K-ADAPT-VQE: Optimizing Molecular Ground State Searches by Chunking Operators

Guido Masella; Oumaya Ladhari; Tatiana A. Bespalova

REVIEW 3 major objections 4 minor 12 references

K-ADAPT-VQE: Optimizing Molecular Ground State Searches by Chunking Operators

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read K-ADAPT-VQE: adding five operators per step cuts quantum function calls by a factor of 4.3 in BeH2 while keeping chemical accuracy.

desk verdict The chunking idea is plausible, but the headline 4.3x savings number does not survive contact with the paper's own call-count arithmetic. read the letter →

arxiv 2506.09658 v1 pith:JXP2W6XB submitted 2025-06-11 quant-ph

Tatiana A. Bespalova , Oumaya Ladhari , Guido Masella This is my paper · ORCID

classification quant-ph

keywords ADAPT-VQEvariationalquantumeigensolveroperatorchunkingmoleculargroundstatechemistryNISQalgorithmsansatzconstructionresourcereduction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

K-ADAPT-VQE modifies ADAPT-VQE by adding the K strongest candidate operators to the trial wavefunction at once instead of one at a time, with K set to 5 in the demonstrations. The paper claims this simple scheduling change substantially cuts the number of VQE optimization iterations and quantum function calls needed to reach chemical accuracy in small molecular ground-state calculations. For linear BeH2 it reports a roughly 4.3-fold reduction in quantum function calls versus the standard one-operator procedure, with comparable or better accuracy across bond lengths, and similar qualitative gains for LiH and N2. A sympathetic reader takes the central insight to be that operator selection by gradient magnitude can be chunked without losing the physical content of the ansatz.

What carries the argument

The load-bearing mechanism is the selection and insertion of a chunk: a group of K=5 operators sorted by absolute gradient magnitude, added together with zero initial parameters, then optimized as part of the full accumulated ansatz. The supporting operator pool is pruned to spin-conserving double excitations that move electrons from Hartree-Fock-occupied orbitals into virtual orbitals, removing single excitations and non-spin-conserving terms before selection begins. The cost accounting that carries the central claim assumes three quantum function evaluations per VQE iteration and counts total evaluations along the convergence path, so the quantum-call reduction is a direct consequence of needing fewer optimization iterations to reach chemical accuracy.

What would settle it

Directly instrument a run of both algorithms on the same molecule, counting every call to the quantum energy function rather than assuming three per iteration, and give both procedures identical total VQE iteration budgets. If 5-ADAPT does not reach the same chemical-accuracy threshold with fewer quantum calls than 1-ADAPT under those matched conditions, the central resource-reduction claim is not established.

Watch

Extended reading notes

Core claim

The paper's demonstrated discovery is that gradient-ranked chunking is an effective resource-reduction heuristic for ADAPT-VQE. At each iteration the algorithm computes the absolute values of the commutator expectation $\langle \Psi | [\hat P_i, H] | \Psi \rangle$ for all pool operators, picks the top K, appends them in decreasing order of gradient with zero initial angles, and re-optimizes all parameters. In the BeH2 comparison, 5-ADAPT reaches an error below $10^{-3}$ Hartree using roughly 3300 quantum function evaluations, while 1-ADAPT needs about 14100, a cost reduction of about 4.3 times; the resulting energies stay close to full configuration interaction. The authors further claim that the ansatz does not suffer when operators within a chunk act on overlapping qubits, which is the feature that separates this strategy from approaches that require disjoint operator supports.

Load-bearing premise

The cost comparison depends on treating the chosen per-step optimization budgets (200 VQE iterations for each 5-ADAPT chunk versus 40 for each 1-ADAPT operator) and the fixed count of three quantum function evaluations per iteration as a fair and accurate measure; if those budgets or the counting rule were changed, the reported 4.3-fold reduction would change.

