus sign was chosen for the physical reason that to suggest that its solution can only be ±2 which is clearly not
we should have ∆n0 → 2 for ∆ε0 → −∞. Reinserting this den- the case for arbitrary values of ∆ε . Therefore, the Hxc poten-
sity into the HK energy functional (28), one can easily confirm tial must lead to total KS energies to be close to the chemical
that one obtains the ground state energy E1̃ of Eq. (19). potential, i.e., the step region of the Fermi functions. This is
Above we have computed the ground state density ∆n0 di- confirmed by looking at the lower panel of Fig. 1 where we
rectly from the HK variational principle. The same result show the self-consistent KS eigenvalues for the parameters of
should, of course, also be obtained by self-consistent solution the upper panel with P = 0.5 at two small temperatures. While
of the KS equations (5). For given single-particle parameters the density already is basically converged at T = 10−3 , the KS
εi , these equations can be conveniently rewritten as eigenvalues are not. In fact, in the T → 0 limit the KS eigen-
values will converge to the chemical potential µ (here taken
∆n = 2 [ f (ε1 + vHxc,1 (∆n)) − f (ε2 + vHxc,2 (∆n))] (31) to be zero) because they have to reproduce the densities. This
can only be achieved by moving the KS eigenvalues onto the
and the solution of this equation has to coincide with the step of the Fermi function whose width is of the order of T .
ground state density of Eq. (30) which we have confirmed Even in this limit, however, the upper and lower KS eigen-
numerically. In the upper panel of Fig. 1 we show the den- values have to converge to slightly different values in order
sity imbalance ∆n as function of ∆ε = ε1 − ε2 for different to give a finite density imbalance ∆n. This highlights the im-
values of the pair hopping parameter P and different temper- portance of working at small but finite temperature for which
atures. The other parameters are U1 = U2 = U12 = 1.0 and the Fermi functions exhibit a step but are not mathematically
J = P and where chosen to ensure that the ground state has discontinuous. This is somewhat related to findings in previ-
the form of Eq. (17) for any ∆ε . As expected, the stronger the ous works28,40 where it was found that Hxc potentials which
pair-hopping interaction, the more the step in ∆n at ∆n = 0 is exhibit step stuctures typically lead to a pinning of KS energy
smoothened. We have solved both the many-body problem as levels to the Fermi energy for a range of single-particle ener-
well as the self-consistent solution of the corresponding KS gies.
 5

III. ESTABLISHING A CONNECTION BETWEEN of the KS orbitals and the occupation numbers functionals of
LATTICE DFT AND REAL-SPACE DFT the KS energy eigenvalues. In the context of DFT, function-
als depending on KS orbitals and KS eigenvalues are implicit
In the present Section and based on the results of the pre- functionals of the density and it is well known that the cor-
vious Section, we will establish a connection between lat- responding (Hxc) potentials (i.e., functional derivatives with
tice DFT and the usual DFT formulation in real space. The respect to the density) can be computed with the Optimized
combination of standard DFT with advanced many-body tre- Effective Potential (OEP) formalism42–44 .
taments of lattice models has a long tradition, possibly starting We now take a look at what our interpretation of the lattice
with what is known as DFT+U method13. A more recent ap- DFT functionals in terms of real-space KS orbitals and occu-
(1)
proach combines dynamical mean-field theory (DMFT) with pation numbers implies for the specific example of EHxc given
DFT15,17,19 . In a somewhat different line, a direct transfer of by Eq. (11). We will not consider general values for n1 and
ideas from lattice DFT to standard quantum chemical methods n2 but instead focus on those points where both occupation
has been suggested recently41. Here we suggest an alternative numbers are integer. For n1 = n2 = 0 as well as for n1 = 1,
connection between lattice DFT and real-space DFT based on (1) (1)
n2 = 0 and n1 = 0, n2 = 1 we see that EHxc vanishes, i.e., EHxc
our results for the two-level system described in Sec. II. is self-interaction free for one electron. For n1 = 2 and n2 = 0
In Mermin’s version of finite-temperature DFT37 , the elec- (1)
we find EHxc (n1 = 2, n2 = 0) = U1 which can be written as
tronic density (in real space) in thermal equilibrium for an in-
teracting many-electron system subject to an electrostatic po- 1 ρ (r)ρ (r′ )
Z Z
(1)
tential v0 (r) is determined by self-consistent solution of the EHxc (n1 = 2, n2 = 0) = d3 r d3 r ′ (37)
4 |r − r′|
KS equation
where we assumed spin-independent KS orbitals such that the
∇2
 
