{"id":"3cab250d-0458-48bc-bef5-45bbe17f2df2","arxiv_id":"2505.03998","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"KD-VQE allocates measurement shots among multiple trial wavefunctions with Boltzmann weights and anneals a virtual temperature, and on a two-site Fermi-Hubbard model it converges to the exact ground state.","lead":"This paper introduces KD-VQE, a variant of the variational quantum eigensolver that spreads measurement shots across several trial wavefunctions and slowly cools a virtual temperature to concentrate resources on the lowest-energy candidate. A smart generalist might read it to see a proposed fix for VQE's sensitivity to poor starting guesses, though the evidence is a single four-qubit toy example.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Finite-shot annealing can prune the true ground-state candidate before its energy is resolved; the paper's single hand-picked run does not establish the claimed reliability.","rationale":"The reader's weakest assumption is that the annealing schedule and initial states ensure the best candidate is not pruned before the global minimum is reached. My stress-test sharpens this into a concrete mechanism: finite-shot noise in the energy estimates feeding the Boltzmann weights can cause premature condensation to a suboptimal candidate, and Eq. (3) provides no protection against this. This is not merely a missing comparison with standard VQE; it is an internal gap between the claimed reliability and the algorithm's own acknowledged failure mode. The paper reports no repeated runs, no error bars, no code, and no sensitivity analysis, so the single successful trajectory in Fig. 2 is consistent with luck rather than robust behavior. The proposed test directly probes whether the annealing and pruning dynamics preserve the true ground-state candidate under shot noise, which would settle whether the concern actually lands. I agree with the reader's REJECT verdict: the central comparative claim is unsupported, and a controlled baseline plus a schedule-sensitivity study would be required before the reliability statements could be accepted.","tokens_in":8570,"tokens_out":5026,"duration_ms":55651,"concrete_test":"Run a shot-level Monte Carlo of the Section III experiment, sweeping the annealing schedule (e.g., initial kT in {10, 25, 50}, per-step reduction in {2%, 5%, 10%}) and the pruning threshold, with 100 independent shot-noise seeds per configuration. Record the fraction of runs in which the final candidate is the exact ground state and compare with standard VQE using the same total shot budget and the same six initial states. If any reasonable schedule within the paper's stated 'too slow / too fast' window prunes the true ground-state candidate or yields a lower success rate than standard VQE, the central reliability claim is contradicted; if all schedules preserve the true candidate with high probability, the concern would not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the central claim of improved convergence and increased reliability over standard VQE, KD-VQE must retain the true low-energy candidate under finite measurement noise. The paper's only mathematical support, Eq. (3), is a variational bound for the exact expectation values of the current ensemble; it does not constrain the shot-reallocation dynamics, and it does not prevent Boltzmann weights from being computed from noisy energy estimates. As T decreases, exp(-epsilon_k/T) becomes exponentially sensitive to estimation error, so a candidate with a favorable early noise fluctuation can receive the bulk of the 10^4 shots and survive the 100-shot pruning threshold, while the true ground-state candidate is starved and pruned. The paper explicitly acknowledges 'premature condensation' as a failure mode in Section III, but gives no criterion for choosing the starting temperature, the 5% reduction rate, or the pruning threshold. The sentence following Eq. (3) overstates the result: a weighted average of local minima is still a local minimum, and Eq. (3) alone does not ensure convergence to the global minimum. The single successful run in Fig. 2 therefore demonstrates that the algorithm can work on one hand-picked instance, not that it reliably avoids the admitted failure mode. Because the abstract and conclusions claim reliability and broad exploration, this unquantified selection dynamic is the load-bearing weak point.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes KD-VQE, a variational quantum eigensolver variant that maintains several trial wavefunctions concurrently, assigns measurement shots according to a Boltzmann distribution with a virtual temperature, and anneals the temperature so that resources concentrate on the lowest-energy candidate. The method is illustrated on the half-filled two-site Fermi-Hubbard model with six initial trial wavefunctions and a total of 10^4 shots, reporting convergence of one candidate to the exact ground-state energy -1.56. The paper claims that, compared with standard VQE, KD-VQE explores a broader region of solution space and offers improved convergence and reliability.","tokens_in":8799,"tokens_out":6188,"duration_ms":62367,"significance":"If the claims were supported, KD-VQE would be a simple, drop-in heuristic for reducing initialization sensitivity in VQE, and the connection to knowledge distillation is a nice conceptual analogy. The paper's Eq. (3) is a correct variational bound for the mixed-state expectation value, and the idea of dynamic shot reallocation is intuitive. However, the current evidence is limited to a