arXiv:2204.12216v1 [physics.

chem-ph] 26 Apr 2022

Chapter 1

Lieb variation principle in density-functional theory

Trygve Helgaker and Andrew M. Teale

Abstract. Lieb’s convex formulation of density-functional theory is presented in a pedagogical

manner, emphasizing its connection to Hohenberg–Kohn theory and to Levy’s constrained-
search theory. The Hohenberg–Kohn and Lieb variation principles are discussed, highlighting
the dual relationship between the ground-state energy and the universal density functional.
Applications of the Lieb variation principle are reviewed, demonstrating how it may be util-
ized to calculate the Kohn–Sham potential of atoms and molecules, to study the exchange–
correlation functional and the adiabatic connection by high-precision many-body methods, and
to calculate the exchange–correlation hole and energy densities of atoms and molecules.

Density-functional theory (DFT) is a hugely successful and powerful theory, under-

lying nearly all electronic-structure calculations of molecules and materials in chem-
istry and physics today. It is also a very beautiful theory – an elegant application
of convex analysis to electronic-structure theory, predicated on the simple obser-
vation that the ground-state electronic energy of an atom, molecule, or material is
continuous and concave in the external potential, as realized in 1983 by Lieb in his
seminal work ‘Density Functionals for Coulomb Systems’ [1]. Unfortunately, nearly
forty years later, Lieb’s convex formulation of DFT is still not widely appreciated
among the practitioners of DFT or even among many of its developers. Likewise, the
teaching of DFT is still mostly based on the Hohenberg–Kohn theorem [2] and Levy’s
constrained-search theory [3]. This is a pity since Lieb’s convex, unifying formulation
of DFT is both accessible and transparent – and, in our experience, likely to intrigue
students.
Lieb’s theory does not only give valuable insight into DFT. It also furnishes us
with a practical tool for computation: the Lieb variation principle, according to which
we may calculate to high accuracy, for any given electron density, the corresponding
external potential and the universal density functional, thereby providing us with a
practical realization of the Hohenberg–Kohn mapping from densities to potentials.
Furthermore, it allows us to study the adiabatic connection and to calculate the uni-
versal density functional of DFT using high-precision quantum-chemical methods,

2020 Mathematics Subject Classification. Primary XXxXX; Secondary YYyYY.

Keywords. Density-functional theory, Hohenberg–Kohn variation principle, Lieb variation
principle, universal density functional theory, Hohenberg–Kohn theorem.
2 T. Helgaker and A. M. Teale

providing both insight and valuable benchmark data for the development of approx-
imate density functionals.
The bulk of this chapter consists of two parts. We begin in Section 1.1 by review-
ing Lieb’s convex formulation of DFT, emphasizing its relationship to Hohenberg–
Kohn theory and to Levy’s constrained-search theory. Next, in Section 1.2, we review
and illustrate the Lieb variation principle as a computational tool in DFT, with applic-
ations to both Kohn–Sham theory and orbital-free DFT. Section 1.3 contains some
concluding remarks.

1.1 Lieb’s convex formulation of density-functional theory

The following presentation of Lieb’s formulation of DFT is intended to be an informal

introduction rather than a rigorous exposition of the theory. Concepts and elements of
convex analysis are introduced as needed, in a hopefully self-contained manner. For
an accessible, rigorous treatment convex analysis, we highly recommend the intro-
ductory text of van Tiel [4].

1.1.1 Hohenberg–Kohn theorem

The Hamiltonian of an atomic or molecular system of 𝑁 electrons in an external
potential 𝑣 can be written in the form:
∑︁ 𝑁
𝐻 (𝑣) = 𝑇 + 𝑊 + 𝑣(r𝑖 ). (1.1)
𝑖=1

Whereas the external potential 𝑣 varies from system to system, the kinetic operator 𝑇 =
ℏ2 Í 𝑁 2 𝑒 2 Í 𝑁 Í𝑖 −1 are
− 2𝑚 e 𝑖=1 ∇r𝑖 and the electron-repulsion operator 𝑊 = 4 𝜋 𝜀0 𝑖=1 𝑗=1 |r𝑖 − r 𝑗 |
the same for all 𝑁-electron systems. We are particularly interested in those potentials
𝑣 that support an electronic ground state and denote the set of these potentials by V𝑁 .
For a given 𝑣 ∈ V𝑁 , the ground-state energy 𝐸 (𝑣) is obtained by solving the
Schrödinger equation,
𝐻 (𝑣)Ψ(x𝑖 ) = 𝐸 (𝑣)Ψ(x𝑖 ). (1.2)
Its solution is a complicated many-body problem, the wave function Ψ(x𝑖 ) depending
on the spatial and spin coordinates x𝑖 = (r𝑖 , 𝜎𝑖 ) of all electrons. By contrast, the
associated ground-state density is a much simpler quantity, depending only on three
spatial coordinates:
∫
𝜌(r) = |Ψ(r, 𝜎, x2 , . . . x 𝑁 )| 2 d𝜎dx2 · · · dx 𝑁 , (1.3)

where we interpret the integration over spin coordinates as a summation over 𝛼 and 𝛽
spins. According to the Hohenberg–Kohn theorem, we may use the density as a fun-
 Lieb variation principle in density-functional theory 3

damental variable in place of the wave function when studying atomic and molecular
systems [2]:
‘Thus 𝑣(r) is (to within a constant) a unique functional of 𝜌(r); since, in
turn, 𝑣(r) fixes 𝐻 (𝑣) we see that the full many-body ground state is a unique
functional of 𝜌(r).’
Equipped with this theorem, we can set up the Hohenberg–Kohn mapping from ground-
state densities to ground-state wave functions via the external potential and the Hamilto-
nian:
𝜌 ↦→ 𝑣𝜌 ↦→ 𝐻 (𝑣𝜌 ) ↦→ Ψ𝜌 . (1.4)
A density that is the ground-state density for some potential 𝑣 ∈ V𝑁 is said to be
𝑣-representable and the set of all 𝑣-representable densities is denoted by A 𝑁 .
The Hohenberg–Kohn theorem (1964) has played an enormously important role in
theoretical chemistry – indeed, it is typically viewed as the cornerstone of DFT, taught
in all introductions to DFT along with Levy’s constrained-search theory (1979). At
the same time, the Hohenberg–Kohn theorem may appear almost mysterious and diffi-
cult to understand intuitively. Although Levy’s constrained-search formulation brings
clarity, the underlying structure and beauty of DFT is best appreciated from the point
of view Lieb’s convex formulation of the theory (1983).
It is interesting to note that many of the elements or concepts of Lieb’s theory
were already present in the early work of Hohenberg and Kohn but were not picked
up on and developed further at the time. We will here highlight these connections,
beginning with the subgradient inequality of the ground-state energy.

1.1.2 Subgradient inequality of the ground-state energy

As part of the proof of the Hohenberg–Kohn theorem [2], the following subgradient
inequality of the ground-state energy at 𝑣 ∈ V𝑁 was established:

𝐸 (𝑢) < 𝐸 (𝑣) + (𝑢 − 𝑣 | 𝜌 𝑣 ), ∀𝑢 ∉ 𝑣 + R,

(1.5)
𝐸 (𝑢) = 𝐸 (𝑣) + (𝑢 − 𝑣 | 𝜌 𝑣 ), ∀𝑢 ∈ 𝑣 + R.

Here 𝜌 𝑣 ∈ A 𝑁 is a ground-state density associated with 𝐻 (𝑣) and we use the short-
hand notation ∫
(𝑢|𝜌) = 𝑢(r) 𝜌(r) dr (1.6)

to denote the interaction of 𝑢 with 𝜌. An intuitive understanding of the subgradient

inequality can be obtained from Figure 1.1, which shows how the inequality arises
from the Rayleigh–Ritz variation principle and the linearity of the Hamiltonian in the
potential. As seen in the figure, the subgradient inequality is a consequence of the
concavity of the ground-state energy. Geometrically speaking, a function is said to be
4 T. Helgaker and A. M. Teale

Figure 1.1. The concavity of the ground-state energy 𝑣 ↦→ 𝐸 (𝑣) in the potential 𝑣 (blue line),
assuming that the potentials 𝑣0 , 𝑣1 ∈ V𝑁 have ground-state wave functions Ψ0 and Ψ1 , respect-
ively. As the potential changes from 𝑣0 to 𝑣1 with the ground-state wave function Ψ0 fixed along
the red line, the energy changes linearly. When the wave function relaxes from Ψ0 to the ground-
state Ψ1 at 𝑣1 , the ground-state energy is lowered, generating a concave curve 𝑣 ↦→ 𝐸 (𝑣). The
slope of the tangent to 𝐸 at 𝑣0 is the density of the ground-state wave function Ψ0 at 𝑣0 . In the
terminology of convex analysis, the ground-state density 𝜌0 is the ‘subgradient’ of the ground-
state at the potential 𝑣0 . The ‘subgradient inequality’ in Eq. (1.5) therefore expresses the notion
that the straight red line in the plot is a tangent to the concave blue curve.

concave if every chord connecting two points on its graph lies on or below the graph.
As we shall see later, concavity of the ground-state energy is the key to DFT.
The subgradient inequality gives the Hohenberg–Kohn theorem directly. Let the
potentials 𝑢, 𝑣 ∈ V𝑁 differ by more than an additive constant and let 𝜌𝑢 , 𝜌 𝑣 ∈ A 𝑁
be the corresponding ground-state densities. From Eq. (1.5), we then obtain the strict
subgradient inequalities 𝐸 (𝑢) < 𝐸 (𝑣) + (𝑢 − 𝑣 | 𝜌 𝑣 ) and 𝐸 (𝑣) < 𝐸 (𝑢) + (𝑣 − 𝑢 | 𝜌𝑢 ).
Adding these together, we conclude that (𝑣 − 𝑢 | 𝜌 𝑣 − 𝜌𝑢 ) < 0 and hence that 𝜌𝑢 ≠ 𝜌 𝑣
in accordance with the Hohenberg–Kohn theorem. The Hohenberg–Kohn theorem is
thus a simple consequence of the (strict) concavity of the ground-state energy in the
external potential, which in turn follows from the Rayleigh–Ritz variation principle
and the linearity of the Hamiltonian in the potential.
 Lieb variation principle in density-functional theory 5

1.1.3 Hohenberg–Kohn and Lieb variation principles

The subgradient inequality of the ground-state energy is also the key to the Hohenberg–
Kohn variation principle. For a 𝑣-representable density 𝜌 ↦→ 𝑣𝜌 , the Hohenberg–Kohn
universal density functional [2] is defined as

𝐹HK (𝜌) = 𝐸 (𝑣𝜌 ) − (𝑣𝜌 |𝜌), (1.7)

where 𝐹HK : A 𝑁 → R and 𝐸 : V𝑁 → R. Combining this definition of 𝐹HK with the

subgradient inequality of Eq. (1.5) expressed as 𝐸 (𝑣𝜌 ) − (𝑣𝜌 |𝜌) ≥ 𝐸 (𝑣) − (𝑣|𝜌), we
obtain
𝐹HK (𝜌) ≥ 𝐸 (𝑣) − (𝑣|𝜌), (1.8)
valid for every 𝜌 ∈ A 𝑁 and for every 𝑣 ∈ V𝑁 . From this inequality, the Hohenberg–
Kohn and Lieb variation principles, respectively, follow directly:

𝐸 (𝑣) = min (𝐹HK (𝜌) + (𝑣|𝜌)) , ∀𝑣 ∈ V𝑁 , (1.9)

𝜌∈A 𝑁

𝐹HK (𝜌) = max (𝐸 (𝑣) − (𝑣|𝜌)) , ∀𝜌 ∈ A 𝑁 . (1.10)

𝑣 ∈V𝑁

We may thus calculate the ground-state energy from the universal density functional
by the Hohenberg–Kohn variation principle [2]; conversely, the universal density
functional is obtained from the ground-state energy by the Lieb variation principle.
Together, these variation principles highlight a symmetry or duality between poten-
tials and ground-state densities and also between the ground-state energy and the
universal density functional. However, this duality was not emphasized in the work
by Hohenberg and Kohn, who considered only the first variation principle Eq. (1.9).
Lieb was the first to study both variation principles rigorously, within the framework
of convex analysis [1].
We note the 𝑣-representability problem of Hohenberg–Kohn theory and the vari-
ation principles established above: the sets A 𝑁 and V𝑁 do not form vector spaces
and are not explicitly known. Also, we have no optimality conditions except by the
Hohenberg–Kohn mapping 𝜌 ↦→ 𝑣𝜌 . These restrictions will be lifted with the devel-
opment of the constrained-search formalism of Levy [3] and the convex formulation
by Lieb [1].

