acceptodds
Under review as a conference paper at ICLR 2027

Generalized Differentiable Hamiltonian Dynamic Nested Sampling

Abstract

Nested sampling reduces the evidence integral to a one-dimensional quadrature over prior volume, at the cost of a complex inner problem: at every iteration, a new sample must be drawn from the prior restricted to a likelihood level set. These constrained draws dominate the running time and can bias the estimator. We present Generalized Differentiable Hamiltonian Dynamic Nested Sampling (GDHDNS), built around three complementary ideas to address this bottleneck. First, the live-point covariance provides a natural local preconditioner: its inverse defines a mass matrix that contracts with the constrained region, costs , and requires no Hessian. Second, a reflected Hamiltonian flow is integrable inside a convex level set, so its endpoint remains correlated with the point seeding it no matter how long the trajectory; this biases the evidence upward as a result. We break that correlation with a preconditioned slice refresh while retaining the flow's gradient-guided efficiency. Third, we show that batched removal has the same volume schedule as sequential removal, permitting unbiased vectorization, and that a run returns with the evidence. We evaluate GDHDNS on sweeps up to , a strong-lensing reconstruction problem, an interatomic potential, and an SDE mixed-effects model. Across the dimension sweeps it is the only method with unbiased evidence at every tested dimension, while running – faster than `dynesty` and up to four orders of magnitude faster than `PolyChord`.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Nested Sampling, Bayesian Evidence, Marginal Likelihood, Hamiltonian Monte Carlo, Slice Sampling, Differentiable Inference, Bayesian Model Comparison

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bma2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Generalized Differentiable Hamiltonian Dynamic Nested Sampling (GDHDNS), a framework for efficient Bayesian evidence estimation and posterior sampling. The proposed method combines covariance-based preconditioning, persistent reflective Hamiltonian dynamics, and preconditioned slice sampling to improve exploration of likelihood-constrained regions. It further introduces batched live-point replacement with a corresponding prior-volume shrinkage schedule and provides evidence gradients by differentiating a weighted likelihood estimator while holding sampler decisions fixed. The authors evaluate the method on synthetic distributions, gravitational lensing reconstruction, Lennard-Jones potentials, and a hierarchical stochastic differential equation model. The experiments demonstrate promising computational efficiency and evidence accuracy across several settings, although limitations remain in high-dimensional and multimodal problems.

Critical strengths

(1) A practically motivated combination of complementary sampling techniques. The proposed framework addresses important computational challenges in nested sampling by combining adaptive geometric preconditioning with reflective Hamiltonian dynamics and slice-based exploration. Using the empirical covariance of live points to construct the mass matrix is an appealing way to adapt to the geometry of the constrained region without requiring Hessian information. The batched replacement strategy also provides a practical mechanism for exploiting parallel computation. These design choices are well motivated by the computational bottlenecks of existing nested sampling methods.

(2) Broad empirical evaluation with meaningful accuracy diagnostics. The paper evaluates evidence estimation against analytically tractable references and considers challenging problems involving anisotropic, hierarchical, and multimodal distributions. Importantly, the experiments assess evidence accuracy in addition to computational efficiency, which is essential because accurate posterior reconstruction does not necessarily imply accurate marginal likelihood estimation. The authors also discuss settings where the method encounters difficulties, providing useful insight into the limitations of the proposed approach.

Critical weaknesses

(1) The stationary-distribution guarantees of the proposed constrained-sampling kernel are not sufficiently justified. The paper claims that the combination of reflective Hamiltonian dynamics and preconditioned slice sampling preserves the constrained prior distribution. However, the described slice-sampling procedure uses bracket doubling followed by shrinkage without specifying the additional acceptability test required by Neal's doubling construction for general disconnected slices. Consequently, the claimed invariance does not directly follow for arbitrary constrained regions. Furthermore, the method carries momenta across sampling iterations while recomputing the mass matrix from the evolving live-point covariance. Partial momentum refreshment does not generally preserve the required Gaussian momentum distribution when the mass matrix changes, unless an appropriate transformation or correction is applied. These issues are consequential because the correctness of the nested sampling evidence estimator depends critically on generating valid samples from the likelihood-constrained prior.

