{"id":"824982f0-d943-4547-866f-eea268cef956","arxiv_id":"2502.02450","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"ST-RCGP, a state-space robust Gaussian process, matches the speed of standard spatio-temporal GPs while staying accurate under outliers, by automatically adapting its outlier-downweighting function during filtering.","lead":"This paper presents a robust version of spatio-temporal Gaussian process regression that stays fast (linear in time) even when data contains outliers. It adapts an existing robust-GP method to state-space form and improves its weaknesses by setting its tuning parameters automatically from the filtering distribution.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 3.3's proof of bounded PIF uses a fixed-weight bound under the false premise that ST-RCGP and RCGP share the same posterior, leaving the central robustness guarantee unproven for the adaptive algorithm.","rationale":"Proposition 3.3 is the paper's central theoretical claim: it promises that the ST-RCGP posterior has bounded influence function and is therefore provably robust to arbitrary single-point contamination. The proof in Appendix B.3 is the only support for this claim. The reader's weakest-assumption analysis correctly identifies that the proof inherits a fixed-weight PIF bound while the method uses adaptive weights. My reading confirms and sharpens this: the proof's own bound depends on |γ(x)| and |c(x)|, and once these are functions of the contaminated data, the bound is not a constant independent of y^c. Moreover, the manuscript itself (Prop. 3.2 and the discussion after it) states that adaptive weights are precisely what distinguishes ST-RCGP from vanilla RCGP, so the 'share the same distribution' premise in B.3 contradicts the paper's own Proposition 3.2. This is an internal inconsistency, not merely a disagreement with external consensus. The concern is load-bearing because if the proof cannot be repaired, the method loses its advertised provable-robustness guarantee; the empirical evaluation would remain useful but the central theoretical contribution would be downgraded. The recommended concrete test—a two-time-step exact computation of the smoothing PIF under a large contaminated first observation—would settle whether the theorem is false or merely unproven. I do not see a reason to change the reader's CONDITIONAL verdict: the practical contribution and experiments are substantial, and the gap is addressable by either a corrected proof or a weakened claim. The β inconsistency in B.3 (σ²/2 vs σ/√2) further indicates the proof was not carefully checked, but it is secondary. Agreement: the reader's weakest assumption and my concern coincide.","tokens_in":33870,"tokens_out":8132,"duration_ms":85201,"concrete_test":"Run the exact ST-RCGP recursion from Prop. 3.1 on a minimal Gauss-Markov model: ns=1, state z_k = z_{k-1} + q_k, y_k = z_k + ε_k, with only two time steps (or three, to include smoothing). Use the IMQ weight with γ_k = \\hat f_k and c_k² = diag(\\hat S_k). For a fixed clean y_2 (and y_3 if used), set y_1 = A for A ∈ {0, 1, 10², 10⁴, 10⁶} and compute the KL divergence between the smoothing posterior under clean y_1 and under contaminated A. If the KL grows without bound as A increases, Prop. 3.3's sup bound fails; if it saturates, the theorem may still be true but the current proof needs to be replaced by one that actually accounts for adaptive weights. As a control, repeat with fixed weights (γ,c constants), where the bound is known to hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Appendix B.3. It applies Altamirano et al.'s fixed-weight PIF bound (Prop. B.2) to ST-RCGP, justified by 'the fact that ST-RCGP and RCGP share the same distribution.' That fact is explicitly false for the proposed algorithm: Prop. 3.2 proves equivalence only for non-adaptive weights, and the paper states that adaptivity is exactly what makes filtering and smoothing differ. In ST-RCGP, γ_k = \\hat f_k and c_k² = diag(\\hat S_k) are the filtering predictive mean and variance. A contaminated observation at time m enters \\hat f_k and \\hat S_k for all k>m through the Kalman recursion, so all subsequent weights are data-dependent. The proof verifies |y| w(x,y)² < ∞ for fixed γ,c, but its own bound is β²(|γ(x)|+|c(x)|), which is not uniform once γ or c depends on y^c. No term in the proof controls the variation of γ,c with the contamination or the resulting change in future posterior updates. Thus the bounded-PIF conclusion does not follow for the adaptive algorithm actually implemented and claimed in Prop. 3.3. A corroborating slip: B.3 sets β=σ²/2 while the main text and Prop. 3.3 set β=σ/√2. The empirical results may still stand, but the paper's headline theoretical