{"id":"fad6730f-fa7b-4f97-951b-9546d5d3462a","arxiv_id":"2412.17485","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"An entropy-based feedback rule, S = k * 2^H, is proposed to adapt the per-iteration shot count in VQAs, claiming about 50% shot savings over fixed-shot training while preserving final cost accuracy.","lead":"This paper proposes a shot-allocation scheme for variational quantum algorithms that sets the number of measurements per training round from the entropy of the previous round's output distribution. It reports roughly 50% fewer total shots than a fixed 1,024-shot baseline at similar final accuracy, but the savings rely on per-benchmark tuning of the method's key constant.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DDS feedback loop uses a finite-sample entropy estimate that is biased downward at low shot counts, and the k constant is tuned per benchmark; the reported ~50% shot savings are not established as a predictive property.","rationale":"I read the paper as proposing a drop-in heuristic: use the previous iteration's empirical distribution entropy to set the next iteration's shot count via Eq. (4), saving roughly half the shots at near-constant accuracy. For that claim to hold, the entropy feedback must be a faithful proxy for measurement difficulty and the constant k must be set independently of the benchmarks used to report performance. Neither condition is met in the manuscript. The plug-in entropy is downward-biased, and because the next shot count is exponential in that biased estimate, the error feeds back; the paper's own Fig. 15 and Sec. IV.B.3 show k is selected per benchmark after seeing performance. The Sec. II.C calibration is against exact distributions and a Hellinger-distance target, not against the empirical entropies actually consumed by Eq. (4), so the theoretical motivation does not quantitatively justify the chosen k values. This does not show the method cannot work; a fixed-k, seeded re-evaluation might rehabilitate it. But the current evidence is insufficient to accept the headline quantitative claims, so I keep the reader's negative verdict. The reader's weakest-assumption identification is correct as far as it goes; I would give the per-benchmark k selection equal weight, hence 'partial' agreement.","tokens_in":19891,"tokens_out":10167,"duration_ms":105696,"concrete_test":"Run the noiseless QAOA and VQE benchmarks with a single fixed a priori k (e.g., k=2 for all circuits), no per-benchmark tuning, and at least 10 seeded repeats per configuration; report mean plus/minus standard deviation of total shots and ARG. In the same runs, instrument Algorithm 1 by also computing the exact statevector entropy H_true at every iteration. If the DDS shot count S_t is systematically below k * 2^(H_true), or if the savings and ARG numbers change materially relative to the tuned-k results, the reported ~50% reduction is an artifact of in-sample k selection and finite-sample entropy bias rather than a property of the adaptive rule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Eq. (4), S = k * 2^H, where H is the plug-in Shannon entropy of the previous iteration's measured counts (Algorithm 1, Sec. III.B). For an m-outcome distribution sampled with N shots, the plug-in entropy is biased downward by roughly (m-1)/(2N ln 2) bits, and Eq. (4) exponentiates that bias: when N is small relative to 2^n, the shot count is depressed because undersampling hides support, not because the distribution is genuinely low-entropy. The paper never analyzes this feedback loop, nor does it connect the empirical H actually used to the exact-distribution calibration in Sec. II.C. That calibration targets a Hellinger distance of 0.05, but the k values used (2, 8, 64) with the same S = k * 2^H do not reproduce that target for the benchmark distributions; for a 12-qubit uniform state, k=2 gives S=8192, which is far from the Hd=0.05 contour. Finally, Sec. IV.B.3 states that k was 'configured ... to identify an optimal value' per model and qubit count using Fig. 15, so the headline 50.69% shot reduction and 6.03% ARG increase are in-sample best-case numbers, not predictions of a fixed adaptive rule.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes distribution-adaptive dynamic shot (DDS), a heuristic that sets the number of measurement shots for each VQA iteration to S = k * 2^H, where H is the Shannon entropy of the empirical output distribution from the previous iteration. The authors motivate the exponential form by simulations of random-number-generation circuits under a Hellinger-distance constraint (Section II.C), and evaluate DDS on QAOA (PL, BA, WS, SK graphs with 4, 8, and 12 qubits) and VQE (H2, LiH, BeH2) in noiseless and noise-injected simulators, comparing to fixed 1,024-shot, linear, and step baselines. They report an average shot reduction of about 50% for QAOA and 47% for VQE at roughly unchanged ARG, and large ARG improvements over tiered allocation baselines.","tokens_in":20140,"tokens_out":12497,"duration_ms":117069,"significance":"If the central claim were established, DDS would be an attractive plug-in shot-reduction scheme for VQAs because it is optimizer-agnostic and requires only the entropy of the previous iteration's counts. The paper also provides a useful empirical observation that required shot counts grow roughly exponentially with output-distribution entropy. However, the reported quantitative