0
$\begingroup$

In stack overflow there are at least two questions asking how to find the straight line in a 2D space nearest to a given set of points $\vec{a}_i$ for $i = 1...N$: question 1 question 2.

In both cases the answer has been use a linear regression. That surprises to me, not only because linear regression doesn't covers the case of a vertical line, but because the measure of error is the "y" distance, not the euclidean distance from point to line. I've decided to check if the answer is the same.

The line will be defined by the vector $\vec{r}$ from origin to the nearest point of line. In polar coordinates lets say $\vec{r}=r \hat{\theta}$.

Error is then defined by the sum for all points of the individual quadratic error $\epsilon_i^2=(\vec{a}_i\hat{\theta}-r)^2$

Partial derivative on $r$ gives $r$ must be the average for all $i$ of $\vec{a}_i\hat{\theta}$.

Partial derivative on $\theta$ gives $ \frac{\partial}{\partial{\theta}}\epsilon_i^2= 2 \epsilon_i \vec{a}_i \frac{\partial \hat{\theta}}{\partial\theta}=2 (\vec{a}_i\hat{\theta}-r) \vec{a}_i \frac{\partial \hat{\theta}}{\partial\theta}$.

where $\frac{\partial \hat{\theta}}{\partial\theta}$ is a normal vector in the direction of the line, orthogonal to $\hat{\theta}$.

I do not know how I can simplify second equation till reach a solution expression that can be applied without need to solve the system using numerical methods (somethign I need for a software application about drawings).

As consequence, I can not answer myself the original question about if linear regression gives the correct optimal result or how far it is from the correct one. By example, if the solution is a vertical line or near than vertical line, is the linear regression result accurate or far from the optimal ?.

$\endgroup$
  • 2
    $\begingroup$ The reasons linear regression models are so ubiquitous are: A. they are analytically tractable and B. Generally you want to know how accurately the "model" predicts the data, so you are really interested in precisely the difference in the $y-$coordinates for fixed $x$. Still, theres no reason not to study different notions of approximation. Rather than try to work out a good analysis of the general problem, I'd fix some data, fit a line to it by regression, then see whether that line is optimal with restpect to the alternate definition. $\endgroup$ – lulu Feb 11 '19 at 18:55
  • $\begingroup$ @lulu: Thanks for your interest in the issue. In my case I'm solving the problem for a draw application, thus, interested in minimize the real euclidean distance. $\endgroup$ – pasaba por aqui Feb 11 '19 at 18:57
  • $\begingroup$ Sure. But, as I said, the linear regression model is often preferred simply because it is so pleasant analytically. To be sure, people do use alternate error measures all the time for various reasons. But you almost always need to use some non-linear optimizer to find the optimal fit for anything other than linear regressions. $\endgroup$ – lulu Feb 11 '19 at 18:59
  • 2
    $\begingroup$ Should have said: I don't imagine that your method coincides with the result obtained by linear regression. Worth checking by explicit example, of course. $\endgroup$ – lulu Feb 11 '19 at 19:01
  • 2
    $\begingroup$ @lulu: read about total least squares. This is what the OP needs. Ordinary regression is anisotropic. $\endgroup$ – Yves Daoust Feb 11 '19 at 19:10
3
$\begingroup$

What you are after is called total least squares.

Let the implicit equation of the line be

$$x\cos\theta+y\sin\theta+p=0$$

and you want to minimize

$$\sum(x\cos\theta+y\sin\theta+p)^2.$$

The optimal solution can be show to be centered on the centroid (by linearity) and the equation can be rewritten

$$(x-\overline x)\cos\theta+(y-\overline y)\sin\theta=0.$$

Now, assuming that the data has been centered, the minimizer of the least squares equation writes

$$0=2\sum(x\cos\theta+y\sin\theta)(-x\sin\theta+y\cos\theta)\\=(\cos^2\theta-\sin^2\theta)\sum2xy-2\sin\theta\cos\theta\sum(x^2-y^2).$$

Using the double angle formulas, you obtain the value of $\tan2\theta$.


Note that the last equation leaves a $\dfrac{k\pi}2$ undeterminacy on $\theta$, because there are two minima and two maxima over a full turn.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.