Defining priors

Priors refer to the prior probability distributions for each model parameter. Typically, these are passed into run_sampler as a regular python dictionary.

The keys of the priors objects should reference the model parameters, in particular, the parameters attribute of the Likelihood. Each key can be either

  • fixed number, in which case the value is held fixed at this value. In effect, this is a Delta-function prior,

  • or a bilby.prior.Prior instance.

If the latter, it will be sampled during the parameter estimation. Here is a simple example that sets a uniform prior for a, and a fixed value for b:

priors = {}
priors['a'] = bilby.prior.Uniform(minimum=0, maximum=10, name='a', latex_label='a')
priors['b'] = 5

Notice, that the latex_label is optional, but if given will be used when generating plots. Latex label strings containing escape characters like \t should either be preceded by r' or include an extra backslash. As an example, either r'$\theta$' or '$\\theta$' is permissible. For a list of recognized escape sequences, see the python docs.

The bilby prior dictionary

The priors passed into run_sampler can just be a regular python dictionary. However, we also provide a class bilby.core.prior.PriorDict which provides extra functionality. For example, to sample from the prior:

>>> priors = bilby.core.prior.PriorDict()
>>> priors['a'] = bilby.prior.Uniform(minimum=0, maximum=10, name='a')
>>> priors['b'] = bilby.prior.Uniform(minimum=0, maximum=10, name='b')
>>> priors.sample()
{'a': 0.1234, 'b': 4.5232}

Available prior classes

We have provided a number of standard priors. An exhaustive list can be found in the API section.

Multivariate Gaussian prior

We provide a prior class for correlated parameters in the form of a multivariate Gaussian distribution. To set the prior you first must define the distribution using the bilby.core.prior.MultivariateGaussianDist class. This requires the names of the correlated variables, their means, and either the covariance matrix or the correlation matrix and standard deviations, e.g.:

>>> names = ['a', 'b']  # set the parameter names
>>> mu = [0., 5.]  # the means of the parameters
>>> cov = [[2., 0.7], [0.7, 3.]]  # the covariance matrix
>>> mvg = bilby.core.prior.MultivariateGaussianDist(names, mus=mu, covs=cov)

It is also possible to define a mixture model of multiple multivariate Gaussian modes of different weights if required, e.g.:

>>> names = ['a', 'b']  # set the parameter names
>>> mu = [[0., 5.], [2., 7.]]  # the means of the parameters
>>> cov = [[[2., 0.7], [0.7, 3.]], [[1., -0.9], [-0.9, 5.]]]  # the covariance matrix
>>> weights = [0.3, 0.7]  # weights of each mode
>>> mvg = bilby.core.prior.MultivariateGaussianDist(names, mus=mu, covs=cov, nmodes=2, weights=weights)

The distribution can also have hard bounds on each parameter by supplying them.

The bilby.core.prior.MultivariateGaussianDist class can then be passed to a bilby.core.prior.MultivariateGaussian prior for each parameter, e.g.:

>>> priors = dict()
>>> priors['a'] = bilby.core.prior.MultivariateGaussian(mvg, 'a')
>>> priors['b'] = bilby.core.prior.MultivariateGaussian(mvg, 'b')

The detailed API information for the distribution and prior classes can be found in the API section.

Defining your own prior

You can define your own by subclassing the bilby.prior.Prior class.

Prior Constraints

Added v. 0.4.3

This allows cuts to be specified in the prior space.

You can provide the PriorDict a conversion_function and a set of Constraint priors to remove parts of the prior space.

Note: after doing this the prior probability will not be normalised.

Simple example

Sample from uniform distributions in two parameters x and y with the condition x >= y.

First thing: define a function which generates z=x-y from x and y.

def convert_x_y_to_z(parameters):
    Function to convert between sampled parameters and constraint parameter.

    parameters: dict
        Dictionary containing sampled parameter values, 'x', 'y'.

    dict: Dictionary with constraint parameter 'z' added.
    converted_parameters = parameters.copy()
    converted_parameters['z'] = parameters['x'] - parameters['y']
    return converted_parameters

Create our prior:

from bilby.core.prior import PriorDict, Uniform, Constraint

priors = PriorDict(conversion_function=convert_x_y_to_z)
priors['x'] = Uniform(minimum=0, maximum=10)
priors['y'] = Uniform(minimum=0, maximum=10)
priors['z'] = Constraint(minimum=0, maximum=10)

Sample from this distribution and plot the samples.

import matplotlib.pyplot as plt

samples = priors.sample(1000000)
plt.hist2d(samples['x'], samples['y'], bins=100, cmap='Blues')