Sample-Size Extrapolation in Simulation-Based Inference via Statistical Scaling

Shiyi Sun; Lei Fang; Geoff K Nicholls

stat.MLOct 12, 2026

Sample-Size Extrapolation in Simulation-Based Inference via Statistical Scaling

Authors: Shiyi Sun, Lei Fang, Geoff K Nicholls

Organizations: Department of Statistics, University of Oxford, Oxford, United Kingdom. · Mathematical Institute, University of Oxford, Oxford, United Kingdom.

Abstract

In simulation-based inference (SBI), neural posterior estimators trained on a bounded range of sample sizes can become inaccurate on larger datasets because their parameterizations do not specify how posterior location, uncertainty and shape must change as observations accumulate. We propose center-estimated NPE (CE-NPE), which combines a statistical center with an explicit contraction rate and learns a local center correction and residual densities in posterior-scale coordinates. The center is a statistical estimate, such as a closed-form statistic or a simulator-based moment match, which does not require an evaluable likelihood. Our analysis explains how contraction amplifies center error and decomposes Gaussian posterior error into shape, mean and covariance contributions. On six regular benchmarks, CE-NPE gives the best reference-posterior agreement of all tested amortized estimators at every out-of-range size, up to five times the largest training size, without retraining; it also fits better within the training range on most of them. Matched controls and nonregular examples characterize the scope of these gains and identify the conditions on which they depend: an appropriate contraction rate, an accurate center, and a stable residual representation.

Figures & tables

Panel of Figure 1 — Figure 1: Progressively removing what the network must extrapolate. All three constructions share a permutation-invariant encoder and differ in which ingredients of the local coordinates $z_{n}=n^{\alpha}(\theta-c_{n})$ are learned from the bounded training range and which are imposed: (i) free conditioning learns scale, center, and shape; (ii) structural scaling imposes the scale; (iii) CE-NPE additionally anchors the center at a statistical estimator.

Panel of Figure 2 — Figure 2: The same location error becomes more serious as uncertainty shrinks.

Figure 3: Results on the six regular benchmarks. Seed means with

\pm 1

seed-SD bands. The shaded band is the training range; dotted lines mark the targets (

1

0.95

, and

0.5

). SD ratio and center error are on log axes.

Figure 4: Posterior contraction on real 1000 Genomes windows. (a) The typical posterior of each estimator as

n

grows from

10

100

, with the moment fit to all windows as the cross. (b) Against

n

: consistency , the fraction of datasets whose

95\%

intervals contain the full-data fit (higher is better); distance , how far the fit lies from the posterior mean in posterior SDs (lower is better); energy distance between posterior-predictive and held-out real windows (lower is better; independent of the anchor). Tables in Appendix G .

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Prior	$\theta\sim\mathcal{N}(0,\,0.1\,I_{10})$
Simulator	$x_{i}\mid\theta\sim\mathcal{N}(\theta,\,0.1\,I_{10})$
Dimensions	$\theta\in\mathbb{R}^{10}$ , $x_{i}\in\mathbb{R}^{10}$
Center	sample mean $\bar{x}$
Reference posterior	conjugate Gaussian posterior
References	Lueckmann et al. (2021)

Appendix

Table 5

Prior	$\theta\sim U[-1,1]^{10}$
Simulator	$x_{i}\mid\theta\sim\mathcal{N}(\theta,\,0.1\,I_{10})$
Dimensions	$\theta\in\mathbb{R}^{10}$ , $x_{i}\in\mathbb{R}^{10}$
Center	sample mean $\bar{x}$
Reference posterior	product of truncated normals
References	Lueckmann et al. (2021)

