This paper addresses a practical problem in simulation-based inference. Modern methods often use neural ratio estimation (NRE) — supervised models that learn the ratio between two probability densities — or flow-based surrogate models, which are generative models that let you both evaluate probability densities and draw samples. Each approach has strengths and weaknesses: a learned ratio alone does not give an explicit normalized density or a way to sample, while flow models can still leave residual errors that hurt precision for complicated simulated data.

The authors propose a hybrid approach they call hybrid neural density estimation. They use a flow to build a parameter-independent reference distribution that can be sampled and evaluated easily. They then train classifiers to estimate the ratio between the true target density (the simulated data distribution) and that reference. Multiplying the learned ratio by the reference density produces an explicit, evaluable surrogate for the target density. The ratio also gives importance weights that let you integrate quantities or resample from the surrogate.

The representation has a few concrete technical payoffs. Because the surrogate is an explicit density, the authors show it defines an exact Asimov dataset — a synthetic dataset whose maximum-likelihood fit returns the parameters that generated the data. The tractable reference distribution also supplies renewable samples for pseudo-experiments, and it makes it possible to apply variance-reduction methods when computing expected test statistics by Monte Carlo.

To demonstrate the idea, the paper applies the construction to a toy statistical model motivated by high-energy physics measurements. For this demonstration the exact densities are known analytically, so the authors can check how well the hybrid surrogate matches the truth. This controlled test helps illustrate the method’s properties without the complications of an opaque simulator.