Time-Dependent DMRG: Methods & Applications

• TD-DMRG is a versatile method that uses matrix-product states to simulate real and imaginary time evolution in many-body quantum systems.
• It integrates techniques like TEBD and TDVP to manage local effective Hamiltonians and control entanglement growth, even for long-range interactions.
• Its algorithmic variants are applied across lattice physics, quantum chemistry, and molecular dynamics to accurately compute spectra, correlation functions, and dynamic responses.

Time-Dependent Density Matrix Renormalization Group (TD-DMRG) denotes a family of matrix-product-state (MPS) methods for real- and imaginary-time propagation of many-body quantum states, together with closely related operator and density-matrix formulations used to compute spectra, correlation functions, response properties, and non-equilibrium dynamics. In the earlier literature, the term often referred specifically to Suzuki–Trotter-based MPS evolution as implemented in time-evolving block decimation (TEBD) and “adaptive TD-DMRG,” whereas later work placed these methods inside a broader time-dependent variational principle (TDVP) framework and connected time evolution directly to sweep-based DMRG optimization on the MPS manifold (Haegeman et al., 2014).

1. Historical scope and terminological evolution

In the literature on one-dimensional quantum many-body systems, TD-DMRG originally meant real-time evolution of MPSs by factorizing the Hamiltonian $H=\sum_i h_i$ into local terms and applying local two-site gates $e^{-i h_{\ell,\ell+1}\Delta t}$, typically through TEBD or adaptive TD-DMRG. This approach is straightforward for nearest-neighbor Hamiltonians, but it becomes inefficient or inaccurate when interactions are long-ranged, when the Hamiltonian does not admit a convenient local decomposition, or when one must approximate long-range exponentials as truncated MPOs (Haegeman et al., 2014).

A second line of development arose in quantum chemistry and molecular physics, where ab initio Hamiltonians contain long-range two-electron terms and where observables of interest are often dynamical correlation functions rather than only state trajectories. In that setting, TD-DMRG and dynamical DMRG (DDMRG) became complementary algorithms: TD-DMRG propagates in real time and Fourier-transforms correlation functions such as $G_{ij}(t)$, whereas DDMRG works directly in the frequency domain through a correction-vector equation at a chosen $\omega$ (Ronca et al., 2017). This distinction matters because TD-DMRG provides broad spectral information in one run and naturally treats observables formulated in real time, while DDMRG is advantageous when only a narrow frequency window is required (Ronca et al., 2017).

A broader conceptual synthesis followed when TDVP was identified as a unifying language for MPS time evolution and optimization. In that formulation, TEBD becomes one instance of a more general projected dynamics, and imaginary-time propagation connects continuously to ground-state DMRG optimization (Haegeman et al., 2014). Later perspective work in chemistry and molecular physics presented TD-DMRG not as a single algorithm, but as a class of MPS/MPO propagation schemes that includes TEBD, Runge–Kutta and Krylov propagators, finite-temperature extensions, and TDVP-based tangent-space integrators particularly suited to ab initio Hamiltonians with long-range interactions (Baiardi et al., 2019).

2. Variational formulation on the MPS manifold

The central variational statement of modern TD-DMRG is the projection of the Schrödinger equation onto the tangent space of the fixed-rank MPS manifold. In its standard form,

$$i\,\partial_t |\Psi(t)\rangle = \mathcal{P}_T H |\Psi(t)\rangle,$$

where $\mathcal{P}_T$ is the orthogonal projector onto the tangent space at the current MPS (Haegeman et al., 2014). In mixed-canonical form, the state may be written with a one-site center tensor $A_C(n)$ or a zero-site bond matrix $C(n)$, and the tangent-space projector admits an explicit gauge-invariant splitting into a sum of one-site projectors and a sum of bond projectors,

$$P_{T|\Psi(A)\rangle\mathrm{MPS}} = \sum_{n=1}^N P_L^{[1:n-1]} \otimes 1_n \otimes P_R^{[n+1:N]} - \sum_{n=1}^{N-1} P_L^{[1:n]} \otimes P_R^{[n+1:N]}.$$

This decomposition is independent of the residual unitary freedom within the Schmidt bases and annihilates pure gauge directions (Haegeman et al., 2014).

