
and the matrix norm is subordinated to the vectorial norm jxj if:
jAj¼max
jAxj
jxj
for any x 6¼ 0 ð7:85Þ
The subordinated norm is the smallest matrix norm compatible with the
norm jxj.
For example, the norm jAj
2
, defined as:
jAj
2
¼
ffiffiffiffiffi
1
p
ð7:86Þ
where
1
is the largest eigenvalue of A
H
A (A
H
¼ hermitic conjugate or trans-
pose of the complex conjugate matrix) is subordinated to the Euclidian norm
jxj
2
but the Frobisher norm:
jAj
F
¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
i
a
2
i
X
j
a
2
j
s
ð7:87Þ
is consistent with the Euclidian norm but not subordinated to it.
Finally, the following norms, which lead to faster calculations, are often
used:
jxj
1
¼
X
i
jx
i
jð7:88Þ
and the corr esponding norm for the matrix A:
jAj
1
¼ max jA
j
jð7:89Þ
where A
j
are the column vectors of A.
Calculation of the upper bound
From the norms of the experimental errors y and A, it is possible to calculate
an upper limit to the norm of the resulting error x by the use of the following
relation.
If j Ij¼1 (it is true for jIj
2
), then the norm of the relative error is:
jxj
jxj
jAj jA
1
j
1 jAj jA
1
j
jyj
jyj
þ
jAj
jAj
ð7:90Þ
The quantity:
condðAÞ¼jAj jA
1
jð7:91Þ
is of great importance in this calculation. It is the so-called condition number of
the matrix A related to the used norm. This number indicates how nearly
singular the matrix is.
If the spectral norm jAj
2
is used, we get the smallest possible condition
number, which is:
cond
2
ðAÞ¼jAj
2
jA
1
j
2
¼
ffiffiffiffiffiffiffiffiffiffi
1
n
p
ð7:92Þ
Common Methods and Techniques 163
Komentáře k této Příručce