VERSION_HISTORY.txt

Current repository:
-Fixed QU fitting using Stokes Q errors instead of U errors when evaluating log_likelihood.

1.4.10:
-added warning to rmsynth3d if users specify both Stokes I model and noise vector (not compatible at present)
-for chunking tools, added optional memmap keyword to allow loading of arrays into memory (speeds up, if there is enough available memory)
-dropped support for Py3.7 (something is broken, and not worth fixing)
-This is the version going into Zenodo as the official version as of time of publication of the RM-Tools paper. Future major patches will also be uploaded to Zenodo.

1.4.9:
-Fixed numpy deprecations in assemble chunks tool, QU-fitting, and 1D synth/clean (thanks to Yingkang Zhang)
-Fixed peakfitting tool from crashing when running on chunks

1.4.8:
-Added m12 model (mixed slab with turbulent foreground).
-forced rmsynth1d outputs to the bit depth specified by the user (previously many outputs were 64bit regardless of user specification).
-Added combined Q/U sigma_add calculation and output. Now produces sigma_add for Q and U separately, and for both together.
-Fixed bug in rmclean3d where multiprocessing pool was not used properly.

1.4.7:
-fixed crash when using Stokes I model cube in rmsynth3d.
-changed all references to speed of light to use astropy constant.
-fixed bug from strange interaction between Prefect, Docker, and multiprocessing.
-added basic tests for helper tools.

1.4.6:
-fixed bug in rmsynth3d when computing RMSF that is not spatially uniform.

1.4.5:
-Fixed 3DIrescale tool to work with new NamedTuple for Stokes I fitting outputs.

1.4.4:
-Fixed calcRMSF tool to work with new NamedTuple system for RMSF calculation/fitting outputs.

1.4.3:
-Fixed bug in RMpeakfit3D which wasn't using the correct channel weights to calculate uncertainties.

1.4.2:
-minor fix to uncertainties generated by RMpeakfit3D

1.4.1:
-changed internal handling of Stokes I models; fixed problem when using user-defined model.

1.4.0:
-updates to do_fitIcube tool: removed deprecated cutoff keywords, fixed tqdm
 progress bars, some minor speed and efficiency improvements.
-relaxed numpy version restriction, since the numpy-astropy version conflict
 has been resolved.
-added tqdm as install dependency.
-removed internal copies of external packages (corner, emcee, nestle) used by QU-fitting
-added pre-commit auto-formatting system
-escaped all LaTeX characters
-implemented auto-formatting, using black, and updated all existing code to that standard.
-removed RMS noise estimate for FDF -- it was inacurrate and misleading
-fixed corMAD noise estimate for FDF -- it was estimating MAD of PI rather than Q and U,
  and thus was biased low. Now uses Q and U of FDF, and reproduces correct value in
  signal-free and RM-cleaned simulations. Still biased high in dirty spectra; no fixing that.
-old versions of core rmsynth and RMSF calculation functions are now officially deprecated
-fixed crashing when trying to generate RMsynth1d plots for source with negative Stokes I model
-improved FITS headers of 3D tools output
-1D: removed RMS noise estimate (for being misleading); removed _chan outputs (for being low value and confusing)
-fixed estimation of empirical FDF noise. (1D RMclean)
-QU-fitting: fixed problems with de-rotated angle wrapping
-Integrated nufft tool for the Fourier transform math (makes rmsynth faster)
-added options to rmsynth3d to decrease RAM use
-new QU-fitting models: m5 (single Burn slab), m6 (double Burn slab), m7 (single mixed emission slab)


1.3.1:
-progress bars updated to use tdqm, which should make them faster.
-Integrated super-resolution method of Cotton & Rudnick 2023, which creates
  narrower RMSF shapes by setting lambda^2_0 = 0.
-Added options to dofitIcube tool to save more outputs (coefficient maps,
  errors, etc).
-Made RMSF theoretical width calculations consistent: Dickey et al. 2019
  now used throughout.

1.3.0:
-merged in bwdepol branch. This does not change any core functionality in the
  existing RM-Tools programs, but adds two new command-line tools: rmtools_bwpredict
  and rmtools_bwdepol, which predict and correct for bandwidth depolarization
  respectively.
-fixed a bug with JSON outputs not handling numpy variables. This was partially
  due to a problem with asscalar() in the utils using a depcreciated numpy function,
  but partially from other causes.
-an improved version of the Stokes I cube fitting routine has been developed
  by Lerato Sebokolody, which includes parallelization.

1.2.0:
-changed Stokes I fitting to always use 64 bit values (even if data is 32 bit).
  This is because the fitter was found to be vulnerable to numerical issues
  when fitting log-polynomials in 32-bits (thanks to Shannon Vanderwoude for
  discovering the problem).
-added parameter errors of Stokes I fit as output to Stokes I fitDict, to
 RMsynth1D outputs, and QU-fitting outputs. These are the 1-sigma errors
 computed by the mpfit routine.
-the renormalize_StokesI_model function is now depreciated, since it cannot
 transform the errors. It is left in for the time being in case it is needed
 for any legacy reasons.
-the make_freq_file.py/rmtools_freqfile tool was found to sometime produce
 truncated output (not all channels included). This was fixed and changed
 to a more rigourous calculation (using WCS, as suggested by Alec Thomson).
-modified do_fitIcube.py/rmtools_fitIcube to output maps of the fit coefficients
 and their errors. Thanks to Biny Sebastian for providing this code.
-merged pull request from Alec Thomson that changes the fractional polarization
 reported by rmsynth1d to use the bias-corrected polarized intensity.
 !!!Previous versions used the non-corrected polarized intensity, from this
    version onwards the corrected polarized intensity will be used!!!
-merged pull request from Alec Thomson that implements windowed cleaning
 and adds a command line flag to set the windowed clean threshold.
-merged pull request from Alec Thomson that changed the QU-fitting routine
 to use the bilby package for the nested sampling. This package allows the
 user to specify which sampler they wish to use -- the default is dynesty,
 but users can choose to use pymultinest or other supported samplers.
-removed support for Python 3.6, as a consequence of the change to bilby.
 From this point forward, Python 3.7 is the minimum version supported
 (at least for the QU-fitting portion of the package).

