Source code for pymaster.covariance

import numpy as np
import healpy as hp
from pymaster import nmtlib as lib
import pymaster.utils as ut
from pymaster import (compute_coupled_cell, NmtBin, NmtWorkspace,
                      NmtFieldCatalog)


def _get_mask_prod_alm(f1, f2):
    # If we have catalog and map, make sure catalog goes
    # first
    fa, fb = (f1, f2) if _is_mask_catalog(f1) else (f2, f1)

    # Check they have the same lmax_mask
    if not f1.is_compatible(f2, strict=False):
        raise ValueError("Fields have incompatible pixelizations.")

    # Check which case we are dealing with
    if _is_mask_catalog(fa):
        if _is_mask_catalog(fb):
            option = 'cat_cat'
        else:
            option = 'cat_map'
    else:
        option = 'map_map'

    if option == 'map_map':
        minfo = fa.minfo
        if fa.is_compatible(fb):
            mask_p = fa.get_mask()*fb.get_mask()
        else:
            mask_a = fa.get_mask()
            if minfo.is_healpix and fb.minfo.is_healpix:
                mask_b = hp.ud_grade(fb.get_mask(), nside_out=minfo.nside)
            else:
                wlm_b = fb.get_mask_alms()
                mask_b = ut.alm2map(np.array([wlm_b]), 0,
                                    minfo, fb.ainfo_mask).squeeze()
            mask_p = mask_a * mask_b
    else:
        # The first field is a catalog
        mask_a, nside_a = fa.get_catalog_mask_map()
        minfo = ut.NmtMapInfo(None, [len(mask_a)])
        if option == 'cat_map':
            if fb.minfo.is_healpix:
                mask_b = hp.ud_grade(fb.get_mask(), nside_out=nside_a)
            else:  # Need to reproject CAR into healpix
                wlm_b = fb.get_mask_alms()
                mask_b = ut.alm2map(np.array([wlm_b]), 0, minfo,
                                    fb.ainfo_mask).squeeze()
            mask_p = mask_a * mask_b
        else:  # cat-cat
            auto = fa is fb
            if auto:
                mask_b, nside_b = mask_a, nside_a
            else:
                mask_b, nside_b = fb.get_catalog_mask_map()
            assert nside_a == nside_b
            mask_p = mask_a * mask_b
            if auto:  # Subtract self-pair contribution
                mask2_a, nside2_a = fa.get_catalog_mask_squared_map()
                assert nside_a == nside2_a
                mask_p -= mask2_a
    mask_p_alm = ut.map2alm(np.array([mask_p]), 0,
                            minfo, fa.ainfo_mask,
                            n_iter=fa.n_iter_mask)[0]
    return mask_p_alm, minfo


[docs] class NmtCovarianceWorkspace(object): """ :obj:`NmtCovarianceWorkspace` objects are used to compute and store the coupling coefficients needed to calculate the Gaussian covariance matrix of angular power spectra under the approximations described in in `Garcia-Garcia et al. 2019 <https://arxiv.org/abs/1906.11765>`_ (see also `Efstathiou et al. 2003 <https://arxiv.org/abs/astro-ph/0307515>`_, and `Couchot et al. 2016 <https://arxiv.org/abs/1609.09730>`_). :obj:`NmtCovarianceWorkspace` objects may be constructed from a set of :obj:`~pymaster.field.NmtField` objects, describing the masks of the fields being correlated, or may be read from a file. We recommend using the class methods :meth:`from_fields` and :meth:`from_file` to create new :obj:`NmtCovarianceWorkspace` objects, rather than using the main constructor. Args: fla1 (:class:`~pymaster.field.NmtField`): First field contributing to the first power spectrum whose covariance you want to compute. fla2 (:class:`~pymaster.field.NmtField`): Second field contributing to the first power spectrum whose covariance you want to compute. flb1 (:class:`~pymaster.field.NmtField`): As ``fla1`` for the second power spectrum. If ``None``, it will be set to ``fla1``. flb2 (:class:`~pymaster.field.NmtField`): As ``fla2`` for the second power spectrum. If ``None``, it will be set to ``fla2``. all_spins (:obj:`bool`): If ``True``, coupling coefficients for all spin combinations will be calculated. Otherwise, only the spin combination determined by the input fields will be considered. The default value is ``True``, but setting it to ``False`` will generally lead to faster results and better memory usage (at the expense of some flexibility). l_toeplitz (:obj:`int`): If a positive number, the Toeplitz approximation described in `Louis et al. 2020 <https://arxiv.org/abs/2010.14344>`_ will be used. In that case, this quantity corresponds to :math:`\\ell_{\\rm toeplitz}` in Fig. 3 of that paper. l_exact (:obj:`int`): If ``l_toeplitz>0``, it corresponds to :math:`\\ell_{\\rm exact}` in Fig. 3 of the paper. Ignored if ``l_toeplitz<=0``. dl_band (:obj:`int`): If ``l_toeplitz>0``, this quantity corresponds to :math:`\\Delta \\ell_{\\rm band}` in Fig. 3 of the paper. Ignored if ``l_toeplitz<=0``. fname (:obj:`str`): Input file name. If not `None`, the values of all input fields will be ignored, and all mode-coupling coefficients will be read from file.""" def __init__(self, fla1, fla2, flb1=None, flb2=None, all_spins=False, l_toeplitz=-1, l_exact=-1, dl_band=-1, fname=None): self.wsp = None self.wsp_SN = None self.wsp_NS = None self.wsp_NN = None if (fname is not None): self._read_from(fname) return if flb1 is None: flb1 = fla1 if flb2 is None: flb2 = fla2 self.all_spins = all_spins self.spin_a1 = fla1.spin self.spin_a2 = fla2.spin self.spin_b1 = flb1.spin self.spin_b2 = flb2.spin self._compute_coupling_coefficients(fla1, fla2, flb1, flb2, all_spins=all_spins, l_toeplitz=l_toeplitz, l_exact=l_exact, dl_band=dl_band)
[docs] @classmethod def from_fields(cls, fla1, fla2, flb1=None, flb2=None, *, all_spins=False, l_toeplitz=-1, l_exact=-1, dl_band=-1): """ Creates an :obj:`NmtCovarianceWorkspace` object containing the mode-coupling coefficients of the Gaussian covariance between the power spectra of two pairs of :class:`~pymaster.field.NmtField` objects (``fla1``, ``fla2``, ``flb1``, and ``flb2``). Note that you can reuse this workspace for the covariance of power spectra between any pairs of fields as long as the fields have the same masks as those passed to this function, and as long as the binning schemes used are also the same. Args: fla1 (:class:`~pymaster.field.NmtField`): First field contributing to the first power spectrum whose covariance you want to compute. fla2 (:class:`~pymaster.field.NmtField`): Second field contributing to the first power spectrum whose covariance you want to compute. flb1 (:class:`~pymaster.field.NmtField`): As ``fla1`` for the second power spectrum. If ``None``, it will be set to ``fla1``. flb2 (:class:`~pymaster.field.NmtField`): As ``fla2`` for the second power spectrum. If ``None``, it will be set to ``fla2``. all_spins (:obj:`bool`): If ``True``, coupling coefficients for all spin combinations will be calculated. Otherwise, only the spin combination determined by the input fields will be considered. l_toeplitz (:obj:`int`): If a positive number, the Toeplitz approximation described in `Louis et al. 2020 <https://arxiv.org/abs/2010.14344>`_ will be used. In that case, this quantity corresponds to :math:`\\ell_{\\rm toeplitz}` in Fig. 3 of that paper. l_exact (:obj:`int`): If ``l_toeplitz>0``, it corresponds to :math:`\\ell_{\\rm exact}` in Fig. 3 of the paper. Ignored if ``l_toeplitz<=0``. dl_band (:obj:`int`): If ``l_toeplitz>0``, this quantity corresponds to :math:`\\Delta \\ell_{\\rm band}` in Fig. 3 of the paper. Ignored if ``l_toeplitz<=0``. """ return cls(fla1=fla1, fla2=fla2, flb1=flb1, flb2=flb2, all_spins=all_spins, l_toeplitz=l_toeplitz, l_exact=l_exact, dl_band=dl_band)
[docs] @classmethod def from_file(cls, fname, fname_SN=None, fname_NS=None, fname_NN=None): """ Creates an :obj:`NmtCovarianceWorkspace` object from the mode-coupling coefficients stored in a FITS file. See :meth:`write_to`. Args: fname (:obj:`str`): Input file name.""" return cls(None, None, fname=fname)
def __del__(self): if self.wsp is not None: if lib.covar_workspace_free is not None: lib.covar_workspace_free(self.wsp) self.wsp = None if self.wsp_SN is not None: if lib.covar_workspace_free is not None: lib.covar_workspace_free(self.wsp_SN) self.wsp_SN = None if self.wsp_NS is not None: if lib.covar_workspace_free is not None: lib.covar_workspace_free(self.wsp_NS) self.wsp_NS = None if self.wsp_NN is not None: if lib.covar_workspace_free is not None: lib.covar_workspace_free(self.wsp_NN) self.wsp_NN = None def _read_from(self, fname): """ Reads the contents of an :obj:`NmtCovarianceWorkspace` object from a FITS file. Args: fname (:obj:`str`): Input file name.""" if self.wsp is not None: lib.covar_workspace_free(self.wsp) self.wsp = None import fitsio as fts f = fts.FITS(fname) print(f) h = f['CWSP_PRIMARY'].read_header() self.lmax = h['LMAX'] self.lmax_mask = h['LMAX_MASK'] if 'LMAX_MASK' in h else self.lmax if 'ALL_SPINS' in h: self.all_spins = h['ALL_SPINS'] self.spin_a1 = h['SPIN_A1'] self.spin_a2 = h['SPIN_A2'] self.spin_b1 = h['SPIN_B1'] self.spin_b2 = h['SPIN_B2'] else: self.all_spins = 1 self.spin_a1 = self.spin_a2 = self.spin_b1 = self.spin_b2 = 0 self.has_SN = np.array([False, False]) self.has_NS = np.array([False, False]) self.has_NN = np.array([False, False]) # Read the coupling coefficients xi_types = ['00_1122', '00_1221', '02_1122', '02_1221', '22P_1122', '22P_1221', '22M_1122', '22M_1221'] xis = {'': {}, 'SN': {}, 'NS': {}, 'NN': {}} # Loop over the different signal-noise combinations for prefix in ['', 'SN', 'NS', 'NN']: xi = xis[prefix] xi_any = False # Read all stored coupling coefficients for n in xi_types: if f'XI{prefix+n}' in f: xi_any = True xi[n] = f[f'XI{prefix + n}'].read() if xi[n].shape != (self.lmax+1, self.lmax+1): raise ValueError(f"XI{prefix + n} shape " f"does not match expected dimensions") xi[n] = xi[n].flatten() else: xi[n] = np.array([0.0]) if not xi_any: xis[prefix] = None # Create all C-level workspaces self.wsp = lib.covar_workspace_init_from_xi( self.spin_a1, self.spin_a2, self.spin_b1, self.spin_b2, self.all_spins, self.lmax, self.lmax_mask, xis['']['00_1122'], xis['']['00_1221'], xis['']['02_1122'], xis['']['02_1221'], xis['']['22P_1122'], xis['']['22P_1221'], xis['']['22M_1122'], xis['']['22M_1221']) if xis['SN'] is not None: if self.wsp_SN is not None: lib.covar_workspace_free(self.wsp_SN) self.wsp_SN = None self.wsp_SN = lib.covar_workspace_init_from_xi( self.spin_a1, self.spin_a2, self.spin_b1, self.spin_b2, self.all_spins, self.lmax, self.lmax_mask, xis['SN']['00_1122'], xis['SN']['00_1221'], xis['SN']['02_1122'], xis['SN']['02_1221'], xis['SN']['22P_1122'], xis['SN']['22P_1221'], xis['SN']['22M_1122'], xis['SN']['22M_1221']) self.has_SN = np.array([self.wsp_SN.has_1122 > 0, self.wsp_SN.has_1221 > 0]) if xis['NS'] is not None: if self.wsp_NS is not None: lib.covar_workspace_free(self.wsp_NS) self.wsp_NS = None self.wsp_NS = lib.covar_workspace_init_from_xi( self.spin_a1, self.spin_a2, self.spin_b1, self.spin_b2, self.all_spins, self.lmax, self.lmax_mask, xis['NS']['00_1122'], xis['NS']['00_1221'], xis['NS']['02_1122'], xis['NS']['02_1221'], xis['NS']['22P_1122'], xis['NS']['22P_1221'], xis['NS']['22M_1122'], xis['NS']['22M_1221']) self.has_NS = np.array([self.wsp_NS.has_1122 > 0, self.wsp_NS.has_1221 > 0]) if xis['NN'] is not None: if self.wsp_NN is not None: lib.covar_workspace_free(self.wsp_NN) self.wsp_NN = None self.wsp_NN = lib.covar_workspace_init_from_xi( self.spin_a1, self.spin_a2, self.spin_b1, self.spin_b2, self.all_spins, self.lmax, self.lmax_mask, xis['NN']['00_1122'], xis['NN']['00_1221'], xis['NN']['02_1122'], xis['NN']['02_1221'], xis['NN']['22P_1122'], xis['NN']['22P_1221'], xis['NN']['22M_1122'], xis['NN']['22M_1221']) self.has_NN = np.array([self.wsp_NN.has_1122 > 0, self.wsp_NN.has_1221 > 0]) def _compute_coupling_coefficients(self, fla1, fla2, flb1, flb2, *, all_spins=False, l_toeplitz=-1, l_exact=-1, dl_band=-1): """ Computes coupling coefficients of the Gaussian covariance between the power spectra of two pairs of :class:`~pymaster.field.NmtField` objects (``fla1``, ``fla2``, ``flb1``, and ``flb2``). Note that you can reuse this workspace for the covariance of power spectra between any pairs of fields as long as the fields have the same masks as those passed to this function, and as long as the binning schemes used are also the same. Args: fla1 (:class:`~pymaster.field.NmtField`): First field contributing to the first power spectrum whose covariance you want to compute. fla2 (:class:`~pymaster.field.NmtField`): Second field contributing to the first power spectrum whose covariance you want to compute. flb1 (:class:`~pymaster.field.NmtField`): As ``fla1`` for the second power spectrum. flb2 (:class:`~pymaster.field.NmtField`): As ``fla2`` for the second power spectrum. all_spins (:obj:`bool`): If ``True``, coupling coefficients for all spin combinations will be calculated. Otherwise, only the spin combination determined by the input fields will be considered. l_toeplitz (:obj:`int`): If a positive number, the Toeplitz approximation described in `Louis et al. 2020 <https://arxiv.org/abs/2010.14344>`_ will be used. In that case, this quantity corresponds to :math:`\\ell_{\\rm toeplitz}` in Fig. 3 of that paper. l_exact (:obj:`int`): If ``l_toeplitz>0``, it corresponds to :math:`\\ell_{\\rm exact}` in Fig. 3 of the paper. Ignored if ``l_toeplitz<=0``. dl_band (:obj:`int`): If ``l_toeplitz>0``, this quantity corresponds to :math:`\\Delta \\ell_{\\rm band}` in Fig. 3 of the paper. Ignored if ``l_toeplitz<=0``. """ self.has_SN = np.array([False, False]) self.has_NS = np.array([False, False]) self.has_NN = np.array([False, False]) if np.any([fla1.anisotropic_mask, fla2.anisotropic_mask, flb1.anisotropic_mask, flb2.anisotropic_mask]): raise NotImplementedError("Covariance matrix estimation not " "implemented for anisotropic weights.") lmax = fla1.ainfo.lmax lmax_mask = fla1.ainfo_mask.lmax self.lmax = lmax self.lmax_mask = lmax_mask ut._toeplitz_sanity(l_toeplitz, l_exact, dl_band, lmax, fla1, flb1) if self.wsp is not None: lib.covar_workspace_free(self.wsp) self.wsp = None def get_wsp(pcl_1122, pcl_1221, has_1122, has_1221): wsp = lib.covar_workspace_init_py(int(fla1.spin), int(fla2.spin), int(flb1.spin), int(flb2.spin), pcl_1122, pcl_1221, int(all_spins), 0, int(has_1122), int(has_1221), int(fla1.ainfo.lmax), int(fla1.ainfo_mask.lmax), l_toeplitz, l_exact, dl_band) return wsp s11_lm, _ = _get_mask_prod_alm(fla1, flb1) s22_lm, _ = _get_mask_prod_alm(fla2, flb2) s12_lm, _ = _get_mask_prod_alm(fla1, flb2) s21_lm, _ = _get_mask_prod_alm(fla2, flb1) pcl_mask_S11_S22 = hp.alm2cl(s11_lm, s22_lm, lmax=lmax_mask) pcl_mask_S12_S21 = hp.alm2cl(s12_lm, s21_lm, lmax=lmax_mask) self.wsp = get_wsp(pcl_mask_S11_S22, pcl_mask_S12_S21, 1, 1) # Compute coupling coefficients for catalog-based field combinations is_catalog_any = (_is_catalog(fla1) or _is_catalog(fla2) or _is_catalog(flb1) or _is_catalog(flb2)) if not is_catalog_any: return has_1122_NS = has_1221_NS = has_1122_SN = has_1221_SN = False has_1122_NN = has_1221_NN = False pcl_mask_N11_S22 = np.zeros_like(pcl_mask_S11_S22) pcl_mask_N12_S21 = np.zeros_like(pcl_mask_S11_S22) pcl_mask_S11_N22 = np.zeros_like(pcl_mask_S11_S22) pcl_mask_S12_N21 = np.zeros_like(pcl_mask_S11_S22) pcl_mask_N11_N22 = np.zeros_like(pcl_mask_S11_S22) pcl_mask_N12_N21 = np.zeros_like(pcl_mask_S11_S22) lmx = fla1.ainfo_mask.lmax n11_lm = None n22_lm = None if ((fla1 is flb1) or (fla1 is flb2)) and _is_catalog(fla1): n11_lm = fla1.get_catalog_variance_alm() if ((fla2 is flb1) or (fla2 is flb2)) and _is_catalog(fla2): if (n11_lm is not None) and (fla2 is fla1): n22_lm = n11_lm else: n22_lm = fla2.get_catalog_variance_alm() # Here's some horrible combinatorics if fla1 is flb1 and _is_catalog(fla1) and _is_catalog(flb1): has_1122_NS = True # Calculate pcl_mask_N11_S22 pcl_mask_N11_S22 = hp.alm2cl(n11_lm, s22_lm, lmax=lmx) if fla2 is flb2 and _is_catalog(fla2) and _is_catalog(flb2): has_1122_NN = True # Calculate pcl_mask_N11_N22 pcl_mask_N11_N22 = hp.alm2cl(n11_lm, n22_lm, lmax=lmx) if fla1 is fla2 and not fla1.is_clustering: # Correct the four-point cumulant prefac = 1/(4*np.pi) corr_noise = prefac * np.sum( (np.sum(fla1.field**2, axis=0)/fla1.nmaps)**2 ) pcl_mask_N11_N22 = pcl_mask_N11_N22 - corr_noise if fla2 is flb2 and _is_catalog(fla2) and _is_catalog(flb2): has_1122_SN = True # Calculate pcl_mask_S11_N22 pcl_mask_S11_N22 = hp.alm2cl(s11_lm, n22_lm) if fla1 is flb2 and _is_catalog(fla1) and _is_catalog(flb2): has_1221_NS = True # Calculate pcl_mask_N12_S21 pcl_mask_N12_S21 = hp.alm2cl(n11_lm, s21_lm, lmax=lmx) if fla2 is flb1 and _is_catalog(fla2) and _is_catalog(flb1): has_1221_NN = True # Calcuate pcl_mask_N12_N21 pcl_mask_N12_N21 = hp.alm2cl(n11_lm, n22_lm, lmax=lmx) if fla1 is fla2 and not fla1.is_clustering: # Correct the four-point cumulant prefac = 1/(4*np.pi) corr_noise = prefac * np.sum( (np.sum(fla1.field**2, axis=0)/fla1.nmaps)**2 ) pcl_mask_N12_N21 = pcl_mask_N12_N21 - corr_noise if fla2 is flb1 and _is_catalog(fla1) and _is_catalog(flb1): has_1221_SN = True # Calculate pcl_mask_S12_N21 pcl_mask_S12_N21 = hp.alm2cl(s12_lm, n22_lm) self.has_NS = np.array([has_1122_NS, has_1221_NS]) self.has_SN = np.array([has_1122_SN, has_1221_SN]) self.has_NN = np.array([has_1122_NN, has_1221_NN]) # TODO: we are not taking advantage of cases # when fla1=fla2 or flb1=flb2 if self.has_NS.any(): self.wsp_NS = get_wsp(pcl_mask_N11_S22, pcl_mask_N12_S21, has_1122_NS, has_1221_NS) if self.has_SN.any(): self.wsp_SN = get_wsp(pcl_mask_S11_N22, pcl_mask_S12_N21, has_1122_SN, has_1221_SN) if self.has_NN.any(): self.wsp_NN = get_wsp(pcl_mask_N11_N22, pcl_mask_N12_N21, has_1122_NN, has_1221_NN)
