{"id":"e4310f51-9d1c-4b7a-b18e-8289151ebe03","arxiv_id":"2507.20379","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Prior-to-posterior maps in Bayesian inverse problems, filtering, and joint state-parameter learning are pointwise globally Lipschitz, producing new non-asymptotic bounds on the error of approximate sequential Bayesian methods.","lead":"A mathematics paper proves that Bayesian updating is stable: if two prior beliefs are close, the posteriors after seeing data stay close under total variation, Hellinger, and Wasserstein distances. This stability is then used to give the first general non-asymptotic error bounds for approximate Bayesian sequential learning methods such as online variational inference and particle filters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 16 as stated omits the assumption Q0=P0; without it the claimed bound omits the initial prior error and is false.","rationale":"The reader's ACCEPT verdict identified the bounded-metric assumption IP-SE-PS.4 as the weakest point. That assumption is indeed a real limitation for the Wasserstein results, but it is explicitly disclosed and does not make any theorem false. The concern raised here is different and more load-bearing: Theorem 16, as literally stated, is false if the approximate method starts from an approximate initial prior Q0 different from the true P0. The proof uses Q0=P0 in the first induction step, but the theorem statement omits this condition. Since Theorem 16 is the vehicle for the paper's central claim that bounded incremental errors imply bounded learning error, an explicit assumption or an initial-error term is required. The paper already handles the inexact-initial-prior case in Section 4.3, so the fix is straightforward, but the main theorem should not be left in a form that admits a counterexample. I recommend CONDITIONAL acceptance: the mathematical framework is sound under the intended Q0=P0 reading, but the statement of Theorem 16 must be corrected to include that assumption or to carry the initial prior error through the bound. My concern is independent of the reader's weak-assumption analysis, hence 'disagree' on the identification of the most load-bearing issue.","tokens_in":67065,"tokens_out":17529,"duration_ms":175591,"concrete_test":"Check Theorem 16's assumptions and base-case proof step: does the statement anywhere impose Q0=P0? If not, evaluate the claimed bound (20) in the constant-likelihood setting with P0 != Q0 and Qk=Q_{k-1} for all k. The bound predicts d(Pk,Qk) <= 0, while the true distance is d(P0,Q0)>0, so the theorem fails as stated. The fix is to add the explicit assumption Q0=P0 (or to add the term K(P0)d(P0,Q0) to the right-hand side), and to make the proof line 'Q1* = F1(P0) = P1' conditional on that assumption.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 16 claims to bound the learning error d(Pk,Qk) solely by the incremental approximation errors d(Qj*,Qj), starting at j=1. The proof's base case relies on the identity Q1* = F1(P0) = P1, which is true only when Q0=P0. The theorem statement, however, only assumes Pi in \\bar P_{i+1} for i in [0,k-1] and Qi in \\bar P_{i+1} for i in [1,k-1]; it never states Q0=P0. If Q0 differs from P0, the bound is missing the propagated initial error d(P0,Q0), and the theorem is false as stated. A concrete counterexample within the paper's assumptions: take a constant likelihood h(yk,x)=1 for all k, so Fk is the identity map, and choose Qk=Qk-1 for all k (no incremental approximation). Then every incremental error d(Qj*,Qj) is zero, Equation (20) gives d(Pk,Qk) <= 0, but the true learning error is d(Pk,Qk)=d(P0,Q0)>0 whenever Q0 != P0. The reader's identified bounded-metric assumption is a disclosed limitation, not an internal gap; the Q0 issue is an actual omission in a central statement. The intended reading appears to be Q0=P0 (the known initial prior), consistent with Section 4.3, but this must be made explicit and the theorem statement corrected or extended with an initial-error term.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a non-asymptotic error analysis framework for approximate Bayesian sequential learning (BSL) in inverse problems, state estimation, and parameter-state estimation. It formalizes the prior-to-posterior map F_k, the exact intermediate posterior Q_k^\\ast, the approximate posterior Q_k, and the incremental approximation error d(Q_k,Q_k^\\ast). The main results are: Theorem 14 establishes pointwise global Lipschitz stability of F_k under total variation, Hellinger, and 1-Wasserstein distances with explicit constants in Table 1; Theorem 16 propagates bounded incremental errors into two sets of bounds on the learning error d(P_k,Q_k); Theorems 19–23 give sufficient conditions for learning-error decay under data assimilation; and Corollaries 25–26 