1.1.4 Vector spaces of densities and potentials

To proceed, we will need to work with vector spaces that contain all densities and all
potentials of interest to us. Lieb introduced the complete vector spaces

𝑋 = 𝐿 3 (R3 ) ∩ 𝐿 1 (R3 ), 𝑋 ∗ = 𝐿 3/2 (R3 ) + 𝐿 ∞ (R3 ) (1.11)

6 T. Helgaker and A. M. Teale

where the space of potentials 𝑋 ∗ is dual to the space of densities 𝑋, thereby ensuring
that all interactions (𝑣|𝜌) with 𝜌 ∈ 𝑋 and 𝑣 ∈ 𝑋 ∗ are finite. While 𝑋 ∗ is sufficiently
large to include all Coulomb potentials, 𝑋 contains all normalized densities arising
from 𝑁-electron states of a finite kinetic energy – such densities are said to be 𝑁-
representable. The set of 𝑁-representable densities I𝑁 can be characterized in a
simple manner,
∫
I𝑁 = {𝜌 ∈ 𝑋 | 𝜌 ≥ 0, 𝜌(r) dr = 𝑁, 𝑇W (𝜌) < +∞}, (1.12)

where 𝑇W : I𝑁 → [0, +∞] is the von Weizsäcker kinetic energy functional:

∫
𝑇W (𝜌) = 2 |∇𝜌 1/2 (r)| 2 dr.
1
(1.13)

It can be shown that I𝑁 is a convex set and that 𝑇W is a convex function, meaning
that, for each pair of different densities 𝜌1 , 𝜌2 ∈ I𝑁 and every 𝜆 ∈ (0, 1), the convex
combination 𝜆𝜌1 + (1 − 𝜆) 𝜌2 also belongs to I𝑁 and that 𝑇W satisfies the convexity
characterization of a convex function:

𝑇W (𝜆𝜌1 + (1 − 𝜆) 𝜌2 ) ≤ 𝜆𝑇W (𝜌1 ) + (1 − 𝜆)𝑇W (𝜌2 ). (1.14)

We note that A 𝑁 ( I𝑁 ( 𝑋.
Even though 𝑋 ∗ contains all Coulomb potentials, it does not contain all potentials
that may be of interest to us – harmonic potentials, for example, support an elec-
tronic ground state but are not contained in 𝑋 ∗ . Indeed, this limitation of the theory
was already indicated in the title of Lieb’s 1983 paper [1]: ‘Density Functionals for
Coulomb Systems’. The restriction is fairly mild, however, and does not reduce the
usefulness of the theory significantly. In keeping with this limitation of the theory, we
redefine V𝑁 to be the set of potentials in 𝑋 ∗ that support a ground state: V𝑁 ( 𝑋 ∗ .

1.1.5 Levy–Lieb constrained-search functional

While Hohenberg–Kohn theory is based on the Schrödinger equation, the constrained-
search theory is based on the Rayleigh–Ritz variation principle for the ground-state
energy 𝐸 : 𝑋 ∗ → R, which for pure states may be written as:

𝐸 (𝑣) = inf hΨ|𝐻 (𝑣)|Ψi. (1.15)

Ψ↦→ 𝑁

The ground-state energy 𝐸 (𝑣) is well defined (finite) for each 𝑣 ∈ 𝑋 ∗ provided the
minimization is over all antisymmetric 𝑁-electron wave functions of a finite kinetic
energy, as indicated by the notation Ψ ↦→ 𝑁. We note, however, that the infimum is not
always achieved – for the oxygen atom, for example, a minimum is obtained for atoms
containing up to nine electrons, but no minimum exists for more than nine electrons.
 Lieb variation principle in density-functional theory 7

In other words, the potential 𝑣 𝑍 (r) = −𝑍𝑒 2 /(4𝜋𝜀 0 𝑟) with 𝑍 = 8 belongs to V𝑁 when
𝑁 ≤ 9 but not when 𝑁 > 9. Nevertheless, 𝐸 (𝑣) is given by Eq. (1.15) is well defined
in all cases and we use the term ‘ground-state energy’ even when the infimum is not
achieved and no ground state exists.
Following the seminal work of Levy [3], we now rewrite the Rayleigh–Ritz vari-
ation principle of Eq. (1.15) in the manner

𝐸 (𝑣) = inf inf hΨ|𝐻 (𝑣)|Ψi, (1.16)

𝜌∈I𝑁 Ψ↦→𝜌

where the short-hand notation Ψ ↦→ 𝜌 indicates that the minimization is restricted to

wave functions with density 𝜌. Introducing the Levy–Lieb constrained-search func-
tional 𝐹LL : I𝑁 → [0, +∞) on the set of 𝑁-representable densities by

𝐹LL (𝜌) = inf hΨ|𝐻 (0)|Ψi, (1.17)

Ψ↦→𝜌

where 𝐻 (0) = 𝑇 + 𝑊, we arrive at the following Hohenberg–Kohn variation principle

for 𝑣 ∈ 𝑋 ∗ [3]:
𝐸 (𝑣) = inf (𝐹LL (𝜌) + (𝑣|𝜌)) . (1.18)
𝜌∈I𝑁

Since I𝑁 and 𝑋∗are explicitly known, the 𝑣-representability problem of Hohenberg–

Kohn theory has been solved.

1.1.6 Discontinuity of the universal density functional

Although the constrained-search theory solves the 𝑣-representability problem, the
question regarding the optimality conditions in the Hohenberg–Kohn variation prin-
ciple remains. Since 𝐹LL is defined on the vector space 𝑋, it is tempting to assume
that we can use the Euler equations for this purpose, since we no longer need to worry
about whether or not the functions in the immediate neighbourhood of optimal func-
tions are 𝑣-representable:

𝛿𝐹LL (𝜌)
= −𝑣(r) + 𝑐, 𝑐 ∈ R. (1.19)
𝛿𝜌(r)

However, use of the Euler equations presumes that the density functional is differ-
entiable. Unfortunately, differentiablity is precluded by the fact that the universal
density-functional is discontinuous [5].
To illustrate, consider a one-electron system, for which the universal density func-
tional has an explicit form, being equal to the von Weizsäcker functional of Eq. (1.13).
A one-electron Gaussian density of unit exponent has a finite kinetic energy:

𝜌(r) = 𝜋 −3/2 exp(−𝑟 2 ), 𝑇W (𝜌) = 3/4𝐸 h . (1.20)

8 T. Helgaker and A. M. Teale

Figure 1.2. A convergent sequence of densities with a divergent von Weizsäcker kinetic energy.

Let now {𝜌 𝑛 }∞
𝑛=1 be a sequence that approaches 𝜌 in the norm of 𝑋:

lim k 𝜌 − 𝜌 𝑛 k 𝑋 = 0, (1.21)
𝑛→+∞

while developing increasingly rapid oscillations of increasingly small amplitude, as

illustrated in Figure 1.2. The kinetic energy 𝑇W (𝜌 𝑛 ) is then driven arbitrarily high in
the sequence, implying that 𝑇W is not continuous:
 
lim 𝑇W (𝜌 𝑛 ) = +∞ ≠ 𝑇W lim 𝜌 𝑛 = 3/4𝐸 h . (1.22)
𝑛→∞ 𝑛→∞

Since it can be shown that 𝑇W ≤ 𝐹LL for each 𝑁 ≥ 1, this result also means that 𝐹LL
is discontinuous and hence not differentiable. Consequently, the Euler equations in
Eq. (1.19) are not well defined. We will return to the problem of identifying minim-
izers in the Hohenberg–Kohn variation principle later, when the proper tools have
been developed.

1.1.7 Rayleigh–Ritz variation principle for ensembles

Our next step is to generalize the constrained-search theory to ensembles, taking as
our starting point the canonical-ensemble Rayleigh–Ritz variation principle:

𝐸 (𝑣) = inf tr (𝛾𝐻 (𝑣)) . (1.23)

𝛾→𝑁

Here each density matrix 𝛾 is a normalized self-adjoint trace-class operator on the

𝑁-electron Hilbert space, which may be given the spectral decomposition
∞
∑︁ ∞
∑︁
𝛾= 𝜆𝑖 |Ψ𝑖 ihΨ𝑖 |, 𝜆𝑖 ≥ 0, 𝜆𝑖 = 1, (1.24)
𝑖=1 𝑖=1

in terms of an orthonormal basis {Ψ𝑖 }∞

𝑖=1 . The expectation value in Eq. (1.23) may
then be written as
∞
∑︁
tr (𝛾𝐻 (𝑣)) = 𝜆𝑖 hΨ𝑖 |𝐻 (𝑣)|Ψ𝑖 i. (1.25)
𝑖=1
 Lieb variation principle in density-functional theory 9

Each 𝑀-degenerate ensemble ground-state density matrix obtained from Eq. (1.23)
using 𝑣 ∈ V𝑁 is a (finite) convex combination of 𝑀 pure ground-state density matrices,
𝑀
∑︁ 𝑀
∑︁
𝛾0 = 𝜆0𝑖 |Ψ0𝑖 ihΨ0𝑖 |, 𝜆0𝑖 ≥ 0, 𝜆0𝑖 = 1, (1.26)
𝑖=1 𝑖=1

where each Ψ0𝑖 is a ground-state wave function obtained from Eq. (1.15) with the
same potential 𝑣. By the same token, each ensemble ground-state density is a convex
combination of pure ground-state densities with the same potential. Generalizing the
concept of 𝑣-representability to ensembles, we say that a density 𝜌 ∈ I𝑁 is ensemble
𝑣-representable if it is the ensemble ground-state density for some external potential
𝑣 ∈ V𝑁 . The set of all ensemble 𝑣-representable densities is denoted by B 𝑁 . Clearly,
A 𝑁 ( B𝑁 .
It should be emphasized that the Rayleigh–Ritz variation principle for pure states
and the Rayleigh–Ritz variation principle for ensembles give the same ground-state
energy for every 𝑣 ∈ 𝑋 ∗ – the only difference is that more solutions (minimizing dens-
ity matrices) are obtained with the ensemble variation principle.

