How do I analyze the correlation of set of points to this complicated parametric equation?

59 Views Asked by At

I am trying to quantify the correlation of a given set of points to the set of points defined by a parametric equation, as well as find the best $l$ to fit the points. However, I am unsure of how to go about doing this.

The parametric equation I am working with is $$x(a)=a\ln \left(\frac{l+\sqrt{l^2+a^2}}{a}\right)$$ $$y(a)=\sqrt{l^2+a^2}-a.$$

I have found $l\approx.373$. The range of the parameter $a$ is $0\lt a \lt \infty$.

The set of points I am trying to analyze is \begin{array}{c|c} \style{font-family:inherit}{x} & \style{font-family:inherit}{\text{y}}\\\hline .372 & .017 \\\hline .343 & .125 \\\hline .308 & .184 \\\hline .267 & .232 \\\hline .234 & .262 \\\hline .192 & .293 \\\hline .147 & .320 \\\hline .110 & .338 \\\hline .065 & .356 \\\hline .030 & .367 \\\hline .003 & .372 \end{array}

A plot of the parametric equation and the set of points suggests a strong correlation for $l\approx.373$. Here is a graph on Desmos showing this: parametric equation and plotted points (The parametric equation is in red and the data points are plotted in black).

Typically, when finding the correlation between a function and a plot of points, the plot is linearized as the function indicates and a linear regression is run to evaluate the correlation. Unfortunately, I have found no way to linearize the set of points. Are there any other ways to find the parametric curve of best fit?

1

There are 1 best solutions below

0
On BEST ANSWER

Plotting the data, just as you did, reveals the strong relation between $x$ and $y$.

Now, what I suspect is that they want you to find the best value of $l$ to improve the fit.

It is impossible to get rid of the parametrization in order to have $y=f(x)$. However, you can do the opposite.

Eliminate $a$ from $y$ $$a=\frac{l^2-y^2}{2 y}\qquad \implies \qquad x=\frac{(l^2-y^2) }{2 y}\,\log \left(\frac{l+y}{l-y}\right)$$

Using the starting value you gave, a basic nonlinear regression gives $$\begin{array}{clclclclc} \text{} & \text{Estimate} & \text{Standard Error} & \text{Confidence Interval} \\ l & 0.372631 & 0.000115 & \{0.37237,0.372892\} \\ \end{array}$$ and you can notice that $0.373$ is not in the confidence interval (even if very close to it).

For sure, this introduces a bias in the problem since what has been measured is $y$ and not $x$.

To compensate, I think that I should use orthogonal regression (total least squares).