INTRODUCTION

TO

LINEAR

ALGEBRA

Fifth Edition

MANUAL FOR INSTRUCTORS

Gilbert Strang
Massachusetts Institute of Technology

math.mit.edu/linearalgebra
web.mit.edu/18.06
video lectures: ocw.mit.edu
math.mit.edu/∼gs
www.wellesleycambridge.com
email: [email protected]

Wellesley - Cambridge Press

Box 812060
Wellesley, Massachusetts 02482
Solutions to Exercises 151

Problem Set 9.1, page 436

1 (a)(b)(c) have sums 4, −2 + 2i, 2 cos θ and products 5, −2i, 1. Note (eiθ )(e−iθ ) = 1.
√ iθ √
2 In polar form these are 5e , 5e2iθ , √1 e−iθ , 5.
5
1
3 The absolute values are r = 10, 100, 10
, and 100. The angles are θ, 2θ, −θ and −2θ.

4 |z × w| = 6, |z + w| ≤ 5, |z/w| = 23 , |z − w| ≤ 5.
√ √ √
3
5 a + ib = 2
+ 12 i, 1
2
+ 2
3
i, i, − 12 + 2
3
i; w12 = 1.

6 1/z has absolute value 1/r and angle −θ; (1/r)e−iθ times reiθ equals 1.
         
a −b c ac − bd real part 1 −3 1 10
7        =  is the matrix
b a d bc + ad imaginary part 3 1 −3 0
form of (1 + 3i)(1 − 3i) = 10.
    
A1 −A2 x1 b
8     =  1  gives complex matrix = vector multiplication (A1 +
A2 A1 x2 b2
iA2 )(x1 + ix2 ) = b1 + ib2 .
1−i
9 2 + i; (2 + i)(1 + i) = 1 + 3i; e−iπ/2 = −i; e−iπ = −1; 1+i
= −i; (−i)103 = i.

10 z + z is real; z − z is pure imaginary; zz is positive; z/z has absolute value 1.

   
a b 0 1
11   includes aI (which just adds a to the eigenvalues and b  . So the
−b a −1 0
eigenvectors are x1 = (1, i) and x2 = (1, −i). The eigenvalues are λ1 = a + bi and
λ2 = a − bi. We see x1 = x2 and λ1 = λ2 as expected for real matrices with complex
eigenvalues.
√
12 (a) When a = b = d = 1 the square root becomes 4c; λ is complex if c < 0
(b) λ = 0 and λ = a + d when ad = bc (c) the λ’s can be real and different.

13 Complex λ’s when (a + d)2 < 4(ad − bc); write (a + d)2 − 4(ad − bc) as (a − d)2 + 4bc

which is positive when bc > 0.

14 The symmetric block matrix has real eigenvalues; so iλ is real and λ is pure imaginary.
√
15 (a) 2eiπ/3 , 4e2iπ/3 (b) e2iθ , e4iθ (c) 7e3πi/2 , 49e3πi (= −49) (d) 50e−πi/4 ,
50e−πi/2 .
152 Solutions to Exercises

π
16 r = 1, angle 2
− θ; multiply by eiθ to get eiπ/2 = i.
√ √
17 a + ib = 1, i, −1, −i, ± √12 ± √i .
2
The root w = w−1 = e−2πi/8 is 1/ 2 − i/ 2.

18 1, e2πi/3 , e4πi/3 are cube roots of 1. The cube roots of −1 are −1, eπi/3 , e−πi/3 .

Altogether six roots of z 6 = 1.

19 cos 3θ = Re[(cos θ + i sin θ)3 ] = cos3 θ − 3 cos θ sin2 θ; sin 3θ = 3 cos2 θ sin θ − sin3 θ.

20 If the conjugate z = 1/z then |z|2 = 1 and z is any point eiθ on the unit circle.

21 ei is at angle θ = 1 on the unit circle; |ie | = 1e ; Infinitely many ie = ei(π/2+2πn)e .

22 (a) Unit circle (b) Spiral in to e−2π (c) Circle continuing around to angle θ = 2π 2 .

Problem Set 9.2, page 443

√ √
1 kuk = 9 = 3, kvk = 3, uH v = 3i + 2, v H u = −3i + 2 (this is the conjugate of
uH v).
 
2 0 1+i  
  3 1
 
2 AH A =  0 2 1 + i  and AAH =   are Hermitian matrices. They
  1 3
1−i 1−i 2
share the eigenvalues 4 and 2.

3 z =multiple of (1+i, 1+i, −2); Az = 0 gives z H AH = 0H so z (not z!) is orthogonal

to all columns of AH (using complex inner product z H times columns of AH ).

