Understanding Density Functional Theory (DFT) and Completing It in Practice
Understanding Density Functional Theory (DFT) and Completing It in Practice
CrossMark
View Export
Online Citation
We review some salient points in the derivation of density functional theory (DFT)
and of the local density approximation (LDA) of it. We then articulate an under-
standing of DFT and LDA that seems to be ignored in the literature. We note
the well-established failures of many DFT and LDA calculations to reproduce the
measured energy gaps of finite systems and band gaps of semiconductors and insula-
tors. We then illustrate significant differences between the results from self consistent
calculations using single trial basis sets and those from computations following the
Bagayoko, Zhao, and Williams (BZW) method, as enhanced by Ekuma and Franklin
(BZW-EF). Unlike the former, the latter calculations verifiably attain the absolute
minima of the occupied energies, as required by DFT. These minima are one of
the reasons for the agreement between their results and corresponding, experimental
ones for the band gap and a host of other properties. Further, we note predictions of
DFT BZW-EF calculations that have been confirmed by experiment. Our subsequent
description of the BZW-EF method ends with the application of the Rayleigh theorem
I. INTRODUCTION
We begin by following the works of Hohenberg and Kohn1 and of Kohn and Sham2 to review
the derivation of density functional theory (DFT) and of the local density approximation (LDA)
of it. The necessity for an overview of this derivation becomes apparent in the last paragraph
of Section II below, where we summarize necessary conditions for the results of ab initio, self
consistent calculations to represent the physical content of DFT or of LDA. Then, in Section III, we
illustrate the “band gap problem,” i.e., the woeful underestimation of measured band gaps, by 30 to
50% or worse, by most DFT and LDA calculations. The tabular presentation of some 99 different
calculations with results significantly lower than corresponding, experimental ones also shows DFT
BZW and BZW-EF calculations whose findings are in agreement with the corresponding, measured
ones. In Section IV, we provide a summary description of the BZW and BZW-EF method. This
description straightforwardly leads to the two key features of the method that guarantee its faithful
rendition of the physical content of DFT. They are (1) the strict adherence to the conditions inherent
to the derivation of DFT, including the verified attainment of the ground state energy, and (2) the
application of the Rayleigh theorem to avoid the destruction of the physical content of low energy
unoccupied levels. With the introduction of the basis set augmentation that is needed for verifiably
reaching the ground state energy, the first feature adds to or completes DFT in practice. The second
feature also does so by identifying, out a potentially infinite number, the calculation that provides
the physical content of DFT.
The authors concluded that the sum of the kinetic and electron-electron interaction energies is a
universal functional of the charge density, since Ψ is a functional of n(r):
Hohenberg and Kohn proceeded to establish that this unique functional of the density attains its
minimum (from above) for the correct ground state density n(r), provided the total number of
particles,
N[n] ≡ n(r)dr = N (5)
is kept constant. Specifically, the actual quote from the Hohenberg and Kohn article1 follows.
“It is well known that for a system of N particles, the energy functional of Ψ ′
Ev [Ψ ′] = (Ψ ′,V Ψ ′) + (Ψ ′, (T+U)Ψ ′) (6)
has a minimum at the ‘correct’ ground state Ψ, relative to arbitrary variations of Ψ ′ in which the
total number of particles is kept constant.”
The authors took the next crucial step consisting of showing that a density n ′(r), associated
with a another potential v ′(r) [different from v(r)], leads to a minimum value of the above energy
functional that is higher than the one obtained with the ground state density n(r) of the system.
Specifically, they established the following inequality:
v(r)n ′(r)dr + F[n ′] > v(r)n(r)dr + F[n] (7)
It should be noted that the potential v(r) is used on both sides of the inequality. This result is gener-
ally referred to as the second Hohenberg and Kohn theorem or as the DFT variational principle.
Hohenberg and Kohn delved into several other considerations that are outside the scope of this
work. These topics include in-depth discussions of the case of a gas with an almost constant density
and that of another with slowly varying density. Gradient expansions emerged in connection with
the latter case.
127104-3 Diola Bagayoko AIP Advances 4, 127104 (2014)
In 1965, Kohn and Sham2 introduced the local density approximation (LDA) of DFT. These
authors rewrote the energy functional in Equations (4) and (6) as
1 n(r)n(r′)
E= v(r)n(r)dr + dr dr′ + Ts [n] + Exc[n], (8)
2 |r−r′|
where Ts [n] is the kinetic energy of a system of non-interacting electrons whose density is n (r),
while Exc [n] is defined as the exchange correlation energy of a system of interacting electrons
with the same density n (r). In the case of slowly varying densities, Kohn and Sham obtained an
expression for Exc[n] in terms of the exchange correlation energy per electron of a uniform gas
whose density is n.