[docs] def write_to(self, fname): """ Writes the contents of an :obj:`NmtCovarianceWorkspace` object to a FITS file. Args: fname (:obj:`str`): Output file name.""" import fitsio as fts # Read header with global information f = fts.FITS(fname, 'rw', clobber=True) h = {'LMAX': self.wsp.lmax, 'LMAX_MASK': self.wsp.lmax_mask, 'ALL_SPINS': self.wsp.all_spins, 'SPIN_A1': self.wsp.spin_a1, 'SPIN_A2': self.wsp.spin_a2, 'SPIN_B1': self.wsp.spin_b1, 'SPIN_B2': self.wsp.spin_b2} f.write(np.ones((1, 1)), header=h, extname='CWSP_PRIMARY') def write_wsp(w, prefix): # This function writes the coupling coefficients of a # workspace to a FITS HDU. if w is None: return for i, n in enumerate(['00_1122', '00_1221', '02_1122', '02_1221', '22P_1122', '22P_1221', '22M_1122', '22M_1221']): exists, xi = lib.get_cw_xi(w, i, (w.lmax+1)**2) if exists: f.write(xi.reshape((w.lmax+1, w.lmax+1)), extname=f'XI{prefix + n}') # Write the coupling coefficients of all workspaces to the FITS file write_wsp(self.wsp, '') write_wsp(self.wsp_SN, 'SN') write_wsp(self.wsp_NS, 'NS') write_wsp(self.wsp_NN, 'NN') f.close()
[docs] def gaussian_covariance(self, cla1b1, cla1b2, cla2b1, cla2b2, wa, wb=None, coupled=False, spins=None): """ Computes the Gaussian covariance matrix for power spectra using the information precomputed in this :class:`NmtCovarianceWorkspace` object). Let us call the four fields used to initialise this workspace `a1`, `a2`, `b1`, and `b2`, corresponding to the two pairs of fields whose power spectra we want the covariance of. These power spectra should have been computed using two :class:`~pymaster.workspaces.NmtWorkspace` objects, ``wa`` and ``wb``, which must be passed as arguments of this method (the power spectrum for fields `a1` and `a2` was computed with ``wa``, and that of `b1` and `b2` with ``wb``). Using the same notation, ``clXnYm`` should be a prediction for the power spectrum between fields `Xn` and `Ym`. These predicted input power spectra should be defined for all multipoles :math:`\\ell` up to the :math:`\\ell_{\\rm max}` with which all fields were constructed. .. note:: Note that, as suggested in `Nicola et al. 2020 <https://arxiv.org/abs/2010.09717>`_ (the so-called "improved narrow-kernel approximation" - iNKA), an optimal choice for the input power spectra would be the mode-coupled version of the true power spectra of the corresponding fields divided by the average of the product of the associated masks across the sky (Eq. 2.36 in the paper). Often, a good substitute for this can be obtained as the pseudo-:math:`C_\\ell` of the associated maps (e.g. computed via :meth:`~pymaster.workspaces.compute_coupled_cell`), divided by the same mean mask product. The convenience function :meth:`get_iNKA_cell` may be used to calculate this spectrum under the iNKA. Args: cla1b1 (`array`): Prediction for the cross-power spectrum between fields `a1` and `b1`. cla1b2 (`array`): As `cla1b1` for fields `a1` and `b2`. cla2b1 (`array`): As `cla1b1` for fields `a2` and `b1`. cla2b2 (`array`): As `cla1b1` for fields `a2` and `b2`. wa (:class:`~pymaster.workspaces.NmtWorkspace`): Workspace containing the mode-coupling matrix for the first power spectrum (that of fields `a1` and `a2`). wb (:class:`~pymaster.workspaces.NmtWorkspace`): As ``wa`` for the second power spectrum (that of fields `b1` and `b2`). If ``None``, ``wa`` will be used instead. coupled (:obj:`bool`): If ``True``, the covariance matrix of the mode-coupled pseudo-:math:`C_\\ell` s will be computed. Otherwise it'll be the covariance of mode-decoupled bandpowers. spins (`array`): A list of 4 integers containing the spins of the fields whose power spectrum covariance one wishes to calculate. Note that you can only select arbitrary spin combinations if you created this object using ``all_spins=True``. If ``None``, the spin combination is determined by the fields used to create this object. """ if spins is not None: if not self.all_spins: if ((spins[0] != self.spin_a1) or (spins[1] != self.spin_a2) or (spins[2] != self.spin_b1) or (spins[3] != self.spin_b2)): raise ValueError( "The input spins do not coincide with those of " "the fields used to initialise this object. If " "you want to use arbitrary spin combinations, " "use `all_spins=True` when initialising this " "class.") if len(spins) != 4: raise ValueError("`spins` must have 4 elements.") spin_a1, spin_a2, spin_b1, spin_b2 = spins else: spin_a1 = self.spin_a1 spin_a2 = self.spin_a2 spin_b1 = self.spin_b1 spin_b2 = self.spin_b2 nm_a1 = 2 if spin_a1 else 1 nm_a2 = 2 if spin_a2 else 1 nm_b1 = 2 if spin_b1 else 1 nm_b2 = 2 if spin_b2 else 1 if wb is None: wb = wa if (wa.wsp.ncls != nm_a1*nm_a2) or (wb.wsp.ncls != nm_b1*nm_b2): raise ValueError("Field spins do not match input workspaces") if (len(cla1b1) != nm_a1*nm_b1) or \ (len(cla1b2) != nm_a1*nm_b2) or \ (len(cla2b1) != nm_a2*nm_b1) or \ (len(cla2b2) != nm_a2*nm_b2): raise ValueError("Field spins do not match input