{"id":"8cf2df98-f5fa-4584-acba-f42616ecab84","arxiv_id":"2607.24491","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"α-gentle tomography of rank-r qudits has minimax Frobenius rate Θ(rd²/(nα²)), with gentleness penalty scaling as ambient dimension d rather than parameter count rd.","lead":"Gentle quantum measurements that barely disturb a d-dimensional state still let you reconstruct it, but cost a factor d/α² in sample complexity versus ordinary tomography. The penalty tracks Hilbert-space dimension, not the number of free parameters—unlike classical differential privacy.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The lower-bound machinery (Prop. 5 → Thm. 11 → Thm. 10) only certifies the rd²/(nα²) rate for α bounded away from 1/2; the regime α∈[1/2,1), where the claim must continuously reduce to the non-gentle product-measurement rate, is never argued in the paper.","rationale":"I read the upper-bound and lower-bound arguments in detail before settling on this concern. The core steps check out: in Thm. 11 the identity Tr[(E_ω/λ_min−I)Δ]=Tr[(E_ω/λ_min)Δ] is valid because Δ=ρ+−ρ− is traceless; A_ω=(E_ω/λ_min−I)/(e^δ−1) has op-norm ≤1 by the eigenvalue-ratio bound; Σ_ω λ_min(E_ω)≤Tr(I)/d=1; and (e^δ−1)²=64α²/(1−2α)⁴ is arithmetically correct. The full-rank packing (Lemma 5–6: positivity for ε≤1/√d via ∥Σν_iV_i∥_op≤√D; Bessel bound Σ_jTr[AV_j]²≤∥A∥²_F≤d) and the low-rank su(d) packing (Lemma 7–8: near-linearity of unitary perturbations, ∥A_12∥²_F≤r/4 giving the 4ε²/r main term) are internally consistent and yield the claimed rd²/(nα²) up to the stated φ(r,d) saturation terms. The upper bound's matrix-Bernstein concentration, the tail-integration (splitting at x(ε) so the log(d/ε) factor survives without an extra d), and the thresholding rank argument are sound up to harmless constant discrepancies (t² stated as 92·log(d/ε)d²/(nα²) in Thm. 8 but derived as 16·…; Prop. 9's threshold drops a square on κα — typo-level). So there is no derivation gap in the proved regime. The single structural weakness is the α<1/2 restriction inherited from Prop. 5, which is also what the reader flagged; I agree it is the load-bearing assumption. It does not overturn the verdict: for every fixed α<1/2 the rate is proved, the regime α∈[1/2,1) is almost certainly patchable via known non-gentle product-measurement lower bounds, and the headline phenomenon (penalty d/α² scaling with ambient dimension) is supported by both sides. Hence ACCEPT should stand, with the paper encouraged to add the one-paragraph monotonicity/non-gentle-bound argument for α≥1/2.","tokens_in":60472,"tokens_out":14660,"duration_ms":1070974,"concrete_test":"Close the regime gap analytically, in two parts. (i) Re-derive Theorem 11's data-processing inequality directly from α-gentleness without routing through Proposition 5's qDP conversion, and check whether the restriction α<1/2 and the factor (1−2α)^4 are removable; if the qDP detour is essential, confirm by constructing (or ruling out) an α-gentle measurement with α≥1/2 whose outcome likelihood ratios are unbounded, which would show the restriction is intrinsic. (ii) Independently, verify the claim in the gap regime: prove that for α∈[1/2,1) the minimax risk over locally-α-gentle product measurements is ≳ rd²/n, most simply by establishing a non-gentle product-measurement lower bound of order rd²/n for squared Frobenius loss (adapting the packing argument of [26]; note trace-norm lower bounds alone yield only dr/n via ∥X∥²_Tr ≤ r∥X∥²_F). If (ii) succeeds, the rate claim is complete for所有α","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the minimax rate under locally α-gentle product measurements is of order rd²/(nα²). The upper bound (Thm. 8) holds uniformly for α∈(0,1) with α-independent constant. The matching lower bound (Thm. 10) rests on the data-processing inequality of Thm. 11, which converts α-gentleness into δ-quantum-differential-privacy via Proposition 5 with δ=2log((1+2α)/(1−2α)). That conversion provably requires α<1/2 (the proof's denominator λ−λ²−2αλ must stay positive), and it injects the factor (1−2α)⁻⁴ into every lower bound. Two consequences are not addressed in the paper:\n\n(1) For α∈[1/2,1) no lower bound is proved at all. Monotonicity does not help: the α-gentle class grows with α, so the infimum risk R(α) is non-increasing and the fixed-α₀<1/2 bound gives only an upper bound on R(α). The claim in this regime reduces to the statement that the non-gentle (α=1, i.e. unconstrained) product-measurement Frobenius² minimax rate is rd²/n, but the paper never invokes or proves such a bound; the cited trace-norm lower bound dr²/n of [26] converts via ∥X∥²_Tr≤r∥X∥²_F to only dr/n, which at α=1 would leave the estimator of Thm. 8 (rate rd²/n) short of proved optimality by a factor r.