dft: a theory full of...

file: a8˙1 Total pages: 13

DFT: A Theory Full of Holes?

Aurora Pribram-Jones,1 David A. Gross,1 and Kieron Burke1

1Department of Chemistry, University of California,1102 Natural Sciences 2, Irvine, CA 92697-2025, USA

(Dated: Friday 1st August, 2014)

This article is a rough, quirky overview of both the history and present state of the art of densityfunctional theory. The field is so huge that no attempt to be comprehensive is made. We focus onthe underlying exact theory, the origin of approximations, and the tension between empirical andnon-empirical approaches. Many ideas are illustrated on the exchange energy and hole. Featuresunique to this article include how approximations can be systematically derived in a non-empiricalfashion and a survey of warm dense matter.

1. WHAT IS THIS ARTICLE ABOUT?

The popularity of density functional theory (DFT) asan electronic structure method is unparalleled, with ap-plications that stretch from biology[1] to exoplanets[2].However, its quirks of logic and diverse modes of practi-cal application have led to disagreements on many frontsand from many parties. Developers of DFT are guidedby many different principles, while applied practitioners(a.k.a. users) are suspicious of DFT for reasons bothpractical (how can I pick a functional with so manychoices? [3]) and cultural (with so many choices, whywould I call this first-principles? ).

A modern DFT calculation[4] begins with the purchaseof a computer, which might be as small as a laptop, anda quantum chemical code. Next, a basis set is chosen,which assigns predetermined functions to describe theelectrons on each atom of the molecule being studied.Finally, a DFT approximation to something called theexchange-correlation energy (XC) is chosen, and the codestarts running. For each guess of the nuclear positions,the code calculates an approximate energy[4]. A geom-etry optimization should find the minimum energy con-figuration. With variations on this theme[5, 6], one canread out all molecular geometries, dissociation energies,reaction barriers, vibrational frequencies, etc. A mod-ern desktop may do a calculation for a 100-atom systemwithin a day. A careful user will repeat the most im-portant parts of the calculation with bigger basis sets, tocheck that answers don’t change significantly.

2. WHERE DOES DFT COME FROM?

Although DFT’s popularity has skyrocketed since ap-plications to chemistry became useful and routine, itsroots stretch back much further[7–9].

Ye olde DFT: Developed without reference tothe Schrodinger equation[10], Thomas-Fermi (TF)theory[11–13] was the first DFT. It is pure DFT, relyingonly on the electronic density, ρ(r), as input. The ki-netic energy was approximated as that of a uniform elec-tron gas, while the repulsion of the electrons was modeledwith the classical electrostatic coulomb repulsion, again

depending only on the electronic density as an input.Mixing in orbitals: John Slater was a master of elec-

tronic structure whose work foreshadowed the develop-ment of DFT. In particular, his Xα method[14] approx-imates the interactions of electrons in ground-state sys-tems and improved upon Hartree-Fock (HF) [15, 16], oneof the simplest ways to capture the Pauli exclusion prin-ciple. One of Slater’s great insights was the importanceof holes, a way of describing the depressed probabilityof finding electrons close to one another. Ahead of histime, Slater’s Xα included focus on the hole, satisfiedexact conditions like sum rules, and considered of thedegree of localization present in the system of interest.

A great logical leap: Although Slater’s methods pro-vided an improvement upon HF, it was not until 1964that Hohenberg and Kohn formulated their famoustheorems[17], which now serve as the foundation of DFT:(i) the ground-state properties of an electronic systemare completely determined by ρ(r), and(ii) there is a one-to-one correspondence between the ex-ternal potential and the density.

We write this by splitting the energy into two pieces:

Eelec[density] = F [density] +NucAtt, (1)

where Eelec is the total energy of the electrons, F is thesum of their exact quantum kinetic and electron-electronrepulsion energies, and NucAtt is their attraction to thenuclei in the molecule being calculated. Square brackets[ ] denote some (very complex) dependence on the one-electron density, ρ(r), which gives the relative probabilityof finding an electron in a small chunk of space aroundthe point r. F is the same for all electronic systems, andso is called universal. For any given molecule, your com-puter simply finds ρ(r) that minimizes Eelec above. Com-pare this to the variational principle in regular quantummechanics. Instead of spending forever searching lots ofwavefunctions that depend on all 3N electronic coordi-nates, you just search over one-electron densities, whichhave only 3 coordinates (and spin).

The pesky thing about the Hohenberg-Kohn theorems,however, is that they tell us such things exist withouttelling us how to find them. This means that to actuallyuse DFT, we must approximate F [density]. We recognizethat the old TF theory did precisely this, with very crude

1


approximations for the two main contributions to F :

F [density] ∼∫d3r ρ5/3(r) + CoulRep, (TF ) (2)

where we’ve not bothered with constants, etc. The firstterm is an approximation to the kinetic energy as a sim-ple integral over the density. It is a local approximation,since the contribution at any point comes from only thedensity at that point. The other piece is the self-repulsionamong electrons, which is simply modeled as the classicalelectrostatic repulsion, often called their Hartree energyor the direct Coulomb energy. Such simple approxima-tions are typically good to within about 10% of the elec-tronic energy, but bonds are a tiny fraction of this, andso are not accurate in such a crude theory[18].

A great calculational leap:: Kohn and Sham proposedrewriting the universal functional in order to approxi-mate only a small piece of the energy. They mapped theinteracting electronic system to a fake non-interactingsystem with the same ρ(r). This requires changingthe external potential, so these aloof, non-interactingelectrons produce the same density as their interactingcousins. The universal functional can now be broken intonew pieces. Where in the interacting system, we had ki-netic energy and electron-electron interaction terms, inthe Kohn-Sham (KS) system, we write the functional

F = OrbKE + CoulRep+XC (3)

where OrbKE is the kinetic energy of the fake KS elec-trons. XC contains all the rest, which includes both ki-netic and potential pieces. Although it is small comparedto the total, ‘nature’s glue’ [19] is critical to getting chem-istry and physics right. The X part is (essentially) theFock exchange from a HF calculation, while C is the cor-relation energy, i.e., that part that traditional methodssuch as coupled cluster usually get very accurately[20].

When you minimize this new expression for the en-ergy, you find a set of orbital equations, the celebratedKS equations. They are almost identical to Hartree-Fockequations, and this showed that Slater’s idea could bemade exact (if the exact functional were known). Thegenius of the KS scheme is that, because it calculates or-bitals and gives their kinetic energy, only XC, a smallfraction of the total energy, needs to be approximated asa density functional. The KS scheme usually producesexcellent self-consistent densities, even with simple ap-proximations like LDA, but approximate potentials forthis non-interacting KS system are typically very differ-ent from the exact KS potential (Fig. 1).

Popular approximations for XC: Despite the over-whelming number of approximations available in the av-erage DFT code, most calculations rely on a few of themost popular approximations. The sequence of these ap-proximations is

XC ∼ XCunif(ρ) (LDA)

∼ XCGGA(ρ, |∇ρ|) (GGA)

∼ a(X −XGGA) +XCGGA (hybrid) (4)

0.0 0.5 1.0 1.5 2.0 2.5 3.0-2

-1

0

1

2

r

vHrL

4Πr2

nHrL He atom

-2�r

KS

GGA

FIG. 1. Radial densities and potentials for the He atom(energies in Hartree, distances in Bohr). The pink line is−2/r, the attraction of real electrons to the nucleus. Theyellow is the exact KS potential. Two fake electrons in the 1sorbital of this potential have the same ground-state density asreal He. The green is the potential of a typical approximationwhich, although inaccurate, yields a highly accurate density.

The first was the third major step from the mid-60s andwas invented in the KS paper[21]. It was the mainstayof solid-state calculations for a generation, and remainspopular for some specific applications even today. It is(almost) never used in quantum chemistry, as it typicallyoverbinds by about 1eV/bond. The local density approx-imation (LDA)[21] assumes that the XC energy dependson the density at each position only, and that dependenceis the same as in a uniform electron gas.