1.1.8 Lieb constrained-search functional

Proceeding as for pure states, we obtain from the ensemble Rayleigh–Ritz variation
principle in Eq. (1.23) the ensemble Hohenberg–Kohn variation principle

𝐸 (𝑣) = inf (𝐹DM (𝜌) + (𝑣|𝜌)) , (1.27)

𝜌∈I𝑁

where the Lieb density-matrix constrained-search functional [1] is given by

𝐹DM (𝜌) = inf tr(𝛾𝐻 (0)). (1.28)

𝛾↦→𝜌

A nontrivial result due to Lieb [1] is that the infimum is always achieved,

𝐹DM (𝜌) = min tr(𝛾𝐻 (0)) = tr(𝛾𝜌 𝐻 (0)), (1.29)

𝛾↦→𝜌

with important implications for the theory. A similar result holds for 𝐹LL .
Since the constrained search over ensembles in 𝐹DM includes the constrained
search over pure states in 𝐹LL , we conclude that

𝐹DM ≤ 𝐹LL . (1.30)

In fact, 𝐹DM ≠ 𝐹LL , since there exist densities 𝜌 ∈ I𝑁 for which 𝐹DM (𝜌) < 𝐹LL (𝜌).
Nevertheless, both functionals are admissible, meaning that both give the correct
ground-state energy for every potential in the Hohenberg–Kohn variation principle.
10 T. Helgaker and A. M. Teale

They differ only in that 𝐹DM gives more minimizing solutions – specifically, each
minimizing density obtained with 𝐹DM is a convex combination of those obtained
with 𝐹LL . The reason for this behaviour is that 𝐹DM is the convex hull of 𝐹LL – that
is, its greatest convex lower bound [1].
Convexity is a powerful property of 𝐹DM , with far-reaching consequences for the
theory. The proof is simple. Let 𝜌1 , 𝜌2 ∈ I𝑁 be different and let 𝜆 ∈ (0, 1). There then
exist by Eq. (1.29) minimizing density matrices 𝛾1 ↦→ 𝜌1 and 𝛾2 ↦→ 𝜌2 in 𝐹DM such
that 𝐹DM (𝜌1 ) = tr(𝛾1 𝐻 (0)) and 𝐹DM (𝜌2 ) = tr(𝛾2 𝐻 (0)). By the definition of 𝐹DM , we
then have

𝐹DM (𝜆𝜌1 + (1 − 𝜆) 𝜌2 ) = min𝛾→𝜆𝜌1 +(1−𝜆)𝜌2 tr(𝛾𝐻 (0))

≤ tr((𝜆𝛾1 + (1 − 𝜆)𝛾2 )𝐻 (0)), (1.31)

where the inequality holds since 𝜆𝛾1 + (1 − 𝜆)𝛾2 ↦→ 𝜆𝜌1 + (1 − 𝜆) 𝜌2 is an allowed (but
not necessarily minimizing) density matrix in the constrained search. Reintroducing
𝐹DM on the right-hand side of Eq. (1.31), we obtain

𝐹DM (𝜆𝜌1 + (1 − 𝜆) 𝜌2 ) ≤ 𝜆𝐹DM (𝜌1 ) + (1 − 𝜆)𝐹DM (𝜌2 ). (1.32)

verifying convexity.
Another important property of 𝐹DM is lower semi-continuity [1], meaning that,
for each 𝜌 ∈ 𝑋 and each sequence {𝜌 𝑛 }∞
𝑛=1 converging to 𝜌, we have

lim inf 𝐹DM (𝜌 𝑛 ) ≥ 𝐹DM (𝜌). (1.33)

𝑛→∞

Roughly speaking, lower semi-continuity at 𝜌 means that 𝐹DM may jump up but not
down as we move away from 𝜌. Functions that are both convex and lower semi-
continuous are said to be closed convex, an important, well-behaved class of functions
to which 𝐹DM belongs.

1.1.9 Minimizing densities

A useful property of convex functions is that every local minimum is a global min-
imum, greatly simplifying their minimization. Since 𝐹DM is convex, 𝜌 ↦→ 𝐹DM +
(𝑣|𝜌) is also convex, implying that every local minimizing density in the (ensemble)
Hohenberg–Kohn variation principle is a global minimizer. Moreover, for a given
potential 𝑣 ∈ 𝑋 ∗ , the set of minimizers constitute a convex set, which is nonempty if
and only if 𝑣 ∈ V𝑁 .
Let us consider the relationship of the minimizing densities in the Hohenberg–
Kohn variation principle to the ground-state densities in the Rayleigh–Ritz variation
 Lieb variation principle in density-functional theory 11

principle. Since 𝐹DM (𝜌) + (𝑣|𝜌) = min𝛾↦→𝜌 tr(𝛾𝐻 (𝑣)), we may write
(
inf 𝛾 tr(𝛾𝐻 (𝑣)), Rayleigh–Ritz variation principle,
𝐸 (𝑣) =
inf 𝜌∈I𝑁 min𝛾→𝜌 tr(𝛾𝐻 (𝑣)), Hohenberg–Kohn variation principle,
(1.34)
where the searches in the two variation principles are over the same complete set of
density matrices. Therefore, if 𝑣 ∈ V𝑁 , then the infimum is achieved by the same
density matrices in both cases:

𝐸 (𝑣) = min tr(𝛾𝐻 (𝑣)) = min min tr(𝛾𝐻 (𝑣)), 𝑣 ∈ V𝑁 . (1.35)

𝛾 𝜌∈I𝑁 𝛾→𝜌

It follows that the minimizing densities in the Hohenberg–Kohn variation principle

are precisely the ground-state densities in the Rayleigh–Ritz variation principle. We
note how this result depends critically on the existence of a minimizing density mat-
rix 𝛾 ↦→ 𝜌 in the constrained-search functional in Eq. (1.29) and hence in Eq. (1.34),
precluding the existence of minimizing densities in the Hohenberg–Kohn variation
principle that are not ground-state densities.
In conclusion, the ensemble Hohenberg–Kohn variation principle is faithful in
the sense that it gives not only the same ground-state energy as does the ensemble
Rayleigh–Ritz variation principle, but also the same ground-state densities, when
these exist. A similar result holds for pure states, with the Levy–Lieb functional.

1.1.10 The Lieb functional and variation principle

Let us now develop DFT within the framework of convex analysis. The key to this
development is the concavity of the ground-state energy 𝐸 : 𝑋 ∗ → R (alluded to in
Section 1.1.2), meaning that the interpolation characterization

𝐸 (𝜆𝑣1 + (1 − 𝜆)𝑣2 ) ≥ 𝜆𝐸 (𝑣1 ) + (1 − 𝜆)𝐸 (𝑣2 ) (1.36)

holds for each pair 𝑣1 , 𝑣2 ∈ 𝑋 ∗ and each 𝜆 ∈ (0, 1). To demonstrate concavity, we
first note that 𝐻 (𝜆𝑣1 + (1 − 𝜆)𝑣2 ) = 𝜆𝐻 (𝑣1 ) + (1 − 𝜆)𝐻 (𝑣2 ) holds by the linearity of
𝑣 ↦→ 𝐻 (𝑣). We may therefore rewrite the Rayleigh–Ritz variation principle (here in a
pure-state formulation) as:

𝐸 (𝜆𝑣1 + (1 − 𝜆)𝑣2 ) = inf h Ψ | 𝐻 (𝜆𝑣1 + (1 − 𝜆)𝑣2 ) | Ψ i

Ψ
= inf (𝜆h Ψ | 𝐻 (𝑣1 ) | Ψ i + (1 − 𝜆)h Ψ | 𝐻 (𝑣2 ) | Ψ i) . (1.37)
Ψ

Minimizing the two expectation values separately, we obtain

𝐸 (𝜆𝑣1 + (1 − 𝜆)𝑣2 ) ≥ inf 𝜆h Ψ | 𝐻 (𝑣1 ) | Ψ i + inf (1 − 𝜆)h Ψ | 𝐻 (𝑣2 ) | Ψ i

Ψ Ψ
= 𝜆 inf h Ψ | 𝐻 (𝑣1 ) | Ψ i + (1 − 𝜆) inf h Ψ | 𝐻 (𝑣2 ) | Ψ i, (1.38)
Ψ Ψ
12 T. Helgaker and A. M. Teale

where the last step follows since 𝜆 and 1 − 𝜆 are both nonnegative. Concavity of the
ground-state energy is thus an immediate consequence of the Rayleigh–Ritz variation
principle and of the linearity of the Hamiltonian in the external potential. We note
that the proof of concavity does not carry through for excited states (except for the
lowest state of each symmetry) since the Rayleigh–Ritz minimization is then sub-
ject to orthogonality constraints that depend on the external potential, precluding the
inequality in Eq. (1.38) from being established.
Since the ground-state energy is concave and everywhere finite, it is also continu-
ous and hence closed concave. We may then apply the Fenchel–Moreau biconjuga-
tion theorem of convex analysis to deduce the existence of a closed convex function
𝐹 : 𝑋 → R such that

𝐸 (𝑣) = inf (𝐹 (𝜌) + (𝑣|𝜌)) 𝑣 ∈ 𝑋 ∗, (1.39)

𝜌∈𝑋

𝐹 (𝜌) = sup (𝐸 (𝑣) − (𝑣|𝜌)) 𝜌 ∈ 𝑋, (1.40)

𝑣 ∈𝑋 ∗

where R = [−∞, ∞] denotes the extended real numbers. The function 𝐹 is the Lieb
universal density functional. The Lieb functional 𝐹 and the ground-state energy 𝐸 are
said to be conjugate functions or Fenchel conjugates. Since 𝐹 can be calculated from
𝐸 and vice versa, the two functions contain the same information, only encoded in
different ways – each property of one function is exactly reflected in some property
of its conjugate function. Fenchel conjugation is a generalization of the Legendre
transformation – the relationship between the ground-state energy and the universal
density functional in DFT is thus similar to the relationship between the Hamiltonian
and the Lagrangian in classical mechanics.
Comparing the Hohenberg–Kohn variation principles in Eq. (1.29) and Eq. (1.39),
we note that the variation is over all 𝑋 in Eq. (1.39) but only over I𝑁 in Eq. (1.29). If
𝐹DM is extended from I𝑁 to 𝑋 by setting it equal to +∞ on 𝑋 \ I𝑁 , then the extended
𝐹DM remains closed convex and satisfies the same Hohenberg–Kohn variation prin-
ciple as does the Lieb functional 𝐹 in Eq. (1.29). However, by the Fenchel–Moreau
theorem, there exists only one closed convex function 𝐹 on 𝑋 that is conjugate to
the closed concave function 𝐸 on 𝑋 ∗ , meaning that the Lieb functional in Eq. (1.40)
and the density-matrix constrained-search functional in Eq. (1.29) are the same func-
tional [1]:
𝐹 = 𝐹DM . (1.41)
Furthermore, this functional is the lower bound to all admissible density functionals
and the most well behaved such functional, being closed convex.
In short, DFT is essentially an exercise in convex analysis predicated on the con-
cavity and continuity of the ground-state electronic energy 𝑣 ↦→ 𝐸 (𝑣). The concepts
and tools of convex analysis are therefore well suited to DFT, from a computational
as well as theoretical point of view.
 Lieb variation principle in density-functional theory 13

2 2 2

1 E '(v) 1 1
F(ρ)

-2 -1 1 2 -2 -1 1 2 -2 -1 1 2

-1 -1 -F '(ρ) -1

E(v)

-2 -2 -2

Figure 1.3. A strictly concave function E : R → (−∞, 0] (left) with a strictly convex conjugate
function F →[0, +∞) (right). Their derivatives are plotted in the middle.

