Adaptive Derivative-Assembled Pseudo-Trotter (ADAPT)

• Adaptive Derivative-Assembled Pseudo-Trotter (ADAPT) is an adaptive variational quantum eigensolver that iteratively builds a compact ansatz using a derivative-based operator selection approach.
• It leverages carefully designed operator pools, symmetry constraints, and pruning methods to reduce measurement overhead and mitigate redundant operator challenges.
• ADAPT supports both ground and excited state simulations by integrating subspace diagonalization and density-matrix techniques for flexible, scalable quantum many-body state construction.

Adaptive Derivative-Assembled Pseudo-Trotter, usually instantiated as ADAPT-VQE, denotes a family of adaptive variational quantum eigensolver schemes in which the ansatz is not fixed a priori but is grown iteratively from an operator pool. In its standard form, the trial state is an ordered product of elementary unitaries,

$$|\Psi\rangle=\prod_{i=1}^{N} e^{\theta_i \hat A_i}|\psi_0\rangle,$$

with the next generator selected from derivative information evaluated on the current state, rather than from a predetermined cluster expansion (Vaquero-Sabater et al., 7 Apr 2025, Liu et al., 2020). The “Derivative-Assembled” component refers to the use of energy derivatives or residual gradients to choose operators; the “Pseudo-Trotter” component refers to the product form $\prod_i e^{\theta_i \hat A_i}$, which is structurally similar to a first-order Trotterization but is assembled adaptively instead of being obtained from a fixed exponential of a full cluster operator (Vaquero-Sabater et al., 7 Apr 2025, Liu et al., 2020).

1. Formal structure and defining principles

In the VQE setting, the objective is to minimize

$$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$

over a parameterized family of states. ADAPT replaces a static ansatz such as UCCSD by an iterative construction in which a reference state $|\psi_0\rangle$, often the Hartree–Fock determinant but in some formulations also a multireference state, is progressively dressed by exponentials of anti-Hermitian generators drawn from a pool (Vaquero-Sabater et al., 7 Apr 2025, Liu et al., 2020). In the original ADAPT formulation summarized in the reduced-density-matrix reformulation, the ordered product may be written as

$$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$

so that only one new operator is appended at each iteration (Liu et al., 2020).

The selection rule is derivative-based. For a candidate operator $\tau_u$, the residual gradient is

$$R_u=\left.\frac{\partial E}{\partial \theta_k}\right|_{\theta_k=0,\tau_k=\tau_u} =\langle \Psi(k-1)|[\hat H,\tau_u]|\Psi(k-1)\rangle.$$

The next operator is chosen by maximizing $|R_u|$ over the pool, and the ansatz is then re-optimized (Liu et al., 2020). In the chemistry-oriented formulation of standard ADAPT-VQE, the same idea appears as the gradient magnitude

$$g_\mu=\left|\frac{\partial E}{\partial \theta_\mu}\right|_{\theta_\mu=0},$$

evaluated on the current ansatz and used to select the operator with the largest magnitude (Vaquero-Sabater et al., 7 Apr 2025).

Convergence is likewise derivative-driven. One common criterion is that the largest gradient magnitude over the pool drops below a tolerance $\epsilon$, while the exact ACSE-inspired formulation uses the residual norm

$\prod_i e^{\theta_i \hat A_i}$0

and stops when $\prod_i e^{\theta_i \hat A_i}$1 (Vaquero-Sabater et al., 7 Apr 2025, Liu et al., 2020). In practice, classical reoptimization is typically warm-started from the previous optimum; BFGS is used in several implementations, and “amplitude recycling” is explicitly identified as part of what makes ADAPT robust against local minima and barren plateaus (Vaquero-Sabater et al., 7 Apr 2025).

2. Operator pools, symmetry structure, and domain-specific realizations

A defining ingredient of ADAPT is the operator pool. In molecular electronic structure, one widely used choice is a fermionic pool of single and double excitations. One formulation uses general one- and two-body anti-Hermitian operators

$\prod_i e^{\theta_i \hat A_i}$2

with

$\prod_i e^{\theta_i \hat A_i}$3

while another chemistry-focused implementation restricts the pool to occupied-to-virtual spin-singlet-adapted single and double excitations mapped to qubits through Jordan–Wigner (Liu et al., 2020, Vaquero-Sabater et al., 7 Apr 2025). These choices preserve particle number and, when spin-adapted, avoid spurious symmetry-breaking states (Vaquero-Sabater et al., 7 Apr 2025).

