Dynamical Mean-Field Theory

• Dynamical Mean-Field Theory (DMFT) is a nonperturbative framework for correlated quantum many-body systems that approximates a lattice problem with interacting impurity problems, embedded in self-consistent, frequency-dependent baths, particularly effective in the infinite-dimensional limit but also applicable to finite-dimensional systems.
• The theory is exact for lattice models with local interactions in the infinite-dimensional limit but becomes approximate in finite dimensions, where application is often most useful in the regime where local temporal fluctuations dominate and spatial correlations are short-ranged.
• Applications of DMFT include the study of quasiparticle renormalization, Hubbard bands, Mott metal–insulator transitions, magnetic and orbital phenomena, and correlation-induced spectral-weight transfer, all within a self-consistent formalism.
• Followup_questions:

Dynamical mean-field theory (DMFT) is a nonperturbative framework for correlated quantum many-body systems that replaces a lattice, molecular, or nanostructure problem by one or more interacting quantum impurity problems embedded in self-consistent, frequency-dependent baths. Its defining approximation is that the interaction self-energy is local in space but fully dynamical in time or frequency, $$\Sigma_{ij,\sigma}(z)\approx\delta_{ij}\Sigma_\sigma(z)$$, or equivalently $\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$. DMFT is exact for lattice models with local interactions in the infinite-dimensional or infinite-coordination limit, with finite-dimensional applications neglecting nonlocal self-energy correlations. The theory describes quasiparticle renormalization, Hubbard bands, local moments, Mott metal–insulator transitions, magnetic and orbital phenomena, and correlation-induced spectral-weight transfer within a unified self-consistent formalism (Vollhardt et al., 2011).

1. Correlated-electron foundation and conceptual structure

Electronic correlations arise when expectation values fail to factorize, for example

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

For narrow $d$- or $f$-bands, the electronic residence time on an atom is long, approximately $\tau\sim\hbar/W$, where $W$ is the bandwidth. The Coulomb repulsion can then be comparable to or larger than the kinetic-energy scale. Their competition produces effective masses, incoherent spectral weight, local moments, quasiparticle destruction, and Mott insulating behavior.

The canonical model is the Hubbard Hamiltonian,

$$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$

where $t_{ij}$ is the hopping amplitude, $U$ is the local repulsion, $\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$0 creates an electron, and $\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$1.

Static Hartree–Fock theory replaces the interaction by an average potential. In the paramagnetic Hubbard model,

$\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$2

This discards fluctuations of the local occupation. The resulting self-energy is essentially frequency independent and cannot properly describe the suppression of double occupancy or the paramagnetic Mott transition.

DMFT preserves the spatial simplification of mean-field theory while retaining temporal quantum fluctuations. Its self-energy can contain lifetimes, quasiparticle renormalization, Hubbard bands, insulating poles, and dynamical changes in local charge and spin configurations. The “mean field” is consequently not a static number but a dynamical bath with memory.

A useful distinction is between:

• Locality: the self-energy has no explicit dependence on lattice momentum or intersite indices.
• Dynamical character: the self-energy remains a nontrivial function of frequency.
• Self-consistency: the bath is generated by the lattice Green’s function and must agree with the impurity Green’s function at convergence.
• Nonperturbative local treatment: the local interaction is retained in the auxiliary impurity problem rather than expanded solely in a small interaction parameter.

2. Infinite coordination and local self-energy

DMFT becomes exact for fermionic lattice models with local interactions in the limit of infinite spatial dimension or coordination number $\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$3. To obtain a nontrivial limit, nearest-neighbor hopping must be scaled according to

$\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$4

with $\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$5 fixed as $\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$6. For a $\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$7-dimensional hypercubic lattice, $\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$8, so the scaling is equivalently $\Sigma_\sigma(\mathbf{k},z)\approx\Sigma_\sigma(z)$9.

Without this quantum scaling, the kinetic energy diverges. A classical scaling $$\langle AB\rangle\neq\langle A\rangle\langle B\rangle.$$0 instead causes the kinetic energy to vanish. With $$\langle AB\rangle\neq\langle A\rangle\langle B\rangle.$$1, individual off-diagonal propagators decrease with distance, schematically as

$$\langle AB\rangle\neq\langle A\rangle\langle B\rangle.$$2

but the number of available paths grows with coordination, so particles remain mobile and the kinetic energy remains finite.

