{"id":"05986b5d-aefd-4d12-803a-a99b2d148f3b","arxiv_id":"2504.18377","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A rejection-sampling algorithm based on Langevin diffusion bridges gives exact samples from equality-constrained product distributions.","lead":"This paper introduces a Monte Carlo algorithm, Constrained Fusion, that draws exact samples from distributions constrained by equality conditions such as a sum constraint, using Langevin diffusions and rejection sampling. The method is aimed at disaggregating low-resolution readings, for instance splitting a daily electricity meter reading into finer time periods, while preserving uncertainty correctly.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2's arbitrary-constraint advantage rests on exact uniform samples from H via an unverified CHMC inner sampler; if CHMC has not converged, the claimed i.i.d. and multimodal behavior is not established.","rationale":"The reader's weakest assumption identifies the same load-bearing point: Algorithm 2 requires well-mixed uniform samples from the constraint manifold, and the paper does not provide guarantees for CHMC. I agree with that assessment. I do not see a comparably serious flaw in the linear-constraint exactness argument: the Girsanov-based rejection construction is standard, and the apparent issue with the normalizing constant ZH in (B.18) is not fatal because acceptance probabilities may be proportional to ZH up to constants that are independent of x and y. The missing minus sign in the printed line 4 of Algorithm 2 is a typographical slip rather than a conceptual error, since the surrounding text states the intended exp(-||y-x||²/2T) acceptance. The empirical application also has a separate concern about the source of Σ_t in the variance-constrained study, but that affects the strength of one data analysis, not the core sampling claim. The central advertised novelty beyond linear constraints is the arbitrary-constraint multimodal capability, and that novelty is only as strong as the uniform sampler feeding Algorithm 2. This is therefore the right thing to test: a direct-uniform replacement is a clean, inexpensive check that either validates the CHMC backend or shows where the method needs additional machinery.","tokens_in":32147,"tokens_out":17397,"duration_ms":203174,"concrete_test":"Re-run the nonlinear simulation of Section 3.3 and Figure 3 using Algorithm 2 with an exact uniform sampler on H instead of CHMC for Step 2. For the mean-variance constraint H={Σx_i=0, (1/3)Σx_i²=8}, H is a circle, so parameterize it and draw θ∼U[0,2π) deterministically. Keep all other settings and the same number of accepted samples. If the direct-uniform accepted samples reproduce the four modes and agree with the CHMC-backed samples in distribution (e.g., by energy distance or a KS-type test on the circle coordinate), the concern is resolved. If they differ, the inner MCMC is the bottleneck and the arbitrary-constraint claim needs to be re-scoped or the CHMC step needs explicit convergence diagnostics and guarantees.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The linear-constraint exactness of Algorithm 1 is well supported by Lemma 2 and Corollary 3. The second half of the central claim—natural handling of multimodal distributions on arbitrary constraints—is carried entirely by Algorithm 2, whose Step 2 requires sampling y ∼ U(H), the uniform distribution on the constraint manifold. In every nonlinear example the paper uses CHMC for this step, and CHMC is an asymptotically correct MCMC sampler. The paper provides no convergence guarantee for CHMC on a general manifold, and it reports no burn-in, thinning, or diagnostics for the Algorithm 2 runs in Figure 3 or the applications. If the CHMC chain has not reached stationarity, the actual proposal distribution is some Q≠U(H), so the rejection ratio in Corollary 3 is evaluated under a different proposal measure and the accepted samples are not distributed as f_H. Remark 4 argues that uniform sampling is easier than sampling the multimodal target, which is plausible for the circle example but is not a guarantee: a disconnected H or a narrow bottleneck can still trap CHMC even when the target density on H is constant. The paper itself flags this reliance in Appendix B.5, so the concern is not that the authors are unaware; it is that the most advertised nonlinear-constraint capability depends on a component whose correctness is asserted rather than verified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces the Constrained Fusion sampler, a rejection sampling method for equality-constrained product densities of the form f_H(y) ∝ ∏_{i=1}^m f_i(y_i) 1_{h(y)=0}. The method represents each f_i as the