4 The four fundamental subspaces are now C(A), N(A), C(AH ), N(AH ). AH and not AT .

5 (a) (AH A)H = AH AHH = AH A again (b) If AH Az = 0 then (z H AH )(Az) = 0.

This is kAzk2 = 0 so Az = 0. The nullspaces of A and AH A are always the same.
 
(a) False 0 1
6 A=Q=  (b) True: −i is not an eigenvalue when S = S H .
(c) False −1 0

7 cS is still Hermitian for real c; (iS)H = −iS H = −iS is skew-Hermitian.

Solutions to Exercises 153
   
0 0 −1 −i
   
   
8 This P is invertible and unitary. P 2 =  −1 0 , P 3 = 
0 −i =
   
0 −1 0 −i
−iI. Then P 100 = (−i)33 P = −iP . The eigenvalues of P are the roots of λ3 = −i,
which are i and ie2πi/3 and ie4πi/3 .

9 One unit eigenvector is certainly x1 = (1, 1, 1) with λ1 = i. The other eigenvectors

are x2 = (1, w, w2 ) and x3 = (1, w2 , w4 ) with w = e2πi/3 . The eigenvector matrix

is the Fourier matrix F3 . The eigenvectors of any unitary matrix like P are orthogonal
(using the correct complex form xH y of the inner product).

10 (1, 1, 1), (1, e2πi/3 , e4πi/3 ), (1, e4πi/3 , e2πi/3 ) are orthogonal (complex inner product!)

because P is an orthogonal matrix—and therefore its eigenvector matrix is unitary.

11 If QH Q = I then Q−1 (QH )−1 = Q−1 (Q−1 )H = I so Q−1 is also unitary. Also

(QU )H (QU ) = U H QH QU = U H U = I so QU is unitary.

12 Determinant = product of the eigenvalues (all real). And A = AH gives det A = det A.

13 (z H AH )(Az) = kAzk2 is positive unless Az = 0. When A has independent columns

this means z = 0; so AH A is positive definite.

    
1 1 −1 + i 2 0 1 1 − i
14 S=√    √1  .
3 1+i 1 0 −1 3 −1 − i 1
    
1 1 −1 − i 2i 0 1 1 1 + i
15 K =(iAT in Problem 14) = √   √  ;
3 1−i 1 0 −i 3 −1 + i 1
λ’s are imaginary.
    
1 1 −i cos θ + i sin θ 0 1 i
16 U= √    √1   has |λ| = 1.
2 −i 1 0 cos θ − i sin θ 2 i 1
 √    √ 
1  1 + 3 −1 + i   1 0 1 1+ 3 1−i √
17 U= √   √  with L2 = 6+2 3.
L 1+i 1+ 3 0 −1 L −1 − i 1+ 3
H
Unitary means |λ| = 1. U = U gives real λ. Then trace zero gives λ = 1 and −1.

18 The v’s are columns of a unitary matrix U , so U H is U −1 . Then z = U U H z =

(multiply by columns) = v 1 (v H H
1 z) + · · · + v n (v n z): a typical orthonormal expansion.
154 Solutions to Exercises

19 z = (1, i, −2) completes an orthogonal basis for C3 . So does any eiθ z.

20 S = A + iB = (A + iB)H = AT − iB T ; A is symmetric but B is skew-symmetric.

21 Cn has dimension n; the columns of any unitary matrix are a basis. For example use

the columns of iI: (i, 0, . . . , 0), . . . , (0, . . . , 0, i)

   
a b + ic w eiφ z with |w|2 + |z|2 = 1
22 [ 1 ] and [ −1 ]; any [ eiθ ];  ;  
b − ic d −z eiφ w and any angle φ

23 The eigenvalues of AH are complex conjugates of the eigenvalues of A: det(A−λI) = 0

gives det(AH − λI) = 0.

24 (I − 2uuH )H = I − 2uuH and also (I − 2uuH )2 = I − 4uuH + 4u(uH u)uH = I.

The rank-1 matrix uuH projects onto the line through u.

25 Unitary U H U = I means (AT −iB T )(A+iB) = (AT A+B T B)+i(AT B−B T A) = I.

AT A + B T B = I and AT B − B T A = 0 which makes the block matrix orthogonal.

H T T T T
 given A+ iB = (A + iB) = A − iB . Then A = A and B = −B . So
26 We are
A −B
that   is symmetric.
B A

27 SS −1 = I gives (S −1 )H S H = I. Therefore (S −1 )H is (S H )−1 = S −1 and S −1 is

Hermitian.

