{"id":"42865c6e-3626-406c-b977-67d4444d962e","arxiv_id":"2502.00724","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A learned Bayesian Cramér-Rao bound that estimates both prior and measurement distributions from data, via a new Fisher score matching objective, converges almost surely to the true bound under regularity conditions.","lead":"This paper develops a data-driven way to compute the Bayesian Cramér-Rao bound, a fundamental lower limit on estimation error, when the measurement model and prior are not known. It learns score functions from samples using a new Fisher score matching technique, enabling benchmarks in problems such as quantized sensors and underwater acoustic frequency estimation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The practical FSM objective (14) replaces the true prior score with a learned one, but the theory (Theorem IV.1 and the finite-sample bounds) analyzes the true-prior objective; the resulting unquantified cross term breaks the finite-sample guarantees and leaves the consistency proof incomplete.","rationale":"The reader correctly identified the prior-score mismatch as a concern, but also listed Assumption V.2 (exact realizability) and the global-minimizer assumption as primary. In my reading, the most load-bearing issue is the internal inconsistency between the objective used in practice and the objective analyzed in the proof: Eq. (14) is what the practitioner optimizes, while Theorem IV.1 and the approximation-error bounds in Theorems V.1-V.2 concern the true-prior-score objective. The cross term arising from the learned prior score is neither bounded in finite samples nor shown to vanish uniformly in the proof of Theorem V.11. This does not mean the method is wrong or the consistency claim is false; the a.s. convergence may be repairable by leveraging prior-score consistency plus compactness of the parameter space. However, as written, the finite-sample relative-error bounds and Corollaries V.12-V.13 are not established for the algorithm actually implemented. The reader's CONDITIONAL verdict remains appropriate: the core idea is sound and the experiments are encouraging, but the central theoretical guarantees need a missing argument before they are accepted as stated.","tokens_in":44233,"tokens_out":7010,"duration_ms":75349,"concrete_test":"Analytically re-derive Theorem V.11 starting from Eq. (14), keeping the term Δ(Ω)=2E_{X,Θ}[s_F(X|Θ;Ω)^T(s_P(Θ;Ω_P^*)−∇_Θ log f_Θ(Θ))], and check whether Assumptions V.2 and V.3 imply sup_Ω |Δ(Ω)|→0 a.s. and a finite-sample bound on L0_F(Ω_F^*) in terms of the empirical FSM loss plus sqrt(L0_P(Ω_P^*)). If this term cannot be controlled, the finite-sample claims must be restricted to the true-prior objective or modified to include an explicit prior-score error term.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The finite-sample and consistency guarantees for the Measurement-Prior Approach are proved for the Fisher Score Matching objective containing the true prior score (Theorem IV.1, Eq. (26)), while the implemented objective (14) uses the learned prior score s_P(θ;Ω_P^*). The equivalence L0_F = L_F + C, which is used to conclude that the global minimizer of the empirical objective drives L0_F to zero (cf. the proof of Theorem V.10 in Appendix H.1 and the claimed analogous proof of Theorem V.11), no longer holds when the true prior score is replaced by the learned one. The additional discrepancy is 2E_{X,Θ}[s_F(X|Θ;Ω)^T(s_P(Θ;Ω_P^*) − ∇_Θ log f_Θ(Θ))]; no stated assumption or theorem bounds this term, nor is it shown to vanish uniformly in Ω. Consequently, Theorem V.2 and Corollary V.7, which are stated in terms of L0_F, do not directly bound the error of the actually trained model, and the proof of Theorem V.11 is incomplete as written. Prior-score consistency (L0_P→0 a.s.) may eventually make the term vanish, so the strong-consistency claim is plausibly repairable, but the finite-sample bounds are not justified for the implemented objective.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a learned Bayesian Cramér-Rao bound (LBCRB) that estimates the Bayesian FIM from i.i.d. parameter-measurement pairs when the prior and measurement distributions are unknown. Two approaches are presented: the Posterior Approach, which learns the posterior score via conditional score matching, and the Measurement-Prior Approach, which learns the prior score and a new Fisher score via a proposed Fisher Score Matching (FSM) objective, optionally using a physics-encoded score neural network. The authors derive finite-sample relative-error bounds for the learned FIM and LBCRB, prove strong consistency of the learned scores and of the LBCRB under realizability and global-minimization assumptions, and validate