Add Gibbsian polar slice sampler

Project.toml

@@ -1,12 +1,13 @@
 name = "SliceSampling"
 uuid = "43f4d3e8-9711-4a8c-bd1b-03ac73a255cf"
-version = "0.2.1"
+version = "0.3.0"
 
 [deps]
 AbstractMCMC = "80f14c24-f653-4e6a-9b94-39d6b0f70001"
 Accessors = "7d9f7c33-5ae7-4f3b-8dc6-eff91059b697"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
 FillArrays = "1a297f60-69ca-5386-bcde-b61e274b549b"
+LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 LogDensityProblems = "6fdf6af0-433a-55f7-b3ed-c6c6e0b8df7c"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 Requires = "ae029012-a4dd-5104-9daa-d747884805df"
@@ -23,6 +24,7 @@ AbstractMCMC = "4, 5"
 Accessors = "0.1"
 Distributions = "0.25"
 FillArrays = "1"
+LinearAlgebra = "1"
 LogDensityProblems = "2"
 Random = "1"
 Requires = "1"

README.md

@@ -1,9 +1,43 @@
 # Implementation of slice sampling algorithms
 
+[![Stable](https://img.shields.io/badge/docs-stable-blue.svg)](https://Red-Portal.github.io/SliceSampling.jl/stable/)
+[![Dev](https://img.shields.io/badge/docs-dev-blue.svg)](https://Red-Portal.github.io/SliceSampling.jl/dev/)
+[![Build Status](https://github.com/Red-Portal/SliceSampling.jl/actions/workflows/CI.yml/badge.svg?branch=main)](https://github.com/Red-Portal/SliceSampling.jl/actions/workflows/CI.yml?query=branch%3Amain)
+[![Coverage](https://codecov.io/gh/Red-Portal/SliceSampling.jl/branch/main/graph/badge.svg)](https://codecov.io/gh/Red-Portal/SliceSampling.jl)
+
+This package implements slice sampling algorithms accessible through the the `AbstractMCMC` [interface](https://github.com/TuringLang/AbstractMCMC.jl).
+For general usage, please refer to [here](https://turinglang.org/SliceSampling.jl/dev/general/).
+
+## Implemented Algorithms
+- Univariate slice sampling algorithms with coordinate-wise Gibbs sampling by R. Neal [^N2003].
+- Latent slice sampling by Li and Walker[^LW2023]
+- Gibbsian polar slice sampling by P. Schär, M. Habeck, and D. Rudolf[^SHR2023].
+
+## Example with Turing Models
+This package supports the [Turing](https://github.com/TuringLang/Turing.jl) probabilistic programming framework:
+
+```@example turing
+using Distributions
+using Turing
+using SliceSampling
+
+@model function demo()
+    s ~ InverseGamma(3, 3)
+    m ~ Normal(0, sqrt(s))
+end
+
+sampler   = LatentSlice(2)
+n_samples = 10000
+model     = demo()
+sample(model, externalsampler(sampler), n_samples; initial_params=[1.0, 0.0])
+```
+
+[^N2003]: Neal, R. M. (2003). Slice sampling. The annals of statistics, 31(3), 705-767.
+[^LW2023]: Li, Y., & Walker, S. G. (2023). A latent slice sampling algorithm. Computational Statistics & Data Analysis, 179, 107652.
+[^SHR2023]: Schär, P., Habeck, M., & Rudolf, D. (2023, July). Gibbsian polar slice sampling. In International Conference on Machine Learning.
+=======
 [![Stable](https://img.shields.io/badge/docs-stable-blue.svg)](https://TuringLang.org/SliceSampling.jl/stable/)
 [![Dev](https://img.shields.io/badge/docs-dev-blue.svg)](https://TuringLang.org/SliceSampling.jl/dev/)
 [![Build Status](https://github.com/TuringLang/SliceSampling.jl/actions/workflows/CI.yml/badge.svg?branch=main)](https://github.com/Red-Portal/SliceSampling.jl/actions/workflows/CI.yml?query=branch%3Amain)
 [![Coverage](https://codecov.io/gh/TuringLang/SliceSampling.jl/branch/main/graph/badge.svg)](https://codecov.io/gh/Red-Portal/SliceSampling.jl)
 
-
-For a working example, please see [here](https://turinglang.org/SliceSampling.jl/dev/general/). 

