{"id":"137dfefe-89e1-4977-9d31-e448fff1f85c","arxiv_id":"1909.00116","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A debiasing and spectral projection procedure constructs asymptotically normal estimators for any linear form of a low-rank matrix from noisy partial observations, enabling confidence intervals and tests.","lead":"This paper develops a general statistical framework for drawing conclusions about sums of entries in a large matrix from noisy, partially observed entries. It provides confidence intervals and hypothesis tests for such linear forms under near-optimal sample sizes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 1 is not instantiated by the paper's own Algorithm 1 for fixed rank: Theorem 3 yields γ_n=O(r^{-1/2}), not o(1), so the debiasing theorem's key input is missing in fixed-rank asymptotics.","rationale":"The reader identified Assumption 1 as the weakest assumption; the stress-test agrees but sharpens the concern. The paper itself concedes that the cited max-norm guarantees (Ma et al. 2017; Chen et al. 2019c) were proved for sampling without replacement, and that an independent-sampling proof is lacking. More decisively, the substitute initial estimator in Algorithm 1 has a stated error bound that, at the paper's own sample size, scales as O(1/√r) for fixed rank. Thus Assumption 1 is not established even for the paper's own pipeline in the usual fixed-rank asymptotics. This is load-bearing because Assumption 1 enters the debiasing expansion (7.1), the perturbation theorem (Theorem 4), and Lemmas 3–6; without it, the normality theorem has no input. The theorem itself remains conditionally valid, so the verdict should stay conditional, but the authors should either prove a max-norm guarantee with γ→0 for Algorithm 1 at the advertised sample size, or add the stronger condition n ≫ r^2 d log^2 d needed for their initializer to satisfy Assumption 1.","tokens_in":54505,"tokens_out":31110,"duration_ms":304074,"concrete_test":"Fix r=3, take d ∈ {10^3, 10^4, 10^5} with λ_i = d and σξ = 1. Run Algorithm 1 with the §5 data splitting at n = 4 r^3 d log^2 d, with m = 2⌈log d⌉ iterations, over 100 trials, and record ‖M̂^(m)−M‖max/σξ. If the ratio does not decrease toward 0 as d grows (Theorem 3 predicts a floor around 1/√r), Assumption 1 fails in fixed-rank asymptotics. If it does decrease, the fixed-rank concern is refuted and only the stated proof needs strengthening.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 1's normality claim is conditional on Assumption 1: an initial estimator with max-norm error o_P(σξ). The paper's constructive route to Assumption 1 is Algorithm 1/Theorem 3. But Theorem 3's bound is ‖M̂^(m)−M‖max ≤ C μκ σξ √(r^2 d1 log^2 d1 / n). Under the theorem's stated sample condition n ≥ C α κ^6 μ^6 r^3 d1 log^2 d1, this ratio is at best O(1/√r). For fixed rank r, it is a positive constant, not o(1). Consequently the γ_{n,d1,d2} in Assumption 1 does not tend to 0 in fixed-rank asymptotics, so the paper's own initialization does not feed the debiasing theorem. Figure 1 confirms this quantitatively: with r=3 and d=2000, log(‖M̂−M‖max/σξ) levels off near −0.4, i.e. a max error of about 0.67σξ rather than a vanishing fraction. The text after Theorem 3 states that Assumption 1 is satisfied with γ=μκ√(r^2 d log^2/n), but for fixed r this is O(r^{-1/2}), not a sequence tending to 0. This is not an internal contradiction of Theorem 1, but it breaks the paper's universality claim: the only initializer supplied fails Assumption 1 in the standard fixed-rank regime.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops a framework for statistical inference on general linear forms tr(M^T T) of a low-rank matrix M from n noisy, uniformly sampled entries of the matrix completion model (1.1). The procedure has four steps: splitting the sample in two, debiasing each half by an additive correction based on the other half, projecting the debiased matrices onto their top-r singular subspaces, and plugging the averaged projection into the linear form (Section 2). Theorem 1 establishes an asymptotic normality result with an explicit Berry-Esseen bound under an initial-estimator condition (Assumption 1: max-norm error o_P(sigma_xi)), an incoherence condition (Assumption 2), a sub-Gaussian noise and SNR condition (Assumption 3), and an alignment condition (Assumption 4). Theorem 2 extends the result to data-driven noise-variance and singular-subspace estimates, yielding confidence intervals and tests. Section 5 introduces a rotation-calibrated Grassmannian gradient descent algorithm (Algorithm 1) intended to supply the initial estimator, with geometric convergence proved in Theorem 3; Section 7 and the appendices contain the proofs, built on a de-localized perturbation expansion of empirical singular vectors (Theorem 4). Simulations with Gaussian and uniform noise and experiments on the Jester and MovieLens datasets illustrate the method.","tokens_in":54833,"tokens_out":20588,"duration_ms":163624,"significance":"If the conditional result holds, this is a substantial contribution to noisy matrix completion inference. The framework covers entries, differences of entries, and other sparse or moderate linear forms at near-optimal sample sizes, and it shows that the rate of the initial estimator is irrelevant provided it is max-norm consistent, which is a useful and somewhat surprising design principle. The proof machinery is a strength: Theorem 1 carries an explicit Berry-Esseen rate, the perturbation analysis in Section 7.1 and Lemmas 1-9 of the appendix provide sharp (2,max)-norm bounds for empirical singular vectors under sub-Gaussian noise, and the double-sample debiasing avoids the efficiency loss of single splitting. The verification of the theoretical input is, however, incomplete: the supplied initializer does not provably satisfy Assumption 1 in the fixed-rank regime, so the end-to-end guarantee of the procedure is not established; this is a load-bearing gap that a major revision should close.","major_comments":[{"comment":"Assumption 1 (Section 2, eq. (2.1)) requires a sequence gamma_{n,d1,d2} -> 0 such that the initial estimators have max-norm error O(gamma sigma_xi), and the Berry-Esseen bound in Theorem 1 together with condition (3.5) requires gamma_n sqrt(log d1) -> 0. The paper's constructive verification, Theorem 3 (Section 5), proves ||Mhat^{(m)} - M||_max <= C6 mu_max kappa_0 sigma_xi sqrt(r^2 d1 log^2 d1 / n) under n >= C1 alpha_d kappa^6 mu^6 r^3 d1 log^2 d1. Substituting the minimal sample size gives ||Mhat^{(m)} - M||_max / sigma_xi = O(1/sqrt(r)) up to constants in mu, kappa, alpha_d; for fixed rank r, the regime used throughout the paper (r = 3 in Section 6.1, r = 2 in Section 6.2), this is a positive constant rather than a vanishing sequence. Consequently the claim on page 18 that 'Assumption 1 is satisfied with gamma_{n,d1,d2} = mu_max kappa_0 sqrt(r^2 d1 log^2 d1 / n)' is incorrect as stated, and the only initializer supplied by the paper does not instantiate Assumption 1 in fixed-rank asymptotics. Figure 1 corroborates this: the curve log(||Mhat - M||_max / sigma_xi) saturates at about -0.4, i.e., a max error near 0.67 sigma_xi, while the text of Section 6.1 states that the debiasing procedure requires o_P(sigma_xi); the histograms in Figures 2 and 3 are therefore not covered by Theorem 1. This gap is load-bearing for the paper's universality claim, because the abstract and Section 2 present Algorithm 1 as the concrete route to Assumption 1. The issue is fixable within scope: either require the sample size in Theorem 3 to satisfy n / (d1 log^2 d1) -> infinity, which makes gamma_n vanish for fixed r, or sharpen Algorithm 1's max-norm rate to sqrt(r d1 log d1 / n) via a leave-one-out analysis; the revision should also reconcile Figure 1 with Assumption 1.","section":"§5 (Theorem 3 and following paragraph); §3 (Assumption 1)"}],"minor_comments":[{"comment":"Several citations contain '[?]' placeholders (e.g., the Chernozhukov et al. reference on page 8 and the data-source descriptions for Jester and MovieLens on pages 19-20); these should be completed before the manuscript goes into production.","section":"§1 and §6.2"},{"comment":"The phrase 'the noise xi is independent with X' should read 'independent of X'; this wording error appears in Assumption 3 and several times in Sections 3 and 7.","section":"Assumption 3 and throughout §3"},{"comment":"The introduction states that 'All proofs are presented in the online supplement,' but the proofs of Theorems 1-4 and Lemmas 1-9 appear in the main text (§7 and Appendices A-I); the sentence should be corrected to match the actual organization.","section":"§1, page 