Numerical Methods for PDEs

2024 – 2025
Correction of exercise session 11

Exercise 11.1
The algorithms are equivalent. To see this, we denote the correction in the unpreconditioned
version with d
e m and relate in brackets the quantities in the unpreconditioned version to those
in the preconditioned one using the same notation in both cases; the relation are

uk = B −1 u
ek , dk = B −1 d
ek .

Stationary iterations for Stationary iterations for

A B −1 u
e = b, u = B −1 u
e Au = b

without preconditioning. with B preconditioner.

Initialization: Initialization:

u0 arbitrary u0 arbitrary
u
e 0 = B u0 r0 = b − A u0
r0 = b − AB −1 u

0
e
For m = 0, 1, . . . :
= b − Au0
dm = B −1 rm
For m = 0, 1, . . . :
um+1 = um + τ dm
d
e m = rm rm+1 = b − Aum
dm = B −1 rm

u
e m+1 = u
em + τ dem


um+1 = um + τ dm
rm+1 = b − AB −1 u

m
e
= b − Aum

End (after M iterations):

uM = B −1 u
eM
Exercise 11.2
The algorithms are equivalent (both for standard and flexible variants). To see this, we
denote the correction and the residual in the unpreconditioned version with d e m and erm
respectively, and relate in brackets the quantities in the unpreconditioned version to those in
the preconditioned one using the same notation in both cases; the relation are

uk = L−T u
ek , dk = L−T d
ek , rk = L e
rk ,

with B −1 = L−T L−1 . For the standard CG both algorithms are given below.
Preconditioned Conjugate Gradient for Preconditioned Conjugate Gradient for

L−1 A L−T u
e = L−1 b , u = L−T u
e Au = b

without preconditioning. with B = LLT preconditioner.

Initialization: Initialization:

u0 arbitrary u0 arbitrary
u
e0 = LT u0 r0 = b − A u0
r0
e = L−1 b − L−1 A L−T u
e0 d−1 = 0


r0 = b − A u0

For m = 0, 1, . . . :
d
e −1 = 0 d−1 = 0
rTm B −1 rm
For m = 0, 1, . . . : βm = (β0 = 0)
rTm−1 B −1 rm−1
rTm e
e rm   dm = B −1 rm + βm dm−1
βem = β0 = 0
e
rTm B −1 rm
rTm−1 e
e rm−1 αm =
rT L−T L−1 rm dTm A dm
= T m −T −1


= βm um+1 = um + αm dm
rm−1 L L rm−1
rm+1 = rm − αm A dm
d
em = e
rm + βem d e m−1
= L−T L−1 rm + βm dm−1


dm
erTm e
rm
α
em =
e T L−1 A L−T d
d em
m
rm L−T L−1 rm
T

= = αm
dTm A dm
u
e m+1 = u
em + α
em d
em


um+1 = um + αm dm
rm+1 = e
e em L−1 A L−T d
rm − α em


rm+1 = rm − αm A dm

End (after M iterations):

uM = L−T u
eM

Both algorithms minimize ∥e e m ∥L−1 AL−T = ∥L−T u

u−u e − L−T u
e m ∥A = ∥u − um ∥A .

2
Exercise 11.3
The Arnoldi method from p.72 of Section 5.4 applied to a symmetric L−1 AL−T is given below
on the left. The transformed algorith is given on the right using the hinted transformations:

vk = L v
ek , wk = L w
ek , rk = L e
rk , zk = L−T v
ek = B −1 vk ,

with B −1 = L−T L−1 .

Arnodli for symmetric matrix A variant of Arnodli for smmetric A and
SPD B .
L−1 A L−T .
Initialization:
Initialization:
r0 arbitrary
r0 arbitrary ž1 = B −1 r0
r0 = L−1 r0 q

žT1 r0
e
β = = ∥e
r0 ∥
β = ∥er0 ∥
e1 = β −1e
v r0 z1 = β −1 ž1
v1 = β −1 r0

For m = 1, 2, . . . :
1. w
emmax(0,m−2)
= L−1 AL−T v
em For m = 1, 2, . . . :
max(0,m−2)
2. For i = max(1, m − 1), m : 1. wm = A zm
(i−1) 2. For i = max(1, m − 1), m :
eiT w
hi m = v em (i−1)
(i) (i−1) hi m = zTi wm
w
em = w
em − hi m v
ei (i) (i−1)
3. hm+1 m = ∥w
(m)
em ∥ wm = wm − hi m vi
(m)
(m) 3. žm+1 = B −1 wm
em+1 = h−1
4. v m+1 m w
em q
(m)
hm+1 m = žTm+1 wm
4. zm+1 = h−1
m+1 m žm+1
(m)
vm+1 = h−1
m+1 m wm

The Krylov space is spanned by zi vectors: z1 = B −1 r0 and

hm m hm−1 m
zm+1 = h−1 −1 −1 −1
m+1,m B (Azm −hm m vm −hm−1 m vm−1 ) = hm+1,m B Azm − zm − zm−1 .
hm+1,m hm+1,m

Note that, setting Vem = [e em ] and Zm = [z1 . . . zm ] one has Vem = LT Zm and
v1 . . . v

VemT L−1 A L−T Vem = Zm

T
A Zm ;

this latter matrix is therefore tri-diagonal.