def test_compare_ecorr_likelihood(self):
"""Compare basis and kernel ecorr methods."""
selection = Selection(selections.nanograv_backends)
ef = white_signals.MeasurementNoise()
ec = white_signals.EcorrKernelNoise(selection=selection)
ec2 = gp_signals.EcorrBasisModel(selection=selection)
tm = gp_signals.TimingModel()
m = ef + ec + tm
m2 = ef + ec2 + tm
pta1 = signal_base.PTA([m(p) for p in self.psrs])
pta2 = signal_base.PTA([m2(p) for p in self.psrs])
params = parameter.sample(pta1.params)
l1 = pta1.get_lnlikelihood(params)
# need to translate some names for EcorrBasis
basis_params = {}
for parname, parval in params.items():
if "log10_ecorr" in parname:
toks = parname.split("_")
basisname = toks[0] + "_basis_ecorr_" + "_".join(toks[1:])
basis_params[basisname] = parval
params.update(basis_params)
l2 = pta2.get_lnlikelihood(params)
msg = "Likelihood mismatch between ECORR methods"
assert np.allclose(l1, l2), msg
def test_efac_backend(self):
"""Test that backend-efac signal returns correct covariance."""
# set up signal and parameters
efac = parameter.Uniform(0.1, 5)
selection = Selection(selections.by_backend)
ef = white_signals.MeasurementNoise(efac=efac, selection=selection)
efm = ef(self.psr)
# parameters
efacs = [1.3, 1.4, 1.5, 1.6]
params = {
"B1855+09_430_ASP_efac": efacs[0],
"B1855+09_430_PUPPI_efac": efacs[1],
"B1855+09_L-wide_ASP_efac": efacs[2],
"B1855+09_L-wide_PUPPI_efac": efacs[3],
}
# correct value
flags = ["430_ASP", "430_PUPPI", "L-wide_ASP", "L-wide_PUPPI"]
nvec0 = np.zeros_like(self.psr.toas)
for ct, flag in enumerate(np.unique(flags)):
ind = flag == self.psr.backend_flags
nvec0[ind] = efacs[ct]**2 * self.psr.toaerrs[ind]**2
# test
msg = "EFAC covariance incorrect."
assert np.all(efm.get_ndiag(params) == nvec0), msg
def __init__(self, psrs, params=None):
print('Initializing the model...')
efac = parameter.Constant()
equad = parameter.Constant()
ef = white_signals.MeasurementNoise(efac=efac)
eq = white_signals.EquadNoise(log10_equad=equad)
tm = gp_signals.TimingModel(use_svd=True)
s = eq + ef + tm
model = []
for p in psrs:
model.append(s(p))
self.pta = signal_base.PTA(model)
# set white noise parameters
if params is None:
print('No noise dictionary provided!...')
else:
self.pta.set_default_params(params)
self.psrs = psrs
self.params = params
self.Nmats = None
def white_noise_block(vary=False):
"""
Returns the white noise block of the model:
1. EFAC per backend/receiver system
2. EQUAD per backend/receiver system
3. ECORR per backend/receiver system
:param vary:
If set to true we vary these parameters
with uniform priors. Otherwise they are set to constants
with values to be set later.
"""
# define selection by observing backend
selection = selections.Selection(selections.by_backend)
# white noise parameters
if vary:
efac = parameter.Uniform(0.01, 10.0)
equad = parameter.Uniform(-8.5, -5)
ecorr = parameter.Uniform(-8.5, -5)
else:
efac = parameter.Constant()
equad = parameter.Constant()
ecorr = parameter.Constant()
# white noise signals
ef = white_signals.MeasurementNoise(efac=efac, selection=selection)
eq = white_signals.EquadNoise(log10_equad=equad, selection=selection)
ec = white_signals.EcorrKernelNoise(log10_ecorr=ecorr, selection=selection)
# combine signals
s = ef + eq + ec
return s
def add_efac(self):
"""Return dictionary containing EFAC white noise signal
attributes
:return: OrderedDict of EFAC signal
"""
efac_dct = dict()
efac = parameter.Uniform(0.001, 5.0)
ef = white_signals.MeasurementNoise(efac=efac,
selection=self.selection)
self.efac_sig = ef(self.psr)
for ii, param in enumerate(self.efac_sig.param_names):
Nvec = self.psr.toaerrs**2 * self.efac_sig._masks[ii]
newsignal = OrderedDict({
'type': 'efac',
'name': param,
'pmin': [0.001],
'pmax': [5.0],
'pstart': [1.0],
'interval': [True],
'numpars': 1,
'Nvec': Nvec
})
efac_dct.update({param: newsignal})
return efac_dct
def test_ephemeris(self):
"""Test physical-ephemeris delay, made three ways: from
marginalized GP, from coefficient-based GP, from
deterministic model."""
ef = white_signals.MeasurementNoise(
efac=parameter.Uniform(0.1, 5.0))
eph = gp_signals.FourierBasisCommonGP_physicalephem(
sat_orb_elements=None)
ephc = gp_signals.FourierBasisCommonGP_physicalephem(
sat_orb_elements=None, coefficients=True)
ephd = deterministic_signals.PhysicalEphemerisSignal(
sat_orb_elements=False)
model = ef + eph
modelc = ef + ephc
modeld = ef + ephd
pta = signal_base.PTA([model(self.psr), model(self.psr2)])
ptac = signal_base.PTA([modelc(self.psr), modelc(self.psr2)])
ptad = signal_base.PTA([modeld(self.psr), modeld(self.psr2)])
cf = 1e-3 * np.random.randn(11)
cf[0] = 1e-5 # this is more sensitive to linearity
bs = pta.get_basis()
da = [np.dot(bs[0], cf), np.dot(bs[1], cf)]
params = {
"B1855+09_efac": 1,
"B1937+21_efac": 1,
"B1855+09_phys_ephem_gp_coefficients": cf,
"B1937+21_phys_ephem_gp_coefficients": cf,
}
db = ptac.get_delay(params=params)
dparams = {
"B1855+09_efac": 1,
"B1937+21_efac": 1,
"frame_drift_rate": cf[0],
"d_jupiter_mass": cf[1],
"d_saturn_mass": cf[2],
"d_uranus_mass": cf[3],
"d_neptune_mass": cf[4],
"jup_orb_elements": cf[5:],
}
dc = ptad.get_delay(params=dparams)
msg = "Reconstructed ephemeris signals differ!"
assert np.allclose(da[0], db[0]), msg
assert np.allclose(da[1], db[1]), msg
# we don't expect an exact match since we are linearizing
assert np.allclose(da[0], dc[0], atol=1e-3), msg
assert np.allclose(da[1], dc[1], atol=1e-3), msg
def efac(self,option="by_backend"):
"""
EFAC signal: multiplies ToA variance by EFAC**2, where ToA variance
are diagonal components of the Likelihood covariance matrix.
"""
if option not in selections.__dict__.keys():
raise ValueError('EFAC option must be Enterprise selection function name')
se=selections.Selection(selections.__dict__[option])
efacpr = interpret_white_noise_prior(self.params.efac)
efs = white_signals.MeasurementNoise(efac=efacpr,selection=se)
return efs
def test_vector_parameter_like(self):
"""Test vector parameter in a likelihood"""
# white noise
efac = parameter.Uniform(0.5, 2)
ef = white_signals.MeasurementNoise(efac=efac)
# red noise
nf = 3
spec = free_spectrum(log10_rho=parameter.Uniform(-20, -10, size=nf))
rn = gp_signals.FourierBasisGP(spec, components=nf)
# timing model
tm = gp_signals.TimingModel()
# combined signal
s = ef + rn + tm
# PTA
pta = signal_base.PTA([s(self.psr)])
# parameters
xs = np.hstack(p.sample() for p in pta.params)
params = {
"B1855+09_red_noise_log10_rho": xs[1:],
"B1855+09_efac": xs[0]
}
# test log likelihood
msg = "Likelihoods do not match"
assert pta.get_lnlikelihood(xs) == pta.get_lnlikelihood(params), msg
# test log prior
msg = "Priors do not match"
assert pta.get_lnprior(xs) == pta.get_lnprior(params), msg
# test prior value
prior = 1 / (2 - 0.5) * (1 / 10)**3
msg = "Prior value incorrect."
