{"id":"c42485e7-3d5b-43e0-b6f4-a430f7ccce48","arxiv_id":"2412.11950","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A supplementary document with proofs and simulations for an asynchronous distributed GP method, whose periodic-kernel Lipschitz constant derivation contains a mathematical error.","lead":"This document is a companion to a paper on asynchronous distributed Gaussian process regression, supplying the proof steps for prediction and tracking error bounds, kernel Lipschitz constants, and extra simulations. It matters to readers who need to audit the mathematical support behind the companion claims, since the supplementary proofs are where the derivations can break.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 9's control bound relies on an unproved spectral assumption: A must be diagonalizable with real negative eigenvalues for inequality (20) and hence bound (24) to hold as stated.","rationale":"The reader's weakest_assumption focused on the unverified inherited premises (Lemma 2, Lemma 3, Algorithm 1) and on the periodic-kernel derivative error. I agree those are important, but the single most load-bearing concern internal to this document is the spectral assumption in Theorem 9: the control guarantee (24) is the headline application, and its proof depends on a clean exponential bound for exp(At). The document neither states nor proves that A has real negative eigenvalues and is diagonalizable. This is an omission, not necessarily a false claim; the theorem may be repairable by switching to real parts and assuming diagonalizability, or by using a more general matrix-exponential bound. But as written, the proof does not cover A that are Hurwitz with complex or defective eigendata. Because the required fix is plausible and localized, the appropriate verdict remains CONDITIONAL rather than REJECT or UNVERDICTED. The periodic-kernel error is real but secondary: the central prediction theorem is stated for kernels satisfying Lemma 3, and the control experiments use ARD-SE, so an incorrect Lipschitz constant for the periodic kernel does not collapse the main claim. I also found the algebraic chain in Theorem 7's proof to be structurally sound given its stated premises, and Corollary 8's monotonicity argument is correct once η_i^k(t) < βσ_f is guaranteed for all weighted predictions. Overall, the reader's conditional verdict is appropriate; my concern sharpens the specific proof gap that should be closed before acceptance.","tokens_in":13754,"tokens_out":11527,"duration_ms":110513,"concrete_test":"Inspect the companion paper's eq. (22) defining A. Compute its eigenvalues and check diagonalizability for the control task. If A has complex-conjugate or repeated defective eigenvalues, re-derive Theorem 9 using the Schur/Jordan form and determine whether bound (24) still holds with an additional constant; if it does not, Theorem 9 as stated must be revised. Separately, recompute the periodic-kernel Lipschitz constant by maximizing |∂κ/∂d| = (4πσ_f²/(pσ_l²)) sqrt(u(1-u)) exp(-2u/σ_l²) over u ∈ [0,1] and compare with Table I.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The control result, Theorem 9, is the applied payoff of the paper, and its proof is not valid as written for a general Hurwitz matrix. In Section III, the proof writes A = QΛQ^{-1} with Λ = diag(Λ_1,...,Λ_n) and Λ_i ∈ R_{<0}, then uses ∥exp(At)∥ ≤ ∥Q∥∥Q^{-1}∥ exp(Λ̄t). No justification is given that the A defined in eq. (22) of the companion paper is diagonalizable with real negative eigenvalues. A Hurwitz matrix can have complex-conjugate eigenvalues or be defective. If A has complex eigenvalues, the statement 'Λ_i ∈ R_{<0}' is false, and the proof would need to use max real parts; the conclusion may survive but only with additional argument. If A is defective, exp(At) includes polynomial-in-t factors, so the clean exponential bound (20) and the ultimate bound (24) are not established. The proof also drops the norm of the input vector b and writes vector inequalities as scalar ones, though these are secondary. A separate but real error is Corollary 4: the derivative of the periodic kernel in eq. (48) omits the cos factor from differentiating sin²(πd/p), and the line 'due to the fact s(x,x') ≥ 1' is backwards (|s| ≤ 1). The periodic kernel's Lipschitz constants in Table I are therefore invalid. This does not by