Exc[n] = n(r)∈xc (n(r)) dr, (9)
where ∈xc (n (r)) is the exchange and correlation energy per electron of a uniform electron gas of
density n. The next pivotal step taken by these authors consisted of using the stationary property
of Equation (8) (i.e., a particular form of Equation (6)) to derive an equation that straightforwardly
leads to the Kohn Sham (KS) equation,
n(r′) ′
1
− ∇2 + v(r) + dr + µxc(n(r)) ψi (r) =∈ i ψi (r), (10)
2 |r − r|
with the charge density computed using only the wave functions of the occupied states as
N
n(r) = |ψi (r)|2. (11)
The referenced derivation of the Kohn Sham equation requires that the total number of particles
(N) be kept constant. Interestingly enough, Equation (11) is sufficient to underscore the fact that
N must be constant. Any change in N means a totally new density and hence new values for all
quantities that are unique functionals of the density! The above KS equation describes a system of
non interacting electrons moving in the potential between brackets in Equation 10. The exchange
correlation contribution to the chemical potential of a uniform gas of density n is µxc (n (r)).
The above overview of some key features of the derivation of DFT and of LDA was to under-
score the following points: (a) In all DFT applications, the total number of particles must be
kept constant; and, (b) for the outcomes of an electronic structure calculation using DFT or LDA
potentials to represent the physical content of DFT or LDA, it is necessary either (b.1) to have, à
priori, the “correct,” tridimensional ground state charge density of the system or (b.2) to verify (as
opposed to assuming) that the calculated, occupied energies have attained their absolute minima
(from above) for the system under study. The content of this paragraph has far reaching implications
as shown below.
TABLE I. Band gaps from ab initio, self consistent DFT calculations. None of the 99 different calculations in Column 3, with
corresponding ranges of calculated gaps in Column 4, produced results acceptably close to the corresponding, experimental
values in Column 5. Details are available on the calculations in the references in Column 6. In stark contrast, the results of
the ab initio, self consistent DFT BZW and BZW-EF calculations in Column 6 agree with corresponding, experimental ones
in Column 5.
Number of Ranges of
DFT Calculated Calculated Experimental DFT BZW or
Material Potential Gaps Gaps Gaps BZW-EF Gaps
Cubic InN LDA & GGA 10 -0.55 - + 0.08 0.61 eV 0.65 eV4
AlAs LDA 5 1.20 – 1.44 eV 2.15 – 2.25 eV 2.14 – 2.15 eV5
zb-ZnS LDA & GGA 5 1.65 – 2.37 eV 3.723 eV 3.725 eV6
w-GaN LDA & GGA 17 1.68 – 2.52 eV 3.3-3.5 eV 3.20 & 3.29 eV7
zb-GaN LDA 8 1.80 – 2.10 eV 3.10 eV 2.90 eV7
Rutile TiO2 LDA & GGA 18 1.67 – 2.12 eV 3.00 – 3.10 2.95 & 3.05 eV8
w-ZnO LDA 12 0.23 – 2.26 eV 3.30 – 3.40 eV 3.39 eV9
zb-BP LDA & GGA 15 1.11 – 1.38 eV 2.02 ± 0.05 eV 2.02 eV10
c-BN LDA & GGA 9 4.20 – 4.47 eV 6.20 – 6.4 ± 0.5 eV 6.48 eV11
(HOMO), respectively. The difference between these two energies is the energy gap (for discrete
energy levels).
Bagayoko et al., using the original BZW method and LDA potentials, predicted the band gaps
and other properties of cubic Si3N412 and c-InN.4 These predictions were confirmed by experiment
in 200313,14 and 2006,15 respectively. Ekuma and Bagayoko9 predicted the indirect band gap of
to the mathematical and physical conditions inherent to the derivation of DFT. This part of the
method allows the attainment of the absolute minima of the occupied energies, as required by DFT,
while keeping the number of particles constant.
The second part (Part II) of our method, begotten from our prior experiences with the contrac-
tion of basis sets,23 enunciates the Rayleigh theorem relative to the eigenvalues of the same equation
solved with two basis sets, with the smaller basis set entirely included in the larger one. It utilizes
this theorem to select, among the potentially infinite number of calculations with large basis sets
leading to the same absolute minima of the occupied energies, the one with the smallest basis set.
