Quantum-Enhanced Auxiliary-Field QMC

Combine a quantum trial wavefunction with classical Auxiliary-Field Quantum Monte Carlo (AFQMC) to push past the accuracy of either method alone.

The problem it solves

One way to reach a ground state is imaginary-time projection: apply the operator e−τH to almost any starting state and, as τ grows, every excited component decays faster than the ground state, which survives. AFQMC realises this projection stochastically, propagating a population of walkers (each a Slater determinant) in imaginary time. The obstacle is that for interacting electrons H contains a two-body interaction — quartic in the fermion operators — and the propagator of a two-body term cannot be applied to a single determinant while keeping it a determinant.

Auxiliary fields — the Hubbard–Stratonovich transformation

The trick that gives the method its name is to rewrite the troublesome two-body propagator as an average over auxiliary fields. The Hubbard–Stratonovich identity expresses e−τ·(two-body) as an integral over fluctuating fields of purely one-body propagators. A one-body propagator does keep a determinant a determinant, so each walker now evolves under a random, time-dependent one-body Hamiltonian, and the two-body physics is recovered by Monte Carlo sampling over the auxiliary fields. Taken to the ideal limit — no sign/phase constraint, a vanishing time step, and an infinite number of samples — this is exact; in practice it is further limited by Trotter time-step error, a finite number of walkers, and the sign/phase problem. The standard phaseless / constrained-path approximation uses a trial wavefunction to control the sign/phase problem, which introduces a trial-state-dependent bias.

The sign (and phase) problem

In practice the auxiliary-field weights can turn negative or complex. Walkers then contribute with competing signs that cancel, and the signal-to-noise ratio decays exponentially with system size and projection time — the notorious fermion sign problem (its complex-valued cousin is the phase problem). The standard cure is the constrained-path / phaseless approximation: use a trial wavefunction |ψT⟩ to constrain the walkers' sign or phase, which tames the exponential variance — but introduces a bias that depends entirely on how good |ψT⟩ is.

Where the quantum computer helps

Since the only remaining error is the constraint bias, the accuracy of phaseless AFQMC is set by the quality of the trial state. Classically you are limited to trial states you can both store and evaluate cheaply — a single determinant, or a modest multideterminant expansion. A quantum computer can prepare a richer trial state (for example from VQE), and the classical AFQMC only needs one thing from it: the overlaps ⟨ψT|φ⟩ between the trial state and each walker, a quantity the quantum device can estimate. A more suitable trial state usually reduces the bias, but this depends on how well it matches the walker distribution, the nodal/phase structure, and the target correlations — it is not judged by the VQE variational energy alone — so a better-quality trial can give a more accurate energy without the quantum computer ever having to run the full Monte Carlo itself.

1. Run VQE to produce a high-quality trial state |ψ_T⟩.
2. Use that state to constrain the sign problem in classical AFQMC.
3. Propagate walkers in imaginary time to refine the ground-state energy.

Seeing it in code

Read the snippet against the pipeline above. The first block is the quantum step: qalgora.vqe(...) optimizes the trial state and qalgora.get_state(...) hands back trial_state = |ψT⟩. The second block is the classical step: run_afqmc(..., trial=trial_state, ...) propagates the walkers in imaginary time (dt, steps) using that quantum trial to impose the phaseless constraint.

import qalgora

# 1. quantum step: optimize a trial state with VQE
energy, params = qalgora.vqe(ansatz, hamiltonian, optimizer,
                             parameter_count=n_params)
trial_state = qalgora.get_state(ansatz, params)

# 2. classical step: AFQMC using the quantum trial for the constraint
afqmc_energy = run_afqmc(molecular_hamiltonian, trial=trial_state,
                         walkers=200, dt=0.005, steps=2000)
print("AFQMC energy:", afqmc_energy)

What it is — and isn't — good for

• The bias is reduced, not removed. A quantum trial shrinks the constraint bias, but the phaseless approximation is still an approximation; the result is only systematically improvable, not exact.
• Overlaps cost shots. Estimating ⟨ψT|φ⟩ needs a dedicated overlap-measurement circuit and carries shot noise; because it must be re-evaluated for many walkers across many imaginary-time steps, it can become the dominant quantum-resource cost.
• Trial-state preparation still has to be affordable. The VQE (or other) circuit that produces |ψT⟩ must be shallow enough to run, which limits how rich the trial can be on near-term devices.
• The heavy lifting stays classical. AFQMC itself runs on classical hardware; the quantum computer contributes a better trial and the overlaps, not the Monte Carlo propagation.