CleanCMB.jl

Algorithms to remove foreground emission and obtain clean maps of the cosmic microwave background
Author komatsu5147
Github
Popularity
5 Stars
Updated Last
5 Months Ago
Started In
June 2020

CleanCMB

This package contains functions to enable extraction of clean maps of the cosmic microwave background (CMB). Some functions can be used to extract non-CMB astrophysical components as well.

Different algorithms exist for extraction of clean maps of the CMB (as well as of astrophysical components). The package currently supports:

  • Internal Linear Combination (ILC) Method
  • Parametric Maximum Likelihood Method

See this note for the relationship between them.

Internal Linear Combination (ILC) Method

  • ilc_weights(cij[, e, ℓid=3]): return weights for the internal linear combination (ILC) method, following Equation (12) of Tegmark et al., PRD, 68, 123523 (2003).
  • cilc_weights(cij, a, b[, ℓid=3]): return weights for the constrained internal linear combination (CILC) method for two components, following Equation (20) of Remazeilles, et al., MNRAS, 410, 2481 (2011).
  • milca_weights(cij, a, B[, ℓid=3]): return weights for the modified internal linear combination algorithm (MILCA) method (CILC for N components), following Equation (15) of Hurier, et al., A&A, 558, A118 (2013).
    • Note: These papers define weights in various domain, including harmonic, wavelet, and pixel domain. You can choose to work in any domain by supplying a covariance matrix cij in the appropriate domain.
  • ilc_clean_cij(cij, w): return a covariance matrix multiplied by weights, w' * cij * w, with w returned by any of the above functions, e.g., w = cilc_weights(cij, a, b).
    • This would be a variance of the extracted component if cij were the same as that used in ilc_weights(), cilc_weights() or milca_weights().
    • If cij is a noise covariance matrix, this function returns the noise variance of the cleaned map for each multipole, band-power bin, pixel, etc.

Arguments

  • cij::Array{<:AbstractFloat,2}: nν-by-nν symmetric covariance matrix, where nν is the number of frequency bands.
    • or, cij::Array{<:AbstractFloat,3}:symmetric covariance matrix with the dimention of (nℓ, nν, nν), (nν, nℓ, nν) or (nν, nν, nℓ) (default). Here, nℓ is the number of elements in the relevant domain, e.g., multipoles, band-power bins, pixels, etc.
    • The functions ilc_weights(), cilc_weights() and milca_weights() are multiple dispatch, which detect the dimension of the input cij automatically.
  • a::Array{<:AbstractFloat,1}: vector of the frequency response, for the component to be extracted. E.g., a=[1,...,1] for CMB. The number of elements is nν.
  • b::Array{<:AbstractFloat,1}: vector of the frequency response, for the component to be nulled. The number of elements is nν.
  • B::Array{<:AbstractFloat,2}: nν-by-nrc matrix of the frequency response, for nrc components to be nulled.

Optional arguments

  • e::Array{<:AbstractFloat,1}: vector of the frequency response, for the component to be extracted. The default value is e=[1,...,1], i.e., CMB. The number of elements is nν.
  • ℓid::Integer=3: location of the index for the nℓ domain, if the dimension of the input cij is (nℓ, nν, nν), (nν, nℓ, nν) or (nν, nν, nℓ). ℓid=1 if cij[nℓ,nν,nν], ℓid=2 if cij[nν,nℓ,nν], and ℓid=3 (the default value) if cij[nν,nν,nℓ].

Parametric Maximum Likelihood Method

  • loglike_beta(nij, A, d) or loglike_beta(nij, A, cij):: return log(likelihood) for frequency response vectors A given a data vector d or data covariance matrix cij, based on Equation (9) of Stompor et al., MNRAS, 392, 216 (2009).

Arguments

  • nij::Array{<:AbstractFloat,2}: nν-by-nν symmetric noise covariance matrix, where nν is the number of frequency bands.
  • A::Array{<:AbstractFloat,2}: nν-by-nc matrix of the frequency response, for nc components in sky.
    • E.g., A = [a B] where a = ones(nν) for CMB and B is a nν-by-nc-1 matrix for the frequency response of foreground components.
  • d::Array{<:AbstractFloat,1}: data vector for a given pixel, (ℓ,m), or any other appropriate domain. The number of elements is nν.
  • cij::Array{<:AbstractFloat,2}: nν-by-nν symmetric covariance matrix for a given multipole, pixel, or any other appropriate domain.

Foreground models

The package contains the following functions to return frequency dependence of foreground components:

  • tsz(νGHz; units="cmb", Tcmb=2.7255): Spectrum of the thermal Sunyaev-Zeldovich effect given in Equation (V) in Appendix of Zeldovich, Sunyaev, Astrophys. Space Sci. 4, 301 (1969).
  • dust1(νGHz; Td=19.6, βd=1.6, νd=353, units="cmb", Tcmb=2.7255): Spectrum of 1-component modified black-body thermal dust emission. The output is normalized to unity at νGHz = νd.
  • synch(νGHz; βs=-3, νs=23, Cs=0, νC=40, units="cmb", Tcmb=2.7255): Spectrum of synchrotron emission. The output is normalized to unity at νGHz = νs.

Arguments

  • νGHz::Real: frequency in units of GHz.

Optional keyword arguments

  • units::String: units of the spectrum. For Rayleigh-Jeans temperature (brightness temperature) units, set units = "rj". The default is the CMB units.
  • Tcmb::Real: present-day temperature of the CMB in units of Kelvin. The default value is 2.7255.
  • Td::Real: dust temperature. The default value is 19.6 (Kelvin).
  • βd::Real: dust emissivity index. The default value is 1.6.
  • νd::Real: frequency at which the output dust spectrum is normalized to unity. The default value is 353 (GHz).
  • βs::Real: synchrotron power-law index. The default is -3.
  • νs::Real: frequency at which the output synchrotron spectrum is normalized to unity. The default is 23 (GHz).
  • Cs::Real: curvature of the synchrotron spectrum. The default value is 0.
  • νC::Real: pivot frequency for curvature of the synchrotron spectrum. The default value is 40 (GHz).

