{"id":"d64b2e13-486f-45d8-bb4d-bad251d25ba5","arxiv_id":"2411.09736","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"NoVa-ADAPT replaces ADAPT-VQE's classical optimization with direct gradient-based parameter updates and reaches comparable measurement cost to ADAPT-VQE on H4 simulations.","lead":"NoVa-ADAPT is a quantum algorithm that builds molecular ground-state circuits by repeatedly adding the operator that lowers the energy fastest, setting gate angles from the same gradient measurements, with no classical optimizer in the loop. On simulated H4 molecules it reaches chemical accuracy at a measurement cost comparable to ADAPT-VQE, at the price of deeper circuits.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cost parity depends on treating a full-pool gradient estimate as one function evaluation; shot-level accounting could reverse the comparison.","rationale":"The reader's weakest assumption is exactly the function-evaluation cost model. I agree that this is the load-bearing point: the paper's own metric is self-consistent, but the jump from 'number of function evaluations' to 'measurement cost' in the abstract is not justified without shot-level accounting, and NoVa's use of gradient magnitudes as both selector and step size makes the gradient shots more demanding than ADAPT's. The absence of error bars and hand-tuned γ are secondary; they weaken generality but are not the central claim's direct dependency. Within the paper's metric the numerical evidence is reproducible in principle (Hamiltonian and pool given), so I would not move the verdict; I would keep CONDITIONAL with the cost-model condition made explicit.","tokens_in":18151,"tokens_out":13148,"duration_ms":148176,"concrete_test":"Simulate H4 with explicit finite-shot noise: for a fixed total shot budget, allocate shots to each energy or pool-gradient estimate using standard quantum estimation (e.g., variance ∝ 1/N_shots), with the update accuracy in NoVa set by the need to keep the rotation error below a target, and plot energy error versus total shots for ADAPT-VQE and NoVa-ADAPT (γ=1 and γ=γ*). If NoVa requires more total shots to reach 1.6 mHartree, the abstract's measurement-cost claim fails. As a secondary check, recompute the γ* curve using Eq. (10) and using γ*=+(d²E/dη²)^{-1}; a difference confirms the published formula is not reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central quantitative claim is 'similar measurement cost' to ADAPT-VQE (Sec. V). The paper counts each energy or gradient estimate as one function evaluation, but Eq. (2) is not one expectation value: for ADAPT-type selection one must estimate i⟨[H,A_i]⟩ for every operator A_i in the pool, so one 'gradient evaluation' is P circuit evaluations. In Fig. 1, NoVa-ADAPT needs roughly an order of magnitude more operators (≈100) than ADAPT-VQE (11), hence ~100 full-pool gradient evaluations vs 11. Even if BFGS energy evaluations dominate ADAPT's total, the NoVa gradient bill scales as P × iterations; for the spin-adapted H4 pool P is O(40), so this is several thousand circuit evaluations that the proxy hides. Furthermore NoVa uses the same gradient magnitude as its update size (Eq. 5), so it needs more shots per gradient to control update bias, whereas ADAPT-VQE only needs enough accuracy to rank operators. Fig. 1(b) parity is therefore not established at the actual shot level. Separately, Eq. (10) for γ* is not the stationary point of Eq. (9) as written (sign is inverted for the usual descent Hessian), so the '2nd-deriv' curves need re-verification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces NoVa-ADAPT, a non-variational adaptive quantum state preparation algorithm. At each iteration, the energy gradient with respect to each operator in an ADAPT-style pool is measured; the operator with the largest gradient magnitude is appended, and the rotation parameter is set to a negative multiplicative constant times the measured gradient, avoiding any classical optimization. The paper compares this algorithm with ADAPT-VQE, feedback-based quantum algorithms, and randomized adaptive approaches on an H4 molecule in the STO-3G basis, using the number of operators in the circuit, the number of function evaluations (as a proxy for measurement cost), and energy error relative to exact diagonalization. The main reported findings are that NoVa-ADAPT reaches chemical accuracy with a similar number of function evaluations as ADAPT-VQE despite using more operators, that it can escape a local minimum where ADAPT-VQE gets trapped for a strongly correlated geometry, and that it shows different noise robustness properties. The authors also explore random initial states and a hybrid ADAPT/NoVa scheme.","tokens_in":18368,"tokens_out":10167,"duration_ms":97234,"significance":"If the central cost-parity claim is correct, NoVa-ADAPT is a useful addition to the toolbox of optimization-free adaptive state preparation methods. The paper is notable for directly comparing against several prior feedback-based approaches and for being transparent about the limitations of its noise model. The numerical evidence is limited to