guarantee is currently unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ST-RCGP, a state-space formulation of robust and conjugate Gaussian process regression (RCGP) that scales linearly in the number of temporal observations while remaining conjugate. The method adapts the inverse-multi-quadric weight function of RCGPs by centering it at the filtering predictive mean and setting its shrinking scale to the filtering predictive variance, which the authors argue resolves three known weaknesses of batch RCGPs: sensitivity to a poorly chosen prior mean, poor uncertainty quantification, and difficult hyperparameter selection. The paper's central theoretical claim, Proposition 3.3, is that the posterior influence function (PIF) of ST-RCGP is bounded, so that any single arbitrarily large outlier has bounded influence on the posterior. The paper also proposes a weighted hyperparameter optimization objective and reports experiments on simulated data, financial flash-crash data, and UK temperature forecasting, alongside reproducible code.","tokens_in":34105,"tokens_out":4291,"duration_ms":46889,"significance":"If the bounded-PIF claim were established for the fully adaptive algorithm, ST-RCGP would be a substantive contribution: it would combine the linear-in-time cost of state-space GPs with provable robustness to outliers, while removing the main practical tuning issues of batch RCGPs. The empirical study is broad and the code is provided, which are concrete strengths. The state-space derivation in Proposition 3.1 is sound, and Proposition 3.2 cleanly characterizes when adaptive filtering/smoothing reduces to batch RCGP. However, the main theoretical guarantee currently rests on a proof that does not analyze the adaptivity of the weights, and the empirical results do not compensate for that gap in a paper whose headline contribution is the robustness guarantee.","major_comments":[{"comment":"The proof of Proposition 3.3 verifies the fixed-weight sufficient condition of Proposition B.2, namely sup_y |y| w(x,y)^2 < infinity for fixed gamma(x) and c(x), but it does not bound the dependence of the adaptive weights on the contaminated observation. In the ST-RCGP algorithm, gamma_k = \\hat f_k and c_k^2 = diag(\\hat S_k) are the filtering predictive mean and variance (Section 3), and for every k > m these quantities are functions of a contaminated observation y^c_{m,j} through the Kalman recursion in Eqs. (5)-(6). Proposition 3.2 establishes equality with batch RCGP only for non-adaptive weights, and Section 3 explicitly states that adaptivity is what makes filtering and smoothing distributions differ. Consequently, Proposition B.2 does not apply to the adaptive posterior: the PIF also receives a contribution from the change in J_{w_k} and \\nabla \\log w_k^2 induced by the change in gamma_k and c_k, and the proof contains no term controlling this contribution. The bounded-PIF conclusion for the algorithm actually implemented and claimed in Proposition 3.3 is therefore unsupported.","section":"Appendix B.3 / Proposition 3.3"},{"comment":"The proof of Proposition 3.3 sets beta = sigma^2/2, while the main text and the statement of Proposition 3.3 set beta = sigma/sqrt(2). This is more than a typo: the constants appearing in the PIF bound of Proposition B.2 depend on beta, and the proof must be carried out for the parameter value used in the algorithm. The discrepancy should be reconciled, and if the proof is meant for a different parameter setting, the proposition statement must be adjusted.","section":"Appendix B.3, Eq. (25)"}],"minor_comments":[{"comment":"The paper presents the removal of hand-picked hyperparameters as a contribution (Issues #1-#3), but the robust hyperparameter objective phi_GB introduces a new quantile parameter delta, set to delta = 0.05 in Appendix C.3 without discussion of its sensitivity. The text should either discuss the choice of delta or provide a sensitivity analysis.","section":"Section 3, phi_GB and Appendix C.3"},{"comment":"The legend in Figure 16 spells 'ST-RTGP', which appears to be a typo for 'ST-RCGP'.","section":"Figure 16"},{"comment":"The superscripts and subscripts for the generalized Bayes posterior mean and covariance are used inconsistently: Proposition 3.1 writes m^GB_{k|k} and P^GB_{k|k}, while Appendix B.1 and parts of the derivation drop the superscript. Unifying this notation would improve readability.","section":"Notation throughout"}],"recommendation":"major_revision","confidential_remarks":"The main issue is a load-bearing proof gap in Proposition 3.3, not a problem with the empirical methodology. The authors should either prove a bounded-PIF result that accounts for the adaptivity of gamma and c, or explicitly restrict the claim to non-adaptive