Course Material

NetworkMaths Graduate Programme

Maynooth 2010

Numerical Linear Algebra

Volker Mehrmann,
TU Berlin,

August 3, 2010
Literature

The material of this course is based on the following textbooks:

G. Golub, C. Van Loan. Matrix computations. Baltimore, 1996.

R. B. Lehoucq, D. C. Sorensen and C. Yang. ARPACK Users Guide: Solution of

Large-Scale Eigenvalue Problems with Implicitly Restarted Arnoldi Methods, Philadel-
phia, 1989.

Y. Saad. Numerical methods for large eigenvalue problems. Manchester, 1992.

L. Trefethen, D. Bau. Numerical linear algebra. Philadelphia, 1997.

D. Watkins. Fundamentals of matrix computations. New York, 2002.

The following books are also useful to complement the material of these notes:

J. Demmel. Applied numerical linear algebra. Philadelphia, 1997.

N.J. Higham. Accuracy and stability of numerical algorithms. Philadelphia, 2002.

G.W. Stewart. Matrix algorithms. Philadelphia, 1998-2001, 2 Volumes.

G.W. Stewart, J.G. Sun. Matrix perturbation theory. Boston, 1990.

2
Chapter 0

Introduction

The main topics of Numerical Linear Algebra are the solution of different classes of eigenvalue
problems and linear systems.
For the eigenvalue problem we discuss different classes.

(a) The standard eigenvalue problem: For a real or complex matrix A Cn,n , determine
x Cn , C, such that
Ax = x.
The standard eigenvalue problem is a special case of the generalized eigenvalue problem:
For real or complex matrices A, E Cm,n , determine x Cn , C, such that

Ax = Ex,

In many applications the coefficient matrices have extra properties such as being real and
symmetric or complex Hermitian.
For linear systems:

Ax = b, x Cn , b Cm
with A Cm,n we again may extra properties for the coefficient matrices.
We will concentrate in this course on the numerical solution of standard and generalized
eigenvalue problems and the solution of linear systems. We will briefly review some of the
standard techniques for small scale problems and put an emphasis on large scale problems.
Applications: Eigenvalue problems arise in

the vibrational analysis of structures and vehicles (classical mechanics);

the analysis of the spectra and energy levels of atoms and molecules (quantum mechan-
ics);

model reduction techniques, where a large scale model is reduced to a small scale model
by leaving out weakly important parts;

many other applications.

Linear systems arise in almost any area of science and engineering such as

3
 (a) frequency response analysis for excited structures and vehicles;

(b) finite element methods or finite difference methods for ordinary and partial differential
equations;

(c) data mining, information retrieval;

(d) and many others.

We will distinguish between small and medium class problems where the full matrices fit into
main memory, these are of today sizes n = 102 105 and large sparse problems, where the
coefficient matrices are stored in sparse formats, and have sizes n = 106 and larger. We will
mainly discuss the case of complex matrices. Many results hold equally well in the real case,
but often the presentation becomes more clumsy. We will point out when the real case is
substantially different.
We will discuss the following algorithms.

A small A large
EVP QR-Algorithm, QZ-Algorithm Lanczos, Arnoldi, Jacobi-Davidson
LS CG, GMRES

4
Chapter 1

Matrix theory

1.1 Basics
1.1.1 Eigenvalues and Eigenvectors
Let A, E Cn,n , then v Cn \ {0} and C that satisfy

Av = Ev

are called eigenvector and eigenvalue of the pair (E, A). In the special case that E is the
n n identity matrix In (I) we have eigenvalues and eigenvectors of the standard eigenvalue
problem.
The sets

(A) := { C| eigenvalue of A}
(E, A) := { C| eigenvalue of (E, A)}

are called spectrum of A (the pair (E, A)), respectively.

1.1.2 Matrix norms

Let A Cm,n , then

||Ax||p
||A||p := sup
x6=0 ||x||p
is the matrix p-norm, p N {} and for invertible matrices A

p (A) := ||A||p ||A1 ||p

is called the p-norm condition number of A.

Special cases:
(a) p = 1 ; the column-sum norm:

m
X
||A||1 = max |aij |
j
i=1

5
 (b) p = ; the row-sum norm:

||A|| = ||AT ||1

(c) p = 2 ; the spectral norm

||A||2 = square root of the largest eigenvalue of AH A

with AH = AT .

Convention:

||A|| = ||A||2 , (A) = 2 (A)

Frobenius norm:
v
u n X
m
uX
||A||F = t |aij |2
j=1 i=1

1.1.3 Isometric and unitary matrices

Definition 1 Let U Cm,n , m n.

(a) U is called isometric if U H U = Ik ;

(b) U is called unitary if U is isometric and n = k.

Theorem 2 Let U Cnk , k n. Then the following are equivalent.

(a) U is isometric;

(b) the columns of U are orthonormal;

(c) hU x, U yi = hx, yi for all x, y Ck (h, i: standard real or complex scalar product);

(d) ||U x|| = ||x|| for all x Ck ;

For k = n, (a)-(d) are equivalent to

(e) U U H = In ;

(f ) U 1 = U H ;

(g) the rows of U are orthonormal.

In this case, furthermore,

||U || = 1 = ||U 1 || = (U ) .

6
1.1.4 Subspaces
Definition 3 A space U Cn is called subspace, if for all x, y U, C we have
x + y U, x U.
Theorem 4 Let U Cn be a subspace with basis (x1 , . . . , xm ) and X = [x1 , . . . , xm ], i.e.
Rank(X) = m.
(a) Then U = R (X) := {Xy| y Cm } (Range or column space of X).
(b) Let Y Cn,m with Rank(Y ) = m, then

R (X) = R (Y ) X = Y B, B Cm,m .
In particular then B is invertible and

XB 1 = Y.
(c) The Gram-Schmidt method for (x1 , . . . , xm ) delivers an orthonormal basis (q1 , . . . , qm )
of U with

Span {q1 , . . . , qj } = Span {x1 , . . . , xj }

for j = 1, . . . , m. This condition is equivalent to:
There exists an upper triangular matrix R Cm,m with X = QR where Q = [q1 , . . . , qm ].
(QR-decomposition)

1.1.5 Invariant subspaces

Definition 5 Let A Cn,n and U Cn . Then U is called A-invariant, if

x U Ax U for all x Cn .
Theorem 6 Let A Cn,n , X Cn,n and U = R (X). Then the following are equivalent:
(a) U is A-invariant;
(b) There exists B Ck,k , such that:

AX = XB.
Furthermore, in this case for C and v Ck :

Bv = v AXv = Xv,
i.e., every eigenvalue von B is also an eigenvalue von A.
Remark 7 If A, X, B satisfy AX = XB and if X has only one column x, then B is a scalar
and we obtain the eigenvalue equation
Ax = x,
i.e., X can be viewed as a generalization of the concept of eigenvector.

7
1.2 Matrix decompositions
1.2.1 Schur decomposition
Theorem 8 (Schur, 1909)
Let A Cn,n . Then there exists U Cn,n unitary such that

T := U H AU
is upper triangular.

Proof: By induction: n = 1 is trivial.

v
n 1 n: Let v Cn be an eigenvector of A to the eigenvalue C. Let q1 := ||v|| and
n
complete q1 to an orthonormal basis (q1 , . . . , qn ) of C . Then Q = [q1 , . . . , qm ] is unitary and
 
1 A12
Q AQ =
0 A22
HA U
By the inductive assumption there exists U22 unitary, such that T22 := U22 22 22 is upper
triangular. Setting  
1 0
U =Q ,
0 U22
then T = U H AU is upper triangular. 2

Remark 9 In the Schur decomposition U can be chosen such that the eigenvalues of A appear
in arbitrary order on the diagonal.

Definition 10 A matrix T Rn,n is called quasi-upper triangular matrix, if T is a block-

upper triangular matrix and the diagonal blocks have size maximal 2 2.

Theorem 11 (Murnaghan, Wintner, 1931)

Let A Rn,n . Then there exists a real orthogonal matrix Q, (i.e., QT Q = I) such that

T = QT AQ

is quasi-upper triangular.

Proof: The proof is similar to that of the Theorem of Schur: If A has a real eigenvector
then we can proceed the induction as in the complex case. Otherwise for a complex
eigenvector v = v1 + iv2 to the complex eigenvalue = 1 + i2 , with v1 , v2 Rn and
1 , 2 R, 2 6= 0 we have from

A (v1 + iv2 ) = (1 + i2 ) (v1 + iv2 )

that
  
Av1 = 1 v1 2 v2 1 2
A [v1 v2 ] = [v1 v2 ] .
Av2 = 1 v2 + 2 v1 2 1

8
Hence Span {v1 , v2 } is an A-invariant subspace. Let (q1 , q2 ) be an orthonormal basis of
Span {v1 , v2 } and let Q = [q1 , q2 , q3 , . . . , qn ] be orthogonal, then

1 2
A 12
QH AQ = 2 1 .
0 A22

The remaining steps of the induction are as in the complex case. 2

Definition 12 A Cn,n is called normal if AAH = AH A.

Example 13 Hermitian matrices, skew-Hermitian matrices and unitary matrices are nor-
mal.

Theorem 14 Let A Cn,n be normal and U Cn,n unitary such that

 
H A11 A12
U AU = ,
0 A22

then A12 = 0.

Proof: Exercise. 2

Corollary 15 If A Cn,n is normal, then there exists U Cn,n unitary, such that U H AU
is diagonal.

Proof: Exercise. 2

Theorem 16 (Generalized Schur form) Let A, E Cn,n be such that the pair (E, A) is
regular, i.e., det E A 6= 0 for all C. Then there exist U, V Cn,n unitary such that

S = U H EV, T := U H AV

are upper triangular.

Proof: Let (Ek ) be a sequence of nonsingular matrices that converges to E. For every let
1
QH
k AEk Qk = Tk

be a Schur decomposition of AEk1 and let ZkH (Ek1 Qk = Sk1 be a QR decomposition. Then
both QH H
k AZk = Rk Sk and Qk Ek Zk = Sk are upper triangular.
Using the Bolzano-Weierstra it follows that the bounded sequence (Qk , Zk ) has a
converging subsequence with a limit (Q, Z), where Q, Z are unitary. Then QH AZ = T and
QH EZ = S are upper triangular. 2

9
1.2.2 The singular value decomposition (SVD)
Theorem 17 (Singular value decomposition, SVD)
Let A Cm,n with Rank (A) = r. Then there exist unitary matrices U Cm,m and V Cn,n
such that
1
.. 0

A = U V H , = .
Cm,n .

r
0 0
Furthermore, 1 = ||A||2 and 1 , . . . , r are uniquely determined.

Proof: Exercise. 2

Definition 18 Let A, U = [u1 , . . . , um ] , V = [v1 , . . . , vn ] , be as in the SVD and k := 0 for

k = r + 1, . . . , min {m, n}. Then
(a) 1 , . . . , min{m,n} are called singular values of A.

(b) u1 , . . . , um are called left singular vectors of A.

(c) v1 , . . . , vn are called right singular vectors of A.

Remark 19 (a) From the SVD one obtains

AH A = V H U H U V H = V H V H = V 2 V H

and
AAH = U V H V H U H = U H U H = U 2 U H ,
i.e. 12 , . . . , r2 are the nonzero eigenvalues of AAH and AH A, respectively.

(b) Since AV = U one has Kernel (A) = Span {vr+1 , . . . , vn } and Image (A) = R (A) =
Span {u1 , . . . , ur }.

(c) The SVD allows optimal low-rank approximation of A, since

H
A = U V

1 0
0 ..
= U + . . . + . H
V

..
. 0
0 n
r
X
= j uj vjH .
j=1

Here uj vjH is a rank one matrix of size m n. For 0 r the matrix

X
A := i uj viH
i=1

10
 is the best rank approximation to A in the sense that

||A A || = inf ||A B|| = +1 ,

BCm,n
Rank (B)

where r+1 := 0.

(d) If A is real, then also U and V can be chosen real.

1.3 Perturbation theory

In the analysis of numerical methods, we will have to study the eigenvalues, eigenvectors
invariant subspaces under small perturbations ?

1.3.1 Canonical angles and vectors

Question: let U, V Cn be subspaces of dimension k. How near are U and V?
Strategy: Compute successively the angles between U and V beginning with the smallest.
Choose normalized vectors x U and y V, such that
!
| hx, yi | = max .
W.l.o.g. we can choose x and y such that their scalar product is real and nonnegative.
Otherwise we can take z C with |z| = 1, so that hzx, yi = z hx, yi is real and nonnegative.
Then | hx, yi | = | hzx, yi |.

(a) Choose x1 U and y1 V with ||x1 || = ||y1 || = 1 such that

hx1 , y1 i = max {Re hx, yi | x U, y V, ||x|| = ||y|| = 1}

Then hx1 , y1 i is real, 1 = arccos hx1 , y2 i is called first canonical angle and x1 , y1 are
called first canonical vectors.

(b) Suppose that we have determined j 1 canonical angles and vectors, i.e.,

x1 , . . . , xj1 U, y1 , . . . , yj1 V

are determined with (x1 , . . . , xj1 ) and (y1 , . . . , yj1 ) orthonormal.

Choose xj U and yj V with xj x1 , . . . , xj1 and yj , y1 , . . . , yj1 and ||xj || =
||yi || = 1, so that
hxj , yj i
has maximal real part. Then hxj , yj i is real,

j := arccos hxj , yj i

is the j-th canonical angle, and xj , yj are j-th canonical vectors. Proceeding inductively
we obtain k canonical angles 0 1 . . . k 2 and orthonormal bases (x1 , . . . , xk )
and (y1 , . . . , yk ) of U, V, respectively.