6"},{"comment":"The statement that the procedure is 'generally robust to reasonable estimate of r' is acknowledged to lack rigorous justification; since all theorems assume r known, this claim should be explicitly labeled as an empirical observation so that the theoretical scope of the paper is unambiguous.","section":"§3.1, page 10"},{"comment":"The term '6 log d1 / d2^1' in Theorem 1's Berry-Esseen bound and the analogous '3/d2^1' terms in Lemma 4 appear to be typesetting artifacts of what should be expressions like 6 log d1 / d1^2; these should be reset in LaTeX so that the bounds are readable and verifiable.","section":"Theorem 1 and §7.2"}],"recommendation":"major_revision","confidential_remarks":"Dear Editor: The conditional theorem and its proof apparatus are, in my reading, sound; the manuscript's central defect is the mismatch between Assumption 1 and Theorem 3, which I regard as fixable within the scope of a major revision rather than grounds for rejection. I recommend checking the paper against the concurrent entrywise-inference work of Chen, Fan, Ma, and Yan (Chen et al., 2019c); the present paper's distinct contribution is the general linear-form framework and the second-order perturbation analysis. The spectral expansion (7.11), imported from Xia (2019b), is load-bearing for Theorem 4; the authors should either state it as a formal lemma with a proof sketch or verify its validity in the present sub-Gaussian, non-symmetric setting. Finally, the manuscript needs editorial cleanup ([?] placeholders and corrupted formula fragments)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Candid take: Theorem 1 is a solid conditional result, but the paper's own initialization does not satisfy the theorem's key assumption in the fixed-rank regime, and the stress-test note lands.\n\nWhat's genuinely good: The double-sample debiasing plus spectral projection framework is a real step beyond entrywise intervals. It yields asymptotic normality for general linear forms, with explicit Berry-Esseen bounds and data-driven variance estimates (Theorem 2). The treatment of alignment via the singular-space projections is illuminating, and the proof structure in Section 7 is careful: the spectral expansion from Xia (2019b) is extended to sub-Gaussian noise, with higher-order terms controlled in (2,max)-norm. For a reader who accepts Assumption 1, the main claim is well-supported.\n\nThe soft spot: Assumption 1 is not instantiated. The paper notes that existing max-norm guarantees were proved under sampling without replacement and PSD matrices, then supplies Algorithm 1. But Theorem 3's bound under its sample condition gives gamma_n = O(1/sqrt(r)), which is not o(1) for fixed r. Worse, Theorem 1 requires gamma_n sqrt(log d1) -> 0; plugging in the supplied gamma gives gamma sqrt(log d1) ~ sqrt(log d1)/sqrt(r) -> infinity. So the paper's own initial estimator, at the stated sample size, does not satisfy the rate condition for the CLT. This is not a contradiction of Theorem 1, but it substantially weakens the 'universal procedure at optimal sample size' claim. The gap is repairable—take n >> r^2 d1 log^3 d1, or sharpen the max-norm analysis—but as written it needs to be stated.\n\nMinor issues: the ROC labels use I(Y>Y') as a noisy proxy, which can inflate the curves; rank r is fixed by a heuristic; a couple of references are placeholders. None fatal.\n\nBottom line: For researchers in high-dimensional inference, the conditional theorem and the debiasing framework are valuable and worth citing with the caveat. The paper deserves a serious referee; I'd send it out and ask the authors to address the initialization gap before acceptance.\n\nRecommendation: engage, but don't let the universality claim pass unchallenged.","headline":"A valuable conditional normality result for linear forms in noisy matrix completion, but the paper's own initial estimator does not satisfy the key assumption in the fixed-rank asymptotic regime.","tokens_in":55330,"tokens_out":5117,"would_cite":true,"duration_ms":47771,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F12","60F05","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Noisy matrix completion can support confidence intervals and tests for any linear summary of the matrix.","keywords":["noisy matrix completion","trace regression","linear forms","double-sample