Propagating sparsely supported states with time-dependent neural quantum states

AI summary

Overview

Research area: Computational many-body quantum physics — specifically time-dependent variational Monte Carlo with neural quantum states (NQSs) applied to quantum dynamics of spin lattice models.

Technical level: Advanced. The paper assumes familiarity with the time-dependent variational principle, Monte Carlo importance sampling, quantum geometric tensors, and neural network ansätze for wave functions.

Scope (one sentence): The paper introduces an "interpolation sampling" scheme, together with continuous-time and Krylov-biased sampling strategies, that lets global-in-time variational Monte Carlo propagate sharply peaked initial states (such as fully polarized or single-magnon states) that standard Born-distribution sampling cannot handle.

What This Paper Is About

Time-dependent neural quantum states are trained by minimizing how badly the trial wave function violates the time-dependent Schrödinger equation, and that minimization is normally estimated by Monte Carlo sampling from the Born distribution |Ψ|². When the initial state is sharply peaked — supported on very few configurations — the wave function has zeros exactly where its time derivative and the Hamiltonian-applied state are nonzero, so the sampled estimate of the loss diverges. The authors propose to sample instead from an interpolation function that blends the initial state with the time-evolved neural network, which has support wherever the time derivative does, and they add two sampling-efficiency improvements.

Key Contributions

1. Interpolation sampling. A new importance-sampling target for global-in-time variational Monte Carlo (gVMC). The wave function is written as Ψ_θ(σ,t) = (1 − f_α(t))Ψ₀(σ) + f_α(t)F_θ(σ,t), with f_α(t) = αt/(T + (α−1)t), and samples are drawn from |F_θ(σ,t)|² rather than |Ψ_θ(σ,t)|². This removes the division by zero that occurs in the local-loss estimator when Ψ has zeros where Ψ̇ and ĤΨ do not.

2. Configuration-time joint sampling. Time is treated as a continuous random variable and sampled jointly with spin configurations in ℋ × [0,T] using a Metropolis–Hastings scheme, replacing the previous discretized sum over fixed time points {t_i} and the bias it introduces.

3. Hybrid Krylov-biased proposal. A proposal rule that mixes random draws from a Krylov set S_K = ∪_{k=0}^K Ĥ^k S₀ with the standard single-spin-flip local rule, with default mixing probability p = 0.5, so that the Markov chain reaches the dynamically relevant region of Hilbert space quickly while still exploring elsewhere.

4. A tNQS architecture with explicit time input. Time is encoded through a SIREN subnetwork using both sine and cosine activations, with the number of learned frequencies set equal to the system size N, and the transformed time variable is concatenated with the spin configuration before a tanh multilayer perceptron producing the real and imaginary parts of log F_θ.

Main Findings

• Support mismatch is real and diagnosable. For the 3×3 transverse-field Ising model (TFIM) at h = h_c, with T = 0.1 and the scaling f₁(t) = t/T, the authors compare the time-local normalized distributions of |Ψ|², |Ψ̇|², and |F|² at t = 0, 0.01, and 0.1 for the z-polarized state |↑⟩^⊗N and the single-magnon state |k_x = 0, k_y = 0⟩, plotting the 250 configurations with the largest probability coefficients. This illustrates that the support of the wave function does not contain the support of its time derivative.

• F_θ interpolates between the initial state and the dynamics. At t = 0 the interpolation function satisfies F_θ(σ,0) = Ψ₀(σ) + Ψ̇_θ(σ,0)·T/α, i.e. it is the first-order Taylor expansion of the time-evolved state about t = 0 evaluated at t = T/α; at t = T it coincides exactly with Ψ_θ(σ,T). It can also be expanded as Ψ_θ(σ,t) plus (1 − f_α(t))/f_α(t) times the sum over k ≥ 1 of t^k/k! times the k-th time derivative of Ψ_θ at t = 0.

• The interpolation target resolves the divergence. Because F_θ has support at least as large as that of Ψ̇_θ and ĤΨ_θ, the local loss L_loc(σ,t) = Ψ̇_θ/F_θ + i(ĤΨ_θ)/F_θ remains finite and the Monte Carlo estimate of the loss and its gradient is well defined and stable.

• tVMC fails through a different mechanism. For an initial state supported on a single configuration σ₀, the quantum geometric tensor S_mn = Cov(O*_m, O_n) and force vector G_m = Cov(O*_m, E_loc) both vanish, since the covariance of any two functions under a delta distribution is zero. With the usual small regularization λ𝟙, the only solution is θ̇ = 0, so the parameters freeze; the paper notes this frozen dynamics was observed in prior work and was circumvented there by an evolve-and-project scheme, which has no direct analogue in the global-in-time framework.

• The hybrid proposal breaks detailed balance, deliberately. If σ ∈ S_K and σ′ ∉ S_K differ by two spin flips, the transition probability from σ to σ′ is zero while the reverse is p/|S_K|. The authors argue the resulting bias is tolerable because the Krylov set captures the most dynamically relevant region and the local loss is driven to zero during training.

• Claimed accuracy gain. The abstract reports that, for the two-dimensional TFIM, interpolation sampling allows accurate propagation of extremely peaked states that standard wave function sampling could not tackle. The specific numerical comparisons of accuracy and efficiency are in the parts of Section IV that are not contained in the provided text, so no error values, energy comparisons, or timing figures can be quoted here.

Methodology in Plain English

The authors do not step the wave function forward in small time increments. Instead, they optimize a neural network over a whole time window [0,T] at once, so the network must satisfy the time-dependent Schrödinger equation at every time in that window — a scheme they call global-in-time variational Monte Carlo, based on the McLachlan variational principle. The quantity being minimized is the squared residual of the Schrödinger equation, averaged over configurations and integrated over time; in practice both the configuration sum and the time integral are estimated by Monte Carlo.

