KDE#

class zfit.prior.KDE(samples, bandwidth=None, name=None)[source]#

Bases: ZfitPrior

Kernel Density Estimate prior from samples.

The KDE prior is a non-parametric way to construct a prior distribution from empirical samples. It estimates the probability density by placing kernels (typically Gaussian) at each sample point. This is particularly useful for hierarchical Bayesian models where posterior samples from one analysis become priors for another.

This prior is suitable for: - Using posterior samples from previous analyses as priors

Implementation details: - Uses exact KDE for < 1000 samples (more accurate but slower) - Switches to grid-based KDE for larger samples (faster approximation) - Automatically adds margin to bounds for numerical stability - Adapts to parameter limits while preserving sample structure

Example

>>> # Create prior from previous posterior samples
>>> posterior_samples = previous_result.samples['param_name']
>>> prior = KDE(posterior_samples)
>>>
>>> # KDE with custom bandwidth
>>> prior = KDE(samples, bandwidth=0.1)
>>>
>>> # Using KDE for hierarchical modeling
>>> group1_posterior = fit1.samples['effect']
>>> group2_prior = KDE(group1_posterior)

Initialize a KDE prior.

Parameters:

  • samples – Array of samples to estimate the density from. Can be a numpy array, list, or tensor.

  • bandwidth (float | str | None) – Bandwidth for kernel density estimation. If None, uses the KDE’s default. Can be float or string like ‘scott’.

  • name (str | None) – Optional name for the prior

Note

The KDE is constructed with a 10% margin beyond the sample range to ensure numerical stability at the boundaries.

__eq__(other)#

Compare two priors for equality.

Parameters:

other – Another ZfitPrior instance to compare with

Returns:

True if the priors are equal

Return type:

bool

__hash__()#

Return hash of the prior based on pdf and name.

Returns:

Hash value for the prior

Return type:

int

log_pdf(value=None)#

Return the log probability of the prior at the given value(s).

Parameters:

value – The parameter value(s) to evaluate the log probability at

Returns:

The log probability

sample(n)#

Sample n values from the prior distribution.

Parameters:

n – Number of samples to draw

Returns:

An array of samples