{"id":"87a12b09-f371-45a7-9eba-6323ec0f89a5","arxiv_id":"2502.01730","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An adaptive estimator based on empirical Bernstein stopping reduces the number of measurements needed to estimate ground-state energies with rigorous error bounds, by up to an order of magnitude in numerical benchmarks.","lead":"This paper uses an adaptive stopping rule, the empirical Bernstein stopping algorithm, to estimate the energy of quantum states from measurements with a rigorous error bound. It shows numerically that this method needs fewer measurement repetitions, up to about ten times fewer, than the standard Hoeffding-based bound in several molecular Hamiltonians.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section IIIA's Hoeffding cap is off by a factor of Ng: after N_Hoeff total rounds the grouped estimator has only N_Hoeff/Ng samples, so the fallback guarantee fails for Ng>1 (e.g., failure probability ≈0.46 instead of δ=0.1 for Ng=3).","rationale":"I read the paper as claiming a rigorous adaptive tail bound that improves on the Hoeffding sample complexity while retaining the 1−δ guarantee. The EBS machinery (Section IIC) is a standard, sound way to use empirical variance, and the numerical experiments are relevant. However, the guarantee chain has a weak link in Section IIIA: the cap that is supposed to handle non-termination is expressed in total measurement rounds, but the Hoeffding number N_Hoeff in Eq. (3) is a sample count for the single-shot estimator. For the grouped estimator, each energy sample costs Ng state preparations, so the cap is a factor Ng too small. The reader's weakest_assumption identifies exactly this gap; my derivation makes the failure probability quantitative. This is an internal correctness issue, not a disagreement with consensus: the claimed bound is simply not a bound for the stated algorithm. I see no reason to doubt the numerical findings for early-stopping runs, and the fix is straightforward, which is why I would not call the underlying idea worthless. But the abstract and Section IIIA assert a rigorous tail bound, and that assertion is false as stated. The reader's REJECT verdict is appropriate; it should stand until the cap is corrected and re-benchmarked.","tokens_in":13064,"tokens_out":7463,"duration_ms":66966,"concrete_test":"Re-derive the capped fallback for the grouped estimator and evaluate it on the H2 example of Fig. 1: with S=∑|h_i|, Ng=3, ε=1.6 mHa, δ=0.1, compute P_fail=2 exp(−N_Hoeff ε²/(2Ng S²)). If P_fail≈0.464>δ, the Section IIIA cap is invalid. For a numerical confirmation, run EBS with the current cap on a high-variance state (e.g., a random state for which the empirical variance does not trigger early stopping) and record whether |mean−E|>ε when stopping at the cap; over 10^4 repetitions, the measured failure rate will exceed δ, while the same simulation with the cap replaced by Ng·N_Hoeff will respect δ.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a rigorous (1−δ)-tail bound for the adaptive estimator. That guarantee relies on the fallback cap in Section IIIA: \"If EBS does not terminate after K checks, we can nevertheless stop sampling after N_Hoeff total measurement rounds to reach the guaranteed accuracy epsilon.\" This is not valid for the grouped estimator (Algorithm 1). One energy sample requires measuring all Ng groups, i.e., Ng state preparations. Equation (3) is derived in Appendix B for the single-shot estimator, where one sample costs one state preparation. For the grouped estimator, the Hoeffding bound for n energy samples with range R=2S, S=∑|h_i|, gives P(|mean−E|≥ε) ≤ 2 exp(−n ε²/(2S²)). To make this ≤δ one needs n ≥ 2S² log(2/δ)/ε² = N_Hoeff samples, i.e., Ng·N_Hoeff total state preparations. The paper caps total rounds at N_Hoeff, so the fallback has only n=N_Hoeff/Ng samples. Substituting yields P_fail = 2 exp(−N_Hoeff ε²/(2Ng S²)) = 2^{1−1/Ng} δ^{1/Ng}, which exceeds δ for every Ng>1 (for Ng=3, δ=0.1 it is ≈0.464). Thus the asserted worst-case guarantee does not hold; the early-stopping Bernstein bound only covers runs that terminate at a check, not the capped non-termination event. The numerical benchmarks may still be valid for low-variance states, but they do not substantiate the rigorous tail bound claimed in the abstract and Section IIIA. The missing factor is concrete and repairable (replace N_Hoeff by Ng·N_Hoeff in the cap), but as written the central claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adaptive energy estimation method for quantum many-body systems that combines the empirical Bernstein stopping (EBS) algorithm with Pauli grouping. The central claim is that the resulting estimator has a rigorous (1−δ) tail bound that exploits the empirical variance of the state, and the authors present numerical benchmarks on small molecular Hamiltonians showing that EBS requires roughly 10–30% of the number of measurement rounds mandated by a Hoeffding-based single-shot estimator.","tokens_in":13406,"tokens_out":8464,"duration_ms":80691,"significance":"If