{"id":"468accf6-c4f0-41b0-bdb3-f3cb6f41cb7f","arxiv_id":"2412.12413","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For quantum state Procrustes optimization, restricting projective measurements to an r-dimensional subspace costs at most c log(r)^4 in objective value, independent of the ambient dimension.","lead":"This paper proves a bound on the accuracy lost when quantum measurement devices are restricted to a low-dimensional subspace: the loss grows only polylogarithmically with that dimension, not with the full system size. The result matters for practical quantum measurement design, where ideal measurements are often impossible to implement.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 2 appears to misapply Lemma 9; the equality displayed around Eq. (92) is not generally true, so the weight-extraction step in Theorem 1 is unproved.","rationale":"The reader's weakest assumption was Proposition 1 / Lemma 2, the combinatorial moment estimate. That is indeed delicate, but I find a more elementary and more directly load-bearing gap in Proposition 2, which is used at the final link of the proof. The equality after Lemma 9 appears to have the wrong rank-one factor; a direct algebraic check shows the two expressions differ unless tildeL^{-2it} commutes with tildeQ_i, which is not assumed. If this is correct, Eq. (95) fails and Theorem 1 is not established by the manuscript. I am not claiming the theorem is false; the gap may be repairable, and several other parts of the proof (Lemma 1, the randomization in Lemma 3, the concentration argument, and the dimension-enlargement step) appear coherent. For that reason I do not recommend outright rejection; I recommend conditional acceptance pending a corrected proof or an independent verification of Proposition 2. The proposed numerical test is cheap and settles whether the specific equality is valid. No machine-checked proof or released code is available to compensate for this gap.","tokens_in":32953,"tokens_out":19025,"duration_ms":156532,"concrete_test":"Numerically check Eq. (92) in dimension 2. Set tildeL=diag(1,1/2), Q_i=|+><+|, rho=|0><0|, tau=I/2, and mu(t)=1/(pi(1+t^2)). Compute LHS = Tr[rho tildeL^2 Q_i tau tildeL^2 Q_i] and RHS = integral_{-inf}^{inf} Tr[rho tildeL^{-2it} Q_i tau tildeL^{-2it} Q_i] mu(dt) to high precision, e.g., with adaptive quadrature. If LHS and RHS differ beyond numerical tolerance (say 1e-12), the equality asserted in Proposition 2's proof is false, and the proof of Theorem 1 is incomplete. Also re-derive Eq. (92) directly from Lemma 9 to confirm the missing rank-one factor is not a typographical artifact.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 1 relies on Section 4.6 Eq. (95), which uses Proposition 2 to replace the random weighted measurement Q_hatL by an unweighted PM Z at a cost ||hatL||^4. In the proof of Proposition 2, after the Hadamard three-lines bound, Lemma 9 is invoked to claim, for each rank-one tildeQ_i: Tr[rho tildeL^2 tildeQ_i tau tildeL^2 tildeQ_i] = integral_R Tr[rho tildeL^{-2it} tildeQ_i tau tildeL^{-2it} tildeQ_i] mu(dt).  (Eq. 92). Lemma 9 with K=tildeL^2, however, gives KAK = integral tildeL^{-2it} A tildeL^{-2it} mu(dt). Taking A=tildeQ_i tau tildeQ_i yields the integrand Tr[rho tildeL^{-2it} tildeQ_i tau tildeQ_i tildeL^{-2it}], not the displayed expression; the extra tildeQ_i after tau is missing. The two integrands are not equal in general because tildeL^{-2it} and tildeQ_i need not commute: for rank-one tildeQ_i=|q><q|, the displayed integrand is <q|rho tildeL^{-2it}|q><q|tau tildeL^{-2it}|q>, whereas the Lemma 9 integrand is <q|tau|q><q|tildeL^{-2it} rho tildeL^{-2it}|q>. A 2x2 example (tildeL=diag(1,1/2), Q_i=|+><+|, rho=|0><0|, tau=I/2) gives different values, including after integration against the Cauchy measure. The later Cauchy-Schwarz step depends on the product form of the displayed expression. Without Proposition 2, the chain (95) cannot extract the weight, so the c log(r)^4 bound in Eq. (21) does not follow from the written proof. This is an internal gap, not a question of matching existing bounds.