{"id":"2a633f82-7496-4194-84b9-251cf8348bef","arxiv_id":"2607.23723","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Double-centered rank-d spectral truncation recovers normalized latent inner products from dense anisotropic Gaussian random geometric graphs at a stable-rank rate matching isotropic SOTA.","lead":"A spectral method recovers hidden pairwise similarities from a dense random geometric graph even when the latent points are anisotropic Gaussians. The error rate depends on the covariance’s stable rank and matches the best known isotropic rates while allowing badly conditioned covariances.","discovery_kind":"extension","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"No structural hole found; the genuine load-bearing risk is the correctness of the paper's new concentration machinery (Proposition 9's L²-operator-norm adaptation of [KRM25] and Lemma 7's self-bounding argument), on which all three nonlinear-term bounds rest — but the full proofs are present and int","rationale":"The reader identified the compact-t assumption as weakest; that is a parameter-regime limitation, not a correctness risk, and the paper is explicit about it. My own read locates the residual risk elsewhere — in the unverified new concentration tools (Proposition 9, Lemma 7) that sharpen [KRM25] and carry the quadratic/cubic/tail bounds — hence \"partial\" agreement rather than full agreement. But this is precisely the ordinary risk profile the reader already described (\"casework absorptions, AI-assisted cubic details\"), and I found the audited steps sound, including subtle ones (the canonical-kernel replacement in Prop. 11; the correlation-matrix identity in Lemma 13 via the Wick-tensor decomposition; the good/bad event split for β̂₁ with the clipping lower bound β̂₁≥cn⁻²). The rate assembly in §4.1 absorbs the auxiliary terms correctly as far as I checked (the τ₁≥√(τ₂r_st) inequality at (17) and the case-split absorptions following (18)). Since no concern I identified rises to load-bearing, the honest output is UNCHANGED with a concrete empirical/analytical verification path that would cheaply raise confidence in the new machinery. ACCEPT at HIGH confidence stands.","tokens_in":34952,"tokens_out":7548,"duration_ms":279925,"concrete_test":"Monte-Carlo audit of the pivotal quadratic-term bound (Proposition 11) plus end-to-end rate check. Take Σ=diag(λ_i), λ_i∝i/d (so r_st≈d/3, diverging condition number), d=100, n∈{500,1000,2000,4000}. (a) Simulate x_i∼N(0,Σ), form Z_ij=⟨x_i,x_j⟩/√τ₂, and estimate (E∥H∆⁽²⁾H∥²_op)^{1/2} over ≥200 trials; regress against n/r_st. If the empirical slope exceeds the predicted n/r_st scaling by a factor growing with n (e.g., an extra √log n or √d_eff), the double-centering analysis — the paper's central mechanism — has a hidden slack or error. (b) Run the full plug-in estimator (3) and check the MSE n⁻²∥Ŝ_d−XXᵀ/√τ₂∥²_F decays like d log n/n + 1/d within constant factors. Agreement on both would empirically confirm the new concentration machinery; a discrepancy localizes the error. Independently re-deriving Lemma 8's constants (the √2∥Σ_A∥op^{1/2} term) is the cheapest analytical cross-check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I read the proof architecture end-to-end and could not find a structural flaw. The reader's flagged weakest assumption (compact interval for t, keeping β₁=φ(t) bounded below) is real but standard for the dense constant-density regime, and the paper never claims otherwise. The places where the theorem would actually break if something were wrong are the two new technical ingredients that are not imported from prior literature: (1) Proposition 9, an adaptation of [KRM25, Thm 1] to L² control of the operator norm with logarithmic factors removed from the first two terms — this sharpening is asserted by comparison in Remark 10 and is the reason the final rate improves on [LS23] by polylog factors; (2) Lemma 7's self-bounding argument E∥YᵀY∥op ≤ C(nµ² + B_n² + µB_n√(n log n)), which feeds Proposition 11 (the quadratic term, the entire motivation for double-centering). Both proofs are given in full and the steps I audited check out (Young's-inequality absorption of M in Lemma 7; the Rademacher matrix-series step in Lemma 8; the Isserlis correlation computation G = 2Y(Σ⊗Σ)Yᵀ). I also verified the footnote-5 trivial bound, which is needed for Theorem 1 to hold unconditionally: it is only valid because ∥Π_d⁺(B)∥_F ≤ ∥B∥_F lets one bound E∥Ŝ_d∥²_F/n² via E[1/β̂₁²] ≤ C_t (using the clipping β̂₁ ≥ cn⁻² and P(E_n^c) ≤ 2e^{-c_K n}); a naive operator-norm route gives C_t·d and fails. This is fine as written but is exactly the kind of quiet step where an error would propagate. Net: ordinary residual risk for a long analytic argument, concentrated in Props 9/11/12/15, not a load-bearing objection.