the rigorous tail bound were established, the work would be a valuable contribution to quantum energy estimation, as it offers a principled way to exploit low empirical variance to reduce sampling overhead. The paper includes reproducible numerical experiments and a public code repository, which are strengths. However, the main correctness claim is undermined by a flaw in the fallback cap described in Section IIIA, so the algorithm as stated does not deliver the promised (1−δ) guarantee.","major_comments":[{"comment":"The cap at N_Hoeff total measurement rounds does not preserve the (1−δ) accuracy guarantee for the grouped energy estimator. Each execution of Algorithm 1 requires Ng state preparations (one per group) to yield a single energy sample, so after N_Hoeff total measurement rounds the algorithm has only N_Hoeff/Ng independent samples. The Hoeffding bound (3), derived in Appendix B for the single-shot estimator where one sample costs one state preparation, requires n ≥ N_Hoeff samples to achieve error ε with probability at least 1−δ. Substituting n = N_Hoeff/Ng into the Hoeffding bound gives a failure probability of 2 exp(−N_Hoeff ε²/(2 Ng S²)) = 2^{1−1/Ng} δ^{1/Ng}, which exceeds δ for every Ng > 1 (e.g., about 0.464 for Ng = 3 and δ = 0.1). Thus the claimed worst-case guarantee fails on the non-termination event; the empirical Bernstein bound in Eq. (4) only covers runs that stop at a check. This invalidates the abstract's claim of a rigorous tail bound for the algorithm as described.","section":"Section IIIA, Eq. (6) and surrounding text"},{"comment":"Even if the cap were changed to Ng·N_Hoeff total measurement rounds, the paper does not account for the union of failure probabilities. The EBS stopping rule allocates an inconfidence sequence (d_i) summing to δ for the checks, while the Hoeffding fallback at the cap is a separate probabilistic statement that carries its own failure probability. Simply stopping at the cap and invoking Hoeffding with the same δ leads to a total failure probability that can be as large as 2δ. The authors need to specify how the overall failure probability δ is partitioned between the EBS stopping rule and the Hoeffding fallback.","section":"Section IIIA, finite inconfidence sequence"},{"comment":"The numerical benchmarks are presented as support for the rigorous guarantee, but because the cap is flawed, the experiments do not substantiate the central claim. For the tested low-variance states the algorithm may terminate before reaching the cap, so the reported improvements could remain valid after a correction, but this needs to be verified. The authors should rerun the benchmarks with a correctly capped algorithm (e.g., with the cap at Ng·N_Hoeff and a proper split of the failure probability) and state whether the claimed reductions in measurement rounds persist.","section":"Section IIIB, Figures 1 and 2"}],"minor_comments":[{"comment":"The notation 'NHoeff !≥ Ng⌊βK⌋' contains a typographical artifact; it should be written as a clear inequality. Also, the text should explicitly define whether 'measurement round' means one state preparation or one execution of Algorithm 1, since the paper uses both notions in different places.","section":"Section IIIA, Eq. (6)"},{"comment":"The sentence 'we find that EBS consistently yields |Ê − E| ≤ ε/4 empirically' reports an observed property, not a rigorous guarantee. It would help to label this explicitly as an empirical observation so readers do not conflate it with the proven tail bound.","section":"Section IIIB, Figure 1 inset"},{"comment":"The term 'inconfidence' is nonstandard; consider using 'failure probability' or 'confidence parameter' consistently to improve readability.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main issue is a concrete and local error: the N_Hoeff cap is off by a factor of Ng for the grouped estimator, and the failure-probability budget is not split between the EBS stopping rule and the Hoeffding fallback. Both are repairable within the paper's scope, so I recommend major revision rather than rejection. The numerical results may still hold for early-terminating states, but the authors must correct the guarantee proof and rerun the benchmarks under the corrected protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on this one. The core idea — applying empirical Bernstein stopping to grouped Pauli measurements for energy estimation — is a sensible new combination, and the paper is clearly written with public code. But the central claim of a rigorous tail bound has a concrete off-by-Ng flaw in the fallback cap in Section IIIA.\n\nEBS draws energy samples via Algorithm 1, which costs Ng state preparations per sample. The Hoeffding bound in Eq. (3) is for the single-shot estimator, where one sample costs one state preparation. The paper caps total measurement rounds at N_Hoeff, but after that many rounds the grouped estimator has only N_Hoeff/Ng samples. The Hoeffding fallback then fails to deliver the promised 1-delta guarantee: for Ng=3 and delta=0.1, the failure probability is about 0.46, not 0.1. So the non-termination event is not covered. This is easily repairable — cap at Ng·N_Hoeff total rounds instead — but as written the guarantee doesn't hold.