This calculation, due to the attainment of the ground state energy, provides the physical content of
DFT not only for the occupied states, but also for low energy unoccupied ones. The basis set for
this particular calculation is referred to as the optimal basis set. The use of basis sets much larger
than the optimal one does not lead to changes in the occupied energies; however, it often lowers
some unoccupied energies, including the lowest ones. We address below the mathematical artifact
responsible for this unphysical, extra-lowering of unoccupied levels while the occupied ones do not
move.
Our calculations have features in common with many others. Specifically, we utilize DFT
potentials, with emphasis on LDA and generalized gradient approximation (GGA)24 ones, like other
calculations. We employ the linear combination of atomic orbital (LCAO) formalism, as done by
most electronic structure calculations. Most of our calculations have been carried out with the
program package developed and refined at the Ames Laboratory of the US Department of Energy,
Ames, Iowa.25 The radial parts of atomic orbitals include Gaussian functions. As done in other
calculations, we begin by obtaining self consistent electronic energies and related wave functions
difference of 0 to 0.2 eV between BZW and BZW-EF results is another manifestation of the limita-
tions of iterative processes. Iterations cannot correct for an inadequate angular representation in the
trial functions, irrespective of the degree of convergence attained by the said iterations. We elaborate
on this point in the discussion section in connection with the work of Franklin et al. on wurtzite
ZnO. Unless otherwise indicated, our calculations henceforth utilize the enhanced BZW-EF version
of the method. With the above understanding of the augmentation procedure for the basis set, the
rest of Part I of the method follows below.
After setting the Fermi levels to zero, the occupied energies from calculations I and II are
compared numerically and graphically. They are generally different, with some occupied ener-
gies from calculation II being lower than their corresponding ones from calculation I. The above
lowering of some occupied energies from calculation II indicates that the initial, small basis set
is inadequate for the rendition of the wave function of the system - in size (dimension), angular
symmetry, and radial functions - for the description of the ground state. Equivalently, the above
difference means that the basis set in calculation I is not complete for the description of the ground
state. With no proof that the outcomes of Calculation II provide the physical description of the
system, we proceed with calculation III, after augmenting the basis set of Calculation II as described
above. The occupied energies of Calculations II and III are compared.
This process of augmenting the basis set and of carrying out self consistent calculations con-
tinues until three successive calculations produce the same occupied energies. At this point, the
calculations are completed, given that we reached the absolute minima of the occupied energies. In
early applications of the method, we used to require that two consecutive calculations produce the
same occupied energies. After our work on zinc blende GaP and BP, where Calculations II and III
Another perspective on the solution of the DFT equations has been suggested by Bagayoko
et al.33 who noted the fact that it is a system of equations that has to be solved. Kohn and Sham
explicitly stated so.2 Once the exchange correlation potential is known, the system of equations
is reduced to the KS Schrodinger type equation and the equation giving the ground state charge
density in terms of the wave functions of the occupied states. For this system, the second equation
is similar to a constraint on the KS equation. With a single trial basis set, with the changes to
linear expansion coefficients only, the iterations solve self consistently the KS equation. However,
these iterations do not solve the system of equations self consistently. To do so, one has to change
the initial input to the charge density equation in terms of the size, angular symmetry, and radial
characteristics of the basis set as done by the BZW-EF method. Additionally, the issue is not to
change the initial input to the charge density equation using arbitrarily selected basis sets. As per
the BZW-EF method, the minimization of the energy functional in Equations (6) or (8) requires that
the basis sets in question be such that, except for the first and the last ones, each one contains the
basis set immediately preceding it in size and is in turn included in the basis set that immediately
follows it in size. Essentially, this minimization that solves the system of equation can be viewed to
comprise iterations for the charge density equations, with progressively augmented basis sets, and
iterations for the KS equation for each choice of basis set. The above embedding of increasing basis
sets is actually needed for the minimization of the energy functional in Equations (6) or (8); as such,
it adds to DFT in practice. These points complete the description of Part I of the BZW-EF (and
BZW) method. Part II of the BZW-EF method follows.
The second part (Part II) of the BZW or BZW-EF method, as given below, does not follow
directly from any derivation of DFT known to us. Further, it turns out to be necessary for most
DFT calculations utilizing the LCAO formalism. The Rayleigh theorem is a key basis for this Part
content of these Hamiltonians, as in Equation (6) (or (8) for LDA), that is a unique functional of
the density, remains the same as that obtained with the optimal basis set. Hence, the lowering of
unoccupied energies cannot be attributed to the content of the Hamiltonians. As noted above, it
is a mathematical artifact that is not an ill-defined “basis set effect”; rather, it is a “basis set and
variational” effect that can be invoked only after verifiably attaining the absolute minima of the
occupied energies.
