dynesty is a Pure Python, MIT-licensed Dynamic Nested Sampling package for estimating Bayesian posteriors
and evidences. See Crash Course and Getting Started
for more information. The latest development version can be found here.
The release paper describing dynesty 1.0 can be found here.
As a multi-purpose sampler,
dynesty is designed to perform
“reasonably well” across a large array of problems but is not optimized for
any single one. In particular, please take caution when applying
estimate Bayesian posteriors and evidences for large-dimensional
(>30 dimensions or so) problems.
dynesty is compatible wit Python 3.6+. It requires
numpy (for arithmetic),
scipy (for special functions),
matplotlib (for plotting), and
While not required,
tqdm also allows for a nice progress bar.
Installing the most recent stable version of the package is as easy as:
pip install dynesty
Alternately, for users who might want newer development versions, it can also be installed directly from a local copy of the repository by running:
python setup.py install
If you find
dynesty useful in your research, please cite the
software and relevant papers.
A list of papers that you
should cite can be generated directly from the
sampler object by calling:
This will return a list of relevant papers and corresponding links to download citation information such as BibTex files. As an example:
import dynesty sampler = dynesty.DynamicNestedSampler(loglike, prior_transform, ndim, bound='balls', sample='rwalk') print(sampler.citations)
Returns the following list of papers that should be cited:
Code and Methods: ================ Speagle (2020): https://ui.adsabs.harvard.edu/abs/2020MNRAS.493.3132S Nested Sampling: =============== Skilling (2004): ui.adsabs.harvard.edu/abs/2004AIPC..735..395S Skilling (2006): projecteuclid.org/euclid.ba/1340370944 Dynamic Nested Sampling: ======================= Higson et al. (2019): doi.org/10.1007/s11222-018-9844-0 Bounding Method: =============== Buchner (2016): ui.adsabs.harvard.edu/abs/2014arXiv1407.5459B Buchner (2017): ui.adsabs.harvard.edu/abs/2017arXiv170704476B Sampling Method: =============== Skilling (2006): projecteuclid.org/euclid.ba/1340370944
If you have utilized some of the error analysis features available through the provided utility functions (see Nested Sampling Errors), you should also cite Chopin & Robert (2010), Higson et al. (2018), and Speagle (2020).
See References and Acknowledgements for additional details.
Bug fix release
- The problem with biased posteriors was fixed when using multi-ellipsoid bounds and rslice and rwalk samplers. Previously the chains did not satisfy detailed balance. (issue #364). Original discovery of the problem and help by Colm Talbot. In the case of complex posteriors, somewhat slower performance may be seen.
- Fix the issue introduced in 1.2.1 when the prior_transform returns a tuple or or a list (rather than numpy array). Now that should be accepted.
Small bug fix release
- The arguments of prior_transform and likelihood function are now explicitely copied, so the sampling can work if those function apply changes to argument vectors ( #362 )
- Fix the compilation of the docs, and update them a bit
This version has multiple changes that should improve stability and speed. The default dynamic sampling behaviour has been changed to focus on the effective number of posterior samples as opposed to KL divergence. The rstagger sampler has been removed and the default choice of the sampler may be different compared to previous releases depending on the dimensionality of the problem. dynesty should now provide 100% reproduceable results if the rstate object is provided. It needs to be a new generation Random Generator (as opposed to numpy.RandomState)
Most of the changes in the release have been contributed by [Sergey Koposov](https://github.com/segasai) who has joined the dynesty project.