A crucial consequence is that each split term is exactly integrable. For the one-site term, the projected dynamics reduces to a linear equation for the center tensor,

$A_C(n,t)=\exp[-H(n)t]\,A_C(n,0),$

while for the bond term one obtains

$e^{-i h_{\ell,\ell+1}\Delta t}$0

with $e^{-i h_{\ell,\ell+1}\Delta t}$1 and $e^{-i h_{\ell,\ell+1}\Delta t}$2 Hermitian effective Hamiltonians built from standard DMRG/MPO environments (Haegeman et al., 2014). This turns time evolution into a sweep over local effective problems whose structure is so close to finite-size DMRG that the same environment construction, orthogonalization, and center-moving machinery can be reused with only minimal code changes (Haegeman et al., 2014).

The same projector logic extends to a two-site tangent space. The resulting two-site TDVP integrator updates a two-site center block, factorizes it, and allows bond-dimension growth followed by truncation. For nearest-neighbor Hamiltonians, $e^{-i h_{\ell,\ell+1}\Delta t}$3 lies exactly in the space of two-site variations, which explains why TEBD and two-site TDVP share the same leading-order behavior in local models even though one splits the Hamiltonian and the other splits the tangent-space projector (Haegeman et al., 2014).

The imaginary-time limit makes the relation to optimization explicit. Under $e^{-i h_{\ell,\ell+1}\Delta t}$4, the one-site update $e^{-i h_{\ell,\ell+1}\Delta t}$5 becomes a ground-state projector, and in the limit $e^{-i h_{\ell,\ell+1}\Delta t}$6 it reproduces the usual one-site DMRG eigenvector update. In this sense, DMRG is obtained as a special case of imaginary-time evolution with infinite time step (Haegeman et al., 2014). Closely related sweep-based equations with exactly integrable local ordinary differential equations were also used for molecular excitonic and vibronic Hamiltonians, where the local evolution is performed in canonical MPS gauges using effective Hamiltonians assembled from left and right boundary contractions (Baiardi et al., 2019).

3. Major algorithmic variants

The oldest and still important class of TD-DMRG algorithms is based on Hamiltonian splitting. TEBD and adaptive TD-DMRG apply Suzuki–Trotter factorizations of $e^{-i h_{\ell,\ell+1}\Delta t}$7 into local gates; for long-range models, alternative adaptive schemes were devised to improve basis construction. A prominent example is the double time window targeting (DTWT) algorithm, which combines pace-keeping and time-step targeting ideas by constructing the reduced density matrix from a double window of future states generated with MSD2, while performing accurate propagation on an embedded single window with RK4. DTWT was used for real-time spin–charge dynamics in the Pariser–Parr–Pople model with long-range electron correlations, where it revealed that the charge velocity is almost twice the spin velocity for standard PPP parameters (Dutta et al., 2010).

In ab initio electronic structure, Feiguin–White time-step targeting was reformulated directly in the renormalized basis. There, an RK4 propagator is combined with multiple time-slice targeting, and the reduced density matrix at each site is built from $e^{-i h_{\ell,\ell+1}\Delta t}$8, $e^{-i h_{\ell,\ell+1}\Delta t}$9, $G_{ij}(t)$0, and $G_{ij}(t)$1 with empirically chosen weights. This approach avoids Trotterization of long-range electronic Hamiltonians, and its improved variants, DDMRG$G_{ij}(t)$2 and TD-DMRG$G_{ij}(t)$3, decouple renormalized bases across states to improve accuracy at fixed bond dimension with nominal $G_{ij}(t)$4 overhead (Ronca et al., 2017).

A more systematic classification emerges from later benchmarks on realistic chemical systems. In that setting, four propagators were compared directly: global Taylor expansion, global Krylov, one-site TDVP, and two-site TDVP. The global Taylor scheme was found to be unreliable in demanding vibronic calculations; global Krylov was accurate but generated highly entangled intermediate states; one-site TDVP required sufficiently large pre-grown bond dimensions; and two-site TDVP provided the most favorable balance of robustness and efficiency for large pyrazine and singlet-fission models (Xie et al., 2019). Related work on the Fenna–Matthews–Olson complex compared propagate-and-compress Runge–Kutta, TDVP with matrix unfolding, and TDVP with projector splitting, and showed that the two TDVP variants agree once the time step is converged, that both are more accurate than P&C-RK4, and that TDVP-PS tolerates much larger time steps (Li et al., 2019).

