Pith. sign in

REVIEW 4 major objections 7 minor 238 references

KNorm-RVB, a two-step reparametrization that normalizes local variables and then reflects away their skewness, makes mean-field variational inference accurate enough to beat current approximations on generalized linear mixed models, mixed m

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 20:11 UTC pith:MEOJCHRG

load-bearing objection KNorm-RVB is a real extension of Tan's RVB with a clean skew-symmetric construction and careful change-of-variable theory; the main caveat is an unquantified Gaussian approximation at the normalization step, which is addressable rather than fatal. the 4 major comments →

arxiv 2607.16698 v1 pith:MEOJCHRG submitted 2026-07-18 stat.ME

Generalized reparametrized variational Bayes with skew-symmetric normalization

Aoxiang Chen , Linda S. L. Tan This is my paper
classification stat.ME MSC 62F1562C1060E05
keywords variational inferencemean-field approximationreparametrizationskew-symmetric densitylatent Gaussian modelslatent non-Gaussian modelshierarchical modelssparse precision matrices
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Mean-field variational inference is cheap but notoriously bad at capturing dependence between local random effects and global parameters; this paper claims the failure is largely residual skewness, not just correlation, and that skewness can be removed before the approximation. The proposed KNorm-RVB reparametrizes each local conditional posterior in two steps: a curvature-based affine map to an approximately standard Gaussian, then a stochastic reflection that symmetrizes the density about an optimally chosen center. The paper proves the reflected variable is even-symmetric about that center, which makes its posterior mean exactly recoverable by mean-field VI and forces local-global posterior covariance to zero, and shows the KL gap to a standard Gaussian shrinks as the reflection group grows. On generalized linear mixed models, mixed multinomial logit models, spatial autoregressive models, and stochastic volatility models, KNorm-RVB reports uniformly better posterior approximation accuracy than existing variational and Laplace-based methods. A sympathetic reader would take away a scalable default for hierarchical models whose local conditionals are mildly non-Gaussian, with a tunable accuracy/compute knob.

Core claim

KNorm-RVB's central claim is that normalizing alone is insufficient: after the standard RVB affine transformation, transformed local variables can remain skewed, and that skewness is what keeps mean-field VI off target. The paper constructs a K-component skew-symmetric density representation in which any conditional posterior can be averaged over a group of blockwise reflections about a point c, giving a symmetrized density that is closer to Gaussian; drawing from it is as simple as applying a uniformly chosen reflection to the normalized draw. The paper proves that this symmetrization strictly reduces the KL divergence to N(c,I), that the symmetrized conditional posterior is even-symmetric

What carries the argument

The K-component skew-symmetric density representation and its stochastic sampler. Any density g on R^N can be written uniquely as K f w, where f is the average of g over a finite abelian group of blockwise reflections (coordinate signs flipped blockwise about a point c) and w is a skewing function that sums to one over the group. Procedure 1 exploits the reverse direction: after normalizing b to tilde b via an affine map built from the Cholesky factor of the sparse conditional precision, KNorm-RVB samples a group element uniformly and sets b* = c + S(tilde b - c), producing a draw from the symmetrized, less skewed density. Theorem 2 shows this step strictly decreases KL divergence to N(c,I),

Load-bearing premise

The whole construction hangs on the second-order Taylor expansion of the local log-likelihood about the conditional posterior mode being a good Gaussian approximation to p(b|θ_g,y); if the true conditional is multimodal, highly skewed at the mode, or the mode is poorly estimated, the normalization and reflection steps may not produce the advertised accuracy gains, and the paper does not quantify this error.

