{"id":"62de26ba-2539-4cd3-a000-1d09c649ab1a","arxiv_id":"2505.07267","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The thesis shows that Bayesian filtering can be made adaptive, provably robust to outliers, and scalable to neural networks via the BONE framework, the WoLF filter, and low-rank Kalman variants.","lead":"This PhD thesis treats online machine learning problems, from forecasting to bandits, as Bayesian filtering tasks where model parameters are updated one observation at a time. It contributes a unifying framework for non-stationary environments, a provably outlier-robust filter that costs the same as a Kalman filter, and scalable low-rank Kalman updates for neural networks.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.5 proof incomplete: the log-determinant bound (T.3) uses a minimum branch that is infinite for W=0 (WoLF-TMD) and whenever H'R^{-1}H is singular; direct bound from Eq. (4.39) repairs it.","rationale":"The paper's central claim is Theorem 4.5, asserting that WoLF-type filters have bounded posterior influence and are therefore outlier-robust. The reader identified a gap in the proof of the T.3 log-determinant term when the weight W can be exactly zero, as in WoLF-TMD. My analysis confirms this and shows the same flaw extends to the case where H'R^{-1}H is singular, which is typical when observation dimension is smaller than parameter dimension. This is load-bearing because the theorem is the basis for the robustness claim, and the written proof fails for one of the paper's own recommended weight functions. However, the failure is local and repairable: Eq. (4.39) itself provides a finite bound without the invalid 'min' step, so the theorem is almost certainly true as stated. Since the reader already assigned a CONDITIONAL verdict, and the concern is a proof gap rather than a false conclusion, I recommend no change to that verdict. The concrete test of re-deriving the bound directly from Eq. (4.39) would settle the issue definitively. No other objections emerged: the computational cost claim is supported by Algorithm 11, the empirical results are consistent with the theory, and the BONE framework is a legitimate contribution.","tokens_in":65521,"tokens_out":15406,"duration_ms":141002,"concrete_test":"Independently re-derive the T.3 bound without the minimum-determinant step: verify that for any W >= 0 (including W=0) and any positive semidefinite H'R^{-1}H (including singular), log|(Sigma_pred^{-1} + W^2 H'R^{-1}H)^{-1}| <= -log|Sigma_pred^{-1}| < inf. If this direct bound holds, replace Eqs. (4.40)-(4.41) with it and confirm that Theorem 4.5 then follows for WoLF-TMD and for singular information matrices.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 4.5 is incomplete at Proposition 4.18 (bound for T.3). Equations (4.39)-(4.41) attempt to bound log|bar Sigma_t|. Although Eq. (4.39) already gives log|bar Sigma_t| <= -log(|Sigma_pred^{-1}| + bar(w)^{2D}|H'R^{-1}H|), which is finite for all bar(w) >= 0 because |Sigma_pred^{-1}| > 0, the proof then takes a 'minimum' of the two determinants. This minimum is 0 when bar(w)=0 (WoLF-TMD, Eq. 4.7) and also when H'R^{-1}H is singular, which occurs whenever the observation dimension is less than the state dimension (e.g., the 2D tracking experiment in Sec. 4.7.1 has D=4, o=2). The resulting -log(0) = +inf does not establish boundedness; the subsequent branch (4.41) invokes -2D log(bar(w)), which is unbounded above when bar(w)=0. Thus the written proof does not cover the WoLF-TMD weight, one of the paper's recommended choices, nor common singular information matrices. The gap is repairable by stopping at (4.39) and noting the RHS is bounded by -log|Sigma_pred^{-1}| < inf, but as written the central robustness claim lacks a complete proof for its own recommended weights.