Time-Dependent Variational Monte Carlo

• Time-dependent variational Monte Carlo is a stochastic method that projects Schrödinger dynamics onto a parametrized manifold of wave functions.
• It employs Monte Carlo sampling to estimate covariance metrics and force terms, ensuring stable and accurate time integration.
• Extensions include imaginary-time evolution, neural-network and tensor-network states, and open-system dynamics to address diverse quantum challenges.

Time-dependent variational Monte Carlo (tVMC, also TDVMC in parts of the literature) is a stochastic realization of the time-dependent variational principle (TDVP) for quantum many-body dynamics. In its standard form, the method restricts evolution to a parametrized manifold of wave functions, projects the Schrödinger flow onto the corresponding tangent space, and evaluates the resulting metric and force terms by Monte Carlo sampling over the instantaneous variational state (Gartner et al., 2022). The same projection structure also underlies imaginary-time evolution, finite-temperature constructions, linear-response formalisms, dissipative trajectory methods, and recent extensions to adaptive parameter selection, unbiased estimators, neural-network states, and two-dimensional tensor-network states (Takai et al., 2015).

1. Variational foundation and projected equations of motion

The central object of tVMC is a time-dependent variational state $|\Psi(\alpha(t))\rangle$, with complex parameters $\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$. Real-time evolution is obtained by minimizing the distance between exact Schrödinger dynamics and its projection onto the variational manifold, following the Dirac–Frenkel or McLachlan TDVP:

$$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$

Introducing logarithmic derivatives

$O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$

and the local energy

$$E_{\mathrm{loc}}(x;t)=\frac{[H\Psi(x;t)]}{\Psi(x;t)},$$

the standard tVMC equations take the covariance form

$i\sum_l S_{kl}\,\dot\alpha_l = F_k,$

with

$$S_{kl}=\langle O_k^* O_l\rangle-\langle O_k^*\rangle\langle O_l\rangle,\qquad F_k=\langle O_k^*E_{\mathrm{loc}}\rangle-\langle O_k^*\rangle\langle E_{\mathrm{loc}}\rangle,$$

where expectations are taken over the $|\Psi|^2$ distribution (Gartner et al., 2022). Equivalent forms are used across lattice, continuum, neural-network, and tensor-network implementations, and are routinely interpreted as stochastic reconfiguration or a natural-gradient step in parameter space (Salioni et al., 10 Jun 2025).

Imaginary-time evolution uses the same geometric structure with a different driving term. In finite-temperature VMC, the imaginary-time TDVP equation is

$\sum_j S_{ij}\,\frac{d\theta_j}{d\tau}=-g_i,$

or in discretized form $\Delta \theta=-\Delta\tau\,S^{-1}g$, with the same covariance metric and energy-correlated force estimators (Takai et al., 2015). In ab initio continuous-space electron dynamics, equivalent equations are written with the quantum geometric tensor $\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$0 and generalized force $\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$1, but the Monte Carlo estimators reduce to the same centered covariance structure (Nys et al., 2024).

The same projection logic extends beyond unitary closed-system dynamics. A Monte Carlo TDVP for Lindblad evolution projects dissipative dynamics onto a pure-state variational manifold by introducing a parameter-space Fokker–Planck equation and associated stochastic differential equation (Transchel et al., 2014). More recently, an open-system tVMC formulation based on quantum trajectories derived nonlinear Stratonovich stochastic equations for the variational parameters, again built from the same metric $\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$2 and Monte Carlo estimators of tangent-space covariances (Apostoli et al., 30 Jun 2025).

2. Variational manifolds and state representations

A distinctive feature of tVMC is that the TDVP equations are agnostic to the specific ansatz, provided logarithmic derivatives and local operator ratios can be evaluated. Early continuous-space bosonic formulations used an exponential Jastrow–Feenberg expansion,

$\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$3

with $\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$4 written as a systematic multi-body series in one-body, two-body, and higher-body correlations (Carleo et al., 2016). In the optical-lattice Bose-gas formulation, the wave function was represented as

$\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$5

with $\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$6 and $\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$7 parameterized by third-order B-splines and constrained to satisfy the Lieb–Liniger cusp at particle coincidence (Gartner et al., 2022).

