GraphMath

Tridiagonalization & Jacobi algorithm

When Householder preprocessing helps — and when it does not

Under construction. This chapter is still being developed and may change as the analysis and experiments are refined.

Can tridiagonalization improve a Jacobi eigensolver even though Jacobi destroys the tridiagonal zeros?

The usual objection is that Householder tridiagonalization costs O(n³) and Jacobi rotations immediately create fill-in. This chapter asks a different question: whether the transformed starting matrix can nevertheless reduce later Jacobi work. It analyzes energy redistribution and parallel pivot structure, then tests six symmetric matrix families with NVIDIA cuSOLVER.

Key ideas

Tridiagonalization does not give Jacobi a permanently sparse matrix, but it changes where the matrix energy is located before the Jacobi iterations begin.

  • Householder tridiagonalization concentrates all off-diagonal entries into the first off-diagonal band while preserving the matrix eigenvalues
  • Jacobi rotations do not preserve that tridiagonal structure; fill-in begins with the first set
  • Rotations acting on nonintersecting pivot pairs commute and can be processed as a parallel Jacobi set
  • For a tridiagonal starting matrix, the standard first Jacobi set acts on alternating first-band entries and, when the first-band energy is approximately evenly distributed, contains approximately half of the total off-diagonal energy
  • Householder similarity can move energy either toward or away from the diagonal; similar diagonal entries are favorable while large diagonal spread can be unfavorable
  • In the reported NVIDIA cuSOLVER experiments, preprocessing reduced Jacobi work in five of six tested matrix families; across all three tested GPUs, the complete eigenvalue-only pipeline was faster than standalone Jacobi for those five families from n = 128 onward
  • Whether sweep savings recover the O(n³) tridiagonalization cost remains matrix- and implementation-dependent

The chapter therefore reaches a conditional conclusion: Householder preprocessing can substantially help a parallel Jacobi eigensolver, but not for every matrix.

Why can tridiagonalization help if Jacobi immediately creates fill-in?

Because the possible benefit is not the permanent preservation of tridiagonal zeros. Tridiagonalization changes the starting distribution of diagonal and off-diagonal energy and concentrates all off-diagonal energy in the first band. That can give the standard first parallel Jacobi set a much stronger collection of pivots, although Householder similarity can also move energy away from the diagonal, so the effect is matrix-dependent.

What did the experiments show?

Experiments with NVIDIA cuSOLVER on six families of dense symmetric matrices showed a consistent separation between five ordinary test families and one deliberately adverse unequal-variance covariance family.

In the eigenvalue-only experiments, Householder preprocessing reduced the subsequent Jacobi work for the five ordinary families. The experiment was run on an RTX 3060, RTX 5060 Ti, and A100; across all three GPUs, the complete preprocessing pipeline was faster than standalone Jacobi from matrix size n = 128 onward. Averaged across the five ordinary families, the complete-pipeline runtime reduction ranged from 7.1% to 10.0% at n = 128, and from 20.8% to 31.6% at n = 1024. On the RTX 3060, repeating the experiment with five different random seeds produced the same qualitative family-level behavior.

A separate experiment requested eigenvectors in both paths. After including formation of the Householder matrix and the back-transformation of the eigenvectors, the same qualitative separation remained, although the crossover occurred at larger matrix sizes.

The unequal-variance covariance family behaved in the opposite direction on all three GPUs. Its large initial diagonal spread was associated with substantial transfer of energy away from the diagonal, more Jacobi work, and longer total runtime. Among the tested families, it was also the clear outlier in initial diagonal standard deviation.

Related chapters

Chapter contents

Was this chapter helpful?

Quick feedback helps us improve the site.