{"id":"906626ba-02fe-4e42-9b92-c4b54dd34448","arxiv_id":"2508.21594","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"QSUT is an adaptive, sequential quantum hypothesis test for composite hypotheses with anytime-valid type I error control and reduced average copy complexity.","lead":"This paper introduces a sequential test for deciding which of two families of quantum states an unknown state belongs to, using fewer copies when the case is easy. The test guarantees the false-alarm probability stays below any chosen level, and simulations on single-qubit states show it can halve copy complexity.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1's anytime type-I guarantee assumes MLEs in Eqs. (12)–(13) exist, but no closedness or compactness is stated; open hypothesis sets in the experiments can violate this.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the proof of Theorem 4.1 depends on the MLEs in Eqs. (12) and (13) being attained maxima, and the paper claims validity under minimal assumptions without stating compactness, closedness, or regularity conditions. This is not merely a technicality: for open hypothesis sets, argmax may fail, and any arbitrary completion of the undefined estimator can break the uniform inequality \\Lambda \\le \\bar\\Lambda for some true state in S0, voiding the anytime type-I guarantee. The experiments in Section 5.4 use the open alternative set H1 : omega in (45,135) union (135,180), where the same issue can arise for the numerator MLE, although the null sets used are finite or simple. The core e-process argument in Appendix A is otherwise sound: the process \\bar\\Lambda is a martingale under H0 for any predictable numerator state, and Ville's inequality gives the stated bound when the MLEs exist. The two-sided extension in Appendix B has garbled displayed inequalities, but the simultaneous-crossing claim can be repaired by observing \\Lambda_0 \\Lambda_1 \\le 1 via the two MLE optimality conditions, so it is not a fatal flaw. Because the identified gap is real but patchable by adding the missing regularity assumptions and clarifying the experimental protocol, the reader's conditional verdict remains appropriate and no further adjustment is needed.","tokens_in":15325,"tokens_out":14993,"duration_ms":146427,"concrete_test":"Run the commuting-qubit counterexample: let S0 = {diag(p,1-p) : p in (0,1)}, S1 = {diag(0,1)}, POVM = computational basis, epsilon0 = 0.05, true state p = 0.95. Follow QSUT and, on the all-0 history (probability 0.95^t), complete the missing argmax by setting \\hat p_t = 0.9. Numerically or analytically evaluate sup_{rho in S0} Pr[D_T = 1] and check whether it exceeds 0.05. If it does, the theorem's assumption-free statement fails; one can also verify directly that no measurable choice \\hat p_t in S0 satisfies prod_i Tr(\\hat p_t^{⊗n_i} M_i^{X_i}) >= prod_i Tr(\\rho^{⊗n_i} M_i^{X_i}) for all rho in S0 on that path. This distinguishes a missing regularity condition from a harmless technicality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee in Theorem 4.1 is valid only if the MLEs in Eqs. (12)–(13) exist at every round, because Appendix A's inequality (29) uses \\hat\\rho_0^t as an exact maximizer to conclude \\Lambda \\le \\bar\\Lambda. The theorem is stated for arbitrary disjoint subsets S0, S1 of D(H) and any measurement policy, with no compactness, closedness, or regularity condition. For open sets, argmax can fail. Example: S0 = {diag(p,1-p) : p in (0,1)}, S1 = {diag(0,1)}, computational-basis measurements. On the all-0 path (positive probability under p = 0.95), sup_{p in S0} p^t = 1 is not attained, so Eq. (12) has no solution; if the rule is completed with any \\hat p_t < 1, there exists a true p in (\\hat p_t, 1) whose likelihood exceeds \\hat p_t^t, so the key inequality \\Lambda \\le \\bar\\Lambda fails for that \\rho and the uniform type-I bound (15) is not delivered. Thus the advertised minimal-assumptions claim is false as stated; at minimum, S0 must be compact, or the MLE must be replaced by an approximate maximizer with an explicit slack factor. The experimental null sets are finite or simple and therefore satisfy this, but the general claim and the open alternative sets in Section 5.4 need clarification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces QSUT, a sequential testing framework for composite quantum hypotheses. At each round t, an arbitrary adaptive measurement policy selects a POVM, and the test