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper studies recovery of latent normalized inner products from a hard-threshold random geometric graph with anisotropic Gaussian latent points x_i ~ N(0,Σ): A_ij = 1{⟨x_i,x_j⟩ ≥ ζ}, with ζ chosen so the edge density is constant order. Because anisotropy amplifies degree fluctuations that are comparable in operator norm to the Gram-matrix signal, the authors analyze a doubly centered adjacency matrix HAH and estimate XX^⊤/√τ₂ via a rank-d positive spectral truncation of a plug-in-scaled version of HAH. Theorem 1 gives a non-asymptotic expected Frobenius bound C_t(dτ₁² log n)/(nτ₂ r_st) + C_t d/r_st², which reduces to O(d log n/n + 1/d) when the stable rank r_st ≍ d — matching the independent isotropic work [FZ26], improving [LS23] by polylogarithmic factors, and permitting diverging condition number. The proof Hermite-expands the doubly centered matrix, treats the quadratic term via canonical (Hoeffding-projected) replacement, the cubic term via an exact tensor-algebra correlation computation, and the degree ≥ 4 residual as a single kernel matrix, all through an L²-operator-norm adaptation (Proposition 9) of the decoupling argument of [KRM25].","tokens_in":35342,"tokens_out":3423,"duration_ms":145465,"significance":"If correct, this is a solid contribution: the first inner-product recovery guarantee for anisotropic Gaussian random geometric graphs whose rate depends on the spectrum of Σ only through effective-dimension quantities (stable rank r_st, effective dimension d_eff, τ₁), thereby covering ill-conditioned covariance with diverging condition number — a regime where the naive (uncentered) spectral method provably fails, as the paper's own lower bound (Eq. 47) demonstrates. Strengths worth naming: the derivation is fully non-asymptotic and self-contained apart from standard inequalities (Hanson–Wright, Rosenthal, Rademacher matrix series, Efron–Stein); the two genuinely new technical ingredients — the L² version of the KRM25 decoupling bound with logarithmic factors removed from the first two terms (Proposition 9, Remark 10) and the self-bounding argument for E‖YᵀY‖_op (Lemma 7) — are proved in full; and the achieved condition n ≫ d log n nearly matches the rate-distortion impossibility result of [MZ24], making the rate near-optimal over any covariance class containing Σ = I_d. The honest treatment of the quadratic term's obstruction (Section 5.3) is a nice touch that justifies the double-","major_comments":[{"comment":"Proposition 9 is the load-bearing new tool (all three nonlinear-term bounds, and hence the polylog improvement over [LS23], pass through it), and the paper carefully re-proves every other step of the adaptation. However, Eq. (45) — the identity E‖G − G₀‖_op = n E[h(x₁)²] used to convert the canonical-kernel bound (40) back to uncentered quantities — is asserted solely by the sentence 'The proof of Theorem 1 of [KRM25] shows that...'. Since this is an equality (not an inequality) feeding a triangle inequality at the final assembly step, and since the target audience of this journal should be able to verify the adaptation without cross-reading a 2025 preprint, I ask the authors to include a short self-contained derivation (a few lines: G − G₀ is a rank-structured matrix built from g(x_i)g(x_j) and h-terms, whose operator norm is deterministic up to the h(x₁) moment). This is a local gap in","section":"§5.2, proof of Proposition 9, Eq. (45)"}],"minor_comments":[{"comment":"Typo: 'Too see this' should be 'To see this'.","section":"§1, third paragraph"},{"comment":"The symbol B_n is overloaded: Eq. (8) defines B_n := τ₁ + √(τ₂ log n) + µ log n, while Proposition 9 and the proofs of Lemmas 7, 12 and 15 use B_n² for the conditional-variance-type quantity E max_j Σ_i {k(z_j,x_i) − E_x k}². The collision is particularly confusing inside the proof of Lemma 7, where both meanings appear ('the quantity B²_n defined in Proposition 9 is given by ...' immediately after B_n was used in sense (8)). Please rename one of them.","section":"§4, Eq. (8) vs. §5.2, Proposition 9"},{"comment":"The notation section already warns that 1 denotes both the all-ones vector and the indicator function; in Eq. (2) and the definition of ∆⁽ᵏ⁾ the two uses appear in the same display. Consider a distinct indicator symbol.","section":"§3.1 / Notation"},{"comment":"The footnote establishing E‖Ŝ_d‖²_F/n² ≤ C_t (needed to state Theorem 1 unconditionally, without the regime condition (9)) is correct but compressed: it silently uses ∥Π_d⁺(B)∥_F ≤ ∥B∥_F together with the clipping bound β̂₁ ≥ cn⁻² and P(E_n^c) ≤ 2e^{−c_K n}. Since this is exactly the step that lets Theorem 1 avoid assuming the rate is already small, one or two sentences of elaboration would help.","section":"§4.3, footnote 5"},{"comment":"The restriction t = ζ/√τ₂ ∈ fixed compact interval is used in at least three distinct places (β₁ = φ(t) bounded away from zero; the Taylor bounds in Lemma 16; the bias bound in Lemma 21). A single remark collecting these uses — and noting explicitly that the sparse regime p → 0 is outside the method because β₁ vanishes — would make the scope of the result clearer to readers arriving from [LS23]/[FZ26].","section":"§3.2, Theorem 1 assumption"},{"comment":"In Lemma 8, the final line