The same adaptive principle has been specialized to qubit-operator pools. In the excited-state convergence-path work, a qubit-excitation pool is built from anti-Hermitian Pauli combinations such as

$\prod_i e^{\theta_i \hat A_i}$4

and corresponding four-qubit operators for doubles, while a more flexible qubit pool uses Pauli strings directly (Zhang et al., 27 Jun 2025). In hardware-oriented molecular calculations, Qubit-ADAPT is used because each unitary is a single Pauli exponential and therefore yields shallower circuits than fermionic exponentials, even though the qubit pool can be larger (Carreras et al., 4 Jun 2025).

Pool construction is also heavily symmetry-dependent. In the light-nuclei shell-model study, ADAPT uses the same anti-Hermitian two-body excitation pool as UCC, but restricted to the $\prod_i e^{\theta_i \hat A_i}$5 subspace and reduced by angular-momentum selection rules and redundancy relations (Carrasco-Codina et al., 18 Jul 2025). In point-group-respecting molecular formulations, HiUCCSD builds a reduced excitation pool directly from the nonzero Hamiltonian integrals. For Abelian point groups this is theoretically equivalent, in terms of which excitations survive, to SymUCCSD; across the studied molecules it reduces the excitation operator pool size for ADAPT-VQE by $\prod_i e^{\theta_i \hat A_i}$6 relative to UCCSD (He et al., 24 Dec 2025). This indicates that ADAPT is not a single ansatz but a framework whose behavior depends strongly on the algebraic and symmetry structure of the chosen generator set.

3. Measurement bottlenecks and efficiency-oriented reformulations

The principal practical obstacle in standard ADAPT-VQE is measurement overhead. In the residual-gradient formulation, the Hamiltonian contains $\prod_i e^{\theta_i \hat A_i}$7 terms, the operator pool contains $\prod_i e^{\theta_i \hat A_i}$8 elements, and naively evaluating all commutators $\prod_i e^{\theta_i \hat A_i}$9 across the full pool leads to $$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$0 measurement scaling (Liu et al., 2020). This bottleneck motivated several reformulations that preserve the adaptive structure while reducing the cost of screening operators.

A central line of work rewrites residual gradients in terms of reduced density matrices. In ADAPT-RDM, one- and two-body residuals are expressed exactly in terms of the $$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$1-, $$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$2-, and $$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$3-RDMs, which reduces the measurement scaling from $$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$4 to $$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$5. In ADAPT-V, the $$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$6-RDM is reconstructed from the $$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$7- and $$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$8-RDMs using Valdemoro’s first-order reconstruction, so that only $$E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle$$9- and $|\psi_0\rangle$0-RDMs are measured and the gradient-measurement scaling becomes $|\psi_0\rangle$1. ADAPT-Vx then combines approximate screening with exact refinement over a reduced auxiliary pool, yielding $|\psi_0\rangle$2 residual-gradient cost while recovering nearly ADAPT-like compactness and accuracy (Liu et al., 2020).

Another line of work replaces commutator-based importance with metrics derived from determinant populations. FAST-VQE samples Slater-determinant populations in the computational basis and uses two heuristic operator-importance metrics, one based on approximate gradients and one based on Selected Configuration Interaction with perturbation theory. In state vector and finite-shot simulations, FAST-VQE using the heuristic metric based on approximate gradients converges at the same rate or faster than ADAPT-VQE and requires dramatically fewer shots (Majland et al., 2023). This suggests that the defining ADAPT principle is adaptive operator ranking rather than any single mandatory estimator of that ranking.

Efficiency improvements also arise from physically motivated state preparation. Using unrestricted-HF natural orbitals and active-space projection, one can optimize state preparation without added computational burden and guide ansatz expansion to yield more concise wavefunctions with expedited convergence toward exact solutions, producing shallower circuits and reduced measurement requirements in H$|\psi_0\rangle$3 models and water (Vaquero-Sabater et al., 2024). A plausible implication is that ADAPT’s practical performance is governed as much by orbital choice and screening infrastructure as by the abstract adaptive loop itself.

