{"id":"0ae6bbde-8ced-4693-ac7d-67ca2e8ea774","arxiv_id":"2602.15202","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Given enough overlapping principal submatrices of a low-rank density matrix, the full matrix can be recovered algebraically via subspace intersection and least squares, using only O(RD) measurement settings.","lead":"This paper proposes an algebraic method to reconstruct an unknown low-rank quantum state from carefully chosen overlapping blocks of its density-matrix entries, using linear algebra instead of expensive optimization. If it works in practice, it offers a fast, deterministic alternative to compressed-sensing tomography for quantum devices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Recovery guarantee rests on an unverified isorank/generic-intersection condition; a simple rank-1 state with support (1,0,1) satisfies isorank but makes the Q_l intersection 2-dimensional, so Algorithm 1's top-R singular vector is not the true column space.","rationale":"The reader's weakest assumption correctly identifies the least secure condition: the isorank and generic-intersection properties are not implied by low rank, are not verified by the algorithm, and can fail for the very structured states the introduction highlights. A concrete low-dimensional counterexample shows the failure can occur even when the observed principal submatrices are isorank, so the problem is not merely rank deficiency. The current CONDITIONAL verdict is therefore appropriate. I did not find an independent fatal flaw that would require rejection; the algebraic idea is sound for generic states, and the numerical demonstration supports it. The secondary issue of the Frobenius-norm identity in the error analysis also points to conditional acceptance but is not the primary collapse point for the recovery claim.","tokens_in":8995,"tokens_out":15727,"duration_ms":162170,"concrete_test":"Run Algorithm 1 with noiseless measurements on the rank-1 state ρ=|v><v| with v=(1,0,1), D=3, R=1, and overlapping blocks r1={1,2}, r2={2,3}. Compute the two principal submatrices (each rank 1), form Q^(1), Q^(2) as in Algorithm 1 line 4, and take the top singular vector of [Q^(1), Q^(2)]. If the result is not v/||v|| (up to phase), the generic-intersection assumption is load-bearing and fails on a simple structured state. Repeat with ρ=diag(1/2,0,0,1/2) on D=4, R=2, blocks {1,2,3} and {2,3,4}, to observe rank-deficient blocks and reconstruction failure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central recovery equality col(ρ) = ∩_l Q_l (Sec. III-B) requires two unverified state-dependent properties: every selected principal submatrix is isorank, and the subspaces Q_l intersect in exactly R dimensions (Sec. III-D). Merely knowing rank(ρ)=R does not imply either. For non-generic but physically motivated states these conditions can fail even when all observed blocks have rank R. Example: D=3, R=1, ρ=|v><v| with v∝(1,0,1), blocks r1={1,2}, r2={2,3}. Both 2×2 principal submatrices have rank 1, so isorank holds, yet Q1=Q2=span(e1,e3); the intersection is 2-dimensional and the top singular vector of [Q^(1),Q^(2)] is not col(ρ). For a diagonal state such as ρ=diag(1/2,0,0,1/2), a 3×3 block can have rank 1<R, which is worse. The paper provides no data-driven certificate for isorank/genericity; in the noisy setting, rank tests are unreliable. The numerical experiment draws Ginibre random states, so it cannot expose this failure. Thus the deterministic guarantee is conditional on a property of the unknown state and is not checked or certified by the algorithm.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an algebraic method for low-rank quantum state tomography. Given a selection pattern of principal submatrices of the density matrix, the algorithm computes the column space of ρ as the intersection of subspaces induced by isorank submatrices: each submatrix is diagonalized, its top-R eigenvectors are padded to the full space, and the top-R left singular vectors of the concatenated padded basis are taken as an estimate of col(ρ). The state is then reconstructed by least squares. The authors claim deterministic recovery guarantees for generic low-rank states, a noisy subspace error bound, O(RD) measurement complexity, and numerical superiority over nuclear-norm minimization and Burer-Monteiro factorization on a five-qubit Ginibre-ensemble example.","tokens_in":9379,"tokens_out":12730,"duration_ms":126745,"significance":"If the recovery guarantee and error analysis were rigorous, the algebraic approach would be a useful contribution: it replaces convex optimization by standard numerical linear algebra, has a simple closed-form reconstruction, and its structured measurement count is competitive with compressed-sensing QST. The paper also sensibly