{"id":"dbfc6493-ae81-4b46-b0ef-e89f9fcfe67e","arxiv_id":"2505.22015","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A debiased, sparsity-adaptive distributed PCA algorithm achieves consistency under finite sixth moments and outperforms prior methods, especially with few machines.","lead":"This paper proposes a debiased distributed PCA method that corrects local eigenvector bias before aggregating across machines, and shows it is consistent under weaker assumptions than prior work. The method detects sparse eigenvectors and gives much lower estimation error when few machines are available.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5's fixed-m consistency for sparse eigenvectors is restricted by Assumption 4 to bounded support: since every strong entry exceeds t0>0 and ||u_i||=1, at most 1/t0^2 coordinates can be strong.","rationale":"The paper is methodologically serious: the aggregate eigenvector bias result in Theorem 1 is plausible under finite sixth moments, and the simulations strongly support the method for eigenvectors with fixed support and well-separated entries. The reader's weakest-assumption diagnosis pointed at Assumption 4, and my stress-test agrees that this is the load-bearing condition, but I would sharpen it. Assumption 4 does not merely require a gap between strong and weak entries; it forces the strong entries to be bounded below by a positive constant t0. Because each population eigenvector has unit norm, this bounds the strong support size by 1/t0^2. Consequently, Theorem 5's 'sparse' result applies to fixed-support sparsity, not to the growing-support sparsity that is usually studied in high-dimensional sparse PCA. If the authors intend the latter, the theorem as stated does not deliver it; if they intend only fixed support, the abstract and introduction should say so explicitly. This explains why all simulations use support sizes 4 and 6, where entries are about 0.4-0.5, far above the threshold t=0.1. The concrete test proposed would settle whether the method fails in the growing-support regime. I would keep the reader's CONDITIONAL verdict, with the additional condition that the authors either prove a growing-support analogue under appropriate scaling or qualify the sparsity claim in the abstract to fixed-support sparsity. The concern is a limitation of scope rather than a detected mathematical contradiction, so it does not warrant rejection.","tokens_in":14350,"tokens_out":16020,"duration_ms":196135,"concrete_test":"Simulate the spiked model with p=5000, n_l=2500 per machine, K=2, and growing-support eigenvectors: u_1 supported on s=ceil(p^{1/2}) coordinates with equal entries s^{-1/2}, u_2 supported on a disjoint set of the same size, and a bulk spectrum satisfying Assumption 3. Run Algorithm 1 with t=0.1 as in Section 5, and also with t=s^{-1/2}/2, for fixed m=3 and m=300. Record the Frobenius error rho=||U_K U_K^T - estimated U_K U_K^T||_F and the fraction of true support recovered in bA(t). If rho does not tend to 0 for fixed m, or if the sparse branch condition ||u_i1||^2 >= 1 - 2/(m^{1/4} p^{1/2}) is never met, then Theorem 5 does not cover growing-support sparsity.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Assumption 4 in Section 4.2 requires a fixed t0>0 with min_{j in A_i} |u_ij| > t0 and max_{j not in A_i} |u_ij| -> 0. Since sum_{j in A_i} u_ij^2 <= ||u_i||^2 = 1, the strong support satisfies |A_i| <= 1/t0^2. Thus, despite the phrasing of Theorem 5, the 'sparse vectors' covered are vectors with O(1) nonzero coordinates, not vectors whose support size grows with p. In the standard growing-support sparse regime, nonzero entries are of order s^{-1/2} -> 0, so no fixed threshold t can separate strong from weak entries. Algorithm 1-2's majority vote will then miss an increasing fraction of the signal, and the sparse branch in Algorithm 1-3 will zero out a non-vanishing fraction of the eigenvector norm. The simulations in Section 5 only use fixed support sizes 4 and 6 (plus a dense weak component in the mixed case), so the claimed advantage 'when the leading eigenvectors are sparse' has not been demonstrated for growing-support sparsity. This is a scope restriction on the central claim, not a disagreement with existing consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies horizontal distributed PCA under a high-dimensional spiked covariance model, where local sample eigenvectors are biased. It proposes a three-stage procedure: local eigendecomposition with a debiasing factor θ; identification of 'strong signal' coordinates by thresholding |\\u005cu005cu005cscap |\\u005cu005cu005chat u_{ij}/θ| and majority voting across machines; and recovery