Chapter 2

Linear models: stability and

redundancy

2.1 Singular Value Decomposition

In this section we shall discuss the Singular Value Decomposition of matrices.

2.1.1 Stability of eigenvalues

We have repeatedly emphasized the importance of eigenvalues for the proper-
ties of linear models. However, sometimes one has matrices where eigenvalues
can only give partial information. Take for example the matrix
 
0.5 5
A= ,
0 0.45
The eigenvalues of A are distinct and equal to 0.5 and 0.45. Therefore, A
can be diagonalized:
A = V DV −1 ,
where    
0.5 0 1.00 −1.00
D= ,V = .
0 0.45 0 0.01
Assume that while computing the eigenvalues of A numerically we made an
11 percent error and estimated the second eigenvalue of A as 0.5 rather than
0.45. This will lead to the estimate of A
à = 0.5 · V · I · V −1 = 0.5 · I,

18
which corresponds to the relative error of 99 percent if L2 matrix norm is
used.
What is the reason for such an instability? As we can see from the
expression for V , eigenvectors of A are very close to being linearly dependent,
||V1 + V2 ||/||V2 || = 10−2 . Consequently, the matrix V is close to a degenerate
matrix causing V −1 to amplify numerical errors in the calculation of matrix
D.
It turns out that there are quantities characterising matrices which remain
stable even in the situation above. Namely, consider matrix B = AT A.
Matrix B is symmetric, therefore its eigenvectors are orthogonal (or can be
chosen to be orthogonal). Consider the decomposition of matrix B,

B = OW OT ,

where

O =

-0.9951 0.0987
0.0987 0.9951

W =

0.0020 0
0 25.4505

As before, let us assume that we made a 10 percent error in the determi-

nation of the largest eigenvalue, 24.45 → 28.00. This leads to the estimate
 
0.0020 0
B̃ = O OT ,
0 28.00

which differs from B by 10 percent in L2 norm. The positive square roots

of eigenvalues of AT A are referred to as singular values of matrix A. We see
that matrix B is stable with respect to numerical errors in determination
of singular values. Is there a decomposition of matrix A itself in terms of
singular values which can take advantage of this stability? As it turns out

19
there is and we will introduce the singular value decomposition in the next
section.
Finally, let us remark on the geometric meaning of singular values: as
we will verify in a moment, a singular value of matrix A is equal to the
corresponding axis of an ellipsoid representing the image of a unit sphere
under the linear map generated by A. Examining matrix W one may notice
that the corresponding ellipse has a large aspect ratio reflecting the near
linear dependence of the eigenvectors of A.

2.1.2 Singular value decomposition.

Let us state without proof that each square matrix A with real matrix ele-
ments can be written in the form
 
w1
 w2 
 tr
A=U V

 ... 
wN

where wi are non-negative real numbers (the singular values) and U and V
are orthogonal matrices.1 This decomposition is not unique, but the singular
values are (up to reordering them). Notice that

AAT = U W V T V W U T = U W W T U T .

Therefore, the singular values of A are indeed the positive square roots of
the eigenvalues of AAT .
Notice also that if u = Av and ||v|| = 1, uT (AAT )−1 u = 1. Therefore
in the two-dimensional case, singular values w1 , w2 are the sizes of the short
and long axis of the ellipse

{Av ; v ∈ R2 with |v| = 1}

(and in the higher dimensional case the numbers wi are also the lengths of
the main axes of the corresponding ellipsoid).
If A is a square symmetric matrix (as is the case with covariance matrices),
then all eigenvalues of A are real, and there is a basis of eigenvectors. So
1
A similar statement holds for complex matrices in which case matrices U and V are
unitary.