itself undermine the control experiments, which use the ARD-SE kernel, but it is a concrete mathematical defect in the kernel catalogue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This complementary document supplies proofs and additional simulations for a companion paper on asynchronous distributed Gaussian process regression. It proves a prediction-error bound |f(x(t)) - fhat(x(t))| <= omega(t) with omega(t) <= beta*sigma_f (Theorem 7, Corollary 8), derives a closed-loop tracking-error bound for a control law with error dynamics e_dot = Ae + b(f - fhat) (Theorem 9, eq. (24)), gives Lipschitz constants for several kernels (Table I, Corollaries 1-4), and reports delay-time and Monte-Carlo control simulations comparing AsyncDGP with BCM, rBCM, POE, gPOE, and MOE. The paper is explicitly a companion to [1] and relies on that paper for definitions of aggregation weights, the information-set rule, and supporting lemmas.","tokens_in":14079,"tokens_out":2558,"duration_ms":25583,"significance":"If the results are correct, the paper provides a formal time-varying error certificate for fusing stale local GP predictions in real-time control, which is a valuable contribution: the prediction bound is derived analytically rather than fitted, and the control bound gives a quantitative ultimate tracking error. The proof of Theorem 7 is internally coherent under the weight definitions inherited from [1], and the supplementary simulations illustrate the claimed advantage of aggregating delayed information. However, the control theorem's proof contains a load-bearing unproved spectral assumption, and the periodic-kernel Lipschitz derivation contains a concrete mathematical error. These issues do not necessarily invalidate the central prediction-bound result, but they prevent the paper from being accepted in its current form.","major_comments":[{"comment":"The proof of Theorem 9 assumes that A admits a decomposition A = Q*Lambda*Q^{-1} with Lambda = diag(Lambda_1, ..., Lambda_n) and Lambda_i in R_{<0}, and then uses ||exp(At)|| <= ||Q|| ||Q^{-1}|| exp(Lambda_bar t). No justification is given that the matrix A defined in eq. (22) of the companion paper is diagonalizable with real negative eigenvalues. A general Hurwitz matrix can have complex-conjugate eigenvalues or be defective; in those cases the stated bound (20) and the ultimate bound (24) do not follow as written. This is load-bearing for the central control claim. The authors should either prove the spectral property for their specific A, add it as an explicit assumption, or redo the proof using the maximal real part of eigenvalues and a Jordan-form bound to handle defective cases.","section":"Section III, eqs. (19)-(24)"},{"comment":"The derivative of the periodic kernel with respect to d_Per is computed incorrectly: differentiating exp(-2 sin^2(pi d/p)/sigma_l^2) produces an additional factor cos(pi d/p), which is omitted in eq. (48). Also, the line 'due to the fact s(x,x\") >= 1' is false for s(x,x\") = sin(pi ||x-x\"||/p); the sine satisfies |s| <= 1. Consequently the Lipschitz constants for the periodic kernel in Table I are not established. The authors need to recompute the derivative, redo the maximization over the appropriate range, and correct the resulting constants. This does not directly affect the control experiments, which use the ARD-SE kernel, but it is a concrete defect in a stated result.","section":"Section IV-F, Corollary 4, eq. (48)"},{"comment":"The proof of Theorem 7 depends essentially on quantities and facts that are not reproduced in this document: the definitions of the aggregation weights omega_i^k(t) and omega_m(t), the definition of omega(t) in eq. (12) of the main paper, the statement of Lemma 2 (posterior concentration), the statement of Lemma 3 (delayed prediction bound), and the information-set rule in Algorithm 1 that guarantees eta_i^k(t) < beta*sigma_f. As a complementary document this is understandable, but for a standalone verification the central prediction bound is not self-contained. I recommend restating the inherited lemmas and the weight definitions (or providing a precise, versioned pointer to the companion paper) so the proof can be checked without guessing.","section":"Section II-B, Theorem 7 proof"}],"minor_comments":[{"comment":"The subsection titled 'Corollaries of the Other Lipschitz constant of