\n\nThe baseline comparison is also a bit unfair. The paper compares EBS's total measurement rounds to N_Hoeff from Eq. (3), which is a single-shot sample count. A grouped Hoeffding baseline would cost Ng·N_Hoeff rounds. Without that, the claimed up-to-tenfold improvement may be overstated by a factor of Ng. The benchmarks might still show an advantage after re-baselining, but the paper doesn't currently demonstrate it.\n\nCredit where due: the adaptation of EBS to grouped measurements is new, the empirical Bernstein inequality is applied correctly, and the numerical exploration is transparent. The flaw is load-bearing but localized.\n\nFor a reading group, this is a decent case study in unit mismatches. I'd send the paper out for review — a competent referee will catch the cap issue, and the practical question (does adaptive stopping reduce cost under a fair comparison?) is worth an answer. Expect heavy revision.","headline":"The central tail-bound claim has a concrete off-by-Ng flaw in the fallback cap, but the core EBS-for-grouped-measurement idea is sound and repairable.","tokens_in":13973,"tokens_out":3597,"would_cite":false,"duration_ms":33058,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"Adaptive variance-aware stopping can cut quantum energy readout shots by up to 10x","keywords":["energy estimation","sampling complexity","empirical Bernstein stopping","quantum many-body systems","Pauli measurements","variance information","quantum chemistry","readout guarantees"],"falsifier":"A direct tail calculation for the grouped estimator after NHoeff total measurement rounds shows its effective sample size is NHoeff/Ng; for any Ng > 1 there exists a state (e.g., one with maximal variance on each group) whose Hoeffding confidence interval has width exceeding epsilon, so the claimed 1-delta guarantee fails unless an additional argument is supplied. A simulation of the capped algorithm on such a state should show a failure rate above delta.","tokens_in":12783,"feed_emoji":"⚛️","tokens_out":9132,"duration_ms":80741,"temperature":0.7,"pith_summary":"This paper aims to cut the sampling cost of estimating ground-state energies of quantum many-body systems while keeping rigorous error bars. It adapts the empirical Bernstein stopping (EBS) algorithm to the grouped Pauli measurement setting: as energy samples are collected, the algorithm tracks their empirical variance and stops as soon as the empirical Bernstein confidence interval fits inside the target accuracy. The authors provide a rigorous tail bound that leverages this empirical variance, and their numerical benchmarks on small molecule Hamiltonians show that EBS consistently needs only 10–30 percent of the measurement rounds required by the previous Hoeffding-based guarantee, sometimes up to an order of magnitude fewer. The practical motivation is that energy readout is a bottleneck in near-term quantum chemistry algorithms, and any provable reduction in sampling complexity matters when high precision is required.","feed_headline":"Variance-aware stopping cuts energy readout shots up to 10x","feed_subtitle":"Empirical-Bernstein stopping exploits a state's low variance and needs only 10-30% of the standard measurement rounds.","key_machinery":"The load-bearing mechanism is the empirical Bernstein stopping algorithm combined with a grouped Pauli measurement scheme. EBS repeatedly evaluates the empirical Bernstein inequality, replacing the unknown variance with the running empirical variance, and stops when the resulting confidence interval width falls below the target accuracy. Geometric sampling checks the stopping condition only at times floor($\\beta$^k) to reduce computational overhead, and a martingale-based mid-interval stopping rule prevents overshooting. The paper caps total measurement rounds at the Hoeffding number NHoeff to guarantee termination, and distributes the failure probability delta over a finite schedule of per-check inconfidences d_i. The grouped estimator is what makes EBS compatible with practical commuting-group measurements: each group is measured once per round, producing one energy sample per Ng state preparations.","core_discovery":"The paper's central claim is that an adaptive estimator built on the empirical Bernstein stopping rule can replace the non-adaptive Hoeffding readout guarantee without sacrificing the (epsilon, delta) accuracy promise. The construction combines a grouped energy estimator, where measuring every commuting Pauli group once yields one unbiased energy sample, with EBS running on those i.i.d. samples. The stopping rule uses the empirical Bernstein inequality, so a state with small actual variance is detected through its small empirical variance and the procedure terminates early; the paper proves the associated tail bound and verifies