A computationally-tractable measure of global sensitivity for sampling-based Bayesian inference

Arina Odnoblyudova; Charita Dellaporta; Fran\c{c}ois-Xavier Briol

arxiv: 2605.28099 · v1 · pith:UUJECKKJnew · submitted 2026-05-27 · 📊 stat.ME · stat.CO

A computationally-tractable measure of global sensitivity for sampling-based Bayesian inference

Arina Odnoblyudova , Charita Dellaporta , Fran\c{c}ois-Xavier Briol This is my paper

Pith reviewed 2026-06-29 11:06 UTC · model grok-4.3

classification 📊 stat.ME stat.CO

keywords global sensitivity analysisFisher divergenceBayesian posteriorsampling-based inferencescore functionsgeneralised Bayesian inferencesimulation-based inference

0 comments

The pith

Fisher divergence measure quantifies global sensitivity of Bayesian posteriors from samples alone.

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

The paper develops a global sensitivity analysis tool for Bayesian models that relies on the Fisher divergence. It takes a reference posterior sample set plus the ability to evaluate score functions and produces a quantity that bounds how much the full posterior distribution can shift under perturbations to the prior or likelihood. The approach targets settings where existing sensitivity methods become impractical due to cost or dimensionality. A reader would care because it opens sensitivity checks for modern workflows such as unnormalised models, time-series inference, and simulation-based neural models.

Core claim

Under mild regularity conditions the Fisher divergence between a reference posterior and a perturbed posterior controls the total variation distance on the entire posterior and supplies explicit bounds on the resulting shifts in the first two moments; the quantity is obtained from posterior samples and score evaluations without requiring additional sampling or optimisation.

What carries the argument

Fisher divergence between reference and perturbed posteriors, which serves as an upper bound on distributional and moment changes.

If this is right

The measure applies directly to generalised Bayesian inference with unnormalised models.
It remains tractable for Bayesian time-series models.
It supports sensitivity checks inside neural simulation-based inference pipelines.
Bounds on the first two posterior moments follow from the same divergence quantity.

Where Pith is reading between the lines

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

Low values would indicate that posterior inferences are stable to modest hyperparameter changes in the prior or likelihood.
The same sampling-plus-score workflow could be reused to compare robustness across several candidate models.
If score functions are cheap, the method could be inserted as a routine diagnostic step before committing to expensive downstream decisions.

Load-bearing premise

Mild regularity conditions must hold on the posteriors and score functions must be evaluable.

What would settle it

A concrete perturbation where the computed Fisher divergence is small yet the total variation distance between posteriors or the moment shifts exceed the claimed bounds.

Figures

Figures reproduced from arXiv: 2605.28099 by Arina Odnoblyudova, Charita Dellaporta, Fran\c{c}ois-Xavier Briol.

**Figure 1.** Figure 1: Synthetic Gaussian location model. Across all panels, symbol ★ denotes the optimal candidate identified via box-constrained optimisation that maximises ˆρ FD m (Π˜ λ ). (a) The FD as a function of the prior hyperparameters µ and σ. The symbol × marks the minimum. (b) The FD is non-convex in the µ, σ parametrisations. (c) In contrast, reparameterising the FD in terms of the natural parameters λ0 = µ/σ2 and … view at source ↗

**Figure 2.** Figure 2: Strength of measure for the multivariate Gaussian model. (a) Same mean, different covariances, (b) Different mean, same covariances. (c) A set of candidate multivariate Gaussian distributions. is the πref and is the optimal π λ . parametrisation. This example therefore illustrates the importance of working in natural exponential family parametrisations, and we will hence focus on this case in the remaind… view at source ↗

**Figure 3.** Figure 3: Finite-sample complexity (top row) and computational cost (bottom row) across dΘ for the multivariate Gaussian model. Results are averaged over 500 runs. The experiments for the Wasserstein distance and the KL divergence were too expensive to run up to m = 10000, so were run for only m = 5000 samples. numerous approaches having been proposed in the literature; see e.g Wu and Martin (2023). It is also well … view at source ↗

**Figure 4.** Figure 4: Ising model. ˆρ FD m (Π˜ λL ) across the neighbourhood Γ with ϵ = 0.05, × is λL,ref. is Matsubara et al. (2024), is Syring and Martin (2019), and is Lyddon et al. (2019). The curve for the approach of Lyddon et al. (2019) is not provided for KSD-Bayes since the approach requires a loss taking the form of a Monte Carlo estimator, whereas KSD-Bayes is a U-statistic loss. modest in all cases. In contrast, the… view at source ↗

**Figure 5.** Figure 5: Kilpisj¨arvi AR(5) model optimisation results. (a) is πref, is π λ inf m,t and , are the optimal π λ sup m , found among all candidate π λm (faint lines). (b) SbFD m (Γj ) contribution to SbFD m (Γ). (c) Optimisation runtime in log-scale (boxplot obtained from 500 repetitions) of our convex boxconstrained (•), decomposed box-constrained (•) and black-box (BBO) (•) optimisation methods. with λ := λπ = [λα,… view at source ↗