computes a non-anticipating sequential split likelihood ratio Lambda_t in Eq. (14) using MLEs over the null and alternative sets; the test rejects H0 as soon as Lambda_t exceeds 1/epsilon0. Theorem 4.1 claims uniform type-I control, sup_{rho in S0} Pr[D_T=1|rho] <= epsilon0, for any measurement policy, with a proof that dominates Lambda_t by a martingale under H0 and applies Ville's inequality. The paper then gives two instantiations, aLHT/aLHT+ based on Helstrom-Holevo measurements and aLVT based on shallow variational circuits, together with a two-sided extension in Section 4.5. Experiments on single-qubit parametric families compare copy complexity with fixed-copy baselines LHT/bLHT and LVT/bLVT.","tokens_in":15615,"tokens_out":14207,"duration_ms":144045,"significance":"If Theorem 4.1 is correct under the stated assumptions, QSUT would be the first non-asymptotic anytime-valid test for composite quantum hypotheses, and the e-process argument is a clean quantum adaptation of the universal inference framework of Wasserman et al. The central proof is short, uses no fitted parameters, and gives an explicit error bound; the paper also openly acknowledges the classical predecessor. The empirical results are suggestive, showing copy-complexity reductions against fixed-copy baselines. However, the advertised 'minimal assumptions' / 'no assumptions' claim is too strong: as stated, Theorem 4.1 requires an exact null MLE to exist at every round, which fails for open hypothesis sets. This is a load-bearing technical gap, although it is repairable with a compactness condition or a supremum-based formulation.","major_comments":[{"comment":"Theorem 4.1 is stated for arbitrary disjoint subsets S0 and S1 of D(H), but the proof requires the MLE rho_hat_0^t in Eq. (12) to exist and to be an exact maximizer, because Appendix A's inequality (29) needs the product likelihood at rho_hat_0^t to be at least the product likelihood at the true state. For open sets S0, the argmax can fail. For example, take S0 = {diag(p,1-p) : p in (0,1)}, S1 = {diag(0,1)}, and computational-basis measurements; on the all-0 path, which has positive probability under p = 0.95, the supremum of p^t over p in (0,1) is 1 and is not attained. Any completion rho_hat_0^t with p_hat < 1 leaves a true p in (p_hat,1) with p^t > p_hat^t, so the key inequality (29) fails and the uniform type-I bound (15) is not delivered. Thus the claimed validity under 'minimal assumptions' or 'no assumptions' is false as stated. The fix is either to require S0 to be compact (or closed) so that the MLE exists, or, preferably, to define the denominator via sup_{rho in S0} product_i Tr(rho^{otimes n_i} M_i^{X_i}); the inequality Lambda_t <= barLambda_t then holds without attainment. The abstract and Section 1.2 should be revised accordingly.","section":"Section 4.1, Eqs. (12)-(14); Appendix A, Eq. (29)"},{"comment":"The proof of Proposition B.1 is garbled. Eq. (35) uses an undefined object rho_hat_t^1 and presents a chain of inequalities that does not follow from the assumption Lambda^t_QSUT,0 >= 1/epsilon0: the displayed lower bound is not what Eq. (34) implies, and the inequalities as written have the wrong direction. Since the two-sided decision rule (22) cites the claimed impossibility of simultaneous crossing, this proof needs to be corrected or the proposition removed. I note that the error-control claim in Theorem 4.2 is not itself invalidated: the rejection event for the two-sided test is exactly {Lambda_QSUT,0 >= 1/epsilon0}, so Theorem 4.1 applies directly, and similarly for the type-II side. The simultaneous-crossing statement is a separate assertion that still requires a correct proof.","section":"Appendix B, Eq. (35)"}],"minor_comments":[{"comment":"The initial estimator rho_hat_1^0 used in the product for i = 1 is never defined; specify that it is any fixed element of S1 or is chosen by a prespecified rule before data collection.","section":"Section 4.1, Eq. (14)"},{"comment":"There is an index typo in Eq. (31): the term Tr((rho_hat_t^1)^{otimes n_t} M_t^x) should presumably read Tr((rho_hat_1^{t-1})^{otimes n_t} M_t^x); as written the estimator index is inconsistent with the rest of the proof.","section":"Appendix A, Eq. (31)"},{"comment":"The block-size description is internally inconsistent: the text says blocks of size k = 10 and then states that the first k - n_H = 6 copies are measured in the computational basis, which does not match the k + 1 round