− + v0 (r) + vHxc [ρ ](r) ϕiσ (r) = εisσ ϕiσ (r) (32) density becomes ρ (r) = 2|ϕ1σ (r)|2 . The same result (in terms
2 of the density) is obtained for n1 = 0, n2 = 2. For full oc-
cupation of the two-level system, n1 = 2, n2 = 2, we obtain
with the KS orbitals ϕiσ and the KS energy eigenvalues εisσ . (1)
The electronic density is given by EHxc (n1 = 2, n2 = 2) = 4U12 + U1 + U2 − 2J which can be
written as
ρ (r) = ∑ niσ |ϕiσ (r)|2 (33) 1
Z Z
ρ (r)ρ (r′ )
(1)
iσ EHxc (n1 = 2, n2 = 2) = d3 r d3 r ′
2 |r − r′|
where the occupation of the KS orbital ϕiσ is given by niσ =
1 2
ϕ ∗ (r)ϕ jσ (r)ϕ2∗σ (r′ )ϕ1σ (r′ )
Z Z
f (εisσ ). In Eq. (32), the Hxc potential is given as functional d3 r d3 r′ iσ
derivative of the Hxc contribution ΩHxc to the grand canonical
− ∑ ∑
2 σ i, j=1 |r − r′|
.
(38)

potential, i.e.,
We recognize that both Eq. (37) and (38) are nothing but the
ΩHxc [ρ ] Hartree plus the exact exchange energy of standard DFT. In
vHxc [ρ ](r) = (34)
δ ρ (r) fact, also for the other integer occupations the functional of
Eq. (11) can with our interpretation (lattice densities identified
which, of course, has to be approximated in practice. as occupations of KS orbitals and interaction parameters de-
In Sec. II we have found the Hxc energy EHxc (i.e., the fined in terms of KS orbitals) be identified as the Hartree plus
zero-temperature limit of ΩHxc ) for the two-level system as exact exchange functionals. However, for occupations n1 = 2,
function of the occupation numbers ni and the interaction n2 = 1 and n1 = 1, n2 = 2 as well as for n1 = n2 = 1 this
parameters Ui , U12 , J, and P. In order to make a connec- identification requires the use of the proper definition of the
tion between these lattice DFT results and real-space DFT, Hartree plus exact exchange energy of ensemble DFT derived
we now propose to interpret the occupation numbers as ni = in Refs. 45 and 46. Based on the recovery of the Hartree plus
∑σ niσ = ∑σ f (εisσ ) and the interaction parameters as two- exact exchange energy for integer occupations, we may infer
electron Coulomb integrals with respect to the KS orbitals. that the energy functional of Eq. (11) is actually the proper
We define a general two-electron integral as generalization of this functional for any non-integer occupa-
Z
ϕi∗σ (r)ϕ jσ (r)ϕk∗σ ′ (r′ )ϕl σ ′ (r′ )
Z tion 0 ≤ ni ≤ 2. Furthermore, from the recovery of a known
(iσ jσ |kσ ′ l σ ′ ) = d3 r d3 r ′ functional from Eq. (11), we also gain confidence that the
|r − r′|
functional of Eq. (25) with our interpretation is a reasonable
(35)
approximation to the exact Hxc functional of a two-level sys-
Then the interaction parameters can be identified in terms of
tem. This will be borne out in the next Section where we apply
these two-electron integrals as
the functional (25) to the description of the H2 molecule.
Ui = (iσ iσ |iσ̄ iσ̄ ) J = (1σ 2σ |2σ 1σ )
U12 = (1σ 1σ |2σ 2σ ) P = (1σ 2σ̄ |1σ̄ 2σ ) (36)
IV. APPLICATION TO THE HYDROGEN MOLECULE
where σ̄ =↓ (↑) for σ =↑ (↓). We note that for real and spin-
independent KS orbitals we have J = P. With this interpreta- In the present Section we describe an application of the for-
tion, the interaction parameters formally become functionals malism presented in Sec. II to the hydrogen molecule. We
 6

