{"id":"3954c681-ccb1-43eb-a21e-5e25be17e37f","arxiv_id":"2501.14080","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A blockwise matrix-sensing measurement design recovers low-rank quantum channels and Lindbladians with near-optimal sample counts and large computational speedups.","lead":"This paper proposes a blockwise measurement scheme that recovers quantum channels and Lindbladians as low-rank matrices from fewer, cheaper measurements. If correct, it offers a unified and much faster route to quantum process tomography for both discrete and open-system dynamics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"First-row blockwise recovery requires Assumption 1 (rank K*_11 = r); without it, physically common low-rank channels such as bit-flip have K*_11 rank < r, and Algorithm 5's inversion of J1^† is underdetermined, so the full-recovery claim is not universal.","rationale":"The paper is honest about Assumption 1 and about the qualitative nature of Remark 3.12, but the abstract and the strongest claim present the blockwise design as a general performance guarantee. The deterministic reconstruction is the only bridge from first-row estimates to full K, and it has no quantified error bound; Assumption 1 is not a harmless technical clause because bit-flip type channels are low-rank, physically standard, and violate it. The reader's verdict already identifies this as the weakest assumption, and I agree. Other gaps, such as the rank mismatch in applying (2r,1/20)-RIP theorems and the mismatch between the analyzed regularized objective and the implemented unregularized ALS, are real proof gaps, but they do not by themselves falsify the measurement-efficiency claim. The load-bearing issue is that the blockwise full-recovery claim requires the first-row rank condition or an explicit row-selection/group-synchronization mechanism; this is exactly what a conditional acceptance should require.","tokens_in":27918,"tokens_out":14995,"duration_ms":137493,"concrete_test":"Analytic/computational test: for the qubit bit-flip channel Kρ = pρ + (1-p)XρX with p = 0.3, compute K = R(mat(K)); verify rank(K) = 2 and rank(K11) = 1. Then generate noiseless first-row data b_{1k} for k = 1, 2 using MO observables satisfying MO ≥ CrN log^6 N, run the first-row blockwise pipeline with Algorithm 5, and compare the reconstructed K to the true K. If the relative Frobenius error is not below 1e-6 because (J1^†)^{-1} in (3.6) is ill-posed, the first-row design fails without Assumption 1. Repeat with row 2 in place of row 1 (K22 has rank 2) to confirm recovery succeeds, isolating the row-choice dependence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Corollary 3.7 only guarantees recovery of the first row K1; the advertised full recovery of all N×N blocks rests on the deterministic reconstruction in Section 3.5. Equation (3.6) requires inverting J1^†, which is only valid under Assumption 1 (rank K*_11 = r). That assumption is not satisfied by physically standard low-rank channels: for the qubit bit-flip channel Kρ = pρ + (1-p)XρX with r = 2, the first diagonal block K*_11 has rank 1, so (3.6) is underdetermined even with exact first-row data and Algorithm 5 cannot recover the remaining rows. The Haar-random justification (Lemma 3.10) applies to Haar-random P*, not to structured physical Vk or Jk, so it does not broaden the scope. Remark 3.12 concedes that the reconstruction error depends on the condition number of J1 and the spectral gap of K1, which are not quantified. Thus the central claim, as stated, is false for an open set of low-rank channels unless Assumption 1 or a row-selection procedure is added.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified low-rank matrix sensing framework for learning quantum superoperators, covering both quantum channels (positive-semidefinite Choi-type reshaped matrices) and Lindbladians (indefinite reshaped matrices). The main contributions are: (i) a random measurement design based on tensor products of Pauli operators, with recovery guarantees derived from the restricted isometry property (RIP) and nonconvex landscape results; (ii) a first-row blockwise measurement design that reduces the full N^2 x N^2 recovery problem to N x N sub-block recovery, combined with parallel or joint ALS algorithms and a deterministic reconstruction step from the first row; and (iii) numerical experiments demonstrating large speedups over full tomography and the random design. The stated sample complexity is M = O(r N^2 log^6 N) measurements for both designs, and the blockwise design is claimed to recover all blocks with the same order of measurements.","tokens_in":28245,"tokens_out":37502,"duration_ms":287646,"significance":"If the guarantees were fully correct, the blockwise design would be a valuable practical contribution to low-rank quantum process tomography and Lindbladian learning, offering order-of-magnitude computational savings relative to full N^2 x N^2 matrix sensing. The paper explicitly compares against the CPT-fit method of [41] and shows competitive recovery rates at a fraction of the runtime. The numerical study is fairly extensive, covering channels and Lindbladians at several system sizes. However, several load-bearing technical gaps affect the printed guarantees: the measurement identity used for the random design is incorrect as stated, the RIP rank parameters in the corollaries do not match the requirements of the cited theorems, the implemented algorithm minimizes an objective different from the one whose landscape is analyzed, and the deterministic block reconstruction relies on a rank assumption that fails for physically common channels. These issues are local and fixable in principle, but they must be addressed before the theoretical claims can be accepted.","major_comments":[{"comment":"The identity (2.4) is false for complex Hermitian ρ0. The proof uses vec(O)vec(ρ0)† = R(ρ0⊗O), but Lemma 2.2 gives R(B⊗C)=vec(C)vec(B)⊤, so R(ρ0⊗O)=vec(O)vec(ρ0)⊤. The correct relation is ⟨ρ1,O⟩ = ⟨K, \\bar ρ0⊗O⟩. A concrete counterexample is the identity channel with ρ0=O=σ_y: the left-hand side of (2.4) is 2, while the right-hand side is -2. Consequently, the random measurement model in Section 3.1, which sets b⋆m0 = tr[(ρm0 ⊗ Om)†K⋆] using (2.4), is inconsistent with the physical measurement tr[O_m^† K(ρ_m0)]. The sensing matrices should be Am = ρm0^⊤ ⊗ Om (or \\bar ρm0 ⊗ Om) rather than ρm0 ⊗ Om. Since the transpose of a Pauli matrix is again a Pauli matrix, the RIP and sample-complexity claims O(rN^2 log^6 N) would survive this correction, but Corollary 3.1 as stated is not valid.","section":"Section 2.1, Proposition 2.7 and Section 3.1"},{"comment":"The corollaries state that the sensing operator satisfies the (r,δ)-RIP with δ ≤ 1/20 and then invoke Theorems 2.1 and 2.2, which require the (2r,1/20)-RIP. An (r,δ)-RIP bound does not imply a (2r,δ)-RIP bound, so the cited theorems do not apply as written. The fix is local: the corollaries should be restated with the (2r,1/20)-RIP. The required sample size from Theorem 2.4 then becomes M ≥ C(2r)d log^6 d = O(rd log^6 d), which is the same order as claimed. This is a technical correction that does not change the scaling conclusions.","section":"Corollaries 3.1, 3.4, 3.7"},{"comment":"The theoretical landscape guarantees (Theorems 2.1 and 2.2) apply to the regularized objective (2.7) with λ = c/8, but the implemented ALS algorithms, including the Nesterov-accelerated version used in Algorithms 3 and 4, minimize the unregularized objective (2.6)/(Q0). Remark 2.14 explicitly acknowledges this omission and cites empirical evidence, but no analysis of the unregularized landscape is provided. As a result, the numerical recoveries are not covered by the paper's theoretical guarantees. The authors should either implement the regularized objective or prove that the unregularized landscape has the same favorable properties.","section":"Section 2.5, Remark 2.14 and Algorithms 1-4"},{"comment":"The deterministic reconstruction from the first row requires Assumption 1 (rank K*_11 = r) in order for (3.6) to invert J1^† and recover all rows. This assumption fails for physically standard low-rank channels. For example, the completely dephasing channel Φ(ρ) = Σ_i |i⟩⟨i| ρ |i⟩⟨i| has Kraus rank r=N and its reshaped Choi matrix is diagonal with N ones, so every diagonal block K*_kk has rank 1 < r. In this case the first-row measurements carry no information about the off-diagonal blocks, and Algorithm 5 cannot recover the full matrix. The paper does state Assumption 1 and Remark 3.11 notes the need for some full-rank diagonal block, but the abstract and introduction do not qualify the blockwise guarantee. The main