{"id":"1c4c7964-5f02-49b2-bc65-3b60d1fe3dea","arxiv_id":"2502.05802","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A distributed Kalman-filtered Gaussian process that reconstructs the full sensor measurement matrix via max/min consensus, enabling linear-in-eigenfunctions communication for wireless sensor networks.","lead":"Wireless sensors can estimate an unknown field like temperature or signal strength without a central hub, using a distributed Gaussian process that runs a Kalman filter on each node. The authors add a new consensus rule that passes along extreme values, so all sensors reach the same map faster and with less data traffic.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified to the central equivalence claim: the dual-extrema consensus preserves Assumption 1 by construction, so the distributed update should reproduce the centralized K-GP update.","rationale":"The reader's weakest_assumption targets Assumption 1 as an unproven structural condition that could break if two sensors contributed nonzero values to the same column. My reading is that this concern does not land: the algorithm is explicitly constructed so that only sensor r writes column r of H_r(0), and the dual-extrema update preserves that 'single owner per column' property. For any matrix entry (i,j), the values across sensors at any iteration are either all zero or a single value v with zeros elsewhere. The max/min combination then returns v if v is present in the neighborhood and 0 otherwise, which is lossless. The proof in Appendix B establishes convergence to Q+(0)+Q-(0), and under the invariant this equals the centralized H. The only gap is that the paper does not explicitly state and prove the invariant, but this is a minor presentational omission rather than a correctness risk. The other issues noted by the reader (overstated complexity, underspecified y_k communication, missing error bars/code) are valid secondary concerns, but they do not threaten the central equivalence claim. Therefore I find no significant objection to the strongest claim, and the reader's conditional verdict can remain unchanged on the basis of those secondary issues rather than Assumption 1.","tokens_in":15473,"tokens_out":9489,"duration_ms":96480,"concrete_test":"Write a small script (e.g., Python) that: (1) builds R=10, E=50 random matrices H_r with a single nonzero column r containing signed random values; (2) iterates Eq. (42)-(44) on a connected random graph for T=diameter steps; (3) after each iteration, for each (i,j), checks that the set of nonzero values across all sensors has size at most 1; (4) checks that after T steps every sensor's matrix equals sum_r H_r(0) to machine precision. If the check ever fails, the protocol is lossy; otherwise the central equivalence is verified computationally.","verdict_should_be":"UNCHANGED","load_bearing_attack":"After checking the proof in Appendix B, I do not find a load-bearing flaw in the central equivalence claim. The structural Assumption 1 is not an extra external restriction; it is guaranteed by the construction in (35)-(36) and is preserved by the dual-extrema update. Initially, sensor r's H_r has nonzero entries only in column r, so for any matrix entry (i,j), the values across H_1..H_R are either all zero or a single value v (from the unique sensor j) with zeros elsewhere. The update (42)-(44) maps each entry to max_j(H_j(t),0)+min_j(H_j(t),0); if any neighbor has v>0 this returns v, if any neighbor has v<0 this returns v, and if no neighbor has v it returns 0. Thus the tuple remains of the same 'single-v-plus-zeros' form, and after the graph diameter the union propagates to all sensors, yielding exactly H = sum_r H_r(0). The paper omits an explicit invariant lemma for Assumption 1, but the step is elementary. The scenario the reader posits (two sensors generating nonzeros in the same column) cannot arise under the stated protocol because each sensor only populates its own pre-assigned column. Therefore I find no significant objection to the central claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributed Gaussian process (GP) algorithm for online scalar-field estimation in wireless sensor networks, called K-DGP. It builds on the Kalman-filter-based GP (K-GP) of Solin et al., which approximates the kernel with E eigenfunctions and performs recursive Bayesian updates. The key distributed element is a new consensus protocol, called dual-extrema consensus, which is designed to share the sparse measurement matrix H_r that each sensor constructs from only its own measurement location. The authors claim that, after consensus, every sensor reconstructs the centralized matrix H exactly, so the distributed update (27)-(30) is identical to the centralized K-GP update. They also extend the framework to dynamic fields by adding a Kalman prediction step. The paper provides a convergence proof in Appendix B and numerical comparisons against MADGP and the centralized GP.","tokens_in":15752,"tokens_out":16746,"duration_ms":175029,"significance":"If the central equivalence holds, the paper contributes a genuinely decentralized, finite-time consensus mechanism for a structured message that arises naturally in recursive GP estimation. The dual-extrema protocol is simple and, under the stated single-nonzero-column structure, provably reconstructs the centralized measurement matrix without the asymptotic convergence and tuning coefficient of average consensus. The extension