{"id":"81e7cd44-49a1-4d25-8603-0205ed9744db","arxiv_id":"2510.19283","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Conditional Brenier-map estimators provably converge to true conditionals, with mean map error decaying like N^{-1/4} (slow) or sqrt(log N / N) (fast), and these rates carry over to an idealized optimal-transport filter.","lead":"This paper derives the first quantitative error bounds for learning conditional optimal-transport (Brenier) maps from finite particle samples and for using them in nonlinear Bayesian filtering. The bounds give explicit convergence rates and reveal how sample size, map complexity, and filter stability interact; experiments on Lorenz 63 and Lorenz 96 show the method tracks non-Gaussian posteriors.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sec. 2.4 example misstates the metric entropy of the Legendre class: CF grows with degree M, so Theorem 2.10's fast rate does not follow for the example.","rationale":"The reader's conditional verdict is sensible, but the most load-bearing weakness is not Assumption 3.1 alone. Theorems 2.7 and 2.10 are stated as finite-sample guarantees with constants independent of N. For Theorem 2.10 to be meaningful, Assumption 2.9(4) must be satisfiable by a concrete class whose complexity constant CF is controlled. The Sec. 2.4 example is the only such demonstration, and its claim CF∝M+1 appears inconsistent with the basis count for d>1. Since the proof of Theorem 2.10 treats CF as fixed, a growing CF changes the optimization of σ and can degrade the rate from N^{−2/(2+γ)} to N^{−1/2} in the simplest case. This is an internal consistency issue in the example, not merely a scope limitation. It does not necessarily invalidate the conditional theorem, so the verdict remains conditional, but the rationale for the condition should be updated to include the entropy-growth problem.","tokens_in":50925,"tokens_out":14696,"duration_ms":117947,"concrete_test":"Take d=1, γ=0, and the class F(M) of degree-M Legendre polynomials on [−1,1] with |φ|≤R and α_min≤∇²_uφ≤β_max. Compute (or bound) log N(δ,F(M),L∞) as a function of M and δ. If the covering-number constant satisfies CF(M)=Ω(M), then substitute M=c(N/log N)^{1/2} into the θ(N,ε,t) bound of Proposition B.8 and attempt to find σ satisfying Eq. (B.21) with σ=(log N/N)^{1/2}. The check fails if the term sqrt(CF(M)/N) dominates σ, showing the fast rate is not achieved by the Sec. 2.4 construction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 2.10 is the central fast-rate claim, and Assumption 2.9(4) is its key sieve condition. The only worked example (Sec. 2.4) claims Assumption 2.6(3) holds with γ=0 and CF∝M+1. But F=F(M) is the set of Legendre polynomials of degree ≤M on Y×U⊂R^{m+n}; its metric entropy constant is proportional to the number of basis functions, (2M+1)^d−1 with d=m+n, not M+1 unless d=1. To meet Assumption 2.9(4), the example sets M=O((N/log N)^{1/(2+γ)}), so CF grows polynomially in N. In the localization proof, Proposition B.8 contains a term (1/N)R·CF(1−γ)^{-2}. If CF∼N^{1/(2+γ)} (for d=1) or worse, this term enters Eq. (B.21) as sqrt(CF/N), which for d=1, γ=0 is ∼N^{−1/4}, larger than the claimed σ∼(log N/N)^{1/2}. The square-rooted localization bound then cannot be made consistent with the fast rate; the induced squared-error rate is at best N^{−1/2}. Thus the example does not substantiate Theorem 2.10 unless CF is truly fixed or the proof is reworked with CF_N explicitly tracked.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops finite-sample error bounds for empirical conditional Brenier maps and applies them to an optimal-transport-based filter. Its main theoretical contributions are Theorem 2.7 (a slow O(N^{-1/2}) bound on the squared L^2 gradient error of the estimated conditional potential) and Theorem 2.10 (a fast O((log N/N)^{2/(2+γ)}) bound under additional Poincaré, smoothness, and sieve-bias assumptions). These map-estimation results are then used in Section 3 to bound the error of an idealized optimal-transport filter: Theorem 3.4 controls an approximate mean filtering error, and Theorem 3.7 extends the bound to the exact mean filtering error under filter stability and Lipschitz regularity conditions. Numerical experiments on bimodal, Lorenz-63, and Lorenz-96 benchmarks are reported in Section 4. The supplementary materials contain the detailed proofs of Theorems 2.7, 2.10, 3.4, and 3.7.","tokens_in":51352,"tokens_out":8368,"duration_ms":92704,"significance":"If the results hold as stated, the paper would constitute a valuable first quantitative analysis of conditional OT map estimation and of OT-based filters, extending the program of Hütter–Rigollet and Divol–Niles-Weed–Pooladian to the conditional/filtering setting. The proof strategy is coherent and the constants are tracked explicitly; the supplementary material is detailed, and the code is made available. However, the fast-rate result currently relies on a worked example whose entropy constant grows with the sieve size, which undermines the advertised application of Theorem 2.10. In addition, the filtering theorems apply to an idealized algorithm that is not the one implemented in the numerical section. These gaps are substantive but potentially repairable, so the manuscript