that combines a debiased rank-one aggregation with sparsity and orthogonality refinements. The main theoretical claims are consistency of the estimated eigenspace as m→∞ under only a finite sixth moment and without symmetric innovations (Theorems 1–4), and consistency for fixed m when the eigenvectors are sparse (Theorem 5); extensions to correlation matrices are also stated. Simulation experiments and a MiniBooNE data analysis report smaller Frobenius errors than the methods of Fan et al. (2019) and Li et al. (2025).","tokens_in":14585,"tokens_out":8726,"duration_ms":87076,"significance":"If the omitted proofs are correct, the paper makes a useful contribution: it removes the symmetric-innovation and sub-Gaussian assumptions for the consistency of distributed eigenspace estimation, provides a debiasing step that visibly improves small-m performance, and gives a clear algorithm with reproducible simulation designs. The paper is also commendable for comparing against two relevant baselines, treating both covariance and correlation matrices, and including a real-data experiment. The main caveat is that the fixed-m sparse result is much narrower than the abstract suggests, and the central proof machinery is not available in the preprint for verification.","major_comments":[{"comment":"Theorem 5, combined with Assumption 4, only covers sparse eigenvectors with bounded support. Because min_{j in A_i} |u_{ij}| > t_0 and ||u_i||=1, the strong support satisfies |A_i| ≤ 1/t_0^2, independent of p. In the standard growing-support sparse regime, the nonzero entries are of order s^{-1/2} -> 0, so no fixed threshold t can separate them from weak entries; the majority vote in Algorithm 1-2 will then misclassify a non-vanishing fraction of the signal, and the sparsity refinement in Algorithm 1-3 will zero out a non-vanishing fraction of the eigenvector norm. The simulations in Section 5 use only fixed supports of size 4 and 6 (plus a dense weak component in the mixed case), so the claimed advantage 'when the leading eigenvectors are sparse' is not demonstrated for growing-support sparsity. The abstract and Theorem 5 should be restricted to the bounded-support case, or the algorithm should be extended to handle support size growing with p.","section":"§4.3 (Theorem 5) and §4.2 (Assumption 4)"},{"comment":"The signal identification step depends on a single user-specified threshold t, and the sparsity refinement uses the hand-set constant 2 m^{-1/4} p^{-1/2} in (5). The paper uses t=0.1 in all simulations and t=0.005 in the real-data experiment, but gives no data-dependent selection rule and no sensitivity analysis. Since Theorems 3–5 require Assumption 4 with t_0>t, the practical validity of the consistency claims is contingent on an unvalidated tuning choice. Please provide a sensitivity study over t and, if possible, an automatic or theoretically guided choice of t and of the constant in (5).","section":"§3.2 and §5 (threshold t and criterion (5))"},{"comment":"All proofs of the main theorems are deferred to the Supplementary Material, which is not included in the submission. The central claims—in particular the finite-sixth-moment consistency of Theorem 1 and the fixed-m sparse consistency of Theorem 5—therefore cannot be verified from the preprint as it stands. The supplementary material must be provided, and at least the key steps (e.g., the decomposition (7) and the control of the averaged residual terms) need to be visible to the reader.","section":"§4 (Theorems 1–6; Supplementary Material)"},{"comment":"There is a gap between the theorem stated in the main text and the estimator whose consistency is claimed. Theorem 1 concerns the average of the raw projection matrices \\u005cu005cu005chat u_i^{(\\u005cuel)} \\u005cu005cu005chat u_i^{(\\u005cuel)\\u005cutop}, whereas Algorithm 1-3 uses the debiased rank-one matrices [\\u005ctheta_i^{(\\u005cuel)}]^{-2} \\u005cu005cu005chat u_i^{(\\u005cuel)} \\u005cu005cu005chat u_i^{(\\u005cuel)\\u005cutop} and then applies thresholding and sparsity refinement. The consistency of \\u005ccheck U_K in Theorem 4 does not follow from Theorem 1 alone unless additional arguments (e.g., uniform control of \\u005ctheta_i^{(\\u005cuel)} and of the threshold indicators) are supplied. Please state explicitly how Theorem 1, or a modified version of it, implies Theorems 4–5, or add the missing intermediate theorem.","section":"§4.1 (Theorem 1) and §3.3 (Algorithm 1-3)"}],"minor_comments":[{"comment":"The sparsity condition is written as u_{i1}^\\utop u_{i1} > 1 - 2 m^{-1/4} p^{-1/2} in (5) but as \\u005c|u_{i1}\\u005c|^2 \\u005cgeq 1 - 2/(m^{1/4} p^{1/2}) in Algorithm 1-3; please make