Adding another level of complexity leads to themore accurate generalized gradient approximations(GGAs)[22, 23], which use information about both thedensity and its gradient at each point. Hybrid approx-imations mix a fraction (a) of HF exact exchange witha GGA[24]. These maneuvers beyond the GGA usuallyincrease the accuracy of certain properties with an afford-able increase in computational cost[25]. (Meta-GGAs tryto use a dependence on the KS kinetic energy density toavoid calculating the Fock exchange of hybrids[26, 27],which can be very expensive for solids.)

Fig. 2 shows that the two most popular functionals,PBE[28, 29] and B3LYP[24, 30], comprise a large frac-tion of DFT citations each year (about 2/3), even thoughthey are now cited only about half the time they areused. PBE is a GGA, while B3LYP is a hybrid[24]. Asa method tied to Hartree-Fock, quantum chemists’ oldstomping grounds, and one with typically higher accu-racy than PBE, B3LYP is more often a chemist’s choice.PBE’s more systematic errors, mathematical rationale,and lack of costly exact exchange, have made it mostpopular in solid-state physics and materials science. Inreality, both are used in both fields and many others aswell.

Cultural wars: The LDA was defined by Kohn andSham in 1965; there is no controversy about how it was

2


0

10

20

30

year

kilo

pa

pe

rs

19

95

1

99

6

19

97

1

99

8

19

99

2

00

0

20

01

2

00

2

20

03

2

00

4

20

05

2

00

6

20

07

2

00

8

20

09

2

01

0

20

11

2

01

2

20

13

OtherDarkBecke93PBE96

FIG. 2. The number of DFT citations has exploded (as haveab initio methods). PBE is the number of citations of Ref.[28], and Becke93 of Ref. [24]. Dark indicates papers usingeither of these approximations without citing the original pa-pers, while other is all other DFT papers. All numbers areestimates. Contrast with Fig. 1 of Ref. [7], which missedalmost 2/3 of these.

designed. On the other hand, adding complexity to func-tional approximations demands choices about how totake the next step. Empirical functional developers fittheir approximations to sets of highly accurate referencedata on atoms and molecules. Non-empirical develop-ers use exact mathematical conditions on the functionaland rely on reference systems like the uniform and slowly-varying electron gases. The PBE GGA is the most popu-lar non-empirical approximation, while the most popularempirical functional approximation is the B3LYP hybrid.Modern DFT conferences usually include debates aboutthe morality of this kind of empiricism.

Both philosophies have been incredibly successful, asshown by their large followings among developers andusers, but each of these successes is accompanied by fail-ures. No single approximation works well enough for ev-ery property of every material of interest. Many userssit squarely and pragmatically in the middle of the twofactions, taking what is best from both of their accom-plishments and insights. Often, empiricists and non-empiricists find themselves with similar end products, agood clue that something valuable has been created withthe strengths of both.

To illustrate this idea, we give a brief allegory froman alternative universe. Since at least the 1960s, accu-rate HF energies of atoms have been available due tothe efforts of Charlotte Froese Fischer and others[31, 32].A bright young chemistry student plots these X ener-gies as a function of Z, the atomic number, and noticesthey behave roughly as Z5/3, as in Fig. 3. She’s anorganic chemistry student, and mostly only cares aboutmain-group elements, so she fits the curve by choosing aconstant to minimize the error on the first 18 elements,finding EX = −0.25Z5/3. Much later, she hears aboutKS DFT, and the need to approximate the XC energy.

5 10 15 20 25-50

-40

-30

-20

-10

0

Atomic Number

Exch

an

ge

Hartree-FockTheorists' LDAChemist's Fit

FIG. 3. Exchange energy (in Hartrees) of atoms from a HFcalculation as a function of Z, atomic number, and two LDAXcalculations, one with the theoretical asymptote, the otherfitted.

A little experimentation shows that if

Xopt = C0

∫d3r ρ4/3(r), (5)

this goes as Z5/3 when Z is large, and choosing C0 =−0.80 makes it agree with her fit.

In our alternate timeline, a decade later, Paul Dirac,a very famous physicist, proves[33] that for a uniformgas, C0 = AX = −(3/4)(3/π)1/3 = −0.738. Worse still,Julian Schwinger proves[34] that inserting the TF densityinto Dirac’s expression becomes exact as Z →∞, so thatEX → −0.2201Z5/3. Thus theirs is the ‘correct’ LDA forX, and our brave young student should bow her head inshame.

Or should she? If we evaluate the mean absolute er-rors in exchange for the first 20 atoms, her functional issignificantly better than the ‘correct’ one[35]. If lives de-pend on the accuracy for those 20 atoms, which wouldyou choose[36]?

This simple fable contains the seeds of our actual cul-tural wars in DFT derivations:

(i) An intuitive, inspired functional need not wait for anofficial derivation. One parameter might be extracted byfitting, and later derived.

(ii) A fitted functional will usually be more accurate thanthe derived version for the cases where it was fitted. Themagnitude of the errors will be smaller, but less system-atic.

(iii) The fitted functional will miss universal propertiesof a derived functional. We see in Sec. 6 that the correctLDA for exchange is a universal limit of all systems, notjust atoms.

(iv) If you want to add the next correction to LDA, start-ing with the wrong constant will make life very difficult(see Sec. 6).

3


3. WHAT’S AT THE FOREFRONT?

Accurate Gaps: Calculating accurate energy gaps andself-interaction errors are notorious difficulties withinDFT[37]. Self-interaction error stems from spurious in-teraction of an electron with itself in the coulomb repul-sion term. Orbital-dependent methods often cure mostof the SIE problem, but they can be expensive to run.The ‘gap problem’ in DFT often stems from treating theKohn-Sham HOMO-LUMO gap as the fundamental gap,but the difference in the HOMO and LUMO of the KSsystem is not the same as the difference between theionization potential and the electron affinity. Ad hocmethods are often used to correct DFT gaps, but thesemethods require expensive additional calculations, em-pirical knowledge of your system, or empirical tuning.However, it has recently been shown that some classesof self-interaction error are really just errors due to poorpotentials leading to poorer densities [38, 39]. Such er-rors are removed by using more accurate densities (Fig.4).

-10

-8

-6

-4

-2

0

0 40 80100 140 180

CCSD(T)PBEHF-PBE

0 40 100 140 180

0

-2

-4

-6

$ (�)

Eb (

kcal

/mol

)

R (Å)

. (�)

$ (�)

-8

FIG. 4. When a DFT calculation is abnormally sensitive tothe potential, the density can go bad. Usually, DFT approx-imate densities are better than HF[40], as in Fig 1. Here,self-consistent PBE results for OH −H2O interactions yieldthe wrong geometry, but PBE on HF-densities fixes this[39].

Range-separated hybrids: Range-separated hybrids[41]improve fundamental gaps calculated via the DFTHOMO-LUMO gap[42]. Screened range-separated hy-brids can even achieve gap renormalization presentwhen moving between gas-phase molecules and molecu-lar crystals[43]. The basic range-separated hybrid schemeseparates the troublesome Coulomb interaction into long-range and short-range pieces. The screened version en-forces exact conditions to determine where this separa-tion occurs and incorporates the dielectric constant as an

adaptive parameter. This technique takes into accountincreased screening as molecules form solids, resulting inreduced gaps critical for calculations geared toward ap-plications in molecular electronics.

Weak Interactions: Another of DFT’s classic failings isits poor treatment of weak interactions[44, 45]. Induceddipoles and the resulting dispersion interactions are notcaptured by the most popular approximations of Eq. 4.This prevents accurate modeling of the vast majority ofbiological systems, as well as a wide range of other phe-nomena, such as surface adsorption and molecular crys-tal packing. GGAs and hybrids are unable to modelthe long-range correlations occurring between fluctua-tions induced in the density. The non-empirical approachbased on the work of Langreth and Lundqvist[46–49] andthe empirical DFT-D of Grimme[50, 51] have dominatedthe advances in this area, along with the more recent, lessempirical approach of Tkatchenko and Scheffler[52, 53].