What would settle it

Take a synthetic hierarchical model whose local conditional posterior is deliberately bimodal or strongly skewed away from a Gaussian fit at the mode (e.g., a logistic random-intercept model with very few observations per subject and strong separation), run KNorm-RVB with K=2 and K=2^n, and compare MMD/ELBO against MCMC and plain R-VB. If KNorm-RVB's accuracy margins collapse or become negative when the Taylor-mode Gaussian is a visibly bad fit, the claim is conditional on near-Gaussian local conditionals; if margins persist, the robustness claim holds.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • On factorized local conditionals (e.g., GLMMs and MMNL models), the cost of the full K=2^n reflection group grows linearly in n rather than exponentially, so the strongest Gaussianization is affordable exactly where it helps most.
  • Larger blockwise reflection groups monotonically improve the KL fit of the transformed local posterior to a standard Gaussian, giving a principled accuracy/compute dial (K=2, 16, 256 in the paper's SAR and SV experiments).
  • When the stated concavity and symmetry conditions hold, the local variational block is provably centered at the true posterior mean and reproduces the true correlation structure up to scale, so residual error is confined to scale and tail behavior.
  • Adding a closed skew normal variational family for global and mixing variables captures skewness that remains after local symmetrization, which is why the CSN variants post the highest MMD scores in the reported comparisons.
  • The same two-step recipe covers latent non-Gaussian models (heavy-tailed driving noise) by conditioning on mixing variables, so one method spans both Gaussian and heavy-tailed latent fields.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The symmetrization acts on the target density, not on the variational family, so it should transport to richer approximators (normalizing flows, importance-weighted VI, full-rank Gaussians) unchanged; the paper only demonstrates Gaussian local families plus CSN globals.
  • The exact mean-recovery theorems assume even-symmetry that KNorm-RVB enforces by construction, but the correlation-recovery theorem assumes elliptical symmetry that it does not enforce; in practice the observed gains may owe as much to a smoother optimization landscape (isotropic, centered ELBO) as to the symmetry mechanism. A testable check: measure ELBO gradient variance after reflection.
  • An adaptive partition choice—splitting coordinates into reflection blocks according to estimated marginal skewness—could replace the manual model-structure-based partition and scale KNorm-RVB to larger latent fields; the paper lists this as future work.
  • The K-component representation itself is a standalone tool: any density with a known mode and sparse dependence structure gets a closed-form, tractable symmetric surrogate with a rejection-free sampler, which could be reused outside variational inference.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes KNorm-RVB, a reparametrized variational Bayes framework for hierarchical models with sparse local precision structure. The method generalizes Tan's RVB in three main directions: (i) joint Cholesky-based normalization of the local conditional posterior, accommodating conditionally dependent local variables (spatial, temporal, subject-level); (ii) a skewness-reduction step that symmetrizes the normalized local variables about an optimized reflection point using a K-component skew-symmetric representation, with exact change-of-variable formulas for the reparametrized log joint density (Theorem 2); and (iii) a closed skew normal (CSN) variational family for global and mixing variables. The paper develops the K-component skew-symmetric construction and representation (Propositions 1–2), stochastic samplers (Proposition 3), an optimal skewing function (Theorem 1), symmetry results showing the transformed conditional posterior is even-symmetric with mean c and zero cross-covariance with global variables (Theorem 3), and conditions under which MFVI recovers the transformed local mean and correlation matrix exactly (Theorems 4–5). The method is evaluated against MCMC, INLA, SVI, GLOSS-VA, and RVB on GLMM, MMNL, SAR, and stochastic volatility models, with improvements reported in MMD-based accuracy, marginal diagnostics, and ELBO.

Significance. KNorm-RVB is a potentially valuable addition to the variational inference toolbox. Its change-of-variable derivations are exact for the chosen transformation, Theorem 2(iv) provides a genuine monotone Gaussianization guarantee as the reflection group grows, and the extension to latent non-Gaussian models via NIG noise broadens the method's applicability. The empirical evaluation is broad and mostly consistent with the paper's claims. The strength of the results, however, is diminished by (i) the unquantified reliance on a local Gaussian approximation to the conditional posterior; (ii) exact-recovery theorems whose assumptions are unverified and whose conclusion concerns the transformed variables; and (iii) an empirical protocol that omits MCMC benchmarks and standard errors in several settings. With clarification and additional diagnostics, the contribution could be made solid; as written, the claims in the abstract are stronger than the evidence and theory can support.