block structure described in Section 4.3. Please clarify the number of single-copy rounds and the total number of copies per block.","section":"Section 5.3"},{"comment":"The experiments do not state the values of r_z and r_x in Eq. (26), nor the grid/discretization used for the MLEs over open sets such as H1 in Section 5.4. These details are needed for reproducibility, especially because the MLE in Eq. (13) may not be attained for open alternative sets.","section":"Section 5"},{"comment":"The paper should state what happens when the stopping time T is infinite, both in the definition of the type-I event and in the experiments: for truncated runs, please specify whether the copy budget n is consumed and how non-rejection is counted in the reported power and average copy complexity.","section":"Section 3 and Section 5"},{"comment":"The sentence claiming that Lambda_t itself is an e-process would benefit from a one-line justification: Lambda_t is dominated by the martingale barLambda_t, so the e-process property follows from the optional stopping of barLambda_t rather than from an intrinsic supermartingale property of Lambda_t.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the MLE-existence gap in Theorem 4.1, which is fixable with a compactness condition or a supremum-based definition. The e-process argument itself appears sound, and Appendix B's errors do not destroy the error-control claims of Theorem 4.2. I do not see a novelty or citation-pattern issue. If the authors do not repair the statement of Theorem 4.1, the paper should not be accepted as it stands."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a clean transplant of classical universal inference into sequential composite quantum hypothesis testing. That is both its main contribution and its main limitation: QSUT is the first protocol I know of that gives non-asymptotic anytime-valid type-I control for composite quantum hypotheses, and the proof is the standard split-likelihood-ratio supermartingale argument. The two instantiations, aLHT and aLVT, are sensible, and the single-qubit experiments show real early-stopping gains.\n\nThe central theorem is correct, with one caveat. The inequality Λ ≤ \\barΛ in Appendix A uses \\hatρ_t^0 as an exact maximizer in (12). For an arbitrary open S0 the argmax can fail to exist, so the \"minimal assumptions\" claim in the abstract is too strong. The experiments avoid this because the null sets in Sections 5.3 and 5.4 are a single point and a finite set. But the theorem is stated for arbitrary disjoint subsets, and the paper should either assume compactness or closedness of S0 or replace the MLE with an approximate maximizer and add a slack factor. This is a patchable technicality, not a flaw in the main idea.\n\nThe other soft spots are minor. Equation (35) in Appendix B is garbled; the two-sided proof can be reconstructed but needs a rewrite. The experiments use 200 runs, no error bars, no code release, and only single-qubit states, so the copy-complexity savings are proof-of-concept. I would also like a clear paragraph on how QSUT differs from Grootveld et al. (2025), which is dispatched in one sentence.\n\nThe citation pattern looks fine; the use of classical universal inference is acknowledged, and the self-citations to Wasserman et al. and Ramdas et al. are appropriate. Theorem 4.1 is a genuine new result in quantum statistics.\n\nI would send this to a serious referee. It deserves to be published after minor revisions.","headline":"A clean and correct adaptation of universal inference to sequential composite QHT, with a patchable MLE-existence gap and proof-of-concept experiments.","tokens_in":16174,"tokens_out":3661,"would_cite":true,"duration_ms":33139,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces QSUT, a sequential universal test that decides between composite sets of quantum states while provably keeping the false-alarm probability at or below a preset level at every stopping time, under any adaptive…","keywords":["quantum hypothesis testing","sequential analysis","composite hypotheses","universal inference","e-processes","anytime-valid inference","Helstrom-Holevo measurement","copy complexity"],"falsifier":"Run QSUT with an open hypothesis class such as the paper's experiment $H_1:\\omega\\in(45^\\circ,135^\\circ)\\cup(135^\\circ,180^\\circ)$ and record a history whose likelihood is maximized only at the excluded boundary $\\omega=135^\\circ$; then the estimator in (13) has