assert np.allclose(pta.get_lnprior(xs), np.log(prior)), msg
# test PTA level parameter names
pnames = [
"B1855+09_efac",
"B1855+09_red_noise_log10_rho_0",
"B1855+09_red_noise_log10_rho_1",
"B1855+09_red_noise_log10_rho_2",
]
msg = "Incorrect parameter names"
assert pta.param_names == pnames
def test_formalism(self):
# create marginalized model
ef = white_signals.MeasurementNoise(
efac=parameter.Uniform(0.1, 5.0))
tm = gp_signals.TimingModel()
ec = gp_signals.EcorrBasisModel(
log10_ecorr=parameter.Uniform(-10, -5))
pl = utils.powerlaw(log10_A=parameter.Uniform(-18, -12),
gamma=parameter.Uniform(1, 7))
rn = gp_signals.FourierBasisGP(spectrum=pl, components=10)
model = ef + tm + ec + rn
pta = signal_base.PTA([model(self.psr)])
# create hierarchical model
tmc = gp_signals.TimingModel(coefficients=True)
ecc = gp_signals.EcorrBasisModel(log10_ecorr=parameter.Uniform(
-10, -5),
coefficients=True)
rnc = gp_signals.FourierBasisGP(spectrum=pl,
components=10,
coefficients=True)
modelc = ef + tmc + ecc + rnc
ptac = signal_base.PTA([modelc(self.psr)])
ps = {
"B1855+09_efac": 1,
"B1855+09_basis_ecorr_log10_ecorr": -6,
"B1855+09_red_noise_log10_A": -14,
"B1855+09_red_noise_gamma": 3,
}
psc = utils.get_coefficients(pta, ps)
d1 = ptac.get_delay(psc)[0]
d2 = (np.dot(pta.pulsarmodels[0].signals[1].get_basis(ps),
psc["B1855+09_linear_timing_model_coefficients"]) +
np.dot(pta.pulsarmodels[0].signals[2].get_basis(ps),
psc["B1855+09_basis_ecorr_coefficients"]) +
np.dot(pta.pulsarmodels[0].signals[3].get_basis(ps),
psc["B1855+09_red_noise_coefficients"]))
msg = "Implicit and explicit PTA delays are different."
assert np.allclose(d1, d2), msg
l1 = pta.get_lnlikelihood(ps)
l2 = ptac.get_lnlikelihood(psc)
# I don't know how to integrate l2 to match l1...
msg = "Marginal and hierarchical likelihoods should be different."
assert l1 != l2, msg
def initialize_pta_sim(psrs, fgw):
# continuous GW signal
s = models.cw_block_circ(log10_fgw=np.log10(fgw), psrTerm=True)
# white noise
efac = parameter.Constant(1.0)
s += white_signals.MeasurementNoise(efac=efac)
# linearized timing model
s += gp_signals.TimingModel(use_svd=True)
model = [s(psr) for psr in psrs]
pta = signal_base.PTA(model)
return pta
def test_efac(self):
"""Test that efac signal returns correct covariance."""
# set up signal and parameters
efac = parameter.Uniform(0.1, 5)
ef = white_signals.MeasurementNoise(efac=efac)
efm = ef(self.psr)
# parameters
efac = 1.5
params = {"B1855+09_efac": efac}
# correct value
nvec0 = efac**2 * self.psr.toaerrs**2
# test
msg = "EFAC covariance incorrect."
assert np.all(efm.get_ndiag(params) == nvec0), msg
def initialize_pta_sim(psrs, fgw,
inc_efac=True, inc_equad=False, inc_ecorr=False,
selection=None,
inc_red_noise=False, noisedict=None):
# continuous GW signal
s = models.cw_block_circ(log10_fgw=np.log10(fgw), psrTerm=True)
# linearized timing model
s += gp_signals.TimingModel(use_svd=True)
# white noise
if selection == 'backend':
selection = selections.Selection(selections.by_backend)
if inc_efac:
efac = parameter.Constant()
s += white_signals.MeasurementNoise(efac=efac, selection=selection)
if inc_equad:
equad = parameter.Constant()
s += white_signals.EquadNoise(log10_equad=equad,
selection=selection)
if inc_ecorr:
ecorr = parameter.Constant()
s += gp_signals.EcorrBasisModel(log10_ecorr=ecorr,
selection=selection)
if inc_red_noise:
log10_A = parameter.Constant()
gamma = parameter.Constant()
pl = utils.powerlaw(log10_A=log10_A, gamma=gamma)
s += gp_signals.FourierBasisGP(pl, components=30)
model = [s(psr) for psr in psrs]
pta = signal_base.PTA(model)
# set white noise parameters
if noisedict is None:
print('No noise dictionary provided!...')
else:
pta.set_default_params(noisedict)
return pta
def test_add_efac_equad(self):
"""Test that addition of efac and equad signal returns
correct covariance.
"""
# set up signals
efac = parameter.Uniform(0.1, 5)
ef = white_signals.MeasurementNoise(efac=efac)
equad = parameter.Uniform(-10, -5)
eq = white_signals.EquadNoise(log10_equad=equad)
s = ef + eq
m = s(self.psr)
# set parameters
efac = 1.5
equad = -6.4
params = {"B1855+09_efac": efac, "B1855+09_log10_equad": equad}
# correct value
nvec0 = efac**2 * self.psr.toaerrs**2
nvec0 += 10**(2 * equad) * np.ones_like(self.psr.toas)
# test
msg = "EFAC/EQUAD covariance incorrect."
assert np.all(m.get_ndiag(params) == nvec0), msg
def test_wideband(self):
ms = white_signals.MeasurementNoise(selection=Selection(selections.by_backend))
dm = gp_signals.WidebandTimingModel(
dmefac=parameter.Uniform(0.9, 1.1),
dmefac_selection=Selection(selections.by_backend),
dmjump=parameter.Normal(0, 1),
dmjump_selection=Selection(selections.by_frontend),
)
model = ms + dm
pta = signal_base.PTA([model(self.psr)])
ps = parameter.sample(pta.params)
pta.get_lnlikelihood(ps)
dmtiming = pta.pulsarmodels[0].signals[1]
msg = "DMEFAC masks do not cover the data."
assert np.all(sum(dmtiming._dmefac_masks) == 1), msg
msg = "DMJUMP masks do not cover the data."
assert np.all(sum(dmtiming._dmjump_masks) == 1), msg
# start with zero DMEFAC and DMJUMP
p0 = {par.name: (1 if "dmefac" in par.name else 0) for par in dmtiming.params}
pta.get_lnlikelihood(params=p0)
phi0 = dmtiming.get_phi(params=p0)
dl0 = dmtiming.get_delay(params=p0)
dm_flags, dme_flags = np.array(self.psr.flags["pp_dm"], "d"), np.array(self.psr.flags["pp_dme"], "d")
delays = np.zeros_like(self.psr.toas)
check = 0
for index, par in enumerate(self.psr.fitpars):
if "DMX" not in par:
msg = "Problem with unbound timing parameters"
assert phi0[index] == 1e40, msg
else:
dmx = self.psr.dmx[par]
which = (dmx["DMXR1"] <= (self.psr.stoas / 86400)) & ((self.psr.stoas / 86400) < dmx["DMXR2"])
check += which
avgdm = np.sum(dm_flags[which] / dme_flags[which] ** 2) / np.sum(1.0 / dme_flags[which] ** 2)
vardm = 1.0 / np.sum(1.0 / dme_flags[which] ** 2)
msg = "Priors do not match"
assert np.allclose(vardm, phi0[index]), msg
delays[which] = (avgdm - self.psr.dm - dmx["DMX"]) / (2.41e-4 * self.psr.freqs[which] ** 2)
msg = "Not all TOAs are covered by DMX"
assert np.all(check == 1)
msg = "Delays do not match"
assert np.allclose(dl0, delays), msg
# sample DMEFACs randomly
p1 = {par.name: (parameter.sample(par)[par.name] if "dmefac" in par.name else 0) for par in dmtiming.params}
pta.get_lnlikelihood(params=p1)
phi1 = dmtiming.get_phi(params=p1)
dl1 = dmtiming.get_delay(params=p1)
sel = Selection(selections.by_backend)(self.psr)
msg = "Problem making selection"
assert np.all(sum(m for m in sel.masks.values()) == 1), msg
dme_flags_var = dme_flags.copy()
for key, mask in sel.masks.items():
dmefac = p1["J1832-0836_" + key + "_dmefac"]
dme_flags_var[mask] *= dmefac
for index, par in enumerate(self.psr.fitpars):
if "DMX" not in par:
msg = "Problem with unbound timing parameters"
assert phi1[index] == 1e40, msg
else:
dmx = self.psr.dmx[par]
which = (dmx["DMXR1"] <= (self.psr.stoas / 86400)) & ((self.psr.stoas / 86400) < dmx["DMXR2"])
avgdm = np.sum(dm_flags[which] / dme_flags_var[which] ** 2) / np.sum(1.0 / dme_flags_var[which] ** 2)
vardm = 1.0 / np.sum(1.0 / dme_flags_var[which] ** 2)
msg = "Priors do not match"
assert np.allclose(vardm, phi1[index]), msg
delays[which] = (avgdm - self.psr.dm - dmx["DMX"]) / (2.41e-4 * self.psr.freqs[which] ** 2)
msg = "Delays do not match"
assert np.allclose(dl1, delays), msg
# GW parameters (initialize with names here to use parameters in common across pulsars)
log10_A_gw_1 = parameter.Uniform(-18, -13)('zlog10_A_gw')
gamma_gw_1 = parameter.Constant(13 / 3)('zgamma_gw')