savings are not yet established as a predictive property: the constant k is selected per benchmark, the entropy feedback from finite samples is not analyzed, and the experimental record contains internal inconsistencies and at least one data error. The paper does not ship code, proofs, or a fixed decision rule that would allow an independent check of the headline numbers.","major_comments":[{"comment":"The text states that 'For each specific QAOA model and qubit count, we configured the constant k to identify an optimal value' using Figure 15. The headline 50.69% shot reduction and 6.03% ARG increase are therefore the best in-sample outcomes of a per-benchmark parameter search, not the behavior of a fixed adaptive rule. A valid test would fix k, or choose it on separate validation data, and report held-out results with multiple random seeds and graph instances.","section":"Section IV.B.3"},{"comment":"The H used in Eq. (4) is the plug-in entropy of a finite-shot histogram whose size S was itself set by Eq. (4). Plug-in entropy is biased downward, and the bias grows as S shrinks; Eq. (4) exponentiates this bias, so the feedback loop can drive shot counts below what the true distribution would require. The paper provides no analysis of this loop, no quantification of the bias, and no comparison between the empirical H values observed during training and the exact-distribution calibration of Section II.C.","section":"Section III.B, Eq. (4)"},{"comment":"The calibration in Section II.C targets a Hellinger distance of 0.05, but the k values actually used (64, 8, and 2 for 4, 8, and 12 qubits) are not connected to that target. For a 12-qubit uniform state, k=2 yields S=8192, whereas achieving Hd=0.05 for that distribution requires on the order of 2*10^5 shots. The paper does not explain why the Hd=0.05 calibration justifies the much smaller shot counts used in the experiments.","section":"Section II.C and Section IV.B.3"},{"comment":"The rows for 4-qubit BA QAOA and 4-qubit WS QAOA are identical in Savg, Iterations, and Average Cuts. This is almost certainly a copy error; the affected entries cannot both be correct, and the aggregate 50.69% reduction figure depends on these rows.","section":"Table I"},{"comment":"All results are single runs with no seeds, repetitions, or error bars, and the random graph ensembles are not specified. Given that COBYLA and shot noise are stochastic, the claimed 6.03% ARG difference from the standard method is within the range where run-to-run variation could change the conclusion.","section":"Section IV.B.1 and Tables I-II"},{"comment":"The parameter guidance is internally inconsistent: Section III.B says larger qubit counts should use a larger k, while Section IV.B.3 says larger entropy requires a smaller k, and the experiments use k=64, 8, and 2 for increasing qubit counts. The k values used for VQE are not reported at all, so the VQE results in Table II are not reproducible from the text.","section":"Section III.B vs. Section IV.B.3"}],"minor_comments":[{"comment":"The pseudocode initializes prev_entropy to 10, which exceeds the maximum entropy of small circuits, and it omits the 1,024-shot upper bound mentioned in Section IV.B.1; the algorithm as written is therefore not a complete specification of the evaluated method.","section":"Algorithm 1"},{"comment":"The formula Hrand = log2 n should be Hrand = n for an n-qubit uniform distribution, unless n is redefined as the number of outcomes rather than the number of qubits.","section":"Appendix C"},{"comment":"An increase in ARG of 26.92% is described in the conclusion as 'maintaining a slight increase in accuracy'; since ARG is an error metric, an increase in ARG is a decrease in accuracy, and the wording should be corrected.","section":"Section IV.B.2 and Conclusion"},{"comment":"The text states that for VQE the DDS shot count increases as the entropy of the distribution grows, which contradicts the motivating claim in Section III.A that entropy generally decreases during VQA training; this tension needs to be resolved.","section":"Section IV.B.1"},{"comment":"The Hellinger distance definition has typesetting errors in the PDF text; the formula should be checked so that the square root and normalization are unambiguous.","section":"Section II.B, Eq. (2)"},{"comment":"The statements that DDS is directly applicable to fault-tolerant quantum computers and that it will reduce costs in cloud-based quantum computing are speculative and not supported by the presented experiments.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would benefit from a validation round with fixed hyperparameters and repeated seeds. The duplicate rows in Table I and the missing VQE k values are the kind of issue that should be caught before any resubmission. The paper fits an empirical quantum-computing venue, but the current evidence does not support the generality of the headline savings."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the idea has legs, the evaluation doesn't. DDS uses the previous iteration's plug-in entropy to set the next iteration's shot count via S = k * 2^H. That is a simple, optimizer-agnostic feedback rule, and the specific adaptive scheme is, as far as I know, new. The paper is clearly written and the literature coverage is honest; the distinction against tiered schedules is real.\n\nThe problem is that the central numbers are not established. Section IV.B.3 says plainly that k was configured per model and qubit count to \"identify an optimal value.