no value, the likelihood ratio is undefined, and the stated minimal-assumption guarantee does not apply. If an implementation substitutes a nearby point and the empirical false-rejection rate exceeds $\\epsilon_0$, the claimed universal control fails.","tokens_in":15091,"feed_emoji":"⚛️","tokens_out":12764,"duration_ms":105129,"temperature":0.7,"pith_summary":"This paper introduces QSUT, a sequential test for deciding whether an unknown quantum state belongs to one composite class of states or another. Unlike fixed-copy protocols, QSUT consumes a variable number of copies and can stop as soon as the evidence is conclusive. The paper proves a non-asymptotic guarantee: for any adaptive measurement strategy, the probability of falsely rejecting the null hypothesis stays at or below a preset level $\\epsilon_0$ at every stopping time. This matters because earlier composite quantum tests could not adapt their copy count, wasting copies on easy states. The paper also gives two practical instantiations and reports single-qubit experiments in which they achieve the error target with lower average copy complexity than fixed-copy baselines.","feed_headline":"Quantum tests can now stop early with error control intact","feed_subtitle":"A sequential split-likelihood test for composite quantum hypotheses keeps type-I error at every stopping time and cuts copy waste.","key_machinery":"The central object is the non-anticipating sequential split likelihood ratio $\\Lambda_t = \\prod_{i=1}^t \\frac{\\mathrm{Tr}((\\hat{\\rho}_1^{i-1})^{\\otimes n_i}M_i^{X_i})}{\\mathrm{Tr}((\\hat{\\rho}_0^t)^{\\otimes n_i}M_i^{X_i})}$, which replaces the alternative state in each factor by the previous-round maximum-likelihood estimate and the null state by the current-round estimate. This ratio is an e-process under the null: because the denominator uses a maximum over the null class, the ratio is bounded by the same expression with the true null state in the denominator, and that expression telescopes into a supermartingale with expectation one. Ville's inequality then converts the crossing of the threshold $1/\\epsilon_0$ into a type-I error bound at any stopping time.","core_discovery":"The paper's central claim is that valid sequential hypothesis testing of composite quantum hypotheses is possible by combining universal inference with arbitrary adaptive measurement policies. The key is to construct the sequential split likelihood ratio from maximum-likelihood estimates of the two classes, using the alternative-class estimate from the previous round in the numerator and the null-class estimate from the current round in the denominator. Under the null hypothesis, replacing the denominator estimate by the true state makes the process a supermartingale with expectation one, so an e-process bound applies: the first time the ratio crosses $1/\\epsilon_0$ has probability at most $\\epsilon_0$ regardless of the measurement policy. The paper further presents two instantiations, one based on Helstrom-Holevo measurements and one based on shallow variational circuits, and reports empirical copy-complexity reductions compared with fixed-copy tests.","pith_inferences":["I infer the same split-likelihood construction could be inverted at multiple thresholds to build anytime-valid confidence regions for an unknown quantum parameter, a direction the paper does not discuss.","I infer the previous-round estimate in the numerator is what prevents the alternative fit from overusing the newest outcome, which is why the e-process property survives arbitrary adaptive measurements.","A directly testable extension is to replace the exact maximum-likelihood estimate by a regularized approximate maximizer so the type-I guarantee remains meaningful when the hypothesis sets are open and no argmax exists.","Since one instantiation uses shallow variational circuits, QSUT is compatible with near-term hardware in principle, but the experiments are simulated and the effect of measurement noise is not analyzed."],"forward_implications":["Composite quantum hypothesis tests no longer need a fixed copy budget: the same protocol provides type-I error control at any stopping time, so resources can be spent adaptively.","Because the error bound holds for every measurement policy, measurement designs can be chosen or tuned purely to make decisions faster, without invalidating the test.","The two-sided version of QSUT uses two simultaneous e-processes and controls both