debiasing","spectral projection","singular subspace perturbation","asymptotic normality","confidence intervals"],"falsifier":"Simulate the procedure with $d_1=d_2=2000$, rank 3, noise $\\sigma_\\xi=0.6$, sample size $n=4r^2 d\\log d$, and a linear form $T$ aligned with a singular vector; the studentized statistic should be approximately standard normal over repeated runs. If the histogram shows detectable bias or the coverage of the Wald intervals misses the nominal level at these sample sizes, the normality claim would be contradicted. A sharper test targets the boundary: for $T$ nearly orthogonal to the singular spaces (small alignment), Assumption 4 fails and the asymptotic normality should break down, which a simulation could confirm.","tokens_in":54319,"feed_emoji":"📊","tokens_out":9179,"duration_ms":79632,"temperature":0.7,"pith_summary":"This paper claims that a matrix completed from noisy, partially observed entries can support much richer statistical inference than point estimation alone. Given any initial estimator whose entry-wise (max-norm) error is smaller than the noise level, the paper's two-step recipe—double-sample debiasing followed by projection onto the estimated low-rank singular subspaces—produces an estimator of any linear form tr(M^T T), a weighted sum of the matrix entries, that is asymptotically normal when standardized by the right variance. This yields confidence intervals and hypothesis tests for entries, entry differences, and more general linear forms at sample sizes that are optimal up to logarithmic factors. The practical point is that users need not design a special estimator for inference; a consistent estimator is enough, and its rate of convergence does not affect the limiting distribution.","feed_headline":"Debiased matrix completion yields tests for any linear form","feed_subtitle":"Confidence intervals and entry comparisons follow from a debiased, projected estimate","key_machinery":"The central object is the double-sample-debiased estimate: split the data in half, form an initial estimate from one half, and add the one-step correction $(d_1d_2/n_0)\\sum_i(Y_i-\\langle \\hat M_{\\mathrm{init}},X_i\\rangle)X_i$ using the other half; repeat with the halves swapped. Projecting the two debiased matrices onto their top-$r$ singular subspaces and averaging gives $\\hat M$. The proof is carried by a sharp perturbation expansion of the empirical singular subspaces, written as a series in powers of the noise matrix, which lets the paper control the $(2,\\max)$-norm errors of the estimated singular vectors under sub-Gaussian noise. To provide the required initial estimator, the paper gives a rotation-calibrated gradient descent on the product of two Grassmann manifolds (rank-$r$ orthonormal frames modulo rotation) that converges geometrically with constant step sizes.","core_discovery":"Under the noisy matrix completion model $Y = \\langle M, X\\rangle + \\xi$ with entries sampled uniformly and sub-Gaussian noise, the paper proves that the debiased and spectrally projected estimator $\\hat M$ satisfies $\\bigl(\\operatorname{tr}(\\hat M^{\\top}T)-\\operatorname{tr}(M^{\\top}T)\\bigr)\\big/\\bigl(\\sigma_\\xi(\\|U^{\\top}T\\|_F^2+\\|TV\\|_F^2)^{1/2}\\bigr)\\cdot\\sqrt{d_1d_2/n}\\xrightarrow{d}N(0,1)$ whenever the initial estimator satisfies Assumption 1 and the incoherence, SNR, and alignment conditions hold. The same convergence remains true with estimated noise level and estimated singular vectors, so the statistic is directly usable for confidence intervals and tests. A notable corollary is that the asymptotic variance of the estimator is governed by how $T$ aligns with the singular spaces of $M$, not by the sparsity of $T$ alone, so dense linear forms are allowed when their alignment is sufficient.","pith_inferences":["A natural next step is to reweight the debiasing sum by inverse sampling probabilities for non-uniform entry sampling; the paper notes the calibration would change but does not develop it.","The alignment condition suggests that tests are intrinsically more powerful for linear forms that lie partially in the row or column singular spaces, while forms wholly in the null space require a different limiting theory.","The paper's numerical evidence suggests the procedure is insensitive to the estimated rank, but no proof is given; one testable