apply the framework to online variational inference. The proofs are collected in appendices, and a one-dimensional inverse problem is used to validate the TV and Hellinger bounds numerically.","tokens_in":67352,"tokens_out":8146,"duration_ms":82914,"significance":"If Theorem 16 is corrected as described below, the framework is a genuinely useful general tool: the Lipschitz constants in Table 1 are explicit functions of likelihood bounds and evidence terms, no coefficients are fitted, and the TV bound improves the corresponding Sprungk (2020) bound by a factor of 1/2 under the same assumptions. The second set of bounds in Theorem 16 is estimable from approximate posteriors and single-step evidence terms, which is attractive for online variational inference and particle-type methods. The paper also gives the first, to my knowledge, global Lipschitz statements under Hellinger and Wasserstein distances, and the VI corollaries show how the framework can be specialized. The main limitation is that all 1-Wasserstein results require the metric space to be bounded (Assumption IP-SE-PS.4), so standard linear-Gaussian state-space models on unbounded state spaces are excluded; this is disclosed in the manuscript but should be stated more prominently in the abstract and conclusions.","major_comments":[{"comment":"Theorem 16 is false as stated because the assumption Q0=P0 is missing. In the base case of the induction, the proof uses the identity Q1^\\ast = F1(P0) = P1 to conclude d(P1,Q1) \\le d(Q1^\\ast,Q1); this identity holds only when Q0=P0, whereas the theorem assumes only Pi\\in\\bar P_{i+1} for i\\in[0,k-1] and Qi\\in\\bar P_{i+1} for i\\in[1,k-1]. A concrete counterexample within the paper's assumptions is the inverse problem on X=[0,1] with Lebesgue density h(y_k,x)=1, so that F_k is the identity map, and the choice Q_k=Q_{k-1} for every k. Then every incremental error d(Q_j^\\ast,Q_j) vanishes and Equation (20) gives d(P_k,Q_k)\\le 0, while the true learning error is d(P_k,Q_k)=d(P_0,Q_0)>0 whenever Q0\\neq P0. The fix is local: add Q0=P0 (and Q0\\in\\bar P_1) to the assumptions, or add the propagated initial-error term to the bounds. Corollaries 25 and 26 inherit the same missing assumption from Theorem 16 and need the same correction.","section":"Section 3.3, Theorem 16 (proof in Appendix C)"}],"minor_comments":[{"comment":"The bounded-diameter condition D<\\infty is load-bearing for every 1-Wasserstein statement in Theorem 14, including the rows for state estimation and parameter-state estimation in Table 1. This is explicitly disclosed, but the abstract and conclusions should state that the Wasserstein results apply only under bounded state/parameter spaces; the standard linear-Gaussian state-space model on an unbounded space is not covered.","section":"Section 3.1, Assumption IP-SE-PS.4"},{"comment":"The theorem is stated only for k\\ge 2 and does not give the k=1 case. With the corrected assumption Q0=P0, the k=1 bound is simply d(P1,Q1)\\le d(Q1^\\ast,Q1); please add this case for completeness.","section":"Theorem 16 statements (20)–(21)"},{"comment":"There is a typo: 'Tonelli's thoerem' should read 'Tonelli's theorem'.","section":"Appendix B.3.1, proof of Lemma 48"},{"comment":"The citation 'Freitas and Neil (2001)' is incomplete; Sequential Monte Carlo Methods in Practice is edited by Doucet, de Freitas, and Gordon, and the author list and year should be corrected accordingly.","section":"References"},{"comment":"The experiment uses the same observation y at every update step, which is not a standard sequential data regime. Please state explicitly that this is a repeated-observation stress test for validating the bounds rather than a sequential-data benchmark.","section":"Section 6, numerical example"}],"recommendation":"major_revision","confidential_remarks":"The missing initial-prior assumption in Theorem 16 is a genuine central error, and the counterexample shows the theorem as printed is false. However, the intended reading Q0=P0 is evident from Section 4.3, and the correction is local; I do not see deeper structural issues in the proofs. The bounded-metric limitation for Wasserstein results should be highlighted more prominently, but it is disclosed. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is worth engaging with. It proves a genuinely new global Lipschitz stability result for the prior-to-posterior map under Hellinger and 1-Wasserstein distances, with explicit constants, and builds a recursive non-asymptotic error framework on top. But Theorem 16 as stated has a missing initial-condition assumption, and the proof of its base case slips on exactly that point. That needs fixing.