Editorial extensions

If this is right

On noisy intermediate-scale hardware, where each quantum function call is expensive, a 4.3-fold reduction in calls makes adaptive ground-state calculations more feasible for small molecules.
The modification is a drop-in change to the ADAPT-VQE selection step, so it composes with other improvements such as pool pruning and better classical optimizers.
K becomes a tunable hyperparameter: larger chunks mean fewer ADAPT iterations but more parameters entering at once, so the optimal K likely depends on the molecule and the optimizer.
The observation that overlapping operators within a chunk do not degrade accuracy contradicts a design assumption of disjoint-support methods and suggests chunking can be paired with denser circuits.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The 4.3x figure is a point estimate tied to the specific comparison budgets; a fairer efficiency measure would hold total optimization effort fixed or measure wall-clock time on a simulator, and the ratio would likely change.
Chunking by gradient magnitude acts like a coarse-grained update on the ansatz landscape, so the idea may transfer to other adaptive variational algorithms beyond ADAPT-VQE, not just molecular ground states.
A direct test would scan K across 1, 2, 3, ... for one molecule and plot quantum calls versus accuracy; the paper fixes K=5 and does not report where the optimum lies.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The thing to know: this paper proposes a simple, plausible tweak to ADAPT-VQE—add the top K operators by gradient magnitude each iteration, without requiring disjoint qubit supports. That is a legitimate variation on TETRIS-ADAPT-VQE, and the simulations across BeH2, LiH, and N2 show 5-ADAPT reaching chemical accuracy at more bond lengths than 1-ADAPT within the same nominal iteration cap. The idea is worth taking seriously.

But the central quantitative claim, the 4.3x reduction in quantum function calls, is not established. The paper states both procedures use 1000 VQE iterations total, yet the call-count formulas in Section 3 use 800 iterations for 5-ADAPT and 3200 for 1-ADAPT. That is a direct contradiction. On top of that, the comparison is asymmetric: 5-ADAPT gets 200 COBYLA iterations per chunk, 1-ADAPT gets 40 per operator. So the per-operator optimization budget differs by a factor of five. The 'three evaluations per COBYLA iteration' is asserted without any basis. If the budgets are matched and the counting made consistent, the 4.3x factor could shrink or even reverse.

The energy convergence plots are more convincing than the cost accounting. They show 5-ADAPT hitting the chemical-accuracy threshold more often across bond lengths. But there are no error bars, no sensitivity analysis for K or the per-step budget, and no code or data to check the specific runs. For an empirical claim, that is thin.

What is genuinely new here is the selection rule—dropping the disjoint-support constraint and simply taking the top K by gradient. That is a small but real step beyond TETRIS-ADAPT-VQE, and the paper cites the relevant prior work. So the contribution is modest but not nil.

Who is this for? Researchers actively working on ADAPT-VQE and near-term quantum chemistry who care about cheap heuristics for reducing circuit construction overhead. It is not a landmark result, but it is a testable idea.

Recommendation: this deserves peer review, because the heuristic is simple, the failure mode is fixable, and the field would benefit from a clean comparison. But a serious referee should demand a matched-budget comparison, consistent arithmetic, error bars, and ideally code/data. Without those, the reported speedup is not credible.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes K-ADAPT-VQE, a modification of ADAPT-VQE in which the K operators with the largest gradient magnitudes are added to the ansatz per iteration, rather than a single operator. It reports classical simulations for BeH2, LiH, and N2 in the STO-3G basis, comparing K=5 (5-ADAPT) with K=1 (1-ADAPT), and claims a roughly 4.3-fold reduction in quantum function calls for BeH2 while attaining chemical accuracy.

Significance. If the central claim is correct, chunking operators by gradient rank is a simple, computationally cheap heuristic with clear practical value for ADAPT-VQE on noisy hardware. The paper has useful strengths: it uses external benchmarks (HF and FCI from PySCF), gives a clear algorithmic pseudocode, and makes a concrete, falsifiable resource-reduction prediction. However, the quantitative evidence for the headline claim is currently compromised by internal inconsistencies in the cost accounting, and no sensitivity analysis is provided for the key protocol choices. The significance therefore rests on a load-bearing calculation that is not yet reproducible as written.

major comments (3)