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This PhD thesis develops Bayesian filtering methods for online machine learning, organized around three goals: adaptivity, robustness, and scalability. Chapter 3 introduces BONE, a framework that unifies existing online-learning algorithms as choices of measurement model, auxiliary variable, conditional prior, posterior approximation, and weighting scheme, and proposes a new method RL[1]-OUPR* for environments with both gradual drift and abrupt changes. Chapter 4 introduces WoLF, a generalized-Bayes filter that replaces the log-likelihood with a weighted version, yielding closed-form KF/EKF-style updates at the same O(D^3) cost as the Kalman filter. The chapter states Theorem 4.5, which claims that WoLF has a bounded posterior influence function under weight conditions satisfied by the proposed IMQ, MD, and TMD weights, and presents experiments on 2D tracking, MLP regression, heavy-tailed regression, and EWMA smoothing. Chapter 5 proposes subspace and low-rank approximations for scaling filtering to neural networks. The central theoretical risk is the proof of Theorem 4.5, which has a gap for exactly-zero weights and singular information matrices; the gap is local and repairable.","tokens_in":65885,"tokens_out":6484,"duration_ms":64616,"significance":"If Theorem 4.5 is correct, WoLF is a practically valuable contribution: outlier-robust Bayesian filtering with closed-form updates, computational cost matching the standard Kalman filter, and a bounded posterior influence function under explicit conditions on the weight function. The empirical comparisons against KF-B, KF-IW, OGD, and RBPF on multiple tasks support the practical relevance, and the experiments are reported with clear metrics and baselines. The BONE framework is a useful descriptive taxonomy that connects Bayesian online changepoint detection, continual learning, and bandit algorithms, although it is not itself a new algorithm. Strengths include explicit assumptions rather than data-fitted claims, reproducible pseudocode, and the fact that the robustness theorem is a parameter-free derivation from stated moment conditions rather than a circular argument.","major_comments":[{"comment":"The proof of Theorem 4.5 is incomplete as written. Equation (4.39) gives the valid bound log|bar Sigma_t| <= -log(|Sigma_pred^{-1}| + bar(w_t)^{2D}|H'R^{-1}H|), whose right-hand side is bounded above by -log|Sigma_pred^{-1}| < infinity because Sigma_pred^{-1} is positive definite. Instead of stopping there, the proof passes to the 'minimum' bound (4.40) and then to (4.41), which invokes -2D log(bar(w_t)). For the threshold weight WoLF-TMD in (4.7), bar(w_t) can be exactly zero, and -log(0) = +infinity; the assertion 'sup bar(w_t) < infinity implies sup log bar(w_t) < infinity' is false when the weight attains zero. The same branch fails whenever H'R^{-1}H is singular, which occurs whenever the observation dimension is less than the state dimension, including the 2D tracking experiment of Section 4.7.1 with D=4 and o=2. The theorem is repairable by stopping at (4.39), but as written the proof does not establish the claimed robustness for the recommended WoLF-TMD weight or for common singular information matrices.","section":"§4.6.2, Proposition 4.18 (bound for T.3)"},{"comment":"The derivation of the bound for the mean-difference term (T.2) contains a division-by-zero issue when weights can be zero. In the chain leading to (4.29), the proof introduces an expression with w_t^{-2} in the denominator; for W = 0, this is undefined. The desired bound can be obtained directly from Lemma 4.14 without this step, because the denominator in the expression before the w^{-2} substitution is at least sigma_min(Sigma_pred^{-1})^2 > 0, so the proof should be rewritten to avoid the undefined intermediate expression. As written, this is a technical error in a load-bearing bound, though it is also locally repairable.","section":"§4.6.2, Lemma 4.15 and Eq. (4.29)"}],"minor_comments":[{"comment":"The theorem states the weight condition as 'sup_y W(y, ŷ) < inf' and 'sup_y W(y, ŷ)^k ||y|| < inf'; presumably these should be '< infinity'. This formulation should be corrected.","section":"§4.5, Theorem 4.5 statement"},{"comment":"The text says the weighting functions satisfy W : R^o x R^o -> R_{++}, but the threshold weight WoLF-TMD in (4.7) takes the value 0. The codomain should be R_{+} or the statement should be adjusted.","section":"§4.4, Eq. (4.7)"},{"comment":"In the paragraph before Lemma 4.13, the goal is stated as showing that 'qLG is outlier robust'; this should read q^{W-LG} to avoid confusion with the non-robust linear Gaussian filter of Theorem 4.4.","section":"§4.6.2, proof of Theorem 4.5"},{"comment":"There are minor notation inconsistencies: the changepoint probability is denoted 'pi' in the pseudocode but 'kappa' in the text, and