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the suboptimality ratio K_{n,r} for a Procrustes problem in which the optimization is over projective measurements aligned with a low-rank subspace S of dimension r inside an ambient space H of dimension n. The main result, Theorem 1 (Eq. (21)), states that K_{n,r} is bounded by c log(r)^4 uniformly in n. The proof has six steps: reduction of a projective measurement on H to a sum of T weighted projective measurements on S (Lemma 1); existence of a permutation and phase choice for which the associated frame Gram operator has operator norm O(log r) (Proposition 1, based on the moment estimate Lemma 2 in Appendix A); randomization with Gaussian weights (Lemma 3); concentration of the random matrix norm via Tropp's inequality (Lemmas 7, 8, Corollary 2); an interpolation inequality that extracts the weight at a cost ||L||_op^4 (Proposition 2); and a final combination, with the small-n case handled by enlarging the ambient space. The paper also includes numerical experiments and a conjecture that K_{n,r}=O(1).","tokens_in":33300,"tokens_out":29659,"duration_ms":278567,"significance":"If correct, Theorem 1 is a strong and clean result: the suboptimality of low-rank-aligned projective measurements is independent of the ambient dimension and only polylogarithmic in the rank r. The proof is essentially self-contained, deriving the bound from standard external results (Lieb's trace convexity, the Hadamard three-lines theorem, and Tropp's matrix concentration inequalities) together with new technical ingredients, notably the probabilistic equipartition of Parseval frames and the operator-norm moment estimates in Lemma 2 and Appendix A. There are no fitted constants and no circular dependence on the target inequality. I also checked the specific stress-test concern about Proposition 2, Eq. (92): the displayed equality is in fact true, although the proof should justify it more explicitly. The main caveat is that several displayed formulas in the combinatorial part are misprinted, which currently makes parts of Appendix A formally incorrect as written; these appear to be typographical and locally fixable.","major_comments":[],"minor_comments":[{"comment":"The step labeled 'Lemma 9' is too terse and, as written, invites the objection that Lemma 9 applies to one copy of K while the integrand contains two powers of \\tilde L^{-2it}. The equality is true, but it should be proved explicitly: for K=\\tilde L^2\\ge 0 with \\|K\\|_{op}\\le 1, the Cauchy measure \\mu(dt)=dt/(\\pi(1+t^2)) satisfies \\int K^{-it}\\otimes K^{-it}\\,\\mu(dt)=K\\otimes K, because -\\log K has nonnegative eigenvalues and the Fourier transform of \\mu is e^{-|x|}. I recommend adding this one-line verification so that the application to the product of two traces is transparent.","section":"§4.5, Eq. (92)"},{"comment":"Several displayed probability and weight formulas are missing a reciprocal denominator. For example, in Lemma 14(ii), Eq. (122) as printed gives a value that can exceed 1; the correct expression is C_{V,\\Gamma}= r\\big/\\big[(\\tbinom{Tr}{n(V)})(\\tbinom{n(\\Gamma)}{z(\\Gamma)})\\big], matching the counting argument in the proof. The same missing reciprocal appears in Eqs. (124), (150), (157), and (158). These are almost certainly typographical, but they must be corrected because the subsequent asymptotic estimates in Lemma 17 rely on the reciprocal form.","section":"Appendix A, Lemma 14 and Lemma 16-17"},{"comment":"The notation in Eq. (45) is ambiguous: the first term should read r k^k (not r^k k). The proof of Proposition 1 only goes through with r k^k, since F(k)^{1/k} is then O(r^{1/k} k)=O(log r). Please make the exponent explicit throughout, including Lemma 19 and the proof of Proposition 1.","section":"§4.2, Lemma 2 and Proposition 1"},{"comment":"The trace term Tr[\\tau Q[\\rho]] appearing on the left-hand side is not quantified. The inequality needs a clearly specified Q (e.g., any fixed feasible Q, or an appropriate extremal choice), otherwise the statement is not well posed.","section":"§3, Corollary 1, Eq. (22)"},{"comment":"The paragraph after Eq. (104) correctly notes that \\hat K_r can fall below 1 and that convergence is not certified; this should be stated as a limitation of the numerics in the main text, not only as an aside, since Figure 1 might otherwise be read as evidence for Conjecture 5 with C=1.","section":"§5, Numerical experiments"},{"comment":"Reference [25] contains corrupted author names ('Micha/suppress l Oszmaniec', 'Zbigniew Pucha/suppress la'); the correct names are Micha\\l{} Oszmaniec and Zbigniew Pucha\\l{}a. Please also check other author names for OCR-style corruption.","section":"References"}],"recommendation":"minor_revision","confidential_remarks":"The paper's central claim appears correct and the proof strategy is substantial. My main worry is presentation, not correctness: the missing denominators in the Appendix A formulas make the formal proof of Lemma 2 invalid as printed, and the use of Lemma 9 in Proposition 2 needs a short justification. Both are local fixes. The numerical section is exploratory and