Hybrid Scan Gibbs Sampler

• Hybrid Scan Gibbs Sampler is a class of MCMC algorithms that integrates systematic latent variable updates with stochastic parameter block updates for enhanced convergence.
• It combines exact Gibbs steps with approximate kernels like Metropolis and NUTS, allowing flexibility in tackling complex Bayesian models.
• The method guarantees reversible ergodicity and efficient convergence diagnostics, making it robust for hierarchical and high-dimensional inverse problems.

A hybrid scan Gibbs sampler is a class of Markov chain Monte Carlo (MCMC) algorithms that merges elements of systematic and random-scan Gibbs sampling in its variable update schedule, specifically tailored for hierarchical Bayesian models with latent variables or highly structured conditionals. The hybrid scan approach encompasses both the hybrid scan (HS) Gibbs samplers introduced for latent variable models, algorithms that alternate between systematic and random updates within each iteration, as well as more general “hybrid” schemes that use a mix of Gibbs steps and other MCMC kernels (e.g., Metropolis, NUTS) within block-specific updates. This paradigm offers both theoretical and practical advantages in terms of reversible ergodic sampling and often facilitates geometric convergence diagnostics or efficiency improvements in complex Bayesian hierarchical settings (Backlund et al., 2018, Qin et al., 2023, Uribe et al., 2021, Sørensen et al., 31 Mar 2026).

1. Algorithmic Structure and Scan Type

In a standard Gibbs sampler, either a systematic scan (SS) or a random scan (RS) is used. The SS update applies all full-conditionals in a fixed order every iteration, while RS selects exactly one block or coordinate (with fixed probability) at random. The hybrid scan Gibbs sampler (“HS” for short, Editor's term) updates the entire set of latent variables systematically at each iteration, while updating (randomly) only a subset of the remaining parameter blocks. Specifically, in the prototypical two-parameter, one-latent-block Bayesian model, the HS Gibbs sampler proceeds as follows (Backlund et al., 2018):

• Update all latent variables $Z$ given current $(X,Y)$.
• With probability $r \in (0,1)$, update $X$ given $(Y,Z)$; otherwise, update $Y$ given $(X,Z)$.

This places HS Gibbs between the SS and RS samplers: latent $Z$ is always updated, but the allocation of parameter block updates is randomized. Extensions to more complex or higher-order block structures naturally generalize this selection scheme, providing algorithmic flexibility.

More broadly, “hybrid” Gibbs in the MCMC literature refers also to blockwise algorithms where the method for drawing from each block-conditional is not homogeneous: e.g., some blocks are updated exactly from the Gibbs full-conditional, while others use an approximate Markov kernel (Metropolis, slice move, Laplace-approximation, or advanced methods such as NUTS). The “hybrid” terminology is appropriate when the algorithm combines deterministic and random-scan logic and/or employs different kernel types for different variable blocks (Qin et al., 2023, Uribe et al., 2021, Sørensen et al., 31 Mar 2026).

2. Full-Conditional Updates and Hybrid Kernels

The operational heart of the hybrid scan Gibbs paradigm lies in its treatment of blockwise full-conditionals. For a parameter partition $U = (X,Y)$ and latent $Z$, the relevant kernels are:

• $(X,Y)$0 (always updated),
• $(X,Y)$1 and $(X,Y)$2 (updated stochastically per scan probability $(X,Y)$3).

In other hybrid Gibbs contexts, especially high-dimensional or nonlinear cases, direct sampling from these conditionals may not be feasible. Here, a hybrid kernel is employed: a block is updated either by

• exact Gibbs (if possible),
• a Metropolis–Hastings move,
• a Laplace approximation independence sampler,
• random-walk Metropolis, or other MCMC kernel designed to leave the corresponding conditional invariant.

For example, in tomographic inverse problems, the image block ($(X,Y)$4) is updated via a Laplace approximation (local normal proposal) based on the current posterior landscape, while hyperparameters are sampled from conjugate or Metropolis-walks as appropriate (Uribe et al., 2021). In recent dynamic structural equation models (DSEMs) with binomial outcomes, one alternates between a Pólya–Gamma Gibbs step for auxiliary latent variables and a NUTS HMC step for marginal state/posterior parameters (see Section 3 below) (Sørensen et al., 31 Mar 2026).

A central theoretical result establishes that hybrid scan Gibbs samplers—where some conditionals are updated with approximate transition kernels but each kernel leaves the corresponding conditional invariant—inherit geometric convergence and $(X,Y)$5-spectral properties under explicit, computable bounds determined by the mixing properties of the blockwise MCMC kernels used (Qin et al., 2023).

3. Specialized Hybrid Scan Sampler Examples

A table of illustrative hybrid scan Gibbs instances:

Model Class / Problem	Block Update Scan/Type	Non-Gibbs Block Kernel
Bayesian latent variable regression	HS: all latents, 1 of $(X,Y)$6	None (full Gibbs)
DSEM with binomial outcomes	Full PG–Gibbs, NUTS step	NUTS HMC for parameters + Kalman
Tomographic reconstruction	Block hybrid Gibbs	Laplace-approx. (image), MH (angles)
High-dim regression, spike-and-slab	Hybrid random scan	Random-walk Metropolis

In the prototypical HS Gibbs problem, updating all latent variables systemically and each main block stochastically yields a reversible, Harris-ergodic Markov chain. The “sandwich” HSS variant introduces an auxiliary intermediate step on the latent space (reversible kernel in $(X,Y)$7) between the data augmentation and $(X,Y)$8 update, further improving mixing and spectral properties (Backlund et al., 2018).

