Real-Time Dynamical Mean-Field Theory

• Real-Time DMFT is a framework that generalizes equilibrium DMFT to simulate ultrafast nonequilibrium dynamics by mapping lattice problems to self-consistent impurity models on real-time contours.
• It utilizes contour-ordered Green's functions and Dyson equations to accurately capture phenomena such as quenches, pump-probe responses, and Mott transitions.
• The method is applied across fermionic, bosonic, and neural network systems, enabling efficient quantum simulations with advanced impurity solvers like Inchworm QMC and tensor-network approaches.

Real-time Dynamical Mean-Field Theory (DMFT) generalizes the equilibrium DMFT formalism to capture the real-time evolution and nonequilibrium dynamics of strongly correlated many-body systems. By mapping the original lattice problem to a quantum impurity coupled self-consistently to a bath, and by formulating the fundamental equations on a suitably chosen real-time contour, this approach enables simulation of ultrafast phenomena, quenches, driven states, and time-dependent responses across a broad spectrum of model systems (fermionic, bosonic, spin, neural network) and materials. Real-time DMFT can operate directly in the time domain, on complex or Keldysh contours, thus circumventing analytic continuation from imaginary time and avoiding severe numerical limitations associated with traditional equilibrium methods. Applications span quantum simulation, tensor-network solvers, nonequilibrium condensed matter, and high-dimensional neural dynamics.

1. Fundamental Formulation and Green's Functions

Real-time DMFT is constructed by mapping the original lattice Hamiltonian to an impurity model embedded in a self-consistently determined bath. The central objects are contour-ordered Green's functions defined on the real-time Keldysh or complex-time contour. For a standard Hubbard model, the local impurity action on the contour $\mathcal{C}$ is

$$S_{\text{imp}} = -i\sum_\sigma \int_\mathcal{C} dt\, H_{\text{loc}}[n_\sigma(t)] - i\sum_\sigma \int_\mathcal{C} dt\, dt' c^\dagger_\sigma(t) \Delta_\sigma(t,t') c_\sigma(t')$$

where $$H_{\text{loc}} = U n_\uparrow n_\downarrow - \mu(n_\uparrow + n_\downarrow)$$ and $\Delta_\sigma(t,t')$ is the hybridization (Weiss) function (Dong et al., 2017).

The impurity Green's function for fermions,

$$G_\sigma(t,t') = -i \langle T_\mathcal{C} c_\sigma(t) c^\dagger_\sigma(t') \rangle$$

serves as the central dynamical observable, with various components (greater, lesser, retarded, advanced) defined by their location on the real-time contour.

For bosonic systems (e.g., Bose-Hubbard), the contour-ordered Green's function generalizes to Nambu space; for neural networks, the order parameters are two-time autocorrelation and causal response functions $C(t,t')$ and $R(t,t')$ (Zou et al., 2023).

2. Real-Time Self-Consistency and Dyson Equations

Self-consistency in real-time DMFT equates the local impurity and lattice Green's functions. For the Bethe lattice with infinite coordination, the hybridization function is set by

$\Delta_\sigma(t, t') = v^2 G_\sigma(t, t')$

or, in the context of quantum simulation,

$\Delta^R(t) = (t^*)^2 G^R_{\text{imp}}(t)$

where $v$ and $t^*$ are hopping or bandwidth parameters (Dong et al., 2017, Rangi et al., 27 Jan 2026).

The Dyson equation for the impurity reads

$$G^R_{\text{imp}}(t,t') = G^R_0(t,t') + \int dt_1\, dt_2\, G^R_0(t, t_1)\, \Sigma^R(t_1, t_2)\, G^R_{\text{imp}}(t_2, t')$$

with $G^R_0$ the noninteracting propagator (determined by $\Delta^R$) and $\Sigma^R$ the retarded self-energy.

Multi-orbital DMFT employs matrix Dyson equations on complex contours (Yu et al., 29 Dec 2025). For bosonic DMFT, the real-time Dyson equation in Nambu space is

$$[i \partial_t \sigma_z + \mu I - \Sigma(t, t') - \Delta(t, t')] \ast G_\text{loc}(\cdot, t') = I \delta_\mathcal{C}(t, t')$$

where $\ast$ denotes contour convolution (Strand et al., 2014).

3. Impurity Solvers: Algorithms for Fermions and Bosons

A range of impurity solvers have been tailored to real-time DMFT.

Inchworm Quantum Monte Carlo: Expands the partition function in powers of the hybridization, but builds partial propagators incrementally ("inching" in time) to overcome the dynamical sign problem. Polynomial scaling of computational cost, typical crossing order $O\sim6$–8 for metallic and insulating regimes, convergence of $G(t)$ up to $t\sim n\delta t$ after $n$ DMFT iterations (Dong et al., 2017).

Tensor-Network Solvers on Complex-Time Contours: For multi-orbital DMFT, time-evolution along complex contours at angle $\alpha$ suppresses entanglement growth, significantly reducing MPS or TTN bond-dimension requirements ($\chi\sim40$–90), while exponential-fitting schemes (ESPRIT) extract real-frequency poles efficiently for spectral function recovery (Yu et al., 29 Dec 2025).

Finite-Chain Mapping: Replaces the continuous bath with a short 1D chain of $N_{\text{bath}}$ orbitals, enabling exact diagonalization or quantum time-propagation. Suitable for near-term quantum hardware, with stable convergence for $U=2$–8 (units of bandwidth), six-site chains sufficient for high-energy Hubbard physics (Rangi et al., 27 Jan 2026).