the method on linear Gaussian, quantized, and underwater-noise frequency-estimation examples. The central theoretical results are plausible and the appendices contain detailed proof sketches, but the practical FSM objective differs from the objective analyzed in the theory in a way that currently leaves a gap in the finite-sample guarantees.","tokens_in":44527,"tokens_out":5557,"duration_ms":58017,"significance":"If the identified gap is closed, this would be a substantial contribution: it provides the first fully learned Bayesian CRB that does not require knowledge of the Fisher score, introduces Fisher score matching as a new tool with potential independent value, and demonstrates strong empirical performance on problems where the BCRB is otherwise inaccessible. The paper ships reproducible code and validates against known bounds in linear Gaussian and quantized cases, which is a clear strength. The theoretical program is ambitious, with explicit non-asymptotic constants and a.s. consistency claims; however, those claims currently rest on an assumption set (realizability, global minimization) that is clearly stated but strong, and on an objective mismatch that must be resolved before the finite-sample results apply to the implemented method.","major_comments":[{"comment":"The implemented FSM objective (14) replaces the true prior score ∇_θ log f_Θ(θ) with the learned score s_P(θ; Ω_P^*), whereas Theorem IV.1 and all subsequent finite-sample results (Theorems V.2 and V.4, Corollaries V.6 and V.7) analyze the objective containing the true prior score. The equivalence L0_F = L_F + C used throughout the proofs no longer holds for the implemented objective; the additional cross term 2 E_{X,Θ}[s_F(X|Θ;Ω)^T(s_P(Θ;Ω_P^*) − ∇_Θ log f_Θ(Θ))] is neither bounded nor shown to vanish uniformly in Ω. Consequently, Theorem V.2 and Corollary V.7 do not directly bound the error of the actually trained model, and the proof of Theorem V.11 (Appendix H.1) is incomplete as written because it invokes the equivalence L0_F = L_F + C. Prior-score consistency (L0_P → 0 a.s.) may salvage the asymptotic consistency claim, but the finite-sample guarantees require either modifying the objective to use the true prior score or adding a quantified bound on this cross term.","section":"Section IV-B.1, Eq. (14) vs Theorem IV.1, Eq. (26)"},{"comment":"Equation (5) defines the prior FIM as F_P = E_Θ[∇_Θ log f_Θ(Θ) ∇_Θ log f_{X|Θ}(X|θ)^T], which is the product of the prior score with the Fisher score. This object is not symmetric positive semidefinite in general and is not the prior FIM; it contradicts the decomposition F_B = n_iid F_M + F_P given in Eq. (3) and the later evaluation in Eq. (16), which correctly uses s_P s_P^T. The second factor should be ∇_Θ log f_Θ(Θ)^T. Although the intended definition is clear from context, as written this is a foundational error that should be corrected.","section":"Section II-B, Eq. (5)"},{"comment":"The claim that with high probability c_B ≥ m_iid c_M + c_P, and hence that the Measurement-Prior Approach enjoys a √m_iid advantage in the empirical-mean error bound, is not rigorously established. Proposition V.5 only compares expectations of squared score norms; the passage from expectations to maxima over the dataset D is asserted (\"suggests also a similar large gap between the maxima\") without a concentration argument or additional assumptions. This is a load-bearing comparison for the claimed advantage of the Measurement-Prior Approach, so it should either be proved or explicitly labeled as a heuristic supported by experiments.","section":"Section V-B, paragraph following Proposition V.5"}],"minor_comments":[{"comment":"The statement defines d_P = intdim(F_M); it should read d_P = intdim(F_P), since the second term of the bound concerns the prior FIM.","section":"Theorem V.2"},{"comment":"In the positive-definiteness condition, the text says \"b_FB ≻ 0\" but the argument concerns the Measurement-Prior learned FIM; it should refer to b_F_MP. Relatedly, the sentence \"By definition we have b_V_MP ⪰ 0 and F_B ≻ 0\" appears to contain a typo (b_V_MP is the bound, not the FIM).","section":"Appendix G, proof of Corollary V.7"},{"comment":"The phrase \"standard divination\" should be \"standard deviation\" in two places.","section":"Section VII-C"},{"comment":"There are minor grammatical errors, e.g., \"within a Bayesian frameworks\" in the abstract; a careful proofreading pass is needed.","section":"Abstract and Introduction"},{"comment":"The consistency theorems rely on Assumption V.2 (NN-realizable score functions) and on global minimization of nonconvex objectives. These assumptions are stated explicitly, but it would help readers if the paper added a short limitations paragraph noting