Next: Locking and Purging in
Up: Orthogonal Deflating Transformation
Previous: Purging .
  Contents
  Index
The deflating orthogonal transformation construction shown in
Algorithm 7.8 is
clearly stable (i.e., componentwise relatively accurate representation of
the transformation one would obtain in exact arithmetic). There is
no question that the similarity transformation
numerically preserves the eigenvalues of . However, there is a serious
question about how well these transformations perform numerically in preserving
Hessenberg form during purging. A modification to the basic
algorithm is needed to assure that if
,
then
is numerically Hessenberg
(i.e., that the entries below the subdiagonal are all tiny relative
to ).
The fact that is
Hessenberg depends upon the term vanishing in the expression
However, on closer examination, we see that
where is the first component of . Therefore,
|| g || = 1| _1 |,
so there may be numerical difficulty when the first component of
is small.
To be specific,
and
thus in exact arithmetic. However, in finite
precision, the computed
.
The error will be on the order of relative to ,
but
so this term may be quite large. It may be as large as order
if
.
This is of serious concern and will occur in practice without the modification
we now propose.
The remedy is to introduce a step-by-step acceptable perturbation and
rescaling of the vector to simultaneously force the conditions
to hold with sufficient accuracy in finite precision.
To accomplish this, we shall devise a scheme to achieve
numerically. As shown in [420], this modification is sufficient to
establish
numerically, relative to
for .
The basic idea is as follows: If, at the th step, the computed
quantity
is not sufficiently small, then it is adjusted
to be small enough by scaling the vector with a number
and the component with a number .
With this rescaling just prior to the computation of , we then
have
and
,
where
.
Certainly, should not be altered with this scaling and this
is therefore required.
This gives the following system of equations to determine
and : if
,
(_j )^2 + (1 - _j^2)^2 &=& 1,
y_j^* H_j q_j + _j _j _j+1 &=&
_M _j+1.
If is on the order of
, then the scaling
may be absorbed into without alteration of and also
without effecting the numerical orthogonality of .
When is modified, it turns out that none of the previously
computed , , need to be altered.
After step , the vector is simply rescaled in subsequent steps, and
the formulas defining
,
are invariant with respect to scaling of .
For complete detail, one should consult [420].
The code shown in Algorithm 7.9 implements this scheme
to compute an acceptable .In practice, this transformation may
be computed and applied in place to obtain
and without storing . However, that implementation
is quite subtle and the construction of is obscured by the details
required to avoid creating additional storage.
This code departs slightly from the above description since the
vector is rescaled to have unit norm only after all of
the columns to have been determined.
There are several implementation issues.
- (3)
- The perturbation shown here avoids problems with exact zero initial entries in
the eigenvector . In theory, this should not happen when is unreduced
but it may happen numerically when a diagonal entry of is small but not zero.
There is a cleaner implementation possible that does not modify zero entries.
This is the simplest (and crudest) correction.
- (10)
- As soon as
is sufficiently large, there is no need for
any further corrections. Deleting this if-clause reverts to the unmodified
calculation of shown in Algorithm 7.8.
- (16)
- This shows one of several possibilities for
modifying to achieve the desired
goal of numerically tiny elements below the subdiagonal of .
More sophisticated strategies would absorb as much of the scaling as possible
into the diagonal element . Here, the branch that does
scale instead of is designed so
that
.
Next: Locking and Purging in
Up: Orthogonal Deflating Transformation
Previous: Purging .
  Contents
  Index
Susan Blackford
2000-11-20