On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Computing the smallest eigenvalueof symmetric tridiagonal matrices
via semiseparable techniques.
Raf Vandebril, Marc Van Barel,Nicola Mastronardi, Steven Delvaux & Yvette Vanberghen
Department of Computer ScienceK.U.Leuven
July 2006
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
Outline
IntroductionGoal of the talkSemiseparable matricesThe QR-algorithm
QR-iterations on the inverse
An implicit approach
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
Outline
IntroductionGoal of the talkSemiseparable matricesThe QR-algorithm
QR-iterations on the inverse
An implicit approach
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
Accurately computing the smallest eigenvalue
I Suppose we have a symmetric tridiagonal n× n matrix T,with one small eigenvalue.
|λ1| ≈ 10−x with 6 ≤ x ≤ 14
|λi| ≈ 1 with 2 ≤ i ≤ n
The smallest one often suffers from loss in accuracy.I Our aim is to compute the eigenvalue λ1 as accurate as
possible, in linear time.I We will base ourselves on the QR-algorithm.I We will make use of semiseparable matrices. These
matrices are the inverses of tridiagonal matrices.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
Outline
IntroductionGoal of the talkSemiseparable matricesThe QR-algorithm
QR-iterations on the inverse
An implicit approach
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
A semiseparable matrix
I A matrix A is called semiseparable if all submatrices takenout of the lower triangular part have rank 1. (Similar for theupper triangular part.)
I All the red blocks are of rank 1.× × × × ×× × × × ×× × × × ×× × × × ×× × × × ×
I The inverse of a tridiagonal matrix is semiseparable and
can be computed in O(n) operations.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
A semiseparable matrix
I A matrix A is called semiseparable if all submatrices takenout of the lower triangular part have rank 1. (Similar for theupper triangular part.)
I All the red blocks are of rank 1.� × × × ×� × × × ×� × × × ×� × × × ×� × × × ×
I The inverse of a tridiagonal matrix is semiseparable andcan be computed in O(n) operations.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
A semiseparable matrix
I A matrix A is called semiseparable if all submatrices takenout of the lower triangular part have rank 1. (Similar for theupper triangular part.)
I All the red blocks are of rank 1.× × × × ×� � × × ×� � × × ×� � × × ×� � × × ×
I The inverse of a tridiagonal matrix is semiseparable andcan be computed in O(n) operations.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
A semiseparable matrix
I A matrix A is called semiseparable if all submatrices takenout of the lower triangular part have rank 1. (Similar for theupper triangular part.)
I All the red blocks are of rank 1.× × × × ×× × × × ×� � � × ×� � � × ×� � � × ×
I The inverse of a tridiagonal matrix is semiseparable andcan be computed in O(n) operations.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
A semiseparable matrix
I A matrix A is called semiseparable if all submatrices takenout of the lower triangular part have rank 1. (Similar for theupper triangular part.)
I All the red blocks are of rank 1.× × × × ×× × × × ×× × × × ×� � � � ×� � � � ×
I The inverse of a tridiagonal matrix is semiseparable andcan be computed in O(n) operations.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
A semiseparable matrix
I A matrix A is called semiseparable if all submatrices takenout of the lower triangular part have rank 1. (Similar for theupper triangular part.)
I All the red blocks are of rank 1.× × × × ×× × × × ×× × × × ×× × × × ×� � � � �
I The inverse of a tridiagonal matrix is semiseparable and
can be computed in O(n) operations.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
InversionSuppose we have a matrix T, inverting this matrix is almoststraightforward using:
I LU-factorization (L is lower bidiagonal, U is upperbidiagonal):
T = LU → S = T−1 = U−1L−1,
with U−1 of upper triangular semiseparable form and L−1
of lower triangular semiseparable form.I QR-factorization (Q is a unitary Hessenberg matrix, R is
upper triangular, with two nonzero superdiagonals):
T = QR → S = T−1 = R−1Q−1,
with Q−1 unitary upper Hessenberg and R−1 an uppertriangular semiseparable matrix of semiseparability rank 2.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
Inversion
Or via direct inversion, with
T =
a1 b1b1 a2 b2
b2 a3 b3b3 a4
and
S =
c1v1 c2s1v1 c3s2s1v1 s3s2s1v1
c2s1v1 c2v2 c3s2v2 s3s2v2c3s2s1v1 c3s2v2 c3v3 s3v3s3s2s1v1 s3s2v2 s3v3 v4
.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
Outline
IntroductionGoal of the talkSemiseparable matricesThe QR-algorithm
QR-iterations on the inverse
An implicit approach
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Goal of the talkSemiseparable matricesThe QR-algorithm
The QR-algorithm
I Suppose we have an n× n matrix A. Computing the matrixA, gives us one step of the QR-algorithm with shift κ:
A− κI = QR
A = RQ + κI.