single noiseless simulation on a four-qubit model with hand-picked parameters, and no comparison to standard VQE is provided. The significance of the paper will depend on whether the authors can supply a fair comparative study and a robustness analysis of the annealing procedure under finite-shot noise.","major_comments":[{"comment":"The central claim of improved convergence and increased reliability relative to standard VQE is not supported by any direct comparison. The numerical section reports only the trajectories of the six trial states and the shot allocation for a single run. A fair baseline would use the same hardware-efficient ansatz, the same total shot budget, and the same number of runs for standard VQE with (i) the best initial state, (ii) random initializations, and (iii) all six initial states, reporting the distribution of final energies. Without such a baseline, the comparative statements in the abstract and conclusions are assertions, not demonstrated results.","section":"Abstract; Section III, Fig. 2"},{"comment":"The sentence 'Eq.(3) ensures that KD-VQE can converge to the ground state through the optimization process' overstates what the inequality proves. Eq. (3) is a valid variational upper bound for the exact expectation value of the instantaneous mixed state, but it does not constrain the shot-reallocation dynamics, the annealing schedule, or the influence of finite measurement noise on the Boltzmann weights. As T decreases, the ensemble condenses onto the candidate with the currently lowest estimated energy, which need not be the global minimum. The role of Eq. (3) should be restated as a per-iteration bound rather than a convergence guarantee.","section":"Section II, after Eq. (3)"},{"comment":"The annealing schedule and pruning threshold are chosen by hand (kT=25, 5% reduction per step, pruning below 100 shots), and no criterion or sensitivity study is provided. The paper explicitly acknowledges 'premature condensation' as a failure mode, but does not quantify its probability. Because the Boltzmann weights are computed from the same noisy energy estimates that the algorithm is meant to improve, a candidate with a favorable early noise fluctuation can capture most of the 10^4 shots and survive pruning while the true ground-state candidate is starved. To support the reliability claim, the authors should report the success probability over many independent noise realizations and show how it depends on the initial temperature, the cooling rate, and the pruning threshold.","section":"Section III, 'The virtual annealing schedule...' paragraph"},{"comment":"The numerical demonstration consists of a single run on one Hamiltonian instance (t=1, U=1) at half filling. This is insufficient to establish the claimed reliability and generalizability. At minimum, the authors should provide statistics over repeated runs with different shot-noise realizations and over a small range of U/t values, and ideally compare against standard VQE under the same conditions.","section":"Section III, Fig. 2"}],"minor_comments":[{"comment":"The word 'Emial' should be 'Email'.","section":"Footnote 1"},{"comment":"The phrase 'As depicted in n Fig.(2d)' contains a typographical error ('n'); it should read 'As depicted in Fig.(2d)'.","section":"Section III, paragraph beginning 'As depicted in'"},{"comment":"The caption does not state which of panels (d)-(g) corresponds to which trial state; please label the panels or state the correspondence explicitly in the caption.","section":"Figure 2 caption"},{"comment":"The penalty coefficient λ and the learning rate η are introduced but their numerical values are never specified; please provide them.","section":"Equation (5) and surrounding text"},{"comment":"The sentence 'εIV are εV are trapped around 0' is ungrammatical, and 'ψII and ψIIIexhibit' is missing a space between 'ψIII' and 'exhibit'.","section":"Section III, paragraph after Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is very short and reads as a preliminary report. The main comparative claim is unsupported by any baseline, and the finite-shot behavior of the annealing procedure is not analyzed. However, the core idea is not obviously invalid, and the missing evidence could in principle be supplied by additional simulations. I recommend major revision with a request for a fair standard-VQE comparison, repeated-run statistics, and a sensitivity analysis of the annealing schedule, and I would ask the authors to provide code/data to verify the single reported simulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things to know about arXiv:2505.03998. It is a modest algorithmic variant: run several VQE trial states, allocate measurement shots according to a Boltzmann weight with a decreasing virtual temperature, and prune states that fall below a shot threshold. This specific combination is new relative to the cited literature, and the variational bound in Eq. (3) is correct. But the paper's central claim—improved convergence and reliability relative to standard VQE—is not established by the evidence, which consists of one hand-picked run on a four-qubit Hubbard dimer with no baseline comparison, no repeated runs, no error bars, and no code.