The distinction between one-site and two-site TDVP remains central. One-site TDVP is symplectic, preserves norm, and for time-independent $G_{ij}(t)$5 preserves energy within the fixed MPS manifold, but it cannot increase bond dimension. Two-site TDVP relaxes that constraint and is therefore better adapted to entanglement growth, at the price of truncation and the loss of exact conservation laws after each SVD (Haegeman et al., 2014).

4. Error structure, conservation laws, and computational scaling

TD-DMRG errors are multi-component. In projector-splitting TDVP, a single left-to-right sweep is first order with local error $G_{ij}(t)$6, while a symmetric left-to-right plus right-to-left composition is second order with local error $G_{ij}(t)$7. Higher-order compositions are available. Even when each local exponential is evaluated accurately, the projected dynamics differs from exact Schrödinger evolution by the variational-manifold residual

$G_{ij}(t)$8

which can be evaluated in practice. In the two-site variant, truncating singular values below a threshold $G_{ij}(t)$9 introduces an additional $\omega$0 contribution. Each local Krylov/Lanczos exponential application scales as $\omega$1, and the overall cost is comparable to a DMRG ground-state sweep, independent of whether the underlying Hamiltonian is short-ranged or long-ranged (Haegeman et al., 2014).

The limiting resource is usually entanglement growth. In the MPS/TTNS setting, accurate representation across a cut with von Neumann entropy $\omega$2 requires $\omega$3, and recent nonadiabatic benchmarks emphasized that the maximum bond entropy is often a stricter convergence indicator than the primary observable itself. The same study showed that nonuniform bond-dimension profiles inherited from early two-site growth can become later bottlenecks, whereas fixed uniform bond dimensions and large-$\omega$4 extrapolation are more reliable (Li et al., 30 Aug 2025).

Recent parameter studies for molecular electron dynamics reinforce that practical convergence behavior depends strongly on algorithmic choices beyond bond dimension alone. For charge migration in chloroacetylene and furfural, projector-splitting TDVP exhibited smooth convergence in both $\omega$5 and $\omega$6, whereas time-step targeting showed non-monotonic time-step convergence. Full complex MPS representations retained norm far better than state-averaged complex ones, and singlet embedding reduced runtimes for non-singlet dynamics while slightly improving $\omega$7-convergence (Wahyutama et al., 2024).

Performance engineering has become a substantial part of the subject. GPU acceleration of contraction-dominated TDVP kernels produced speedups of up to 73 times in benchmarks on the FMO complex, whereas propagate-and-compress schemes benefited less because their bottlenecks remained QR and SVD factorizations on the CPU (Li et al., 2019). Parallel real-space TDVP extended long-range time evolution to 32 processes with parallel efficiencies as high as 86%, and enabled dynamical correlation functions for a 201-site Heisenberg chain with $\omega$8 interactions that would be difficult to obtain sequentially (Secular et al., 2019).

5. Domains of application

In lattice many-body physics, TD-DMRG has been used for quenches, transport, and dynamical correlation functions in both short-range and long-range models. A benchmark example is the long-range XY Hamiltonian

$\omega$9

for which the two-site TDVP integrator with a fourth-order composition and $$i\,\partial_t |\Psi(t)\rangle = \mathcal{P}_T H |\Psi(t)\rangle,$$0 on $$i\,\partial_t |\Psi(t)\rangle = \mathcal{P}_T H |\Psi(t)\rangle,$$1 spins reproduced light-cone behavior and leakage patterns for $$i\,\partial_t |\Psi(t)\rangle = \mathcal{P}_T H |\Psi(t)\rangle,$$2, and remained applicable in the regime $$i\,\partial_t |\Psi(t)\rangle = \mathcal{P}_T H |\Psi(t)\rangle,$$3 where TEBD-based calculations are not directly applicable (Haegeman et al., 2014). In the PPP model for conjugated polymers, DTWT resolved spin–charge separation and showed that long-range electron correlations make both charge and spin propagate faster than in the Hubbard model, with $$i\,\partial_t |\Psi(t)\rangle = \mathcal{P}_T H |\Psi(t)\rangle,$$4–$$i\,\partial_t |\Psi(t)\rangle = \mathcal{P}_T H |\Psi(t)\rangle,$$5 for $$i\,\partial_t |\Psi(t)\rangle = \mathcal{P}_T H |\Psi(t)\rangle,$$6–40 (Dutta et al., 2010).