claims should be restated as conditional on this assumption, or the method should be augmented with a row-selection procedure that is guaranteed to find a full-rank diagonal block (which may not exist for certain channels).","section":"Section 3.4-3.5, Assumption 1 and Algorithm 5"},{"comment":"The paper asserts that Theorems 2.1 and 2.2, originally proved for real-valued matrices, extend to the complex case with the same proof, but no details or reference are given. Since the quantum sensing problem is intrinsically complex, this extension is load-bearing for all theoretical guarantees. The authors should either provide a proof (e.g., via Wirtinger calculus) or cite a published complex version of the landscape theorems.","section":"Remark 2.10"}],"minor_comments":[{"comment":"There is a typo in the statement of the involution property: 'R((R(A))' should read 'R(R(A))'.","section":"Lemma 2.2"},{"comment":"The line '⟨ρ⊤0 ⊗ O†, K†⟩ = ⟨Ekl ⊗ O, K⟩' omits the adjoint on O; it is only correct when O is Hermitian. Please clarify the notation and the assumptions in that derivation.","section":"Section 3.2"},{"comment":"The text says 'r+, r− > 0', but for quantum channels r− = 0. The condition should be r− ≥ 0.","section":"Section 2, Eq. (2.1)"},{"comment":"The performance guarantee for Algorithm 5 is stated in terms of the condition number of J1 and the spectral gap of K1, but these quantities are not bounded under any stated assumptions. The authors should either quantify them or state clearly that the deterministic reconstruction has no finite-sample guarantee.","section":"Remark 3.12"},{"comment":"The numerical experiments use random density matrices and random Hermitian observables from QuTiP, which are not the incoherent Pauli measurements treated in the theoretical sections. The paper acknowledges this in remarks, but the theory-numerics gap should be made explicit in the main text.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper contains interesting ideas and the blockwise measurement design is a potentially useful practical scheme. However, the theoretical sections have several fixable but significant issues: the incorrect measurement identity in Proposition 2.7, the RIP rank mismatch in the corollaries, the gap between the regularized objective analyzed and the unregularized ALS implemented, and the restrictive Assumption 1 that excludes common channels such as complete dephasing. I would encourage the authors to revise with these points in mind; in particular, the abstract and introduction should state the conditional nature of the blockwise guarantee, and the measurement model for the random design must be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth taking seriously. The blockwise first-row measurement design is a real idea: it splits the N^2-by-N^2 superoperator recovery into N low-rank N-by-N sensing problems, and the reported speedups over full ALS and CPT-fit are large and reproducible in structure. The explicit handling of indefinite reshaped operators (VV†−UU†) for Lindbladians is also a genuine extension beyond PSD channel tomography. The numerics are honest, with error bars, timing, and a comparison to a credible baseline.\n\nThe soft spots are real but mostly fixable. Proposition 2.7's equation (2.4) is wrong as printed: for Hermitian ρ0, O, K the correct identity has ρ0^T (equivalently O⊗conj(ρ0)), not ρ0⊗O. The second equality in that proposition is right, so this looks like a typo, but it matters because the measurement model in Section 3.1 uses the wrong sensing matrix. The good news is that the corrected sensing matrices still form an incoherent basis, so the RIP-based theory can survive with a changed constant.\n\nThe corollaries invoke (r,δ)-RIP while the landscape theorems from Ge-Jin-Zheng require (2r,1/20)-RIP. That is a gap in the printed statements, though it only changes constants in the sample complexity. More substantively, the implemented ALS drops the regularizer that the theory analyzes; the authors say so in Remark 2.14. So the global-convergence guarantee does not cover the actual code. The numerics suggest the omission is benign, but that is an empirical claim, not a proven one.