{"id":"14795254-86ad-4af8-bf54-ef30a46f53d3","arxiv_id":"2501.09336","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"In high-SNR settings AJIVE's shared-subspace error is minimax-optimal and decays like 1/√K in the number of matrices, while in low-SNR settings a non-diminishing error floor appears even for an oracle-aided spectral estimator.","lead":"This paper proves when AJIVE, a standard method for finding the common pattern shared across several noisy data matrices, gets better as you add more matrices and when it stalls. The authors also prove matching lower bounds showing AJIVE is near-optimal in high-signal settings, and that a low-signal error floor persists even for an idealized oracle-aided version of the method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Low-SNR 'fundamental barrier' rests only on an algorithm-specific lower bound; no information-theoretic limit is proved, so the abstract overclaims.","rationale":"The reader identified the gap between the abstract's 'fundamental barrier' claim and the proven algorithm-specific lower bound as the weakest assumption. I agree: the load-bearing step is Section 5, where the paper explicitly states it cannot deliver an information-theoretic limit. Theorem 3 applies only to Algorithm 2, a one-step spectral estimator initialized at the true U⋆, and the constructed hard instance has a specific cross-loading structure (V⋆ᵀW⋆ = 0.6I). This demonstrates a limitation of the spectral approach, but not of the estimation problem itself. The high-SNR analysis, by contrast, is internally coherent: Theorem 1's first-order term matches Theorem 2's minimax lower bound, and the appendix strengthens the lower bound with second-order terms. The numerical experiments are consistent but do not close the logical gap. Therefore the reader's CONDITIONAL verdict is appropriate; no new objection beyond the identified one is needed, and the condition is that the fundamental-barrier claim be reconciled with the proven algorithm-specific statement.","tokens_in":63705,"tokens_out":7479,"duration_ms":81766,"concrete_test":"Attempt an information-theoretic lower bound using Le Cam's two-point method on the Theorem 3 hard instance: fix the same unique subspaces {U⋆ₖ}, V⋆, W⋆, and choose two shared subspaces U⋆ and U⋆′ with projection distance Ω(σ⁴n²/σₘᵢₙ⁴); compute the KL divergence between the two induced Gaussian observation laws as a function of K. If the KL divergence can be bounded by a constant (or driven to 0) while the separation remains Ω(σ⁴n²), then the non-diminishing error is truly fundamental and the abstract is supported; if the KL divergence is too large to yield a constant lower bound, then Theorem 3's algorithmic bound is not evidence of an information-theoretic barrier and the abstract must be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that AJIVE's non-diminishing low-SNR error is a fundamental limitation depends on Theorem 3, but Theorem 3 is a lower bound for the oracle-aided spectral estimator (Algorithm 2), not a minimax lower bound over all estimators. The paper itself concedes in Section 5 ('we are unable to deliver an information-theoretic limit') and Section 7 ('might be a fundamental limit') that the barrier is not proven information-theoretically. Moreover, Theorem 3's construction is adversarial: it sets V⋆ᵀW⋆ = 0.6I and uses a fourth-order spectral expansion whose bias comes from the cross term α₄·K⁻¹Σₖ(U⋆U⋆ₖᵀ + U⋆ₖU⋆ᵀ) in E[Q]. Nothing rules out an estimator outside this spectral/alternating-minimization class—for example, a joint likelihood approach with multiple random starts, a penalized estimator, or a non-spectral method—that could have vanishing error as K grows. Since Algorithm 2 is initialized at the true U⋆ and still fails, the result is evidence about the algorithm, not about the statistical problem. Therefore the abstract's phrase 'fundamental barrier' is stronger than what Theorem 3 proves, and the paper's own Section 7 wording is the accurate statement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the AJIVE algorithm for estimating the shared column subspace in the JIVE model with K noisy data matrices. The main theoretical contribution is a two-regime characterization: in a high-SNR regime, the estimation error of AJIVE is shown to decay as σ/σmin times roughly sqrt(n/K) plus