draws on the existing algebraic matrix-completion literature rather than re-deriving it. However, the central recovery condition is state-dependent and is only asserted as 'generic', and the noisy perturbation bound contains a false norm identity and an unquantified spectral quantity. As it stands, the significance of the claimed deterministic guarantee is not established to the standard required for publication.","major_comments":[{"comment":"The central recovery equality col(ρ)=∩_l Q_l is asserted under 'generic conditions' cited from [25,27,28], but the two state-dependent hypotheses—(i) every measured principal submatrix is isorank and (ii) dim ∩_l Q_l = R—are neither derived from rank(ρ)=R nor checked by Algorithm 1. They can fail for physically meaningful states even when all observed blocks have rank R. Example: D=3, R=1, ρ=|v><v| with v=(1,0,1)^T/√2, and blocks r1={1,2}, r2={2,3}. Both 2×2 principal submatrices have rank 1, the blocks share one row and cover [3], so the stated pattern conditions are met, yet Q1=Q2=span{e1,e3}, the intersection has dimension 2, and the top singular vector of [Q^(1),Q^(2)] is not col(ρ). The wording in III-D ('This condition is generically satisfied when ...') is too weak to support the subsequent 'deterministic' recovery claim; the hypotheses need to be stated as explicit assumptions an","section":"Sec. III-B and III-D"},{"comment":"The displayed identity ∥P̃_tot − P_tot∥_F^2 = Σ_l ∥P̃_Ql − P_Ql∥_F^2 is incorrect. From the definition P_tot = Σ_l Q^(l)Q^(l)H, the difference is a sum, so the squared Frobenius norm contains cross terms Σ_{l≠k} Tr((P̃_Ql − P_Ql)(P̃_Qk − P_Qk)); these are not generally zero. Consequently the stated noise accumulation bound is unproved. This is load-bearing because the paper's deterministic subspace-error guarantee is one of its central claims. The authors should replace this step with a valid perturbation analysis, for example by bounding the cross terms directly or by treating the sum as a single Hermitian matrix and using a sin-Θ theorem for the top-R invariant subspace of P̃_tot.","section":"Sec. III-B, Error Analysis"},{"comment":"The final chordal-distance bound divides by σ_min(P̃_tot), but no lower bound on this quantity is provided or related to the local gap δ, L, or the overlap pattern. Moreover, P_tot = Σ_l Q^(l)Q^(l)H is a sum of projectors and is not itself an orthogonal projection (unless L=1), so the invocation of [33, Thm. 1.1]—a perturbation bound for orthogonal projections—does not apply as written. Without an explicit spectral-gap assumption on P̃_tot and a statement of the theorem's actual hypotheses, the displayed guarantee can be vacuous.","section":"Sec. III-B, final bound"}],"minor_comments":[{"comment":"The typesetting `[S_rl U^(l) S_rc_l]` should read `[S_{r_l}U^{(l)}, S_{r_l^c}]`; as printed it looks like a matrix product rather than a concatenation.","section":"Algorithm 1, line 4"},{"comment":"Introduce σ_min(·) explicitly, and clarify that P_tot is called an 'aggregate projector' but is not an orthogonal projection matrix in general.","section":"Notation"},{"comment":"The experiments report medians over 20 trials without error bars, and only one system size, one rank, and one SNR are tested. The Ginibre ensemble satisfies the generic assumptions almost surely, so the experiment cannot detect the failure mode illustrated in the major comment above; a structured-state example would be informative.","section":"Sec. IV"},{"comment":"The shadow-tomography discussion is a short remark with no analysis or experiments. If it is intended as a contribution, more detail is needed; otherwise it should be labeled as outlook.","section":"Sec. III-E"},{"comment":"The measurement-complexity formula M = (R+d)^2 + (L−1)((R+d)^2−R^2) counts real parameters of the union of Hermitian blocks; the text should state this explicitly and define L in terms of D, R, and d for the pattern in Fig. 1.","section":"Sec. III-D"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper leans heavily on the authors' own previous work for the algebraic completion conditions, but that is not circular; the genuinely new content is the QST-specific algorithm, error analysis, and complexity count. I would not reject on self-citation grounds. The main blockers are the invalid Frobenius-norm step and the unquantified σ_min term in the error bound, together with the need to state the generic recovery condition as an explicit, testable hypothesis. If those can be repaired, the paper could be suitable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper applies an existing algebraic matrix-completion idea to low-rank QST: measure a few fully observed principal submatrices, compute local column spaces, intersect them to get the global column space, then fill in the rest by least squares. That combination is genuinely new, and the O(RD) measurement count for rank-R D×D states is attractive. The experiments on random Ginibre states show the method is fast and accurate against CVX and Burer–Monteiro baselines.