weights. The beta inconsistency in Appendix B.3 also needs fixing. If the robustness theorem is withdrawn or weakened, the paper's contribution becomes substantially more empirical, and the introduction and abstract should be revised accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things up front. The algorithmic contribution is real: the paper turns RCGP into a linear-in-time state-space method and solves the known tuning problems with adaptive IMQ centering and shrinking functions plus a robust hyperparameter objective. That is genuinely useful, and the empirical study is solid. The second thing is that the headline robustness theorem (Prop 3.3) is not actually proved for the adaptive algorithm. The proof in Appendix B.3 applies Altamirano et al.'s fixed-weight PIF bound to ST-RCGP, justified by the claim that ST-RCGP and RCGP share the same distribution. But the paper itself shows that is only true for non-adaptive weights (Prop 3.2). In the algorithm, gamma_k and c_k are the filtering predictive mean and variance, so a contaminated observation changes the weights at all later times. The proof verifies a condition on w(x,y) for fixed gamma and c, and the resulting bound (beta^2 times |gamma|+|c|) is not uniform once gamma and c depend on the contamination. No term controls that dependence, so the bounded-PIF conclusion does not follow. There is also a smaller slip: B.3 sets beta = sigma^2/2 while the main text and Prop 3.3 use beta = sigma/sqrt(2). That is fixable, but the gap is load-bearing, not cosmetic.\n\nWhat the paper does well: the Kalman-filter derivation is careful and correct, Prop 3.2 is a clean equivalence result, and the adaptive weight selection is a nice idea that demonstrably fixes the RCGP failure modes. The experiments are more thorough than usual: finance and weather examples, sensible baselines, and the computational cost story is convincing. The code is available, which helps reproducibility. I also want to credit the authors for being honest about the limitations of prior RCGP tuning.\n\nThe soft spots, in proportion: the unproven robustness guarantee is the main one. The experimental comparisons have a few small weaknesses, such as fitting hyperparameters on outlier-free data for some baselines in Table 1 and using different kernels across methods in the DJIA experiment, but these are minor and do not undermine the main empirical conclusions. The free parameter delta in the robust objective is mentioned only in the appendix and deserves more discussion.\n\nWho this is for: anyone working on robust or scalable GP methods, applied spatio-temporal statisticians, and practitioners with outliers in weather or finance data. They will get value from the algorithm even if the theory needs repair. I would send this to peer review, but I would ask the authors to either provide a proper PIF analysis for the fully adaptive weights or restate the claim to match what is actually proved.","headline":"Useful and practical state-space extension of RCGP, but the proof of the headline robustness guarantee ignores the adaptivity of the weights and needs to be fixed before I trust the claim.","tokens_in":34700,"tokens_out":1825,"would_cite":true,"duration_ms":20376,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G15","62F35","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that robust-and-conjugate Gaussian process regression can be run as a state-space filter, giving outlier-resistant spatio-temporal GP predictions at linear-in-time cost.","keywords":["spatio-temporal Gaussian processes","generalized Bayes","state-space models","Kalman filtering","outlier robustness","inverse multi-quadric weights","posterior influence function","robust hyperparameter optimization"],"falsifier":"Simulate a temporal GP with one outlier of growing magnitude planted at an early time step, and compute the KL divergence between the ST-RCGP filtering posterior and the posterior obtained when that observation is removed; if the divergence increases without bound as the outlier magnitude grows, the adaptive version of Proposition 3.3 is false. A sharper check is to compute $\\sup_{y^c}|\\mathrm{PIF}(y^c,D)|$ numerically for the fully adaptive weights rather than for weights fixed at their clean-data values.","tokens_in":33598,"feed_emoji":"📈","tokens_out":7628,"duration_ms":74069,"temperature":0.7,"pith_summary":"This paper tries to establish that the robust-and-conjugate Gaussian process (RCGP), previously a batch method with cubic cost and fragile hyperparameters, can be rewritten as a state-space filter that keeps Gaussian posteriors and linear-in-time computation while resisting outliers. The key move is to let the robustness weight's center and width be the mean and