docs/Project.toml

@@ -6,6 +6,7 @@ FillArrays = "1a297f60-69ca-5386-bcde-b61e274b549b"
 LogDensityProblems = "6fdf6af0-433a-55f7-b3ed-c6c6e0b8df7c"
 MCMCDiagnosticTools = "be115224-59cd-429b-ad48-344e309966f0"
 PDMats = "90014a1f-27ba-587c-ab20-58faa44d9150"
+Plots = "91a5bcdd-55d7-5caf-9e0b-520d859cae80"
 PrettyTables = "08abe8d2-0d0c-5749-adfa-8a2ac140af0d"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 SliceSampling = "43f4d3e8-9711-4a8c-bd1b-03ac73a255cf"

docs/make.jl

@@ -15,10 +15,11 @@ makedocs(;
         assets=String[],
     ),
     pages=[
-        "Home"                      => "index.md",
-        "General Usage"             => "general.md",
-        "Univariate Slice Sampling" => "univariate_slice.md",
-        "Latent Slice Sampling"     => "latent_slice.md"
+        "Home"                          => "index.md",
+        "General Usage"                 => "general.md",
+        "Univariate Slice Sampling"     => "univariate_slice.md",
+        "Latent Slice Sampling"         => "latent_slice.md",
+        "Gibbsian Polar Slice Sampling" => "gibbs_polar.md"
     ],
 )
 

docs/src/gibbs_polar.md

@@ -0,0 +1,76 @@
+
+# [Gibbsian Polar Slice Sampling](@id polar)
+
+## Introduction
+Gibbsian polar slice sampling (GPSS) is a recent vector-valued slice sampling algorithm proposed by P. Schär, M. Habeck, and D. Rudolf[^SHR2023].
+It is an computationally efficient variant of polar slice sampler previously proposed by Roberts and Rosenthal[^RR2002].
+Unlike other slice sampling algorithms, it operates a Gibbs sampler over polar coordinates, reminiscent of the elliptical slice sampler (ESS).
+Due to the involvement of polar coordinates, GPSS only works reliably on more than one dimension.
+However, unlike ESS, GPSS is applicable to any target distribution.
+
+
+## Description
+For a $$d$$-dimensional target distribution, GPSS utilizes the following augmented target distribution:
+```math
+\begin{aligned}
+    p(x, T)      &= \varrho_{\pi}^{(0)}(x) \varrho_{\pi}^{(1)}(x) \, \operatorname{Uniform}\left(T; 0, \varrho^1(x)\right) \\
+    \varrho_{\pi}^{(0)}(x) &= {\lVert x \rVert}^{1 - d} \\
+    \varrho_{\pi}^{(1)}(x) &= {\lVert x \rVert}^{d-1} \pi\left(x\right)
+\end{aligned}
+```
+As described in Appendix A of the GPSS paper, sampling from $$\varrho^{(1)}(x)$$ in polar coordinates magically targets the augmented target distribution.
+
+In a high-level view, GPSS operates a Gibbs sampler in the following fashion:
+```math
+\begin{aligned}
+T_n      &\sim \operatorname{Uniform}\left(0, \varrho^{(1)}\left(x_{n-1}\right)\right) \\
+\theta_n &\sim \operatorname{Uniform}\left\{ \theta \in \mathbb{S}^{d-1} \mid \varrho^{(1)}\left(r_{n-1} \theta\right) > T_n \right\} \\
+r_n      &\sim \operatorname{Uniform}\left\{ r \in \mathbb{R}_{\geq 0} \mid \varrho^{(1)}\left(r \theta_n\right) > T_n \right\} \\
+x      &= \theta r,
+\end{aligned}
+```
+where $$T_n$$ is the usual acceptance threshold auxiliary variable, while $$\theta$$ and $$r$$ are the sampler states in polar coordinates.
+The Gibbs steps on $$\theta$$ and $$r$$ are implemented through specialized shrinkage procedures.
+
+The only tunable parameter of the algorithm is the size of the search interval (window) of the shrinkage sampler for the radius variable $$r$$.
+
+!!! info
+    Since the direction and radius variables are states of the Markov chain, this sampler is **not reversible** with respect to the samples of the log-target $$x$$.
+
+## Interface
+
+!!! warning
+    By the nature of polar coordinates, GPSS only works reliably for targets with dimension at least $$d \geq 2$$.
+
+!!! warning
+    When initializing the chain (*e.g.* the `initial_params` keyword arguments in `AbstractMCMC.sample`), it is necessary to inialize from a point $$x_0$$ that has a sensible norm $$\lVert x_0 \rVert > 0$$, otherwise, the chain will start from a pathologic point in polar coordinates. If it is smaller than `1e-5`, the current implementation automatically sets the initial radius as `1e-5`.
+
+
+```@docs
+GibbsPolarSlice
+```
+
+## Demonstration
+As illustrated in the original paper, GPSS shows good performance on heavy-tailed targets despite being a multivariate slice sampler:
+```@example gpss
+using Distributions
+using Turing
+using SliceSampling
+using LinearAlgebra
+using Plots
+
+@model function demo()
+    x ~ MvTDist(1, zeros(10), Matrix(I,10,10))
+end
+model = demo()
+
+n_samples = 10000
+chain = sample(model, externalsampler(GibbsPolarSlice(10)), n_samples; initial_params=ones(10))
+histogram(chain[:,1,:], xlims=[-10,10])
+savefig("cauchy_gpss.svg")
+```
+![](cauchy_gpss.svg)
+
+
+[^SHR2023]: Schär, P., Habeck, M., & Rudolf, D. (2023, July). Gibbsian polar slice sampling. In International Conference on Machine Learning.
+[^RR2002]: Roberts, G. O., & Rosenthal, J. S. (2002). The polar slice sampler. Stochastic Models, 18(2), 257-280.