\n\nThe strength is that the noiseless recovery logic follows from prior work on algebraic completion. If the local subspaces intersect in exactly R dimensions, then the column space is identified, and the rest is standard linear algebra. That part is solid for generic states.\n\nNow the soft spots. First, the deterministic guarantee is not what the abstract promises. The condition that the intersection of the Q_l subspaces has dimension R is not implied by the isorank condition or by rank ρ = R. A simple counterexample: D=3, R=1, ρ = |v><v| with v ∝ (1,0,1). The two principal submatrices on {1,2} and {2,3} are both rank 1, but the padded subspaces Q1 = Q2 = span(e1,e3), so their intersection is 2-dimensional, and Algorithm 1's top singular vector of Q_tot is not col(ρ). The paper calls the condition 'generic' and cites prior papers; that is fine as a formal statement, but the algorithm never checks it, and the experiment samples only Ginibre states, so the failure mode is invisible. For a QST method, the user cannot know in advance whether the state is generic.\n\nSecond, the noisy error bound in Section III-B is wrong as written: it asserts ||P̃_tot − P_tot||_F² = Σ_l ||P̃_Ql − P_Ql||_F². That identity does not hold because cross terms between different l are dropped. The final bound also depends on σ_min(P̃_tot), which is unquantified and can be near zero when the intersection condition fails.\n\nMinor issues: no code or data, the physicality of the least-squares reconstruction is not pinned down, and the measurement operators for off-diagonal entries may need nonlocal circuits. These are addressable.\n\nOverall: the core idea is worth refereeing. The paper is a good starting point, but the deterministic guarantee needs to be scoped to generic states, the error bound needs fixing, and the experiment should include non-generic states. I'd send it to review and ask for a major revision.","headline":"A clean idea with a real proof gap: the recovery guarantee fails for a simple non-generic state, and the error bound drops cross terms.","tokens_in":9828,"tokens_out":5260,"would_cite":false,"duration_ms":42296,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A rank-R density matrix is uniquely fixed by a few of its principal submatrices, and the full state then follows from least squares.","keywords":["quantum state tomography","low-rank density matrix","matrix completion","subspace estimation","isorank","algebraic reconstruction","classical shadows","deterministic recovery"],"falsifier":"Take the noiseless rank-2 state ρ = 1/2(|00⟩⟨00| + |01⟩⟨01|) on two qubits and use the overlapping block pattern with blocks {00,01,10} and {01,10,11} (R=2, overlap two indices). The second block has matrix diag(0.5,0,0), rank 1, violating isorank, so the subspace intersection computed from the two blocks fails to recover the |00⟩ component and the algorithm's output does not equal ρ despite zero noise.","tokens_in":8920,"feed_emoji":"⚛️","tokens_out":5138,"duration_ms":50891,"temperature":0.7,"pith_summary":"The paper aims to show that full quantum state tomography for low-rank states can be done deterministically by measuring a structured set of principal submatrices, not by solving a costly convex program. If the chosen submatrices each have the same rank as the full state, their padded column spaces contain the state's column space, and when their intersection has exactly R dimensions the column space is recovered exactly. An orthonormal basis for that intersection is obtained as the top-R left singular vectors of a concatenated matrix, after which every unknown entry is filled by least-squares regression. This yields O(RD) measurement settings and a reconstruction based only on standard linear algebra, with an explicit bound on how noise in local measurements propagates to the global subspace estimate.","feed_headline":"One subspace trick cuts quantum tomography to O(RD) measurements","feed_subtitle":"Measure just a few overlapping blocks of a low-rank density matrix; the full state then falls out of linear algebra.","key_machinery":"The load-bearing object is the isorank principal submatrix: a square submatrix on an index set r_l whose rank equals the rank R of the full density matrix. Because ρ_obs^(l) = U_l Σ_l U_l^H, padding its top-R eigenvectors back into the D-dimensional space gives a subspace Q_l that contains col(ρ). Intersecting these local subspaces across blocks recovers the global column space exactly when the intersection has dimension R; the intersection is computed by the