{"id":"4678ce2a-7c77-44dd-b563-17961da095c9","arxiv_id":"2411.09653","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"An optimal transport reformulation of Bayes' rule lets neural networks approximate posterior samples in nonlinear filtering, with error controlled by the optimization gap.","lead":"Bayes' formula, the rule for updating beliefs with new measurements, is notoriously hard to compute. This chapter presents an optimal transport formulation that turns the update into a machine-learning-friendly optimization problem.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The error bounds in Prop. 5/7 are not attached to the implemented algorithm: Remark 3 admits c-concavity is not imposed on f, and the OT filter experiments never verify the alpha-strong-convexity assumption.","rationale":"The reader's CONDITIONAL verdict is well supported, and my stress-test identifies the same load-bearing assumption: the alpha-strong-convexity/c-concavity condition is required for the central error bound but is neither enforced nor verified in the numerical implementation. The paper itself flags this in Remark 3, which strengthens rather than weakens the concern: the authors know the assumption is unconstrained and propose a post-hoc check, but no such check appears for the figures that support the central claim. I did not find an independent fatal flaw: the OT formulation is internally consistent, Prop. 6's expansion is plausible, and the simulations are illustrative. The gap is that the theoretical guarantee may not apply to the actual method as implemented; this makes the article a strong proposal whose central claim is conditional on an unverified hypothesis. That supports the existing CONDITIONAL verdict, so I recommend no change. The concrete test is inexpensive: evaluate the Hessian of the learned f and, if needed, compare against a constrained c-concave architecture.","tokens_in":19337,"tokens_out":4822,"duration_ms":54122,"concrete_test":"For each trained f checkpoint used to produce Figs. 3-5, use automatic differentiation to compute the smallest eigenvalue of Hess_x(1/2|x|^2 - f(x,y)) at the transported particle locations and observed values y used in the experiments. If any eigenvalue is negative or is smaller than the alpha asserted in Prop. 7, then Assumption 3 fails and Eq. (13)/(27) does not bound that run. As a complementary check, retrain the same experiments with f constrained to be an input-convex network and compare the resulting MMD/MSE; if performance is unchanged while the assumption is satisfied, the violating runs are merely unguaranteed, and if performance degrades, the paper should report a valid error bound for the unconstrained procedure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The single most load-bearing gap is that the theoretical guarantee does not cover the algorithm actually run. Prop. 5 (Eq. 13) and Prop. 7 (Eq. 27) assume x -> 1/2|x|^2 - f(x,y) is alpha-strongly convex for every y. The implemented max-min procedure (15) optimizes over an unconstrained neural network class F; Remark 3 explicitly states that this c-concavity is not imposed and says it can be checked a posteriori, yet no such check is reported for the experiments in Sec. 4.2 or Sec. 6. If the trained f violates the condition, the bound d(...,...) <= sqrt(4 eps / alpha) does not apply, so the numerical success is not protected by the chapter's central quantitative claim. This is not a matter of style or consensus; without c-concavity the saddle-point reduction of the Monge problem is not justified, and the error bound is the only formal bridge between the max-min objective (11)/(15) and accuracy of the resulting posterior approximation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This chapter argues that the Bayes update can be implemented by solving a max-min optimal transport problem over a transport map T and a c-concave potential f, with T(.,y) the OT map from the prior P_X to the posterior P_{X|Y}(.