arxiv: 2605.13551 · v1 · submitted 2026-05-13 · 💻 cs.LG

Recognition: unknown

Mixed neural posterior estimation for simulators with discrete and continuous parameters

Jan Boelts , Cornelius Schr\"oder , Jonas Beck , Jakob H. Macke , Michael Deistler , Daniel Gedon

Authors on Pith no claims yet

Pith reviewed 2026-05-14 20:13 UTC · model grok-4.3

classification 💻 cs.LG

keywords neural posterior estimationsimulation-based inferencemixed parametersdiscrete and continuousposterior calibrationlikelihood-free inferencescientific simulatorsautoregressive classifier

0 comments

The pith

Neural posterior estimation extends to simulators with mixed discrete and continuous parameters through joint factorization and training.

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

This paper develops an extension of neural posterior estimation that handles simulators whose parameters mix discrete choices with continuous values. It factorizes the target posterior into an autoregressive classifier for the discrete dimensions and a generative model for the continuous dimensions. Both pieces are trained together on data from the simulator under one shared simulation-based objective. The resulting approximations are shown to be accurate and calibrated on toy cases where the true posterior is known and on several real scientific simulators. The extension matters because many models in science combine both parameter types, so fast likelihood-free inference now applies to a wider class of problems.

Core claim

By factorizing the joint posterior into discrete and continuous components, pairing an autoregressive classifier for the discrete parameters with a generative model for the continuous parameters, and optimizing the combined network under a single simulation-based objective, the method produces accurate and calibrated posterior approximations for both tractable toy examples and real-world scientific simulators.

What carries the argument

The inference network that factorizes the joint posterior into an autoregressive classifier for discrete parameters and a generative model for continuous parameters, trained jointly under a single simulation-based objective.

Load-bearing premise

The factorization of the joint posterior into discrete and continuous components combined with joint training under a single simulation-based objective will produce accurate and calibrated approximations without additional constraints or post-training adjustments.

What would settle it

On a simulator whose true posterior is known exactly, draw samples from the learned approximation and check whether they match the true distribution or pass a calibration diagnostic; systematic mismatch would refute the claim.

Figures

Figures reproduced from arXiv: 2605.13551 by Cornelius Schr\"oder, Daniel Gedon, Jakob H. Macke, Jan Boelts, Jonas Beck, Michael Deistler.

**Figure 1.** Figure 1: Inference method overview. MNPE is trained on a dataset of parameter–simulation pairs {(θ, x)} where θ is sampled from a mixed prior over discrete and continuous parameters. The inference network consists of a subnetwork for the discrete dimensions (orange), implemented as a masked autoregressive density estimator (MADE), and one for the continuous dimensions (violet), which can be a standard inference ne… view at source ↗

**Figure 2.** Figure 2: Tractable Gaussian benchmark. (a) MNPE posteriors for three observations spanning different regimes: xo = −0.5 (discrete state θd=0 dominates, unimodal), xo = 1.0 (uncertain θd, bimodal), and xo = 2.5 (θd=1 dominates). Histograms show the continuous marginal posterior p(θc | xo) from MNPE samples (violet), with the analytical ground-truth density overlaid (black). Inset bar charts compare the discrete post… view at source ↗

**Figure 3.** Figure 3: Tandem queueing simulator. (a) Posterior predictive simulations for arrival counts narr and queue length q1, comparing MNPE (green), MCMC (blue), and prior predictive (grey). Dashed line marks xo. (b) C2ST for joint (grey) and marginal comparisons (violet) against the MCMC reference across training samples. Error bars show standard error of the mean over five MNPE training repetitions. Dashed black line: i… view at source ↗

**Figure 4.** Figure 4: MNPE on a Hodgkin-Huxley simulator. (a) One observation (black) and two posterior predictive voltage traces (green) for a posterior trained on 100k samples. (b) MSE (mode ±std.) on 1k test samples. (c) Simulation-based calibration for a posterior trained on 100k samples, Left: for continuous parameters, Right: for two discrete parameters. (d) Simulation-based calibration across training samples. Mean error… view at source ↗

read the original abstract

Neural Posterior Estimation (NPE) enables rapid parameter inference for complex simulators with intractable likelihoods. NPE trains an inference network to estimate a probability density over parameters given data, typically assumed to be \emph{continuous}. However, many scientific models involve parameter spaces that are \emph{mixed}, that is, they contain both discrete and continuous dimensions. We address this limitation by extending NPE to mixed parameter spaces through an inference network that jointly handles discrete and continuous parameters. The inference network factorizes the joint posterior into discrete and continuous components, combining an autoregressive classifier for the discrete parameters with a generative model for the continuous parameters, trained jointly under a single simulation-based objective. In addition, we propose a diagnostic tool to assess the calibration of the mixed posterior approximation. Across tractable toy examples and real-world scientific simulators, our joint inference approach yields accurate and calibrated posteriors. The inference framework is available in the \texttt{sbi} Python package.