a misalignment-dependent term, and a matching minimax lower bound is established; in a low-SNR regime, a non-diminishing error term is identified. Since the paper does not prove an information-theoretic lower bound for this low-SNR barrier, it instead proves a lower bound for an oracle-aided spectral estimator (Theorem 3) and presents numerical evidence. The paper also contains a detailed proof appendix based on SVD perturbation expansions, concentration inequalities, and Fano/packing arguments.","tokens_in":63949,"tokens_out":7999,"duration_ms":83393,"significance":"The high-SNR results are a solid contribution: Theorem 1 provides a nearly complete rate characterization of AJIVE, and Theorem 2 shows the first-order rate is minimax-optimal. The dependence on the misalignment parameter θ and the explicit benefit of multiple matrices (1/sqrt(K) rate) are new and quantitatively useful. The paper also ships detailed, self-contained proofs, including a nontrivial Fano construction for the θ-dependent term that may be of independent interest. The low-SNR part is suggestive but not at the same level of rigor: Theorem 3 is an algorithm-specific lower bound for an oracle-initialized spectral estimator, not a minimax lower bound, and the paper's own caveats in Sections 5 and 7 are more accurate than the abstract. If the claims are appropriately qualified, the paper is a valuable contribution to the theory of multi-matrix subspace estimation.","major_comments":[{"comment":"The abstract and Section 1 state that in low-SNR settings AJIVE exhibits a non-diminishing error as K grows, but Theorem 1 does not prove this. Theorem 1 is an upper bound that holds only under the high-SNR conditions (6a)-(6b); when σ√n/σmin is not ≪ √θ, those conditions fail and the bound, including the simplified expression (8), is not a valid consequence. The discussion immediately after (8) interprets the second-order term E2 as a low-SNR non-diminishing error, but this is an extrapolation outside the theorem's assumptions. Please either prove a low-SNR lower bound for AJIVE itself or explicitly label this statement as a conjecture supported by the oracle-estimator analysis and the simulations.","section":"Section 3, Eqs. (6)-(8)"},{"comment":"The phrase 'fundamental barrier' in the abstract and in the contribution list is stronger than what Theorem 3 establishes. Theorem 3 is a lower bound for the oracle-aided spectral estimator (Algorithm 2) on a specially constructed configuration with V⋆ᵀW⋆ = 0.6I; it is not a minimax lower bound over all estimators, and it does not rule out non-spectral methods that could have vanishing error as K grows. The paper itself concedes this in Section 5 ('we are unable to deliver an information-theoretic limit') and Section 7 ('might be a fundamental limit'). The abstract, Section 1, and the concluding remarks should be aligned with these qualifications; the current wording overstates the strength of Theorem 3.","section":"Section 5, Theorem 3"},{"comment":"The connection to Neyman and Scott's problem is presented as if the oracle estimator being 'one step of alternating minimization starting from the ground truth' makes it representative of the MLE. This is an analogy, not a proof: Algorithm 2 is neither the nonconvex least-squares estimator nor a full alternating-minimization run, and the paper does not show inconsistency of the MLE itself. The text should clearly state that the Neyman-Scott connection is heuristic, or provide an analysis of the actual nonconvex estimator.","section":"Section 5.3"}],"minor_comments":[{"comment":"The abstract's phrase 'a fundamental barrier' is inconsistent with Section 7's accurate wording 'might be a fundamental limit'; please use the qualified version throughout.","section":"Abstract"},{"comment":"The statement 'Without loss of generality, assume n is divisible by 3 and K is even' is not a true WLOG; rounding or a padding argument should be supplied, or the assumption should be stated as a technical condition.","section":"Section D.3.1"},{"comment":"In the proof of (22) and (23), the sentence 'Moreover, this truncation is tight...' is followed by a probability statement for a deterministic expectation; the phrase 'with probability at least 1 - O(N^-11)' around (23b) is misleading since (23b) is an exact identity, not a high-probability