Entanglement-Aware Product-Formula Simulation

Use the correlation structure of the evolving state to estimate the Trotter error of a product-formula simulation more tightly, and adapt the step size to reach a target accuracy with fewer steps where that structure allows.

The idea

Product-formula (Trotter) simulation splits a Hamiltonian into easy-to-exponentiate partitions and applies them in sequence. The worst-case Trotter error bound typically depends only on commutator norms of those partitions, and can be pessimistic. An entanglement-aware analysis goes further by accounting for the correlation structure of the current state: in some many-body evolutions the actual Trotter error stays well below the worst-case bound and closer to the average case, so the same accuracy can be reached with fewer steps. This does not always hold — it depends on the Hamiltonian and the state.

Partitioned Trotter step

import qalgora

@qalgora.kernel
def pf1_step(q: qalgora.qview, dt: float):
    # first-order product formula, partitioned by interaction graph
    apply_partition_a(q, dt)
    apply_partition_b(q, dt)

Adaptive error estimation

To adapt the step size you need a cheap, local estimate of the Trotter error. One way is to evolve the same interval twice at different resolutions and compare: a coarse trajectory at step size dt, and a fine / reference trajectory at dt/2 with twice the steps. The difference between the two estimates the local Trotter error, which then drives the step-size (or partitioning) choice. This two-trajectory, adaptive error estimate is the whole of the scheme — there is no separate physical process.

import qalgora

# evolve the same interval at two resolutions and compare
state_coarse = qalgora.get_state(evolve_kernel, steps, dt)
state_fine   = qalgora.get_state(evolve_kernel, 2 * steps, dt / 2)
error_est    = estimate_trotter_error(state_fine, state_coarse)
entropy      = qalgora.tensor.entanglement_entropy(state_coarse, cut=n // 2)
print("S =", entropy, " err =", error_est)

Limitations

• It does not change the worst-case complexity, and does not guarantee fewer steps for an arbitrary Hamiltonian — the saving appears only when the state's entanglement / correlation structure makes the true error smaller than the worst-case bound.
• The error estimate itself has a cost — the extra fine-resolution evolution (or the measurements that replace it on hardware) — which must be weighed against the steps it saves.