Optimal Slice-Adaptive Tuning of Hybrid Slice Sampling
stat.CO · cs.LG · stat.ML
Quick verdict
Rigorous exact-cost analysis yields automated, near-optimal slice-adaptive tuning for slice sampling with guarantees; simulations only on synthetic targets.
Summary
This paper develops a rigorous cost analysis of one-dimensional hybrid slice sampling and converts it into fully automated window tuning. Under a contiguity assumption on slices, the author derives exact closed-form expressions for the expected number of target density evaluations per iteration as a function of the initial window width w, covering stepping out and two novel doubling variants (cached and lazy cached doubling); the shrinkage cost is characterized via a partial differential equation. The analysis shows the optimal window scales with the slice width λ: w = 1.358λ for stepping out (4.715 evaluations on average), 3.211λ for cached doubling (5.901), and 2.277λ for lazy cached doubling (5.410). Building on tractable surrogate costs, the paper introduces one-shot and gradient-based tuning rules with formal (ε,δ) suboptimality bounds, plus an online scheme that estimates conditional slice-width distributions from Markov chain draws and provably converges in probability under geometric ergodicity. Simulations on univariate targets, 256-dimensional products, mixtures, and Gamma/t targets with varying tail weight show near-oracle cost after roughly 100 iterations, insensitive to initial w spanning six orders of magnitude. Main limitations: the theory covers only contiguous (unimodal) slices, tuning relies on unanalyzed bounds on the unobserved slice width, and no code or empirical comparison with competing adaptive slice samplers is provided.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- For targets with contiguous slices, the expected per-iteration cost of stepping out, cached doubling, and lazy cached doubling is exactly (or asymptotically, for lazy) characterized as a function of w/λ, with oracle optima w* = 1.358λ (cost 4.715), 3.211λ (5.901), and 2.277λ (5.410) respectively.
- Lazy cached doubling saves roughly 15-25% of evaluations over Neal's original doubling at well-tuned w and is at most ~25% costlier than optimally tuned stepping out in the worst case, while stepping out can be arbitrarily worse in relative terms.
- Closed-form one-shot tuners and tractable gradient-based tuners carry formal (ε,δ) suboptimality guarantees (stepping out: factor 3; doubling: +4 or +0.7 evaluations) and require no user input.
- An online tuner estimating conditional slice-width distributions from Markov chain draws converges in probability to the surrogate-optimal cost under geometric ergodicity, with tuning overhead vanishing relative to sampling cost.
Strengths
Exact, closed-form cost characterization
Theorems 3.6 and 3.8 give exact expected per-iteration evaluation counts for stepping out and cached doubling as functions of w/λ; the shrinkage recursion is solved via a generating-function PDE (Lemma 3.3) yielding the clean formula E[Cshrink] = 1 + 2E[log((br−r)/λ + 1)] (Proposition 3.4).
Genuine algorithmic novelty with quantified payoff
Lazy cached doubling (Algorithms 9-10) short-circuits unnecessary evaluations; Fig 1 reports savings of ~1-2 evaluations at well-tuned w (~15-25% total cost reduction over Neal's original doubling), and Fig 3b shows it is at most ~25% costlier than stepping out while stepping out can be arbitrarily worse.
Tuning with real guarantees, not heuristics
Automation-ready for probabilistic programming
Honest scoping of assumptions
Weaknesses
Synthetic-only empirical evidence
Theory-practice gap on λ
Convergence is to surrogate optimum under extra preconditions
No statistical uncertainty in simulations
No code or data availability statement
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets the computational cost of Neal-style hybrid slice sampling on one-dimensional directions (also covering hit-and-run and Gibbs directions when slices are contiguous), measured as the expected number of target density evaluations per iteration, decomposed into slice-finding, shrinkage, and acceptance costs (Section 3). Central claims: (i) exact cost characterizations for stepping out (Theorem 3.6) and cached doubling (Theorem 3.8), plus a coarse continuity/asymptotic characterization for lazy cached doubling (Theorem 3.9); (ii) oracle optima w* ∝ λ with constants 1.358 (stepping out, 4.715 evaluations), 3.211 (cached doubling, 5.901), and 2.277 (lazy cached doubling, 5.410); (iii) closed-form one-shot (Eq. 8) and gradient-based (Eq. 9) tuners with formal suboptimality bounds (Theorem 4.2); (iv) an online tuner from chain draws converging in probability (Theorem 4.3); and (v) simulations showing convergence after roughly 100 iterations regardless of initial w ∈ [10⁻³, 10³].
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the optimal initial window width w for slice sampling?
This paper proves the optimal width is proportional to the exact slice width λ at the current slice level: w = 1.358λ for stepping out (about 4.715 density evaluations per iteration on average), w = 3.211λ for cached doubling (5.901), and w = 2.277λ for lazy cached doubling (5.410). Since λ changes every iteration, the paper builds automatic tuning schemes that adapt w as a function of the slice variable u.
How many target density evaluations does slice sampling cost per iteration?
When optimally tuned, roughly 4.7 evaluations per iteration with stepping out and 5.4-5.9 with the doubling variants, according to the paper's exact analysis. Lazy cached doubling (5.410) is recommended because it degrades most gracefully under mistuning.
Does this slice-adaptive tuning work for multimodal targets?
The theory assumes slices are contiguous intervals, which is guaranteed for unimodal targets. Simulations on two-component mixtures (Uniform, Normal, Laplace, Cauchy) suggest the tuners remain near-optimal in practice, but there is no theoretical guarantee for non-contiguous slices.
Is code available for this paper?
The paper text contains no code or data availability statement. It does provide complete pseudocode (Algorithms 1-10 in Appendix C) and all constants needed to reimplement the samplers and tuning schemes.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-8 appear only as captions plus in-text descriptions in the extracted text; quantitative curve values could not be independently verified from the figure rasters, so figure-based judgments rely on captions and surrounding prose.
- The extraction otherwise appears complete: main text, references [1]-[58], Appendices A-C, ending with Algorithm 10 — no truncation detected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.