CH 8 Sol

Solutions — Chapter 8
sds 8.1. Simple Dynamical Systems.
8.1.1. Solve the following initial value problems:

du du du
(a) = 5 u, u(0) = − 3, (b) = 2 u, u(1) = 3, (c) = − 3 u, u(−1) = 1.
dt dt dt
Solution: (a) u(t) = − 3 e5 t , (b) u(t) = 3 e2(t−1) , (c) u(t) = e− 3(t+1) .
8.1.2. Suppose a radioactive material has a half-life of 100 years. What is the decay rate γ?
Starting with an initial sample of 100 grams, how much will be left after 10 years? 100
years? 1, 000 years?
Solution: γ = log 2/100 ≈ 0.0069. After 10 years: 93.3033 gram; after 100 years 50 gram; after
1000 years 0.0977 gram.
8.1.3. Carbon-14 has a half-life of 5730 years. Human skeletal fragments discovered in a cave
are analyzed and found to have only 6.24% of the carbon-14 that living tissue would have.
How old are the remains?
Solution: Solve e− (log 2)t/5730 = .0624 for t = − 5730 log .0624/ log 2 = 22, 933 years.
8.1.4. Prove that if t? is the half-life of a radioactive material, then u(n t? ) = 2−n u(0). Ex-
plain the meaning of this equation in your own words.
? “ ”t/t?
Solution: By (8.6), u(t) = u(0) e− (log 2)t/t = u(0) 21 = 2−n u(0) when t = n t? . After
every time period of duration t? , the amount of material is cut in half.
8.1.5. A bacteria colony grows according to the equation du/dt = 1.3 u. How long until the
colony doubles? quadruples? If the initial population is 2, how long until the population
reaches 2 million?
Solution: u(t) = u(0) e1.3 t . To double, we need e1.3 t = 2, so t = log 2/1.3 = 0.5332. To
quadruple takes twice as long, t = 1.0664. To reach 2 million needs t = log 106 /1.3 = 10.6273.
du
8.1.6. Deer in Northern Minnesota reproduce according to the linear differential equation =
dt
.27 u where t is measured in years. If the initial population is u(0) = 5, 000 and the environ-
ment can sustain at most 1, 000, 000 deer, how long until the deer run out of resources?
Solution: The solution is u(t) = u(0) e.27 t . For the given initial conditions, u(t) = 1, 000, 000
when t = log(1000000/5000)/.27 = 19.6234 years.
du
♦ 8.1.7. Consider the inhomogeneous differential equation = a u + b, where a, b are constants.
dt
(a) Show that u? = − b/a is a constant equilibrium solution. (b) Solve the differential
equation. Hint: Look at the differential equation satisfied by v = u − u? . (c) Discuss the
stability of the equilibrium solution u? .
Solution:
evv 9/9/04 415 °

c 2004 Peter J. Olver
b du
(a) If u(t) ≡ u? = − , then = 0 = a u + b, hence it is a solution.
a dt
dv b
(b) v = u − u? satisfies = a v, so v(t) = c ea t and so u(t) = c ea t − .
dt a
(c) The equilibrium solution is asymptotically stable if and only if a < 0 and is stable if
a = 0.
8.1.8. Use the method of Exercise 8.1.7 to solve the following initial value problems:
du du du
(a) = 2 u − 1, u(0) = 1, (b) = 5 u + 15, u(1) = − 3, (c) = − 3 u + 6, u(2) = − 1.
dt dt dt
Solution: (a) u(t) = 1
2 + 1
2 e2 t , (b) u(t) = − 3, (c) u(t) = 2 − 3 e− 3(t−2) .
8.1.9. The radioactive waste from a nuclear reactor has a half-life of 1000 years. Waste is con-
tinually produced at the rate of 5 tons per year and stored in a dump site. (a) Set up an
inhomogeneous differential equation, of the form in Exercise 8.1.7, to model the amount
of radioactive waste. (b) Determine whether the amount of radioactive material at the
dump increases indefinitely, decreases to zero, or eventually stabilizes at some fixed amount.
(c) Starting with a brand new site, how long until the dump contain 100 tons of radioac-
tive material?
du log 2
Solution: (a) =− u + 5 ≈ − .000693 u + 5. (b) Stabilizes at the equilibrium solution
dt 1000
5000 h “
log 2
”i
u? = 5000/ log 2 ≈ 721 tons. (c) The solution is u(t) = 1 − exp − 1000 t which equals
! log 2
1000 log 2
100 when t = − log 1 − 100 ≈ 20.14years.
log 2 5000
♥ 8.1.10. Suppose that hunters are allowed to shoot a fixed number of the Northern Minnesota
deer in Exercise 8.1.6 each year. (a) Explain why the population model takes the form
du
= .27 u − b, where b is the number killed yearly. (Ignore the season aspects of hunt-
dt
ing.) (b) If b = 1, 000, how long until the deer run out of resources? Hint: See Exercise
8.1.7. (c) What is the maximal rate at which deer can be hunted without causing their ex-
tinction?
Solution:
(a) The rate of growth remains proportional to the population, while hunting decreases the
population by a fixed amount. (This assumes hunting is done continually throughout
the year, which is not what
“
happens in ”
real life.)
(b) The solution is u(t) = 5000 − .27 e.27 t + 1000
1000
.27 . Solving u(t) = 100000 gives t =
1 1000000 − 1000/.27
log = 24.6094 years.
.27 5000 − 1000/.27
(c) We need the equilibrium u? = b/.27 to be less than the initial population, so b < 1350
deer.
du
♦ 8.1.11. (a) Prove that if u1 (t) and u2 (t) are any two distinct solutions to = a u with a > 0,
dt
then | u1 (t) − u2 (t) | → ∞ as t → ∞. (b) If a = .02 and u1 (0) = .1, u2 (0) = .05, how long
do you have to wait until | u1 (t) − u2 (t) | > 1, 000?
Solution:
(a) | u1 (t) − u2 (t) | = ea t | u1 (0) − u2 (0) | → ∞ when a > 0.
(b) t = log(1000/.05)/.02 = 495.17.
du 2 1
8.1.12. (a) Write down the exact solution to the initial value problem = u, u(0) = .
dt 7 3
evv 9/9/04 416 °

(b) Suppose you make the approximation u(0) = .3333. At what point does your solution
differ from the true solution by 1 unit? by 1000? (c) Answer the same question if you also
du
approximate the coefficient in the differential equation by = .2857 u.
dt
Solution:
(a) u(t) = 31 e2 t/7 . h i
(b) One unit: t = log 1/(1/3 − .3333) /(2/7) = 36.0813;
h i
1000 units: t = log 1000/(1/3 − .3333) /(2/7) = 60.2585
(c) One unit: t ≈ 30.2328 solves 13 e2 t/7 − .3333 e.2857 t = 1. (Use a numerical equation
solver.) 1000 units: t ≈ 52.7548 solves 31 e2 t/7 − .3333 e.2857 t = 1000.
♦ 8.1.13. Let a be complex. Prove that u(t) = c ea t is the (complex) solution to our scalar ordi-
nary differential equation (8.1). Describe the asymptotic behavior of the solution as t → ∞,
and the stability properties of the zero equilibrium solution.
Solution: The solution is still valid as a complex solution. If Re a > 0, then u(t) → ∞ as t →
∞, and the origin is an unstable equilibrium . If Re a = 0, then u(t) remains bounded t → ∞,
and the origin is a stable equilibrium. If Re a < 0, then u(t) → 0 as t → ∞, and the origin is an
asymptotically stable equilibrium.
ev 8.2. Eigenvalues and Eigenvectors.
8.2.1. Find the eigenvalues0and eigenvectors

1 of the following matrices:
! ! !
1 −2 B 1 − 23 C 3 1 1 2
(a) , (b) @ A, (c) , (d) ,
−2 1 1 1 −1 1 −1 1
0 1 2 0− 6 1 0 1 0 1
3 −1 0 −1 −1 4 1 −3 11 2 −1 −1
(e) B
@ −1 2 −1 C B
A, (f ) @ 1 3 −2 C B
A, (g) @ 2 −6 16 C
A, (h)
B
@ −2 1 1CA,
0 −1 3 0
1 1 −1
1 0
1 −3 7 1 1 0 1
0 1 3 4 0 0 4 0 0 0
−4 −4 2 B C B
B C B4 3 0 0C B 1 3 0 0C C
(i) @ 3 4 −1 A, (j) B
@0 0 1 3A
C, (k) B
@ −1
C.
1 2 0A
−3 −2 3
0 0 4 5 1 −1 1 1
Solution: ! !
−1 1
(a) Eigenvalues: 3, −1; eigenvectors: , .
1 1
! !
1 1 4 1
(b) Eigenvalues: 2 , 3 ; eigenvectors: , .
3 1
!
−1
(c) Eigenvalues: 2; eigenvectors: .
1
√ √ √ ! √ !
(d) Eigenvalues: 1 + i 2, 1 − i 2; eigenvectors: − i 2 , i 2 .
0 1 0
1
1 0 1
1
1 −1 1
(e) Eigenvalues: 4, 3, 1; eigenvectors: B C B C B C
@ −1 A, @ 0 A, @ 2 A.
1 1 1
evv 9/9/04 417 °

0 √ 1 0 √ 1
0 1 3 3
√ √ 2 B −1 +√√2 C B −1 −√√2 C
(f ) Eigenvalues: 1, 6, − 6; eigenvectors: B C B
@ 0 A, B
C B
3 C, B
@ 2+ √ A @ 2− √ A
C
3 C.
1 2 2
0 1 0
1 1 0
1 1
3 3 − 2i 3 + 2i
(g) Eigenvalues: 0, 1 + i , 1 − i ; eigenvectors: B C B C B
@ 1 A , @ 3 − i A, @ 3 + i A.
C
0 1 0
01 1 1
1 −1
(h) Eigenvalues: 2, 0; eigenvectors: B C B
@ −1 A, @ −3 A
C
0
1 1 1 0 1 0 1
1
2 −2
B C B3C B 3C
(i) −1 simple eigenvalue, eigenvector @ −1 A; 2 double eigenvalue, eigenvectors @ 0 A, @ 1 A
1 1 0
0 1 0 1 0 1 0 1
1 0 1 0
B
−1 C B
C B 0C
C B C B C
B1C B0C
(j) Two double eigenvalues: −1, eigenvectors B B C, B
@ 0 A @ −3 A
C and 7, eigenvectors B C, B C.
@0A @1A
0
0 1 0 1 0 1 0 1
2 0 2
0 0 0 1
B C B C B C B C
0C B0C B1C B1C
(k) Eigenvalues: 1, 2, 3, 4; eigenvectors: B
B C , B C, B C, B C .
@0A @1A @1A @0A
1 1 0 0
!
cos θ − sin θ
8.2.2. (a) Find the eigenvalues of the rotation matrix Rθ = . For what values
sin θ cos θ
of θ are the eigenvalues real? (b) Explain why your answer gives an immediate solution to
Exercise 1.5.7c.
!
±iθ 1
Solution: (a) The eigenvalues are cos θ ± i sin θ = e with eigenvectors . They are real
∓i
only for θ = 0 and π. (b) Because Rθ − a I has an inverse if and only if a is not
!
an eigenvalue.
cos θ sin θ
8.2.3. Answer Exercise 8.2.2a for the reflection matrix Fθ = .
sin θ − cos θ
Solution: The eigenvalues are ±1 with eigenvectors ( sin θ, cos θ ∓ 1 )T .
8.2.4. Write down (a) a 2 × 2 matrix that has 0 as one of its eigenvalues and ( 1, 2 ) T as a cor-
responding eigenvector; (b) a 3 × 3 matrix that has ( 1, 2, 3 )T as an eigenvector for the
eigenvalue −1.
Solution: (a) O, and (b) − I , are trivial examples. 0 1
0 1 0
B
8.2.5. (a) Write out the characteristic equation for the matrix 0 @ 0 1C A.
α β γ
(b) Show that, given any 3 numbers a, b, and c, there is a 3 × 3 matrix with characteristic
equation − λ3 + a λ2 + b λ + c = 0.
Solution:
(a) The characteristic equation is − λ3 + a λ2 + b λ + c = 0.
(b) Use the matrix in part (a). 0 1
0 c −b
8.2.6. Find the eigenvalues and eigenvectors of the cross product matrix A = B
@ −c 0 aCA.
b −a 0
√
Solution: The eigenvalues are 0, ± i a2 + b2 + c2 . If a = b = c = 0 then A = O and all vectors
evv 9/9/04 418 °

