Density functional theory (DFT) is a framework in quantum mechanics for calculating the properties of interacting particles from their spatial density rather than from their complete many-particle wave function. Its most widely used form describes electrons in atoms, molecules, and solids. In this setting, the central variable is the electron density, a function of three spatial coordinates, even when the system contains many electrons.
DFT is a principal method of computational chemistry and condensed-matter physics because it combines a rigorous theoretical foundation with computational approximations that are practical for substantial systems. It is used to determine structures, bonding, reaction energies, magnetic behavior, and many other properties. Its foundational theorems are exact, but the functionals used in routine calculations are approximate. Understanding the distinction between the exact theory, its computational implementation, and the limitations of a chosen approximation is essential to interpreting its predictions.
Historical development and intellectual origins
The earliest density-based descriptions of electrons preceded modern DFT by several decades. In 1927, Llewellyn Thomas and Enrico Fermi independently developed a statistical model in which the electronic kinetic energy was approximated locally using the density of a uniform electron gas. Their model combined this kinetic energy with electron–nucleus attraction and classical electron–electron repulsion. It avoided explicit electronic orbitals and offered a tractable description of atoms, but its local treatment of kinetic energy could not adequately describe shell structure or ordinary molecular binding.
Paul Dirac added a local expression for exchange energy in 1930. Exchange is a consequence of the antisymmetry of an electronic wave function under interchange of electrons, rather than an additional classical force. Carl Friedrich von Weizsäcker introduced a density-gradient contribution to the kinetic energy in 1935. These developments established several ingredients that later became central to density-functional approximations: the uniform electron gas, local energy expressions, and corrections involving spatial variations of the density.
The decisive conceptual step came in 1964, when Pierre Hohenberg and Walter Kohn established that the ground-state density determines the external potential, subject to qualifications concerning degeneracy and an arbitrary additive constant. They also formulated a variational principle for the energy as a functional of the density. Unlike the Thomas–Fermi model, this theory did not depend on assuming a particular approximate kinetic-energy expression. It established that an exact density-based description was possible in principle.
In 1965, Kohn and Lu Jeu Sham introduced an auxiliary system of noninteracting electrons. This construction retained explicit one-electron orbitals to calculate most of the kinetic energy accurately, while collecting the remaining many-electron effects into an exchange–correlation functional. The resulting Kohn–Sham equations became the basis of most practical electronic DFT.
Later developments strengthened both the mathematics and the approximations. Mel Levy’s constrained-search formulation in 1979 and Elliott Lieb’s convex formulation in 1983 clarified the domain and variational structure of the universal functional. During the 1980s and 1990s, gradient-dependent and hybrid approximations made DFT increasingly useful for molecular chemistry as well as solids. The 1998 Nobel Prize in Chemistry was awarded jointly to Walter Kohn, for developing density-functional theory, and John Pople, for developing computational methods in quantum chemistry. Their contributions were distinct but helped establish electronic-structure calculation as a widely used research instrument.
Density, universal functionals, and the variational principle
For an -electron system with fixed nuclei, the electronic Hamiltonian can be written as
where is the kinetic-energy operator, is the electron–electron interaction, and is the external potential. In an isolated molecule, this potential usually describes attraction to the nuclei. Nuclear–nuclear repulsion is a separate constant for a fixed nuclear geometry.
The density gives the expected number of electrons per unit volume and satisfies
A many-electron wave function depends on the spatial and spin coordinates of all electrons. By contrast, the density has only three spatial variables. This reduction does not mean that the electrons cease to interact; rather, the difficulty of their interaction is transferred into the form of a density functional.
The Hohenberg–Kohn theorems establish, in their usual nondegenerate ground-state formulation, that the ground-state density determines the external potential up to an additive constant. With the interaction and particle number fixed, it therefore determines the Hamiltonian and ground-state properties. The associated energy functional is
Here is universal: it does not depend on the particular external potential, although it does depend on the kinetic operator and particle interaction. Universality allows the same functional framework to describe different atoms, molecules, and materials.
The constrained-search construction gives a direct definition,
with suitable extensions to statistical ensembles. The search is over admissible antisymmetric wave functions yielding the specified density. Minimizing over admissible normalized densities gives the ground-state energy. Not every arbitrary function is an acceptable density, and questions of representability—whether a density comes from a wave function or from a ground state in some potential—are important parts of the mathematical theory.