invariant distribution of a Langevin diffusion, proposes Brownian-bridge endpoints that satisfy the constraint (for linear constraints via a constrained Gaussian, and for general constraints via uniform samples from the manifold H), and applies a path-space rejection step based on the Girsanov density. The paper proves the acceptance ratio in Lemma 2 and Corollary 3, gives two algorithms, compares against importance sampling, random-walk Metropolis-Hastings, and CHMC in simulation studies, and applies the method to disaggregation of electricity consumption and medicine price data.","tokens_in":32400,"tokens_out":13454,"duration_ms":141097,"significance":"The core linear-constraint exactness claim is well supported: the derivation of the acceptance ratio is coherent, the algorithm produces i.i.d. samples when its ingredients are exact, and the simulation studies demonstrate practical gains in mean and variance estimation and in exploring the four modes of a circle-constrained target. The decoupling of the target distribution from the constraint is an elegant idea with potential value in disaggregation and imputation. The paper is considerably weaker for the advertised 'arbitrary constraints' claim because Algorithm 2 inherits correctness from an unverified uniform sampler on H; this part needs substantial additional support or a careful re-scoping of the claims.","major_comments":[{"comment":"The exactness of the arbitrary-constraint sampler requires y to be drawn exactly from U(H), the uniform distribution on the constraint manifold. The implementation uses CHMC as the inner sampler, but no convergence guarantee, burn-in, thinning, or stationarity diagnostics are reported for any of the nonlinear experiments (Fig. 3, Fig. 10, and Study 3 in Appendix E.3). If the inner chain has not reached stationarity, the proposal is some Q ≠ U(H), and the rejection ratio in Steps 4–8 of Algorithm 2 is no longer the Radon–Nikodym ratio of Corollary 3; accepted draws are then not i.i.d. from f_H, and the claimed natural handling of multimodal targets on arbitrary constraints is not established. This is load-bearing for the second half of the paper's central claim. Please provide verifiable diagnostics or mixing guarantees for the specific manifolds used, or explicitly state that the nonlinear algorithm is approximate.","section":"Section 3.2 (Case 2), Algorithm 2, Appendix B.5"},{"comment":"The nonlinear constraint is written as ∑_i Y_i^{(t)} = S_t and ∑_i (Y_i^{(t)} − S_t)^2 = Σ_t. If S_t is the daily total, the variance of the three within-day readings is proportional to ∑_i (Y_i^{(t)} − S_t/3)^2, so the displayed constraint is not the 'variance constraint' described in the text. Please correct the formula and confirm which constraint was used for Figure 6b; the interpretation of the uncertainty comparisons in Study 1 depends on this.","section":"Section 4.2, Study 1, display of H_t"},{"comment":"The exactness of the rejection step is conditional on Condition 1 holding for every f_i, but the paper does not verify the condition for the Student-t and Generalized Logistic densities used in the experiments, and it does not state explicitly whether the exact layered Brownian bridge construction of Appendix B.2 was used when φ_i is unbounded. Please add these verifications or state the additional assumptions; otherwise the 'exact' label on the simulation results is not fully supported.","section":"Section 3.1, Lemma 2 and Condition 1"}],"minor_comments":[{"comment":"The printed acceptance condition `log U1 ≤ ∥y−x∥^2/(2T)` is missing the minus sign; it should be `log U1 ≤ −∥y−x∥^2/(2T)` or equivalently `U1 ≤ exp(−∥y−x∥^2/(2T))`.","section":"Algorithm 2, Step 4"},{"comment":"The constants l_i disappear silently between equation (6) and equation (7); please state explicitly that they are absorbed into the proportionality constant of the Radon–Nikodym ratio.","section":"Equations (6) and (7)"},{"comment":"The displayed identity should read tr(Σ*) − ∑ σ_i^2 = −(1/w^2) ∑ σ_i^4; as written the sign is wrong, although the stated reduction is positive.","section":"Appendix F, Proposition 5"},{"comment":"The dimension of the FGM copula parameter θ is stated as R^{2m−m−1}; the standard dimension for an m-variate FGM copula is 2^m − m − 1.","section":"Appendix E.2, Definition 6"},{"comment":"Study 2 is not a product-density example; the text should explain how the single-target-density formulation maps onto Proposition 1 and the algorithm for product densities.","section":"Appendix E.2, Study 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable fit for stat.CO. The main risk is overclaiming for arbitrary nonlinear constraints; the authors should either supply a rigorous treatment