[Section 3, BeH2 cost comparison] The call-count arithmetic is internally inconsistent. The text states: 'For both the 5- and 1-ADAPT procedures, the total number of VQE iterations is 1000 (200 iterations for 5-ADAPT chunks, 40 iterations for each 1-ADAPT operator).' Immediately below, however, the 5-ADAPT formula 3×800 + 180×25/5 assumes 800 optimizer iterations, and the 1-ADAPT formula 3×3200 + 180×25 assumes 3200; neither matches the stated 1000. The terms 180×25/5 and 180×25 are never defined, and the factor 'three evaluations per COBYLA iteration' is asserted without justification or measurement. Since the 4.3x reduction is computed from these numbers, the headline claim is not reproducible as written. The inconsistency is also visible in Fig. 1(b), whose x-axis extends to 3200 iterations although the text says both runs used 1000.
[Section 3 and Fig. 1 caption] The efficiency comparison uses a single hand-selected pair of per-step budgets (200 VQE iterations per 5-ADAPT step, 40 per 1-ADAPT step) and no sensitivity analysis. Even if the total iteration counts are meant to be equal, the optimization schedule differs between the two protocols, and the reported cost ratio depends on these choices. The paper states that K=5 was chosen empirically; without tests for other K values and for matched per-operator or per-step budgets, the general claim that K-ADAPT-VQE 'substantially reduces' calls is not established. The authors should report the cost ratio for several (K, budget) combinations or provide an argument that the advantage is insensitive to these choices.
[Section 4, Conclusion] The conclusion says 5-ADAPT achieves 'comparable accuracy' to 1-ADAPT, but Fig. 1(a) and the accompanying text report that 1-ADAPT errors are roughly an order of magnitude larger. If 1-ADAPT does not reach chemical accuracy at 1000 iterations, then the phrase 'to achieve chemical accuracy' in the call-count comparison is ambiguous or false; if it does reach chemical accuracy only after 3200 iterations, the statement that both procedures used 1000 iterations is wrong. The paper must clarify what the reported call counts are actually measuring: reaching chemical accuracy, reaching a fixed ansatz size, or some other stopping criterion.

minor comments (4)

[Fig. 1 caption] The caption labels the N2 panel as '(b)', but (b) is the BeH2 convergence plot; the N2 panel should be (d). The same typo appears in the text where 'Fig 1(c) shows results for N2.'
[Section 2.4, Algorithm step 4e] The convergence threshold epsilon is never given a numerical value; the Results section should state the threshold used, since it directly affects the reported iteration counts.
[Section 3] The paper should define what counts as a 'quantum function call' (e.g., one circuit execution, one expectation-value estimation, or one energy evaluation including shot noise) and state explicitly whether the MIMIQ simulations are noiseless. This is central to the resource metric.
[Section 2.4, Algorithm step 4d] The maximum number of VQE iterations for 1-ADAPT is given only in the Fig. 1 caption, not in the algorithm description; it should be stated in the main text alongside the 200-iteration choice for 5-ADAPT.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: K-ADAPT-VQE's resource-reduction claim is an empirical benchmark result, not a quantity derived from its inputs.

full rationale

The paper's central comparison (K-ADAPT vs 1-ADAPT) is measured on external benchmarks: HF and FCI energies come from PySCF, and the STO-3G Hamiltonians are standard. The number of VQE iterations and quantum function calls are reported outcomes of running the algorithms, not fitted parameters. There is no equation in which the claimed reduction is defined in terms of itself, and no uniqueness theorem is imported from the authors' prior work. The only self-citation with potential relevance is the MIMIQ emulator [4], but it is used as a simulation tool, not as a load-bearing proof step; the algorithm's logic does not reduce to 'MIMIQ says so.' The choice of K=5 and the iteration caps (200 vs 40) are arbitrary protocol choices, and the call-count arithmetic in Section 3 is internally inconsistent (1000 stated vs 800/3200 used), but these are correctness/fairness issues, not circularity: changing the budgets would change the reported ratio without making the derivation equivalent to its input. Therefore no circular step can be exhibited under the required standard.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced; the method is a selection heuristic. The ledger is dominated by heuristic choices: K, optimizer budgets, the three-evaluations-per-iteration assumption, and the top-K ranking assumption.

free parameters (3)

K (number of operators added per chunk) = 5
Chosen empirically ('In the following, we use K = 5', Section 2.3) without a reported sweep or sensitivity analysis; all efficiency claims are for this value.
VQE optimizer budget per ADAPT step = 200 COBYLA iterations for 5-ADAPT, 40 for 1-ADAPT
These budgets are baked into the cost comparison in Section 3 and the Fig. 1 caption; the 4.3x ratio is not shown to be stable across budgets.
Quantum function evaluations per COBYLA iteration = 3
Assumed in Section 3 ('Assuming three evaluations per COBYLA iteration') to convert iterations into quantum function calls; not measured, and COBYLA evaluation counts can vary.