major comments (4)
  1. [Section 3.1, Eq. (3)] The central claim of improved accuracy rests on the unquantified Gaussian approximation p(b|θ_g,y) ≈ N(b̂,(Q+H(b̂))^{-1}) from the second-order expansion. For binary, count, and heavy-tailed likelihoods the conditional local posterior can be skewed or heavy-tailed; then b̃ = L^T(b−b̂) is not approximately N(0,I) and retains dependence on θ_g. Theorem 2(iv) only guarantees that p(b*|θ_g,y) is closer to N(c,I) than p(b̃|θ_g,y) in KLD, not that the MFVI approximation to the original posterior is accurate. The paper gives no bound, diagnostic, or simulation study of this expansion error for any of the four model classes. Please add a quantitative assessment of the normalization quality (e.g., a remainder bound for log-concave conditionals or an empirical Gaussianity diagnostic per model class).
  2. [Section 6, Theorems 3–5] The exact-recovery claims are partly built into the construction. Procedure 1 symmetrizes b* about c, so E(b*|θ_g,y)=c by design, and Algorithm 1 sets c to the variational mean. Theorem 4 then establishes a stationary point at ν=c only under concavity conditions that are not verified for the applications and are not inherited from p(b|θ_g,y) (the symmetrized density is a sum of reflected densities and need not be log-concave). Theorem 5 assumes elliptical symmetry a.e., which the paper admits is not guaranteed. Recovery is for the transformed b*, conditional on unverified conditions, so the abstract's 'recovers the local posterior mean and correlation matrix exactly' overstates the scope. Please verify the conditions or reframe the theorems as motivational symmetry analysis.
  3. [Section 7, M* protocol and Table 1] The MMD comparison between 'variational approximation and MCMC' does not state the parameterization in which the draws are compared. Since RVB and KNorm-RVB use different transformations, comparing in their respective reparametrized spaces would conflate reparametrization effects with approximation quality; the draws presumably need to be mapped to a common space (e.g., via Procedure S1), and this should be stated. The RBF kernel bandwidth is also not reported. In addition, Table 1 reports ELBO differences of about 1.4 nats without standard errors; as Monte Carlo estimates, these differences may not be significant. Please add standard errors and an MCMC-based accuracy measure for at least one configuration of the MMNL latent non-Gaussian model.
  4. [Section 7.2, latent non-Gaussian MMNL] The abstract's general claim of improved posterior accuracy for mixed multinomial logit models is supported for the latent Gaussian case by Figure 5, but for the latent non-Gaussian case the only evidence is the unquantified ELBO in Table 1; no MCMC benchmark is reported. The paper acknowledges this, but the general claim in the abstract is not supported for this setting. Similarly, for the latent non-Gaussian SAR and stochastic volatility models only marginal diagnostics are reported, not the multivariate M* measure used elsewhere. Please either supply the missing benchmarks or temper the general claim to match the metrics available for each model class.
minor comments (7)
  1. [Supplement S3.3] 'The strict concavity property of f_{λ_g}(J)' should read 'strict convexity'; log|J| is concave, so −log|J| is convex.
  2. [Supplement S3.3] The statement 'the integral over A^c is equal to zero' is justified only by the a.e.-w.r.t.-q_{λ_g} assumption in Theorem 5; this should be stated at the point of use.
  3. [Section 4.2, Eq. (7)] The notation p_{b,y|θ_g} is nonstandard; please define it explicitly as the conditional joint density of (b,y) given θ_g.
  4. [Figures 3–8] Boxplots of M* are not accompanied by numeric values; given the modest differences between some methods, report medians and IQRs in the text or tables.
  5. [Section 3.2] Specify the inverse Gaussian parametrization (mean and variance) so the mean-0, variance-1 calibration of the NIG noise is checkable.
  6. [Section 3.1, Algorithm 1] The per-iteration cost of evaluating the reparametrized density requires recomputing the conditional mode b̂(θ_g) and Cholesky factor L(θ_g) for each sampled θ_g. The paper does not state whether these are recomputed every iteration or cached/approximated, nor how this interacts with the stochastic gradient estimator. A brief implementation note would be valuable for reproducibility and scalability.
  7. [General] No code or data links are provided; for a methods paper with this many implementation details (reflection groups, CSN subclass, Procedure S1), a reproducibility artifact would substantially increase confidence.

Circularity Check

1 steps flagged

Exact mean-recovery theorem is built into the symmetrization and the choice c=μ; empirical MMD benchmarks remain external.

specific steps
  1. self definitional [Section 5 (Algorithm 1) and Section 6 (Theorems 3–4)]
    "The reflection point c is set to the mean (µ⊤1,...,µ⊤n)⊤ of the Gaussian variational approximation for b∗, so c is updated jointly with the other variational parameters in Algorithm 1."

    Theorem 4 states that KL{q_{ν,λ}(θ*)||p(θ*|y)} has a stationary point at ν=c, and Theorem 3 concludes E(b*|θg,y)=c from TGm,c-invariance. But that invariance is manufactured by Procedure 1 (b* is a uniformly random reflection of ~b about c), and c is defined as the variational mean μ. Thus 'exact recovery of the local posterior mean' is true by construction: the symmetrized variable's conditional mean is the reflection point, and the reflection point is the optimized variational mean. The theorem does not independently validate approximation of the original local posterior mean; it certifies the construction's own centering.