is promising rather than ready.","major_comments":[{"comment":"The Legendre example is not a valid verification of the fast-rate assumptions. Assumption 2.6(3) fixes C_F as a constant independent of N, but Section 2.4 states that for the Legendre class F(M) the entropy bound holds with 'γ=0 and C_F ∝ M+1' and then sets M = O((N/log N)^{1/(2+γ)}). Hence C_F grows with N. This is not just a notational issue: the localization proof in SM B.3 uses C_F as a constant. In Proposition B.8, θ(N,ε,t) contains ε N^{-1/2} sqrt(C_F) ε^{-γ/2} sqrt(log(1+ε^{-1})) and N^{-1} C_F (1-γ)^{-2}. For γ=0 and C_F ~ N^{1/2}, choosing ε ~ (log N/N)^{1/2} gives θ ~ N^{-1/4} log^{3/4} N, which is much larger than ε. The fixed-point step leading to (B.23) then cannot recover the claimed squared-error rate; it would only give at best N^{-1/4}. Thus the only worked example that is supposed to satisfy Assumption 2.9(4) does not substantiate Theorem 2.10. The proof must either tra","section":"Section 2.4, Assumption 2.6(3), SM B.3"},{"comment":"The filtering theorems are proved for the idealized algorithm (3.26) and, with an extra resampling error, for (3.28). The text explicitly states that the interacting particle system (3.27) 'makes our theoretical analysis not applicable.' However, Algorithm 4.1 and the numerical experiments in Section 4 implement the interacting version: particles are propagated through previous optimal transport maps without regeneration, and in the paragraph around Eq. (4.4) the v_i^t are formed by shuffling the u_i^t. Consequently the Lorenz-63 and Lorenz-96 benchmarks do not test the end-to-end rates in Theorems 3.4 and 3.7. The manuscript should either state clearly in the abstract and conclusion that the error analysis covers an idealized filter and that the numerical algorithm is an interacting approximation, or provide a propagation-of-chaos or resampling-error analysis bridging the two. This is a","section":"Section 3.4 vs. Section 4"},{"comment":"The exact mean filtering error bound (3.14) contains the factor max(1, ϱ^t) with ϱ = 2 L_Y C'_D C_D C_stab + 1 - λ. For the bound to be uniform in time, (3.16) must hold, namely 2 L_Y C'_D C_D C_stab ≤ λ. This is a genuine contractivity condition on the product of the map Lipschitz constant, the coupling constant, the stability constant of the divergence, and the filter stability rate. The numerical section does not verify this condition, and for neural-network parameterizations the constants L_Y and C'_D are not computed. While the authors acknowledge in Remark 3.10 that this may be an artifact, the theorem as stated does not provide a uniform filtering rate for the reported benchmarks. The manuscript should state this limitation prominently and, ideally, provide a nontrivial example where (3.16) is verified.","section":"Theorem 3.7 and Remark 3.10"}],"minor_comments":[{"comment":"The sentence 'directly apply to the time-invariant case' appears to mean 'time-varying'; please correct the typo.","section":"Section 1.1"},{"comment":"The text refers to 'Assumption 2.9(6)', but Assumption 2.9 has only four items. The reference should be to Assumption 2.9(4).","section":"Proof of Theorem 2.10, SM B.3"},{"comment":"Even in the slow-rate Theorem 2.7, the statement 'Assumption 2.6 with γ=0 and C_F ∝ M+1' conflicts with Assumption 2.6(3), which requires C_F to be fixed. For clarity, define C_F(M) and state how the bound depends on it when M is allowed to grow.","section":"Section 2.4"},{"comment":"There are several spelling and notation issues, e.g., 'Lipshitz' for 'Lipschitz', and the notation eϕ in (B.4) is introduced as an argmin that may not exist; the text mentions a minimizing sequence but should make that explicit at first use.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The referee report above is my own assessment. The stress-test concern about the Legendre example lands: the fast-rate theorem is internally consistent as a conditional result, but the only worked verification of Assumption 2.9(4) is invalid because C_F grows with the sieve size. I would ask the authors to either rework the localization argument with C_F(N) explicitly tracked, or replace the example with a class whose entropy constant is independent of N while still achieving the required bias decay. The gap between the idealized algorithm analyzed and the interacting algorithm implemented is also important for the paper's framing. The manuscript has clear strengths—detailed supplementary proofs, explicit constants, reproducible code—so I lean toward major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first quantitative error analysis for conditional Brenier maps and for an OT-based filter. That is a real step beyond the existing unconditional OT map theory (Hütter–Rigollet, Chewi et al.), and the paper does it honestly: the slow-rate theorem is clean, the fast-rate proof in the supplement is a genuine adaptation of the one-shot localization argument, and the authors flag their own strong assumptions instead of burying them. The code is online and the experiments are reproducible. Worth a serious referee.