These results establish existence and a variational framework, not an explicit inexpensive formula for . They also do not make every observable a simple expression of the density. Although the exact ground-state density determines the underlying system under the theorem’s assumptions, recovering excitation spectra or detailed pair correlations may require additional theory. Moreover, an approximate functional does not generally provide an upper bound to the exact ground-state energy. The exact variational principle and the behavior of an approximate calculation must therefore be distinguished.
The Kohn–Sham construction and self-consistent solution
The Kohn–Sham construction introduces a noninteracting reference system whose density matches that of the interacting system. Its orbitals are auxiliary mathematical objects: they make the dominant kinetic-energy contribution accessible without requiring the full interacting wave function. In atomic units, the energy is decomposed as
where is the noninteracting kinetic-energy functional and
is the classical Hartree repulsion. The exchange–correlation functional contains exchange, the nonclassical part of the interaction energy, and the difference between interacting and noninteracting kinetic energies.
Variation with respect to orthonormal orbitals gives
with
The occupations include the relevant spin multiplicities or explicitly resolved spin occupations. The equations resemble one-electron Schrödinger equations, but they are coupled through the density-dependent potential.
A calculation therefore proceeds self-consistently. An initial density generates an effective potential; the orbital equations are solved; a new density is constructed; and the process is repeated until specified convergence criteria are met. Simply replacing the old density by the new one can cause oscillations, especially in metals and large cells. Density mixing, preconditioning, and extrapolation methods are used to stabilize this nonlinear procedure.
For spin-polarized systems, the functional commonly depends on separate spin densities and . This permits different occupations and effective potentials for the two spin channels. More general formulations describe noncollinear magnetization and include spin–orbit coupling. A converged result may nevertheless correspond to a metastable electronic or magnetic solution rather than the lowest-energy state, so convergence alone does not establish that the global ground state has been found.
The total energy is not simply the sum of occupied orbital eigenvalues. Such a sum counts interaction terms incorrectly unless appropriate corrections are made. Nor are all directly measurable electron-removal or addition energies. For an isolated finite system, the highest occupied eigenvalue of exact Kohn–Sham theory equals minus the first ionization energy under the standard conditions, but this special relation does not give every orbital energy a comparable interpretation.
Exchange–correlation approximations
Practical DFT depends chiefly on approximating the exchange–correlation functional. The local density approximation (LDA) uses the exchange–correlation energy per electron of a uniform electron gas evaluated at the local density:
Its spin-dependent version uses the local densities of both spin channels. Uniform-electron-gas exchange is known analytically, while widely used correlation expressions draw on accurate numerical calculations. LDA can work well beyond slowly varying systems, partly because its treatment of the exchange–correlation hole incorporates important physical constraints.
A generalized gradient approximation (GGA) adds dependence on density gradients. Prominent examples include the Becke 1988 exchange approximation and the Perdew–Burke–Ernzerhof functional of 1996, usually called PBE. GGAs distinguish environments that have the same local density but different spatial variation. They often improve bond energies and equilibrium structures relative to LDA, although improvements are property-dependent: a functional that gives better molecular atomization energies need not give better solid-state lattice constants.
Meta-GGAs incorporate further local information, commonly the positive orbital kinetic-energy density,
This quantity helps distinguish single-orbital, covalent, metallic, and weak-overlap environments. The strongly constrained and appropriately normed functional, SCAN, introduced in 2015, is a prominent example. Because many meta-GGAs depend explicitly on orbitals through , they are commonly implemented using generalized Kohn–Sham operators rather than a purely multiplicative local potential.
A hybrid functional mixes a fraction of orbital-dependent exact exchange with semilocal exchange and correlation. B3LYP became widely used in molecular chemistry during the 1990s; PBE0 uses one-quarter exact exchange with PBE ingredients. Screened hybrids such as the Heyd–Scuseria–Ernzerhof family restrict exact exchange to shorter ranges, making periodic calculations more manageable. Range-separated hybrids divide the Coulomb interaction into short- and long-range parts and can improve phenomena that depend strongly on long-range electron transfer.
Long-range dispersion requires additional treatment. Semilocal approximations cannot generally reproduce the asymptotic attraction between separated neutral fragments arising from correlated fluctuations. Common approaches include atom-pair dispersion corrections, nonlocal correlation functionals, and many-body dispersion models. These are particularly consequential for molecular crystals, layered materials, and adsorption.