A converges to the largest eigenvalue, and on the bottomright to the smallest eigenvalue, w.r.t. the shift!!
I Repeating this procedure gives us the standardQR-algorithm. This method lets A converge to a diagonalmatrix, with the eigenvalues on the diagonal.
I One step of the QR-method applied on a semiseparable oron a tridiagonal matrix costs O(n) operations.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
QR-algorithm with shiftQR-algorithm without shift
Outline
Introduction
QR-iterations on the inverseQR-algorithm with shiftQR-algorithm without shift
An implicit approach
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
QR-algorithm with shiftQR-algorithm without shift
Outline
Introduction
QR-iterations on the inverseQR-algorithm with shiftQR-algorithm without shift
An implicit approach
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
QR-algorithm with shiftQR-algorithm without shift
QR-algorithm with shift on the inverse
I QR-algorithm with shift on the matrix T and on its inverse S.
I Standard QR-algorithm (with shift) computes first thelargest eigenvalues.
I This results in a lot of flops before the smallest one iscomputed in the tridiagonal case.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
QR-algorithm with shiftQR-algorithm without shift
Results of QR-algorithm with shiftI The smallest eigenvalue via the semiseparable approach,
is a little bit more accurate (λ1 ≈ 10−12).I But, we do loose accuracy when inverting the tridiagonal
matrix. (Blue is semiseparable, Green is tridiagonal)
1 2 3 4 5 610
−5
10−4
10−3
10−2
10−1
100
Experiments (size*100)
Rel
ativ
e A
ccur
acy
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
QR-algorithm with shiftQR-algorithm without shift
Outline
Introduction
QR-iterations on the inverseQR-algorithm with shiftQR-algorithm without shift
An implicit approach
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
QR-algorithm with shiftQR-algorithm without shift
QR-algorithm without shiftI Without shift means no forced convergence to the largest
eigenvalue (λ1 ≈ 10−12).I Both approaches perform equally well.
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 610
−9
10−8
10−7
Experiments (size*100)
Rel
ativ
e ac
cura
cy
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Outline
Introduction
QR-iterations on the inverse
An implicit approachThe explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea to compute the smallest eigenvalue
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Outline
Introduction
QR-iterations on the inverse
An implicit approachThe explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea to compute the smallest eigenvalue
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
The QR-algorithm, the explicit approachSuppose we have a shift κ, and we calculate theQR-decomposition of the matrix A− κI:
A− κI = QR
and let us denote with A the matrix:
A = RQ + κI.
The matrix A is the result of performing one step of theQR-algorithm on the matrix A.
When we perform the operations as presented here, this iscalled the explicit approach.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Outline
Introduction
QR-iterations on the inverse
An implicit approachThe explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea to compute the smallest eigenvalue
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
The implicit approach
We can rewrite the formulas from before as:
(A− κI) = QR ⇒ Q−1 (A− κI) = R
substituting R in the next equation
A = RQ + κI
gives usA = Q−1AQ.
When we apply the transformations Q and Q−1 at the sametime, we have an implicit algorithm.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Why an implicit algorithm?
I If only the eigenvalues are desired, we do not need to storeQ.
I We do not lose accuracy due to the adding and subtractingof the shift matrix κI, compared to the explicitQR-algorithm with shift.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Outline
Introduction
QR-iterations on the inverse
An implicit approachThe explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea to compute the smallest eigenvalue
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Combined QR-iterations
I Inverting causes loss in accuracy!I Suppose we have a tridiagonal matrix T with its inverse
S = T−1.I If S converges to the dominant eigenvalue in S = QHSQ, we
have that:
T = S−1
QHTQ = QHS−1Q =(QHSQ
)−1[T ε
εT λ1
]=
[S ε
εT 1/λ1
]−1
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
A first attempt
I As seen before, when using the QR-algorithm with shift,we get quick convergence to the largest eigenvalue of S.
I Using now this Q to compute:
T = QHTQ,
would force the matrix T to converge to its smallesteigenvalue.
I Numerical results are not satisfactory.I Due to small rounding errors, causing the factor Q, not to
be exactly equal to the factor Q for T we wanted.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Outline
Introduction
QR-iterations on the inverse
An implicit approachThe explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea to compute the smallest eigenvalue
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
The QR-factorization of tridiagonal matrices
The QR-factorization of the tridiagonal matrix A− κI, consists ofn− 1 Givens transformations performed from the bottom to thetop.