For strongly correlated electrons on lattices, a standard tVMC ansatz is the generalized pair-product state dressed by correlation operators and quantum-number projections,

$\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$8

where $\alpha(t)=\{\alpha_1(t),\dots,\alpha_P(t)\}$9 and the correlators include Gutzwiller and long-range Jastrow factors (Ido et al., 2015). In finite-temperature VMC, the truncated manifold is enlarged to a linear combination of Gutzwiller–Jastrow–Pfaffian–backflow components, with thermal pure quantum states generated by imaginary-time TDVP within that manifold (Takai et al., 2015).

Neural-network wave functions have become a major variational class for tVMC. Adaptive real-time tVMC has been benchmarked with spin-Jastrow and restricted Boltzmann machine (RBM) states in the transverse-field Ising model (Salioni et al., 10 Jun 2025). Ab initio electron dynamics has been formulated with FermiNet-type wave functions, and a neural-basis tVMC construction freezes the nonlinear feature extractor while evolving only the last linear layer, so that the dynamics is confined to a compact manifold spanned by a fixed neural basis (Fu et al., 4 Jun 2026). Continuous-space fermionic dynamics has also been formulated with time-dependent Slater–Jastrow–backflow ansätze, including neural-network parameterizations of backflow transformations (Nys et al., 2024).

Tensor-network manifolds likewise fit the same framework. A recent two-dimensional formulation specializes tVMC to projected entangled pair states (PEPS), treating the PEPS tangent space stochastically and analytically removing gauge redundancies before solving the stochastic reconfiguration equations (Wu, 7 Dec 2025). Earlier dissipative TDVP Monte Carlo work had already shown that pure-state variational classes such as matrix product states can be used to simulate Lindblad evolution by sampling stochastic parameter trajectories (Transchel et al., 2014).

3. Monte Carlo estimators, linear algebra, and time integration

At each time step, tVMC samples configurations from the instantaneous probability distribution associated with the variational state and accumulates Monte Carlo estimates of the metric and force. In bosonic continuum simulations this is typically done with Metropolis–Hastings sampling over particle coordinates (Gartner et al., 2022). In lattice-fermion implementations, the sampled configurations are occupation-number basis states and the local energy exploits the sparse structure of the Hamiltonian (Ido et al., 2015). In flow-based continuous-variable formulations, exact sampling is obtained from a pushforward distribution induced by the flow, and the same covariance estimator is interpreted as the pullback of the quantum Fisher geometry (Stokes et al., 2022).

Because $$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$0 is an empirical covariance matrix, numerical conditioning is a central issue. Stabilization strategies explicitly reported in the literature include diagonal shifts, truncated singular-value decompositions, pseudoinverses, and diagonal rescalings (Salioni et al., 10 Jun 2025). In the optical-lattice Bose-gas study, the covariance matrix was rescaled by its diagonal entries and regularized with $$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$1; the linear system was solved by QR decomposition and the parameters were propagated with fourth-order Runge–Kutta using $$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$2 (Gartner et al., 2022). In many-variable tVMC for lattice fermions, fourth-order Runge–Kutta was also used, with adaptive time steps proportional to $$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$3 or $$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$4 depending on the interaction scale (Ido et al., 2015). In neural-basis tVMC for electron dynamics, a second-order Heun predictor–corrector scheme combined with an SVD pseudoinverse was found substantially more stable than Cholesky factorization (Fu et al., 4 Jun 2026).