that the a.s. guarantees are conditional on realizability and global optimization, which are strong requirements in practice.","section":"Section V-D and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is real and lands: the practical objective (14) uses the learned prior score, while the theory analyzes the true-prior objective, leaving the finite-sample bounds without direct justification. This is repairable but is more than a presentation issue. The paper's other strengths—external validation, reproducible code, the PeSNN idea, and the breadth of experiments—make it worth a major revision rather than rejection. I would also ask the authors to fix Eq. (5) and to either prove or de-emphasize the c_B ≥ m_iid c_M + c_P comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline here is a genuinely useful new method: the Fisher Score Matching objective (Theorem IV.1) is a fresh score-matching variant for learning the Fisher score, and the Physics-encoded Score Neural Network combined with the Measurement-Prior approach is a practical, well-tested recipe. The paper ships code and validates against analytic BCRBs in several places, which is real evidence.\n\nThe main thing to know before citing or relying on it: there is a documented gap between the theory and the implemented objective. The practical FSM loss (14) replaces the true prior score with the learned prior score, but the theory — Theorem IV.1 and the finite-sample bounds in Theorems V.2–V.4 and Corollaries V.6–V.7 — is proved for the objective with the true prior score. The equivalence L0_F = L_F + C that underlies the consistency proof no longer holds for the implemented loss, and the extra cross term 2E[s_F^T(s_P^* − ∇ log f_Θ)] is not bounded. This doesn't sink the strong-consistency claim; it is plausibly repairable if prior-score consistency gives the needed vanishing. But as written, the finite-sample bounds are not justified for the actually trained model. A referee should ask for a rigorous treatment of this term.\n\nThe strong realizability and global-minimizer assumptions (V.2, V.10, V.11) are fairy-tale in practice, though standard in the learning theory literature; the paper could do more to discuss what happens when they fail. The numerical results are careful and align with analytic benchmarks, but the real-data examples (underwater noise) have no error bars, which makes the comparison less convincing.\n\nWho is this for? Anyone working on learned bounds for estimation problems where the likelihood is intractable or unknown, especially in signal processing. It deserves a serious referee; the core idea is sound, but the proof of Theorem V.11 is incomplete as written and the finite-sample guarantees need qualification.\n\nRecommendation: engage with it, but require a revision that either fixes the proof or clearly labels the finite-sample bounds as conditional on the true-prior-score objective.\n\nBest.","headline":"Useful new method for learned Bayesian CRBs, with a real gap between the theory and the implemented objective.","tokens_in":45002,"tokens_out":1880,"would_cite":true,"duration_ms":18850,"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 paper claims that the Bayesian Cramér-Rao bound can be learned entirely from measurement-parameter pairs by learning score functions, and proves that, under realizability assumptions, the learned bound converges almost surely to the…","keywords":["Bayesian Cramér-Rao bound","score matching","Fisher score matching","physics-encoded neural network","parameter estimation","Bayesian Fisher information","learned bounds","quantized measurements"],"falsifier":"In the linear-Gaussian example of Section VI-A with known matrices A and Σ, train the Posterior and Measurement-Prior LBCRBs with increasing dataset size ND and compare the learned Bayesian Fisher information matrix with the analytic expression; if the relative error does not approach zero or does not follow the predicted square-root decay as ND grows, the consistency and finite-sample claims are contradicted.","tokens_in":44003,"feed_emoji":"📊","tokens_out":7014,"duration_ms":69112,"temperature":0.7,"pith_summary":"The Bayesian Cramér-Rao bound (BCRB) says how accurately any estimator can perform, but it requires full knowledge of the prior and measurement distributions, which many real problems do not provide. This paper proposes a Learned BCRB (LBCRB) computed from a dataset of parameter-measurement pairs by learning score functions (gradients of log-densities) instead of the densities themselves. Two routes are given: the Posterior Approach learns the score of the posterior, while the Measurement-Prior Approach learns the prior score and the Fisher score separately and combines them through the standard decomposition