Decompose original rows by cyclic one-sided Jacobi column rotations.
The method rotates A and accumulates V; it never forms A transpose A.
A common maximum-entry scale protects the ordinary pair dot products.
Numerical null columns remain in V, including when there are fewer rows than
columns. Their left directions and relative accuracy remain unresolved.
This bounded kernel does not implement LAPACK's preconditioning, scaled-pair
rotations or high-relative-accuracy guarantees for extreme column disparities.
The one-sided construction is described by LAPACK DGESVJ:
https://www.netlib.org/lapack/lapack_routine/dgesvj.f
Decompose original rows by cyclic one-sided Jacobi column rotations. The method rotates A and accumulates V; it never forms A transpose A. A common maximum-entry scale protects the ordinary pair dot products. Numerical null columns remain in V, including when there are fewer rows than columns. Their left directions and relative accuracy remain unresolved. This bounded kernel does not implement LAPACK's preconditioning, scaled-pair rotations or high-relative-accuracy guarantees for extreme column disparities. The one-sided construction is described by LAPACK DGESVJ: https://www.netlib.org/lapack/lapack_routine/dgesvj.f