4. REDUCING COST: IS LESS MORE?

No matter how much progress is made in improvingalgorithms to reduce the computational cost of DFT cal-culations, there will always be larger systems of interest,and even the fastest calculations become prohibitivelyexpensive. The most glaring example is molecular dy-namics (MD) simulation in biochemistry. With classicalforce fields, these can be run for nano- to milli-seconds,with a million atoms, with relative ease. But when bondsbreak, a quantum treatment is needed, and the first ver-sions of these were recognized in last year’s Nobel prizein chemistry[54–56]. These days, many people run Car-Parrinello MD[57, 58], with DFT calculations inside theirMD, but this reduces tractable system sizes to a few hun-dred atoms.

Because of this, there remains a great deal of interestin finding clever ways to keep as much accuracy as neededwhile simplifying computational steps. One method fordoing so involves circumventing the orbital-dependentKohn-Sham step of traditional DFT calculations. Al-ternatively, one can save time by only doing those costlysteps (or even more expensive procedures) on a system’smost important pieces, while leaving the rest to be cal-culated using a less intensive method. The key to bothapproaches is to achieve efficiency without sacrificing pre-cious accuracy.

Removing the orbitals: Orbital-free methods[6, 59–63]like TF reduce computational costs, but are often notaccurate enough to compete with KS DFT calculations.Current methods search for a similar solution, by workingon non-interacting kinetic energy functionals that allowcontinued use of existing XC functionals[64]. (An intrigu-ing alternative is to use the potential as the basic variable[65, 66] – see Secs. 6 and 7.)

Embedding: Partitioning and embedding are similarprocedures, in which calculations on isolated pieces of amolecule are used to gain understanding of the molecule

4


as a whole[67]. One might want to separate out molecularregions to look more closely at pieces of high interest orto find a better way to approximate the overall energywith density functionals. Parsing a molecule into chunkscan also allow for entirely new computational approachesnot possible when dealing with the molecule as a whole.

Partition DFT[68] is an exact embedding methodbased on density partitioning[69, 70]. Because it usesensemble density functionals[71, 72], it can handle non-integer electron numbers and spins[73]. Energy of thefragments is minimized using effective potentials consist-ing of a fragment’s potential and a global partition poten-tial that maintains the correct total density. This break-down into fragment and partition energies allows approx-imations good for localized systems to be used alongsidethose better for the extended effects associated with thepartition potential.

While partition DFT uses DFT methods to break upthe system, projector-based wavefunction-theory-in-DFTembedding techniques combine wavefunction and DFTmethods[74, 75]. This multiscale approach leverages theincreased accuracy of some wavefunction methods forsome bonds, where high accuracy is vital, without ex-tending this computational cost to the entire system.Current progress in this field has been toward the reduc-tion of the errors introduced by the mismatch of meth-ods between subsystems, and the method has been re-cently applied to heterolytic bond cleavage and conju-gated systems[76]. Density matrix embedding theory onlattices[77] and its extension to full quantum mechanicalchemical systems[78] use ideas from the density matrixrenormalization group (DMRG)[79, 80], a blazingly fastway to exactly solve low-dimensional quantum mechanicsproblems. This shifts the interactions between fragmentsto a quantum bath instead of dealing with them througha partition potential.

5. WHAT IS THE UNDERLYING THEORYBEHIND DFT APPROXIMATIONS?

Given the Pandora’s box of approximate functionals,many found by fitting energies of systems, most usersimagine DFT as an empirical hodge-podge. Ultimately,if we end up with a different functional for every system,we will have entirely defeated the idea of first-principlescalculations. However, in the days prior to the mid-90s,many decades of theory were developed to better under-stand the local approximation and how to improve onit[44]. Here we summarize the most relevant points.

The joint probability of finding one electron in a lit-tle chunk of space around point A and another in someother chunk of space around point B is called the pairprobability density. The exact quantum repulsion amongelectrons is then

ElecRep =1

2

∫dA

∫dB

P (A,B)

|rA − rB |. (6)

But we can also write

P (A,B) = ρ(A) ρcond(A,B). (7)

where ρ(A) is the density at rA and ρcond(A,B) is theprobability of finding the second electron at B, given thatthere’s one at A. (If you ignore the electron at A, this isjust ρ(B), and Eq. 6 gives the coulomb repulsion in Eqs.2 and 3). We write this conditional probability as

ρcond(A,B) = ρ(B) + ρXC(A,B). (8)

where nXC(A,B) is called the hole around A. It is mostlynegative and represents a missing electron (it integratesto -1), since the conditional probability integrates toN − 1. With a little math trick (called the adiabaticconnection[81, 82]), we can fold the kinetic correlationinto the hole so that

XC =1

2

∫dA

∫dB

ρ(A) ρXC(A,B)

|rA − rB |. (9)

−6 −4 −2 0 2 4 6−1

0

1

2

3

4

5

B

ρ(B

)

A = 2

FIG. 5. Cartoon of a one-dimensional 10-electron density(solid red), the conditional density (dot-dashed blue) givenan electron at A = 2, and its hole density (dashed green).

Because the XC hole tends to follow an electronaround, i.e., be centered on A as in Fig. 5, its shape isroughly a simple function of ρ(A). If one approximatesthe hole by that of a uniform gas of density ρ(A), Eq. 9above yields the LDA for the XC energy. So the LDA ap-proximation for XC can be thought of as approximatingthe hole by that of a uniform gas[44, 83].

But while the XC is roughly approximated by LDA, theenergy density at each point in a system is not, especiallyin systems of low symmetry. But from Eq. 9, the energydepends only on the average of the XC hole over thesystem, and Fig. 6 shows such a system-averaged holefor the He atom. (Integrate over A and the angular partsof B in Eq. (9).) The LDA hole is not deep enough, and

5


0 1 2−0.5

−0.4

−0.3

−0.2

−0.1

0

u

2πu〈ρ

x(u)〉

HF

LDA

GGA

He atom

FIG. 6. Representation of system-averaged radial exchangeholes for the helium atom[84], weighted by the Coulumb re-pulsion, so that the area equals the X energy. The LDA hole(dashed green) is not deep enough, reflecting the LDA under-estimate of the magnitude of the X energy. The GGA hole(dot-dashed blue) is substantially better, but a little too deep.

neither is the LDA energy. This is the effect that leadsto LDA overbinding of molecules.

GGA Made Briefer: The underlying idea behind thePerdew series of GGAs was to improve on the LDAhole[85]. Adding gradient corrections to the hole violatescertain sum rules (negativity of the X hole and integra-tion to -1, and integration to 0 for the correlation hole),so the real-space cutoff procedure was designed to restorethese conditions. This is an effective resummation of thegradient expansion, producing the numerical GGA. Thepopular functional PBE was derived from imposing ex-act conditions on a simple form[28, 29], but should bebelieved because it mimics the numerical GGA. In Fig.6, we show how the GGA hole roughly improves on LDA,reducing typical energy errors by a factor of three.

GGAs don’t only show how important good hole mod-els can be. They also demonstrate that good approxi-mations can satisfy different exact conditions, so pick-ing which to satisfy is non-trivial. For instance, B88[23],PW91[86, 87], and PBE[28, 29] give similar values for ex-change energy when densities do not get too small or varytoo quickly. However, once they do, each behaves verydifferently. Each approximation was sculpted to satisfydifferent exact conditions in this limit. Becke decideda good energy density for exponential electronic densi-ties was important. Perdew first thought that a partic-ular scaling behavior was important[88], then that satis-fying a certain bound was better[28]. Without a system-atic way to improve our approximations, these difficultchoices guide our progress. But starting from a modelfor the XC hole is an excellent idea, as such a model canbe checked against the exact XC hole[89].

XDM: A recent, parameter-free approach to cap-turing dispersion is the exchange-hole dipole moment(XDM) method[90–93], where perturbation theory yieldsa multipole-multipole interaction, and quantum effectsare included through the dipole moment of the elec-tron with its exchange hole. Using these in concert

with atomic polarizabilities and dipole moments gener-ates atomic pair dispersion coefficients that are within4% of reference C6 values[94]. Such a model has the ad-vantage over the more popular methods of Sec. 3 becauseits assertions about the hole can be checked.

