ASIS: Ancillarity-Sufficiency Interweaving Strategy

• ASIS is a strategy that alternates between ancillary and sufficient augmentations within each MCMC iteration to improve sampler mixing.
• It deterministically maps between centered and non-centered parameterizations, preserving the posterior while reducing dependence among samples.
• ASIS has been effectively applied in stochastic volatility, time-varying parameters, and gravitational lensing, yielding significant efficiency gains.

Searching arXiv for recent and foundational papers on ASIS to ground the article. arXiv search query: "Ancillarity-Sufficiency Interweaving Strategy ASIS stochastic volatility" Ancillarity–Sufficiency Interweaving Strategy (ASIS) is a Markov chain Monte Carlo acceleration technique proposed by Yu and Meng (2011) that interweaves two discordant parameterizations of a hierarchical model within a single Gibbs iteration: one in which the augmented variables are ancillary for a target parameter, and another in which they are sufficient. Across stochastic volatility, time-varying parameter, Student-$t$, gravitational-lensing, and panel-data settings, ASIS preserves the posterior target under an invertible deterministic mapping between parameterizations, but alters the path of the chain through parameter space so as to reduce dependence among successive iterates, improve mixing, and decrease autocorrelation (Kastner et al., 2017, Hosszejni et al., 2019, Tak et al., 2016, Nakakita et al., 24 Jul 2025).

1. Conceptual basis

In the ASIS framework, ancillary augmentation (AA) denotes a parameterization in which the augmented latent variables have a distribution that does not depend on the parameter of interest. By contrast, sufficient augmentation (SA) denotes a parameterization in which the augmented variables contain all the information needed to estimate that parameter. The central idea is not to choose between these augmentations once and for all, but to alternate between them inside each MCMC iteration (Tak et al., 2016, Kastner et al., 2017).

This distinction is operational rather than merely terminological. In the non-centered parameterization of the stochastic volatility model, the standardized latent process $z_t = (h_t-\mu)/\sigma$ has dynamics independent of $\mu$ and $\sigma$, so $z$ is near-ancillary for those parameters. In the centered parameterization, the latent log-volatility sequence $h_t$ depends directly on $\mu$ and $\sigma$ through the state equation and is near-sufficient for them. The same centered/non-centered dichotomy recurs in time-varying parameter models, where the centered latent increments are informative for the process variances, whereas the standardized non-centered states are ancillary for those variances (Kastner et al., 2017, Bitto et al., 2016).

The theoretical motivation is that the two parameterizations typically fail in complementary regimes. In stochastic volatility, the centered parameterization breaks down when the volatility-of-volatility parameter is small, whereas the non-centered parameterization shows deficiencies for highly persistent latent variable series. In the Student-$t$ degrees-of-freedom problem, the paper conjectures that ancillarity DA is progressively more efficient than sufficiency DA as $\nu$ increases, with the break-even point near $z_t = (h_t-\mu)/\sigma$0. In hierarchical panel models, SA is faster if $z_t = (h_t-\mu)/\sigma$1, while AA is faster if $z_t = (h_t-\mu)/\sigma$2 (Kastner et al., 2017, Hosszejni, 2021, Nakakita et al., 24 Jul 2025).

A common misconception is that ASIS is itself a new statistical model. It is instead a strategy for traversing the same posterior distribution through two equivalent augmentations. The target does not change; only the Markov transition does. This is why the literature repeatedly describes ASIS as “combining best of different worlds” rather than replacing model-specific latent-state or parameter updates (Kastner et al., 2017, Hosszejni et al., 2019).

2. Interweaving mechanism and parameter mappings

The algorithmic structure of ASIS is consistent across applications. A baseline sampler operates in one parameterization, then the current state is deterministically mapped into the alternative parameterization, one or more parameters are resampled there, and the chain is mapped back. Because the mappings are one-to-one and deterministic, no Jacobian correction is needed in the implementations described in the supplied papers (Tak et al., 2016, Kastner et al., 2017).

In the basic stochastic volatility model, the mapping is

$z_t = (h_t-\mu)/\sigma$3

In time-varying parameter models, the mapping is

$z_t = (h_t-\mu)/\sigma$4

In the gravitational-lensing time-delay model, the ancillary latent process $z_t = (h_t-\mu)/\sigma$5 and the sufficient latent process $z_t = (h_t-\mu)/\sigma$6 are related by

$z_t = (h_t-\mu)/\sigma$7

$z_t = (h_t-\mu)/\sigma$8

These mappings are linear and bijective with unit Jacobian (Bitto et al., 2016, Tak et al., 2016).