\n\nThe idea is sensible on its face, and the writing is clear. Eq. (3) is a legitimate upper bound for a mixed state at a given step; it does not do what the text says it does. The sentence after Eq. (3) claims it 'ensures' convergence to the ground state. That is an overstatement: a weighted average of local minima is still a local minimum, and the bound alone says nothing about the shot-reallocation dynamics. The stress-test concern is real. As T decreases, exp(-epsilon/T) becomes exponentially sensitive to the noisy energy estimates, so a candidate with a favorable early fluctuation can capture most of the 10^4 shots and survive the 100-shot pruning threshold, while the true ground-state candidate is starved and pruned. The paper acknowledges 'premature condensation' in Section III but offers no criterion for choosing the initial temperature, the 5% reduction rate, or the pruning threshold. So the single successful trajectory is a proof-of-principle that the algorithm can work on a friendly instance, not a demonstration of reliability.\n\nThe citation pattern is fine; the paper covers the relevant VQE and knowledge-distillation literature, and the one self-citation is an ordinary quantum-circuit reference.\n\nWho benefits: a reader collecting ideas for multi-start VQE or resource allocation will find a clearly stated scheme worth thinking about. A reader wanting evidence that this scheme helps on any non-trivial problem will be disappointed.\n\nRecommendation: this deserves referee attention only if the venue is willing to demand major revision. The author should add a standard-VQE baseline with the same ansatz and shot budget, multiple seeds and error bars, a sensitivity analysis of the schedule, and ideally code. Without that, the comparative claim in the abstract should be removed. If the editor expects a quick, rigorous outcome, desk rejection is defensible; if the journal tolerates incremental algorithm papers with the expectation that the author will add benchmarks, send it to review.","headline":"Coherent but unproven: the single four-qubit demo cannot carry the claimed reliability, and the annealing dynamics need analysis rather than assertion.","tokens_in":9332,"tokens_out":3957,"would_cite":false,"duration_ms":40718,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68"],"pacs":["03.67.Ac","03.67.Lx"],"model":"deepseek-v4-flash","headline":"KD-VQE runs several trial wavefunctions at once and uses a virtual annealing temperature to shift measurement shots toward the lowest-energy candidate, improving VQE's convergence reliability on the two-site Fermi-Hubbard model.","keywords":["variational quantum eigensolver","knowledge distillation","virtual annealing","Boltzmann distribution","measurement shot allocation","Fermi-Hubbard model","hybrid quantum-classical algorithm","quantum optimization"],"falsifier":"Run KD-VQE on a Hamiltonian whose known ground state is reached only by a trial wavefunction that starts with the highest initial energy of the set; under the paper's schedule (kT=25, 5% decrease per step), if that candidate is pruned before its energy estimate drops below the others, the central claim of reliable convergence to the global minimum fails.","tokens_in":8300,"feed_emoji":"⚛️","tokens_out":6649,"duration_ms":64161,"temperature":0.7,"pith_summary":"The paper proposes an extension of the variational quantum eigensolver (VQE) called KD-VQE, which runs several trial wavefunctions at once and distributes measurement shots among them according to a Boltzmann distribution at a virtual temperature. As the temperature is lowered step by step, shots are pulled away from high-energy candidates and concentrated on the lowest-energy one, a process the paper calls virtual annealing. The claim is that this filters out suboptimal states, lets the optimization explore a broader region of the solution space, and makes convergence to the ground state more reliable and less sensitive to initialization. The method is demonstrated on the two-site Fermi-Hubbard model, where it converges to the exact ground-state energy after pruning five of six initial trial states.","feed_headline":"Virtual annealing steers VQE to the ground state","feed_subtitle":"Running several trial states and shifting shots to the best one as the temperature cools improves VQE convergence.","key_machinery":"The load-bearing object is the Boltzmann-weighted mixed state $\\rho(T) = \\frac{1}{Z}\\sum_k \\exp(-\\varepsilon_k/T)|\\psi_k(\\theta_k)\\rangle\\langle\\psi_k(\\theta_k)|$ with partition function $Z$, together with a virtual annealing schedule that lowers $T$ (from $kT=25$ by 5% per step). This construction turns measurement-shot allocation into a resource-allocation policy: candidate $k$ receives $N_s \\exp(-\\varepsilon_k/T)/Z$ of the total $N_s$ shots, so the optimization automatically spends more measurements on promising states as the temperature drops. The variational principle $\\mathrm{tr}(\\rho H) \\ge E_{gs}$ guarantees the mixed-state energy is an upper bound on the ground-state energy, so the annealing process can only improve the bound. The demonstration also uses a reduced subspace fixed by particle-number conservation (half filling), a penalty term to enforce the particle number, and a hardware-efficient ansatz whose initial states are eigenstates of the hopping term obtained by Fourier transformation.","core_discovery":"KD-VQE's central claim is that a VQE optimizer need not commit to a single trial wavefunction. Instead, a mixed state of several candidates, weighted by Boltzmann factors $\\exp(-\\varepsilon_k/T)/Z$ at a virtual temperature $T$, serves as the ansatz; each candidate is optimized with standard gradient descent while the shot budget is reallocated in proportion to those weights. As $T$ is annealed from a high value toward zero, the algorithm progressively removes candidates whose estimated energies are high, pruning them once their shot allocation falls