0 1 0 1 0 1
a b i −ac
B
are eigenvectors. Otherwise, the eigenvectors are @ bC B C
A, @ −a A ∓ √
B C
@ − b c A.
2 2
a +b +c 2 2 2
c 0 a +b
8.2.7. Find all eigenvalues and eigenvectors of the following complex0matrices: 1
! ! ! 1 + i −1 − i 1− i
i 1 2 i i −2 i +1
(a) , (b) , (c) , (d) B
@ 2 −2 − i 2 − 2i C
A.
0 −1 + i − i −2 i +2 i −1
−1 1+ i −1 + 2 i
Solution: ! !
1 −1
(a) Eigenvalues: i , −1 + i ; eigenvectors: , .
0 1
√ √ !
(b) Eigenvalues: ± 5; eigenvectors: i (2 ± 5) .
1! !
−1 3 1
(c) Eigenvalues: −3, 2 i ; eigenvectors: , 5 + 5 i .
1 1
0 1 0 1 0 1
−1 −1 + i 1+ i
(d) −2 simple, eigenvector B C
@ −2 A; i double, eigenvectors @
B
0 C B
A, @ 1
C
A.
1 1 0
8.2.8. Find all eigenvalues and eigenvectors of (a) the n × n zero matrix O; (b) the n × n
identity matrix I .
Solution:
(a) Since O v = 0 = 0 v, we conclude that 0 is the only eigenvalue; all nonzero vectors are
eigenvectors.
(b) Since I v = v = 1 v, we conclude that 1 is the only eigenvalue; all nonzero vectors are
eigenvectors.
8.2.9. Find the eigenvalues and eigenvectors of an n × n matrix with every entry equal to 1.
Hint: Try with n = 2, 3, and then generalize.
! !
−1 1
Solution: For n = 2, the eigenvalues are 0, 2, and the eigenvectors are , and . For
1 1
0 1 0 1 0 1
−1 −1 1
n = 3, the eigenvalues are 0, 0, 3, and the eigenvectors are B C B C B C
@ 0 A, @ 1 A, and @ 1 A. In gen-
1 0 1
eral, the eigenvalues are 0, with multiplicity n − 1, and n, which is simple. The eigenvectors
corresponding to the 0 eigenvalues are of the form ( v1 , v2 , . . . , vn )T where v1 = −1 and vj = 1
for j = 2, . . . , n. The eigenvector corresponding to n is ( 1, 1, . . . , 1 )T .
♦ 8.2.10. Let A be a given square matrix. (a) Explain in detail why any nonzero scalar multiple
of an eigenvector of A is also an eigenvector. (b) Show that any nonzero linear combina-
tion of two eigenvectors v, w corresponding to the same eigenvalue is also an eigenvector.
(c) Prove that a linear combination c v + d w with c, d 6= 0 of two eigenvectors correspond-
ing to different eigenvalues is never an eigenvector.
Solution:
(a) If A v = λ v, then A(c v) = c A v = c λ v = λ (c v) and so c v satisfies the eigenvector
equation with eigenvalue λ. Moreover, since v 6= 0, also c v 6= 0 for c 6= 0, and so c v is a
bona fide eigenvector.
(b) If A v = λ v, A w = λ w, then A(c v + d w) = c A v + d A w = c λ v + d λ w = λ (c v + d w).
(c) Suppose A v = λ v, A w = µ w. Then v and w must be linearly independent as other-
wise they would be scalar multiples of each other and hence have the same eigenvalue.
Thus, A(c v + d w) = c A v + d A w = c λ v + d µ w = ν(c v + d w) if and only if c λ = c ν
and d µ = d ν, which, when λ 6= µ, is only possible if either c = 0 or d = 0.
evv 9/9/04 419 °

8.2.11. True or false: If v is a real eigenvector of a real matrix A, then a nonzero complex
multiple w = c v for c ∈ C is a complex eigenvector of A.
Solution: True — by the same computation as in Exercise 8.2.10(a), c v is an eigenvector for the
same (real) eigenvalue λ.
“ ”T
♦ 8.2.12. Define the shift map S: C n → C n by S v1 , v2 , . . . , vn−1 , vn = ( v 2 , v3 , . . . , v n , v1 ) T .
(a) Prove that S is a linear map, and write down its matrix representation A.
(b) Prove that A is an orthogonal matrix.
(c) Prove that the sampled exponential vectors ω0 , . . . , ωn−1 in (5.84) form an eigenvector
basis of A. What are the eigenvalues?
Solution: 0 1
0 1 0 0 0
B 0 1 0 0C
B C
B C
B
B
0 1 0 C
C
(a) A = B
B
.. .. .. C
C.
B . . . C
B C
B
B
0 1 0C
C
@0 0 1A
1 0 0
(b) AT A = I by direct computation, or, equivalently, the columns of A are the standard
orthonormal basis vectors en , e1 , e2 , . . . , en−1 , written in a slightly different order.
(c) Since
„ «T
ωk = 1, e2 k π i /n , e4 k π i /n , . . . , e2 (n−1) k π i /n ,
„ «T
S ωk = e2 k π i /n , e4 k π i /n , . . . , e2 (n−1) k π i /n,1 = e2 k π i /n ωk ,
and so ωk is an eigenvector with corresponding eigenvalue e2 k π i /n .
0 1
1 4 4
8.2.13. (a) Compute the eigenvalues and corresponding eigenvectors of A = B @ 3 −1 0 A.
C
0 2 3
(b) Compute the trace of A and check that it equals the sum of the eigenvalues. (c) Find
the determinant of A, and check that it is equal to to the product of the eigenvalues.
“ ”T
Solution: (a) Eigenvalues: −3, 1, 5; eigenvectors: ( 2, −3, 1 )T , − 23 , −1, 1 , ( 2, 1, 1 )T .
(b) tr A = 3 = −3 + 1 + 5 (c) det A = −15 = (−3) · 1 · 5.
8.2.14. Verify the trace and determinant formulae (8.24–25) for the matrices in Exercise 8.2.1.
Solution:
(a) tr A = 2 = 3 + (−1); det A = −3 = 3 · (−1).
(b) tr A = 56 = 21 + 31 ; det A = 16 = 12 · 13 .
(c) tr A = 4 = 2 + 2; √ det A = 4 = 2√· 2. √ √
(d) tr A = 2 = (1 + i 2) + (1 − i 2); det A = 3 = (1 + i 2) · (1 − i 2).
(e) tr A = 8 = 4 + 3√+ 1; det √A = 12 = 4 · 3 · 1. √ √
(f ) tr A = 1 = 1 + 6 + (− 6); det A = −6 = 1 · 6 · (− √ 6). √
(g) tr A = 2 = 0 + (1 + i ) + (1 − i ); det A = 0 = 0 · (1 + i 2) · (1 − i 2).
(h) tr A = 4 = 2 + 2 + 0; det A = 0 = 2 · 2 · 0.
(i) tr A = 3 = (−1) + 2 + 2; det A = −4 = (−1) · 2 · 2.
(j) tr A = 12 = (−1) + (−1) + 7 + 7; det A = 49 = (−1) · (−1) · 7 · 7.
evv 9/9/04 420 °

(k) tr A = 10 = 1 + 2 + 3 + 4; det A = 24 = 1 · 2 · 3 · 4.
8.2.15. (a) Find the explicit formula for the characteristic polynomial det(A − λ I ) = − λ 3 +
a λ2 − b λ + c of a general 3 × 3 matrix. Verify that a = tr A, c = det A. What is the
formula for b? (b) Prove that if A has eigenvalues λ1 , λ2 , λ3 , then a = tr A = λ1 + λ2 + λ3 ,
b = λ1 λ2 + λ1 λ3 + λ2 λ3 , c = det A = λ1 λ2 λ3 .
Solution:
(a) a = a11 + a22 + a33 = tr A, b = a11 a22 − a12 a21 + a11 a33 − a13 a31 + a22 a33 − a23 a32 ,
c = a11 a22 a33 + a12 a23 a31 + a13 a21 a 32 − a11 a23 a32 − a12 a21 a33 − a13 a22 a31 = det A
(b) When the factored form of the characteristic polynomial is multiplied out, we obtain
− (λ − λ1 )(λ − λ2 )(λ − λ3 ) = − λ3 + (λ1 + λ2 + λ3 )λ2 − (λ1 λ2 + λ1 λ3 + λ2 λ3 )λ − λ1 λ2 λ3 ,
giving the eigenvalue formulas for a, b, c.
8.2.16. Prove that the eigenvalues of an upper triangular (or lower triangular) matrix are its
diagonal entries.
Solution: If U is upper triangular,Qso is U − λ I , and hence p(λ) = det(U − λ I ) is the product
of the diagonal entries, so p(λ) = (uii − λ), and so the roots of the characteristic equation are
the diagonal entries u11 , . . . , unn .
♦ 8.2.17. Let Ja be the n × n Jordan block matrix (8.22). Prove that its only eigenvalue is λ = a
and the only eigenvectors are the nonzero scalar multiples of the standard basis vector e 1 .
Solution: Since Ja −λ I is an upper triangular matrix with λ−a on the diagonal, its determinant
is det(Ja − λ I ) = (a − λ)n and hence its only eigenvalue is λ = a, of multiplicity n. (Or use
Exercise 8.2.16.) Moreover, (Ja − a I )v = ( v2 , v3 , . . . , vn , 0 )T = 0 if and only if v = c e1 .
♦ 8.2.18. Suppose that λ is an eigenvalue of A. (a) Prove that c λ is an eigenvalue of the scalar
multiple c A. (b) Prove that λ + d is an eigenvalue of A + d I . (c) More generally, c λ + d is
an eigenvalue of B = c A + d I for scalars c, d.
Solution: parts (a), (b) are special cases of part (c): If A v = λ v then Bv = (c A + d I )v =
(c λ + d) v.
8.2.19. Show that if λ is an eigenvalue of A, then λ2 is an eigenvalue of A2 .
Solution: If A v = λ v then A2 v = λA v = λ2 v, and hence v is also an eigenvector of A2 with
eigenvalue λ2 .
8.2.20. True or false: (a) If λ is an eigenvalue of both A and B, then it is an eigenvalue of the
sum A + B. (b) If v is an eigenvector of both A and B, then it is an eigenvector of A + B.
! !
0 1 0 0
Solution: (a) False. For example,
!
0 is an eigenvalue of both and , but the
0 0 1 0
0 1
eigenvalues of A + B = are ± i . (b) True. If A v = λ v and B v = µ v, then (A+B)v =
1 0
(λ + µ)v, and so v is an eigenvector with eigenvalue λ + µ.
8.2.21. True or false: If λ is an eigenvalue of A and µ is an eigenvalue of B, then λ µ is an
eigenvalue of the matrix product C = A B.
Solution: False in general, but true if the eigenvectors coincide: If A v = λ v and B v = µ v, then
A B v = (λ µ)v, and so v is an eigenvector with eigenvalue λ µ.
♦ 8.2.22. Let A and B be n × n matrices. Prove that the matrix products A B and B A have the
same eigenvalues.
Solution: If A B v = λ v, then B A w = λ w, where w = B v. Thus, as long as w 6= 0, it
evv 9/9/04 421 °

is an eigenvector of B A with eigenvalue λ. However, if w = 0, then A B v = 0, and so the
eigenvalue is λ = 0, which implies that A B is singular. But then so is B A, which also has 0 as
an eigenvalue. Thus every eigenvalue of A B is an eigenvalue of B A. The converse follows by
the same reasoning. Note: This does not imply their null eigenspaces or kernels have the same
dimension; compare Exercise 1.8.18. In anticipation of Section 8.6, even though A B and B A
have the same eigenvalues, they may have different Jordan canonical forms.
♦ 8.2.23. (a) Prove that if λ 6= 0 is a nonzero eigenvalue of A, then 1/λ is an eigenvalue of A −1 .
(b) What happens if A has 0 as an eigenvalue?
Solution: (a) Starting with A v = λ v, multiply both sides by A−1 and divide by λ to obtain
A−1 v = (1/λ) v. Therefore, v is an eigenvector of A−1 with eigenvalue 1/λ.
(b) If 0 is an eigenvalue, then A is not invertible.
♦ 8.2.24. (a) Prove that if det A > 1 then A has at least one eigenvalue with | λ | > 1. (b) If
| det A | < 1, are all eigenvalues | λ | < 1? Prove or find a counter-example.
Solution: (a) If all | λj | ≤ 1 then so is their product 1 ≥ | λ1 . . . λn | = | det A |, which is a
!
2 0
contradiction. (b) False. A = has eigenvalues 2, 13 while det A = 32 .
0 31
8.2.25. Prove that A is a singular matrix if and only if 0 is an eigenvalue.
Solution: Recall that A is singular if and only if ker A 6= {0}. Any v ∈ ker A satisfies A v = 0 =
0 v. Thus ker A is nonzero if and only if A has a null eigenvector.
8.2.26. Prove that every nonzero vector 0 6= v ∈ R n is an eigenvector of A if and only if A is a
scalar multiple of the identity matrix.
Solution: Let v, w be any two linearly independent vectors. Then A v = λ v and A w = µ w for
some λ, µ. But v + w is an eigenvector if and only if A(v + w) = λ v + µ w = ν(v + w), which
requires λ = µ = ν. Thus, A v = λ v for every v, which implies A = λ I .
8.2.27. How many unit eigenvectors correspond to a given eigenvalue of a matrix?
Solution: If λ is a simple real eigenvalue, then there are two real unit eigenvectors: u and − u.
For a complex eigenvalue, if u is a unit complex eigenvector, so is e i θ u, and so there are in-
finitely many complex unit eigenvectors. (The same holds for a real eigenvalue if we also allow
complex eigenvectors.) If λ is a multiple real eigenvalue, with eigenspace of dimension greater
than 1, then there are infinitely many unit real eigenvectors in the eigenspace.
8.2.28. True or false: (a) Performing an elementary row operation of type #1 does not change
the eigenvalues of a matrix. (b) Interchanging two rows of a matrix changes the sign of its
eigenvalues. (c) Multiplying one row of a matrix by a scalar multiplies one of its eigenval-
ues by the same scalar.
Solution: All false. Simple 2 × 2 examples suffice to disprove them.
8.2.29. (a) True or false: If λ1 , v1 and λ2 , v2 solve the eigenvalue equation (8.12) for a given
matrix A, so does λ1 + λ2 , v1 + v2 . (b) Explain what this has to do with linearity.
Solution: False. The eigenvalue equation A v = λ v is not linear in the eigenvalue and eigenvec-
tor since A(v1 + v2 ) 6= (λ1 + λ2 )(v1 + v2 ) in general.
8.2.30. An elementary reflection matrix has the form Q = I − 2 u uT , where u ∈ R n is a unit

vector. (a) Find the eigenvalues and eigenvectors for the elementary reflection matrices
evv 9/9/04 422 °

