Error when fitting zero-degree polynomial

Description

When trying to fit data with a zero-degree 1-dimensional polynomial, a ValueError is raised.

Expected behavior

I expected a model object to be returned that represents a constant at the average of the data points.

Actual behavior

A ValueError is raised. The message is dependent on the fitter used.

Steps to Reproduce

Simply run the following script:

#!/usr/bin/env python3

import numpy as np

from astropy.modeling.models import Polynomial1D
from astropy.modeling.fitting import TRFLSQFitter, LevMarLSQFitter, DogBoxLSQFitter

if __name__ == '__main__' :
    model = Polynomial1D(0, c0=0)
    fitter = TRFLSQFitter() # Also fails with LevMarLSQFitter or DogBoxLSQFitter
    
    # Should work
    # Actually produes an error
    # The exact error depends on the fitter used
    # TRFLSQFitter    : ValueError: The return value of `jac` has wrong shape: expected (10, 1), actual (10, 10).
    # LevMarLSQFitter : ValueError: The array returned by a function changed size between calls
    # DogBoxLSQFitter : ValueError: The return value of `jac` has wrong shape: expected (10, 1), actual (10, 10).
    fit = fitter(model, np.arange(10, dtype=float), np.ones((10,)), weights=np.ones(10,))
    print('fit :', fit)

It seems that the error is in _NonLinearLSQFitter._wrap_deriv method. There is a try/except block. The except section returns a value that gives the correct behavior. However, the try section gives incorrect behavior.

System Details

>>> import platform; print(platform.platform())
Windows-10-10.0.19044-SP0
>>> import sys; print("Python", sys.version)
Python 3.10.4 (tags/v3.10.4:9d38120, Mar 23 2022, 23:13:41) [MSC v.1929 64 bit (AMD64)]
>>> import numpy; print("Numpy", numpy.__version__)
Numpy 1.22.0
>>> import erfa; print("pyerfa", erfa.__version__)
pyerfa 2.0.0.1
>>> import astropy; print("astropy", astropy.__version__)
astropy 5.1
>>> import scipy; print("Scipy", scipy.__version__)
Scipy 1.9.1
>>> import matplotlib; print("Matplotlib", matplotlib.__version__)
Matplotlib 3.5.3

[EdBehn]

@pllim - That’s a good point. I didn’t know about Const1D. However, I still think this bug should be fixed.

[WilliamJamieson]

I stumbled on this by accident.

Thanks, we are always looking for the corner cases.

I’m creating a tool where the user can compose their own model as the sum of simpler components. (ie several gaussians and a polynomial background). That’s why I wasn’t using LinearLSQFitter.

No worries, I often see users wanting to always use the non-linear fitters because they “always work”; however, whenever possible one should use the linear fitter because it will be faster, more stable, and more accurate.

For your application it makes sense to use a more general tool, as your problems will generally be non-linear.