20
in this case one can choose U to be the matrix of the orthogonal column
eigenvectors, and set V = U (but with the sign reversed in some columns
in order to get the terms wi > 0). Since U tr = U −1 this gives the above
decomposition. For example the symmetric matrix
 
1 2
A=
2 1

has eigenvalues 3 and −1, and singular values 3 and 1. A SVD decomposition
is  √ √   √ √ ′
1/√2 1/ √2 3 0 1/√2 −1/√ 2
A= .
1/ 2 −1/ 2 0 1 1/ 2 1/ 2
If A is non symmetric then the relationship between eigenvalues and sin-
gular values is less obvious. For example,
 
1 1
A=
0 1
√
has eigenvalues 1 (twice), and singular values (3 ± 5)/2.
Singular value decomposition (SVD) holds for rectangular matrices as
well. Namely, we have the following

Theorem 2.1.1. Let A be a real m × n matrix with n ≥ m. Then there exist

an m × m orthogonal matrix U and orthogonal n × n matrix V such that
 
w1 0 ... 0
 w2 0 ... 0 
A = UW V T = U 
 tr
V

. .
 . 0 ... 0 
wm 0 . . . 0

where diagonal entries of m × n matrix W , w1 ≥ w2 ≥ . . . wm ≥ 0 are

non-negative real numbers.

Clearly, SVD can be applied to a rectangular matrix A with m > n. To

prove this, apply the above theorem to AT .
Let us briefly discuss the problem of calculation of singular values. For
simplicity, let us assume that rank(A) = m. Then all singular values are
positive. Then we see from the above formula, that singular values can be
computed either as square roots of eigenvalues of m × m matrix AAT =

21
U W W T U T or as square roots of non-zero eigenvalues of n × n matrix AT A =
V WTWV T.
Let us re-iterate that:

AAT = U W W T U T , AT A = V W T W V T .

Therefore, columns of m × m orthogonal matrix U are normalized orthogonal

eigenvectors of AAT , columns of n × n orthogonal matrix V are normalized
orthogonal eigenvectors of matrix AT A. Signs of the columns of U and V
are chosen in such a way that matrix W is non-negative. Notice that in
general matrices U and V are not unique even up to the sign inversion of the
columns.
It also possible to define U and V directly in terms of matrix A: Let Ui
and Vi be the i-th columns of U and V correspondingly. Using SVD,

AVi = wi Ui , AT Ui = wi Vi , i ≤ m; AVj = 0, j > m.

The triplet (wi , Ui , Vi ) is referred to as singular value, left singular vector
and right singular vector associated with wi . Theorem 2.1.1 essentially
states that one can always find orthogonal bases of left and right singular
vectors.
Matlab can be used for computing the singular value decomposition of a
matrix.

A =[1,10;0,1]
1 10
0 1
svd(A)
ans =
10.0990
0.0990
[U,S,V]=svd(A)
U =
0.9951 -0.0985
0.0985 0.9951
S =
10.0990 0
0 0.0990
V =

22
 0.0985 -0.9951
0.9951 0.0985
eig(A)
ans =
1
1
[U,V]=eigenv(A)
U =
1.0000 -1.0000
0 0.0000
V =
1 0
0 1

Note that matrix A in the above example cannot be diagonalised by a rota-

tion.

2.1.3 Application of the singular value decomposition

to solving linear equations
Let us demonstrate the usefulness of SVD for solving linear systems of equa-
tions which are degenerate within numerical precision. In this case some
singular values of the matrix which determines the system are zero within
numerical precision.
First, consider a non-degenerate case. Note that for orthogonal matrices
U, V one has U −1 = U tr and V −1 = V tr . So this implies that
 −1 
w1
−1
 w2−1 
 tr
A =V  U

. .
 . 
−1
wN

(assuming that A is invertible, i.e. assuming that wi 6= 0). Therefore in the

invertible case,
Aa = b

23
is equivalent to
 
w1−1

−1
w2−1 
 tr
a=A b=V   U b.