**Figure 6.** Figure 6: Kilpisj¨arvi model. × are true observations. Mean and 95% credible regions. by evaluating the quadratic-form objective in all 2dΛ = 214 corner points of the box constraints, and finally a method which uses the decomposition of the sensitivity measure as a sum of quadratic forms to consider only PdΘ j=1 2 dΛj = 28 corner configurations. The black-box dual annealing approach runs for substantially longer tha… view at source ↗

**Figure 7.** Figure 7: Prior sensitivity analysis for the Turin model with introduced prior dependence. ⋆ denotes the worst-case candidate identified via (a) FD-based optimisation and (b) informal sensitivity analysis. × marks the evaluated candidates. The y-axis is in log-scale. parameters θi , θj that we suspect of being correlated, we consider candidate priors of the form π(θ | λ) = πref(θ | λprior)cλc (θi , θj ), where λ := … view at source ↗

read the original abstract

Bayesian inference can often be sensitive to the choice of hyperparameters of the prior or likelihood, yet defining and quantifying this sensitivity in a principled and computationally feasible way remains challenging in practice. Unfortunately, existing sensitivity methods are rarely applicable in modern Bayesian workflows due to their high computational cost and poor performance in moderate to high dimensions. To address these limitations, we introduce a new approach to global sensitivity analysis based on the Fisher divergence. Our method only requires a set of samples from a reference posterior and the ability to evaluate score functions, making it broadly computationally tractable. Under mild regularity conditions, it controls changes in the whole posterior, and provides a bound on the impact of perturbations on the first two moments. We demonstrate these strengths on challenging Bayesian inference problems which are practically out of reach of existing approaches, including generalised Bayesian inference for unnormalised models, inference in Bayesian models of time series, and neural simulation-based inference.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a Fisher-divergence sensitivity measure that runs from posterior samples plus score evaluations and claims bounds on the full posterior plus first two moments.

read the letter

The core idea is to use Fisher divergence between a reference posterior and a perturbed one as a global sensitivity diagnostic. It only needs samples from the reference and the ability to evaluate the score, which sidesteps the usual cost of re-running MCMC or variational inference for every hyperparameter change. That setup directly targets the cases the abstract flags: generalized Bayes with unnormalized densities, time-series models, and neural simulation-based inference.

What stands out is the claim that, under mild regularity conditions, the divergence controls changes across the entire posterior and supplies explicit bounds on shifts in the mean and variance. If those bounds hold with reasonable constants, the method could slot into existing sampling workflows without much extra machinery. The demonstrations on problems that are currently out of reach for other global sensitivity tools are the practical test.

The main uncertainty is how restrictive the regularity conditions turn out to be once you look at the proofs and whether the bounds remain informative rather than vacuous on realistic posteriors. The abstract does not spell out the derivation steps or any error analysis, so the tightness of the moment bounds and the behavior near the boundary of the conditions are still open. No circularity is visible; the measure is built from the standard Fisher divergence without reducing to fitted parameters.

This is aimed at statisticians and machine-learning researchers who already produce posterior samples and need a cheap way to audit sensitivity to prior or likelihood choices. It is worth sending to a serious referee because the computational tractability claim addresses a documented gap and the information-theoretic grounding looks coherent on its face.

Referee Report

1 major / 0 minor

Summary. The paper introduces a Fisher divergence-based global sensitivity measure for sampling-based Bayesian inference. It requires only reference posterior samples and score function evaluations, making it tractable in moderate-to-high dimensions. Under mild regularity conditions the measure is claimed to control changes to the full posterior and to bound perturbations to the first two moments; demonstrations are provided for generalised Bayesian inference with unnormalised models, time-series models, and neural simulation-based inference.

Significance. If the stated bounds hold, the contribution would be practically useful: it supplies a computationally feasible sensitivity diagnostic that existing methods cannot scale to the regimes the authors target. The grounding in the standard Fisher divergence (rather than ad-hoc or fitted quantities) and the explicit focus on sampling-based workflows are strengths. Reproducible demonstrations on otherwise intractable problems would further strengthen the case.

major comments (1)

[Abstract / theoretical section] Abstract and theoretical development: the central claim that the Fisher divergence 'controls changes in the whole posterior' and 'provides a bound on the impact of perturbations on the first two moments' is stated without derivation details, explicit regularity conditions, or error analysis. A load-bearing section (presumably the main theoretical result) must supply the precise statement of the conditions, the proof strategy, and any counter-example checks before the claim can be verified.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their detailed review and for highlighting the need for greater transparency in the theoretical claims. The single major comment concerns the presentation of the central results on posterior control and moment bounds. We address it directly below.

read point-by-point responses

Referee: [Abstract / theoretical section] Abstract and theoretical development: the central claim that the Fisher divergence 'controls changes in the whole posterior' and 'provides a bound on the impact of perturbations on the first two moments' is stated without derivation details, explicit regularity conditions, or error analysis. A load-bearing section (presumably the main theoretical result) must supply the precise statement of the conditions, the proof strategy, and any counter-example checks before the claim can be verified.