\n\n(2) Even as α→1/2⁻, the (1−2α)⁴ constant means the sharp α-dependence of the rate is established only for α bounded away from 1/2 by a fixed constant; the paper does not exclude a different interpolation of the gentleness penalty near the breakdown of the qDP conversion.\n\nThis is precisely the soft spot the reader identified; my reading confirms it is the genuinely load-bearing one rather than a cosmetic constant issue, because the restriction is structural to the proof technique, not an artifact of bookkeeping.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper studies estimation of a d-dimensional, rank-r quantum state from n separately measured copies when each measurement is locally α-gentle, meaning that every post-measurement state is within trace distance α of the input state. The authors construct a quantum label-switch measurement by gently randomizing each basis in a complete set of mutually unbiased bases, extend it using mutually unbiased measurements, and analyze an unbiased linear estimator followed by projection onto the state space and spectral thresholding. They obtain a Frobenius mean-squared-error upper bound of order rd² log(d/ε)/(nα²), together with a rank-consistency guarantee. Lower bounds are developed through an Assouad reduction and a new data-processing inequality for locally gentle measurements, with separate packings for full-rank, low-rank, and pure states, giving a matching rate up to a logarithmic factor when α is bounded away from 1/2. The paper also treats gentle estimation of the probability vector of a pure state and gives an ancilla-and-controlled-gate implementation of the basic label-switch mechanism.","tokens_in":61009,"tokens_out":9011,"duration_ms":338253,"significance":"If the stated optimality holds over the claimed parameter range, the paper provides a sharp quantification of the information–disturbance trade-off in high-dimensional quantum tomography. The ambient-dimensional gentleness penalty d/α², rather than a penalty proportional to the number of parameters rd, is a notable contrast with classical local differential privacy. The manuscript has substantial strengths: an explicit unbiased estimator; implementable gentle label-switch measurements; MUB/MUM constructions; matrix-concentration analysis with projection and rank-consistent spectral thresholding; and a new gentle data-processing inequality combined with full-rank, low-rank, and pure-state packings. These ideas should be useful beyond tomography. The significance is currently tempered by the two scope gaps identified below.","major_comments":[{"comment":"The advertised optimality is not established over the full range α∈(0,1). Proposition 5 requires α<1/2, and Theorem 11 inherits this restriction; Theorem 10 therefore assumes α bounded away from 1/2 and its constants contain (1−2α)^4. For α∈[1/2,1) no matching lower bound is given, and monotonicity only implies that the risk is no larger than at a smaller α. At α=1, where all product measurements are allowed, the claimed rate is rd²/n, but the manuscript does not prove or identify a matching Frobenius-loss product-measurement lower bound; the cited trace-norm lower bound [26] does not yield this by norm equivalence alone. Thus the Abstract's unqualified optimal rate and sharp 1/α² dependence are currently proved only for α≤1/2−c. Please either restrict the main claims accordingly or provide a lower-bound argument covering α∈[1/2,1], and clarify the behavior as α↑1/2.","section":"Theorem 10; Proposition 5; Theorem 11; Abstract"},{"comment":"Theorem 8 uses a complete set of MUBs, which is guaranteed here only in prime-power dimensions. For arbitrary d, Proposition 9 uses MUMs but the upper bound becomes rd²/(nα²κ²). The text says the optimal rate extends when κ∼1, yet no construction or cited theorem is given ensuring complete MUMs with efficiency κ bounded below by a dimension-independent constant. If κ(d) tends to zero, the displayed all-d rate is not attained by the proposed estimator. Please supply a dimension-uniform lower bound on κ for a concrete MUM family, restrict the main theorem to dimensions where this is known, or state the additional κ-dependent penalty in the main claims.","section":"§5.3, Proposition 9"}],"minor_comments":[{"comment":"The displayed threshold t²(ϵ)=92 log(d/ϵ)d²/(κnα) appears to be missing squares. Appendix D says to replace α by ακ in Theorem 8, which would give t²(ϵ) proportional to d²/(nα²κ²), not d²/(κnα).","section":"Proposition 9"},{"comment":"With σ²≤10d²/(nα²) and R≤3d/(nα), the stated Matrix-Bernstein theorem appears to give an exponent −t²nα²/(24d²) (for t≤2), rather than the displayed −t²nα²/(4d²). This does not affect the rate, but the constants should be reconciled. The