should not be used as evidence for the conjecture. If the author corrects the displayed formulas and clarifies Eq. (92), I would support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the main theorem is credible and the proof chain holds together — including the step the stress-test flagged. Eq. (92) is actually true. The author misapplies Lemma 9 there: with K = L̃², Lemma 9 applied to A = Q̃τQ̃ gives KQ̃τQ̃K, a different integrand. But the displayed identity is correct by a direct spectral argument: with 0 ≤ λ_a ≤ 1, λ_aλ_b = ∫(λ_aλ_b)^{-it}μ(dt) term by term because the Cauchy measure's Fourier transform is e^{-|x|}. The stress-test's own 2×2 example confirms this — both the left side and the integrated displayed integrand equal 5/32. So the conclusion that weight extraction fails and Theorem 1 collapses does not survive reading the paper. The real problem is the justification, not the statement: the author should replace the wrong Lemma 9 citation with the spectral argument.\n\nWhat's new and good: first general upper bound on K_{n,r} for r ≥ 2, polylog in r and independent of n; prior work only settled r = 2. The equipartition-of-Parseval-frames technique looks original, the fourth-moment/log⁴ interpolation is sound, and the citation pattern is clean.\n\nSoft spots, in proportion. Lemma 2/Appendix A is the weakest load-bearing piece: a long combinatorial moment computation, not machine-checked, and Proposition 1 rests on it; that deserves independent verification. The Lemma 9 slip in Proposition 2 is real and should be fixed. Minor: the M0 case in Proposition 2 is dismissed as 'analogous' without details; Lemma 7's tail is stated sharper than what Tropp's theorem gives as written (the constant doesn't matter, but the adaptation should be shown); the numerics are exploratory, code on request, and the paper says so itself.\n\nWho it's for: people working on measurement restrictions and Procrustes problems, and anyone who wants a new frame-equipartition tool. The conjecture (constant upper bound) is not resolved, and the paper is honest about that.\n\nRecommendation: send to a serious referee. Ask for a corrected proof of Eq. (92) and a hard look at Appendix A. My verdict would be accept after those.","headline":"The bound is real and Eq. (92) survives the stress-test; the paper's actual flaw is a wrong justification, not a false statement.","tokens_in":33933,"tokens_out":23591,"would_cite":true,"duration_ms":179797,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P15","60B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that restricting projective measurements to a low-rank subspace costs at most a polylogarithmic factor in the subspace dimension, with no dependence on the ambient dimension.","keywords":["projective measurements","low-rank subspace","suboptimality ratio","Parseval frames","Procrustes problem","random matrix concentration","trace inequalities","quantum measurement"],"falsifier":"Exhibit, for arbitrarily large $r$, a Parseval frame of $S$ for which every permutation and phase choice yields an equipartition whose averaged overlap matrix has operator norm $\\omega(\\log r)$; Proposition 1 would be false and Theorem 1 would not follow from the submitted proof. A concrete route: take the frame with one block equal to the identity ($L_1=1_S$), where the paper's own analysis gives expected operator norm $\\approx \\log r/\\log\\log r$, and check whether the supremum over partitions crosses the $\\log r$ threshold.","tokens_in":32655,"feed_emoji":"⚛️","tokens_out":6844,"duration_ms":60664,"temperature":0.7,"pith_summary":"In a quantum Procrustes problem, one measures a fixed state and wants the expected distance to another fixed target state to be as small as possible. This paper shows that if only projective measurements aligned with a low-rank subspace $S$ of dimension $r$ are allowed, the loss compared with using all projective measurements on the full space is at most $c\\log(r)^4$, for a universal constant $c$ and for all ambient dimensions $n\\ge r\\ge 2$. This matters because measurement instruments often cannot implement arbitrary measurements, and the result says that restricting to the small subspace where the two states live is nearly optimal regardless of how large the surrounding Hilbert space is. The proof combines projective-measurement decompositions, Parseval frames, probabilistic and random-matrix estimates, and new trace inequalities. Numerical experiments suggest the ratio is close to $1$ for small $r$, though the paper notes that more complex behavior is not ruled out.","feed_headline":"Low-rank restriction costs