(2) The empirical contribution of the proposed components is not sufficiently isolated from existing methods. Although the paper reports substantial computational improvements, it does not directly compare against Gradient-Guided Nested Sampling (GGNS), which is presented as the closest methodological predecessor and a special case of the proposed framework. Moreover, the experiments lack systematic ablations separating the effects of covariance preconditioning, reflective dynamics, slice refreshment, momentum persistence, and batched replacement. Without these comparisons, it is difficult to determine which components are responsible for the reported improvements and whether the gains arise from the proposed sampling mechanism or primarily from implementation-level optimizations. Additionally, the differentiable evidence estimator relies on a standard posterior-weighted score identity with fixed sampling decisions, making its distinct theoretical contribution less clear.

Questions for authors

(1) Correctness of the slice-sampling update. Does the implementation include the additional acceptability test required by Neal's bracket-doubling procedure? The described algorithm appears to accept a candidate immediately upon satisfying the likelihood constraint, without checking whether the same bracket could have been generated from the proposed state. For disconnected slices, this can violate detailed balance. Could the authors clarify the exact procedure and provide a justification for invariance under general non-convex or disconnected constraints?

(2) Invariance under adaptive mass matrices and persistent momentum. How does the proposed kernel preserve its joint stationary distribution when the empirical mass matrix changes between generations while momentum is retained? If the previous momentum follows N(0,Mold)\mathcal{N}(0,M_{\mathrm{old}}) and the new refreshment noise follows N(0,Mnew)\mathcal{N}(0,M_{\mathrm{new}}), the partially refreshed momentum generally has covariance cos⁡2(δ)Mold+sin⁡2(δ)Mnew\cos^2(\delta)M_{\mathrm{old}}+\sin^2(\delta)M_{\mathrm{new}}, rather than MnewM_{\mathrm{new}}. Is an additional momentum transformation or resampling step applied in the implementation? If not, could the authors explain why this does not affect the claimed stationary-distribution guarantee?

(3) Comparison with the closest predecessor and component-level contributions. Could the authors provide a direct comparison with GGNS under matched computational budgets and a systematic ablation of the proposed components? In particular, comparing variants with and without preconditioning, slice refreshment, and momentum persistence would help determine whether the improvements arise from better constrained-prior exploration or primarily from vectorization and implementation choices. Such results would materially affect the assessment of the paper's methodological contribution.

Reviewer_ExMMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposed GDHDNS, a nested sampling scheme that is a particular numerical framework adapted from GGNS (Lemos et al. 2024). GDHDNS maintains a reflected Hamiltonian flow at every inner problem iteration, but apply a live-point covariance estimation to maintain an adaptive Hamiltonian geometry as well as batch removal of low-likelihood points at each outer loop to refresh the evidence. Comparing with dynesty, PolyChord and Tempered SMC under different low-to-high dimensional tasks, GDHDNS shows its bias removal ability and descently fast computation time, indicating the geometric-aware reflected Hamiltonian flow based sampling can be more powerful than original GGNS setup.

Critical strengths

GDHDNS improves the inner constrained sampler of nested sampling by combining geometry-adaptive Hamiltonian transport with slice refreshment. The result is mostly convincing, with reasonable ablation and plot.

Critical weaknesses

In all table and figures, comparison with GGNS is missing. I believe it is relevant since two challenging points (L063-077) and proposed changes in numerical scheme is claimed to be a direct improvement from GGNS. Therefore, missing comparison with GGNS really hurts the completeness of the paper as readers might eager to understand what are the additional gains from original GGNS scheme.

Questions for authors
  • Following from the major weakness, can author faithfully compare with GGNS and discuss the advantages of improvement from GGNS scheme?
  • What is relationship between Eq. (3), Eq. (5) and L788 in A1? It seems that cholesky factored matrices is multiplied in the actual algorithm but in Eq. (5) it is not.
  • In Figure 2, why GDHDNS claimed to be potentially succeed in reconstructing the from lensing data? I am unfamiliar with cosmology but can authors compare with some methods at least to show why and how it works compared with prior methods?
  • In Figure 8, why it is a good visualization to indicate GDHDNS has advantages over Dynesty?
Reviewer_7VsxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the problem of approximating the Bayesian posterior’s normalizing constant, ZZ, a.k.a. the Bayesian evidence. An established approach is nested sampling, which gradually increases a likelihood threshold λ\lambda and draws replacement samples from the constrained prior, π(θ∣L(θ)>λ)\pi(\theta\mid L(\theta)>\lambda). This work proposes a reflective-HMC-based method for sampling from this constrained distribution.
Building on earlier reflective HMC approaches to nested sampling, particularly Gradient-Guided Nested Sampling (GGNS) by Lemos et al. (2024), Improving Gradient-Guided Nested Sampling for Posterior Inference. Proceedings of Machine Learning Research.