of the inner uniform sampler or soften the abstract's claim to 'exact for linear constraints and approximate for nonlinear constraints.' The linear-constraint part is sound enough to build on, so I would not reject, but the nonlinear claim needs work before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's the take: this paper is worth reading if you work on exact constrained sampling. It extends Monte Carlo Fusion from a single coalescence point to arbitrary equality constraints, and for linear constraints the construction is genuinely exact and well motivated. The core Radon-Nikodym argument in Lemma 2 and Corollary 3 is coherent: the f_i^2 initial distribution cancels the Girsanov factor, leaving a rejection step on Brownian bridges, and the constrained Gaussian proposal is the right vehicle. The toy example and the linear-constraint simulation honestly match the theory.\n\nThe soft spots are in the nonlinear half. The stress-test worry is correct: Algorithm 2 samples y uniformly on H via CHMC, and if that chain is not converged the rejection ratio in Corollary 3 is against the wrong proposal. The authors acknowledge this in Appendix B.5, but the abstract still promises 'naturally deals with multimodal distributions on arbitrary constraints,' which is only as good as the inner uniform sampler. A disconnected or bottlenecked H can trap CHMC even with a flat target.\n\nI also found something the stress test missed. In Algorithm 2, line 4 reads: accept if log U1 ≤ ||y−x||^2/(2T). Since log U1 is negative and the right side is positive, this accepts everything. It should be log U1 ≤ −||y−x||^2/(2T), or U1 ≤ exp(...). As typeset, the nonlinear experiments are not running the algorithm the theory describes. That may be a typo, but it has to be fixed and the experiments rechecked.\n\nMinor points: Σ_t in the variance-constrained study appears without a source; if it is computed from the ground-truth high-frequency data, that should be said plainly. The code footnote has no visible URL. No burn-in or diagnostics are reported for the CHMC runs.\n\nNet: this is a serious paper with a real method for linear constraints, and a plausible but currently unproven endpoint for nonlinear constraints. It should go to peer review—a good referee will push for the sign fix, a crisp statement of what exactness means when the inner sampler is MCMC, and clearer reporting. I'd cite it once the nonlinear claims are cleaned up.","headline":"A genuinely new and mostly sound exact sampler for linear equality constraints; the nonlinear extension is advertised too strongly and, as printed, one acceptance line in Algorithm 2 appears to have a missing minus sign.","tokens_in":32935,"tokens_out":5537,"would_cite":true,"duration_ms":58013,"reading_group":"yes","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 introduces the Constrained Fusion sampler, which returns exact i.i.d.","keywords":["statistical disaggregation","equality constraints","exact sampling","Langevin diffusion","rejection sampling","Brownian bridge","time series imputation","constrained Monte Carlo"],"falsifier":"Run Algorithm 2 on a nonlinear constraint where the prescribed uniform sampler provably fails to mix within a fixed budget—say a thin curved constraint connecting two distant lobes—and compare the output with ground truth from independent long-run MCMC or numerical integration. If the Constrained Fusion output fails to place mass on both lobes, the claim that it is exact on arbitrary constraints is falsified in that regime; if it still succeeds, the inner sampler's non-convergence turns out to be harmless.","tokens_in":31938,"feed_emoji":"⚡","tokens_out":8279,"duration_ms":79789,"temperature":0.7,"pith_summary":"This paper tackles a common but intractable statistical task: given several independent component densities $f_1,\\dots,f_m$ and an equality constraint that ties their values together, draw samples from the joint density restricted to the constraint. Standard rejection sampling fails because the constraint has zero Lebesgue measure, and Markov chain methods on the constraint can miss disconnected modes. The authors' Constrained Fusion sampler instead couples the components through Brownian bridges whose endpoints are forced to satisfy the constraint, then applies an exact path-space rejection step. If the constraint is linear, the method returns exact independent samples; on nonlinear constraints it reduces the problem to drawing uniform samples from the constraint manifold, and the simulations show it can discover all modes of a multimodal target where a constrained Hamiltonian sampler gets stuck. A sympathetic reader would care because this makes disaggregation and imputation problems, such