docs/src/latent_slice.md

@@ -15,9 +15,9 @@ where $$y$$ is the parameters of the log-target $$\pi$$, $$s$$ is the width of t
 Naturally, the sampler operates as a blocked-Gibbs sampler 
 ```math
 \begin{aligned}
-l &\sim \operatorname{Uniform}\left(l; \; y - s/2,\, y + s/2\right) \\
-s &\sim p(s \mid y, l) \\
-y &\sim \operatorname{slice-sampler}\left(y \mid s, l\right),
+l_n &\sim \operatorname{Uniform}\left(l; \; y_{n-1} - s_{n-1}/2,\, y_{n-1} + s_{n-1}/2\right) \\
+s_n &\sim p(s \mid y_{n-1}, l_{n}) \\
+y_n &\sim \operatorname{shrinkage}\left(y \mid s_n, l_n\right),
 \end{aligned}
 ```
 where $$y$$ is updated using the shrinkage procedure by Neal[^N2003] using the initial interval formed by $$(s, l)$$.

src/SliceSampling.jl

@@ -5,6 +5,7 @@ using AbstractMCMC
 using Accessors
 using Distributions
 using FillArrays
+using LinearAlgebra
 using LogDensityProblems
 using SimpleUnPack
 using Random
@@ -48,6 +49,13 @@ Return the initial sample for the `model` using the random number generator `rng
 """
 function initial_sample end
 
+function exceeded_max_prop(max_prop::Int)
+    error("Exceeded maximum number of proposal $(max_prop).\n", 
+          "Here are possible causes:\n",
+          "- The model might be broken or pathologic.\n",
+          "- There might be a bug in the sampler.")
+end
+
 # Univariate Slice Sampling Algorithms
 export Slice, SliceSteppingOut, SliceDoublingOut
 
@@ -64,9 +72,12 @@ include("doublingout.jl")
 
 # Latent Slice Sampling 
 export LatentSlice
-
 include("latent.jl")
 
+# Gibbsian Polar Slice Sampling 
+export GibbsPolarSlice
+include("gibbspolar.jl")
+
 # Turing Compatibility
 
 if !isdefined(Base, :get_extension)

src/doublingout.jl

@@ -1,20 +1,29 @@
 
 """
-    SliceDoublingOut(max_doubling_out, window)
-    SliceDoublingOut(window)
+    SliceDoublingOut(window; max_doubling_out, max_proposals)
 
 Univariate slice sampling by automatically adapting the initial interval through the "doubling-out" procedure (Scheme 4 by Neal[^N2003])
 
-# Fields
-- `max_doubling_out`: Maximum number of "doubling outs" (default: 8).
+# Arguments
 - `window::Union{<:Real, <:AbstractVector}`: Proposal window.
+
+# Keyword Arguments
+- `max_doubling_out`: Maximum number of "doubling outs" (default: 8).
+- `max_proposals::Int`: Maximum number of proposals allowed until throwing an error (default: `typemax(Int)`).
 """
 struct SliceDoublingOut{W <: Union{<:AbstractVector, <:Real}} <: AbstractGibbsSliceSampling
-    max_doubling_out::Int
     window          ::W
+    max_doubling_out::Int
+    max_proposals   ::Int
 end
 
-SliceDoublingOut(window::Union{<:AbstractVector, <:Real}) = SliceDoublingOut(8, window)
+function SliceDoublingOut(
+    window          ::Union{<:AbstractVector, <:Real};
+    max_doubling_out::Int = 8,
+    max_proposals   ::Int = typemax(Int),
+)
+    SliceDoublingOut(window, max_doubling_out, max_proposals)
+end
 
 function find_interval(
     rng  ::Random.AbstractRNG,

src/gibbs.jl

@@ -38,6 +38,7 @@ function AbstractMCMC.step(
     state  ::GibbsSliceState;
     kwargs...,
 )
+    max_prop = sampler.max_proposals
     logdensitymodel = model.logdensity
     w = if sampler.window isa Real
         Fill(sampler.window, LogDensityProblems.dimension(logdensitymodel))
@@ -48,15 +49,15 @@ function AbstractMCMC.step(
     θ  = copy(state.transition.params)
     @assert length(w) == length(θ) "window size does not match parameter size"
 
-    total_props = 0
+    n_props = zeros(Int, length(θ))
     for idx in shuffle(rng, 1:length(θ))
         model_gibbs = GibbsObjective(logdensitymodel, idx, θ)
         θ′idx, ℓp, props = slice_sampling_univariate(
-            rng, sampler, model_gibbs, w[idx], ℓp, θ[idx]
+            rng, sampler, model_gibbs, w[idx], ℓp, θ[idx], max_prop,
         )
-        total_props += props
-        θ[idx] = θ′idx
+        n_props[idx] = props
+        θ[idx]       = θ′idx
     end
-    t = Transition(θ, ℓp, (num_proposals=total_props,))
+    t = Transition(θ, ℓp, (num_proposals=n_props,))
     t, GibbsSliceState(t)
 end