1.1.2:
-modified the generation of maxPI and peakRM maps in RMsynth3D so that
  it occurs in its own function rather than as part of the file writing
  function; this allows it to be run inside a notebook.
-modified RMpeakfit_3D to be more stable when FDF cubes have NaNs present.
-Stokes I fit coefficients no longer become NaNs when negative Stokes I
  models are renormalized to a new reference frequency.
-added test for unreasonably large peak RM (compared to computed range)
  as this can happen when peak is on edge of spectrum. Changes RM to NaN in
  these cases.
-Fixed crash when sigmaAdd calculations received NaNs in inputs.
-added experimental tool to compute QU models from clean components,
  requested by Luke Pratley. This currently has a normalization error of some
  kind, which I've asked Luke to chase down.
-dropping support for Python3.5, since the QUfitting code has problems with it
 and it's already end-of-life.
-improved the calculate_RMSF script to not crash when given width smaller than
 first sidelobes. Also corrected it to produce the correct width: previously
 when width was specified by user it would be 2x wider than specified.
-fixed problem with do_fitIcube.py which would crash when fed 4D Stokes I cube.
-fixed undesired behaviour in rmsynth1d and rmclean1d where spectra plots
 wouldn't be saved when plotting to screen was turned on.

1.1.1:
-updated calculate_RMSF to use same theoretical width as rmsynth1D
  (the formula from Dickey et al. 2019)
-fixed problem where very small Stokes I model values (~1E-60) would
  cause numeric overflow in Q/I, U/I, creating infinities that were
  not handled gracefully. Now converting infinities to NaNs.
-fixed peakfitcube not being callable from the command line.
-fixed typos in rmtools_calcRMSF.
-fixed pip installation so that QUfitting models are correctly installed.

1.1.0: New feature update
-QU-fitting updates from Alec Thomson: sigma-clipping of posteriors,
    saving evidence to output file, option to not restart from scratch.
-Fixed bug in chunk reassembly script (assemble_chunks.py) where it would
    not figure out the cube size correctly when fed the last chunk.
-Updated RMclean1D to use new derotated angle equation for Observed noise
    (previously was using old flawed equation fixed in v1.0.2).
-Fixed problem with handling of degenerate FITS axes in RMclean3D
-Created new tool: RMpeakfit_3D.py,which applies the 1D peak fitting to
    each pixel in a 3D FDF cube. Outputs 2D maps for the resulting
    parameters.
-New Stokes I fitting. Two changes:
    1) Logarithmic polynomial fitting function added. This is essentially
        a 5-th order polynomial in log-log space, modified into the
        standard power-law with curvature formulation. Added flags to
        RMsynth1D, do_QUfitting, and do_fitIcube to let user select fitting
        function.
    2) Dynamic order selection. If user enters a negative polynomial order,
        it will now trying fitting successively higher order polynomials
        (starting with 0) until the AIC no longer improves.
-default Stokes I fitting is now 2nd order polynomial
-Plotting code has been made more consistent, in terms of no longer
    requiring user to press 'Enter', for all routines.
-QUfitting now has .dat output like that of RMsynth1D.
-All supporting tools now have command-line hooks, which all begin with
    'rmtools_'.

1.0.5: RMSF FWHM calculation fix   (13 Aug 2020)
-Theoretical estimate for FWHM now uses formula from Dickey et al 2019
    instead of Brentjens & de Bruyn 2005, as this is more accurate.
-RMSF fitting improved to be more accurate (previously would over-estimate)
-fitReal option for RMSF fitting now works as intended (previously crashed)

1.0.4: QUfitting tweaks + RMsynth3D crash fixes  (10 Aug 2020)
-Minor QUfitting fixes by Alec Thomson.
-Fixed crash when writing 3D cubes where all pixels are NaNs.
-changed output names for the QUfitting multinest files.
-Fixed crashes in RMsynth3D when input cubes are all NaNs.
-Fixed crash when RMsynth3D normalizes output cubes by Stokes I model.

1.0.3: Noise propagation improvements   (23 Jul 2020)
-Minor fixes to new plot-saving code.
-Changed theoretical noise estimation equation: Shinsuke Ideguchi pointed
    out that it wasn't correctly computing the error in the fractional
    polarization. Modified to normalize the channel errors by the model
    Stokes I (errors in the model are not accounted for). This will alter
    the noise behaviour for sources with steep Stokes I spectra.

1.0.2: Fixing derotated angle errors + Multimodel QUfitting  (6 Jul 2020)
-RMsynth1D: added option to save plots (pull request by Lenart Heino)
-Error formula for the derotated polarization angle (dPolAngle0Fit/Chan)
    was wrong (too small by a factor of sqrt(Nchan)). Corrected formula
    incorporated.
-Multi-modal posterior handling added to QU-fitting (by Alec Thomson).

1.0.1: Minor bug fixes   (12 May 2020)
-For RMsynth1D, the default Phimax value was previously fixed at 600,
    which was not ideal for high-frequency data. Modified to now be
    dynamic: 10x RMSF FWHM.
-RMclean1D was not using the noise properly when calculating derived
    properties. RMclean1D error bars prior to this verison likely have
    issues.

1.0: Official release (1 May 2020)