as recovering three daily electricity readings from a single daily meter total, statistically principled and gives calibrated uncertainty intervals.","feed_headline":"Exact draws from constrained distributions, no burn-in needed","feed_subtitle":"Brownian-bridge fusion makes linear constraints exact and still finds multimodal targets.","key_machinery":"The engine is the augmented target–proposal pair in Proposition 1: target $g_H(x^{(1:m)}, y^{(1:m)})\\propto \\prod_i f_i^2(x^{(i)}) p_i(y^{(i)}|x^{(i)})/f_i(y^{(i)})$ on $\\mathbb{R}^{md}\\times H$ and proposal $h_H\\propto \\prod_i f_i(x^{(i)}) \\cdot N(y^{(i)};x^{(i)},T I_d)$ restricted to $H$. Their Radon–Nikodym derivative, Corollary 3, is the expectation over Brownian bridges of $\\exp(-\\sum_i \\int_0^T \\phi_i(\\omega_s^{(i)})ds)$ with $\\phi_i(u)=\\frac12(\\|\\nabla\\log f_i(u)\\|^2+\\nabla\\cdot\\nabla\\log f_i(u))-l_i\\ge 0$, so rejection sampling on this derivative corrects any constrained-Gaussian proposal into the target. The second key piece is that the constrained Gaussian endpoint distribution is tractable: under a linear constraint $Ay=c$ the endpoints are Gaussian with an explicit normalising constant $Z_H(x^{(1:m)})\\propto \\exp(-(c-Ax)^\\top(AA^\\top)^{-1}(c-Ax)/(2T))$, which gives the first-stage acceptance probability. For nonlinear constraints the same rejection identity holds if endpoints are drawn uniformly from $H$, decoupling the distribution from the geometry of the constraint.","core_discovery":"The central claim is that sampling from $f_H(y^{(1:m)})\\propto \\prod_i f_i(y^{(i)}) \\mathbb{I}_{y^{(1:m)}\\in H}$ can be done exactly by simulating $m$ independent Langevin diffusions whose invariant laws are $f_i^2$, drawing Brownian-bridge proposals from initial points $x^{(i)}\\sim f_i$ to endpoints $y^{(i)}$ on $H$ distributed as a constrained Gaussian (or uniform, in the general case), and accepting with probability given by the expectation of $\\exp(-\\sum_i \\int_0^T \\phi_i(\\omega_s^{(i)})ds)$ over Brownian bridges. Proposition 1 identifies this proposal as a valid augmented distribution whose $y$-marginal is exactly $f_H$, and the rejection step is implemented exactly through Poisson-thinning when the $\\phi_i$ are bounded, with a layered Brownian-bridge construction for unbounded $\\phi_i$. For linear constraints Algorithm 1 is exact; for arbitrary smooth manifolds Algorithm 2 is exact provided one can sample uniformly from $H$.","pith_inferences":["The advertised exactness on arbitrary nonlinear constraints inherits fully from the assumption that the inner uniform sampler on $H$ mixes; because the paper gives no convergence guarantee for that step, in practice the nonlinear version is exact only up to the quality of the CHMC uniform samples.","An immediate testable extension is to swap the CHMC uniform sampler for a different uniform-on-manifold sampler, or an exact sampler when one exists, and check whether the fused output remains correct; this would isolate where the nonlinear method's error enters.","The MSE analysis suggests a practical decision rule for imputation: impose the constraint when residual variance dominates bias, since otherwise per-component errors can grow even when total variance shrinks — a caveat for naive users of constrained imputation.","Because the acceptance probability decays when the constraint is far from the typical set of the product density, a sequential Monte Carlo version of the fusion step — analogous to the paper's own proposed future work — is the natural route to scaling the method to high-dimensional or far-away constraints."],"forward_implications":["For any linear constraint, Algorithm 1 produces exact i.i.d. samples from the constrained product density, so the returned samples need no burn-in or convergence diagnostics.","Adding a sum constraint always reduces the total variance of a Gaussian model, and when model uncertainty exceeds its bias the constrained imputation also has lower mean squared error; simulations indicate the same holds for heavy-tailed non-Gaussian models.","The method turns time-series disaggregation into sequential constrained sampling: imputing high-resolution readings from low-frequency totals gives tighter, non-inflating uncertainty bands than unconstrained forecasting.","On nonlinear constraints such as mean-plus-variance constraints, the algorithm explores multiple modes of the constrained distribution with far fewer samples than a constrained Hamiltonian sampler, because the distribution and the constraint are handled separately.","Only one tuning parameter $T$ affects efficiency, not accuracy; the user does not have to pick step sizes, mass matrices or proposal scales for the target