# 44 larger its weight gets smaller and smaller with

• Lesson Plan
• 66

This preview shows pages 44–46. Sign up to view the full content.

– 44 – larger, its weight gets smaller and smaller. With this, in nonlinear fitting all datapoints are always included and their weights automatically adjust as the fit parameters home into their correct values. And you can’t get into the chasing-tail syndrome that can happen with the strict on-off inclusion. 12.1. Stetson’s sliding weight Stetson recommends using a sliding weight. To explain this, we review the ML concept of chi-square fitting. We define chi-square as χ 2 = M - 1 summationdisplay m =0 ( y m a · f ( x m )) 2 σ 2 m , (12.1a) and we minimize χ 2 by setting its derivative with respect to each parameter a n equal to zero: 2 da n = 2 M - 1 summationdisplay m =0 f n ( x m y m σ 2 m . (12.1b) Here Δ y m = ( y m a · f ( x m )). For each coefficient a n setting this to zero gives M - 1 summationdisplay m =0 f n ( x y m σ 2 m = 0 . (12.1c) Now we wish to modify this equation by introducing a weight w ( | Δ y m | ) that makes datapoints with large | Δ y m | contribute less, so it reads like this: M - 1 summationdisplay m =0 w ( | Δ y m | ) f n ( x m y m σ 2 m = 0 . (12.2) It’s clear that we need the following properties for w y m ) : 1. w y m ) = w ( | Δ y m | ) , meaning simply that it should depend on the absolute value of the residual and not bias the solution one way or the other. 2. w ( | Δ y m | ) 1 as | Δ y m |→ 0, meaning that datapoints with small residuals contribute their full weight. 3. w ( | Δ y m | ) 0 as | Δ y m |→∞ , meaning that datapoints with large residuals contribute nothing.

This preview has intentionally blurred sections. Sign up to view the full version.

– 45 – Stetson recommends w ( | Δ y m | ) = 1 1 + parenleftBig | Δ y | ασ parenrightBig β . (12.3) This function w ( | Δ y m | ) has the desired properties. Also, for all β it equals 0.5 for | Δ y m | = ασ . As β →∞ the cutoff gets steeper and steeper, so in this limit it becomes equivalent to a complete cutoff for | Δ y m | > ασ . Stetson recommends α = 2 to 2.5, β = 2 to 4 on the basis of years of experience. Stetson is a true expert and we should take his advice seriously; he provides a vibrant discussion to justify these choices in real life, including an interesting set of numerical experiments. However, for large M I see a problem with the choice α = 2 to 2.5. For large β , for which the cutoff is sharp, it seems to me that the cutoff should duplicate Chauvenet’s criterion. Referring to equation 6.5, this occurs by setting α = 2 erf - 1 parenleftbigg 1 1 2 M parenrightbigg (12.4) and I recommend making this change, at least for problems having reasonably large M ; this makes α larger than Stetson’s choice. I’m more of a purist than Stetson, probably because I’m a radio astronomer and often fit thousands of spectral datapoints that are, indeed, characterized mainly by Gaussian statistics. Stetson is an optical astronomer and probably sees a lot more departures from things like cosmic rays. Nevertheless, in a CCD image with millions of pixels, of which only a fraction are characterized by non-Gaussian problems such as cosmic ray hits, it seems to me only reasonable to increase α above Stetson’s recommended values by using equation 12.4.
This is the end of the preview. Sign up to access the rest of the document.

{[ snackBarMessage ]}

### What students are saying

• As a current student on this bumpy collegiate pathway, I stumbled upon Course Hero, where I can find study resources for nearly all my courses, get online help from tutors 24/7, and even share my old projects, papers, and lecture notes with other students.

Kiran Temple University Fox School of Business ‘17, Course Hero Intern

• I cannot even describe how much Course Hero helped me this summer. It’s truly become something I can always rely on and help me. In the end, I was not only able to survive summer classes, but I was able to thrive thanks to Course Hero.

Dana University of Pennsylvania ‘17, Course Hero Intern

• The ability to access any university’s resources through Course Hero proved invaluable in my case. I was behind on Tulane coursework and actually used UCLA’s materials to help me move forward and get everything together on time.

Jill Tulane University ‘16, Course Hero Intern