\n\nWhat's new: Sprungk (2020) had pointwise global Lipschitz continuity under total variation and only local results under Hellinger and Wasserstein. Theorem 14 upgrades Hellinger and 1-Wasserstein to pointwise global constants, halves the TV constant, and covers inverse problems, state estimation, and parameter-state estimation. The constants are explicit functions of likelihood bounds and evidence terms, with no fitted coefficients and no circular dependence. Appendix B lays out the TV and Hellinger proofs in detail, and the Wasserstein proofs are complete under the stated assumptions. That part looks solid.\n\nSoft spots, in order of importance. Theorem 16 claims to bound d(P_k,Q_k) by the incremental errors d(Q_j^*,Q_j) starting from j=1. The theorem only assumes P_i is admissible for i in [0,k-1] and Q_i for i in [1,k-1]; it never states Q0=P0. The proof's base case uses Q_1^*=F_1(P_0)=P_1, which is only true when Q0=P0. Take a constant likelihood and Q_k=Q_{k-1}; all incremental errors vanish, but d(P_k,Q_k)=d(P0,Q0) can be positive. So the theorem is false as stated. The intended reading is almost certainly Q0=P0—Section 4.3 shows the authors know how to add an initial-prior error term—but the statement and proof need correction.\n\nThe bounded-metric assumption IP-SE-PS.4 is also worth flagging, though it is disclosed: every 1-Wasserstein statement in Theorem 14 needs it, so linear Gaussian state-space models with unbounded state space are excluded. The TV and Hellinger results do not need it.\n\nSome of the Hellinger and Wasserstein variants for state/parameter-state estimation are proved by analogy rather than fully written out. No formal verification, and no code or data, so reproducibility rests on checking the appendix. I did not find a reason to doubt the main argument once Theorem 16 is fixed.\n\nWho should read it: people analyzing online variational inference, particle filters, and Gaussian filters, and anyone working on non-asymptotic stability of Bayesian updating. It deserves a serious referee. My recommendation: send to peer review, and require the authors to fix Theorem 16's initial-condition gap, then re-check the base cases for all three distances.","headline":"Genuinely new global Lipschitz stability results under Hellinger and Wasserstein distances, but Theorem 16 as stated needs the assumption Q0=P0 (or an initial-error term) before the recursive error bounds are correct.","tokens_in":67877,"tokens_out":3515,"would_cite":false,"duration_ms":36149,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62M20","60B10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The prior-to-posterior map is pointwise globally Lipschitz under total variation, Hellinger, and 1-Wasserstein distances, so bounded per-step approximation errors yield a bounded learning error in sequential Bayesian learning.","keywords":["Bayesian sequential learning","global Lipschitz stability","posterior stability","total variation distance","Hellinger distance","Wasserstein distance","online variational inference","learning error bounds"],"falsifier":"On $X=\\mathbb{R}$ with a Gaussian observation model (bounded and Lipschitz likelihood), fix a base prior $\\mu$ and take $\\mu'$ to be Gaussian with the same mean and variance $\\sigma'^2\\to\\infty$; compute the ratio $W_1(F_k(\\mu),F_k(\\mu'))/W_1(\\mu,\\mu')$ at a fixed step $k$. If the ratio is unbounded as $\\sigma'^2$ grows, then the pointwise global Lipschitz claim under $W_1$ genuinely needs a bounded space, confirming the role of Assumption IP-SE-PS.4; if the ratio stays bounded, the bounded-diameter assumption could be relaxed and the Wasserstein results would extend to unbounded settings.","tokens_in":66854,"feed_emoji":"📏","tokens_out":12665,"duration_ms":111792,"temperature":0.7,"pith_summary":"This paper proves that the Bayesian update used at each step of sequential learning — taking a prior distribution to the posterior — is globally stable: the distance between two posteriors grows at most linearly with the distance between the two priors, uniformly over all priors, for total variation, Hellinger, and 1-Wasserstein distances. The stability constant depends only on the base prior's evidence and on problem constants, so it stays finite even when the approximate posterior is arbitrarily far from the true posterior. From this, the paper derives the first general, non-asymptotic error analysis for approximate Bayesian sequential methods: the learning error at time $k$ is bounded by a weighted sum of per-step incremental approximation errors, and a second bound can be computed from the approximate posteriors alone. It also identifies sufficient conditions under which assimilating a new datum shrinks the learning error. If correct, this turns error analysis of any new sequential algorithm into a per-step verification task, with a finite accuracy certificate for the whole run.","feed_headline":"Global Lipschitz stability bounds every approximate Bayesian