def test_adapt_t_process_prior(self):
"""Test that red noise signal returns correct values."""
# set up signal parameter
pr = gp_priors.t_process_adapt(
log10_A=parameter.Uniform(-18, -12),
gamma=parameter.Uniform(1, 7),
alphas_adapt=gp_priors.InvGamma(),
nfreq=parameter.Uniform(5, 25),
)
basis = gp_bases.createfourierdesignmatrix_red(nmodes=30)
rn = gp_signals.BasisGP(priorFunction=pr,
basisFunction=basis,
name="red_noise")
rnm = rn(self.psr)
# parameters
alphas = scipy.stats.invgamma.rvs(1, scale=1, size=1)
log10_A, gamma, nfreq = -15, 4.33, 12
params = {
"B1855+09_red_noise_log10_A": log10_A,
"B1855+09_red_noise_gamma": gamma,
"B1855+09_red_noise_alphas_adapt": alphas,
"B1855+09_red_noise_nfreq": nfreq,
}
# basis matrix test
F, f2 = gp_bases.createfourierdesignmatrix_red(self.psr.toas,
nmodes=30)
msg = "F matrix incorrect for free spectrum."
assert np.allclose(F, rnm.get_basis(params)), msg
# spectrum test
phi = gp_priors.t_process_adapt(f2,
log10_A=log10_A,
gamma=gamma,
alphas_adapt=alphas,
nfreq=nfreq)
msg = "Spectrum incorrect for free spectrum."
assert np.all(rnm.get_phi(params) == phi), msg
# inverse spectrum test
msg = "Spectrum inverse incorrect for free spectrum."
assert np.all(rnm.get_phiinv(params) == 1 / phi), msg
# test shape
msg = "F matrix shape incorrect"
assert rnm.get_basis(params).shape == F.shape, msg
def dm_noise_block(gp_kernel='diag',
psd='powerlaw',
nondiag_kernel='periodic',
prior='log-uniform',
dt=15,
df=200,
Tspan=None,
components=30,
gamma_val=None,
coefficients=False):
"""
Returns DM noise model:
1. DM noise modeled as a power-law with 30 sampling frequencies
:param psd:
PSD function [e.g. powerlaw (default), spectrum, tprocess]
:param prior:
Prior on log10_A. Default if "log-uniform". Use "uniform" for
upper limits.
:param dt:
time-scale for linear interpolation basis (days)
:param df:
frequency-scale for linear interpolation basis (MHz)
:param Tspan:
Sets frequency sampling f_i = i / Tspan. Default will
use overall time span for indivicual pulsar.
:param components:
Number of frequencies in sampling of DM-variations.
:param gamma_val:
If given, this is the fixed slope of the power-law for
powerlaw, turnover, or tprocess DM-variations
"""
# dm noise parameters that are common
if gp_kernel == 'diag':
if psd in ['powerlaw', 'turnover', 'tprocess', 'tprocess_adapt']:
# parameters shared by PSD functions
if prior == 'uniform':
log10_A_dm = parameter.LinearExp(-20, -11)
elif prior == 'log-uniform' and gamma_val is not None:
if np.abs(gamma_val - 4.33) < 0.1:
log10_A_dm = parameter.Uniform(-20, -11)
else:
log10_A_dm = parameter.Uniform(-20, -11)
else:
log10_A_dm = parameter.Uniform(-20, -11)
if gamma_val is not None:
gamma_dm = parameter.Constant(gamma_val)
else:
gamma_dm = parameter.Uniform(0, 7)
# different PSD function parameters
if psd == 'powerlaw':
dm_prior = utils.powerlaw(log10_A=log10_A_dm, gamma=gamma_dm)
elif psd == 'turnover':
kappa_dm = parameter.Uniform(0, 7)
lf0_dm = parameter.Uniform(-9, -7)
dm_prior = utils.turnover(log10_A=log10_A_dm,
gamma=gamma_dm,
lf0=lf0_dm,
kappa=kappa_dm)
elif psd == 'tprocess':
df = 2
alphas_dm = gpp.InvGamma(df / 2, df / 2, size=components)