extension is a rigorous rank-estimation scheme inside the debiasing loop.","One could apply the same double-sample debiasing to other trace-regression designs with known second moment of the design, since the debiasing formula only needs $E[\\operatorname{vec}(X)\\operatorname{vec}(X)^\\top] = (d_1d_2)^{-1} I$."],"forward_implications":["Entrywise confidence intervals for the entries of $M$ become available at near-optimal sample sizes, $n$ on the order of $r d_1 \\log d_1$ up to constants and log factors.","Comparisons such as $H_0: M(i,j_1)=M(i,j_2)$, including group versions that aggregate over users, can be tested with asymptotic level control.","The initial estimator only needs max-norm consistency $o_P(\\sigma_\\xi)$; its convergence rate does not enter the limiting distribution, so a suboptimal estimator can lead to optimal inference.","The asymptotic variance of the linear-form estimator is determined by the alignment of $T$ with $M$'s singular spaces, so dense linear forms are allowed when their singular-space alignment is not small.","The integrated mean squared error of the debiased projected estimator is $\\sigma_\\xi^2 r d_1 d_2 (d_1 + d_2)/n$ up to a $(1+o(1))$ factor, matching the minimax lower bound."],"supporting_citations":[{"why":"Provides the max-norm consistency rate for a nonconvex estimator under sampling without replacement that motivates Assumption 1.","marker":"Ma et al. (2017)"},{"why":"Concurrent entrywise inference and max-norm guarantees cited as another basis for Assumption 1.","marker":"Chen et al. (2019c)"},{"why":"Supplies the double-sample-splitting debiasing device that avoids the efficiency loss of naive sample splitting.","marker":"Chernozhukov et al. (2018)"},{"why":"Developed the spectral representation of empirical singular vectors that the paper extends from Gaussian to sub-Gaussian noise.","marker":"Xia (2019b)"},{"why":"Earlier normal approximation of singular subspaces in low-rank matrix regression used as a technical template.","marker":"Xia (2019a)"},{"why":"Gives the nuclear-norm estimator rate and minimax benchmark whose Frobenius rate the paper's estimator matches.","marker":"Koltchinskii et al. (2011)"}],"fun_headline_variants":["Debiased matrix completion yields tests and intervals for linear forms","Inference on linear forms from noisy matrices, debiased and projected","Normal inference for linear forms via double-sample debiased completion","Matrix completion: prove linear form confidence intervals with debiasing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument collapses without an initial estimator whose maximum entry-wise error is $o_P(\\sigma_\\xi)$ under the paper's independent uniform sampling; the paper itself notes that such guarantees were previously proved mainly for sampling without replacement and positive semidefinite matrices, and supplies its own initial estimator only under stronger sample-size and SNR conditions.","fun_headline_variants_meta":{"raw":{"variants":["Debiased matrix completion yields tests and intervals for linear forms","Inference on linear forms from noisy matrices, debiased and projected","Normal inference for linear forms via double-sample debiased completion","Matrix completion: prove linear form confidence intervals with debiasing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000263,"raw_usage":{"total_tokens":1560,"prompt_tokens":864,"completion_tokens":696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":623}},"tokens_in":480,"tokens_out":696,"duration_ms":7221,"temperature":1.0,"reasoning_tokens":623,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T06:01:51.862527+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the procedure with $d_1=d_2=2000$, rank 3, noise $\\sigma_\\xi=0.6$, sample size $n=4r^2 d\\log d$, and a linear form $T$ aligned with a singular vector; the studentized statistic should be approximately standard normal over repeated runs. If the histogram shows detectable bias or the coverage of the Wald intervals misses the nominal level at these sample sizes, the normality claim would be contradicted. A sharper test targets the boundary: for $T$ nearly orthogonal to the singular spaces (small alignment), Assumption 4 fails and the asymptotic normality should break down, which a simulation could confirm.","supporting_citations":[],"review_version":1}