numerically that the number of measurement rounds is 10–30 percent of the Hoeffding baseline for the ground states of H2, LiH, BeH2, H2O, and NH3 Hamiltonians under several fermion-to-qubit encodings.","pith_inferences":["Allocating measurements among groups according to per-group empirical variance could lower the variance of the grouped estimator and further reduce total rounds; the paper mentions this as future work.","The equal division of the failure probability $\\delta$ among checks is a design choice; alternative schedules may reduce the expected sample count without changing the tail bound.","The same empirical-Bernstein machinery could apply to other readout strategies, such as randomized classical shadows, if a suitable empirical Bernstein inequality is derived for the estimator.","For highly mixed states the adaptive estimator offers little over the simple bound, so a practical user should measure the state's concentration before choosing between EBS and a fixed-sample strategy."],"forward_implications":["For low-variance states, the expected sample complexity scales as $\\max(\\sigma^2/\\epsilon^2, R/\\epsilon)$ up to logarithmic factors, giving up to a quadratic improvement over the Hoeffding scaling $(\\sum_i |h_i|)^2/\\epsilon^2$.","The numerical benchmarks show EBS uses 10–30 percent of the Hoeffding measurement rounds to reach chemical accuracy for several small molecules, with the advantage persisting as the qubit count grows to 16.","The advantage appears only at high precision (epsilon below about $10^{-2}$ Hartree); at moderate precision the constant overhead of the grouped estimator cancels the benefit.","Because the guarantee degrades to the Hoeffding bound in the worst case, the method is safe to use as a drop-in readout subroutine for hybrid quantum-classical algorithms such as VQE."],"supporting_citations":[{"why":"Supplies the single-shot estimator and the Hoeffding bound NHoeff that serves as both the baseline and the cap on total measurement rounds.","marker":"[11]"},{"why":"Supplies the empirical Bernstein stopping algorithm, including the geometric sampling schedule and the expected sample complexity bound.","marker":"[20]"},{"why":"Supplies the empirical Bernstein inequality used to form the confidence interval in the stopping rule.","marker":"[21]"},{"why":"Supplies the grouping method used to partition Hamiltonians into commuting groups for the numerical benchmarks and the grouped estimator.","marker":"[18]"},{"why":"Supplies the martingale-based mid-interval stopping rule that controls overshooting in the geometric sampling schedule.","marker":"[22]"}],"fun_headline_variants":["Variance-aware stopping slashes energy readout shots 10x","Adaptive estimator uses empirical variance to cut measurement rounds","Empirical Bernstein stopping needs only 10-30% of readout shots","Exploit state variance to estimate energy with 10x fewer measurements"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee relies on the unproven assertion that stopping after NHoeff total measurement rounds still delivers the promised accuracy, even though the grouped estimator consumes Ng state preparations per energy sample and therefore has only NHoeff/Ng samples by that point.","fun_headline_variants_meta":{"raw":{"variants":["Variance-aware stopping slashes energy readout shots 10x","Adaptive estimator uses empirical variance to cut measurement rounds","Empirical Bernstein stopping needs only 10-30% of readout shots","Exploit state variance to estimate energy with 10x fewer measurements"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000532,"raw_usage":{"total_tokens":2523,"prompt_tokens":869,"completion_tokens":1654,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":1580}},"tokens_in":485,"tokens_out":1654,"duration_ms":12450,"temperature":1.0,"reasoning_tokens":1580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:40:07.622990+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct tail calculation for the grouped estimator after NHoeff total measurement rounds shows its effective sample size is NHoeff/Ng; for any Ng > 1 there exists a state (e.g., one with maximal variance on each group) whose Hoeffding confidence interval has width exceeding epsilon, so the claimed 1-delta guarantee fails unless an additional argument is supplied. A simulation of the capped algorithm on such a state should show a failure rate above delta.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the empirical Bernstein stopping algorithm, including the geometric sampling schedule and the expected sample complexity bound."},{"cited_title":"Audibert, R","cited_arxiv_id":null,"evidence_quote":"Supplies the empirical Bernstein inequality used to form the confidence interval in the stopping rule."},{"cited_title":"Mnih,Efficient stopping rules, Master’s thesis, Univer- sity of Alberta (2008)","cited_arxiv_id":null,"evidence_quote":"Supplies the martingale-based mid-interval stopping rule that controls overshooting in the geometric sampling schedule."}],"review_version":1}