dm_prior = gpp.t_process(log10_A=log10_A_dm,
gamma=gamma_dm,
alphas=alphas_dm)
elif psd == 'tprocess_adapt':
df = 2
alpha_adapt_dm = gpp.InvGamma(df / 2, df / 2, size=1)
nfreq_dm = parameter.Uniform(-0.5, 10 - 0.5)
dm_prior = gpp.t_process_adapt(log10_A=log10_A_dm,
gamma=gamma_dm,
alphas_adapt=alpha_adapt_dm,
nfreq=nfreq_dm)
if psd == 'spectrum':
if prior == 'uniform':
log10_rho_dm = parameter.LinearExp(-10, -4, size=components)
elif prior == 'log-uniform':
log10_rho_dm = parameter.Uniform(-10, -4, size=components)
dm_prior = gpp.free_spectrum(log10_rho=log10_rho_dm)
dm_basis = utils.createfourierdesignmatrix_dm(nmodes=components,
Tspan=Tspan)
elif gp_kernel == 'nondiag':
if nondiag_kernel == 'periodic':
# Periodic GP kernel for DM
log10_sigma = parameter.Uniform(-10, -4)
log10_ell = parameter.Uniform(1, 4)
log10_p = parameter.Uniform(-4, 1)
log10_gam_p = parameter.Uniform(-3, 2)
dm_basis = gpk.linear_interp_basis_dm(dt=dt * const.day)
dm_prior = gpk.periodic_kernel(log10_sigma=log10_sigma,
log10_ell=log10_ell,
log10_gam_p=log10_gam_p,
log10_p=log10_p)
elif nondiag_kernel == 'periodic_rfband':
# Periodic GP kernel for DM with RQ radio-frequency dependence
log10_sigma = parameter.Uniform(-10, -4)
log10_ell = parameter.Uniform(1, 4)
log10_ell2 = parameter.Uniform(2, 7)
log10_alpha_wgt = parameter.Uniform(-4, 1)
log10_p = parameter.Uniform(-4, 1)
log10_gam_p = parameter.Uniform(-3, 2)
dm_basis = gpk.get_tf_quantization_matrix(df=df,
dt=dt * const.day,
dm=True)
dm_prior = gpk.tf_kernel(log10_sigma=log10_sigma,
log10_ell=log10_ell,
log10_gam_p=log10_gam_p,
log10_p=log10_p,
log10_alpha_wgt=log10_alpha_wgt,
log10_ell2=log10_ell2)
elif nondiag_kernel == 'sq_exp':
# squared-exponential GP kernel for DM
log10_sigma = parameter.Uniform(-10, -4)
log10_ell = parameter.Uniform(1, 4)
dm_basis = gpk.linear_interp_basis_dm(dt=dt * const.day)
dm_prior = gpk.se_dm_kernel(log10_sigma=log10_sigma,
log10_ell=log10_ell)
elif nondiag_kernel == 'sq_exp_rfband':
# Sq-Exp GP kernel for DM with RQ radio-frequency dependence
log10_sigma = parameter.Uniform(-10, -4)
log10_ell = parameter.Uniform(1, 4)
log10_ell2 = parameter.Uniform(2, 7)
log10_alpha_wgt = parameter.Uniform(-4, 1)
dm_basis = gpk.get_tf_quantization_matrix(df=df,
dt=dt * const.day,
dm=True)
dm_prior = gpk.sf_kernel(log10_sigma=log10_sigma,
log10_ell=log10_ell,
log10_alpha_wgt=log10_alpha_wgt,
log10_ell2=log10_ell2)
elif nondiag_kernel == 'dmx_like':
# DMX-like signal
log10_sigma = parameter.Uniform(-10, -4)
dm_basis = gpk.linear_interp_basis_dm(dt=dt * const.day)
dm_prior = gpk.dmx_ridge_prior(log10_sigma=log10_sigma)
dmgp = gp_signals.BasisGP(dm_prior,
dm_basis,
name='dm_gp',
coefficients=coefficients)
return dmgp
def red_noise_block(psd='powerlaw',
prior='log-uniform',
Tspan=None,
components=30,
gamma_val=None,
coefficients=False,
select=None,
modes=None,
wgts=None,
break_flat=False,
break_flat_fq=None):
"""
Returns red noise model:
1. Red noise modeled as a power-law with 30 sampling frequencies
:param psd:
PSD function [e.g. powerlaw (default), turnover, spectrum, tprocess]
:param prior:
Prior on log10_A. Default if "log-uniform". Use "uniform" for
upper limits.