11
Lemma 20 For i, j = 1, . . . , k and i 6= j the canonical vectors satisfy hxi , yj i = 0.

Proof: Exercise. 2

Corollary 21 Let X = [x1 , . . . , xk ] and Y = [y1 , . . . , yk ]. Then

cos 1 0
X H Y = (hxi , yj i) =
..
.
0 cos k
with cos 1 . . . cos k 0 and this is SVD.

Practical computation of canonical angles and vectors

(a) Determine orthonormal bases of U and V, i.e., isometric matrices P, Q Cn,k with

R (P ) = U, R (Q) = V.

(b) Compute the SVD of P H Q

P H Q = U V H
with the diagonal matrix
H H
=U QV
| {zP } |{z}
XH Y

(c) Set U = [u1 , . . . , uk ] and V = [v1 , . . . , vk ]. Then

(a) j = arccos j , j = 1, . . . , k are the canonical angles and

(b) P uj , Qvj , j = 1, . . . , k are the canonical vectors.

1.3.2 Distance between subspaces

Definition 22 Let U, V Cn be subspaces of dimension k.

(a) For x U we call

d (x, V) := min ||x y||
yV

the distance from x to V and

(b)
d (U, V) := max d (x, V)
xU
||x||=1

the distance of U and V.

Theorem 23 Let U, V Cn be subspaces of dimension k with canonical angles 1 . . . k ,

then
d (U, V) = sin k .

12
Proof: See Stewart/Sun. Matrix perturbation theory. Boston, 1990. 2

Lemma 24 Let    
Im Im
U =R and U = R ,
0 X
with X C(nm)m be m-dimensional subspaces of Cn and let 1 , . . . , m be the canonical
angles between U and U. Then
tan 1 , . . . , tan n
are the singular values of X, in particular then ||X|| = tan m .

Proof: See Stewart/Sun. Matrix perturbation theory. Boston, 1990. 2

13
Chapter 2

Eigenvalue problems with full

matrices

Situation: A Cn,n , where n is small enough so that the matrix A can be fully stored and
that we can manipulate the whole matrix by similarity transformations.

2.1 The power method

Idea: Take an arbitrary q Cn \{0} and form the sequence q, Aq, A2 q, . . .. What will happen?
Assumption: A is diagonalizable. Let 1 , . . . , n with |1 | . . . |n | be the eigenvalues
of A and let (v1 , . . . , vn ) be a basis of eigenvectors. Then there exist c1 , . . . , cn with

q = c1 v1 + . . . + cn vn .

Further assumption: c1 6= 0 (this happens with probability 1 if q is random). Then,

Aq = c1 1 v1 + . . . + cn n vn ,
Ak q = c1 k1 v1 + . . . + cn kn vn .

For |1 | > 1 the powers |k1 | will grow, so we scale as

 k  k
1 k 2 n
A q = c1 v1 + c2 v2 + . . . + cn vn .
k1 1 1

Third assumption: |1 | > |2 | . . . |n |. Then,

 k  k
1 k
A q c1 v1 |c2 |  2  ||v2 || + . . . + |cn |  n  ||vn ||
   
k  1   1 
1
   k
2 k
|c2 |||v2 || + . . . + |cn |||vn ||   0,
1

and hence limk 1

k1
Ak q = c1 v1 and the convergence is linear with convergence rate r | 21 |.

14
Definition 25 A sequence (xk ) converges linearly to x, if there exists r with 0 < r < 1 such
that
||xk+1 x||
lim = r.
k ||xk x||

Then r is called the convergence rate of the sequence.

We say that the convergence (xk ) x is of order m 2 if

||xk+1 x||
lim = c 6= 0.
k ||xk x||m

If m = 2 then we speak of quadratic convergence and if m = 3 of cubic convergence.

In practice we do not know 1/k1 , thus we normalize differently and divide by the largest (in
modulus) component of Ak q.
Algorithm: (Power method)
Computes the dominant eigenvalue 1 and the associated eigenvector v1 .

(a) Choose q0 Cn \ {0}

(b) Iterate, for k = 1, 2, . . . to convergence

1
qk := Aqk1 ,
k

where k is the largest (in modulus) component of Aqk1 .

The power method can also be used for large scale problem where only matrix vector mul-
tiplication is available. It computes only one eigenvalue and eigenvector and is for example
used in the Google page rank method. By the presented analysis we have proved the following
theorem.

Theorem 26 Suppose that A Cn,n has the eigenvalues 1 , . . . , n with |1 | > |2 | . . .

|n |. If q0 Cn \{0} has a component in the invariant subspace associated to 1 , (i.e., c1 6= 0),
then the sequence (qk ) defined in the power method converges to an eigenvector associated with
1 . The convergence is linear with rate r | 12 |. Furthermore, the sequence (k ) converges
to the eigenvalue 1 .

Remark 27 (a) This theorem also holds for non-diagonalizable matrices.

(b) Forming the full products Aqk costs 2n2 flops and the scaling O(n) flops. Hence m
iterations will cost 2n2 m flops.

(c) If (as is very common) | 21 | 1 then the convergence is very slow.

15
2.2 Shift-and-Invert and Rayleigh-Quotient-Iteration
Observations: Let A Cn,n and (, v) C Cn with Av = v. Then
(a) A1 v = 1 v for A invertible, and
(b) (A %I)v = ( %)v for all % C.
If 1 , . . . , n are again the eigenvalues of A with |1 | . . . |n |, then we can perform the
following iterations.

Inverse Iteration This is the power method applied to A1 . If |n | < |n1 |, then the
inverse iteration converges to an eigenvector to n with convergence rate | n1
n
| (which is
small if |n |  |n1 |).

Shift and Invert Power Method This is the power method applied to (A %I)1 . Let
j , k be the eigenvalues that are closest to %, and suppose that |j %| < |k %|. Then the
power method for (A %I)1 converges to an eigenvector associated with j with rate
 
 j % 
 k %  .
 

This is small if |j %|  |k %| and j % would be optimal, but where do we get such

good shifts? To answer this question we need some results on residuals and backward errors.
Definition 28 Let A Cn,n and (, w) C Cn . Then Aw w is called the eigenvalue
residual of (, w) with respect to A.
Theorem 29 Let C, > 0, A Cnn , and w Cn with ||w|| = 1. If ||Aw w|| = ,
then there exists a matrix ( the backward error matrix) E Cn,n with ||E|| such that
(A + E)w = w.
Proof: Let r := Aw w and E = rwH . Then
H
(A + E)w = Aw r w
| {zw} = w
=1

and ||E|| = ||rwH || ||r||||wH || = ||r|| = .

2

Small residual implies small backward error in eigenvalue/eigenvector pair.

The idea to determine a good eigenvalue approximation (shift) from a given eigenvector
approximation is to minimize the residual ||Aw w||. Consider the over-determined linear
system
w = Aw
with the n 1-Matrix w, the unknown vector and the right hand side Aw. We can use the
normal equations to solve ||Aw w|| = min!, i.e., we use
wH Aw
wH w = wH Aw respect. = .
wH w

16
Definition 30 Let A Cn,n and w Cn \ {0}. Then

wH Aw
r(w) :=
wH w
is called the Rayleigh-quotient of w with respect to A.

The following theorem gives an estimate for the distance of the Rayleigh-quotient from an
eigenvalue.

Theorem 31 Let A Cn,n and (, v) C Cn be an eigenvalue/eigenvector pair of A with

kvk = 1. Then for w Cn with kwk = 1 the following estimate holds

| r(w)| 2kAk kv wk.

This gives an the idea for an iteration to iterate computing an approximate eigenvector and
from this a Rayleigh-quotient, i.e., the following algorithm:
Algorithm: Rayleigh-Quotient-Iteration (RQI)
This algorithm computes an eigenvalue/eigenvector pair (, v) C Cn of the matrix A
Cn,n .

(a) Start: Choose q0 Cn with ||q0 || = 1 and set 0 := q0H Aq0 .

(b) Iterate for k = 1, 2, . . . to convergence

(a) Solve the linear system (A k1 I)x = qk1 for x.

x
(b) qk :=
||x||
(c) k := qkH Aqk

Remark 32 (a) It is difficult to analyze the convergence of this algorithm but one observes
practically that it almost always converges. The convergence rate is typically quadratic.
For Hermitian matrices A = AH there is more analysis and one can even show cubic
convergence.

(b) The Rayleigh-quotient iteration can also be applied to very large matrices provided that
a linear system solver is available (we get back to this later).

(c) Costs: O(n3 ) flops per step if the linear system is solved with full Gaussian elimination.
The costs are O(n2 ) for Hessenberg matrices (see Chapter 2.4.2) and they can be even
smaller for banded or other sparse matrices.

(d) Ak1 I is almost singular, i.e., if k1 is close to an eigenvalue, then linear systems
with A k1 I are generally ill-conditioned. But if we use a backward stable method
then we get a small backward error, i.e.,

(A + A k1 I)x = qk1

with ||A|| small. Thus we can expect good results only if the eigenvalue/eigenvector
computation is well conditioned.

17
Conditioning of eigenvalues: if is a simple eigenvalue of A, i.e., the algebraic multiplicity
is 1 and if v, w are normalized right and left eigenvectors, i.e.,

Av = v, wH A = wH , kvk = 1 = kwk,

then for small perturbations A we have in first approximation that A+A has an eigenvalue
+ with
1
|| H kAk.
|w v|
We then have that 1/|wH v| is a condition number for simple eigenvalues. For normal matrices
we have v = w and thus |wH v| = 1. Normal matrices thus have well-conditioned eigenvalues.

2.3 Subspace iteration

To compute several eigenvalues and the associated invariant subspace, we can generalize the
power method to the subspace iteration. Consider A Cn,n with eigenvalues 1 , . . . , n ,
where |1 | . . . |n |.
Idea: Instead of q0 Cn , consider a set of linearly independent vectors {w1 , . . . , wm } Cn .
Set
W0 := [w1 , . . . , wm ] Cnm ,
and form the sequence W0 , AW0 , A2 W0 , . . . via
h i
Wk := Ak W0 = Ak w1 , . . . , Ak wm , k 1.

In general, we expect R(Wk ) to converge to the invariant subspace U associated with the m
eigenvalues 1 , . . . , m . This iteration is called subspace iteration.

Theorem 33 Let A Cn,n with eigenvalues 1 , . . . , n satisfy

|1 | . . . |m | > |m+1 | . . . |n |.

Let U, V be the invariant subspaces associated with 1 , . . . , m , and m+1 , . . . , n respectively.

Furthermore, let W Cnm with Rank (W ) = m and R(W )  V =  {0}. Then for the
 m+1 
iteration W0 := W , Wk+1 = AWk for k 0 and for every % with  m  < % < 1, there exists
a constant c such that
d(R(Wk ), U) c %k , k 1.

Proof: We prove the theorem for the case that A is diagonalizable. We perform a similarity
transformation  
1 A1 0
Anew = S Aold S = ,
0 A2
with A1 = diag (1 , . . . , m ) and A2 = diag (m+1 , . . . , n ). Then A1 is nonsingular, since
|1 | . . . |m | > 0. Set
 
1 Im
Unew = S Uold = R
0

18
and  
1 0
Vnew = S Vold = R .
Inm
Furthermore, let  
Z1
Wnew = S 1 Wold =
Z2
for some Z1 Cm,m and Z2 Cnm,m . Then

(a) d(R(Wnew ), Unew ) (S)d(R(Wold ), Uold ) (Exercise)

(Here (S) = kSkkS 1 k is the condition number of S with respect to inversion.)

(b) R(Wnew ) Vnew = {0} Z1 is nonsingular (Exercise)

In the following we drop the index new. Then

     
Z1 Im I
W = = Z1 = Z1
Z2 Z2 Z11 X0

with X0 = Z2 Z11 , and hence

 
Im
R(W ) = R
X0
as well as  
k k I
R(Wk ) = R(A W ) = R(A ).
X0
Then it follows that

Im
Ak1 0 Ak1
      
I I k k
Ak = = = A2 X0 A1 Ak1 ,
X0 0 Ak2 X0 Ak2 X0 | {z }
=:Xk

and thus,  
Im
R(Wk ) = R .
Xk
It remains to show that
d(R(Wk ), U) 0.
(k)
Let m be the largest canonical angle between R(Wk ) and U. Then
k
d(R(Wk ), U) = sin (k) (k) k
m tan m = ||Xk || ||A2 ||||X0 ||||A1 ||
= |km+1 |||X0 |||k
m |,

which implies that

 m+1 k
 
d(R(Wk ), U) c   .
m 
Undoing the similarity transformation we obtain the desired result. For the diagonalizable
case we do not need the bound %. This will be only needed in the non-diagonalizable case. 2

19
Remark 34 For W0 = [w1 , . . . , wm ] we have
h i
Ak W 0 = Ak w1 , . . . , A k wm ,

(j)
i.e., we perform the iteration not only for W0 but simultaneously also for all W0 = [w1 , . . . , wj ],
since h i
k (j) k k
A W 0 = A w1 , . . . , A wj .

Under appropriate assumptions, we then have convergence of

n o
Span Ak w1 , . . . , Ak wj

to the invariant subspace associated with 1 , . . . , j for all j = 1, . . . , m. For this reason one
often speaks of simultaneous subspace iteration.

Problems with subspace iteration in finite precision arithmetic:

Theory: Ak W0 = Ak w1 , . . . , Ak wm in general has Rank m (for generic starting values).

 

Practice: Unfortunately, in finite precision arithmetic rounding errors lead to linear depen-
dence in R(Wk ) already after few iterations.

The basic idea to cope with this problem is to orthonormalize the columns in every step.

Step 1: Factor W0 = [w1 , . . . , wm ] = Q0 R0 with Q0 Cn,m isometric and R0 Cm,m upper

triangular. Then R(W0 ) = R(Q0 ) and furthermore,
n o
(0) (0)
Span {w1 , . . . , wj } = Span q1 , . . . , qj , j = 1, . . . , m,
h i
(0) (0)
where Q0 = q1 , . . . , qm . (This follows form the triangular form of R0 .)

