PosteriorStacker
Combines Bayesian analyses from many datasets.
- Introduction
- Method
- Tutorial
- Output plot and files
Introduction
Fitting a model to a data set gives posterior probability distributions for a parameter of interest. But how do you combine such probability distributions if you have many datasets?
This question arises frequently in astronomy when analysing samples, and trying to infer sample distributions of some quantity.
PosteriorStacker allows deriving sample distributions from posterior distributions from a number of objects.
Method
The method is described in Appendix A of Baronchelli, Nandra & Buchner (2020).
The inputs are posterior samples of a single parameter, for a number of objects. These need to come from pre-existing analyses, under a flat parameter prior.
The hierarchical Bayesian model (illustrated above) models the sample distribution as a Gaussian with unknown mean and standard deviation. The per-object parameters are also unknown, but integrated out numerically using the posterior samples.
Additional to the Gaussian model (as in the paper), a histogram model (using a flat Dirichlet prior distribution) is computed, which is non-parametric and more flexible. Both models are inferred using UltraNest.
The output is visualised in a publication-ready plot.
Synopsis of the program:
$ python3 posteriorstacker.py --help usage: posteriorstacker.py [-h] [--verbose VERBOSE] [--name NAME] filename low high nbins Posterior stacking tool. Johannes Buchner (C) 2020-2021 Given posterior distributions of some parameter from many objects, computes the sample distribution, using a simple hierarchical model. The method is described in Baronchelli, Nandra & Buchner (2020) https://ui.adsabs.harvard.edu/abs/2020MNRAS.498.5284B/abstract Two computations are performed with this tool: - Gaussian model (as in the paper) - Histogram model (using a Dirichlet prior distribution) The histogram model is non-parametric and more flexible. Both models are computed using UltraNest. The output is plotted. positional arguments: filename Filename containing posterior samples, one object per line low Lower end of the distribution high Upper end of the distribution nbins Number of histogram bins optional arguments: -h, --help show this help message and exit --verbose VERBOSE Show progress --name NAME Parameter name (for plot) Johannes Buchner (C) 2020-2021
Licence
AGPLv3 (see COPYING file). Contact me if you need a different licence.
Install
Clone or download this repository. You need to install the ultranest python package (e.g., with pip).
Tutorial
In this tutorial you will learn:
- How to find a intrinsic distribution from data with asymmetric error bars and upper limits
- How to use PosteriorStacker
Lets say we want to find the intrinsic velocity dispersion given some noisy data points.
Our data are velocity measurements of a few globular cluster velocities in a dwarf galaxy, fitted with some model.
Preparing the inputs
For generating the demo input files and plots, run:
$ python3 tutorial/gendata.py
Visualise the data
Lets plot the data first to see what is going on:
Caveat on language: These are not actually "the data" (which are counts on a CCD). Instead, this is a intermediate representation of a posterior/likelihood, assuming flat priors on velocity.
Data properties
This scatter plot shows:
- large, sometimes asymmetric error bars
- intrinsic scatter
Resampling the data
We could also represent each data point by a cloud of samples. Each point represents a possible true solution of that galaxy.
Running PosteriorStacker
We run the script with a range limit of +-100 km/s:
$ python3 posteriorstacker.py posteriorsamples.txt -80 +80 11 --name="Velocity [km/s]" fitting histogram model... [ultranest] Sampling 400 live points from prior ... [ultranest] Explored until L=-1e+01 [ultranest] Likelihood function evaluations: 114176 [ultranest] Writing samples and results to disk ... [ultranest] Writing samples and results to disk ... done [ultranest] logZ = -20.68 +- 0.06865 [ultranest] Effective samples strategy satisfied (ESS = 684.4, need >400) [ultranest] Posterior uncertainty strategy is satisfied (KL: 0.46+-0.08 nat, need <0.50 nat) [ultranest] Evidency uncertainty strategy is satisfied (dlogz=0.14, need <0.5) [ultranest] logZ error budget: single: 0.07 bs:0.07 tail:0.41 total:0.41 required:<0.50 [ultranest] done iterating. logZ = -20.677 +- 0.424 single instance: logZ = -20.677 +- 0.074 bootstrapped : logZ = -20.676 +- 0.123 tail : logZ = +- 0.405 insert order U test : converged: False correlation: 377.0 iterations bin1 0.051 +- 0.046 bin2 0.052 +- 0.051 bin3 0.065 +- 0.058 bin4 0.062 +- 0.057 bin5 0.108 +- 0.085 bin6 0.31 +- 0.14 bin7 0.16 +- 0.10 bin8 0.051 +- 0.050 bin9 0.047 +- 0.044 bin10 0.048 +- 0.047 bin11 0.047 +- 0.045 fitting gaussian model... [ultranest] Sampling 400 live points from prior ... [ultranest] Explored until L=-4e+01 [ultranest] Likelihood function evaluations: 4544 [ultranest] Writing samples and results to disk ... [ultranest] Writing samples and results to disk ... done [ultranest] logZ = -47.33 +- 0.07996 [ultranest] Effective samples strategy satisfied (ESS = 1011.4, need >400) [ultranest] Posterior uncertainty strategy is satisfied (KL: 0.46+-0.07 nat, need <0.50 nat) [ultranest] Evidency uncertainty strategy is satisfied (dlogz=0.17, need <0.5) [ultranest] logZ error budget: single: 0.13 bs:0.08 tail:0.41 total:0.41 required:<0.50 [ultranest] done iterating. logZ = -47.341 +- 0.440 single instance: logZ = -47.341 +- 0.126 bootstrapped : logZ = -47.331 +- 0.173 tail : logZ = +- 0.405 insert order U test : converged: False correlation: 13.0 iterations mean -0.3 +- 4.7 std 11.6 +- 5.2 Vary the number of samples to check numerical stability! plotting results ...
Notice the parameters of the fitted gaussian distribution above. The standard deviation is quite small (which was the point of the original paper). A corner plot is at posteriorsamples.txt_out_gauss/plots/corner.pdf
Visualising the results
Here is the output plot, converted to png for this tutorial with:
$ convert -density 100 posteriorsamples.txt_out.pdf out.png
In black, we see the non-parametric fit. The red curve shows the gaussian model.
The histogram model indicates that a more heavy-tailed distribution may be better.
The error bars in gray is the result of naively averaging the posteriors. This is not a statistically meaningful procedure, but it can give you ideas what models you may want to try for the sample distribution.
Output files
- posteriorsamples.txt_out.pdf contains a plot,
- posteriorsamples.txt_out_gauss contain the ultranest analyses output assuming a Gaussian distribution.
- posteriorsamples.txt_out_flexN contain the ultranest analyses output assuming a histogram model.
- The directories include diagnostic plots, corner plots and posterior samples of the distribution parameters.
With these output files, you can:
- plot the sample parameter distribution
- report the mean and spread, and their uncertainties
- split the sample by some parameter, and plot the sample mean as a function of that parameter.
If you want to adjust the plot, just edit the script.
If you want to try a different distribution, adapt the script. It uses UltraNest for the inference.
Take-aways
- PosteriorStacker computed a intrinsic distribution from a set of uncertain measurements
- This tool can combine arbitrarily pre-existing analyses.
- No assumptions about the posterior shapes were necessary -- multi-modal and asymmetric works fine.