Adaptive Ansätze (ADAPT-VQE)

Fixed ansätze force a single uncomfortable choice. Make the circuit too shallow and it cannot reach the ground state; make it too deep and you hit barren plateaus and noise. ADAPT-VQE, introduced by Grimsley et al., sidesteps the dilemma: instead of fixing the circuit up front, it grows the ansatz one operator at a time, adding only the gate that most reduces the energy. The circuit ends up exactly as expressive as the problem demands, and no more.

The operator pool

ADAPT-VQE starts from a pool of candidate generators $\{A_m\}$ — anti-Hermitian operators whose exponentials $e^{\theta_m A_m}$ are the allowed circuit blocks. In quantum chemistry these are the fermionic single- and double-excitation operators of unitary coupled cluster; in the hardware-friendly qubit-ADAPT variant they are individual Pauli strings. The pool defines the vocabulary from which the algorithm builds its circuit.

The growth loop

Starting from a reference state $|\psi^{(0)}\rangle$ (typically the Hartree–Fock state), each iteration does the following:

Score every candidate by its gradient. For each pool operator $A_m$ , compute the energy gradient that adding it at zero angle would produce. A standard identity gives this as a commutator expectation value on the current state:
$\left.\frac{\partial E}{\partial\theta_m}\right|_{\theta_m=0} = \langle\psi^{(k)}|\,[\,H, A_m\,]\,|\psi^{(k)}\rangle.$
Pick the largest. Select the operator with the biggest gradient magnitude — the steepest available descent direction — and append $e^{\theta_m A_m}$ to the circuit with a fresh parameter.
Re-optimize all parameters. Run a full VQE over every angle accumulated so far, not just the new one.
Check convergence. If the largest remaining gradient is below a threshold $\varepsilon$ , stop. Otherwise return to step 1.

Why this beats a fixed circuit

Compactness. ADAPT typically reaches chemical accuracy with far fewer parameters and a shallower circuit than a fixed unitary-coupled-cluster ansatz, because it never includes operators that do not help.
Problem-tailored expressibility. The circuit's reachable set is grown toward the actual ground state, so expressibility is spent where it matters.
Plateau resistance. Because each operator is added in a high-gradient direction and the circuit grows gradually rather than starting deep and random, ADAPT largely avoids the initialization-on-a-plateau problem.

Costs and caveats

Adaptivity is not free. Every iteration must measure the gradient of every pool operator, which can mean many measurements per step; the pool itself can be large for big molecules. There is also a risk of getting stuck if the threshold $\varepsilon$ is set too loosely. Variants such as qubit-ADAPT shrink the pool to individual Pauli strings to cut measurement and gate overhead, at the cost of needing more (cheaper) operators in the final circuit.

Key takeaways

ADAPT-VQE builds the ansatz dynamically, appending the highest-gradient operator from a pool each iteration.
Candidate operators are scored by the commutator gradient $\langle[H, A_m]\rangle$ on the current state — no need to add them first.
The result is a compact, problem-tailored, plateau-resistant circuit, at the price of extra gradient measurements per step.