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

This extends NPE to mixed discrete-continuous parameters via a factorized posterior trained jointly, which works on their tests but leaves the calibration dynamics under strong coupling unexamined.

read the letter

The paper's core move is to split the posterior into an autoregressive classifier for discrete parameters and a conditional generative model for the continuous ones, then train both pieces together on simulations from the simulator. They also supply a post-training diagnostic for checking calibration of the combined approximation. This is packaged in the sbi library and demonstrated on both simple toy cases and a couple of real scientific simulators, where the posteriors come out accurate and calibrated according to their metrics.

Referee Report

1 major / 1 minor

Summary. The paper extends Neural Posterior Estimation (NPE) to simulators with mixed discrete and continuous parameters. It factorizes the joint posterior into an autoregressive classifier for the discrete parameters and a conditional generative model for the continuous parameters, with both components trained jointly under a single simulation-based objective. A diagnostic is proposed to assess calibration of the mixed posterior. The method is evaluated on tractable toy examples and real-world scientific simulators, claiming to yield accurate and calibrated posteriors, and is released in the sbi package.

Significance. If the joint training produces calibrated approximations for both discrete and continuous components without additional constraints or post-hoc adjustments, the work would be a useful practical extension of NPE to a common class of scientific models. The open-source implementation in sbi supports reproducibility. The significance is tempered by the lack of explicit analysis showing that the single-objective training does not allow continuous-component gradients to degrade discrete marginal calibration.

major comments (1)

[Methods (joint training objective)] The central claim that joint training under a single simulation-based objective yields calibrated posteriors for both components rests on the unexamined assumption that gradients from the continuous density estimator will not dominate or bias the autoregressive classifier for discrete parameters. No analysis, weighting scheme, alternating optimization, or marginal calibration term is described to guard against this coupling, particularly when discrete choices depend strongly on continuous values.

minor comments (1)

[Abstract] The abstract states that the diagnostic assesses calibration but does not specify the exact procedure or metrics (e.g., whether it checks marginal calibration for discrete parameters separately).

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive review and for identifying a key point regarding the joint training procedure in our extension of NPE to mixed discrete-continuous parameter spaces. We address the major comment below and describe the revisions we will implement.

read point-by-point responses

Referee: The central claim that joint training under a single simulation-based objective yields calibrated posteriors for both components rests on the unexamined assumption that gradients from the continuous density estimator will not dominate or bias the autoregressive classifier for discrete parameters. No analysis, weighting scheme, alternating optimization, or marginal calibration term is described to guard against this coupling, particularly when discrete choices depend strongly on continuous values.

Authors: We appreciate the referee highlighting the need to examine potential gradient interference in the joint objective. Our current manuscript relies on extensive empirical validation across toy models and scientific simulators, where the proposed calibration diagnostic confirms accurate marginal posteriors for both discrete and continuous components even under strong interdependencies. To directly address this concern, the revised manuscript will include a new subsection on training dynamics. This will feature (i) monitoring of discrete marginal calibration throughout joint optimization, (ii) additional experiments that systematically vary the strength of dependence between discrete and continuous parameters, and (iii) a brief discussion of why explicit weighting or alternating optimization was not required in the evaluated settings. These additions will provide concrete evidence that continuous-component gradients do not degrade discrete calibration under the conditions tested. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method is a direct extension trained on independent simulations

full rationale

The paper introduces a factorization of the joint posterior into an autoregressive classifier for discrete parameters and a conditional generative model for continuous parameters, trained jointly under a single simulation-based objective. This is presented as a straightforward methodological extension of standard NPE, with empirical validation on independent toy and scientific simulators. No self-definitional steps, fitted inputs renamed as predictions, or load-bearing self-citations appear in the derivation; the central claim rests on the joint training procedure and post-hoc calibration diagnostics rather than reducing to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the approach relies on standard assumptions of neural density estimation and simulator access.

pith-pipeline@v0.9.0 · 5475 in / 1147 out tokens · 57978 ms · 2026-05-14T20:13:23.183601+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 9 canonical work pages · 1 internal anchor

[1]

Ryan Prescott Adams and David J. C. MacKay. Bayesian online changepoint detection.arXiv preprint arXiv:0710.3742,

work page internal anchor Pith review Pith/arXiv arXiv
[2]

Investigating the impact of model misspecification in neural simulation-based inference.arXiv preprint arXiv:2209.01845,

Patrick Cannon, Daniel Ward, and Sebastian M Schmon. Investigating the impact of model misspecification in neural simulation-based inference.arXiv preprint arXiv:2209.01845,

work page arXiv
[3]