\n\nWhat I actually learned: The conditional setting breaks convex conjugacy in the u-coordinate only, and the paper handles that correctly. Lemma B.1 and the empirical-process argument in Proposition B.8 are the load-bearing pieces, and the slow rate (Theorem 2.7) is solid. The fast rate (Theorem 2.10) is the real prize, and it is a theorem about any class satisfying Assumption 2.9, not a numerical observation.\n\nNow the soft spots, in proportion.\n\nThe filtering half (Section 3) rests on Assumption 3.1, uniform geometric stability of the true filter. The authors say it themselves: this is relatively strong and only verified for compact U or minorization-type conditions. That limits the practical reach of the end-to-end bound, but it is a standard type of assumption in the SMC/particle filter literature, not an artifact.\n\nMore worrying is the gap between the idealized algorithm (3.26)/(3.28) and the interacting algorithm actually benchmarked (3.27). The paper says the interacting version is outside the analysis. That is an honest scope limitation, but it means the numerics do not demonstrate the theory.\n\nThe sharpest issue is the worked example in Sec. 2.4. The paper claims Assumption 2.6(3) holds with CF ∝ M+1, then chooses M = O((N/log N)^{1/(2+γ)}) to satisfy the bias assumption. If CF genuinely grows with M, the localization bound in Proposition B.8 contains CF under a square root, and the fast rate degrades. The authors never track CF as a function of N in the proof. I have not fully verified whether the covering-number constant should really be CF ∝ M+1 or something smaller for the Legendre class with bounded derivatives, but as written the example does not substantiate Theorem 2.10. The theorem may still hold for classes with fixed CF, or with a reworked proof that tracks CF_N explicitly. This needs a fix before publication.\n\nBottom line: the core contribution is significant, the proofs are mostly careful, and the limitations are mostly stated. The unfixed example and the idealized-vs-implemented gap are the things I would ask the authors to address. A serious editor should send this to referees; I expect major revision or a reworked example.","headline":"First real error bounds for conditional Brenier maps and an OT-based filter; the central theorem is a genuine contribution, but the worked Legendre example likely does not deliver the advertised fast rate and the numerics use an algorithm outside the analysis.","tokens_in":51748,"tokens_out":1893,"would_cite":true,"duration_ms":18760,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G35","49Q22","65M32","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes finite-sample error bounds for conditional Brenier maps and derives end-to-end posterior error rates for an optimal-transport filtering algorithm.","keywords":["conditional optimal transport","Brenier maps","nonlinear filtering","sample complexity","Bayesian inference","simulation-based inference","empirical process theory","data assimilation"],"falsifier":"Run the OT filter on a hidden Markov model where the filter is not uniformly stable—for example, nearly uninformative observations with a transition kernel that lacks minorization—and test whether the empirical posterior error fails to decay with N as predicted. More directly, in a static conditional OT problem with log-concave conditionals, compute the empirical squared gradient error for N = 10^3, 10^4, 10^5 and check whether it follows (log N / N)^(2/(2+γ)); if it stalls at N^(-1/2) under conditions designed to trigger the fast rate, the localization argument or Poincaré assumptions are ins","tokens_in":50836,"feed_emoji":"📊","tokens_out":6135,"duration_ms":53563,"temperature":0.7,"pith_summary":"The paper's central project is to turn a heuristic—replacing Bayes' rule in filtering by a learned transport map—into a theorem with explicit convergence rates. It shows that a conditional Brenier map, estimated by solving an empirical dual optimal transport problem over a restricted function class, has squared-gradient error decaying like (log N / N)^(2/(2+γ)) under regularity assumptions, and at least like N^(-1/2) under weak assumptions. These are among the first quantitative statistical guarantees for conditional optimal transport maps, and they matter for simulation-based inference and Bayesian conditioning. The paper then feeds these rates into a filtering recursion, bounding the posterior error of an optimal-transport filter with a slow N^(-1/4) rate and a fast (log N / N)^(1/2) rate, provided the true filter is uniformly geometrically stable. If correct, practitioners get concrete sample-size guidance and a principled alternative to bootstrap particle filters.","feed_headline":"First sample-error rates proven for conditional transport maps","feed_subtitle":"A dual-OT analysis gives OT filters concrete convergence guarantees, with Lorenz tests to back them.","key_machinery":"The central object is the conditional Brenier map T(y,·) = ∇_u φ†(y,·), which pushes the reference measure η_U to the conditional posterior ν(·|y). The potential φ† is the minimizer of the conditional Kantorovich dual S(φ) = ∫[φ(y,v) + φ*(y,u)] dη dν. The analysis hinges on a stability lemma: the excess dual risk S(φ)−S(φ†) is sandwiched between (2β_max)^{-1} and (2α_min)^{-1} times the squared L² gradient error, so statistical control of the empirical dual objective implies control of the transport