Dynamical Mean-Field Theory (DMFT)

• DMFT is a non-perturbative framework that reduces many-body lattice problems to a self-consistent quantum impurity model, preserving full frequency-dependent local dynamics.
• It accurately captures key phenomena such as the Mott–Hubbard transition, quasiparticle renormalization, and spectral weight redistribution through a local self-energy.
• Modern applications extend DMFT to combine with DFT, cluster methods, and advanced impurity solvers to model correlated electron materials and multiorbital systems.

Dynamical mean-field theory (DMFT) is a non-perturbative mean-field theory for correlated lattice fermions that becomes exact in the limit of infinite lattice coordination or spatial dimension. It reduces the full lattice many-body problem to a self-consistent quantum impurity problem, namely a single interacting site embedded in a dynamical bath representing the rest of the lattice. In this formulation the self-energy is local but fully frequency dependent, so local quantum fluctuations are retained exactly while spatial dependence beyond one site is neglected. DMFT therefore occupies an intermediate position between static mean-field theories and fully nonlocal many-body treatments, and it has become a standard framework for correlated models, realistic materials, and several controlled extensions beyond the single-site approximation (Vollhardt, 2010, Vollhardt, 2019, Vollhardt et al., 2011).

1. Physical content and scope

The motivating notion is electronic correlation in the many-body sense: expectation values do not generally factorize, so that

$$\langle AB\rangle \neq \langle A\rangle \langle B\rangle.$$

For the Hubbard model this appears already in the on-site double occupancy,

$$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$

When the local Coulomb repulsion $U$ is comparable to or larger than the bandwidth $W$, these correlations strongly suppress double occupancy and qualitatively reorganize the electronic spectrum. Static mean-field decouplings, such as Hartree or Hartree–Fock,

$$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$

replace the interaction by a static site-dependent potential and therefore miss intrinsically dynamical phenomena such as the Mott–Hubbard metal–insulator transition, mass renormalization, heavy quasiparticles, and the coexistence of a narrow quasiparticle peak with lower and upper Hubbard bands (Vollhardt, 2010, Vollhardt, 2019).

The canonical lattice model is the single-band spin-$\tfrac12$ Hubbard Hamiltonian,

$$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$

with hopping $t_{ij}$ and on-site repulsion $U$. At half filling and strong $U/W$, this model displays the Mott–Hubbard transition, Hubbard bands, and strong quasiparticle renormalization, all of which are central targets of DMFT. The same local-correlation logic extends to density–density interactions and multiorbital Hamiltonians with intra-orbital $$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$0, inter-orbital $$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$1, and Hund’s coupling $$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$2, which is why DMFT is naturally applicable to transition-metal and rare-earth systems with narrow $$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$3 or $$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$4 bands (Vollhardt, 2010).

A common misconception is that DMFT is merely a phenomenological ansatz. For models with purely local interactions, its locality assumption is not ad hoc in the large-dimension or large-coordination limit. The method is best understood as a controlled mean-field theory for local correlations, with a systematic small parameter $$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$5 or $$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$6, rather than as a static mean-field approximation in the Hartree–Fock sense (Vollhardt, 2010, Vollhardt et al., 2011).

2. High-dimensional construction and locality of the self-energy

The formal construction starts from the high-dimensional limit. On a hypercubic lattice with nearest-neighbor hopping,

$$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$7

one must scale the hopping as

$$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$8

or equivalently $$\langle \hat n_{i\uparrow}\hat n_{i\downarrow} \rangle \neq \langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle.$$9, so that kinetic and interaction energies remain comparable as $U$0. With this quantum scaling, the noninteracting density of states remains finite; for the hypercubic lattice it tends to a Gaussian,

$U$1

The off-diagonal propagators then vanish with distance,

$U$2

so diagrams with spatially separated external vertices are suppressed (Vollhardt, 2010, Vollhardt et al., 2011).

