Dynamic Parametrization (quenching)

Every variational driver accepts a quenching flag that controls how many parameters the classical optimizer varies at each step.

ADAPTVQE(basis="FAO", quenching=True)    # default: re-optimize everything
ADAPTVQE(basis="FAO", quenching=False)   # freeze the past, tune only the newest

quenching=True (default)

quenching=False

Adaptive drivers

re-optimize all parameters each growth step

optimize only the newest; earlier angles frozen at their previous optimum

Fixed ansatz (VQE)

one joint minimization

sweep parameters one at a time, in index order

Cost per step

\(k\)-dimensional optimization

1-dimensional line search

Energy

variationally lowest

an upper bound to the above

Adaptive drivers

quenching=True is textbook ADAPT-VQE: after appending \(e^{\theta_k A_k}\) with \(\theta_k = 0\), the optimizer is handed the whole parameter vector, warm-started from the previous optimum. That re-optimization is what lets ADAPT-VQE reach FCI, and it is why the flag defaults to True.

quenching=False quenches each angle into place: parameters \(\theta_1 \dots \theta_{k-1}\) are held fixed and only \(\theta_k\) is varied. Each step is then a one-dimensional line search — far cheaper per operator, at the cost of variational freedom.

seen = []
driver = ADAPTVQE(pool="qeb", load_hamiltonian="lih.parquet", quenching=False)
driver.run(callback=lambda info: seen.append(info["parameters"].copy()))

# Every step appends exactly one parameter and leaves the earlier ones untouched.
for earlier, later in zip(seen, seen[1:]):
    assert later.size == earlier.size + 1
    assert (later[:earlier.size] == earlier).all()

The two policies agree on the very first operator — with a single parameter there is nothing to freeze.

Fixed ansätze

VQE has no growth loop, so quenching=False takes the natural analogue: a sequential sweep. Parameter \(k\) is optimized alone, with \(0 \dots k-1\) already at their optimized values and \(k+1 \dots\) at their starting values.

This is the same trade: cheaper individual optimizations, a weaker variational result. The joint minimization is always at least as good.

Which to use

Keep the default quenching=True for production energies — it is the standard algorithm and the one validated against FCI throughout the test suite.

quenching=False is useful when the per-step optimization cost dominates: deep ansätze with many parameters, expensive cost functions (shot-based hardware evaluation, for instance), or when you want to study how much of ADAPT-VQE’s accuracy comes from re-optimization rather than from operator selection.

Where it applies

quenching is implemented once on VariationalDriver and is honoured by every driver that inherits from it:

  • _optimize_grown — the growth loop of ADAPTVQE, VASQE, the deflation excited-state growth, and SubspaceADAPTVQE;

  • _optimize_all — VQE.run, its deflated excited states, and SubspaceVQE.