{"id":"7fce40cc-2ba5-4dd4-a874-291f69f21822","arxiv_id":"2611.17008","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The sticky manifold PDMP provides a principled, modular framework for trans-dimensional MCMC that is rejection-free and retains the non-reversible, ballistic nature of component PDMP dynamics.","lead":"This paper introduces a new framework called the sticky manifold piecewise deterministic Markov process (SM-PDMP) for Bayesian model averaging. It allows efficient, rejection-free sampling across different submodels of varying dimensionality without requiring additional likelihood evaluations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Velocity compatibility condition (3.2) is stated but not verified for any non-trivial example beyond Gaussian or unit-velocity cases; the paper provides no general construction or existence proof for Q_stick and Q_unstick when the ambient and manifold velocity distributions differ in form (e.g.…","rationale":"The reader's weakest_assumption correctly identifies the velocity compatibility condition (3.2) as a non-trivial assumption that must be verified for each specific choice of dynamics and kernels. The paper provides examples only for cases where the ambient and manifold velocity distributions are essentially the same (Gaussian or uniform on spheres). The concern I raise is more specific: even when the manifold PDMP is derived from a coordinate chart (Algorithm 1), the velocity distribution on the manifold is the pushforward of the ambient velocity distribution through the chart derivative. If the chart is not isometric, this pushforward distribution differs from the ambient distribution, and the paper does not show how to construct Q_stick/Q_unstick satisfying (3.2) in that case. This is a concrete gap in the claimed generality. The paper acknowledges the issue implicitly (Section 3.5 discusses the cone example but does not provide a general solution) and states that the velocity compatibility condition is 'easily verifiable' (abstract), but no verification is given for non-isometric charts. The concrete test I propose would settle whether this gap is real or merely apparent. I agree with the reader's assessment that this is a significant concern, hence 'agree'. The verdict should remain CONDITIONAL because the paper is otherwise sound and the numerical examples are convincing, but this gap needs to be addressed before the claim of general applicability can be accepted.","tokens_in":71932,"tokens_out":815,"duration_ms":8726,"concrete_test":"Take a simple non-isometric chart, e.g., the stereographic projection of S^2 onto R^2 (Section 2.5). Specify the ambient PDMP with standard Gaussian velocity and the manifold PDMP with the pushforward of the Gaussian through the chart (which is not Gaussian). Construct explicit Q_stick and Q_unstick kernels that satisfy (3.2). If no such construction is provided, or if the construction requires resampling the velocity (thereby losing the persistence property claimed as an advantage), then the generality of the framework is limited to isometric charts or trivial velocity distributions.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The invariance proof (Theorem 3.5) rests on the velocity compatibility condition (3.2), which requires that Q_stick maps the incoming ambient velocity distribution to the prescribed joint distribution of the manifold velocity and the stored normal component, and Q_unstick does the reverse. The paper gives examples where this holds (Gaussian velocities, unit velocities), but these are essentially the same distribution on both sides. In the more general setting where the ambient and manifold velocity distributions differ—for instance, when the manifold PDMP uses a coordinate chart that is not isometric (Section 2.5, stereographic projection), the pushforward of the ambient Gaussian is not Gaussian on the manifold—the paper does not provide a construction of Q_stick/Q_unstick that satisfies (3.2). The authors acknowledge this gap in Section 3.5 (\"two charts of the quadratic cone\") but do not resolve it. Without a general method to construct these kernels, the applicability of the framework to non-isometric charts is unsubstantiated. This is a load-bearing concern because the central claim of the paper is that the SM-PDMP leaves the target invariant for any choice of component dynamics and unsticking kernels, subject to \"easily verifiable velocity compatibility conditions\" (abstract). If those conditions cannot be verified for a practically relevant class of manifold PDMPs, the claimed generality is overstated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes a sticky manifold piecewise deterministic Markov process (SM-PDMP) for Bayesian model averaging, targeting mixed-dimensional distributions that combine continuous and discrete model uncertainty. The key idea is to construct a PDMP that moves between an ambient Euclidean space and a lower-dimensional submanifold (or a nested sequence of submanifolds) by deterministically sticking to the submanifold upon intersection and unsticking after a random time. The unsticking rate is derived from the prior ratio and the normal velocity component, and the process is shown to leave the target invariant under a velocity compatibility condition. The paper also develops a manifold PDMP algorithm based on coordinate charts and a blow-up construction to handle singular sets. Numerical examples include Gaussian mixture models, quantum state tomography, Bayesian PCA, and survival extrapolation.","tokens_in":72314,"tokens_out":1364,"duration_ms":12636,"significance":"The paper addresses a challenging problem in Bayesian model averaging: efficient trans-dimensional sampling without reversible-jump proposals. The sticky manifold construction is principled, modular, and rejection-free, requiring no additional likelihood evaluations. The invariance proof is rigorous and general, accommodating non-linear flows and non-differentiable test functions via a flow-wise Gauss-Green formula. The numerical examples demonstrate competitive performance against reversible-jump methods, particularly in settings with varying posterior geometry. The paper also introduces Rao-Blackwellisation techniques to improve efficiency. The work is likely to be of interest to researchers in computational statistics and Bayesian inference.","major_comments":[{"comment":"The velocity compatibility condition (3.2) is central to the invariance proof but is only verified for Gaussian and unit-velocity cases. The paper acknowledges this gap in Section 3.5 but does not provide a general construction for Q_stick and Q_unstick when the ambient and manifold velocity distributions differ, e.g., under non-isometric charts. This limits the claimed generality of the framework. The authors should either provide a general method to construct such kernels or clarify the scope of applicability.","section":"Section 