event.","section":"Section C.1.1"},{"comment":"The figure caption states the y-axis is in log scale but does not specify whether the x-axis is √K or log √K; adding the axis scaling in the caption would improve readability.","section":"Section 6, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The high-SNR theorems and proofs are the strongest part of the paper and are likely publishable after the overstatements are corrected. The main concern is that the abstract and contribution bullets commit to an information-theoretic 'fundamental barrier' that the paper does not actually prove; the body text is more careful. I would encourage the editor to ask for a revision that temper the low-SNR claims, rather than a rejection, because the core technical work is sound and the caveats are already present in Sections 5 and 7."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new results are worth a careful read: Theorem 1 gives AJIVE guarantees with explicit dependence on the number of matrices K and the misalignment parameter theta, and Theorem 2 matches the first-order high-SNR rate with a minimax lower bound. The appendix even strengthens the lower bound to include second-order terms, and the packing construction behind it looks like real work. That part of the paper is in good shape.\n\nThe soft spot is exactly where you'd expect: the low-SNR story. The body is admirably honest—Section 5 says outright that no information-theoretic limit is proved, and Section 7 calls the non-diminishing error something that \"might be\" fundamental. But the abstract and parts of the introduction sell Theorem 3 as a fundamental barrier, when Theorem 3 is an algorithm-specific lower bound for a spectral estimator initialized at the true U*. That is a real gap between packaging and proof. It is not a fatal flaw; it is a scope mismatch. The Neyman-Scott connection is suggestive, but suggestive is not a theorem.\n\nThe simulations are fine as sanity checks, no error bars and no code, but the theory stands on its own. The homoscedastic Gaussian assumption and known ranks are restrictive, but they're standard for this literature and clearly stated.\n\nMy take: the high-SNR analysis is a solid contribution that closes a genuine gap in JIVE theory. The low-SNR 'barrier' is a well-defined open problem, not a closed result. A serious referee should ask the authors to rewrite the abstract and introduction to match the proven scope, and to state explicitly that the oracle lower bound does not rule out non-spectral estimators.\n\nThis deserves peer review. I'd send it out, and I'd tell the authors that the main revision is aligning the claims with the theorems.","headline":"Solid high-SNR minimax theory for AJIVE; the low-SNR 'fundamental barrier' is honestly labeled in the body but overclaimed in the abstract.","tokens_in":64519,"tokens_out":899,"would_cite":true,"duration_ms":12387,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H25","62C20","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes a two-regime theory for AJIVE shared-subspace estimation: high signal-to-noise gives a minimax-optimal $1/\\sqrt K$ error decay, while low signal-to-noise leaves a non-diminishing error that also plagues an…","keywords":["JIVE","AJIVE","shared subspace estimation","multi-view data integration","spectral method","minimax lower bound","subspace misalignment","incidental parameter problem"],"falsifier":"Run a low-SNR sequence with $K$ from 25 to 10,000 on the paper's hard configuration ($V_k^{\\star\\top}W_k^\\star = 0.6I$, $\\sigma\\sqrt n/\\sigma_{\\min}$ small, $r\\le n/3$) and test a full joint nonconvex estimator, not the oracle-initialized one-step estimator: if its error converges to 0, the claimed fundamental barrier is false. Since the paper explicitly lacks an information-theoretic lower bound in this regime, a minimax calculation that forces positive error for all estimators would be the definitive confirmation.","tokens_in":63444,"feed_emoji":"📊","tokens_out":11930,"duration_ms":110734,"temperature":0.7,"pith_summary":"Integrative data analysis frequently asks whether several noisy matrices share one low-dimensional signal subspace, with each matrix carrying its own private variation on top. This paper analyzes AJIVE, a two-stage spectral method for estimating that shared subspace, and establishes a