threshold constant also changes between Theorem 8 (92) and Proposition 14 (16).","section":"Appendix C, Eqs. (37)–(38)"},{"comment":"The phrase “α bounded away from 1/2” is ambiguous because it can include α>1/2. Since Theorem 11 applies only for α<1/2, write α∈(0,1/2−c] explicitly.","section":"Theorem 10"},{"comment":"Section 4.4 gives an ancilla-plus-CNOT implementation for the probability-vector measurement. If the same physical implementability is claimed for the full MUB/MUM tomography measurement, add a sentence or proposition explaining the reduction via a basis change before the label-switch circuit.","section":"§4.4; Abstract"},{"comment":"There are several typographical/grammatical errors, including “proof is allows,” “a again,” “standing state tomography,” “thought” for “through,” “finish of,” “chosed,” and “meausrements.” In Proposition 9, Pρ(rank(ρ̂n=r) is also missing a closing parenthesis.","section":"General"},{"comment":"The notation ρ_{±j,i}, ρ_{ν,i}, and ρ_{ν^{j±}} is quite close and denotes different averaged and padded states. A short displayed reminder before Theorem 11 would make the lower-bound chain easier to verify.","section":"Definition 7 and Theorem 11"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The headline result is clean: under local α-gentleness the Frobenius minimax rate for rank-r states is order rd²/(nα²), so the gentleness penalty is d/α² and tracks ambient dimension, not the parameter count rd. That is genuinely different from classical local DP and is the thing worth remembering.\n\nWhat they do well is complete. They build an explicit MUB-wise (or MUM-wise) gentle label-switch measurement, give a physical ancilla+CNOT realization, run a projected least-squares plus spectral thresholding estimator, and get matching upper bounds via Matrix-Bernstein. The lower-bound side is new machinery: Assouad on carefully packed full-rank, low-rank and pure families plus a quantum DPI that controls sums of symmetrized KLs under gentleness. Appendices are written out; constants are tracked; self-citations are to their own prior gentle-qubit and certification tools, not circular. The dimensional scaling appears on both sides of the argument, so it is not an artifact of one construction.\n\nThe soft spot the stress-test flags is real and structural, not cosmetic. Proposition 5 converts gentleness to qDP only for α<1/2, and Theorem 11 therefore injects (1−2α)⁻⁴. Theorem 10 explicitly assumes α bounded away from 1/2. For α∈[1/2,1) there is no lower bound, and monotonicity does not close the gap to the unconstrained product-measurement rate. Near α=1/2 the sharp α-dependence is also unproved. Upper bounds hold for all α∈(0,1), so the estimator is fine; only optimality is incomplete in that regime. There is also the usual log(d) gap and the prime-power caveat for MUBs (handled by MUMs when κ∼1). None of these sink the paper.\n\nThis is for people who work on quantum statistics, gentle measurements, or the interface with differential privacy. A serious referee should see it. I would bring it to reading group, cite the rate and the DPI, and expect acceptance after the α-range is either extended or clearly scoped.","headline":"Solid minimax theory for gentle tomography with a real dimensional surprise, but the matching lower bound only holds for α bounded away from 1/2.","tokens_in":60236,"tokens_out":543,"would_cite":true,"duration_ms":19372,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62C20","81P50","62F12"],"pacs":[],"model":"grok-4.5","headline":"Locally gentle measurements raise the minimax cost of estimating a rank-r quantum state to order rd²/(nα²), with a gentleness penalty that scales only with ambient dimension d.","keywords":["quantum state tomography","gentle measurements","local differential privacy","minimax rates","mutually unbiased bases","Frobenius norm","quantum data-processing inequality","low-rank density matrices"],"falsifier":"Exhibit either a locally α-gentle product measurement and estimator whose Frobenius risk is o(rd²/(nα²)) for a sequence of rank-r states, or a family of rank-r states on which every such measurement incurs risk ω(rd²/(nα²)), for α bounded away from 1/2.","tokens_in":59880,"feed_emoji":"⚛️","tokens_out":1224,"duration_ms":29356,"temperature":0.7,"pith_summary":"This paper asks how much harder it becomes to learn a d-dimensional quantum state when every measurement is forced to disturb the state by at most a fixed amount α. Without that constraint the best Frobenius risk scales like rd/n for rank-r states; with local gentleness the optimal rate becomes rd²/(nα²). The extra factor d/α² is surprising because classical local differential privacy typically multiplies risk by the full number of parameters, here order rd. The