The above description of Part II of the BZW or BZW-EF method shows that it adds to or
completes DFT, inasmuch as its correct application is concerned. Indeed, the above referenced
extra-lowering destroys the physical content of the lowest laying unoccupied energies in atoms,
molecules, some nanostructures, semiconductors, and insulators. It provides a simple explanation
of the rampant underestimations of measured band gaps by most single basis set DFT (and LDA)
calculations. The importance of Part II of the method resides in the preservation of the physical con-
tent of low laying unoccupied energies for non metals. For binding energies and band widths, Part I
of the BZW-EF method is needed even for the description of metals. As noted above, the embedding
of basis sets, as done in the BZW-EF method, seems to be necessary for the minimization of the
energy content of the Hamiltonian as in Equations (6) and (8).
are less than half of corresponding results from similar SIC calculations with the 4d electrons in
the valence.35 The drastically different outcomes of two similar SIC calculations suggest that issues
other than DFT may affect computational results. We do not discuss the adjustments of LDA results
by Furthmüller et al. for some quasiparticle properties in light of our focus on DFT, a ground state
theory. Additionally, Furthmüller et al. found differences of 0.76 and 0.79 eV between the band
gaps at Γ, for w-InN and c-InN, depending on whether or not the 4d electrons are explicitly treated
(i.e., included in the valence). However, the corresponding changes in band gaps, for DFT BZW
calculations,36 are just 0.13 eV for both w-InN and c-InN. The rather large (0.76 &0.79 eV) and
small (0.13 eV) changes in the band gaps, as calculated by Furthmüller et al.35 and Bagayoko
et al.36, respectively, strongly suggests differences between the results of these calculations other
than those due to DFT. In stark difference with the small change (0.13 eV) in the band gaps,
when the 4d electrons are moved from the core to the valence, LDA BZW valence band widths
increase by more than 1 eV for both materials. The inclusion of the 4d electrons in the valence is
needed for a correct description of InN4,27,36. This need to include the d electrons in the valence is
a key result obtained by Wei and Zunger37 in their study of II-VI semiconductors. This case of InN
illustrates our contention, borne out of the derivation of DFT, that there may be shortfalls of some
calculations, and not of DFT, that are ascribed to DFT.
The importance of the band gap partly explains our emphasis on it in this overview of DFT
and our description of the BZW-EF method. Many applications of semiconductors and insulators
hinge on the nature (direct or indirect) and numeric value of the band gap. Additionally, once
calculations produce an inaccurate band gap, then their outputs cannot be utilized to calculate
dielectric functions, densities of states, or effective masses that agree with experiments – without
adjustments. Peaks in the two former quantities will disagree with corresponding, experimental
a local spin density approximation (LSDA) potential and linear muffin tin orbitals (LMTO). A
central difficulty in determining the relevance of our discussions to this important development
resides in the fact that we only addressed ab-initio calculations for the ground states of systems with
fixed numbers of particles, while the intricate determination of relaxation energies, as described by
Jarlborg et al., necessarily entails calculations of the total energies of systems with (N-1), N, and
(N+1) electrons. The relaxed excitations energies for an electron and a hole are respectively the
differences between EtN−1 and EtN and between EtN+1 and EtN, where the E’s are total energies and
the superscripts are the numbers of electrons. The relaxation energy for the excitation process is
defined as the difference between these excitation energies and εn0, the nth Kohn Sham orbital en-
ergy. Despite the noted comparison difficulties, it appears that future work could aid in elucidating
a possible convergence. Indeed, we expect the relaxation energies obtained by using the BZW-EF
method throughout the above calculations to be different from the ones resulting from calculations
that utilize a single basis set. Even if the BZW-EF method is not employed in the calculations
of the total energies, significant differences are expected depending on whether or not εn0 is from
a BZW-EF calculation or not. The above suggested comparisons are expected to be particularly
informative if calculations with and without the BZW-EF method utilize the same DFT potential.
We have studied magnetic and semi metallic materials in the past. We can therefore carry out spin
polarized calculations in the future.