# Second GW parameters
log10_A_gw_2 = parameter.Uniform(-18, -13)('zlog10_A_other_gw')
gamma_gw_2 = parameter.Constant(7 / 3)('zgamma_other_gw')
##### Set up signals #####
# timing model
tm = gp_signals.TimingModel()
# white noise
ef = white_signals.MeasurementNoise(efac=efac, selection=selection)
eq = white_signals.EquadNoise(log10_equad=log10_equad, selection=selection)
ec = white_signals.EcorrKernelNoise(log10_ecorr=log10_ecorr,
selection=selection)
# red noise (powerlaw with 30 frequencies)
pl = utils.powerlaw(log10_A=red_noise_log10_A, gamma=red_noise_gamma)
rn = gp_signals.FourierBasisGP(spectrum=pl, components=30, Tspan=Tspan)
cpl_1 = utils.powerlaw(log10_A=log10_A_gw_1, gamma=gamma_gw_1)
cpl_2 = utils.powerlaw(log10_A=log10_A_gw_2, gamma=gamma_gw_2)
#Common red noise process with no correlations
crn_1 = gp_signals.FourierBasisGP(spectrum=cpl_1,
components=30,
Tspan=Tspan,
parfiles = sorted(glob.glob('data/*.par'))
timfiles = sorted(glob.glob('data/*.tim'))
psrs = []
for p, t in zip(parfiles, timfiles):
psr = Pulsar(p, t)
psrs.append(psr)
##### parameters and priors #####
# Uniform prior on EFAC
# efac = parameter.Uniform(0.1, 5.0)
efac = parameter.LinearExp(0.1, 5.0)
# white noise
ef = white_signals.MeasurementNoise(efac=efac)
# timing model
tm = gp_signals.TimingModel()
# full model is sum of components
model = ef + tm
# initialize PTA
pta = signal_base.PTA([model(psrs[0])])
priors = bilby_warp.get_bilby_prior_dict(pta)
print(priors)
parameters = dict.fromkeys(priors.keys())
likelihood = bilby_warp.PTABilbyLikelihood(pta, parameters)
label = 'test_bilby'
def gwb_ul(psrs_cut, num_points):
# find the maximum time span to set GW frequency sampling
tmin = [p.toas.min() for p in psrs_cut]
tmax = [p.toas.max() for p in psrs_cut]
Tspan = np.max(tmax) - np.min(tmin)
# define selection by observing backend
selection = selections.Selection(selections.by_backend)
# white noise parameters
# we set these ourselves so we know the most likely values!
efac = parameter.Constant(1)
# quad = parameter.Constant(0)
# ecorr = parameter.Constant(0)
# red noise parameters
log10_A = parameter.LinearExp(-20, -11)
gamma = parameter.Uniform(0, 7)
# GW parameters (initialize with names here to use parameters in common across pulsars)
log10_A_gw = parameter.LinearExp(-18, -12)('log10_A_gw')
gamma_gw = parameter.Constant(4.33)('gamma_gw')
# white noise
ef = white_signals.MeasurementNoise(efac=efac, selection=selection)
# eq = white_signals.EquadNoise(log10_equad=equad, selection=selection)
# ec = white_signals.EcorrKernelNoise(log10_ecorr=ecorr, selection=selection)
# red noise (powerlaw with 30 frequencies)
pl = utils.powerlaw(log10_A=log10_A, gamma=gamma)
rn = gp_signals.FourierBasisGP(spectrum=pl, components=30, Tspan=Tspan)
# gwb (no spatial correlations)
cpl = utils.powerlaw(log10_A=log10_A_gw, gamma=gamma_gw)
gw = gp_signals.FourierBasisGP(spectrum=cpl,
components=30,
Tspan=Tspan,
name='gw')
# timing model
tm = gp_signals.TimingModel(
use_svd=True) # stabilizing timing model design matrix with SVD
s = ef + rn + gw + tm
# intialize PTA
models = []
for p in psrs_cut:
models.append(s(p))
pta = signal_base.PTA(models)
outDir = './chains/psrs/{0}'.format(psrs_cut[0].name)
sample = sampler.setup_sampler(pta, outdir=outDir)
x0 = np.hstack([p.sample() for p in pta.params])
# sampler for N steps
N = int(
num_points) # normally, we would use 5e6 samples (this will save time)
sample.sample(
x0,
N,
SCAMweight=30,
AMweight=15,
DEweight=50,
)
chain = np.loadtxt(os.path.join(outDir, 'chain_1.txt'))
pars = np.loadtxt(outDir + '/pars.txt', dtype=np.unicode_)
ind = list(pars).index('log10_A_gw')
UL, unc = model_utils.ul(chain[:, ind])
return UL, unc
def test_conditional_gp(self):
ef = white_signals.MeasurementNoise(efac=parameter.Uniform(0.1, 5.0))
tm = gp_signals.TimingModel()
ec = gp_signals.EcorrBasisModel(log10_ecorr=parameter.Uniform(-10, -5))
pl = utils.powerlaw(log10_A=parameter.Uniform(-18, -12),
gamma=parameter.Uniform(1, 7))
rn = gp_signals.FourierBasisGP(spectrum=pl,
components=10,
combine=False)
model = ef + tm + ec + rn
pta = signal_base.PTA([model(self.psr), model(self.psr2)])
p0 = {
"B1855+09_basis_ecorr_log10_ecorr": -6.051740765663904,
"B1855+09_efac": 2.9027266737466095,
"B1855+09_red_noise_gamma": 6.9720332277819725,
"B1855+09_red_noise_log10_A": -16.749192700991543,
"B1937+21_basis_ecorr_log10_ecorr": -9.726747733721872,
"B1937+21_efac": 3.959178240268702,
"B1937+21_red_noise_gamma": 2.9030772884814797,
"B1937+21_red_noise_log10_A": -17.978562921948992,
}
c = utils.ConditionalGP(pta)
cmean = c.get_mean_coefficients(p0)
# build index for the global coefficient vector
idx, ntot = {}, 0
for l, v in cmean.items():
idx[l] = slice(ntot, ntot + len(v))
ntot = ntot + len(v)
# repeat the computation using the common-signal formalism
TNrs = pta.get_TNr(p0)
TNTs = pta.get_TNT(p0)
phiinvs = pta.get_phiinv(p0, logdet=False, method="cliques")
TNr = np.concatenate(TNrs)
Sigma = sps.block_diag(TNTs, "csc") + sps.block_diag(
[np.diag(phiinvs[0]), np.diag(phiinvs[1])])
ch = cholesky(Sigma)
mn = ch(TNr)
iSigma = sps.linalg.inv(Sigma)
# check mean values
msg = "Conditional GP coefficient value does not match"
for l, v in cmean.items():
assert np.allclose(mn[idx[l]], v, atol=1e-4, rtol=1e-4), msg
# check variances
par = "B1937+21_linear_timing_model_coefficients"
c1 = np.cov(
np.array([cs[par] for cs in c.sample_coefficients(p0, n=10000)]).T)
c2 = iSigma[idx[par], idx[par]].toarray().T
msg = "Conditional GP coefficient variance does not match"
assert np.allclose(c1, c2, atol=1e-4, rtol=1e-4), msg
# check mean processes
proc = "B1937+21_linear_timing_model"
p1 = c.get_mean_processes(p0)[proc]
p2 = np.dot(pta["B1937+21"]["linear_timing_model"].get_basis(),
mn[idx[par]])
msg = "Conditional GP time series does not match"
assert np.allclose(p1, p2, atol=1e-4, rtol=1e-4), msg
# check mean of sampled processes
p2 = np.mean(np.array(
[pc[proc] for pc in c.sample_processes(p0, n=1000)]),
axis=0)
msg = "Mean of sampled conditional GP processes does not match"
assert np.allclose(p1, p2, atol=1e-4, rtol=1e-4)
# now try with a common process
crn = gp_signals.FourierBasisCommonGP(spectrum=pl,
orf=utils.hd_orf(),
components=10,
combine=False)
model = ef + tm + ec + crn
pta = signal_base.PTA([model(self.psr), model(self.psr2)])
p0 = {
"B1855+09_basis_ecorr_log10_ecorr": -5.861847220080768,
"B1855+09_efac": 4.588342210948306,
"B1937+21_basis_ecorr_log10_ecorr": -9.151872649912377,
"B1937+21_efac": 0.8947815819783302,
"common_fourier_gamma": 6.638289750637263,
"common_fourier_log10_A": -15.68180643904114,
}
c = utils.ConditionalGP(pta)
cmean = c.get_mean_coefficients(p0)
idx, ntot = {}, 0
for l, v in cmean.items():
idx[l] = slice(ntot, ntot + len(v))
ntot = ntot + len(v)
TNrs = pta.get_TNr(p0)
TNTs = pta.get_TNT(p0)
phiinvs = pta.get_phiinv(p0, logdet=False, method="cliques")
TNr = np.concatenate(TNrs)
Sigma = sps.block_diag(TNTs, "csc") + sps.csc_matrix(phiinvs)
ch = cholesky(Sigma)
mn = ch(TNr)
msg = "Conditional GP coefficient value does not match for common GP"
for l, v in cmean.items():
assert np.allclose(mn[idx[l]], v)
def model_simple(psrs, psd='powerlaw', efac=False, n_gwbfreqs=30,
components=30, freqs=None,
vary_gamma=False, upper_limit=False, bayesephem=False,
select='backend', red_noise=False, Tspan=None, hd_orf=False,
rn_dropout=False, dp_threshold=0.5):
"""
Reads in list of enterprise Pulsar instance and returns a PTA
instantiated with the most simple model allowable for enterprise:
per pulsar:
1. fixed EFAC per backend/receiver system at 1.0
2. Linear timing model.
3. Red noise modeled as a power-law with
30 sampling frequencies. Default=False
global:
1.Common red noise modeled with user defined PSD with
30 sampling frequencies. Available PSDs are
['powerlaw', 'turnover' 'spectrum']
2. Optional physical ephemeris modeling.
:param psd:
PSD to use for common red noise signal. Available options
are ['powerlaw', 'turnover' 'spectrum']. 'powerlaw' is default
value.
:param gamma_common:
Fixed common red process spectral index value. By default we
vary the spectral index over the range [0, 7].
:param upper_limit:
Perform upper limit on common red noise amplitude. By default
this is set to False. Note that when performing upper limits it
is recommended that the spectral index also be fixed to a specific
value.
:param bayesephem:
Include BayesEphem model. Set to False by default
"""
amp_prior = 'uniform' if upper_limit else 'log-uniform'
# find the maximum time span to set GW frequency sampling
if Tspan is None:
Tspan = model_utils.get_tspan(psrs)
# timing model
model = gp_signals.TimingModel()
#Only White Noise is EFAC set to 1.0
selection = selections.Selection(selections.by_backend)
if efac:
ef = parameter.Uniform(0.1,10.0)
else:
ef = parameter.Constant(1.00)
model += white_signals.MeasurementNoise(efac=ef, selection=selection)
# common red noise block
if upper_limit:
log10_A_gw = parameter.LinearExp(-18,-12)('gw_log10_A')
else:
log10_A_gw = parameter.Uniform(-18,-12)('gw_log10_A')
if vary_gamma:
gamma_gw = parameter.Uniform(0,7)('gw_gamma')
else:
gamma_gw = parameter.Constant(4.33)('gw_gamma')
pl = signal_base.Function(utils.powerlaw, log10_A=log10_A_gw,
gamma=gamma_gw)
if hd_orf:
if freqs is None:
gw = gp_signals.FourierBasisCommonGP(spectrum=pl,
orf=utils.hd_orf(),
components=n_gwbfreqs,
Tspan=Tspan,
name='gw')
else:
gw = gp_signals.FourierBasisCommonGP(spectrum=pl,
orf=utils.hd_orf(),
modes=freqs,
name='gw')
model += gw
else:
if freqs is None:
crn = gp_signals.FourierBasisGP(spectrum=pl, components=n_gwbfreqs,
Tspan=Tspan, name='gw')
else:
crn = gp_signals.FourierBasisGP(spectrum=pl, modes=freqs,
name='gw')
model += crn
if red_noise and rn_dropout:
if amp_prior == 'uniform':
log10_A = parameter.LinearExp(-20, -11)
elif amp_prior == 'log-uniform':
log10_A = parameter.Uniform(-20, -11)
else:
log10_A = parameter.Uniform(-20, -11)
gamma = parameter.Uniform(0, 7)
k_drop = parameter.Uniform(0, 1)
if dp_threshold == 6.0:
dp_threshold = parameter.Uniform(0,1)('k_threshold')
pl = dropout.dropout_powerlaw(log10_A=log10_A, gamma=gamma,
k_drop=k_drop, k_threshold=dp_threshold)
rn = gp_signals.FourierBasisGP(pl, components=components,
Tspan=Tspan, name='red_noise')
model += rn
elif red_noise:
# red noise
model += models.red_noise_block(prior=amp_prior, Tspan=Tspan,
components=components)
# ephemeris model
if bayesephem:
model += deterministic_signals.PhysicalEphemerisSignal(use_epoch_toas=True)
# set up PTA
pta = signal_base.PTA([model(p) for p in psrs])
return pta
def white_noise_block(vary=False,
inc_ecorr=False,
gp_ecorr=False,
efac1=False,
select='backend',
name=None):
"""
Returns the white noise block of the model:
1. EFAC per backend/receiver system
2. EQUAD per backend/receiver system
3. ECORR per backend/receiver system
:param vary:
If set to true we vary these parameters
with uniform priors. Otherwise they are set to constants
with values to be set later.
:param inc_ecorr:
include ECORR, needed for NANOGrav channelized TOAs
:param gp_ecorr:
whether to use the Gaussian process model for ECORR
:param efac1:
use a strong prior on EFAC = Normal(mu=1, stdev=0.1)
"""
if select == 'backend':
# define selection by observing backend
backend = selections.Selection(selections.by_backend)
# define selection by nanograv backends
backend_ng = selections.Selection(selections.nanograv_backends)
else:
# define no selection
backend = selections.Selection(selections.no_selection)
# white noise parameters
if vary:
if efac1:
efac = parameter.Normal(1.0, 0.1)
else:
efac = parameter.Uniform(0.01, 10.0)
equad = parameter.Uniform(-8.5, -5)
if inc_ecorr:
ecorr = parameter.Uniform(-8.5, -5)
else:
efac = parameter.Constant()
equad = parameter.Constant()
if inc_ecorr:
ecorr = parameter.Constant()
# white noise signals
ef = white_signals.MeasurementNoise(efac=efac,
selection=backend,
name=name)
eq = white_signals.EquadNoise(log10_equad=equad,
selection=backend,
name=name)
if inc_ecorr:
if gp_ecorr:
if name is None:
ec = gp_signals.EcorrBasisModel(log10_ecorr=ecorr,
selection=backend_ng)
else:
ec = gp_signals.EcorrBasisModel(log10_ecorr=ecorr,
selection=backend_ng,
name=name)
else:
ec = white_signals.EcorrKernelNoise(log10_ecorr=ecorr,
selection=backend_ng,
name=name)
# combine signals
if inc_ecorr:
s = ef + eq + ec
elif not inc_ecorr:
s = ef + eq
return s
def test_pta(self):
# get parameters from PAL2 style noise files
params = get_noise_from_pal2(datadir + "/B1855+09_noise.txt")
params2 = get_noise_from_pal2(datadir + "/J1909-3744_noise.txt")
params.update(params2)
# setup basic model
efac = parameter.Constant()
equad = parameter.Constant()
ecorr = parameter.Constant()
log10_A = parameter.Constant()
gamma = parameter.Constant()
selection = Selection(selections.by_backend)
ms = white_signals.MeasurementNoise(efac=efac,
log10_t2equad=equad,
selection=selection)
ec = white_signals.EcorrKernelNoise(log10_ecorr=ecorr,
selection=selection)
pl = utils.powerlaw(log10_A=log10_A, gamma=gamma)
rn = gp_signals.FourierBasisGP(pl)
s = ms + ec + rn
pta = s(self.psrs[0]) + s(self.psrs[1])
# set parameters
pta.set_default_params(params)
# get parameters
efacs, equads, ecorrs, log10_A, gamma = [], [], [], [], []
for pname in [p.name for p in self.psrs]:
efacs.append([
params[key] for key in sorted(params.keys())
if "efac" in key and pname in key
])
equads.append([
params[key] for key in sorted(params.keys())
if "equad" in key and pname in key
])
ecorrs.append([
params[key] for key in sorted(params.keys())
if "ecorr" in key and pname in key
])
log10_A.append(params["{}_red_noise_log10_A".format(pname)])
gamma.append(params["{}_red_noise_gamma".format(pname)])
# correct value
tflags = [sorted(list(np.unique(p.backend_flags))) for p in self.psrs]
cfs, logdets, phis = [], [], []
for ii, (psr, flags) in enumerate(zip(self.psrs, tflags)):
nvec0 = np.zeros_like(psr.toas)
for ct, flag in enumerate(flags):
ind = psr.backend_flags == flag
nvec0[ind] = efacs[ii][ct]**2 * (
psr.toaerrs[ind]**2 +
10**(2 * equads[ii][ct]) * np.ones(np.sum(ind)))
# get the basis
bflags = psr.backend_flags
Umats = []
for flag in np.unique(bflags):
mask = bflags == flag
Umats.append(
utils.create_quantization_matrix(psr.toas[mask])[0])
nepoch = sum(U.shape[1] for U in Umats)
U = np.zeros((len(psr.toas), nepoch))
jvec = np.zeros(nepoch)
netot = 0
for ct, flag in enumerate(np.unique(bflags)):
mask = bflags == flag
nn = Umats[ct].shape[1]
U[mask, netot:nn + netot] = Umats[ct]
jvec[netot:nn + netot] = 10**(2 * ecorrs[ii][ct])
netot += nn
# get covariance matrix
cov = np.diag(nvec0) + np.dot(U * jvec[None, :], U.T)
cf = sl.cho_factor(cov)
logdet = np.sum(2 * np.log(np.diag(cf[0])))
cfs.append(cf)
logdets.append(logdet)
F, f2 = utils.createfourierdesignmatrix_red(psr.toas, nmodes=20)
phi = utils.powerlaw(f2, log10_A=log10_A[ii], gamma=gamma[ii])
phis.append(phi)
# tests
Ns = pta.get_ndiag(params)
pphis = pta.get_phi(params)
pphiinvs = pta.get_phiinv(params)
Ts = pta.get_basis(params)
zipped = zip(logdets, cfs, phis, self.psrs, Ns, pphis, pphiinvs, Ts)
for logdet, cf, phi, psr, N, pphi, pphiinv, T in zipped:
msg = "EFAC/ECORR logdet incorrect."