H4 (8 qubits) at two geometries, so the generality of the conclusions remains open. The algorithm's basic premise is clearly stated and the results are reproducible in principle from the provided Hamiltonian in Appendix A.","major_comments":[{"comment":"The claim of 'similar measurement cost to ADAPT-VQE' rests on counting each energy or gradient estimate as one function evaluation. Under this proxy, a full gradient evaluation over the operator pool is treated as a single unit, but Eq. (2) shows that selecting an operator requires estimating i⟨[H,A_i]⟩ for every A_i in the pool, i.e., P separate circuit expectation values for a pool of size P. Because NoVa-ADAPT uses about 100 operators while ADAPT-VQE uses 11 in Fig. 1, the gradient-evaluation bill for NoVa is already ~100P versus ~11P before including ADAPT's BFGS energy evaluations. For a spin-adapted fermionic pool with P~40, this can reverse the apparent parity under shot-level accounting. The authors should either report a shot-level metric (e.g., total number of Pauli expectation-value estimates) or multiply the gradient-evaluation count by the pool size and demonstrate that the conclusion is unchanged.","section":"Sec. V, Fig. 1(b); abstract"},{"comment":"Equation (10) for γ* has the wrong sign. Combining Eq. (5), η_n = -γ g with g = ∂E/∂η_n at η_n=0, and the quadratic expansion in Eq. (9), the stationary point with respect to γ is γ* = (∂²E/∂η_n²)^{-1}, not -(...)^{-1} as printed. As written, the formula is inconsistent with Eq. (9). Since the '2nd-deriv' results in Figs. 1, 2, 3, 4, 5, 6, 7, and 8 use γ from this equation, the authors must verify the implementation, correct the equation, and confirm the affected numerical curves.","section":"Sec. III, Eq. (10)"}],"minor_comments":[{"comment":"The noise model does not inject rotational or gradient errors during the classical optimization steps of ADAPT-VQE, which the text acknowledges underestimates ADAPT-VQE's sensitivity. Please make this caveat more prominent, for example in the figure captions, since the figures otherwise suggest a more definitive robustness comparison than the model supports.","section":"Sec. VI, Figs. 6 and 7"},{"comment":"For the random initial state results, the lower panel x-axis is labeled 'number of function evaluations' but the caption describes 'Energy error'; please clarify whether the plotted quantity is energy or energy error, and consider adding error bars or shaded regions for the 100-sample averages.","section":"Sec. V, Fig. 3"},{"comment":"The description of the ACSE comparison would benefit from a sentence stating how the step size ϵ is optimized with BFGS and whether that optimization cost is included in the function-evaluation counts of Fig. 5.","section":"Sec. IV C, Fig. 5"},{"comment":"The lower-bounding γ in Eq. (8) is derived from Ref. [36]; the factor 1/4 in Eq. (8) leads to a 1/8 in Eq. (7). Please check that the constants are quoted consistently with the source.","section":"Sec. III, Eqs. (7) and (8)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope. The main unresolved point is the measurement-cost proxy; if the authors can show parity under a more realistic cost model, the paper would be a solid contribution. The sign error in Eq. (10) is readily fixable but requires re-verification of the affected numerical results. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper describes NoVa-ADAPT, a non-variational version of ADAPT-VQE that uses the measured energy gradient to set both the operator choice and the gate parameter, eliminating the classical optimizer. The numerics on H4 are honest and the comparison to FQA, randomized adaptive, and quantum ACSE is useful, but the abstract's claim about matching ADAPT-VQE's measurement cost is not yet supported. One 'gradient evaluation' here is actually a full-pool measurement (P circuit evaluations), not a single expectation value, and NoVa-ADAPT uses roughly an order of magnitude more iterations than ADAPT-VQE. At shot level the parity in Fig. 1(b) is not established. There's also a concrete sign error in Eq. (10): for a standard descent model gamma* should be +1/H, not -1/H, so the 2nd-derivative curves need re-verification.\n\nWhat is genuinely new is the combination of ADAPT's gradient-ranked operator selection with a non-variational update. The authors correctly place this between FQA (fixed Hamiltonian generators) and the randomized adaptive algorithm (random operators), and the hybrid is a natural thing to try. The paper also shows an interesting effect in Fig. 2: a constant step size can walk past a local minimum that traps ADAPT-VQE. The full H4 Hamiltonian is in the appendix, so the experiments are reproducible. I also credit the authors for explicitly saying that their noise model underestimates ADAPT-VQE's error—that's a limitation, but an honest one.\n\nThe soft spots are proportionate. The measurement-cost proxy is the load-bearing weakness. The absence of error bars on the stochastic comparisons is minor since they do 100-run averages. The hand-tuned constant gamma is a hyperparameter, but they also test a second-derivative version. The demonstration is one small molecule, so the general claim is provisional.