# Point source likelihood

icecube_tools also provides an interface to the likelihoods often used in point source searches of the neutrino data (see this paper by Braun et al.).

$\mathcal{L} = \prod_{i=1}^N \Bigg[ \frac{n_s}{N} \mathcal{S}(\theta_i, E_i, \gamma) + (1-\frac{n_s}{N}) \mathcal{B}(\theta_i, E_i) \Bigg],$

where $$N$$ is the total number of detected neutrino events, $$n_s$$ is the expected number of source events, $$\theta$$ is the neutrino direction, $$E$$ is the reconstructed neutrino energy and $$\gamma$$ is the source spectral index.

The point source likelihood is a mixture model with two components: one representing possible astrophysical neutrino sources, $$\mathcal{S}(\theta, E)$$, and the other known background, $$\mathcal{B}(\theta, E)$$. Each component has terms depending on the directional or spatial source–neutrino relationship and also the energy of the neutrinos, as higher energy neutrinos are more likely to come from astrophysical sources. Depending on the search, the energy dependence may be omitted. Also, there may be a temporal dependence added, but this is not yet implemented in icecube_tools.

Here we implement a simple likelihood and apply it to some simulated data. There are several likelihoods available, and more information can be found in the API documentation.

[1]:

import numpy as np
from matplotlib import pyplot as plt
import h5py

from icecube_tools.point_source_likelihood.spatial_likelihood import (
EventDependentSpatialGaussianLikelihood,
)
from icecube_tools.point_source_likelihood.energy_likelihood import (
MarginalisedEnergyLikelihood2021,
MarginalisedIntegratedEnergyLikelihood,
)
from icecube_tools.point_source_likelihood.point_source_likelihood import (
PointSourceLikelihood,
TimeDependentPointSourceLikelihood,
)

from icecube_tools.detector.effective_area import EffectiveArea
from icecube_tools.detector.r2021 import R2021IRF
from icecube_tools.detector.detector import IceCube
from icecube_tools.utils.data import Events, SimEvents, RealEvents


## Spatial likelihood

We can start with the spatial/directional term. Let’s use the energy dependent spatial likelihood. It is build from a Gaussian with an event-wise uncertainty sampled from the IRF data. The background case will simply be an isotropic distribution on the sphere.

[2]:

angular_resolution = 5
spatial_likelihood = EventDependentSpatialGaussianLikelihood(angular_resolution)


We show the likelihood profile for a single event with an assumed uncertainty of 1 degree.

[3]:

source_coord = (np.pi, np.deg2rad(30))
test_ra = np.pi + np.linspace(-0.1, 0.1, 100)
test_dec = np.full((100), source_coord[1])

fig, ax = plt.subplots()
ax.axvline(
)
ax.set_xlabel("RA [deg]")
ax.set_ylabel("Spatial likelihood")
ax.legend()

[3]:

<matplotlib.legend.Legend at 0x7ff296af28b0>


## Energy likelihood

Now let’s think about the energy-dependent term. The way this is handled is to marginalise over the true neutrino energies, to directly connect the reconstructed neutrino energies to the spectral index of a simple power-law source model.

Doing this properly requires a knowledge of the relationship between the true and reconstructed energies as well as the details of the power law model. The most straightforward way to implement this is to simulate the a large number of events using the Simulator and build a likelihood using the output of this simulation and MarginalisedEnergyLikelihoodFromSim. We do exactly this with pre-computed lists of events, to be found in the data subdirectory: sim_output_{index}.h5. These were simulated using point sources with spectral index index at 45 degrees declination. The likelihood is restricted to a small band of declination around the assumed source. Using the same declination for our test source, this is fine. For different source declinations further simulations would be needed to account for the declination dependence of the detector acceptance.

[4]:

aeff = EffectiveArea.from_dataset("20210126", period="IC86_II")
irf = R2021IRF.from_period("IC86_II")
# new_reco_bins = irf.reco_energy_bins[12, 2]
new_reco_bins = np.linspace(1, 9, num=25)
detector = IceCube.from_period("IC86_II")
energy_likelihood = MarginalisedIntegratedEnergyLikelihood(
detector, new_reco_bins, max_index=4.5
)
# energy_likelihood = MarginalisedEnergyLikelihood2021([1.5, 2.0, 2.5, 3.0, 3.5, 3.7, 4.0], 'data', 'sim_output', np.pi/4,)
# the likelihood class is backwardscompatible with the "older" simulation-based energy likelihood

[5]:

new_reco_bins

[5]:

array([1.        , 1.33333333, 1.66666667, 2.        , 2.33333333,
2.66666667, 3.        , 3.33333333, 3.66666667, 4.        ,
4.33333333, 4.66666667, 5.        , 5.33333333, 5.66666667,
6.        , 6.33333333, 6.66666667, 7.        , 7.33333333,
7.66666667, 8.        , 8.33333333, 8.66666667, 9.        ])