RPA and other methods: Originally put forth in the1950s as a method for the uniform electron gas, therandom phase approximation (RPA) can be viewed asa simplified wavefunction method or a nonlocal densityfunctional approach that uses both occupied and unoc-cupied KS states to approximate the correlation energy.RPA correlation performs extremely well for noncovalent,weak interactions between molecules and yields the cor-rect dissociation limit of H2[95], two of the major failuresof traditional DFT approximations[96].

Though computational expense once hindered its wideuse, resolution-of-identity implementations[97, 98] haveimproved its efficiency, making RPA accessible to re-searchers interested in large molecular systems. RPAgives good dissociation energy for catalysts involving thebreaking of transition-metal-ligand and carbon-carbonbonds in a system of over 100 atoms[99]. ThoughRPA handles medium- and long-range interactions verywell, its trouble with short-range correlations invitesdevelopment of methods that go ‘beyond RPA.’ RPAused in quantum chemistry usually describes only theparticle-hole channel of the correlation, but anotherrecent approach to RPA is the particle-particle RPA(pp-RPA)[100]. pp-RPA is missing some correlation,which causes errors in total energies of atoms and smallmolecules. This nearly cancels out in reaction energy cal-culations and yields fairly accurate binding energies[101].

RPA and variations on it will likely lead to methodsthat work for both molecules and solids, and their com-putational cost will be driven down by algorithmic devel-opment. However, RPA is likely to remain substantiallymore expensive than a GGA calculation for the indefinitefuture. While it may rise to fill an important niche inquantum chemistry, producing comparably accurate en-ergetics to modern functionals without any empiricism,such methods will not replace DFT as the first run formany calculations. Moreover, as with almost all ‘better’methods than DFT, there appears to be no way to buildin the good performance of older DFT approximations.

6. IS THERE A SYSTEMATIC APPROACH TOFUNCTIONAL APPROXIMATION?

A huge intellectual gap in DFT development has beenin the theory behind the approximations. This, as de-tailed above, has allowed the rise of empirical energyfitting. Even the most appealing non-empirical devel-opment seems to rely on picking and choosing which ex-act conditions the approximation should satisfy. Lately,even Perdew has resorted to one or two parameters inthe style of Becke[102, 103], in order to construct a meta-GGA. Furthermore, up until the mid 1990s, many good

6


approximations were developed as approximations to theXC hole, which could then be tested and checked for sim-ple systems.

However, in fact, there is a rigorous way to developdensity functional approximations. Its mathematicalfoundations were laid down 40 years ago by Lieb andSimon[104–106]. They showed the fractional error in theenergy in any TF calculation vanishes as Z → ∞, keep-ing N = Z. Their original proof is for atoms, but appliesto any molecule or solid, once the nuclear positions arescaled by Z1/3 also. Their innocuous statement is in factvery profound. This very complicated many-body quan-tum problem, in the limit of large numbers of electrons,has an almost trivial (approximate) solution. And al-though the world finds TF theory too inaccurate to beuseful, and performs KS calculations instead, the equiv-alent statement (not proven with rigor) is that the frac-tional error in the LDA for XC vanishes as Z →∞. XC,like politics, is entirely local in this limit.

These statements explain many of the phenomena wesee in modern DFT:(i) LDA is not just an approximation that applies for uni-form or slowly varying systems, but is instead a universallimit of all electronic systems.(ii) LDA is the leading term in an asymptotic expansionin powers of ~, i.e., semiclassical. Such expansions arenotoriously difficult to deal with mathematically.(iii) The way in which LDA yields an ever smaller erroras Z grows is very subtle. The leading corrections areof several origins. Often the dominant error is a lack ofspatial quantum oscillations in the XC hole. However,as Z grows, these oscillations get faster, and so their neteffect on the XC energy becomes smaller. Thus, even asZ grows, LDA should not yield accurate energy densitieseverywhere in a system (and its potential is even worse,as in Fig. 1), but the integrated XC energy will becomeever more accurate.(iv) The basic idea of the GGA as the leading correc-tion to LDA makes sense. The leading corrections tothe LDA hole should exist as very sophisticated func-tionals of the potential, but whose energetic effects canbe captured by simple approximations using the densitygradient. This yields improved net energetics, but energydensities might look even worse, especially in regions ofhigh gradients, such as atomic cores.

Next, we continue the allegory from Sec. 2. To do so,we subtract the LDA X energy from our accurate ones,so we can see the next correction, and plot this, per elec-tron, in Fig. 7. Now, a bright young chemist has heardabout the GGA, cooks up an intuitive correction to LDA,and fits one parameter to the Nobel gas values. Later,some physicsts derive a different GGA, which happensto also give the correct value. Later still, a derivationof the correction for large Z is given, which can be usedto determine the parameter (and turns out to match theempirical value within 10%). The only difference withthe orginal allegory is that this is all true. The chemistwas Axel Becke; his fitted functional is B88[23]. The

10 20 30 40 50

-0.20

-0.15

-0.10

-0.05

0.00

Atomic Number

HX-

LD

AXL�

Z

Large Z FitHartree-FockPBE CorrectionB88 Correction

FIG. 7. The non-local exchange energy (exchange minusLDAX) per electron of atoms with atomic number Z (com-pare with Fig. 3). The PBE functional tends to the theoret-ical limit (Z → ∞) (horizontal green line), but B88 is moreaccurate for Z < 50 because of fitting[107].

derived functional is PBE[28], and the derivation of theparameter in B88 is given in Ref. [107].

This true story both validates Becke’s original proce-dure and the semiclassical approach to density functionalapproximation. Note that even the correction is evalu-ated on the TF density to find the limiting behavior.The PBE exchange functional also yields the leading thecorrection to the exchange energy of atoms. By throwingthis away and restoring the (different) gradient expansionfor slowly-varying gases, PBEsol was created[108].

Semiclassical approximations: New approximationsdriven by semiclassical research can be divided into den-sity approaches and potential approaches. In the den-sity camp, we find innovations like the approximationsby Armiento and Mattsson[109–111], which incorporatesurface conditions through their semiclassical approach.In the potential functional camp, we find highly ac-curate approximations to the density, which automati-cally generate approximations to non-interacting kineticenergies[65, 66, 112]. Since these approaches use poten-tial functionals, they are orbital-free and incredibly ef-ficient, but only apply in one dimension (see also Sec.7). Current research is focused on extension to three di-mensions, semiclassical approximations in the presence ofclassical turning points, as well as semiclassical approxi-mations to exchange and correlation energies.

7. WARM DENSE MATTER: A HOT NEWAREA?

Though we do not live at icy absolute zero, most chem-istry and physics happens at low enough temperaturesthat electrons are effectively in their ground state. Mostresearchers pretend to be at zero temperature for theirDFT work with inpunity. But some people, either thoseworking at high enough temperatures and pressures or

7


those interested in low-energy transitions, can’t ignorethermal effects. Those of us caught up in these warmerpursuits must tease out where temperature matters forour quantum mechanical work.

Mermin proved a finite-temperature version of theHohenberg-Kohn theorem in 1965[113], and the finite-temperature LDA was shown in the original Kohn-Shampaper[21]. However, many people continue to rely on thezero-temperature approximations, though they populatestates at higher energy levels using finite-temperatureweightings. Better understanding and modeling of thefinite-temperature XC hole could lead to improvement insome of the finer details of these calculations, like opticaland electronic properties[114].

WDM and MD: One area that has seen great recentprogress with DFT is the study of warm dense mat-ter (WDM)[115, 116]. WDM is intermediate to solidsand plasmas, inhabiting a world where both quantumand classical effects are important. It is found deepwithin planetary interiors, during shock physics exper-iments, and on the path to ignition of inertial confine-ment fusion. Lately, use of DFT MD has been a boonto researchers working to simulate these complicatedmaterials[117–121]. Most of these calculations are per-formed using Kohn-Sham orbitals with thermal occupa-tions, ignoring any temperature dependence of XC, inhopes that the kinetic and Coulomb energies will capturemost of the thermal effects. Agreement with experimenthas been excellent, though there is great interest in see-ing if temperature-dependent XC approximations affectsthese results.