\n\nOn the deterministic reconstruction: the stress-test's bit-flip counterexample does not hold up. For a bit-flip channel Kρ = pρ + (1−p)XρX, the first diagonal block K11 is diag(p, 1−p), which has rank 2 = r for generic p. So that example is wrong. The broader concern is valid, though: there are open sets of low-rank channels where the first block has rank < r (e.g., all Kraus operators annihilate the first coordinate), and the authors explicitly concede this. Assumption 1 is a real structural condition, not a generic nicety, and the error of Algorithm 5 is left unquantified. Finally, the Lindbladian story assumes noisy derivative data are already available, which the authors flag honestly.\n\nBottom line: this deserves a serious referee. Fix the Proposition 2.7 typo, align the RIP rank in the corollaries, and either analyze the unregularized ALS or present the regularizer as a heuristic with numerical support. After that, the paper is a solid contribution to quantum process tomography.","headline":"A genuinely useful blockwise measurement design for low-rank quantum superoperator learning, with printed theory that needs several honest fixes before its guarantees can be quoted as-is.","tokens_in":28702,"tokens_out":10571,"would_cite":true,"duration_ms":86720,"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 claims that any low-rank quantum superoperator, including both quantum channels and Lindbladians, can be recovered from O(r N^2 log^6 N) random Pauli measurements, and that a blockwise first-row design achieves the same scaling…","keywords":["quantum process tomography","Lindbladian learning","low-rank matrix sensing","restricted isometry property","Pauli measurements","alternating least squares","quantum channel","blockwise measurement design"],"falsifier":"Take a rank-2 reshaped superoperator $K^*$ that is diagonal in the canonical basis with two nonzero entries, generate noiseless Pauli measurements from it, run the blockwise pipeline (first-row ALS plus Algorithm 5), and check whether the relative Frobenius error fails to reach machine precision while the random-design ALS on the same data succeeds; this would directly demonstrate that Assumption 1, rather than the RIP or the optimizer, is the load-bearing condition.","tokens_in":27691,"feed_emoji":"⚛️","tokens_out":8261,"duration_ms":69578,"temperature":0.7,"pith_summary":"This paper tries to establish that an unknown low-rank quantum superoperator, written as $K\\rho = \\sum_{k=1}^{r_+} V_k \\rho V_k^\\dagger - \\sum_{k=1}^{r_-} U_k \\rho U_k^\\dagger$, can be reconstructed from far fewer measurements than full quantum process tomography. The authors show that roughly $O(r N^2 \\log^6 N)$ random Pauli-type measurements suffice, with explicit recovery guarantees in the presence of noise, and that this scaling holds for quantum channels (where $r_- = 0$) and for Lindbladians (where the reshaped matrix carries negative eigenvalues). They then propose a blockwise design that learns only the first row of $N \\times N$ blocks of the reshaped operator and reconstructs the rest deterministically. A sympathetic reader would care because this unifies two tasks, quantum channel tomography and Lindbladian learning, under one low-rank matrix sensing framework with a tractable optimization landscape.","feed_headline":"Blockwise design learns channels and Lindbladians from O(rN²) data","feed_subtitle":"A first-row block design cuts computation by up to three orders of magnitude while keeping near-optimal sample complexity.","key_machinery":"The load-bearing objects are the reshaping operator $R$, which converts the superoperator into a Hermitian matrix $K$ with possibly negative eigenvalues, and the first-row blockwise sensing map. $R$ maps Kronecker products to rank-one outer products, giving $K = VV^\\dagger - UU^\\dagger$, which is what lets matrix sensing apply to Lindbladians. Measurement is expressed as $\\langle \\rho_1, O \\rangle = \\langle \\rho_0 \\otimes O, K \\rangle$, and the invalid input states $E_{lk}$ are emulated by four valid density matrices. Random Pauli observables make the sensing map satisfy the restricted isometry property, and the landscape theorem imported from the matrix sensing literature then guarantees that all local minima of the regularized objective recover the true matrix. The blockwise design uses the same observables for all initial states, so the stacked operator inherits the RIP, alternating least squares with Nesterov acceleration solves the factored