at most a polylog factor","feed_subtitle":"In a quantum Procrustes problem, subspace-aligned measurements stay within polylog of the full optimum, independent of ambient size.","key_machinery":"The central object is a decomposition of a restricted projective measurement into weighted projective measurements over the low-rank subspace $S$. Given a projective measurement $P$ on $H$, its projectors projected onto $S$ form a Parseval frame $\\{|v_i\\rangle\\}$ with $\\sum_i |v_i\\rangle\\langle v_i| = 1_S$; splitting this frame into $T$ equal blocks gives weight matrices $L_t = \\sum_j |v_{r(t-1)+j}\\rangle\\langle e_j|$, so that $\\operatorname{Tr}[\\rho P[\\tau]] = \\sum_t \\operatorname{Tr}[\\rho Q_{L_t}[\\tau]]$ with $\\sum_t L_t L_t^\\dagger = 1_S$. The argument then randomizes: with i.i.d. Gaussians $g_t$, the operator $\\hat L = \\sum_t g_t L_t$ has variance controlled by the operator norm of $L = \\sum_t L_t^\\dagger L_t$, which the probabilistic method bounds by $O(\\log r)$ through a delicate moment count of cycles in random partitions. A complex-interpolation inequality extracts the weight at cost $\\|L\\|_{\\mathrm{op}}^4$, and Gaussian concentration supplies $\\mathbb{E}\\|\\hat L\\|_{\\mathrm{op}}^4 \\le c\\log(r)^4$.","core_discovery":"Define $K_{n,r}$ as the smallest constant such that for every pair of density matrices $\\rho,\\tau$ supported on an $r$-dimensional subspace $S$ of an $n$-dimensional Hilbert space, the optimal projective-measurement overlap over all of $H$ is at most $K_{n,r}$ times the optimal overlap using only projective measurements aligned with $S$. The paper establishes $K_{n,r}\\le c\\log(r)^4$ for all $n\\ge r\\ge 2$. The proof decomposes an arbitrary measurement on $H$ into $T=\\lceil n/r\\rceil$ weighted measurements on $S$ whose weight matrices come from an equipartition of a Parseval frame, shows via the probabilistic method that a permutation and phase choice make the averaged Gram matrix have operator norm $O(\\log r)$, randomizes the weights with independent Gaussians to decouple the $T$ terms, uses Gaussian concentration to control the resulting random operator, and finally strips the weights with a new complex-interpolation trace inequality. Each step preserves the projective-measurement structure, which is what makes the bound about measurements rather than about general quantum channels.","pith_inferences":["If the paper's conjecture of a bounded $K_{n,r}$ holds, the polylogarithmic slack in Theorem 1 is an artifact of the proof rather than a real cost; a concrete test would push numerical optimization to $r\\gg 20$ with global search or semidefinite bounds to check whether $K_r>1$ appears.","The interpolation step (Proposition 2) is stated for projective measurements but the proof uses only that the $Q_i$ are symmetric rank-one operators, so the same suboptimality bound should extend to rank-one POVMs and symmetric Kraus representations; if so, similar guarantees would hold for measurement implementations that simulate POVMs by postselection.","The combinatorial moment bound (Lemma 2) controls the averaged Gram matrix of arbitrary Parseval frames under equipartitions, independent of the Procrustes objective, so it could be reused in other frame-restricted quantum optimization problems such as state discrimination or tomography with block-structured instruments."],"forward_implications":["For the Procrustes problem, optimizing over subspace-aligned measurements approximates the full optimum up to a factor $c\\log(r)^4$, so measurement policies can be chosen in the low-rank space.","The bound is independent of the ambient dimension $n$: adding unused dimensions to the Hilbert space does not worsen the suboptimality ratio.","Corollary 1 gives a concrete, though not sharp, approximation inequality for the expected Frobenius-distance minimization.","The polylogarithmic exponent $4$ arises from the dimensional factor in Gaussian matrix concentration, and the paper notes that improving the bound would require exploiting extra low-rank structure in the weight matrices.","Numerical experiments with manifold gradient ascent suggest $K_r\\approx 1$ for $r\\le 20$, and the paper poses the conjecture that $K_{n,r}\\le C$ for a universal constant $C\\ge 1$ (possibly $C=1$)."],"supporting_citations":[{"why":"Supplies the quantum-control formulation of maximizing $\\operatorname{Tr}[\\rho P[\\tau]]$ over projective measurements, the objective whose suboptimality ratio is studied.","marker":"[26]"},{"why":"Gives first-order stationary conditions for the unitary optimization and proves the exact $K_{n,2}=1$ case for pure