Appendix

Table 6

Prior	$\theta\sim U[-10,10]^{2}$
Simulator	$x_{i}\mid\theta\sim\tfrac{1}{2}\mathcal{N}(\theta,I)+\tfrac{1}{2}\mathcal{N}(\theta,0.01\,I)$
Dimensions	$\theta\in\mathbb{R}^{2}$ , $x_{i}\in\mathbb{R}^{2}$
Center	ten EM steps from the coordinate-wise median
Reference posterior	grid integration, $201\times 201$ points around the MAP
References	Sisson et al. (2007) ; Beaumont et al. (2009)

Appendix

Table 7

Prior	$\theta=(\beta,f)$ : $\beta\sim\mathcal{N}(0,2)$ , $f\sim\mathcal{N}\bigl(0,(F^{\top}F)^{-1}\bigr)$ , with $F_{i,i-2}=1$ , $F_{i,i-1}=-2$ , $F_{i,i}=1+\sqrt{(i-1)/9}$ , $F_{i,j}=0$ otherwise, $1\leq i,j\leq 9$
Simulator	$x_{i}=(x_{i,1},\dots,x_{i,10})$ , $x_{i,1}=\sum_{t=1}^{T}z_{t}$ , $x_{i,2:10}=\frac{1}{x_{i,1}}Vz$ , $z_{t}\sim\mathrm{Bern}\bigl(\eta(v_{t}^{\top}f+\beta)\bigr)$ , $\eta(\cdot)=\exp(\cdot)/(1+\exp(\cdot))$ , frozen input between time bins $t-8$ and $t$ : $v_{t}\sim\mathcal{N}(0,I)$ , $V=[v_{1},\dots,v_{T}]$
Dimensions	$\theta\in\mathbb{R}^{10}$ , $x_{i}\in\mathbb{R}^{10}$
Fixed parameters	duration $T=100$ time bins
Center	simulator method of moments: damped Gauss–Newton steps matching $X^{\top}\sigma(X\theta)$ to the replicate mean
Reference posterior	adaptive importance sampling from the Newton MAP
References	Lueckmann et al. (2017) ; Gonçalves et al. (2020)

Appendix

Table 8

Prior	$\theta\sim U[-3,3]^{5}$
Simulator	$x_{i}=(x_{i,1},\dots,x_{i,4})$ , $x_{i,j}\sim\mathcal{N}(m_{\theta},S_{\theta})$ , where $m_{\theta}=(\theta_{1},\theta_{2})^{\top}$ , $S_{\theta}=\bigl(\begin{smallmatrix}s_{1}^{2}&\rho s_{1}s_{2}\\ \rho s_{1}s_{2}&s_{2}^{2}\end{smallmatrix}\bigr)$ , $s_{1}=\theta_{3}^{2}$ , $s_{2}=\theta_{4}^{2}$ , $\rho=\tanh\theta_{5}$
Dimensions	$\theta\in\mathbb{R}^{5}$ , $x_{i}\in\mathbb{R}^{8}$
Center	moment estimator on the branch $\theta_{3},\theta_{4}>0$
Reference posterior	importance sampling on a fundamental domain
References	Papamakarios et al. (2019) ; Greenberg et al. (2019) ; Hermans et al. (2020) ; Durkan et al. (2020)

Appendix

Table 9

Prior	$\theta\sim U[-1,1]^{2}$
Simulator	$x_{i}\mid\theta=\begin{bmatrix}r\cos\alpha+0.25\\ r\sin\alpha\end{bmatrix}+\begin{bmatrix}-\|\theta_{1}+\theta_{2}\|/\sqrt{2}\\ (-\theta_{1}+\theta_{2})/\sqrt{2}\end{bmatrix}$ , where $\alpha\sim U(-\pi/2,\pi/2)$ , $r\sim\mathcal{N}(0.1,\,0.01^{2})$
Dimensions	$\theta\in\mathbb{R}^{2}$ , $x_{i}\in\mathbb{R}^{2}$
Center	moment estimator on the branch $\theta_{1}+\theta_{2}\geq 0$
Reference posterior	grid integration on a fundamental domain, expanded by the symmetry group
References	Greenberg et al. (2019)