subproblems, and the final deterministic step uses randomized SVD of the first row plus the pseudo-inverse of $J_1^\\dagger$ to complete the full matrix.","core_discovery":"The central claim is that learning a superoperator of the form (2.1) reduces to low-rank matrix sensing of the Hermitian reshaped matrix $K = R(\\mathrm{mat}(K)) = VV^\\dagger - UU^\\dagger \\in \\mathbb{C}^{N^2 \\times N^2}$, which is not positive semidefinite in general. Under random Pauli measurements, the sensing map satisfies a $(2r, 1/20)$-restricted isometry property once $M \\ge C r N^2 \\log^6 N$, and then all local minima of the regularized factored objective equal the true $K$ in the noiseless case, with explicit noisy error bounds. The blockwise design instead isolates the first-row blocks $K_{1k}$: with $M_O \\ge C r N \\log^6 N$ shared Pauli observables and $(3N-2)M_O$ total measurements, the stacked sensing map has the same RIP constant, so joint alternating least squares recovers the entire first row. Algorithm 5 then reconstructs all other block rows from a randomized SVD of the first row and the pseudo-inverse of $J_1^\\dagger$, provided the first diagonal block $K^*_{11}$ has rank $r$. Numerical tests show channel and Lindbladian recovery at comparable accuracy, with the blockwise algorithms running roughly 100 to 1000 times faster than the direct random design at the sizes tested.","pith_inferences":["An implication the paper only gestures at is that the Lindbladian half of the unified claim is not yet an end-to-end experimental protocol, because the paper explicitly sets aside the derivative-estimation step that would supply noisy time-derivative data from actual trajectories.","The deterministic completion step inverts the first diagonal block, so Assumption 1 is also a structural prior on the superoperator; one could test how often structured physical noise models, such as diagonal dephasing or sparse jump operators, violate it, since the Haar-random justification covers generic but not structured operators.","The same blockwise plus matrix-completion pipeline could be applied to other non-Hermitian low-rank linear maps whenever the measurement model gives access to subblocks; the paper's extension to pseudo-unitary group synchronization points in that direction.","A concrete next experiment would run the subset-based joint ALS on data from a real quantum processor with a known Lindbladian noise model, to see whether the measurement-count and runtime advantages survive experimental state-preparation and measurement errors."],"forward_implications":["If the bounds are right, learning a low-Kraus-rank channel or a low-rank Lindbladian requires $O(r N^2 \\log^6 N)$ measurements instead of the $O(N^4)$ of full process tomography.","The blockwise first-row design keeps the same measurement scaling as the random design but reduces each ALS subproblem to size $N \\times N$, enabling parallel execution and cutting runtime by orders of magnitude (roughly 100 times at $N=8$ and 1000 times at $N=16$ in the reported experiments).","Because the recovery guarantees rest on the restricted isometry property rather than positive semidefiniteness, the same algorithm applies to quantum channels ($r_- = 0$) and Lindbladians ($r_- = 1$) with no separate treatment.","In the noisy case the error bounds scale polynomially in the noise level and system dimension, and numerical experiments confirm that the Frobenius error decays as the number of observables grows.","The first-row assumption can be relaxed: any row whose diagonal block has rank $r$ works, and randomly chosen blocks can in principle be handled through group synchronization."],"supporting_citations":[{"why":"Supplies the restricted isometry property for random Pauli measurements, which is the engine behind the near-optimal sample bounds.","marker":"[51]"},{"why":"Supplies the no-spurious-local-minima and noisy recovery landscape theorems used in the paper's main corollaries.","marker":"[25]"},{"why":"Provides the prior guaranteed-recovery quantum process tomography result that this paper extends to non-positive-semidefinite superoperators, and the numerical baseline it compares against.","marker":"[41]"},{"why":"Provides the incoherence condition and matrix completion analogy that motivate Assumption 1 on the first diagonal block.","marker":"[12]"},{"why":"Provides