assert np.allclose(N.solve(psr.residuals, logdet=True)[1],
logdet,
rtol=1e-10), msg
msg = "EFAC/ECORR D1 solve incorrect."
assert np.allclose(N.solve(psr.residuals),
sl.cho_solve(cf, psr.residuals),
rtol=1e-10), msg
msg = "EFAC/ECORR 1D1 solve incorrect."
assert np.allclose(
N.solve(psr.residuals, left_array=psr.residuals),
np.dot(psr.residuals, sl.cho_solve(cf, psr.residuals)),
rtol=1e-10,
), msg
msg = "EFAC/ECORR 2D1 solve incorrect."
assert np.allclose(N.solve(psr.residuals, left_array=T),
np.dot(T.T, sl.cho_solve(cf, psr.residuals)),
rtol=1e-10), msg
msg = "EFAC/ECORR 2D2 solve incorrect."
assert np.allclose(N.solve(T, left_array=T),
np.dot(T.T, sl.cho_solve(cf, T)),
rtol=1e-10), msg
# spectrum test
msg = "Spectrum incorrect for GP Fourier signal."
assert np.all(pphi == phi), msg
# inverse spectrum test
msg = "Spectrum inverse incorrect for GP Fourier signal."
assert np.all(pphiinv == 1 / phi), msg
def test_single_pulsar(self):
# get parameters from PAL2 style noise files
params = get_noise_from_pal2(datadir + "/B1855+09_noise.txt")
# setup basic model
efac = parameter.Constant()
equad = parameter.Constant()
ecorr = parameter.Constant()
log10_A = parameter.Constant()
gamma = parameter.Constant()
selection = Selection(selections.by_backend)
ms = white_signals.MeasurementNoise(efac=efac,
log10_t2equad=equad,
selection=selection)
ec = white_signals.EcorrKernelNoise(log10_ecorr=ecorr,
selection=selection)
pl = utils.powerlaw(log10_A=log10_A, gamma=gamma)
rn = gp_signals.FourierBasisGP(pl)
s = ms + ec + rn
m = s(self.psrs[0])
# set parameters
m.set_default_params(params)
# get parameters
efacs = [params[key] for key in sorted(params.keys()) if "efac" in key]
equads = [
params[key] for key in sorted(params.keys()) if "equad" in key
]
ecorrs = [
params[key] for key in sorted(params.keys()) if "ecorr" in key
]
log10_A = params["B1855+09_red_noise_log10_A"]
gamma = params["B1855+09_red_noise_gamma"]
# correct value
flags = ["430_ASP", "430_PUPPI", "L-wide_ASP", "L-wide_PUPPI"]
nvec0 = np.zeros_like(self.psrs[0].toas)
for ct, flag in enumerate(np.unique(flags)):
ind = flag == self.psrs[0].backend_flags
nvec0[ind] = efacs[ct]**2 * (
self.psrs[0].toaerrs[ind]**2 +
10**(2 * equads[ct]) * np.ones(np.sum(ind)))
# get the basis
bflags = self.psrs[0].backend_flags
Umats = []
for flag in np.unique(bflags):
mask = bflags == flag
Umats.append(
utils.create_quantization_matrix(self.psrs[0].toas[mask])[0])
nepoch = sum(U.shape[1] for U in Umats)
U = np.zeros((len(self.psrs[0].toas), nepoch))
jvec = np.zeros(nepoch)
netot = 0
for ct, flag in enumerate(np.unique(bflags)):
mask = bflags == flag
nn = Umats[ct].shape[1]
U[mask, netot:nn + netot] = Umats[ct]
jvec[netot:nn + netot] = 10**(2 * ecorrs[ct])
netot += nn
# get covariance matrix
cov = np.diag(nvec0) + np.dot(U * jvec[None, :], U.T)
cf = sl.cho_factor(cov)
logdet = np.sum(2 * np.log(np.diag(cf[0])))
# test
msg = "EFAC/ECORR logdet incorrect."
N = m.get_ndiag(params)
assert np.allclose(N.solve(self.psrs[0].residuals, logdet=True)[1],
logdet,
rtol=1e-10), msg
msg = "EFAC/ECORR D1 solve incorrect."
assert np.allclose(N.solve(self.psrs[0].residuals),
sl.cho_solve(cf, self.psrs[0].residuals),
rtol=1e-10), msg
msg = "EFAC/ECORR 1D1 solve incorrect."
assert np.allclose(
N.solve(self.psrs[0].residuals, left_array=self.psrs[0].residuals),
np.dot(self.psrs[0].residuals,
sl.cho_solve(cf, self.psrs[0].residuals)),
rtol=1e-10,
), msg
msg = "EFAC/ECORR 2D1 solve incorrect."
T = m.get_basis(params)
assert np.allclose(
N.solve(self.psrs[0].residuals, left_array=T),
np.dot(T.T, sl.cho_solve(cf, self.psrs[0].residuals)),
rtol=1e-10,
), msg
msg = "EFAC/ECORR 2D2 solve incorrect."
assert np.allclose(N.solve(T, left_array=T),
np.dot(T.T, sl.cho_solve(cf, T)),
rtol=1e-10), msg
F, f2 = utils.createfourierdesignmatrix_red(self.psrs[0].toas,
nmodes=20)
# spectrum test
phi = utils.powerlaw(f2, log10_A=log10_A, gamma=gamma)
msg = "Spectrum incorrect for GP Fourier signal."
assert np.all(m.get_phi(params) == phi), msg
# inverse spectrum test
msg = "Spectrum inverse incorrect for GP Fourier signal."