itself."],"supporting_citations":[{"why":"Supplies the diffusion-bridge Radon–Nikodym derivative used to derive the rejection ratio in Lemma 2.","marker":"[Besko et al., 2006]"},{"why":"Gives the invariant-density result that $f_i^2$ is the stationary law of the Langevin diffusion, anchoring the augmented target construction.","marker":"[Hansen et al., 2003]"},{"why":"Introduces rejection sampling based on Langevin diffusion bridges, the precursor this paper extends to product densities with arbitrary constraints.","marker":"[Dai, 2017]"},{"why":"Establishes the Monte Carlo Fusion framework of coalescing diffusions, which the Constrained Fusion sampler generalises.","marker":"[Dai et al., 2019]"},{"why":"Provides the layered Brownian-bridge construction that makes path-space rejection exact even when $\\phi_i$ is unbounded.","marker":"[Besko et al., 2008]"},{"why":"Defines constrained Hamiltonian Monte Carlo, both the baseline comparator and the uniform-on-manifold sampler used for nonlinear constraints.","marker":"[Brubaker et al., 2012]"},{"why":"Supplies the concrete constrained HMC implementation used in Algorithm 2 to draw uniform samples from the constraint manifold.","marker":"[Lelièvre et al., 2019]"},{"why":"Provides linearly constrained multivariate normal sampling used for the endpoint proposal in the linear-constraint case.","marker":"[Vrins, 2018]"}],"fun_headline_variants":["Brownian bridges crack constrained sampling exactly","Langevin bridges make equality-constrained sampling exact","Exact Monte Carlo for equality-constrained models","Statistical disaggregation with exact Langevin sampling","Exact draws from constrained models without burn-in"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, for nonlinear constraints, the step that draws uniform points from the constraint surface can be done exactly or at least mixed well enough; the paper does not prove this for its inner sampler, so if that sampler misses parts of the surface the whole exact-sampling claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Brownian bridges crack constrained sampling exactly","Langevin bridges make equality-constrained sampling exact","Exact Monte Carlo for equality-constrained models","Statistical disaggregation with exact Langevin sampling","Exact draws from constrained models without burn-in"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00083,"raw_usage":{"total_tokens":3623,"prompt_tokens":941,"completion_tokens":2682,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":2612}},"tokens_in":557,"tokens_out":2682,"duration_ms":21464,"temperature":1.0,"reasoning_tokens":2612,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:17:33.265870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 2 on a nonlinear constraint where the prescribed uniform sampler provably fails to mix within a fixed budget—say a thin curved constraint connecting two distant lobes—and compare the output with ground truth from independent long-run MCMC or numerical integration. If the Constrained Fusion output fails to place mass on both lobes, the claim that it is exact on arbitrary constraints is falsified in that regime; if it still succeeds, the inner sampler's non-convergence turns out to be harmless.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion-bridge Radon–Nikodym derivative used to derive the rejection ratio in Lemma 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the invariant-density result that $f_i^2$ is the stationary law of the Langevin diffusion, anchoring the augmented target construction."},{"cited_title":"(2017) A New Rejection Sampling Method Without Using Hat Function","cited_arxiv_id":null,"evidence_quote":"Introduces rejection sampling based on Langevin diffusion bridges, the precursor this paper extends to product densities with arbitrary constraints."},{"cited_title":"and Roberts, G","cited_arxiv_id":null,"evidence_quote":"Establishes the Monte Carlo Fusion framework of coalescing diffusions, which the Constrained Fusion sampler generalises."},{"cited_title":"and Roberts, G","cited_arxiv_id":null,"evidence_quote":"Provides the layered Brownian-bridge construction that makes path-space rejection exact even when $\\phi_i$ is unbounded."},{"cited_title":"and Urtasun, R","cited_arxiv_id":null,"evidence_quote":"Defines constrained Hamiltonian Monte Carlo, both the baseline comparator and the uniform-on-manifold sampler used for nonlinear constraints."},{"cited_title":"(2018) Sampling the Multivariate Standard Normal Distribution Under a Weighted sum Constraint","cited_arxiv_id":null,"evidence_quote":"Provides linearly constrained multivariate normal sampling used for the endpoint proposal in the linear-constraint case."}],"review_version":1}