1.1.11 Convex conjugation: a one-dimensional differentiable model

To understand the conjugate relationship between the ground-state energy and the
universal density functional better, we will consider two one-dimensional models of
the energy E : R → R and the density functional F : R → R that satisfy the Fenchel
conjugations

E (𝑣) = inf (F (𝜌) + (𝑣|𝜌)), F (𝜌) = sup (E (𝑣) − (𝑣|𝜌)), (1.42)

𝜌∈R 𝑣 ∈R

where the interaction between the potential and density is now the simple scalar
product (𝑣|𝜌) = 𝑣𝜌. We begin by considering in this section the differentiable func-
tions plotted in Figure 1.3. In agreement with the true ground-state energy, we assume
that E ≤ 0 with E (0) = 0. In addition, we assume that E is twice differentiable with
E 00 < 0 and lim𝑣→±∞ E 0 (𝑣) = ∓∞. Strict concavity cannot be correct since it implies
that E (𝑣) < 0 for repulsive potentials 𝑣 > 0, but we accept this unphysical behaviour
for the time being in order to explore the consequences of strict concavity.
Consider now the condition for a maximizing potential in the Lieb variation prin-
ciple in Eq. (1.42). By the derivative assumptions on E, a maximizing potential 𝑣 ∈ R
exists for every 𝜌 ∈ R as the solution to the stationarity condition 𝜌 = E 0 (𝑣) – in this
case, even for the nonphysical negative densities. Since E 0 is strictly decreasing, this
mapping from potentials to densities can be inverted to give the Hohenberg–Kohn
mapping from densities to potentials, which we recognize as the stationary condition
in the Hohenberg–Kohn variation principle: 𝑣 = −F 0 (𝜌). We thus have two equivalent
optimality conditions:

E 0 (𝑣) = 𝜌 ⇐⇒ −F 0 (𝜌) = 𝑣. (1.43)

The ground-state energy and the universal density functional are therefore functions
whose derivatives (to within a negative sign) are each other’s inverse functions. These
14 T. Helgaker and A. M. Teale

2 2 2

1 E '(v) 1 1
F(ρ)

-2 -1 1 2 -2 -1 1 2 -2 -1 1 2

-1 -1 -F '(ρ) -1

E(v)

-2 -2 -2

Figure 1.4. A concave function E : R → (−∞, 0] (left) and its convex conjugate function
F : R → [0, +∞] (right) with their derivatives plotted in the middle.

inverse relationships are illustrated in Figure 1.3, which shows how we may generate
F by differentiation of E (left plot), inversion of E 0 to obtain −F 0 (middle plot),
followed by integration of F 0 to yield F (right plot). Although simple, this example
illustrates the essence of DFT.

1.1.12 Convex conjugation: a one-dimensional nondifferentiable model

We have seen how, in a simple model, the Hohenberg–Kohn mapping follows from
strict concavity of the ground-state energy. The true energy is not strictly concave,
however, only concave. For example, when a scalar 𝜇 is added to the potential, the
energy changes linearly, in the manner 𝜇 ↦→ 𝐸 (𝑣 + 𝜇) = 𝐸 (𝑣) + 𝜇𝑁 where 𝑁 is the
number of electrons. Also, we have 𝐸 (𝑣) = 0 for all repulsive potentials 𝑣 ≥ 0, thereby
violating strict concavity. There may be more ways in which strict concavity is viol-
ated, but these two violations can easily be explored by a one-dimensional model.
In Figure 1.4, we have plotted a model energy E : R → (−∞, 0] that is linear on
[−1, −1/2] with slope 3/2 and linear on [0, +∞) with slope 0. Also, E is nondifferen-
tiable at −1 and −1/2.
The effect on the density functional of this modification F : R → R is dramatic.
First, F (𝜌) = +∞ for nonphysical densities 𝜌 < 0, which is a consequence of the fact
that E (𝑣) = 0 for all repulsive potentials 𝑣 > 0:

F (𝜌) = sup (E (𝑣) − (𝑣|𝜌)) ≥ sup (E (𝑣) − (𝑣|𝜌)) = sup (−(𝑣|𝜌)) = +∞. (1.44)
𝑣 ∈R 𝑣>0 𝑣>0

Second, F becomes nondifferentiable at 𝜌 = 0 and 𝜌 = 1/2 – that is, at the negative

slopes of the linear segments of E. From the middle plots in Figure 1.4, we note
that each horizontal segment of 𝑣 ↦→ (𝑣, E 0 (𝑣)) corresponds to a vertical segment of
𝜌 ↦→ (𝜌, F 0 (𝜌)) and vice versa, meaning that each linear segment of 𝑣 ↦→ (𝑣, E (𝑣))
 Lieb variation principle in density-functional theory 15

corresponds to a kink of 𝜌 ↦→ (𝜌, F (𝜌)) and vice versa. Roughly speaking, therefore,
we may associate nondifferentiability of F with nonstrict concavity of E.
Since 𝐸 and 𝐹 are nondifferentiable, we cannot express the optimality conditions
in terms of derivatives, as done in Eq. (1.43). Instead, we introduce for a convex func-
tion 𝑓 : R → R the subdifferential of 𝑓 at 𝑥 by

𝜕 𝑓 (𝑥) = [ 𝑓−0 (𝑥), 𝑓+0 (𝑥)], (1.45)

where 𝑓−0 (𝑥) and 𝑓+0 (𝑥) are the left and right derivatives, respectively, of 𝑓 at 𝑥. Each
𝑚 ∈ 𝜕 𝑓 (𝑥) is then the slope of a supporting line to 𝑓 and 𝑥, known as a subgradient
of 𝑓 at 𝑥. Returning to Figure 1.4, we see that 𝜕F (𝜌) = {F 0 (𝜌)} everywhere except
at the nondifferentiable points, where 𝜕F (0) = (−∞, 0] and 𝜕F (3/2) = [1/2, 1]. At
these points, we have plotted supporting lines to F for illustration. For the ground-
state energy, we have 𝜕E (𝑣) = {E 0 (𝑣)} everywhere except 𝜕E (−1) = [3/2, 2] and
𝜕E (−1/2) = [1, 3/2].
Subgradients allow global minima (maxima) of convex (concave) functions to be
identified in a simple manner: for a convex (concave) function 𝑓 : R → R, a point
𝑥 ∈ R is a global minimizer (maximizer) if and only if 0 ∈ 𝜕 𝑓 (𝑥). The condition for
a minimum in the Hohenberg–Kohn variation principle in Eq. (1.42) is therefore that
the subdifferential of the convex objective function 𝜌 ↦→ F (𝜌) + (𝑣|𝜌) contains zero.
Since 𝜕 (F (𝜌) + (𝑣|𝜌)) = 𝜕F (𝜌) + 𝑣, this condition is equivalent to −𝑣 ∈ 𝜕F (𝜌);
likewise, the condition for a global maximum in the Lieb variation principle becomes
𝜌 ∈ 𝜕E (𝑣). In terms of subgradients, the optimality conditions of the Hohenberg–
Kohn and Lieb variation principles are therefore

− 𝑣 ∈ 𝜕F (𝜌) ⇐⇒ 𝜌 ∈ 𝜕E (𝑣), (1.46)

which reduce to the stationary conditions in Eq. (1.43) when both functions are dif-
ferentiable. The ground-state energy E and the density functional F are therefore
functions whose subdifferential mappings 𝜕E : 𝑋 ∗ ⇒ 𝑋 and 𝜕F : 𝑋 ⇒ 𝑋 ∗ are each
other’s inverse multifunctions, as illustrated in the middle plot of Figure 1.4.

1.1.13 Hohenberg–Kohn and Lieb optimality conditions

To set up the optimality conditions of the Hohenberg–Kohn and Lieb variation prin-
ciples for the exact ground-state energy 𝐸 and universal density functional 𝐹, we must
generalize subdifferentials and subgradients to functions on vector spaces. For the
concave ground-state energy 𝐸 : 𝑋 ∗ → R and the convex density functional 𝐹 : 𝑋 →
R, the subdifferentials at 𝑣 ∈ 𝑋 ∗ and 𝜌 ∈ 𝑋 are, respectively, given by

𝜕𝐸 (𝑣) = {𝜌 ∈ 𝑋 | 𝐸 (˜𝑣) ≤ 𝐸 (𝑣) + (˜𝑣 − 𝑣|𝜌), ∀˜𝑣 ∈ 𝑋 ∗ , 𝐸 (𝑣) ∈ R}, (1.47)

∗
𝜕𝐹 (𝜌) = {𝑣 ∈ 𝑋 | 𝐹 ( 𝜌)
˜ ≥ 𝐹 (𝜌) + (𝑣| 𝜌˜ − 𝜌), ∀ 𝜌˜ ∈ 𝑋, 𝐹 (𝜌) ∈ R}. (1.48)
16 T. Helgaker and A. M. Teale

The subdifferentials 𝜕𝐸 (𝑣) ⊂ 𝑋 and 𝜕𝐹 (𝜌) ⊂ 𝑋 ∗ are convex sets, which may or may
not be empty. In Eq. (1.47), we recognize the subgradient inequality from Eq. (1.5),
where it was defined on V𝑁 ⊂ 𝑋. Also, it is a simple exercise to check that 𝜕𝐹 (𝜌) in
Eq. (1.48) reduces to that of Eq. (1.45) for convex functions on the real axis.
Proceeding as in the one-dimensional case, we find that the optimality conditions
of the Hohenberg–Kohn and Lieb variation principles are given by

𝐸 (𝑣) = 𝐹 (𝜌) + (𝑣|𝜌) ⇐⇒ −𝑣 ∈ 𝜕𝐹 (𝜌) ⇐⇒ 𝜌 ∈ 𝜕𝐸 (𝑣). (1.49)

The subdifferential mappings 𝜕𝐹 (𝜌) : 𝑋 ⇒ 𝑋 ∗ and 𝜕𝐸 (𝑣) : 𝑋 ∗ ⇒ 𝑋 clearly play an

important role in DFT, setting up the mappings between densities and potentials.
It is instructive to consider in detail the subgradient condition on the density func-
tional. Using the constrained-search expression for 𝐹DM = 𝐹 given in Eq. (1.29), we
obtain

−𝑣 ∈ 𝜕𝐹 (𝜌) ⇐⇒ (∀ 𝜌˜ ∈ 𝑋) : 𝐹 ( 𝜌)
˜ ≥ 𝐹 (𝜌) − (𝑣| 𝜌˜ − 𝜌)
⇐⇒ (∀ 𝜌˜ ∈ 𝑋) : 𝐹 ( 𝜌)
˜ + (𝑣| 𝜌)
˜ ≥ 𝐹 (𝜌) + (𝑣|𝜌)
⇐⇒ (∀ 𝜌˜ ∈ 𝑋) : min𝛾↦→𝜌˜ tr (𝛾𝐻 (𝑣)) ≥ tr min𝛾↦→𝜌 (𝛾𝐻 (𝑣))