two-regime picture. When the noise is weak enough, $\\sigma\\sqrt n/\\sigma_{\\min} \\ll \\min\\{\\sqrt\\theta,\\sqrt{K\\theta}\\}$, the estimation error decays like $(\\sigma/\\sigma_{\\min})\\sqrt{n/K + r/(K\\theta)}$, matching a minimax lower bound, so pooling more matrices genuinely helps. When the noise is strong, a second-order bias term $\\frac{1}{\\theta(1\\wedge K\\theta)}\\frac{\\sigma^2 n}{\\sigma_{\\min}^2}$ does not vanish as $K\\to\\infty$, and the paper shows the same non-vanishing error occurs for an oracle-aided spectral estimator. The result matters because it circumscribes when multi-view integration can and cannot improve shared-subspace recovery.","feed_headline":"More matrices help only at high signal-to-noise","feed_subtitle":"Shared-subspace error falls like 1/√K in the high-SNR regime, then hits a floor that survives oracle information.","key_machinery":"The load-bearing object is the averaged projection matrix $M := \\frac{1}{K}\\sum_{k=1}^K \\widehat U_k \\widehat U_k^\\top$, formed from per-matrix estimates of the top $(r+r_k)$ left singular subspaces; AJIVE outputs the top-$r$ eigenspace of $M$. The analysis separates the noiseless leading term $U^\\star U^{\\star\\top} + \\frac{1}{K}\\sum_k U_k^\\star U_k^{\\star\\top}$ from a perturbation $\\Delta$ built from first-order SVD expansions, and the misalignment parameter $\\theta$ controls the eigengap of that leading term. For the oracle lower bound the machinery is a fourth-order polynomial approximation $Q$ of the oracle matrix $M$, whose expectation contains a cross term $\\alpha_4 \\frac{1}{K}\\sum_k (U^\\star U_k^{\\star\\top} + U_k^\\star U^{\\star\\top})$ with $\\alpha_4 \\asymp \\sigma^4 nd$; because $V^{\\star\\top}W^\\star = 0.6I$ in the hard configuration, this cross term tilts the leading eigenspace away from $U^\\star$ and the tilt does not shrink as $K$ grows.","core_discovery":"On the paper's own terms, the discovery is a complete first-order characterization of AJIVE for the JIVE model $A_k = U^\\star V_k^{\\star\\top} + U_k^\\star W_k^{\\star\\top} + E_k$ with i.i.d. Gaussian noise of variance $\\sigma^2$. Theorem 1 (simplified at equation (8)) says that under the high-SNR condition $\\sigma\\sqrt n/\\sigma_{\\min} \\ll \\min\\{\\sqrt\\theta,\\sqrt{K\\theta}\\}$ the estimation error $\\|\\widehat U\\widehat U^\\top - U^\\star U^{\\star\\top}\\|$ is bounded by $(\\sigma/\\sigma_{\\min})(\\sqrt{n/K} + \\sqrt{r/(K\\theta)})$ plus a second-order term; Theorem 2 shows the first-order rate is minimax optimal, with the rate separating a large-$\\theta$ regime where unique subspaces barely interfere from a small-$\\theta$ regime where their alignment degrades the rate. In the complementary low-SNR regime the second-order term dominates and converges to $\\frac{1}{\\theta}\\frac{\\sigma^2 n}{\\sigma_{\\min}^2}$ as $K\\to\\infty$. Theorem 3 constructs a worst-case loading configuration $V_k^{\\star\\top}W_k^\\star = 0.6I$ on which even the oracle estimator, namely one alternating-minimization step initialized at the true $U^\\star$, has error bounded below by $C\\sigma^4 n^2/\\sigma_{\\min}^4$, so the non-diminishing error is presented as a structural barrier rather than an artifact of AJIVE.","pith_inferences":["A likely testable extension is that full alternating minimization, which re-estimates $U$ jointly with the unique components rather than stopping after one step from the truth, may behave differently from the oracle estimator in the low-SNR regime; the paper only analyzes the one-step version.","Because the theory assumes equal noise variance across matrices and known ranks, an immediate extension would allow per-matrix $\\sigma_k^2$; the second-stage average would then need inverse-variance weighting, which may shift or remove the non-diminishing term.","The hard configuration uses adversarial loadings with $V^{\\star\\top}W^\\star = 0.6I$; the numerical experiments show the low-SNR floor under shared loading while random loadings show the $1/\\sqrt K$ decay, suggesting that alignment between $V_k^\\star$ and $W_k^\\star$, not just subspace misalignment $\\theta$, controls whether the floor appears.","A practical diagnostic suggested by the theory: estimate $\\theta$ and the SNR before deciding whether to pool more matrices; when $\\theta$ and SNR are both small, the shared subspace