In class we were discussing the geometry of $\rho$-correlated, $n$-dimensional Gaussians; call them $\vec{g}$, $\vec{h}$.

We have that $\| \vec{g} \|^2 = \sum_{i=1}^n g_i^2$. This is the sum of $n$ very well-behaved independent random variables with mean $1$ and variance $2$, so its distribution value will be $n \pm O(\sqrt{n})$ with very high probability. (To be more accurate, say $n \pm O(\sqrt{n} \log n)$ with probability at least $1 - n^{-O(1)}$.) Hence $\|\vec{g}\|$ will be $\sqrt{n}(1 + O(1/\sqrt{n})$ with high probability. Since $\vec{h}$ is also distributed as a standard $n$-dimensional Gaussian, the same is true of $\vec{h}$.

Now imagine $\vec{g}$ is fixed and we choose $\vec{h}$ to be $\rho$-correlated to $\vec{g}$. Then $\vec{g} \cdot \vec{h} = \sum_{i=1}^n g_i h_i = \sum_{i=1}^n (\rho g_i^2 + \sqrt{1-\rho^2} g_i g_i')$, where the $g_i'$ random variables are independent standard normals. I.e., the dot product is $\rho \|\vec{g}\|^2$ plus an independent one-dimensional normal, $\sqrt{1-\rho^2} \cdot N(0, \sum_i g_i^2)$. Since $\|\vec{g}\|^2 = \sum_i g_i^2 = n \pm O(\sqrt{n})$ with high probability, the dot product is $\rho n \pm O((\rho + \sqrt{1-\rho^2})\sqrt{n})$ with high probability.

So the cosine of the angle between $\vec{g}$ and $\vec{h}$ will be $(\rho n \pm O(\sqrt{n}))/(n \pm O(\sqrt{n})) = \rho \pm O(1/\sqrt{n})$, and hence the angle will be $\arccos \rho \pm O(1/\sqrt{n})$ with high probability (assuming $\rho$ is treated as a constant in $(0,1)$).

Now this doesn't really yet prove that $\vec{g}$ and $\vec{h}$ are distributed like a random pair on the surface of the $\sqrt{n}$-radius sphere with angle $\arccos \rho$. We just know that their angle will be $\arccos \rho$ and they will both be on the sphere. We should really look at the distribution of $\vec{h} - \rho \vec{g}$. But this is just a scaled $n$-dimensional Gaussian, so its distribution is spherically symmetric. I guess this more or less justifies the overall claim.

Actually, I believe that $\vec{h} - \rho \vec{g}$ will be close to orthogonal to $\rho \vec{g}$ with high probability. The dot product between the two will be distributed like $\sqrt{1-\rho^2} \cdot N(0, \sum_i g_i^2)$, which will probably be on the order of $\sqrt{1-\rho^2}\sqrt{n}$. However the product of the lengths of the two vectors will be like $\rho \sqrt{1-\rho^2} n$. So the cosine of the angle between them will be like $1/(\rho \sqrt{n})$, and the angle will be close to $90$ degrees.

## Thursday, April 05, 2007

Subscribe to:
Post Comments (Atom)

## No comments:

Post a Comment