Bayesian posterior for IceCube 7-year point-source search with neutrino-count statistics


The presence of a population of point sources in a dataset modifies the underlying neutrino-count statistics from the Poisson distribution. This deviation can be exactly quantified using the non-Poissonian template fitting technique, and in this work we present the first application of this approach to the IceCube high-energy neutrino dataset. Using this method, we search in 7 years of IceCube data for point-source populations correlated with the disk of the Milky Way, the Fermi bubbles, the Schlegel, Finkbeiner, and Davis dust map, or with the isotropic extragalactic sky. In addition, a northern-sky-only spatial template is also considered. In order to enhance the flexibility of the results, we publish the full posterior from our analysis, which can be used to establish limits on specific population models that would contribute to the observed IceCube neutrino flux.

A full description of the method can be found in

Data release

Suggested citation for this dataset:

IceCube Collaboration (2019): Bayesian posterior for IceCube 7-year point-source search with neutrino-count statistics. DOI:10.21234/dtrs-e557

Click here to download (.zip, 12MB)

Included in the download are the following files:

Data files

This release contains the posterior and a short example, written in Python, on how to load and use the posterior. We emphasize that our analysis and hence these posteriors are constructed with the assumption that the astrophysical population produces neutrinos with an E^(-2) energy spectrum.

posterior.h5: Within this HDF5 formatted file, five tables named IsotropicGalactic_diskFermi_bubbleSFD_dust, and Northern_sky contain the posterior samples for their respective templates.

Each table describes equally weighted samples using five columns. Four columns, labeled ln_Aln_Fbn1, and n2, contain the coordinates for the sample in natural logarithmic parameter space for the differential source-count function normalization A and break F_b, while the power indices n_1 and n_2 are in linear space. The fifth column — labeled loglikelihood — gives the natural logarithm of the likelihood function at the location of the corresponding sample. In addition, each table has two attributes named P_log_evidence and NP_log_evidence that contain the natural logarithm of the evidence integral for the Poissonian model and non-Poissonian model, respectively.

The root node of the HDF5 file also contains a series of attributes named units_ln_Aunits_ln_Fbunits_n1, and units_n2 that specify the units that the posterior sample coordinates are given in. Another series of root attributes named prior_ln_Aprior_ln_Fbprior_n1, and prior_n2 give the probability density of the uniform priors for each of the model parameters. An example program that loads the posterior samples, computes the Bayes factor at specific model locations in parameter space, and generates a reproduction of Figure 10 in the accompanying paper. This program requires the numpyscipymatplotlib and pytables (named tables in pip) python libraries to be installed.

For any questions about this data release, please write to