3.1, Eq. (3.2)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The assumption that M is of codimension one, orientable, and closed in R^d is restrictive. While the blow-up construction in Section 3.5 and Remark 3.12 extend to nested strata and overlapping manifolds, the paper does not fully address higher codimensions or non-orientable manifolds. The authors should discuss the practical implications of these restrictions and whether they can be relaxed.","section":"Section 3.2, Assumption 3.4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The RMSE for R=1 and d=10,20,50 shows that the sampler fails to make any trans-dimensional jumps, resulting in RMSE=0.5 (the prior weight). This is a significant limitation: when the submanifold is in a low-density region, the process may never visit it, leading to reducible dynamics. The authors acknowledge this but do not propose a remedy. This issue should be discussed more prominently as a practical limitation.","section":"Section 4.2, Table 1","authors_acknowledge":"The smaller sphere is reached rarely and produces longer residence episodes. marginal model-index process can be limited by the longer sojourn, even if the mixture weights of the target distribution are equalω= 1−ω= 1/2.","authors_acknowledge_p":0.69},{"comment":"The blow-up construction for the cone apex is elegant, but the paper does not verify the velocity compatibility condition for the polar coordinate chart. The Jacobian factor in Remark 3.14 suggests that the prior must diverge sufficiently rapidly for the process to hit the apex. This is a non-trivial condition that may not hold in practice. The authors should provide a concrete example where this condition is satisfied or discuss how to handle cases where it is not.","section":"Section 3.5, Example 3.13","authors_acknowledge":"If, instead, a prior does not diverge sufficiently rapidly in the original space, a na¨ ıve application of PDMP dynamics on the blow-up coordinates will have zero probability of hitting the front face, producing reducible dynamics.","authors_acknowledge_p":0.83}],"minor_comments":[{"comment":"The abstract states 'easily verifiable velocity compatibility conditions', but the paper does not provide a general verification procedure. Consider rephrasing to reflect the current state of the examples.","section":"Abstract","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The discussion of the stereographic projection for S^1 is clear, but the extension to higher-dimensional spheres (Remark 2.4) notes that no isometric chart exists. The paper could mention that the geodesic flow approach is used instead, which is fine, but the transition from chart-based to geodesic-based dynamics could be smoother.","section":"Section 2.5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The SM-Zigzag with velocity refreshment is described, but the paper does not specify how the refreshment rate is chosen or whether it affects the invariance. A brief comment on the rate selection would be helpful.","section":"Section 3.1, Example 3.3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The RMSE comparison uses different budgets (CPU time vs. number of evaluations). While both are informative, the CPU time comparison may be hardware-dependent. The authors should clarify that the conclusions are robust across both metrics.","section":"Section 4.3, Table 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The RJ-HMC results show high sensitivity to step size, as noted. The paper could include a brief discussion on how to tune HMC in practice for such problems, or recommend using PDMP dynamics as a more robust alternative.","section":"Section 4.4, Table 3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The discussion of length-scale differences and HMC acceptance rates is insightful. Consider moving some of this material to the main text to support the claim that PDMP dynamics are more robust to multiscale structure.","section":"Appendix C.5.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper is well-written and the notation is consistent. However, the length is substantial; some technical details in the appendices could be streamlined or referenced more concisely.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Hi,\n\nThis paper generalises the sticky PDMP idea (Bierkens et al. 2023a) from coordinate-aligned linear subspaces to general embedded submanifolds of codimension one. The core contribution is the sticky manifold PDMP (SM-PDMP), a modular framework that combines a Euclidean PDMP in the ambient space with a manifold PDMP on the submanifold, with sticking and unsticking kernels that preserve the mixed-dimensional target. The invariance proof (Theorem 3.5) is rigorous, using a blow-up construction and a flow-wise Gauss–Green formula. The paper also develops a coordinate-chart-based manifold PDMP (Algorithm 1) and a Rao–Blackwellised unsticking rate that can improve efficiency. The numerical examples are well-chosen and demonstrate the method on several problems (quantum state tomography, Bayesian PCA, survival extrapolation).\n\nThe main soft spot is the velocity compatibility condition (3.2). The paper states it as an assumption and gives examples where it holds (Gaussian velocities, unit velocities), but these are essentially the same distribution on both sides. The paper acknowledges in Section 3.5 that for non-isometric charts (e.g., stereographic projection) the pushforward of the ambient Gaussian is not Gaussian on the manifold, and does not provide a general construction for Q_stick/Q_unstick that satisfies (3.2) in that setting. The stress-test note is correct: this is a load-bearing concern because the claimed generality of the framework depends on these conditions being \"easily verifiable\" (abstract). The paper also restricts to codimension-one, orientable, closed submanifolds, which is a significant limitation for many practical problems. The blow-up procedure for higher codimensions is discussed but not fully developed, and the paper notes that it may limit the range of priors (Remark 3.14).\n\nOverall, this is a solid methodological contribution with a clean mathematical core. The limitations are acknowledged, but they are real and not minor. The paper deserves a serious referee; the authors should be asked to address the velocity compatibility condition more thoroughly, either by providing general constructions or by clarifying the scope of applicability.\n\nBest,\n[Your name]","headline":"A principled, modular framework for trans-dimensional PDMP that generalises sticky PDMPs to curved submanifolds; the core invariance proof is solid, but the velocity compatibility condition is a genuine practical bottleneck that the paper acknowledges but does not resolve.","tokens_in":72910,"tokens_out":570,"duration_ms":9323},"pith_extraction":null,"created_at":"2026-11-13T07:32:39.751312+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":null,"supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}