full rationale

The paper's practical accuracy claims are grounded in MMD and ELBO comparisons against MCMC, which are external benchmarks and not circular. The main circularity is confined to the theoretical 'exact recovery' motivation in Section 6: the symmetry that drives Theorem 3 is imposed by Procedure 1's uniform reflection, and the reflection point c is set to the variational mean μ. Consequently, Theorem 4's stationary point at ν=c reduces to 'the optimal variational location equals the location we already optimized.' The abstract's claim that 'under symmetry conditions, MFVI recovers the local posterior mean and correlation matrix exactly' is therefore partly self-definitional for the mean; the correlation-recovery theorem is explicitly acknowledged by the authors as requiring elliptical symmetry that KNorm-RVB does not guarantee. The normalization hinge in Section 3.1 is an unquantified Gaussian approximation and a correctness risk, not a circular step. Citations to Tan (2021) and Tan & Chen (2025) are building blocks rather than load-bearing self-citation chains. Overall the empirical contribution remains independently supported, but the central theoretical justification contains a by-construction reduction, warranting a 6.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The ledger lists the hand-set/optimization-dependent quantities and unproved background assumptions the central claims are built on; no invented physical entities are introduced.

free parameters (3)
  • Reflection point c = c = (μ_1,...,μ_n)^T (current variational means), Algorithm 1
    Defines the center about which the normalized local variables are symmetrized. Theorems 3–4 guarantee the transformed posterior mean equals c, so the 'exact mean recovery' statement is tied to an optimized variational quantity rather than a pre-specified model constant.
  • Blockwise reflection group G_m (m and block partition) = m=n for factorized conditionals (K=2^n); m in {1,...,16} for SAR/SV; block partition chosen by model structure
    Controls the amount of skewness reduction versus computational cost. The paper says the partition is selected based on model structure and empirical performance; no adaptive rule is provided.
  • NIG tail parameter prior rate α_η = 1 (Exp(1) prior on η)
    Set to 1 following Cabral et al. (2023) to mitigate overfitting; it is a hand-set prior that influences the latent non-Gaussian extension, though not the main framework.
axioms (5)
  • domain assumption Joint density factorizes as p(y,θ)=p(θ_g)p(b|θ_g)∏ p(y_i|b_i,θ_g) with b|θ_g Gaussian or NIG-driven and sparse precision Q.
    Defines the model class (Eq. 1, Sections 3.1–3.2).
  • ad hoc to paper log p(y_i|b_i,θ_g) is twice differentiable and the second-order Taylor expansion about the conditional posterior mode b̂ gives a Gaussian approximation with precision Q+H(b̂).
    Eq. (3) in Section 3.1; the whole normalization rests on this local quadratic approximation.
  • standard math The blockwise reflection group G_m is a finite abelian group whose elements have determinant ±1, and probability densities transform by unit Jacobian.
    Used in Propositions 1–3 and Theorem 2; standard group theory (Gallian 2025).
  • ad hoc to paper For Theorems 4–5, log p(b*|θ_g,y) is concave a.e. w.r.t. q_{λ_g} and strictly concave on a nonempty set, and p(b*|θ_g,y) is elliptically symmetric about c with a fixed scale matrix a.e.; these conditions are not verified for the applications.
    Section 6; the paper explicitly notes KNorm-RVB does not guarantee the elliptical symmetry needed for correlation recovery.
  • domain assumption The variational family q(b*) is a location-scale family with spherically symmetric base density (Gaussian in practice); q(θ_g) is a CSN subclass.
    Defines the optimization class; Theorems 4–5 depend on even-symmetry of the base density.
invented entities (1)
  • Blockwise reflection group G_m and auxiliary reflection index S* no independent evidence
    purpose: Stochastic symmetrization of normalized local variables to reduce skewness before mean-field approximation.
    A mathematical construction (not a physical entity). It generalizes Wang et al. (2004)'s two-point reflection to 2^m reflections; it has no falsifiable handle outside the paper itself, but it is not a hidden physical assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 32174 in / 18711 out tokens · 170028 ms · 2026-08-01T20:11:46.760619+00:00 · methodology

0 comments
read the original abstract

Bayesian hierarchical models with high-dimensional latent structure require scalable posterior approximations that preserve key dependencies while remaining computationally tractable. Mean-field variational inference (MFVI) is efficient, but can be unreliable when local variables are strongly correlated or tightly coupled to global variables. We propose KNorm-RVB, a generalized reparametrized variational Bayes framework for latent Gaussian and latent non-Gaussian models with sparse local precision matrices. KNorm-RVB maps the conditional posterior of local variables toward a standard Gaussian via normalization followed by skewness reduction, enabled by a novel K-component skew-symmetric density representation. This reparametrization centers the transformed conditional local posterior at an optimized reflection point and decorrelates local and global variables, making MFVI much more effective. Under symmetry conditions, we show that MFVI recovers the local posterior mean and correlation matrix exactly, motivating KNorm-RVB's normalization and symmetrization of the conditional local posterior before applying MFVI. We combine a Gaussian variational family for reparametrized local variables with a flexible closed skew normal family for the remaining variables. Across generalized linear mixed models, mixed multinomial logit models, spatial autoregressive models, and stochastic volatility models, KNorm-RVB improves posterior approximation accuracy over existing methods.