assert np.all(m.get_phiinv(params) == 1 / phi), msg
gw_log10_A = parameter.Uniform(-18, -11)('gw_log10_A')
gw_gamma = parameter.Constant(args.gamma_gw)('gw_gamma')
# White noise parameter priors
efac = parameter.Constant()
equad = parameter.Constant()
ecorr = parameter.Constant()
Nf = args.nfreqs
freqs = np.linspace(1 / Tspan, Nf / Tspan, Nf)
# # white noise
selection = selections.Selection(selections.nanograv_backends)
ef = white_signals.MeasurementNoise(efac=efac,
log10_t2equad=equad,
selection=selection)
ec = white_signals.EcorrKernelNoise(log10_ecorr=ecorr, selection=selection)
# red noise (powerlaw with 30 frequencies)
pl = utils.powerlaw(log10_A=log10_A, gamma=gamma)
rn = gp_signals.FourierBasisGP(spectrum=pl, modes=freqs)
# timing model
tm = gp_signals.TimingModel()
# gw (powerlaw with 5 frequencies)
gw_pl = utils.powerlaw(log10_A=gw_log10_A, gamma=gw_gamma)
gw_pshift = gp_signals.FourierBasisGP(
spectrum=gw_pl,
def compute_like(self, npsrs=1, inc_corr=False, inc_kernel=False):
# get parameters from PAL2 style noise files
params = get_noise_from_pal2(datadir + "/B1855+09_noise.txt")
params2 = get_noise_from_pal2(datadir + "/J1909-3744_noise.txt")
params.update(params2)
psrs = self.psrs if npsrs == 2 else [self.psrs[0]]
if inc_corr:
params.update({"GW_gamma": 4.33, "GW_log10_A": -15.0})
# find the maximum time span to set GW frequency sampling
tmin = [p.toas.min() for p in psrs]
tmax = [p.toas.max() for p in psrs]
Tspan = np.max(tmax) - np.min(tmin)
# setup basic model
efac = parameter.Constant()
equad = parameter.Constant()
ecorr = parameter.Constant()
log10_A = parameter.Constant()
gamma = parameter.Constant()
selection = Selection(selections.by_backend)
ef = white_signals.MeasurementNoise(efac=efac, selection=selection)
eq = white_signals.EquadNoise(log10_equad=equad, selection=selection)
ec = white_signals.EcorrKernelNoise(log10_ecorr=ecorr,
selection=selection)
pl = utils.powerlaw(log10_A=log10_A, gamma=gamma)
rn = gp_signals.FourierBasisGP(pl)
orf = utils.hd_orf()
crn = gp_signals.FourierBasisCommonGP(pl,
orf,
components=20,
name="GW",
Tspan=Tspan)
tm = gp_signals.TimingModel()
log10_sigma = parameter.Uniform(-10, -5)
log10_lam = parameter.Uniform(np.log10(86400), np.log10(1500 * 86400))
basis = create_quant_matrix(dt=7 * 86400)
prior = se_kernel(log10_sigma=log10_sigma, log10_lam=log10_lam)
se = gp_signals.BasisGP(prior, basis, name="se")
# set up kernel stuff
if isinstance(inc_kernel, bool):
inc_kernel = [inc_kernel] * npsrs
if inc_corr:
s = ef + eq + ec + rn + crn + tm
else:
s = ef + eq + ec + rn + tm
models = []
for ik, psr in zip(inc_kernel, psrs):
snew = s + se if ik else s
models.append(snew(psr))
pta = signal_base.PTA(models)
# set parameters
pta.set_default_params(params)
# SE kernel parameters
log10_sigmas, log10_lams = [-7.0, -6.5], [7.0, 6.5]
params.update({
"B1855+09_se_log10_lam": log10_lams[0],
"B1855+09_se_log10_sigma": log10_sigmas[0],
"J1909-3744_se_log10_lam": log10_lams[1],
"J1909-3744_se_log10_sigma": log10_sigmas[1],
})
# get parameters
efacs, equads, ecorrs, log10_A, gamma = [], [], [], [], []
lsig, llam = [], []
for pname in [p.name for p in psrs]:
efacs.append([
params[key] for key in sorted(params.keys())
if "efac" in key and pname in key
])
equads.append([
params[key] for key in sorted(params.keys())
if "equad" in key and pname in key
])
ecorrs.append([
params[key] for key in sorted(params.keys())
if "ecorr" in key and pname in key
])
log10_A.append(params["{}_red_noise_log10_A".format(pname)])
gamma.append(params["{}_red_noise_gamma".format(pname)])
lsig.append(params["{}_se_log10_sigma".format(pname)])
llam.append(params["{}_se_log10_lam".format(pname)])
GW_gamma = 4.33
GW_log10_A = -15.0
# correct value
tflags = [sorted(list(np.unique(p.backend_flags))) for p in psrs]
cfs, logdets, phis, Ts = [], [], [], []
for ii, (ik, psr, flags) in enumerate(zip(inc_kernel, psrs, tflags)):
nvec0 = np.zeros_like(psr.toas)
for ct, flag in enumerate(flags):
ind = psr.backend_flags == flag
nvec0[ind] = efacs[ii][ct]**2 * psr.toaerrs[ind]**2
nvec0[ind] += 10**(2 * equads[ii][ct]) * np.ones(np.sum(ind))
# get the basis
bflags = psr.backend_flags
Umats = []
for flag in np.unique(bflags):
mask = bflags == flag
Umats.append(
utils.create_quantization_matrix(psr.toas[mask])[0])
nepoch = sum(U.shape[1] for U in Umats)
U = np.zeros((len(psr.toas), nepoch))
jvec = np.zeros(nepoch)
netot = 0
for ct, flag in enumerate(np.unique(bflags)):
mask = bflags == flag
nn = Umats[ct].shape[1]
U[mask, netot:nn + netot] = Umats[ct]
jvec[netot:nn + netot] = 10**(2 * ecorrs[ii][ct])
netot += nn
# get covariance matrix
cov = np.diag(nvec0) + np.dot(U * jvec[None, :], U.T)
cf = sl.cho_factor(cov)
logdet = np.sum(2 * np.log(np.diag(cf[0])))
cfs.append(cf)
logdets.append(logdet)
F, f2 = utils.createfourierdesignmatrix_red(psr.toas,
nmodes=20,
Tspan=Tspan)
Mmat = psr.Mmat.copy()
norm = np.sqrt(np.sum(Mmat**2, axis=0))
Mmat /= norm
U2, avetoas = create_quant_matrix(psr.toas, dt=7 * 86400)
if ik:
T = np.hstack((F, Mmat, U2))
else:
T = np.hstack((F, Mmat))
Ts.append(T)
phi = utils.powerlaw(f2, log10_A=log10_A[ii], gamma=gamma[ii])
if inc_corr:
phigw = utils.powerlaw(f2, log10_A=GW_log10_A, gamma=GW_gamma)
else:
phigw = np.zeros(40)
K = se_kernel(avetoas,
log10_sigma=log10_sigmas[ii],
log10_lam=log10_lams[ii])
k = np.diag(
np.concatenate((phi + phigw, np.ones(Mmat.shape[1]) * 1e40)))
if ik:
k = sl.block_diag(k, K)
phis.append(k)
# manually compute loglike
loglike = 0
TNrs, TNTs = [], []
for ct, psr in enumerate(psrs):
TNrs.append(np.dot(Ts[ct].T, sl.cho_solve(cfs[ct], psr.residuals)))
TNTs.append(np.dot(Ts[ct].T, sl.cho_solve(cfs[ct], Ts[ct])))
loglike += -0.5 * (
np.dot(psr.residuals, sl.cho_solve(cfs[ct], psr.residuals)) +
logdets[ct])
TNr = np.concatenate(TNrs)
phi = sl.block_diag(*phis)
if inc_corr:
hd = utils.hd_orf(psrs[0].pos, psrs[1].pos)
phi[len(phis[0]):len(phis[0]) + 40, :40] = np.diag(phigw * hd)
phi[:40, len(phis[0]):len(phis[0]) + 40] = np.diag(phigw * hd)
cf = sl.cho_factor(phi)
phiinv = sl.cho_solve(cf, np.eye(phi.shape[0]))
logdetphi = np.sum(2 * np.log(np.diag(cf[0])))
Sigma = sl.block_diag(*TNTs) + phiinv
cf = sl.cho_factor(Sigma)
expval = sl.cho_solve(cf, TNr)
logdetsigma = np.sum(2 * np.log(np.diag(cf[0])))
loglike -= 0.5 * (logdetphi + logdetsigma)
loglike += 0.5 * np.dot(TNr, expval)
method = ["partition", "sparse", "cliques"]
for mth in method:
eloglike = pta.get_lnlikelihood(params, phiinv_method=mth)
msg = "Incorrect like for npsr={}, phiinv={}".format(npsrs, mth)
assert np.allclose(eloglike, loglike), msg
def _ecorr_test_ipta(self, method="sparse"):
"""Test of sparse/sherman-morrison ecorr signal and solve methods."""