0 1 0 1 0 1
! 0 √1
3 B 2 C
1 B5C
corresponding to the following unit vectors: (i) , (ii) @ A, (iii) B C
@ 1 A, (iv ) B 0 C.
@ A
0 4
5 0 − √1
2
(b) What are the eigenvalues and eigenvectors of a general elementary reflection matrix?
Solution: ! !
−1 0 1
(a) (i) Q = . Eigenvalue −1 has eigenvector: ; Eigenvalue 1 has eigenvector:
0 1 0
0 1 0 1
! 7 24 3
0
. (ii) Q = B@ 25 − 25 C A. Eigenvalue −1 has eigenvector: B5C
@ A; Eigenvalue
1 24 7 4
−
0 25 1
− 25 0 1 5
4 1 0 0
1 has eigenvector: B C B
@ 5 A. (iii) Q = @ 0 −1 0 A.
C
Eigenvalue −1 has eigenvector:
3
0 1
−5 0
0 1 0 1
0 1 0 1
0 1 0 0 1 0
B C B C B C B
@ 1 A;
Eigenvalue 1 has eigenvectors: @ 0 A , @ 0 A.
(iv ) Q = 1 0C @0
A. Eigenvalue 1
0 0 1
0 1 1
0 1 0 1
0 0
−1 1 0
B
has eigenvector: @ 0 A; Eigenvalue −1 has eigenvectors: B
C C B C
@ 0 A, @ 1 A .
1 1 0
(b) u is an eigenvector with eigenvalue −1. All vectors orthogonal to u are eigenvectors
with eigenvalue +1.
♦ 8.2.31. Let A and B be similar matrices, so B = S −1 A S for some nonsingular matrix S.
(a) Prove that A and B have the same characteristic polynomial: pB (λ) = pA (λ). (b) Explain
why similar matrices have the same eigenvalues. (c) Do they have the same eigenvectors?
If not, how are their
!
eigenvectors
!
related? (d) Prove that the converse is not true by show-
2 0 1 1
ing that and have the same eigenvalues, but are not similar.
0 2 −1 3
Solution: h i
det(B − λ I ) = det(S −1 A S − λ I ) = det S −1 (A − λ I )S
(a)
= det S −1 det(A − λ I ) det S = det(A − λ I ).
(b) The eigenvalues are the roots of the common characteristic equation.
(c) Not usually. If w is an eigenvector of B, then v = S w is an eigenvector of A and con-
versely. ! !
1 1 −1 2 0
(d) Both have 2 as a double eigenvalue. Suppose = S S, or, equiv-
−1 3 0 2
! ! !
1 1 2 0 x y
alently, S = S for some S = . Then, equating entries,
−1 3 0 2 z w
we must have x − y = 2 x, x + 3 y = 0, z − w = 0, z + 3 w = 2 w which implies
x = y = z = w = 0, and so S = O, which is not invertible.
8.2.32. Let A be a nonsingular n × n matrix with characteristic polynomial pA (λ). (a) Explain
how to construct the characteristic polynomial pA−10(λ) of its inverse
1
directly from pA (λ).
! 1 4 4
1 2
(b) Check your result when A = (i) , (ii) B
@ −2 −1 0 A.
C
3 4
0 2 3
Solution: ! !
−1 −1 1 (− λ)n 1
(a) pA−1 (λ) = det(A − λ I ) = det λ A I −A = p .
λ det A A λ
evv 9/9/04 423 °

Or, equivalently, if
pA (λ) = (−1)n λn + cn−1 λn−1 + · · · + c1 λ + c0 ,
then, since c0 = det A 6= 0,
" #
n n c1 n−1 cn−1 1
pA−1 (λ) = (−1) λ + λ + ··· + λ + .
c0 c0 c0
!
−2 1
(b) (i) A−1 = 3 . Then pA (λ) = λ2 − 5 λ − 2, while
2 − 12
!
λ2 2 5
pA−1 (λ) = λ2 + 5
2 λ− 1
2 = − − +1 .
0 1
2 λ2 λ
3
B − 5 − 45 4
5 C
B C
(ii) A−1 = B
B
6 3
− 58 C.
C Then pA (λ) = − λ3 + 3 λ2 − 7 λ + 5, while
@ 5 5 A
4
− 5 − 25 7
5 !
3 7 2 3 1 − λ3 1 3 7
pA−1 (λ) = − λ + 5 λ − 5 λ+ 5 = − 3 + 2 − +5 .
5 λ λ λ
♥ 8.2.33. A square matrix A is called nilpotent if Ak = O for some k ≥ 1. (a) Prove that the
only eigenvalue of a nilpotent matrix is 0. (The converse is also true; see Exercise 8.6.13.)
(b) Find examples where Ak−1 6= O but Ak = O for k = 2, 3, and in general.
Solution:
(a) If A v = λ v then 0 = Ak v = λk v and hence λk = 0.
(b) If A has size (k + 1) × (k + 1) and is all zero except for ai,i+1 = 1 on the supra-diagonal,
i.e., a k × k Jordan block J0,k with zeros along the diagonal.
♥ 8.2.34. (a) Prove that every eigenvalue of a matrix A is also an eigenvalue of its transpose A T .
(b) Do they have the same eigenvectors? (c) Prove that if v is an eigenvector of A with
eigenvalue λ and w is an eigenvector of AT with a different eigenvalue µ 6= λ, then v and w
are orthogonal vectors with respect to the dot product.
!
0 1
5 −4 2
0 −1
(d) Illustrate this result when (i) A = , (ii) A = B @ 5 −4 1CA.
2 3
−2 2 −3
Solution:
(a) det(AT − λ I ) = det(A − λ I )T = det(A − λ I ), and hence A and AT have the same
characteristic polynomial.
(b) No. See the examples.
(c) λ v · w = (A v)T w = vT AT w = µ v · w, so if µ 6= λ, v · w = 0 and the vectors are
orthogonal. ! !
−1 1
(d) (i) The eigenvalues are 1, 2; the eigenvectors of A are v1 = , v2 = − 2 ;
1 1
! !
T 2 1
the eigenvectors of A are w1 = , w2 = , and v1 , w2 are orthogonal, as
1 1
0 1
1
are v2 , w1 . The eigenvalues are 1, −1, −2; the eigenvectors of A are v1 = B C
@ 1 A, v 2 =
0 1 0 1 0 1 0 1
0 0 1
1 0 3 2 1
B C B1C T B C B C B C
@ 2 A, v3 = @ 2 A; the eigenvectors of A are w1 = @ −2 A, w2 = @ −2 A, w3 = @ −1 A.
1 1 1 1 1
Note that vi is orthogonal to wj whenever i 6= j.
8.2.35. (a) Prove that every real 3 × 3 matrix has at least one real eigenvalue. (b) Find a real
evv 9/9/04 424 °

4 × 4 matrix with no real eigenvalues. (c) Can you find a real 5 × 5 matrix with no real
eigenvalues?
Solution: (a) The characteristic
0
equation of a13 × 3 matrix is a cubic polynomial, and hence has
0 1 0 0
B
B −1 0 0 0C C
at least one real root. (b) B@ 0 0
C has eigenvalues ± i . (c) No, since the charac-
0 1A
0 0 −1 0
teristic polynomial is degree 5 and hence has at least one real root.
8.2.36. (a) Show that if A is a matrix such that A4 = I , then the only possible eigenvalues of
A are 1, −1, i and − i . (b) Give an example of a real matrix that has all four numbers as
eigenvalues.
Solution:
(a) If
0
A v = λ v, then1v = A4 v = λ4 v, and hence any eigenvalue must satisfy λ4 = 1.
1 0 0 0
B
B 0 −1 0 0CC
(b) B@0
C.
0 0 −1 A
0 0 1 0
8.2.37. A projection matrix satisfies P 2 = P . Find all eigenvalues and eigenvectors of P .
Solution: If P v = λ v then P 2 v = λ2 v. Since P v = P 2 v, we find λ v = λ2 v and so, since
v 6= 0, 2 = λ so the only eigenvalues are λ = 0, 1. All v ∈ rng P are eigenvectors with eigenvalue
1 since if v = P u, then P v = P 2 u = P u = v, whereas all w ∈ ker P are null eigenvectors.
8.2.38. True or false: All n × n permutation matrices have real eigenvalues.
0 1
0 0 1 √
Solution: False. For example, B
@1 0 0C 1
A has eigenvalues 1, − 2 ± 2
3
i.
0 1 0
8.2.39. (a) Show that if all the row sums of A are equal to 1, then A has 1 as an eigenvalue.
(b) Suppose all the column sums of A are equal to 1. Does the same result hold? Hint: Use
Exercise 8.2.34.
Solution:
(a) According to Exercise 1.2.29, if z = ( 1, 1, . . . , 1 )T , then A z is the vector of row sums of
A, and hence, by the assumption, A z = z so z is an eigenvector with eigenvalue 1.
(b) Yes, since the column sums of A are the row sums of AT , and Exercise 8.2.34 says that
A and AT have the same eigenvalues.
8.2.40. Let Q be an orthogonal matrix. (a) Prove that if λ is an eigenvalue, then so is 1/λ.
(b) Prove that all its eigenvalues are complex numbers of modulus | λ | = 1. In particular,
the only possible real eigenvalues of an orthogonal matrix are ±1. (c) Suppose v = x + i y
is a complex eigenvector corresponding to a non-real eigenvalue. Prove that its real and
imaginary parts are orthogonal vectors having the same Euclidean norm.
Solution:
(a) If Q v = λ v, then QT v = Q−1 v = λ−1 v and so λ−1 is an eigenvalue of QT . Exercise
8.2.34 says that a matrix and its transpose have the same eigenvalues.
(b) If Q v = λ v, then, by Exercise 5.3.16, k v k = k Q v k = | λ | k v k, and hence | λ | = 1.
Note that this proof also applies to complex eigenvalues/eigenvectors using the Hermi-
tian norm.
(c) If e i θ = cos θ + i sin θ is the eigenvalue, then Q x = cos θx − sin θy, Q y = sin θx + cos θy,
evv 9/9/04 425 °

while Q−1 x = cos θx − sin θy, Q−1 y = sin θx + cos θy. Thus, cos2 θk x k2 + 2 cos θ sin θx ·
y+sin2 θy2 = k Q−1 x k2 = k x k2 = k Qx k2 = cos2 θk x k2 −2 cos θ sin θx·y+sin2 θy2 and
so for x · y = 0 provided θ 6= 0, ± 12 π, ± π, . . . . Moreover, k x k2 = cos2 θk x k2 + sin2 θy2 ,
while, by the same calculation, k y k2 = sin2 θk x k2 + cos2 θy2 which imply k x k2 =
k y k2 . For θ = 21 π, we have Q x = −y, Q y = x, and so x · y = −(Q y) · (Q x) = x · y is
also zero.
♦ 8.2.41. (a) Prove that every 3×3 proper orthogonal matrix has +1 as an eigenvalue. (b) True
or false: An improper 3 × 3 orthogonal matrix has −1 as an eigenvalue.
Solution:
(a) According to Exercise 8.2.35, a 3 × 3 orthogonal matrix has at least one real eigenvalue,
which by Exercise 8.2.40 must be ±1. If the other two eigenvalues are complex conju-
gate, µ± i ν, then the product of the eigenvalues is ±(µ2 +ν 2 ). Since this must equal the
determinant of Q, which by assumption, is positive, we conclude that the real eigenvalue
must be +1. Otherwise, all the eigenvalues of Q are real, and they cannot all equal −1
as otherwise its determinant would be negative.
(b) True. It must either have three real eigenvalues of ±1, of which at least one must be −1
as otherwise its determinant would be +1, or a complex conjugate pair of eigenvalues
λ, λ, and its determinant is −1 = ± | λ |2 , so its real eigenvalue must be −1 and its com-
plex eigenvalues ± i .
♦ 8.2.42. (a) Show that the linear transformation defined by 3 × 3 proper orthogonal matrix cor-
responds to rotating through an angle around a line through the origin in R 3 — the axis of
0 1
the rotation. Hint: Use Exercise 8.2.41(a). 3 4
B 5 0 5 C
B C
B
(b) Find the axis and angle of rotation of the orthogonal matrix B − 4 12 3 C.
13 C 13 13 A
@
48 5 36
− 65 − 13 65
Solution: (a) The axis of the rotation is the eigenvector v corresponding to the eigenvalue +1.
Since Q v = v, the rotation fixes the axis, and hence must rotate around it. Choosing an or-
thonormal basis u1 , u2 , u3 , where u1 is a unit eigenvector in the direction of the axis of rota-
tion, while0u2 + i u3 is a complex
1
eigenvector for the eigenvalue e i θ . In this basis, Q has ma-
1 0 0
trix form B C
@ 0 cos θ − sin θ A, where θ is the angle of rotation. (b) The axis is the eigenvec-
0 1
0 sin θ cos θ
2 √
tor B C 7 2 30
@ −5 A for the eigenvalue 1. The complex eigenvalue is 13 + i 13 , and so the angle is
1
7
θ = cos−1 13 ≈ 1.00219.
8.2.43. Find all invariant subspaces, cf. Exercise 7.4.32, of a rotation in R 3 .

