virgil.radio

Build from arrays; scalars broadcast over the points.

ddid and obs are the original labels, relabelled 0..n-1 in sorted order. Each ddid must belong to one observation.

Read a CASA measurement set (needs the radio extra).

Cross-correlations only; one sample per (row, channel), kept if the row and both parallel-hand polarizations (XX, YY or RR, LL) are unflagged and both weights positive. The two polarizations are averaged with their weights (Stokes I for equal weights) and the weight is their sum. WEIGHT_SPECTRUM is used when present and non-empty, else WEIGHT. u = UVW[0] ν / c, v = UVW[1] ν / c per channel. One ddid is read at a time, so memory is that of the largest ddid at the given stride.

Parameters:

The points where boolean mask is true, labels unchanged.

Labels are kept as they are (not relabelled), so a CalibrationNuisance built for the full data still fits.

Points of the given ddids and/or observations.

ddids and obs are the original ids (ddid_ids, obs_ids). Labels are kept; see subset.

The complex model M_k at every point (all at once).

M = F g_e (ν/ν₀)^α exp(-2πi(δα_e u + δδ_e v)) T(u, v) V_scene(u, v); see CalibrationNuisance. scene is a SourceModel (normalized to unit flux; cal.flux sets the scale), evaluated at wavelength c / ν of each point so that chromatic scenes work.

A likelihood term for fit's likelihoods.

Parameters:

Precompute the dirty beams and images of a one-Image fit.

Returns a ToeplitzVisData, whose term gives the χ² of term for a scene that is one Image with the layout of image, at a cost independent of the number of points. Provisional.

Parameters:

The error scales s_d per ddid at values: the profiled ŝ of CalibrationNuisance(sigmas="profile"), the fitted cal.sigma with sigmas=True, ones otherwise. Provisional, as the module.

χ² of the data with the σ scales applied, per ddid.

Returns (chi2, n): float64 arrays of length n_ddid, n the number of unmasked points of each ddid (χ² sums two terms per point). chunk_size is that of term.

Model-free estimate of the error scale from time-adjacent samples.

Samples of one baseline, channel and scan (and ddid label) are paired in time order without overlap; ŝ² is the mean over pairs of ΔRe² / (1/w_i + 1/w_j), likewise for Im, and ŝ averages the two.

This is an upper bound. Sky, gain or phase change within a pair adds to |ΔV|², so ŝ is biased high for rapidly varying data (long gaps, atmospheric phase), exact only if the signal is constant. Samples without a partner are not used. Needs the time, ant1, ant2 and scan fields filled by read_ms; with stride > 1 the rows read are no longer adjacent.

Parameters:

Returns:

The dirty beam Σ w cos(2π(u x + v y)) / Σ w on the pixel grid.

The response to a unit point source at the phase centre, normalized to 1 there. See dirty_image for the parameters; values and nuisance together divide the weights by the fitted s_d² (as use_sigma=True). On an even npix the phase centre lies between four pixels (the convention of Image), so the sampled peak is slightly below 1; use an odd npix for a pixel centred on it.

Dirty image Σ w Re[V e^{+2πi(u x + v y)}] / Σ w in Jy/beam.

A display diagnostic, never part of the likelihood: a type-1 NUFFT at the raw (u, v), in the orientation of Image (East left, North up). Details: Dirty and residual images.

Parameters:

Returns:

Default priors of the fitted sites, a dict for fit.

bounds override the ends (lo, hi) of a site's prior, keyed by flux, gain, offset, sigma, alpha, gain_amp_width or gain_phase_width (the last two are LogUniform(1e-3, 10) by default). Without a bound, the gains and σ scales are improper Jeffreys priors (flat in ln x, no box) and the offsets are uniform over ±offset_beams beams of their observation (2π b_k, see the class docstring). offset=(-1/(2 q_min), 1/(2 q_min)) in mas, with q_min the shortest baseline (offset_bound_mas), is the old default: a shift of the phase centre beyond that is less than half a turn of phase on every baseline.

Starting values of the fitted sites.