variance of the one-step predictive distribution, so the method adapts itself as it filters and does not need a hand-picked outlier proportion or a well-chosen prior mean. The authors prove that under inverse multi-quadric weights a single corrupted observation has bounded posterior influence, and they replace standard likelihood-based hyperparameter fitting with a weighted objective so outliers do not distort the kernel and noise estimates. They demonstrate the method on financial crash data and UK temperature forecasts, where it matches expensive robust alternatives in accuracy at a fraction of their cost.","feed_headline":"Outlier-proof Gaussian process forecasts at linear cost","feed_subtitle":"A state-space reworking of robust conjugate GPs keeps spatio-temporal prediction accurate when observations are corrupted.","key_machinery":"The load-bearing object is the inverse multi-quadric weight function $w_{\\mathrm{IMQ}}(x,y)=\\beta(1+(y-\\gamma(x))^2/c(x)^2)^{-1/2}$ inside the weighted score-matching loss; it is a bump centered at $\\gamma(x)$ that decays like $1/|y|$ as $y$ moves away, and its squared tail decay is exactly what makes the posterior influence function finite. Writing that loss into a state-space GP keeps the update quadratic in the latent state, so the filtering posterior remains Gaussian and the Kalman predict-update recursions carry the computation. The adaptivity of the weights, with $\\gamma$ and $c^2$ read off the one-step predictive distribution, is what distinguishes ST-RCGP from vanilla RCGP and what makes the hyperparameters self-tuning.","core_discovery":"The central claim is Proposition 3.3: when the weight function is $w_{\\mathrm{IMQ}}(x,y)=\\beta(1+(y-\\gamma(x))^2/c(x)^2)^{-1/2}$ with $\\beta=\\sigma/\\sqrt{2}$, $\\gamma(x_k)=\\hat{f}_k$ (the filtering predictive mean), and $c^2(x_k)=\\mathrm{diag}(\\hat{S}_k)$ (the filtering predictive variance), the posterior influence function is bounded for any single contaminated observation, so the posterior cannot be pulled without limit by one extreme outlier. The paper further claims this robustness comes at no loss of conjugacy or speed: Proposition 3.1 shows the generalized-Bayes posterior stays Gaussian and is updated by Kalman filtering and smoothing, so the method inherits the linear-in-time cost of spatio-temporal GPs, and Proposition 3.2 shows that the non-adaptive special case reproduces the original RCGP exactly. On the algorithmic side, the adaptive centering and shrinking functions remove the RCGP's sensitivity to the prior mean and eliminate the need to preselect a shrinking constant.","pith_inferences":["A full proof of bounded influence for the fully adaptive filter would need to control how much the data-dependent $\\gamma_k$ and $c_k$ move when an early observation is corrupted; the paper's Proposition 3.3 holds those weights fixed, so closing that gap is a natural next step.","The same adaptive-centering scheme could be used with other weight families that decay at least as fast as an inverse multi-quadric, and the robustness condition would still suggest which decay rates preserve bounded influence.","Because the filter already produces one-step predictive means and variances, the weight assigned to each incoming point is a ready-made online outlier score, which could be exported to monitoring and anomaly-detection systems.","The exactness of the filter relative to approximate Student-$t$ methods suggests that the main remaining bottleneck is spatial dimension, where variational or inducing-point approximations would be the natural extension."],"forward_implications":["The posterior at each filtering step is Gaussian, so predictions, smoothing, and uncertainty intervals are available in closed form at the same order of cost as a plain spatio-temporal GP.","A single observation contaminated to an arbitrarily large value cannot drive the predictive mean or variance without bound, in the fixed-weight setting for which the proof is written.","The robust weighted objective $\\varphi_{\\mathrm{GB}}(\\theta)$ prevents kernel lengthscales, amplitudes, and noise levels from being fitted to outliers, closing a failure mode of leave-one-out hyperparameter selection.","In well-specified data without outliers the weight ratio stays near one, so ST-RCGP recovers ordinary spatio-temporal GP behavior rather than sacrificing efficiency for robustness.","On financial crash data and temperature forecasting with clustered outliers, ST-RCGP's RMSE and NLPD are comparable to Student-$t$ sequential filters while its wall-clock time stays close to that of a plain state-space GP."],"supporting_citations":[{"why":"Supplies the robust-and-conjugate GP construction and the fixed-weight posterior-influence bound that