\" So the 50.69% savings and 6.03% ARG increase are in-sample best-case outcomes, not predictions of a fixed rule. There are no error bars, apparently single runs per configuration, and the baselines (lin fn, step fn) are given fixed parameters while DDS's constant is tuned per benchmark. That asymmetry alone would cap the strength of any claim.\n\nThe stress-test note about the feedback loop is also on point. The entropy fed back is the plug-in entropy from the previous iteration's finite sample. At low shot counts that estimate is biased downward, and the bias is exponentiated by 2^H in Eq. (4). The paper never analyzes whether this loop is stable or whether it just hides support. The calibration in Sec. II.C is for static random circuits with Hd=0.05; the values of k used (2, 8, 64) don't reproduce that target for the actual VQA distributions, so the connection from calibration to deployment is loose.\n\nTwo smaller things: Table I has duplicate rows for 4-qubit BA and 4-qubit WS, with identical Savg, iterations, and average cuts. That looks like a copy-paste error and needs correction. And the \"DDS achieves a 24.39% reduction in ARG for VQE\" claim sits oddly with the noiseless QAOA ARG increase; it is plausible but again rests on the same tuned setup.\n\nWho is this for? People working on VQA resource optimization would find the entropy-feedback idea worth a look. The paper could become useful after major revision: fixed a priori k (or a principled way to set it), repeated seeded runs with error bars, tuned baselines, and an analysis of the entropy-bias feedback loop. As presented, I would not trust the headline numbers. But the idea deserves a serious referee rather than a desk reject—the fixes are substantial but well-posed.","headline":"A plausible entropy-feedback heuristic whose headline shot savings are fitted numbers, not predictions—worth a careful referee but not acceptance as is.","tokens_in":20735,"tokens_out":2513,"would_cite":false,"duration_ms":23452,"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":"A variational quantum algorithm can cut its total measurement shots by roughly half by deriving each round's shot budget from the entropy of the previous round's output distribution.","keywords":["variational quantum algorithms","dynamic shot allocation","information entropy","Hellinger distance","measurement shot reduction","QAOA","VQE","quantum resource estimation"],"falsifier":"Run DDS on a VQA while also computing the exact output distribution by statevector simulation, and compare the entropy estimated from the finite shot sample with the true entropy at the low shot counts DDS actually assigns; if the estimated entropy systematically understates the true entropy, part of the reported saving is an artifact of the estimator's bias. The sharper test is to drive the same shot schedule with exact statevector entropy and check whether the roughly 50% shot saving and the accuracy changes survive.","tokens_in":19654,"feed_emoji":"⚛️","tokens_out":10223,"duration_ms":89889,"temperature":0.7,"pith_summary":"The paper tries to establish a cheap feedback rule for variational quantum algorithms (VQAs): spend measurement shots according to how concentrated the circuit's output distribution has become, rather than fixing the same budget every round. It reports that the shots needed to approximate a distribution to a target Hellinger distance grow roughly exponentially with the distribution's information entropy. On that basis it proposes DDS, which sets the next round's shot count from the entropy of the previous round's measured outcomes, capped at the conventional 1,024-shot budget. In simulations of QAOA and VQE, the rule is claimed to cut total shots by roughly half while keeping final accuracy close to the fixed-shot baseline, and to beat tiered schedules that cut shots without looking at the state. The practical payoff, if the claim holds, is cheaper VQA training on near-term hardware without redesigning the optimizer.","feed_headline":"Quantum variational training halves measurement costs","feed_subtitle":"DDS reads the output distribution's entropy each round, cutting total shots ~50% with accuracy nearly unchanged.","key_machinery":"The load-bearing object is the mapping $S = k \\times 2^{H}$, a calibration curve that the paper fits to static random-number-generation circuits: for a target Hellinger distance of 0.05, the required shots grow exponentially with the information entropy $H$ of the output distribution. In training, DDS estimates $H$ from the counts of the previous iteration and substitutes it into this curve, with $k$ chosen by qubit count and circuit complexity. The mechanism works because VQA output distributions typically start near-uniform, which means high entropy and many shots, and become concentrated on a few high-probability states as parameters converge, which means low entropy and few shots.","core_discovery":"The paper's central claim is that the number of measurement shots a VQA needs in a given training round tracks the information entropy of the circuit's output distribution through an approximately exponential relation $S = k \\times 2^{H}$, where $k$ is a per-circuit constant. DDS computes the empirical entropy of the previous round's measurement outcomes and feeds it into this relation to set the next round's shot