Solution: In general, besides the trivial invariant subspaces {0} and R 3 , the axis of rotation and
its orthogonal complement plane are invariant. If the rotation is by 180◦ , then any line in the
orthogonal complement plane is also invariant. If R = I , then every subspace is invariant.
8.2.44. Suppose Q is an orthogonal matrix. (a) Prove that K = 2 I − Q − QT is a positive
semi-definite matrix. (b) Under what conditions is K > 0?
Solution:
(a) (Q − I )T (Q − I ) = QT Q − Q − QT + I = 2 I − Q − QT = K and hence K is a Gram
matrix, which is positive semi-definite by Theorem 3.28.
(b) The Gram matrix is positive definite if and only if ker(Q − I ) = {0}, which means that
evv 9/9/04 426 °

Q does not have an eigenvalue of 1.
♦ 8.2.45. Prove that every proper affine isometry F (x) = Q x + b of R 3 , where det Q = +1,
is one of the following: (a) a translation x + b, (b) a rotation centered at some point of
R 3 , or (c) a screw consisting of a rotation around an axis followed by a translation in the
direction of the axis. Hint: Use Exercise 8.2.42.
Solution: If Q = I , then we have a translation. Otherwise, we can write F (x) = Q(x − c) + c
in the form of a rotation around the center point c provided we can solve (Q − I )c = b. By the
Fredholm alternative, this requires b to be orthogonal to coker(Q − I ) which is also spanned by
the rotation axis v, i.e., the eigenvector for the eigenvalue +1 of QT = Q−1 . More generally, we
write F (x) = Q(x − c) + c + t v and identify the affine map as a screw around the axis in the
direction of v passing through c.
♥ 8.2.46. Let Mn be the n × n tridiagonal matrix hose diagonal entries are all equal to 0 and
whose sub- and super-diagonal entries all equal 1. (a) Find the eigenvalues and eigenvec-
tors of M2 and M3 directly. (b) Prove that the eigenvalues and eigenvectors of Mn are ex-
plicitly given by
!T
kπ kπ 2kπ nkπ
λk = 2 cos , vk = sin , sin , ... sin , k = 1, . . . , n.
n+1 n+1 n+1 n+1
How do you know that there are no other eigenvalues?
Solution: 0 1
! !! 0 1 0
0 1 −11
(a) For M2 = : eigenvalues 1, −1; eigenvectors , . For M3 = B
@1 0 1CA:
1 0 1
1
0 1 0 1 0 1
0 1 0
√ √ √ 1 −1 √ 1
eigenvalues − 2, 0, 2; eigenvectors B C B C B
@ − 2 A, @ 0 A, @ 2 A.
C
1 1 1
(b) The j th entry of the eigenvalue equation Mn vk = λk vk reads
(j − 1) k π (j + 1) k π kπ jkπ
sin + sin = 2 cos sin ,
n+1 n+1 n+1 n+1
α−β α+β
which is a standard trigonometric identity: sin α + sin β = 2 cos sin . These
2 2
are all the eigenvalues because an n × n matrix has at most n eigenvalues.
♦ 8.2.47. Let a, b be fixed scalars. Determine the eigenvalues and eigenvectors of the n × n tridi-
agonal matrix with all diagonal entries equal to a and all sub- and super-diagonal entries
equal to b. Hint: See Exercises 8.2.18, 46.
Solution: We have A = a I +b Mn , so by Exercises 8.2.18 and 8.2.46 it has the same eigenvectors
kπ
as Mn , while its corresponding eigenvalues are a + b λk = a + 2 b cos for k = 1, . . . , n.
n+1
♥ 8.2.48. Find a formula for the eigenvalues of the tricirculant n × n matrix Z n that has 1’s on
the sub- and super-diagonals as well as its (1, n) and (n, 1) entries, while all other entries
are 0. Hint: Use Exercise 8.2.46 as a guide.
Solution:
!T
2kπ 2kπ 4kπ 6kπ 2 (n − 1) k π
λk = 2 cos , vk = cos , cos , cos , . . . , cos , 1 ,
n n n n n
for k = 1, . . . , n.
8.2.49. Let A be an n × n matrix with eigenvalues λ1 , . . . , λk , and B an m × m matrix with
evv 9/9/04 427 °

!
A O
eigenvalues µ1 , . . . , µl . Show that the (m+n)×(m+n) block diagonal matrix D =
O B
has eigenvalues λ1 , . . . , λk , µ1 , . . . , µl and no others. How are the eigenvectors related?
! ! ! !
v Av v v
Solution: Note first that if A v = λ v, then D = = λ , and so is
0 0 0 0
an eigenvector for D with!eigenvalue λ. Similarly, each eigenvalue µ and eigenvector w of B
0
gives an eigenvector of D. Finally, to check that D has no other eigenvalue, we compute
w
! ! !
v Av v
D = =λ and hence, if v 6= 0, then λ is an eigenvalue of A, while if w 6= 0,
w Bw w
then it must also be an eigenvalue for B.
!
a b
♥ 8.2.50. Let A = be a 2 × 2 matrix. (a) Prove that A satisfies its own characteristic
c d
equation, meaning pA (A) = A2 − (tr A) A + (det A) I = O. Remark : This result is a special
case of the Cayley–Hamilton Theorem, to be developed in Exercise 8.6.16.
(tr A) I − A
(b) Prove the inverse formula A−1 = when det A 6= 0. (c) Check the Cayley–
det A
!
2 1
Hamilton and inverse formulas when A = .
−3 2
Solution:
(a) Follows by direct computation.
(b) Multiply the characteristic equation by A−1 and rearrange terms.
(c) tr A = 4, det A = 7 and one checks A2 − 4 A + 7 I = O.
♥ 8.2.51. Deflation: Suppose A has eigenvalue λ and corresponding eigenvector v. (a) Let b be
any vector. Prove that the matrix B = A − v bT also has v as an eigenvector, now with
eigenvalue λ − β where β = v · b. (b) Prove that if µ 6= λ − β is any other eigenvalue of A,
then it is also an eigenvalue of B. Hint: Look for an eigenvector of the form w + c v, where
w is the eigenvector of A. (c) Given a nonsingular matrix A with eigenvalues λ1 , λ2 , . . . , λn
and λ1 6= λj , j ≥ 2, explain how to construct a deflated matrix B whose eigenvalues are
0 1
! 3 −1 0
3 3 B
0, λ2 , . . . , λn . (d) Try out your method on the matrices and @ −1 2 −1 C A.
1 5
0 −1 3
Solution:
(a) B v = (A − v bT )v = A v − (b · v)v = (λ − β)v.
(b) B (w + c v) = (A − v bT )(w + c v) = µ w + (c (λ − β) − b · w)v = µ (w + c v) provided
c = b · w/(λ − β − µ).
(c) Set B = A−λ1 v1 bT where v1 is the first eigenvector of A and b is any vector such that
b · v1 = 1. For example, we can set b = v1 /k v1 k2 . (Weilandt deflation, [ 12 ], chooses
b = rj /(λ1 v1,j ) where v1,j is any nonzero entry of v1 and rj is the corresponding row
of A.) ! !
1 −3
(d) (i) The eigenvalues of A are 6, 2 and the eigenvectors , . The deflated ma-
1 1
! ! !
λ v vT 0 0 1 0
trix B = A − 1 1 21 = has eigenvalues 0, 2 and eigenvectors , .
k v1 k −2 2 1 1
0 1 0 1 0 1
1 −1 1
(ii) The eigenvalues of A are 4, 3, 1 and the eigenvectors @ −1 A, @ 0 C
B C B B C
A @ 2 A. The de-
,
1 1 1
evv 9/9/04 428 °

0 1
5 1
B − 43C
λ v vT B 3 3 C
flated matrix B = A − 1 1 21 = B
B
1 2 1C has eigenvalues 0, 3, 1 and the
k v1 k @ 3 3 3CA
4 1 5
0 1 0 1 0 1
− 3 3 3
1 −1 1
eigenvectors B C B C B C
@ −1 A, @ 0 A, @ 2 A.
1 1 1
diag 8.3. Eigenvector Bases and Diagonalization.
8.3.1. Which of the following are complete eigenvalues for the

! indicated matrix? !
What is the
2 1 2 1
dimension of the associated eigenspace? (a) 3, , (b) 2, ,
1 2 0 2
0 1 0 1 0 1
0 0 −1 1 1 −1 −1 −4 −4
(c) 1, B
@1 1 0CA, (d) 1, B
@1 1 0CA, (e) −1, B @ 0 −1 0CA,
1 0 0 1 1 2 0 4 3
0 1
0 1
0 1 1 −1 −1 −1 −1
1 0 −1 1 B0
−i 1 0 B 1 −1 −1 −1 C
B C B 0 1 0 1CC
B
B
C
(f ) − i , @−i 1 −1 A, (g) −2, B
@ −1
C, (h) 1, B0 0 1 −1 −1 C
C.
1 1 0A B
@0
C
0 0 −i 0 0 1 −1 A
1 0 0 1
0 0 0 0 1
Solution:
(a) Complete. dim = 1 with basis ( 1, 1 )T .
(b) Not complete. dim = 1 with basis ( 1, 0 )T .
(c) Complete. dim = 1 with basis ( 0, 1, 0 )T .
(d) Not an eigenvalue.
(e) Complete. dim = 2 with basis ( 1, 0, 0 )T , ( 0, −1, 1 )T .
(f ) Complete. dim = 1 with basis ( i , 0, 1 )T .
(g) Not an eigenvalue.
(h) Not complete. dim = 1 with basis ( 1, 0, 0, 0, 0 )T .
8.3.2. Find the eigenvalues and a basis for the each of the eigenspaces of the following matri-
ces. Which are complete? 0 1
! ! ! ! 4 −1 −1
4 −4 6 −8 3 −2 i −1
(a) , (b) , (c) , (d) , (e) B
@0 3 0CA,
1 0 4 −6 4 −1 1 i
0 1
1 0
−1 2 1
0 1 0 1 1 0 0 0 −1 0 1 2
−6 0 −8 −2 1 −1 B
B 0 1 0 0C B
B 0 1 0 1C
(f ) B C B
@ −4 2 −4 A, (g) @ 5 −3 6CA, (h) B
@ −1 1 −1 0 A
C
C, (i) B
@ −1 −4 1 −2 A
C
C.
4 0 6 5 −1 4
1 0 −1 0 0 1 0 1
Solution: !
2
(a) Eigenvalue: 2. Eigenvector: . Not complete.
1
! !
2 1
(b) Eigenvalues: 2, −2. Eigenvectors: , . Complete.
1 1
evv 9/9/04 429 °