Appendix

Table 10

Prior	$\beta\sim\mathrm{LogNormal}(\log 0.4,\,0.5)$ , $\gamma\sim\mathrm{LogNormal}(\log 0.125,\,0.2)$ , truncated to $R_{0}=\beta/\gamma\in[1.5,8]$ (about $12\%$ of prior mass removed; below $1.5$ no epidemic occurs within the horizon, above $8$ the peak precedes the first observation and $\beta$ is weakly identified). Inference is in $(\log\beta,\log\gamma)$ .
Simulator	$x_{i}=(x_{i,1},\dots,x_{i,10})$ , $x_{i,t}\sim\mathrm{Binomial}\bigl(1000,\,\frac{I(t)}{N}\bigr)$ stored as fractions, where $I$ is simulated from $\frac{dS}{dt}=-\beta\frac{SI}{N}$ $\frac{dI}{dt}=\beta\frac{SI}{N}-\gamma I$ $\frac{dR}{dt}=\gamma I$
Dimensions	$\theta\in\mathbb{R}^{2}$ , $x_{i}\in\mathbb{R}^{10}$
Fixed parameters	population $N=10^{6}$ , duration $T=160$ , initial conditions $(S(0),I(0),R(0))=(N-1,1,0)$
Center	simulator method of moments: coarse grid, then three damped Gauss–Newton steps matching the simulator’s mean output to the replicate mean
Reference posterior	grid integration in whitened Laplace coordinates; Gaussian to high accuracy, fitted contraction exponent $0.51$
References	Kermack and McKendrick (1927)

Appendix

Table 11

Prior	$\alpha\sim\mathrm{LogNormal}(-0.125,\,0.5)$ , $\beta\sim\mathrm{LogNormal}(-3,\,0.5)$ , $\gamma\sim\mathrm{LogNormal}(-0.125,\,0.5)$ , $\delta\sim\mathrm{LogNormal}(-3,\,0.5)$ , truncated to parameters for which neither species falls below $e^{-5}$ individuals at any observation time (about $8\%$ of prior mass removed; an effectively extinct species carries no information about its rates). Inference is in log-rate coordinates.
Simulator	$x_{i,1,t}\sim\mathrm{LogNormal}(\log X(t),\,0.1)$ , $x_{i,2,t}\sim\mathrm{LogNormal}(\log Y(t),\,0.1)$ stored as log-counts, where $X,Y$ are simulated from $\frac{dX}{dt}=\alpha X-\beta XY$ $\frac{dY}{dt}=-\gamma Y+\delta XY$
Dimensions	$\theta\in\mathbb{R}^{4}$ , $x_{i}\in\mathbb{R}^{20}$
Fixed parameters	duration $T=20$ time units, initial conditions $(X(0),Y(0))=(30,1)$
Center	simulator method of moments: $13^{4}$ grid, then four damped Gauss–Newton steps with multi-start
Reference posterior	adaptive importance sampling with a Student- $t$ proposal at the Gauss–Newton MAP; Gaussian to high accuracy, fitted contraction exponent $0.50$
References	Lotka (1920)

Appendix

Table 12

Gaussian linear
	C2ST				SD ratio				$95\%$ joint coverage
estimator	$n=10$	$n=20$	$n=50$	$n=100$	$n=10$	$n=20$	$n=50$	$n=100$	$n=10$	$n=20$	$n=50$	$n=100$
Free-G	0.57	0.61	0.75	0.89	1.02	1.09	1.24	1.41	0.96	0.97	0.97	0.86
Scale-G	0.55	0.57	0.65	0.71	1.01	1.00	1.00	1.00	0.96	0.95	0.91	0.87
CE-G	0.51	0.51	0.55	0.57	1.00	1.00	1.03	1.06	0.96	0.96	0.96	0.97
Free-NSF	0.55	0.58	0.72	0.87	1.01	1.06	1.26	1.53	0.96	0.98	0.99	1.00

Appendix

Table 9: Results on the six regular benchmarks at every evaluation size (three seeds, except the mixture: five for C2ST and coverage, two for SD ratio; training

n\in\{5,\dots,20\}

). The two symmetric multimodal benchmarks are in Table 13 . Bold marks the best entry in each column and block (closest to the target where the metric has one).