..
 . 
−1
wN

If wi = 0 for some i’s, then Aa = b might have no solution or infinitely many,

but even so let us define
 −1 
w1
+
 w2−1 
 tr
a = A b := V  U b

. .
 . 
−1
wN

where we replace wi−1 by 0 if wi = 0. (So we replace 1/0 by 0: a dream come

true!!) Here A+ stands for the generalized inverse of A. Of course at
this point it is not clear why this strange choice is made, but the following
proposition provides the motivation:
Proposition 2.1.1. Let a = A+ b. Then a is the minimal length vector out
of the set of vectors minimizing ||Aa − b||.
Proof:
Here is the sketch of the proof of this important result. Let us view A as
a map from a linear space S to a linear space T, a ∈ S, b ∈ T. Let us choose
the basis in S consisting of right singular vectors of A. Let us choose the
basis in T consisting of left singular vectors of A. In these bases, A = W .
Then X X
||a||2 = a2k + a2k ,
k:wk 6=0 k:wk =0
X X
||Aa − b||2 = (wk ak − bk )2 + b2k .
k:wk 6=0 k:wk =0

We see from the second formula that the minimum of ||Aa − b|| is achieved
for a vector a: ak = bk /wk for all k : wk 6= 0. Then it follows from the
second formula that the shortest such vector has ak = 0 for all k : wk = 0.
Combining the last two conclusions, we find that a = W + b or

a = A+ b

24
in the arbitrary basis.
It’s worth noting the following corollary of Proposition 2.1.1: if Aa = b
has infinitely many solutions, the minimal norm solution is a0 = A+ b.
As one can see from the proof, the proven
  result holds for rectangular
1
matrices as well. For example, take A = . Then the singular value
1
decomposition is
   √ √  √ 
1 1/√2 −1/√ 2 2
A= = (1)′ .
1 1/ 2 1/ 2 0

So the generalized inverse of A is

 √ T  √ √ 
+ 1/ 2 1/ √2 1/√2

A = (1) = 1/2 1/2 .
0 −1/ 2 1/ 2
 
0
So if we try to solve Aa = b := (which of course has no solution), we
1
get as a ‘good’ choice
   
+ 0
0
a=A = 1/2 1/2 = 1/2.
1 1

It is easy to check that a = 1/2 is the ‘shortest’ solution, which minimises

||Aa − b||.
Here is another example: Consider a symmetric matrix
 
1 0.9
A=
0.9 1

The solution to Aa = b, where b = (1, 1)T is

>> a=inv(A)*b

a =

0.5263
0.5263

25
 Suppose that A is a matrix we measured experimentally and due to nu-
merical errors A(1, 2) = A(2, 1) ≈ 1. The resulting matrix is degenerate:
 
1 1
à = .
1 1
As the result, the system Ãa = b has infinitely many solutions. However,
 
′ + 0.5
a = Ã b =
0.5

is still close to the exact solution.

2.1.4 Least square methods

Suppose you need to find β so that some data vector y fits the linear regression
model y = Xβ + ǫ where X is given beforehand. Here y, β, ǫ are vectors and
X is a matrix. The idea is to choose β so that it ’minimizes’ the error vector
ǫ, i.e. it minimizes its length ǫtr ǫ. This is equal to

β̂ = argminβ {ǫtr ǫ}
= argminβ {(y − Xβ)tr (y − Xβ)}
= argminβ {||y − Xβ||}.

By the theorem concerning generalized inverses,

β̂ = X + y.

achieves the minimal error.

2.1.5 Further applications of SVD: Principle Compo-

nent Analysis
As before let A be decomposed as
 
w1 0 ... 0
 w2 0 ... 0 
 tr
A=U V ,

..
 . 0 ... 0 
wm 0 ... 0

26
(where we assume that A is m × n matrix and n ≥ m). Here U is a m × m
matrix and V a n × n matrix. Let u1 , . . . , um be the columns of U and
v1 , . . . , vn those of V . Then the above means gives

A = w1 u1 v1tr + w2 u2 v2tr + · · · + wm um vm
tr
.

