I don't have the book at hand at present, so I can't be sure what the authors meant and in what context. I will surmise that they are speaking of the most straightforward evidence. And the key is that they - I suppose - talk about some single linear combination, that is, single variable derived from two variables. "a linear combination of two variables would provide a better discriminator if they were correlated than if they were uncorrelated".
Correlated variables like on the left mean that the classes are aligned roughly along one line. It is this line that pretends to be the single discriminant (discriminator). As the cloud approaches straight sausage in shape the class centroids approach the red line. SSbetween for that one discriminant becomes maximal and almost exhausts the overal 2D dispersion of the centroids.
If the cloud is about round like a concretion of classes - the right pic - so that no correlation between the variables occur, no single line could pretend to be the sole discriminator. A red discriminant 1 must be supplemented by a green discriminant 2 in order to explain the full scatter of the centroids. Moreover, because the configuration is round, a discriminant's direction in space is not so uniquely determined as on the left: if we rotate it a bit it won't lose much of its as it is weak discerning power.
The key point is the shape of the total cloud rather than correlation per se. Below there is no correlation but the sausage persists, and the discrimination by a single discriminant is as good as on the first pic. On the first pic, both V1 and V2 variables contribute about equally to the linear combination which the discriminant is; on the last pic, only V1 contributes to it much, actually - coincides with it.
The issue we were speaking is the issue how effective is discriminant analysis as the dimensionality reducer for this or that dataset.