0.0 0.2 0.4 0.6 0.8 1.00.0

0.5

1.0

1.5

2.0

x

Ele

ctr

on

De

nsity

TepidWarmHotSizzling

FIG. 8. The density of a single electron in a flat box spreadstoward the infinite walls as temperatures rise.

Exact conditions: Exact conditions have beenderived[114, 122–124] for finite-temperature systems thatseem very similar to their ground-state counterparts.However, a major difference in thermal systems is thatwhen one squeezes or compresses the length scale of thesystem, one sees an accompanying scaling of the temper-ature. This is further reflected in the thermal adiabaticconnection, which connects the non-interacting Kohn-

Sham system to the interacting system through scalingof the electron-electron interaction. At zero tempera-ture, this allows us to write the XC energy in terms ofthe potential alone, as long as it is accompanied by ap-propriate squeezing or stretching of the system’s lengthscale (see Sec. 5). With the temperature-coordinate scal-ing present in thermal ensembles, the thermal adiabaticconnection requires not only length scaling, but also thecorrect temperature scaling.

0

0.2

0.4

0.6

0.8

1

1.2

0 2 4 6 8 10

De

nsity

x

ex, zero TexPFATF

FIG. 9. Eight electrons in the potential −2 sin2 (πx/10) ina 1d box. At zero temperature (gray), the density exhibitssharp quantum oscillations, which wash out as the tempera-ture increases (black). This effect is much weaker near theedges. TF is used in many warm simulations, but (green)misses all oscillations, vital for accurate chemical effects. Theorbital-free, finite-temperature potential functional approxi-mation of Ref. [125] is almost exact here (red).

OF Methods: Orbital-free methods, discussed in Sec.4, are of particular interest in the WDM community.Solving the KS equations with many thermally populatedorbitals is repeated over and over in quantum molecu-lar dynamics, leading to prohibitive cost as temperaturesrise. The focus on free energies for thermal ensembles hasled to two different approaches to orbital-free approxi-mations. One uses two separate forms, one for kineticcontributions and one for entropic contributions[124].Following this path, one can either make approxima-tions empirically[126] or non-empirically[127]. Anotherapproach enforces a particular type of response in theuniform gas limit[128]. If one wishes to approximatethe kentropy[114] as a whole, one can use temperature-dependent PFT to generate highly accurate approxima-tions from approximate densities generated semiclassi-cally or stochastically[125, 129]. Fig. 9 shows the accu-racy of a semiclassical density approximation, which cap-tures the quantum oscillations missed by Thomas-Fermiand still present as temperatures rise.

8


8. WHAT CAN WE GUESS ABOUT THEFUTURE?

The future of DFT remains remarkably bright. As Fig.2 shows, the number of applications continues to grow ex-ponentially, with three times as much activity than pre-viously realized (Fig. 1 of [7]). While empiricism hasgenerated far too many possible alternatives, the stan-dard well-derived approximations continue to dominate.

To avoid losing insight, it is important to further de-velop the systematic path to approximations, which es-chews all empiricism and expands the functional in pow-ers of ~, Planck’s constant. This will ultimately tell uswhat we can and cannot do with local-type approxima-tions. There is huge room for development in this area,and any progress could impact all those applications.

Meanwhile, new areas have been (e.g. weak in-teractions) or are being developed (warm dense mat-

ter). New methods, such as using Bayesian statisticsfor error analysis[130] or machine learning for findingfunctionals[63, 131], are coming on line. Such methodswill not suffer the limitations of local approximations,and should be applicable to strongly correlated electronicsystems, an arena where many of our present approxima-tions fail. We have little doubt that DFT will continueto thrive for decades to come.

9. ACKNOWLEDGMENTS

APJ thanks the U.S. Department of Energy (DE-FG02-97ER25308), and KB thanks the National ScienceFoundation (CHE-1112442). We are grateful to CyrusUmrigar for data on the helium atom and to Min-CheolKim for Fig. 4 and Attila Cangi for Fig. 9.

[1] Shaun D. Wong, Martin Srnec, Megan L. Matthews,Lei V. Liu, Yeonju Kwak, Kiyoung Park, Caleb B.Bell III, E. Ercan Alp, Jiyong Zhao, Yoshitaka Yoda,Shinji Kitao, Makoto Seto, Carsten Krebs, J. MartinBollinger, and Edward I. Solomon. Elucidation of theFe(iv)=O intermediate in the catalytic cycle of the halo-genase SyrB2. Nature, 499(7458):320–323, 07 2013.

[2] M. D. Knudson, M. P. Desjarlais, R. W. Lemke, T. R.Mattsson, M. French, N. Nettelmann, and R. Redmer.Probing the interiors of the ice giants: Shock compres-sion of water to 700 gpa and 3.8 g/cm3. Phys. Rev.Lett., 108:091102, Feb 2012.

[3] D. Rappoport, N. R. M. Crawford, F. Furche, andK. Burke. Which functional should i choose? In E.I.Solomon, R.B. King, and R.A. Scott, editors, Compu-tational Inorganic and Bioinorganic Chemistry. Wiley,John & Sons, Inc., 2009.

[4] Kieron Burke and Lucas O. Wagner. Dft in a nutshell.Int. J. Quant. Chem., 2012.

[5] R. G. Parr and W. Yang. Density Functional Theory ofAtoms and Molecules. Oxford University Press, 1989.

[6] R. M. Dreizler and E. K. U. Gross. Density Func-tional Theory: An Approach to the Quantum Many-Body Problem. Springer–Verlag, Berlin, 1990.

[7] K. Burke. Perspective on density functional theory. J.Chem. Phys., 136, 2012.

[8] Robert O. Jones. Density functional theory: A personalview. In Adolfo Avella and Ferdinando Mancini, editors,Strongly Correlated Systems, volume 171 of Springer Se-ries in Solid-State Sciences, pages 1–28. Springer BerlinHeidelberg, 2012.

[9] Andrew Zangwill. The education of walter kohn andthe creation of density functional theory. Archive forHistory of Exact Sciences, pages 1–74, 2014.

[10] E. Schrodinger. An undulatory theory of the mechanicsof atoms and molecules. Phys. Rev., 28:1049–1070, Dec1926.

[11] L. H. Thomas. The calculation of atomic fields. Math.Proc. Camb. Phil. Soc., 23(05):542–548, 1927.

[12] E. Fermi. Rend. Acc. Naz. Lincei, 6, 1927.

[13] E. Fermi. Eine statistische Methode zur Bestimmungeiniger Eigenschaften des Atoms und ihre Anwendungauf die Theorie des periodischen Systems der Elemente(a statistical method for the determination of someatomic properties and the application of this methodto the theory of the periodic system of elements).Zeitschrift fur Physik A Hadrons and Nuclei, 48:73–79,1928.

[14] J. C. Slater. Phys. Rev., 81:385, 1951.[15] V. Fock. Naherungsmethode zur losung des quanten-

mechanischen mehrkorperproblems. Z. Phys., 61:126–148, 1930.

[16] D. R. Hartree and W. Hartree. Self-consistent field, withexchange, for beryllium. Proceedings of the Royal So-ciety of London. Series A - Mathematical and PhysicalSciences, 150(869):9–33, 1935.

[17] P. Hohenberg and W. Kohn. Inhomogeneous electrongas. Phys. Rev., 136(3B):B864–B871, Nov 1964.

[18] Edward Teller. On the stability of molecules in theThomas-Fermi theory. Rev. Mod. Phys., 34(4):627–631,Oct 1962.

[19] Stefan Kurth and John P. Perdew. Role of the exchange-correlation energy: Nature’s glue. International Journalof Quantum Chemistry, 77(5):814–818, 2000.

[20] Rodney J. Bartlett and Monika Musial. Coupled-clustertheory in quantum chemistry. Rev. Mod. Phys., 79:291–352, Feb 2007.