	T1: parameter-dependent support				T2: boundary parameter
method	KL	SD ratio	center err.	C2ST	KL	SD ratio	center err.	C2ST
Free-G	12.27	1.09	4.72	0.92	0.12	1.00	0.59	0.61
Scale-G	1.31	2.69	2.02	0.82	0.24	0.83	0.39	0.61
CE-G	1.11	2.71	0.54	0.78	0.10	0.81	0.18	0.58
Free-NSF	0.65	2.02	0.87	0.70	0.17	1.02	0.52	0.61
CE-NSF	0.49	2.82	1.81	0.68	0.18	0.84	0.24	0.59

Appendix

Table 10: Nonregular prototypes at

n=100

(

5\times

the largest training size; ten seeds). Each block violates a different assumption of the structured representation.

	CE-G					CE-NSF
imposed $\alpha$	SD ratio	center err.	cov. 90%	C2ST	KL	SD ratio	center err.	cov. 90%	C2ST	KL
$1/2$ (default)	2.71	0.54	0.97	0.78	1.10	2.82	1.81	0.97	0.68	0.49
$0.8$ (estimated)	1.40	0.25	0.91	0.72	0.67	1.42	0.59	0.92	0.59	0.14
$1$ (oracle ablation)	0.92	0.15	0.86	0.69	0.48	0.92	0.23	0.85	0.56	0.03

Appendix

Table 11: Result on T1 at

n=100

	T2, $\theta_{0}=0$ stratum				T3	T4
estimator	SD ratio	center err.	cov. 90%	KL	ridge dist.	mode dist.
Exact posterior	1.00 $\to$ 1.00	0.00 $\to$ 0.00	0.96 $\to$ 0.94	0.00 $\to$ 0.00	–	–
Free-G	0.85 $\to$ 0.74	0.19 $\to$ 0.29	0.70 $\to$ 0.93	0.11 $\to$ 0.11	3.4 $\to$ 7.9	9.5 $\to$ 21.2
Scale-G	0.83 $\to$ 0.86	0.19 $\to$ 0.61	0.53 $\to$ 0.03	0.13 $\to$ 0.43	3.4 $\to$ 7.9	9.5 $\to$ 21.2
CE-G	0.83 $\to$ 0.79	0.19 $\to$ 0.22	0.60 $\to$ 0.58	0.11 $\to$ 0.16	3.4 $\to$ 6.3	9.2 $\to$ 12.7
Free-NSF	1.01 $\to$ 1.14	0.03 $\to$ 0.58	0.77 $\to$ 0.26	0.01 $\to$ 0.23	3.4 $\to$ 7.8	9.5 $\to$ 21.0

Appendix

Table 12: Fixed-

\theta_{0}

strata,

n=20\to 100

(ten seeds). T2 at the boundary (

\theta_{0}=0

), with the exact posterior as reference; T3/T4, distance from each estimator’s posterior mean to the nearest ridge or mode in units of its width, which grows with

n

for any single-center summary, the exact posterior mean included. Accuracy on T3 and T4 is read from Table 10 .

SLCP
	C2ST				MMD				$95\%$ joint coverage
estimator	$n=10$	$n=20$	$n=50$	$n=100$	$n=10$	$n=20$	$n=50$	$n=100$	$n=10$	$n=20$	$n=50$	$n=100$
Free-G	0.98	0.98	0.99	0.99	0.064	0.073	0.084	0.093	0.96	0.99	0.94	0.87
Scale-G	0.98	0.98	1.00	1.00	0.058	0.142	0.363	0.532	0.97	0.92	0.20	0.00
CE-G	0.98	0.98	0.98	0.99	0.040	0.072	0.144	0.208	0.95	0.94	0.87	0.69
Free-NSF	0.75	0.79	0.83	0.91	0.004	0.004	0.007	0.013	0.92	0.98	0.96	0.82

Appendix

Table 13: The two symmetric multimodal sbibm benchmarks.

