ModelResult.eval_uncertainty does not work for complex models

[newville]

Description

ModelResult.eval_uncertainty is not working properly for models that generate complex data. Running the method gives a warning about casting complex to float, and returns a real array of the same length as the model.

See https://groups.google.com/g/lmfit-py/c/tqaLwSBH-Ts/m/HBDP0y4HEAAJ

A Minimal, Complete, and Verifiable example

import numpy as np
from lmfit import Model

def cmplx(f, omega, areal, aimag, off, sigma):
    return (areal*np.cos(f*omega + off) + 1j*aimag*np.sin(f*omega + off))*np.exp(-f/sigma)

f = np.linspace(0, 10, 501)
dat = cmplx(f, 4, 10, 5, 0.2, 4.5) + (0.1+0.2j)*np.random.normal(scale=0.25, size=len(f))


mod = Model(cmplx)
params= mod.make_params(omega=5, areal=5, aimag=5,
                        off={'value': 0.5, 'min': -2, 'max':2},
                        sigma={'value': 3, 'min': 1.e-5, 'max': 1000})

result = mod.fit(dat,  params=params, f=f)
print(result.fit_report())

dfit = result.eval_uncertainty()
print('Data  : ', len(dat), dat.dtype)
print('Fit   : ', len(result.best_fit), result.best_fit.dtype)
print('D_Fit : ', len(dfit), dfit.dtype)

Fit report:

[[Model]]
    Model(cmplx)
[[Fit Statistics]]
    # fitting method   = leastsq
    # function evals   = 49
    # data points      = 1002
    # variables        = 5
    chi-square         = 1.59641755
    reduced chi-square = 0.00160122
    Akaike info crit   = -6444.87518
    Bayesian info crit = -6420.32641
    R-squared          = (0.9996115675746108+2.8498175850731404e-05j)
[[Variables]]
    omega:  3.99961585 +/- 2.4279e-04 (0.01%) (init = 5)
    areal:  9.99808550 +/- 0.00755744 (0.08%) (init = 5)
    aimag:  4.99883013 +/- 0.00590546 (0.12%) (init = 5)
    off:    0.20069043 +/- 7.0226e-04 (0.35%) (init = 0.5)
    sigma:  4.50042395 +/- 0.00497052 (0.11%) (init = 3)
[[Correlations]] (unreported correlations are < 0.100)
    C(omega, off)   = -0.7328
    C(areal, sigma) = -0.6982
    C(aimag, sigma) = -0.4351
    C(areal, aimag) = +0.3021
/Users/Newville/Codes/lmfit-py/lmfit/model.py:1602: ComplexWarning: Casting complex values to real discards the imaginary part
  fjac[key][i] = (res1[key] - res2[key]) / (2*dval)
Data  :  501 complex128
Fit   :  501 complex128
D_Fit :  501 float64

Version information

Python: 3.10.10 | packaged by conda-forge | (main, Mar 24 2023, 20:17:34) [Clang 14.0.6 ]

lmfit: 1.2.1.post49+g78c9994, scipy: 1.11.0, numpy: 1.25.0,asteval: 0.9.30, uncertainties: 3.1.7

Link(s)

See https://groups.google.com/g/lmfit-py/c/tqaLwSBH-Ts/m/HBDP0y4HEAAJ

[newville]

I have a proof-of-principle fix but will wait until #888 is merged to make a PR.

[newville]

This is now fixed in the master branch, with a test added from the example code above.