- Saving likelihood. It is now possible to save likelihood calls history during sampling into HDF5 file (this is not compatible with parallel sampling yet). The relevant options are save_history=False, history_filename=None (#235)
- add_batch() function now has the mode parameter that allows you to manually chose the logl range for the batch (#328)
- More testing with code coverage of >90% + validation on test problems
- Internal refactoring reducing code duplication (saved_run, integral calculations, different samplers etc)
- Multiple speedups: ellipsoid bounds, bootstrap, jitter_run (#239, #256, #329)
- Exception is raised if unknown arguments are provided for static/dynamic samplers (#295)
- Migrate to new numpy random generator functionality from RandomState (#280)
- Make dynesty fully deterministic if random state is provided (#292)
- Remove the pointvol parameter used in internal calculations, such as ellipsoid splits (#284)
- Get rid of vol_dec parameter (#286)
- Improve bounding ellipsoids algorithms, for example how we are dealing with degenerate ellipsoids (#264, #268)
- Introduce more stable multi-ellipsoidal splitting using BIC (#286)
- Do not use KL divergence function for stopping criteria based on posterior, instead use the criterion based on the number of effective samples. The old behaviour can still be achieved by using the dynesty.utils.old_stopping_function (#332)
- Fix bugs in dynamic sampling that can lead to sampler not finding points in the interval (#244)
- Major refactor of rslice/slice sampling increasing its stability (#269, #271)
- Disable ncdim when slice sampling (#271)
- Change the defaults for slices/walks/bootstrap (vs number of dimensions) (#297)
- Change default samplers (vs ndim, i.e. use rslice for high dimensions) (#286)
- Remove rstagger sampler, as it was producing incorrect results/based on non-Markovian chains (#322)
- Fix rwalk sampler. Previously the chains were not Markovian (#319, #323, #324)
- Change step adaptation of rwalk and rslice (#260, #323)
- Change the calculation of evidence uncertainties, remove factor the unnecessary factor two, and improve numerical stability of calculations (#306, #360)
- Refactor the addition of batches in dynamic sampling, preventing possibly infinite loop (#326)
- Improve stability of resample_equal (#351)
- New Results interface with a dedicated object, rather than a wrapper around the dictionary (#330)
- Improved behavior and stability of the bounding distributions (with Sergey Koposov and Johannes Buchner).
- Added support for specifying the number of clustering dimensions (
'ncdim') in case these may differ from the number of prior dimensions (
'npdim') (with Colm Talbot).
- Fixed a bug where
dynestywas not properly enforcing nested sampling’s monotonically-increasing likelihood condition when sampling (with Colm Talbot).
- Improved ability to save sampler objects to disk to backup progress (with Patrick Sheehan and Alex Nitz).
- Limited support for user-defined proposal strategies (with Gregory Ashton).
- Additional small bugfixes, references, and documentation updates.
- Small quality-of-life improvements to plotting.
- Added citation tool.
- Ensemble bounds can now adapt to elongated distributions (with Johannes Buchner).
- Random walks now behave differently near boundaries (with Gregory Ashton).
- Pickling sampler states should now work better in Python 3 (with Dustin Lang.
- Doubled output errors in default approximation in line with theoretical expectations.
- Small bugfixes and docfixes (with Patricio Cubillos).
- Added support for periodic boundary conditions.
- Set up basic tests for continuous integration.
- Added a logo!
- Updated and reorganized documentation and demos.
- Added proper support for gradients.
- Changed defaults and added several “quality of life” improvements.
- Updated documentation.
- Modified re-scaling behavior to better deal with inefficient proposals.
- Improved stability of the current ellipsoid decomposition algorithm.
- Added new
'auto'options and changed a number of defaults to make things easier for general users.
- Plotting now defaults to 95% credible intervals instead of 68%.
- Added in a fast approximation option for
- Modified the default stopping heuristic. It now evaluates significantly faster but is a less accurate probe of the “true” KL divergence.
'rwalk'behavior to better deal with edge cases.
- Changed defaults so performance should now be more stable (albiet slower) for the average user.
- Improved the stability of bounding ellipsoids.
- Fixed performance issues with
- Small plotting improvements.
- Fixed a minor bootstrapping bug that affected performance for some users.
- Fixed a serious bug associated with the new singular decomposition algorithm and changed its behavior so it no longer auto-kills user runs when it fails.
dynestyis now on PyPI!
- Added two new slice sampling options (
- Changed internals to allow user to access quantities during dynamic batch allocation. WARNING: Breaks some aspects of backwards compatibility for advanced users utilizing generators.
- Simplified parallelism options.
- Fixed a singular decomposition bug that occasionally appeared during runtime.
- Small plotting/utility improvements.
- Fixed additional Python 2/3 compatibility bugs.
- Added the ability to pass user-specified custom print functions.
- Added importance reweighting.
- Small improvements to plotting utilities.
- Small changes to improve user outputs and basic functionality.
mapbugs that broke compatibility between Python 2 and 3.
- Fixed a bug where the sampler could break during the first update from the
unit cube when using a
- Introduced a function wrapper for
loglikelihoodfunctions to allow users to pass
- Fixed a small bug that could cause bounding ellipsoids to fail.
- Introduced a stability fix to the default
weight_functionwhen computing evidence-based weights.
Initial beta release.