Slice Sampling¶
An MCMC method that adds a height beneath an unnormalized density and updates within its level-set slice using a transition that preserves the target distribution.
Core Idea¶
Slice sampling targets a density proportional to \(f(x)\) by adding a height \(y\) below \(f(x)\) and forming the slice \(S_y=\{x:f(x)>y\}\). An ideal update draws \(y\) uniformly from \((0,f(x))\) and then \(x\) uniformly from the full slice, giving \(f(x)/\int f\) as the stationary marginal. Practical methods may instead use another transition that preserves the slice's uniform conditional law; they do not necessarily draw an independent point from an entire disconnected slice.[^ref-c0c9330e46f7]
Scope of Application¶
Neal's scalar method randomly positions an interval around the current state, steps outward and shrinks after rejected proposals. It can update a nonconjugate conditional known only up to normalization. Murray, Adams and MacKay's elliptical specialization instead uses a Gaussian-prior direction and a likelihood threshold along an ellipse for latent Gaussian models. The bracket width of the scalar method and the Gaussian-prior assumption of the elliptical method should not be interchanged.[ref-c0c9330e46f7][ref-a5389ac8c2b7]
Clarity¶
The full mathematical slice may have several components; a practical interval is only a constructed search region. Neal explicitly shows that stepping out with fixed width can fail irreducibility across zero-density gaps. Stationarity of a valid kernel is therefore distinct from independent full-slice draws, global reachability and fast finite-run mixing.[^ref-c0c9330e46f7]
Manages Complexity¶
The auxiliary height converts density sampling into level-set membership and a valid conditional move. Stepping out and shrinkage can adapt to local width, reducing reliance on a single fixed proposal scale, but each density evaluation costs work and separated modes remain challenging. The user must report the actual transition rule and diagnostics rather than only the slice-sampling label.[^ref-c0c9330e46f7]
Abstract Reasoning¶
Check that \(f\) is nonnegative and integrable, derive the under-graph marginal by integrating out \(y\), and specify whether the horizontal update is ideal full-slice Gibbs or a practical invariant kernel. For a Gaussian-shaped \(f(x)=e^{-x^2/2}\), the illustrative height \(e^{-1/2}\) produces \(S_y=(-1,1)\). Then check reachability and mixing for the chosen practical implementation, particularly across density-zero regions.[^ref-c0c9330e46f7]
Knowledge Transfer¶
The height-threshold and invariant-transition logic transfers from a scalar bracket method to latent-Gaussian elliptical sampling, but the horizontal geometry and assumptions do not. The live Monte Carlo Simulation prime is the proposed broad parent; Slice Sampling adds the exact under-graph construction. An unrelated data “slice” is only a shared word.[ref-c0c9330e46f7][ref-a5389ac8c2b7]
[^ref-c0c9330e46f7]: Radford M. Neal, “Slice Sampling”, Annals of Statistics 31(3), 2003, original paper §3–4 and rejoinder pp.759–760. [^ref-a5389ac8c2b7]: Iain Murray, Ryan Prescott Adams and David J. C. MacKay, “Elliptical Slice Sampling”, AISTATS 2010, original paper §2 and Figure 2.
Relationships to Other Abstractions¶
Current abstraction Slice Sampling Domain-specific
Parents (1) — more general patterns this builds on
-
Slice Sampling is a kind of Monte Carlo Simulation Prime
Slice sampling is a Markov-chain Monte Carlo construction with an auxiliary height and slice-invariant update.
Hierarchy paths (4) — routes to 4 parentless roots
- Slice Sampling → Monte Carlo Simulation → Approximation → Representation → Abstraction
- Slice Sampling → Monte Carlo Simulation → Iteration
- Slice Sampling → Monte Carlo Simulation → Probability → Measure → Set and Membership
- Slice Sampling → Monte Carlo Simulation → Probability → Measure → Aggregation → Micro Macro Linkage
Neighborhood in Abstraction Space¶
Slice Sampling sits in a sparse region of the domain-specific corpus (81st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Numerical Root-Finding & Quadrature Methods (7 abstractions)
Nearest neighbors
- Epigraph — 0.85
- Space-Filling Curve — 0.83
- Ridders' Method — 0.82
- Jensen's Inequality — 0.82
- Interval Contractor — 0.81
Computed from structural-signature embeddings · 2026-10-08