Figures

Figures reproduced from arXiv: 2607.16698 by Aoxiang Chen, Linda S. L. Tan.

Figure 1
Figure 1. Figure 1: (left) depicts θ and its point-reflection 2ˆθ − θ for d = 2, along with two points obtained by reflecting θ about ˆθ in one coordinate while holding the other fixed. Averaging over these four reflections gives the four-component symmetrized density, f4,θˆ(θ) = 1 4    g   θ1 θ2   + g   2 ˆθ1 − θ1 θ2   + g   θ1 2 ˆθ2 − θ2   + g   2 ˆθ1 − θ1 2 ˆθ2 − θ2      . To illustrate th… view at source ↗
Figure 2
Figure 2. Figure 2: Bernoulli GLMM. (a) Conditional posterior p(b1 | θg, y) and its Gaussian and skew￾symmetric approximations. (b) Q-Q plot of draws of ˜b1 and b ∗ 1 . A closely related approach is GLOSS-VA (Kock et al. 2026) for latent Gaussian mod￾els with conditionally independent local variables. It builds on a conditionally structured Gaussian variational family, q(θg) Q i p(bi | θg) (Tan et al. 2020), and introduces sk… view at source ↗
Figure 3
Figure 3. Figure 3: Polypharmacy latent Gaussian model. Marginal posteriors for selected global [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Polypharmacy latent non-Gaussian model. Boxplots of absolute mean difference, [PITH_FULL_IMAGE:figures/full_fig_p026_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Latent Gaussian model for electricity data. Marginal posteriors for selected global variables, boxplot for M∗ . constrained optimization, write B = LBL ⊤ B and define ζ = vech(L ∗ B), where L ∗ B,ii = log LB,ii and L ∗ B,ij = LB,ij for i ̸= j. The global variables are θg = (β ⊤, ζ⊤, log a1, . . . , log ar) ⊤. We analyze the Electricity data from the mlogit R package (see e.g., Tan 2017), com￾prising n = 36… view at source ↗
Figure 6
Figure 6. Figure 6: SIDS latent Gaussian model: Boxplots of M∗ and marginal posteriors for b30, β0 and β1. independent NIG as in Section 3.2 [PITH_FULL_IMAGE:figures/full_fig_p029_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: SIDS latent non-Gaussian model: Marginal distributions of the mixing variable log(v1), together with boxplots of the absolute mean differences, standard deviation ratios for log(v). 29 [PITH_FULL_IMAGE:figures/full_fig_p029_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: GBP latent Gaussian model. Marginal posteriors for [PITH_FULL_IMAGE:figures/full_fig_p031_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: GBP latent non-Gaussian model: Marginal distributions of log( [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

238 extracted references · 6 canonical work pages

  1. [1]

    Advances in Neural Information Processing Systems 34 (NeurIPS 2021) , pages =

    Amortized Variational Inference for Simple Hierarchical Models , author =. Advances in Neural Information Processing Systems 34 (NeurIPS 2021) , pages =

    2021

  2. [2]

    arXiv preprint arXiv:1511.07367 , year=

    Black box variational inference for state space models , author=. arXiv preprint arXiv:1511.07367 , year=

    Pith/arXiv arXiv

  3. [3]

    Neural Computation , pages =

    Amari, Shunichi , title =. Neural Computation , pages =. 1998 , volume =

    1998

  4. [4]

    2016 , publisher=

    Information Geometry and Its Applications , author=. 2016 , publisher=

    2016

  5. [5]

    Bayesian Conjugacy in Probit, Tobit, Multinomial Probit and Extensions: A Review and New Results , journal =

    Niccol\`. Bayesian Conjugacy in Probit, Tobit, Multinomial Probit and Extensions: A Review and New Results , journal =. 2023 , doi =

    2023

  6. [6]

    , Title =

    Attias, H. , Title =. Proceedings of the 15th Conference on Uncertainty in Artificial Intelligence , pages=. 1999 , address=

    1999

  7. [7]

    Arellano-Valle and Adelchi Azzalini , journal =

    Reinaldo B. Arellano-Valle and Adelchi Azzalini , journal =. On the Unification of Families of Skew-Normal Distributions , volume =

  8. [8]

    2008 , doi =

    The centred parametrization for the multivariate skew-normal distribution , journal =. 2008 , doi =

    2008

  9. [9]

    Azzalini , journal =

    A. Azzalini , journal =. A Class of Distributions Which Includes the Normal Ones , volume =

  10. [10]

    Azzalini and A

    A. Azzalini and A. Capitanio , journal =. Statistical Applications of the Multivariate Skew Normal Distribution , volume =