selection = Selection(selections.nanograv_backends)
efac = parameter.Uniform(0.1, 5)
ecorr = parameter.Uniform(-10, -5)
ef = white_signals.MeasurementNoise(efac=efac)
ec = white_signals.EcorrKernelNoise(log10_ecorr=ecorr,
selection=selection,
method=method)
tm = gp_signals.TimingModel()
s = ef + ec + tm
m = s(self.ipsr)
# set parameters
efacs = [1.3]
ecorrs = [-6.1, -6.2, -6.3, -6.4, -7.2, -8.4, -7.1, -7.9]
params = {
"J1713+0747_efac": efacs[0],
"J1713+0747_ASP-L_log10_ecorr": ecorrs[0],
"J1713+0747_ASP-S_log10_ecorr": ecorrs[1],
"J1713+0747_GASP-8_log10_ecorr": ecorrs[2],
"J1713+0747_GASP-L_log10_ecorr": ecorrs[3],
"J1713+0747_GUPPI-8_log10_ecorr": ecorrs[4],
"J1713+0747_GUPPI-L_log10_ecorr": ecorrs[5],
"J1713+0747_PUPPI-L_log10_ecorr": ecorrs[6],
"J1713+0747_PUPPI-S_log10_ecorr": ecorrs[7],
}
# get EFAC Nvec
nvec0 = efacs[0]**2 * self.ipsr.toaerrs**2
# get the basis
flags = [
"ASP-L", "ASP-S", "GASP-8", "GASP-L", "GUPPI-8", "GUPPI-L",
"PUPPI-L", "PUPPI-S"
]
bflags = self.ipsr.backend_flags
Umats = []
for flag in np.unique(bflags):
if flag in flags:
mask = bflags == flag
Umats.append(
utils.create_quantization_matrix(self.ipsr.toas[mask],
nmin=2)[0])
nepoch = sum(U.shape[1] for U in Umats)
U = np.zeros((len(self.ipsr.toas), nepoch))
jvec = np.zeros(nepoch)
netot, ct = 0, 0
for flag in np.unique(bflags):
if flag in flags:
mask = bflags == flag
nn = Umats[ct].shape[1]
U[mask, netot:nn + netot] = Umats[ct]
jvec[netot:nn + netot] = 10**(2 * ecorrs[ct])
netot += nn
ct += 1
# get covariance matrix
wd = Woodbury(nvec0, U, jvec)
# test
msg = "EFAC/ECORR {} logdet incorrect.".format(method)
N = m.get_ndiag(params)
assert np.allclose(N.solve(self.ipsr.residuals, logdet=True)[1],
wd.logdet(),
rtol=1e-8), msg
msg = "EFAC/ECORR {} D1 solve incorrect.".format(method)
assert np.allclose(N.solve(self.ipsr.residuals),
wd.solve(self.ipsr.residuals),
rtol=1e-8), msg
msg = "EFAC/ECORR {} 1D1 solve incorrect.".format(method)
assert np.allclose(
N.solve(self.ipsr.residuals, left_array=self.ipsr.residuals),
np.dot(self.ipsr.residuals, wd.solve(self.ipsr.residuals)),
rtol=1e-8,
), msg
msg = "EFAC/ECORR {} 2D1 solve incorrect.".format(method)
T = m.get_basis()
assert np.allclose(N.solve(self.ipsr.residuals, left_array=T),
np.dot(T.T, wd.solve(self.ipsr.residuals)),
rtol=1e-8), msg
msg = "EFAC/ECORR {} 2D2 solve incorrect.".format(method)
assert np.allclose(N.solve(T, left_array=T),
np.dot(T.T, wd.solve(T)),
rtol=1e-8), msg
def _ecorr_test(self, method="sparse"):
"""Test of sparse/sherman-morrison ecorr signal and solve methods."""
selection = Selection(selections.by_backend)
efac = parameter.Uniform(0.1, 5)
ecorr = parameter.Uniform(-10, -5)
ef = white_signals.MeasurementNoise(efac=efac, selection=selection)
ec = white_signals.EcorrKernelNoise(log10_ecorr=ecorr,
selection=selection,
method=method)
tm = gp_signals.TimingModel()
s = ef + ec + tm
m = s(self.psr)
# set parameters
efacs = [1.3, 1.4, 1.5, 1.6]
ecorrs = [-6.1, -6.2, -6.3, -6.4]
params = {
"B1855+09_430_ASP_efac": efacs[0],
"B1855+09_430_PUPPI_efac": efacs[1],
"B1855+09_L-wide_ASP_efac": efacs[2],
"B1855+09_L-wide_PUPPI_efac": efacs[3],
"B1855+09_430_ASP_log10_ecorr": ecorrs[0],
"B1855+09_430_PUPPI_log10_ecorr": ecorrs[1],
"B1855+09_L-wide_ASP_log10_ecorr": ecorrs[2],
"B1855+09_L-wide_PUPPI_log10_ecorr": ecorrs[3],
}
# get EFAC Nvec
flags = ["430_ASP", "430_PUPPI", "L-wide_ASP", "L-wide_PUPPI"]
nvec0 = np.zeros_like(self.psr.toas)
for ct, flag in enumerate(np.unique(flags)):
ind = flag == self.psr.backend_flags
nvec0[ind] = efacs[ct]**2 * self.psr.toaerrs[ind]**2
# get the basis
bflags = self.psr.backend_flags
Umats = []
for flag in np.unique(bflags):
mask = bflags == flag
Umats.append(
utils.create_quantization_matrix(self.psr.toas[mask],
nmin=2)[0])
nepoch = sum(U.shape[1] for U in Umats)
U = np.zeros((len(self.psr.toas), nepoch))
jvec = np.zeros(nepoch)
netot = 0
for ct, flag in enumerate(np.unique(bflags)):
mask = bflags == flag
nn = Umats[ct].shape[1]
U[mask, netot:nn + netot] = Umats[ct]
jvec[netot:nn + netot] = 10**(2 * ecorrs[ct])
netot += nn
# get covariance matrix
wd = Woodbury(nvec0, U, jvec)
# test
msg = "EFAC/ECORR {} logdet incorrect.".format(method)
N = m.get_ndiag(params)
assert np.allclose(N.solve(self.psr.residuals, logdet=True)[1],
wd.logdet(),
rtol=1e-10), msg
msg = "EFAC/ECORR {} D1 solve incorrect.".format(method)
assert np.allclose(N.solve(self.psr.residuals),
wd.solve(self.psr.residuals),
rtol=1e-10), msg
msg = "EFAC/ECORR {} 1D1 solve incorrect.".format(method)
assert np.allclose(
N.solve(self.psr.residuals, left_array=self.psr.residuals),
np.dot(self.psr.residuals, wd.solve(self.psr.residuals)),
rtol=1e-10,
), msg
msg = "EFAC/ECORR {} 2D1 solve incorrect.".format(method)
T = m.get_basis()
assert np.allclose(N.solve(self.psr.residuals, left_array=T),
np.dot(T.T, wd.solve(self.psr.residuals)),
rtol=1e-10), msg
msg = "EFAC/ECORR {} 2D2 solve incorrect.".format(method)
assert np.allclose(N.solve(T, left_array=T),
np.dot(T.T, wd.solve(T)),
rtol=1e-10), msg
def test_add_efac_equad_backend(self):
"""Test that addition of efac-backend and equad-backend signal returns
correct covariance.
"""
selection = Selection(selections.by_backend)
efac = parameter.Uniform(0.1, 5)
equad = parameter.Uniform(-10, -5)
ef = white_signals.MeasurementNoise(efac=efac, selection=selection)
eq = white_signals.EquadNoise(log10_equad=equad, selection=selection)
s = ef + eq
m = s(self.psr)
# set parameters
efacs = [1.3, 1.4, 1.5, 1.6]
equads = [-6.1, -6.2, -6.3, -6.4]
params = {
"B1855+09_430_ASP_efac": efacs[0],
"B1855+09_430_PUPPI_efac": efacs[1],
"B1855+09_L-wide_ASP_efac": efacs[2],
"B1855+09_L-wide_PUPPI_efac": efacs[3],
"B1855+09_430_ASP_log10_equad": equads[0],
"B1855+09_430_PUPPI_log10_equad": equads[1],
"B1855+09_L-wide_ASP_log10_equad": equads[2],
"B1855+09_L-wide_PUPPI_log10_equad": equads[3],
}
# correct value
flags = ["430_ASP", "430_PUPPI", "L-wide_ASP", "L-wide_PUPPI"]
nvec0 = np.zeros_like(self.psr.toas)
for ct, flag in enumerate(np.unique(flags)):
ind = flag == self.psr.backend_flags
nvec0[ind] = efacs[ct]**2 * self.psr.toaerrs[ind]**2
nvec0[ind] += 10**(2 * equads[ct]) * np.ones(np.sum(ind))
logdet = np.sum(np.log(nvec0))
# test
msg = "EFAC/EQUAD covariance incorrect."
assert np.all(m.get_ndiag(params) == nvec0), msg
msg = "EFAC/EQUAD logdet incorrect."