Kernel' contains no corollaries and appears to be a placeholder; either fill it with content or remove it.","section":"Section IV-B"},{"comment":"The rational quadratic kernel is written as sigma_f^2 (1 + ||x-x'||^2/(2 alpha sigma_l^2)) without the exponent -alpha. The later derivative in eq. (41) is consistent with the usual definition with exponent -alpha, so eq. (38) appears to be a typo and should be corrected.","section":"Eq. (38)"},{"comment":"The notation d_x(x, x') is used in the SE kernel proof without a clear definition; it should be defined explicitly as the chosen distance d_SE = ||x - x'||, or the notation should be unified with the rest of the section.","section":"Section IV-A, eqs. (25)-(27)"},{"comment":"The control simulation paragraph states that the reference follows a uniform distribution between [0.4, 0.6] and [3, 5] for each simulation; it would be clearer to specify whether these are the ranges for the two components of the reference signal or two different simulation settings.","section":"Section V-C"},{"comment":"Reference [1] is cited as 'The 39th Annual AAAI Conference on Artificial Intelligence, 2024' with an OpenReview link; if the paper has since been published, the final venue and DOI should be provided.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's main theorem and control result are attractive, but the control proof's unproved diagonalizability assumption and the periodic-kernel derivative error are real technical defects that need to be fixed or explicitly scoped. The paper is a companion document, so some dependence on [1] is expected; however, the degree of self-containment is low for Theorem 7. I would recommend major revision rather than rejection because the central prediction-bound argument appears coherent and the control result is likely salvageable by adding an explicit spectral assumption or a Jordan-form analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a supplementary proof document, not a standalone paper. If you're reading it to check the claims in the main paper, the important takeaway is that the prediction-error bound (Theorem 7) looks structurally sound, while the control bound (Theorem 9) and the periodic-kernel Lipschitz constant are not fully established as written.\n\nThe paper does useful work: it spells out the aggregation-weight proof, gives correct derivations for the SE, ARD-SE, and RQ kernel constants, and provides extra delay-time and Monte-Carlo simulations. The proof of Theorem 7 is a clean Cauchy-Schwarz argument, and Corollary 8 follows directly. That part holds up.\n\nThe soft spots are real but localized. Theorem 9 assumes A = QΛQ^{-1} with real negative eigenvalues, and uses this to bound ||exp(At)||. For a general Hurwitz matrix this is not guaranteed—you can have complex or defective eigenvalues—and no justification is given for the specific A in eq. (22) of the main paper. The proof also drops the norm of b and writes vector inequalities as scalar ones. These are fixable with a more careful argument, but as written the ultimate bound (24) is not proven for a general Hurwitz A. Separately, Corollary 4's periodic-kernel derivative omits the cos factor from differentiating sin^2(πd/p), and the claim 's(x,x') ≥ 1' is backwards. The resulting constant is invalid. That doesn't affect the simulations, which use ARD-SE, but it's a concrete defect in the kernel catalogue.\n\nThe paper is not coherent enough to be taken as-is, but it's not incoherent either. The main prediction result is the core, and it's plausible. The errors are in the control extension and one side lemma—both worth fixing, not fatal to the whole project.\n\nWho is this for? Researchers working on distributed GP regression and control who want the detailed derivations behind the companion paper. It deserves a serious referee: the proof of Theorem 7 is worth checking carefully, and the control proof and periodic-kernel constant need correction. I'd send it to peer review as a companion to the main paper, with clear requests for revision.","headline":"Companion proof document for an AAAI paper: the core prediction bound is plausible, but the control theorem has a spectral assumption gap and the periodic-kernel Lipschitz constant is wrong.","tokens_in":14591,"tokens_out":3817,"would_cite":true,"duration_ms":34334,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M20","68T05","60G15"],"pacs":[],"model":"deepseek-v4-flash","headline":"An