Amplified Loaded Dice Roller (ALDR)

• Amplified Loaded Dice Roller (ALDR) is an adaptive rejection sampling method for univariate densities expressed as weighted mixtures, ensuring exact variate generation.
• It partitions the target density's support into adaptive strips, using tight majorizers and minorizers to compute explicit rejection bounds and refine proposals.
• ALDR is applied in Bayesian computation and MCMC, demonstrating low rejection rates and enhanced efficiency, even for non-log-concave weight functions.

The Amplified Loaded Dice Roller (ALDR)—a term not directly appearing but conceptually linked to the "Vertical Weighted Strips" (VWS) framework—refers to a class of adaptive rejection sampling schemes for drawing exactly from univariate densities that can be expressed in the form of a weighted product of a tractable base density and a nonnegative weight function. Originating in the literature on VWS, ALDR methods construct proposal distributions as finite mixtures, partitioning the target's support into adaptively chosen intervals ("strips"), and compute sharp, explicit bounds on the rejection probability. These schemes facilitate exact variate generation for statistically relevant, possibly unnormalized, weighted densities, particularly in Bayesian computation and Markov chain Monte Carlo contexts (Raim et al., 2024, Raim et al., 21 Sep 2025).

1. Mathematical Structure of Weighted Densities and Partitioning

The foundational ALDR methodology is developed for densities of the form

$$f(x) = \frac{w(x)g(x)}{\psi}, \quad \psi = \int_{\Omega} w(x)g(x)\, d\nu(x),$$

where $g(x)$ is a tractable base density (with $\Omega$ its support), $w(x) \ge 0$ is a weight function, and $\psi$ is an intractable normalizing constant in many applications. Generating draws from $f(x)$ is challenging due to the analytically unavailable normalization.

The ALDR/VWS scheme proceeds by partitioning $\Omega$ into $K$ disjoint strips $S_k = (\alpha_{k-1},\alpha_k]$ and, for each, constructing a majorizer $\overline{w}_k(x)$ satisfying $g(x)$0 for all $g(x)$1. Proposal densities are then given as mixture components,

$g(x)$2

leading to the overall proposal,

$g(x)$3

Sampling then consists of drawing a component $g(x)$4, then $g(x)$5. Acceptance occurs with probability $g(x)$6, independently within each strip (Raim et al., 2024).

2. Adaptivity and Proposal Refinement

To achieve high acceptance rates, the ALDR approach adaptively refines the partition. The key metric is the upper bound on the overall rejection rate,

$g(x)$7

where $g(x)$8 is derived using any minorizer $g(x)$9. Regions with high contributions to the rejection bound are iteratively split (typically at the midpoint or rejected values), focusing computational effort where $\Omega$0 fluctuates most relative to the base $\Omega$1. This recursive partitioning, tracked by stripwise $\Omega$2, rapidly decreases $\Omega$3 (Raim et al., 2024, Raim et al., 21 Sep 2025).

Within Markov chain Monte Carlo—particularly Gibbs sampling—where the target density changes across iterations, a self-tuned variant of VWS amortizes the cost of proposal construction. The proposal mixture is updated over repeated draws with knot insertion (to split strips) or deletion (to coarsen strips) driven by observed acceptance rates and stripwise contributions to $\Omega$4 (Raim et al., 21 Sep 2025).

3. Theoretical Efficiency and Exactness

Exactness of the ALDR scheme follows by construction: the proposal $\Omega$5 is everywhere an envelope for $\Omega$6, so samples have the correct target distribution when the accept-reject step is applied. The accept-reject constant is $\Omega$7, as $\Omega$8 for all $\Omega$9. Explicit, computable bounds on the rejection rate allow efficiency to be predicted and tuned without trial, a unique feature compared to non-adaptive rejection samplers. The acceptance rate approaches $w(x) \ge 0$0 as the partition is refined; for log-concave weights $w(x) \ge 0$1, linear majorizers can be used in place of constants, improving efficiency ("linear-VWS") (Raim et al., 2024).

4. Applications in Statistical Simulation

ALDR/VWS methods are particularly valuable when standard adaptive rejection sampling is not applicable, for example when $w(x) \ge 0$2 is not log-concave or the conditional distributions are unfamiliar. Applications include:

• Sampling from the von Mises–Fisher distribution on the sphere, where $w(x) \ge 0$3 and $w(x) \ge 0$4 is the exponential of a linear function, with demonstration that linear-VWS can achieve rejection rates as low as $w(x) \ge 0$5–$w(x) \ge 0$6 for moderate $w(x) \ge 0$7 (Raim et al., 2024).
• Within Bayesian small area estimation, sampling conditional latent variances where the density is the product of inverse-gamma and log-normal forms. Empirical evaluation demonstrates that self-tuned VWS within Gibbs greatly improves effective sample size and computational efficiency compared to independent Metropolis-Hastings, with typical knots $w(x) \ge 0$8 and tolerances $w(x) \ge 0$9, $\psi$0 (Raim et al., 21 Sep 2025).

5. Algorithmic and Implementation Considerations

Efficient implementation of ALDR schemes involves maintaining, for each conditional or target density, a sorted set of strip boundaries ("knots"), arrays of mixture weights, and CDF differences for the base density $\psi$1. Insertions and deletions of knots are $\psi$2 for search, and $\psi$3 for updating strip weights, with $\psi$4 typically less than $\psi$5. Initialization begins with a single strip, refining only as needed based on rejections. Numerical issues such as generating draws from truncated $\psi$6 on strips are commonly addressed by precomputing CDF bounds and inverting via quantiles (Raim et al., 21 Sep 2025).

6. Connections to Weighted Function Spaces and Beyond

Analogous weighted strip constructions appear in the theory of partial differential equations, notably in weighted Sobolev spaces for the Laplace operator in periodic infinite strips (Milisic et al., 2013). There, VWS terminology denotes function spaces with weights controlling algebraic growth/decay, yielding existence/uniqueness results and explicit solution representation via periodized Green functions. A plausible implication is that similar partitioning strategies and weight management principles underlie both the simulation (ALDR/VWS) and analytic (weighted PDE) domains, with the former emphasizing probabilistic approximation and the latter regularity theory in infinite or punctured domains (Kruse, 2019, Milisic et al., 2013).

7. Limitations and Extensions

Current ALDR/VWS approaches are developed for univariate densities; extension to higher dimensions is nontrivial due to the exponential growth of partitions with dimensionality. The efficiency of the scheme depends on the ability of the majorizers to tightly bound the weight function within each strip; extreme non-smooth structure in $\psi$7 can increase required $\psi$8 and thus proposal complexity. Open directions include the introduction of more flexible partition geometries (e.g., noncontiguous, adaptive to the base measure), treatment of multivariate targets, and incorporation of additional analytic information about the weight function (such as higher-order convexity or analyticity) to further compress the partition structure (Raim et al., 2024, Raim et al., 21 Sep 2025).