Situation after step k 1: R(Wk1 ) = R(Qk1 ) with

h i
(k1) (k1)
Qk1 = q1 , . . . , qm Cn,m

isometric and
n o n o
k1 k1 (k1) (k1)
Span A w1 , . . . , A wj = Span q1 , . . . , qj , j = 1, . . . , m.

Step k: Let AQk1 = Qk Rk be a QR decomposition with Qk Cn,m isometric and Rk

Cm,m upper triangular. Then

R(Wk ) = R(AWk1 ) = R(AQk1 ) = R(Qk ),

and moreover
n o n o
(k) (k)
Span Ak w1 , . . . , Ak wj = Span q1 , . . . , qj , j = 1, . . . , m.

Algorithm: Unitary Subspace Iteration

20
 (a) Start: Choose Q0 Cn,m isometric.

(b) Iterate. For k = 1, 2, . . . to convergence:

(a) Compute Zk = AQk1

(b) Compute QR-decomposition Zk = Qk Rk .

Remark 35 Theoretically the convergence behavior of the unitary subspace iteration is as for
the subspace iteration but the described problems in finite precision arithmetic do not arise.

2.4 The Francis QR Algorithm

2.4.1 Simple QR Algorithm Without Shifts
Let A Cn,n have the eigenvalues 1 , . . . , n where |1 | . . . |n |.
Idea: hUse the n-dimensional
i unitary subspace iteration, i.e., choose m = n and Q0 = In . If
(k) (k)
Qk = q1 , . . . , qn , then for every 1 m n
n o
(k) (k)
Span q1 , . . . , qm
 
 m+1 
converges to the invariant subspace associated with 1 , . . . , m with a rate  m  provided
that |m+1 | < |m | and one does not run into an exceptional situation.
(k) (k)
To observe the convergence, we form Ak = Q1 k AQk . If Span{q1 , . . . , qm } converges to an
invariant subspace, then we expect that in the matrix

m nm
 
m A11 A12
Ak =
nm A21 A22

the block A21 converges to 0 for k . Since this happens for all m simultaneously, it
follows that Ak converges to block-upper triangular matrix.
Another question is whether we can directly move from Ak1 to Ak ? To see this, observe that

Ak1 = Q1
k1 AQk1
Ak = Q1
k AQk

and hence
Ak = Q1 1 1
k Qk1 Ak1 Qk1 Qk = Uk Ak1 Uk .
| {z }
=:Uk

Thus we can reformulate the k-th step of the unitary subspace iteration

AQk1 = Qk Rk

as
Ak1 = Q1 1
k1 AQk1 = Qk1 Qk Rk = Uk Rk .

21
This is a QR decomposition of Ak1 and we have

Ak = Uk1 Ak1 Uk = Uk1 Uk Rk Uk = Rk Uk .

Algorithm: (QR Algorithm)(Francis and Kublanovskaya 1961)

For a given matrix A Cn,n this algorithm constructs a sequence (Ak ) of similar matrices
that converges to block upper-triangular form.

(a) Start with A0 = A

(b) Iterate for k = 1, 2, . . . to convergence.

(a) Compute a QR-decomposition of Ak1 : Ak1 = Uk Rk

(b) Compute Ak via Ak = Rk Uk .

Theorem 36 (Convergence of the QR algorithm)

Let A Cn,n have eigenvalues 1 , . . . , n , where |1 | . . . |m | > |m+1 | . . . |n |. Let
V Cn be the invariant subspace associated with m+1 , . . . , n , and let (Ak ) be the sequence
generated by the QR Algorithm. If

Span {e1 , . . . , em } V = {0}

and
m nm
(k) (k)
" #
m A11 A12
Ak = (k) (k) ,
nm A21 A22
 
then for every % with  m+1  < % < 1 there exists a constant c such that
 
m

(k)
||A21 || c%k .

Proof: (Sketch) Let U be the invariant subspace associated with 1 , . . . , m and

n o
(k) (k)
Uk = Span q1 , . . . , qm ,

where h i
(k)
Qk = q1 , . . . , qn(k)

is the unitary matrix with Q1

k AQk = Ak from the unitary subspace iteration. One first
shows that
(k)
||A21 || 2 2||A||d(U, Uk )
Then using the convergence results for the subspace iteration there exists a constant c > 0
with
d(U, Uk ) c%k .

Then choose c := 2 2||A||c. 2

In the special case that A is Hermitian, the sequence Ak converges to a diagonal matrix.

22
Remark 37 In the presented form the algorithm has two major disadvantages:

(a) It is expensive, since it costs O(n3 ) flops per iteration step.

(b) The convergence is slow (only linear).

A way to address the two problems is the Hessenberg reduction and the use of shifts.

2.4.2 Hessenberg reduction

Definition 38 A matrix A = [aij ] is called Hessenberg matrix or in Hessenberg form, if

aij = 0 for i > j + 1. A Hessenberg matrix A is called unreduced if ai+1,i 6= 0 for all
i = 1, . . . , n 1.

Householder transformations
The QR decomposition and other unitary transformations can be realized via Householder
transformations
2
P =I vv H
vH v

for v Cn \ {0}. Householder Transformations are Hermitian and unitary. (Exercise) Mul-
tiplication with Householder transformations is geometrically a reflection of a vector x Cn
at the hyperplane Span {v} . (Exercise.)
A typical task: Reflect x Cn \ {0} to a multiple of the first unit vector, i.e., determine v
and from this P such that

P x = ||x||e1 .

To determine such a v we make an ansatz v = x + e1 and obtain

v = x ||x||e1 and P x = ||x||e1 . (Exercise)

For numerical reasons we take


x + ||x||e1 , x1 0,
v=
x ||x||e1 , x1 < 0.

Advantage: The multiplication with Householder transformations is cheap. For B Cn,m

the computation of P B only needs 4mn flops (instead of 2n2 m flops for classical ma-
trix/matrix multiplication).
Givens rotations

23
Another tool for unitary operations are Givens rotations

1
..
.




1


c s


1

Gi,j (c, s) =
.. ,

.

1


s c


1

..
.
1

where |c|2 +|s|2 = 1 and where the matrix differs from an identity only in positions (i, i), (i, j), (j, i), (j, j).
Multiplication of a matrix with a Givens rotation allows to zero an element in any position.
E.g. choose
 
c s
G1,2 (c, s) =
s c

with |c|2 + |s|2 = 1 such sa11 + ca21 = 0, then we have

   
a11 a12
G1,2 (c, s) = .
a21 a22 0

(a) For a Hessenberg matrix H Cn,n the QR decomposition can be performed in O(n2 )
flops using Givens rotations.

(b) Hessenberg matrices are invariant under QR iterations.

Theorem 39 (Implicit Q Theorem) Let A Cn,n and let Q = [q1 , . . . , qn ] , U = [u1 , . . . , un ]

be unitary matrices such that

H = Q1 AQ = (hij ) and G = U 1 AU = [gij ]

are Hessenberg matrices. If q1 = u1 and H is unreduced, then

qi = ci ui

for ci C with |ci | = 1 and |hi,i1 | = |gi,i1 | for i = 2, . . . , n, i.e., Q is determined already
essentially uniquely by q1 .

Proof: Exercise. 2

24
2.4.3 The Francis QR Algorithm with Shifts
Deflation: Let H Cn,n be in Hessenberg form. If H is not unreduced, i.e., if hm+1,m = 0
for some m, then
m nm
 
m H11 H12
H=
nm 0 H22
i.e., we can split our problem into two subproblems H11 , H22 .
Algorithm (QR Algorithm with Hessenberg reduction and shifts)
Given: A Cn,n :
(a) Compute U0 unitary such that
H0 := U0H AU0
is in Hessenberg form. We may assume that H0 is unreduced, otherwise we can deflate
right away.
(b) Iterate for k = 1, 2, . . . until deflation happens, i.e.,
(k)
hm+1,m = O(eps)(|hm,m | + |hm+1,m+1 |)
for some m and the machine precision eps.
(i) Choose shift k C.
(ii) Compute a QR decomposition Hk1 k I = Qk Rk of Hk1 k I.
(iii) Form Hk = Rk Qk + k I.

Remark 40 (a) Steps (ii) and (iii) of this algorithm correspond to a QR iteration step for
H0 k I.
 
(k) k 
(b) The sub-diagonal entry hm+1,m in Hk converges with rate  m+1  to 0.

m k

(k) (k)
(c) If hm+1,m = 0 or hm+1,m = O(eps), then we have deflation and we can continue with
smaller problems.
(d) If k is an eigenvalue then deflation happens immediately after one step.

Shift strategies:
(a) Rayleigh-quotient shift: For the special case that A is Hermitian, the sequence Ak
(k)
converges to a diagonal matrix. Then qn is a good approximation to an eigenvector
and a good approximation to the eigenvalue is the Rayleigh-quotient
r(qn(k) ) = (qn(k) )H Aqn(k)
(k)
which is just the n-th diagonal entry an,n of QH
k AQk .
(k)
Heuristic: We expect in general an,n to be a good approximation to an eigenvalue and
therefore may choose
(k)
k = an,n
(k)
With this choice hn,n1 typically converges quadratically to 0.

25
 (b) Wilkinson-shift: Problems with the Rayleigh-quotient shift arise when the matrix is
real and has nonreal eigenvalues, e.g., for
 
0 1
A= .
1 0

A QR iteration for A0 = A yields Q0 = A0 , R0 = I and hence,

R0 Q0 = IA0 = A0 ,

i.e., the algorithm stagnates. To avoid such situations, for A Cn,n in the k-th step,
one considers the submatrix B in
n2 2
 
n2
Ak =
2 B
(k)
and chooses the k as shift that is nearest to ann .

(c) Another strategy, the double-shift will be discussed below.

(d) On the average 2-3 iterations are needed until a 1 1 or 2 2 block deflates.

2.4.4 Implicit Shifts and Bulge-Chasing

Let H Cn,n be a Hessenberg matrix and 1 , . . . , l C. Carry out l steps of the QR
Algorithm with shifts 1 , . . . , l .

H 1 I = Q1 R1
H1 = R1 Q1 + 1 I
..
.
Hl1 l I = Ql Rl
Hl = Rl Ql + l I.

Then
Hl = QH H H H
l Ql Rl Ql + l Ql Ql = Ql (Ql Rl + l I)Ql = Ql Hl1 Ql

and thus per induction

Hl = QH H
Q1 . . . Ql = QH HQ.
l . . . Q1 H | {z }
=:Q

This opens the question whether we can compute Q directly without carrying out l QR-
iterations.

Lemma 41 M := (H l I) . . . (H 1 I) = Q1 . . . Ql Rl . . . R1 = QR.
| {z }
=:R

Proof: By induction we show that

(H j I) . . . (H 1 I) = Q1 . . . Qj Rj . . . R1 , j = 1, . . . , l.

26
j = 1: This is just the first step of the QR algorithm.

j 1 j:

Q1 . . . Qj Rj . . . R1
= Q1 . . . Qj1 (Hj1 j I)Rj1 . . . R1
 
= Q1 . . . Qj1 QHj1 . . . QH
1 HQ 1 . . . Qj1 j I Rj1 . . . R1
= (H j I)Q1 . . . Qj1 Rj1 . . . R1
I.A.
= (H j I)(H j1 I) . . . (H 1 I).

This leads to the idea to compute M and then the Householder QR decomposition of M , i.e.,
M = QR, and to set
H = QH RQ = Hl .
This means that one just needs one QR decomposition instead of l QR decompositions in
each QR step. On the other hand we would have to compute M , i.e., l 1 matrix-matrix
multiplications. But this can be avoided by computing H directly from H using the implicit
Q Theorem.

Implicit shift-strategy:

(a) Compute
M e1 = (H l I) . . . (H 1 I)e1 ,
the first column of M . Then the first l + 1 entries are in general nonzero. If l is not too
large, then this costs only O(1) flops.

(b) Determine a Householder matrix P0 such that P0 (M e1 ) is a multiple of e1 . Transform

H with P0 as

l+1 nl1
 
l+1 0
P0 =
nl1 0 I
l+2 nl2
 
l+2
P0 HP0 =
nl2 0 H

P0 changes only rows and columns 1, . . . , l + 1 of H. This gives a Hessenberg matrix

with a bulge.

(c) Determine Householder matrices P1 , . . . , Pn2 to restore the Hessenberg form. This is
called bulge chasing, since we chase the bulge the down the diagonal. This yields

H := Pn2 . . . P1 P0 HP0 . . . Pn2

that is again in Hessenberg form and Pk e1 = e1 for k = 1, . . . , n 2.

27
 (d) P0 has the same first column as Q. As in the first step for P0 we have Pk e1 = e1 , then
also
P0 P1 . . . Pn2
has the same first column as P0 and Q, respectively. With the implicit Q Theorem then
also Q and P0 . . . , Pn2 and therefore also H and QH HQ are essentially equal, thus we
have computed H and Hl directly from H.
Algorithm (Francis QR algorithm with implicit double-shift strategy): (Francis
1961)
Given A Cn,n :

(a) Determine U0 unitary so that H0 := U0H AU0 is in Hessenberg form.

(b) Iterate for k = 1, 2, . . . to convergence (deflation):

(a) Compute the eigenvalues 1 , 2 of the lower right 2 2 submatrix of Hk1 .

(b) Compute (with the implicit shift strategy) for l = 2 the matrix Qk that one obtains
with 2 steps of the QR Algorithm with shifts 1 , 2 .
(c)
Hk := QH
k Hk1 Qk

Remark 42 (a) The empirical costs for the computation of all eigenvalues of A are approx.
10n3 flops, If also the transformation matrix Q is needed then this leads to approximately
25n3 flops.