\n\nThis paper is a reasonable contribution to the adaptive-simulation subfield. It deserves peer review, but a referee should ask for shot-level measurement accounting and a corrected Eq. (10) before the central claim can be accepted. I wouldn't cite the cost-parity conclusion as is, but the hybrid algorithm and the local-minimum-escape observation are worth remembering.","headline":"NoVa-ADAPT is a plausible hybrid of ADAPT-VQE and feedback-based methods, but the headline claim of measurement-cost parity rests on a proxy that hides the pool-size factor and needs a shot-level accounting.","tokens_in":18910,"tokens_out":6601,"would_cite":false,"duration_ms":62937,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"NoVa-ADAPT removes the classical optimizer from ADAPT-VQE and still matches its measurement cost to chemical accuracy.","keywords":["non-variational algorithm","ADAPT-VQE","adaptive quantum circuits","ground state preparation","measurement cost","quantum chemistry","noise robustness","feedback-based quantum algorithms"],"falsifier":"Simulate both algorithms on H4 with finite measurement shots, tallying the total shot count including every gradient estimate, and compare the shot budget needed to reach chemical accuracy; if NoVa-ADAPT's total exceeds ADAPT-VQE's, the claimed parity is an artifact of counting each energy or gradient estimate as one function evaluation. A hardware experiment with characterized gate over-rotation errors could also test the predicted robustness difference.","tokens_in":17959,"feed_emoji":"⚛️","tokens_out":11764,"duration_ms":91492,"temperature":0.7,"pith_summary":"Adaptive quantum algorithms for molecular ground states typically pair gradient-based operator selection with a classical optimization loop over circuit parameters. This paper asks whether that loop is necessary and proposes NoVa-ADAPT, which sets each new parameter directly from the measured energy gradient of the operator selected. In simulations of the H4 molecule, NoVa-ADAPT reaches chemical accuracy with a measurement cost similar to ADAPT-VQE, despite building deeper circuits. The paper also finds that the optimization-free update is more tolerant of gate rotation errors than ADAPT-VQE, while being more sensitive to errors in the gradient measurements. If these results hold, removing the variational optimizer is a viable way to cut measurement overhead in adaptive quantum chemistry simulations.","feed_headline":"Optimizer-free ADAPT matches VQE measurement cost","feed_subtitle":"Deeper circuits but measurement costs comparable to ADAPT-VQE on small molecules.","key_machinery":"The machinery is the gradient-scaled exponential update. At each iteration the energy derivative with respect to every pool operator $A_i$ is measured, $\\left.\\partial E/\\partial\\theta\\right|_{\\theta=0} = i\\langle\\psi|[H,A_i]|\\psi\\rangle$; the operator $A_n$ with the largest magnitude is selected and applied as $e^{i\\eta_n A_n}$, with $\\eta_n = -\\gamma\\, \\left.\\partial E_{A_n}/\\partial\\theta\\right|_{\\theta=0}$. The paper uses three prescriptions for $\\gamma$: a constant, the descent-lemma bound $1/(4\\|H\\|^2\\|A_n\\|^2)$, and the second-derivative optimum $\\gamma^* = -(\\partial^2 E/\\partial\\eta^2)^{-1}$. This construction transfers the role of the classical optimizer to a fixed, measurement-determined update while preserving ADAPT's problem-tailored circuit growth.","core_discovery":"The central claim is that a non-variational version of ADAPT-VQE can match the variational original in measurement cost down to chemical accuracy. Starting from the Hartree–Fock state, the algorithm measures the energy gradients of every operator in the pool, appends the operator with the largest gradient magnitude as $e^{i\\eta A}$, and chooses $\\eta = -\\gamma$ times that gradient. With $\\gamma$ taken as a constant, as a descent-lemma lower bound, or as the inverse second derivative, the update lowers the energy at each step without any classical parameter optimization. On the H4 molecule, NoVa-ADAPT uses more operators than ADAPT-VQE but a comparable number of function evaluations to reach chemical accuracy; on a strongly correlated H4 geometry it even escapes the local minimum that traps ADAPT-VQE. The paper presents this as evidence that gradient-based operator selection alone, without a variational subroutine, is sufficient for efficient adaptive ground-state preparation in small molecular systems.","pith_inferences":["The function-evaluation metric likely understates the true sampling cost of gradient estimates; a shot-level comparison on larger molecules is needed to see whether the parity survives outside the small, noiseless setting.","The constant-gamma strategy's ability to sidestep a local minimum hints that curvature information (the second-derivative update) guides the path into the trap; a tunable mix of first- and second-order information could give a more robust update rule.","The same non-variational gradient update could be applied with other operator pools, such as