count, starting from a default of 1,024 shots and never exceeding it. The reported results are that this rule cuts the total shot budget by 50.69% on average for QAOA and 47.09% for VQE, with the approximation-ratio gap only 6.03% higher than the fixed-shot baseline on QAOA and 24.39% lower on VQE; under simulated device noise the saving is about 31% with minimal accuracy loss. Compared with tiered schedules that simply reduce shots from 1,000 downward, DDS uses more shots but converges to substantially lower final costs, improving the approximation-ratio gap by 63.34% on QAOA and 37.78% on VQE.","pith_inferences":["Editorial: reading the shot rule as a fixed oversampling factor over the distribution's effective support size suggests the same entropy feedback could apply to any distributional sampling task, not only VQA training.","Editorial: because the reported savings are averages over a small set of circuits and one classical optimizer, a sweep over seeds, optimizers, and larger qubit counts is the natural next test of whether the roughly 50% figure holds.","Editorial: the entropy that drives DDS is itself estimated from the shot sample DDS chose, so the schedule and its own estimator form a feedback loop; comparing DDS against a version driven by exact statevector entropy would separate the saving due to the entropy-shots relation from any bias of the finite-sample estimator."],"forward_implications":["On four QAOA graph models, DDS cuts total shots by an average of 50.69% while increasing the approximation-ratio gap by only 6.03% relative to the 1,024-shot fixed baseline.","On VQE molecular benchmarks, DDS cuts total shots by 47.09% and lowers the approximation-ratio gap by 24.39% relative to the fixed baseline.","Under noise models with real-device gate-error rates, DDS still reduces total shots by about 31.15%, and with the shot cap removed it uses 4.30% fewer shots than the fixed baseline while improving accuracy by 18.05%.","DDS leaves the total number of training iterations essentially unchanged, so the savings come from per-iteration measurement cost rather than from stopping early.","Against tiered linear- and step-function schedules, DDS trades additional shots for much better final accuracy: 63.34% ARG improvement on QAOA, 37.78% on VQE, and roughly 68% in the noisy setting."],"supporting_citations":[{"why":"Supplies the tiered shot-allocation baselines (linear and step functions) and the fixed-shot convention that DDS is compared with.","marker":"[30]"},{"why":"Defines the approximation-ratio gap (ARG), the fidelity metric used to compare final accuracy across shot-allocation methods.","marker":"[31]"},{"why":"Defines information entropy, the feedback metric that drives the dynamic shot allocation.","marker":"[32]"},{"why":"Defines the Hellinger distance, the accuracy target used to establish the exponential entropy-versus-shots relation.","marker":"[34]"},{"why":"Supplies the random-number-generation circuits whose entropy-versus-shots data calibrate the exponential shot rule.","marker":"[35]"},{"why":"Provides the simulation backend and the default 1,024-shot budget against which DDS reductions are measured.","marker":"[36]"}],"fun_headline_variants":["Entropy-guided shots halve quantum training costs","Adaptive shot counts from entropy cut VQA shots by half","Entropy-curved shots: quantum training with 50% fewer measurements","DDS: entropy-driven shot cuts halve VQA resource use","Quantum training cuts shots by half with entropy guide"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the entropy computed from one round's finite shot sample reliably predicts how many shots the next round genuinely needs, and that the exponential calibration measured on static random circuits carries over unchanged to VQA circuits.","fun_headline_variants_meta":{"raw":{"variants":["Entropy-guided shots halve quantum training costs","Adaptive shot counts from entropy cut VQA shots by half","Entropy-curved shots: quantum training with 50% fewer measurements","DDS: entropy-driven shot cuts halve VQA resource use","Quantum training cuts shots by half with entropy guide"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000906,"raw_usage":{"total_tokens":3943,"prompt_tokens":1041,"completion_tokens":2902,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":2819}},"tokens_in":657,"tokens_out":2902,"duration_ms":20290,"temperature":1.0,"reasoning_tokens":2819,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:27:20.739645+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DDS on a VQA while also computing the exact output distribution by statevector simulation, and compare the entropy estimated from the finite shot sample with the true entropy at the low shot counts DDS actually assigns; if the estimated entropy systematically understates the true entropy, part of the reported saving is an artifact of the estimator's bias. The sharper test is to drive the same shot schedule with exact statevector entropy and check whether the roughly 50% shot saving and the accuracy changes survive.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the approximation-ratio gap (ARG), the fidelity metric used to compare final accuracy across shot-allocation methods."},{"cited_title":"Phalak and S","cited_arxiv_id":null,"evidence_quote":"Provides the simulation backend and the default 1,024-shot budget against which DDS reductions are measured."}],"review_version":1}