authors construct explicit gentle measurements—built from mutually unbiased bases and a quantum label-switch kernel—that attain the new rate, show they can be realized by entangling the system with an ancillary register via controlled-NOT gates, and prove matching lower bounds via a new quantum data-processing inequality on carefully chosen families of states. The result quantifies the unavoidable information–disturbance trade-off for high-dimensional quantum tomography and shows that the geometry of quantum state space can be kinder to privacy than its classical counterpart.","feed_headline":"Gentle quantum tomography costs an extra factor d/α²","feed_subtitle":"The privacy penalty tracks Hilbert-space dimension, not the number of free parameters—unlike classical local DP.","key_machinery":"A quantum Assouad reduction that converts estimation into simultaneous testing of 2^D local alternatives, controlled by a new data-processing inequality: the sum of symmetrized KL divergences of outcome laws after any locally α-gentle product measurement is bounded by a multiple of nα² times a sum of squared trace distances between neighboring states. Matching upper bounds are obtained from gentle-ized mutually unbiased bases (or MUMs) followed by projected least-squares and spectral thresholding.","core_discovery":"Under the constraint that each of n independent measurements may change the unknown state by at most α in trace distance, the minimax risk of estimating a rank-r density matrix on ℂ^d in Frobenius norm is of order rd²/(nα²) (up to logs). The same rate holds for the full-rank case r = d, giving d³/(nα²). Thus the statistical price of gentleness is a multiplicative factor d/α² that depends on the ambient Hilbert-space dimension rather than on the intrinsic number of free parameters rd.","pith_inferences":["The ambient-dimension penalty suggests that other quantum estimation tasks with local gentleness (e.g., shadow tomography or entanglement witnessing) may likewise escape the classical parameter-count barrier.","Because the lower-bound construction works for any derived parameter that separates under Hamming distance on the same hypercube, the framework immediately yields gentleness lower bounds for many linear functionals of ρ.","If the α-away-from-1/2 restriction can be removed, the same rates would describe the entire gentle regime down to nearly non-disturbing measurements."],"forward_implications":["Any quantum algorithm that re-uses post-measurement states under a strict disturbance budget of α must pay at least a d/α² sample overhead for full tomography.","The same gentle MUB construction yields an optimal estimator of the probability vector of a pure state with risk of order d/(nα²).","Classical local DP mechanisms can be lifted to quantum measurements via Naimark dilation with an ancillary register and controlled-NOT gates while preserving both privacy and gentleness.","When a complete set of MUMs with efficiency κ ~ 1 exists, the same rate holds in every finite dimension, not only prime-power dimensions."],"fun_headline_variants":["Gentle quantum tomography pays extra d/α² factor","Minimax rate for gentle state estimation hits rd²/(nα²)","Gentleness multiplies tomography risk by d/α² not rd","Optimal gentle measurements cost Hilbert-dimension factor","Quantum estimation under α-gentleness scales as d³/(nα²)"],"cache_read_input_tokens":49280,"weakest_assumption_plain":"The matching lower bound requires the gentleness parameter α to stay a fixed distance below one-half, so that the constants that blow up as α approaches one-half remain bounded.","fun_headline_variants_meta":{"raw":{"variants":["Gentle quantum tomography pays extra d/α² factor","Minimax rate for gentle state estimation hits rd²/(nα²)","Gentleness multiplies tomography risk by d/α² not rd","Optimal gentle measurements cost Hilbert-dimension factor","Quantum estimation under α-gentleness scales as d³/(nα²)"]},"model":"grok-4.5","effort":"low","cost_usd":0.004356,"raw_usage":{"total_tokens":1341,"prompt_tokens":868,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":43564000,"prompt_tokens_details":{"text_tokens":868,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":402,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":868,"tokens_out":71,"duration_ms":8024,"temperature":1.0,"reasoning_tokens":402,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T13:19:43.110905+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Exhibit either a locally α-gentle product measurement and estimator whose Frobenius risk is o(rd²/(nα²)) for a sequence of rank-r states, or a family of rank-r states on which every such measurement incurs risk ω(rd²/(nα²)), for α bounded away from 1/2.","supporting_citations":[],"review_version":1}