Authors: We agree that the current manuscript states these claims at a high level without supplying the full set of regularity conditions, the complete proof outline, or an accompanying error analysis in a single load-bearing section. In the revised version we will add an expanded theoretical subsection (new Section 3.2) that (i) lists the precise regularity conditions (twice-differentiability of the log-density, integrability of the score, and a uniform bound on the Hessian), (ii) sketches the proof strategy via the data-processing inequality for Fisher divergence followed by a Taylor expansion of the perturbed posterior, and (iii) includes a short discussion of the resulting error terms together with a simple counter-example check confirming that the bounds fail when the regularity conditions are violated. These additions will make the claims directly verifiable without altering the main results. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation grounded in standard Fisher divergence

full rationale

The paper defines a global sensitivity measure directly from the Fisher divergence (a standard information-theoretic quantity) and derives its properties (posterior control and moment bounds) under stated mild regularity conditions using samples and score evaluations. No load-bearing step reduces by construction to a fitted parameter, self-citation chain, or renamed input; the central claims remain independent of the target result and are externally falsifiable via the divergence definition itself. This is the most common honest non-finding for papers that import a known divergence and apply it to sensitivity analysis.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the definition of Fisher divergence applied to posteriors and the invocation of mild regularity conditions for bounds; no free parameters or invented entities are mentioned.

axioms (1)

domain assumption Mild regularity conditions on the posteriors and score functions
Required for the control of posterior changes and moment bounds to hold, as stated in the abstract.

pith-pipeline@v0.9.1-grok · 5697 in / 1046 out tokens · 35884 ms · 2026-06-29T11:06:11.949349+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references

[1]

Then, cFDm(˜Πref∥˜Πλ)−FD( ˜Πref∥˜Πλ) = 1 m mX i=1 gtest(θi; ˜Πλ)

By construction,E ˜Πλ[gtest(θ; ˜Πλ)] = 0. Then, cFDm(˜Πref∥˜Πλ)−FD( ˜Πref∥˜Πλ) = 1 m mX i=1 gtest(θi; ˜Πλ). By Theorem 1 (forq= 1) and Remark 1 of Durmus et al. (2024), it follows that ifg test ∈L √ V , then under Assumption 2 the following moment inequality holds E   1 m mX i=1 gtest(θi; ˜Πλ) !2  ≤ C( ˜Πλ) m ,(25) for some C( ˜Πλ)≤K ′ gtest(·; ˜Πλ) 2...

2024
[2]

Since ˜ΠΛ ∈ PΓ we have: ∆(θ; ˜Πλ) = ∇θT(θ) ⊤(λπref −λ π) + (λL −λ Lref )∇ θl(θ;x 1:n) 2 2 . Using the fact that (a+b) 2 ≤2(a 2 +b 2) and the operator norm inequality,∥Ax∥ 2 ≤ ∥A∥ 2∥x∥2, we obtain: ∆(θ; ˜Πλ)≤2∥∇ θT(θ) ⊤(λπref −λ π)∥2 2 + 2 (λL −λ Lref )2 ∥∇θl(θ;x 1:n)∥2 2 ≤2∥λ πref −λ π∥2 2 ∥∇θT(θ)∥ 2 2 + 2 (λL −λ Lref )2 ∥∇θl(θ;x 1:n)∥2 2. Dividing by p V...

2018

[1] [1]

Then, cFDm(˜Πref∥˜Πλ)−FD( ˜Πref∥˜Πλ) = 1 m mX i=1 gtest(θi; ˜Πλ)

By construction,E ˜Πλ[gtest(θ; ˜Πλ)] = 0. Then, cFDm(˜Πref∥˜Πλ)−FD( ˜Πref∥˜Πλ) = 1 m mX i=1 gtest(θi; ˜Πλ). By Theorem 1 (forq= 1) and Remark 1 of Durmus et al. (2024), it follows that ifg test ∈L √ V , then under Assumption 2 the following moment inequality holds E   1 m mX i=1 gtest(θi; ˜Πλ) !2  ≤ C( ˜Πλ) m ,(25) for some C( ˜Πλ)≤K ′ gtest(·; ˜Πλ) 2...

2024

[2] [2]

Since ˜ΠΛ ∈ PΓ we have: ∆(θ; ˜Πλ) = ∇θT(θ) ⊤(λπref −λ π) + (λL −λ Lref )∇ θl(θ;x 1:n) 2 2 . Using the fact that (a+b) 2 ≤2(a 2 +b 2) and the operator norm inequality,∥Ax∥ 2 ≤ ∥A∥ 2∥x∥2, we obtain: ∆(θ; ˜Πλ)≤2∥∇ θT(θ) ⊤(λπref −λ π)∥2 2 + 2 (λL −λ Lref )2 ∥∇θl(θ;x 1:n)∥2 2 ≤2∥λ πref −λ π∥2 2 ∥∇θT(θ)∥ 2 2 + 2 (λL −λ Lref )2 ∥∇θl(θ;x 1:n)∥2 2. Dividing by p V...

2018