⇐⇒ inf 𝛾 tr (𝛾𝐻 (𝑣)) ≥ tr 𝛾𝜌 𝐻 (𝑣) , (1.50)

where 𝛾𝜌 is a ground-state density matrix of 𝐻 (𝑣), thereby verifying that the con-
ditions −𝑣 ∈ 𝜕𝐹 (𝜌) ⇐⇒ 𝜌 ∈ 𝜕𝐸 (𝑣) hold if and only if 𝜌 ∈ B 𝑁 is an ensemble
ground-state density of 𝑣 ∈ V𝑁 . In particular, 𝜌 ∉ B 𝑁 if and only if 𝜕𝐹 (𝜌) = ∅.
By a general result of convex analysis, 𝐹 is subdifferentiable on a dense subset of
I𝑁 . The set of ensemble 𝑣-representable densities is therefore dense in the set of 𝑁-
representable densities.
Next, we verify that 𝜕𝐹 (𝜌) is unique up to a scalar. Let 𝜌 be a ground-state density
associated with the external potential 𝑣𝜌 ∈ V𝑁 so that −𝑣𝜌 ∈ 𝜕𝐹 (𝜌). Using Eq. (1.49)
and the expression for the subgradient inequality given in Eq. (1.5), we obtain

−𝑣𝜌 ∈ 𝜕𝐹 (𝜌) ⇐⇒ 𝜌 ∈ 𝜕𝐸 (𝑣𝜌 )

(
𝐸 (𝑢) < 𝐸 (𝑣𝜌 ) + (𝑢 − 𝑣𝜌 |𝜌), ∀𝑢 ∉ 𝑣𝜌 + R,
⇐⇒
𝐸 (𝑢) = 𝐸 (𝑣𝜌 ) + (𝑢 − 𝑣𝜌 |𝜌), ∀𝑢 ∈ 𝑣𝜌 + R,
(
𝐸 (𝑣𝜌 ) > 𝐸 (𝑢) + (𝑣𝜌 − 𝑢|𝜌), ∀𝑢 ∉ 𝑣𝜌 + R,
⇐⇒
𝐸 (𝑣𝜌 ) = 𝐸 (𝑢) + (𝑣𝜌 − 𝑢|𝜌), ∀𝑢 ∈ 𝑣𝜌 + R.
 Lieb variation principle in density-functional theory 17

After this rearrangement of the subgradient inequality, we use Eq. (1.47) to give
(
𝜌 ∉ 𝜕𝐸 (𝑢), ∀𝑢 ∉ 𝑣𝜌 + R,
−𝑣𝜌 ∈ 𝜕𝐹 (𝜌) ⇐⇒
𝜌 ∈ 𝜕𝐸 (𝑢), ∀𝑢 ∈ 𝑣𝜌 + R,
(
−𝑢 ∉ 𝜕𝐹 (𝜌), ∀𝑢 ∉ 𝑣𝜌 + R,
⇐⇒ (1.51)
−𝑢 ∈ 𝜕𝐹 (𝜌), ∀𝑢 ∈ 𝑣𝜌 + R,

in accordance with the Hohenberg–Kohn theorem. Our results from Eqs. (1.50) and (1.51)
are summarized as follows:
(
∅, 𝜌 ∉ B𝑁 ,
𝜕𝐹 (𝜌) = (1.52)
−𝑣𝜌 + R, 𝜌 ∈ B 𝑁 ,
(
∅, 𝑣 ∉ V𝑁 ,
𝜕𝐸 (𝑣) = Í 𝑀 Í𝑀 (1.53)
𝑖=1 𝜆 𝑖 𝜌 𝑣𝑖 𝑖=1 𝜆 𝑖 = 1, 𝜆 𝑖 ≥ 0 , 𝑣 ∈ V𝑁 ,

where 𝑣𝜌 is an external potential such that 𝐻 (𝑣𝜌 ) supports a ground state with ground-
state density 𝜌 ∈ B 𝑁 , whereas {𝜌 𝑣𝑖 }𝑖=1
𝑀 are the 𝑀 degenerate pure ground-state dens-

ities of 𝑣 ∈ V𝑁 . Note that 𝜕𝐸 (𝑣) is empty when 𝑣 does not support a ground state.

1.1.14 Moreau–Yosida regularization

Although nondifferentiability is an unavoidable feature of the density functional con-
jugate to the exact ground-state energy, we may instead work with the modified energy
functional
1
𝐸 𝛾 (𝑣) = 𝐸 (𝑣) − 𝛾k𝑣k 2 (1.54)
2
with 𝛾 > 0. Modified in this manner, the ground-state energy becomes strictly convex
with a differentiable conjugate density functional
 1 
𝐹𝛾 (𝜌) = min 𝐹 ( 𝜌) ˜ + k 𝜌˜ − 𝜌k 2 , (1.55)
𝜌˜ 2𝛾
which is the Moreau–Yosida regularization of the Lieb functional [6]. The smoothing
parameter 𝛾 > 0 can be adjusted, noting that 𝐹𝛾 (𝜌) approaches 𝐹 (𝜌) pointwise from
below as 𝛾 → 0 from above. It is an attractive regularization in that, for each 𝑣, the
true energy 𝐸 (𝑣) can be recovered exactly from 𝐸 𝛾 (𝑣) according to Eq. (1.54).
A slight complication of the Moreau–Yosida regularizaton is that Coulomb poten-
tials are not square-integrable, as required to construct 𝐸 𝛾 . This problem may be
avoided by putting the system in a large box or by truncating the Coulomb potential at
large separations. In any case, Moreau–Yosida regularization shows that, even though
the exact density functional is not differentiable, it is ‘almost’ differentiable, justifying
the construction of differentiable density-functional approximation.
18 T. Helgaker and A. M. Teale

Figure 1.5. The four-way correspondence for orbital-free DFT (left) and DFT in a magnetic
field (right). The black horizontal and vertical arrows represent the relationships between each
functional by bi-conjugation of both variables simultaneously. The solid diagonal arrows indic-
ate skew conjugations of one of the variables independently. The dashed blue and red arrows
indicate the dual relationships between the relevant variables.

1.1.15 Four-way correspondence of density-functional theory

An important aspect of Lieb’s approach to DFT is its flexibility in describing exten-
ded density functional theories. In particular, we consider two cases, orbital-free
DFT and DFT in the presence of magnetic fields. In both of these cases, the energy
becomes a bifunctional with dependence on additional variables. In orbital-free DFT,
the density functional in the context of grand canonical ensemble DFT depends on
both the charge density 𝜌 and the chemical potential 𝜇. For DFT in the presence
of an external magnetic field, two approaches have been suggested: magnetic-field
density-functional theory (BDFT) [7, 8], in which the density functional depends on
the charge density 𝜌 and the external vector potential associated with the magnetic
field A; and current-density functional theory (CDFT) [9, 10], in which the density
functional depends on the charge density 𝜌 and the paramagnetic component of the
current density jp . A useful concept for these extended theories is the four-way corres-
pondence of DFT, which helps to identify the relationships between the bifunctionals
in each of these approaches.
Figure 1.5 shows the four-way correspondence for orbital-free DFT in the left
panel and for DFT in a magnetic field in the right panel. In orbital-free DFT, the
energy E (𝑣, 𝑁) is a functional of the external potential 𝑣 and the particle number 𝑁.
The corresponding density functional depends on the charge density 𝜌 and the chem-
ical potential 𝜇. This highlights the nature of the optimization problem in orbital-free
DFT as a convex–concave saddle function. In Ref. 11, an optimization scheme was
developed explicitly accounting for the saddle nature of the optimization, resulting in
more rapid convergence for finite systems with all electrons included.
 Lieb variation principle in density-functional theory 19

For DFT in the presence of magnetic fields, the four-way correspondence in Fig-
ure 1.5 clarifies the connections between the functionals involved in BDFT and CDFT.
In particular, the variation principle used in BDFT corresponds to the partial conjug-
ation 𝜌 → 𝑢 from F to E, whilst in CDFT the variation principle corresponds to the
full conjugation (𝜌, jp ) → (𝑢, A) from G to E – see Ref. 12 for a more detailed dis-
cussion of these functionals. The full mathematical characterization of CDFT using
convex analysis similar to Lieb’s original formulation for standard DFT has recently
been completed [13, 14].
In both of these cases, the extended density functional theories can be described
using the techniques of convex analysis as in Lieb’s pioneering 1983 paper [1]. In
doing so, key features of the mathematical structure of the approaches and their inter-
relations are clearly revealed.

1.2 Applications of the Lieb variation principle

The work of Lieb to establish the variation principle of Eq. (1.40) has led not only
to a clear and rigorous framework for understanding the concepts underpinning DFT,
but also to a wide range of practical applications. In this section, we review various
applications of the Lieb variation principle, demonstrating how it may be utilized to
calculate the Kohn–Sham potential of atoms and molecules, to study the adiabatic
connection and calculate the exchange–correlation functional from high-precision
many-body methods, and to study the exchange–correlation hole and energy dens-
ities of atoms and molecules.
We here pragmatically treat the density functional as differentiable, both as a func-
tion of the density and of the interaction strength. We justify this practice by noting
that all approximate density functionals are taken to be differentiable and that the
exact density functional is ‘almost differentiable’ in the sense that it may be approx-
imated to any accuracy by a differentiable Moreau–Yosida regularized functional [6].

1.2.1 Studying the Kohn-Sham noninteracting system

Consider a simple generalization of the the Hamiltonian in Eq (1.1) to include an
interaction-strength dependent two-electron interaction,
𝑁
∑︁
𝐻𝜆 (𝑣) = 𝑇 + 𝑊𝜆 + 𝑣(r𝑖 ) (1.56)
𝑖=1
20 T. Helgaker and A. M. Teale

where the interaction strength between two electrons 𝑖 and 𝑗 is modulated by a para-
meter 𝜆 ∈ R,
∑︁
𝑊𝜆 = 𝑤𝜆 (𝑟 𝑖 𝑗 ), 𝑤0 (𝑟 𝑖 𝑗 ) = 0, 𝑤1 (𝑟 𝑖 𝑗 ) = 1/𝑟 𝑖 𝑗 , (1.57)
𝑖< 𝑗

here chosen such that 𝜆 = 0 corresponds to a noninteracting system of electrons and

𝜆 = 1 to the Hamiltonian for the usual physical system. The ground-state energy
associated with this Hamiltonian is 𝐸 𝜆 and the corresponding interaction-strength
dependent Lieb variation principle [1] is

𝐹𝜆 (𝜌) = sup (𝐸 𝜆 (𝑣) − (𝑣|𝜌)) (1.58)