  11. [11]

    Proceedings of the 35th International Conference on Machine Learning , pages =

    Dissecting Adam: The Sign, Magnitude and Variance of Stochastic Gradients , author =. Proceedings of the 35th International Conference on Machine Learning , pages =. 2018 , editor =

    2018

  12. [12]

    A comparison of centring parameterisations of

    Bass, Mark R and Sahu, Sujit K , journal=. A comparison of centring parameterisations of. 2017 , publisher=

    2017

  13. [13]

    Journal of Computational and Graphical Statistics , volume=

    Dynamically updated spatially varying parameterizations of hierarchical Bayesian models for spatial data , author=. Journal of Computational and Graphical Statistics , volume=. 2019 , publisher=

    2019

  14. [14]

    , title=

    Baker, Stuart G. , title=. Journal of the Royal Statistical Society: Series D (The Statistician) , volume=. 1994 , doi=

    1994

  15. [15]

    Journal of political economy , volume=

    The pricing of options and corporate liabilities , author=. Journal of political economy , volume=. 1973 , publisher=

    1973

  16. [16]

    Spatial and spatio-temporal models with

    Blangiardo, Marta and Cameletti, Michela and Baio, Gianluca and Rue, H. Spatial and spatio-temporal models with. Spatial and spatio-temporal epidemiology , volume=. 2013 , publisher=

    2013

  17. [17]

    Blei and Alp Kucukelbir and Jon D

    David M. Blei and Alp Kucukelbir and Jon D. McAuliffe , title =. Journal of the American Statistical Association , volume =. 2017 , doi =

    2017

  18. [18]

    Booth, J. G. and Hobert, J. P. , title =. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume =

  19. [19]

    2023 , eprint=

    Sample Average Approximation for Black-Box VI , author=. 2023 , eprint=

    2023

  20. [20]

    and Girolami, Mark and Mackey, Lester , booktitle =

    Barp, Alessandro and Briol, François-Xavier and Duncan, Andrew B. and Girolami, Mark and Mackey, Lester , booktitle =. Minimum. 2019 , publisher =

    2019

  21. [21]

    Current trends in Bayesian methodology with applications , volume=

    Hamiltonian Monte Carlo for hierarchical models , author=. Current trends in Bayesian methodology with applications , volume=. 2015 , publisher=

    2015

  22. [22]

    Clinical Trials , volume=

    Bayesian hierarchical modeling of patient subpopulations: efficient designs of phase II oncology clinical trials , author=. Clinical Trials , volume=. 2013 , publisher=

    2013

  23. [23]

    Breslow, N. E. and D. G. Clayton , Title =. Journal of the American Statistical Association , Pages =

  24. [24]

    Journal of the American Statistical Association , volume =

    Michael Braun and Jon McAuliffe , title =. Journal of the American Statistical Association , volume =. 2010 , doi =

    2010

  25. [25]

    and Jordan, Michael I

    Broderick, Tamara and Boyd, Nicholas and Wibisono, Andre and Wilson, Ashia C. and Jordan, Michael I. , title =. Proceedings of the 26th International Conference on Neural Information Processing Systems - Volume 2 , series =. 2013 , location =

    2013

  26. [26]

    Browne and Fiona Steele and Mousa Golalizadeh and Martin J

    William J. Browne and Fiona Steele and Mousa Golalizadeh and Martin J. Green , journal =. The Use of Simple Reparameterizations to Improve the Efficiency of Markov Chain Monte Carlo Estimation for Multilevel Models with Applications to Discrete Time Survival Models , volume =

  27. [27]

    Controlling the flexibility of non-

    Cabral, Rafael and Bolin, David and Rue, H. Controlling the flexibility of non-. Bayesian Analysis , volume=. 2023 , publisher=

    2023

  28. [28]

    Fitting latent non-

    Cabral, Rafael and Bolin, David and Rue, H. Fitting latent non-. Journal of the American Statistical Association , volume=. 2024 , publisher=

    2024

  29. [29]

    Journal of Statistical Software, Articles , volume =

    Bob Carpenter and Andrew Gelman and Matthew Hoffman and Daniel Lee and Ben Goodrich and Michael Betancourt and Marcus Brubaker and Jiqiang Guo and Peter Li and Allen Riddell , title =. Journal of Statistical Software, Articles , volume =. 2017 , pages =

    2017

  30. [30]

    Proceedings of the 41st International Conference on Machine Learning , editor =

    Batch and Match: Black-Box Variational Inference with a Score-Based Divergence , author =. Proceedings of the 41st International Conference on Machine Learning , editor =. 2024 , publisher =

