Hello.

This is a small module that finds sensible priors for correlation functions (to be fit using the corrfitter package) in an automatic and pure-Bayesian way.

To get the full documentation, go into this directory and type

[ firefox | chrome | etc. ] doc/_build/html/index.html

Traditionally, one would design priors for a correlator fit using some combination of QCD-inspired rules of thumb ("the spectrum is separated by about Lambda_QCD" ), and empirical-Bayes type arguments (like eyeing the effective mass of correlators averaged over all configurations).

While empirical Bayes approaches are often fine since they approximate a pure Bayesian result, they do technically involve double-counting the information avaliable to us, and this makes some people feel on-edge. By double-counting, I mean information from the same data is used to determine both the prior and the likelihood distributions, when we're doing this we're on theoretically shaky ground.

This module produces priors in a "pure-Bayesian" way (as opposed to empirical Bayesian), meaning that the prior and likelihood distributions are found from independant bits of data.

With a gv.Dataset.Dataset full of correlators dset, you can run

prior, new_dset = corrbayes.get_prior( dset, 1, nexp )

This line will shave off a (randomly chosen) single datapoint for each key (e.g. a correlator for each key on a single configuration), and return new_dset, the same as dset but with that single point taken out. Then that point is used to deduce sensible priors for a fit using new_dset. No information from new_dset has been used to determine these priors, hence no double-counting of information.

If you are also doing fits to 3-point correlators, this can be handled too. All you need to do is pass an argument into get_priors called currents, this must be a list of strings giving the name you gave the current in your datatag convention (See tag conventions section).

The second argument of get_prior in the above code segment gives the number of points to be used for working out priors (therefore also the number of points shaved off the datset). One is fine in many cases, but if your correlators are noisy (e.g. vector mesons), then cranking this up to 10 or 20 would make the process more stable.

If you find the priors generated are too tight, you can loosen them by passing get_prior the optional argument loosener. If for example you pass in loosener=0.5, this adds a 50% error onto the priors. The default loosener is 0.3.

This thing should be compatable with at least python2.7-3.6.

I should also mention; this only works for single source/sink combinations, e.g. if you're planning on using a matrix of smearings, this won't work properly. If you want to use this with smearings, let me know and I'll adapt the code so it can handle that kind of thing.