applies Young's inequality to pass from M ≤ ‖Σ_A‖_op + C√(log N) b √M to the stated bound; the intermediate inequality √M ≤ √(2‖Σ_A‖_op) + C√(log N) b is only valid after absorbing the (1/2)M term, so displaying one more line would avoid the appearance of a sign error.","section":"§5.2, Lemma 8"},{"comment":"Equation (47) lower-bounds E‖∆⁽²⁾_op by cn/√d_eff and then compares to the signal n/√r_st; the conclusion 'noise and signal are comparable whenever d_eff ≍ r_st' would be sharper if stated as d_eff ≲ r_st (which holds always, Eq. (7)), i.e., the obstruction is generic rather than restricted to well-conditioned Σ.","section":"§5.3, Eq. (47)"}],"recommendation":"minor_revision","confidential_remarks":"Two process notes for the editor. (1) One author is a co-author of the cited [KRM25], whose decoupling theorem is the main external input; the paper adapts and re-proves it rather than citing it as a black box, and the overlap is transparently handled, so I see no issue — but the independence of the verification of Eq. (45) is the one place where the paper leans on the authors' own prior preprint without proof (see Major Comment 1). (2) The manuscript contains an explicit declaration that GPT-5.5/GPT-5.6 assisted with preliminary calculations and with completing proof details in Proposition 12. The disclosure is commendable and the proofs I audited are internally consistent, but the editor may wish to confirm the journal's policy on AI-assisted proof generation. I also note the simultaneous independent work [FZ26] is cited and properly scoped; novelty relative to it (anisotropic Σ, Hermite-based rather than spherical-harmonic treatment of the noise) seems genuine."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is that this closes the dense anisotropic hard-threshold recovery gap left by LS23 and the concurrent isotropic sharpening in FZ26. Theorem 1 gives an explicit MSE in the stable rank of Σ that recovers the isotropic rate when r_st ≍ d and still works when the condition number diverges (e.g., eigenvalues i/d). That is the actual contribution.\n\nWhat they do well is mechanical and clear. Degree fluctuations from the quadratic form x_i^⊤ Σ x_i are real and amplified by anisotropy; double-centering HAH kills the rank-one obstruction that otherwise makes the quadratic Hermite term as large as the signal. They then split the residual into quadratic, cubic, and whole-tail pieces, control each with an L^{2} adaptation of the KRM25 decoupling (their Prop. 9), and convert operator-norm denoising into Frobenius recovery by standard rank-d truncation. The reduction chain (Props. 2–3, Lemmas 4 and 20) is clean. The lower-bound calculation showing that uncentered Δ^(2) is too big is honest motivation, not decoration. Citations are appropriate; the overlap with KRM25 is disclosed and the tool is re-proved for the L^{2} setting they need.\n\nSoft spots are ordinary for a long analytic argument, not structural. Everything rests on the new concentration pieces (Prop. 9’s log-factor removal, Lemma 7’s self-bound for the quadratic feature matrix, and the cubic/tail correlation estimates). I audited the Young absorption, the Isserlis step, and the Rademacher matrix-series bound; they check. The compact-interval assumption on t is real but standard for constant-density dense analyses and is never hidden. No experiments, as expected. Instance-optimal rates for a fixed Σ sequence are left open, which they say.\n\nThis is for people who already care about high-dimensional geometric graphs and spectral kernel approximation. It deserves a serious referee. I would cite the stable-rank statement and the double-centering observation. Send it out.","headline":"Solid anisotropic extension of spectral inner-product recovery; stable-rank rate is the right object and the double-centering + Hermite/decoupling argument holds up on the page.","tokens_in":35817,"tokens_out":521,"would_cite":true,"duration_ms":10518,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H12","05C80","60B20"],"pacs":[],"model":"grok-4.5","headline":"Double-centering plus a rank-d spectral cut recovers latent inner products from anisotropic Gaussian geometric graphs at the isotropic rate, even when the covariance is badly conditioned.","keywords":["random geometric graphs","latent inner-product recovery","anisotropic Gaussian","double centering","spectral estimator","Hermite expansion","stable rank","decoupling"],"falsifier":"Generate anisotropic Gaussian geometric graphs with stable rank ~ d, constant edge density, and n slightly larger than d log n; compute the normalized Frobenius error of the doubly-centered rank-d spectral estimator. If the error fails to tend to zero while the isotropic comparator succeeds, the claimed rate is false.","tokens_in":35541,"feed_emoji":"📐","tokens_out":967,"duration_ms":17206,"temperature":0.7,"pith_summary":"This paper asks whether you can reconstruct the hidden pairwise inner products of high-dimensional Gaussian points from a single hard-threshold geometric graph, when the points are