(b) The convergence analysis is difficult, no global convergence proof is known.

(c) The method works also for real problems in real arithmetic, since the double shift can
be chosen as complex conjugate pairs.

2.5 The QZ algorithm

For the solution of the generalized eigenvalue problem Ex = Ax with regular (E, A), E, A
Cn,n we can extend the ideas of the QR algorithm. The basic idea is to apply the QR algorithm
implicitly to E 1 A, without really computing the inverse and the product.
The first step is a transformation to Hessenberg-triangular form, i.e., one computes unitary
matrices U0 , V0 such that
U0H EV0 = S0 , U0H AV0 = T0 ,
with S0 upper triangular and T0 upper Hessenberg. If S0 was invertible, then S01 T0 would
be of upper Hessenberg form.

Algorithm. Hessenberg-triangular reduction

For a given regular pair (E, A) with E, A Cn,n the algorithm computes unitary matrices
U0 , V0 such that
U0H EV0 = S0 , U0H AV0 = T0 ,
with S0 upper triangular and T0 upper Hessenberg.

28
Use the QR decomposition to compute U such that U H E is upper triangular and set A =
U H A = [aij ], E = U H E = [eij ].

For j = 1, . . . , n 2

for i = n, n 1, . . . , j + 2
Determine a 2 2 Householder or Givens matrix P such that
   
ai1,j
P =
aij 0

and set A := P A = [aij ], E = P E = [eij ], where P = diag (Ii2 , P , Ini ).

Determine a 2 2 Householder or Givens matrix Q such that
   
ei,i1 aii Q = 0

and set A := AQ, E = EQ, where Q = diag (Ii2 , Q, Ini ).

end

end

This algorithm costs about 5n3 flops plus extra 23/6n3 if U , V are desired.
If A is not unreduced then we can again deflate the problem into subproblems, i.e., as
   
E11 E12 A11 A12
E A =
0 E22 0 A22

and continue with the subproblems.

Every 0 element in the diagonal of E corresponds to an infinite eigenvalue, i.e. a zero eigen-
value of A E. If an element on the diagonal of E is zero then we can introduce a 0 in the
position (n, n 1) of A and move the 0 to the bottom of the diagonal of E (Exercise). This
can be repeated until all zero elements are in the bottom part of the diagonal of E and the
corresponding part of A is triangular. Then we have deflated all the infinite eigenvalues and
have obtained
   
H E11 E12 H A11 A12
U0 EV0 = S0 = , U0 AV0 = T0 = ,
0 E22 0 A22

with E22 strictly upper triangular and A22 is nonsingular upper triangular by the regularity
of (E, A).
In the top pair (E11 , A11 ) then E11 is invertible and in principle we can imply the implicit
1
QR algorithm to E11 A11 . This leads to the QZ algorithm of Moler and Stewart from 1973
(Exercise).

29
Chapter 3

Eigenvalue problems with large

sparse matrices

Situation: Given a matrix A Cn,n with n very large (e.g., n 106 , 107 , . . .) and A sparse,
i.e., A has only very few nonzero elements. A sparse matrix A Cm,n is described by 6
parameters.

(a) m: no. of rows;

(b) n: no. of columns;

(c) nnz: no. of nonzero elements;

(d) a: list of nonzero elements;

(e) irow: list of row indices;

(f) jcol: list of column indices.

The programming environment MATLAB for example has the data structure sparse.

(a) For x Cn we can compute the product Ax, this often works with O(n) flops instead
of O(n2 ).

(b) We cannot apply standard similarity transformations, since the transformed matrix in
general is not sparse any more.

Often even A is not given but just a subroutine, that computes the product Ax for a given
x Cn , i.e.,
x black box Ax.

This is then the only possibility to obtain information about the matrix A.

Example 43 (a) Discretized Laplace operator on a uniform grid.

30
 For a given function u(x, y), we get on a two-dimensional grid in each grid point (i, j)
approximations uij = u(xi , yj ) for i = 1, . . . , m and j = 1, . . . , n. Here the approxima-
tion to v = u in the grid point (i, j) obtained via the 5-point difference star satisfies

-1
|
-1 4 -1 vij = 4uij ui,j+1 ui,j1 ui+1,j ui1,j ,
|
-1

together with further boundary conditions. This we can write as matrix equation.

v = [v11 , . . . , v1n , v21 , . . . , v2n , . . . , vm1 , . . . , vmn ] = Au

u = [u11 , . . . , u1n , u21 , . . . , u2n , . . . , um1 , . . . , umn ]

Instead of explicitly storing A we can write a subroutine that computes v = Au from u

via

for i=1,...,m
for j=1,...,n
v[i,j]=4*u[i,j]-u[i,j+1]-u[i,j-1]-u[i+1,j]-u[i-1,j]
end
end

(b) Toeplitz matrices:

a0 a1 . . . an

a1
..
.
..
. ...

A=
.. .. .. ..


. . . .
an ... a1 a0
A is uniquely determined by the row vector

[an , . . . , a1 , a0 , a1 , . . . , an ]

and is suffices to store this vector and to write a subroutine to compute the product Ax.

Then the question arises, how to compute some eigenvalues and eigenvectors of A? In general
we are only interested in a few eigenvalues, e.g., the ones that are largest or smallest in
modulus. (We would certainly not be able to store all eigenvectors because that would need
O(n3 ) storage.)
1. Idea: Unitary subspace iteration
h i
(0) (0) (0) (0)
(a) Start with m  n orthonormal vectors q1 , . . . , qm and set Q0 = q1 , . . . , qm .

(b) Compute AQk1 = Qk , hence m-matrix vector products. For sparse matrices this costs
O(mn) flops instead of O(mn2 ) for general matrices.

31
 (c) Compute a QR decomposition:
Qk = Qk Rk ,
where Qk Cn,m is isometric and Rk Cm,m is upper triangular.

In general this method does not work very well.

2. Idea: Petrov-Galerkin projection methods: Make use of the fact that for (, v) C Cn \
{0} we have

(, v) is an eigenvalue/eigenvector pair Av v = 0, Av v Cn .

(a) Thus we may construct K Cn , the search space.

(b) Choose a second subspace L Cn , the test space. Then determine (, v) C K with

Av v L.

Often one uses L = K, this is called Galerkin projection. Hope: Some of the (, v) are
good approximations to eigenvalue/eigenvectors pairs of A.

3.1 Krylov Spaces

Question: How do we find a subspace K that contains good approximation to eigenvectors
of A?

As motivation we consider the special case that A Rn,n is symmetric with eigenvalues
1 . . . n . Then

xT Ax
1 = max r (x) and n = min r (x) , where r(x) = (Exercise)
x6=0 x6=0 xT x

Let R(Qk ) be given via Qk = [q1 , . . . , qk ] Rn,k and let

y T QTk AQk y
Mk := max r (x) = max ,
xR(Qk )\{0} y6=0 y T QTk Qk y

and analogously
mk := min r (x) .
xR(Qk )\{0}

Then clearly 1 Mk mk n and, in particular, Mk is an approximation to 1 and mk

to n . It is then our goal to construct the vectors q1 , q2 , . . . , qk , . . . such that Mk and mk
quickly approximate 1 and n , i.e., we want

Mk 1 and mk n for small k.

Let q1 , . . . , qk be already constructed and choose uk , wk Span {q1 , . . . , qk } such that

Mk = r (uk ) and mk = r (wk ) .

32
It is clear that r (x) grows fastest in the direction of the gradient

2  
r(x) = Ax r(x)x (Exercise).
xT x

Then Mk+1 > Mk if r (uk ) 6= 0 and

r (uk ) Span {q1 , . . . , qk , qk+1 } . (3.1)

Analogously mk+1 < mk if r (wk ) 6= 0 and

r (wk ) Span {q1 , . . . , qk , qk+1 } . (3.2)

Since r (x) Span {x, Ax}, the conditions 3.1 and 3.2 are satisfied if

Span {q1 , q2 } = Span {q1 , Aq1 } ,

Span {q1 , q2 , q3 } = Span q1 , Aq1 , A2 q1 ,


..
. n o
Span {q1 , . . . , qk+1 } = Span q1 , Aq1 , A2 q1 , . . . , Ak q1 .

Definition 44 Let A Cn,n , x Cn and l N.

(a) Kl (A, x) := x, Ax, A2 x, . . . , Al1 x is called Krylov matrix for A and x.

 

(b)
n o
Kl (A, x) := R (Kl (A, x)) = Span x, Ax, A2 x, . . . , Al1 x

is called Krylov space for A and x.

We have just observed that for symmetric matrices A Rn,n already after few matrix vector
multiplications, Krylov spaces yield good approximations to eigenvalues 1 and n , i.e., eigen-
values at the exterior of the spectrum. We expect that a similar property holds for general
matrices A Cn,n . Heuristic: Krylov spaces are good search spaces!

In the following we present a few properties of Krylov spaces. In particular, we construct the
relationship to minimal polynomials and Hessenberg matrices.

Reminder: Let A Cn,n , x Cn , then there exists a unique normalized polynomial p of

smallest degree such that

0 = p (A) x = Am x + m1 Am1 x + . . . + 1 Ax + 0 x.

Then p is called minimal polynomial of x with respect to A.

33
Lemma 45 Let A Cn,n , x Cn and let be the degree of the minimal polynomial of x
with respect to A. Then

(a) dim Km (A, x) = m m .

(b) K (A, x) ist A-invariant.

(c) Km (A, x) = K (A, x) fur m

Proof: Exercise. 2

Lemma 46 Let A Cn,n and g1 Cn be such that g1 , Ag1 , . . . , Am1 g1 are linearly inde-
pendent. Suppose that g2 , . . . , gn are such that G = [g1 , g2 , . . . , gn ] is nonsingular and let
B = G1 AG = [bij ]. Then the following are equivalent

(a) bjk = 0 for k = 1, . . . , m 1 and j = k + 2, . . . , n, i.e.,

b11 ... ... ... ... b1n
..

b21 .

.. ..

0 . .

B= .. .. .

. . bm,m1 bmm . . . bmn
.. .. ..

..
. 0 . . .
.. .. .. ..

..
.

. . . .
0 ... 0 bn,m . . . bnn

(b) Span {g1 , . . . , gl } = Kl (A, g1 ) for l = 1, . . . , m

If one of the conditions is satisfied and if m = n, then B is an unreduced Hessenberg matrix.

Proof: Exercise. 2

3.2 The Arnoldi algorithm

Let A Cn,n be given. We already know the decompositions

QR decomposition Hessenberg reduction

A = QR A = QHQH
Householder Householder
Gram-Schmidt new: Arnoldi

34
In the transformation with Householder matrices we construct the structure (triangular/Hessenberg)
by applying orthogonal transformations to the whole matrix, in the Gram-Schmidt/Arnoldi
method we construct the structure (triangular/Hessenberg) column by column without ever
transforming the matrix itself.

Ansatz: Let Q = [q1 , . . . , qn ] be unitary and H a Hessenberg matrix. We want QH AQ = H

or
h11 . . . ... h1n
..
h21 .
A [q1 , . . . , qn ] = [q1 , . . . , qn ] . ..
,

.. .


0 hn,n1 hnn
respectively. Comparing the k-th columns we obtain

Aqk = h1k q1 + . . . + hkk qk + hk+1,k qk+1 .

Thus,
k
!
1 X
qk+1 = Aqk hik qi .
hk+1,k
i=1
Due to the orthonormality of the qi we get

qjH Aqk = hjk , j = 1, . . . , k

Thus we can determine qk+1 from q1 , . . . , qk if hk+1,k 6= 0, which holds if H is unreduced.

Algorithm (Arnoldi, 1951):

Determines for x Cn \ {0} and A Cn,n a unitary matrix Q = [q1 , . . . , qn ], such that
Q1 AQ = H is in Hessenberg form.
x
1) Start: q1 = ||x|| .

2) For k = 1, 2, . . . , n 1
k
X
(a) qk+1 := Aqk hik qi , hik = qiH Aqk .
i=1
(b) hk+1,k := ||qk+1 ||.
1
(c) qk+1 = qk+1 .
hk+1,k

Remark 47 (a) The algorithm stops if hm+1,m = 0 for some m ( good breakdown). Then
k+1
X
Aqk = hjk qj
j=1

for k = 1, . . . , m 1 and
m
X
Aqm = hjm qj .
j=1

35
 Thus,
h11 ... h1,m1 h1m

h21 ..
.

A [q1 , . . . , qm ] = [q1 , . . . , qm ] 0 .. ..
,

. .
.. ..

..
. . .
0 . . . hm,m1 hmm
| {z }
=:Hm

i.e., the subspace Span{q1 , . . . , qm } is A-invariant.

(b) If hm+1,m 6= 0 then

h11 ... h1m
.. ..
h21 . .
A [q1 , . . . , qm ] = [q1 , . . . , qm+1 ]
.. ..
.

0 . .
0 . . . hm+1,m

In other words we have

AQm = Qm Hm + qm+1 [0, . . . , 0, hm+1,m ]

= Qm Hm + hm+1,m qm+1 eTm .

This is called the Arnoldi relation. Due to the orthonormality of the qi we also have QH
m AQm =
Hm .

Consequence: Due to the relation between Hessenberg matrices and Krylov spaces it follows
that
Span {q1 , . . . , ql } = Kl (A, x)
for l = 1, . . . , m + 1. Thus the Arnoldi algorithm computes orthonormal bases of Krylov
spaces.
Application: The Arnoldi algorithm as projection method for A Cn,n , n large.

1) For a given start vector x 6= 0 compute the vectors q1 , q2 , . . . with the Arnoldi method.

2) Stop after m  n steps

(a) either due a good breakdown in step m;

(b) or because we cannot store more vectors or because we have convergence. Note
that the computation of
Xk
qk+1 = Aqk hjk qj
j=1

is more and more expensive in each further step (concerning flops and storage).

