virgil.ensemble

Regularizer families to draw from, uniformly: "tv" (TV), "tsv" (TSV), "maxent" (MaxEntropy) and "starlet" (StarletL1).

For each family, the range (low, high) of the weight per data point (the weight divided by the total number of independent data). Weights are drawn log-uniformly in it, the invariant prior for a scale, one in each of n_weights equal bins of log w so that every sweep spans the range. The defaults are wide enough to hold the corners of the contest bench's datasets with a window's width to spare below them; if the L-curve corners fall at the edges, move the range.

Weights per group, at least three (the L-curve's corner needs them).

Pixels per Nyquist pixel (nyquist_pixel_scale), drawn uniformly.

The field is field_of_view(data) times one of these, drawn uniformly. A field larger than the interferometric field of view lets the image hold flux the shortest baselines resolve out, which on the contest bench gave better images than a field of 1 or less (field 4 was the best of its arms). On a uv lattice (AMI, data with a uv_grid) a field above 1 aliases, so there the factors are capped at 1.

Starting images, drawn uniformly: "moments" and "dirty" (see starting_image; a dirty start needs data with phases) and "flat".

Largest number of pixels on a side. A field that would need more keeps its size and takes coarser pixels instead, so its pixels per Nyquist pixel fall below oversample, but never below one: if even Nyquist pixels cannot cover it, the field is cropped.

Width of the window of weights kept in each group, in dex, from corner / 10**window_dex up to the L-curve's corner: as in MYTHRA, the weights just before the turnover, where the fit to the data has stopped improving but the image is not yet over-smoothed. Stronger weights, past the corner, trade fit for smoothness fast.

Drop a member if the raw χ² per data point of any dataset exceeds this.

Drop a member if, on any dataset, its χ² per data point exceeds this multiple of the best member's on that dataset. With miscalibrated errors no member reaches χ²/N ≈ 1, so a relative threshold is the one that bites.

Then drop members whose total χ² per data point lies more than this many robust standard deviations (1.4826 times the median absolute deviation) above the median.

Without a star, the members are recentred on the best one, searching shifts up to this (default: the beam's major axis). With a star, the star fixes the position and they are not shifted.

A member joins the mean if, on every dataset, the mean's χ² stays within this fraction of the best member's. The default, None, is each dataset's χ²/N noise, √(2/N): a mean that fits as well as the best member, up to the noise. 0 keeps the mean's χ² at or below the best member's, which on data the best member fits to the noise often keeps that member alone (and the spread is then zero); on the contest bench it kept 1 of 72 members on most datasets.

combine warns if fewer members than this are kept.

Position of the group in the ensemble.

The regularizer family.

Pixels on a side.

Pixel size in mas.

The starting image.

The regularizer weights, largest first.

The group's settings.

Its fits, one per weight, largest weight first.

The settings of its group.

Its regularizer weight.

The fit.

Raw χ² per data point of each dataset (with the quoted errors).

Whether it is in the mean image.

Why it was left out: "window" (outside its group's L-curve window), "diverged" (a non-finite χ²: the fit failed), "chi2" (a dataset fitted too badly), "outlier" (total χ² far above the others) or "mean" (adding it made the mean fit a dataset worse). None if kept.

The mean scene, exactly: the kept members' images on their own grids (member0, member1, ...), each carrying its share of the flux, and with a star a star component. chi2_red and trace are of this model.

The mean extended emission, on the common grid (the finest pixels and the largest field of the kept members). Its brightness sums to one; with a star its flux is the mean flux relative to the star. Resampling smooths the members' images a little, so this image is for display and scoring; it fits the data less well than model.

Standard deviation across the kept members of the same quantity as mean.brightness, pixel by pixel.

Raw χ² per data point of the mean scene on each dataset.

chi2_red of the running mean after each accepted member, from the best member alone; on every dataset it stays within spec.mean_rtol of the first entry.

Every reconstruction, kept or not.

The groups, with their L-curves.