A compact summary of the application-specific augmentations appears below.

Model class	Ancillary augmentation	Sufficient augmentation
Stochastic volatility	standardized latent $z_t = (h_t-\mu)/\sigma$9 or $\mu$0	latent log-volatility $\mu$1
TVP models	standardized states $\mu$2	centered states $\mu$3
Gravitational lensing	latent OU states $\mu$4	transformed latent states $\mu$5
Student-$\mu$6 degrees of freedom	$\mu$7, $\mu$8	latent $\mu$9 or $\sigma$0
Gaussian panel data	$\sigma$1	latent random effects $\sigma$2

This repeated structure suggests that ASIS is most natural when two conditions hold simultaneously: the model admits equivalent centered and non-centered formulations, and at least one parameter block exhibits regime-dependent mixing under those formulations. That implication is explicit in the stochastic-volatility, Student-$\sigma$3, and panel-data treatments (Kastner et al., 2017, Hosszejni, 2021, Nakakita et al., 24 Jul 2025).

3. Canonical role in stochastic volatility models

The stochastic volatility literature is the most systematic development of ASIS in the supplied sources. The baseline model is

$\sigma$4

$\sigma$5

with $\sigma$6. The centered parameterization works directly with $\sigma$7; the non-centered parameterization uses $\sigma$8, yielding parameter-free latent dynamics $\sigma$9 (Kastner et al., 2017).

For latent-state sampling, the paper uses the 10-component normal-mixture approximation to $z$0 and AWOL banded-Cholesky sampling. Parameter updates are then carried out in both parameterizations within the same iteration. The paper recommends 2-block interweaving to avoid over-conditioning. Empirically, the computational overhead is negligible: the reported time per 1,000 iterations is about $z$1 s for the centered sampler and $z$2 s for GIS-NC, with nearly constant cost across parameter values and linear scaling in $z$3 (Kastner et al., 2017).

The empirical gains are regime-specific but broad. In a simulation design with $z$4 parameter constellations, the interwoven samplers are reported as always as good as or better than the best raw parameterization at negligible computational cost. For daily EUR/USD exchange rates, the inefficiency factors are reported as follows: centered $z$5, $z$6, $z$7; non-centered $z$8, $z$9, $h_t$0; GIS-C $h_t$1, $h_t$2, $h_t$3 (Kastner et al., 2017).

The leverage extension adds correlation between return shocks and volatility innovations: $h_t$4 Here the paper alternates centered and non-centered Random-Walk Metropolis–Hastings updates, and finds that repeating the interweaving multiple times per iteration—“ASISx5”—often outperformed a single interweaving step. In a grid with $h_t$5 DGPs and $h_t$6, RWMH-ASISx5 is reported to have ESR in $h_t$7 across DGPs, whereas AUX ranges from $h_t$8 to $h_t$9. For $\mu$0, the reported wall-clock minutes are $\mu$1–$\mu$2 for RWMH, $\mu$3–$\mu$4 for RWMH-ASISx5, and $\mu$5–$\mu$6 for AUX (Hosszejni et al., 2019).

These results do not imply that ASIS uniformly dominates all other samplers on all metrics. The leverage paper states that no universally best method exists, and that AUX can have lower inefficiency factors than RWMH-ASISx5, but RWMH-ASISx5 compensates via speed and produces markedly more stable effective sampling rates (Hosszejni et al., 2019).

4. Implementation in astronomical time-delay inference

In “Bayesian Estimates of Astronomical Time Delays between Gravitationally Lensed Stochastic Light Curves,” ASIS is embedded in a Metropolis–Hastings within Gibbs sampler for estimating time delays between lensed quasar light curves. The intrinsic quasar magnitude $\mu$7 is modeled as a continuous-time Ornstein–Uhlenbeck process,

$\mu$8

with irregularly spaced transitions

$\mu$9

Microlensing is modeled by a polynomial regression $\sigma$0, and the time delay $\sigma$1 enters through the alignment $\sigma$2 (Tak et al., 2016).