hardware-efficient qubit-ADAPT pools, potentially transferring the measurement-cost result to circuits with different noise characteristics.","If the predicted robustness to rotational errors holds on real devices, the update rule could be used as a building block inside variational or hybrid algorithms to make parameter optimization less sensitive to control noise."],"forward_implications":["NoVa-ADAPT reaches chemical accuracy for the H4 molecule with a number of function evaluations comparable to ADAPT-VQE, so the classical optimization subroutine is not essential for this adaptive algorithm's measurement efficiency.","Because NoVa-ADAPT adds more operators than ADAPT-VQE, its circuits are deeper, so on noisy hardware the savings in optimization must be weighed against increased gate-error exposure.","The optimization-free update is less sensitive to rotational (parameter) errors, since a range of update sizes still reduce the energy, making it a candidate for devices with imprecise gate control.","On strongly correlated systems where ADAPT-VQE settles into a local minimum, a constant-step non-variational update can keep lowering the energy and reach chemical accuracy.","Starting with a few NoVa-ADAPT iterations and then switching to ADAPT-VQE slightly reduces the measurement cost relative to running ADAPT-VQE alone, suggesting a practical hybrid workflow."],"supporting_citations":[{"why":"Defines ADAPT-VQE, the adaptive variational algorithm whose operator-selection step NoVa-ADAPT inherits and whose performance it benchmarks against.","marker":"[17]"},{"why":"Introduces qubit-ADAPT-VQE and the qubit operator pool used in the strongly correlated H4 simulation where ADAPT-VQE is trapped.","marker":"[18]"},{"why":"Establishes feedback-based quantum algorithms (FQALON), an optimization-free approach that NoVa-ADAPT is compared with.","marker":"[23]"},{"why":"Extends feedback-based algorithms to ground state preparation, providing the FQA baseline for the molecular-simulation comparisons.","marker":"[33]"},{"why":"Introduces randomized adaptive quantum state preparation and supplies the descent lemma that justifies the lower-bounding choice of gamma.","marker":"[36]"},{"why":"Shows how recycling the BFGS Hessian across ADAPT-VQE iterations reduces its measurement cost; the paper adopts this scheme in its ADAPT-VQE simulations.","marker":"[45]"},{"why":"Presents the quantum algorithm for solving contracted eigenvalue equations (quantum ACSE), another optimization-free iterative method compared here.","marker":"[49]"}],"fun_headline_variants":["No-variational ADAPT hits chemical accuracy at VQE cost","Optimizer-free ADAPT rivals VQE measurement cost","Gradient-only ADAPT matches VQE cost, escapes traps","ADAPT without optimizer: same cost, deeper circuits","Non-variational ADAPT equals VQE measurement burden"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The measurement-cost comparison assumes that estimating the energy and estimating a gradient each cost one function evaluation; if gradient estimates require many more measurement shots than energy estimates, the claimed cost parity between NoVa-ADAPT and ADAPT-VQE could reverse.","fun_headline_variants_meta":{"raw":{"variants":["No-variational ADAPT hits chemical accuracy at VQE cost","Optimizer-free ADAPT rivals VQE measurement cost","Gradient-only ADAPT matches VQE cost, escapes traps","ADAPT without optimizer: same cost, deeper circuits","Non-variational ADAPT equals VQE measurement burden"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000145,"raw_usage":{"total_tokens":1130,"prompt_tokens":847,"completion_tokens":283,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":201}},"tokens_in":463,"tokens_out":283,"duration_ms":3098,"temperature":1.0,"reasoning_tokens":201,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:21:21.941522+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate both algorithms on H4 with finite measurement shots, tallying the total shot count including every gradient estimate, and compare the shot budget needed to reach chemical accuracy; if NoVa-ADAPT's total exceeds ADAPT-VQE's, the claimed parity is an artifact of counting each energy or gradient estimate as one function evaluation. A hardware experiment with characterized gate over-rotation errors could also test the predicted robustness difference.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces qubit-ADAPT-VQE and the qubit operator pool used in the strongly correlated H4 simulation where ADAPT-VQE is trapped."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes feedback-based quantum algorithms (FQALON), an optimization-free approach that NoVa-ADAPT is compared with."},{"cited_title":"Wiersema and N","cited_arxiv_id":null,"evidence_quote":"Introduces randomized adaptive quantum state preparation and supplies the descent lemma that justifies the lower-bounding choice of gamma."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows how recycling the BFGS Hessian across ADAPT-VQE iterations reduces its measurement cost; the paper adopts this scheme in its ADAPT-VQE simulations."}],"review_version":1}