    2024

  31. [31]

    and Gower, Robert M

    Cai, Diana and Modi, Chirag and Margossian, Charles C. and Gower, Robert M. and Blei, David M. and Saul, Lawrence K. , booktitle =. 2024 , publisher =

    2024

  32. [32]

    Affine Independent Variational Inference , volume =

    Challis, Edward and Barber, David , booktitle =. Affine Independent Variational Inference , volume =

  33. [33]

    An Extended Simplified

    Cristian Chiuchiolo and Janet van Niekerk and Håvard Rue , year=. An Extended Simplified. 2203.14304 , archivePrefix=

    Pith/arXiv arXiv

  34. [34]

    Gauss--Newton hessian matrix , author=

    Hessian matrix vs. Gauss--Newton hessian matrix , author=. SIAM Journal on Numerical Analysis , volume=. 2011 , publisher=

    2011

  35. [35]

    Nott and Linda S

    Aoxiang Chen and David J. Nott and Linda S. L. Tan , year=. Weighted Fisher divergence for high-dimensional. 2503.04246 , archivePrefix=

    arXiv

  36. [36]

    Rotated Mean-Field Variational Inference and Iterative

    Yifan Chen and Sifan Liu , year=. Rotated Mean-Field Variational Inference and Iterative. 2510.07732 , archivePrefix=

    Pith/arXiv arXiv

  37. [37]

    Christensen and Gareth O

    Ole F. Christensen and Gareth O. Roberts and Martin Sköld , journal =. Robust Markov Chain Monte Carlo Methods for Spatial Generalized Linear Mixed Models , volume =

  38. [38]

    and Clark, James S

    Cressie, Noel and Calder, Catherine A. and Clark, James S. and Hoef, Jay M. Ver and Wikle, Christopher K. , title =. Ecological Applications , volume =. doi:10.1890/07-0744.1 , year =

    work page doi:10.1890/07-0744.1

  39. [39]

    2015 , publisher=

    Statistics for spatial data , author=. 2015 , publisher=

    2015

  40. [40]

    Crowder , journal =

    Martin J. Crowder , journal =. Beta-Binomial Anova for Proportions , volume =

  41. [41]

    Density estimation using Real

    Laurent Dinh and Jascha Sohl-Dickstein and Samy Bengio , editor =. Density estimation using Real. 5th International Conference on Learning Representations , year=

  42. [42]

    De Backer, M. and C. Twelve weeks of continuous oral therapy for toenail onychomycosis caused by dermatophytes: A double-blind comparative trial of terbinafine 250 mg/day versus itraconazole 200 mg/day , Journal =

  43. [43]

    A. P. Dempster and N. M. Laird and D. B. Rubin , title =. J. R. Statist.Soc. B , year =

  44. [44]

    J. L. Doob , title =. 1949 , pages =

    1949

  45. [45]

    Duchi, J. and E. Hazan and Y. Singer , Title =. Journal of Machine Learning Research , Pages =

  46. [46]

    Statistical Science , pages =

    Daniele Durante and Tommaso Rigon , title =. Statistical Science , pages =. 2019 , doi =

    2019

  47. [47]

    Daniele Durante and Francesco Pozza and Botond Szabo , year=. Skewed. 2301.03038 , archivePrefix=

    Pith/arXiv arXiv

  48. [48]

    Neural Spline Flows , volume =

    Durkan, Conor and Bekasov, Artur and Murray, Iain and Papamakarios, George , booktitle =. Neural Spline Flows , volume =

  49. [49]

    Probability:

    Durrett, Rick , edition =. Probability:. 2019 , publisher=

    2019

  50. [50]

    Journal of Computational and Graphical Statistics , volume =

    Shourya Dutta and Janet van Niekerk and HÃ¥vard Rue , title =. Journal of Computational and Graphical Statistics , volume =. 2026 , doi =

    2026

  51. [51]

    Proceedings of The 24th International Conference on Artificial Intelligence and Statistics , pages =

    Fisher Auto-Encoders , author =. Proceedings of The 24th International Conference on Artificial Intelligence and Statistics , pages =. 2021 , editor =

    2021

  52. [52]

    Faes and J

    C. Faes and J. T. Ormerod and M. P. Wand , title =. Journal of the American Statistical Association , volume =

  53. [53]

    Bias Reduction of Maximum Likelihood Estimates , volume =

    David Firth , journal =. Bias Reduction of Maximum Likelihood Estimates , volume =

  54. [54]

    Biometrika , volume =

    Fasano, Augusto and Durante, Daniele and Zanella, Giacomo , title = ". Biometrika , volume =. 2022 , doi =

    2022

  55. [55]

