TODO

* handle correlated parameters: use pseudo-inverse for covariance matrix

* API: add all_datasets(): for n, title in F:all_datasets() do ...
  (or a public class Dataset?)

* API: access to confidence intervals

* GUI: Tools > Macros

* document how to get Gaussian/Lorentzian width from Voigt

* info cov/errors - do not specify dataset

* samples: add CeO2 from http://mysite.du.edu/~balzar/s-s_rr.htm

* XRPD - db of popular materials? but which are the most popular?

* GUI: plot histogram of residuals:
  http://www.itl.nist.gov/div898/handbook/eda/section3/normprpl.htm

* Use more exact Voigt approximation from Faddeeva Package (by SGJ) or libcerf

* FFT data transformation using FFTW library,
  fft[-inv]-xx, where xx is re, im, amp, e.g. @n = fft-re(@3)

* convolution and deconvolution of two datasets or a dataset with Gaussian:
  convolve(@0, @1); convolve(@0, 2.3); deconvolve(@0,3.4)
  OR                conv(@0, gauss(2.3))

* GUI: draw points (handles) for dragging peak width at half height

* CLI/GUI: use sigaction() instead of signal() to handle SIGINT
  print a message from interrupt_handler()

* refactor sidebar; use wxDataViewCtrl for the lists at the sidebar,
  rewrite ListWithColors::populate(), add sorting by clicking on column
  header (use case: sort functions by centers)

* syntax enhancements:  $min = @*.min(y if a)
                        print all: F[*](x) # print all components as columns

* finish the powder diffraction add-on (Pawley fitting)

* add functions cot and coth

* allow all the functions from data transformation in variable/UDF definition

* GUI: there is ugly vertical line (a part of the fitted curve) in points
  where the value is undefined (e.g. in ReadShockley function for x=0).

* GUI: when dataset is changed and the active function (i.e. the function
  in the bottom of the sidebar) is not in the new active dataset, the active
  function should also be changed (to the function that was active last time
  this dataset was active)

* (?) GUI: possibility to plot horizontal/vertical lines on the main plot
  (for example to show theoretical peak positions in XRD or to mark simulation
  restarts in data from MD. Syntax: plot line x=42.2 label="foo"

* GUI: a simple (for the user) way to interrupt fitting. Requires two threads.

* confidence bands

* data smoothing (Savitzky-Golay and/or other methods)

* splines: fitting, guessing, handling knots in the GUI

* user defined functions:
  - calculating limits, width, area, etc. for UDF
  - ?? use LLVM for faster calculation of user-defined functions

* info dpeak [range]: show parameters of selected data range assuming that
  this range contains a peak: area, center, FWHM, max, peak-area, backgr-area
  (like in the Xview program)

* more sophisticated algorithms for peak detection
    - compare methods from origin pkfind, peakfit, fitzpeak and other programs
    - research papers: http://dx.doi.org/10.1016/j.chroma.2004.11.073
                       http://dx.doi.org/10.1002/cem.1005
                       http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2631518/

* "anonymous" function types: %f = ~1/(1+exp(11*(~0.663-x)))  -> Anon1(a0,a1)

* fitting - external libraries to be considered/tested:
  - ceres-solver (or http://dlib.net/optimization.html)
  - libgencurvefit (Differential Evolution),
  - PaGMO library that collects many optimisation methods.
  - GSL: https://www.sourceware.org/ml/gsl-discuss/2014-q1/msg00025.html

* fitting: in addition to the least squares fit, add so-called robust fit 
  (it is called robust fit in the Numerical Recipes book), i.e. optimize
  assuming the error distribution is other then Gaussian, eg. Lorenzian...;
  there is also a literature about so called 'maximum-likelihood method'
  in which the exact Poisson error distribution is handled.