Note that ui vitr is a m × n rank-1 matrix whose j-th column is equal to ui

times j-th component of vi .
Principal component analysis is based on the idea that in certain situa-
tions the leading singular value w1 is much larger than the others. So in this
case
A is approximately equal to w1 u1 v1tr .
Note that v1tr x is equal to ± times the length of the orthogonal projection
of x onto the vector v1 and that w1 u1 v1tr x is a multiple of the vector u1 . So
Ax is approximately equal to w1 u1 times the weighted average of x (with the
weight described by the vector v1 ).
We have dealt with such a situation in Chapter 1 in the context of the
growth model: We have seen that LN for large N is proportional to the matrix
baT for some vectors a and b. Vector a determines the column weights which
were important for the total growth value. Now we can identify this matrix
as the principle component of matrix LN and vectors a and b with left and
right singular vectors of LN associated with the largest singular value.

2.1.6 Pairs trade by exploiting correlations in the stock

market.
There are stock market price forecasting methods which employ correlations
between prices for various shares or group of shares.
It is well known that daily increments of an individual stock’s price are
independent random variables. Thus forecasting price by observing the move-
ment of a single stock is not a very gratifying task. However, price variations
of groups of stocks are usually correlated. Thus having observed the move-
ment of the stock A we may be in a better position to predict the price
movement for the stock B correlated with stock A. If we see that the prices
for A move, but prices for stock B do not follow the correlation we can base
our trading strategy on the bet that price increments will soon re-converge
according to the correlation. This strategy is referred to as pairs trade.

27
 Background note. This is what Wikipedia has to say about pairs trade:

The pairs trade was developed in the late 1980s by quantitative analysts.
They found that certain securities, often competitors in the same sector, were
correlated in their day-to-day price movements. When the correlation broke
down, i.e. one stock traded up while the other traded down, they would sell
the outperforming stock and buy the underperforming one, betting that the
”spread” between the two would eventually converge.
Some real-life examples of potentially correlated pairs:

• Coca-Cola (KO) and Pepsi (PEP)

• Wal-Mart (WMT) and Target (TGT)

• Dell (DELL) and Hewlett-Packard (HPQ)

• Ford (F) and General Motors (GM)

The pairs trade helps to hedge sector- and market-risk. For example, if
the market as a whole crashes and your two stocks plummet along with it,
you should experience a gain on the short position and a negating loss on
the long position, leaving your profit close to zero in spite of the large move.
In a pairs trade, you are not making a bet on the direction of the stocks in
absolute terms, but on the direction of the stocks relative to each other.
Some mathematical algorithms relevant for pairs trading are discussed in
see Y. Zhu, D. Shasha, Fast approaches to simple problems in financial
time series streams.
The mathematical problem we need to solve in order to be able to trade
pairs is this: what is the most likely price increment of B given the observed
price increment of A.
To be more specific, let us assume that joint probability distribution of
logarithmic price increments a of stock A and logarithmic price increment
b of stock B is Gaussian with covariance matrix C,
1 − 1 (a,b)C −1 (a,b)T
̺(a, b) = e 2 ,
Z
where Z is the normalisation constant. Further in the course we will see how
Gaussianity of logarithmic increments follows from linear stock price models.
Note that a, b can be either price increments or vectors of price increments

28
in the case we are trying to correlate groups of stocks. So, what is the most
probable value of b given a? This can be computed as

b0 = argmaxb (̺(b | a)) = argmaxb (̺(a, b)/̺(a))

= argmaxb (̺(a, b)) = argminb (−log(̺(a, b))).

Therefore,
−1 T
b0 = argminb ((a, b)C −1 (a, b)T ) = argminb (2bCba −1 T
a + bCbb b ).

Therefore,
−1 T −1 T
Cbb b0 = −Cba a
Matrix C −1 is the inverse of covariance matrix. The latter is measured exper-
imentally, which leads to inaccuracy in determining small correlation com-
ponents. Therefore, it is often the case that only the principle component of
C can be determined accurately. As the result, C −1 should be interpreted
as a generalized inverse and the above system may have to be solved in the
sense of minimizing the fitting error as in Section 2.1.4.
In general, the analysis of covariance matrices of large size using principal
components is an important part of the study of movements of financial
markets. You will learn more about in the course on financial time series.