start with the problem treated with a minimal basis set before H2 molecule in the minimal basis (mb) takes the form
investigating larger basis sets.
2
ĤHmb
2
= ∑ εi n̂i + Enuc(R)
i=1
A. Minimal basis set 2
1
+ ∑ ∑ (iσ jσ ′ |kσ ′ l σ )ĉ†iσ ĉ†jσ ′ ĉkσ ′ ĉl σ (46)
In the minimal basis for the hydrogen molecule (H2 ), for 2 σ ,σ ′ i j,k,l=1
each atom we have only one single s-type basis function. Let
g(r)(= g(−r)) be such an s-type normalized (real) basis func- where the single-particle energies are given by
tion localized at the origin r0 = 0. If we take the hydro-
gen atoms to be located at ±R/2, from the localized basis εi = hiσ |ĥ0 (r)|iσ i (47)
functions g1/2 (r) = g(r ± R/2) we can construct two normal-
and the off-diagonal matrix elements of ĥ0 (r) vanish due to
ized and orthogonal (spin) orbitals, one bonding and one anti-
bonding, which take the form the symmetry ĥ0 (−r) = ĥ0 (r) together with the symmetry
(41) of the basis functions, i.e.,
1 Z
ϕ1 (r) = ϕ1σ (r) = p (g1 (r) + g2(r))
hiσ |ĥ0 (r)| jσ ′ i = δσ σ ′ δi j d3 r ϕi∗σ (r)ĥ0 (r)ϕiσ (r) = δσ σ ′ δi j εi .
2(1 + S)
1 (48)
ϕ2 (r) = ϕ2σ (r) = p (g1 (r) − g2(r)) (39) Again, due to the symmetry (41) of the basis functions, it is
2(1 − S)
easy to show that all those two-electron integrals vanish for
where which three of the four indices {i, j, k, l} are equal and the
Z Hamiltonian for H2 in the minimal basis becomes
S= d3 r g1 (r)g2 (r) (40) 2
ĤHmb
2
= ∑ εi n̂i + Ŵ2 + Enuc(R) . (49)
is the overlap integral of the two localized basis functions. We i=1
note that by construction the ϕk (r) are eigenfunctions of the Since this Hamiltonian now has exactly the form of the one
parity operator, i.e., they satisfy the symmetry relations studied in the Sec. II (plus the additive nuclear repulsion en-
ergy), the total energy functional of H2 takes the form
ϕ1 (−r) = ϕ1 (r) ϕ2 (−r) = −ϕ2 (r) (41)
2
When written in terms of field operators, the Hamiltonian of EH2 (∆n) = ∑ εi ni + F(∆n) + Enuc(R) (50)
the hydrogen molecule is given by i=1

with the HK functional of Eq. (28). The direct minimization

Z
ĤH2 = ∑ d3 r ψ̂σ† (r)ĥ0 (r)ψ̂σ (r) + Enuc(R) for ∆n as well as the solution of the corresponding KS equa-
σ
tion (31) proceed as described in Sec. II and the resulting dif-
1 1
Z Z
+ ∑ d3 r d3 r′ ψ̂σ† (r)ψ̂σ† ′ (r′ ) ψ̂ ′ (r′ )ψ̂σ (r)
(42) ference in occupation numbers is given by Eq. (30). Note that
2 σ ,σ ′ |r − r′ | σ here we are only minimizing the occupation number differ-
ence while the orbitals of Eq. (39) remain fixed.
with the single-particle Hamiltonian We have mapped out the binding energy curve for the
H2 molecule in minimal basis with the open source PySCF
∇2 code47. We calculated, for each internuclear distance R, the
ĥ0 (r) = − + v(r) (43)
2 interaction parameters Ui and P as well as the corresponding
matrix elements (47) which enter in the evaluation of the oc-
where v(r) is the attractive potential due to the protons cupation numbers (Eq. (30)). The total energy was then com-
puted by adding the internuclear repulsion to the total energy
1 1
v(r) = − − , (44) functional (28). In the minimal basis, our approach becomes
|r − R/2| |r + R/2| exact and therefore equivalent to full configuration interaction
(FCI) which can be confirmed analytically48. In Fig. 2 we
and the nuclear electrostatic repulsion energy Enuc (R) = 1/R
show the binding energy curves from our approach and com-
with R = |R|.
pare with standard spin-restricted Hartree-Fock (HF) as well
In the minimal basis, the field operators can be written as
as DFT calculations using the LDA and PBE functionals49.
2 As expected, our approach recovers the full CI results while
ψ̂σ† (r) = ∑ ϕi∗σ (r)ĉ†iσ (45) both spin-restricted HF and standard DFT, as is well-known,
i=1 do not recover the correct large separation limit (see, how-
ever, the partition DFT of Refs. 50 and 51 which also captures
where the ĉ†iσ are the creation operators for an electron in or- dissociation without breaking the spin symmetry). The rea-
bital ϕiσ . When inserted into Eq. (42) the Hamiltonian of the son why our approach does indeed recover this limit simply
 7