× ×⊗ × ×
× × ×× × ×
× ×
I Apply a Givens transformation
on the left.I Involving row 1 and 2.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
The QR-factorization of tridiagonal matrices
The QR-factorization of the tridiagonal matrix A− κI, consists ofn− 1 Givens transformations
× × ×× ×⊗ × ×
× × ×× ×
I Apply a Givens transformation
on the left.I Involving row 2 and 3.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
The QR-factorization of tridiagonal matrices
The QR-factorization of the tridiagonal matrix A− κI, consists ofn− 1 Givens transformations
× × ×× × ×
× ×⊗ × ×
× ×
I Apply a Givens transformation
on the left.I Involving row 3 and 4.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
The QR-factorization of tridiagonal matrices
The QR-factorization of the tridiagonal matrix A− κI, consists ofn− 1 Givens transformations
× × ×× × ×
× × ×× ×⊗ ×
I Apply a Givens transformation
on the left.I Involving row 4 and 5.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
The QR-factorization of tridiagonal matrices
The QR-factorization of the tridiagonal matrix A− κI, consists ofn− 1 Givens transformations
× × ×× × ×
× × ×× ×
×
I We have now an uppertriangular matrix R.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Implicit QR-algorithm for tridiagonal matrices
As mentioned before we want to apply the transformations atboth sides at the same time
× ×× × ×
× × ×× × ×
× ×
I Apply the first Givens
transformation on both sidesof the matrix.
I Involving row 1 and 2.I Involving column 1 and 2.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Implicit QR-algorithm for tridiagonal matrices
As mentioned before we want to apply the transformations atboth sides at the same time
× × ⊗× × ×⊗ × × ×
× × ×× ×
I Apply a Givens transformation
to annihilate the bulge.I Involving row 2 and 3.I Involving column 2 and 3.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Implicit QR-algorithm for tridiagonal matrices
As mentioned before we want to apply the transformations atboth sides at the same time
× ×× × × ⊗
× × ×⊗ × × ×
× ×
I Apply a Givens transformation
to annihilate the bulge.I Involving row 3 and 4.I Involving column 3 and 4.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Implicit QR-algorithm for tridiagonal matrices
As mentioned before we want to apply the transformations atboth sides at the same time
× ×× × ×
× × × ⊗× × ×⊗ × ×
I Apply a Givens transformation
to annihilate the bulge.I Involving row 4 and 5.I Involving column 4 and 5.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Implicit QR-algorithm for tridiagonal matrices
As mentioned before we want to apply the transformations atboth sides at the same time
× ×× × ×
× × ×× × ×
× ×
I We have again a tridiagonalmatrix.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
The implicit Q-theorem
I In the implicit approach, it seemed that only one Givenstransformation was needed to start the process.
I So the algorithm is as follows:I Compute the first Givens transformation of the
QR-factorization of T − κI.I Apply it on both sides of the matrix T.I Chase the bulge.
I The implicit Q-theorem proves that this approach performsa step of the QR-algorithm on T.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Outline
Introduction
QR-iterations on the inverse
An implicit approachThe explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea to compute the smallest eigenvalue
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
The idea
I We know that if S converges to the largest via theQR-algorithm, that T converges to the smallest.
I Determine the first Givens transformation of the implicitapproach for T, via the QR-factorization of S.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Algorithm 1
A first algorithm has the following structure:1. Invert the matrix T. Denote S = T−1.2. Compute the initial Givens transformation for S.3. Perform the implicit QR-step on T.
Store this in T.4. Perform the implicit QR-step on S.
Store this in S.5. Start again at step 2.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Remarks on algorithm
Remarks on this algorithm:1. Due to small rounding errors, we get more and more
divergence between the intial Givens of S and the initialGivens we need in T.
2. This can result in slow convergence.3. Slow convergence causes loss in accuracy.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Algorithm 2
A more stable approach1. Invert the matrix T. Denote S = T−1.2. Compute the initial Givens transformation for S.3. Perform the implicit QR-step on T.
Store this in T.4. Start again at step 1.
The complexity is similar to the complexity of the first algorithm.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Remarks on algorithm
Remarks on this algorithm:1. More stable due to the forced inversion of T.2. This leads to a more accurate first Givens transformation.3. Faster convergence.4. Less loss in accuracy.
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
The explicit approachThe implicit approachCombined QR-iterations without shiftThe implicit QR-algorithm of a tridiagonalAlgoritmic idea
Remarks on algorithmBlue is the new approach, green represents the traditionalapproach.
1 2 3 4 510
−5
10−4
10−3
10−2
10−1
100
Experiments (size*100)
Rel
ativ
e A
ccur
acy
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Outline
Introduction
QR-iterations on the inverse
An implicit approach
Conclusions & Future work
On the orthogonal reductions to semiseparable form
On computing the smallesteigenvalue of
symmetric tridiagonalmatrices
IntroductionQR-iterations on the inverse
An implicit approachConclusions & Future work
Conclusions
I No dramatic increase in accuracy.I Improving convergence speed can be done via
I More accurate inversion of the tridiagonal.I Using shifted QR for the semiseparable to determine the
first Givens.I Improving accuracy can be done via
I Iterative refinement.I More accurate first Givens transformation in the implicit
approach, i.e. more accurate inversion.
On the orthogonal reductions to semiseparable form