The decisive consequence is the collapse of the irreducible self-energy onto a single site: $U$3 Thus the self-energy loses all momentum dependence but retains its full frequency dependence. This is the central technical simplification behind DMFT. Locality makes the interacting lattice Green’s function

$U$4

and the local Green’s function

$U$5

depend only on the noninteracting density of states and on the local self-energy (Vollhardt, 2010).

Because time remains one-dimensional, the limit is still a genuine dynamical many-body problem. DMFT is therefore not a static large-$U$6 approximation but a local dynamical theory. That distinction underlies its ability to connect Fermi-liquid renormalization at low energies to incoherent Hubbard bands at high energies within a single self-consistent framework (Vollhardt, 2019).

3. Impurity mapping and self-consistency equations

Locality of $U$7 implies that the lattice problem can be represented by a single interacting site embedded in an effective medium. In the cavity construction, one removes a site, integrates out the rest of the lattice, and obtains a local effective action

$U$8

where $U$9 is the Weiss field. Equivalently, the impurity may be written as an Anderson model with bath levels $W$0, hybridizations $W$1, and hybridization function

$W$2

This impurity problem captures the full local dynamics: electrons hop into and out of the impurity and interact locally via $W$3 (Vollhardt, 2010, Vollhardt et al., 2011).

The impurity Green’s function is

$W$4

and the impurity self-energy is defined by the local Dyson equation

$W$5

The lattice Green’s function with that same self-energy is

$W$6

and its local part is

$W$7

The DMFT self-consistency condition is

$W$8

which determines $W$9 or, equivalently, $$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$0 (Vollhardt, 2010, Vollhardt et al., 2011).

Operationally, the paramagnetic DMFT loop begins with a guess for $$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$1 or $$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$2, computes $$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$3 and $$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$4, updates the Weiss field through

$$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$5

solves the impurity problem, extracts a new self-energy,

$$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$6

and iterates to convergence. For the Bethe lattice with semicircular density of states,

$$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$7

the self-consistency simplifies to

$$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$8

or equivalently

$$\hat n_{i\uparrow}\hat n_{i\downarrow}\to \hat n_{i\uparrow}\langle \hat n_{i\downarrow}\rangle+ \hat n_{i\downarrow}\langle \hat n_{i\uparrow}\rangle -\langle \hat n_{i\uparrow}\rangle\langle \hat n_{i\downarrow}\rangle,$$9

which is one of the best-known closed forms in single-site DMFT (Vollhardt, 2010).

4. Impurity solvers and computational realizations

The impurity solver is the computational bottleneck. Quantum Monte Carlo, including Hirsch–Fye and CT-QMC, is numerically exact on the imaginary-time or Matsubara axis and handles arbitrary interaction strength, but it requires analytic continuation to obtain real-frequency spectra, and that continuation is ill-posed. Numerical renormalization group has very high resolution at low energies and is particularly effective for zero-temperature and low-temperature properties, but is less accurate at high frequencies and more difficult in multiorbital settings. Exact diagonalization discretizes the bath with a finite number of levels and provides direct access to real-frequency information, but finite bath size leads to coarse spectral functions and finite-size effects. Iterated perturbation theory is a semi-analytic approximation that interpolates between weak and strong coupling, while DMRG-based impurity solvers are efficient for one-dimensional bath representations (Vollhardt, 2010, Vollhardt, 2019, Vollhardt et al., 2011).

The practical structure of the solver determines which part of the DMFT loop is most expensive. In exact-diagonalization-type approaches the hybridization function is approximated by a finite set of bath levels, so bath fitting and bath discretization errors must be controlled directly. The ASCI-DMFT approach adapts adaptive sampling configuration interaction to this role and is reported to be several orders of magnitude faster than the current best published ground state DMFT simulations, making it possible to study bath discretization error and cluster sizes beyond the current state of the art in zero-temperature DMFT (Mejuto-Zaera et al., 2017). In a different direction, a periodic ab initio DMFT formulation based on Hartree–Fock, IAO+PAO local orbitals, and coupled-cluster Green’s function impurity solvers treats impurities comprising the full unit cell or a supercell, avoids double counting by using Hartree–Fock as the low-level theory, and handles impurity problems with several hundred orbitals (Zhu et al., 2019).

