Delensed Covariance¶
Measures the empirical per-mode covariance of the delensed field, and its finite difference in theta. It is the field block of the Fisher forecast, which cannot run without it.
Each job uses one seed and runs at every stencil point \(\theta_0\) and \(\theta_0 \pm h_i\):
load_sim(theta)at the chosenpol(common random numbers: at one seed the fields differ only through the cosmology),- inverse-lens the noiseless lensed field by the Wiener-filtered true phi \(W\phi\), with \(W = C_\phi / (C_\phi + N_\phi)\) built once at \(\theta_0\),
- store the per-mode covariance \(\mathrm{Re}(X Y^*) / n_\text{side}^2\) of every component on the rfft grid.
pol |
Stored components |
|---|---|
I |
TT |
P |
EE, BB |
IP |
TT, TE, EE, BB |
The merge is the realization mean at each stencil point.
Choosing N_phi for the Wiener Filter¶
| Driver Setting | What it Controls |
|---|---|
phi_noise="", qe_filter=unlensed (default) or lensed |
the box quadratic-estimate \(N^{(0)}\) (the TT estimator for I, the EB estimator for P and IP), with the unlensed or lensed spectra in its filter; the response is always unlensed |
phi_noise=<path> |
the nphi of a merged map_joint or sampler phi-noise file |
Match it to the forecast it will feed: nothing checks that the file's W was built from the same N_phi as the forecast's phi block.
Running it¶
bash get_delensed_field_covariance/get_delensed_field_covariance.sh
#copy delensed_covariance_output/ to your machine, then
python get_delensed_field_covariance/merge_delensed_field_covariance.py --covariance_dir <delensed_covariance_output>
Other settings in the driver: pol (I, P or IP; the forecast must use the same), derivative_params (the forecast's --params must be drawn from this list), step_sigma (0.05, in units of PARAM_SIGMA; accurate at nside 128 / 2.5' / 2.5 to 10 uK-arcmin, so for another box or noise level rerun with the step halved and check that the forecast does not change substantially), num_realizations, seed_prefix, out_dir_name (the output folder's name inside get_delensed_field_covariance/). A 3-point temperature job at nside 128 takes ~20 s.
Every setting has a default in the driver and can be overridden on the command line as --<name>=<value>, e.g.
bash get_delensed_field_covariance/get_delensed_field_covariance.sh --noise_level=10 --pol=IP \
--derivative_params="omch2 logA" --out_dir_name=delensed_covariance_output_IP_10uk
A misspelled name is ignored and the default used. Each run writes the submission date and every setting to submission_settings.txt in the output folder (replacing any earlier copy); use a new out_dir_name when you change settings.
Output: delensed_covariance.npz¶
| Key | Content |
|---|---|
pol, components |
the data's polarization and the stored components, in order |
delensed_fid |
the delensed covariance at \(\theta_0\), shape (n_components, nside, nside // 2 + 1) |
delensed_plus, delensed_minus |
the same at \(\theta_0 \pm h_i\), shape (n_params, n_components, nside, nside // 2 + 1) |
names, steps |
the differentiated parameters (in the package's standard order) and their steps \(h_i\) |
| metadata | the box, the stencil points, the seeds and how W was built |
Merged temperature files written before 2026-11-11 have no components key and no component axis; the forecast still reads them.