:param Tspan:
Sets frequency sampling f_i = i / Tspan. Default will
use overall time span for indivicual pulsar.
:param components:
Number of frequencies in sampling of red noise
:param gamma_val:
If given, this is the fixed slope of the power-law for
powerlaw, turnover, or tprocess red noise
:param coefficients: include latent coefficients in GP model?
"""
# red noise parameters that are common
if psd in [
'powerlaw', 'powerlaw_genmodes', 'turnover', 'tprocess',
'tprocess_adapt', 'infinitepower'
]:
# parameters shared by PSD functions
if prior == 'uniform':
log10_A = parameter.LinearExp(-20, -11)
elif prior == 'log-uniform' and gamma_val is not None:
if np.abs(gamma_val - 4.33) < 0.1:
log10_A = parameter.Uniform(-20, -11)
else:
log10_A = parameter.Uniform(-20, -11)
else:
log10_A = parameter.Uniform(-20, -11)
if gamma_val is not None:
gamma = parameter.Constant(gamma_val)
else:
gamma = parameter.Uniform(0, 7)
# different PSD function parameters
if psd == 'powerlaw':
pl = utils.powerlaw(log10_A=log10_A, gamma=gamma)
elif psd == 'powerlaw_genmodes':
pl = gpp.powerlaw_genmodes(log10_A=log10_A, gamma=gamma, wgts=wgts)
elif psd == 'turnover':
kappa = parameter.Uniform(0, 7)
lf0 = parameter.Uniform(-9, -7)
pl = utils.turnover(log10_A=log10_A,
gamma=gamma,
lf0=lf0,
kappa=kappa)
elif psd == 'tprocess':
df = 2
alphas = gpp.InvGamma(df / 2, df / 2, size=components)
pl = gpp.t_process(log10_A=log10_A, gamma=gamma, alphas=alphas)
elif psd == 'tprocess_adapt':
df = 2
alpha_adapt = gpp.InvGamma(df / 2, df / 2, size=1)
nfreq = parameter.Uniform(-0.5, 10 - 0.5)
pl = gpp.t_process_adapt(log10_A=log10_A,
gamma=gamma,
alphas_adapt=alpha_adapt,
nfreq=nfreq)
elif psd == 'infinitepower':
pl = gpp.infinitepower()
if psd == 'spectrum':
if prior == 'uniform':
log10_rho = parameter.LinearExp(-10, -4, size=components)
elif prior == 'log-uniform':
log10_rho = parameter.Uniform(-10, -4, size=components)
pl = gpp.free_spectrum(log10_rho=log10_rho)
if select == 'backend':
# define selection by observing backend
selection = selections.Selection(selections.by_backend)
elif select == 'band' or select == 'band+':
# define selection by observing band
selection = selections.Selection(selections.by_band)
else:
# define no selection
selection = selections.Selection(selections.no_selection)
if break_flat:
log10_A_flat = parameter.Uniform(-20, -11)
gamma_flat = parameter.Constant(0)
pl_flat = utils.powerlaw(log10_A=log10_A_flat, gamma=gamma_flat)
freqs = 1.0 * np.arange(1, components + 1) / Tspan
components_low = sum(f < break_flat_fq for f in freqs)
if components_low < 1.5:
components_low = 2
rn = gp_signals.FourierBasisGP(pl,
components=components_low,
Tspan=Tspan,
coefficients=coefficients,
selection=selection)
rn_flat = gp_signals.FourierBasisGP(pl_flat,
modes=freqs[components_low:],
coefficients=coefficients,
selection=selection,
name='red_noise_hf')
rn = rn + rn_flat
else:
rn = gp_signals.FourierBasisGP(pl,
components=components,
Tspan=Tspan,
coefficients=coefficients,
selection=selection,
modes=modes)
if select == 'band+': # Add the common component as well
rn = rn + gp_signals.FourierBasisGP(
pl, components=components, Tspan=Tspan, coefficients=coefficients)
return rn