assumptions (4)

domain assumption The gradient magnitude of a single operator, computed with all other parameters fixed, remains a valid ranking for choosing K operators to add and optimize jointly.
Invoked in Section 2.3 and step 4b of the algorithm; if joint optimization changes the ordering, the heuristic lacks support.
domain assumption The operator-pool restrictions (excluding singles, non-spin-conserving jumps, and jumps into occupied or virtual orbitals) do not compromise the ability to reach chemical accuracy.
Pool construction in Section 2.2 removes many operators; no test shows these removals are harmless for all systems.
domain assumption COBYLA with the assigned iteration budgets and a 10^-3 Hartree tolerance is a representative optimizer for the efficiency comparison.
Used in all runs in Section 3; different optimizers or tolerances could change the iteration counts.
domain assumption Noiseless MIMIQ simulation of the quantum function is an adequate proxy for the NISQ resource counts being compared.
All results come from a simulator [4]; no shot noise or hardware noise is included, although the paper frames the method for NISQ devices.

how reviews work

0 comments

Cite this review

Pith. "Pith review of K-ADAPT-VQE: Optimizing Molecular Ground State Searches by Chunking Operators." pith.science (2026). https://pith.science/paper/JXP2W6XB

@misc{pith2026250609658,
  author       = {Pith},
  title        = {Pith review of: K-ADAPT-VQE: Optimizing Molecular Ground State Searches by Chunking Operators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JXP2W6XB}},
  note         = {Machine review of arXiv:2506.09658}
}

read the original abstract

Classical simulation of molecular systems is limited by exponential scaling, a hurdle quantum algorithms like Variational Quantum Eigensolvers (VQEs) aim to overcome. Although ADAPT-VQE enhances VQEs by dynamically building ans\"atze, it can remain computationally intensive. This work presents K-ADAPT-VQE, which improves efficiency by adding operators in chunks of K at each iteration. Our results from simulating small molecular systems show that K-ADAPT-VQE substantially reduces the total number of VQE iterations and quantum function calls required to achieve chemical accuracy in molecular ground state calculations.

Figures

Figures reproduced from arXiv: 2506.09658 by the authors.

**Figure 2.** The resulting structure of the ansatz for BeH [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 2 canonical work pages

[1]

Anastasiou, P.G., Chen, Y., Mayhall, N.J., Barnes, E., Economou, S.E.: Tetris-adapt-vqe: An adaptive algorithm that yields shallower, denser circuit ansätze. Phys. Rev. Res. 6, 013254 (Mar 2024). https://doi.org/10.1103/PhysRevResearch.6.013254

doi:10.1103/physrevresearch.6.013254 2024
[2]

International Journal of The- oretical Physics 21(6–7), 467–488 (1982)

Feynman, R.P.: Simulating physics with computers. International Journal of The- oretical Physics 21(6–7), 467–488 (1982). https://doi.org/10.1007/BF02650179

doi:10.1007/bf02650179 1982
[3]

Nature Communications 10(1) (Jul 2019)

Grimsley, H.R., Economou, S.E., Barnes, E., Mayhall, N.J.: An adaptive varia- tional algorithm for exact molecular simulations on a quantum computer. Nature Communications 10(1) (Jul 2019). https://doi.org/10.1038/s41467-019-10988-2

doi:10.1038/s41467-019-10988-2 2019
[4]

Comparative Benchmarking of Utility-Scale Quantum Emulators

Leonteva, A., Masella, G., Outteryck, M., Orioli, A.P., Whitlock, S.: Comparative benchmarking of utility-scale quantum emulators. arXiv preprint arXiv:2504.14027 (2025). https://doi.org/10.48550/arXiv.2504.14027

work page Pith review arXiv doi:10.48550/arxiv.2504.14027 2025
[5]

Reviews of Modern Physics92(1), 015003 (2020)

McArdle, S., Endo, S., Aspuru-Guzik, A., Benjamin, S.C., Yuan, X.: Quantum computational chemistry. Reviews of Modern Physics92(1), 015003 (2020)

work page 2020
[6]

Quantum Science and Technology 5(3), 034014 (jun 2020)

McClean, J.R., et al.: Openfermion: the electronic structure package for quan- tum computers. Quantum Science and Technology 5(3), 034014 (jun 2020). https://doi.org/10.1088/2058-9565/ab8ebc

doi:10.1088/2058-9565/ab8ebc 2020
[7]

PRX Quantum 5, 030333 (Aug 2024)

Parella-Dilmé, T., Kottmann, K., Zambrano, L., Mortimer, L., Kottmann, J.S., Acín, A.: Reducing entanglement with physically in- spired fermion-to-qubit mappings. PRX Quantum 5, 030333 (Aug 2024). https://doi.org/10.1103/PRXQuantum.5.030333 K-ADAPT-VQE 9

doi:10.1103/prxquantum.5.030333 2024
[8]

Nature Communications 5(1) (Jul 2014)

