Neutrally stable unreduced fixed points under QR iteration - PowerPoint PPT Presentation

1 / 10
About This Presentation
Title:

Neutrally stable unreduced fixed points under QR iteration

Description:

Non-convergence at fixed points: Batterson and Smillie 1989, Day unpublished ... Near reduced matrices: flag manifold structure. Away from reduced matrices: ... – PowerPoint PPT presentation

Number of Views:43
Avg rating:3.0/5.0
Slides: 11
Provided by: math97
Category:

less

Transcript and Presenter's Notes

Title: Neutrally stable unreduced fixed points under QR iteration


1
Neutrally stable unreduced
fixed points under QR iteration
  • Foundations of Computational Mathematics
  • 8 August 2002
  • David M. Day
  • Sandia National Laboratory

DOE insignia and the acknowledgement statement
MUST be used on the title slideof all
presentation material distribute outside of
Sandia.
Sandia is a multiprogram laboratory operated by
Sandia Corporation, a Lockheed Martin
Company,for the United States Department of
Energy under contract DE-AC04-94AL85000.
2
Outline
  • QR Iteration for algebraic eigenvalue problems
  • Shifts Generalized Raleigh quotient, Exceptional
  • Fixed Points
  • Differentiability
  • Neutral Stability
  • New Shift

3
QR iteration
  • HQR convergence problems seldom arise
  • Non-convergence due to unachievable convergence
    criteria du Croz,
  • Non-convergence at fixed points Batterson and
    Smillie 1989, Day unpublished
  • Domain unreduced Hessenberg matrices
  • Iteration Orthogonal similarity transformation

4
Shifts
  • Basic (unshifted) QR Not competitive
  • Raleigh quotient shift real
  • Double implicit generalized Raleigh quotient
  • HQR Every 10th iteration uses an exceptional
    shift to break symmetry
  • Ex Wilkinson and Reinsch, EISPACK, LAPACK all
    use different exceptional shifts
  • HQR neutrally stable fixed points
    with either Ex shift

5
Fixed Points
  • B QR, B Q Q B
  • Sequence of shifts p(B) QR, shift polynomial
    p
  • Fixed points p(B) Orthogonal scalar
  • Extension of Parlett 1966 to degree(p) gt 1
  • Fixed points not detectible
  • Implicit shifts R not computed
  • Efficiency iteration is in place
  • Theory characterize p() and B p(B) in O(n)?

6
Differentiability
  • Near reduced matrices flag manifold structure
  • Away from reduced matrices Euclidian structure

7
Neutral Stability
  • Reduced map orthogonality similarity class
  • Basic QR Both the osc map and the matrix map
    have spectrum on the boundary of the unit disc at
    any unreduced fixed point
  • Center manifold Dynamics depend on Riemannian
    metric
  • HQR Nonconvergence Need both osc map and the
    matrix map to have spectrum on the boundary of
    the unit disc (Experimental observation)

8
HQR Work
  • Coordinates on OSC Intrinsic
  • Criterion of Ammar Mehrmann 91 on symplectic
    orthogonal Schur decompositions
  • 4 by 4 matrices dim monic p() p(B) O
    scalar 1.
  • Potential Extensions Extrinsic coords, use
    existence proofs of symp orth Schur decomp by
    Mehl, Lin, and many others.

9
New Shift
  • Goal Make fixed points observable
  • Batterson 1999 Generalized Raleigh quotient
    shifts min norm last row p(B)
  • New max(norm first column/norm last row)
  • Near Target New converges to Old
  • Fixed points submatrices NW SE (positive
    subdiagonal) implies not a fixed points

10
Conclusions
  • QR iteration not globally convergent
  • Neutrally stable fixed points recurrent problem
  • Exceptional shifts do not eliminate all fixed
    points
  • Fixed points hard to detect
  • New shift fixed points observable
Write a Comment
User Comments (0)
About PowerShow.com