The computational bottlenecks depend on both the estimator construction and the variational family. Dense covariance assembly scales as $$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$5 in the number of parameters $$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$6 and samples $$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$7, while direct dense linear solves scale as $$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$8 (Salioni et al., 10 Jun 2025). In the PEPS-tVMC formulation, tensor locality and a “small-o” storage strategy reduce memory costs, while Cholesky decomposition becomes viable after gauge removal renders the stochastic reconfiguration matrix well conditioned (Wu, 7 Dec 2025). In continuous-space ab initio electron dynamics, evaluating local energies requires Laplacians of many-body wave functions and was identified as an $$\min \left\| (i\partial_t-H)|\Psi(\alpha(t))\rangle \right\|.$$9 cost driver, while inversion of the quantum geometric tensor scales as $O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$0 (Nys et al., 2024).

These implementations underscore an important structural point: the Monte Carlo part of tVMC is not a post-processing layer added to TDVP, but the mechanism by which the tangent-space metric, projected force, and, in open systems, diffusion terms are estimated and propagated at each step (Apostoli et al., 30 Jun 2025).

4. Real-time dynamics and dynamical response

A major application area of tVMC is real-time non-equilibrium dynamics. In continuous-space Lieb–Liniger bosons, tVMC based on a systematic Jastrow expansion reproduces ground-state properties with high accuracy and captures post-quench unitary dynamics benchmarked against exact Bethe-ansatz and quench-action results (Carleo et al., 2016). In one-dimensional and two-dimensional Hubbard models, many-variable tVMC reproduces the time evolution of energy, momentum distribution, spin structure factor, and superconducting correlations during and after interaction quenches (Ido et al., 2015). In ab initio electronic dynamics, time-dependent Jastrow and backflow correlations reveal dynamical many-body effects beyond mean-field in driven molecules and quenched quantum dots (Nys et al., 2024).

The extraction of dynamical spectra is another established use case. In a one-dimensional Bose gas in an optical lattice, tVMC was used to compute the dynamic structure factor

$O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$1

by applying a weak, Gaussian-in-time, multi-mode spatial perturbation and monitoring the density response (Gartner et al., 2022). In linear response, the dynamic structure factor was obtained from

$O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$2

allowing the full $O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$3 to be extracted from a single broadband weak-pulse simulation (Gartner et al., 2022). In the deep-lattice regime, the resulting dispersions agreed well with exact-diagonalization results for the Bose–Hubbard model, whereas in shallow lattices the simulations detected multi-band effects beyond the single-band approximation (Gartner et al., 2022).

The same work also examined nonlinear response. A strong single-mode pulse at low wave number generated higher harmonics in the density-fluctuation power spectrum and produced a broadened but informative reconstruction of the excitation spectrum from a single drive (Gartner et al., 2022). A further noteworthy result was that noise-only propagation, with no external perturbation, generated power-spectrum peaks at the same excitation energies as the weak-pulse linear-response calculation. The authors explicitly noted, however, that the corresponding spectral weights are not proportional to those of $O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$4 because the sampling noise is colored; quantitative spectral weights still require a calibrated weak probe (Gartner et al., 2022).

Linear-response theory can also be formulated directly inside VMC. Linear-response VMC around an optimized ground state leads to a generalized eigenvalue problem formally analogous to the Tamm–Dancoff approximation, and has been used to compute electronic excitation energies and oscillator strengths with Jastrow–Slater wave functions (Mussard et al., 2017). This establishes a bridge between the explicitly time-dependent TDVP formulation and small-amplitude response theory within the same variational tangent space (Mussard et al., 2017).

5. Imaginary time, finite temperature, and open-system extensions

Imaginary-time tVMC is the basis of several finite-temperature and state-preparation methods. In finite-temperature VMC for strongly correlated electrons, the imaginary-time TDVP is implemented as stochastic reconfiguration in a truncated manifold of Pfaffian-based states, generating thermal pure quantum states starting from randomized high-temperature initializations (Takai et al., 2015). Benchmarks on one-dimensional and two-dimensional Hubbard models showed that energy per site, double occupancy, and nearest-neighbor spin correlations match full-space TPQ calculations or reference estimates within error bars in several regimes, with systematic improvement as the number of Pfaffian components is increased (Takai et al., 2015).