asynchronous distributed Gaussian-process predictor that fuses stale local predictions still carries a time-varying error bound, and a closed-loop controller using it keeps tracking error within a computable limit.","keywords":["asynchronous distributed Gaussian processes","online learning","prediction error bound","delayed predictions","tracking control","kernel Lipschitz constants","Gaussian process regression"],"falsifier":"Recompute the Lipschitz constant of the periodic kernel by differentiating $\\sigma_f^2\\exp(-2\\sin^2(\\pi\\|x-x'\\|/p)/\\sigma_l^2)$ with respect to $\\|x-x'\\|$ while keeping the chain-rule cosine factor; if the maximum is not the two-case value in Corollary 4, the delayed-prediction bound for that kernel is unsupported and the global certificate does not apply to it.","tokens_in":13556,"feed_emoji":"🤖","tokens_out":8975,"duration_ms":71775,"temperature":0.7,"pith_summary":"This companion document supplies the proofs behind an asynchronous distributed Gaussian-process regression method for online learning and dynamical systems. It aims to establish that even when local GP models send back predictions computed from outdated data, their weighted fusion still satisfies a time-varying prediction-error bound that never exceeds $\\beta\\sigma_f$. It then shows that a closed-loop controller built on this predictor has tracking error converging to a computable limit proportional to that bound. If these results hold, real-time control need not wait for fresh GP updates: stale but certified predictions can be fused safely, with the price of staleness explicitly quantified. The document also derives Lipschitz constants for several common kernels, which are used to bound the error contributed by delays.","feed_headline":"Distributed GP predictions stay error-bounded despite stale data","feed_subtitle":"A companion proof shows stale local GP forecasts fuse into a certified prediction bound and a closed-loop tracking guarantee.","key_machinery":"The load-bearing object is the aggregation rule and the associated time-varying bound $\\omega(t)$ defined in the companion paper. The weights $\\omega_i^k(t)$ are chosen so that the weighted sum of delayed-error bounds can be compressed by Cauchy-Schwarz into $\\omega(t)$, and the information-set rule sets $\\rho_i^k(t)=0$ whenever $\\eta_i^k(t) \\ge \\beta\\sigma_f$, which is what forces the global bound below $\\beta\\sigma_f$. For the control half, the machinery is the matrix-exponential estimate $\\|\\exp(At)\\| \\le \\|Q\\|\\|Q^{-1}\\|e^{\\bar{\\Lambda}t}$ under the Hurwitz assumption on $A$; that estimate is what carries the prediction certificate into the control guarantee.","core_discovery":"The central claim is Theorem 7 of the companion paper: for the asynchronous distributed GP predictor $\\hat{f}(x(t)) = \\sum_{i}\\sum_{k}\\omega_i^k(t)\\mu_i(x(t_k^i)) + \\omega_m(t)m(x(t))$, the prediction error satisfies $|f(x(t)) - \\hat{f}(x(t))| \\le \\omega(t)$, where $\\omega(t)$ is a time-varying quantity built from the delayed-error bounds $\\eta_i^k(t)$ and the prior bound $\\beta\\sigma_f$. Corollary 8 sharpens this to $\\omega(t) \\le \\beta\\sigma_f$ by exploiting the information-set rule that admits only delayed predictions with $\\eta_i^k(t) < \\beta\\sigma_f$. Theorem 9 feeds this bound into the error dynamics $\\dot{e} = Ae + b(f - \\hat{f})$ and obtains the ultimate tracking bound $\\lim_{t\\to\\infty}\\|e(t)\\| \\le \\|Q\\|\\|Q^{-1}\\|\\bar{\\omega}/|\\bar{\\Lambda}|$. The contribution is a formal certificate that fusing stale local GP predictions in real time remains safe, with the cost of asynchrony measured by $\\omega(t)$.","pith_inferences":["The same weighting argument would likely survive replacing Gaussian-process experts with any base learner that supplies a per-expert error bound, since the proof consumes only the bounds $\\eta_i^k(t)$ and $\\beta\\sigma_f$ rather than the GP likelihood itself.","A direct stress test of the certificate is to run the algorithm with artificially growing delay patterns while enforcing the information-set rule; the empirical prediction error should never cross $\\omega(t)$.","If a simpler but looser certificate is preferred, the monotonicity in Corollary 8 suggests replacing the time-varying $\\omega(t)$ by the constant $\\beta\\sigma_f$ in the tracking bound, at the price of a larger ultimate bound.","The periodic-kernel Lipschitz constant in Corollary 4 is not established by the given derivation, because that derivation drops the chain-rule cosine factor when differentiating the periodic kernel with respect to the distance; until this is corrected, the certificate should not be claimed for periodic kernels."],"forward_implications":["A control designer can use asynchronous distributed GP predictions with stale data and still quote a formal worst-case prediction error at each time step, instead of treating staleness heuristically.","The ultimate tracking error scales linearly with the worst-case prediction bound and with the condition number $\\|Q\\|\\|Q^{-1}\\|$ of the error dynamics, so better GP accuracy directly improves the tracking guarantee.","The information-set rule that discards delayed predictions whose bound reaches $\\beta\\sigma_f$ yields a uniform cap on the prediction error, making the certificate independent of how large the delay grows.","The kernel Lipschitz constants give an explicit way to compute $\\eta_i^k(t)$ for linear, squared-exponential, ARD-SE, and rational-quadratic kernels, so the bound can be evaluated in practice.","If the periodic-kernel Lipschitz constant were established, the same delayed-prediction bound would extend the certificate to periodic kernels as well."],"supporting_citations":[{"why":"Defines the AsyncDGP algorithm, the aggregation weights, the quantity $\\omega(t)$, and the information-set rule on which all proofs in this document rely.","marker":"[1]"},{"why":"Supplies the reproducing-property inequality used to bound the delayed prediction error in the proof of Lemma 3.","marker":"[42]"},{"why":"Supplies the matrix-exponential norm estimate used to turn the prediction bound into the closed-loop tracking bound in Theorem 9.","marker":"[43]"}],"fun_headline_variants":["Stale GP data still yields bounded prediction error","Asynchronous GP fusion keeps error bounds tight","Distributed GP error bounded despite delayed updates","Certified bounds for stale Gaussian process predictions","Asynchronous distributed GP: error stays bounded"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument depends on the companion paper's per-local-model error bounds and on the algorithm's rule admitting only delayed predictions with a sufficiently small error bound; for the periodic kernel that foundation is missing because the stated Lipschitz constant comes from a derivative that omits the chain-rule cosine factor.","fun_headline_variants_meta":{"raw":{"variants":["Stale GP data still yields bounded prediction error","Asynchronous GP fusion keeps error bounds tight","Distributed GP error bounded despite delayed updates","Certified bounds for stale Gaussian process predictions","Asynchronous distributed GP: error stays bounded"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1164,"prompt_tokens":790,"completion_tokens":374,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":406,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":406,"tokens_out":374,"duration_ms":3952,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:25:30.107019+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the Lipschitz constant of the periodic kernel by differentiating $\\sigma_f^2\\exp(-2\\sin^2(\\pi\\|x-x'\\|/p)/\\sigma_l^2)$ with respect to $\\|x-x'\\|$ while keeping the chain-rule cosine factor; if the maximum is not the two-case value in Corollary 4, the delayed-prediction bound for that kernel is unsupported and the global certificate does not apply to it.","supporting_citations":[{"cited_title":"Asynchronous distributed gaussian process regression,","cited_arxiv_id":null,"evidence_quote":"Defines the AsyncDGP algorithm, the aggregation weights, the quantity $\\omega(t)$, and the information-set rule on which all proofs in this document rely."},{"cited_title":"Learning-based Symbolic Abstractions for Nonlinear Control Systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the reproducing-property inequality used to bound the delayed prediction error in the proof of Lemma 3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the matrix-exponential norm estimate used to turn the prediction bound into the closed-loop tracking bound in Theorem 9."}],"review_version":1}