of the Bayesian Fisher information. The paper introduces Fisher Score Matching, a variant of score matching for gradients with respect to the conditioning variable, and a Physics-encoded score neural network that builds known physical structure into the learned score. The result is a data-only bound with finite-sample error bounds, almost-sure convergence to the true BCRB, and demonstrated use on problems where the BCRB was previously unavailable.","feed_headline":"Score networks recover the Bayesian Cramér-Rao bound from data","feed_subtitle":"Two score-matching methods estimate the bound from parameter-measurement pairs, with proven convergence and error rates.","key_machinery":"The engine is Fisher Score Matching (FSM), a new score-matching objective (equation 26) that targets the Fisher score, the gradient of the log-likelihood with respect to the conditioning parameter rather than the data. Classical and conditional score matching cannot do this; FSM adds a cross term involving the true prior score that cancels the intractable part and makes the objective depend only on samples, once the prior score has been learned. The Measurement-Prior Approach then uses the exact decomposition of the Bayesian Fisher information into the number of i.i.d. measurements times the single-sample measurement Fisher information plus the prior Fisher information. When the measurement distribution is known to depend on the parameter through a physical model, the Physics-encoded score neural network restricts the Fisher score model to the chain-rule form involving that known model, which reduces network complexity and sample complexity.","core_discovery":"The central claim is that the Bayesian Fisher information matrix can be estimated consistently from i.i.d. parameter-measurement pairs by plugging learned scores into either the posterior identity or the decomposition into measurement and prior components, and that the relative error of the resulting bound is controlled by the score-matching objective. Formally, Corollaries V.12 and V.13 state that under the paper's assumptions the learned Fisher information matrix converges almost surely to the true one, so the LBCRB converges almost surely to the true BCRB. Theorems V.1–V.4 give non-asymptotic bounds in which the approximation error is bounded by a constant times the square root of the intrinsic dimension times the relative score error, and the empirical-mean error decays like a constant divided by the square root of the number of training samples, with a sample-size threshold that grows with the intrinsic dimension and the condition number of the bound.","pith_inferences":["The consistency proof assumes the true prior, posterior, and Fisher score functions lie exactly in the chosen neural network family; in practice this realizability is only approximate, so the convergence should be read as an idealization rather than a guarantee for any finite architecture.","The implemented Fisher score objective substitutes the learned prior score for the true prior score, while the theoretical analysis is carried out with the true prior score, leaving an unquantified bias in the practical algorithm that the current theorems do not cover.","The LBCRB is an approximation to the BCRB, not a guaranteed lower bound; a natural extension is to add error-correction or model-selection mechanisms to make it a valid bound, which the paper lists as future work.","The Fisher score learned by FSM is independent of the bound and could be reused for non-Bayesian Cramér-Rao bounds or for building learned maximum-likelihood estimators, directions the paper leaves open."],"forward_implications":["For any problem with a training set of parameter-measurement pairs, the LBCRB can be computed even when the prior and measurement distributions are completely unknown, including quantized and correlated-noise settings where density-based generative models fail.","The Measurement-Prior Approach can evaluate the bound for any desired number of i.i.d. measurements after a single training run, because it learns a single-sample Fisher score and scales it linearly.","With sufficient network capacity and a global minimizer, the learned bound converges almost surely to the true BCRB, so in the infinite-data limit the approximation error vanishes.","Finite-sample guarantees show the relative error of the learned bound scales with the square root of the intrinsic dimension divided by the number of training samples, and with the condition number of the true bound, so harder problems need more training samples.","Domain knowledge encoded through a physics-encoded score neural network reduces the network complexity and the sample complexity needed to reach a given accuracy."],"supporting_citations":[{"why":"Defines