Open quantum dynamics has been incorporated into tVMC by combining variational projection with stochastic unravelings of the Lindblad equation. A Monte Carlo TDVP for dissipative systems represents the density matrix as an ensemble over a pure-state variational manifold, derives a Fokker–Planck equation for the parameter distribution, and samples the associated stochastic differential equation for the parameters (Transchel et al., 2014). This construction enables the use of pure-state classes such as matrix product states for dissipative evolution and is explicitly described as “embarrassingly parallel” because trajectories are independent (Transchel et al., 2014).

A more recent trajectory-based framework combines tVMC with quantum state diffusion for open many-body systems. For a normalized variational state $O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$5, the projected Stratonovich stochastic equation for the variational parameters is

$O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$6

where $O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$7 is the same covariance metric, $O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$8 is built from an effective local energy, and $O_k(x)=\partial_{\alpha_k}\ln \Psi(x;t),$9 encodes the tangent-space projection of the collapse operators (Apostoli et al., 30 Jun 2025). This formalism has been validated for the locally dissipative long-range Ising model, reproducing non-equilibrium magnetization and spin squeezing dynamics relevant to trapped-ion and Rydberg platforms (Apostoli et al., 30 Jun 2025).

The scope of TDVP-based Monte Carlo projection has also extended beyond quantum many-body Hamiltonians. A generalized TD-VMC framework for first-order-in-time partial differential equations projects $$E_{\mathrm{loc}}(x;t)=\frac{[H\Psi(x;t)]}{\Psi(x;t)},$$0 onto a parametrized manifold and estimates the resulting Galerkin system by Monte Carlo sampling, with the time-dependent Schrödinger equation recovered as a special case (Zhao et al., 2022). In that work, the method was applied to the multi-asset Black–Scholes equation for option pricing, illustrating that the stochastic-projection viewpoint is not confined to conventional quantum dynamics (Zhao et al., 2022).

6. Bias, conditioning, adaptive control, and current directions

Although the covariance equations of tVMC are standard, recent work has shown that their conventional Monte Carlo estimators are not universally reliable. A detailed analysis proved that standard tVMC is affected by a systematic statistical bias, or else exponential sample complexity, when the wave function has exact or approximate zeros that are under-sampled by the Born distribution (Sinibaldi et al., 2023). The origin is a support mismatch: the Born distribution excludes configurations at nodes, while the exact TDVP sums can still receive nonzero contributions from derivatives or Hamiltonian matrix elements on those configurations (Krinitsin et al., 5 May 2026). This is especially relevant for fermionic systems, anti-symmetric ansätze, and protocols with measurements that dynamically create zeros in the wave function (Sinibaldi et al., 2023).

Two complementary remedies have been proposed. One is projected quantum evolution, or p-tVMC, which replaces the instantaneous TDVP linear solve by a variational projection of the exactly Trotterized short-time evolution onto the ansatz manifold through an infidelity minimization problem (Sinibaldi et al., 2023). The other is an unbiased real-time TDVMC based on self-normalized importance sampling with respect to a cutoff-deformed distribution $$E_{\mathrm{loc}}(x;t)=\frac{[H\Psi(x;t)]}{\Psi(x;t)},$$1, which restores full support and removes the nodal support-mismatch bias while retaining asymptotic consistency (Krinitsin et al., 5 May 2026). In the latter formulation, practical values $$E_{\mathrm{loc}}(x;t)=\frac{[H\Psi(x;t)]}{\Psi(x;t)},$$2–$$E_{\mathrm{loc}}(x;t)=\frac{[H\Psi(x;t)]}{\Psi(x;t)},$$3 were reported to remove nodal bias without large variance inflation in the tested lattice-spin quenches (Krinitsin et al., 5 May 2026).

A separate line of work addresses overparameterization and ill-conditioning rather than estimator bias. Adaptive tVMC introduces the local-in-time error (LITE),

$$E_{\mathrm{loc}}(x;t)=\frac{[H\Psi(x;t)]}{\Psi(x;t)},$$4

as an intrinsic diagnostic of the projection error, and uses relevance scores derived from the TDVP quantities already computed at each step to freeze or unfreeze parameters dynamically (Salioni et al., 10 Jun 2025). This approach was benchmarked in overparameterized spin-Jastrow and RBM states, where it improved numerical stability and reduced the need for strong regularization (Salioni et al., 10 Jun 2025).