|y). After a historical survey covering the Kalman filter, SIR particle filters, the ensemble Kalman filter, and the feedback particle filter, it presents the OT formulation (Eq. 11), an error bound under an assumed optimality gap and alpha-strong convexity (Prop. 5), a recursive OT filter with a stability-based error bound (Prop. 7), an asymptotic connection to the feedback particle filter (Prop. 6), and numerical demonstrations on a static bimodal example, a dynamic bimodal example, and the Lorenz-63 model.","tokens_in":19590,"tokens_out":22131,"duration_ms":212791,"significance":"If the stated guarantees applied to the algorithm actually implemented, the chapter would provide a useful simulation-based (likelihood-free) methodology for posterior sampling and nonlinear filtering that avoids the weight degeneracy of SIR and the Gaussian bias of EnKF. The chapter's strengths are its unified discrete/continuous treatment, its historical synthesis, the clear statement of the semi-dual max-min formulation, the explicit bounded-Lipschitz error bound in Eq. (13), and the numerical comparisons using MMD and MSE. However, the formal guarantees currently rest on assumptions that are neither imposed nor verified in the experiments, and there is an apparent sign inconsistency in the central objective; these issues need to be resolved before the central claim is fully supported.","major_comments":[{"comment":"As printed, Eq. (11) defines J(f,T) = E_{P_{X,Y}}[f(X,Y)] − E_{P_X⊗P_Y}[f(T(X,Y),Y) + c(T(X,Y),X)], which gives the transport cost a negative sign. With this sign, the max-min problem is not the semi-dual of the constrained Monge problem in Eq. (10): in the elementary case P_X=δ_0, P_Y=δ_1, P_{X|Y}=δ_1, the printed objective yields value 0 (or is unbounded for f at the c-concavity boundary) instead of the OT cost 1/2. The displayed empirical objective in Eq. (14) has the cost term with a plus sign, and the expansion in Prop. 6 also uses the plus sign. Please correct the sign in Eq. (11) so that the cost term is added, e.g., J = E_{P_{X,Y}}[f(X,Y)] − E_{P_X⊗P_Y}[f(T(X,Y),Y)] + E_{P_X⊗P_Y}[c(T(X,Y),X)].","section":"§4, Eq. (11), Eq. (14), Prop. 6"},{"comment":"The quantitative error bounds in Eq. (13) (Prop. 5) and Eq. (27) (Prop. 7) both assume that x ↦ 1/2|x|^2 − f_t(x,y) is alpha-strongly convex for all y and t. The implemented algorithm in Eq. (15) optimizes over an unrestricted neural-network class F, and Remark 3 explicitly concedes that c-concavity is not imposed, suggesting only an a posteriori check. No such check is reported for the experiments in Sec. 4.2 or Sec. 6. Consequently, Eqs. (13) and (27) do not cover the algorithm actually run, and the numerical success is not protected by the chapter's central quantitative claim. The chapter should either impose c-concavity/alpha-strong-convexity (e.g., via input-convex network architectures), verify and report the condition a posteriori, or state a version of Prop. 5 that holds without this assumption.","section":"§4 Remark 3; §5.2 Prop. 5 and Prop. 7"},{"comment":"The filter error bound in Prop. 7 also assumes uniform geometric stability of the filter (Def. 2). Remark 5 correctly acknowledges that this condition is strong and is only guaranteed under a minorization condition. However, the numerical sections that support the OT filter, including the dynamic bimodal example (28) and the Lorenz-63 example, do not check or discuss whether these models satisfy the stability assumption. If the chapter aims to present the OT filter as an algorithm with a proven guarantee, Assumption 1 of Prop. 7 should be checked for the reported examples or explicitly stated as an unverified hypothesis in the experimental sections.","section":"§5.2 Def. 2, Prop. 7, Remark 5; §6"},{"comment":"Proposition 6 is stated as a result, but the asymptotic form f(x;y)=phi(x)y+psi(x)Δt and T(x,y)=x+K(x)y+u(x)Δt is assumed rather than derived. The appendix proof is a formal Taylor expansion for functions of that form; it does not establish that the solution of the max-min problem (11) has this asymptotic structure, nor does it give conditions for the remainder to be O(Δt^3) uniformly. Please label Eq. (18) explicitly as an ansatz/assumption and state the smoothness and limit-exchange conditions needed for the expansion, so that the claimed recovery of the FPF update is a clearly qualified statement.","section":"§4.3, Prop. 6, Eq. (18)"}],"minor_comments":[{"comment":"The notation for the shuffled sample is inconsistent: the text writes “i=2,...,n” where it should be “i=1,...,N,” and in the cost term it is unclear whether the second argument of c should be the shuffled sample X^i or the paired sample Xi. Please clarify so that Eq. (14) is an unambiguous empirical estimate of the objective in Eq. (11).","section":"§4.1, Eq. (14)"},{"comment":"The heading “Lorentz-63” should be “Lorenz-63” to match the standard name of the model.","section":"§6.0.2"},{"comment":"The minorization constant epsilon in the condition a(x|x') ≥ epsilon rho(x) uses the same symbol as the optimality gap epsilon in Prop. 7. Please use different symbols to avoid confusion.","section":"§5.2, Remark 5"},{"comment":"The reference “Doucet A.and Johansen AM” has a formatting error (missing space and second initial format), and the Acknowledgments section is empty; the latter should either be filled or removed.