the Bayesian Cramér-Rao bound and the regularity conditions on the joint density that the whole analysis assumes.","marker":"[1]"},{"why":"Provides score matching, the method extended here to learn the prior score and, via Fisher score matching, the Fisher score.","marker":"[23]"},{"why":"Gives the Bayesian Fisher information decomposition and the standard Bayesian bound used as ground truth.","marker":"[40]"},{"why":"Presents the prior-score-based BCRB estimation approach that LBCRB extends by also learning the measurement distribution.","marker":"[22]"},{"why":"Introduces the generative Cramér-Rao bound, the main prior method for learned bounds that requires tractable densities such as normalizing flows.","marker":"[17]"},{"why":"Shows the difficulty of applying normalizing-flow-based learned bounds to quantized measurements, motivating score-based learning.","marker":"[21]"},{"why":"Supplies the approximation-error and empirical-mean-error decomposition used to structure the finite-sample bounds.","marker":"[45]"},{"why":"Provides matrix concentration inequalities with intrinsic dimension used in the empirical-mean-error theorems.","marker":"[46]"},{"why":"Gives the strong uniform law of large numbers that underlies the almost-sure consistency proofs for the learned scores.","marker":"[55]"}],"fun_headline_variants":["Score networks learn Bayesian Cramér-Rao bounds from data","Data-driven BCRB from score matching for unknown models","Learning the BCRB without full model knowledge via scores","Estimating Bayesian Fisher information with score networks","Learned BCRB: score networks replace known distributions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The almost-sure convergence and the non-asymptotic bounds rest on the assumptions that the true prior, posterior, and Fisher score functions lie exactly in the chosen neural network family and that training finds a global minimizer, and the practical objective also uses a learned prior score while the theory uses the true prior score, with the error from that substitution left unquantified.","fun_headline_variants_meta":{"raw":{"variants":["Score networks learn Bayesian Cramér-Rao bounds from data","Data-driven BCRB from score matching for unknown models","Learning the BCRB without full model knowledge via scores","Estimating Bayesian Fisher information with score networks","Learned BCRB: score networks replace known distributions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000233,"raw_usage":{"total_tokens":1500,"prompt_tokens":958,"completion_tokens":542,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":463}},"tokens_in":574,"tokens_out":542,"duration_ms":5828,"temperature":1.0,"reasoning_tokens":463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T17:57:19.072941+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the linear-Gaussian example of Section VI-A with known matrices A and Σ, train the Posterior and Measurement-Prior LBCRBs with increasing dataset size ND and compare the learned Bayesian Fisher information matrix with the analytic expression; if the relative error does not approach zero or does not follow the predicted square-root decay as ND grows, the consistency and finite-sample claims are contradicted.","supporting_citations":[{"cited_title":"Bayesian bounds for parameter estimation and nonlinear filtering/tracking,","cited_arxiv_id":null,"evidence_quote":"Gives the Bayesian Fisher information decomposition and the standard Bayesian bound used as ground truth."},{"cited_title":"Bayesian Cram ´er-Rao bound estimation with score-based models,","cited_arxiv_id":null,"evidence_quote":"Presents the prior-score-based BCRB estimation approach that LBCRB extends by also learning the measurement distribution."},{"cited_title":"Learning to bound: A generative Cram´er-Rao bound,","cited_arxiv_id":null,"evidence_quote":"Introduces the generative Cramér-Rao bound, the main prior method for learned bounds that requires tractable densities such as normalizing flows."},{"cited_title":"A generative Cram´er-Rao bound on frequency estimation with learned measurement distribution,","cited_arxiv_id":null,"evidence_quote":"Shows the difficulty of applying normalizing-flow-based learned bounds to quantized measurements, motivating score-based learning."},{"cited_title":"An introduction to matrix concentration inequalities,","cited_arxiv_id":null,"evidence_quote":"Provides matrix concentration inequalities with intrinsic dimension used in the empirical-mean-error theorems."},{"cited_title":"Generic uniform convergence,","cited_arxiv_id":null,"evidence_quote":"Gives the strong uniform law of large numbers that underlies the almost-sure consistency proofs for the learned scores."}],"review_version":1}