For realistic multiorbital materials, modern DFT+DMFT implementations often rely on CT-QMC and compact imaginary-time representations. A multi-orbital continuous-time QMC solver in a compact Legendre representation was introduced specifically to generate accurate quasi-continuous imaginary-time Green’s functions, remove errors inherent to standard discrete-time approaches, and integrate the impurity solve into a full DFT+DMFT workflow in CASTEP (Sheridan et al., 2018). In nonequilibrium steady-state problems, an auxiliary impurity problem embedded in a Markovian environment can be formulated through a Lindblad equation, providing a nonequilibrium extension of exact-diagonalization-based DMFT on the real-frequency axis (Arrigoni et al., 2012).

From the standpoint of methodology, these developments do not alter the formal core of DMFT; they broaden the class of impurity models and frequency regimes that can be solved accurately enough to sustain the self-consistency loop in practice.

5. Characteristic results for correlated electrons

For the half-filled Hubbard model, DMFT yields the canonical three-peak evolution of the spectral function

$\tfrac12$0

At small $\tfrac12$1, the spectrum resembles the noninteracting density of states with weak renormalization. At intermediate $\tfrac12$2, a narrow quasiparticle peak appears at $\tfrac12$3 between lower and upper Hubbard bands near $\tfrac12$4. At large $\tfrac12$5, the quasiparticle peak disappears and a Mott gap opens, producing a paramagnetic Mott insulator. The low-energy quasiparticle weight is

$\tfrac12$6

and the mass enhancement is $\tfrac12$7; $\tfrac12$8 decreases as the Mott transition is approached from the metallic side and vanishes at the interaction where the metallic solution disappears (Vollhardt, 2010, Vollhardt et al., 2011).

At finite temperature, single-site DMFT yields a coexistence region $\tfrac12$9 in which both metallic and insulating solutions exist. The first-order Mott transition line $$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$0 lies inside this region and terminates at a finite-temperature critical endpoint $$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$1. Above $$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$2, there is no true phase transition but only a crossover between bad-metallic and bad-insulating regimes. Within the insulating solution, increasing temperature fills the gap with spectral weight; within the metallic solution, the quasiparticle peak is no longer pinned to its zero-temperature value. On the Bethe lattice at $$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$3, NRG yields $$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$4 for the disappearance of the metallic solution (Vollhardt, 2010).

Several further signatures are characteristic of DMFT. At $$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$5, a metallic solution obeys Luttinger pinning, so the spectral function at the Fermi level equals the noninteracting density of states at that energy. The real part of the DMFT self-energy may develop kinks below the Hubbard-band scale and above the strict Fermi-liquid regime, producing kinks in the effective dispersion $$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$6; in DMFT this mechanism is purely electronic and does not require coupling to phonons or other bosons. In frustrated or otherwise suitable lattices, DMFT also finds metallic ferromagnetism as an intermediate-coupling correlation effect, with a Curie–Weiss susceptibility and a strongly reduced Curie temperature compared to simple Stoner theory (Vollhardt, 2010, Vollhardt, 2019).

One often-cited limitation of single-site DMFT becomes especially visible in the Mott phase diagram. In the half-filled Hubbard model, the first-order transition line has a negative slope in single-site DMFT. This is an artifact of the missing magnetic exchange energy: in the large-$$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$7 limit the superexchange $$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$8, so the insulating entropy remains too large down to low temperature (Vollhardt, 2010, Vollhardt et al., 2011). The point is not that DMFT fails to capture the Mott transition, but that purely local dynamics does not suffice for all aspects of low-temperature thermodynamics.

