Matrix Theory

A.Holst, V.Ufnarovski
15.5. HINTS AND ANSWERS 291

15.5 Hints and answers

1.1. There are two different approaches. In the first one write A as a block
of rows and note that in B = Eij A all rows different from i are equal to zero.
What is in the row i? For C = AEij one need P to consider the columns instead.
In the second approach write A as A = k,l akl Ekl and use Eij Ekl = jk Eil .
Note that for the equality B = C we need almost all elements in those
matrices to be equal zero (the element on the intersection of the row i and
column j is the only possible exclusion).
Answer: for equality one need ajl = 0 for l 6= j; aki = 0 for k 6= i and
ajj = aii .
1.2. It is sufficient to use X = Eij and apply the previous exercise.
1.3. Example was in the section 1.7. V AB = O B = O the contradic-
tion.
1.4. Use the previous exercise taking B C as B.
1.5. The approach as (A1 , . . . , An )B = (A1 B, . . . , An B) is wrong, because
the product Ai B does not exist if n > 1. P P
Consider instead the first column.PIts ith element is k aik bk1 = k bk1 aik
thus the column can be written as k bk1 Ak , which is a linear combination of
columns Ak in A.
1
1.6. Study instead A(A + B)1 B and use (XY Z)1 = Z 1 Y 1 X 1 .
1.7. One approach for the inverse can be found in the section 1.8. Another
approach: write the inverse X in the triangle form with unknown variables,
consider AX = I as a system of equations and show that it has a solution.
1.8. Try to find the inverse in the form I + dXY T and use that Y T X is a
number that can be moved free.
c T
Answer: d = 1+cY T X if 1 + cY X 6= 0, otherwise the inverse does not exist
(which is less trivial and most simple can be proved using the determinant in
the next section).

1 1
1.9. Write this as I + cA1 XY T A and use the answer from the
previous exercise.
1.10. The diagonals in the product of the triangular matrices multiplies by
themselves and to get zero in some power it should be zero itself. Two different
proofs that the triangular matrix with zero main diagonal is nilpoten can be
found in the section 1.7.
1.11. Find the inverse in the form
 1 
A X
.
O B 1

Answer: X = A1 CB 1 .
1.12. Use that b is a number that can be moved free. Use b(Y T A1 X =
ba 1 to simplify the calculations.
For the (upper) triangular matrix use that Y = O and apply the induction.
292

2.1. Basic variables: x2 , x3 , x5 , thus r(A) = 3, free variables: x1 , x4 . No

left or right inverse because r(A) 6= 4, 5. If b4 = 0 then x5 = b3 , x4 = t, x3 =
b2
2 2b3 , x2 = 2t 3b3 + b1 , x1 = s, otherwise no solutions.
2.2. Try to use a permutation matrix P that moves the zero row direct to
the end (and it works here).
One possible decomposition is:

0 1 0 0 1 0 0 0 0 1 3 3 2
0 0 1 0 2 1 0 0 0 0 0 3 1
P = 0 0 0 1 , L = 1 2 1 0 , U = 0 0 0 0 0 .

1 0 0 0 0 0 0 1 0 0 0 0 0

Basic variables: x2 , x4 , thus r(A) = 2, free variables: x1 , x3 , x5 . No left or right

inverse because r(A) 6= 4, 5.
Because AX = O P AX = O LU X = O U X = O we get the
solutions: x5 = t, x4 = 31 t, x3 = s, x2 = 3s t, x1 = u.
2.3. For example, x + y + z = 0, x + y + z = 1.
2.4. For example,
     
0 1 1 0 i 0
P = ,L = ,U = .
1 0 0 1 0 i

2.5. Because we do not need P in LU decomposition we get

1 0 0 1 1 0
A = LU = 2 1 0 0 1 2 .
1 1 1 0 0 1

2.6. In one direction: the product of the invertible matrices is invertible

itself.
The opposite direction is more complicated. First show that the permutation
matrix which permutes two rows only is an elementary matrix. Second prove
that any permutation matrix is a product of such matrices. Third prove that
invertible diagonal matrix with a single element different from 1 is an elementary
matrix. Prove that any invertible diagonal matrix is a product of such matrices
and apply Exercise 2.8 to finish the proof.
2.7. Find an LU decomposition and find first all possible inverses to U.
Answer: left inverse  
3 1 a a
1 b .
1 2 2 b
2.8. Start from LU decomposition and continue in the following way, work-
ing with the columns in U. First multiply by a permutation matrix Q from the
right to get all pivot elements on the main diagonal and all basic columns in
the beginning. Second use elementary operations with the columns to clean the
rest of the rows after the pivot elements. This can be done by the multiplication
with elementary matrices from the right hand side. They create R.
15.5. HINTS AND ANSWERS 293