The crucial diagrammatic consequence is that irreducible self-energy diagrams with distinct external sites vanish as $$\langle AB\rangle\neq\langle A\rangle\langle B\rangle.$$3. The self-energy becomes local: $$\langle AB\rangle\neq\langle A\rangle\langle B\rangle.$$4 or in momentum space,

$$\langle AB\rangle\neq\langle A\rangle\langle B\rangle.$$5

For the hypercubic lattice,

$$\langle AB\rangle\neq\langle A\rangle\langle B\rangle.$$6

and the scaled noninteracting density of states tends to a Gaussian,

$$\langle AB\rangle\neq\langle A\rangle\langle B\rangle.$$7

Locality does not imply localization, absence of hopping, or static behavior. The lattice Green’s function remains momentum dependent through the bare dispersion: $$\langle AB\rangle\neq\langle A\rangle\langle B\rangle.$$8 The local Green’s function is obtained by momentum averaging,

$$\langle AB\rangle\neq\langle A\rangle\langle B\rangle.$$9

In finite dimensions, single-site DMFT replaces the exact momentum-dependent self-energy by a local approximation. Its accuracy is greatest when local temporal fluctuations dominate and spatial correlations are short-ranged, including regimes of high temperature, high excitation energy, high doping, strong frustration, or substantial orbital degeneracy (Georges, 2011).

3. Impurity mapping and self-consistency equations

The locality of the self-energy permits a cavity construction. One site, conventionally site $d$0, is removed from the lattice, leaving a cavity system. Electrons can leave the selected site, propagate through the cavity, and return. The cavity is represented by a frequency-dependent hybridization function $d$1.

The effective impurity action is

$d$2

where $d$3, $d$4 and $d$5 are Grassmann fields, and $d$6 is the Weiss or bath Green’s function.

In Matsubara frequency,

$d$7

with fermionic Matsubara frequencies

$d$8

The hybridization has the cavity representation

$d$9

where $f$0 is the Green’s function of the lattice with site $f$1 removed. It describes the amplitude and lifetime for an electron to leave the impurity, propagate through the environment, and return.

Solving the impurity problem gives

$f$2

The impurity self-energy is defined by

$f$3

DMFT imposes

$f$4

where $f$5 is the local lattice Green’s function calculated with the same self-energy. Equivalently,

$f$6

and the bath update is

$f$7

or

$f$8

For the infinite-coordination Bethe lattice with semicircular density of states, the self-consistency relation simplifies to

$f$9

with the precise prefactor depending on the bandwidth convention.

The same structure has a rigorous finite-system formulation. For a partition into impurity clusters, the self-energy is block diagonal,

$\tau\sim\hbar/W$0

and the bath of cluster $\tau\sim\hbar/W$1 is the remainder of the lattice dressed by the self-energies of the other clusters: $\tau\sim\hbar/W$2 A mathematical analysis of IPT-DMFT establishes existence of fixed points under specified finite-system, half-filled, paramagnetic, translation-invariant assumptions, while not proving uniqueness or convergence of the practical iteration (Cancès et al., 2024).

4. Numerical solution and impurity solvers

A conventional equilibrium DMFT calculation consists of a self-consistency loop:

1. Specify the lattice and parameters: choose $\tau\sim\hbar/W$3 or $\tau\sim\hbar/W$4, $\tau\sim\hbar/W$5, $\tau\sim\hbar/W$6, temperature, symmetry assumptions, and an initial bath or self-energy.
2. Construct the lattice Green’s function:

$\tau\sim\hbar/W$7

1. Perform the lattice Hilbert transform:

$\tau\sim\hbar/W$8

2. Define the Weiss field:

$\tau\sim\hbar/W$9

1. Solve the interacting impurity model to obtain $W$0.
2. Extract the self-energy:

$W$1

1. Mix and iterate until the impurity and lattice quantities converge.

Linear mixing is often used: $W$2 A typical convergence condition compares successive Green’s functions, self-energies, or hybridization functions below a prescribed tolerance.