the strict/weak inequality and the notation for the threshold consistent.","section":"§3.3"},{"comment":"The displayed expression contains 'U_KUk', which should be 'U_K U_K^\\utop'.","section":"§4.3, Theorem 6"},{"comment":"The symbol \\u005casymp is used informally in the definition of strong signals; since Assumption 4 builds on this distinction, please define the notation precisely before the assumptions.","section":"§2.3 and §4.2"},{"comment":"The statement that existing distributed PCA methods require sub-Gaussianity and symmetric innovations is not reconciled with He et al. (2024), which is cited as working 'without moment constraints'; please clarify the exact comparison class of assumptions.","section":"§1 and §4.1"},{"comment":"Figure 6 reports only the average AR over 1000 replications; adding boxplots or standard errors would make the claimed superiority over the baselines more convincing.","section":"§6, Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The central idea seems promising and the simulations are useful, but the sparse-eigenvector claim is considerably narrower than advertised and the proof supplement is essential. In a revision, the authors should also clarify the relation to He et al. (2024) and to the prior debiasing result of Yin and Zhou (2024), since the reader needs to see which components of the algorithm are genuinely new relative to those works."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know upfront. First, this paper is a serious contender: it fixes a real bias problem in distributed PCA when the number of machines m is small, and the proposed debiasing-and-sparsity-detection scheme is genuinely new, not a repackaging of Fan et al. (2019) or Li et al. (2025). Second, its headline claim about sparse eigenvectors is narrower than it reads. The sparse consistency result (Theorem 5) rests on Assumption 4, which demands a fixed threshold t0 separating strong from weak entries; since the eigenvector has unit norm, the strong support can have at most 1/t0^2 coordinates. So the result covers eigenvectors with O(1) nonzeros, not the growing-support case where entries decay as s^{-1/2}. The simulations only use supports of size 4 and 6, so the practical advantage for genuinely high-dimensional sparse vectors is not demonstrated. That is a scope restriction, not a fatal flaw.\n\nWhat the paper does well: the local debiasing step using the Yin-Zhou (2024) relation is a natural and plausible fix, and the majority-vote for signal identification is a reasonable way to aggregate local information. The main consistency theorem (Theorem 1) improves on Fan et al. (2019) by requiring only a finite sixth moment and dropping the symmetric-innovation assumption — a real generalization. The simulations agree: the method reduces error substantially when m is small, and it also handles exponential data, which is more than many competitors can claim.\n\nWhere I would push back: all proofs are deferred to a supplementary file that is not in the arXiv version, so the central claims cannot be verified from the preprint. The tuning parameters are set by hand (t = 0.1 in simulation, t = 0.005 on real data) with no guidance on how to choose them. No code or standard errors are provided. None of this is disqualifying for a preprint, but it matters for reproducibility.\n\nThe debiasing factor comes from the authors' own prior work; that is fine because the relation is cited and not fitted to the current problem. The circularity concern in the reader's report does not land once you read the algorithm.\n\nMy bottom line: I would send this to a serious referee. The method is likely correct and fills a real gap, and the theory is testable. The referee should ask for the supplementary proofs, a clear statement of the sparsity regime, and code. Whoever reviews it should push on Theorem 5's interpretation.","headline":"A genuinely new debiased distributed PCA with a real but bounded sparsity claim; worth serious referee time, but the sparse consistency result overstates its scope.","tokens_in":15137,"tokens_out":2454,"would_cite":true,"duration_ms":23168,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H25","60B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Distributed PCA can be made consistent in high-dimensional spiked models by debiasing each machine's eigenvector estimates before aggregating, and for sparse eigenvectors this consistency holds even with a fixed number of machines.","keywords":["distributed PCA","high-dimensional spiked model","eigenvector bias correction","random matrix theory","sparse eigenvectors","finite sixth moment","correlation matrix","majority