28 If U has (complex) orthonormal columns, then U H U = I and U is unitary. If those

columns are eigenvectors of A, then A = U ΛU −1 = U ΛU H is normal. The direct test

for a normal matrix (which is AAH = AH A because diagonals could be real!) and ΛH
surely commute:

AAH=(U ΛU H )(U ΛH U H )=U (ΛΛH )U H = U (ΛH Λ)U H = (U ΛH U H )(U ΛU H ) =AH A.

An easy way to construct a normal matrix is 1 + i times a symmetric matrix. Or take

A = S + iT where the real symmetric S and T commute (Then AH = S − iT and
AAH = AH A).
Solutions to Exercises 155

Problem Set 9.3, page 450

1 Equation (3) (the FFT) is correct using i2 = −1 in the last two rows and three columns.
     
1 1 1 1 1
     
     
 1  1 1 i2 1 1 1 1 H
2 F −1 =

 
2
 
 
= F .
 4
 1   1 1  2 1 −1 
     
2
1 1 i −i i
   
1 1 1 1 1
   
   
 1   1 i2  1 1
3 F =





 permutation last.

 1  1 1  1 −1 
   
1 1 i2 −i i
   
1 1 1 1
   
   
4 D= e2πi/6  (note 6 not 3) and F3  1 e2πi/3 e4πi/3 .
   
e4πi/6 1 e4πi/3 e2πi/3

5 F −1 w = v and F −1 v = w/4. Delta vector ↔ all-ones vector.

 
4 0 0 0
 
 
 0 0 0 4 
2 
6 (F4 ) =   and (F4 )4 = 16I. Four transforms recover the signal!

0 0 4 0
 
0 4 0 0
               
1 1 2 2 0 0 0 2
               
               
0 1 0 0 1 0 0   0
7 c=                
  →   →   →   =F c. Also C =   →   →   →   =F C.
1 0 0 2  0   1   2   −2 
               
0 0 0 0 1 1 0 0
Adding c + C gives (1, 1, 1, 1) to (4, 0, 0, 0) = 4 (delta vector).

8 c → (1, 1, 1, 1, 0, 0, 0, 0) → (4, 0, 0, 0, 0, 0, 0, 0) → (4, 0, 0, 0, 4, 0, 0, 0) = F8 c.

C → (0, 0, 0, 0, 1, 1, 1, 1) → (0, 0, 0, 0, 4, 0, 0, 0) → (4, 0, 0, 0, −4, 0, 0, 0) = F8 C.

√
9 If w64 = 1 then w2 is a 32nd root of 1 and w is a 128th root of 1: Key to FFT.
156 Solutions to Exercises

10 For every integer n, the nth roots of 1 add to zero. For even n, they cancel in pairs. For

any n, use the geometric series formula 1 + w + · · · + wn−1 = (wn − 1)/(w − 1) = 0.

√ √
In particular for n = 3, 1 + (−1 + i 3)/2 + (−1 − i 3)/2 = 0.

11 The eigenvalues of P are 1, i, i2 = −1, and i3 = −i. Problem 11 displays the eigen-

vectors. And also det(P − λI) = λ4 − 1.

 
0 1 0
 
 
12 Λ = diag(1, i, i2 , i3 ); P =  0 0 1  and P T lead to λ3 − 1 = 0.
 
1 0 0
13 e1 = c0 + c1 + c2 + c3 and e2 = c0 + c1 i + c2 i2 + c3 i3 ; E contains the four eigenvalues

of C = F EF −1 because F contains the eigenvectors.

14 Eigenvalues e1 = 2 − 1 − 1 = 0, e2 = 2 − i − i3 = 2, e3 = 2 − (−1) − (−1) = 4,

e4 = 2 − i3 − i9 = 2. Just transform column 0 of C. Check trace 0 + 2 + 4 + 2 = 8.

15 Diagonal E needs n multiplications, Fourier matrix F and F −1 need 21 n log2 n multi-

plications each by the FFT. The total is much less than the ordinary n2 for C times x.

16 The row 1, wk , w 2k , . . . in F is the same as the row 1, wN−k , wN−2k , . . . in F because

wN−k = e(2πi/N)(N−k) is e2πi e−(2πi/N)k = 1 times w k . So F and F have the same

rows in reversed order (except for row 0 which is all ones).

17 0 000 reverses to 000 = 0

1 001 reverses to 100 = 4

2 010 reverses to 010 = 2 Now evens come before odds !

3 011 reverses to 110 = 6

4 100 reverses to 001 = 1

5 101 reverses to 101 = 5

6 110 reverses to 011 = 3

7 111 reverses to 111 = 7