false rejection and false acceptance, provided at least one target error is below 1.","In the single-qubit experiments, the sequential tests reach the same power as fixed-copy baselines with substantially fewer copies once the available budget is large enough."],"supporting_citations":[{"why":"Supplies the universal-inference framework and e-process theory from which QSUT's sequential split likelihood ratio is derived.","marker":"[Wasserman et al., 2020]"},{"why":"The fixed-copy learned Helstrom-Holevo test that QSUT builds on and compares against.","marker":"[Fujiki et al., 2025]"},{"why":"Establishes sequential quantum hypothesis testing for simple hypotheses, the starting point QSUT extends to composite hypotheses.","marker":"[Martínez Vargas et al., 2021]"},{"why":"Defines the Helstrom-Holevo measurement used by the aLHT instantiation.","marker":"[Helstrom, 1969]"},{"why":"Supplies Ville's inequality applied to the supermartingale in the proof of type-I error control.","marker":"[Ville, 1939]"},{"why":"Provides the power-one sequential testing framework and type-I error formulation that the paper adopts.","marker":"[Robbins and Siegmund, 1972]"},{"why":"Gives the informationally complete POVM used for state estimation inside each aLHT block.","marker":"[Renes et al., 2004]"}],"fun_headline_variants":["Sequential quantum hypothesis tests slash copy waste","Quantum tests stop early without inflating errors","Adaptive quantum testing: fewer copies, same guarantees","Universal inference enables sequential quantum tests"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the maximum-likelihood estimates in (12) and (13) exist and attain their maxima over the relevant hypothesis classes at every round, so that the sequential split likelihood ratio is always defined.","fun_headline_variants_meta":{"raw":{"variants":["Sequential quantum hypothesis tests slash copy waste","Quantum tests stop early without inflating errors","Adaptive quantum testing: fewer copies, same guarantees","Universal inference enables sequential quantum tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000557,"raw_usage":{"total_tokens":2642,"prompt_tokens":932,"completion_tokens":1710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":1655}},"tokens_in":548,"tokens_out":1710,"duration_ms":12924,"temperature":1.0,"reasoning_tokens":1655,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:39:57.969306+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run QSUT with an open hypothesis class such as the paper's experiment $H_1:\\omega\\in(45^\\circ,135^\\circ)\\cup(135^\\circ,180^\\circ)$ and record a history whose likelihood is maximized only at the excluded boundary $\\omega=135^\\circ$; then the estimator in (13) has no value, the likelihood ratio is undefined, and the stated minimal-assumption guarantee does not apply. If an implementation substitutes a nearby point and the empirical false-rejection rate exceeds $\\epsilon_0$, the claimed universal control fails.","supporting_citations":[{"cited_title":"Universal inference","cited_arxiv_id":null,"evidence_quote":"Supplies the universal-inference framework and e-process theory from which QSUT's sequential split likelihood ratio is derived."},{"cited_title":"Quantum hypothesis testing for composite alternative hypotheses","cited_arxiv_id":null,"evidence_quote":"The fixed-copy learned Helstrom-Holevo test that QSUT builds on and compares against."},{"cited_title":"Quantum sequential hypothesis testing","cited_arxiv_id":null,"evidence_quote":"Establishes sequential quantum hypothesis testing for simple hypotheses, the starting point QSUT extends to composite hypotheses."},{"cited_title":"Quantum detection and estimation theory","cited_arxiv_id":null,"evidence_quote":"Defines the Helstrom-Holevo measurement used by the aLHT instantiation."},{"cited_title":"Étude critique de la notion de collectif","cited_arxiv_id":null,"evidence_quote":"Supplies Ville's inequality applied to the supermartingale in the proof of type-I error control."},{"cited_title":"A class of stopping rules for testing parametric hypotheses","cited_arxiv_id":null,"evidence_quote":"Provides the power-one sequential testing framework and type-I error formulation that the paper adopts."},{"cited_title":"Symmetric informationally complete quantum measurements","cited_arxiv_id":null,"evidence_quote":"Gives the informationally complete POVM used for state estimation inside each aLHT block."}],"review_version":2}