Gaussian linear
	$n=20$ (training boundary)				$n=100$ ( $5\times$ )
estimator	SD ratio	center err.	cov.	C2ST	SD ratio	center err.	cov.	C2ST
Free-G	1.09	0.11	0.97	0.61	1.41	0.59	0.86	0.89
Free-G (best effort)	1.01	0.07	0.96	0.54	0.98	0.81	0.39	0.93
Free-NSF	1.06	0.11	0.98	0.58	1.53	0.53	1.00	0.87
Free-NSF (best effort)	1.03	0.08	0.96	0.54	1.42	0.44	1.00	0.83

Appendix

Table 14: Best-effort free baselines beside the standard estimators at the training boundary (

n=20

) and at

5\times

(

n=100

); seed means over three seeds. Bold marks the best entry in each column and block (closest to the target where the metric has one).

		C2ST				SD ratio				$95\%$ joint coverage
$N$	estimator	$0.5N$	$1N$	$2.5N$	$5N$	$0.5N$	$1N$	$2.5N$	$5N$	$0.5N$	$1N$	$2.5N$	$5N$
20	Free-G	0.88	0.92	0.97	0.99	2.27	3.07	4.37	7.43	0.94	0.96	0.61	0.36
20	Scale-G	0.88	0.90	0.96	0.99	2.30	2.60	3.09	4.46	0.95	0.92	0.46	0.15
20	CE-G	0.57	0.55	0.64	0.76	1.02	1.00	1.14	1.83	0.94	0.92	0.95	0.96
20	Free-NSF	0.74	0.79	0.92	0.96	1.46	1.80	2.44	3.78	0.95	0.97	0.63	0.38
20	CE-NSF	0.55	0.53	0.61	0.75	1.47	1.26	1.77	2.40	0.94	0.93	0.93	0.70

Appendix

Table 15: SIR at two training maxima against the relative size

n/N

method	C2ST at $n=20$	KL at $n=20$	C2ST at $n=100$	KL at $n=100$	cov. at $n=100$
Free-G (standardized, tuned)	0.898	2.317	0.942	2.980	0.966
CE-G ( $\log n$ residual)	0.552	0.121	0.537	0.096	0.870
Fixed EM center, free scale	0.574	0.174	0.623	0.328	0.974
Fixed EM center, imposed scale	0.558	0.137	0.513	0.033	0.909

Appendix

Table 16: Fixed-center controls on the Gaussian mixture at the training boundary (

n=20

) and

5\times

(

n=100

); means across three seeds.

	SIR				Lotka–Volterra
estimator	$n=10$	$n=20$	$n=50$	$n=100$	$n=10$	$n=20$	$n=50$	$n=100$
Fixed center, free scale	0.653	0.642	0.789	0.897	0.768	0.806	0.886	0.940
Fixed center, imposed scale	0.634	0.572	0.647	0.750	0.729	0.704	0.656	0.704
CE-G	0.640	0.579	0.663	0.772	0.738	0.713	0.671	0.737
CE-NSF	0.602	0.559	0.614	0.740	0.707	0.665	0.655	0.757

Appendix

Table 17: Fixed-center controls on SIR and Lotka–Volterra: C2ST at every evaluation size (training sizes

5

20

); means across three seeds, lowest per column in bold.

benchmark	imposed rate	true asymptotic rate	estimated $\alpha$
Gaussian linear	$1/2$	$1/2$	$0.50$
Gaussian linear (uniform)	$1/2$	$1/2$	$0.48$
Gaussian mixture	$1/2$	$1/2$	$1.38$
Bernoulli GLM	$1/2$	$1/2$	$0.46$
SIR	$1/2$	$1/2$	$0.56$
Lotka–Volterra	$1/2$	$1/2$	$0.48$

