Suggestion: "recalculate" option for ModelResult.conf_interval()

[jagerber48]

If I generate a model, give it data, fit it, and then calculate the confidence interval using result.conf_interval() the confidence interval will be calculated with sigmas = [1, 2, 3] by default. If I realize I want sigmas = [1], or sigmas=[4] after the fact I must manually set result.ci_out = None prior to making a second call to result.conf_interval().

My suggestion would be to have a recalculate (or force_recalculate?) kwarg in resut.conf_interval which, when True, always recalculates the confidence interval even if it had been calculated previously. Backwards compatibility would suggest the kwarg defaults to False though it's not obvious to me that's the most expected behavior. It is strange to do, for example result.conf_interval(sigmas=[1]) then result.conf_interval(sigmas=[1, 2]) and get the same result both times (with the latter being incorrect in my opinion). So that would suggest the recalculate keyword should default to True.

Perhaps a more sophisticated solution would be to "memoize" not only the ci_out, but also the function call signature that was used to generate ci_out. Then if result.conf_interval is called with an identical signature the memoized version is used, but if it is called with a different signature the calculation is repeated with the new signature.

Here's a minimal example showing the current (in my opinion, strange) behavior.

import numpy as np
from numpy import exp, loadtxt, pi, sqrt

from lmfit import Model
from lmfit.printfuncs import report_ci

def gaussian(x, amp, cen, wid):
    """1-d gaussian: gaussian(x, amp, cen, wid)"""
    return (amp / (sqrt(2*pi) * wid)) * exp(-(x-cen)**2 / (2*wid**2))

x = np.linspace(0, 10, 25)
y = gaussian(x, 5, 5, 1) + np.random.normal(0, 0.2, len(x))
yerr = np.full_like(x, 0.2)
gmodel = Model(gaussian)
result = gmodel.fit(y, x=x, amp=5, cen=5, wid=1, weights=1/yerr)

report_ci(result.conf_interval(sigmas=[1, 2, 3]))
report_ci(result.conf_interval(sigmas=[1]))
result.ci_out = None
report_ci(result.conf_interval(sigmas=[1]))

[jagerber48]

Oops, this is a duplicate of this issue: #791