哪个多项式回归次数显着？ 取决于点数或其他参数？

Question

I am studying the stability of numerical derivatives as a function of the step I take to compute these derivatives. With a derivative with 15 points (obtained by the finite difference method), I get the following plot (each multipole " l " corresponds to a parameter whose depends the derivative but it doesn't matter):

Now, I would like to compare this derivative of 15 points with the derivative computed with 3, 5 and 7 points. For this, I have just plotted the relative difference like (with absolute differences):

abs(f'_15_pts - f'_3_pts)/f'_3_pts for comparison between 15 and 3 points
abs(f'_15_pts - f'_5_pts)/f'_5_pts for comparison between 15 and 5 points
abs(f'_15_pts - f'_7_pts)/f'_7_pts for comparison between 15 and 7 points

And my issue occurs when I want to do a polynomial regression on the relative variations above with the multipole l=366.42 (the problem remains right for other multipoles).

For example, when I do a cubic regression (3 degrees), I get the following plot:

I don't know exactly how to interpret these results: maybe it means that I have a relative error maximal between 3 points and 15 points derivatives and less between 5 and 15 like between 7 and 15 points.

Then, If I want to do for example a polynomial regression of degree 10, I get the following plot:

As you can see, this is totally different from cubic regression above.

So I don't know which degree to take for polynomial regression, I mean which degree is relevant to get valid physical results: 3, 4, 6 or maybe 10 . If I take a too large degree, results aren't valid since I have dirac peaks and straight lines.

I guess the right polynomial degree to keep depends of the initial number of points of the interpolated curve (140 points for the first figure) and maybe others parameters.

As conclusion, could anyone tell me if there is criteria to determine which polynomial degree to apply?, I mean the degree which will be the most relevant from a relative error point of view.

If I don't do regression, I have the following plot which is blurred to interpret:

That's why I would like to interpolate these data, to see more clearly the differences between different relative evolutions.

PS: here the code snippets of polynomial regression:

stepForFit = np.logspace(-8.0,-1.0,10000)
coefs_3_15 = poly.polyfit(np.log10(stepNewArray), np.log10(errorRelative_3_15), 10)
ffit_3_15 = poly.polyval(np.log10(stepForFit), coefs_3_15)
coefs_5_15 = poly.polyfit(np.log10(stepNewArray), np.log10(errorRelative_5_15), 10)
ffit_5_15 = poly.polyval(np.log10(stepForFit), coefs_5_15)
coefs_7_15 = poly.polyfit(np.log10(stepNewArray), np.log10(errorRelative_7_15), 10)
ffit_7_15 = poly.polyval(np.log10(stepForFit), coefs_7_15)

# Plot interpolation curves
plt.plot(stepForFit[stepArrayId], np.power(10,ffit_3_15[stepArrayId]), colorDerPlot[0])
plt.plot(stepForFit[stepArrayId], np.power(10,ffit_5_15[stepArrayId]), colorDerPlot[1])
plt.plot(stepForFit[stepArrayId], np.power(10,ffit_7_15[stepArrayId]), colorDerPlot[2])v

UPDATE 1: Given I have not hypothesis (or a model) on the value of relative error, I can't put constraints apriori on the degree of polynome that must best-fit the data.

But maybe I have an clue since the derivatives that I have computed are 3, 5, 7 and 15 points. So I have an uncertainty of level respectively of O(h^2), O(h^4), O(h^6) and O(h^14).

For example, for the 3 points derivatives, I have:

and so the final expression of derivatives:

By the way, I don't understand why we pass from $O(h^4)$ to $O(h^2)$ between the expression.

But main issue is that I have not for instant hypothesis on the polynomial degree that I have to apply.

Maybe, I should test a range of polynomial degree and compute at each time the chi2, so minimal chi2 will give to me the right degree to take into account.

What do you think about this? Does Numpy or Python already have this kind of study with specific functions?

UPDATE 2: I tried to determine over a range of 1-15 degree of polynomial that could best fit the data. My criterion was to fit with a polynome for each degree and then compute the chi2 between "interpolation computed data" and "experimental data". If new chi2 is lower than previous chi2, I update the degree to choose to do a polynomial regression.

Unfortunately, for each of the 3,5 and 7 points derivatives, I always get by this research of "ideal degree", the maximum degree which corresponds to the max of degree-interval explored.

Ok, the chi2 is minimal for the highest degree but this doesn't correspond to physical results. One doesn't forget that below 10^-4, behavior of Cl' is chaotic, so I don't expect a physical interpretation on the convergence of the derivatives as an increasing number of points for derivatives.

But the interesting area is above 10^-4 where I have more stability.

Given my method of selecting best degree as a function of chi2 doesn't work (it always gives the maximal degree of the range explored), is there another method to get a nice fit? I know this is difficult since chaotic region for small steps.

Last thing, the cubic regression (3 degrees) gives nice curves but I don't understand why this only occurs for degree 3 and not for higher degree.

As someone said in the comments, for higher degree, regression is overfitted: how to fix this?

Answer 1

I have to say that I find your question formulation very confusing, so I can only help you with a bit general answer. Perhaps you could split your big question in several smaller ones next time.

To start with, I assume that your question is: how does the amount of points in a differentiation stencil matter, when I will do polynomial interpolation on the derivative afterwards?

The number of points in a stencil generally increases the accuracy of the computation of the derivative. You can see that by filling in Taylor expansions for the variables in the numerical derivative. After terms cancels you are left with some higher-order term that gives you a lower bound on the error that you make. The underlying assumption however is, that the function (in your case C ) of which you compute the derivative is smooth on the interval that you compute the derivatives on. Meaning that if your function is not nicely behaved on your 15-point stencil, then that derivative is essentially worthless.

The order of the polynomial in polynomial regression is usually a free parameter chosen by the user because the user might know that their series is behaved like a polynomial up to certain degree, but unaware of the polynomial coefficients. If you know something about your data, you can set the degree yourself. For instance if you know your data is linear correlated with step, you could set the degree to 1 and you have linear regression. In this case you don't want to specify any higher degree because your data will likely fit to a polynomial, which you know is not the case. In a similar way, if you know your data behaves like a polynomial of some degree, you certainly don't want to fit any higher. If you have really no clue what degree the polynomial should be, then you should make an educated guess. A good strategy is just to plot the polynomial going through the data points, upping the polynomial one degree at the time. You then want the line to go in between the points, and not diverge towards specific points. In case you have many outliers, there exists method that are better suited than least-squares.

Now onward to your problem specifically.

• There is no way to compute an optimal degree, unless you have more information about your data. Degree is a hyperparameter. If you want an optimum for it, you need to put additional prior information, such as "I want the lowest degree polynomial that fits the data with an error epsilon."
• Overfitting is simply fixed by choosing a lower degree polynomial. If that does not fix it, then least-squares regression is not for you. You need to look into a regression method that chooses a different metric, or you need to preprocess your data, or you need a non-polynomial fit (fit a function of certain shape, then use Levenberg-Marquardt for instance).
• 15-step derivative looks very questionable, you do likely not have this kind of smoothness in your data. If you have a good reason for this, tell us, otherwise just use 2 points for the first derivative, or 3 or 5 for the second.
• The expression with the Landau symbol (big-O) is not switching fourth order to second. If you subtract the two equations and divide by h^2 the O(h^4)/h^2 becomes O(h^2) .