𝑣 ∈𝑋 ∗

This interaction-strength dependent functional provides a powerful tool to study elec-

tronic systems at any interaction strength. In particular, in Kohn–Sham theory, we are
often concerned with a noninteracting system of electrons (𝜆 = 0) having the same
density as that of the physical system (𝜆 = 1). The properties of this system can be
readily studied using Eq. (1.58) simply by setting 𝜆 = 0 and using an accurate density
for the physical system as input to 𝐹0 (𝜌).
Savin and Colonna [15, 16] presented the first practical calculations of the Lieb
variation principle in finite basis sets for atomic systems. In their approach, the optim-
ization of the external potential 𝑣 was carried out directly using the derivative-free
Nelder–Mead method. To facilitate this optimization, the external potential 𝑣ext (r)
was expressed in terms of a basis-set expansion with an additional 𝐶/𝑟 term, whose
constant 𝐶 was determined to account for the asymptotic form of the potential with
increasing distance 𝑟 from the nucleus. Following this work, Wu and Yang [17]
employed a similar expansion for the external potential,
∑︁
𝑣(r) = 𝑣ext (r) + 𝑣ref (r) + 𝑏 𝑡 𝑔𝑡 (r) (1.59)
𝑡

where 𝑣ext (r) is the external potential due to the nuclei as usually evaluated, 𝑣ref (r) is
a choice of reference potential selected to capture the long-range asymptotic decay of
the potential, and the final term is a linear combination of Gaussian basis functions
𝑔𝑡 (r) with coefficients 𝑏 𝑡 that are to be determined using the Lieb variation prin-
ciple of Eq. (1.58). This expansion is convenient for implementation in finite-basis
molecular codes and allows for direct optimization with respect to 𝑏 𝑡 using derivative-
based (quasi-)Newton approaches. Calculations at the 𝜆 = 0 limit were presented
in Ref. 17, with significantly accelerated convergence relative to the Nelder–Mead
approach of Ref. 15. Many choices of reference potential have been considered for
use in Eq. (1.59). Commonly, the Fermi–Amaldi potential [18] has been used. How-
ever, since this potential is not size-consistent, alternatives such as the Slater [19],
 Lieb variation principle in density-functional theory 21

Krieger–Li–Iafrate [20], and localized-Hartree–Fock [21] exchange potentials have

also been utilized as size-consistent alternatives.
The noninteracting case of Eq. (1.58) is somewhat computationally simpler than
its (partially) interacting counterparts. To see this, consider the 𝜆 = 0 limit of 𝐻𝜆 (𝑣) in
Eq. (1.56); in this limit, the electronic Schrödinger equation becomes separable into a
set of one-electron equations
 1 2 
− 2 ∇𝑖 + 𝑣(r) 𝜑𝑖 (r) = 𝜀𝑖 𝜑𝑖 (r). (1.60)

which may be solved by diagonalization to give the canonical molecular orbitals 𝜑𝑖 (r)
and orbital energies 𝜀𝑖 = 𝜑𝑖 | − 12 ∇𝑖2 + 𝑣(r)|𝜑𝑖 . The total noninteracting energy is
then simply 𝐸 0 (𝑣) = 𝑖𝑛occ 𝜀𝑖 with the associated noninteracting wave function Φ0 =
Í
det 𝜑1 . . . 𝜑 𝑛occ. . Substitution of this expression into Eq. (1.58) for a 𝑣-representable
density 𝜌 leads to
∑︁𝑛occ 
𝐹0 (𝜌) = max∗ 𝜑𝑖 − 21 ∇𝑖2 𝜑𝑖 − (𝑣|𝜌Φ0 − 𝜌) , (1.61)
𝑣 ∈𝑋 𝑖=1

where 𝜌Φ0 is the density associated with Φ0 . The computational advantage of this
form for 𝐹0 (𝜌) is clear since no two-electron integrals are required and no choice of
(partially) interacting wave-function ansatz is required. Instead, all operations amount
to evaluating matrix elements of one-electron operators and solving Eq. (1.60) with
the potential of Eq. (1.59) at each step of the optimization procedure used to evaluate
Eq. (1.61). At convergence, 𝜌Φ0 = 𝜌 and the Lieb functional becomes the noninter-
acting Kohn–Sham kinetic energy, 𝐹0 (𝜌) = 𝑇s (𝜌). Also, for 𝜆 = 0, the derivative
𝜕 2 𝐹0 (𝜌)/𝜕𝑏 𝑡 𝜕𝑏 𝑢 takes a simple form and so second-order optimization schemes such
as the trust-region Newton method can be readily employed, further accelerating con-
vergence.
Given a sufficiently accurate input density 𝜌, which may be obtained from wave-
function-based quantum-chemical methods, Eq. (1.61) yields accurate estimates of
the Kohn–Sham noninteracting kinetic energy 𝑇s (𝜌), the Kohn–Sham orbitals 𝜑𝑖 and
eigenvalues 𝜀𝑖 , and the Kohn–Sham effective potential 𝑣s (r), which may be identified
with the potential of Eq. (1.59) at convergence. However, two complications arise for
finite (orbital) basis-set calculations.
Firstly, as pointed out by Harriman [22, 23], in a finite orbital basis, there is no
Hohenberg–Kohn theorem. As a result, the optimizing potential in Eq. (1.58) may not
be unique and strict convexity of 𝐹𝜆 (𝜌) cannot be assumed. In a complete orbital
basis, the potential is determined up to a constant, leading to a well-defined shape
of the Kohn–Sham potential from Eq. (1.61), for example, and the constant can then
be chosen arbitrarily, usually so that the potentials vanish asymptotically. In finite
basis sets, the extra nonuniqueness of the potential manifests itself in a near singular
behaviour of 𝜕 2 𝐹0 (𝜌)/𝜕𝑏 𝑡 𝜕𝑏 𝑢 and convergence to oscillatory potentials.
22 T. Helgaker and A. M. Teale

Secondly, in finite basis sets, an input density arising from a correlated wave func-
tion may not be exactly represented by the density of a single Slater determinant in the
same basis set; see Refs. 22, 23. In practical calculations, this may manifest itself as
|𝜌Φ0 (r) − 𝜌(r)| > 0 for some points r. It is therefore necessary to carefully check the
results of calculations to ensure that any residual density differences are sufficiently
small. In practice, we find that, for small atomic and molecular systems, Gaussian
basis sets of augmented triple-zeta quality are adequate to obtain close approxima-
tions to correlated input densities and that this problem is essentially removed for
basis sets of quadruple-zeta quality and beyond.
To avoid oscillatory potentials in calculations in finite basis sets, it is neces-
sary to perform calculations in a regularized manner. Strategies for regularization
include using approaches such as singular-value decomposition or Tikhonov regular-
ization in second-order optimizations to avoid issues with the singular components
of 𝜕 2 𝐹0 (𝜌)/𝜕𝑏 𝑡 𝜕𝑏 𝑢 . An alternative approach, which emphasizes smoothness of the
calculated potentials is the smoothing-norm approach of Ref. 24. In this approach, a
quadratic penalty function is applied to the energy that penalizes oscillatory solutions.
This approach has the advantage that the objective function, gradient, and Hessian
are then consistently modified leading to simple implementation with essentially any
optimization procedure. Interestingly, this approach is closely related to the Moreau–
Yosida regularization of convex analysis; see the discussion in Ref. 6. Finally, we note
that a similar regularization can be achieved by tailoring the basis set chosen for use in
the expansion of Eq. (1.59) – see, for example, Ref. 25. Whilst all of these approaches
are effective in avoiding oscillatory solutions, the results of the calculations must be
carefully assessed to ensure that the density 𝜌Φ0 (r) remains sufficiently close to the
input density 𝜌 and that the associated potentials are not over-regularized.
Figure 1.6 shows the exchange–correlation potentials obtained from the Lieb vari-
ation principle at 𝜆 = 0. The main features of the potential are clearly captured includ-
ing the −1/𝑟 asymptotic decay and inter-shell structure. The inset shows the valence
region, where the potentials reproducing the electron density calculated from Hartree–
Fock theory and coupled-cluster single–doubles–perturbative-triples (CCSD(T)) the-
ory are most different.

1.2.2 Lieb calculations of the adiabatic connection

The adiabatic connection is a central concept in DFT, connecting the noninteract-
ing Kohn–Sham system with the fully interacting physical system. The density-fixed
adiabatic connection, in particular, can be determined by considering how the 𝜆-
interacting Lieb functional may be expressed in terms of its noninteracting limit and
 Lieb variation principle in density-functional theory 23

Figure 1.6. The exchange–correlation potentials obtained by performing the Lieb optimization
at 𝜆 = 0 for the neon atom using the Hartree–Fock density (blue) and CCSD(T) density (green)
in the uncontracted aug-cc-pCVQZ basis set. The inset highlights the differences in the valence
region arising from the treatment of electron correlation. The potentials are determined using
the smoothing-norm approach with a regularization parameter of 10−5 a.u.

a correction term,
∫ 𝜆 ∫ 𝜆
𝐹𝜆 (𝜌) = 𝐹0 (𝜌) + 𝐹𝜈0 (𝜌)d𝜈 = 𝑇s (𝜌) + W𝜈 (𝜌)d𝜈 (1.62)
0 0

where the prime indicates differentiation with respect to interaction strength 𝜈. Express-
ing 𝐹𝜆 (𝜌) in its constrained-search form,
𝜌

𝐹𝜆 (𝜌) = min tr (𝛾𝐻𝜆 (0)) = tr 𝛾𝜆 𝐻𝜆 (0) , (1.63)
𝛾↦→𝜌

differentiating with respect to the interaction strength, and using the Hellmann–Feynman
theorem, we obtain for the linear adiabatic-connection path with 𝑤𝜆 (𝑟 𝑖 𝑗 ) = 𝜆/𝑟 𝑖 𝑗 in
Eq. (1.57) the adiabatic-connection integrand

W𝜆 (𝜌) = tr 𝛾𝜆 𝑊𝜆0 .
𝜌

(1.64)

The Lieb variation principle therefore allows us to calculate the density-fixed adia-
batic connection directly, by performing calculations for 0 ≤ 𝜆 ≤ 1 using Eqs. (1.58)
and (1.59) as discussed in Refs. 15–17, 26–30.
24 T. Helgaker and A. M. Teale

The integrand W𝜈 (𝜌) accounts for all electron–electron interactions. It is custom-

ary to decompose this quantity further into the classical Coulomb energy,
∬
1
𝐽𝜆 (𝜌) = 𝑤𝜆 (𝑟 12 ) 𝜌(r1 ) 𝜌(r2 ) dr1 dr2 , (1.65)
2
the exchange energy
𝜌

𝐸 x,𝜆 (𝜌) = tr 𝛾0 𝑊𝜆 − 𝐽𝜆 (𝜌), (1.66)

and the correlation energy

∫ 𝜆
Wc,𝜈 (𝜌) = tr (𝛾 𝜈 − 𝛾0 )𝑊𝜈0 ,
𝜌 𝜌

𝐸 c,𝜆 (𝜌) = Wc,𝜈 (𝜌) d𝜈, (1.67)
0

so that

𝐹𝜆 (𝜌) = 𝐹0 (𝜌) + 𝐽𝜆 (𝜌) + 𝐸 x,𝜆 (𝜌) + 𝐸 c,𝜆 (𝜌) (1.68)

Since the integrand is the derivative of each component with respect to the interaction
strength, we see that, for the linear adiabatic connection path with 𝑤𝜆 (𝑟 𝑖 𝑗 ) = 𝜆/𝑟 𝑖 𝑗 ,
the contributions from the classical Coulomb energy and the exchange energy are
constants. The shape of the adiabatic connection is therefore entirely determined by
the correlation contribution of Eq. (1.67).
In fact, the shape of the linear-path adiabatic connection can be well understood
from a perturbative analysis at 𝜆 = 0 using Görling–Levy (GL) perturbation the-
ory [31]. The correlation energy can be expanded as
∞
∑︁
(𝑛)
𝐸 c,𝜆 (𝜌) = 𝜆 𝑛 𝐸 GL (𝜌) (1.69)
𝑛=2

