Correlation

``Correlation" is a concept similar to cosine. A cosine measures the angle between two vectors. It is given by the dot product of the two vectors divided by their magnitudes:

$$c \eq { ({\bf x} \cdot {\bf y}) \over \sqrt{ ({\bf x} \cdot {\bf x}) ({\bf y} \cdot {\bf y}) } }$$ (48)

This is the sample normalized correlation we first encountered on page  as a quality measure of fitting one image to another.

Formally, the normalized correlation is defined using x and y as zero-mean, scalar, random variables instead of sample vectors. The summation is thus an expectation instead of a dot product:

$$c \eq {\E(xy) \over \sqrt{ \E(x^2)\, \E(y^2) } }$$ (49)

A practical difficulty arises when the ensemble averaging is simulated over a sample. The problem occurs with small samples and is most dramatically illustrated when we deal with a sample of only one element. Then the sample correlation is

$$\hat{c} \eq {xy \over \vert x\vert \, \vert y\vert} \eq \pm 1$$ (50)

regardless of what value the random number x or the random number y should take. For any n, the sample correlation $\hat{c}$ scatters away from zero. Such scatter is called ``bias." The topic of bias and variance of coherency estimates is a complicated one, but a rule of thumb seems to be to expect bias and variance of $\hat{c}$ of about $1/\sqrt{n}$ for samples of size n. Bias, no doubt, accounts for many false ``discoveries,'' since cause-and-effect is often inferred from correlation.