8

I am currently reading a paper that claims that the correlation coefficient for a uniform distribution on the interior of an ellipse

$$f_{X,Y} (x,y) = \begin{cases}\text{constant} & \text{if} \ (x,y) \ \text{inside the ellipse} \\ 0 & \text{otherwise} \end{cases}$$

is given by

$$\rho = \sqrt{1- \left(\frac{h}{H}\right)^2 }$$

where $h$ and $H$ are the vertical heights at the center and at the extremes respectively.

enter image description here

The author does not reveal how he reaches that and instead only says that we need to change scales, rotate, translate and of course integrate. I would very much like to retrace his steps but I am a bit lost with all that. I would therefore be grateful for some hints.

Thank you in advance.

Oh and for the record

Châtillon, Guy. "The balloon rules for a rough estimate of the correlation coefficient." The American Statistician 38.1 (1984): 58-60

It's quite amusing.

Dilip Sarwate
  • 41,202
  • 4
  • 94
  • 200
JohnK
  • 18,298
  • 10
  • 60
  • 103
  • 2
    Could you write down an expression for the ellipse? The _"height at the extreme"_ doesn't make sense for a standard ellipse: $$\frac{x^2}{a^2} + \frac{y^2}{b^2} = 1$$ since it has height $0$ at the extremes. Indeed, if $(X,Y)$ is uniformly distributed on the interior of the standard ellipse, then $\rho = 0$. – Dilip Sarwate Nov 17 '15 at 23:29
  • @DilipSarwate Yeah, I tried the standard case and computed $\rho = 0$, no problems there. What about the other cases though, where you need to scale, rotate etc? – JohnK Nov 17 '15 at 23:36
  • 2
    The entire operation appears fundamentally mistaken. The "change scales" part destroys the uniformity. A truly uniform distribution is achieved either as the limiting distribution within narrow (Euclidean) buffers of the curve or is uniform by arclength. In either case the normalizing constant is a complete elliptic function and won't possibly simplify to the expression given here. I'm not sure what $h$ and $H$ mean, but--as an example--the correlation coefficient for an ellipse with a major axis twice the minor axis, tilted at an angle of $\pi/6$, will be $0.78004$. – whuber Nov 17 '15 at 23:45
  • @whuber I have included a figure from the paper that explains what $h$ and $H$ stand for, I hope that this makes it clearer. – JohnK Nov 17 '15 at 23:52
  • The picture helps immensely: not only does it show what $h$ and $H$ are, it shows that the distribution is *not* over the ellipse, but over its *interior*! (The ellipse, by definition, is the curve.) – whuber Nov 18 '15 at 02:32
  • @whuber I apologise for the misunderstanding. – JohnK Nov 18 '15 at 08:21
  • 2
    If you would like a full, complete answer, then you will find it in my post at http://stats.stackexchange.com/a/71303/919. After all, when the ellipse is a circle the uniform is (obviously) circularly symmetric, so just about everything in that answer applies directly. In particular, by viewing the ellipse not as a rotation of a horizontal ellipse, but as a *skew transformation,* the formula for $\rho$ becomes obvious, because $\sqrt{1-\rho^2}=\lambda = h/H$ (using the notation in the section on "How to Create Ellipses"). – whuber Nov 18 '15 at 16:03
  • @whuber Wow! I wish I had read your answer to the other question before spending so much time on my own answer to this question. Thanks for the learning experience. – Dilip Sarwate Nov 18 '15 at 16:34

1 Answers1

10

Let $(X,Y)$ be uniformly distributed on the interior of the ellipse $$\frac{x^2}{a^2} + \frac{y^2}{b^2} = 1$$ where $a$ and $b$ are the semi-axes of the ellipse. Then, $X$ and $Y$ have marginal densities \begin{align} f_X(x) &= \frac{2}{\pi a^2}\sqrt{a^2-x^2}\,\,\mathbf 1_{-a,a}(x),\\ f_X(x) &= \frac{2}{\pi b^2}\sqrt{b^2-y^2}\,\,\mathbf 1_{-b,b}(y), \end{align} and it is easy to see that $E[X] = E[Y] = 0$. Also, \begin{align} \sigma_X^2 = E[X^2] &= \frac{2}{\pi a^2}\int_a^a x^2\sqrt{a^2-x^2}\,\mathrm dx\\ &= \frac{4}{\pi a^2}\int_0^a x^2\sqrt{a^2-x^2}\,\mathrm dx\\ &= \frac{4}{\pi a^2}\times a^4 \frac 12\frac{\Gamma(3/2)\Gamma(3/2)}{\Gamma(3)}\\ &= \frac{a^2}{4}, \end{align} and similarly, $\sigma_Y^2 = \frac{b^2}{4}$. Finally, $X$ and $Y$ are uncorrelated random variables.

Let \begin{align} U &= X\cos \theta - Y \sin \theta\\ V &= X\sin \theta + Y \cos \theta \end{align} which is a rotation transformation applied to $(X,Y)$. Then, $(U,V)$ are uniformly distributed on the interior of an ellipse whose axes do not coincide with the $u$ and $v$ axes. But, it is easy to verify that $U$ and $V$ are zero-mean random variables and that their variances are \begin{align} \sigma_U^2 &= \frac{a^2\cos^2\theta + b^2\sin^2\theta}{4}\\ \sigma_V^2 &= \frac{a^2\sin^2\theta + b^2\cos^2\theta}{4} \end{align} Furthermore, $$\operatorname{cov}(U,V) = (\sigma_X^2-\sigma_Y^2)\sin\theta\cos\theta = \frac{a^2-b^2}{8}\sin 2\theta$$ from which we can get the value of $\rho_{U,V}$.

Now, the ellipse on whose interior $(U,V)$ is uniformly distributed has equation

$$\frac{(u \cos\theta + v\sin \theta)^2}{a^2} + \frac{(-u \sin\theta + v\cos \theta)^2}{b^2} = 1,$$ that is, $$\left(\frac{\cos^2\theta}{a^2} + \frac{\sin^2\theta}{b^2}\right) u^2 + \left(\frac{\sin^2\theta}{a^2} + \frac{\cos^2\theta}{b^2}\right) v^2 + \left(\left(\frac{1}{a^2} - \frac{1}{b^2}\right)\sin 2\theta \right)uv = 1,$$ which can also be expressed as $$\sigma_V^2\cdot u^2 + \sigma_U^2\cdot v^2 -2\rho_{U,V}\sigma_U\sigma_V\cdot uv = \frac{a^2b^2}{4}\tag{1}$$ Setting $u = 0$ in $(1)$ gives $\displaystyle h = \frac{ab}{\sigma_U}$. while implicit differentiation of $(1)$ with respect to $u$ gives $$\sigma_V^2\cdot 2u + \sigma_U^2\cdot 2v\frac{\mathrm dv}{\mathrm du} -2\rho_{U,V}\sigma_U\sigma_V\cdot \left(v + u\frac{\mathrm dv}{\mathrm du}\right) = 0,$$ that is, the tangent to the ellipse $(1)$ is horizontal at the two points $(u,v)$ on the ellipse for which $$\rho_{U,V}\sigma_U\cdot v = \sigma_v\cdot u.$$ The value of $H$ can be figured out from this, and will (in the unlikely event that I have made no mistakes in doing the above calculations) lead to the desired result.

Dilip Sarwate
  • 41,202
  • 4
  • 94
  • 200