Goncalves, and Jakob H

Michael Deistler, Pedro J. Goncalves, and Jakob H. Macke. Truncated proposals for scalable and hassle-free simulation-based inference. InAdvances in Neural Information Processing Systems, 2022a. Michael Deistler, Jakob H Macke, and Pedro J Gonçalves. Energy-efficient network activity from disparate circuit parameters.Proceedings of the National Academy of...

work page arXiv
[4]

Simultaneous identification of changepoints and model parameters in switching dynamical systems.bioRxiv, pp

Xiaoming Fu, Kai Fan, Heinrich Zozmann, Lennart Schüler, and Justin M Calabrese. Simultaneous identification of changepoints and model parameters in switching dynamical systems.bioRxiv, pp. 2024.01.30.577909,

2024
[5]

Do diffusion models dream of electric planes? Discrete and continuous simulation-based inference for aircraft design.arXiv preprint arXiv:2603.13284,

Aurelien Ghiglino, Daniel Elenius, Anirban Roy, Ramneet Kaur, Manoj Acharya, Colin Samplawski, Brian Matejek, Susmit Jha, Juan Alonso, and Adam Cobb. Do diffusion models dream of electric planes? Discrete and continuous simulation-based inference for aircraft design.arXiv preprint arXiv:2603.13284,

work page arXiv
[6]

Manuel Gloeckler, J. P. Manzano-Patrón, Stamatios N. Sotiropoulos, Cornelius Schröder, and Jakob H. Macke. Scalable simulation-based model inference with test-time complexity control.arXiv preprint arXiv:2603.15292,

work page arXiv
[7]

Kelly, David J

Ryan P. Kelly, David J. Warne, David T. Frazier, David J. Nott, Michael U. Gutmann, and Christo- pher Drovandi. Simulation-based Bayesian inference under model misspecification.arXiv preprint arXiv:2503.12315,

work page arXiv
[8]

BayesFlow 2.0: Multi-backend amortized Bayesian inference in Python.arXiv preprint arXiv:2602.07098,

Lars Kühmichel, Jerry M Huang, Valentin Pratz, Jonas Arruda, Hans Olischläger, Daniel Habermann, Simon Kucharsky, Lasse Elsemüller, Aayush Mishra, Niels Bracher, et al. BayesFlow 2.0: Multi-backend amortized Bayesian inference in Python.arXiv preprint arXiv:2602.07098,

work page arXiv
[9]

Composable effects for flexible and acceler- ated probabilistic programming in numpyro.arXiv preprint arXiv:1912.11554, 2019

Du Phan, Neeraj Pradhan, and Martin Jankowiak. Composable effects for flexible and accelerated proba- bilistic programming in NumPyro.arXiv preprint arXiv:1912.11554,

work page arXiv 1912
[10]

Scuola Medica Salernitana,

Sean Talts, Michael Betancourt, Daniel Simpson, Aki Vehtari, and Andrew Gelman. Validating Bayesian inference algorithms with simulation-based calibration.arXiv preprint arXiv:1804.06788,

work page arXiv
[11]

joint marginal

13 Appendix Contents A Calibration checks 15 A.1 Empirical baselines for discrete calibration checks . . . . . . . . . . . . . . . . . . . . . . . . . 15 A.2 Alternative Calibration checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 B Additional examples 16 B.1 Coal mining disaster changepoint inference . . . . . . . . . . ....

2019
[12]

The observation is the full sequence of annual disaster countsx= (y 1851,...,y 1961)∈N111

and the continuous parameters areθc = (λearly,λlate)∈R2 >0. The observation is the full sequence of annual disaster countsx= (y 1851,...,y 1961)∈N111. The discrete switchpoint is analytically marginalized reducing inference to a purely continuous problem amenable to NUTS (Abril-Pla et al., 2023; Hoffman & Gelman, 2014). The discrete posteriorp(s|x) is the...

1961
[13]

compresses the observation before the density estimator. We also applied a variance-stabilizing√·transform to the Poisson counts before training, which produces a more uniform variance across rate regimes and improves z-scoring during training. We evaluate MNPE across seven training budgets from103 to10 5 simulations, with five independent training runs p...

1905
[14]

Both networks have 3 hidden layers with 32 hidden features each

C Experimental details C.1 Gaussian example MNPE uses a neural spline flow with 5 coupling transforms and 4 blocks for the continuous parameters, and a MADE for the discrete parameters. Both networks have 3 hidden layers with 32 hidden features each. Training uses up toN= 10,000simulations with learning rate set toη= 5e−4, validation set fraction of ten p...

2008
[15]

(2020) to cover biological meaningful ranges and are listed in Tab

and Gonçalves et al. (2020) to cover biological meaningful ranges and are listed in Tab. C-1. We used a step currentIinj of2µA/cm2 for1000msand run the simulation for1450ms. This stimulus and recording protocol corresponds to the voltage recordings from the Allen Institute for Brain Science (2016). 20 Table C-1: Parameter bounds and values that were used ...

2020