By way of light relief, this section will concentrate on the simple and practical problem of modelling the distribution of points using the multivariate gaussian function. I have already indicated how this can be used for making decisions as to what category a point belongs to, and I shall go into this matter in more detail shortly. At present, let us settle for an attempt to fit a good gaussian function to a set of data. Whether we see this as a form of data compression which replaces detailed knowledge of the data with some rather vaguer approximation thereto, or an estimate of some deeper underlying process responsible for having generated the data (by some ineffable mechanisms it is better not to eff) does not much matter. Either way we go through a certain computation. The terminology tends to suggest the estimation of some underlying ineffable whatnot, but it would be a mistake to suppose that the accompanying metaphysical baggage is either necessary or desirable, still less that the writer has a commitment to it.
To focus ideas, the reader might like to recall the problem of telling the guys from the gals by fitting gaussian distributions to the points on the height-weight diagram of Figure 1.2. back in chapter one, or of telling `aah' from `ooh' in twelve dimensions if he has had a look at the data of the disk or played with the fview program.