From Bareiss' Algorithm to the Stable Computation of Partial Correlations
Abstract
This paper presents the derivation of a new algorithm for the stable computation of sample partial correlation coefficients. The authors start with Bareiss' algorithm for the solution of linear system of equations with (nonsymmetric) Toeplitz coefficient matrix and show how to generalize it to matrices that are not Toeplitz. The so generalized Bareiss algorithm computes the LU and UL factorizations of those matrices whose contiguous principal submatrices are all non-singular. For symmetric positive-definite matrices A, which naturally satisfy this condition, the normalized version of Bareiss' algorithm is just the Hyperbolic Cholesky algorithm, which computes the upper and lower triangular Cholesky factors U and L of A by means of 2 x 2 hyperbolic rotations. Guided by the data flow graph of the Hyperbolic Cholesky algorithm, we show that there exists one sequence of 2 x 2 orthogonal rotations that effects the transformation from U to L such that the sines of these rotations equal the hyperbolic tangents of the hyperbolic rotations. From the connection to the Hyperbolic Cholesky algorithm it also follows that, if A is a sample covariance matrix, the sines ar sample partial correlation coefficients.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 1988
- Accession Number
- ADA198696
Entities
People
- Ilse C. Ipsen
- Jean-marc Delosme
Organizations
- Yale University