Functional design balances exact constraints, reference data, computational cost, and transferability. Empirical fitting is not itself evidence of poor quality, just as satisfying many exact constraints does not guarantee accuracy for every system. Labels such as “higher rung” describe added ingredients and flexibility, not a universal ranking of predictive reliability.
Numerical representation and computational implementation
A density functional does not by itself specify a complete numerical calculation. The orbitals, density, potentials, and integrals must be represented with finite computational resources. Localized basis sets built from Gaussian functions are common in molecular chemistry because many required integrals can be evaluated efficiently. Numerical atomic orbitals provide another localized representation. Plane waves are especially natural for periodic solids, while real-space grids, finite elements, and wavelets provide alternative systematic discretizations.
In a plane-wave calculation, an energy cutoff limits the largest reciprocal-space vectors included in the basis. Increasing the cutoff improves completeness but enlarges the calculation. Localized bases instead require systematic enlargement of orbital sets, including polarization and diffuse functions where appropriate. Basis-set incompleteness can alter energy differences and geometries. In weakly bound complexes, basis-set superposition error can arise because each fragment benefits artificially from basis functions centered on the other fragment.
Core electrons introduce rapidly varying orbitals near nuclei. A pseudopotential replaces the nucleus and selected core electrons with an effective interaction acting on valence electrons. Norm-conserving and ultrasoft pseudopotentials are widely used variants. The projector augmented-wave method, introduced by Peter Blöchl in 1994, combines smooth computational functions with atom-centered reconstruction of all-electron information. Such methods require validation for the relevant bonding, pressure, and oxidation environments; a core approximation that works near ambient conditions may become inadequate when nominal core states participate in bonding.
For crystals, Bloch electronic states are labeled by wave vectors, and integrals over the Brillouin zone are approximated using finite meshes. Metals require particular care because occupations change at the Fermi energy. Smearing schemes improve numerical integration and convergence, but their widths influence computed energies and must be treated consistently. A periodically repeated molecular or surface calculation also requires enough separation, or suitable electrostatic corrections, to control interactions between artificial images.
The computational cost depends on the functional, basis, solver, and physical system. Conventional orbital-based methods often approach cubic scaling with system size in important computational stages, although no single scaling law describes every implementation. Sparse and linear-scaling methods exploit locality of electronic structure and are particularly effective in sufficiently large insulating systems. Exact exchange and advanced correlation treatments generally add substantial cost.
Numerical convergence and physical accuracy are separate issues. A tightly converged calculation can still be inaccurate because of the functional, magnetic state, neglected relativistic effects, or an inappropriate structural model. Reproducible results therefore require specifying the functional, core treatment, basis or cutoff, reciprocal-space sampling, electronic occupations, convergence tolerances, and treatment of spin and dispersion.
Energies, forces, response, and observable properties
Ground-state total energies are the principal output of conventional DFT. Differences between them yield quantities such as molecular binding energies, reaction energies, defect formation energies, and relative stability of crystal structures. Nuclear positions are usually treated within the Born–Oppenheimer approximation, which separates electronic calculations from nuclear motion. The electronic ground-state energy plus nuclear repulsion defines a potential-energy surface on which equilibrium geometries and reaction pathways can be studied.
Atomic forces are derivatives of this energy with respect to nuclear coordinates. For an exact stationary electronic solution in a suitable representation, the Hellmann–Feynman expression relates these forces to the derivative of the Hamiltonian. Basis functions that move with the atoms introduce additional Pulay terms. Consistent forces enable structural optimization and molecular dynamics, in which nuclei evolve under forces recalculated from the electronic structure. Parrinello and Car’s 1985 extended-Lagrangian method provided a different influential route to first-principles molecular dynamics by evolving auxiliary electronic degrees of freedom alongside the nuclei.
Second derivatives determine harmonic force constants and phonon frequencies. These can be obtained by finite atomic displacements or by density-functional perturbation theory, which computes the response to small disturbances without separately displacing every atom in a large supercell. Vibrational calculations support interpretation of infrared and Raman spectra and provide harmonic zero-point and thermal contributions. A zero-temperature electronic energy alone is not a finite-temperature free energy: nuclear vibrations, configurational disorder, and other thermodynamic contributions may be required.
