Time-Dependent Variational Principle (TDVP)

• TDVP is a variational method that restricts quantum evolution to a chosen manifold, preserving symmetries and minimizing variational error.
• It employs algorithmic frameworks like matrix product states and tensor networks to efficiently simulate high-dimensional, strongly correlated quantum systems.
• Recent advances extend TDVP to dissipative and open quantum dynamics, enhancing numerical stability and scalability in practical simulations.

The time-dependent variational principle (TDVP) is a foundational method for simulating quantum dynamics in high-dimensional and strongly correlated systems by restricting time evolution to a chosen variational manifold. Originally formulated for wavefunction dynamics, the TDVP provides a mathematically rigorous path to optimal, symmetry-preserving approximations for time evolution and yields diverse algorithmic frameworks, most notably for matrix product states (MPS), tensor networks, and Gaussian state spaces (Haegeman et al., 2011, Haegeman et al., 2014, Bauernfeind et al., 2019, Rowan et al., 2019, Guaita et al., 2019, Lihm et al., 2020). The TDVP has been extended to dissipative dynamics, mixed states, and open quantum systems, with implementations ranging from analytic Gaussian ansätze to neural-network-based representations of quantum states (Reh et al., 2021, Kraus et al., 2012). This article details its variational architecture, tangent-space projection, algorithmic forms, error structure, application breadth, and recent methodological advances.

1. Variational Principle and Projected Dynamics

The TDVP is a variational projection of exact quantum dynamics onto a tractable submanifold of states, defined by parameters $\theta = (\theta_1, \ldots, \theta_N)$. For Schrödinger evolution,

$$i\,\partial_t\,|\Psi(\theta)\rangle = \hat H\,|\Psi(\theta)\rangle,$$

TDVP demands that $|\Psi(\theta(t))\rangle$ evolve such that its time derivative is the orthogonal projection of $\hat H\,|\Psi(\theta)\rangle$ onto the manifold’s tangent space: $$i\,P_{T_\Psi}\,\partial_t|\Psi(\theta)\rangle = P_{T_\Psi}\,\hat H\,|\Psi(\theta)\rangle,$$ where $P_{T_\Psi}$ is the projector onto the tangent space at $|\Psi(\theta)\rangle$ (Haegeman et al., 2011).

This is variationally justified via Dirac-Frenkel stationarity of the action

$$S[\Psi(\theta)] = \int dt\,\langle\Psi(\theta)|i \partial_t - \hat H|\Psi(\theta)\rangle,$$

leading to Euler–Lagrange-type equations

$$\langle\partial_j\Psi|\partial_i\Psi\rangle\,\dot{\theta}^i = -\langle\partial_j\Psi|\hat H| \Psi\rangle,$$

with $\partial_i = \frac{\partial}{\partial \theta^i}$ and the Gram matrix $$G_{ji} = \langle\partial_j\Psi|\partial_i\Psi\rangle$$. After gauge fixing, $G$ becomes invertible and the flow reduces to a system of explicit ODEs (Haegeman et al., 2011, Haegeman et al., 2014).

In the McLachlan variant, one minimizes the norm $\|i\,\partial_t|\Psi\rangle - \hat H|\Psi\rangle\|$ over tangent vectors, resulting in equivalent projective equations (Rowan et al., 2019).

2. Tangent Space, Gauge, and Projector Construction

For a variational manifold parameterized by tensors or wavefunction coefficients, the tangent space $T_\Psi$ consists of all infinitesimal variations generated by the parameters. The orthogonal projector can be written as

$$P_{T} = \sum_{i,j} |\partial_i \Psi\rangle\, (G^{-1})^{ij} \langle\partial_j \Psi|,$$

where $G$ is the metric/Grothendieck (Gram) matrix. This formalism ensures that only physical, gauge-fixed tangent vectors contribute to the dynamics. When redundancies due to parameterization (such as MPS gauge freedom) exist, gauge fixing—e.g., by canonical forms—renders $G$ invertible and the equations numerically stable (Haegeman et al., 2011, Haegeman et al., 2014, Bauernfeind et al., 2019).