power" "spectrum shapes") if (len(cla1b1[0]) < self.wsp.lmax + 1) or \ (len(cla1b2[0]) < self.wsp.lmax + 1) or \ (len(cla2b1[0]) < self.wsp.lmax + 1) or \ (len(cla2b2[0]) < self.wsp.lmax + 1): raise ValueError("Input C_ls have a weird length. " f"Expected {self.wsp.lmax+1}, but got " f"({len(cla1b1[0])}, {len(cla1b2[0])}, " f"{len(cla2b1[0])}, {len(cla2b2[0])}).") if coupled: len_a = wa.wsp.ncls * (self.wsp.lmax+1) len_b = wb.wsp.ncls * (self.wsp.lmax+1) wa.check_unbinned() wb.check_unbinned() covar_SS = lib.comp_gaussian_covariance_coupled( self.wsp, int(spin_a1), int(spin_a2), int(spin_b1), int(spin_b2), wa.wsp, wb.wsp, cla1b1, cla1b2, cla2b1, cla2b2, 0, 0, 0, 0, len_a * len_b ) covar_NN = covar_NS = covar_SN = np.zeros_like(covar_SS) if self.has_NN.any(): covar_NN = lib.comp_gaussian_covariance_coupled( self.wsp_NN, int(spin_a1), int(spin_a2), int(spin_b1), int(spin_b2), wa.wsp, wb.wsp, np.ones_like(cla1b1), np.ones_like(cla1b2), np.ones_like(cla2b1), np.ones_like(cla2b2), 1, 1, 1, 1, len_a * len_b) if self.has_NS.any(): covar_NS = lib.comp_gaussian_covariance_coupled( self.wsp_NS, int(spin_a1), int(spin_a2), int(spin_b1), int(spin_b2), wa.wsp, wb.wsp, np.ones_like(cla1b1), np.ones_like(cla1b2), cla2b1, cla2b2, 1, 1, 0, 0, len_a * len_b) if self.has_SN.any(): covar_SN = lib.comp_gaussian_covariance_coupled( self.wsp_SN, int(spin_a1), int(spin_a2), int(spin_b1), int(spin_b2), wa.wsp, wb.wsp, cla1b1, cla1b2, np.ones_like(cla2b1), np.ones_like(cla2b2), 0, 0, 1, 1, len_a * len_b) else: len_a = wa.wsp.ncls * wa.wsp.bin.n_bands len_b = wb.wsp.ncls * wb.wsp.bin.n_bands covar_SS = lib.comp_gaussian_covariance( self.wsp, int(spin_a1), int(spin_a2), int(spin_b1), int(spin_b2), wa.wsp, wb.wsp, cla1b1, cla1b2, cla2b1, cla2b2, 0, 0, 0, 0, len_a * len_b ) covar_NN = covar_NS = covar_SN = np.zeros_like(covar_SS) if self.has_NN.any(): covar_NN = lib.comp_gaussian_covariance( self.wsp_NN, int(spin_a1), int(spin_a2), int(spin_b1), int(spin_b2), wa.wsp, wb.wsp, np.ones_like(cla1b1), np.ones_like(cla1b2), np.ones_like(cla2b1), np.ones_like(cla2b2), 1, 1, 1, 1, len_a * len_b) if self.has_NS.any(): covar_NS = lib.comp_gaussian_covariance( self.wsp_NS, int(spin_a1), int(spin_a2), int(spin_b1), int(spin_b2), wa.wsp, wb.wsp, np.ones_like(cla1b1), np.ones_like(cla1b2), cla2b1, cla2b2, 1, 1, 0, 0, len_a * len_b) if self.has_SN.any(): covar_SN = lib.comp_gaussian_covariance( self.wsp_SN, int(spin_a1), int(spin_a2), int(spin_b1), int(spin_b2), wa.wsp, wb.wsp, cla1b1, cla1b2, np.ones_like(cla2b1), np.ones_like(cla2b2), 0, 0, 1, 1, len_a * len_b) covar = covar_SS+covar_SN+covar_NS+covar_NN return covar.reshape([len_a, len_b])
[docs] class NmtCovarianceWorkspaceFlat(object): """ :obj:`NmtCovarianceWorkspaceFlat` objects are used to compute and store the coupling coefficients needed to calculate the Gaussian covariance matrix of angular power spectra using a flat-sky version of the approximations described in `Garcia-Garcia et al. 2019 <https://arxiv.org/abs/1906.11765>`_. When initialized, this object is practically empty. The information describing the coupling coefficients must be computed or read from a file afterwards. """ def __init__(self): self.wsp = None def __del__(self): if self.wsp is not None: if lib.covar_workspace_flat_free is not None: lib.covar_workspace_flat_free(self.wsp) self.wsp = None
[docs] def read_from(self, fname): """ Reads the contents of an :obj:`NmtCovarianceWorkspaceFlat` object from a FITS file. Args: fname (:obj:`str`): Input file name. """ if self.wsp is not None: lib.covar_workspace_flat_free(self.wsp) self.wsp = None self.wsp = lib.read_covar_workspace_flat(fname)
[docs] def compute_coupling_coefficients(self, fla1, fla2, bin_a, flb1=None, flb2=None, bin_b=None): """ Computes coupling coefficients of the Gaussian covariance between the power spectra of two pairs of :class:`~pymaster.field.NmtFieldFlat` objects (``fla1``, ``fla2``, ``flb1``, and ``flb2``). Note that you can reuse this workspace for the covariance of power spectra between any pairs of fields as long as the fields have the same masks as those passed to this function, and as long as the binning schemes used are also the same. Args: fla1 (:class:`~pymaster.field.NmtFieldFlat`): First field contributing to the first power spectrum whose covariance you want to compute. fla2 (:class:`~pymaster.field.NmtFieldFlat`): Second field contributing to the first power spectrum whose covariance you want to compute. bin_a (:class:`~pymaster.bins.NmtBinFlat`): Binning scheme for the