36
 This yields an orthonormal basis of the Krylov spaces:

Kl (A, x) = Span{q1 , . . . , ql } = R (Ql ) , l = 1, . . . , m

3) If hm+1,m = 0, then R (Qm ) = Km (A, x) is invariant.

4) If hm+1,m 6= 0, then choose Km (A, x) = R (Qm ) as search and test space for a projection
method, i.e., determine C and v R (Qm ) \ {0} with

Av v R (Qm ) .

Hope: Since R(Qm ) is a Krylov space, some of the (, v) are good approximations to eigen-
value/eigenvector pairs of A.

Definition 48 Let A Cn,n and Cn K = 6 {0} be a subspace. Then (, v) C, Cn is called

a Ritz pair of A with respect to K, is called Ritz value and v Ritz vector if

v K \ {0} und Av v K.

The Ritz pairs are obtained from the eigenvalues of Hm = QH

m AQm C
m,m . This follows

form the following lemma.

Lemma 49 Let A Cn,n , let Qm Cn,m be isometric, C, z Cm and v = Qm z. Then

QH
m AQm z = z Av v R (Qm ) .

Proof:

QH H
m AQm z = z = Qm Qm z QH
m (Av v) = 0 Av v R (Qm ) .

Remark 50 We can compute the eigenvalues of Hm by the Francis QR algorithm and here
we can exploit that Hm is already in Hessenberg form. To compute the eigenvectors, we carry
out one step of inverse iteration with a computed eigenvalue of Hm as shift, i.e., we choose
a start vector w0 Cm , ||w0 || = 1, solve

(Hm Im ) w1 = w0
w1
for w1 and set w1 = ||w 1 ||
. Since is already a good eigenvalue approximation, i.e., a good
shift, one step is usually enough.

To check whether a Ritz pair is a good approximation, we can compute the residual. A
small residual means a small backward error, i.e., if Av v is small and the eigenvalue is
well-conditioned, then (, v) is a good approximation to an eigenvalue/eigenvector pair of A.
Theorem 51 Let A, Qm , Hm , hm+1,m be the results of m steps of the Arnoldi algorithm.
Furthermore, let z = [z1 , . . . , zm ]T Cm be an eigenvector of Hm associated with C.
Then (, v), with v = Qm z, is a Ritz pair of A with respect to R (Qm ) and

||Av v|| = |hm+1,m ||zm |.

37
Proof: Using the Arnoldi relation, it follows that

Av v = AQm z Qm z
Qm Hm + hm+1,m qm+1 eTm z Qm z


=
= Qm (Hm z z) +hm+1,m zm qm+1
| {z }
=0
||Av v|| = |hm+1,m ||zm |.

Remark 52 (a) For the computation of the residual Av v we do not need to determine
the Ritz vector v explicitly.
(b) After some iterations in finite precision arithmetic the orthonormality of the qi dete-
riorates. This happens in particular when |hm+1,m | is small. Then the Ritz values
deteriorate as well. This can be fixed by re-orthonormalization using the modified Gram-
Schmidt method for q1 , . . . , qm .
(c) We know how to detect good approximations but we are not sure that they occur.

3.3 The Symmetric Lanczos Algorithm

Special case: A = AH Cn,n is Hermitian.
Let H = QH AQ be in Hessenberg form, then T := H is tridiagonal. Suppose we have
computed T with the Arnoldi algorithm. Then

1 1 0 ... 0
. . . . . . ..
1
. . . .

T = 0
. .
.. .. . .. Rn,n

0
.. . . . .

..
. . . . n1
0 . . . 0 n1 n
is even real, since the diagonal of an Hermitian matrix is real and, furthermore ,we have
m = hm+1,m = ||qm+1 || for m = 1, . . . , n 1. With Q = [q1 , . . . , qm ], comparing the columns
in AQ = QT , we obtain that

Aq1 = 1 q1 + 1 q2
Aqk = k1 qk1 + k qk + k qk+1 , k = 2, . . . , n 1
Aqn = n1 qn1 + n qn

This is called a 3-term recursion. Since q1 , . . . , qn are orthonormal we have, furthermore, that

k = qkH Aqk .

Algorithm (Lanczos, 1950)

Given AH = A Cn,n .

38
 x
1) Start: Choose x 6= 0. Then set q0 := 0, q1 := and 0 := 0.
||x||
2) For k = 1, 2, . . . , m

(a) k := qkH Aqk ,

(b) rk = Aqk k1 qk1 k qk ,
1
(c) If rk = 0 STOP. Otherwise set k := ||rk || and qk+1 := rk .
k

Remark 53 (a) The Lanczos algorithm is essentially the Arnoldi algorithm for A = AH .

(b) As in the Arnoldi algorithm, after m steps we have (due to Lemma 46) that

Span {q1 , . . . , ql } = Kl (A, x) , l = 1, . . . , m

The eigenvalue/eigenvector pairs of the matrix

1 1 0
..
1 2 .
Tm = .. ..


. . m1
0 m1 m

yield Ritz pairs for A with respect to R (Qm ).

(c) As in the Arnoldi method, in finite precision arithmetic the orthonormality of q1 , . . . , qm

deteriorates and re-orthonormalization is necessary. Otherwise one obtains spurious
eigenvalues.

(d) Due to the 3-term recursion we do no need to store more than three vectors qi , thus
we can choose m much bigger than in the Arnoldi algorithms. However, if we do not
store the qi then we cannot re-orthonormalize. The it is necessary to detect the spurious
eigenvalues and remove them. (Cullum-Willoughby method 1979.)

3.4 The Nonsymmetric Lanczos Algorithm

A disadvantage of Arnoldi algorithm is that the recursion becomes more and more expensive
the more steps we perform. So we might ask whether there also exists a 3-term recursion for
A 6= AH ?

Idea: Transform A to tridiagonal form and allow the transformation matrix to become non-
unitary, i.e., we want to determine
X 1 AX = T
with T tridiagonal, or equivalently

1 1 0
..
1 2 .
AX = X
.. ..
.

. . n1
0 n1 n

39
If we write X as
X = [x1 , . . . , xk ] ,
then we obtain
Axk = k1 xk1 + k xk + k xk+1
for k = 1, . . . , n 1 and with 0 := 0, x0 := 0. Furthermore,
T H = X H AH X = Y 1 AH Y
with Y := X . Then clearly Y H X = I, i.e., yjH xi = ij with
Y = [y1 , . . . , yn ] .
Such families of vectors (x1 , . . . , xn ) and (y1 , . . . , yn ) are called bi-orthogonal. We again com-
pare columns in AH Y = Y T H . This means that
AH yk = k1 yk1 + k yk + k yk+1
for k = 1, . . . , n 1 and 0 := 0, y0 := 0. Since Y H X = I, it follows that
k = ykH Axk .
Moreover,
k xk+1 = Axk k1 xk1 k xk =: rk ,
and
k yk+1 = AH yk k1 yk1 k yk =: sk .
Furthermore, we have
H 1 H
1 = yk+1 xk+1 = s rk
k k k
for all k 1. But there is still freedom in the computation of k , k . We could in principle
use one of the variants
k = ||rk ||, k = ||sk ||, k = k , . . .

Algorithm (Nonsymmetric Lanczos algorithm, with choice k = ||rk ||)

Given: A Cn,n , x1 , y1 Cn with y1H x1 = 1.
1) Start: 0 := 0, 0 := 0, x0 := 0, y0 := 0
2) For k = 1, 2, . . .
k := ykH Axk ,
rk := Axk k1 xk1 k xk ,
sk := AH yk k1 yk1 k yk ,
If sH
k rk = 0, then STOP. Otherwise

k = ||rk ||,
1 H
k = s rk ,
k k
1
xk+1 = rk ,
k
1
yk+1 = sk .
k

40
Remark 54 1) After m steps, if we do not have a breakdown (since we cannot divide by
0), we have

1 1 0
1 . . . ..

.

A [x1 , . . . , xm ] = [x1 , . . . , xm , xm+1 ]
.. .. .
| {z } . . m1

:=Xm m1 m
0 m

If we denote the submatrix consisting of the first m rows with Tm , then we have

AXm = Xm Tm + m xm+1 eTm

and analogously
AH Ym = Ym Tm
H H
+ m ym+1 eTm .

2) Due to the relationship with Krylov spaces and Hessenberg matrices, we have

Span {x1 , . . . , xl } = Kl (A, x1 ) , l = 1, . . . , m,

and
Span {y1 , . . . , yl } = Kl AH , y1 ,


l = 1, . . . , m.

3) For A = AH and x1 = y1 , i.e., ||x1 || = 1, the algorithm is identical to the symmetric

Lanczos algorithm.

4) The nonsymmetric Lanczos algorithm breaks down with a serious breakdown if sH

k rk = 0
or is very small in modulus.

(a) If rk = 0, then Span{x1 , . . . , xk } is an A-invariant subspace (good breakdown).

(b) If sk = 0, then Span{y1 , . . . , yk } is an AH -invariant subspace.
(c) If sH
k rk = 0, but sk , rk 6= 0, then the algorithm breaks down without delivering
information about invariant subspaces. There are tricks to fix this problem in some
but not all cases (look ahead Lanczos, see below), but in general this algorithm has
to be used with great care.

5) The nonsymmetric Lanczos algorithm is not numerically stable, whenever sH

k rk 0.

6) The Lanczos algorithm as projection method. After m  n steps we have

AXm = Xm Tm + m xm+1 eTm .

Since YmH Xm = Im , we have

Tm = YmH AXm .
Some of the eigenvalues of this matrix are typically good approximations to an eigenvalue
of A. Let C and z Cm , v = Xm z. If (, z) is an eigenvalue/eigenvector pair of
Tm , then
YmH AXm z = Tm z = z = YmH Xm z

41
 and this is equivalent to
YmH (AXm z Xm z) = 0.
We thus have YmH (Av v) = 0 or

Av v R (Ym ) .

Such a pair (, v) is often called Petrov pair. Here we use R (Xm ) = Km (A, x1 ) as
search space and R (Ym ) = K AH , y1 as test space.

7) The Lanczos algorithm in practice

(a) In the look-ahead-variant one weakens the bi-orthonormality to avoid the serious
breakdowns (QMR algorithm).
(b) The bi-orthonormality deteriorates due to round-off errors in finite precision arith-
metic.

3.5 Convergence of Krylov Space Methods

For A Cn,n and x Cn Krylov space methods construct Xm = [x1 , . . . , xm ] Cn,m such
that
Span {x1 , . . . , xl } = Kl (A, x1 ) , l = 1, . . . m
We stop the method after m  n steps and choose K = Km (A, x1 ) as search space together
with a test space L Cn and then we compute pairs

C, v K \ {0} ,

such that
Av v L.
In the Arnoldi and the symmetric
Lanczos algorithm we choose L = K, in the nonsymmetric
Lanczos method L = Km AH , y1 .

Then the obvious question is whether in the computed pairs (, v) there are good approxi-
mations to eigenvalue/eigenvector pairs?

Lemma 55 Denoting by m1 the set of polynomials of degree less than or equal to m 1,

then
Km (A, x) = {p(A)x | p m1 } .

Proof: Let w Km (A, x). Then there exist 0 , . . . , m1 C with

w = 0 x + 1 Ax + . . . + m1 Am1 x = p (A) x,

where p (t) = 0 + 1 t + . . . + m1 tm1 . 2

The convergence is typically very good for eigenvalue in the outer part of the spectrum and
very slow for the eigenvalues in the interior. Quantitative results using Lemma 55 are based
on optimal polynomial approximations, but a complete convergence analysis in all cases is an
open problem.

42
3.6 The Implicitely Restarted Arnoldi algorithm
The Arnoldi algorithm becomes more and more expensive the more iterations one performs,
and the alternative nonsymmetric Lanczos is unstable. Can we resolve the problem with
Arnoldi algorithm?

Ideas:
1) Use restarts: After m steps of the Arnoldi algorithm with startvector x choose p
m1 with 
large for desiredi ,
|p (i ) | =
small for undesiredi .
Then choose p(A)x as new start vector and restart the Arnoldi algorithm. Since the start
vector has been enlarged in the components of the direction of the desired eigenvectors,
we except that the Krylov spaces contain, in particular, good approximations to the
desired eigenvectors.
The disadvantage is that we start with a single new vector, we loose a lot of already
obtained information. It would be ideal if we choose the new start vector p (A) x opti-
mally in the sense that we keep as much information as possible, i.e., maximally many
approximation to desired eigenvalues. This is again an open problem.
2) To preserve more information, we proceed as follows. If k eigenvalues are desired then
we run the Arnoldi algorithm for m = k + l steps. Then keep k vectors and throw away
l vectors. This leads to the

IRA (implicitly restarted Arnoldi method ), Sorensen, 1992.

The strategy of the implicitly restarted Arnoldi algorithm is as follows:

1) After m steps of the Arnoldi algorithm (without good breakdown) we have the Arnoldi
relation
AQm = Qm Hm + hm+1,m qm+1 eTm .
Qm = [q1 , . . . , qm ] is isometric and Hm Cm,m is in Hessenberg form.
2) Choose l Shifts 1 , . . . , l C. (Details will be described later!)
3) Carry out l steps of the QR algorithm with shifts 1 , . . . , l .

H (1) = Hm
For j = 1, . . . , l,
H (j) j I = Uj Rj (QR decomposition)
H (j+1) = Rj Uj + j I
end
Hm = H (l+1)
U = U1 . . . Ul
(i) (i)
Then Hm = U H Hm U and furthermore every Ui has the form Ui = G12 . . . Gm1,m with
(i) (i)
Givens rotation matrices G12 , . . . , Gm1,m . Every Ui is a Hessenberg matrix, thus U is
a product of l Hessenberg matrices and hence a band matrix with lower bandwidth l.