learner","feed_subtitle":"Bounded per-step approximation errors guarantee bounded learning error for any prior, in three problem settings.","key_machinery":"The load-bearing object is the prior-to-posterior map $F_k$ together with its pointwise global Lipschitz constant $K(\\mu;y_k) = C(y_k)/Z_k(\\mu)$, where $Z_k(\\mu)$ is the evidence (marginal likelihood) of the base prior and $C(y_k)$ is a supremum of the likelihood, or of a product of the observation and transition models, or (for the Wasserstein statements) a Lipschitz constant of those models. The recursive mechanism is a triangle-inequality decomposition of the learning error at step $k$ into a propagated part, $d(F_k(P_{k-1}),F_k(Q_{k-1}))$, which the Lipschitz inequality compresses to a multiple of the previous step's error, plus the incremental approximation error $d(F_k(Q_{k-1}),Q_k)$. Recursing from step $k$ down to step 1 telescopes these factors into the weighted sums of Theorem 16, in which each incremental error is multiplied by a product of constants over the evidence, and one of the two resulting bounds replaces the intractable evidence of the true posteriors with evidence of the approximate posteriors, making it estimable in practice. The bounded-diameter assumption on the underlying space is what keeps the Wasserstein constants finite.","core_discovery":"The central discovery is the pointwise global Lipschitz continuity of the prior-to-posterior map $F_k$: for every admissible prior $\\mu$, there is a constant $K(\\mu;y_k)$ such that $d(F_k(\\mu),F_k(\\mu')) \\le K(\\mu;y_k)\\,d(\\mu,\\mu')$ for every other admissible prior $\\mu'$, where $d$ is total variation, Hellinger, or 1-Wasserstein distance, in inverse problems, state estimation, and parameter-state estimation. The constant has the form $C(y_k)/Z_k(\\mu)$, where $Z_k(\\mu)$ is the evidence of the base prior and $C(y_k)$ collects boundedness or Lipschitz constants of the likelihood and transition models; it never involves $\\mu'$. Because the constant does not depend on $\\mu'$, the bound applies no matter how different the two priors are — precisely the regime of approximate sequential learning, where the approximate posterior $Q_{k-1}$ can be far from the true $P_{k-1}$. The theorem also halves the best known total-variation bound under the same assumptions. Theorem 16 converts this stability into an error analysis: $d(P_k,Q_k)$ is at most a sum over steps of the incremental approximation errors $d(Q_j,Q^*_j)$, each weighted by a product of ratios of likelihood constants to evidence terms, giving one linear-in-error stability bound and one bound computable from the approximate posteriors alone. Theorems 19–23 then give sufficient conditions under which the propagated part of the error shrinks, $d(P_k,Q^*_k) \\le d(P_{k-1},Q_{k-1})$, offering a first general explanation of learning error decay.","pith_inferences":["The bounded-diameter assumption (IP-SE-PS.4) does real work: on unbounded spaces such as linear Gaussian state-space models the 1-Wasserstein constants become vacuous ($D=\\infty$), so the Wasserstein half of Theorem 14 is, as stated, a bounded-space result; the total-variation and Hellinger claims carry the unbounded settings.","Read as a template, the framework reduces the analysis of any new sequential algorithm to one method-specific task, bounding the per-step error $d(Q_j,Q^*_j)$, after which Theorem 16 supplies the whole-run certificate; the paper's online-VI corollaries show exactly how such per-step bounds are produced.","Because the Lipschitz constant is $C(y_k)/Z_k(\\mu)$, an informative datum that raises the evidence automatically shrinks the factor propagating prior error, suggesting that learning error decay is generic whenever observations are informative, beyond the specific sufficient conditions the paper states.","A testable extension suggested by the framework: on a bounded state space, run an ensemble or particle filter from deliberately corrupted initial priors, record per-step incremental errors, and compare the simulated learning error with the estimable bound of Equation (21) to gauge how tight the certificate is in a realistic regime."],"forward_implications":["Bounded-input, bounded-output stability: if every step's incremental approximation error is bounded, the overall learning error $d(P_k,Q_k)$ stays bounded in total variation, Hellinger, or 1-Wasserstein distance, regardless of how large the prior error was.","Two runs of any approximate method whose per-step incremental errors are both bounded produce approximate posteriors $Q_k$ and $\\tilde Q_k$ that remain within a bounded distance of each other and of the true posterior.","The second set of bounds is computable online: it depends only on the approximate posteriors and on single-datum evidences $Z_i(Q_{i-1})$, never on the intractable