3.1. Follows directly from the formula for the determinant on the page
40The permutation matrix P(1,2) corresponding to the transposition (1, 2) is a
possible counterexample for n > 3..
3.2.
 
 1 + x1 y1
 x1 y2 x1 yn 
 x2 y1 1 + x2 y2 x2 yn 
 ...................................  =

 
 xn y1 xn y2 1 + xn yn 
   
 1
 0 0   x1 y1
  x1 y2 x1 yn 
 x2 y1 1 + x2 y2 x2 yn   x2 y1 1 + x2 y2 x2 yn 
 ...............................  +  ............................... .

   
 xn y1 xn y2 1 + xn yn   xn y1 xn y2 1 + xn yn 
In the first determinant use the expansion in the first row and induction, in the
second carry out x1 and use the remaining in the first row to clean the rest.
3.3. In one direction: use 1 = det A det A1 . In the opposite direction use
theorem 3.16.
3.4. 26 .
3.5. For odd n : use det(A) = (1)n det A. For n = 4 use direct calcula-
tions to get:  
 0 a b c 

 
 a 0 d e 
=


 b d 0 f 
 
 
 c e f 0 
2
a2 f 2 + 2 adf c 2 aebf + b2 e2 2 bedc + d2 c2 = (af + dc be) .
For arbitrary n = 2k one can prove that |A| = (P fk )2 , where the Pfaffian P fk
is defined as
 
X 1 2 3 4 2k 1 2k
sign ai1 j1 ai2 j2 aik jk ,
i1 j1 i2 j2 ik jk

where the sum extends over all possible distinct terms subject to the restric-
tion 1 is < js 2k, 1 s k.
3.6. Expand in the last column.
3.7. If C = O then this is theorem 3.18. Try to reduce the general
 case to 
I O
this one by the multiplication from the left with the block matrix
X I
choosing the correct block X.
3.8. Use the previous exercise!
294

4.1. A standard way toPget the basis vectors Xi is to solve the system
AX = O and to write X = ti Xi , were ti are free variables. But there exists
other ways.
Answer: Endast a). One possible basis: (2, 1, 0)T ; (1, 0, 1)T .
4.2. Use the hint from the exercise 3.1 to get a basis. One possible basis:
(0, 1, 0, 31 , 1)T ; (0, 3, 1, 0, 0)T , (1, 0, 0, 0)T .
4.3. To find a basis one need to know from calculus that every solution
can be written as y(t) = C1 et + C2 tet + C3 t2 et . Without such knowledge use
substitution y(t) = x(t)et and get the equation for x(t) which is easy to solve
directly.
Answer: one possible basis is et , tet , t2 et , thus the dimension is equal to 3.
4.4. Try to express the basis with the help of Eij .
Answer: a) The basis Eij Eji with i < j gives the dimension n(n1) 2 .
b) The basis Eij Ein with j < n gives the dimension n(n 1).
c),d),e) are not vector spaces.
15.5. HINTS AND ANSWERS 295

5.1. a) is not linear. If the basis is e1 = E11 , e2 = E12 , e3 = E21 , e4 = E22

then the answer is:

1 0 1 0 1 0 0 2 1 0 0 0
1 1
0 1 0 1 0 0 0 0 0
2 2 0
b)
0 0
, c) , d) 1 1
.
1 0 0 0 0 0 0
2 2 0
0 0 0 1 1 0 0 2 0 0 0 1

5.2. Use the Gauss elimination or LU decomposition. Use the hint from
4.1 for the kernels. Answer: r(A) = 2. Possible bases are:
for Ker A : (1, 0, 0, 0, 0)T , (0, 3, 1, 0, 0)T , (0, 3, 0, 1, 3)T .
for Im A : (0, 1, 2, 1)T , (0, 3, 9, 3)T .
for Ker AT : (1, 0, 0, 0)T , ((0, 5, 2, 1)T .
for Im A : (0, 1, 3, 32)T , (0, 0, 0, 3, 1)T . The second vector is not a row in A,
but the second non-zero row in U. Try to understand why the non-zero rows in
U can be used as a basis for AT , though it does not work for the columns.
5.3. The basis 1, x, . . . , xn shows that the dimension is equal to n + 1. In
this basis the matrix for D : Pn Pn looks as

0 1 0 0 0
0 0 2 0 0

0 0 0 0 0

........................