43
 4) We have AQm = Qm Hm + fm+1 eTm with fm+1 = hm+1,m qm+1 . Hence,

A Qm U = Qm U U H Hm U +fm+1 eTm U ,
| {z } | {z } |{z}
=:Qm Hm =:uT
m

where uTm is the last column of U . Thus, we have

AQm = Qm Hm + fm+1 uTm

with
uTm = [0, . . . , 0, , , . . . , ].
| {z } | {z }
k1 l

for some C. We then partition Qm as

 
Qm = Qk Ql = [q1 , . . . , qm ]

and set Qj = [q1 , . . . , qj ] for j = 1, . . . , m. In the same way we partition Hm . Then

i 
Hk
h i h
A Qk Ql = Qk Ql + fm+1 [0, . . . , 0, , , . . . , ] .
e1 eTk Hl

We then delete the last l columns and obtain

AQk = Qk Hk + Ql e1 eTk + fm+1 [0, . . . , 0, ]

 
= Qk Hk + Ql e1 + fm+1 eTk
| {z }
=:fk+1

= Qk Hk + fk+1 eTk .

This is again an Arnoldi relation, since one easily see that QH

k fk+1 . In particular, this
Arnoldi relation is the same as that after k steps of the restarted Arnoldi algorithm
with the restart vector q1 . This is the reason for the terminology implicit restart.

5) Then we perform l further Arrnoldi steps and begin again with 1) until sufficiently many
eigenvalues are found.

This concept works very nicely but some questions remain.

1) How to choose the shifts 1 , . . . , l ?

2) What is the relation between q1 and q1 ?

3) What is the relation between the corresponding Krylov spaces?

Lemma 56 Let p (t) = (t 1 ) . . . (t l ). Then, with the notation and assumptions above,

p (A) Qm = Qm U R + Fm ,
 
where Fm = 0 Fm and R = Rk . . . R1 .

44
Proof: We have already shown that

p (Hm ) = (Hm l I) . . . (Hm 1 I) = U R.

Then we show via induction on l that

p (A) Qm = Qm p (Hm ) + Fm .

l = 1: We have the Arnoldi relation AQm = Qm Hm + hm+1,m qm+1 eTm . Then

(A 1 I) Qm = Qm (Hm 1 I) + hm+1,m qm+1 eTm .

| {z }
=:F1

l 1 l: We have

p (A) Qm
= (A 1 I) (A 2 I) . . . (A l I) Qm
 
I.V.
= (A 1 I) Qm (Hm 2 I) . . . (Hm l I) + Fm1
Qm (Hm 1 I) + hm+1,m qm+1 eTm (Hm 2 I) . . . (Hm l I) + (A 1 I) Fm1
 
=
h i
= Qm p (Hm ) + hm+1,m qm+1 eTm (Hm 2 I) . . . (Hm l I) + m l + 1 l 1R 0 (A 1 I) Fm1 ,
| {z }
has band widthl1
|
{z
}

=0,...,0,, . . . ,

| {z }
l
| {z }
=:Fm

2
 
where Fm = 0 Fm has the desired form.

Theorem 57 With the notation and assumptions above

1) q1 = p (A) q1 for a C.

2) R(Qj ) = p (A) R (Qj ) = R (p (A) Qj ), for j = 1, . . . , m.

Proof:
1) One has

p (A) Qm = Qm U R + Fm = Qm R + Fm
| {z }
=Qm

r11 . . . r1m
= Qm
.. .. +  0 F  .
. . m
0 rmm

Comparing the first columns yields

p (A) q1 = q1 r11 .

45
 Therfore, choose = 1/r11 . Here r11 6= 0, otherwise we would have p (A) q1 = 0, i.e.,
the minimal polynomial of q1 with respect to A would have degree l. But then the
Krylov space Kl (A, q1 ) is invariant, i.e., the Arnoldi would have had a good
breakdown after l steps which is a contradiction to the fact that we have done m > l
steps without breakdown.
2) We could proceed as in 1) but we use the following approach: Since we know already
that
R (Qj ) = Kj (A, q1 ) , j = 1, . . . , m
and
R(Qj ) = Kj (A, q1 ) , j = 1, . . . , m
it follows that
R(Qj ) = Span p (A) q1 , Ap (A) q1 , . . . , Aj1 p (A) q1


= Span p (A) q1 , p (A) Aq1 , . . . , p (A) Aj1 q1



= p (A) Span q1 , Aq1 , . . . , Aj1 q1



= p (A) R (Qj )

As a consequence we obtain that R(Qj ) = (A 1 I) . . . (A l I) R (Qj ) , j = 1, . . . , m.

This correspond to l steps of subspace iteration with shifts 1 , . . . , l . Then we can expect
that Kj (A, q1 ) contains better approximations to eigenvectors than Kj (A, q1 ).
Choice of shifts: Suppose that A is diagonalizable and (v1 , . . . , vn ) is a basis of eigenvectors
to eigenvalues 1 , . . . , n . Then
q1 = c1 v1 + . . . + cn vn with ci C.
Since q1 = p(A)q1 it follows that
q1 = c1 p (1 ) v1 + . . . + cn p (n ) vn .
But then |p (i ) | is large (small) if i if far from j (respectively near to a) j . The component
to eigenvalues from 1 , . . . , l will be enlarged q1 . Choose 1 , . . . , l far away of the desired
eigenvalues.
Example 58 Search k eigenvalues C. Compute the m = k + l eigenvalues of Hm and
choose the l eigenvalues of Hm furthest away from as shifts.
Remark 59 1) Locking and purging. If (, v) is a converged Ritz pair, then we would like
to avoid that further Ritz pairs converge to this one. This can be done in different ways.
(a) If is desired then v is locked by working in the orthogonal complement.
(b) If is undesired then v is removed (purged).
Details can be found in Lehoucq/Sorensen 1996 Deflation techniques for an IRA iter-
ation.
2) The function eigs in Matlab is based on the ARPACK package that contains the best
implementation of the implicitly restarted Arnoldi method, see also the ARPACK user
guide.

46
3.7 The Jacobi-Davidson Algorithm
If only a specific pair (, u) of a large sparse matrix A Cn,n is desired (e.g., the eigenvalue
largest in modulus or the one nearest to C, then alternative methods can be considered.

1. Idea. Apply the Arnoldi algorithm with shift-and-invert, i.e., apply the Arnoldi algorithm
to (A I)1 . This has the disadvantage that per step we have to solve a linear system
with A I. In practice, this has to be done very accurately, otherwise the convergence
deteriorates. If a sparse LR decomposition of (A I)1 can be determined, then this is
acceptable otherwise this is problematic.

2. Idea. We use again a projection method. We choose an isometric matrix Qm Cn,m

with m  n and use R (Qm ) as search and test space. Then we have again
QH
mA |
Qm z = z Av v R (Qm ) ,
{z }
=:v

i.e., we obtain Ritz pairs (, v) from QH

m AQm .
The new idea is that R (Qm ) does not have to be a Krylov space.

This leads to the following task: For q1 Cn , ||q1 || = 1 construct an isometric matrix
Qm = [q1 , . . . , qm ] ,
such that (, u) is quickly approximated by Ritz pairs of A with respect too R(Qm ). When
q1 , . . . , qk are constructed, then we compute Ritz pairs of A with respect to R(Qk ). Let
(k , vk ) be the Ritz pair with ||vk || = 1 that approximates (, u) best.
Ansatz:
u = vk + x, with x vk ,
= k + .
Here the unknown u is scaled so that (u vk ) vk . This can be achieved by determining an
approximation to x and computing qk+1 from x by orthonormalizing against q1 , . . . , qk .
Computation of x. Consider
P = I vk vkH .
This is the orthogonal projection to Span (vk ) . Furthermore, let
rk := Avk k vk
be the residual of (k , vk ) with respect to A. Then
P vk = 0, P x = x, P rk = rk ,
since (k , vk ) is a Ritz pair, and hence rk R (Qk ) 3 vk , i.e., rk vk . Moreover,
Au = u
A (vk + x) = (vk + x)
(A I) x = (A I) vk
= (A k I) vk + vk = rk + vk
P (A I) x = P rk = rk
P (A I) P x = rk and x vk .

47
Unfortunately we cannot solve the last equation, since we do not know . Therefore, we
replace by the Ritz value k ,

P (A k I) P x = rk , x vk . (3.3)

This equation is called the Jacobi correction equation.

Remark 60 P (A k I) P is the orthogonal projection of A k I on Span {vk } .

Algorithm: (Jacobi-Davidson algorithm, Sleijpen/van der Vorst, 1996)

For a given matrix A Cn,n , this algorithm computes Qm Cn,m isometric, such that
Mm = Q Hm AQm contains good approximations to a given pair (, u).

1) Start: Choose q1 Cn , ||q1 || = 1.

2) Iterate for k = 1, 2, . . . to convergence

(a) Qk = [q1 , . . . , qk ] , wk = Aqk , Wk = [w1 , . . . , wk ] = AQk and Mk = QH

k Wk .
(b) Compute the Ritz pairs (i.e., the eigenvalue/eigenvector pairs of Mk ) and choose
the Ritz pair (k , vk ), that approximates (, u). Then vk = Qk z for z Ck .
(c) Compute rk = Wk z k Qk z. (This is the residual, since rk = Wk z k Qk z =
AQk z k Qk z = Avk k vk .)
(d) If ||rk || is sufficiently small, then we stop the iteration since we have a converged
eigenvalue/eigenvector pair. Otherwise solve the Jacobi correction equation (3.3)
for x.
(e) Compute qk+1 from x by orthonormalization against q1 , . . . , qk .

Interpretation: Since P x = x it follows from the Jacobi correction equation (3.3) that

(A k I) x = rk + vk

with = vkH (A k I) x. Hence,

x = (A k I)1 rk + (A k I)1 vk
= vk + (A k I)1 vk

Since vk is already in the search space, we have extended R (Qk ) by the scaled vector
(A k I)1 vk . Since
v H Avk
vkH (Avk k vk ) = 0, k = kH ,
vk vk
this corresponds to the vector that one obtains by applying one step of the Rayleigh quotient
iteration to vk , which converges cubically for A = AH and at least quadratically otherwise.

As a consequence in general, the Jacobi-Davidson method converges quadratically to an eigen-

value/eigenvector pair (, u) if the Jacobi correction equation is solved exactly in every step.
The Jacobi-Davidson method can be interpreted as a Newton method.

Remark 61 1) In general, it is too expensive to solve the correction equation exactly,

instead one uses approximations such as

48
 (a) P (A k I) P I, hence x = rk . This is formally equivalent to the Arnoldi
algorithm (Exercise).
(b) P (A k I) P (D k I) where D is the diagonal part of A.
(This is the Davidson method in quantum chemistry.) It works very well for diag-
onally dominant matrices.
(c) Solve the correction equation with iterative methods (see next chapter).

2) The algorithm is called Jacobi-Davidson since as in the Davidson algorithm it uses

a stepwise extension of the search space. For the approximation P (A k I) P
(D k I) we get back the Davidson algorithm.
Otherwise the following ansatz is due to Jacobi.

cT
      
1 1 1
A = = .
z b F z z

This is equivalent to

= + cT z,
(F I) z = b.

Jacobi suggested to solve (F I)z = b iteratively and to compute an improved eigen-

value from = + cT z approximation. The basic idea is the search for corrections in
the orthogonal complement, i.e., [0, z T ]T is orthogonal to the ansatz vector [1, 0]T . In
the Jacobi-Davidson-algorithm we search orthogonal to the last approximation.

3) We can compare the Jacobi-Davidson method and the Arnoldi algorithm with shift-and-
Invert for the computation of a desired eigenvalue/eigenvector pair. Jacobi-Davidson
consists in the solution of the Jacobi correction equation P (A k I) P x = rk and the
computation of the new search direction. Then we must determine the new eigenvalue of
Mk by applying the operator A in the step Wk = AQk . We can solve the Jacobi correc-
tion equation approximatively, since it is only used for the search direction. The spectral
information of the operator A will be injected in the step Wk = AQk where Mk and its
eigenvalues are computed. When we solve Jacobi correction equation approximatively
the convergence may be slowed down.

(a) In the Arnoldi algorithm in both steps, the computation of the search direction and the
application of the operator are combined in the equation
k
qk+1 = (A I)1 qk
X
hik qi .
i=1

The solution of the linear system has to be done very accurately. If this is not done
accurately enough, then the information about the operator is not injected well enough.
Furthermore, restarts, locking and purging is easier in the Jacobi-Davidson method,
since we do not need to preserve an Arnoldi relation.

(b) On the other hand the Jacobi-Davidson algorithm approximates only one pair at a time
while the Arnoldi method determines several eigenvalues at the same time.

49
3.8 Large Scale Generalized Eigenvalue Problems
For large scale generalized eigenvalue problems Ex = Ax with regular pairs (E, A) we
can apply all the methods from the previous sections whenever a solution of linear systems
(0 E A)x = b are feasible for chosen shifts 0 .
Set = ( + 0 )1 and replace Ex = Ax by x = (0 E A)1 Ex =: Ax and apply the
chosen algorithm to the matrix A. Whenever a matrix-vector multiplication with A is needed
we have to carry out a matrix-vector multiplication with E and solve a linear system with
(0 E A).
If an eigenvalue/eigenvector pair (, x) has been computed then it corresponds to an eigen-
value = 1 0 with the same eigenvector, using the spectral transformation =
( + 0 )1 . Using this relation we can make any eigenvalue an exterior eigenvalue by choosing
an appropriate 0 .
This spectral transformation can also be used in the case E = I to achieve fast convergence
near any chosen shift.

50
Chapter 4

Iterative methods for Large Sparse

Linear Systems

Situation: A Cn,n sparse, n large, b Cn .

