Hi,

I am thinking about writing a package that implements a generalized least squares fitting method or perhaps seek to contribute somewhere where it might fit. Before I write more than my prototype, I wanted to double check that I am not overlooking some already existing project.

The algorithm that I want to implement (or have implemented ) is described in this paper:

The important feature is that it can handle covariance in all input parameters. This allows, for example, to fit a curve to measurement data with uncertainties in x- and y-values AND have correlations in uncertainties accounted for. This can be of relevance in calibration, where a certain standard is used several times (e.g. a standard solution that is diluted several times). The covariance matrix of the ouput parameters is a natural byproduct of the algorithm. Also, the algorithm does not really differentiate between “x” and “y” values, it knows only measured values and constrain functions (which can be used to implement a curve fit) so it is more versatile than just a method for curve fitting.

To find potential candidates for packages that implement a least squares fit like the described method I was looking through github, juliahub.com, this discourse and other websites but had no luck. LsqFit.jl has a section in the docs describing how to take care of correlations in the y-errors, but its does not seem to work (see this PR). But even if it did, it would only be half the solution, since it would not account for uncertainties in x and would be limited to curve fitting.

Another package that showed up was TotalLeastSquares.jl. It handles the full covariance matrix of input parameters but seems to be limited to linear models and I don’t know how one would get the covariance matrix of the output parameters (although it is perhaps possible and maybe not even that hard, I am just missing the mathematical background).

Then of course there is JuMP, but it does not seem to handle uncertainty at all, or at least it is not obvious to me how it does so.

OK, before this post gets even longer:

Is there some project that already provides the functionality I described above, or would that be a useful addition to the Julia toolbox that is currently missing?

Thanks!