map itself. Empirical process chaining and a one-shot localization argument produce the N rates. In the filtering half, the same map is recomputed at every time step from particles, and uniform g","core_discovery":"The load-bearing result is Theorem 2.10: under strong convexity and smoothness of the model class, Poincaré inequalities on the reference and target conditionals, and a class rich enough to approximate the true potential to order (log N / N)^(1/(2+γ)), the empirical conditional Brenier potential achieves E‖∇φ̂ − ∇φ†‖²_{L²η} ≤ C (log N / N)^(2/(2+γ)). Reading the gradient as the transport map, this is a finite-sample, parameterized estimator converging at nearly the parametric rate. The paper also proves a slow rate, Theorem 2.7, requiring only boundedness and covering-number control of the function class. For filtering, Theorems 3.4 and 3.7 convert these map errors into posterior-divergence","pith_inferences":["The uniform geometric stability assumption is likely the main practical barrier; the paper notes it holds only for restricted systems, such as compact state spaces. A natural extension the authors leave open is replacing it with expectation-style or detectability-based stability to reach broader filtering models.","The exact-error rate is bottlenecked by Lipschitzness of Brenier maps in the observation y, which is open on noncompact domains. A testable extension is to prove or disprove a local or Hölder version; the paper shows Hölder exponents below 1 make its Grönwall argument diverge.","Because the fast rate is driven by localization, one could check whether (log N / N)^(2/(2+γ)) is minimax for conditional OT map estimation by adapting known lower bounds for unconditional OT maps; the paper does not address optimality.","The EnKF-reference trick suggests a broader design principle: choose a data-dependent reference close to the target so the learned correction is small, potentially extending to other linear-Gaussian approximations and reducing training difficulty."],"forward_implications":["Conditional OT maps used in simulation-based inference inherit a sample complexity: N^(-1/2) squared-gradient error under weak assumptions, and (log N / N)^(2/(2+γ)) under log-concavity and Poincaré conditions.","OT-based filters get end-to-end guarantees: expected posterior divergence decays at N^(-1/4) generally and (log N / N)^(1/2) under the fast-rate conditions, with errors accumulating only through a geometrically discounted sum.","The theory identifies the quantities practitioners should monitor: the filter stability constant λ, strong convexity and smoothness bounds of the potential class, Poincaré constants, and Lipschitz constants in the observation coordinate.","For systems with uniformly log-concave updates and quadratic-in-u potentials, the required regularity conditions persist over time, so the fast rates can be maintained uniformly in time.","An EnKF-based reference measure makes the learned map near-identity in near-Gaussian settings; numerical benchmarks show the OT filter matching or beating common filters on Lorenz 63 and Lorenz 96."],"fun_headline_variants":["First finite-sample error rates for conditional transport maps","OT filters get convergence guarantees with new error analysis","Triangular optimal transport filtering: now with proven rates","Sample-error bounds for Brenier maps in filtering scenarios"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the true filtering map forgets its starting distribution at a uniform geometric rate no matter what observations arrive—a condition the paper itself calls relatively strong and hard to verify outside compact state spaces; for the exact-error theorem, an additional fragile premise is that the conditional Brenier maps are uniformly Lipschitz in the observation, which is currently open for noncompact observation spaces.","fun_headline_variants_meta":{"raw":{"variants":["First finite-sample error rates for conditional transport maps","OT filters get convergence guarantees with new error analysis","Triangular optimal transport filtering: now with proven rates","Sample-error bounds for Brenier maps in filtering scenarios"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000156,"raw_usage":{"total_tokens":999,"prompt_tokens":634,"completion_tokens":365,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":378,"completion_tokens_details":{"reasoning_tokens":304}},"tokens_in":378,"tokens_out":365,"duration_ms":4121,"temperature":1.0,"reasoning_tokens":304,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T08:42:50.832675+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the OT filter on a hidden Markov model where the filter is not uniformly stable—for example, nearly uninformative observations with a transition kernel that lacks minorization—and test whether the empirical posterior error fails to decay with N as predicted. More directly, in a static conditional OT problem with log-concave conditionals, compute the empirical squared gradient error for N = 10^3, 10^4, 10^5 and check whether it follows (log N / N)^(2/(2+γ)); if it stalls at N^(-1/2) under conditions designed to trigger the fast rate, the localization argument or Poincaré assumptions are ins","supporting_citations":[],"review_version":1}