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The chapter is largely a survey with several propositions imported from Al-Jarrah et al. (2023, 2024b) rather than proved in full. That is acceptable for a chapter, but the imported statements must be self-consistent with the notation used here; the sign error in Eq. (11) and the unverified alpha-strong-convexity assumption are the main obstacles. The numerical evidence is suggestive but not decisive because the error bounds are not attached to the implemented network class. A revision that fixes the sign, verifies or imposes c-concavity, and clearly labels the stability assumption as unverified for the experiments would make the chapter publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Useful book chapter. The central OT max-min formulation of the Bayes update is not new here — the authors credit it to Taghvaei and Hosseini and to Al-Jarrah et al. — but the chapter does a genuinely good job of connecting that formulation to the historical filtering literature and to the feedback particle filter. The genuinely new piece, Prop. 6, expands the objective in Delta t and recovers the FPF gain and drift, including the optional divergence-free term and an explanation of why it does not affect the density. That derivation is clean and internally consistent. Prop. 5 and Prop. 7 are also honest: they state exactly which assumptions the bounds need.\n\nThe stress-test note lands. Prop. 5's bound requires x -> 1/2|x|^2 - f(x,y) to be alpha-strongly convex for every y. Remark 3 admits the neural network class does not impose c-concavity; the authors say it can be checked a posteriori, but no such check is reported for the Sec. 4.2 or Sec. 6 experiments. So the paper's central quantitative claim is not attached to the algorithm actually run. That is the main soft spot. Second, Prop. 7 also assumes uniform geometric stability (Def. 2), which the chapter itself acknowledges is strong and only guaranteed under a minorization condition; the experiments do not verify it. Third, the finite-particle error analysis is explicitly left open in Remark 6. Fourth, the numerics are illustrative rather than decisive: no code or data are shipped, there are no error bars, and the 'exact' posterior in Fig. 4b is approximated by a 1e5-particle SIR run, which is a reasonable stand-in but worth a sentence of caveat.\n\nNone of this is fatal for a survey chapter. The authors are transparent about the gap between the theory and the implementation, and the OT formulation itself rests on standard duality rather than on a fitted circular argument. The literature survey is solid and the citation pattern is appropriate, including to the authors' own prior work, which is correctly identified as the source of the main formulation.\n\nThe chapter is best read by graduate students or researchers entering nonlinear filtering who want a single coherent statement of the OT approach, its relationship to FPF and EnKF, and a map of the open problems. It deserves serious peer review; a referee should push for either a posteriori verification of c-concavity on the reported experiments or a revised statement that clearly separates the ideal OT bound from the practical neural-network algorithm.","headline":"A clear expository synthesis of an OT-based likelihood-free Bayes update, worth refereeing, but its advertised error bound does not yet cover the neural network implementation.","tokens_in":20086,"tokens_out":1860,"would_cite":true,"duration_ms":19715,"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":"This chapter argues that the Bayes update in nonlinear filtering can be implemented by solving a max-min optimal transport problem whose solution is a map from prior to posterior samples, yielding a likelihood-free algorithm that captures…","keywords":["Bayes formula","nonlinear filtering","optimal transport","posterior sampling","feedback particle filter","particle filter curse of dimensionality","simulation-based inference","machine learning"],"falsifier":"On the static example $Y=\\tfrac12 X\\odot X+\\lambda_w W$ with $n=2$, train $f$ and $T$, compute the optimality gap $\\epsilon$ from (12), estimate $\\alpha$ as the minimum eigenvalue of the Hessian of $x\\mapsto\\tfrac12|x|^2-f(x,y)$ across $y$, and compare the empirical bounded-Lipschitz distance between $T(\\cdot,y)_\\#P_X$ and $P_{X|Y}(\\cdot|y)$ with $\\sqrt{4\\epsilon/\\alpha}$; a violation would refute Proposition 5's claim.","tokens_in":19085,"feed_emoji":"🎯","tokens_out":12816,"duration_ms":104932,"temperature":0.7,"pith_summary":"The paper's central claim is that Bayes' formula — the step that turns a prior belief into a posterior after an observation — can be recast as a max-min optimal transport problem. The solution is a map $T(\\cdot,y)$ that pushes samples from the prior $P_X$ onto the conditional posterior $P_{X|Y}(\\cdot|y)$, for every observation $y$ at once. Because the objective is evaluated on samples from the joint law $(X,Y)$, the method is simulation-based: no likelihood formula is ever needed, which suits modern machine-learning practice. If the claim holds, nonlinear filters gain a principled way to represent multimodal posteriors, avoiding both the Gaussian bias of the ensemble Kalman filter and the weight degeneracy of particle filters. The chapter also shows that in the continuous-time limit the same optimization recovers the feedback particle filter, so the new update is an exact time discretization of an earlier algorithm.","feed_headline":"Bayes update becomes a max-min optimal transport problem","feed_subtitle":"Neural nets learn a transport map that captures bimodal posteriors where particle filters collapse and Kalman filters go Gaussian.","key_machinery":"The load-bearing object is the block-triangular transport map $(x,y)\\mapsto(T(x,y),y)$ selected by the quadratic-cost Monge problem under the coupling constraint $(T(X,Y),Y)\\sim P_{X,Y}$. The max-min objective (11) is its Kantorovich dual: $f$ is a c-concave potential, meaning $x\\mapsto\\tfrac12|x|^2-f(x,y)$ is convex, and $T$ is the map that transports the independent coupling $P_X\\otimes P_Y$ toward the joint law $P_{X,Y}$. This structure does three jobs: it pins down the unique Bayes-consistent map, it converts the consistency condition into an optimization amenable to neural-network parameterization, and it yields a quantitative error certificate through the optimality gap $\\epsilon$ and the strong-convexity constant $\\alpha$. In the recursive filter, the same step is iterated, and uniform geometric stability of the filter (Definition 2) converts the per-step bound into the global bound (27).","core_discovery":"The core discovery is the characterization of the Bayes update as a conditional Monge problem: among all maps $T$ with $(T(X,Y),Y)\\sim P_{X,Y}$, choose the one minimizing $\\mathbb{E}[\\tfrac{1}{2}|T(X,Y)-X|^2]$. Kantorovich duality turns this into the max-min problem $\\max_{f\\text{ c-concave}_x}\\min_T J(f,T;P_{X,Y})$ with $J(f,T;P_{X,Y})=\\mathbb{E}_{(X,Y)\\sim P_{X,Y}}[f(X,Y)]-\\mathbb{E}_{(X,Y)\\sim P_X\\otimes P_Y}[f(T(X,Y),Y)+\\tfrac{1}{2}|T(X,Y)-X|^2]$. For absolutely continuous priors with convex support, the pair $(f,T)$ exists and is essentially unique, and $T(\\cdot,y)$ is the quadratic-cost optimal transport map from $P_X$ to $P_{X|Y}(\\cdot|y)$. When the pair is only approximately optimal, the bounded-Lipschitz error between the transported prior and the true posterior is at most $\\sqrt{4\\epsilon/\\alpha}$, where $\\epsilon$ is the optimality gap and $\\alpha$ is the strong-convexity constant of $x\\mapsto\\tfrac12|x|^2-f(x,y)$; in a recursively stable filter this becomes a time-uniform bound $C/\\lambda\\sqrt{4\\epsilon/\\alpha}$.","pith_inferences":["The optimality gap (12) could be tracked during training as a live diagnostic: with an estimate of $\\alpha$, a trained filter would come with a data-dependent, checkable error bound rather than an unverified heuristic.","The learned map is amortized: after training on joint samples, conditioning on a new observation $y$ costs one forward pass through $T$, which points toward cheap conditional generation in inverse problems and simulation-based inference.","The bias-variance split the paper sketches suggests that the OT approach can beat the curse of dimensionality exactly when the transport map has exploitable structure; high-dimensional image-style observation models are the natural stress test of that expectation.","If the uniform geometric-stability assumption is relaxed to asymptotic stability, the bound likely loses its time-uniformity, predicting that on detectable-but-not-minorizing systems the error stays finite over each horizon but may drift with time."],"forward_implications":["Bayes updates become likelihood-free: a simulator that draws $(X_i,Y_i)$ from the joint model is enough to train $f$ and $T$, so the method applies when the likelihood has no closed form.","Multimodal posteriors