below a threshold, and finally condenses onto the surviving state. On the half-filled two-site Fermi-Hubbard model with six initial states, the method prunes the five states that start at or relax to higher energies and converges to the exact ground-state energy of $-1.56$ with all shots assigned to the best candidate $\\psi_{VI}$.","pith_inferences":["A natural extension the paper leaves implicit is an adaptive schedule: choosing the temperature drop from the spread of estimated energies would remove the hand-tuned $kT=25$, 5%-per-step choice and could make the method portable to new problems.","Because the method is essentially simulated annealing over a discrete set of ansatze, it could be connected to population-based VQE strategies; a formal comparison of shot overhead versus single-ansatz VQE would clarify when the broader search is worth the cost.","On noisy hardware, energy estimates fluctuate more, which could cause premature pruning of the true best candidate; a sensitivity analysis of the pruning threshold would be a concrete test of the method's robustness.","The two-site Hubbard demonstration uses only six initial states; testing on a system where the ground state is not representable by any single initial ansatz would probe whether the annealing procedure can still steer the ensemble correctly."],"forward_implications":["A user of KD-VQE only needs to allocate shots by Boltzmann weights and anneal the temperature; no new ansatz or circuit structure is required.","The variational principle for mixed states means the energy estimate produced by KD-VQE remains a valid upper bound on the ground-state energy throughout the annealing process.","KD-VQE remains compatible with extensions developed for standard VQE, so improvements such as better optimizers or error mitigation can be applied after the annealing phase.","In the demonstrated two-site Hubbard case, KD-VQE filters out all five suboptimal candidates and concentrates all shots on the state that reaches the exact ground-state energy."],"supporting_citations":[{"why":"Introduces the variational quantum eigensolver that KD-VQE extends.","marker":"[1]"},{"why":"Provides the theoretical basis for the variational principle used for mixed states.","marker":"[2]"},{"why":"Supplies the knowledge-distillation softening idea that motivates the virtual temperature.","marker":"[27]"},{"why":"Introduces the hardware-efficient ansatz used to prepare trial wavefunctions.","marker":"[17]"},{"why":"Defines the Fermi-Hubbard Hamiltonian used in the demonstration.","marker":"[37]"},{"why":"Reviews VQE optimization and supplies the gradient-descent update rule.","marker":"[11]"},{"why":"Gives the penalty method used to enforce half filling.","marker":"[45]"},{"why":"Documents VQE sensitivity to initialization, the problem KD-VQE targets.","marker":"[35]"}],"fun_headline_variants":["Cranking down virtual temperature finds quantum ground state","VQE anneals to pick the fittest trial state","Virtual annealing prunes bad VQE candidates efficiently","Temperature-guided shots boost VQE convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the virtual annealing schedule and the hand-selected initial wavefunctions ensure the best candidate is not pruned before it reaches the global minimum, and no procedure is given for choosing either on a new problem.","fun_headline_variants_meta":{"raw":{"variants":["Cranking down virtual temperature finds quantum ground state","VQE anneals to pick the fittest trial state","Virtual annealing prunes bad VQE candidates efficiently","Temperature-guided shots boost VQE convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00027,"raw_usage":{"total_tokens":1584,"prompt_tokens":861,"completion_tokens":723,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":664}},"tokens_in":477,"tokens_out":723,"duration_ms":6714,"temperature":1.0,"reasoning_tokens":664,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:40:31.681183+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run KD-VQE on a Hamiltonian whose known ground state is reached only by a trial wavefunction that starts with the highest initial energy of the set; under the paper's schedule (kT=25, 5% decrease per step), if that candidate is pruned before its energy estimate drops below the others, the central claim of reliable convergence to the global minimum fails.","supporting_citations":[{"cited_title":"The first step of KD-VQE is to examine the conservation laws and symmetry properties of the given Hamil- tonian H","cited_arxiv_id":null,"evidence_quote":"Introduces the variational quantum eigensolver that KD-VQE extends."},{"cited_title":"Once the reduced subspace is determined, the next step is to construct the trial wavefunctions and prepare the mixed state according to the Boltzmann distri- bution","cited_arxiv_id":null,"evidence_quote":"Provides the theoretical basis for the variational principle used for mixed states."},{"cited_title":"Wecker, M","cited_arxiv_id":null,"evidence_quote":"Supplies the knowledge-distillation softening idea that motivates the virtual temperature."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the hardware-efficient ansatz used to prepare trial wavefunctions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Fermi-Hubbard Hamiltonian used in the demonstration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reviews VQE optimization and supplies the gradient-descent update rule."},{"cited_title":"Mitarai, T","cited_arxiv_id":null,"evidence_quote":"Gives the penalty method used to enforce half filling."}],"review_version":1}