Strong-Coupling NCA for Bosons: Formulation via pseudo-particle fields and NCA bubble diagrams; captures dynamical transitions, damping, and thermalization in quenched Bose-Hubbard models, including amplitude (Higgs) oscillations inaccessible to simple mean-field (Strand et al., 2014).

4. Analytic Continuation and Spectral Function Reconstruction

Direct real-time simulation eliminates the need for ill-posed analytic continuation. Spectral functions $A(\omega) = -\frac{1}{\pi}\,\text{Im} G^R(\omega)$ are obtained by Fourier transformation of the retarded Green's function, often supplemented by linear prediction or exponential fitting when the time window $t_{\text{max}}$ is limited (Dong et al., 2017, Yu et al., 29 Dec 2025).

Complex-time evolution followed by exponential fitting (ESPRIT or SVD of Hankel matrices) yields high-resolution spectra by reconstructing the real-frequency poles with negligible extra cost. Typical tunable parameters: time-step $\Delta t$, contour angle $\alpha=0.2$–$0.4$ (balances entanglement and spectral fidelity), SVD cutoff $\epsilon_{\text{svd}}\sim10^{-6}$ (Yu et al., 29 Dec 2025).

For real-time DMFT in quantum simulation, cubic-spline interpolation is essential to avoid Fourier artifacts on coarse time grids (Rangi et al., 27 Jan 2026).

5. Applications to Nonequilibrium, Quantum Simulation, and Neural Networks

Real-time DMFT is the only numerically controlled framework for simulating nonequilibrium and time-dependent strongly correlated phenomena on the lattice.

Quenches, Pulsed Fields, Pump-Probe: Real-time DMFT applies unchanged to nonequilibrium problems; the two-time Dyson self-consistency captures non-time-translation-invariant dynamics. Key observables include $G^{>}(t)$, which tracks metal-insulator transitions as $U$ increases, and the real-time evolution of the self-energy and spectral weight (Dong et al., 2017).

Quantum Simulation (NISQ Devices): Finite-chain mapping and time-domain iteration schemes enable direct implementation on quantum hardware via Trotterization or variational circuits. Six-site chains, time windows $t\in[0,20]$, and accurate Hubbard band recovery are demonstrated with only twelve qubits and shallow circuits (Rangi et al., 27 Jan 2026).

Bosonic Dynamics: BDMFT describes interaction quenches in the Bose-Hubbard model, capturing rapid thermalization, long-lived plateaus, and collapse-and-revival oscillations characteristic of nontrivial nonequilibrium bosonic regimes (Strand et al., 2014).

Neural Network Dynamics: Dynamical mean-field theory yields closed self-consistent equations for two-time autocorrelation and response functions $C(t,t')$, $R(t,t')$ in randomly connected networks with bidirectional correlations. The transition from fixed point to chaos is determined exactly by the gain and asymmetry, $g_c(\eta)=1/(1+\eta)$ (Zou et al., 2023).

6. Extensions: Multiorbital, TDDFT, and Complex-Time Formulations

Multi-orbital DMFT in real time leverages complex-time contours and tensor-network solvers for efficient simulation of Hund's metals and Kanamori models, recovering quasiparticle peaks, Hubbard bands, and multiplet structure at reduced computational cost (Yu et al., 29 Dec 2025).

TDDFT extensions embed DMFT-derived exchange-correlation potentials $v_{xc}^{\text{DMFT}}(n)$ into real-time Kohn-Sham equations under the adiabatic local density approximation (ALDA). This enables simulation of carrier dynamics (e.g., Bloch oscillations in 3D Hubbard models), for which the DMFT $v_{xc}$ exhibits discontinuity at half-filling—a direct signature of the Mott transition (Karlsson et al., 2010). ALDA captures coherent field-driven acceleration and local Mott physics but lacks true damping, motivating future research into memory-dependent xc functionals.

7. Physical Insights and Regimes

Key physical phenomena elucidated by real-time DMFT include:

• Metal-insulator transitions and Mott gap formation: spectral function evolution shows three-peak structures and gap opening as $U$ increases (Dong et al., 2017, Rangi et al., 27 Jan 2026).
• Nonequilibrium relaxation and thermalization: BDMFT identifies windows of rapid and slow relaxation, prethermal plateaus, and dynamical transitions in both normal and superfluid regimes (Strand et al., 2014).
• Quantum many-body coherence: TDDFT–DMFT reproduces interaction-induced sidebands ("beats") in Bloch oscillations (Karlsson et al., 2010).
• Neural network instability and chaos: Real-time DMFT gives exact criteria for transitions among fixed-point, oscillatory, and chaotic regimes; the fluctuation-dissipation theorem is recovered only for fully symmetric couplings (Zou et al., 2023).
• Computational performance: Inchworm QMC and complex-time tensor networks offer polynomial rather than exponential scaling, enabling simulations at previously inaccessible time scales (Dong et al., 2017, Yu et al., 29 Dec 2025).

Real-time DMFT thus provides a unified framework for simulating, analyzing, and interpreting ultrafast, nonequilibrium dynamics in correlated quantum matter and high-dimensional random systems, with demonstrable applicability to quantum simulation platforms and ab initio materials modeling.