are representable: in the static and dynamic examples, the OT filter keeps both modes of a bimodal posterior, whereas SIR collapses into one mode and EnKF forces a Gaussian.","The error is controlled by training quality: whenever the learned pair has gap $\\epsilon$ and $f$ is $\\alpha$-strongly c-concave, the posterior error is no larger than $\\sqrt{4\\epsilon/\\alpha}$, and $C/\\lambda\\sqrt{4\\epsilon/\\alpha}$ for uniformly stable filters.","The OT update is an exact time discretization of the feedback particle filter: the leading terms of a small-time expansion of the max-min objective reproduce the FPF gain and Poisson equation, with a divergence-free correction that does not change the density evolution.","With a resampling stage, the finite-particle version carries an additional $1/\\sqrt{N}$ sampling error, keeping the usual Monte Carlo rate while avoiding weight-based degeneracy."],"supporting_citations":[{"why":"Supplies the rigorous derivation of the max-min formulation, the justification of Prop. 4, and the neural-network implementation used in the numerical experiments.","marker":"Al-Jarrah et al. (2024b)"},{"why":"Its Theorem 2.3 is the basis for existence and uniqueness of the solution to the max-min problem in Prop. 4.","marker":"Carlier et al. (2016)"},{"why":"Its Theorem 4.3 is the bound that Prop. 5 extends to the approximate optimality-gap setting.","marker":"Rout et al. (2022)"},{"why":"Its Theorem 3.6 provides a second source for the optimality-gap error estimate used in Prop. 5.","marker":"Makkuva et al. (2020)"},{"why":"The earlier paper where the optimal transport formulation of Bayes' law first appeared.","marker":"Taghvaei and Hosseini (2022)"},{"why":"Its Prop. 2 is the stated source of the filter error bound in Prop. 7.","marker":"Al-Jarrah et al. (2023)"},{"why":"Its Theorem 2.4 justifies replacing the conditioning constraint with the coupling condition (9).","marker":"Kovachki et al. (2020)"},{"why":"Defines the feedback particle filter whose update the OT formulation recovers in the continuous-time limit.","marker":"Yang et al. (2016)"}],"fun_headline_variants":["Bayes as an optimal transport game","Neural transport maps for Bayes update","Bayes via Monge: a max-min trick","ML learns Bayes with transport maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The error guarantees rest on two unverified conditions: the filter must be uniformly geometrically stable, meaning past errors decay at a fixed exponential rate from any starting point, and the learned $f$ must make $x\\mapsto\\tfrac12|x|^2-f(x,y)$ strongly convex for every $y$; the paper's experiments do not check either.","fun_headline_variants_meta":{"raw":{"variants":["Bayes as an optimal transport game","Neural transport maps for Bayes update","Bayes via Monge: a max-min trick","ML learns Bayes with transport maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1218,"prompt_tokens":969,"completion_tokens":249,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":196}},"tokens_in":585,"tokens_out":249,"duration_ms":3772,"temperature":1.0,"reasoning_tokens":196,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:25:16.045808+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the static example $Y=\\tfrac12 X\\odot X+\\lambda_w W$ with $n=2$, train $f$ and $T$, compute the optimality gap $\\epsilon$ from (12), estimate $\\alpha$ as the minimum eigenvalue of the Hessian of $x\\mapsto\\tfrac12|x|^2-f(x,y)$ across $y$, and compare the empirical bounded-Lipschitz distance between $T(\\cdot,y)_\\#P_X$ and $P_{X|Y}(\\cdot|y)$ with $\\sqrt{4\\epsilon/\\alpha}$; a violation would refute Proposition 5's claim.","supporting_citations":[{"cited_title":"title Vector quantile regression: an optimal transport approach","cited_arxiv_id":null,"evidence_quote":"Its Theorem 2.3 is the basis for existence and uniqueness of the solution to the max-min problem in Prop. 4."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Its Theorem 4.3 is the bound that Prop. 5 extends to the approximate optimality-gap setting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Its Theorem 3.6 provides a second source for the optimality-gap error estimate used in Prop. 5."},{"cited_title":"Inproceedings","cited_arxiv_id":null,"evidence_quote":"The earlier paper where the optimal transport formulation of Bayes' law first appeared."},{"cited_title":"C @nEZ =PZ6' (|tZRoť Ɏ Z-[ 2n>9Z캼w]","cited_arxiv_id":null,"evidence_quote":"Defines the feedback particle filter whose update the OT formulation recovers in the continuous-time limit."}],"review_version":1}