0 0 0 n 1 0

0 0 0 0 n
0 0 0 0 0

The rank is n 1 and we have neither left nor right inverse. If we take away the
last row we get the matrix for D : Pn Pn1 and find that the right inverse
should exist and of course this is the integration.
5.4. The map has a nontrivial kernel and cannot be injective. This example
shows that the situation for the linear operators in the infinite dimensional case
is more complicated: injectivity does not follows from the surjectivity.
5.5. a) 2. b) 2. c) 3.
5.6. See theorem 5.6.
5.7. See theorem 5.7.
5.8. Try to imitate the dimensions arguments used for the Lagrange inter-
polation. Later we will prove more general case - see Hermite interpolation.
296

6.1. See theorem 6.1.

6.2. Use that pA (0) = |A|. Another approach: according to theorem 5.9
a square matrix is invertible if and only if Ker A = O AX = 0X has only
trivial solutions.
6.3. For example 2 2 matrices E11 and E22 with = = 0.
6.4. As example one can take a triangular matrix with ones on the main
diagonal.
6.5. AX = X A2 X = 2 X (2 )X = 0, but X 6= O = 0, 1.
6.6. Use AX = X Ak X = k X. Answer: = 0.
6.7. Use that det(I) = (1)3 det I = 1. Examples: iI and E12 E21 .
6.8.
1. A1 ABA = BA.
2. Corollary 6.7.
3. Suppose that A is not invertible. Consider A for very small 6= 0.
According to Exercise 6.2 the matrix A is invertible and therefor A B
and BA have the same characteristic polynomial pA B = pBA . Let 0
and get the result using the fact that the characteristic polynomial depends
continuously on the matrix.
The disadvantage of this solution is that it uses continuous arguments that
does not work for other fields (e.g. finite fields). There exists two other
solutions that works for arbitrary field. In both of them it is sufficient to
consider the cases where both matrices are not invertible (otherwise AB
and BA are similar).
In the first approach we use Im A and Im B. Note that both this subspaces
are different from Kn , where n is the size of matrices.
If one of them is inside another (e.g. Im A Im B) then we can choose
the basis in the largest one and complete it arbitrary. Then both images
lies in the subspace generated by e1 , . . . , en1 and the matrices looks as
block matrices
   
A1 A2 B1 B2
A= ,B = ,
O 0 O 0

where A1 , B1 have size (n 1) (n 1), and A2 , B2 are columns of the

size n 1. Then
 
A1 B 1
AB = , pAB () = pA1 B1 ();
0 O
 
B 1 A1
BA = , pBA () = pB1 A1 ()
O 0
and we can use the induction.
T
Otherwise dim (Im A Im B) n 2 and we can we choose the T basis {ei }
starting from this intersection and complete it such that Im A Im B
15.5. HINTS AND ANSWERS 297

he2 , . . . , en1 i, Im A he2 , . . . , en i and Im B he1 , . . . , en1 i. Then the

matrices looks as block matrices
   
O 0 B2 B1
A= ,B = ,
A1 A2 0 O

where A1 , B1 have size (n 1) (n 1), and A2 , B2 are columns of the

size n 1. Then
 
0 O
AB = , pAB () = pA1 B1 ();
A1 B 1
 
B 1 A1
BA = , pBA () = pB1 A1 ()
O 0
and we can use the induction.
Another approach uses the kernels instead of the images try to find it!

