Rounding Errors and Stability of Algorithms
Rounding Errors and Stability of Algorithms
Rounding Errors and Stability of Algorithms
(G + δG)yc = b (6)
where δG is lower (upper) triangular such that
|δG| ≤ 2nu|G| + O(u 2 ) (7)
Backward error analysis of backward and forward
substitution
Some notation:
I For any two matrices C = [cij ]n×m and F = [fij ]n×m , we write
C ≤ F if cij ≤ fij for all 1 ≤ i ≤ n, 1 ≤ j ≤ m.
I For any C = [cij ]n×m , |C| = [|cij |]n×m .
(G + δG)yc = b (6)
where δG is lower (upper) triangular such that
|δG| ≤ 2nu|G| + O(u 2 ) (7)
Here (7) means |δgij | ≤ 2nu|gij | + O(u 2 ) for all 1 ≤ i, j ≤ n and O(u 2 )
stands for a matrix whose entries are O(u 2 ).
Backward error analysis of backward and forward
substitution
Some notation:
I For any two matrices C = [cij ]n×m and F = [fij ]n×m , we write
C ≤ F if cij ≤ fij for all 1 ≤ i ≤ n, 1 ≤ j ≤ m.
I For any C = [cij ]n×m , |C| = [|cij |]n×m .
(G + δG)yc = b (6)
where δG is lower (upper) triangular such that
|δG| ≤ 2nu|G| + O(u 2 ) (7)
Here (7) means |δgij | ≤ 2nu|gij | + O(u 2 ) for all 1 ≤ i, j ≤ n and O(u 2 )
stands for a matrix whose entries are O(u 2 ).
For a proof see pp. 159-161 of Fundamentals of Matrix Computations, 2nd Edition.
Backward error analysis of backward and forward
substitution
Exercise Given any two matrices C = [cij ]n×m and F = [fij ]n×m , show
that
1. |C| ≤ |F | ⇒ kCkp ≤ kF kp where p = 1, F , ∞.
2. kCkp = k|C|kp for p = 1, F , ∞.
Backward error analysis of backward and forward
substitution
Exercise Given any two matrices C = [cij ]n×m and F = [fij ]n×m , show
that
1. |C| ≤ |F | ⇒ kCkp ≤ kF kp where p = 1, F , ∞.
2. kCkp = k|C|kp for p = 1, F , ∞.
(G + δG)yc = b
where δG is lower (upper) triangular such that
(G + δG)yc = b
where δG is lower (upper) triangular such that
A + δA = Lc Uc
where
|δA| ≤ 2nu|Lc ||Uc | + O(u 2 ) (9)
and consequently,
A + δA = Lc Uc
where
|δA| ≤ 2nu|Lc ||Uc | + O(u 2 ) (9)
and consequently,
(A + E)xc = b
where
|E| ≤ 6nu|Lc ||Uc | + O(u 2 ) (11)
and consequently,
kUc k∞ √
For GEPP it is observed that for almost all matrices in practice, kAk∞
≈ n.
Backward error analysis of Gaussian Elimination
Therefore the process is backward stable if
kLc k∞ kUc k∞
kAk∞
is small.
In GENP, the presence of small pivots can make kLc k∞ very large.
Therefore, GENP is unstable.
In GEPP and GECP, kLc k∞ ≤ n. Therefore,
kUc k∞ √
For GEPP it is observed that for almost all matrices in practice, kAk∞
≈ n.
kUc k∞ 2n−1
However, kAk∞
= n
when A is the n × n Wilkinson matrix in GEPP.
Backward error analysis of Gaussian Elimination
Therefore the process is backward stable if
kLc k∞ kUc k∞
kAk∞
is small.
In GENP, the presence of small pivots can make kLc k∞ very large.
Therefore, GENP is unstable.
In GEPP and GECP, kLc k∞ ≤ n. Therefore,
kUc k∞ √
For GEPP it is observed that for almost all matrices in practice, kAk∞
≈ n.
kUc k∞ 2n−1
However, kAk∞
= n
when A is the n × n Wilkinson matrix in GEPP.
Therefore GEPP is conditionally backward stable as stabiity is subject to
kUc k∞
kAk∞
being small.
Backward error analysis of Gaussian Elimination
Therefore the process is backward stable if
kLc k∞ kUc k∞
kAk∞
is small.
In GENP, the presence of small pivots can make kLc k∞ very large.
Therefore, GENP is unstable.
In GEPP and GECP, kLc k∞ ≤ n. Therefore,
kUc k∞ √
For GEPP it is observed that for almost all matrices in practice, kAk∞
≈ n.
kUc k∞ 2n−1
However, kAk∞
= n
when A is the n × n Wilkinson matrix in GEPP.
Therefore GEPP is conditionally backward stable as stabiity is subject to
kUc k∞
kAk∞
being small.
kUc k∞
For GECP it can be proved that kAk∞
= O(n).
Backward error analysis of Gaussian Elimination
Therefore the process is backward stable if
kLc k∞ kUc k∞
kAk∞
is small.
In GENP, the presence of small pivots can make kLc k∞ very large.
Therefore, GENP is unstable.
In GEPP and GECP, kLc k∞ ≤ n. Therefore,
kUc k∞ √
For GEPP it is observed that for almost all matrices in practice, kAk∞
≈ n.
kUc k∞ 2n−1
However, kAk∞
= n
when A is the n × n Wilkinson matrix in GEPP.
Therefore GEPP is conditionally backward stable as stabiity is subject to
kUc k∞
kAk∞
being small.
kUc k∞
For GECP it can be proved that kAk∞
= O(n).
Hence GECP is unconditionally backward stable.
GEPP: The algorithm of choice for solving square
systems
A + δA = GcT Gc
A + δA = GcT Gc
Consequently,
kδAkF 2n3/2 u
≤ + O(u 2 ).
kAkF 1 − 2n3/2 u
Backward error analysis of Cholesky factorization
Theorem Let A be an n × n positive definite matrix and Gc be the
Cholesky factor computed in floating point arithmetic via some
version of Cholesky factorization algorithm. Then
A + δA = GcT Gc
Consequently,
kδAkF 2n3/2 u
≤ + O(u 2 ).
kAkF 1 − 2n3/2 u
A + δA = GcT Gc
Consequently,
kδAkF 2n3/2 u
≤ + O(u 2 ).
kAkF 1 − 2n3/2 u