Example Julia Codes

  • examples/CleanWMAP.jl
    • This code applies ilc_weights() in pixel domain to produce a clean map of CMB from the temperature maps of Wilkinson Microwave Anisotropy Probe (WMAP) at five frequency bands.
    • The code requires Healpix.jl.

Pipelines

Here we provide example codes for pipelines, which do everything from generation of simulated maps to estimation of the cosmological parameter (tensor-to-scalar ratio of the primordial gravitational waves) in one go.

  • examples/ILCPipelineSOSAT.jl

    • This code applies ilc_weights() in harmonic domain to produce power spectra of clean polarisation maps of the CMB.
    • The code requires Healpix.jl and Libsharp.jl. It also calls python wrappers pymaster and classy via PyCall.
    • This is a simulation pipeline for the Small Aperture Telescope (SAT) of the Simons Observatory. The code performs the following operations:
      1. Read in a hits map and calculate weights for inhomogeneous noise.
      2. include("compute_cl_class.jl") (see examples/compute_cl_class.jl) to generate theoretical scalar- and tensor-mode power spectra of the CMB using CLASS.
      3. Read in and smooth foreground maps by beams of the telescope.
      4. Generate Gaussian random realisations of the CMB and noise.
      5. Calculate covariance matrices cij of the foreground, noise, and CMB+foreground+noise maps.
      6. Calculate ILC weights in harmonic domain using ilc_weights().
      7. Calculate power spectra of the clean CMB maps using ilc_clean_cij().
      8. Show results for visual inspection, if showresults = true.
      9. Calculate the tensor-to-scalar ratio and its uncertainty from the simulated realisations.
      10. Write out the results (tensor-to-scalar ratios with and without residual foreground marginalisation) to a file ilc_results_sosat.csv and create a PDF figure ilc_clbb_sim_sosat.pdf showing the cleaned B-mode power spectrum, minus noisebias, and minus the residual foreground.
    • For your reference, the results from 300 realisations starting with the initial seed of 5147 are given in examples/results/ilc_results_sosat_300sims_seed5147.csv. You can compute the mean and standard deviation of the tensor-to-scalar ratios. You should find, for 30 < ℓ < 260:
      • Without FG marginalisation: r = (2.0 ± 1.6) x 10-3
      • With marginalisation: r = (-1.0 ± 2.7) x 10-3

  • examples/MILCAPipelineSOSAT.jl

    • This code performs a hybrid of the maximum likelihood and ILC methods. See this note for the relationship between them.

      • The code applies loglike_beta() in harmonic domain to find the best-fitting synchrotron and dust spectral indices (βs and βd), and calculates weights using the N-component constrained ILC milca_weights().
      • When calculating the weights, it uses the total covariance matrix cij (like for the ILC) rather than the noise covariance matrix nij (like for the maximum likelihood). This minimises further the foreground contribution that is not modeled.

    • The code performs the same operations as the above ILC pipeline for the Small Aperture Telescope (SAT) of the Simons Observatory, but replaces the steps f and g with

      1. Calculate the best-fitting βs and βd by minimising -loglike_beta() with respect to them.
      2. Calculate weights using milca_weights() with the best-fitting βs and βd, and calculate power spectra of the clean CMB maps using ilc_clean_cij().

    • More detail of the procedure:

      • The code finds the best-fitting βs and βd for each band-power at multipoles below the "switching" multipole, ℓ ≤ ℓswitch (the default value is ℓswitch = 30). This may better handle complex foreground properties on large angular scales.
      • For higher multipoles, the code finds the global βs and βd using the covariance matrix smoothed to a common resolution specified by smooth_FWHM (in units of degrees; the default value is smooth_FWHM = 0.5). This is equivalent to finding βs and βd on smoothed maps with the common resolution.

    • For your reference, the results from 300 realisations starting with the initial seed of 5147 are given in examples/results/milca_results_sosat_300sims_seed5147.csv. You can compute the mean and standard deviation of the tensor-to-scalar ratios. You should find, for 30 < ℓ < 260:

      • Without FG marginalisation: r = (1.2 ± 2.7) x 10-3
      • With marginalisation: r = (-0.8 ± 5.1) x 10-3

  • examples/ILCPipelineSOSATCCATp.jl

  • examples/MILCAPipelineSOSATCCATp.jl

Splitting Pipelines

The above codes do everything in one go. However, sometimes it is convenient to split the processes. For example, we do not have to re-generate maps and their covariance matrices everytime we make a small modification to the rest of the pipeline.

Here we provide example codes for splitting the pipelines into two pieces: (1) Generation of simulated maps and their covariance matrices, and (2) Application of foreground cleaning methods to the covariance matrices.

  • examples/GenerateCovMatrices.jl: Perform the steps (a)-(e) of the pipeline and write out the covariance matrices to binary files in arrays of (nν, nν, nbands) where nν is the number of frequency channels and nbands is the number of band-powers. It also writes out the binned scalar and tensor power spectra used to generate the simulations to text files.
  • examples/PerformILC.jl and examples/PerformMILCA.jl: Perform the steps (f)-(j) of of the pipeline and write out the estimated tensor-to-scalar ratios to a csv file.

Acknowledgment

Development of the functions provided in this package was supported in part by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC-2094 - 390783311 and JSPS KAKENHI Grant Number JP15H05896. The Kavli IPMU is supported by World Premier International Research Center Initiative (WPI), MEXT, Japan.