The ancillary augmentation is the latent OU state vector $\sigma$3, which is ancillary for the microlensing coefficients $\sigma$4 because the OU prior does not involve $\sigma$5. The sufficient augmentation is the transformed latent process $\sigma$6, which embeds $\sigma$7 into the latent-state dynamics and is sufficient for $\sigma$8. Inside each iteration, the sampler first updates $\sigma$9, then performs an AA Gaussian update for $t$0, constructs $t$1, performs an SA Gaussian update for $t$2, and finally maps back to $t$3. The proposal scales for $t$4 and $t$5 are adapted every $t$6 iterations to target acceptance rates in $t$7 (Tak et al., 2016).

The empirical benefit is highly parameter-specific. For the Q0957+561 dataset with a curve-shifted model $t$8, the reported $t$9 values are $\nu$0 for CMHwG, $\nu$1 for MHwG, and $\nu$2 for MHwG+ASIS; by contrast, $\nu$3 is $\nu$4, $\nu$5, and $\nu$6, respectively. The paper therefore states that ASIS substantially improves $\nu$7 mixing but has little impact on $\nu$8 in this application. The authors also report that the collapsed sampler requires roughly three times more CPU per iteration than the non-collapsed MHwG sampler, while ASIS is layered on MHwG and keeps this efficiency (Tak et al., 2016).

This application is notable because it makes explicit that ASIS may chiefly accelerate a nuisance or regression block rather than the headline parameter. The paper’s recommendation is correspondingly hybrid: use the profile likelihood to identify dominant mode(s) of $\nu$9, then run the ASIS-enhanced Bayesian sampler near those modes for coherent joint inference (Tak et al., 2016).

5. Shrinkage priors, robust tails, and parameter learning

In time-varying parameter models with shrinkage, ASIS is used as “boosting based on ASIS.” The model is built in a non-centered form

$z_t = (h_t-\mu)/\sigma$00

with double gamma shrinkage on $z_t = (h_t-\mu)/\sigma$01 and normal-gamma shrinkage on $z_t = (h_t-\mu)/\sigma$02. The baseline chain runs in the non-centered parameterization, then interweaves temporarily into the centered parameterization to resample the process variances $z_t = (h_t-\mu)/\sigma$03 and fixed coefficients $z_t = (h_t-\mu)/\sigma$04, after which it maps back. The full conditional for $z_t = (h_t-\mu)/\sigma$05 in the centered form is generalized inverse Gaussian, and the full conditional for $z_t = (h_t-\mu)/\sigma$06 is Gaussian (Bitto et al., 2016).

The gains reported for the EU inflation application are large. Under hierarchical double gamma shrinkage, the inefficiency factor for $z_t = (h_t-\mu)/\sigma$07 falls from $z_t = (h_t-\mu)/\sigma$08 without ASIS to $z_t = (h_t-\mu)/\sigma$09 with ASIS; for $z_t = (h_t-\mu)/\sigma$10, from $z_t = (h_t-\mu)/\sigma$11 to $z_t = (h_t-\mu)/\sigma$12; and for $z_t = (h_t-\mu)/\sigma$13, from $z_t = (h_t-\mu)/\sigma$14 to $z_t = (h_t-\mu)/\sigma$15. Under the hierarchical Lasso, $z_t = (h_t-\mu)/\sigma$16 falls from $z_t = (h_t-\mu)/\sigma$17 to $z_t = (h_t-\mu)/\sigma$18. The paper describes these changes as markedly improved mixing and shows faster movement across parameter space in Figure 1 (Bitto et al., 2016).

In the Student-$z_t = (h_t-\mu)/\sigma$19 degrees-of-freedom problem, ASIS interweaves a sufficient augmentation based on latent $z_t = (h_t-\mu)/\sigma$20 or $z_t = (h_t-\mu)/\sigma$21 with an ancillary augmentation based on $z_t = (h_t-\mu)/\sigma$22, where $z_t = (h_t-\mu)/\sigma$23 a priori. The sufficient augmentation permits exact rejection sampling for $z_t = (h_t-\mu)/\sigma$24 under an exponential prior, while the ancillary augmentation uses Metropolis updates on $z_t = (h_t-\mu)/\sigma$25 with a Jacobian factor $z_t = (h_t-\mu)/\sigma$26. The paper states that ancillarity DA becomes progressively more efficient as $z_t = (h_t-\mu)/\sigma$27 increases, with the break-even point near $z_t = (h_t-\mu)/\sigma$28, and that ASIS combines the benefits of both DAs (Hosszejni, 2021).