[21] W. Kohn and L. J. Sham. Self-consistent equations in-cluding exchange and correlation effects. Phys. Rev.,140(4A):A1133–A1138, Nov 1965.

[22] J.P. Perdew. Density functional approximation for thecorrelation energy of the inhomogeneous gas. Phys. Rev.B, 33:8822, 1986.

[23] A. D. Becke. Density-functional exchange-energy ap-proximation with correct asymptotic behavior. Phys.Rev. A, 38(6):3098–3100, Sep 1988.

[24] Axel D. Becke. Density-functional thermochemistry. iii.the role of exact exchange. The Journal of ChemicalPhysics, 98(7):5648–5652, 1993.

9


[25] J. P. Perdew and K. Schmidt. In V. E. Van Doren,K. Van Alsenoy, and P. Geerlings, editors, DensityFunctional Theory and Its Applications to Materials.American Institute of Physics, Melville, NY, 2001.

[26] John P. Perdew and Stefan Kurth. Density Functionalsfor Non-relativistic Coulomb Systems in the New Cen-tury, pages 1–55. Springer, Berlin / Heidelberg, 2003.

[27] Jianmin Tao, John P. Perdew, Viktor N. Staroverov,and Gustavo E. Scuseria. Climbing the density func-tional ladder: Nonempirical meta–generalized gradientapproximation designed for molecules and solids. Phys.Rev. Lett., 91(14):146401, Sep 2003.

[28] John P. Perdew, Kieron Burke, and Matthias Ernzer-hof. Generalized gradient approximation made simple.Phys. Rev. Lett., 77(18):3865–3868, Oct 1996. ibid. 78,1396(E) (1997).

[29] K. Burke J.P. Perdew and M. Ernzerhof. Perdew, burke,and ernzerhof reply. Phys. Rev. Lett., 80:891, 1998.

[30] Chengteh Lee, Weitao Yang, and Robert G. Parr. Devel-opment of the colle-salvetti correlation-energy formulainto a functional of the electron density. Phys. Rev. B,37(2):785–789, Jan 1988.

[31] C.F. Fischer. A multi-configuration Hartree-Fock pro-gram. Computer Physics Communications, 1(3):151 –166, 1969.

[32] C.F. Fischer. Hartree–Fock method for atoms. A nu-merical approach. Wiley, Jan 1977.

[33] P. A. M. Dirac. Note on exchange phenomena in theThomas atom. Mathematical Proceedings of the Cam-bridge Philosophical Society, 26(03):376–385, 1930.

[34] Julian Schwinger. Thomas-fermi model: The secondcorrection. Phys. Rev. A, 24(5):2353–2361, Nov 1981.

[35] Nicholas C. Handy and Aron J. Cohen. Left-right corre-lation energy. Molecular Physics, 99(5):403–412, 2001.

[36] In fact, sadly, the young chemist is unable to find a per-manent position, and she ends up selling parametrizedfunctionals for food on the streets. On the other hand,the physicists all celebrate their triumph over empiri-cism with a voyage on a brand new ship, which has beendesigned with materials whose properties have been cal-culated using DFT. Because the local approximation,as given above, underestimates the magnitude of theexchange energy, the brittle transition temperature isoverestimated. When the new ship sails through icy wa-ters, its hull is weakened and damaged by an iceberg,so all of them drown. (The interested reader may findmore information on the ductile-to-brittle transition inRef. [132] and other works by Kaxiras.).

[37] Carlos Fiolhais, F. Nogueira, and M. Marques. APrimer in Density Functional Theory. Springer-Verlag,New York, 2003.

[38] Min-Cheol Kim, Eunji Sim, and Kieron Burke. Un-derstanding and reducing errors in density functionalcalculations. Phys. Rev. Lett., 111:073003, Aug 2013.

[39] Min-Cheol Kim, Eunji Sim, and Kieron Burke. Ionsin solution: Density corrected density functional the-ory (dc-dft). The Journal of Chemical Physics,140(18):18A528, 2014.

[40] E. J. Baerends, O. V. Gritsenko, and R. van Meer. TheKohn-Sham gap, the fundamental gap and the opti-cal gap: the physical meaning of occupied and virtualKohn-Sham orbital energies. Phys. Chem. Chem. Phys.,15:16408–16425, 2013.

[41] Jochen Heyd, Gustavo E. Scuseria, and MatthiasErnzerhof. Hybrid functionals based on a screenedcoulomb potential. The Journal of Chemical Physics,118(18):8207–8215, 2003.

[42] Leeor Kronik, Tamar Stein, Sivan Refaely-Abramson,and Roi Baer. Excitation gaps of finite-sized systemsfrom optimally tuned range-separated hybrid function-als. Journal of Chemical Theory and Computation,8(5):1515–1531, 2012.

[43] Sivan Refaely-Abramson, Sahar Sharifzadeh, ManishJain, Roi Baer, Jeffrey B. Neaton, and Leeor Kronik.Gap renormalization of molecular crystals from density-functional theory. Phys. Rev. B, 88:081204, Aug 2013.

[44] R.O. Jones and O. Gunnarsson. The density functionalformalism, its applications and prospects. Rev. Mod.Phys., 61:689, 1989.

[45] J.F. Dobson E.K.U. Gross and M. Petersilka. Densityfunctional theory of time-dependent phenomena. Topicsin Current Chemisty, 181:81, 1996.

[46] D.C. Langreth Y. Andersson and B.I. Lunqvist. van derwaals interactions in density-functional theory. Phys.Rev. Lett., 76:102, 1996.

[47] M. Dion, H. Rydberg, E. Schroder, D. C. Langreth, andB. I. Lundqvist. Van der Waals density functional forgeneral geometries. Phys. Rev. Lett., 92(24):246401, Jun2004.

[48] Jose M Soler, Emilio Artacho, Julian D Gale, AlbertoGarcıa, Javier Junquera, Pablo Ordejon, and DanielSanchez-Portal. The SIESTA method for ab initioorder-N materials simulation. Journal of Physics: Con-densed Matter, 14(11):2745, 2002.

[49] Kyuho Lee, Andre K. Kelkkanen, Kristian Berland,Stig Andersson, David C. Langreth, Elsebeth Schroder,Bengt I. Lundqvist, and Per Hyldgaard. Evaluation of adensity functional with account of van der waals forcesusing experimental data of h2 physisorption on cu(111).Phys. Rev. B, 84:193408, Nov 2011.

[50] Stefan Grimme. Semiempirical gga-type density func-tional constructed with a long-range dispersion correc-tion. Journal of Computational Chemistry, 27(15):1787–1799, 2006.

[51] Petr Jurecka, Jiri Sponer, Jiri Cerny, and Pavel Hobza.Benchmark database of accurate (MP2 and CCSD(T)complete basis set limit) interaction energies of smallmodel complexes, DNA base pairs, and amino acidpairs. Phys. Chem. Chem. Phys., 8:1985–1993, 2006.

[52] Alexandre Tkatchenko and Matthias Scheffler. Accuratemolecular van der waals interactions from ground-stateelectron density and free-atom reference data. Phys.Rev. Lett., 102:073005, Feb 2009.

[53] Guo-Xu Zhang, Alexandre Tkatchenko, Joachim Paier,Heiko Appel, and Matthias Scheffler. van der waals in-teractions in ionic and semiconductor solids. Phys. Rev.Lett., 107:245501, Dec 2011.

[54] A. Warshel and M. Levitt. Theoretical studies ofenzymic reactions: Dielectric, electrostatic and stericstabilization of the carbonium ion in the reaction oflysozyme. Journal of Molecular Biology, 103(2):227 –249, 1976.

[55] Michael Levitt. The birth of computational structuralbiology. Nature Structural Biology, 8:392–393, 2001.

[56] Martin Karplus. Spinach on the ceiling: A theoreti-cal chemist’s return to biology. Annual Review of Bio-physics and Biomolecular Structure, 35(1):1–47, 2006.

10


[57] R. Car and M. Parrinello. Unified approach for molecu-lar dynamics and density-functional theory. Phys. Rev.Lett., 55(22):2471–2474, Nov 1985.

[58] Radu Iftimie, Peter Minary, and Mark E. Tuckerman.Ab initio molecular dynamics: Concepts, recent devel-opments, and future trends. Proceedings of the NationalAcademy of Sciences of the United States of America,102(19):6654–6659, 2005.

[59] Y. A. Wang and E. A. Carter. Orbital-free kinetic-energy density functional theory. In S. D. Schwartz,editor, Theoretical Methods in Condensed Phase Chem-istry, chapter 5, page 117. Kluwer, Dordrecht, 2000.

[60] V. V. Karasiev, R. S. Jones, S. B. Trickey, and Frank E.Harris. Properties of constraint-based single-point ap-proximate kinetic energy functionals. Phys. Rev. B,80:245120, Dec 2009.

[61] V. V. Karasiev, R. S. Jones, S. B. Trickey, and Frank E.Harris. Erratum: Properties of constraint-based single-point approximate kinetic energy functionals [phys. rev.b 80, 245120 (2009)]. Phys. Rev. B, 87:239903, Jun2013.

[62] V.V. Karasiev and S.B. Trickey. Issues and challengesin orbital-free density functional calculations. ComputerPhysics Communications, 183(12):2519 – 2527, 2012.

[63] John C. Snyder, Matthias Rupp, Katja Hansen, Klaus-Robert Mueller, and Kieron Burke. Finding densityfunctionals with machine learning. Phys. Rev. Lett.,108:253002, 2012.

[64] W. Koch and M. C. Holthausen. A Chemist’s Guideto Density Functional Theory. Wiley-VCH, Weinheim,second edition, 2002.

[65] Attila Cangi, Donghyung Lee, Peter Elliott, KieronBurke, and E. K. U. Gross. Electronic structure viapotential functional approximations. Phys. Rev. Lett.,106:236404, Jun 2011.

[66] Attila Cangi, E. K. U. Gross, and Kieron Burke. Poten-tial functionals versus density functionals. Phys. Rev.A, 88(6), 2013.

[67] Hai Lin and DonaldG. Truhlar. Qm/mm: what have welearned, where are we, and where do we go from here?Theoretical Chemistry Accounts, 117(2):185–199, 2007.

[68] Peter Elliott, Kieron Burke, Morrel H. Cohen, andAdam Wasserman. Partition density-functional theory.Phys. Rev. A, 82(2):024501, Aug 2010.

[69] Morrel Cohen and Adam Wasserman. On hardness andelectronegativity equalization in chemical reactivity the-ory. Journal of Statistical Physics, 125:1121–1139, 2006.10.1007/s10955-006-9031-0.

[70] Morrel H. Cohen and Adam Wasserman. On the foun-dations of chemical reactivity theory. The Journal ofPhysical Chemistry A, 111(11):2229–2242, 2007.

[71] L.N. Oliveira E.K.U. Gross and W. Kohn. Density-functional theory for ensembles of fractionally occupiedstates. i. basic formalism. Phys. Rev. A, 37:2809, 1988.

[72] A. Pribram-Jones, Z.-H. Yang, J. R. Trail, K. Burke,R. J. Needs, and C. A. Ullrich. Excitations and bench-mark ensemble density functional theory for two elec-trons. J. Chem. Phys., 140:18A541, 2014.

[73] Rougang Tang, Jonathan Nafziger, and Adam Wasser-man. Fragment occupations in partition density func-tional theory. Phys. Chem. Chem. Phys., 14:7780–7786,2012.

[74] Frederick R. Manby, Martina Stella, Jason D. Good-paster, and Thomas F. Miller. A simple, exact density-

functional-theory embedding scheme. Journal of Chem-ical Theory and Computation, 8(8):2564–2568, 2012.

[75] Taylor A. Barnes, Jason D. Goodpaster, Frederick R.Manby, and Thomas F. Miller. Accurate basis set trun-cation for wavefunction embedding. The Journal ofChemical Physics, 139(2), 2013.

[76] Jason D. Goodpaster, Taylor A. Barnes, Frederick R.Manby, and Thomas F. Miller. Accurate and systemat-ically improvable density functional theory embeddingfor correlated wavefunctions. The Journal of ChemicalPhysics, 140(18), 2014.

[77] Gerald Knizia and Garnet Kin-Lic Chan. Density ma-trix embedding: A simple alternative to dynamicalmean-field theory. Phys. Rev. Lett., 109:186404, Nov2012.

[78] Gerald Knizia and Garnet Kin-Lic Chan. Density ma-trix embedding: A strong-coupling quantum embeddingtheory. arXiv:1212.2679, 2013.

[79] Steven R. White. Density matrix formulation forquantum renormalization groups. Phys. Rev. Lett.,69(19):2863–2866, Nov 1992.

[80] Steven R. White. Density-matrix algorithms for quan-tum renormalization groups. Phys. Rev. B, 48:10345–10356, Oct 1993.

[81] D.C. Langreth and J.P. Perdew. The exchange-correlation energy of a metallic surface. Solid StateCommun., 17:1425, 1975.

[82] O. Gunnarsson and B.I. Lundqvist. Exchange and cor-relation in atoms, molecules, and solids by the spin-density-functional formalism. Phys. Rev. B, 13:4274,1976.

[83] K. Burke M. Ernzerhof and J.P. Perdew. Long-rangeasymptotic behavior of ground-state wavefunctions. J.Chem. Phys., 105:2798, 1996.

[84] M. Ernzerhof and J.P. Perdew. Generalized gradientapproximation to the angle- and system-averaged ex-change hole. J. Chem. Phys., 109:3313, 1998.

[85] Ad Becke. Density-functional thermochemistry .5. sys-tematic optimization of exchange-correlation function-als. Journal Of Chemical Physics, 107(20):8554–8560,Nov 22 1997.

[86] S. H. Vosko K. A. Jackson M. R. Pederson D.J. SinghJ. P. Perdew, J. A. Chevary and C. Fiolhais. Atoms,molecules, solids, and surfaces: Applications of the gen-eralized gradient approximation for exchange and cor-relation. Phys. Rev. B, 46:6671, 1992.

[87] Jp Perdew, Ja Chevary, Sh Vosko, Ka Jackson, Mr Ped-erson, Dj Singh, and C Fiolhais. Atoms, molecules,solids, and surfaces - applications of the generalized gra-dient approximation for exchange and correlation (vol46, pg 6671, 1992). Physical Review B, 48(7):4978, Aug15 1993.

[88] Kieron Burke, John P. Perdew, and Y. Wang. Deriva-tion of a generalized gradient approximation: The PW91density functional, page 81. Plenum, NY, 1997.

[89] A. C. Cancio and C. Y. Fong. Scaling properties ofexchange and correlation holes of the valence shell ofsecond-row atoms. Phys. Rev. A, 85:042515, Apr 2012.

[90] Erin R. Johnson and Axel D. Becke. Van der waals in-teractions from the exchange hole dipole moment: Ap-plication to bio-organic benchmark systems. ChemicalPhysics Letters, 432(4 - 6):600 – 603, 2006.

[91] Axel D. Becke and Erin R. Johnson. A unifieddensity-functional treatment of dynamical, nondynami-

11


cal, and dispersion correlations. The Journal of Chem-ical Physics, 127(12), 2007.

[92] Axel D. Becke and Erin R. Johnson. Exchange-holedipole moment and the dispersion interaction revisited.The Journal of Chemical Physics, 127(15), 2007.

[93] Alberto Otero-de-la Roza, Blessing Huynh Cao, Ivy K.Price, Jason E. Hein, and Erin R. Johnson. Predictingthe relative solubilities of racemic and enantiopure crys-tals by density-functional theory. Angewandte ChemieInternational Edition, 53(30):7879–7882, 2014.

[94] A.D. Becke and E.R. Johnson. Exchange-hole dipolemoment and the dispersion interaction: High-order dis-persion coefficients. J. Chem. Phys., 124:014104, 2006.

[95] M. Fuchs, Y.-M. Niquet, X. Gonze, and K. Burke. De-scribing static correlation in bond dissociation by kohn–sham density functional theory. The Journal of Chem-ical Physics, 122(9):094116, 2005.

