How to translate estimated model parameters when fitting centered and scaled data?

59 Views Asked by At

I routinely use a non-linear curve fitting tool to fit data according to a user prescribed model / function. One piece of advice that I commonly see around non-linear curve fitting is about data conditioning, specifically scaling and centering. Consider the simple 2D model:

$y = \dfrac{a}{x^{2}} + b$

The scaled (and centered in the case of the $x$ array) arrays which can be passed to the curve fitting tool are:

$x' = \left[(x - c_{0}) / c_{1}\right] - 0.5$

$y' = (y - k_{0}) / k_{1}$

where $c_{0}$ and $k_{0}$ are minima in each series and $c_{1}$ and $k_{1}$ are the ranges. When we perform the fit, we are fitting the model

$y' = \dfrac{a'}{x'^{2}} + b'$

The curve fitting tool returns $a'$ and $b'$. It is straightforward to un-scale $b'$ to find $b$.

Thoughts are welcome on how to translate $a'$ to $a$.