The reported Relative Numerical Efficiency values make the complementarity concrete. At $z_t = (h_t-\mu)/\sigma$29, AA rises from about $z_t = (h_t-\mu)/\sigma$30 at $z_t = (h_t-\mu)/\sigma$31 to about $z_t = (h_t-\mu)/\sigma$32 at $z_t = (h_t-\mu)/\sigma$33, while SA falls from about $z_t = (h_t-\mu)/\sigma$34 at $z_t = (h_t-\mu)/\sigma$35 to about $z_t = (h_t-\mu)/\sigma$36 at $z_t = (h_t-\mu)/\sigma$37. ASIS is reported at about $z_t = (h_t-\mu)/\sigma$38 for $z_t = (h_t-\mu)/\sigma$39 and up to about $z_t = (h_t-\mu)/\sigma$40 for $z_t = (h_t-\mu)/\sigma$41. The same paper also notes that AA can experience numerical failures in inverse-CDF evaluation for extremely heavy-tailed data and extreme initializations, whereas ASIS avoids these failures and retains higher RNE (Hosszejni, 2021).

6. Convergence theory, practical scope, and limitations

The most explicit recent convergence theory in the supplied material is for Gaussian hierarchical panel models. With

$z_t = (h_t-\mu)/\sigma$42

the paper derives a trade-off identity implying that SA is faster if $z_t = (h_t-\mu)/\sigma$43, whereas AA is faster if $z_t = (h_t-\mu)/\sigma$44. Under the asymptotic condition $z_t = (h_t-\mu)/\sigma$45, the $z_t = (h_t-\mu)/\sigma$46-chain under ASIS becomes approximately i.i.d., and the spectral radius tends to $z_t = (h_t-\mu)/\sigma$47 (Nakakita et al., 24 Jul 2025).

The reported Monte Carlo results support the theoretical ordering even for small panels. For $z_t = (h_t-\mu)/\sigma$48, the MCSE values for $z_t = (h_t-\mu)/\sigma$49 are $z_t = (h_t-\mu)/\sigma$50 for SA, $z_t = (h_t-\mu)/\sigma$51 for AA, and $z_t = (h_t-\mu)/\sigma$52 for ASIS under Pattern 1; $z_t = (h_t-\mu)/\sigma$53, $z_t = (h_t-\mu)/\sigma$54, and $z_t = (h_t-\mu)/\sigma$55 under Pattern 2; and $z_t = (h_t-\mu)/\sigma$56, $z_t = (h_t-\mu)/\sigma$57, and $z_t = (h_t-\mu)/\sigma$58 under Pattern 3. For the U.S. cigarette panel, the reported MCSE values are $z_t = (h_t-\mu)/\sigma$59 for SA, $z_t = (h_t-\mu)/\sigma$60 for AA, and $z_t = (h_t-\mu)/\sigma$61 for ASIS (Nakakita et al., 24 Jul 2025).

Several limitations are also explicit in the supplied sources. ASIS does not necessarily improve every parameter equally: in astronomical time-delay estimation it chiefly accelerates $z_t = (h_t-\mu)/\sigma$62 and has little effect on $z_t = (h_t-\mu)/\sigma$63 (Tak et al., 2016). It is not always worth applying ASIS to every sampler component: in the leverage SV paper it is not applied to AUX because the collapsed parameter step already accounts for about $z_t = (h_t-\mu)/\sigma$64 of runtime and interweaving would add large overhead but little benefit (Hosszejni et al., 2019). Finite-sample behavior can deviate from approximate AR(1) theory in panel models, and AA can encounter numerical inverse-CDF difficulties in the Student-$z_t = (h_t-\mu)/\sigma$65 setting (Nakakita et al., 24 Jul 2025, Hosszejni, 2021).

Taken together, these results characterize ASIS as a general strategy for hierarchical and latent-variable models in which centered and non-centered parameterizations exchange the roles of ancillarity and sufficiency. The recurring empirical pattern is not uniform dominance of one augmentation, but complementarity between them. This suggests that the central value of ASIS lies in robustness across parameter regimes and model classes rather than in any single, fixed efficiency guarantee (Kastner et al., 2017, Hosszejni et al., 2019, Nakakita et al., 24 Jul 2025).