In binomial DSEM inference, the hybrid NUTS–Gibbs sampler alternates a full (Devroye-style) Gibbs update of all Pólya–Gamma latent variables (exactly rendering binomial mixed-logit observations conditionally Gaussian) with a single NUTS step targeting the marginal posterior over population- and between-level parameters, integrating out state-space latent variables via a Kalman filter (Sørensen et al., 31 Mar 2026). The computational benefit is a reduction in effective dimension for the NUTS step (from $(X,Y)$9 to $r \in (0,1)$0 variables), yielding large gains in mixing efficiency and wall-clock time.

In high-dimensional tomographic Bayesian reconstruction with unknown view angles, the algorithm samples each block with a method suited to its structure: a Laplace-approximation-driven independence sampler for the image, componentwise random-walk Metropolis for the angle variables, and conjugate or approximate-Gamma draws for scale hyperparameters (Uribe et al., 2021). Such heterogeneity in kernel specialization within the same Gibbs cycle is a hallmark of the hybrid scan paradigm.

4. Theoretical Properties and Convergence

Hybrid scan Gibbs samplers display important theoretical properties concerning convergence rates and spectral gaps. Fundamental results are as follows (Backlund et al., 2018, Qin et al., 2023):

• For generic latent variable models, HS Gibbs is reversible and Harris-ergodic under mild conditions; geometric ergodicity can be established using drift and minorization criteria. Notably, geometric ergodicity for SS and RS Gibbs is unknown for several latent-variable models where it is easily proved for HS Gibbs.
• If a block is updated by an approximate, reversible kernel (e.g., Metropolis or slice step), and each such kernel mixes rapidly (spectral norm bounded away from one), then the overall hybrid chain’s spectral gap $r \in (0,1)$1 is bounded within a constant factor of the corresponding exact Gibbs ($r \in (0,1)$2). Explicitly, $r \in (0,1)$3, where $r \in (0,1)$4 is an upper bound on all the blockwise spectral norms.
• The sandwich-augmented HS (HSS) Gibbs algorithm always leads to operator improvement in both asymptotic variance and spectral gap, provided the sandwich kernel is positive and reversible with respect to the appropriate marginal.

These guarantees yield tractable central limit theorems for ergodic averages, provide guidance for algorithm design/tuning (e.g., block choice, proposal step size), and show that HS Gibbs (and its sandwich HSS variant) are often theoretically preferable for complex latent-variable models (Backlund et al., 2018, Qin et al., 2023).

5. Practical Applications and Efficiency

Hybrid scan Gibbs methodologies have been applied to a diverse array of problems:

• In Student’s $r \in (0,1)$5 location-scale models and heavy-tailed regression, HS Gibbs sampling with sandwich-steps significantly reduces the integrated autocorrelation time for key parameters, with almost no additional computational burden (Backlund et al., 2018).
• In shrinkage/mixed-effect regression, hybrid scan samplers are shown empirically to trade off slight loss in autocorrelation performance relative to systematic Gibbs for guaranteed geometric ergodicity and ease of theoretical verification (Backlund et al., 2018).
• For dynamic structural equation models with binomial data, the hybrid NUTS–Gibbs approach outperforms both Metropolis-within-Gibbs and pure HMC by factors of 5–16× in effective sample size per second, especially in large $r \in (0,1)$6, $r \in (0,1)$7 regimes, and enables efficient GPU parallelization (Sørensen et al., 31 Mar 2026).
• In edge-preserving tomographic reconstructions with angle uncertainty, hybrid Gibbs with block-specific kernel specialization achieves joint uncertainty quantification (image and angles) and strong mixing in high-dimensional settings, where fixed-angle methods suffer large increases in estimation error (Uribe et al., 2021).

A “plausible implication” is that the hybrid scan approach should be considered the default option for models with latent variables and block heterogeneity, especially when systematic scan geometric ergodicity is unproven or blockwise specialized kernels allow maximal exploitation of computational structure.

6. Recommendations and Limitations

• HS Gibbs has the same implementation complexity as systematic or random-scan Gibbs but often enables analytic convergence results unavailable for other scan types (Backlund et al., 2018).
• In settings where some blockwise conditionals admit exact Gibbs updates and others require approximate or Metropolis moves, the hybrid scan paradigm provides a theoretically sound integration of both, retaining overall reversibility, guaranteeing spectral gap preservation under explicit bounds, and supporting modular algorithm construction (Qin et al., 2023).
• Choice of scan probability and block structure can be tuned for computational balance, with sandwich steps further enhancing mixing at marginal cost.
• The main limitation occurs when blockwise MCMC kernels are poorly tuned (i.e., spectral norm close to one), in which case overall mixing decelerates notably; this demonstrates the necessity of monitoring the efficiency of each block update kernel.

Hybrid scan Gibbs samplers and their sandwich augmentations provide a principled, efficient, and demonstrably robust family of algorithms for posterior sampling in latent-variable and high-dimensional Bayesian inverse problems, with strong theoretical backing and documented practical superiority in numerous complex inference scenarios (Backlund et al., 2018, Qin et al., 2023, Sørensen et al., 31 Mar 2026, Uribe et al., 2021).