Specifying absolute errors in curve_fit

[pv]

Continued from gh-448

An option absolute_sigma is added to scipy.optimize.curve_fit, to accommodate the common cases:

• sigma = relative weights, pcov variance estimated from data
• sigma = one standard deviation errors on ydata, determines pcov variances

Also, make the returned pcov always be a 2D array. In indeterminate cases, it used to return a scalar, which may not be desirable.

[josef-pkt]

Yes, I like returning the not available pcov in the correct shape better.
I don't know whether it would be better full of NaNs or full of infs. In either case users should get suspicious about their results.

[coveralls]

Coverage remained the same when pulling fffae00 on pv:curve-fit-yerr into 0e7eb52 on scipy:master.

[pv]

Ok, no objections, so merging.

[djpine]

While people are modifying curve_fit, how about improving the docstring? Right now, it has a rather misleading example that works without providing initial guesses only because all the fitting parameters are close to one, the default values. I think users would benefit greatly from seeing an example that needs good initial guesses. I would be happy to work on this if no one else wants to.

[pv]

@djpine: an improved the docstring would be welcome. If you already know how it should be improved, please go ahead.

[josef-pkt]

This reminds me about the starting values.

If we change them slightly from 1.0 we avoid the list confusion, issue ???

[Tillsten]

I would suggest to make the starting values a necessary argument. Giving it a default of one is probably a big source of bugs.

[djpine]

I agree.

---

David Pine
Professor of Physics

Center for Soft Matter Research
Department of Physics
New York University

Office: 601 Meyer
Department Office: 424 Meyer

Mailing address:
4 Washington Place
New York, NY 10003

E-mail: pine@nyu.edu
Tel: 212-998-7744
Web: http://www.physics.nyu.edu/pine/Home.html
http://www.physics.nyu.edu/pine/Home.html

---

On Wed, Dec 4, 2013 at 10:12 PM, Till Stensitzki
notifications@github.comwrote:

I would suggest to make the starting values a necessary argument. Giving
it a default of one is probably a big source of bugs.

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/3098#issuecomment-29846034
.