N = m.get_ndiag(params)
assert np.allclose(N.solve(self.psr.residuals, logdet=True)[1],
logdet,
rtol=1e-10), msg
msg = "EFAC/EQUAD D1 solve incorrect."
assert np.allclose(N.solve(self.psr.residuals),
self.psr.residuals / nvec0,
rtol=1e-10), msg
msg = "EFAC/EQUAD 1D1 solve incorrect."
assert np.allclose(
N.solve(self.psr.residuals, left_array=self.psr.residuals),
np.dot(self.psr.residuals / nvec0, self.psr.residuals),
rtol=1e-10,
), msg
msg = "EFAC/EQUAD 2D1 solve incorrect."
T = self.psr.Mmat
assert np.allclose(N.solve(self.psr.residuals, left_array=T),
np.dot(T.T, self.psr.residuals / nvec0),
rtol=1e-10), msg
msg = "EFAC/EQUAD 2D2 solve incorrect."
assert np.allclose(N.solve(T, left_array=T),
np.dot(T.T, T / nvec0[:, None]),
rtol=1e-10), msg
def test_common_red_noise(self):
"""Test of a coefficient-based common GP."""
pl = utils.powerlaw(log10_A=parameter.Uniform(-18, -12),
gamma=parameter.Uniform(1, 7))
ef = white_signals.MeasurementNoise(
efac=parameter.Uniform(0.1, 5.0))
Tspan = max(self.psr.toas.max(), self.psr2.toas.max()) - min(
self.psr.toas.max(), self.psr2.toas.max())
pl = utils.powerlaw(log10_A=parameter.Uniform(-18, -12),
gamma=parameter.Uniform(1, 7))
rn = gp_signals.FourierBasisCommonGP(spectrum=pl,
orf=utils.hd_orf(),
components=20,
Tspan=Tspan)
model = ef + rn
rnc = gp_signals.FourierBasisCommonGP(spectrum=pl,
orf=utils.hd_orf(),
components=20,
Tspan=Tspan,
coefficients=True)
modelc = ef + rnc
pta = signal_base.PTA([model(self.psr), model(self.psr2)])
ptac = signal_base.PTA([modelc(self.psr), modelc(self.psr2)])
params = {
"B1855+09_efac": 1.0,
"B1937+21_efac": 1.0,
"common_fourier_gamma": 5,
"common_fourier_log10_A": -15,
}
# get GP delays in two different ways
cf, cf2 = np.random.randn(40), np.random.randn(40)
bs = pta.get_basis(params)
da = [np.dot(bs[0], cf), np.dot(bs[1], cf2)]
params.update({
"B1855+09_common_fourier_coefficients": cf,
"B1937+21_common_fourier_coefficients": cf2
})
db = ptac.get_delay(params)
msg = "Implicit and explicit GP delays are different."
assert np.allclose(da[0], db[0]), msg
assert np.allclose(da[1], db[1]), msg
cpar = [p for p in ptac.params if "coefficients" in p.name]
def shouldfail():
return cpar[0].get_logpdf(params)
self.assertRaises(NotImplementedError, shouldfail)
def test_pta_phiinv_methods(self):
ef = white_signals.MeasurementNoise(efac=parameter.Uniform(0.1, 5))
span = np.max(self.psrs[0].toas) - np.min(self.psrs[0].toas)
pl = utils.powerlaw(log10_A=parameter.Uniform(-16, -13),
gamma=parameter.Uniform(1, 7))
orf = utils.hd_orf()
vrf = utils.dipole_orf()
rn = gp_signals.FourierBasisGP(spectrum=pl, components=30, Tspan=span)
hdrn = gp_signals.FourierBasisCommonGP(spectrum=pl,
orf=orf,
components=20,
Tspan=span,
name='gw')
vrn = gp_signals.FourierBasisCommonGP(spectrum=pl,
orf=vrf,
components=20,
Tspan=span,
name='vec')
vrn2 = gp_signals.FourierBasisCommonGP(spectrum=pl,
orf=vrf,
components=20,
Tspan=span * 1.234,
name='vec2')
# two common processes, sharing basis partially
model = ef + rn + hdrn # + vrn
pta = signal_base.PTA([model(psr) for psr in self.psrs])
ps = parameter.sample(pta.params)
phi = pta.get_phi(ps)
ldp = np.linalg.slogdet(phi)[1]
inv1, ld1 = pta.get_phiinv(ps, method='cliques', logdet=True)
inv2, ld2 = pta.get_phiinv(ps, method='partition', logdet=True)
inv3, ld3 = pta.get_phiinv(ps, method='sparse', logdet=True)
if not isinstance(inv3, np.ndarray):
inv3 = inv3.toarray()
for ld in [ld1, ld2, ld3]:
msg = "Wrong phi log determinant for two common processes"
assert np.allclose(ldp, ld, rtol=1e-15, atol=1e-6), msg
for inv in [inv1, inv2, inv3]:
msg = "Wrong phi inverse for two common processes"
assert np.allclose(np.dot(phi, inv),
np.eye(phi.shape[0]),
rtol=1e-15,
atol=1e-6), msg
for inva, invb in itertools.combinations([inv1, inv2, inv3], 2):
assert np.allclose(inva, invb)
# two common processes, no sharing basis
model = ef + rn + vrn2
pta = signal_base.PTA([model(psr) for psr in self.psrs])
ps = parameter.sample(pta.params)
phi = pta.get_phi(ps)
ldp = np.linalg.slogdet(phi)[1]
inv1, ld1 = pta.get_phiinv(ps, method='cliques', logdet=True)
inv2, ld2 = pta.get_phiinv(ps, method='partition', logdet=True)
inv3, ld3 = pta.get_phiinv(ps, method='sparse', logdet=True)
if not isinstance(inv3, np.ndarray):
inv3 = inv3.toarray()
for ld in [ld1, ld2, ld3]:
msg = "Wrong phi log determinant for two common processes"
assert np.allclose(ldp, ld, rtol=1e-15, atol=1e-6), msg
for inv in [inv1, inv2, inv3]:
msg = "Wrong phi inverse for two processes"
assert np.allclose(np.dot(phi, inv),
np.eye(phi.shape[0]),
rtol=1e-15,
atol=1e-6), msg
for inva, invb in itertools.combinations([inv1, inv2, inv3], 2):
assert np.allclose(inva, invb)
# three common processes, sharing basis partially
model = ef + rn + hdrn + vrn
pta = signal_base.PTA([model(psr) for psr in self.psrs])
ps = parameter.sample(pta.params)
phi = pta.get_phi(ps)
ldp = np.linalg.slogdet(phi)[1]
inv1, ld1 = pta.get_phiinv(ps, method='cliques', logdet=True)
inv2, ld2 = pta.get_phiinv(ps, method='partition', logdet=True)
inv3, ld3 = pta.get_phiinv(ps, method='sparse', logdet=True)
if not isinstance(inv3, np.ndarray):
inv3 = inv3.toarray()
for ld in [ld1, ld3]:
msg = "Wrong phi log determinant for two common processes"
assert np.allclose(ldp, ld, rtol=1e-15, atol=1e-6), msg
for inv in [inv1, inv3]:
msg = "Wrong phi inverse for three common processes"
assert np.allclose(np.dot(phi, inv),
np.eye(phi.shape[0]),
rtol=1e-15,
atol=1e-6), msg
for inva, invb in itertools.combinations([inv1, inv3], 2):
assert np.allclose(inva, invb)
# four common processes, three sharing basis partially
model = ef + rn + hdrn + vrn + vrn2
pta = signal_base.PTA([model(psr) for psr in self.psrs])
ps = parameter.sample(pta.params)
phi = pta.get_phi(ps)
ldp = np.linalg.slogdet(phi)[1]
inv1, ld1 = pta.get_phiinv(ps, method='cliques', logdet=True)
inv2, ld2 = pta.get_phiinv(ps, method='partition', logdet=True)
inv3, ld3 = pta.get_phiinv(ps, method='sparse', logdet=True)
if not isinstance(inv3, np.ndarray):
inv3 = inv3.toarray()
for ld in [ld1, ld3]:
msg = "Wrong phi log determinant for two common processes"
assert np.allclose(ldp, ld, rtol=1e-15, atol=1e-6), msg
for inv in [inv1, inv3]:
msg = "Wrong phi inverse for four processes"
assert np.allclose(np.dot(phi, inv),
np.eye(phi.shape[0]),
rtol=1e-15,
atol=1e-6), msg
for inva, invb in itertools.combinations([inv1, inv3], 2):
assert np.allclose(inva, invb)
