Fisher Forecast¶
A Gaussian Fisher forecast of the sampler's posterior widths, for temperature (I), polarization (P) or joint (IP) data, from per-mode covariance blocks on the sampler's own rfft grid:
with \(w_k\) the rfft mode weights (2 per bulk mode, 1 on the self-conjugate columns). For a single-number block the trace is \(\partial_i \ln C_k \, \partial_j \ln C_k\).
The blocks¶
--pol |
Field blocks | Phi block |
|---|---|---|
I |
\(C_{TT} + N_{TT}\) | \(C_\phi + N_\phi\) |
P |
\(C_{EE} + N_{EE}\); \(C_{BB} + N_{BB}\) | same |
IP |
\(\begin{pmatrix} C_{TT} + N_{TT} & C_{TE} \\ C_{TE} & C_{EE} + N_{EE} \end{pmatrix}\); \(C_{BB} + N_{BB}\) | same |
- The field blocks are always measured. Their signal at \(\theta_0\) and \(\theta_0 \pm h_i\) is the empirical delensed covariance of
--delensed_covariance, measured at the same box, noise level and--pol. The finite-difference steps \(h_i\) come from that file. Without the file the forecast is an error: there is no analytic delensing. - The noise is analytic and independent of the cosmology, with \(N_{EE} = N_{BB} = 2 N_{TT}\).
- The phi block uses \(C_\phi(\theta)\) from direct CAMB calls and an \(N_\phi\) held at \(\theta_0\).
Choosing N_phi¶
| Option | N_phi |
|---|---|
| default | the quadratic-estimate \(N^{(0)}\) of the box, built like load_sim's (the TT estimator for I, the EB estimator for P and IP), with the unlensed spectra in both its response and its filter |
--qe_filter lensed |
the same, with the lensed spectra in the filter; the response stays unlensed |
--phi_noise <npz> |
the nphi of a merged map_joint_phi_noise.npz or sampler_phi_noise.npz; unmeasured (NaN) modes carry no phi information |
Build the delensed covariance's Wiener filter from the same N_phi the forecast uses; nothing checks it.
Running it¶
python -m cmb_lensing.fisher_forecast --noise 5 --delensed_covariance delensed_covariance.npz
python -m cmb_lensing.fisher_forecast --noise 5 --delensed_covariance delensed_covariance.npz --qe_filter lensed
python -m cmb_lensing.fisher_forecast --noise 5 --delensed_covariance delensed_covariance.npz --phi_noise map_joint_phi_noise.npz
python -m cmb_lensing.fisher_forecast --noise 5 --pol IP --delensed_covariance delensed_covariance_IP.npz
python -m cmb_lensing.fisher_forecast --noise 5 --pol IP --delensed_covariance delensed_covariance_IP.npz --no_bb
Other flags: --nside, --theta_pix, --l_knee, --beam_fwhm, --params (default logA omch2 theta_MC_100), --camb_lmax (default 4000, the support of the simulations).
Outputs go to cmb_lensing/fisher_output/ as fisher_matrix, covariance_matrix, correlation_matrix (PNG) and fisher (npz), each tagged _<pol>_<nphi>_<noise>_uk, e.g. fisher_I_qe_unlensed_5_uk.npz. Parameters are in the package's standard order: theta_MC_100, logA, ns, ombh2, omch2.
What the forecast leaves out¶
- The blocks are added as independent. The residual lensing left in the delensed field tracks \(C_\phi(\theta)\), so the field blocks count some of the lensing information the phi block already has.
- The delensed BB is pure residual lensing at \(r = 0\). Its modes are neither Gaussian nor independent, so counting them as such is optimistic;
--no_bbleaves the BB block out and shows how much of aPorIPforecast rests on it. - The correlations between the fields and phi are ignored.
- For
IPthe \(N^{(0)}\) is the EB estimator's, as in the sampler; a combined temperature and polarization estimator would be lower. - Parameters not in
--paramsare held fixed, not marginalized over.
The analytic-delensing options of the earlier forecast (CAMB's partially delensed spectrum, the iterated QE filter, the gradient response) were removed on 2026-11-11; the last version with them is commit b50d4cc.