the weight 'nu_t(r(1))' is missing a subscript in some places. These should be harmonized for readability.","section":"§3.4, Algorithm 10 (RL[1]-OUPR*)"}],"recommendation":"major_revision","confidential_remarks":"The thesis is a compilation of previously published work, which is disclosed in the front matter. For a journal submission, the incremental contribution beyond Duran-Martin et al. (2024) and related papers should be clarified. The main concern is the proof gap in Theorem 4.5; it is repairable, but the current version should not be published without the fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid thesis rather than a new breakthrough. The main content—BONE, WoLF, and the scalability tools—is already out in the author's earlier papers, so don't go looking for a fresh result. What the thesis does well is put the pieces in one place, write the background carefully, and make the connections explicit. The BONE taxonomy genuinely helps organize the online-learning and changepoint literature, and RL[1]-OUPR* is a nice concrete instantiation. The WoLF idea—replacing the likelihood with a weighted version while keeping the Kalman filter computational complexity—is attractive, and the experiments (2D tracking, corrupted UCI regression, EWMA) support the practical claims.\n\nThe soft spot is exactly where the reader put it: Theorem 4.5's proof has an incomplete bound on the log-determinant term. Equation (4.39) already gives log|Σ̄_t| ≤ -log(|Σ^{-1}_{t|t-1}| + w̄^{2D}|H'R^{-1}H|), which is finite for all w̄ ≥ 0 because the first determinant is positive. Instead of stopping there, the proof passes to a minimum of the two determinants, and that minimum branch can be zero for the paper's own WoLF-TMD weight, and also when H'R^{-1}H is singular—which happens whenever the observation dimension is smaller than the state dimension, as in the 2D tracking experiment. So the written proof does not cover the theorem's intended scope. This is a genuine gap, but a small one: treating W=0 as the prior-update case, or simply stopping at (4.39), repairs it. The claim is very likely true; the paper needs a corrected proof.\n\nLess important: no code or data artifacts are shipped with the thesis. The experiments are described and some notebooks are referenced, but nothing with hashes. For a methods thesis, reproducible artifacts would be a significant plus. And since the chapters are explicitly based on the author's own papers, the novelty of the thesis as such is mostly organizational.\n\nWho is this for? Someone working on generalized Bayes, online or changepoint methods, or robust filtering who wants a single coherent picture. The thesis deserves a serious referee; it is not a desk reject. I would send it out with a request to fix Theorem 4.5's proof and add the zero-weight case explicitly.","headline":"A well-written thesis that compiles the author's already-published Bayesian filtering work; the flagship robustness theorem has a real but easily patchable proof gap around exactly-zero weights.","tokens_in":66435,"tokens_out":3384,"would_cite":false,"duration_ms":35759,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62M20","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Bayesian filtering can serve as a single principled framework for online learning, and a new weighted-likelihood filter makes it provably robust to outliers at the computational cost of a standard Kalman filter.","keywords":["Bayesian filtering","online learning","generalised Bayes","outlier robustness","posterior influence function","changepoint detection","Kalman filter","low-rank approximation"],"falsifier":"Take the one-dimensional linear Gaussian state-space model from Section 4.7.1, use the threshold weight (4.7) with a finite threshold $c$, and send the contaminated observation $y_t^c$ to infinity. Since the weight becomes exactly zero for $\\|y_t^c-\\hat y_t\\|>c$, the theorem's bound on the log-determinant term (T.3) may fail; compute the KL divergence between the contaminated and uncontaminated posteriors numerically as $y_t^c$ grows, and check whether it remains bounded as Theorem 4.5 would require.","tokens_in":65302,"feed_emoji":"🛡️","tokens_out":9804,"duration_ms":84652,"temperature":0.7,"pith_summary":"This thesis makes the case that Bayesian filtering—recursively updating a posterior distribution