the eigendecomposition trick used to evaluate the superoperator on scaled Pauli inputs using valid density matrices.","marker":"[24]"},{"why":"Provides the Nesterov-accelerated alternating least squares scheme used in the optimization algorithms.","marker":"[55]"},{"why":"Provides the randomized SVD used in the deterministic reconstruction of the full matrix from the first row.","marker":"[54]"}],"fun_headline_variants":["Blockwise design: 1000x faster channel and Lindbladian learning","Near-optimal samples, 1000x speedup: blockwise quantum superoperator learning","Unified blockwise matrix sensing learns channels and Lindbladians fast","RIP-based blockwise design: near-optimal samples, 1000x speedup","Blockwise sensing: 1000x faster, same sample complexity for quantum maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole deterministic completion step presupposes that the first diagonal block of the reshaped superoperator has full rank $r$; if that block is degenerate, for example a low-rank diagonal operator with zero off-diagonal blocks, estimating the first row cannot determine the rest of the matrix and the method breaks down.","fun_headline_variants_meta":{"raw":{"variants":["Blockwise design: 1000x faster channel and Lindbladian learning","Near-optimal samples, 1000x speedup: blockwise quantum superoperator learning","Unified blockwise matrix sensing learns channels and Lindbladians fast","RIP-based blockwise design: near-optimal samples, 1000x speedup","Blockwise sensing: 1000x faster, same sample complexity for quantum maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000722,"raw_usage":{"total_tokens":3270,"prompt_tokens":1006,"completion_tokens":2264,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":2158}},"tokens_in":622,"tokens_out":2264,"duration_ms":14761,"temperature":1.0,"reasoning_tokens":2158,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:25:00.077911+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a rank-2 reshaped superoperator $K^*$ that is diagonal in the canonical basis with two nonzero entries, generate noiseless Pauli measurements from it, run the blockwise pipeline (first-row ALS plus Algorithm 5), and check whether the relative Frobenius error fails to reach machine precision while the random-design ALS on the same data succeeds; this would directly demonstrate that Assumption 1, rather than the RIP or the optimizer, is the load-bearing condition.","supporting_citations":[{"cited_title":"No spurious local minima in nonconvex low rank problems: A unified geometric analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the no-spurious-local-minima and noisy recovery landscape theorems used in the paper's main corollaries."},{"cited_title":"Universal low-rank matrix recovery from Pauli measurements","cited_arxiv_id":null,"evidence_quote":"Supplies the restricted isometry property for random Pauli measurements, which is the engine behind the near-optimal sample bounds."},{"cited_title":"Guaranteed recovery of quantum pro- cesses from few measurements","cited_arxiv_id":null,"evidence_quote":"Provides the prior guaranteed-recovery quantum process tomography result that this paper extends to non-positive-semidefinite superoperators, and the numerical baseline it compares against."},{"cited_title":"Exact matrix completion via convex optimization","cited_arxiv_id":null,"evidence_quote":"Provides the incoherence condition and matrix completion analogy that motivate Assumption 1 on the first diagonal block."},{"cited_title":"Quantum tomography via compressed sensing: Error bounds, sample complexity and efficient estimators","cited_arxiv_id":null,"evidence_quote":"Provides the eigendecomposition trick used to evaluate the superoperator on scaled Pauli inputs using valid density matrices."},{"cited_title":"Nesterov acceleration of alternating least squares for canonical tensor decomposition: Momentum step size selection and restart mechanisms","cited_arxiv_id":null,"evidence_quote":"Provides the Nesterov-accelerated alternating least squares scheme used in the optimization algorithms."},{"cited_title":"A randomized algorithm for the decom- position of matrices","cited_arxiv_id":null,"evidence_quote":"Provides the randomized SVD used in the deterministic reconstruction of the full matrix from the first row."}],"review_version":1}