[96] Aron J. Cohen, Paula Mori-Snchez, and Weitao Yang.Insights into current limitations of density functionaltheory. Science, 321(5890):792–794, 2008.

[97] Filipp Furche. Developing the random phase ap-proximation into a practical post-kohnsham correlationmodel. The Journal of Chemical Physics, 129(11), 2008.

[98] Henk Eshuis, Julian Yarkony, and Filipp Furche. Fastcomputation of molecular random phase approximationcorrelation energies using resolution of the identity andimaginary frequency integration. The Journal of Chem-ical Physics, 132(23):234114, 2010.

[99] Henk Eshuis and Filipp Furche. A parameter-free den-sity functional that works for noncovalent interactions.The Journal of Physical Chemistry Letters, 2(9):983–989, 2011.

[100] Helen van Aggelen, Yang Yang, and Weitao Yang.Exchange-correlation energy from pairing matrix fluc-tuation and the particle-particle random phase approx-imation. The Journal of Chemical Physics, 140(18),2014.

[101] Degao Peng, Stephan N. Steinmann, Helen van Agge-len, and Weitao Yang. Equivalence of particle-particlerandom phase approximation correlation energy andladder-coupled-cluster doubles. The Journal of Chemi-cal Physics, 139(10), 2013.

[102] A. Zupan P. Blaha J.P. Perdew, S. Kurth. Accurate den-sity functional with correct formal properties: A stepbeyond the generalized gradient approximation. Phys.Rev. Lett., 82:2544, 1999.

[103] Jianwei Sun, Robin Haunschild, Bing Xiao, Ireneusz W.Bulik, Gustavo E. Scuseria, and John P. Perdew.Semilocal and hybrid meta-generalized gradient approx-imations based on the understanding of the kinetic-energy-density dependence. The Journal of ChemicalPhysics, 138(4), 2013.

[104] E.H. Lieb and B. Simon. Thomas-fermi theory revisited.Phys. Rev. Lett., 31:681, 1973.

[105] Elliott H Lieb and Barry Simon. The thomas-fermi the-ory of atoms, molecules and solids. Advances in Math-ematics, 23(1):22 – 116, 1977.

[106] Elliott H. Lieb. Thomas-fermi and related theories ofatoms and molecules. Rev. Mod. Phys., 53:603–641, Oct1981.

[107] Peter Elliott and Kieron Burke. Non-empirical deriva-tion of the parameter in the b88 exchange functional.Can. J. Chem. Ecol., 87(10):1485–1491, Oct 2009.

[108] M. Ernzerhof and G. E. Scuseria. Assessmentof the Perdew–Burke–Ernzerhof exchange-correlationfunctional. J. Chem. Phys, 110:5029, 1999.

[109] R. Armiento and A.E. Mattsson. Functional designedto include surface effects in self-consistent density func-tional theory. Phys. Rev. B, 72:085108, 2005.

[110] Ann E. Mattsson and Rickard Armiento. Implementingand testing the am05 spin density functional. Phys. Rev.B, 79:155101, Apr 2009.

[111] Ann E. Mattsson and Rickard Armiento. The subsystemfunctional scheme: The armiento-mattsson 2005 (am05)functional and beyond. International Journal of Quan-tum Chemistry, 110(12):2274–2282, 2010.

[112] Attila Cangi, Donghyung Lee, Peter Elliott, and KieronBurke. Leading corrections to local approximations.Phys. Rev. B, 81(23):235128, Jun 2010.

[113] N. D. Mermin. Thermal properties of the inhomogenouselectron gas. Phys. Rev., 137:A: 1441, 1965.

[114] Aurora Pribram-Jones, Stefano Pittalis, E.K.U. Gross,and Kieron Burke. Thermal density functional theoryin context. In Frank Graziani, Michael P. Desjarlais,Ronald Redmer, and Samuel B. Trickey, editors, Fron-tiers and Challenges in Warm Dense Matter, volume 96of Lecture Notes in Computational Science and Engi-neering, pages 25–60. Springer International Publishing,2014.

[115] National Research Council Committee on High En-ergy Density Plasma Physics Plasma Science Com-mittee. Frontiers in High Energy Density Physics:The X-Games of Contemporary Science. The NationalAcademies Press, 2003.

[116] Frank Graziani, Michael P. Desjarlais, Ronald Red-mer, and Samuel B. Trickey, editors. Frontiers andChallenges in Warm Dense Matter, volume 96 of Lec-ture Notes in Computational Science and Engineering.Springer International Publishing, 2014.

[117] Thomas R. Mattsson and Michael P. Desjarlais. Phasediagram and electrical conductivity of high energy-density water from density functional theory. Phys. Rev.Lett., 97:017801, Jul 2006.

[118] Bastian Holst, Ronald Redmer, and Michael P. Desjar-lais. Thermophysical properties of warm dense hydrogenusing quantum molecular dynamics simulations. Phys.Rev. B, 77:184201, May 2008.

[119] Andre Kietzmann, Ronald Redmer, Michael P. Desjar-lais, and Thomas R. Mattsson. Complex behavior offluid lithium under extreme conditions. Phys. Rev. Lett.,101:070401, Aug 2008.

[120] M. D. Knudson and M. P. Desjarlais. Shock compressionof quartz to 1.6 tpa: Redefining a pressure standard.Phys. Rev. Lett., 103:225501, Nov 2009.

[121] Seth Root, Rudolph J. Magyar, John H. Carpenter,David L. Hanson, and Thomas R. Mattsson. Shock com-pression of a fifth period element: Liquid xenon to 840gpa. Phys. Rev. Lett., 105(8):085501, Aug 2010.

[122] Helmut Eschrig. T > 0 ensemble-state density func-tional theory via legendre transform. Phys. Rev. B,82:205120, Nov 2010.

[123] S. Pittalis, C. R. Proetto, A. Floris, A. Sanna,C. Bersier, K. Burke, and E. K. U. Gross. Exact con-ditions in finite-temperature density-functional theory.Phys. Rev. Lett., 107:163001, Oct 2011.

[124] James W. Dufty and S. B. Trickey. Scaling, bounds, andinequalities for the noninteracting density functionals at

12


finite temperature. Phys. Rev. B, 84:125118, Sep 2011.[125] Attila Cangi and Aurora Pribram-Jones, 2014. in prepa-

ration.[126] Valentin V. Karasiev, Travis Sjostrom, and S. B.

Trickey. Generalized-gradient-approximation noninter-acting free-energy functionals for orbital-free densityfunctional calculations. Phys. Rev. B, 86:115101, Sep2012.

[127] Valentin V. Karasiev, Debajit Chakraborty, Olga A.Shukruto, and S. B. Trickey. Nonempirical general-ized gradient approximation free-energy functional fororbital-free simulations. Phys. Rev. B, 88:161108, Oct2013.

[128] Travis Sjostrom and Jerome Daligault. Nonlocal orbital-free noninteracting free-energy functional for warmdense matter. Phys. Rev. B, 88:195103, Nov 2013.

[129] Roi Baer, Daniel Neuhauser, and Eran Rabani. Self-averaging stochastic kohn-sham density-functional the-

ory. Phys. Rev. Lett., 111:106402, Sep 2013.[130] Andrew J. Medford, Jess Wellendorff, Aleksandra Vo-

jvodic, Felix Studt, Frank Abild-Pedersen, Karsten W.Jacobsen, Thomas Bligaard, and Jens K. Norskøv. As-sessing the reliability of calculated catalytic ammoniasynthesis rates. Science, 345(6193):197–200, 2014.

[131] John C. Snyder, Matthias Rupp, Katja Hansen, LeoBlooston, Klaus-Robert Mller, and Kieron Burke.Orbital-free bond breaking via machine learning. TheJournal of Chemical Physics, 139(22), 2013.

[132] Georg Schusteritsch and Efthimios Kaxiras. Sulfur-induced embrittlement of nickel: a first-principles study.Modeling and Simulation in Materials Science and En-gineering, 20:065007, 2012.

13

dft: a theory full of...

Documents