(𝑛)
where 𝐸 GL is the 𝑛th-order GL correlation energy, for which explicit expressions
may be readily derived and implemented. In particular, we see that the expansion of
Wc,𝜆 (𝜌) = W𝜆 (𝜌) − 𝐽𝜆 (𝜌) − 𝐸 x,𝜆 (𝜌) is given by
∞
∑︁ ∞
∑︁
(𝑛) (𝑛+1)
Wc,𝜆 (𝜌) = 𝜆 𝑛 WGL (𝜌) = 𝜆 𝑛 (𝑛 + 1)𝐸 GL (𝜌), (1.70)
𝑛=1 𝑛=1

(2)
from which is it clear that the slope of the linear-path adiabatic connection is 2𝐸 GL (𝜌).
For dynamically correlated systems, where low-order perturbative expansions provide
a good approximation to 𝐸 c,𝜆 (𝜌), plotting 𝜆 ↦→ W𝜆 (𝜌) gives smooth, almost linear
curves. For systems with more significant static correlation (i.e., where a single Kohn–
Sham determinant is not close to the physical multi-determinantal wave function), the
corresponding plots display significantly more curvature.
 Lieb variation principle in density-functional theory 25

The Lieb variation principle is therefore a powerful tool to study not only the
Kohn–Sham noninteracting system, but also the entire adiabatic connection, giving
access to all of the associated 𝜆-interacting energies, wave functions and associated
density matrices. To perform practical studies of the adiabatic connection for many-
electron systems in finite basis sets, a key step is to choose an appropriate, sufficiently
accurate, ansatz to determine the input density 𝜌 in Eq. (1.58) and to calculate the
energy 𝐸 𝜆 at each interaction strength. A natural choice for few-electron systems is
full configuration-interaction (FCI) theory, as employed in Refs. 15, 26. However,
the entire repertoire of quantum-chemical methods can be utilized to study larger
systems. Some care is required, however, when setting up optimization procedures to
perform the Lieb maximization in Eq. (1.58) when the ansatz for 𝐸 𝜆 is nonvariational,
such as for Møller–Plesset and coupled-cluster theories. In particular, the required
derivatives are most conveniently evaluated using the Lagrangian densities for these
approaches [32]; see Refs. 26, 27 for further details.

1.2.3 The adiabatic connection for H2

A simple illustrative example, for which FCI calculations can be readily performed, is
the H2 molecule. At short internuclear separations 𝑅, this system has a physical wave
function that is well represented by a single Slater determinant. Correlation effects,
whilst important for chemical accuracy, are relatively subtle and the CI expansion has
one large coefficient on the Hartree–Fock determinant, with minor contributions from
excited determinants – characteristic of dynamic correlation. At large 𝑅, the system
has a FCI wave function with large coefficients on more than one determinant – char-
acteristic of significant static correlation. At the dissociation limit, the physical wave
function would be well represented by just two determinants, one for each hydro-
gen atom. Intermediate bond lengths capture the regime where the both dynamic and
static correlation effects play a role.
The adiabatic-connection curves for 𝜆 ↦→ Wc,𝜆 (𝜌) calculated using the Lieb vari-
ation principle of Eq. (1.58) at the FCI level in the uncontracted aug-cc-pCVQZ basis
(for both the orbital and potential expansions) are presented in Figure 1.7 for 𝑅 = 1.4,
3.0, 5.0 and 7.0 bohr. These plots capture the transition from near the equilbirum geo-
metry to almost dissociated hydrogen atoms and the corresponding transition from
almost purely dynamic to static correlation.
As expected from the perturbative analysis of Eq. (1.70), the plot at equilibrium
geometry displays only subtle curvature – characteristic of low-order dynamical cor-
relation effects. As 𝑅 increases, the plots exhibit more rapid curvature – indicative
of more significant higher-order correlation contributions and the relevance of static
correlation. Finally, at large 𝑅, close to dissociation, we see a characteristic ‘L’ curve,
a feature of (almost) pure static correlation. This shape arises because the restricted
Kohn–Sham single determinant cannot well represent the 𝜆-interacting wave func-
26 T. Helgaker and A. M. Teale

Wλ, c / hartree
λ
0.2 0.4 0.6 0.8 1.0

-0.05

-0.10

-0.15

-0.20

-0.25

Figure 1.7. The linear adiabatic connection for the H2 molecule at internuclear separations
𝑅 = 1.4, 3.0, 5.0 and 7.0 bohr (top to bottom).

tion, which consists of essentially one determinant for each hydrogen atom at this
geometry. Since the hydrogen atoms are widely separated, there is no dynamical cor-
relation between the electrons. As a result, the curve abruptly changes for small 𝜆
values and then is essentially flat, reflecting the fact that, at the dissociation limit
for a restricted Kohn–Sham reference wave function, 𝐸 x (𝜌) = −1/2𝐽 (𝜌). Therefore,
to cancel the electron–electron interactions at the dissociation limit, we must have
𝐸 c (𝜌) = −1/2𝐽 (𝜌). The curve presented at 𝑅 = 7.0 bohr shows that the behaviour
with respect to 𝜆 is already approximately constant, reflecting the linear behaviour of
𝐽 (𝜌) with respect to 𝜆 in Eq. (1.65).

1.2.4 An alternative path: Generalized adiabatic connections

As indicated in Eq. (1.57), the noninteracting and physical systems can be connected
by any number of alternative paths – such adiabatic connections have been called
generalized adiabatic connections [33]. The Lieb variation principle of Eq. (1.58) can
for such connections be applied in the same manner as described in Section 1.2.2
but with 𝐸 𝜆 (𝜌) calculated using 𝑊𝜆 (𝜌) defined with a different modification of the
electron–electron interaction.
In Ref. 28, some common adiabatic paths of relevance to range-separated meth-
ods were examined. In range-separated methods [34], a distinction is made between
 Lieb variation principle in density-functional theory 27

Wλ / hartree

1.0

0.8

0.6

0.4

0.2

λ
0.2 0.4 0.6 0.8 1.0

Figure 1.8. The error-function generalized adiabatic connection for the H2 molecule at inter-
nuclear separations 𝑅 = 1.4, 3.0, 5.0 and 7.0 bohr

long-ranged and short-ranged interactions of electrons. These interactions are typic-

ally partitioned by modifying the electron–electron interaction operator. A common
approach is to use the error function to modulate the electron–electron interactions,

𝜆
𝑤𝜆 (𝑟 𝑖 𝑗 ) = erf 𝜇(𝜆)𝑟 𝑖 𝑗 /𝑟 𝑖 𝑗 , 𝜇(𝜆) = . (1.71)
1−𝜆
For small values of 𝜆, the long-range interactions are captured, with the shorter-range
interactions being emphasized as 𝜆 increases. The value of the Lieb functional varies
between 𝐹0 (𝜌) = 𝑇s (𝜌) and 𝐹1 (𝜌) in a manner that reflects the relative importance
of different ranges of electron–electron interaction. In Figure 1.8, the contributions
become progressively localized to small 𝜆 values as the electron–electron interactions
become predominantly long-ranged at large internuclear separation.
The Lieb variation principle provides insight into the extent to which dynamic and
static correlation can be divided according to the range of interaction. By exploiting
the generalized adiabatic-connection paths in conjunction with knowledge of how the
generalized adiabatic connection behaves for accurate calculations, it may be possible
to devise new approximations to the exchange–correlation energy that more effect-
ively exploit range-separation techniques.
28 T. Helgaker and A. M. Teale

1.2.5 The exchange–correlation hole and energy densities

The exchange–correlation hole is a central concept in DFT and its analysis has given
insight into why commonly used semi-local density-functional approximations can
be so successful – see, for example, Ref. 35. The exchange–correlation energy can be
expressed in terms of the exchange–correlation energy density 𝑤xc,𝜆 as

ℎxc,𝜆 (r, r0) 0

∫ 1∫ ∫
1
𝐸 xc (𝜌) = 𝜌(r)𝑤xc,𝜆 (r) dr d𝜆, 𝑤xc,𝜆 (r) = dr . (1.72)
0 2 |r − r0 |

The exchange–correlation hole at interaction strength 𝜆 is given by

𝑃2,𝜆 (r, r0)

ℎxc,𝜆 (r, r0) = − 𝜌(r0) (1.73)
𝜌(r)

where the pair density 𝑃2,𝜆 (r, r0) is obtained using the Lieb variation principle. At
𝜆 = 0, the exchange–correlation hole reduces to the Fermi (exchange) hole and so the
Coulomb (correlation) hole can be identified as

ℎc,1 (r, r0) = ℎxc,1 (r, r0) − ℎxc,0 (r, r0). (1.74)

In Figure 1.9, the exchange–correlation holes for the H2 molecule at internuclear sep-
arations 𝑅 = 1.4, 3.0, 5.0 and 7.0 bohr are presented, plotted along the bond axis with
the position of the reference electron 0.3 bohr to the left of the second hydrogen atom,
as done in Ref. 35. All required quantities were calculated using the Lieb variation
principle at the FCI level of theory.
At 𝑅 = 1.4 bohr, it is clear that the exchange hole is delocalized with a significant
amplitude on both atoms. The Coulomb hole is negative close to the reference point
and positive close to the other hydrogen nucleus, leading to an exchange–correlation
hole that is more localized around the reference point than either of its constituent
components. As the bond length increases, this effect becomes more and more pro-
nounced, with the total exchange–correlation hole being relatively strongly localized
around the reference point already at 𝑅 = 3.0 bohr, with only relatively modest amp-
litude on the left hydrogen atom. For the longer bond lengths, the localization of the
exchange–correlation hole is essentially complete, with cancellation of the Fermi and
Coulomb holes far from the reference point.
This behaviour of the exchange–correlation hole rationalizes to some extent the
success of popular semi-local exchange–correlation functionals. Semi-local approx-
imations such as generalized-gradient approximations cannot be expected to capture
the strong nonlocality of the Fermi and Coulomb holes. However, when exchange and
correlation are treated together, the overall exchange–correlation hole is more local-
ized and can be effectively modelled by these simple density functionals. For more
detailed discussion of these ideas, see Ref. 35
 Lieb variation principle in density-functional theory 29

hxc (r,r') / a.u. hxc (r,r') / a.u.

0.05 0.05

z / bohr z / bohr
-6 -4 -2 2 4 6 -6 -4 -2 2 4 6

-0.05 -0.05

-0.10 -0.10

-0.15 -0.15

hxc (r,r') / a.u. hxc (r,r') / a.u.

0.05 0.05

z / bohr z / bohr
-6 -4 -2 2 4 6 -6 -4 -2 2 4 6

-0.05 -0.05

-0.10 -0.10

-0.15 -0.15

Figure 1.9. The Fermi (orange, 𝜆 = 0), Coulomb (green, 𝜆 = 1) and exchange–correlation (blue,
𝜆 = 1) holes for H2 at internuclear separations 𝑅 = 1.4 (top left), 3.0 (top right), 5.0 (lower left)
and 7.0 (lower right) bohr plotted along the internuclear axis with the position of the reference
electron chosen to be 0.3 bohr left of the right-hand hydrogen nucleus.