to dynamic fields via a Kalman prediction step is also a useful and relatively rare capability for distributed GP methods. The paper is transparent about the assumptions that sensors know R and their own IDs, which are the same assumptions used by MADGP. However, the advertised complexity advantages are overstated, and the sharing of the measurement vector y_k is not fully specified. With corrections, the underlying algorithmic idea is sound and likely of interest to the WSN and multi-robot estimation community.","major_comments":[{"comment":"The complexity claims are inaccurate. The paper states that the K-DGP computational complexity is O(R^3) and, in the abstract/conclusion, that it scales linearly with E. However, after the dual-extrema consensus reconstructs the full E x R matrix H, the update (27)-(30) requires forming H^T P H, which costs O(E^2 R) or O(E R^2) operations, and this term dominates O(R^3) when E >> R. The comparison with MADGP's O(E^3) is therefore still favorable, but the improvement is more modest than claimed. Similarly, Section 5's 'communication complexity to O(E)' is inconsistent with Section 3.1's statement of message size O(E * R); the manuscript should clarify whether the zero columns of H_r are actually transmitted or only the non-zero column plus the sensor index.","section":"§3.1, Eq. (27); §5"},{"comment":"The consensus protocol (42)-(44) is defined only for the matrices H_r, but the Kalman update (29) also requires the full measurement vector y_k. The text says 'This characterization is also applicable to yk' but does not provide an explicit dual-extrema update for y_k, and the comment in Algorithm 1 line 6 ('construct with yk') is ambiguous. The paper should specify exactly how the scalar measurements y_r are exchanged and combined, for example by encoding y as a 1 x R row matrix and applying the same max/min protocol, so that each sensor obtains the centralized vector y_k.","section":"§3.1–§3.2, Algorithm 1"},{"comment":"The step from (B.18) to (B.19) is asserted without proof. Because the update is a difference of two max-plus products, it does not obviously compose as A^t. The proof should include an induction, using the fact that Assumption 1 is preserved by the update, showing that each nonzero entry propagates along paths of the graph; then the closed form in (B.19) follows. As written, Theorem 2's proof is incomplete, although the gap appears repairable.","section":"Appendix B, Eqs. (B.18)–(B.19)"}],"minor_comments":[{"comment":"The squared exponential kernel is missing the minus sign and the square on the distance; it should read exp(-||x - x'||^2 / (2 l^2)).","section":"Eq. (4)"},{"comment":"The spectral density for a 2-D squared exponential kernel should be sigma_s^2 (2 pi l^2) exp(-l^2 lambda_e / 2); the factor (2 pi l)^{3/2} appears to be a typo that changes the scale of the initial prior covariance P_0.","section":"Eq. (47)"},{"comment":"Clarify that the max and min operations are elementwise and that the zero matrix E is included as one of the candidates in the max; the current notation is easy to misread as a matrix-level max.","section":"§3.2, Eqs. (43)–(44)"},{"comment":"The stopping criterion based on RMSE between successive H matrices may stop before the finite-time consensus target is reached if the threshold is not tight; state whether the theoretical equivalence assumes theta_th = 0 or full T_max iterations.","section":"Algorithm 1, line 7"},{"comment":"The dynamic-field results are only qualitative (Fig. 8); quantitative RMSE over time would make the claimed capability more convincing.","section":"§4.3"},{"comment":"There is a repeated word 'whose whose' in the description of the zero matrix.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The core algorithmic idea is sound and the finite-time consensus mechanism is the main novelty. The revision should focus on correcting the complexity analysis and specifying the y_k consensus; the proof gap in Appendix B should also be patched. The paper seems appropriate for a letters-style venue in distributed estimation or multi-robot systems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe short version: this is a credible extension of K-GP to distributed sensor networks, with a genuinely new consensus protocol, and the central equivalence claim survives closer inspection. It is not a field-reshaping result, but it is not a throwaway either.\n\nWhat is actually new is the observation that each sensor's measurement matrix H_r in the K-GP update is column-sparse, so the global H can be reconstructed without averaging the dense E x E sufficient statistics that MADGP exchanges. The dual-extrema consensus is a neat construction: by taking the max of the non-negative part and the negative of the max of the negated matrix, you preserve the intrinsic column and propagate it over the graph. The proof in Appendix B is basic max-plus algebra, and the structural Assumption 1 is indeed preserved by the update, as the stress-test note confirms. The dynamic-field extension via the Kalman prediction step is also a plausible addition, rarely reported in DGP work. Reference handling is honest: it builds on Solin, Jang, and the max-consensus literature without overselling novelty.