[6]:

# test_energies = np.geomspace(10, 1e7) # GeV
test_indices = [2.0, 2.5, 3, 3.5]
energy = np.logspace(2, 7.66, num=1000, endpoint=False)
fig, ax = plt.subplots()
for index in test_indices:
ax.plot(
energy,
label=f"{index:.1f}",
)
ax.set_xscale("log")
ax.set_yscale("log")
ax.set_xlabel("E_reco [GeV]")
ax.set_ylabel("Energy likelihood")
ax.legend()

[6]:

<matplotlib.legend.Legend at 0x7ff28a61c0d0>


## Point source likelihood

Now we can bring together the spatial and energy terms to build a full PointSourceLikelihood. First, let’s load some data from the simulation notebook to allow us to demonstrate.

[7]:

events = SimEvents.load_from_h5("h5_test.hdf5")

Resetting rng


Now lets put our likelihood structure and data in together, along with a proposed source location:

[8]:

likelihood = PointSourceLikelihood(
spatial_likelihood,
energy_likelihood,
events.ra[events.periods[0]],
events.dec[events.periods[0]],
events.reco_energy[events.periods[0]],
events.ang_err[events.periods[0]],
source_coord,
which="both",
)


The likelihood will automatically select a declination band around the proposed source location. Because of the Gaussian spatial likelihood, neutrinos far from the source will have negligible contribution. We can control the width of this band with the optional argument band_width_factor. Let’s see how many events are there in total, in the selected declination band and in vicinity of the proposed source:

[9]:

likelihood.Ntot, likelihood.N, likelihood.Nprime

[9]:

(204, 175, 45)


We also note that the background likelihood is implemented automatically, for more information on the options here, check out the API docs. This is just a function of energy, with a constant factor to account for the isotropic directional likelihood.

[10]:

fig, ax = plt.subplots()
energy = np.logspace(new_reco_bins[0], new_reco_bins[-1], num=1000, endpoint=False)
ax.step(
energy,
label=f"index 3.7",
)
ax.set_xscale("log")
ax.set_yscale("log")
ax.set_xlabel(r"$\log_{10}{(E_\mathrm{reco} / \mathrm{GeV})}$")
ax.set_ylabel("Background likelihood")

[10]:

Text(0, 0.5, 'Background likelihood')


# Time dependent point source analysis

Encompassing multiple data seasons, although shown only for one (the same as above), uses a single ns, which is distributed among the seasons with the number of expected events for gamma and the provided observational times (times) as relative weight.

[28]:

events = SimEvents.load_from_h5("h5_test.hdf5")
# energy_likelihood = MarginalisedIntegratedEnergyLikelihood(irf, aeff, new_reco_bins)
# index_list = list(np.arange(1.5, 4.25, 0.25))
tllh = TimeDependentPointSourceLikelihood(
source_coords,
["IC86_II"],
events.ra,
events.dec,
events.reco_energy,
events.ang_err,
which="both",
times={"IC86_II": 1},
sigma=2,
band_width_factor=5,
)

m = tllh._minimize()

m

Resetting rng

[28]:

FCN = -142.9 Nfcn = 58
EDM = 3.39e-08 (Goal: 0.0001) time = 0.2 sec
Valid Minimum No Parameters at limit
Below EDM threshold (goal x 10) Below call limit
Covariance Hesse ok Accurate Pos. def. Not forced
 Name Value Hesse Error Minos Error- Minos Error+ ns 26 4 0 106 index 2.54 0.21 1.5 5
 ns index 19.7 0.00195 0.00195 0.0447
[29]:

_ = m.draw_profile("index")

[30]:

_ = m.draw_profile("ns")


There is also a method to use only the events within the selected declination band and their energy to fit the background index. Default setting is one background index (index_atmo), using two at once does not lead to a converging fit right now.

[31]:

tllh._minimize_bg()

[31]:

FCN = 102.1 Nfcn = 11
EDM = 1.25e-05 (Goal: 0.0001)
Valid Minimum No Parameters at limit
Below EDM threshold (goal x 10) Below call limit
Covariance Hesse ok Accurate Pos. def. Not forced
 Name Value Hesse Error Minos Error- Minos Error+ Limit- weight 0.0 0.1 0 1 yes index_astro 2.5 0.1 1.5 5 yes index_atmo 3.33 0.20 1.5 5
 weight index_astro index_atmo 0 0 0 0 0 0 0 0 0.0385
[32]:

_ = tllh.m.draw_profile("index_atmo", bound=(1.6, 3.9))

[ ]: