Revision 177e66fb023362fceeb3f99605c9433c864d8529 authored by Christopher M. Biwer on 07 June 2017, 17:06:29 UTC, committed by Collin Capano on 07 June 2017, 17:06:29 UTC
1 parent 705c6da
sensitivity.py
""" This module contains utilities for calculating search sensitivity
"""
import numpy
from . import bin_utils
def compute_search_efficiency_in_bins(
found, total, ndbins,
sim_to_bins_function=lambda sim: (sim.distance,)):
"""
Calculate search efficiency in the given ndbins.
The first dimension of ndbins must be bins over injected distance.
sim_to_bins_function must map an object to a tuple indexing the ndbins.
"""
bins = bin_utils.BinnedRatios(ndbins)
# increment the numerator and denominator with found / found+missed injs
[bins.incnumerator(sim_to_bins_function(sim)) for sim in found]
[bins.incdenominator(sim_to_bins_function(sim)) for sim in total]
# regularize by setting denoms to 1 to avoid nans
bins.regularize()
# efficiency array is the ratio
eff = bin_utils.BinnedArray(bin_utils.NDBins(ndbins), array=bins.ratio())
# compute binomial uncertainties in each bin
err_arr = numpy.sqrt(eff.array * (1-eff.array)/bins.denominator.array)
err = bin_utils.BinnedArray(bin_utils.NDBins(ndbins), array=err_arr)
return eff, err
def compute_search_volume_in_bins(found, total, ndbins, sim_to_bins_function):
"""
Calculate search sensitive volume by integrating efficiency in distance bins
No cosmological corrections are applied: flat space is assumed.
The first dimension of ndbins must be bins over injected distance.
sim_to_bins_function must maps an object to a tuple indexing the ndbins.
"""
eff, err = compute_search_efficiency_in_bins(
found, total, ndbins, sim_to_bins_function)
dx = ndbins[0].upper() - ndbins[0].lower()
r = ndbins[0].centres()
# volume and errors have one fewer dimension than the input NDBins
vol = bin_utils.BinnedArray(bin_utils.NDBins(ndbins[1:]))
errors = bin_utils.BinnedArray(bin_utils.NDBins(ndbins[1:]))
# integrate efficiency to obtain volume
vol.array = numpy.trapz(eff.array.T * 4. * numpy.pi * r**2, r, dx)
# propagate errors in eff to errors in V
errors.array = numpy.sqrt(
((4 * numpy.pi * r**2 * err.array.T * dx)**2).sum(axis=-1)
)
return vol, errors
def volume_to_distance_with_errors(vol, vol_err):
""" Return the distance and standard deviation upper and lower bounds
Parameters
----------
vol: float
vol_err: float
Returns
-------
dist: float
ehigh: float
elow: float
"""
dist = (vol * 3.0/4.0/numpy.pi) ** (1.0/3.0)
ehigh = ((vol + vol_err) * 3.0/4.0/numpy.pi) ** (1.0/3.0) - dist
delta = numpy.where(vol >= vol_err, vol - vol_err, 0)
elow = dist - (delta * 3.0/4.0/numpy.pi) ** (1.0/3.0)
return dist, ehigh, elow
def volume_montecarlo(found_d, missed_d, found_mchirp, missed_mchirp,
distribution_param, distribution, limits_param,
min_param=None, max_param=None):
"""
Compute sensitive volume and standard error via direct Monte Carlo integral
Injections should be made over a range of distances such that sensitive
volume due to signals closer than D_min is negligible, and efficiency at
distances above D_max is negligible
TODO : Replace this function by Collin's formula given in Usman et al .. ?
OR get that coded as a new function?
Parameters
-----------
found_d: numpy.ndarray
The distances of found injections
missed_d: numpy.ndarray
The distances of missed injections
found_mchirp: numpy.ndarray
Chirp mass of found injections
missed_mchirp: numpy.ndarray
Chirp mass of missed injections
distribution_param: string
Parameter D of the injections used to generate a distribution over
distance, may be 'distance', 'chirp_distance'.
distribution: string
form of the distribution over the parameter, may be
'log' (uniform in log D)
'uniform' (uniform in D)
'distancesquared' (uniform in D**2)
'volume' (uniform in D***3)
limits_param: string
Parameter Dlim specifying limits inside which injections were made
may be 'distance', 'chirp distance'
min_param: float
minimum value of Dlim at which injections were made; only used for
log distribution, then if None the minimum actually injected value
will be used
max_param: float
maximum value of Dlim out to which injections were made; if None
the maximum actually injected value will be used
Returns
--------
volume: float
Volume estimate
volume_error: float
The standard error in the volume
"""
d_power = {
'log' : 3.,
'uniform' : 2.,
'distancesquared' : 1.,
'volume' : 0.
}[distribution]
mchirp_power = {
'log' : 0.,
'uniform' : 5. / 6.,
'distancesquared' : 5. / 3.,
'volume' : 15. / 6.