!
1± i
(c) Eigenvalues: 1 ± 2 i . Eigenvectors: . Complete.
2
! !
−i i
(d) Eigenvalues: 0, 2 i . Eigenvectors: , . Complete.
1 1
0 1 0 1
1 1
(e) Eigenvalue 3 has eigenspace basis B C B C
@ 1 A, @ 0 A. Not complete.
0
0 1 0 1 1 0 1
−1 0 −2
(f ) Eigenvalue 2 has eigenspace basis B C B C B C
@ 0 A, @ 1 A. Eigenvalue −2 has @ −1 A. Complete.
0 1
1 0 0 1
1
0 −1
(g) Eigenvalue 3 has eigenspace basis B C B C
@ 1 A. Eigenvalue −2 has @ 1 A. Not complete.
0 1
1 0 1
1 0 1 0 1
0 0 1 1
B C B C B C B C
B0C B0C B3C B1C
(h) Eigenvalue 0 has B @0A
C. Eigenvalue −1 has B C. Eigenvalue 1 has B C, B C. Complete.
@1A @1A @0A
1 0 1 0
1
1 0
01 1
1 2 −1
B C B
0 C B −1 C B
B 1C
C
(i) Eigenvalue 0 has eigenspace basis B B C, B
@1A @ 0A
C
C. Eigenvalue 2 has B
@ −5 A
C. Not complete.
0 1 1
8.3.3. Which of the following matrices admit eigenvector bases of R n ? For those that do, ex-
n
hibit such a basis. If not, what is the dimension of the subspace of
0 R spanned1by the
! ! ! 1 −2 0
1 3 1 3 1 3
eigenvectors? (a) , (b) , (c) , (d) B
@ 0 −1 0CA,
3 1 −3 1 0 1
40 −4 −1 1
0 1 0 1 0 1 0 0 −1 1
1 −2 0 2 0 0 0 0 −1 B
B 0 −1 0 1C
(e) B
@ 0 −1 0C B
A, (f ) @ 1 −1 1C B
A, (g) @ 0 1 0CA, (h) B
@1
C
C.
0 −1 0 A
0 −4 −1 2 1 −1 1 0 0
1 0 1 0
Solution: ! !
−1 1
(a) Eigenvalues: −2, 4; the eigenvectors , form a basis for R 2 .
1 1
! !
i −i
(b) Eigenvalues: 1−3 i , 1+3 i ; the eigenvectors , , are not real, so the dimension
1 1
is 0. !
1
(c) Eigenvalue: 1; there is only one eigenvector v1 = spanning a one-dimensional sub-
0
space of R 2 . 0 1 0 1
1 1
B C B C
(d) The eigenvalue 1 has eigenvector @ 0 A, while the eigenvalue −1 has eigenvectors @ 1 A,
0 1
2 0
0
B C 3
@ −1 A. The eigenvectors form a basis for R .
0 0 1 0 1
1 0
(e) The eigenvalue 1 has eigenvector B C
@ 0 A, while the eigenvalue −1 has eigenvector
B C
@ 0 A.
0 1
The eigenvectors span a two-dimensional subspace of R 3 .
evv 9/9/04 430 °

0 1 0 1 0 1
0 0 8
B C B C B C
(f ) λ = −2, 0, 2. The eigenvectors are @ −1 A, @ 1 A, and @ 5 A, forming a basis for R 3 .
1 1 7
1 0 0 1 0 1
−i i 0
(g) the eigenvalues are i , − i , 1. The eigenvectors are B C
@ 0 A,
B
@ 0CA and
B C
@ 1 A. The real
1 1 0
eigenvectors span only a one-dimensional subspace of R 3 . 0 1 0 1 0 1 0 1
0 4 −1 −1
B C B C B C B C
B1C B3C B i C B−i C
(h) The eigenvalues are −1, 1, − i −1, − i +1. The eigenvectors are B
@0A
C, B C,
@2A
B C, B
@−i A @ i A
C.
0 6 1 1
The real eigenvectors span a two-dimensional subspace of R 4 .
8.3.4. Answer Exercise 8.3.3 with R n replaced by C n .
Solution: Cases (a,b,d,f,g,h) all have eigenvector bases of C n .
8.3.5. (a) Give an example of a 3 × 3 matrix that only has 1 as an eigenvalue, and has only one
linearly independent eigenvector. (b) Give an example that has two linearly independent
eigenvectors.
0 1 0 1
1 1 0 1 0 0
Solution: (a) B
@0 1 1CA, (b)
B
@0 1 1CA.
0 0 1 0 0 1
8.3.6. True or false: (a) Every diagonal matrix is complete. (b) Every upper triangular ma-
trix is complete.
Solution: (a) True. The standard basis vectors are eigenvectors. (b) False. The Jordan matrix
1 101 is incomplete since e1 is the only eigenvector.
8.3.7. Prove that if A is a complete matrix, so is c A+d I , where c, d are any scalars. Hint: Use
Exercise 8.2.18.
Solution: According to Exercise 8.2.18, every eigenvector of A is an eigenvector of c A + d I with
eigenvalue c λ + d, and hence if A has a basis of eigenvectors, so does c A + d I .
8.3.8. (a) Prove that if A is complete, so is A2 . (b) Give an example of an incomplete matrix
A such that A2 is complete.
Solution: (a) Every eigenvector of A is an eigenvector of!A2 with eigenvalue λ2 , and hence if A
0 1
has a basis of eigenvectors, so does A2 . (b) A = with A2 = O.
0 0
♦ 8.3.9. Suppose v1 , . . . , vn forms an eigenvector basis for the complete matrix A, with λ1 , . . . , λn
the corresponding eigenvalues. Prove that every eigenvalue of A is one of the λ 1 , . . . , λn .
n
X n
X
Solution: Suppose A v = λv. Write v = ci vi . Then A v = ci λi vi and hence, by linear
i=1 i=1
independence, λi ci = λ ci . THus, either λ = λi or ci = 0. Q.E.D.
n n
8.3.10. (a) Prove that if λ is an eigenvalue of A, then λ is an eigenvalue of A . (b) State
and prove a converse if A is complete. Hint: Use Exercise 8.3.9. (The completeness hy-
pothesis is not essential, but this is harder, relying on the Jordan canonical form.)
Solution: (a) If A v = λ v, then, by induction, An v = λn v, and hence v is an eigenvector
with eigenvalue λn . (b) √
Conversely, if A is complete and An has eigenvalue µ, then there is a
th
(complex) n root λ = n µ that is an eigenvalue of A. Indeed, the eigenvector basis of A is an
evv 9/9/04 431 °

eigenvector basis of An , and hence, using Exercise 8.3.9, every eigenvalue of An is the nth power
of an eigenvalue of A.
♦ 8.3.11. Show that if A is complete, then any similar matrix B = S −1 A S is also complete.
Solution: As in Exercise 8.2.31, if v is an eigenvector of A then S −1 v is an eigenvector of B.
Moreover, if v1 , . . . , vn form a basis, so do S −1 v1 , . . . , S −1 vn ; see Exercise 2.4.21 for details.
8.3.12. Let U be an upper triangular matrix with all its diagonal entries equal. Prove that U
is complete if and only if U is a diagonal matrix.
Solution: According to Exercise 8.2.16, its only eigenvalue is λ, the common value of its diag-
onal entries, and so all eigenvectors belong to ker(U − λ I ). Thus U is complete if and only if
dim ker(U − λ I ) = n if and only if U − λ I = O. Q.E.D.
8.3.13. Show that each eigenspace of an n × n matrix A is an invariant subspace, as defined in
Exercise 7.4.32.
Solution:
Let V = ker(A − λ I ). If v ∈ V , then A v ∈ V since (A − λ I )A v = A(A − λ I )v = 0. Q.E.D.
♦ 8.3.14. (a) Prove that if v = x ± i y is a complex conjugate pair of eigenvectors of a real matrix
A corresponding to complex conjugate eigenvalues µ ± i ν with ν 6= 0, then x and y are
linearly independent real vectors. (b) More generally, if vj = xj ± i yj , j = 1, . . . , k are
complex conjugate pairs of eigenvectors corresponding to distinct pairs of complex conju-
gate eigenvalues µj ± i νj , νj 6= 0, then the real vectors x1 , . . . , xk , y1 , . . . , yk are linearly
independent.
Solution:
(a) Let µ ± i ν be the corresponding eigenvalues. Then the complex eigenvalue equation
A(x + i y) = (µ + i ν)(x + i y) implies that
A x = µx − ν y and A y = ν x + µ y.
Now, suppose c x + d y = 0 for some (c, d) 6= (0, 0). Then
0 = A(c x + d y) = (c µ + d ν)x + (− c ν + d µ)y
The determinant of the coefficient ! matrix of these two sets of equations for x, y is
c d
det = −(c2 + d2 )ν 6= 0 because we are assuming the eigenvalues
cµ + dν −cν + dµ
are truly complex. This implies x = y = 0, which contradicts the fact that we have an
eigenvector.
(b) This is proved by induction. Suppose we know x1 , . . . , xk−1 , y1 , . . . , yk−1 are lin-
early independent. If x1 , . . . , xk , y1 , . . . , yk were linearly dependent, there would ex-
ist (ck , dk ) 6= (0, 0) such that . zk = ck xk + dk yk is some linear combination of
x1 , . . . , xk−1 , y1 , . . . , yk−1 . But then A zk = (ck µ + dk ν)xk + (− ck ν + dk µ)yk is also a
linear combination, as is A2 zk = (ck (µ2 − ν 2 ) + 2 dk µ ν)xk + (−2 ck µ ν + dk (µ2 − ν 2 ))yk .
Since the coefficient matrix of the first two such vectors is nonsingular, a suitable linear
combination would vanish: 0 = a zk + b A zk + c A2 zk , which would give a vanishing
linear combination of x1 , . . . , xk−1 , y1 , . . . , yk−1 , which, by the induction hypothesis,
must be trivial. A little more work demonstrates that this implies zk = 0, and so, in
contradiction to part (b), would imply xk , yk are linearly dependent. Q.E.D.
! ! !
3 −9 5 −4 −4 −2
8.3.15. Diagonalize the following matrices: (a) , (b) , (c) ,
2 −6 2 −1 5 2
evv 9/9/04 432 °

0 1 0 1 0 1 0 1
−2 3 1 8 0 −3 3 3 5 2 5 5
B
(d) @ 0 1 −1 C
A, (e)
B
@ −3 0 −1 C
A, (f )
B
@ 5 6 5CA, (g)
B
@0 2 0CA,
0 0 3 3 0 −2 −5 −8 −7 0 −5 −3
0 1 0 1 0 1
1 0 −1 1 0 0 1 0 2 1 −1 0
B
B0 2 −1 1CC
B
B0 0 0 1CC
B
B −3 −2 0 1CC
(h) B
@0
C, (i) B C, (j) B C.
0 −1 0A @1 0 0 0A @ 0 0 1 −2 A
0 0 0 −2 0 1 0 0 0 0 1 −1
Solution: ! !
3 3 0 0
(a) S = ,D= .
1 2 0 −3
! !
2 1 3 0
(b) S = ,D= .
1 1 0 1
! !
3 1 3 1 −1 + i 0
(c) S = − 5 + 5 i − 5 − 5 i ,D= .
1 1 0 −1 − i
0 1 0 1
1
B
1 1 − 10 C
−2 0 0
(d) S =B 1 C, D = B @ 0 1 0C A.
@0 1 −2 A
0 0 11 0 0 3
0 0 1
0 21 1 0 0 0
(e) S =B C
@ 1 −10 6 A, D = @ 0 7
B
0CA.
0
0 7 3 0 0 1
−1 0 1
− 51 − 53 i − 15 + 53 i −1 2 + 3i 0 0
(f ) S =B@ −1 −1
C
0 A, D = @ 0
B
2 − 3i 0 CA.
0
1 1
1 0
1 1
0 0 −2
1 0 −1 2 0 0
B C B
(g) S = @ 0 −1 0 A, D = @ 0 2 0CA.
0
0 1 1 1 0
0 0 −3 1
−4 3 1 0 −2 0 0 0
B C B C
−3 2 0 1 C B 0 −1 0 0 C
(h) S =BB
@ 0 6 0 0A
C, D = B
@ 0
C.
0 1 0A
0
12 0 0 0 1 0
0 0 0 2 1
0 −1 0 1 −1 0 0 0
B C B
−1 0 1 0 0 −1 0 0C
(i) S =BB
@ 0
C
C, D = B
B C
C.
1 0 1A @ 0 0 1 0A
0
1 0 1 0 0 10 0 1
3 0 1
−1 1 2i − 23 i 1 0 0 0
B C B
B 1 −3 − 1 − 2 i − 21 + 2 i C C, D = B 0 −1 0 0CC
(j) S =BB 2 C B C.
@ 0 0 1+ i 1− i A @ 0 0 i 0 A
0 0 1 1 0 0 0 −i
!
1 1
8.3.16. Diagonalize the Fibonacci matrix F = .
1 0
0 √ √ 10 √ 10 √ 1
! √1
1+ 5 1− 5 1+ 5
0 CB − 1−√ 5
1 1 B C B 5 2 √5 C
Solution: = @ 2 2 A@ 2 √ AB@
C.
1 0 1− 5 1 1+√ 5 A
1 1 0 2 −√
5 2 5
!
0 −1
8.3.17. Diagonalize the matrix of rotation through 90◦ . How would you interpret
1 0
the result?
! !
i −i i 0
Solution: S = ,D = . A rotation does not stretch any real vectors, but
1 1 0 −i
evv 9/9/04 433 °