4. Example: A = E11 , B = E21 , in fact any A, B, where AB = O, BA 6= O.

5. Note that |I C| = PC (1).
6.9. Use Exercise 3.8 to get pM () = ( 1)n2 (2 (d T
 + 1) C B).
p
Eigenvalues are = 1 and = 21 d + 1 (d 1)2 + 4C T B .
298

7.1.

a) E1 , E2 , E3 , E4 , E5 , E8 .

b) E4 , E7 , E8 .

c) E4 , E8 .

d) One can cut rows and columns with indices 6, 7, 9 and get

4 1 0 0 0 0
0 4 0 0 0 0

0 0 2 0 0 0
.
0 0 0 0 1 0

0 0 0 0 0 0
0 0 0 0 0 0

A4I A4I A2I A A A

e) O E1 E2 ; O E3 ; O E4 E5 E6 ;
A A A
O E7 ; O E8 E9 .

7.2. In one direction it is exercise 6.6. For another direction: it is sufficient

to prove this for the Jordan form. In fact much more is proved in theorem 7.11.
7.3. Direct approach: use SJS = J T , S = S 1 for

0 0 1
0 1 0
S= ............. .

1 0 0

See corollary 7.9 for another approach.

7.4. According to Exercise 7.3 it is sufficient to prove that if Ai are similar
to Bi then the block matrices

A1 O O B1 O O
O A2 O O B2 O
.................. , ..................


O O Ak O O Bk

are similar as well.

7.5. Both matrices have different eigenvalues and are digonalizable. Note
that B has rank 1. Answer:

1 0 0  
0 3 0 , 0 0
.
0 5
0 0 6

7.6. Yes, they have the same (diagonal) Jordan form.

15.5. HINTS AND ANSWERS 299

1 1 0 1 1 1
7.7. For example, J = 0 1 1 ,S = 0 1 0 for A,
0 0 1 1 1 0

0 0 0 0 1 1
J = 0 0 1 ,S = 0 2 1 for B and for C we have
0 0 0 1 1 1

0 0 0 0 0 1 1 0 3 0
0 1 1 0 0 1 0 1 0 0

J = 0 0 1 0 ,S = 1 0 1
0 0 1
.

0 0 0 1 0 0 0 1 0 1
0 0 0 0 1 0 0 0 2 0
7.8. Study the case n = 2 first. For general case conjugate first with
1
the permutation matrix
 S = P = (E11 + E2n + E32 +  E4n1 + ) for the
1 2 3 4 5 6
permutation = to get a block matrix.
1 n 2 n 1 3 n 2
Consider the cases with even and odd n separately.
More details: S 1 AS =
T
1 0 0 0 0 0 0 a1 1 0 0 0
0 0 0 1 0 a2 0
0 0 0 0 1


0 1 0 0 ....................... 0 1 0 0 =

0 0 1 0 0 an1 0 0 0 0 1 0
................ an 0 0 0 ................

0 a1 0 0
A1 O O
an 0
0 0
O A2 O
0
0 0 a2 = .................. ,

0 0 an1 0
O O Ak
.......................
where n = 2kor n = 2k 1. If n = 2k then all blocks looks as Ai =

0 ai
, but if n = 2k 1 then the last block Ak is different it
ani+1 0
is 1 1 block (ak ). For the block Ai of size 2 2 there are three possibilies. If
ai = ani+1 = 0 then the Jordan form is O. If exactly one of ai , ani+1 is equal
to zero, then the Jordan form is J2 and if both are
different from zero then the

the Jordan form is diag ai ani+1 , ai ani+1 .
7.9.The Jordan form is

2 1 0 0 0
0 2 1 0 0

J = 0 0 2 0 0 .

0 0 0 2 0
0 0 0 0 2

7.10. The Jordan form should be idempotent as well. But Jn ()2 = (I +

Jn )2 = 2 I + 2Jn + Jn2 cannot be equal to Jn () if n > 1.
300

7.11. Use that multiplication with D from the left multiplies the rows by i
and from the right the columns. Compare the result outside the main diagonal.
7.12. Use the previous result.
15.5. HINTS AND ANSWERS 301

8.1. a) x 1; b) (x 1)(x 2) (x n).

8.2. We have pA (x) = (x 0)n = xn , thus An = 0.
8.3. Use that the minimal polynomial divides both xk and the characteristic
polynomial if Ak = 0. Thus it has form xm and m n.
8.4. A (x) = (x 1)2 if 6= 0 and A (x) = (x 1) if = 0 thus it does not
depend continuously. On the other hand the characteristic polynomial has the
same degree and (as the determinant) depends continuously on the matrix.
8.5. For example E12 and the zero matrix O of the same size.
8.6. Study the Jordan form. Answer: for n 3. Counterexample for n = 4 :

0 1 0 0 0 1 0 0
0 0 0 0 0 0 0 0
0 0 0 1 , 0 0 0 0 .