Responses to electric fields, strain, and magnetic perturbations yield dielectric properties, elastic constants, piezoelectric coefficients, and magnetic response parameters. Care is needed when translating electronic response into experimentally measured quantities. For example, the static dielectric response of an insulating crystal may include both electronic polarization and ionic displacement, whereas a clamped-ion calculation includes only the electronic contribution.
Electronic densities can also be analyzed to describe bonding and charge redistribution. Density differences, orbital projections, and real-space partitions help identify covalent bonding, charge transfer, and changes caused by adsorption. However, assigning an atomic charge inside a molecule or solid requires a partitioning convention. Different population analyses can produce different numerical charges without disagreeing about the underlying total density.
Comparison with experiment likewise requires matching conditions. Calculated equilibrium lattice constants at negligible temperature are not identical to measurements influenced by thermal expansion. Spectral peaks can involve excitations and lifetimes absent from a ground-state calculation. DFT supplies an electronic-structure foundation, but the connection to an observable often requires a further model of vibrations, temperature, disorder, or the measurement process.
Applications in chemistry and materials science
In molecular chemistry, DFT is used to explore equilibrium conformations, bond strengths, acid–base equilibria, and reaction mechanisms. A reaction pathway can contain several intermediates and transition states whose relative energies determine plausible kinetic sequences. Calculated barriers connect electronic structure to chemical kinetics, but comparisons with solution experiments may also require solvation, thermal corrections, and treatment of conformational populations. Small errors in activation free energies can produce substantial errors in predicted rate constants because rates depend exponentially on those barriers.
Catalysis is a particularly important application. Calculations on transition-metal complexes investigate ligand effects, spin states, and competing bond-making or bond-breaking pathways. Surface calculations examine adsorption sites, dissociation barriers, and reaction intermediates on metals, oxides, and supported particles. For example, models of nitrogen activation in ammonia synthesis or carbon monoxide oxidation connect local electronic structure with elementary surface reactions. A predicted adsorption energy is one input to catalytic behavior, not a direct measurement of catalytic activity, which also depends on coverage, transport, and the operating state of the catalyst.
In solid-state physics, DFT provides equilibrium crystal structures, elastic properties, charge densities, and electronic bands. It helps explain differences between metals, semiconductors, and insulators and allows comparison of competing structural phases under pressure. Defect calculations address vacancies, interstitial atoms, substitutional impurities, and their charge states. Reliable defect energetics require chemical-potential conventions, charge-state corrections, and suitable treatment of band edges, especially when periodic supercells contain charged defects.
Energy-related applications include battery electrodes, solid electrolytes, photovoltaic absorbers, and electrocatalysts. In battery materials, calculations can estimate insertion voltages from energy differences and examine ion migration through transition-state searches. In layered solids, dispersion-aware calculations describe interlayer binding and exfoliation trends. For electrochemical interfaces, applied potential, solvent structure, and ionic screening introduce additional complications beyond a neutral calculation in vacuum.
High-throughput workflows extend these methods to large collections of candidate materials. Automated calculations can populate databases of structures, formation energies, and electronic properties, while phase-stability analyses identify compounds stable against specified decomposition pathways. Such screening is constrained by the assumed structures and chemistry: finding a low calculated energy does not establish that a material can be synthesized readily or remains stable under every operating condition.
DFT calculations also provide data for machine-learning interatomic potentials and property models. These models can extend simulation to larger sizes and longer times, but they inherit errors and coverage limitations from their reference calculations. A model trained on one functional reproduces that functional’s potential-energy surface within its learned domain; it does not automatically recover the exact electronic energy or remain reliable for unfamiliar bonding environments.
Systematic errors and difficult electronic regimes
Many common functional errors can be related to how an approximation treats electron number, spin, and localization. The classical Hartree expression contains an electron’s interaction with its own charge distribution. For a one-electron system, exact exchange cancels this contribution completely and correlation vanishes. Semilocal approximations generally do not achieve the exact cancellation, producing self-interaction error. Related many-electron delocalization errors can favor densities that spread charge too broadly, affecting molecular dissociation, charge transfer, defect localization, and reaction barriers.