The impurity solver determines the computational cost and accessible observables. Common solvers include quantum Monte Carlo, continuous-time quantum Monte Carlo, Hirsch–Fye quantum Monte Carlo, numerical renormalization group, exact diagonalization, density-matrix renormalization group, iterated perturbation theory, non-crossing approximation, fluctuation-exchange approximation, and strong-coupling expansions.

• Quantum Monte Carlo is effective at finite temperature but produces imaginary-time or Matsubara data; real-frequency spectra require analytic continuation.
• Exact diagonalization provides direct real-frequency information but represents the bath with a finite number of levels and has exponentially growing Hilbert-space cost.
• Numerical renormalization group is powerful for low-energy impurity dynamics and real-frequency spectra.
• Perturbative solvers are computationally efficient but reliable only in appropriate interaction and temperature regimes.
• Machine-learning-assisted solvers use ensembles of approximate impurity solutions to learn corrections to the impurity Green’s function. Data-driven DMFT has been demonstrated for the single-band Hubbard model and the Mott transition, while its reliability remains restricted by the training domain and reference data (Sheridan et al., 2021).
• Influence-functional matrix-product-state solvers represent the bath-induced temporal memory as an MPS and provide equilibrium and nonequilibrium impurity solutions, with errors controlled by time step, contour length, memory cutoff, and bond dimension (Nayak et al., 4 Mar 2025).

The exact equilibrium formalism can also be formulated measure-theoretically. Hybridization and self-energy functions are represented by positive spectral measures, making causality explicit through Pick-function properties such as

$W$3

Finite-bath parametrizations generally cannot contain nontrivial IPT-DMFT fixed points because the number of poles proliferates under the impurity and self-consistency maps; the appropriate fixed-point space may require continuous or infinitely supported spectral measures (Cancès et al., 2024).

5. Spectra, quasiparticles, and the Mott transition

The spectral function is

$W$4

For $W$5, it reduces to the noninteracting density of states. With interactions, it reveals coherent and incoherent excitations.

In a correlated Fermi liquid,

$W$6

at low temperature and frequency. The quasiparticle weight is

$W$7

and the effective mass satisfies approximately

$W$8

Increasing $W$9 reduces $$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$0, narrows the coherent quasiparticle peak, increases the effective mass, and lowers the coherence scale.

At intermediate or strong coupling, the local spectrum commonly contains:

1. a narrow quasiparticle peak near the Fermi level;
2. a lower Hubbard band at negative energy;
3. an upper Hubbard band at positive energy.

The Hubbard bands are incoherent, approximately atomic excitations associated with adding or removing an electron in the presence of $$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$1. At half filling, their separation is of order $$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$2. The central peak represents coherent itinerant motion persisting between atomic regimes.

The Mott–Hubbard transition is driven by competition between kinetic energy, which favors delocalization and double occupancy, and local repulsion, which suppresses double occupancy and favors localization. At sufficiently large $$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$3, the quasiparticle peak disappears and a gap opens.

At zero temperature, the metallic solution disappears at an upper critical interaction $$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$4, while the insulating solution disappears at a lower critical interaction $$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$5. At finite temperature, metallic and insulating solutions coexist in a hysteresis region bounded by these critical interactions. The first-order line ends at a critical endpoint; above it, the metal–insulator change becomes a crossover.

DMFT also captures local moments, Kondo-like screening, bad-metal behavior, entropy effects, magnetic ordering, orbital order, charge order, and quasiparticle destruction. In the paramagnetic Mott insulator of infinite-dimensional single-site DMFT, the exchange scale $$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$6 vanishes with dimension, leaving an extensive local-moment entropy approximately $$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$7 per electron down to low temperature. This is partly an artifact of neglecting nonlocal magnetic exchange (Vollhardt et al., 2011).

The high-energy and intermediate-energy regimes are physically significant. Atomic multiplets and local charge or spin excitations dominate at high energies; hybridization broadens and screens these states at intermediate energies; and a coherent heavy quasiparticle may emerge at low energies. DMFT therefore describes correlated matter as a sequence of crossovers between atomic, incoherent, and itinerant regimes rather than as a simple binary distinction between localized and itinerant electrons (Georges, 2011).