4. Redundancy, pruning, and compact ansatz construction

Although ADAPT is often described as producing compact ansätze, detailed analyses show that standard gradient-based growth can accumulate redundant operators. In stretched linear H$|\psi_0\rangle$4, three distinct mechanisms were identified: poor operator selection, operator reordering, and fading operators. Poor operator selection refers to operators that are chosen because of a large gradient but acquire a very small optimized amplitude and remain negligible; operator reordering refers to repeated selection of the same excitation at later positions, causing an earlier copy to collapse toward zero amplitude; fading operators are initially important operators whose amplitudes decay close to zero as the ansatz grows (Vaquero-Sabater et al., 7 Apr 2025).

These effects motivate pruning, but naive amplitude thresholding is not sufficient because some operators are cooperatively important: an operator may appear with a small amplitude when it is added and later become one of the most important operators after subsequent insertions (Vaquero-Sabater et al., 7 Apr 2025). Pruned-ADAPT-VQE addresses this by introducing a decision factor

$|\psi_0\rangle$5

where $|\psi_0\rangle$6 is the normalized position in the ansatz. A dynamic threshold

$|\psi_0\rangle$7

is then used to decide whether the most suspicious operator should be deleted after optimization, without any additional quantum circuit evaluations and without a re-optimization after deletion (Vaquero-Sabater et al., 7 Apr 2025). The refinement process is described as cost-free, and on the systems examined it reduces ansatz size and accelerates convergence, especially in flat energy landscapes (Vaquero-Sabater et al., 7 Apr 2025).

A second compactness strategy is to replace the gradient criterion itself. Param-ADAPT-VQE selects excitation operators according to the magnitude of the optimized local parameter $|\psi_0\rangle$8 rather than the traditional gradient-based metric. It supplements this with a sub-Hamiltonian technique, in which each candidate excitation is screened against only those Hamiltonian terms sharing orbital indices with that excitation, and a hot-start global VQE step in which the newly added parameter is initialized to its locally optimized value rather than to zero (He et al., 4 Feb 2026). Numerical experiments report improvements in computational accuracy, ansatz size, and measurement costs, while preserving the basic ADAPT framework (He et al., 4 Feb 2026).

Symmetry-informed pool reduction serves the same compactness objective at a different level. HiUCCSD constructs symmetry-respecting excitation sets directly from nonzero electronic integrals and, in the studied molecules, reduces ADAPT excitation pools by $|\psi_0\rangle$9 relative to UCCSD while avoiding the non-Abelian subgroup failures observed for SymUCCSD (He et al., 24 Dec 2025). Taken together, pruning, parameter-based selection, and Hamiltonian-informed pool design show that “compact ADAPT” is itself a research area rather than an automatic property of the original algorithm.

5. Excited states, density matrices, and subspace generalizations

ADAPT has expanded far beyond ground-state VQE. One route is variational quantum deflation. In the RDM-based formulation, ADAPT-V and ADAPT-Vx are generalized to excited states by replacing $$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$0 with a deflated Hamiltonian

$$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$1

or equivalently by optimizing the VQD functional with overlap penalties against previously obtained states (Liu et al., 2020). This preserves the ADAPT ansatz-growth logic while targeting orthogonal excited states.

A second route uses the convergence path itself as a basis for quantum subspace diagonalization. If $$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$2 denotes the ADAPT state after $$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$3 iterations, the span of $$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$4 is used as a compact subspace, and low-lying eigenstates are obtained from the generalized eigenvalue problem

$$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$5

This ADAPT-QSD construction yields approximate ground and excited states with only a small overhead beyond the ground-state ADAPT run and can also lower the ground-state estimate relative to the best individual ADAPT state (Zhang et al., 27 Jun 2025).

A third route is density-matrix-based. TEPID-ADAPT-VQE variationally diagonalizes a truncated low-temperature Gibbs state,

$$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$6

by minimizing the free energy

$$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$7

with a shared adaptive ansatz $$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$8 acting on several computational-basis reference states (Saroni et al., 28 Jun 2026). On the molecules studied, both TEPID-ADAPT and a modified MORE-ADAPT reproduce excited-state spectra and potential energy curves within chemical accuracy, but TEPID-ADAPT uses a single physically motivated hyperparameter, the temperature, whereas MORE-ADAPT uses multiple hyperparameters whose optimal values depend sensitively on the target problem (Saroni et al., 28 Jun 2026).