The two developments discussed serve in part to delineate the domain of validity of the con-
tent of this article. Indeed, as for the TB-mBJ potentials, this work does not apply to calculations
using potentials that are not derivable entirely from DFT. Other examples of such potentials are
LDA+C,46 LDA+U,47 DFT+U+V,48 and several others. This work, as per the derivation of DFT
reviewed in Section II, only applies strictly to studies of the ground states of materials. So, it does
VI. CONCLUSION
We followed the works of Hohenberg and Kohn and of Kohn and Sham to articulate the
derivation of density functional theory (DFT) and of its local density approximation (LDA). The
derivation elucidated clear conditions, summarized at the end of Section II, to which electronic
127104-11 Diola Bagayoko AIP Advances 4, 127104 (2014)
structure calculations and related ones have to adhere in order for their results to represent the phys-
ical content of DFT, in general, or of LDA, in particular. Stark differences between results of some
99 calculations that woefully underestimated band gaps of materials and those obtained with the
BZW or BZW-EF method posed a question. This question was partly answered by the lack of any
evidence, as per the contents of the 99 publications, of a strict adherence to the referenced, neces-
sary conditions. Hence, disagreements between the results of these calculations and corresponding,
experimental ones do not necessarily signify any failures or shortfalls of DFT. The referenced,
necessary conditions have to be met by a calculation before its results can represent the physical
content of DFT. As to the reasons that DFT BZW and BZW-EF calculations accurately reproduce
experimental results and make predictions confirmed by experiment, the first part of the answer was
the strict adherence of these calculations to the conditions inherent to the derivation of DFT and of
LDA. In retrospect, Part I of the method relative to the minimization of the energy was prescribed
by Hohenberg and Kohn in 1964, except for the needed use of successively augmented basis sets!
By showing the necessity for the variations associated with that of size and other attributes of the
basis set, in addition to variations of linear expansion coefficients in the self-consistency process,
Part I add to DFT in practice. Part II of the BZW and BZW-EF method consists of utilizing the
Rayleigh theorem to avoid mathematical artifacts that arise when basis sets much larger than the
optimal ones are employed in ab-initio, self consistent calculations. This Part II of the method is
critical in the study of all materials that are not known to be metals, due to their energy or band
gaps. Hence, the study of atoms, molecules, clusters and other finite structures, nanostructures,
semiconductors, and insulators rests in part on this second part of our method that complements its
first part. By preserving the physical content of low laying energy levels, Part II of the method also
adds to or completes DFT in practice.
Fitch, and D. J. Frost, Phys. Rev. B, Rapid Communications 65, 161202(R) (2002).
15 J. Schörmann, D. J. As, K. Lischka, P. Schley, R. Goldhahn, S. F. Li, W. Löffler, M. Hetterich, and H. Kalt, Appl. Phys. Lett.
24 David C. Langreth and John P. Perdew, Phys. Rev. B 21, 5469 (1980).
25 B. N. Harmon, W. Weber, and D. R. Hamann, Phys. Rev. B 25(2), 1109–1115 (1982).
26 D. Bagayoko, G. L. Zhao, J. D. Fan, and J. T. Wang, J. of Physics: Condensed Matter 10, 5645 (1998).
27 G. L. Zhao, D. Bagayoko, and T. D. Williams, Physical Review B60, 1563 (1999).
28 C. E. Ekuma, Master’s Degree Thesis in Physics (Southern University and A&M College in Baton Rouge (SUBR, Baton
Rouge, LA 70813).
29 C. E. Ekuma, M. Jarrell, J. Moreno, and D. Bagayoko, AIP Advances 2, 012189 (2012).
30 D. Bagayoko, in Proceedings, Malian Symposium of Applied Sciences (MSAS), Bamako, Mali (2004) pp. 53–58.
31 C. E. Ekuma, M. Jarrell, J. Moreno, G. L. Zhao, and D. Bagayoko, Phys. Lett. A Vol. 377(34-36), 2172–2176, 1 November
2013 - http://authors.elsevier.com/sd/article/S0375960113005318.
32 I. H. Nwigboji, J. I. Ejembi, Y. Malozovsky, B. Khamala, L. Franklin, G. L. Zhao, C. E. Ekuma, and D. Bagayoko, manuscript
(2014).
42 D. Koller, F. Tran, and P. Blaha, Phys. Rev. B 83, 195134 (2011).
43 J. A. Carmago-Martinez and R. Baquero, Phys. Rev. B 86, 195106 (2012).
44 H. Jiang, J. Chem. Phys. 138, 134115 (2013).
45 B. Ul Haq, R. Ahmed, S. Goumri-Said, A. Shaari, and A. Afaq, Phase Transitions: A Multinational Journal Vol. 86(No. 12),
1167 (2013).
46 Su-Huai Wei and Alex Zunger, Phys. Rev. B 57, 8983 (1998).
47 V. I. Anisimov, J. Zaanen, and Ole K. Andersen, Phys. Rev. B 44, 943 (1991).