0 0 0 0 0 0 0 0

8.7. X should be nilpotent, but X 2 = O according to exercise 8.3.

8.8. X is nilpotent, thus dim Ker X 2 2, and X 2 cannot have rank n 1.
8.9.The matrix looks as

0 0 0 0 0 a0
1 0 0 0 0 a1

0 1 0 0 0 a2

A= 0 0 1 0 0 a3 .
............................

0 0 0 0 0 an2
0 0 0 0 1 an1

Note that Ai E1 = Ei+1 if i < n and

An E1 = AAn1 E1 = AEn = a0 E1 a1 E2 an1 En

p(A)E1 = An E1 + an1 An1 E1 + . . . + a1 AE1 + a0 IE1 = O,

which gives p(A)Ei+1 = p(A)Ai E1 = Ai p(A)E1 = O as well, thus p(A) = O.
On the other hand q(A) 6= O if deg q < n, because q(A)E1 6= O. This shows
that p(x) is the minimal polynomial.
Thus every polynomial xn + an1 xn1 + . . . + a0 is the minimal polynomial
for some matrix A.
8.10. Use ideas from exercise 6.8. If AB = O, but BA 6= O this is already
a good counterexample.
302

T
9.1. The map p(x) (p(2), p(1), p0 (1), p00 (1)) has the matrix

1 2 4 8

1 1 1 1
A= .

0
1 2 3

0 0 2 6

Because
1 26 8 1
27 27 9 3
1
19 2 1

A1

= 9 3 2 ,
1


9 19 13 0

1 1
27 27 19 61
we have:
1 1
(1 + 3x + 3x2 + x3 ); L20 (x) = 26 3x 3x2 x3


L10 (x) =
27 27
1 1
(8 + 6x 3x2 x3 ); L22 (x) = 2 + 3x x3 .


L21 (x) =
9 6
The polynomial corresponding to etx is

p(x) = e2t L10 (x) + et L20 (x) + tet L21 (x) + t2 et L22 (x) =
  
1 2t t t 2 t

2t t t 27 2
e + 26e + 24te + 9t e + 3e 3e + 18te + t x+
27 2
  
9
3e2t 3et 9t2 et x2 + e2t et 3tet t2 et x3 .


2
1
e2t + 26et
+ 24tet + 9t2 et I + 3e2t 3et
+ 18te t 27 2



etA = 27 + 2 t A
+ 3e2t 3et 9t2 et A2 + e2t et 3tet 92 t2 et A3 .

9.2. Note that P is a permutation matrix. Compute P 2 , P 3 . . . and get that

A = a1 I + a2 P + a3 P 2 + P+ an P n1

.
1
9.3. Because (1+x) = k=0 k xk for |x| < 1 the function f (x) = (1+x) 2

is defined on A if eigenvalues of A satisfies || < 1 and besides that f (A)2 =

I + A. This definitely works for A = N.
9.4. No, because f 0 (x) = 2 1
x
is not defined for x = 0. Nevertheless the
2
equation X = N can have solutions for some nilpotent matrices N .
9.5. If i are all eigenvalues for A then ei are all eigenvalues for eA (which
is evident for the Jordan form). It remains to apply theorem 7.10.
9.6. If all eigenvalues i of tA satisfy |i | < 1 then the function f (x) =
P k
ln(1 x) = k=1 xk is defined on tA. Thus there exists B = f (A) such that
I tA = eB . It remains to use the Jacobi identity above for B.
15.5. HINTS AND ANSWERS 303

11.1.kXk 2
1 = 3 + , kXk2 = 5 + , kXk = for the first vector and
kXk1 = 2 + 13 + 4, kXk2 = 31, kXk = 4 for the second one. p