Goal: Determine x Cn with Ax = b.

4.1 Splitting-Methods
The basic idea of splitting methods is to split the matrix A in two summands A = M N
and to transfer the linear system to a fix-point equation

M x = N x + b.

This immediately leads to an iterative method via the recursion

M x(k+1) = N x(k) + b.

(a) If A, M are nonsingular and % M 1 N < 1, where

Remark 62

% (B) := max ||,

(B)

then the iteration converges for every start vector x0 to A1 b (Exercise).

(b) The convergence is linear with convergence rate % M 1 N . This is typically too slow

compared with other methods so that splitting methods are today used mostly only in the
context of preconditioning.

0 ... 0 0 0
Example 63 We split A as A = . . . ... + .. .. . .
+ . . .

.
0 0 0 ... 0
| {z } | {z } | {z }
=:L =:D =:R

(a) Set M = D and N = L R. This is called Jacobi-method.

(b) Set M = L + D and N = R. This is called Gau-Seidel method.

51
4.2 The Conjugate Gradient Method (CG)
Special case: A Rn,n symmetric and positive definite, b Rn .

Basic idea: Consider

1
: Rn R, x 7 (x) = xT Ax xT b.
2
Then the gradient is (x) = Ax b the Hessian is Hess (x) = A, i.e., our linear system
corresponds to a minimization problem, since x = A1 b is the unique global minimum of .
1
(x) = bT A1 b.
2
Thus we may hope that a stepwise iterative minimization method for will converge to a
solution of the linear system.

4.2.1 Steepest Descent

Idea: decreases most in the direction of the negative gradient

(x) = b Ax.

Definition 64 Let A Cn,n and x, b Cn . Then

r = b Ax

is called residual of x with respect to A and b.

For r 6= 0 we have that (x + r) < (x) for some > 0. Thus we can decrease by
choosing the parameter to minimize the residual.

Lemma 65 The minimum in 7 (x + r) is given for

rT r
= .
rT Ar
Proof: Exercise. 2

Algorithm (Steepest Descent)

Determines for A Rn,n symmetric positive definite and b Rn the solution x = A1 b of
Ax = b.

1) Start: Choose x0 Rn .

2) Iterate for k = 1, 2, . . . to convergence

(a) rk1 = b Axk1

T r
rk1 k1
(b) If rk1 = 0 then stop and use xk1 = A1 b. Otherwise set k = T Ar
.
rk1 k1

52
 (c) xk = xk1 + k rk1 .

Remark 66 One can show that

  
1 1 1
(xk+1 ) + bT A1 b 1 (xk ) + bT A1 b .
2 2 (A) 2

We thus have global convergence for all start vectors. But the method has many disadvan-
tages.

(a) The convergence is very slow if 2 (A) is large.

(b) Furthermore, even if becomes small very quickly, then this is not automatically true
for the residual.

The main reasons for these disadvantages are the following.

1) We minimize only in one search direction rk , but we have many more directions than
one (namely r0 , . . . , rk ).

2) The search directions are not different enough.

4.2.2 A-Conjugate Search Directions

To improve the convergence behavior of the descent method we add a little modification. The
basic idea is to choose in every step instead of the negative gradient a direction p Rn with
p 6 r. Then we also find in this direction a decrease of . Thus in every step instead of rk
we choose a search direction pk with pTk rk 6= 0.
We require the following conditions for pk+1 and xk+1 .

R1) p1 , . . . , pk+1 are linearly independent.

R2) (xk+1 ) = min (x), where Rk+1 := x0 + Span {p1 , . . . , pk+1 }.

xRk+1

R3) xk+1 is easily computed from xk .

The first two conditions R1) and R2) together guarantee convergence in at most n steps in
exact arithmetic, since we minimize over the whole space Rn .

To compute pk+1 and xk+1 , suppose that the search directions p1 , . . . , pk Rn and xk with
(xk ) = min (x) are already computed and then determine pk+1 and xk+1 with (xk+1 ) =
xRk
min (x), such that all three conditions R1)-R3) are satisfied.
xRk+1

In order to achieve this we set xk = x0 + Pk yk with Pk = [p1 , . . . , pk ] and yk Rk and make

the ansatz
xk+1 = x0 + Pk y + pk+1

53
for y Rk , R. Then we determine the parameters y and . We have
1
(xk+1 ) = (x0 + Pk y + pk+1 )T A (x0 + Pk y + pk+1 ) (x0 + Pk y + pk+1 )T b
2
1
= (x0 + Pk y) + pTk+1 A (x0 + Pk y) pTk+1 b + 2 pTk+1 Apk+1
2
1 2 T
= (x0 + Pk y) + pk+1 APk y + pk+1 Apk+1 pTk+1 r0 .
T
| {z
nur y
} |2 {z }
nur
If the mixed term was not there then we could minimize separately over the two variables.
Thus we choose pk+1 so that
pTk+1 APk = 0
and obtain
 
1 2 T
min (x) = min (x0 + Pk y) + min pk+1 Apk+1 pTk+1 r0 .
xRk+1 yRk R 2
| {z } | {z }
Sol. y=yk pT r
k+1 0
Sol. k+1 = T
p Apk+1
k+1

The first minimization problem is solved by y = yk , since xk = x0 + Pk yk satisfies

(xk ) = min (x) .

xRk

pT
k+1 r0
The second minimization is just a scalar minimization and solved by k+1 = T
pk+1 Apk+1
. Thus
we have satisfied conditions R2) and R3).

As a consequence, we choose A-conjugate search directions pk , i.e., we choose

pk+1 Span {Ap1 , . . . , Apk } , k = 1, 2, . . . .

Then
pTi Apj = 0, i 6= j, i, j = 1, . . . , k
i.e., p1 , . . . , pk are orthogonal with respect to the scalar product

hx, yiA := y T Ax.

Then the question arises whether we can always find A-conjugate search directions.

Lemma 67 If rk = b Axk 6= 0, then there exists pk+1 Span {Ap1 , . . . , Apk } with
pTk+1 rk 6= 0.

Proof: For k = 0 this is clear (choose e.g. p1 = r0 ). For k 1 then with rk 6= 0 it follows
that
A1 b 6 Rk = x0 + Span {p1 , . . . , pk } ,
since A1 b is the unique minimum, which however is not reached yet, since rk 6= 0.
Therefore,
b 6 Ax0 + Span {Ap1 , . . . , Apk }

54
or
r0 = b Ax0 6 Span {Ap1 , . . . , Apk } .
Thus there exists pk+1 Span {Ap1 , . . . , Apk } with pTk+1 r0 6= 0. Since
xk x0 + Span {p1 , . . . , pk }, we have

rk = b Axk r0 + Span {Ap1 , . . . , Apk }

and thus also

pTk+1 rk = pTk+1 r0 6= 0.
2

Remark 68 From the proof of the Lemma 67 we have the following observation. Since
pT rk = pT r0 for p Span {Ap1 , . . . , Apk }T , we have, in particular, that pTk+1 rk = pTk+1 r0 , and
thus
pT r0 pT rk
k+1 = T k+1 = T k+1 .
pk+1 Apk+1 pk+1 Apk+1

We then can finally show that also the first requirement R1) is satisfied.

Lemma 69 The search directions p1 , . . . , pk are linearly independent.

Proof: The matrix PkT APk = diag pT1 Ap1 , . . . , pTk Apk is invertible, since A is positive

definite. Thus Pk has full rank, i.e., the columns p1 , . . . , pk are linearly independent. 2

Algorithm (A-conjugate search directions)

The algorithm computes for A Rn,n symmetric positive definite and b Rn the solution
x = A1 b of Ax = b.

1) Start: Choose x0 Rn

2) Iterate for k = 1, 2, . . . until convergence

(a) rk = b Axk ,
(b) If rk = 0 then stop and use xk = A1 b as solution. Otherwise choose pk+1
Span {Ap1 , . . . , Apk } with pTk+1 rk 6= 0 and compute

pTk+1 rk
k+1 = .
pTk+1 Apk+1

(c) xk+1 = xk + k+1 pk+1 .

Note that we still have freedom in the choice of pk+1 .

55
4.2.3 The Conjugate Gradient Algorithm, CG
We have seen that the choice of A-conjugate search directions has many advantages (an easy
computation of xk+1 from xk and a guaranteed convergence in at most n steps in exact
arithmetic). On the other hand we would like to keep the advantage of steepest descent
that the function decreases maximally in the direction of the negative gradient, i.e., this is
heuristically a good search direction. The idea is then to use the freedom in pk+1 to choose
that pk+1 which is nearest to rk , the direction of the negative gradient, i.e., to choose pk+1 so
that
||pk+1 rk || = min ||p rk ||. (4.1)
pSpan{Ap1 ,...,Apk }

At first sight this looks strange, since we wanted to choose directions that allow an easy
solution of the optimization problem and here we introduce another optimization problem.
We will see now that this optimization problem is easy to solve since it will turn out that
pk+1 is just a linear combination of pk and rk .

In the following, under the same assumptions as before, we choose the A-conjugate search
directions to minimize (4.1) for k = 0, . . . , m. Let Pk = [p1 , . . . , pk ] and show then that
pk+1 Span {pk , rk }.

Lemma 70 Let k {1, . . . , m} and zk Rk , such that

||rk APk zk || = min ||rk AP z||.

zRk

Then pk+1 = rk APk zk .

Proof: Let p := APk zk , then by assumption p is the orthogonal projection of rk to

R (APk ) . Hence
||p rk || = min ||p rk ||
pR(APk )

and therefore p = pk+1 . 2

Theorem 71 If rk 6= 0 for k = 0, . . . , m, then for k = 0, . . . , m the following statements

hold.

1) rk+1 = rk k+1 Apk+1 ,

2) Span {p1 , . . . , pk+1 } = Span {r0 , . . . , rk } = Kk+1 (A, r0 ),

3) rk+1 rj for j = 0, . . . , k,

4) pk+1 Span {pk , rk } for k 1.

Proof:
1) Since xk+1 = xk + k+1 pk+1 , it follows that

rk+1 = b Axk+1 = b Axk k+1 Apk+1 .

| {z }
= rk

56
2) Applying 1) inductively we obtain
Span {Ap1 , . . . , Apk } Span {r0 , . . . , rk } , k = 1, . . . , m.
We have already shown that for all k = 0, . . . , m we have
pk+1 = rk APk zk Span {r0 , . . . , rk } .
Thus, we have
Span {p1 , . . . , pk+1 } Span {r0 , . . . , rk }
for k = 0, . . . , m. Moreover, with 1) it follows that
rk+1 Span {rk , Apk+1 } Span {rk , Ar0 , . . . , Ark }
for k = 0, . . . , m. Therefore,
r1 Span {r0 , Ar0 } ,
r2 Span {r0 , Ar0 , Ar1 } Span r0 , Ar0 , A2 r0 ,


..
. .
By induction we then have finally
Span {p1 , . . . , pk+1 } Span {r0 , . . . , rk } Kk+1 (A, r0 ) .
Equality follows by a dimension argument.
3) We show that PkT rk = 0 i.e., p1 , . . . , pk rk for all k = 1, . . . , m. By 2) we then also
have r0 , . . . , rk1 rk desired. We have xk+1 = x0 + Pk yk , where yk minimizes the
function
1
(x0 + Pk y) = (x0 + Pk y)T A(x0 + Pk y) (x0 + Pk y)T b
2
1
= (x0 ) + y T PkT (Ax0 b) + y T PkT APk y.
2
The gradient of y 7 (x0 + Pk y) therefore vanishes for y = yk , i.e.,
PkT APk yk + PkT (Ax0 b) = 0.
This is equivalent to 0 = PkT (b Ax0 APk yk ) = PkT (b Axk ) = PkT rk .
4) If k = 1, then by 2), it follows that p2 Span{r0 , r1 }. Since p1 = r0 we then have
p2 Span{p1 , r1 }. For k > 1 we partition zk from Lemma 70 as
 
w
zk = , w Rk1 , R.

with rk = rk1 k Apk . By 1) we then obtain from Lemma 70 that
pk+1 = rk APk zk
= rk APk1 w Apk

= rk APk1 w + (rk rk1 )
k
 

= 1+ rk + sk ,
k

57
 where

sk = rk1 APk1 w
k
Span{rk1 , APk1 w}
Span{rk1 , Ap1 , . . . , Apk1 }
Span{r0 , . . . , rk1 }.

(Observe that k is nonzero by construction!) By 3) then rk and sk are orthogonal.

Then we can solve the optimization problem in Lemma 4.4 by determining w and
such that
2
 
kpk+1 k2 = 1 + krk k2 + ksk k2
k
is minimal. Then, in particular, sk so that ksk k is minimal (for fixed and variable
w). But krk1 APk1 zk will be minimized (see Lemma 70) via z = zk1 and one
obtains pk = rk1 APk1 zk1 . Hence, sk is a multiple of pk , and therefore

pk+1 Span{rk , sk } Span{rk , pk }.

Corollary 72 After scaling pk+1 we have

pk+1 = rk + k pk .

Since pTk Apk+1 = 0, we furthermore have

pTk Ark
k = .
pTk Apk

Hence pk+1 can be constructed directly from pk and rk without solving a minimization problem.

Algorithm: (CG, conjugate-gradient method - Hesteness/Stiefels, 1952)

For A Rn,n symmetric, positive definite and b Rn the algorithm computes the solution
x = A1 b of Ax = b.
1) Start: x0 Rn , r0 = b Ax0 , p1 = r0 .

2) Iterate for k = 1, 2, . . . until convergence

pTk rk1
(a) k = ,
pTk Apk
(b) xk = xk1 + k pk ,
(c) rk = b Axk ,
pTk Ark
(d) k+1 = ,
pTk Apk
(e) pk+1 = rk + k+1 pk .