}[distribution]
# establish maximum physical distance: first for chirp distance distribution
if limits_param == 'chirp_distance':
mchirp_standard_bns = 1.4 * 2.**(-1. / 5.)
all_mchirp = numpy.concatenate((found_mchirp, missed_mchirp))
max_mchirp = all_mchirp.max()
if max_param is not None:
# use largest actually injected mchirp for conversion
max_distance = max_param * \
(max_mchirp / mchirp_standard_bns)**(5. / 6.)
elif limits_param == 'distance':
if max_param is not None:
max_distance = max_param
else:
# if no max distance given, use max distance actually injected
max_distance = max(found_d.max(), missed_d.max())
else:
raise NotImplementedError("%s is not a recognized parameter"
% limits_param)
# volume of sphere
montecarlo_vtot = (4. / 3.) * numpy.pi * max_distance**3.
# arrays of weights for the MC integral
if distribution_param == 'distance':
found_weights = found_d ** d_power
missed_weights = missed_d ** d_power
elif distribution_param == 'chirp_distance':
# weight by a power of mchirp to rescale injection density to the
# target mass distribution
found_weights = found_d ** d_power * \
found_mchirp ** mchirp_power
missed_weights = missed_d ** d_power * \
missed_mchirp ** mchirp_power
else:
raise NotImplementedError("%s is not a recognized distance parameter"
% distribution_param)
all_weights = numpy.concatenate((found_weights, missed_weights))
# measured weighted efficiency is w_i for a found inj and 0 for missed
# MC integral is volume of sphere * (sum of found weights)/(sum of all weights)
# over injections covering the sphere
mc_weight_samples = numpy.concatenate((found_weights, 0 * missed_weights))
mc_sum = sum(mc_weight_samples)
if limits_param == 'distance':
mc_norm = sum(all_weights)
elif limits_param == 'chirp_distance':
# if injections are made up to a maximum chirp distance, account for
# extra missed injections that would occur when injecting up to
# maximum physical distance : this works out to a 'chirp volume' factor
mc_norm = sum(all_weights * (max_mchirp / all_mchirp) ** (5. / 2.))
# take out a constant factor
mc_prefactor = montecarlo_vtot / mc_norm
# count the samples
if limits_param == 'distance':
Ninj = len(mc_weight_samples)
elif limits_param == 'chirp_distance':
# find the total expected number after extending from maximum chirp
# dist up to maximum physical distance
if distribution == 'log':
# only need minimum distance in this one case
if min_param is not None:
min_distance = min_param * \
(numpy.min(all_mchirp) / mchirp_standard_bns) ** (5. / 6.)
else:
min_distance = min(numpy.min(found_d), numpy.min(missed_d))
logrange = numpy.log(max_distance / min_distance)
Ninj = len(mc_weight_samples) + (5. / 6.) * \
sum(numpy.log(max_mchirp / all_mchirp) / logrange)
else:
Ninj = sum((max_mchirp / all_mchirp) ** mchirp_power)
# sample variance of efficiency: mean of the square - square of the mean
mc_sample_variance = sum(mc_weight_samples ** 2.) / Ninj - \
(mc_sum / Ninj) ** 2.
# return MC integral and its standard deviation; variance of mc_sum scales
# relative to sample variance by Ninj (Bienayme' rule)
vol = mc_prefactor * mc_sum
vol_err = mc_prefactor * (Ninj * mc_sample_variance) ** 0.5
return vol, vol_err
def volume_binned_pylal(f_dist, m_dist, bins=15):
""" Compute the sensitive volume using a distance binned efficiency estimate
Parameters
-----------
f_dist: numpy.ndarray
The distances of found injections
m_dist: numpy.ndarray
The distances of missed injections
Returns
--------
volume: float
Volume estimate
volume_error: float
The standard error in the volume
"""
def sims_to_bin(sim):
return (sim, 0)
total = numpy.concatenate([f_dist, m_dist])
ndbins = bin_utils.NDBins([bin_utils.LinearBins(min(total), max(total), bins),
bin_utils.LinearBins(0., 1, 1)])
vol, verr = compute_search_volume_in_bins(f_dist, total, ndbins, sims_to_bin)
return vol.array[0], verr.array[0]
def volume_shell(f_dist, m_dist):
""" Compute the sensitive volume using sum over spherical shells.
Parameters
-----------
f_dist: numpy.ndarray
The distances of found injections
m_dist: numpy.ndarray
The distances of missed injections
Returns
--------
volume: float
Volume estimate
volume_error: float
The standard error in the volume
"""
f_dist.sort()
m_dist.sort()
distances = numpy.concatenate([f_dist, m_dist])
dist_sorting = distances.argsort()
distances = distances[dist_sorting]
low = 0
vol = 0
vol_err = 0
for i in range(len(distances)):
if i == len(distances) - 1:
break
high = (distances[i+1] + distances[i]) / 2
bin_width = high - low
if dist_sorting[i] < len(f_dist):
vol += 4 * numpy.pi * distances[i]**2.0 * bin_width
vol_err += (4 * numpy.pi * distances[i]**2.0 * bin_width)**2.0
low = high
vol_err = vol_err ** 0.5
return vol, vol_err
![swh spinner](/static/img/swh-spinner.gif)
Computing file changes ...