Parallel algorithms for canonical variates computation

A generalization of the singular value decomposition is introduced, and its suitability for canonical variate analysis is demonstrated. A generalization called the HK singular value decomposition (HK-SVD), which involves the simultaneous diagonalization of three matrices, is presented. The four basic matrix operations for canonical correlations are found. For an n×n matrix, these procedures all require O(n³) FLOPS. It is shown that, given a computer with O(n²) processors, the execution time can be cut down to O(n); a viable candidate for the computing is the connection machine (CM). The floating point processing units on the CM are described