FIG. 2. Binding energy of the hydrogen molecule in minimal basis FIG. 3. Left panel: Binding energy of the hydrogen molecule in cc-
as function of the internuclear separation R for various spin-restricted pvtz basis as function of internuclear separation R using HF, LDA,
approaches: Hartree-Fock (HF) as well as the DFT approaches using and PBE (spin-restricted) self-consistent orbitals to evaluate the en-
the LDA and PBE functionals and our functional (LDFT). The full ergy with our approach in comparison to full CI results. Upper right
CI result is given as reference. In the inset we show the occupation panel: occupation numbers of the HOMO and LUMO KS orbitals as
number difference in full CI as function of R which is identical to our function of R computed from different self-consistent orbitals. Lower
LDFT result. All quantities given in atomic units. right panel: interaction parameters as function of R from different or-
bitals. All quantities given in atomic units.

lies in the fact that both the ground-state Slater determinant as

well as the doubly excited Slater determinant contribute, i.e., for the spin-restricted self-consistent calculations. The closest
the occupation numbers n1 and n2 are not strictly integer. In results to the full CI reference in this limit is achieved with
the context of our DFT approach this is possible because we PBE, followed by LDA and HF. Around the equilibrium bond
are working with the equilibrium grand-canonical ensemble distance the three sets of orbitals lead to similar results, very
of non-interacting KS wavefunctions and not just with the KS close to self-consistent HF. LDA orbitals give significantly
ground state wavefunction. improved total energies as compared to self-consistent LDA
results, while PBE orbitals give slightly worse energies than
self-consistent PBE ones.
B. Beyond the minimal basis In the upper right panel of Fig. 3 we compare orbital occu-
pations from our approach with the three sets of orbitals. They
We have also evaluated the binding energy curves of the all follow reasonably well the exact reference results with HF
hydrogen molecule within our approach but using a larger ba- performing slightly worse. The interaction parameters (lower
sis. To this end, we performed self-consistent HF, LDA, and right panel of Fig. 3) from the three sets of orbitals are reason-
PBE calculations with the larger cc-pvtz basis47 . From the ably similar among each other, again with HF showing more
resulting lowest-lying molecular KS orbitals, the HOMO or- pronounced differences to the DFT results, especially for the
bital ϕH (r) = ϕ1σ (r) and orbital energy εH = ε1 as well as the interaction parameter UL corresponding to the LUMO orbital.
LUMO orbital ϕL (r) = ϕ2σ (r) with orbital energy εL = ε2 we As discussed above, the results of the present Section for
evaluate both the single-particle matrix elements of Eq. (47) our approach were not computed self-consistently but with or-
as well as all the interaction parameters. Note that the orbital bitals obtained either from HF or other DFT functionals. In
energies εi are not KS energy eigenvalues but the expectation principle our two-level functional can be read as an orbital-
values of the core Hamiltonian in the KS orbitals, as defined dependent functional which requires the OEP method for cal-
by Eq. (47). Once these parameters are determined, we eval- culation of the corresponding self-consistent potential which,
uate the total energy according to Eq. (50), i.e., we find the however, is beyond the scope of the present work.
orbital occupations ni which minimize the total energy accord-
ing to our two-level model while keeping the orbitals fixed.
In the left panel of Fig. 3 we show the binding energy V. SUMMARY AND CONCLUSIONS
curves obtained in our approach from the three different sets
of self-consistent molecular (spin-restricted) orbitals obtained The present work was motivated by earlier work27 on a
with the three different functionals in comparison to the cor- simple model of a double quantum dot on a lattice subject
responding standard spin-restricted and full CI results. As a to different types of electron-electron interactions. We aimed
common theme, for the three sets of orbitals the large sep- at finding possible connections between lattice DFT (here in
aration limit is captured correctly with our approach unlike the framework of Mermin’s finite-temperature version of DFT
 8