Peruzzo, A., et al.: A variational eigenvalue solver on a pho- tonic quantum processor. Nature Communications 5(1) (Jul 2014). https://doi.org/10.1038/ncomms5213

doi:10.1038/ncomms5213 2014

Show all 12 references

[9]

Quantum2, 79 (Aug 2018)

Preskill, J.: Quantum computing in the nisq era and beyond. Quantum2, 79 (Aug 2018). https://doi.org/10.22331/q-2018-08-06-79

2018 doi
[10]

npj Quantum Information11(1), 1–19 (2025)

Ramôa, M., Anastasiou, P.G., Santos, L.P., Mayhall, N.J., Barnes, E., Economou, S.E.: Reducing the resources required by adapt-vqe using coupled exchange op- erators and improved subroutines. npj Quantum Information11(1), 1–19 (2025). https://doi.org/10.48550/arXiv.2407.08696

2025 doi arXiv
[11]

Romero, J., Babbush, R., McClean, J.R., Hempel, C., Love, P., Aspuru-Guzik, A.: Strategies for quantum computing molecular energies using the unitary coupled cluster ansatz (2018), https://arxiv.org/abs/1701.02691

2018 arXiv
[12]

WIREs Computational Molecular Science 8(1), e1340 (2018)

Sun, Q., et al.: Pyscf: the python-based simulations of chemistry frame- work. WIREs Computational Molecular Science 8(1), e1340 (2018). https://doi.org/10.1002/wcms.1340

2018 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Anastasiou, P.G., Chen, Y., Mayhall, N.J., Barnes, E., Economou, S.E.: Tetris-adapt-vqe: An adaptive algorithm that yields shallower, denser circuit ansätze. Phys. Rev. Res. 6, 013254 (Mar 2024). https://doi.org/10.1103/PhysRevResearch.6.013254

doi:10.1103/physrevresearch.6.013254 2024

[2] [2]

International Journal of The- oretical Physics 21(6–7), 467–488 (1982)

Feynman, R.P.: Simulating physics with computers. International Journal of The- oretical Physics 21(6–7), 467–488 (1982). https://doi.org/10.1007/BF02650179

doi:10.1007/bf02650179 1982

[3] [3]

Nature Communications 10(1) (Jul 2019)

Grimsley, H.R., Economou, S.E., Barnes, E., Mayhall, N.J.: An adaptive varia- tional algorithm for exact molecular simulations on a quantum computer. Nature Communications 10(1) (Jul 2019). https://doi.org/10.1038/s41467-019-10988-2

doi:10.1038/s41467-019-10988-2 2019

[4] [4]

Comparative Benchmarking of Utility-Scale Quantum Emulators

Leonteva, A., Masella, G., Outteryck, M., Orioli, A.P., Whitlock, S.: Comparative benchmarking of utility-scale quantum emulators. arXiv preprint arXiv:2504.14027 (2025). https://doi.org/10.48550/arXiv.2504.14027

work page Pith review arXiv doi:10.48550/arxiv.2504.14027 2025

[5] [5]

Reviews of Modern Physics92(1), 015003 (2020)

McArdle, S., Endo, S., Aspuru-Guzik, A., Benjamin, S.C., Yuan, X.: Quantum computational chemistry. Reviews of Modern Physics92(1), 015003 (2020)

work page 2020

[6] [6]

Quantum Science and Technology 5(3), 034014 (jun 2020)

McClean, J.R., et al.: Openfermion: the electronic structure package for quan- tum computers. Quantum Science and Technology 5(3), 034014 (jun 2020). https://doi.org/10.1088/2058-9565/ab8ebc

doi:10.1088/2058-9565/ab8ebc 2020

[7] [7]

PRX Quantum 5, 030333 (Aug 2024)

Parella-Dilmé, T., Kottmann, K., Zambrano, L., Mortimer, L., Kottmann, J.S., Acín, A.: Reducing entanglement with physically in- spired fermion-to-qubit mappings. PRX Quantum 5, 030333 (Aug 2024). https://doi.org/10.1103/PRXQuantum.5.030333 K-ADAPT-VQE 9

doi:10.1103/prxquantum.5.030333 2024

[8] [8]

Nature Communications 5(1) (Jul 2014)

Peruzzo, A., et al.: A variational eigenvalue solver on a pho- tonic quantum processor. Nature Communications 5(1) (Jul 2014). https://doi.org/10.1038/ncomms5213

doi:10.1038/ncomms5213 2014