somehow corresponds to two complex stretches. 0 1
0 1 5 12
0 −1 0 B 13
0 13 C
8.3.18. Diagonalize the rotation matrices (a) B
@1 0 0CA, (b) B
B 0
@
1 0 C
C
A
.
0 0 1 12
− 13 0 5
13
Solution:
0 1 0 10 10 1
i 1
0 −1 0 i −i 0 i 0 0 B−2 2 0
C
(a) B
@1 0 0C B
A=@1 1 0C B
A@ 0 −i 0CAB i
@ 2
1
2 0CA,
0 0 1 0 0 1 0 0 1 0 0 1
2
0 1 0 10 10 1
5 12 5+12 i i 1
B 13 0 13 C B −i i 0CB 0 0C
CB 2 0 2 C
B C B CBB 13 CB C
B
(b) B
@
0 1 0CC
A
=B
B
@
0 0 1C
CB
AB 0 5−12 i CB
0C B − i
2 0 1
2
C.
C
@ 13 A@ A
− 12
13 0 5
13 1 1 0 0 0 1 0 1 0
! !
1 1 2 −2
8.3.19. Which of the following matrices have real diagonal forms? (a) , (b) ,
1 −3 1 4
0 1
0 1 0 1 0 1 1 0 0 0
0 1 0 0 3 2 3 −8 2 B
B 0 1 0 0C
(c) @ −1 0 1C B
A, (d) @ −1 1 −1 C B
A, (e) @ −1 2 2C B
A, (f ) B
@ 0
C
C.
0 1 0A
1 1 0 1 −3 −1 1 −4 2
−1 −1 −1 −1
Solution:
(a) Yes, distinct real eigenvalues −3,
√ 2.
(b) No, complex eigenvalues 1 ± i 6.√
(c) No, complex eigenvalues 1, − 21 ± 25 i .
(d) No, incomplete eigenvalue 1 (and complete eigenvalue −2).
(e) Yes, distinct real eigenvalues 1, 2, 4.
(f ) Yes, complete real eigenvalues 1, −1.
8.3.20. Diagonalize the following complex matrices: 0 1
! ! ! −i 0 1
i 1 1− i 0 2− i 2+ i
(a) , (b) , (c) , (d) B
@−i 1 −1 C
A.
1 i i 2+ i 3− i 1+ i
1 0 −i
Solution: ! !
1 −1 1+ i 0
(a) S = ,D= .
1 1 0 −1 + i
! !
−2 + i 0 1− i 0
(b) S = ,D= .
1 1 0 2+ i
! !
1 1 4 0
(c) S = 1 −2 − 2 i , D = .
1 1 0 −1
0 1 0 1
0 1 −1 1 0 0
(d) S =B@ 1 −1 − i
3
5 + 1 C
5 i A , D = B
@ 0 1 − i 0 C
A.
0 1 1 0 0 −1 − i
8.3.21. Write down a real matrix that has (a) eigenvalues −1, 3 and corresponding
0 1 0
eigenvec-
1 0 1
! ! −1 2 0
−1 1
tors , , (b) eigenvalues 0, 2, −2 and associated eigenvectors @ 1 A, @ −1 C
B C B B C
A, @ 1 A ;
2 1
! !
0 1 3
2 1
(c) an eigenvalue of 3 and corresponding eigenvectors , ; (d) an eigenvalue
−3 2
evv 9/9/04 434 °

!
1+ i
−1 + 2 i and corresponding eigenvector ; (e) an eigenvalue −2 and correspond-
3i
0 1 0 1
2 1
ing eigenvector B C
@ 0 A; (f ) an eigenvalue 3 + i and corresponding eigenvector
B
@ 2i C
A.
−1 −1 − i
Solution: Use A = S Λ S −1 . For parts (e), (f ) you can choose 0
any other eigenvalues
1
and eigen-
! 6 6 −2 !
7 4 3 0
vectors you want to fill in S and Λ. (a) , (b) B@ −2 −2 0 C
A , (c) ,
−8 −5 0 3
0 1
6 6 0
−4 1
! 0 0 4 1
4 3 2 0C
(d) 1 − 3 , (e) example: @ 0 1 B C B
0 A, (f ) example: @ −2 3 0 A.
6 −3 0 0 −2 −2 −2 !
0 !
1 2
8.3.22. A matrix A has eigenvalues −1 and 2 and associated eigenvectors and .
2 3
Write down the matrix form of the linear transformation L[ u ] = A u in terms of (a) !
the!
1 3
standard basis e1 , e2 ; (b) the basis consisting of its eigenvectors; (c) the basis , .
1 4
0 1
! !
11 −6 −1 0 B −4 −6 C
Solution: (a) (b) , (c) @ A.
18 −10 0 2
3 5
♦ 8.3.23. Prove that two complete matrices A, B have the same eigenvalues (with multiplicities)
if and only if they are similar, i.e., B = S −1 A S for some nonsingular matrix S.
Solution: Let S1 be the eigenvector matrix for A and S2 the eigenvector matrix for B. Thus, by
the hypothesis S1−1 A S1 = Λ = S2−1 B S2 and hence A = S1 S2−1 B S2 S1−1 = S −1 B S where
S = S2 S1−1 .
8.3.24. Let B be obtained from A by permuting both its rows and columns using the same
permutation π, so bij = aπ(i),π(j) . Prove that A and B have the same eigenvalues. How are
their eigenvectors related?
Solution: Let v = ( v1 , v2 , . . . , vn )T be an eigenvector for A. Let v
e be obtained by applying the
permutation to the entries of v, so vei = vπ(i) . Then the ith entry of B v e is
n
X n
X n
X
bij vej = aπ(i),π(j) vπ(j) = aπ(i),j vj = λ vπ(i) = λ vei ,
j =1 j =1 j =1
and hence v
e is an eigenvector of B with eigenvalue λ. Q.E.D.
8.3.25. True or false: If A is a complete upper triangular matrix, then it has an upper trian-
gular eigenvector matrix S.
Solution: True. Let λj = ajj denote the j th diagonal entry of A, which is the same as the j th
eigenvalue. We will prove that the corresponding eigenvector is a linear combination of e 1 , . . . , ej ,
which is equivalent to the eigenvector matrix S being upper triangular. We use induction on
the size n. Since A is upper triangular, it leaves the subspace V spanned by e1 , . . . , en−1 invari-
ant, and hence its restriction is an (n − 1) × (n − 1) upper triangular matrix. Thus, by induction
and completeness, A possesses n − 1 eigenvectors of the required form. The remaining eigenvec-
tor vn cannot belong to V (otherwise the eigenvectors would be linearly dependent) and hence
must involve en . Q.E.D.
8.3.26. Suppose the n × n matrix A is diagonalizable. How many different diagonal forms does
it have?
evv 9/9/04 435 °

Solution: The diagonal entries are all eigenvalues, and so are obtained from each other by per-
mutation. If all eigenvalues are distinct, then there are n ! different diagonal forms — otherwise,
n!
if it has distinct eigenvalues of multiplicities j1 , . . . jk , there are distinct diagonal
j1 ! · · · j k !
forms.
8.3.27. Characterize all complete matrices that are their own inverses: A −1 = A. Write down
a non-diagonal example.
Solution:!A2 = I if and only if D 2 = I , and so all its! eigenvalues
! are ±1. Examples: A = !
3 −2 1 1 0 1
with eigenvalues 1, −1 and eigenvectors , ; or, even simpler, A = .
4 −3 1 2 1 0
♥ 8.3.28. Two n × n matrices A, B are said to be simultaneously diagonalizable if there is a non-
singular matrix S such that both S −1 A S and S −1 B S are diagonal matrices. (a) Show
that simultaneously diagonalizable matrices commute: A B = B A. (b) Prove that the
converse is valid provided one of the matrices has no multiple eigenvalues. (c) Is every pair
of commuting matrices simultaneously diagonalizable?
Solution: (a) If A = S Λ S −1 and B = S D S −1 where Λ, D are diagonal, then A B = S Λ D S −1 =
S D Λ S −1 = B A, since diagonal matrices commute. (b) According to Exercise 1.2.12(e), the
only matrices that commute with an n × n diagonal matrix with distinct entries is another di-
agonal matrix. Thus, if A B = B A, and A = S Λ S −1 where all entries of Λ are distinct, then
!
0 1
D = S −1 B S commutes with Λ and hence is a diagonal matrix. (c) No, the matrix
0 0
commutes with the identity matrix, but is not diagonalizable. See also Exercise 1.2.14.
evsym 8.4. Eigenvalues of Symmetric Matrices.
8.4.1. Find the eigenvalues and an orthonormal eigenvector basis for the following symmetric
matrices: 0 1 0 1
! ! ! 1 0 4 6 −4 1
2 6 5 −2 2 −1 B C B
(a) (b) , (c) , (d) @ 0 1 3 A, (e) @ −4 6 −1 C A.
6 −7 −2 5 −1 5
4 3 1 1 −1 11
Solution: ! !
1 2 1 1
(a) eigenvalues: 5, −10; eigenvectors: √ ,√ ,
5 ! 1 5 !−2
1 −1 1 1
(b) eigenvalues: 7, 3; eigenvectors: √ ,√ .
2 1 2 1
√ √
7 + 13 7 − 13
(c) eigenvalues: , ;
2 2 0 √ 1 0 √ 1
2 3− 13 2 3+ 13
B C B C
eigenvectors: q √ @ 2 A, q √ @ 2 A.
26 − 6 13 1 26 + 6 13 1
evv 9/9/04 436 °

0 1 0 1
4 0 1 4
B √ 3 − √
B 5 2C B − 5C B
C
B 5 2C C
B C
3 C, B B C B 3 C
(d) eigenvalues: 6, 1, −4; eigenvectors: B √ 4 C, B − √ C.
B 5 2C B 5C B 5 2C
B C @ A B C
@ A @ A
√1 0 √1
0 2 1 0 1 0 2
1
B √1 C B − √1 C √1
B 6C B 3C B 2C
B C B C B C
(e) eigenvalues: 12, 9, 2; eigenvectors: B
B
1 C
− √ C, B B 1
√ C ,B
B 1
√ C
C
.
B 6C B 3CC B 2C
@ A @ A @ A
√2 √1 0
6 3
8.4.2. Determine whether the following symmetric matrices are positive definite by computing
their eigenvalues. Validate your conclusions by using the methods from Chapter 4.
0 1 0 1
! ! 1 −1 0 4 −1 −2
2 −2 −2 3 B C B
(a) (b) , (c) @ −1 2 −1 A, (d) @ −1 4 −1 C A.
−2 3 3 6
0 −1 1 −2 −1 4
Solution: √
(a) eigenvalues 52 ± 21 17; positive definite
(b) eigenvalues −3, 7; not positive definite
(c) eigenvalues 0, 1, 3;√
positive semi-definite
(d) eigenvalues 6, 3 ± 3; positive definite.
8.4.3. Prove that a symmetric matrix is negative definite if and only if all its eigenvalues are
negative.
Solution: Use the fact that K = − N is positive definite and so has all positive eigenvalues. The
eigenvalues of N = − K are − λj where λj are the eigenvalues of K. Alternatively, mimic the
proof in the book for the positive definite case.
8.4.4. How many orthonormal eigenvector bases does a symmetric n × n matrix have?
Solution: If all eigenvalues are distinct, there are 2n different bases, governed by the choice of
sign in the unit eigenvectors ± uk . If the eigenvalues are repeated, there are infinitely many,
since any orthonormal basis of each eigenspace will contribute to an orthonormal eigenvector
basis of the matrix. !
a b
8.4.5. Let A = . (a) Write down necessary and sufficient conditions on the entries
c d
a, b, c, d that ensures that A has only real eigenvalues. (b) Verify that all symmetric 2 × 2
matrices satisfy your conditions.
Solution:
(a) The characteristic equation p(λ) = λ2 − (a + d)λ + (a d − b c) = 0 has real roots if and
only if its discriminant is non-negative: 0 ≤ (a + d)2 − 4 (a d − b c) = (a − d)2 + 4 b c, which
is the necessary and sufficient condition for real eigenvalues.
(b) If A is symmetric, then b = c and so the discriminant is (a − d)2 + 4 b2 ≥ 0.
♥ 8.4.6. Let AT = − A be a real, skew-symmetric n × n matrix. (a) Prove that the only possi-
ble real eigenvalue of A is λ = 0. (b) More generally, prove that all eigenvalues λ of A are
purely imaginary, i.e., Re λ = 0. (c) Explain why 0 is an eigenvalue of A whenever n is
odd. (d) Explain why, if n = 3, the eigenvalues of A 6= O are 0, i ω, − i ω for some real ω.
(e) Verify these facts for
0
the particular
1
matrices0 1
0 1
! 0 0 2 0
0 3 0 0 1 −1 B
0 −2 0 0 0 −3 C
(i) , (ii) B@ −3 0 −4 A,
C
(iii) B
@ −1 0 −1 A,
C
(iv ) B
B
C
C.
2 0 @ −2 0 0 0A
0 4 0 1 1 0
0 3 0 0
evv 9/9/04 437 °