states, the baseline the theorem extends.","marker":"[36]"},{"why":"Provides the trace-convexity inequality (Lemma 5) used to pass from a biweighted measurement to single weighted measurements in the randomization step.","marker":"[23]"},{"why":"Supplies the Gaussian matrix concentration inequality (Lemma 7) that controls the operator norm of the randomized weight operator $\\hat L$ and yields the fourth-moment bound.","marker":"[34]"},{"why":"Provides the method of moments used to estimate the operator norm of $L^\\pi(\\Theta)$ through trace moments.","marker":"[33]"},{"why":"Gives the tight balls-into-bins analysis used to argue the near-sharpness of the $\\log r/\\log\\log r$ order in the extreme frame case.","marker":"[28]"},{"why":"Supplies the Stirling-number counting formula used in Lemma 13 to bound how many permutations can realize a given factor in the moment expansion.","marker":"[31]"},{"why":"Provides the Parseval frame normalization and basic frame identities that underlie the decomposition into weighted measurements.","marker":"[35]"}],"fun_headline_variants":["Restricted measurements stay polylog-close to optimal","Suboptimality ratio immune to ambient dimension","Polylog cost for low-rank projective measurements","Measurements on low-rank subspaces: polylog overhead","Optimality gap shrinks to polylog in subspace dimension"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof rests on the claim that the vectors obtained by projecting a measurement onto the low-rank subspace can always be reordered and phase-shifted so that, when split into equal blocks, their averaged overlap matrix has operator norm at most a constant times $\\log r$; if the combinatorial counting argument behind this estimate fails, the variance control and with it the polylog bound collapse.","fun_headline_variants_meta":{"raw":{"variants":["Restricted measurements stay polylog-close to optimal","Suboptimality ratio immune to ambient dimension","Polylog cost for low-rank projective measurements","Measurements on low-rank subspaces: polylog overhead","Optimality gap shrinks to polylog in subspace dimension"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1303,"prompt_tokens":960,"completion_tokens":343,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":264}},"tokens_in":576,"tokens_out":343,"duration_ms":3451,"temperature":1.0,"reasoning_tokens":264,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:07:33.412781+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Exhibit, for arbitrarily large $r$, a Parseval frame of $S$ for which every permutation and phase choice yields an equipartition whose averaged overlap matrix has operator norm $\\omega(\\log r)$; Proposition 1 would be false and Theorem 1 would not follow from the submitted proof. A concrete route: take the frame with one block equal to the identity ($L_1=1_S$), where the paper's own analysis gives expected operator norm $\\approx \\log r/\\log\\log r$, and check whether the supremum over partitions crosses the $\\log r$ threshold.","supporting_citations":[{"cited_title":"Quantum control by von neumann measurements","cited_arxiv_id":null,"evidence_quote":"Supplies the quantum-control formulation of maximizing $\\operatorname{Tr}[\\rho P[\\tau]]$ over projective measurements, the objective whose suboptimality ratio is studied."},{"cited_title":"Quantum state transformation by optimal projective measurements","cited_arxiv_id":null,"evidence_quote":"Gives first-order stationary conditions for the unitary optimization and proves the exact $K_{n,2}=1$ case for pure states, the baseline the theorem extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the trace-convexity inequality (Lemma 5) used to pass from a biweighted measurement to single weighted measurements in the randomization step."},{"cited_title":"Tropp et al","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian matrix concentration inequality (Lemma 7) that controls the operator norm of the randomized weight operator $\\hat L$ and yields the fourth-moment bound."},{"cited_title":"Balls into bins","cited_arxiv_id":null,"evidence_quote":"Gives the tight balls-into-bins analysis used to argue the near-sharpness of the $\\log r/\\log\\log r$ order in the extreme frame case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Stirling-number counting formula used in Lemma 13 to bound how many permutations can realize a given factor in the moment expansion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Parseval frame normalization and basic frame identities that underlie the decomposition into weighted measurements."}],"review_version":1}