11.2. kAk(1) = 10, kAk(2) = 30, kAk() = 4, kAk1 = 6, kAk2 = 15 + 221,

kAk = 7 for the first matrix and kAk(1) = 6+ 13+ 17, kAk(2) = 56, kAk() =
p
5, kAk1 = 5 + 13, kAk2 = 28 + 18 2, kAk = 5 + 17 for the second one.
11.3. kAk(1) = 2 + , kAkF = 2 + 2 , kAk() = max(1, ), kAk1 = 1 +
r q q 2
2 2 2
, kAk2 = 1+ 2 + 1+ 4 , kAk = 1 + , (A) = 1 + 2 + 1 + 4 .
P
11.4. p IfP|xk | = max{|x
p i |} thenkXk1 = |xi | n|xk | = nkXk and
kXk2 = |xi |2 n|xk |2 = nkXk . Besides that kXk = |xk | =
2
p pP
|xi |2 = kXk2 . The obvious inequality P (|xi |2 ) P
P P
|xk |2 ( |xi |)
pP kXk2 kXk
gives 1 . At last the inequality kXk1 = |xi | = |xi | 1
|xi |2 n = nkXk2 follows from the Cauchy-Schwarz inequality.
11.5. Use that kAk2F = tr PA A.
H
2
11.6. Use that kAkF = i , where i are eigenvalues of AH A to estimate
the Frobenius norm. The exercise 11.4 explain how the root expressions appears.
Use kAk2 |AEi | |aij | to get kAk() kAk2 and 1m kAk1 kAk2 . Use

kAk2 kAkF and kAk F mnkAk() (also following from the the exercise
11.4 ) to get kAk2 mn kAk() . P P
If |AX| =kAk2 with |X| = 1 then |AX| |xi | |AEi | |xi |kAk1 =
kXk1 kAk1 n|X|kAk1 gives kAk2 nkAk1 .
The shortest way to the last pair of inequalities is to use that A and AH have
the same norm, because they have the same singular values, but this we will
prove much later. Otherwise consider a vector Y with kY k() = 1 such that
kAY k() = kAk() and estimate |Y | and |AY | and kAXk() using exercise
11.4.
11.7. x1 = 0, x2 = 0.001 in the first case and x2 = 1, x2 = 0 in the second.
The relative error is kXk 1+0.001 kBk1 0+0.001
kXk1 = 0.001 = 1001, though kBk1 = 1+0.001 0.001
1

is rather small.
11.8. see the section 11.4.
11.9. The second statement follows directly from the first because A +
B = A(I + A1 B). Use the spectral radius of A to show that no eigenvalue of
I + A are equal to zero. Alternative approach: consider the geometric series
I A + A2 A3 + and show that it converges and gives (1 + A)1 .
304

12.1. See section 12.3.

12.2. All axioms can be checked directly, e.g. (v|v) = 0 (Av, Av) = 0
Av = 0 v = 0, because A is invertible.
12.3. G = AH A. The rest follows from theorem 11.6.
12.4. See theorem
 5.3.   
1 5 12 1 169 129
12.5. Q = 13 , R = 13 .
12 5 0 73
12.6. Two possible decompositions are:
     
1 0 0 1 0 1 0 2
A= = .
0 1 0 2 1 0 0 1

12.7. A = cXX H , for some complex column vector X of length 1 and real
constant c.
Such matrix is obviously Hermitian. To get that this is the only choice note
that by theorem 12.11 A = U DU H for a diagonal matrix D of rank 1. If c = dii
is the only non-zero element in D then X = U Ei .
Another longer solution is to write A as A = XY H . Consider first the case
when |X| = |Y | = 1 and use that A = AH = Y X H to get |xi |2 = |yi |2 . Put
xi = i yi and get more from AH = A.
15.5. HINTS AND ANSWERS 305

13.1. Nontrivial is that AAH and AH A can have different size. Use theorem
13.3. If A = U SV H is the SV decomposition, then AAH = U S 2 U H is also a
SV decomposition. What about
AH A?
1i 3
13.2. Eigenvalues: a) 2 ; b)1; c)0, 1; d)0, 1 + i.
q
Singular values: a),b) 32 5 ; c) 2; d)2.
13.3. Use that the Frobenius norm does not change after the multiplication
with a unitary matrix.
13.4. If A = U SV H is the SV decomposition then it is sufficient to show
that A and S are the matrices of the same linear map in different orthonormal
bases.
13.5. First, using the SV decomposition reduce the problem to the case
A = S, thus suppose that all singular values are on the main diagonal and the
rest is zero. Second, we need to find a matrix of rang k such that kA Xk2 =
k+1 this is easily done, replace all j in S with j > k by zeros.
The last and most difficult part is to show that for any matrix of rang k we
have kA Xk2 k+1 . Here we use the dimensions arguments. Consider two
subspaces: Ker X and U , consisting of vectors having last n k 1 coordinates
equal to zero, i.e. of form v = (v1 , . . . , vk , vk+1 , 0 . . . , 0)T . Because dim Ker X +
dim U nk +(k +1) > n we should have a non-zero r vector v in their intersection.
Pk+1 2
|(AX)v| |Av| i=1 (i vi )
For this vector we have |v| = |v| = P k+1 2
k+1 .
i=1 (vi )