Solution:
(a) If A v = λ v and v 6= 0 is real, then
λk v k2 = (A v) · v = (A v)T v = vT AT v = − vT A v = − v · (A v) = − λk v k2 ,
and hence λ = 0.
(b) Using the Hermitian dot product,
λ k v k2 = (A v) · v = vT AT v = − vT A v = − v · (A v) = − λ k v k2 ,
and hence λ = − λ, so λ is purely imaginary.
(c) Since det A = 0, cf. Exercise 1.9.10, at 0
least one of the1eigenvalues of A must be 0.
0 c −b
(d) The characteristic polynomial of A = B @ −c 0 aC 3 2 2
A is − λ + λ(a + b + c ) and
2
√ b −a 0
2 2 2
hence the eigenvalues are 0, ± i a + b + c , and so are √ all zero if and only if A = O.
(e) The eigenvalues are: (i) ± 2 i , (ii) 0, ± 5 i , (iii) 0, ± 3 i , (iv ) ± 2 i , ± 3 i .
♥ 8.4.7. (a) Prove that every eigenvalue of a Hermitian matrix A, so AT = A as in Exercise
3.6.49, is real. (b) Show that the eigenvectors corresponding to distinct eigenvalues are
orthogonal under the Hermitian dot product on C n . (c) Find the eigenvalues and eigen-
vectors of the following Hermitian matrices, and verify orthogonality:
0 1
! ! 0 i 0
2 i 3 2− i
(i) , (ii) , (iii) B
@−i 0 iC A.
− i −2 2+ i −1
0 −i 0
Solution:
(a) Let A v = λ v. Using the Hermitian dot product,
λk v k2 = (A v) · v = vT AT v = vT A v = v · (A v) = λk v k2 ,
and hence λ = λ, which implies that the eigenvalue λ is real.
(b) Let A v = λ v, A w = µ w. Then
λv · w = (A v) · w = vT AT w = vT A w = v · (A w) = µv · w,
since µ is real. Thus, if λ 6= µ then v · w = 0. √ ! √ !
√ (2 − 5) i (2 + 5) i
(c) (i) eigenvalues ± 5; eigenvectors: , .
!
1 !
1
2− i −2 + i
(ii) eigenvalues 4, −2; eigenvectors: , .
1 5
0 1 0 1 0 1
√ 1 −1
√ −1
√
(iii) eigenvalues 0, ± 2; eigenvectors: B C B C B
@ 0 A, @ i 2 A , @ − i 2 A.
C
1 1 1
♥ 8.4.8. Let M > 0 be a fixed positive definite n × n matrix. A nonzero vector v 6= 0 is called a
generalized eigenvector of the n × n matrix K if
K v = λ M v, v 6= 0, (8.31)
where the scalar λ is the corresponding generalized eigenvalue. (a) Prove that λ is a gen-
eralized eigenvalue of the matrix K if and only if it is an ordinary eigenvalue of the ma-
trix M −1 K. How are the eigenvectors related? (b) Now suppose K is a symmetric ma-
trix. Prove that its generalized eigenvalues are all real. Hint: First explain why this does
not follow from part (a). Instead mimic the proof of part (a) of Theorem 8.20, using the
weighted Hermitian inner product h v , w i = v T M w in place of the dot product. (c) Show
that if K > 0, then its generalized eigenvalues are all positive: λ > 0. (d) Prove that the
eigenvectors corresponding to different generalized eigenvalues are orthogonal under the
weighted inner product h v , w i = v T M w. (e) Show that, if the matrix pair K, M has n
distinct generalized eigenvalues, then the eigenvectors form an orthogonal basis for R n . Re-
evv 9/9/04 438 °

mark : One can, by mimicking the proof of part (c) of Theorem 8.20, show that this holds
even when there are repeated generalized eigenvalues.
Solution: (a) Rewrite (8.31) as M −1 K v = λ v, and so v is an eigenvector for M −1 K with
eigenvalue λ. (b) If v is an generalized eigenvector, then since K, M are real matrices, K v =
λ M v. Therefore,
λ k v k2 = λ vT M v = (λ M v)T v = (K v)T v = vT (K v) = λ vT M v = λ k v k2 ,
and hence λ is real. (c) If K v = λ M v, K w = µ M w, with λ, µ and v, w real, then
λ h v , w i = (λ M v)T w = (K v)T w = vT (K w) = µ vT M w = µ h v , w i,
and so if λ 6= µ then h v , w i = 0, proving orthogonality. (d) If K > 0, then λ h v , v i =
vT (λ M v) = vT K v > 0, and so, by positive definiteness of M , λ > 0. (e) Part (b) proves
that the eigenvectors are orthogonal with respect to the inner product induced by M , and so
the result follows immediately from Theorem 5.5.
8.4.9. Compute the generalized eigenvalues and eigenvectors, as in (8.31), for the following
matrix pairs. Verify orthogonality of the eigenvectors under the appropriate inner product.
! ! ! !
3 −1 2 0 3 1 2 0
(a) K = , M= , (b) K = , M= ,
−1 2 0 3 1 1 0 1
0 1 0 1
! ! 6 −8 3 1 0 0
2 −1 2 −1 B
(c) K = , M= , (d) K = @ −8 24 −6 C B
A, M = @0 4 0CA,
−1 4 −1 1
3 −6 99 0 0 9
0 1 0 1 0 1 0 1
1 2 0 1 1 0 5 3 −5 3 2 −3
(e) K = B
@2 8 2C B
A, M = @1 3 1C B
A , (f ) K = @ 3 3 −1 C B
A, M = @ 2 2 −1 C
A.
0 2 1 0 1 1 −5 −1 9 −3 −1 5
Solution: ! !
−3 1
5 1
(a) eigenvalues: 3, 2; eigenvectors: , 2 ;
1 1!
!
1 1
1
(b) eigenvalues: 2, 2 ; eigenvectors: , − 2 ;
1 1
! !
1 1
(c) eigenvalues: 7, 1; eigenvectors: 2 , ;
10 0
1 0 1 0 1
6 −6 2
(d) eigenvalues: 12, 9, 2; eigenvectors: @ −3 A, @ 3 C
B C B B C
A, @ 1 A ;
0
41 0 21 0 0 1
1 −1 1
(e) eigenvalues: 3, 1, 0; eigenvectors: B C B C B 1C
@ −2 A, @ 0 A, @ − 2 A;
1 1
0 1 0
1 1
1 −1
(f ) 2 is a double eigenvalue with eigenvector basis B C B C
@ 0 A, @ 1 A, while 1 is a simple eigen-
0 1
1 0
2
value with eigenvector B C
@ −2 A. For orthogonality you need to select an M orthogonal
1
basis of the two-dimensional eigenspace, say by using Gram–Schmidt.
♦ 8.4.10. Let L = L∗ : R n → R n be a self-adjoint linear transformation with respect to the inner
product h · , · i. Prove that all its eigenvalues are real and the eigenvectors are orthogonal.
Hint: Mimic the proof of Theorem 8.20 replacing the dot product by the given inner product.
evv 9/9/04 439 °

Solution: If L[ v ] = λ v, then, using the inner product,
λ k v k2 = h L[ v ] , v i = h v , L[ v ] i = λ k v k2 ,
which proves that the eigenvalue λ is real. Similarly, if L[ w ] = µ : w, then
λ h v , w i = h L[ v ] , w i = h v , L[ w ] i = µ h v , w i,
and so if λ 6= µ, then h v , w i = 0.
♥ 8.4.11. The difference map ∆: C n → C n is defined as ∆ = S − I , where S is the shift map of
Exercise 8.2.12. (a) Write down the matrix corresponding to ∆. (b) Prove that the sam-
pled exponential vectors ω0 , . . . , ωn−1 from (5.84) form an eigenvector basis of ∆. What
are the eigenvalues? (c) Prove that K = ∆T ∆ has the same eigenvectors as ∆. What are
its eigenvalues? (d) Is K positive definite? (e) According to Theorem 8.20 the eigenvec-
tors of a symmetric matrix are real and orthogonal. Use this to explain the orthogonality of
the sampled exponential vectors. But, why aren’t they real?
Solution:
0
−1 1 0 0 01
B 0 −1 1 0 0C
B C
B C
B 0 −1 1
0 C
B C
(a) B
B .. ..
. .. ... C
C.
B . . C
B C
B
B 0 −1 1 0C
C
@ 0 0 −1 1A
1 0 0 −1
(b) Using Exercise 8.2.12(c),
“ ”
∆ωk = (S − I ) ωk = 1 − e2 k π i /n ωk ,
and so ωk is an eigenvector of ∆ with corresponding eigenvalue 1 − e2 k π i /n .
(c) Since S is an orthogonal matrix, S T = S −1 , and so S T ωk = e− 2 k π i /n ωk . Therefore,
K ωk = (S T − I )(S − I )ωk = (2 I − S − S T )ωk
!
“ ” kπ i
−2 k π i /n − 2 k π i /n
= 2 −e −e ωk = 2 − 2 cos ωk ,
n
kπ i
and hence ωk is an eigenvector of K with corresponding eigenvalue 2 − 2 cos .
n
T
(d) Yes, K > 0 since its eigenvalues are all positive; or note that K = ∆ ∆ is a Gram
matrix, with ker ∆ = {0}.
kπ i (n − k) π i
(e) Each eigenvalue 2 − 2 cos = 2 − 2 cos for k 6= 12 n is double, with a two-
n n
dimensional eigenspace spanned by ωk and ωn−k = ωk . The correpsonding real eigen-
vectors are Re ωk = 12 ωk + 21 ωn−k and Im ωk = 21i ωk − 21i ωn−k . On the other hand,
if k = 21 n (which requires that n be even), the eigenvector ωn/2 = ( 1, −1, 1, −1, . . . )T
is real.
0
c0 c1 c2 c3 . . . cn−1 1
Bc c0 c1 c2 . . . cn−2 C
B n−1 C
B C
B cn−2 cn−1 c0 c1 . . . cn−3 C
♥ 8.4.12. An n × n circulant matrix has the form C = B B .
C,
C
B . .. .. .. .. C
@ . . . . . A
c1 c2 c3 c4 . . . c0
in which the entries of each succeeding row are obtained by moving all the previous row’s
entries one slot to the right, the last entry moving to the front. (a) Check that the shift
matrix S of Exercise 8.2.12, the difference matrix ∆, and its symmetric product K of Ex-
ercise 8.4.11 are all circulant matrices. (b) Prove that the sampled exponential vectors
evv 9/9/04 440 °

ω0 , . . . , ωn−1 , cf. (5.84) are eigenvectors of C. Thus, all circulant matrices have the same eigenvectors!
What are the eigenvalues? (c) Prove that Ω−1 n C Ωn = Λ where Ωn is the Fourier matrix
in Exercise 7.1.62 and Λ is the diagonal matrix with the eigenvalues of C along the diago-
nal. (d) Find the eigenvalues and eigenvectors
0
of the following
1
circulant
0
matrices: 1
0 1 1 −1 −1 1 2 −1 0 −1
! 1 2 3 B C B
1 2 B 1 1 −1 −1 C B −1 2 −1 0C
(i) , (ii) B C
@ 3 1 2 A, (iii) B C, (iv ) B
C
C.
2 1 @ −1 1 1 −1 A @ 0 −1 2 −1 A
2 3 1
−1 −1 1 1 −1 0 −1 2
(e) Find the eigenvalues of the tricirculant matrices in Exercise 1.7.13. Can you find a gen-
eral formula for the n × n version? Explain why the eigenvalues must be real and positive.
Does your formula reflect this fact? (f ) Which of the preceding matrices are invertible?
Write down a general criterion for checking the invertibility of circulant matrices.
Solution:
(a) The eigenvector equation
C ωk = (c0 + c1 e2 k π i /n + c2 e4 k π i /n + · · · + +cn−1 e2 (n−1) k π i /n ) ωk

can either be proved directly, or by noting that
C = c0 I + c1 S + c2 S 2 + · · · + cn−1 S n−1 ,
and using Exercise 8.2.12(c). ! !