in the grand-canonical ensemble37) and real-space DFT. This 8 P. Gütlich, A. B. Gaspar, and Y. Garcia, Beilstein J. Org. Chem. 9, 342
connection could be found if the interaction paramaeters of (2013).
9 C. Ahn, A. Cavalleri, A. Georges, S. Ismail-Beigi, A. J. Millis, and J.-M.
the lattice DFT model are read as two-electron Coulomb inte-
Triscone, Nat. Mater. 20, 1462 (2021).
grals with respect to the KS orbitals of real-space DFT and the 10 A. J. Cohen, P. Mori-Sánchez, and W. Yang, Chem. Rev. 112, 289 (2012).
lattice “densities” are interpreted as the occupation numbers 11 N. Q. Su and X. Xu, Annu. Rev. Phys. Chem. 68, 155 (2017).
12 J. P. Perdew, A. Ruzsinszky, J. Sun, N. K. Nepal, and A. D. Kaplan, Proc.
of these KS orbitals. For the interactions studied in Ref. 27
we found that, for integer occupation of the orbitals, this in- Natl. Acad. Sci. USA 118, e2017850118 (2021).
13
terpretation leads to a recovery of the exact-exchange energy V. I. Anisimov, J. Zaanen, and O. K. Andersen, Phys. Rev. B 44, 943 (1991).
14 L. Wang, T. Maxisch, and G. Ceder, Phys. Rev. B 73, 195107 (2006).
functional. 15 G. Kotliar, S. Y. Savrasov, K. Haule, V. S. Oudovenko, O. Parcollet, and
We also studied an additional term to the interaction in C. A. Marianetti, Rev. Mod. Phys. 78, 865 (2006).
16 K. Held, Adv. Phys. 56, 829 (2007).
the lattice model, the pair-hopping interaction, for which we
17 D. Jacob, K. Haule, and G. Kotliar, Phys. Rev. B 82, 195115 (2010).
found the analytical form of the exact HK energy functional. 18 M. Karolak, T. O. Wehling, F. Lechermann, and A. I. Lichtenstein, J. Phys.:
As a crucial difference to the previously studied interactions, Condens. Matter 23, 085601 (2011).
we found that the pair-hopping term leads to non-integer oc- 19
C. Weber, D. J. Cole, D. D. O’Regan, and M. C. Payne, Proc. Nat. Acad.
cupations of the KS orbitals even in the limit of zero tem- Sci. 111, 5790 (2014).
20 N. Zaki, H. Park, R. M. Osgood, A. J. Millis, and C. A. Marianetti, Phys.
perature. Another connection to real-space DFT could be es-
tablished by showing that the Hamiltonian of the hydrogen Rev. B 89, 205427 (2014).
21 W. Huang, D.-H. Xing, J.-B. Lu, B. Long, W. H. E. Schwarz, and J. Li, J.
molecule (H2 ), when treated in a minimial basis of localized Chem. Theory Comput. 12, 1525 (2016).
basis functions, has exactly the form of our lattice model with 22 O. Gunnarsson and K. Schönhammer, Phys. Rev. Lett. 56, 1968 (1986).
23 K. Schönhammer, O. Gunnarsson, and R. M. Noack, Phys. Rev. B 52, 2504
the pair-hopping term included. Knowing the exact HK en-
ergy functional (and therefore also the exact Hxc energy func- (1995).
24 N. A. Lima, M. F. Silva, L. N. Oliveira, and K. Capelle, Phys. Rev. Lett. 90,
tional) for this model, the binding energy curve of H2 was
146402 (2003).
found to coincide with the full CI one in the minimal basis 25 D. Carrascal, J. Ferrer, J. C. Smith, and K. Burke, J. Phys.: Condens. Matter
for all internuclear distances and without the need of breaking 27, 393001 (2015).
26 S. Kurth and G. Stefanucci, J. Phys.: Condens. Matter 29, 413002 (2017).
spin symmetry. This is achieved by both the KS ground state
27 N. Sobrino, S. Kurth, and D. Jacob, Phys. Rev. B 102, 035159 (2020).
and the doubly excited KS determinant having finite weight 28
G. Stefanucci and S. Kurth, Phys. Rev. Lett. 107, 216401 (2011).
in the KS ensemble (in the zero-temperature limit) which is 29 J. P. Bergfield, Z.-F. Liu, K. Burke, and C. A. Stafford, Phys. Rev. Lett. 108,
equivalent to saying that both HOMO and LUMO KS or- 066801 (2012).
bitals have non-integer occupation in this ensemble. Finally, 30 P. Tröster, P. Schmitteckert, and F. Evers, Phys. Rev. B 85, 115409 (2012).
31 D. Karlsson, A. Privitera, and C. Verdozzi, Phys. Rev. Lett. 106, 116401
we suggested a post-SCF evaluation of our energy functional
for larger basis sets allowing to recover the correct large- (2011).
32 G. Stefanucci and S. Kurth, Nano Lett. 15, 8020 (2015).
separation limit without spin symmetry breaking. This latter 33 S. Kurth and G. Stefanucci, Phys. Rev. B 94, 241103(R) (2016).
approach still requires a self-consistent treatment, e.g., within 34
D. Jacob and S. Kurth, Nano Lett. 18, 2086 (2018).
the OEP approach, which will be the subject of future work. 35 N. Sobrino, R. D’Agosta, and S. Kurth, Phys. Rev. B 100, 195142 (2019).
36 D. Jacob, G. Stefanucci, and S. Kurth, Phys. Rev. Lett. 125, 216401 (2020).
37 N. Mermin, Phys. Rev. 137, A1441 (1965).
38 J.P. Perdew, R.G. Parr, M. Levy, and J.L. Balduz, Phys. Rev. Lett. 49, 1691