voting"],"falsifier":"Simulate a spiked model whose first eigenvector has entries decaying as $u_j \\propto j^{-\\alpha}$ with small $\\alpha$, so that no fixed threshold separates strong from weak entries, and run the algorithm with several thresholds $t$ at large $N$ and small $m$; if the estimated support $\\hat A(t)$ continues to misclassify coordinates and the error $\\rho$ fails to vanish while an oracle estimator with the true support is consistent, the core separation premise fails.","tokens_in":14129,"feed_emoji":"📊","tokens_out":10848,"duration_ms":94117,"temperature":0.7,"pith_summary":"Under a high-dimensional spiked model—a covariance structure with a few separated leading eigenvalues—sample eigenvectors deviate from their population counterparts by a persistent angle, so averaging local PCA results across machines inherits a bias unless the number of machines diverges. The paper proposes to remove that bias locally: each machine rescales its sample eigenvectors by a correction factor derived from its own sample eigenvalues before the estimates are aggregated. The central claim is that this debiased aggregation is consistent as the number of machines tends to infinity under only a finite sixth moment and without symmetric innovations, conditions much weaker than those required by existing distributed PCA methods. When the leading eigenvectors are sparse, the paper further claims consistency for any fixed number of machines, which is the regime where the bias problem is most severe.","feed_headline":"Debiased distributed PCA needs few machines and no symmetric data","feed_subtitle":"A local bias correction plus a majority-vote sparsity step removes the need for many machines and symmetric data.","key_machinery":"The load-bearing object is the correction factor $\\theta_i^{(\\ell)}$ of equation (2), a function of the local sample eigenvalues that quantifies the asymptotic shrinkage of sample eigenvector entries relative to population entries; it enters the identity $|\\hat u_{ij}/\\theta_i|=|u_{ij}|+o_p(1)$, which the algorithm uses to debias each machine. The second mechanism is the decomposition of each local projection matrix into a signal part $(u_i^\\top\\hat u_i)^2 u_i u_i^\\top$ and a residual $R_n$, with the property that residual averages vanish as $m\\to\\infty$. The third mechanism is the signal-identification step: threshold indicators on the rescaled entries are sent to a central node, which takes a majority vote to decide which coordinates are strong; strong coordinates are then estimated from the debiased local projections, while weak coordinates are filled in from the aggregated matrix and a sparsity criterion decides whether to keep a sparse support. These three pieces together carry the consistency argument.","core_discovery":"The paper's central discovery is an entrywise bias-correcting relation: under the spiked model, $|\\hat u_{ij}/\\theta_i|=|u_{ij}|+o_p(1)$, where $\\hat u_{ij}$ is a local sample eigenvector entry, $u_{ij}$ is the corresponding population entry, and $\\theta_i$ is a correction factor computed from local sample eigenvalues. Dividing by $\\theta_i$ before aggregation converts biased local projections into approximately unbiased rank-one estimates of $u_i u_i^\\top$. The paper proves that the residual terms in the projection-matrix decomposition cancel in the cross-machine average, yielding consistency as $m\\to\\infty$ (Theorem 4), that the majority-vote signal-identification step recovers the strong coordinates for any $m$ (Theorem 3), and that for sparse eigenvectors the final estimator is consistent even for fixed $m$ (Theorem 5). The same guarantees are extended to the sample correlation matrix.","pith_inferences":["Editorial inference: the threshold-separation structure of Assumption 4, not the moment condition, is what makes signal identification work; datasets whose eigenvector entries decay smoothly rather than in two separated groups could violate the majority-vote assumption even when all moment conditions hold.","Editorial inference: the same entrywise identity suggests a natural way to build debiased distributed sparse PCA—use $\\hat A(t)$ as an initial support estimate and then run a local refinement step on the strong coordinates, which may improve downstream sparse loading recovery beyond the paper's current estimator.","Editorial inference: because the proof mechanism is a resolvent-plus-contour-integration decomposition of projection matrices, the debiasing strategy should extend to other spectral functionals of the spiked model, such as subspace distances or linear statistics of eigenvectors, not only to the top-$K$ projection."],"forward_implications":["Distributed