6. Materials, finite systems, transport, and nonequilibrium extensions

Materials and nanostructures

DFT+DMFT combines a material-specific one-electron Hamiltonian from LDA or GGA with local Coulomb and Hund interactions. The lattice Green’s function becomes matrix valued: $$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$8 The self-energy is nonzero only in the selected correlated subspace.

A double-counting correction is required because the DFT exchange-correlation functional already contains an average interaction among localized orbitals. Its precise form is not uniquely determined and affects level positions, occupancies, and magnetic moments. DFT+DMFT has been applied to transition-metal oxides, rare-earth compounds, Mott insulators, heavy quasiparticles, heterostructures, and topological correlated systems.

For molecules and finite clusters, the bath may be discrete rather than continuous. Cellular DMFT partitions selected molecular orbitals into clusters and retains intra-cluster self-energy components. Hydrogen-chain and hydrogen-ring calculations show that single-site and cluster DMFT can provide competitive energies in intermediate- and strong-correlation regimes, including bond stretching and dissociation. Two-site CDMFT improves on single-site DMFT by incorporating short-range intersite correlations, although finite bath size, cluster geometry, and treatment of long-range Coulomb terms remain important (Lin et al., 2010).

Real-space DMFT assigns different local self-energies to inequivalent atoms: $$\hat H= \sum_{ij,\sigma}t_{ij}\hat c^\dagger_{i\sigma}\hat c_{j\sigma} +U\sum_i\hat n_{i\uparrow}\hat n_{i\downarrow},$$9 It is therefore applicable to molecules, nanostructures, surfaces, interfaces, and magnetic clusters. In Fe and FePt clusters, dynamical fluctuations reduce magnetic moments relative to DFT+$t_{ij}$0 because local charge and spin configurations fluctuate in time rather than remaining frozen (Turkowski et al., 2011).

Molecular electronics and transport

In nanoscopic conductors, DFT supplies geometry, electrode embedding, and weakly correlated states, while DMFT treats localized $t_{ij}$1- or $t_{ij}$2-orbitals. The interacting device Green’s function includes lead embedding self-energies and local correlation self-energies: $t_{ij}$3 The coherent transmission is

$t_{ij}$4

For Ni nanocontacts, dynamical correlations generate narrow quasiparticle resonances near the Fermi level. Their interference with broad conduction channels produces Fano-shaped conductance features. The relevant orbitals can be mixed valent rather than purely Kondo-like, demonstrating that a low-energy resonance need not imply a pure Kondo mechanism (Jacob et al., 2010).

DFT+DMFT transport calculations for Cu/Co spin valves show that the imaginary part of the self-energy suppresses coherent transmission away from the Fermi energy through finite lifetimes, while the real part shifts correlated levels and modifies elastic $t_{ij}$5–$t_{ij}$6 scattering. At the Fermi energy, Fermi-liquid behavior makes the imaginary part vanish, so correlation-induced transmission changes are primarily associated with real-part level shifts (Droghetti et al., 2022).

Bosonic and spin dynamical mean-field theories

Bosonic DMFT extends the impurity mapping to interacting bosons in optical lattices. It combines a static condensate field with a dynamical bath generated by virtual hopping. In a $t_{ij}$7 expansion, the leading term gives the Gutzwiller condensate mean field and the second-order connected term generates the dynamical hybridization. In Nambu space, normal and anomalous Green’s functions, self-energies, Weiss fields, and hybridizations are matrix valued. BDMFT describes superfluid–Mott transitions, spin-ordered insulating phases, supersolids, and virtual exchange processes absent from simple Gutzwiller theory (Snoek et al., 2010).

SpinDMFT applies a related local-dynamical construction to dense spin systems at infinite temperature. For spin-$t_{ij}$8 Heisenberg models at large effective coordination, the environment of a spin becomes a time-dependent classical Gaussian random field whose covariance is fixed self-consistently by spin autocorrelations. This provides a real-time stochastic analogue of the fermionic impurity mapping (Gräßer et al., 2021).