    Improving the

    Ferkingstad, Egil and Rue, H. Improving the

  56. [56]

    Biostatistics , volume =

    Fong, Youyi and Rue, Håvard and Wakefield, Jon , title = ". Biostatistics , volume =

  57. [57]

    2025 , publisher=

    Contemporary Abstract Algebra , author=. 2025 , publisher=. doi:10.1201/9781003488026 , isbn=

    work page doi:10.1201/9781003488026 2025

  58. [58]

    2020 , eprint=

    On the Promise of the Stochastic Generalized Gauss-Newton Method for Training DNNs , author=. 2020 , eprint=

    2020

  59. [59]

    Gart , journal =

    John J. Gart , journal =. Statistical Analyses of the Relative Risk , volume =

  60. [60]

    Covariances, robustness, and variational

    Giordano, Ryan and Broderick, Tamara and Jordan, Michael I , journal=. Covariances, robustness, and variational

  61. [61]

    Journal of Machine Learning Research , volume=

    Black box variational inference with a deterministic objective: Faster, more accurate, and even more black box , author=. Journal of Machine Learning Research , volume=

  62. [62]

    Gelfand and Sujit K

    Alan E. Gelfand and Sujit K. Sahu and Bradley P. Carlin , journal =. Efficient parametrisations for normal linear mixed models , volume =

  63. [63]

    Gelfand and Sujit K

    Alan E. Gelfand and Sujit K. Sahu and Bradley P. Carlin , editor =. Efficient parametrisations for generalized linear mixed models , booktitle =. 1996 , publisher =

    1996

  64. [64]

    Journal of the American statistical Association , volume=

    Gibbs sampling , author=. Journal of the American statistical Association , volume=. 2000 , publisher=

    2000

  65. [65]

    Data Analysis Using Regression and Multilevel/Hierarchical Models , DOI=

    Gelman, Andrew and Hill, Jennifer , year=. Data Analysis Using Regression and Multilevel/Hierarchical Models , DOI=

  66. [66]

    Genton and M.G

    Marc G. Genton and M.G. Loperfido , title =. Annals of the Institute of Statistical Mathematics , volume =

  67. [67]

    Genton and David E

    Marc G. Genton and David E. Keyes and George Turkiyyah , title =. Journal of Computational and Graphical Statistics , volume =

  68. [68]

    Proceedings of the 29th International Conference on Machine Learning , year=

    Samuel Gershman and Matt Hoffman and David Blei , Title =. Proceedings of the 29th International Conference on Machine Learning , year=

  69. [69]

    Geyer , journal =

    Charles J. Geyer , journal =. [Practical Markov Chain Monte Carlo]: Rejoinder , volume =

  70. [70]

    The Closed Skew-Normal Distribution , booktitle=

    Graciela Gonz\'. The Closed Skew-Normal Distribution , booktitle=. 2004 , editor=

    2004

  71. [71]

    Bayesian Analysis , volume=

    Fast and accurate estimation of non-nested binomial hierarchical models using variational inference , author=. Bayesian Analysis , volume=. 2022 , publisher=

    2022

  72. [72]

    Biometrika , volume=

    Partially factorized variational inference for high-dimensional mixed models , author=. Biometrika , volume=. 2025 , publisher=

    2025

  73. [73]

    Flexible Variational

    Gunawan, David and Kohn, Robert and Nott, David , journal=. Flexible Variational. 2023 , publisher=

    2023

  74. [74]

    The Review of Economic Studies , volume=

    Multivariate stochastic variance models , author=. The Review of Economic Studies , volume=. 1994 , publisher=

    1994

  75. [75]

    1970 , publisher=

    Monte Carlo sampling methods using Markov chains and their applications , author=. 1970 , publisher=

    1970

  76. [76]

    Skew-symmetric distributions and Fisher information — a tale of two densities , volume =

    Marc Hallin and Christophe Ley , journal =. Skew-symmetric distributions and Fisher information — a tale of two densities , volume =

  77. [77]

    Bernoulli , pages =

    Marc Hallin and Christophe Ley , title =. Bernoulli , pages =

  78. [78]

    Han, S. and X. Liao and D. B. Dunson and L. C. Carin , Title =. Proceedings of the 19th International Conference on Artificial Intelligence and Statistics , pages=. 2016 , editor=

    2016

  79. [79]

    BMC medical research methodology , volume=

    joineRML: a joint model and software package for time-to-event and multivariate longitudinal outcomes , author=. BMC medical research methodology , volume=. 2018 , publisher=

    2018

  80. [80]

    Journal of Machine Learning Research , volume=

    Stochastic variational inference , author=. Journal of Machine Learning Research , volume=

Showing first 80 references.