PCA implementations can drop the symmetric-innovation and sub-Gaussian assumptions; finite sixth moments suffice for consistency of the debiased estimator as $m\\to\\infty$.","For sparse leading eigenvectors, a fixed number of machines is enough: the estimation error $\\rho$ converges to zero as the local sample sizes grow, removing the usual requirement that the machine count diverge.","The correction step uses only each machine's own eigenvalues and eigenvectors and communicates only threshold votes and rescaled projections, so the communication cost stays comparable to existing averaging methods.","The same debiased aggregation applies to the sample correlation matrix, so the method can be used after standardization without new theory."],"supporting_citations":[{"why":"Derives the entrywise relation between rescaled sample eigenvector entries and population entries that the debiasing step exploits.","marker":"Yin and Zhou (2024)"},{"why":"Establishes the almost-sure limit of the inner product between sample and population eigenvectors, quantifying the persistent bias the method corrects.","marker":"Benaych-Georges and Nadakuditi (2011)"},{"why":"Defines the averaging-based distributed PCA baseline whose consistency the paper extends and whose sub-Gaussian and symmetric-innovation assumptions are relaxed.","marker":"Fan et al. (2019)"},{"why":"Provides the two-round distributed PCA baseline compared in simulations and real-data experiments.","marker":"Li et al. (2025)"},{"why":"Gives the asymptotic behavior of sample eigenvectors in the spiked covariance model that motivates the bias-correction framework.","marker":"Paul (2007)"},{"why":"Supplies deterministic equivalents for sample eigenvalues and eigenvectors used in deriving the correction factor.","marker":"Mestre (2008)"}],"fun_headline_variants":["Debiased PCA: tiny clusters, no symmetry, still consistent","Local bias correction makes sparse PCA work on few machines","Distributed PCA that needs neither symmetry nor many machines","Bias-corrected PCA aggregator: survives with small machine count","Sparse-friendly debiased PCA: fewer machines, fewer assumptions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that a single fixed threshold can separate strong from weak eigenvector entries—strong entries uniformly above $t_0$ and weak entries uniformly tending to zero—so that the majority vote across machines classifies coordinates correctly.","fun_headline_variants_meta":{"raw":{"variants":["Debiased PCA: tiny clusters, no symmetry, still consistent","Local bias correction makes sparse PCA work on few machines","Distributed PCA that needs neither symmetry nor many machines","Bias-corrected PCA aggregator: survives with small machine count","Sparse-friendly debiased PCA: fewer machines, fewer assumptions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1606,"prompt_tokens":925,"completion_tokens":681,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":595}},"tokens_in":541,"tokens_out":681,"duration_ms":6904,"temperature":1.0,"reasoning_tokens":595,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:17:32.031194+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a spiked model whose first eigenvector has entries decaying as $u_j \\propto j^{-\\alpha}$ with small $\\alpha$, so that no fixed threshold separates strong from weak entries, and run the algorithm with several thresholds $t$ at large $N$ and small $m$; if the estimated support $\\hat A(t)$ continues to misclassify coordinates and the error $\\rho$ fails to vanish while an oracle estimator with the true support is consistent, the core separation premise fails.","supporting_citations":[{"cited_title":"and Zhou, W","cited_arxiv_id":null,"evidence_quote":"Derives the entrywise relation between rescaled sample eigenvector entries and population entries that the debiasing step exploits."},{"cited_title":"and Nadakuditi, R","cited_arxiv_id":null,"evidence_quote":"Establishes the almost-sure limit of the inner product between sample and population eigenvectors, quantifying the persistent bias the method corrects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the averaging-based distributed PCA baseline whose consistency the paper extends and whose sub-Gaussian and symmetric-innovation assumptions are relaxed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the asymptotic behavior of sample eigenvectors in the spiked covariance model that motivates the bias-correction framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies deterministic equivalents for sample eigenvalues and eigenvectors used in deriving the correction factor."}],"review_version":1}