For MPS, the canonical site decomposition and tangent-space construction permit an explicit splitting of the projector into sums of one-site and bond projectors (Haegeman et al., 2011, Haegeman et al., 2014), providing the algebraic backbone for practical local-update algorithms.

3. Algorithmic Realizations: MPS, Tensor Networks, Gaussians

TDVP admits algorithmic implementations across various variational classes:

• Matrix Product States (MPS): For 1D quantum chains, the TDVP is most often implemented on the MPS manifold, either in uniform (infinite) or finite-size settings. The dynamics is realized by sweeping local (one-site or two-site) updates, in which each tensor (or pair) is evolved under an effective Hamiltonian constructed from the network’s environment. Imaginary-time TDVP delivers efficient ground-state optimization resembling DMRG, while real-time TDVP is symplectic, energy-conserving, and numerically stable (Haegeman et al., 2011, Haegeman et al., 2014).
• Tree Tensor Networks (TTNs): For arbitrary loop-free tensor networks, the TDVP constructs tangent projectors respecting the network’s gauge and locality. Updates take the form of single-site or multi-site evolutions interleaved with canonicalization sweeps. The flexibility of TTN geometry allows application to impurity problems and hybridization with Matrix Product Operators (MPOs) (Bauernfeind et al., 2019).
• Gaussian States and Bosonic/Fermionic Systems: TDVP on Gaussian (coherent/squeezed) manifolds generates coupled ODEs for displacements and covariances, capturing both mean-field and quantum fluctuation effects. The formalism provides accurate dynamics for field theories, bosonic lattice systems, and optimal approximations for ground states and linear response (Rowan et al., 2019, Guaita et al., 2019, Lihm et al., 2020).

4. Error Structure, Conservation Laws, and Complexity

TDVP evolution at every step is optimal within the chosen manifold: the residual $i\partial_t|\Psi\rangle - \hat H|\Psi\rangle$ is minimized in norm or orthogonally projected out. The only source of deviation from exact dynamics is the geometric (variational) error—how much $\hat H|\Psi\rangle$ fails to lie in $T_\Psi$—which can be monitored at each time step. In contrast to Trotter-based algorithms, there is no Trotter error; time step errors enter only through the ODE solution (Haegeman et al., 2011, Haegeman et al., 2014).

Crucially, TDVP exactly conserves norm and energy (for time-independent Hamiltonians) and any symmetry generator commuting with $\hat H$, a property absent in TEBD or split-step integrators (Haegeman et al., 2011, Goto et al., 2018). The computational complexity for each time step scales as $O(D^3)$ for MPS of bond dimension $D$, matching TEBD but often with substantially reduced truncation-induced error, especially in real-time evolution (Haegeman et al., 2011, Haegeman et al., 2014, Sander et al., 13 Aug 2025). For Gaussian manifolds, the ODEs are of size $O(N^2)$ for $N$-mode systems (Guaita et al., 2019).

5. Extensions: Dissipative, Mixed-State, and Open Quantum Dynamics

TDVP generalizes to mixed-state and dissipative dynamics governed by Lindblad equations. In these settings, the equation of motion for the density matrix $\rho$,

$\dot{\rho} = \mathcal{L}[\rho],$

is projected onto the tangent space of a variational manifold of mixed states. For pure-state ansatzes (e.g., MPS with statistical sampling), a Fokker–Planck or stochastic differential equation is derived for the variational parameters, with the drift and diffusion determined by projected Lindbladian action (Transchel et al., 2014).

In the general mixed-state setting, the projection depends on a choice of monotone Riemannian metric (e.g., quantum Fisher, Bures–Helstrom), inducing a family of TDVP flows. For fermionic Gaussian states, all such metrics coincide and the TDVP reduces to “Gaussification”—preserving Gaussianity under dissipative evolution (Kraus et al., 2012). Algorithms extend to neural-network ansatzes for open quantum systems by recasting the density operator as a POVM-based probability distribution parameterized by deep autoregressive networks and using TDVP to project the Lindblad flow, with stochastic sampling for force and metric evaluation (Reh et al., 2021).