6. Materials, extensions, and limits

The combination of band theory with DMFT, usually denoted LDA+DMFT or GGA+DMFT, starts from a one-particle Hamiltonian $$\hat{H} = \sum_{\mathbf{R}_i , \mathbf{R}_j} \sum_{\sigma} t_{ij}\hat{c}_{i \sigma}^{\dagger}\hat{c}_{j \sigma} + U \sum_{\mathbf{R}_i} \hat{n}_{i \uparrow}\hat{n}_{i \downarrow},$$9 or $t_{ij}$0, selects a correlated subspace such as transition-metal $t_{ij}$1 or rare-earth $t_{ij}$2 orbitals, adds local Coulomb and Hund terms, and solves the resulting multiorbital problem within DMFT. In this way DMFT reduces to pure DFT when interactions are weak but produces quasiparticle renormalization and Hubbard bands when local correlations are strong. Representative examples include SrVO$t_{ij}$3 and CaVO$t_{ij}$4, where LDA+DMFT yields a three-peak $t_{ij}$5 spectrum with a lower Hubbard band near $t_{ij}$6 eV, a quasiparticle peak at the Fermi level, and an upper Hubbard band near $t_{ij}$7 eV; NiO, where LDA+DMFT reproduces the insulating gap and the strong $t_{ij}$8 weight at the top of the valence band; KCuF$t_{ij}$9, where GGA+DMFT finds a strong cooperative Jahn–Teller distortion and a total-energy minimum in excellent agreement with experiment; and Fe, where DFT+DMFT recovers the bcc–fcc structural transition above the Curie temperature and realistic phonon properties (Vollhardt, 2010, Vollhardt, 2019, Zhu et al., 2019).

Single-site DMFT neglects nonlocal correlations, so it cannot describe short-range antiferromagnetic or superconducting fluctuations, pseudogap states driven by nonlocal correlations, long-wavelength critical behavior, or a momentum-dependent self-energy. Several extensions address this limitation. Cluster DMFT, CDMFT, and DCA embed a finite cluster rather than a single site and thereby capture short-range correlations and momentum selectivity. EDMFT includes intersite interactions through additional bosonic baths. Diagrammatic extensions such as D$U$0A and dual fermion add nonlocal corrections in terms of local two-particle vertices. DMFT+fRG starts the renormalization-group flow from the DMFT solution rather than from the bare action. More recently, diagrammatic Monte Carlo has been reformulated around the DMFT fixed point so that the expansion contains only explicitly nonlocal corrections, and a fluctuating dynamical mean-field theory has been proposed in which collective fluctuations of auxiliary impurity models across different sites are incorporated through functional integration (Vollhardt, 2019, Carlström, 2023, Semenov et al., 27 Feb 2026).

Disorder and nonequilibrium require distinct generalizations. Arithmetic disorder averaging reduces the noninteracting limit to the coherent potential approximation and cannot capture Anderson localization, whereas geometric averaging of the local density of states in typical-medium theory yields a mean-field description of localization and can be combined with DMFT to study the Anderson–Hubbard model. Statistical DMFT keeps a full distribution of local environments, and EDMFT has also been used to describe glassy Coulomb phases and self-organized criticality in the Coulomb glass (Miranda et al., 2011). Nonequilibrium DMFT formulates the self-consistency on the Keldysh contour and has been applied to interaction quenches, strong fields, and time-resolved spectroscopy, while auxiliary Lindblad-based impurity formulations provide a steady-state route for transport through correlated structures (Vollhardt, 2019, Arrigoni et al., 2012).

The conceptual reach of DMFT extends beyond correlated electrons. Bosonic DMFT treats the Bose–Hubbard model with distinct normal and condensed baths; a dynamical mean-field approximation has been constructed for lattice $U$1 theory and shown to work best in higher dimensions; and the same single-degree-of-freedom-in-a-self-consistent-bath logic has been used to formulate aging dynamics in fully connected disordered $U$2-spin models (Vollhardt et al., 2011, Akerlund et al., 2013, Altieri et al., 2020). These generalizations reinforce the central meaning of DMFT: a local dynamical embedding theory whose exactness is tied to a large-connectivity limit and whose practical value depends on how completely the relevant physics is encoded in local quantum dynamics.