Proposition 3.3 adapts to the state-space setting.","marker":"Altamirano et al. (2024)"},{"why":"Provides the SDE representation of spatio-temporal GPs that gives the linear-in-time Kalman filtering formulation used for ST-RCGP.","marker":"Solin (2016)"},{"why":"Establishes the Kalman filtering and smoothing equations for temporal GP regression that Proposition 3.1 extends to the weighted loss.","marker":"Hartikainen & Särkkä (2010)"},{"why":"Introduces the weighted score-matching divergence whose IMQ-weighted form is the RCGP loss.","marker":"Barp et al. (2019)"},{"why":"Provides the relevance-pursuit robust GP baseline and documents the prior-mean sensitivity of RCGPs that adaptive centering is designed to fix.","marker":"Ament et al. (2024)"},{"why":"Supplies the influence-function definition of robustness used for the posterior influence function guarantee.","marker":"Huber (1981)"},{"why":"Supplies the Markov sequential-GP baselines with Student-t errors that ST-RCGP matches in accuracy at lower cost in the crash experiments.","marker":"Wilkinson et al. (2023)"},{"why":"Provides the generalized-Bayes posterior-updating rule used to define the robust loss-based posterior.","marker":"Bissiri et al. (2016)"}],"fun_headline_variants":["Robust spatio-temporal GP: linear cost, no tuning","Outlier-proof GP: adaptive, linear time","Spatio-temporal GP that shrugs off outliers","Bounded-influence GP for spatio-temporal data","Conjugate GP: robust, adaptive, linear in time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the bounded-influence proof, which treats the robust weight function as fixed while one observation is corrupted, still describes the algorithm in which that weight function is recomputed from the data, including the corrupted observation.","fun_headline_variants_meta":{"raw":{"variants":["Robust spatio-temporal GP: linear cost, no tuning","Outlier-proof GP: adaptive, linear time","Spatio-temporal GP that shrugs off outliers","Bounded-influence GP for spatio-temporal data","Conjugate GP: robust, adaptive, linear in time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1580,"prompt_tokens":909,"completion_tokens":671,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":593}},"tokens_in":525,"tokens_out":671,"duration_ms":6616,"temperature":1.0,"reasoning_tokens":593,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T12:05:24.697807+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a temporal GP with one outlier of growing magnitude planted at an early time step, and compute the KL divergence between the ST-RCGP filtering posterior and the posterior obtained when that observation is removed; if the divergence increases without bound as the outlier magnitude grows, the adaptive version of Proposition 3.3 is false. A sharper check is to compute $\\sup_{y^c}|\\mathrm{PIF}(y^c,D)|$ numerically for the fully adaptive weights rather than for weights fixed at their clean-data values.","supporting_citations":[{"cited_title":"Robust and conjugate G aussian process regression","cited_arxiv_id":null,"evidence_quote":"Supplies the robust-and-conjugate GP construction and the fixed-weight posterior-influence bound that Proposition 3.3 adapts to the state-space setting."},{"cited_title":"Stochastic differential equation methods for spatio-temporal G aussian process regression","cited_arxiv_id":null,"evidence_quote":"Provides the SDE representation of spatio-temporal GPs that gives the linear-in-time Kalman filtering formulation used for ST-RCGP."},{"cited_title":"and S \\\"a rkk \\\"a , S","cited_arxiv_id":null,"evidence_quote":"Establishes the Kalman filtering and smoothing equations for temporal GP regression that Proposition 3.1 extends to the weighted loss."},{"cited_title":"Minimum S tein discrepancy estimators","cited_arxiv_id":null,"evidence_quote":"Introduces the weighted score-matching divergence whose IMQ-weighted form is the RCGP loss."},{"cited_title":"Robust G aussian processes via relevance pursuit","cited_arxiv_id":null,"evidence_quote":"Provides the relevance-pursuit robust GP baseline and documents the prior-mean sensitivity of RCGPs that adaptive centering is designed to fix."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the influence-function definition of robustness used for the posterior influence function guarantee."},{"cited_title":"J., S \\\"a rkk \\\"a , S., and Solin, A","cited_arxiv_id":null,"evidence_quote":"Supplies the Markov sequential-GP baselines with Student-t errors that ST-RCGP matches in accuracy at lower cost in the crash experiments."},{"cited_title":"G., Holmes, C","cited_arxiv_id":null,"evidence_quote":"Provides the generalized-Bayes posterior-updating rule used to define the robust loss-based posterior."}],"review_version":1}