2.2 MATLAB exercises for Week 3.

2.2.1 Lesley Matrices
Say the growth 
 of a populations of three generations is described by L =
0 2 4
 1/4 0 0 .
0 1/2 0

• Determine the eigenvalues and the corresponding eigenvectors of L. What will

happen with the population, will it become stationary?

• Is it enough to know the total population at time 0 to find the total population
at time t? For example, assume that the population in year 0 is distributed as

29
  
0
 100 . What will be the population in 10 years? What if the population
900  
900
starts with  100 ?
0
• What has the age distribution of these to do with the eigenvector corresponding
to the largest eigenvector?

• Calculate the distribution of population after 1000 years for the above two initial
conditions without calculating L1000 directly.

• What would happen if the 4 in the matrix was replaced by 5 (check eigenvalues)?
Plot the growth of the populations over the period 100 years. Consider using
’semilogy’ command. Analyze the dependence of your answers on the initial
distribution of population.

2.2.2 Markov matrices

Consider the population growth model with the following transition matrix:
 
1/2 0 0 0 0
 0 1/2 0 1/3 1/4 
 
A=  0 0 1/2 0 1/4  .
 0 1/2 0 1/3 1/4 
1/2 0 1/2 1/3 1/4

• What will be the asymptotic distribution of the population?

• Does the asymptotic distribution depend on the initial distribution of population

or just on the total initial population?

• What will happen to the evolution of the population if you replace in the first
column 1/2 and 1/2 by 1/3 and 2/3?

30
2.2.3 Solving equations
 
3 2 −1
Suppose you want to solve Ax = b where A =  −1 3 2  and b =
  1 −1 −1
10
 5  . There are several ways of solving this. For example,
−1
x=inv(A)*b and x=pinv(A)*b
The 2nd method is based on generalized inverses and is generally more suit-
able for calculations involving experimental data. (Type ’help pinv’ into
command window to learn more about pinv.)  
1 1 0
Now suppose you want to solve Ax = b where A =  1 0 0  and
  0 1 0
10
b =  5  . Can you still use both methods? Does A−1 exist? The
−1
solution
x=pinv(A)*b
is based on the least square approach discussed in Sections 2.1.3 and 2.1.4.
Compare the norm of the residual

r = A ∗ x − b.

to the norm of b.

2.3 Ill-conditioned systems: general theory.

We have already considered examples of linear systems Ax = b for which
small relative errors in the matrix A and/or right hand side b lead to large
relative errors in the solution. Such systems are called ill-conditioned. Is
there a simple criterion which would allow us to decide whether the linear
system at hand is ill-conditioned?

Definition. The condition number K(A) of an invertible matrix A is the

31
maximum ratio of the relative error in A−1 b divided by the relative error in
b, where the maximum is taken over all b 6= 0.

It follows from the definition that K(A) >> 1 describes ill-conditioned sys-
tems for which the relative error in the knowledge of the right hand side is
amplified by the factor of K(A). Is it easy to calculate K(A)?

Proposition 2.3.1.
K(A) =|| A || · || A−1 || .
Proof: Let e be the error in the determination of vector b, b′ = b + e.
Then the error in x = A−1 b is A−1 e. Therefore,
def ||A−1 e||/||A−1 b||
K(A) = maxb6=0,e6=0
||e||/||b||
||A−1 e|| ||b||
= maxb6=0,e6=0
||e|| ||A−1 b||
||A−1 e|| ||Ab||
= maxb6=0,e6=0 = ||A|| · ||A−1 ||.
||e|| ||b||

The definition of condition number given above is based on errors in

the determination of the r. h. s. of Ax = b. In fact it also applies to
quantification of errors in the knowledge of matrix A is well: Let A′ = A+δA,
where δA is the error in the definition of A. Then A′−1 ≈ A−1 − A−1 δAA−1
and the system A′ x = b reduces to Ax = b′ , where

b′ = b + e = b − δAA−1 b

Therefore large K(A) signals the enhancement of the relative error in the
solution as the result of errors both in determination of equation matrix A
and equation r. h. s. b.
Using singular value decomposition of matrix A it is easy to show that

K(A) = M ax singular value of A/M in singular value of A.