\n\nThe soft spots are real but not load-bearing. First, the complexity claims are overstated. The paper advertises O(R^3) computation and O(E) communication, but the covariance update P_k = K S K^T costs O(E^2 R) and typically dominates the O(R^3) inversion unless E is tiny. The O(E) communication is correct only if a sensor transmits just its non-zero column; the paper never says that it does, and Section 3.1 explicitly states O(E*R) for the message. Second, the global measurement vector y_k is used in the update but no consensus protocol is given for it; Algorithm 1 says 'construct with y_k' without explaining how a sensor obtains all R measurements. That is a concrete gap. Third, the simulations report no error bars and ship no code; the dynamic-field demonstration is a single qualitative run. These are fixable with a revision.\n\nAll in all, the central argument holds up, the consensus idea is worth knowing, and the paper deserves a serious referee. I would accept it for review, with demands for a corrected complexity analysis and a spelled-out y_k sharing mechanism.","headline":"A credible distributed GP paper with a genuinely new column-preserving consensus protocol; the central equivalence holds up, but the complexity and y_k communication claims need fixing before publication.","tokens_in":16251,"tokens_out":3873,"would_cite":true,"duration_ms":39558,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes a distributed Kalman-filter Gaussian process whose dual-extrema consensus protocol lets every sensor reconstruct the centralized measurement matrix, so the distributed update is identical to the centralized K-GP update.","keywords":["distributed Gaussian process","Kalman filter","wireless sensor networks","scalar field estimation","consensus protocol","max-plus algebra","kernel approximation"],"falsifier":"For $R=2$, set $H_1(0)_{11}=1$ and $H_2(0)_{11}=3$ so that two sensors write different nonzero values into the same matrix position, violating Assumption 1. The dual-extrema update returns 3 at that position, while the centralized matrix would have entries 1 and 3 in separate columns; any mismatch between the consensus limit and the centralized $H$ falsifies the paper's equivalence claim.","tokens_in":15294,"feed_emoji":"📡","tokens_out":8221,"duration_ms":78036,"temperature":0.7,"pith_summary":"This paper proposes a distributed Kalman-filter Gaussian process (K-DGP) for estimating an unknown scalar field with a wireless sensor network. The central claim is that by giving each sensor a column of the measurement matrix and applying a new dual-extrema consensus protocol, every sensor reconstructs the same centralized measurement matrix $H$. Once that happens, the recursive update equations are identical to the centralized Kalman-filter GP, so the whole network shares one posterior without a central fusion node. The paper also shows that the approach scales as $O(R^3)$ computation and $O(E \\cdot R)$ communication per sensor, versus $O(E^3)$ and $O(E^2)$ for the prior basis-function MADGP, and demonstrates online estimation of a dynamic diffusion-convection field.","feed_headline":"Distributed GP now matches centralized estimation at linear cost","feed_subtitle":"A dual-extrema consensus protocol lets every sensor rebuild the full measurement matrix, cutting communication and computation costs.","key_machinery":"The load-bearing object is the modified measurement matrix $H \\in \\mathbb{R}^{E\\times R}$ whose $r$-th column holds the eigenfunction vector $\\Phi(x^r_k)$ evaluated at sensor $r$'s location and whose other columns are zero in each sensor's local message $H_r$. The dual-extrema consensus protocol, written in max-plus algebra as $H_r(t+1)=\\max_{j\\in\\{r,\\mathcal{N}_r\\}}\\{H_j(t)|\\bar E\\}-\\max_{j\\in\\{r,\\mathcal{N}_r\\}}\\{-H_j(t)|\\bar E\\}$, preserves the intrinsic elements by treating positive and negative parts separately; the max expression extracts the non-negative component and the negated max extracts the non-positive component, so their sum equals the original column value rather than an average. This protocol is what lets each sensor reconstruct the centralized $H$ in finite iterations, making the distributed Kalman update (27)-(30) identical to the centralized one.","core_discovery":"The paper's central discovery is an exact structural equivalence: with the setup of Section 3, the distributed K-DGP update (27)-(30) is the same as the centralized K-GP update. The enabling choice is to replace the shared matrix $\\Phi$ by a column-decomposed measurement matrix $H=[\\Phi(x^1_k),\\ldots,\\Phi(x^R_k)]$, where each sensor $r$ owns column $\\Psi^r_r=\\Phi(x^r_k)$ and zero columns elsewhere, so $H=H_1+\\cdots+H_R$. The dual-extrema consensus (42)-(44) is designed to preserve these intrinsic elements: the max operation on messages joined with a zero matrix keeps only non-negative entries, the min operation keeps only non-positive entries, and their sum recovers exactly the entry that belongs to that column. Because the protocol converges to $Q^+(0)+Q^-(0)$, and because Assumption 1 guarantees those two operations do not mix distinct nonzero values, the consensus value is precisely the centralized $H$. The paper proves this convergence in Appendix B and demonstrates in simulation that it reaches consensus faster than the average-consensus MADGP and supports dynamic-field estimation through the Kalman prediction step.","pith_inferences":["Editorial inference: because dual-extrema converges exactly in a number of iterations equal to the network diameter, one could trigger transmission only when the local matrix changes beyond a threshold, cutting radio traffic further in