The obstacle is that the natural sampling distribution, |Ψ|², is zero on configurations where the residual is not zero, producing a division by zero for peaked states. Their fix is to sample from a different function, F_θ, constructed so that at early times it already carries the information in the time derivative of the wave function, and at the end of the window it equals the wave function itself. Sampling from |F|² instead of |Ψ|² makes the estimator finite.

Two practical refinements reduce cost. First, the time variable is sampled continuously together with spin configurations in a single Metropolis–Hastings chain, so no discretized time grid is needed and no time-discretization bias is introduced. Second, the proposal step is split: half the time a new configuration is drawn from a Krylov set built by repeatedly applying the Hamiltonian to the initial state's support (which is exactly the region reached by short-time dynamics, since the Taylor expansion of e^{−iĤt} stays in a finite Krylov subspace at early times), and half the time from ordinary single-spin flips. This gets the chain into the right corner of Hilbert space quickly without trapping it there.

For benchmarking, the authors use NetKet, JAX, and Flax for the neural quantum states, and TeNPy for matrix product state references at bond dimensions 64 and 128, propagated with TDVP at a time step of 0.001. Networks are trained with the continuous resilient (CoRe) optimizer using an exponentially decaying learning rate η = η₀·r^{m/M} with defaults η₀ = 0.005, r = 0.5, and M = 1000. Observables on the 4×4 lattice (Hilbert space dimension 2^16) are computed by explicit summation rather than Monte Carlo, so that measurement error is eliminated and only optimization error remains. The TFIM is studied on a two-dimensional square lattice at its critical field h_c = 3.04438, starting from three initial states of different support: the z-polarized state (support on one configuration), a single magnon with k_x = k_y = 0 (support on all single-spin-flip configurations, represented with suppression exponent R = 30), and the x-polarized state (uniform support over the full Hilbert space).

Why This Matters

Impact on research. The paper challenges the default assumption in the neural-quantum-state community that the Born distribution, or a close variant, is the right sampling basis for variational Monte Carlo. It extends time-dependent NQS methods to initial states that are ubiquitous in quench-dynamics studies — Néel states, domain wall states, and antiferromagnetically ordered states in cold-atom experiments — all of which violate the positivity/broad-support requirement. It also sidesteps the singular quantum geometric tensor inversion and parameter freezing that plague sequential time-dependent variational Monte Carlo.

Real-world applications (as motivated by the classes of problem the method targets):

• Quench dynamics of antiferromagnetically ordered states in cold-atom quantum simulators.
• Real-time dynamics of spin lattices relevant to frustrated magnetism and nonequilibrium phases.
• Benchmarking and cross-checking tensor-network time-evolution results where bond dimension must grow prohibitively to cross the entanglement barrier.
• Many-electron and quantum-emitter dynamics, following the tNQS architectures the paper cites as prior applications.

Industry relevance. The work sits in the emerging quantum-simulation software stack: it is implemented on open-source, accelerator-friendly tooling (NetKet, JAX, Flax) and is directly relevant to groups using neural-network wave functions as a classical emulator for quantum hardware experiments and for verifying quantum dynamics on near-term devices.

Future Directions

• Quantifying the detailed-balance bias. The hybrid Krylov proposal provably violates detailed balance; establishing a quantitative bound on the resulting sampling bias, or designing a corrected accept/reject step that restores balance while keeping the efficiency gain, is an obvious next step.
• Systematic hyperparameter study. The roles of the interpolation rate α, the interval length T, the Krylov order K, and the mixing probability p (fixed at 0.5 as a default here) are not resolved by a single demonstration and invite controlled sweeps.
• Scaling beyond exact summation. Observables here were computed by explicit summation on a 4×4 lattice with a 2^16-dimensional Hilbert space; extending to larger lattices requires accurate Monte Carlo estimation of observables, not only of the loss.
• Transfer to other Hamiltonian classes. The Krylov-set construction depends on the sparsity structure of Ĥ and the initial support; testing whether it remains efficient for Hubbard-type or long-range models, and for the many-electron real-space settings the paper cites, is an open question.

Target Audience

Researchers and graduate students in computational quantum many-body physics and quantum chemistry who work on real-time dynamics with variational Monte Carlo or neural quantum states; practitioners of tensor-network methods interested in the entanglement barrier and in cross-validation references; and machine-learning researchers interested in how the choice of sampling distribution, not just the network architecture, determines what a variational ansatz can actually learn. Readers should be comfortable with variational principles, local energies, and Markov chain Monte Carlo.

Authors’ abstract

Neural quantum states (NQSs) have emerged as a powerful ansatz for quantum dynamics. Their high entanglement capacity promises to overcome the entanglement barrier. However, existing time-dependent NQS methods are ill-equipped to propagate sparsely supported initial states due to the inherent mismatch between the supports of such states and their time derivatives. To address this issue, we introduce the interpolation sampling method as a novel form of importance sampling, where samples are drawn from a distribution that interpolates between the wave function and its time derivative. Combined with a global-in-time variational principle, we demonstrate with the example of the two-dimensional transverse-field Ising model that interpolation sampling allows for accurate propagation of extremely peaked states that standard wave function sampling could not tackle. We further improve sampling efficiency by introducing (i) a configuration-time joint sampling scheme where spin configurations and time are both treated as random variables, and (ii) a hybrid strategy for sample proposal that incorporates the knowledge of a Krylov subspace. Our work extends the scope of time-dependent NQS methods to a wider range of physically relevant scenarios, while challenging the prevailing reliance on the Born distribution and its close variants as the default basis for sampling.