Therefore, K(A) can be interpreted as the aspect ratio of the ellipsoid {Av |
||v|| = 1}.

32
 For the two examples of ill-conditioned systems already considered in this
Chapter,
 
0.5 5
K ≈ 113 >> 1.
0 0.45
 
1 0.9
K ≈ 19
0.9 1
The beauty of the condition number is that it can be used to correctly
identify an ill-conditioned system irrespective of the precise nature of numer-
ical instability present.
For instance, in the example of Section 2.1.1, we identified the origin of
instability as the near-degeneracy of eigenvectors A.
Here is the example of an ill-conditioned system with a different kind of
instability:
   
100 −200 100
A= and Ax = .
−200 401 −100

The exact solution of this is x = (201, 100)′ . However, if we change A11 by

1% and A22 by 1/4% and take
   
101 −200 100
x=
−200 400 −100

then we get as a solution x = (50, 24.75)′ . This is a huge change of 75%

percent. Why is this? Well, the matrix A is symmetric, so the eigenvectors
are orthogonal and moreover the eigenvalues and the singular values are
the same. So the sensitivity to error is not caused by eigenvectors almost
coinciding. However, K(A) is large, which can be easily verified in Matlab:

A=[100,-200;-200,401]; B=[101,-200;-200,400]; cond(A), cond(B),

eig(A), eig(B)

ans =
2.5080e+003
ans =
625.5009
ans =

33
 0.1997
500.8003
ans =
0.7997
500.2003

The smallest eigenvalues of A and B differ substantially in size which

explains the difference in two estimates of the solution.
A simple rule of thumb is that for a system with condition number K(A)
one expect a loss of roughly log10 K(A) decimal places in the accuracy of the
solution. This can understood as follows: Consider the equation Ax = b.
Let 10−nrhs be the estimate of the relative error in b (i. e. b is accurate to
nrhs − 1 decimal places. Let 10−nsol be the estimate of the relative error of
the solution x (i. e. x is accurate to nsol − 1 decimal places. By definition of
K(A),
10−nsol
K(A) ∼ −n .
10 rhs
Therefore,
nsol = nrhs − log10 (K(A)),
as claimed.
We established that the large value of K(A) means that the system at
hand is ill-conditioned. We also found a simple formula for K(A).
The question of how to study an ill-conditioned system numerically is
system-dependent. As we demonstrated above, an SVD based algorithm or
the use of generalized inverses can stabilize the numerical algorithm. Some-
times, a simple transformation of the system can be useful. For example,
before attempting to solve
   
−1 1 1
x= ,
10000 20000 1000

one should re-scale the 2nd row by a 1000 to make it the same order of
magnitude as the first one.

34
2.4 Numerical computations with matrices
Matlab has very good routines for solving systems of linear equations. If you
need to code one of these routines in C, check for example the book by Press
at al. This is much safer than writing your own code. However, in order
to choose the right code for a given ill-conditioned system it is important to
have a general idea about the underlying algorithms.

Gauss Elimination Algorithm is a process of row eliminations aimed

at rewriting the matrix in upper-diagonal form for which the equation can
be easily solved. Here is an example:

(i) Row elimination.