The paper’s proposed contributions include:

  1. Adaptive preconditioning: Using the covariance of the live points to adjust Hamiltonian trajectories.
  2. Reflective HMC combined with slice sampling: Combining reflected trajectories with slice updates to improve exploration and reduce dependence between samples.
  3. Parallel replacement: Replacing multiple live points simultaneously.
Critical strengths
  1. The experimental results show promising computational performance. On the Gaussian and funnel benchmarks, the method reports substantially shorter runtimes than dynesty and PolyChord while maintaining evidence estimates close to the reference values. This suggests practical value for computationally demanding evidence estimation.

  2. The experiments cover different geometries and scientific models and use repeated runs. The reported limitations on high-dimensional and multimodal problems help identify where the method works well and where it remains inadequate.

Critical weaknesses
  1. Section 4.2 uses the non-ergodicity of deterministic billiard trajectories to motivate slice sampling, but this does not establish non-ergodicity of the full HMC process with momentum randomization. Appendix C explicitly acknowledges that momentum refreshment breaks the billiard invariant. So, the claimed theoretical necessity of slice sampling is not established. They could instead provide empirical evidence that adding slice sampling improves the performance, but this is missing in the current paper.
  2. Volume preservation is not established for the implemented finite-bisection reflection procedure.
  3. Propositions 1 and 2 appear to have limited novelty. Proposition 1 formalizes a batched-removal procedure and volume-accounting method already used in parallel nested sampling, including Henderson and Goggans (2014) and Yallup et al. (2026 Appendix D.2), and Proposition 2 applies the established Fisher score identity using posterior-weighted samples (see Poyiadjis et al., 2011, Section 2.1). Its implementation is potentially useful, but the proposition does not establish a new gradient-estimation principle or additional statistical guarantees.
  4. The main potential advance in the current work seems to be the combined sampling algorithm. However, the experiments omit GGNS—the method explicitly being extended—and as I mentioned earlier, the experiments do not test the effect of removing either the slice-sampling or reflected-HMC component. The reported speedups therefore do not establish whether the combination improves upon GGNS or a simpler sampler with the same preconditioning and parallel implementation.
Questions for authors
  1. What guarantees that locating a boundary crossing with eight bisections, reflecting, and continuing for the remaining time produces a reversible, volume-preserving map? If these properties hold only approximately, how is the resulting sampling error controlled or corrected?

  2. Are matched-computational-budget comparisons available against GGNS, together with removing either the slice updates or the reflected-HMC stage while retaining the same preconditioning and batching?

Comments on the paper

1 comment

Thank you for the detailed response and additional experiments. The GGNS comparison and ablations address my concern about missing empirical comparisons. I also appreciate the clarification that the score identity is not claimed as novel. However, my main concern about the correctness guarantees remains unresolved.

Slice update. The independently expanding bracket differs from Neal’s standard doubling construction, so it requires a separate invariance argument for the implemented procedure. The absence of detectable evidence bias over 20 seeds is reassuring, but does not establish invariance for general disconnected slices. This remains relevant because §4.2, lines 228–230, claims invariance without assumptions on the region’s shape. Could the authors justify the actual update, or replace it with a validated construction and reassess the results?

Adaptive mass matrix and persistent momentum. I accept the pure-scaling argument under an exact reflected flow and a suitable momentum distribution. However, this does not resolve general changes in metric shape. The factor cos⁡2(δ)≈0.13\cos^2(\delta) \approx 0.13 attenuates the covariance mismatch but does not bound its effect on the position distribution or evidence estimates. Replacing only a fraction of the live points also does not guarantee sufficiently small metric changes. Full refreshment removes this particular mismatch, but its empirical agreement with the default variant does not establish the latter’s invariance. The manuscript should distinguish the fixed-metric argument from the guarantees justified for the adaptive implementation.

Ablations. The new results support the importance of slice refreshment, but provide limited evidence for the additional benefit of the full combination. On the 32-D funnel, removing the reflected flow yields a similar mean evidence error and a smaller observed across-seed standard deviation, while reducing runtime from 31.47 s to 8.82 s. The Gaussian result shows a variance benefit at additional cost. These findings warrant more qualified claims about the benefits of the individual components and the full method.

I truly appreciate the substantive additions, but empirical robustness does not resolve the outstanding invariance questions. I therefore maintain my current rating, while remaining open to a justified or corrected kernel with corresponding validation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.