over model parameters—is a single principled foundation for online learning tasks such as continual learning, prequential forecasting, and contextual bandits. It contributes three toolkits: the BONE framework, which unifies many existing adaptive methods under a five-component design space; the WoLF filter, which replaces the likelihood in a Kalman-style update with a weighted log-likelihood and is proven to bound the influence of arbitrarily large outliers; and scalable approximations (subspace EKF, PULSE, low-rank EKF) that make filtering practical for deep neural networks. The central theoretical result is Theorem 4.5, which shows the WoLF posterior has a bounded posterior influence function under weighting schemes like inverse multi-quadratic and Mahalanobis weights. If the claims hold, robust Bayesian online learning can be achieved at the computational cost of a standard Kalman filter rather than expensive variational iterations.","feed_headline":"Blocks any single outlier in Kalman filtering","feed_subtitle":"Weighted-likelihood WoLF filter provably bounds outlier influence at Kalman speed.","key_machinery":"The central object carrying the robustness argument is the weighted observation likelihood: the loss (4.2) that replaces the standard log-likelihood in the recursive posterior update (4.1). Concretely, the update equals the Kalman filter with the measurement precision $R_t^{-1}$ replaced by $W^2(y_t,\\hat y_t) R_t^{-1}$, so a single scalar weight controls how much the current observation moves the posterior. The proof of Theorem 4.5 decomposes the posterior influence function—the KL divergence between the posterior under a contaminated observation and the uncontaminated posterior—into three terms ((T.1), (T.2), (T.3)) and bounds each using spectral-norm inequalities; the key condition is that the weight decays fast enough relative to the growth of the contaminated point. For non-stationarity, the auxiliary variable $\\psi_t$ (e.g., a runlength) is the object that encodes regime information, and the BONE framework's five components parameterise the design space. For scalability, the machinery is low-rank structure: a projection matrix $A \\in \\mathbb{R}^{D\\times d}$ for subspace filters, and a diagonal-plus-low-rank precision matrix for LoFi.","core_discovery":"On the paper's own terms, the discovery is that generalised Bayesian inference can be imported into filtering without sacrificing closed-form updates: defining the loss as $\\ell_t(\\theta_t) = -W^2(y_t,\\hat y_t)\\log p(y_t|\\theta_t)$ and using it in place of the log-likelihood yields the same Kalman precision update with $R_t^{-1}$ scaled by $W^2$. Theorem 4.5 proves that whenever the weight satisfies $\\sup_y W(y,\\hat y)<\\infty$ and $\\sup_y W(y,\\hat y)^k \\|y-\\hat y\\|<\\infty$ for $k\\ge 2$, the posterior influence function is bounded, so a single arbitrarily large contaminated observation has bounded effect. The thesis also claims that its BONE framework, with components (M.1) measurement model, (M.2) auxiliary variable, (M.3) conditional prior, (A.1) posterior algorithm, and (A.2) weighting, subsumes a wide range of prior online-learning methods and that a new instantiation, RL[1]-OUPR*, handles both gradual and abrupt change; and that the low-rank/subspace filters make neural-network-scale filtering practical. These are presented as a unified toolkit for Bayesian online learning rather than as separate tricks.","pith_inferences":["The W=0 gap for WoLF-TMD suggests a general design principle for future robust filters: either choose weights bounded strictly away from zero, or handle the zero-weight branch explicitly so that the log-determinant bound remains valid.","The bounded-influence condition $\\sup_y W(y,\\hat y)^k \\|y-\\hat y\\|<\\infty$ for $k\\ge 2$ can be used as a recipe for designing new weighting functions beyond the IMQ and threshold families; smooth compactly-supported weights would need a zero-branch treatment.","Combining WoLF with online changepoint detection (as done in Section 4.7.4) opens a testable hypothesis: outlier contamination and genuine regime change can be disentangled, so a robust filter should reduce false changepoint alarms on datasets with labelled outliers and labelled changepoints.","The BONE framework's design space suggests that adaptive methods' performance