Appendix

Table 18: Contraction exponent estimated from the training range.

residual context	KL	SD ratio	90% cov.	center error	SBC KS	exponent
$\log n$	$0.0018$	$0.991$	$0.886$	$0.061$	$0.039$	$0.500$
$1/n$	$0.0012$	$0.998$	$0.889$	$0.059$	$0.038$	$0.500$
none	$0.0023$	$0.992$	$0.888$	$0.065$	$0.041$	$0.500$

Appendix

Table 19: Matched residual-context comparison at

n=10{,}000

(means over five seeds; coverage is the nominal

90\%

region).

Gaussian linear
	C2ST				SD ratio
estimator	$n=10$	$n=20$	$n=50$	$n=100$	$n=10$	$n=20$	$n=50$	$n=100$
Free-NSF	0.55	0.58	0.72	0.87	1.01	1.06	1.26	1.53
PF-NPSE (Langevin)	0.56	0.57	0.57	0.62	1.11	1.12	1.03	0.98
PF-NPSE (GAUSS)	0.51	0.52	0.55	0.59	0.99	0.99	1.00	0.98
CE-NSF	0.51	0.51	0.54	0.56	1.02	1.01	0.99	0.98

Appendix

Table 20: Score-based compositional estimator on the six regular benchmarks. PF-NPSE (three seeds) with the published annealed Langevin sampler and with the GAUSS sampler of Linhart et al. (2026) . Coverage requires a density and is not available for PF-NPSE.

Prior	$\theta\sim U[0.2,\,6]$ per kb; growth parameter $\sim U[0,\,2]$ (raw rate $U[0,8]$ )
Simulator	coalescent tree of $k=20$ under exponential growth (time-changed merging); mutations Poisson on branches, class counts $\xi_{b}$ , then $S=\sum_{b}\xi_{b}$ , $\pi=\sum_{b}b(k-b)\xi_{b}/\binom{k}{2}$ , $\eta=\xi_{1}+\xi_{k-1}$
Dimensions	$\theta\in\mathbb{R}^{2}$ , $x_{i}\in\mathbb{R}^{3}$
Center	simulator method of moments in the Watterson tradition ( Watterson, 1975 ) : the growth rate from the two observed ratios $(\bar{\pi}/\bar{S},\bar{\eta}/\bar{S})$ against moment curves precomputed once by simulation, then $\theta$ from the scale $\bar{S}$ ; a grid lookup with quadratic refinement at query time, likelihood-free and contracting at $n^{-1/2}$
Reference posterior	none: the likelihood is intractable
References	Kingman (1982) ; Watterson (1975)

Appendix

Table 25

	$n=10$			$n=20$			$n=50$			$n=100$
	rad.	cov.	KS	rad.	cov.	KS	rad.	cov.	KS	rad.	cov.	KS
Free-G	1.04	0.94	0.052	0.87	0.98	0.063	0.82	1.00	0.087	0.82	0.98	0.131
Free-NSF	0.98	0.95	0.033	0.92	0.98	0.062	0.92	0.97	0.110	1.01	0.92	0.224
Scale-G	1.00	0.95	0.042	1.06	0.96	0.077	1.34	0.87	0.167	1.64	0.75	0.181
CE-G	0.97	0.95	0.046	0.98	0.96	0.053	1.14	0.94	0.091	1.24	0.87	0.075
CE-NSF	1.00	0.95	0.046	0.95	0.95	0.048	1.04	0.91	0.039	1.10	0.86	0.068

Appendix

Table 22: In-model protocol, every size (seed means): rms standardized truth distance (radius; target

1

), joint

95\%

HPD coverage, and maximum marginal SBC KS; bold: within one seed SD of the best per column.