In molecular and vibronic dynamics, TD-DMRG has moved from reduced models to realistic Hamiltonians. It has been used to compute nearly exact absorption and fluorescence spectra of molecular aggregates at zero and finite temperature, exploiting the strict locality of the zero-exciton Hamiltonian for fluorescence and RK4 propagation in the one-exciton manifold for absorption (Ren et al., 2018). Sweep-based TDVP formulations were then applied to excitonic and vibronic Hamiltonians with more than 20 degrees of freedom, including pyrazine and aggregate models, by exploiting exactly integrable local equations in an MPS/MPO representation (Baiardi et al., 2019). Subsequent chemical benchmarks on pyrazine internal conversion and singlet fission with continuous phonon baths showed that two-site TDVP can describe the full quantum dynamics accurately and efficiently, while bond dimensions for realistic 24-mode and 183-mode problems grow into the hundreds or thousands depending on the truncation threshold (Xie et al., 2019).

In ab initio electronic dynamics, TD-DMRG has been used for molecular dynamical polarizabilities and hyperpolarizabilities, electronic absorption spectra, and ultrafast ionization dynamics under the exact non-relativistic electronic Hamiltonian. Real-time tangent-space TD-DMRG was demonstrated for systems including up to 20 electrons and 32 orbitals, and applied to benzene charge migration, decacene absorption, and BH dynamical response (Baiardi, 2020). Imaginary-time TD-DMRG has also been repurposed as an optimizer for non-Hermitian transcorrelated Hamiltonians: first for the two-dimensional Fermi–Hubbard model with a Gutzwiller correlator, and later for explicitly correlated molecular Hamiltonians with two- and three-body transcorrelated MPOs, where it improved convergence toward the complete-basis-set limit in atoms and first-row diatomics (Baiardi et al., 2020, Baiardi et al., 2022).

6. Current methodological issues and directions

One persistent misconception is that disagreement between TD-DMRG and other tensor-network dynamics methods necessarily signals a conceptual inconsistency. A recent comparison of TD-DMRG and ML-MCTDH for exciton dissociation showed that previously reported discrepancies of up to 60% were caused primarily by insufficient bond dimensions and suboptimal tensor-network topologies. With uniform large bond dimensions, the difference dropped below 10%; extrapolation reduced it to about 4%; and an optimized TreeX topology lowered it further to approximately 2–3%, establishing agreement at the numerically exact level (Li et al., 30 Aug 2025).

A second misconception is that ground-state heuristics transfer unchanged to real-time dynamics. Recent electron-dynamics benchmarks found that this is not generally true. Split localization reduced entanglement and resource usage in some systems, but natural orbitals converged faster with bond dimension in others. Likewise, a density-matrix-averaged orbital selection was sometimes inferior to a hole-density-matrix-adapted construction tailored to the actual dynamics, and the choice between full complex and state-averaged complex MPS representations had a direct impact on norm retention and stability (Wahyutama et al., 2024).

The present state of TD-DMRG is therefore best understood as a technically differentiated toolbox rather than a single algorithm. TEBD remains natural for short-range Hamiltonians; time-step targeting remains attractive when it can be integrated into existing quantum-chemistry codes; one-site TDVP is preferred when exact manifold conservation is essential; two-site TDVP is preferred when entanglement growth must be accommodated; and imaginary-time variants connect time evolution to ground-state or effective non-Hermitian optimization (Haegeman et al., 2014, Baiardi et al., 2019). Across these variants, the decisive practical ingredients are MPO construction, symmetry exploitation, site ordering or tensor topology, bond-dimension control, and explicit monitoring of truncation, time-step, and variational-manifold errors.