58
Remark 73 There are many theoretical results behind this simple algorithm, for example the
convergence after at most n steps in exact arithmetic, since the CG-algorithm is a special
case of the algorithm of A-conjugate search directions. The iterate xk satisfies
(xk ) = min (x) ,
xRk

where (x) = 21 xT Ax xT b and Rk = x0 + Span{p1 , . . . , pk }. But Span{p1 , . . . , pk } =

Kk (A, r0 ) by Theorem 71, i.e., we minimize over the affine Krylov space x0 + Kk (A, r0 ).
Hence xk satisfies
(xk ) = min (x) .
xx0 +Kk (A,r0 )

For this reason one calls the CG-algorithm a Krylov space method.

4.2.4 Convergence properties of the CG method

Using the relationship of the CG algorithm with Krylov spaces allows a detailed convergence
analysis. For this we introduce a special norm.
Definition 74 Let A Rn,n be symmetric and positive definite. Then the norm defined by

||x||A := xT Ax
on Rn is called the A-Norm or energy norm.
We would like to estimate the error
ek := A1 b xk = A1 (b Axk ) = A1 rk
where (xk ) is the sequence generated by the CG algorithm.
Theorem 75 (Optimality of CG in the A-Norm) Let A Rn,n be symmetric and
positive definite and let (xk ) be the CG sequence generated from a starting vector x0 . If
rk1 6= 0, then
||ek ||A = ||A1 b xk ||A < ||A1 b x||A
for all x x0 + Kk (A, r0 ) with xk 6= x.
Proof: We know that xk x0 + Kk (A, r0 ). Let x x0 + Kk (A, r0 ) be arbitrary and
x = xk x, i.e., x Kk (A, r0 ), and moreover
e := A1 b x = A1 b (xk x) = ek + x
Then
||e||2A = eT Ae = (ek + x)T A (ek + x)
= eTk Aek + 2eTk Ax + xT Ax
and
2eTk Ax = 2rkT A1 Ax = 2rkT x = 0
since x Kk (A, r0 ) = Span {r0 , . . . , rk1 } and rk rj for j = 0, . . . , k 1 due to
Theorem 71. Thus we obtain
||e||2A = ||ek ||2A + ||x||2A > ||ek ||2A , if x 6= 0.
2

59
Corollary 76 Let e k := {p k |p (0) = 1}. With the notation and assumptions of Theo-
rem 75 (in particular rk1 6= 0), there exists a unique polynomial pk
e k with

||pk (A) e0 ||A = min ||p (A) e0 ||A .

p
ek

Furthermore, ek = pk (A) e0 and

||ek ||A ||p (A) e0 ||A

= min inf max |p () | (4.2)
||e0 ||A p
ek ||e0 ||A pe k (A)

Proof: There exists xk x0 + Kk (A, r0 ), i.e.,

xk = x0 + pk1 (A) r0

for some pk1 k1 . Furthermore,

rk = b Axk = b Ax0 Apk1 (A) r0 .

| {z }
=:r0

Thus, we have

ek = A1 rk = A1 r0 pk1 (A) r0 = e0 pk1 (A) Ae0 = (I pk1 (A) A) e0 .

| {z } | {z }
=e0 =pk (A)k

Then the uniqueness of pk , and the first equality in (4.2) follows from Theorem 75. To prove
the inequality in (4.2), let (v1 , . . . , vn ) be an orthonormal basis of eigenvectors of A to the
eigenvalues 1 , . . . , n . Furthermore let p k and

e0 = c1 v1 + . . . + cn vn with c1 , . . . , cn R.

Then,
p (A) e0 = c1 p (1 ) v1 + . . . + cn p (n ) vn .
By the orthogonality of the vi we obtain
n
X
||e0 ||2A = eT0 Ae0 = c2i i
i=1

and
n
X n
X
||p (A) e0 ||2A = c2i p (i )2 i max p () 2
c2i i .
(A)
i=1 i=1

But this implies that

||p (A) e0 ||2A
2 max |p () |2 .
||e0 ||A (A)

60
Remark 77 1) From Corollary 76 we conclude that the CG algorithm converges fast if A
has an appropriate spectrum, i.e., one for which there exist polynomials p with p (0) = 1
and small degree such that |p () | is small for all (A). This is e.g. the case if

(a) the eigenvalues occur in clusters,

(b) all eigenvalues are far from the origin,
(then 2 (A) = max
min
is not too large).

2) With the help of Chebysheff polynomials one can prove that

 k
||ek ||A 1
2 ,
||e0 ||A +1

where := 2 (A) and

 k
||ek ||2 1
2 .
||e0 ||2 +1

3) We can improve the convergence of the CG algorithm via preconditioning:

(a) For general linear systems Ax = b consider

M 1 Ax = M 1 b

where M 1 A has an appropriate spectrum and M z = c is easy to solve.

(b) For Ax = b with A symmetric and positive definite consider

C 1 AC T C T x = C 1 b

where C 1 AC T has an appropriate spectrum and C T z = d is easy to solve.

C 1 AC T is again symmetric and positive definite.

4.2.5 The CG and the Lanczos Algorithm

In this section we use the same notation as in previous sections. Consider the matrices

1 2 0
..
1 .
Rk = [r0 , . . . , rk1 ], Pk = [p1 , . . . , pk ], Bk =

.
.
.
. n
0 1

Using the equations p1 = r0 and pi = ri1 + i pi1 for i = 2, . . . , n (see Section 4.2.3) we
obtain
Rk = Pk Bk .
Then the matrix RkT ARk is tridiagonal, since

pT1 Ap1

0
RkT ARk = BkT PkT APk Bk = BkT
.. Bk .

.
0 pTk Apk

61
Furthermore, we know from Theorem 75, that the r0 , . . . , rk1 are orthogonal and span a
r
Krylov space, i.e., krr00 k , . . . , krk1
k1 k
is an orthonormal basis of Kk (A, r0 ).
This leads to an interesting conclusion. If q1 := krr00 k and if q1 , . . . , qk are the vectors generated
by the Lanczos algorithm, then by the implicit Q Theorem
rj1
qj = , j = 1, . . . , k.
krj1 k

Thus, the tridiagonal matrix generated by the Lanczos algorithm is (except for signs) the
matrix RkT ARk , i.e.,
CG Lanczos
Application: In the course of the CG algorithm we can generate the tridiagonal matrix
RkT ARk and obtain information about extremal eigenvalues of A and the condition number
2 (A) = max
min
.

4.2.6 The GMRES algorithm

Situation: A Cn,n general and invertible, n large, A sparse, b Cn .

Goal: Determine x Cn with Ax = b.

In Section 4.2.4 we have noticed that certain affine Krylov spaces are good search spaces.
This suggests to use again a Krylov space method. In the CG algorithm we have used that
the solution x = A1 b is the unique minimum of = 21 xT Ax xT b. This, however, holds in
general only if A Rn,n is symmetric positive definite.

Idea: For a given starting vector x0 Cn and r0 := b Ax0 , determine xk via

kb Axk k2 = min kb Axk2 . (4.3)

xx0 +Kk (A,r0 )

A Hermitian ; MINRES (mininmal residuals), Paige/Saunders 1975

A general ; GMRES (generalized minimal residuals), Saad/Schultz 1986

This means that we have to solve in each step the least-squares problem

kb Axk k2 = min kb Axk2 .

xx0 +Kk (A,r0 )

In Section 4.2.5 we have seen that the CG algorithm corresponds to the Lanczos algorithm.
We expect that in the general case

GMRES Arnoldi

After k steps of the Arnoldi algorithm (without breakdown) we have the Arnoldi relation

AQk = Qk Hk + hk+1,k qk+1 eTk = Qk+1 Hk+1,k ,

62
with Qk = [q1 , . . . qk ], Qk+1 = [Qk , qk+1 ] isometric and

h11 . . . ... h1k
h21 . .
. ..
.

. . ..
Hk+1,k = 0 .. .. Ck+1,k .

.
..

..
. . hk,k1 hkk
0 ... 0 hk+1,k
r0
If q1 = kr0 k , then Span{q1 , . . . , qk } = Kk (A, r0 ). Let x x0 + Kk (A, r0 ), i.e., x = x0 + Qk y
for an y Ck . Then
kb Axk = kb A(x0 + Qk y)k
= kr0 AQk yk
= kr0 Qk+1 Hk+1,k yk
= kQH r0 Hk+1,k yk since Qk+1 is isometric,
k+1
r0
= kr0 k e1 Hk+1,k y since q2 , . . . , qk+1 q1 = kr0 k . (4.4)

!
Reminder. For the solution of least-squares problems kc M yk = min, with M Ck,n ,
k n we may
1) compute QR decomposition of M ,
Q Cn,n unitary,
 
M = QR, R= R1 0 .

2) Since Q is unitary, we have

c1 R1 y 2
 
2 H 2
QH c =
 
kc M yk = kQ c Ryk = , where c1 c2 .
c2
If R1 is invertible, then this is minimal if R1 y = c. Thus we solve R1 y = c.
In the least-squares problem (4.3), the matrix Hk+1,k is in Hessenberg form and we need to
solve this problem for every k. Suppose that we have solved the problem for k 1, i.e., we
already have a QR decomposition for Hk,k1 ,
 
Rk1
Hk,k1 = Qk Rk , Qk unitary, Rk1 =
e e e e , Rk1 upper triangular.
0
Then
e H hkk
      
eH
Q 0 eH
Q 0 Hk,k1 hkk R
ek1 Q
k Hk+1,k = k = k
0 1 0 1 0 hk+1,k 0 hk+1,k
 
Rk1 0
= .
0 hk+1,k
Thus, the element hk+1,k can be eliminated by a single Givens rotation and we obtain the QR
decomposition of Hk+1,k from Hk,k1 by this Givens rotation (this costs only O(n) flops).

Algorithm (GMRES)
For A Cn,n invertible, b Cn , and a starting vector x0 Cn , the algorithm computes the
solution x = A1 b of Ax = b.

63
1) Start: r0 = b Ax0 , h10 = kr0 k.

2) Iterate for k = 1, 2, . . . to convergence

rk
a) qk = ,
hk,k1

k
X
b) rk = Aqk hjk qj with hjk = QH
j rk ,
j=1

c) hk+1,k = krk k,

d) Determine yk such that kr0 k e1 Hk+1,k yk is minimal,

e) xk = x0 + Qk yk .

64
Remark 78 As the CG algorithm, also GMRES can be analyzed via polynomial approxima-
e k = {p k | p(0) = 1}.
tion in

x = x0 + p(A)r0 for p k1 ,

since x x0 + Kk (A, r0 ). Therefore,

 
rk := b Axk = b Ax0 Ap(A)r0 = I Ap(A) r0 = p(A)r0 for p
e k.

Then we can reformulate GMRES as

Determine p
e k , such that kp(A)r0 k is minimal.

If pk
e k is such that rk = pk (A)r0 , then

krk k = kpk (A)r0 k kp(A)r0 k

for all p
e k.

Theorem 79 Let A Cn,n be diagonalizable and V 1 AV = diagonal. Then,

krk k  
(V ) inf max p().
 
kr0 k pe k (A)

Proof: For every p

e k we have

kp(A)k = kp(V V 1 k = kV p()V 1 k

kV k kp()k kV 1 k = (V ) kp()k,

and, furthermore,
kp()k = max |p()|,
(A)

since is diagonal. Thus, we obtain

krk k = kpk (A)r0 k inf kp(A)r0 k inf kp(A)k kr0 k

p
ek p
ek

kr0 k (V ) inf max |p()|.

e k (A)
p

Consequence The GMRES algorithm (in most cases) converges fast if

1) the spectrum A is appropriate and

2) (V ) is small, i.e., if A is not too far from a normal matrix (since for a normal matrix
V can be chosen unitary, i.e., with condition number 1).

Remark 80 Convergence acceleration can again be achieved via preconditioning, i.e., instead
of Ax = b we solve M 1 Ax = M 1 b, where M y = c is easy to solve and chosen such that the
spectrum is appropriate.

65
Remark 81 Other methods for the solution of Ax = b, A invertible with A 6= AH :

1) CGN . Instead of Ax = b consider the normal equations, i.e, AH Ax = AH b with positive

definite AH A and apply to this the CG algorithm without forming the product.
Disadvantage. The condition number is squared (AH A) = (A)2 .
Advantage. The eigenvalues of AH A are the squares of the singular values of A.
Therefore CGN is a good approach for matrices A with bad spectrum but good singular
values.

2) BiCG (Bi-conjugate gradients).

CG: The computed xk x0 + Kk (A, r0 ) delivers rk r0 , . . . , rk1 ,
hence rk Kk (A, r0 ).
BiCG: choose s0 with sH
0 r0 = 1 and determine xk x0 + Kk (A, r0 )
with rk Kk (A, s0 )
BICG corresponds to the nonsymmetric Lanczos algorithm.

Ax = x Ax = b
A= AH Lanczos CG
A 6= AH Arnoldi GMRES
Lanczos BiCG

3) Survey of Krylov space methods

Common Krylov space: K = Kk (A, r0 )

Important quantity: rk = b Axk (residual)

a) Ritz-Galerkin-Ansatz: choose xk x0 + K such that rk K

; CG, FOM, GENCG
b) Minimal-residual-Ansatz: choose xk x0 + K such that krk k is minimal
; MINRES, GMRES, ORTHODIR
c) Petrov-Galerkin-Ansatz: choose xk x0 + K such that rk L, L Cn , dim L = k
; BiCG, QMR
d) Minimal error-Ansatz: choose xk x0 + K such that kxk A1 bk is minimal.
; SYMMLQ, GMERR

There are further hybrid methods, like (CGS,Bi-CGSTAB,. . . )

But none of the methods is really efficient without preconditioning. To obtain a good pre-
conditioner depends very much on the problem and usually it has to be chosen based on
knowledge about the background of the problem.

66