The exchange–correlation hole is a rather complex quantity, depending explicitly

on the coordinates of two electrons. It is therefore desirable to consider quantities
that reduce the complexity, whilst retaining enough information to determine the
exchange–correlation energy. One possibility is to consider the energy density 𝑤xc,𝜆
directly as a target for functional development, since one electronic coordinate has
already been removed. Furthermore, it is possible to interchange the order of integ-
ration in Eq. (1.72) to generate a coupling-constant averaged energy density, thereby
also removing the 𝜆 dependence. This simpler function can then be analysed in a
similar manner to the exchange–correlation potentials generated by approximate func-
tionals – see, for example, Ref. 36. Another alternative is to consider the angle- and
system-averaged exchange–correlation hole, which takes advantage of the isotropic
nature of the Coulomb interaction,
∫ ∫
¯ℎxc (𝑢) = 1 𝜌(r) ℎxc (r, r + u) dΩu dr (1.75)
𝑁 4𝜋
30 T. Helgaker and A. M. Teale

where u = r0 − r and the integration over the solid angle Ωu averages over all directions
for u. The exchange–correlation energy can then be expressed as
∫ ∞
ℎ¯ xc (𝑢)
𝐸 xc = 𝑁 4𝜋𝑢 2 d𝑢 (1.76)
0 2𝑢
Both of these alternatives provide simpler functions that can be targeted in functional
development. Again, the Lieb variation principle provides access to all of the quant-
ities required to compute these functions accurately using ab-initio methods. The
resulting data can then serve as a benchmark for approximate models.

1.2.6 Beyond the ground state

The Gross–Olivera–Kohn (GOK) approach to excitation energies [37–39] in ensemble
DFT can also be studied using the Lieb variation principle [40]. In the GOK approach,
a density matrix of a statistical ensemble is formed as a convex combination of the 𝑀
lowest states of the Hamiltonian,
𝑀
∑︁
𝛾𝜆,𝑤 = 𝑤𝑖 |Ψ𝜆,𝑖 ihΨ𝜆,𝑖 |, (1.77)
𝑖=1
Í
where the ensemble weights 𝑤𝑖 satisfy 0 ≤ 𝑤𝑖+1 ≤ 𝑤𝑖 and 𝑖 𝑤𝑖 = 1. The ground state
is |Ψ𝜆,0 i and |Ψ𝜆,𝑖 ≥1 i are excited states. For an implementation of the Lieb variation
principle for such ensembles, see Ref. 41. Excitation energies can be evaluated as
𝜕𝐸 xc,𝑤
𝐸 1 − 𝐸 0 = (𝜀1 − 𝜀0 ) + . (1.78)
𝜕𝑤 𝜌=𝜌𝑤

Here 𝜀1 − 𝜀 0 is a Kohn–Sham eigenvalue difference, which is corrected by a partial-

derivative term evaluated as
𝜕𝐸 xc,𝑤 𝜕
= (𝐹1,𝑤 (𝜌) − 𝐹0,𝑤 (𝜌)) − 𝐽 (𝜌 𝑤 ). (1.79)
𝜕𝑤 𝜌=𝜌 𝑤 𝜕𝑤 𝜌=𝜌𝑤

In this way, the Lieb variation principle may also offer insights into the behaviour
of the correction term in the GOK approach to excitation energies; see Ref. 41 for
further details.
Other approaches to excited states have been developed using perturbation theory
along the linear and generalized adiabatic connections [42], which were discussed
in Sections 1.2.2 and 1.2.4. In these cases, the Lieb variation principle was used to
determine the ground-state adiabatic connections and then excitation energies estim-
ated by perturbation at each value of the interaction strength. The Lieb variation
principle may provide a useful tool for further study of time-independent approaches
to excitation energies in the future.
 Lieb variation principle in density-functional theory 31

1.3 Conclusions

We have seen how DFT is predicated on the concavity and continuity of the ground-
state energy in the external potential – from this fact, DFT follows by application
of convex analysis. In particular, the universal density functional and the ground-
state energy are conjugate functions, containing the same information but encoded
in different ways. While the ground-state energy can be obtained from the universal
density functional by the Hohenberg–Kohn variation principle, the density functional
can be calculated from the ground-state energy by the Lieb variation principle.
The Lieb variation principle is not only theoretically important, giving insight
into the structure of DFT, but it is also a practical computational tool. It allows us
to calculate, for any 𝑁-representable density, the universal density functional and the
external potential that supports this density (if such a potential exists). The utility of
this approach has been briefly demonstrated here for determining the Kohn–Sham
potential, the adiabatic connection and detailed information on the exchange and cor-
relation holes, to high accuracy using ab-initio many-body wave function approaches.
This tool can give valuable insight into the numerical behaviour of the exact universal
density functional, which may serve as a useful benchmark to guide the development
of approximate models.
It is our view that the beautiful convex formulation of DFT first put forward by
Lieb [1] has not yet received the attention it deserves, as a framework for teaching
DFT and as a tool for further development of DFT – we hope the present overview
goes some way towards rectifying this situation. We also highlighted how Lieb’s
approach can be utilized for extended density-functional theories such those required
for systems in a magnetic field and for orbital-free density functional theory. This
further illustrates how Lieb’s convex formulation of DFT can be used to unify the
presentation of different variants and extensions of DFT and clarify the relationships
between them. Lieb’s convex formulation of DFT continues to illuminate the devel-
opment of state-of-the-art approaches in DFT almost forty years after its publication.

Acknowledgements. We thank Dr. Tom J. P. Irons and Dr. Bang Huynh for useful
discussions in the preparation of this manuscript.

Funding. This work was partially supported by the Research Council of Norway
through its Centres of Excellence scheme, project number 262695. We acknowledge
financial support from the European Research Council under H2020/ERC Consolid-
ator Grant top DFT (Grant No. 772259).
References

[1] E. H. Lieb, Int. J. Quantum Chem. 24, 243 (1983).

[2] P. Hohenberg and W. Kohn, Phys. Rev. 136, B864 (1964).

[3] M. Levy, Proc. Natl. Acad. Sci. U.S.A. 76, 6062 (1979).

[4] J. van Tiel, Convex Analysis: An Introductory Text, Wiley, 1984.

[5] P. E. Lammert, Int. J. Quantum Chem. 107, 1943 (2007).

[6] S. Kvaal, U. Ekstrom, A. M. Teale, and T. Helgaker, J. Chem.Phys. 140, 18A518

(2014).

[7] C. J. Grayce and R. A. Harris, Phys. Rev. A 50, 3089 (1994).

[8] F. R. Salsbury and R. A. Harris, J. Chem. Phys. 107, 7350 (1997).

[9] G. Vignale and M. Rasolt, Phys. Rev. Lett. 59, 2360 (1987).

[10] G. Vignale and M. Rasolt, Phys. Rev. B 37, 10685 (1988).

[11] M. S. Ryley, M. Withnall, T. J. P. Irons, T. Helgaker, and A. M. Teale, J. Phys.

Chem. A 125, 459 (2021).

[12] S. Reimann, A. Borgoo, E. I. Tellgren, A. M. Teale, and T. Helgaker, J. Chem.

Theory Comput. 13, 4089 (2017).

[13] E. I. Tellgren, S. Kvaal, E. Sagvolden, U. Ekstrom, A. M. Teale, and T. Helgaker,

Phys. Rev. A 86, 062506 (2012).

[14] S. Kvaal, A. Laestadius, E. Tellgren, and T. Helgaker, J. Phys. Chem. Lett. 12,
1421 (2021).

[15] F. Colonna and A. Savin, J. Chem.Phys. 110, 2828 (1999).

[16] A. Savin, F. Colonna, and R. Pollet, Int. J. Quantum Chem. 93, 166 (2003).

[17] Q. Wu and W. Yang, J. Chem.Phys. 118, 2498 (2003).

[18] E. Fermi and E. Amaldi, Mem. Accad. Italia 117, 117 (1934).

[19] J. C. Slater, Physical Review 81, 385 (1951).

[20] J. B. Krieger, Y. Li, and G. J. Iafrate, Phys. Lett. A 146, 256 (1990).

[21] F. D. Sala and A. Görling, J. Chem. Phys. 115, 5718 (2001).

34 REFERENCES

[22] J. E. Harriman, Phys. Rev. A 27, 632 (1983).

[23] J. E. Harriman, Phys. Rev. A 34, 29 (1986).

[24] T. Heaton-Burgess, F. A. Bulat, and W. Yang, Phys. Rev. Lett. 98, 256401
(2007).

[25] A. Heßelmann, A. W. Götz, F. Della Sala, and A. Görling, J. Chem. Phys. 127,
054102 (2007).

[26] A. M. Teale, S. Coriani, and T. Helgaker, J. Chem.Phys. 130, 104111 (2009).

[27] A. M. Teale, S. Coriani, and T. Helgaker, J. Chem.Phys. 132, 164115 (2010).

[28] A. M. Teale, S. Coriani, and T. Helgaker, J. Chem.Phys. 133, 164112 (2010).

[29] M. D. Stromsheim, N. Kumar, S. Coriani, E. Sagvolden, A. M. Teale, and T. Hel-

gaker, J. Chem. Phys. 135, 194109 (2011).

[30] A. M. Teale, S. Coriani, and T. Helgaker, Range-dependent Adiabatic Connec-

tions, in International conference of computational methods in sciences and
engineering 2009 (ICCMSE 2009), edited by T. Simos and G. Maroulis, volume
1504 of AIP Conference Proceedings, pp. 92–99, European Soc Computat
Methods Sci, Engn & Technol, 2012, 7th International Conference on Com-
putational Methods in Science and Engineering (ICCMSE), Rhodes, GREECE,
SEP 29-OCT 04, 2009.

[31] A. Görling and M. Levy, Phys.Rev. A 50, 196 (1994).

[32] T. Helgaker and P. Jørgensen, Theor. Chim. Acta 75, 111 (1989).

[33] W. Yang, J. Chem. Phys. 109, 10107 (1998).

[34] A. Savin, Density Functionals for Coulomb Systems, in Recent Developments

and Applications of Modern Density Functional Theory, p. 327, Elsevier, Ams-
terdam, 1996.

[35] E. J. Baerends and O. V. Gritsenko, J. Phys. Chem. A 101, 5383 (1997).

[36] T. J. P. Irons and A. M. Teale, Mol. Phys. 114, 484 (2016).

[37] E. K. U. Gross, L. N. Oliveira, , and W. Kohn, Phys. Rev. A 37, 2805 (1988).

[38] L. N. Oliveira, E. K. U. Gross, and W. Kohn, Phys. Rev. A 37, 2821 (1988).

[39] E. K. U. Gross, L. N. Oliveira, , and W. Kohn, Phys. Rev. A 37, 2809 (1988).

[40] E. H. Lieb, Density Functionals for Coulomb Systems, in Nato ASI Series vol.
123, edited by R. M. Dreizler, pp. 31–80, Plenum, New York, 1985.
 REFERENCES 35

[41] A. Borgoo, A. M. Teale, and T. Helgaker, Excitation Energies from Ensemble

DFT, in International conference of computational methods in sciences and
engineering 2015 (ICCMSE 2015), edited by T. Simos, Z. Kalogiratou, and
T. Monovasilis, volume 1702 of AIP Conference Proceedings, p. 090049, 2015,
International Conference of Computational Methods in Sciences and Engineer-
ing (ICCMSE), Athens, GREECE, MAR 20-23, 2015.

[42] E. Rebolini, J. Toulouse, A. M. Teale, T. Helgaker, and A. Savin, Mol. Phys.

113, 1740 (2015).