= AAT + 1. Eigenvalues are 1 (with the

13.6. Let A = (a0 , . . . , an1 ), b
b b2 4a2
multiplicity (n 2)) and 1,2 = 0
. The singular values are 1 (n 2

2
b+ b2 4a20
times) and 1 , 2 . The condition number is 2|a0 | .
13.7. Use that 1+k +(k )2 + +(k )n1 = 0 (consider this as a geometric
sum). This gives F H F = I. If F = (F1 , . . . , Fn ) is a column representation then
P Fj = j Fj , thus P F = F diag(1, , . . . , n1 ) = F D and F 1 P F = D.
Note that Circ(x1 , . . . , xn ) = x1 I + x2 P + xn P n1 = f (P ) is a poly-
nomial in P. Because F diagonalizes any polynomial in P as well we have
F 1 Circ(x1 , . . . , xn )F = diag(f (1), f (), . . . , f (n1 )). Thus f (j ) are eigen-
values and their product is the determinant.
13.8. Combine Schurs lemma with theorems 12.7 and 13.7.
13.9. Use theorem 13.8.
13.10. Non-zero nilpotent matrices are not diagonalizable and therefore are
not normal.
13.11. Matrices in a) and b) are Hermitian (and normal). Their eigenvalues
are real. Matrix in c) is skew-Hermitian (and normal). Its eigenvalues are pure
imaginary. Matrices in d), e) are normal (direct control of the definition), but
their eigenvalues are nether real nor imaginary. Matrix in f) is not normal.
13.12. No.
13.13. Because A is real its characteristic polynomial has real coefficients.
Because a real polynomial of degree 3 has at least one real root we get that
at least one eigenvalues is real. Because A is unitary we have |i | = 1 for all
eigenvalues. Thus if all eigenvalues are real they are equal to 1 and two of then
306

are equal which give us = 0 or = . Otherwise one of eigenvalues is equal to

= 1 and two other are complex. Let = ei is one of them. AX = X gives
AX = X thus if X = U1 + iU2 then AU1 = 21 A(X + X) = 21 (ei X + ei X) =
1
2 (cos + i sin )(U1 + iU2 ) + (cos i sin )(U1 iU2 ) = cos U1 sin U2 .
AU2 = sin U1 + cos U2 can be obtained in a similar way. What if U1 = 0?
13.14. Use the previous exercise. Use that det A is the product of eigenvalues
and tr A is their sum.
13.15. Prove more that the exponent of a skew-symmetric matrix is
unitary. For this consider this first matrices of size 1 1, then the diagonal
matrices and use that normal matrices are diagonalizable.
13.16. Use the hint from the previous exercise.    
0 1 1 0
13.17. a),c) No. Common counterexample: A = ,B = .
1 0 0 0
b)Yes: (AB)H (AB) = B H AH AB = B H B = I.
15.5. HINTS AND ANSWERS 307

4.1. a) No. For example, A = E11 E22 , B = E12 + E21 . b) Yes. Consider
first the case when A = C 2 , where C is a diagonal matrix with positive elements
on the diagonal.
More details: we have that AB = C 2 B is similar to C 1 C 2 BC = CBC. But
C = C H thus CBC = C H BC is congruent to B (which is positive definite) and
therefore is positive definite as well and therefor has positive eigenvalues. For
general case let U H AU = D be a diagonal matrix for some unitary matrix U.
Then U H BU = B1 is positive definite as well and DB1 = U H ABU is similar
to AB. But D = C 2 because A is positive definite and we can use the previous
case.
14.2. In the Gaussian elimination we get d1 = d2 = 1, d3 = 2. All numbers
are
 positive
 and the answer is yes. Alternatively, 1 = a11 = 1 > 0, 2 =
 1 1 
 1 2  = 1 > 0 and 3 = det A > 0 thus the matrix is positively definite.
 