closed-form solution of minimizing the sum of squared chordal distances to the Q_l, namely the top-R left singular vectors of the concatenated basis matrix Q_tot. This machinery converts a tomographic inversion problem into two standard","core_discovery":"The central claim is that for a rank-R density matrix ρ, full recovery is guaranteed from fully observed isorank principal submatrices whose padded eigen-subspaces Q_l satisfy dim(∩_l Q_l) = R. Under this condition, col(ρ) equals the intersection ∩_l Q_l, and a basis can be computed in closed form as the top-R left singular vectors of Q_tot = [Q^(1), ..., Q^(L)]. The paper proves the global subspace error obeys d_c(Ũ,U) ≤ ε sqrt(2 Σ_l |r_l|)/(δ σ_min(P̃_tot)) under a spectral-gap assumption, and shows that a simple overlapping block-diagonal pattern with pairwise overlap at least R and full coordinate coverage satisfies the required generic conditions. From the recovered column space, each c","pith_inferences":["If the isorank condition is not verified in practice, a state whose support is aligned with the measurement blocks (e.g., a superposition supported on only one block) can silently violate rank(ρ_obs^(l)) = R; one could detect this by comparing the block ranks to the assumed R and adaptively enlarging blocks.","The subspace-recovery formulation suggests an incremental or streaming tomography protocol: blocks can be measured one at a time and the global column space updated by adjusting the SVD of Q_tot, which may be useful when measurement settings are scheduled adaptively.","The same intersection-of-subspaces principle extends naturally to tensor-train or matrix-product-state completions from fiberwise observations, where the local subspaces come from matricized blocks rather than principal submatrices.","The error bound invites a design heuristic: choose the overlap d that minimizes the noise-amplification term σ_min(P̃_tot) for a fixed measurement budget, rather than the minimal d=1 pattern."],"forward_implications":["Selective QST with the proposed overlapping block pattern needs O(RD) measurement settings, achieving deterministic (non-probabilistic) recovery for generic low-rank mixed states.","The reconstruction is purely algebraic — EVD, SVD, and least squares — so it avoids the computational cost of nuclear-norm SDPs and non-convex factorization methods.","The error bound d_c(Ũ,U) ≤ ε√(2Σ|r_l|)/(δσ_min(P̃_tot)) shows that measurement noise is amplified by a factor controlled by the local spectral gap δ and the conditioning of the aggregate projector; larger overlaps reduce the error.","The same structured-entry data can be supplied by classical shadow estimation instead of per-entry measurements, combining low-rank recovery with shadow tomography's logarithmic measurement complexity per observable.","For rank-1 states the measurement count reduces to 3D−2, matching known directly-complete pure-state measurements."],"fun_headline_variants":["Algebraic quantum tomography: fewer measurements, exact recovery","Low-rank quantum states recovered with linear algebra alone","Matrix completion slashes quantum tomography measurements","Deterministic recovery for low-rank quantum states","Quantum state tomography cut to O(RD) via closed-form linear algebra"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire recovery guarantee rests on every measured principal submatrix having rank exactly R (the isorank property) and on the generic overlap conditions holding; the paper offers no practical test to confirm these properties for the unknown state.","fun_headline_variants_meta":{"raw":{"variants":["Algebraic quantum tomography: fewer measurements, exact recovery","Low-rank quantum states recovered with linear algebra alone","Matrix completion slashes quantum tomography measurements","Deterministic recovery for low-rank quantum states","Quantum state tomography cut to O(RD) via closed-form linear algebra"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1126,"prompt_tokens":615,"completion_tokens":511,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":359,"completion_tokens_details":{"reasoning_tokens":436}},"tokens_in":359,"tokens_out":511,"duration_ms":6332,"temperature":1.0,"reasoning_tokens":436,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T22:55:48.965930+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the noiseless rank-2 state ρ = 1/2(|00⟩⟨00| + |01⟩⟨01|) on two qubits and use the overlapping block pattern with blocks {00,01,10} and {01,10,11} (R=2, overlap two indices). The second block has matrix diag(0.5,0,0), rank 1, violating isorank, so the subspace intersection computed from the two blocks fails to recover the |00⟩ component and the algorithm's output does not equal ρ despite zero noise.","supporting_citations":[],"review_version":1}