	$n=10$			$n=20$			$n=50$			$n=100$
	cons.	rad.	energy	cons.	rad.	energy	cons.	rad.	energy	cons.	rad.	energy
Free-G	0.79	1.32	0.162	0.81	1.32	0.132	0.77	1.32	0.094	0.66	1.59	0.085
Free-NSF	0.80	1.28	0.158	0.79	1.52	0.123	0.72	1.63	0.090	0.58	1.92	0.077
Scale-G	0.83	1.21	0.158	0.77	1.44	0.129	0.58	1.87	0.095	0.28	2.79	0.085
CE-G	0.83	1.23	0.153	0.81	1.32	0.122	0.85	1.22	0.087	0.83	1.23	0.079
CE-NSF	0.83	1.18	0.160	0.83	1.28	0.121	0.86	1.22	0.088	0.85	1.28	0.082

Appendix

Table 23: 1000 Genomes windows, every size (seed means): consistency with and standardized distance to the full-data moment fit, and posterior-predictive energy distance to held-out windows; bold: within one seed SD of the best per column.

Explore similar work

May 27, 2026

Conservative neural posterior estimation via distributionally robust training

Simulation-based inference with neural posterior estimation (NPE) often yields overconfident and unreliable posteriors under limited simulation budgets. To address this, we propose DRO-NPE, a distributionally robust approach that replaces the standard NPE objective with a worst-case loss over a Wasserstein ambiguity set. We introduce KL-based metrics for miscoverage and miscalibration, and use these to show that the DRO-NPE objective controls overfitting and reduces posterior overconfidence. Our method is tractable, parallelisable, and readily integrates with standard normalising flows. Across benchmark SBI tasks, DRO-NPE consistently improves coverage and calibration, while narrowing the gap between empirical and population NPE loss, leading to more reliable inference in low-simulation regimes.

William Laplante, Yuga Hikida, Charita Dellaporta +2

Department of Statistical Science, University College London, UK · The Alan Turing Institute, UK · Department of Computer Science, Aalto University, Finland

Simulation-Based Inference Neural Posterior Estimation Probability Calibration Distributionally Robust Optimization Wasserstein Distributionally Robust Optimization Robust Estimation

Ask PDF

May 13, 2026

Mixed neural posterior estimation for simulators with discrete and continuous parameters

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 continuous. However, many scientific models involve parameter spaces that are 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 sbi Python package.

Jan Boelts, Cornelius Schröder, Jonas Beck +3

appliedAI Institute for Europe · Machine Learning in Science, University of Tübingen · Tübingen AI Center +3

Simulation-Based Inference Parameter Estimation Neural Posterior Estimation

Ask PDF

Sep 29, 2026

High-Dimensional Simulation-Based Inference in Latent Spaces

Neural simulation-based inference (SBI) has been widely successful in inferring a relatively small number of interpretable parameters from potentially high-dimensional observations, such as images or time series. Accordingly, representation learning in SBI has focused almost exclusively on compressing the observations used to condition the posterior. More recently, however, SBI has begun to target increasingly high-dimensional parameter spaces, raising the complementary question of whether the inference target itself should be compressed. Our answer is a practical merger of SBI and latent generative modeling, which learns a low-dimensional representation of the simulator parameters, performs posterior inference directly in this latent space, and maps posterior samples back to the original parameter space. We characterize the conditions under which latent-space inference recovers the desired target posterior and systematically study its empirical trade-offs. Across four case studies and three generative families, we compare latent and standard estimators while controlling for network capacity, regularization, optimization, and training compute. At matched training compute, latent-space inference achieves accuracy and marginal calibration comparable to direct target-space inference while sampling up to more than an order of magnitude faster.

Lars Kühmichel, Stefan T. Radev, Bhanu Prasanna Koppolu +3

Department of Statistics, TU Dortmund University, Germany · Department of Cognitive Science, Rensselaer Polytechnic Institute

Simulation-Based Inference Neural Posterior Estimation Efficient Inference Latent Variable Models

Ask PDF