14.3. Consider the determinant of U H AU = D.

14.4. Multiply the matrices and get the equation for every block. Answer:

H
B11 = A11 , X = A21 A1 1
11 , Y = A11 A12 ; B22 = A22 A21 A1 11 A12 . If A is
Hermitian then A1 11 is Hermitian as well, and A H
12 = A 21 , which gives X = Y.
14.5. See Theorem 14.5.  
2 0
14.6. A = GGH , where G = 1 19 .
2 2
14.7. Note that B need not be a square matrix!
Let f (X) = X H B H ABX be a corresponding quadratic form. If BX = O
with X 6= O then f (X) = 0 thus the form cannot be positive definite. Otherwise,
let Ker B = O. Put Y = BX. Because A is positive definite we have that
f (X) = Y H AY 0 and we have the equality only when Y = O, thus X = O.
14.8. Because U is Hermitian all eigenvalues are real thus Re = . Let
V = 3U. What we need to know is how many eigenvalues of V are between 0 and
3
2 . Using Gaussian elimination we find that two eigenvalues of V are positive
and the same is true for V 32 I. Thus there are no eigenvalues in the given
interval.
14.9. Consider C() = A B and study p() = det C(). Show that if S
exists then p() should have real roots. Check that this is not the case for given
matrices.
More details. If S H AS = diag(a, b) and S H BS = diag(c, d) then
 
H
 a d
det(S C()S) = det diag(a c, b d) = bc
c b

and we see that p() has real roots. On the other hand
 
 1 
p() =   = 2 1.
1 

14.10. To make the calculations easier note that AH A = A2 = abs(A)
and we can use Lagrange interpolation.
308
   
1
5 6 28 3
Answer: U = , P = 161 .
61 6 5 3 33
14.11. Consider vector space of polynomials R 1 of degree less than n and check
that we have a scalar product (f (x), g(x)) = 0 f (x)g(x)dx here. Check that in
the basis 1, x, . . . xn1 we have (xi , xj ) = aij . The corresponding quadratic form
(f, f ) has matrix A in this basis and is positive definite, because it corresponds
to a scalar product.
14.12. Consider a vector space V consisting of all functions v(x) that are
ikt
linear combinations of the R functions e for k = 0, 1 . . . , n. It is sufficient to
show that (u(x), v(x)) = u(x)v(x)f (x)dx is a scalar product here.
14.13. Consider the matrix B =P A n I. It is positive semidefinite and
X H BX 0. Take X = Ei and X = Eii and see what this means for A.
14.14. We need only to check the number of positive eigenvalues for A, A2I
and A 5I, which can be done with the help of Gaussian elimination.
15.5. HINTS AND ANSWERS 309

15.1. Use theorem 15.10 and the fact that B is arbitrary.

More details. Theorem 15.10 gives XB = A+ B + Y AXB = AA+ B for
any column B. Therefore AX = AA+ by theorem 1.2 and it remains to use the
properties of A+ .
  3 3 2 2
0 0 1 1 6 2
15.2. A+ = , B + = 18 .
1 1 1 1 2 6
1 1 2 2
15.3. X = (0, 4, 7)T .
15.4. x = Pv + (v Pv). Note that P(v Pv)L = Pv P 2 v = 0, thus
w =T v Pv Ker P. Moreover, we have V = Im P Ker P, because if x
Im P Ker P then
x = Py = P 2 y = Px = 0.
We do not need the condition P = P to prove all this.
15.5. Let P be the corresponding matrix. If P H = P then P is Hermitian
and diagonalizable. Then P 2 = P iff eigenvalues are real and 2 = = 0
or = 1. In particularly, P + = P.
15.6. If AXA = A then P 2 = AXAX = AX = P. The opposite is less
trivial. Change the basis and use is that Im AX = Im A.
Moredetails. 
According to exercise 15.5 we can restrict ourself by the case
I O A1 A2
AX = . Let A = . Because Im AX = Im A we get
O O A3 A4
A3 = O, A4 = O. Then
    
I O A1 A2 A1 A2
(AX)A = = = A.
O O O O O O