slowly varying fields.","Editorial inference: the column-decomposition idea is not tied to Gaussian processes; any distributed filter where each agent privately owns a disjoint block of the global measurement matrix could use the same max/min protocol to reconstruct the full matrix.","Editorial inference: the protocol needs every sensor to know the network size $R$ and its own ID; applying it to self-configuring or anonymous networks would first require a discovery mechanism, which the paper does not address."],"forward_implications":["After consensus converges, every sensor holds the same $m_k$ and $P_k$, so predictions at any query point match the centralized K-GP exactly.","Per-sensor communication drops from $O(E^2)$ to $O(E \\cdot R)$ and matrix inversion from $O(E^3)$ to $O(R^3)$, as long as the number of basis functions $E$ dominates the number of sensors $R$.","Dual-extrema consensus reaches the centralized matrix in finite iterations bounded by the graph diameter, without tuning a consensus coefficient $\\gamma$.","Online estimation of dynamic fields becomes possible by inserting the Kalman prediction step (40)-(41), avoiding the need for analytic eigenfunctions of spatio-temporal kernels.","In the reported simulations, K-DGP reaches the centralized GP accuracy with fewer eigenfunctions than MADGP and converges in fewer consensus iterations."],"supporting_citations":[{"why":"Supplies the single-agent recursive Kalman-filter GP that the paper extends to multiple agents.","marker":"Solin et al. (2018)"},{"why":"Provides the Hilbert-space eigenfunction and eigenvalue approximations used to build the basis in (45)-(46).","marker":"Solin & Särkkä (2014)"},{"why":"Gives the finite-dimensional eigenfunction approximation framework behind the distributed GP formulation.","marker":"Pillonetto et al. (2019)"},{"why":"Defines the MADGP baseline with average consensus whose communication and computation costs are compared against.","marker":"Jang et al. (2020)"},{"why":"Supplies the average-consensus protocol used as the baseline in the consensus comparison.","marker":"Saber & Murray (2003)"},{"why":"Provides the max-consensus convergence proof that Appendix B extends to dual extrema.","marker":"Nejad et al. (2009)"},{"why":"Derives the temporal Kalman prediction step used for dynamic field estimation.","marker":"Hartikainen & Särkkä (2010)"}],"fun_headline_variants":["Kalman GP: linear cost, central-level field estimation","Distributed GP with Kalman: central accuracy, linear scale","Dual-extrema consensus: the trick to central-level distributed GP","K-DGP: fast consensus, central-grade accuracy, linear cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The equivalence only holds while Assumption 1 is true: for every position in the message matrix, all sensors that send a nonzero value must send exactly the same value, meaning each column of the centralized matrix can be owned by only one sensor; if two sensors ever place different nonzero values in the same position, the dual-extrema rule will combine them into a value that no longer equals the centralized matrix.","fun_headline_variants_meta":{"raw":{"variants":["Kalman GP: linear cost, central-level field estimation","Distributed GP with Kalman: central accuracy, linear scale","Dual-extrema consensus: the trick to central-level distributed GP","K-DGP: fast consensus, central-grade accuracy, linear cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000348,"raw_usage":{"total_tokens":1960,"prompt_tokens":1061,"completion_tokens":899,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":677,"completion_tokens_details":{"reasoning_tokens":827}},"tokens_in":677,"tokens_out":899,"duration_ms":8731,"temperature":1.0,"reasoning_tokens":827,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:53:02.631123+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For $R=2$, set $H_1(0)_{11}=1$ and $H_2(0)_{11}=3$ so that two sensors write different nonzero values into the same matrix position, violating Assumption 1. The dual-extrema update returns 3 at that position, while the centralized matrix would have entries 1 and 3 in separate columns; any mismatch between the consensus limit and the centralized $H$ falsifies the paper's equivalence claim.","supporting_citations":[{"cited_title":"B., & S¨arkk¨a, S","cited_arxiv_id":null,"evidence_quote":"Supplies the single-agent recursive Kalman-filter GP that the paper extends to multiple agents."},{"cited_title":"Hilbert Space Methods for Reduced-Rank Gaussian Process Regression","cited_arxiv_id":"1401.5508","evidence_quote":"Provides the Hilbert-space eigenfunction and eigenvalue approximations used to build the basis in (45)-(46)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the finite-dimensional eigenfunction approximation framework behind the distributed GP formulation."},{"cited_title":"Y ., Kim, D., & Kim, H","cited_arxiv_id":null,"evidence_quote":"Defines the MADGP baseline with average consensus whose communication and computation costs are compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the average-consensus protocol used as the baseline in the consensus comparison."},{"cited_title":"M., Attia, S","cited_arxiv_id":null,"evidence_quote":"Provides the max-consensus convergence proof that Appendix B extends to dual extrema."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Derives the temporal Kalman prediction step used for dynamic field estimation."}],"review_version":1}