differences often reduce to the choice of auxiliary variable and prior-reset rule, which could allow a systematic empirical comparison rather than the current ad-hoc benchmarking."],"forward_implications":["A single weighting function converts the standard Kalman filter into an outlier-robust filter with the same $O(D^3)$ update cost, so robust filtering no longer requires repeated variational inner iterations.","The BONE framework provides a common language for adaptive online learning methods, so existing algorithms like BOCD, runlength priors, and changepoint-probability models become interchangeable design choices.","The new hybrid rule RL[1]-OUPR* gives a single-hypothesis algorithm that tracks both gradual drift and abrupt changepoints, which the experiments show outperforms pure runlength and pure OU methods on several benchmarks.","The scalable variants—subspace EKF, PULSE, and LoFi—reduce memory and time to the point where Bayesian filtering can update a deep neural network online; LoFi maintains a diagonal-plus-low-rank precision matrix to achieve this."],"supporting_citations":[{"why":"Supplies the generalised Bayesian posterior (loss-function update) that WoLF's update (4.1) builds on.","marker":"Bissiri et al. (2016)"},{"why":"Defines the posterior influence function and the robustness criterion used to state Theorem 4.5.","marker":"Matsubara et al. (2022)"},{"why":"Provides the inverse multi-quadratic weighting function and the bounded-influence framework WoLF adopts.","marker":"Altamirano et al. (2023b)"},{"why":"Classical robustness framework that the paper invokes to define outlier robustness.","marker":"Huber (1981)"},{"why":"Gives the detect-and-reject taxonomy and a variational robust filter that serves as a baseline and motivates the threshold weighting.","marker":"Wang et al. (2018)"},{"why":"Weighted-likelihood idea that inspires the W^2-scaled likelihood in the WoLF loss.","marker":"Barp et al. (2019)"}],"fun_headline_variants":["Provably bounds any single outlier's effect in Kalman filtering","Robust Kalman filtering with provable outlier bound","Bounded outlier influence at Kalman speed via Bayesian filter","Scalable Bayesian filter that's robust to any single outlier"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of the robustness theorem assumes that no observation is ever assigned weight exactly zero; the threshold weight WoLF-TMD does assign weight exactly zero to sufficiently far-out observations, so the written proof has a gap there, although the result can be repaired by treating zero-weight observations as making the posterior equal to the prior.","fun_headline_variants_meta":{"raw":{"variants":["Provably bounds any single outlier's effect in Kalman filtering","Robust Kalman filtering with provable outlier bound","Bounded outlier influence at Kalman speed via Bayesian filter","Scalable Bayesian filter that's robust to any single outlier"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00144,"raw_usage":{"total_tokens":5824,"prompt_tokens":989,"completion_tokens":4835,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":4768}},"tokens_in":605,"tokens_out":4835,"duration_ms":32953,"temperature":1.0,"reasoning_tokens":4768,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:21:41.321931+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the one-dimensional linear Gaussian state-space model from Section 4.7.1, use the threshold weight (4.7) with a finite threshold $c$, and send the contaminated observation $y_t^c$ to infinity. Since the weight becomes exactly zero for $\\|y_t^c-\\hat y_t\\|>c$, the theorem's bound on the log-determinant term (T.3) may fail; compute the KL divergence between the contaminated and uncontaminated posteriors numerically as $y_t^c$ grows, and check whether it remains bounded as Theorem 4.5 would require.","supporting_citations":[{"cited_title":"Robust generalised B ayesian inference for intractable likelihoods","cited_arxiv_id":null,"evidence_quote":"Defines the posterior influence function and the robustness criterion used to state Theorem 4.5."},{"cited_title":"Robust G aussian K alman filter with outlier detection","cited_arxiv_id":null,"evidence_quote":"Gives the detect-and-reject taxonomy and a variational robust filter that serves as a baseline and motivates the threshold weighting."}],"review_version":1}