In open system molecular dynamics, constrained TDVP variants using the square-root NOSSE formulation enforce trace conservation and recover energy conservation in the closed-system limit, eliminating basis-induced artificial dissipation (Joubert-Doriol et al., 2015).

6. Numerical Strategies and Modern Extensions

Recent advances have focused on improving numerical stability, flexibility, and applicability:

• Controlled Bond Expansion (CBE): One-site TDVP is supplemented with local bond-dimension increases whenever the projection error or variance bound exceeds a prescribed threshold, improving accuracy in entangling dynamics but retaining O($D^3$) scaling (Li et al., 2022).
• Stochastic Adaptive Bond Growth: The stochastic adaptive 1-TDVP (SA-1TDVP) algorithm grows MPS bond dimension by sampling new singular values according to level-spacing statistics of the entanglement Hamiltonian, achieving near-2TDVP accuracy at a fraction of the computational cost (Xu et al., 2021).
• Krylov-Enriched TDVP: Ancilla-Krylov-enriched TDVP augments the tangent space dynamically with global Krylov vectors, dramatically reducing projection error and allowing larger time steps without sacrificing unitarity (Yang et al., 2020).
• Quantum Circuit Simulation: In circuit contexts, TDVP has been tailored for quantum circuits by “local” projector splitting, efficiently handling long-range gates and entanglement diffusion, and outperforming TEBD for large-scale circuits by spreading bond growth globally (Sander et al., 13 Aug 2025).
• Clifford and Neural Network Augmentations: Clifford-dressed and Clifford-circuit-augmented TDVP interleaves periodically applied entanglement-cooling Clifford gates with TDVP sweeps, transferring stabilizer entanglement to the Clifford layer and reducing the necessary MPS bond dimension for accurate long-time simulation. Neural network-based TDVP attaches variational learning to density-matrix propagation using neural architectures (Mello et al., 2024, Qian et al., 2024, Reh et al., 2021).

7. Applications, Performance, and Scope

TDVP underpins state-of-the-art simulations across many areas:

• Quantum Lattice Dynamics: TDVP-MPS is optimal and stable for global quantum quenches, transport, and real-time correlation functions in 1D and quasi-1D systems, including those with long-range or time-dependent Hamiltonians (Haegeman et al., 2011, Haegeman et al., 2014, Li et al., 2022).
• Ground State and Imaginary-Time Optimization: Imaginary-time TDVP unifies time evolution and ground-state search, effectively recovering the DMRG algorithm in the infinite-time-step limit (Haegeman et al., 2014).
• Non-equilibrium and Finite-Temperature Dynamics: Gaussian TDVP frameworks provide rigorous access to spectral, linear response, and thermal correlations in Bose-Hubbard, anharmonic lattice, and related bosonic systems (Guaita et al., 2019, Lihm et al., 2020).
• Open Quantum Systems: Lindblad-TDVP variants capture dissipative Many-Body dynamics far beyond exact diagonalization sizes (Reh et al., 2021). For Gaussian states, TDVP provides unique, geometry-independent evolution (Kraus et al., 2012).
• Tensor Network Generalization: TDVP extends seamlessly to tree tensor networks (TTN), tree-like tensor structures for impurity problems, and quantum chemistry applications (Bauernfeind et al., 2019).
• Quantum Circuits and Benchmarking: Quantum circuit simulation using local-TDVP achieves new limits on 49-qubit circuit depth and complexity, outperforming TEBD especially for hardware-efficient and long-range circuits (Sander et al., 13 Aug 2025).

Overall, the time-dependent variational principle offers a robust, systematically optimal, and symmetry-respecting framework for quantum dynamics and optimization, adaptable to a wide range of variational manifolds and system classes, with broad impact on condensed matter, quantum information, and quantum chemistry (Haegeman et al., 2011, Haegeman et al., 2014, Reh et al., 2021, Li et al., 2022).