first power spectrum. flb1 (:class:`~pymaster.field.NmtFieldFlat`): As ``fla1`` for the second power spectrum. If ``None``, it will be set to ``fla1``. flb2 (:class:`~pymaster.field.NmtFieldFlat`): As ``fla2`` for the second power spectrum. If ``None``, it will be set to ``fla2``. bin_b (:class:`~pymaster.bins.NmtBinFlat`): Binning scheme for the second power spectrum. If ``None``, ``bin_a`` will be used. """ if flb1 is None: flb1 = fla1 if flb2 is None: flb2 = fla2 if bin_b is None: bin_b = bin_a if (fla1.fl.fs.nx != fla2.fl.fs.nx) or \ (fla1.fl.fs.ny != fla2.fl.fs.ny) or \ (fla1.fl.fs.nx != flb1.fl.fs.nx) or \ (fla1.fl.fs.ny != flb1.fl.fs.ny) or \ (fla1.fl.fs.nx != flb2.fl.fs.nx) or \ (fla1.fl.fs.ny != flb2.fl.fs.ny): raise ValueError("Everything should have the same resolution!") if self.wsp is not None: lib.covar_workspace_flat_free(self.wsp) self.wsp = None self.wsp = lib.covar_workspace_flat_init_py(fla1.fl, fla2.fl, bin_a.bin, flb1.fl, flb2.fl, bin_b.bin)
[docs] def write_to(self, fname): """ Writes the contents of an :obj:`NmtCovarianceWorkspaceFlat` object to a FITS file. Args: fname (:obj:`str`): Output file name. """ if self.wsp is None: raise ValueError("Must initialize workspace before writing") lib.write_covar_workspace_flat(self.wsp, "!"+fname)
[docs] def gaussian_covariance(self, spin_a1, spin_a2, spin_b1, spin_b2, larr, cla1b1, cla1b2, cla2b1, cla2b2, wa, wb=None): """ As :meth:`NmtCovarianceWorkspace.gaussian_covariance` but for the flat-sky versions of all quantities involved. The only difference with is that all power spectra must have been sampled at the input multipoles ``larr``, and the spins of all fields must be specified. """ nm_a1 = 2 if spin_a1 else 1 nm_a2 = 2 if spin_a2 else 1 nm_b1 = 2 if spin_b1 else 1 nm_b2 = 2 if spin_b2 else 1 if wb is None: wb = wa if (wa.wsp.ncls != nm_a1*nm_a2) or (wb.wsp.ncls != nm_b1*nm_b2): raise ValueError("Input spins do not match input workspaces") if (len(cla1b1) != nm_a1*nm_b1) or \ (len(cla1b2) != nm_a1*nm_b2) or \ (len(cla2b1) != nm_a2*nm_b1) or \ (len(cla2b2) != nm_a2*nm_b2): raise ValueError("Input spins do not match input power" "spectrum shapes") if ( (len(cla1b1[0]) != len(larr)) or (len(cla1b2[0]) != len(larr)) or (len(cla2b1[0]) != len(larr)) or (len(cla2b2[0]) != len(larr)) ): raise ValueError("Input C_ls have a weird length. " f"Expected {len(larr)}, but got " f"({len(cla1b1[0])}, {len(cla1b2[0])}, " f"{len(cla2b1[0])}, {len(cla2b2[0])}).") len_a = wa.wsp.ncls * self.wsp.bin.n_bands len_b = wb.wsp.ncls * self.wsp.bin.n_bands covar1d = lib.comp_gaussian_covariance_flat( self.wsp, spin_a1, spin_a2, spin_b1, spin_b2, wa.wsp, wb.wsp, larr, cla1b1, cla1b2, cla2b1, cla2b2, len_a * len_b) covar = np.reshape(covar1d, [len_a, len_b]) return covar
[docs] def gaussian_covariance(cw, spin_a1, spin_a2, spin_b1, spin_b2, cla1b1, cla1b2, cla2b1, cla2b2, wa, wb=None, coupled=False): """ Computes the Gaussian covariance matrix for power spectra using the information precomputed in cw (a :class:`NmtCovarianceWorkspace` object). ``cw`` should have been initialized using four :class:`~pymaster.field.NmtField` objects (let's call them `a1`, `a2`, `b1`, and `b2`), corresponding to the two pairs of fields whose power spectra we want the covariance of. These power spectra should have been computed using two :class:`~pymaster.workspaces.NmtWorkspace` objects, ``wa`` and ``wb``, which must be passed as arguments of this function (the power spectrum for fields `a1` and `a2` was computed with ``wa``, and that of `b1` and `b2` with ``wb``). Using the same notation, ``clXnYm`` should be a prediction for the power spectrum between fields `Xn` and `Ym`. These predicted input power spectra should be defined for all multipoles :math:`\\ell` up to the :math:`\\ell_{\\rm max}` with which all fields were constructed. .. warning:: This function is deprecated and will be removed in a future version of NaMaster. Use the :meth:`NmtCovarianceWorkspace.gaussian_covariance` method instead. Args: cw (:obj:`NmtCovarianceWorkspace`): Workspace containing the precomputed coupling coefficients. spin_a1 (:obj:`int`): Spin of field `a1`. spin_a2 (:obj:`int`): Spin of field `a2`. spin_b1 (:obj:`int`): Spin of field `b1`. spin_b2 (:obj:`int`): Spin of field `b2`. cla1b1 (`array`): Prediction for the cross-power spectrum between fields `a1` and `b1`. cla1b2 (`array`): As `cla1b1` for fields `a1` and `b2`. cla2b1 (`array`): As `cla1b1` for fields `a2` and `b1`. cla2b2 (`array`): As `cla1b1` for fields `a2` and `b2`. wa (:class:`~pymaster.workspaces.NmtWorkspace`): Workspace containing the mode-coupling matrix for the first power spectrum (that of fields `a1` and `a2`). wb (:class:`~pymaster.workspaces.NmtWorkspace`): As ``wa`` for the second power spectrum (that of fields `b1` and `b2`). If ``None``, ``wa`` will be used instead. coupled (:obj:`bool`): If ``True``, the covariance matrix of the mode-coupled pseudo-:math:`C_\\ell` s will be computed. Otherwise it'll be the covariance of mode-decoupled bandpowers. """ return cw.gaussian_covariance(cla1b1, cla1b2, cla2b1, cla2b2, wa, wb=wb, coupled=coupled, spins=[spin_a1, spin_a2, spin_b1, spin_b2])