Nonequilibrium DMFT

Nonequilibrium DMFT replaces equilibrium frequency-dependent quantities by contour-time Green’s functions and self-energies. For steady-state transport, the impurity Green’s function is represented in Keldysh space and the self-consistent bath includes retarded, advanced, and Keldysh components.

An auxiliary quantum master-equation approach embeds a finite interacting impurity cluster in Markovian reservoirs. The reduced density matrix obeys a Lindblad equation, and the steady state is obtained from the zero-eigenvalue Liouvillian mode. Intermediate bath sites transform the Markovian external reservoirs into a structured, effectively non-Markovian hybridization for the impurity. This approach provides direct real-frequency steady-state observables without analytic continuation (Arrigoni et al., 2012).

Influence-functional MPS methods likewise treat temporal bath memory directly and have been applied to nonequilibrium steady states of photo-doped Mott systems with long-lived doublon and holon populations (Nayak et al., 4 Mar 2025).

7. Extensions, limitations, and developments beyond single-site DMFT

The principal limitation of single-site DMFT is the local-self-energy approximation: $t_{ij}$9 It neglects nonlocal self-energy and vertex correlations, including short-range antiferromagnetism, intersite singlets, momentum-selective pseudogaps, nonlocal pairing, long-wavelength critical fluctuations, and some low-dimensional collective phenomena.

Cluster DMFT and the dynamical cluster approximation restore short-range spatial structure by replacing the single impurity with a cluster. Cellular DMFT yields a cluster self-energy $U$0, while DCA introduces momentum-sector dependence. These methods have been used for antiferromagnetism, pseudogaps, unconventional superconductivity, and momentum differentiation.

Diagrammatic extensions, including the dynamical vertex approximation and dual-fermion theory, generate nonlocal self-energy and vertex corrections from local DMFT vertices. Functional-renormalization approaches can begin from a converged DMFT solution and generate nonlocal correlations during the flow.

For disordered systems, conventional averaged DMFT cannot describe Anderson localization because the arithmetic average of the local density of states can remain finite at a mobility edge. Statistical DMFT retains a distinct hybridization function for every site,

$U$1

and captures distributions of local densities of states, quasiparticle weights, Kondo scales, rare regions, and inhomogeneous Mott–Anderson transitions. Typical medium theory instead uses the geometric average

$U$2

which can vanish at the Anderson transition. Extended DMFT introduces a retarded bosonic bath and replica order parameters to treat intersite density fluctuations, Coulomb gaps, electronic Griffiths phases, and glassy behavior (Miranda et al., 2011).

Real-space DMFT captures local Kondo screening in inhomogeneous systems but neglects nonlocal self-energy feedback from indirect exchange. Benchmarks against DMRG for a two-impurity Anderson model show excellent agreement in Kondo-screened regimes and failure when RKKY correlations dominate, where the exact state is a nonlocal singlet but real-space DMFT can develop artificial symmetry breaking (Titvinidze et al., 2012).

A diagrammatic Monte Carlo expansion around DMFT has been proposed to treat nonlocal corrections systematically. Local interaction processes are absorbed into impurity vertices, while the remaining expansion is performed in explicitly nonlocal dressed hopping. If the resulting series is convergent or resummable, the method is asymptotically exact beyond DMFT (Carlström, 2023).

The formal exactness of DMFT must therefore be qualified:

• Exact in the infinite-coordination limit for local-interaction lattice models with appropriately scaled hopping.
• Approximate in finite dimensions because nonlocal self-energy and vertex correlations are omitted.
• Solver dependent in practical calculations, with errors from statistical sampling, bath discretization, perturbative truncation, analytic continuation, tensor-network truncation, or machine-learning extrapolation.
• Parameter dependent in materials applications, particularly through the choice of correlated subspace, interaction parameters, charge self-consistency, and double-counting prescription.

DMFT’s central achievement is the controlled separation of local temporal dynamics from nonlocal spatial correlations. It treats the former through a self-consistent interacting impurity and approximates the latter through a local self-energy. This construction explains the coexistence of coherent quasiparticles, incoherent Hubbard excitations, local moments, dynamical screening, and Mott localization, while providing a basis for systematic extensions to clusters, disorder, transport, bosonic systems, nonequilibrium dynamics, and diagrammatic treatments of nonlocal correlations.