true posteriors, so practitioners can certify a finite accuracy envelope during learning.","Under the sufficient conditions of Theorems 19–23, assimilating the next datum can strictly shrink the learning error, giving a first general theoretical account of the error decay observed empirically in sequential Monte Carlo and other learners.","In state estimation with transition-observation products bounded above and below, composing the Lipschitz stability with forgetting results implies that a well-chosen alternative initial prior can beat the true prior, and that a larger immediate incremental error can be worthwhile if it improves later steps."],"supporting_citations":[{"why":"Provides the local stability results and the evidence Lipschitz lemma (its Theorem 15) that Theorem 14 globalizes and halves.","marker":"Sprungk (2020)"},{"why":"Earlier global TV stability for state estimation under bounded likelihoods; the bound that Theorem 14 cuts in half.","marker":"Rebesschini and van Handel (2015)"},{"why":"Same baseline: TV stability of the prior-to-posterior map in filtering under likelihood bounds.","marker":"Law et al. (2015)"},{"why":"Same baseline for TV stability in filtering under likelihood bounds, compared in Appendix A.2.","marker":"Sanz-Alonso et al. (2023)"},{"why":"Supplies Lemma 44, the Hellinger bound for scaled measures used in the Hellinger proof, and a tensor-train sequential learning analysis the framework absorbs.","marker":"Zhao and Cui (2024)"},{"why":"Local posterior stability result with evidence-dependent constant that motivates the pointwise-global form.","marker":"Basu (2000)"},{"why":"Posterior perturbation analysis under integral probability metrics whose prior-pair-dependent constant highlights the need for a constant depending on only one prior.","marker":"Garbuno-Inigo et al. (2023)"},{"why":"Forgetting result used to show alternative initial priors or approximate posteriors can improve long-run learning error.","marker":"Le Gland and Oudjane (2004)"}],"fun_headline_variants":["Lipschitz stability tames approximate Bayesian inference","Posterior map is globally Lipschitz for BSL","Approx Bayesian learners bounded by Lipschitz stability","Global Lipschitz posterior bounds approximate BSL error","One Lipschitz constant governs all approximate BSL errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the 1-Wasserstein results the space on which the unknown state or parameter lives must be bounded, with a finite maximum distance $D$ between any two points; this fails for common unbounded models such as linear Gaussian state-space systems, while the total-variation and Hellinger results do not need it.","fun_headline_variants_meta":{"raw":{"variants":["Lipschitz stability tames approximate Bayesian inference","Posterior map is globally Lipschitz for BSL","Approx Bayesian learners bounded by Lipschitz stability","Global Lipschitz posterior bounds approximate BSL error","One Lipschitz constant governs all approximate BSL errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000653,"raw_usage":{"total_tokens":3085,"prompt_tokens":1128,"completion_tokens":1957,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":744,"completion_tokens_details":{"reasoning_tokens":1879}},"tokens_in":744,"tokens_out":1957,"duration_ms":12972,"temperature":1.0,"reasoning_tokens":1879,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:44:35.112507+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On $X=\\mathbb{R}$ with a Gaussian observation model (bounded and Lipschitz likelihood), fix a base prior $\\mu$ and take $\\mu'$ to be Gaussian with the same mean and variance $\\sigma'^2\\to\\infty$; compute the ratio $W_1(F_k(\\mu),F_k(\\mu'))/W_1(\\mu,\\mu')$ at a fixed step $k$. If the ratio is unbounded as $\\sigma'^2$ grows, then the pointwise global Lipschitz claim under $W_1$ genuinely needs a bounded space, confirming the role of Assumption IP-SE-PS.4; if the ratio stays bounded, the bounded-diameter assumption could be relaxed and the Wasserstein results would extend to unbounded settings.","supporting_citations":[{"cited_title":"Data Assimilation: A Mathematical Introduction, volume 62","cited_arxiv_id":null,"evidence_quote":"Same baseline: TV stability of the prior-to-posterior map in filtering under likelihood bounds."},{"cited_title":"Tensor-train methods for sequential state and parameter learning in state-space models","cited_arxiv_id":null,"evidence_quote":"Supplies Lemma 44, the Hellinger bound for scaled measures used in the Hellinger proof, and a tensor-train sequential learning analysis the framework absorbs."},{"cited_title":"Uniform stability of posteriors","cited_arxiv_id":null,"evidence_quote":"Local posterior stability result with evidence-dependent constant that motivates the pointwise-global form."}],"review_version":2}