Two-dimensional tensor-network tVMC addresses a related but distinct conditioning problem: the null directions generated by PEPS gauge redundancy. There, a one-time QR projection eliminates all analytically identifiable null modes in tangent space, yielding a numerically well-conditioned stochastic reconfiguration matrix that can be solved by Cholesky decomposition with only a tiny diagonal regulator (Wu, 7 Dec 2025). In highly expressive neural-network tVMC for electron dynamics, an alternative stabilization strategy is to freeze the nonlinear feature extractor entirely and evolve only a compact linear coefficient layer, thereby bypassing the complex-plane instabilities of full-network evolution (Fu et al., 4 Jun 2026).

These developments clarify a recurring misconception. The main numerical obstacle in tVMC is not merely Monte Carlo noise in a generic sense, but the interaction between stochastic estimation, tangent-space conditioning, ansatz redundancy, and, in some cases, nodal structure. Current research directions therefore emphasize unbiased estimators, geometry-aware regularization, adaptive reduction of active parameter sets, and more expressive yet numerically controllable variational manifolds (Salioni et al., 10 Jun 2025).

7. Scope, comparative position, and outlook

Across its variants, tVMC occupies a distinctive position among many-body simulation methods. It treats continuous-space and lattice models within a unified TDVP framework, naturally accommodates strong interactions and nonlinear drives, and is not limited by the entanglement-growth bottlenecks that constrain tensor-network time evolution in higher dimensions (Gartner et al., 2022). At the same time, its quantitative accuracy is limited by the expressivity of the chosen variational manifold, and in one-dimensional lattice settings more specialized methods such as tDMRG may remain more precise for some observables when the tVMC ansatz is too restricted (Gartner et al., 2022).

The range of demonstrated applications is broad: continuous-space bosons in optical lattices and Lieb–Liniger gases (Gartner et al., 2022); nonequilibrium Hubbard dynamics (Ido et al., 2015); finite-temperature thermodynamics in Hubbard models (Takai et al., 2015); linear-response excitation energies in atoms (Mussard et al., 2017); ab initio electron dynamics in atoms and molecules (Nys et al., 2024); flow-based continuous-variable simulations (Stokes et al., 2022); dissipative many-body dynamics via Lindblad trajectories (Apostoli et al., 30 Jun 2025); and real-time PEPS dynamics on $$E_{\mathrm{loc}}(x;t)=\frac{[H\Psi(x;t)]}{\Psi(x;t)},$$5 and $$E_{\mathrm{loc}}(x;t)=\frac{[H\Psi(x;t)]}{\Psi(x;t)},$$6 lattices up to $$E_{\mathrm{loc}}(x;t)=\frac{[H\Psi(x;t)]}{\Psi(x;t)},$$7–$$E_{\mathrm{loc}}(x;t)=\frac{[H\Psi(x;t)]}{\Psi(x;t)},$$8 using a single GPU (Wu, 7 Dec 2025).

A plausible implication of this body of work is that tVMC is evolving from a single algorithmic recipe into a family of projection-based stochastic methods whose common core is the TDVP covariance equation, but whose practical success increasingly depends on estimator design, geometric conditioning, and ansatz engineering. The documented extensions to backflow, neural-network quantum states, hybrid Jastrow–Slater manifolds, gauge-fixed PEPS, and neural bases indicate that the formalism is compatible with progressively richer state representations (Fu et al., 4 Jun 2026). The recent emphasis on unbiased sampling, adaptive expressivity control, and open-system trajectory formulations further suggests that future progress will be driven less by changing the basic TDVP equation than by refining how the tangent-space projection is represented, sampled, and stabilized (Krinitsin et al., 5 May 2026).