snowline module

Calculates the Bayesian evidence and posterior samples of arbitrary monomodal models.

class snowline.ReactiveImportanceSampler(param_names, loglike, transform=None, resume=True, run_num=None, num_test_samples=2)

Bases: object

Sampler with reactive exploration strategy.

Storage & resume capable, optionally MPI parallelised.

Initialise importance sampler.

Parameters

• param_names (list of str, names of the parameters.) – Length gives dimensionality of the sampling problem.

• loglike (function) – log-likelihood function. Receives multiple parameter vectors, returns vector of likelihood.

• transform (function) – parameter transform from unit cube to physical parameters. Receives multiple cube vectors, returns multiple parameter vectors.

• log_dir (str) – where to store output files

• resume (continue previous run if available.) –

• num_test_samples (int) – test transform and likelihood with this number of random points for errors first. Useful to catch bugs.

run(num_global_samples=400, num_gauss_samples=400, max_ncalls=100000, min_ess=400, max_improvement_loops=4, heavytail_laplaceapprox=True, verbose=True)

Sample at least min_ess effective samples have been drawn.

The steps are:

1. Draw num_global_samples from prior. The highest likelihood point is chosen.

2. Optimize to find maximum likelihood point.

3. Estimate local covariance with finite differences.

4. Importance sample from Laplace approximation (with num_gauss_samples).

5. Construct Gaussian mixture model from samples

6. Simplify Gaussian mixture model with Variational Bayes

7. Importance sample from mixture model

Steps 5-7 are repeated (max_improvement_loops times). Steps 2-3 are performed with MINUIT, Steps 3-6 are performed with pypmc.

Parameters

• min_ess (float) – Number of effective samples to draw

• max_ncalls (int) – Maximum number of likelihood function evaluations

• max_improvement_loops (int) – Number of times the proposal should be improved

• num_gauss_samples (int) – Number of samples to draw from initial Gaussian likelihood approximation before improving the approximation.

• num_global_samples (int) – Number of samples to draw from the prior to

• heavytail_laplaceapprox (bool) – If False, use laplace approximation as initial gaussian proposal If True, use a gaussian mixture, including the laplace approximation but also wider gaussians.

• verbose (bool) – whether to print summary information to stdout

run_iter(num_gauss_samples=400, max_ncalls=100000, min_ess=400, max_improvement_loops=4, heavytail_laplaceapprox=True, verbose=True)

Iterative version of run(). See documentation there. Returns current samples on each iteration.

init_globally(num_global_samples=400)

Sample num_global_samples points from the prior and store the best point.

laplace_approximate(num_global_samples=400, verbose=True)

Find maximum and derive a Laplace approximation there.

Parameters

• num_global_samples (int) – Number of samples to draw from the prior to find a good starting point (see init_globally).

• verbose (bool) – If true, print out maximum likelihood value and point

print_results()

Give summary of marginal likelihood and parameters.

plot(**kwargs)