We study the permutational limit distribution of goodness-of-fit statistics computed in various forms of generalized canonical correlation analysis. In simple cases of canonical analysis the exact distribution can be computed. For somewhat more complicated forms approximations have been tabulated by Krishnaiah and others. For the most general forms of canonical analysis we must use resampling methods to generate the permutation distribution. Our approach is illustrated by examples of varying degree of complexity. For small examples we can actually study the quality of our approximations. For more complicated examples this is not possible, but the approximate permutation distributions themselves are a very valuable data analytical tool. (A)
Abstract