ADAPT ideas have also been reinterpreted as adaptive subspace construction rather than direct nonlinear optimization. In ADAPT-GCIM, a non-orthogonal basis of generator-coordinate states built from UCC excitation generators is used to solve a generalized eigenvalue problem $$|\Psi(k)\rangle = e^{\theta_k \tau_k}|\Psi(k-1)\rangle,\qquad |\Psi(0)\rangle = |\Psi_0\rangle,$$9, thereby replacing much of the nonlinear constrained optimization by linear algebra in an adaptively constructed subspace (Zheng et al., 2023). Conversely, in a Hamiltonian-independent setting, an ADAPT-style pseudo-Trotter ansatz combined with simulated annealing is used to decide whether a target $\tau_u$0-body matrix is $\tau_u$1-representable by minimizing the Hilbert–Schmidt distance to a reachable $\tau_u$2-RDM (Massaccesi et al., 21 Mar 2025). These extensions show that the ADAPT idea has become a general template for adaptive operator selection in both pure-state and density-matrix variational algorithms.

6. Applications, limitations, and current research directions

ADAPT has been applied across molecular chemistry, nuclear structure, and model Hamiltonians. In quantum chemistry, it is repeatedly benchmarked on stretched and strongly correlated geometries, where adaptive growth is most beneficial. In nuclear shell-model calculations from $\tau_u$3He to $\tau_u$4B, ADAPT uses fewer total operations than UCC for nuclei with small valence spaces and appears more efficient near magic numbers, whereas UCC becomes more resource-efficient toward the mid shell (Carrasco-Codina et al., 18 Jul 2025). In the Extended Lipkin Model, ADAPT-VQE reproduces ground-state energies across a phase diagram containing both first- and second-order quantum phase transitions, with slower convergence near criticality and faster convergence deep inside a phase (Baid et al., 2024).

The practical limitations are equally well documented. A hardware study on benzene shows that even after active-space reduction, Hamiltonian compression, symmetry-restricted pools, qubit-operator ansätze, circuit-layout optimization, and error mitigation, current IBM hardware does not yield molecular energies accurate enough for reliable quantum-chemical conclusions; noise overwhelms the small energy differences ADAPT is designed to exploit (Carreras et al., 4 Jun 2025). Under realistic thermal-relaxation simulations, achieving energy errors below $\tau_u$5 mHa for ansätze up to seven operators required coherence times about two orders of magnitude longer than current typical values, and full noisy Qubit-ADAPT optimization became chemically meaningful only when coherence improvements reached roughly $\tau_u$6–$\tau_u$7 in the same scaling model (Carreras et al., 4 Jun 2025). This establishes measurement noise, gate fidelity, and coherence as central external constraints on ADAPT’s usefulness.

Mapping and symmetry can also qualitatively change ADAPT behavior. Under Bravyi–Kitaev mapping, fixed-ansatz UCCSD can exhibit zero-gradient initialization traps caused by global phase cancellations in stretched or highly polarized geometries, while ADAPT-VQE remains effective because the commutator gradients $\tau_u$8 are nonzero and isolate the dominant symmetry-breaking operators (Dipojono, 4 Jun 2026). In the reported LiH, HF, and H$\tau_u$9O examples, an accelerated ADAPT implementation reached the exact active-space FCI energy in the first macro-cycle, whereas the fixed BK-UCCSD optimization stayed at zero energy shift (Dipojono, 4 Jun 2026). This suggests that ADAPT’s adaptive operator-level screening can be structurally more robust than fixed cluster expansions when fermion-to-qubit mappings induce nonlocal cancellation patterns.

The broader trajectory of the field points toward self-regularizing, measurement-aware, and symmetry-informed ADAPT schemes. Pool reduction through Hamiltonian structure, parameter-based ranking, pruning, RDM-based screening, convergence-path subspace methods, and density-matrix formulations all retain the core ADAPT principle—iterative ansatz growth from derivative- or response-based operator diagnostics—while targeting the dominant obstacles of redundant operators, excessive measurements, and hardware noise (He et al., 24 Dec 2025, He et al., 4 Feb 2026, Vaquero-Sabater et al., 7 Apr 2025). A plausible implication is that ADAPT is best understood not as a single algorithm but as a modular architecture for adaptive quantum many-body state construction.