[docs] def gaussian_covariance_flat(cw, spin_a1, spin_a2, spin_b1, spin_b2, larr, cla1b1, cla1b2, cla2b1, cla2b2, wa, wb=None): """ As :meth:`gaussian_covariance` but for the flat-sky versions of all quantities involved. The only difference with :meth:`gaussian_covariance` is that all power spectra must have been sampled at the input multipoles ``larr``. .. warning:: This function is deprecated and will be removed in a future version of NaMaster. Use the :meth:`NmtCovarianceWorkspaceFlat.gaussian_covariance` method instead. """ return cw.gaussian_covariance(spin_a1, spin_a2, spin_b1, spin_b2, larr, cla1b1, cla1b2, cla2b1, cla2b2, wa, wb=wb)
def _is_catalog(f): return isinstance(f, NmtFieldCatalog) def _is_mask_catalog(f): if isinstance(f, NmtFieldCatalog): if f.mask is not None: return False return True return False
[docs] def get_iNKA_cell(fla, flb, cl_guess=None, w=None): """ Returns the power spectrum that should be used in the calculation of the Gaussian covariance matrix according to the improved Narrow-Kernel Approximation (iNKA) of `Nicola et al. 2020 <https://arxiv.org/abs/2010.09717>`_. This can then be used, for instance, as input for :meth:`NmtCovarianceWorkspace.gaussian_covariance`. The two fields whose power spectra we need must be compatible. This means that, at least, they must be represented in harmonic space up to the same maximum multipole. If they are also compatible at the map level, the effective sky fraction used in the iNKA will be calculated from the product of their masks. Otherwise, their harmonic-space spectrum will be used. Args: fla (:class:`~pymaster.field.NmtField`): First field whose power spectrum we want to calculate. flb (:class:`~pymaster.field.NmtField`): Second field whose power spectrum we want to calculate. cl_guess (`array`): A guess for the true power spectra between ``fla`` and ``flb``. The number of power spectra must correspond to the spins of the two fields in question. If ``None``, the pseudo-:math:`C_\\ell` between the two fields will be used instead. w (:class:`~pymaster.workspaces.NmtWorkspace`): Workspace containing the mode-coupling matrix for these two fields. This is only required if ``cl_guess`` is not ``None``. If needed but ``None``, the mode-coupling matrix will be calculated on the fly. Returns: (`array`): power spectrum to be used in covariance calculations. """ if not fla.is_compatible(flb, strict=False): raise ValueError("Fields have incompatible pixelizations") # 1. Compute fsky as the mean of the mask product. # If both fields are compatible at the map level, just take # the product of their maps and average. Otherwise use # Parseval's theorem and do it from their harmonic spectrum. use_map_product = fla.is_compatible(flb) if use_map_product: wawb = np.mean(fla.get_mask()*flb.get_mask()) else: lmax = fla.ainfo_mask.lmax walm = fla.get_mask_alms() wblm = flb.get_mask_alms() clw = hp.alm2cl(walm, wblm, lmax=lmax) ls = np.arange(lmax+1) # Correct for catalogs if _is_catalog(fla) and _is_catalog(flb): phi_a = 1 if fla.mask is not None else fla.get_cloud_kernel(lmax) phi_b = 1 if flb.mask is not None else flb.get_cloud_kernel(lmax) # Subtract shot noise if fla is flb: clw = clw - fla.Nw # Multiply by kernels clw = clw * phi_a * phi_b wawb = np.sum((2*ls+1)*clw)/(4*np.pi) # 2. Compute pseudo-Cl # If no guess Cl is provided, compute it from the data. if cl_guess is None: pcl_ab = compute_coupled_cell(fla, flb) # Note that we don't need to worry abot catalogs # here, since the function above already subtracts # the shot-noise contribution. else: # We'll need to calculate the MCM if not available if w is None: # Just some token bins that go to the right lmax b = NmtBin.from_lmax_linear( fla.ainfo.lmax, nlb=int(fla.ainfo.lmax//10)) w = NmtWorkspace.from_fields(fla, flb, b) pcl_ab = w.couple_cell(cl_guess) # 3. Return ratio return pcl_ab / wawb