L1 = x1 + 2x2 + 3x3 = 1
L2 = 2x1 + 2x2 + x1 = 1
L3 = 3x1 + 2x2 + x3 = 1

L1 = x1 + 2x2 + 3x3 = 1
L2 − 2L1 = −2x2 − 5x3 = −1
L3 − 3L1 = −4x2 − 8x3 = −2

L1 = x1 + 2x2 + 3x3 = 1
L2 − 2L1 = −2x2 − 5x3 = −1
L3 − 3L1 − 2(L2 − 2L1 ) = 2x3 = 0

(ii) Backward solution:

x3 = 0,
x2 = 1/2,
x1 = 0.

35
Some general remarks:

(i) It takes approximately 2n3 /3 additions/multiplications to solve an n × n

system using Gaussian elimination. Really, to eliminate variable x1 we need
(2n + 2)(n − 1) operations. Hence the total number of operations is
n
X
(2k 2 − 2) = 2/3n3 + lower order terms in n.
k=1

(ii) Gaussian elimination is numerically stable for positive definite or diago-

nally dominant matrices.

The notion of diagonally dominant matrices is important:P

Def. A matrix A is diagonally dominant if |(A)ii | > j6=i |(A)ij | for all row
indices i.2
Jacobi Algorithm. Gauss elimination algorithm is exact, but expensive
for matrices of large size. Iterative algorithms can often find a good approx-
imation to the solution at a lower cost. Jacobi algorithm is based on the
following elegant idea: solution is obtained as the fixed point of the recursion
derived from the original equation. Namely, to solve Ax = b, let us re-write

A = L + D + U,

where L, D and U are lower-triangular, diagonal and upper triangular com-

ponents of matrix A. Then

Dx = (−(L + U )x + b)

Consider the following recursion:

xk+1 = D−1 (−(L + U )xk + b) , k ≥ 0. (R)

We have the following:

Proposition 2.4.1. Let A be a diagonally dominant matrix. Then recursion
(R) converges to the unique solution of the equation Ax = b in the limit
k → ∞ for any initial vector x0 .
2
Some textbooks refer to such matrices as strictly diagonally dominant as opposed to
weakly dominant matrices defined using ”≥” rather than the strict inequality.

36
 Proof:
By Levy−Desplanques theorem diagonally dominant matrix A is non-
singular. Therefore, Ax = b has a unique solution. Clearly, x is the fixed
point of the recursion (R). Let xk = x + δxk . Then

δxk+1 = −D−1 (L + U )δxk .

Matrix A is diagonally dominant. Consequently, as it is easy to show, all

eigenvalues of matrix T = −D−1 (L + U ) lie inside the unit circle. Hence
δxk = T k δx0 → 0 as k → ∞ in virtue of Theorem 1. 1. 1. Hence xk → x as
k → ∞ for any initial vector x0

Jacobi algorithm consists of running recursion (R) starting with some initial
condition until sufficient convergence has been achieved. Note that the com-
plexity of one iteration step is approximately 2n2 for n >> 1. Given that
the number of iterations required to achieve the desired accuracy is much
smaller than n, Jacobi algorithm costs much less than Gaussian elimination
algorithm. It is clear that the rate of convergence is determined by the largest
eigenvalue of matrix T introduced above.

Example. Consider

4x − y + z = 7
4x − 8y + z = −21
−2x + y + 5z = 15

This can be rewritten as

x = (7 + y − z)/4
y = (21 + 4x + z)/8
z = (15 + 2x − y)/5

Then taking
xk+1 = (7 + yk − zk )/4
yk+1 = (21 + 4xk + zk )/8
zk+1 = (15 + 2xk − yk )/5
and any choice for (x0 , y0 , z0 ) we get that (xk , yk , zk ) is very close to the
solution (2, 4, 3) for k large. We have sufficient theoretical background to

37
explain why Jacobi algorithm converges for the case at hand: the matrix
 
0 1/4 −1/4
 4/8 0 1/8 
2/5 −1/5 0

has eigenvalues within the unit circle as the original matrix

 
4 −1 1
 4/8 −8 1 
−2 1 5

is clearly diagonally dominant.