For an isolated system described by the appropriate ensemble extension, the exact ground-state energy is piecewise linear between adjacent integer electron numbers. Many approximate functionals instead display curvature. Convex curvature favors fractional charge spreading, while excessive concavity can favor localization. This behavior helps explain why two fragments separated by a large distance may acquire unphysical fractional charges in an approximate calculation.
The fundamental energy gap is
where is the ionization energy and the electron affinity. In exact multiplicative Kohn–Sham theory it can be expressed as
where is the exchange–correlation derivative discontinuity. Consequently, the auxiliary orbital gap is not generally the fundamental gap, even with the exact functional. Common semilocal calculations frequently underestimate semiconductor and insulator gaps. Optical gaps introduce the additional effect of attraction between the excited electron and the hole.
Strong correlation presents another major difficulty. When several electronic configurations have comparable importance, a single auxiliary determinant combined with a simple functional can give poor energies or incorrect qualitative behavior. Stretched hydrogen molecules illustrate the problem of static correlation, while transition-metal oxides combine localized orbitals, magnetism, and competing electronic states. Allowing broken spin symmetry can lower an approximate energy and improve some structural predictions, but the resulting spin density may not have the symmetry of the exact finite-system state.
The DFT+U method adds an orbital-dependent correction for a selected localized subspace, often involving transition-metal or rare-earth states. It can improve localization and insulating behavior, but results depend on the subspace definition, interaction parameters, and subtraction of interactions already represented approximately. Hybrid functionals, density-functional embedding, and combinations with dynamical mean-field theory address related problems through different approximations. None is a universal correction.
Error assessment therefore benefits from multiple kinds of evidence: convergence tests, comparisons across physically motivated approximations, benchmark calculations, and relevant experiments. Agreement for one observable can result from cancellation between errors and does not establish comparable accuracy for another.
Time dependence, finite temperature, and extensions of the framework
Ground-state electronic DFT is part of a broader family of density-based theories. Time-dependent density functional theory (TDDFT) extends the framework to electronic dynamics and excitations. The Runge–Gross theorem of 1984 established a density–potential mapping for specified initial conditions under suitable assumptions. Time-dependent Kohn–Sham equations then propagate auxiliary orbitals in an effective potential. Linear-response formulations calculate excitation energies and oscillator strengths, while real-time methods follow the response to explicit time-dependent fields.
In practice, TDDFT often uses an adiabatic approximation: the exchange–correlation potential at a given time is evaluated from a ground-state functional at the instantaneous density. This neglects explicit memory dependence. Linear-response TDDFT is widely used for molecular spectra, but common approximations have difficulties with long-range charge-transfer excitations and states with substantial double-excitation character. Optical excitations in extended systems can also require exchange–correlation kernels with nonlocal features absent from simple local approximations.
Finite-temperature DFT was placed on a formal basis by N. David Mermin in 1965. It replaces ground-state energy minimization with an appropriate equilibrium free-energy principle and includes electronic entropy. This framework matters when electronic thermal excitations are physically significant, as in warm dense matter and high-temperature plasmas. Fractional occupations used merely to improve convergence in an ordinary metallic calculation should be distinguished from a controlled treatment of a material at a specified electronic temperature. Electronic heating also does not by itself include nuclear thermal motion.
Orbital-free DFT seeks to approximate the noninteracting kinetic-energy functional directly, avoiding the explicit Kohn–Sham orbitals. It can reduce computational cost greatly, particularly for some simple metals and high-temperature systems. Its central challenge remains the construction of transferable kinetic-energy functionals that capture shell structure, strongly inhomogeneous densities, and different bonding environments with comparable accuracy.
Additional formulations expand the choice of basic variables or the class of particles. Current-density functional theory addresses systems in which magnetic fields and electronic currents require information beyond an ordinary scalar density. Superconducting density-functional theories introduce an anomalous pair density. Nuclear energy-density functional methods describe protons and neutrons through functionals adapted to nuclear interactions. These approaches share density-based variational ideas but are not identical to conventional electronic Kohn–Sham DFT.
Research on new electronic functionals increasingly combines exact constraints, accurate reference calculations, and data-driven representations. Nonlocality, dependence on electronic history, strong correlation, and reliable uncertainty estimates remain active issues. Machine-learned functionals must satisfy additional requirements beyond fitting reference energies, including stable functional derivatives and sensible behavior during self-consistent density optimization.