(1982).
ACKNOWLEDGMENTS 39 G. Stefanucci and S. Kurth, Phys. Stat. Sol. (b) 250, 2378 (2013).
40 S. Kurth and G. Stefanucci, Phys. Rev. Lett. 111, 030601 (2013).
41
J. P. Coe, Phys. Rev. B 99, 165118 (2019).
We acknowledge financial support through 42 J.D. Talman and W.F. Shadwick, Phys. Rev. A 14, 36 (1976).
Grant PID2020-112811GB-I00 funded by 43 T. Grabo, T. Kreibich, S. Kurth, and E.K.U. Gross, in Strong Coulomb

MCIN/AEI/10.13039/501100011033 as well as by grant Correlations in Electronic Structure Calculations: Beyond Local Density
IT1453-22 “Grupos Consolidados UPV/EHU del Gobierno Approximations, edited by V. Anisimov (Gordon and Breach, Amsterdam,
2000), p. 203.
Vasco”. 44 S. Kümmel and L. Kronik, Rev. Mod. Phys. 80, 3 (2008).
45 T. Gould and S. Pittalis, Phys. Rev. Lett. 119, 243001 (2017).
46 T. Gould and S. Pittalis, Phys. Rev. Lett. 123, 016401 (2019).
1 P. 47 Q. Sun, T. C. Berkelbach, N. S. Blunt, G. H. Booth, S. Guo, Z. Li, J. Liu,
Hohenberg and W. Kohn, Phys. Rev. 136, B864 (1964).
2 W. Kohn and L. Sham, Phys. Rev. 140, A1133 (1965). J. D. McClain, E. R. Sayfutyarova, S. Sharma, S. Wouters, and G. K.-L.
3 R.G. Parr and W. Yang, Density-Functional Theory of Atoms and Molecules Chan, Wiley Interdiscip. Rev. Comput. Mol. Sci. 8, e1340 (2018).
48 A. Szabo and N. S. Ostlund, Modern Quantum Chemistry (McGraw-Hill,
(Oxford University Press, New York, 1989).
4 R.M. Dreizler and E.K.U. Gross, Density Functional Theory (Springer, New York, 1989).
49 J.P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. 77, 3865 (1996);
Berlin, 1990).
5 K. Burke, J. Chem. Phys. 136, 150901 (2012). ibid. 78, 1396 (1997)(E).
6 M. Imada, A. Fujimori, and Y. Tokura, Rev. Mod. Phys. 70, 1039 (1998). 50 J. Nafziger and A. Wasserman, J. Chem. Phys. 143, 234105 (2015).
7 P. A. Lee, N. Nagaosa, and X.-G. Wen, Rev. Mod. Phys. 78, 17 (2006). 51 Y. Shi, Y. Shi, and A. Wasserman, arXiv:2305.13545 (2023).