Let us make the final remark about the choice of the initial condition
for the recursion (R). Jacobi algorithm works well for matrices close (in the
precise sense explained above) to diagonal matrices. Therefore, it is natural
to choose
x0 = D−1 b
as the initial approximation to the solution.

Gauss-Seidel algorithm is another iterative algorithm for solving systems

of linear equations. It is a modification of Jacobi algorithm which immedi-
ately uses the updated components (xk+1 )1 , (xk+1 )2 . . . (xk+1 )m to calculate
(xk+1 )m+1 (xk+1 )m+2 . . .. The aim of this trick is to improve the speed of
convergence of the recursion (R).
Let is illustrate Gauss-Seidel algorithm using the previous example: The
recursion is
xk+1 = (7 + yk − zk )/4
yk+1 = (21 + 4xk+1 + zk )/8
zk+1 = (15 + 2xk+1 − yk+1 )/5
It differs from Jacobi recursion by replacing xk by xk+1 in the second equation
and xk , yk by xk+1 , yk+1 in the third equation.
In general, Gauss-Seidel recursion is given by

Dxn+1 = −Lxn+1 + b − U xn

As before, this recursion is guaranteed to converge to the solution of Ax = b

if A is diagonally dominant.

38
2.5 Matlab exercises for week 4 (Linear Al-
gebra)
2.5.1 Ill posed systems
 
10 7 8 7
 7 5 6 5 
Suppose you want to solve Ax = b where A = 
 8
 and
6 10 9 
7 5 9 10
 
32
 23 
b= 33  . Solve for x using for example


31
x=A \ b or x=inv(A)*x
 
32.1
 22.9 
What would the answer be if b was changed to b′ =   32.9 ? Our aim


31.1
is to explain the reason for large difference between the two solutions corre-
sponding to a small relative change of b. Compute in both cases the residual
Ax − b. Notice that these residuals are very small compared with norms of
Ax and b. Therefore, we are in the situation when the problem Ax ≈ b is ill
posed.

• What is the condition number of A?

• What are the singular values of A?

• Confirm that the condition number of A is the ratio of the largest and the smallest
singular values of A. What is the geometric meaning of the largest and of the
smallest singular values?

• What is the ratio of the relative error in the solution to the relative error in b?
What is the maximal error magnification factor? How well do these two numbers
agree to each other? (Take the decimal logarithm of the numbers.)

39
2.5.2 MATLAB capabilities investigation: sparse ma-
trices
When solving PDE’s numerically one often has to deal matrices with lots of
zeros, but of very high dimension. Such matrices are called sparse. Matlab
has commands for efficient operations with sparse matrices
• What does the matlab commands ’sparse’ and ’full’ do? (Try ’help sparse’ to
find out. Use ’sparse(eye(n))’ for sufficiently large integer n as an example.)
• Investigate ’sparse’ and ’full’ commands a bit deeper by running through the fol-
lowing sequence of commands pausing regularly to analyze intermediate results:

n=6
format rat
P=sparse(1,1,1,n,n)
full(P)
Q=sparse(n,n,1,n,n)
R=P+Q
full(R)
S=sparse(1:n-1,2:n,1/3,n,n)
full(S)
T=P+S
full(T)

2.5.3 Solving Ax = b by iteration

Create a MATLAB function which solves a linear system Ax = b using Jacobi
iteration scheme from the notes. Organise your function m-file as follows:
function x=jacobi(A,b,xstart)
%%% This is the Jacobi iteration scheme
%%% to solve AX=b
%%% using the starting point xstart
%%%%%%%%%%%%%%%%%%%%%%%%%%
Put your iteration here
%%%%%%%%%%%%%%%%%%%%%%%%%%
end

• Check that your code works for the matrix used as an example in the lecture.

40
 • Find an example of non-diagonally-dominant matrix and an initial condition for
which the iteration does not converge.

• Create another MATLAB function for solving a linear system using Gauss-Seidel
iteration scheme.

You may need to hand in these codes at the end of the term.

41