anisotropic rather than spherical. The obstacle is that anisotropy (and even random norms in the isotropic case) creates large degree fluctuations that can drown the Gram-matrix signal. The authors remove those fluctuations by doubly centering the adjacency matrix, then take its top-d spectral approximation. They prove a mean-squared error bound controlled by the stable rank of the latent covariance: whenever that stable rank is order d, the rate matches the best known isotropic rate and vanishes as soon as n is a little larger than d log n. The same bound continues to hold for covariances whose condition number diverges, so the method does not require well-conditioned geometry. The argument expands the centered threshold kernel in Hermite polynomials and controls the nonlinear remainder by a recent decoupling technique rather than the classical trace method.","feed_headline":"Spectral recovery works for anisotropic geometric graphs","feed_subtitle":"Double-centering matches the isotropic rate even when the latent covariance is badly conditioned","key_machinery":"The doubly centered adjacency matrix HAH together with its entrywise Hermite expansion: double centering kills the large quadratic degree term, the linear Hermite piece supplies the Gram signal, and a decoupling argument bounds the quadratic, cubic, and whole higher-order residual in operator norm.","core_discovery":"For i.i.d. latent points x_i ~ N(0, Σ) and a hard-threshold graph of constant edge density, the rank-d spectral approximation of the doubly centered adjacency matrix recovers the normalized Gram matrix XX^⊤/√τ₂ with expected Frobenius error at most C_t (d τ₁² log n)/(n τ₂ r_st) + C_t d/r_st². In particular, when the stable rank r_st is comparable to d the error is O(d log n / n + 1/d), matching the isotropic state of the art and remaining valid for ill-conditioned Σ.","pith_inferences":["The stable-rank dependence suggests that recovery thresholds for other anisotropic latent-space models (e.g., spiked or low-effective-rank covariances) may be governed by effective dimension rather than ambient d.","Because the argument never uses well-conditioning beyond the stable rank, the same double-centering step is a candidate preprocessing step for spectral community detection or kernel clustering under heterogeneous degree patterns.","An instance-optimal rate that tracks the full spectrum of Σ, rather than only its stable rank, remains open and would sharpen the minimax picture."],"forward_implications":["Strong recovery of normalized latent inner products is possible under the same n ≫ d log n condition that is nearly information-theoretically necessary in the isotropic case.","The estimator remains valid for covariances whose condition number diverges, provided the stable rank stays order d.","Degree correction by explicit double centering is sufficient; one need not discard eigenpairs adaptively.","The same Hermite-plus-decoupling analysis supplies an operator-norm denoising bound that converts directly into Frobenius recovery after rank-d truncation."],"fun_headline_variants":["Double-centering recovers latent products in anisotropic geometric graphs","Rank-d spectral estimate matches isotropic rate for anisotropic graphs","Doubly centered adjacency recovers Gram matrix despite ill-conditioned Σ","Stable-rank rate for inner-product recovery from anisotropic RGG","Hermite expansion yields isotropic-rate recovery under anisotropy"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The normalized threshold that sets the edge density must stay inside a fixed compact interval, so the linear Hermite coefficient stays bounded away from zero; if the threshold drifts the signal coefficient vanishes and the stated rate no longer holds.","fun_headline_variants_meta":{"raw":{"variants":["Double-centering recovers latent products in anisotropic geometric graphs","Rank-d spectral estimate matches isotropic rate for anisotropic graphs","Doubly centered adjacency recovers Gram matrix despite ill-conditioned Σ","Stable-rank rate for inner-product recovery from anisotropic RGG","Hermite expansion yields isotropic-rate recovery under anisotropy"]},"model":"grok-4.5","effort":"low","cost_usd":0.003992,"raw_usage":{"total_tokens":1283,"prompt_tokens":872,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":39924000,"prompt_tokens_details":{"text_tokens":872,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":327,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":872,"tokens_out":84,"duration_ms":7765,"temperature":1.0,"reasoning_tokens":327,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T14:46:55.121737+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Generate anisotropic Gaussian geometric graphs with stable rank ~ d, constant edge density, and n slightly larger than d log n; compute the normalized Frobenius error of the doubly-centered rank-d spectral estimator. If the error fails to tend to zero while the isotropic comparator succeeds, the claimed rate is false.","supporting_citations":[],"review_version":1}