should be treated as only partially identified."],"forward_implications":["In the high-SNR regime, averaging more matrices improves shared-subspace estimation at the minimax-optimal rate $1/\\sqrt K$, so additional views give real accuracy gains.","The optimal rate worsens as the unique subspaces become more aligned: the $r/(K\\theta)$ term shows that near-identifiability, not noise alone, controls difficulty.","In the low-SNR regime, AJIVE's error has a component $\\frac{1}{\\theta}\\frac{\\sigma^2 n}{\\sigma_{\\min}^2}$ that survives $K\\to\\infty$; beyond a threshold the number of matrices is irrelevant to the error floor.","An oracle-aided spectral estimator, given the true unique components in the first stage, still fails on a worst-case alignment, indicating the floor is shared by spectral methods rather than a quirk of AJIVE.","The persistent error fits the classical incidental-parameter phenomenon: the shared subspace is the structural parameter and the per-matrix unique subspaces are incidental parameters that contaminate likelihood-type estimation."],"supporting_citations":[{"why":"Defines the AJIVE two-stage spectral algorithm whose error guarantees are the subject of the paper.","marker":"[FJHM18]"},{"why":"Supplies the singular-subspace perturbation expansion used for the first-order AJIVE analysis and for the fourth-order oracle-estimator approximation.","marker":"[Xia21]"},{"why":"The concurrent stacked-SVD analysis whose noiseless failure motivates AJIVE and whose assumptions the paper avoids.","marker":"[MM24]"},{"why":"Introduces the JIVE model and the rank and identifiability assumptions underlying the estimation problem.","marker":"[LHMN13]"},{"why":"The classical incidental-parameter result that the paper invokes to interpret the oracle estimator's persistent bias.","marker":"[NS48]"},{"why":"Provides the packing construction and Fano-style argument adapted to build the minimax lower bounds.","marker":"[CMW13]"},{"why":"Supplies the generalized Fano method used to convert KL-divergence bounds into estimation lower bounds.","marker":"[Yu97]"},{"why":"Gives the high-dimensional spiked-model detection threshold used to show the SNR condition is tight when K is constant.","marker":"[EAJ18]"},{"why":"Establishes the single-matrix rate-optimal subspace perturbation baseline that the multi-matrix lower bound builds on.","marker":"[CZ18]"}],"fun_headline_variants":["AJIVE gains from more matrices only at high SNR","Even oracle can't fix AJIVE's low-SNR error floor","Multi-matrix integration helps AJIVE only when signal is strong","High SNR: error drops with K; low SNR: it flattens"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the oracle-aided spectral estimator, namely one alternating-minimization step initialized at the true shared subspace, represents the fundamental low-SNR limit; the paper states it cannot prove an information-theoretic lower bound, so any estimator outside this spectral class that drives the error to zero would overturn the barrier conclusion.","fun_headline_variants_meta":{"raw":{"variants":["AJIVE gains from more matrices only at high SNR","Even oracle can't fix AJIVE's low-SNR error floor","Multi-matrix integration helps AJIVE only when signal is strong","High SNR: error drops with K; low SNR: it flattens"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000488,"raw_usage":{"total_tokens":2496,"prompt_tokens":1128,"completion_tokens":1368,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":744,"completion_tokens_details":{"reasoning_tokens":1295}},"tokens_in":744,"tokens_out":1368,"duration_ms":9776,"temperature":1.0,"reasoning_tokens":1295,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:07:09.590239+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a low-SNR sequence with $K$ from 25 to 10,000 on the paper's hard configuration ($V_k^{\\star\\top}W_k^\\star = 0.6I$, $\\sigma\\sqrt n/\\sigma_{\\min}$ small, $r\\le n/3$) and test a full joint nonconvex estimator, not the oracle-initialized one-step estimator: if its error converges to 0, the claimed fundamental barrier is false. Since the paper explicitly lacks an information-theoretic lower bound in this regime, a minimax calculation that forces positive error for all estimators would be the definitive confirmation.","supporting_citations":[],"review_version":1}