1 1
(b) (i) Eigenvalues 3, −1; eigenvectors , .
1 −1
0 1 0 1
0 1
√ √ 1 B 1 C B 1 C
B C B √ C B √ C
(ii) Eigenvalues 6, − 23 − 2
3
i , , − 32 + 2
3
i; eigenvectors @ 1 A, B 1
B− 2 + 3 C, B
i C B−
1
− 3
i C.
C
@ √2 A @ 2 √2 A
1
0 1 0
−1
1 02
−1230i+ 1
−i 1
2 2
3
1 1 1 1
B C B
1C B i C B C B
C B −1 C B − i C
C
(iii) Eigenvalues 0, 2 − 2 i , 0, 2 + 2 i ; eigenvectors B
B C, B C, B
@ 1 A @ −1 A @ 1 A @ −1 A
C, B C.
0 1 0
1
1 0
−1i 0 −1 1
i
1 1 1 1
B C B
1C B i C B C B
C B −1 C B − i C
C
(iv ) Eigenvalues 0, 2, 4, 2; eigenvectors B B C, B C, B
@ 1 A @ −1 A @ 1 A @ −1 A
C, B C.
1 −i −1 √ i√ √ √
7+ 5 7+ 5 7− 5 7− 5
(c) The eigenvalues are (i) 6, 3, 3; (ii) 6, 4, 4, 2; (iii) 6, 2 , 2 , 2 , 2 ; (iv ) in
2kπ
the n × n case, they are 4 + 2 cos for k = 0, . . . , n − 1. The eigenvalues are real and
n
positive because the matrices are symmetric, positive definite.
(d) Cases (i,ii) in (d) and all matrices in part (e) are invertible. In general, an n × n circu-
lant matrix is invertible if and only if none of the roots of the polynomial c0 +c1 x+· · ·+
cn−1 xn−1 = 0 is an nth root of unity: x 6= e2 k π/n .
8.4.13. Write out the spectral decomposition of the0following matrices:

1 0 1
! ! 1 1 0 3 −1 −1
−3 4 2 −1
(a) , (b) , (c) B @ 1 2 1 A,
C
(d) B
@ −1 2 0CA.
4 3 −1 4
0 1 1 −1 0 2
Solution:
evv 9/9/04 441 °

0 10 10 1
! 1
B √5 − √2 √1 √2 C
−3 4 B 5 CB 5
C 0 CBB 5 5C
(a) = @ 2 A@ A@ A.
4 3 1
−√ 2 1
√ √ 0 −5 √
05√ 5 √ 1 5 5 0 √ 1
0 1
!
B √ 1− 2
√ √1+ 2
√ C
√ √ 1− 2
√ √ 1 √
2 −1 B 4−2 2 4+2 2 CB 3 + 2 0 CBB 4−2 2 4−2 2
C
C
(b) =B C@ √ AB √ C.
−1 4 @
√ 1 √ 1 A @ 1+ 2
√ √ 1 √ A
√ √ 0 3− 2 √
0 4−2 2 4+2 2
1 0 4+2 2 1 4+2 2
1 1 1 0 1 1 2 1
0 1 B √6 − √2 √ C
3 C B 3 0 0 CB
√
6
√
6
√ C
6C
1 1 0 B
B C B C
B
B C
B C 2 1 1 1
(c) @1 2 1A = B B √6 0 −√ C C
B
B 0 1 0 CB − √
C B 0 √ C .
B 3 C@ AB 2 2C C
0 1 1 @ A @ A
√1 √1 √1 0 0 0 √1 − √1 √1
60 2 3 1 03 3 3 1
2 1 0 1 2 1 1
0 1 B − √ 0 √ − √ √ √
3 −1 −1 B 6 3CCB 4 0 0 C B
B 6 6 6CC
B C B 1 1 1 CB CB 1 1 C
(d) @ −1 2 0A = B B √ −√ C B
√ CB 0 2 0 C B C B 0 −√ √ C .
B 6 2 3 C@ AB 2 2CC
−1 0 2 @ A @ A
√1 √1 √1 0 0 1 √1 √1 √1
6 2 3 3 3 3
8.4.14. Write out the spectral factorization of the matrices listed in Exercise 8.4.1.
Solution: 0 10 10 1
! 2 √1 C 2 1
2 6 B √5 5 CB 5 0 CB √
5
√
5C
(a) = B @ AB C.
6 −7 @ 1 2 A @ 1 2 A
√ − √ 0 −10 √ − √
05 5 10 10
5 5 1
! 1 1 1 √1 C
− √ √ C − √
5 −2 B 2 2 CB 5 0 CB
B 2 2C
(b) =B@ A@ A@ A.
−2 5 1 1 1 1
√
2
√
2
0 −10 √
2
√
2
0 √ √ 1 0 √ 1
3− 13 3+ 13 0 √ 1
!
B √ √ √ √ C 7+ 13 B √3− 13 √ √ 2 √ C
2 −1 B 26−6 13 26+6 13 CB 2 0 C B 26−6 13 26−6 13 C
(c) =B C@ √ AB √ C,
−1 5 @
√ 2 √ 2 A 7− 13 @
√ 3+ 13 √ 2 √ A
√ √ 0 2 √
0 26−6 13 26+6 1 13 26+6 13 26+6 13
4 3 4 0 10 4 3 1
1
0 1 B 5 √ 2 − 5 − 5 √2 C 6 0 0 CB √ √ √
1 0 4 B CB B 5 2 5 2 2CC
B C B 3 4 3 C B C B C
(d) @0 1 3A = B B 5√ 2 5
√ C C
B
B 0 1 0 CB
C − 3 4
0 C,
B 5 2 C@ A@ B 5 5 C
4 3 1 @ A A
4 3 1
√1 0 √1 0 0 −4 − √ − √ √
20 2 5 2 5 2 2
1 0 1
1 1 1 0 1 1 1 2
0 1 B √ − √ √ √ − √ √
6 −4 1 B 6 3 2CCB 12 0 0 CB
B 6 6 6CC
B C B 1 1 1 CB CB 1 1 1 C
(e) @ −4 B
6 −1 A = B − √ √ C B
√ CB 0 9 0 CB − √ C B √ √ C .
B 6 3 2 C@ AB 3 3 3CC
1 −1 11 @ A @ A
√2 √1 0 0 0 2 √1 √1 0
6 3 2 2
8.4.15. Construct a symmetric matrix with the following eigenvectors and eigenvalues, or ex-
“ ”T “ ”T
3 4
plain why none exists: (a) λ1 = 1, v1 = 5, 5 , λ2 = 3, v2 = − 54 , 3
5 ,
(b) λ1 = −2, v1 = ( 1, −1 ) , λ2 = 1, v2 = ( 1, 1 ) , (c) λ1 = 3, v1 = ( 2, −1 )T ,
T T
λ2 = −1, v2 = ( −1, 2 )T , (d) λ1 = 2, v1 = ( 2, 1 )T , λ2 = 2, v2 = ( 1, 2 )T .

0 1 0 1
57 24 1 3
B 25 − 25 C B − 2 2C
Solution: (a) @ A; (b) @ A; (c) none, since eigenvectors are not orthog-
24 43 3
−
! 25 25 2 − 12
2 0
onal; (d) . Even thought the given eigenvectors are not orthogonal, one can construct
0 2
evv 9/9/04 442 °

an orthogonal basis of the eigenspace.
8.4.16. Use the spectral factorization to diagonalize the following quadratic forms:
(a) x2 − 3 x y + 5 y 2 , (b) 3 x2 + 4 x y + 6 y 2 , (c) x2 + 8 x z + y 2 + 6 y z + z 2 ,
(d) 23 x2 − x y − x z + y 2 + z 2 , (e) 6 x2 − 8 x y + 2 x z + 6 y 2 − 2 y z + 11 z 2 .
Solution: „ «2 „ «2
(a) 1
2
√3 √1 y
x+ + 112 − √1 x + √3 y = 201
(3 x + y)2 + 11 2
20 (− x + 3 y) ,
10 10 10 10
„ «2 „ «2
(b) 7 √1 x + √2 y + 11
2 − √2 x + √1 y = 75 (x + 2 y)2 + 52 (− 2 x + y)2 ,
5 5 5 5
„ «2 “ ”2 „ «2
4 3
(c) −4 √ x+ √ y − √1 z + − 53 x + 45 y + 6 4
√ x+ √ 3
y + √1 z
5 2 5 2 2 5 2 5 2 2
= − 252
(4 x + 3 y − 5 z)2 + 25
1
(− 3 x + 4 y)2 + 25 3
(4 x + 3 y + 5 z)2 ,
„ «2 „ «2 „ «2
1 √1 x + √1 y + √1 z √1 y + √1 z √2 x + √1 y + √1 z
(d) 2 + − + 2 −
3 3 3 2 2 6 6 6
= 61 (x + y + z)2 + 21 (− y + z)2 + 31 (− 2 x + y + z)2 ,
„ «2 „ «2 „ «2
(e) 2 √1 x + √1 y + 9 − √1 x + √1 y + √1 z + 12 √1 x − √1 y + √2 z
2 2 3 3 3 6 6 6
2 2 2
= (x + y) + 3 (− x + y + z) + 2 (x − y + 2 z) .
0 1
2 1 −1
B
♥ 8.4.17. (a) Find the eigenvalues and eigenvectors of the matrix A = 1 2 1CA. (b) Use @
−1 1 2
the eigenvalues to compute the determinant of A. (c) Is A positive definite? Why or why
not? (d) Find an orthonormal eigenvector basis of R 3 determined by A or explain why
none exists. (e) Write out the spectral factorization of A if possible. (f ) Use orthogonal-
ity to write the vector ( 1, 0, 0 )T as a linear combination of eigenvectors of A.
Solution: 0 1 0 1 0 1
1 −1 1
(a) λ1 = λ2 = 3, v1 = B C B C B
@ 1 A, v2 = @ 0 A, λ3 = 0, v3 = @ −1 A;
C
0 1 1
(b) det A = λ1 λ2 λ3 = 0.
(c) Only positive semi-definite;
0 1
not positive
0
definite
1
since it has a zero eigenvalue.
0 1
1 1
√1 B − √6 C B √3 C
B 2C B C B C
B C B
(d) u1 = B 1 C
B √ C, u 2 =B 1 CC, u
B
=B 1 C
C;
B √ − √
B 2C
@ A B 6CC 3
B
B 3C
C
@ A @ A
0 √2 √1
0 6 13 0 1
1 1 1 0 1
0 1 B √2 − √6 √ C √1 √1 0C
2 1 −1 B 3 CB 3 0 0 CB
B 2 2 C
B B 1 CB CB C
(e) @ 1 2 1CA=BB √2 √1 − √1 C B
B0 3 0C B − √1 √1 √2 C;
B 6 3CC@
CB
AB 6 6 6 C
C
−1 1 2 @ A @ A
0 √2 √1 0 0 0 √1 − √1 √1
0 1 6 3 3 3 3
1
B C √1 √1 √1
(f ) @0A = u1 − u2 + u3 .
2 6 3
0
8.4.18. True or false: A matrix with a real orthonormal eigenvector basis is symmetric.
Solution: True. If Q is the orthogonal matrix formed by the eigenvector basis, then A Q = Q Λ
where Λ is the diagonal eigenvalue matrix. Thus, A = Q Λ Q−1 = Q Λ QT , which is symmetric.
evv 9/9/04 443 °

♦ 8.4.19. Let u1 , . . . , un be an orthonormal basis of R n . Prove that it forms an eigenvector basis
for some symmetric n × n matrix A. Can you characterize all such matrices?
Solution: The simplest is A = I . More generally, any matrix of the form A = S T Λ S, where
S = ( u1 u2 . . . un ) and Λ is any real diagonal matrix.
0 1
n
T 2@ X 2
8.4.20. Prove that any quadratic form can be written as x A x = k x k λi cos θi A,
i=1
where λi are the eigenvalues of A and θi = <
) (x, vi ) denotes the angle between x and the
th
i eigenvector.
n
X
Solution: Using the spectral decomposition, we have xT A x = (QT x)T Λ(QT x) = λi yi2 ,
i=1
where yi = ui · x = k x k cos θi denotes the ith entry
0 of QT1x. Q.E.D.
3 1 2
8.4.21. An elastic body has stress tensor T = B C
@ 1 3 1 A. Find the principal stretches and
principal directions of stretch. 2 1 3
√ √
Solution: Principal stretches = eigenvalues: 4 + 3, 4 − 3, 1;
“ √ ”T “ √ ”T
directions = eigenvectors 1, −1 + 3, 1 , 1, −1 − 3, 1 , ( −1, 0, 1 )T .
♦ 8.4.22. Given a solid body spinning around its center of mass, the eigenvectors of its positive
definite inertia tensor prescribe three mutually orthogonal principal directions of rotation,
while0the corresponding
1 eigenvalues are the moments of inertia. Given the inertia tensor
2 1 0
T =B C
@ 1 3 1 A, find the principal directions and moments of inertia.
0 1 2
Solution: Moments of inertia: 4, 2, 1; principal directions: ( 1, 2, 1 )T , ( −1, 0, 1 )T , ( 1, −1, 1 )T .

♦ 8.4.23. Let K be a positive definite 2 × 2 matrix. (a) Explain why the quadratic equation
xT K x = 1 defines an ellipse. Prove that its principal axes are the eigenvectors of K, and
the semi-axes are the reciprocals of the square roots of the eigenvalues. (b) Graph and de-
scribe the following curves: (i) x2 +4 y 2 = 1, (ii) x2 +x y+y 2 = 1, (iii) 3 x2 + 2 x y + y 2 = 1.
(c) What sort of curve(s) does xT K x = 1 describe if K is not positive definite?
Solution:
(a) Let K = QΛ QT be its spectral decomposition. Then xT K x = yT Λ y where x = Q y.
The ellipse yT Λ y = λq1 y12 + λ2 y12 = 1 has principal axes aligned with the coordinate
axes and semi-axes 1/ λi , i = 1, 2. The map x = Q y serves to rotate the coordinate
axes to align with the columns of Q, i.e., the eigenvectors, while leaving the semi-axes
unchanged.
(b)
1
0.5
(i) -1 -0.5 0.5 1 ellipse with semi-axes 1, 12 and principal axes ( 1, 0 )T , ( 0, 1 )T .

-0.5
-1
evv 9/9/04 444 °

1.5
1
√ q
0.5 ellipse with semi-axes 2, 23 ,
(ii) -1.5 -1 -0.5 0.5 1 1.5 and principal axes ( −1, 1 )T , ( 1, 1 )T .
-0.5
-1
-1.5
1.5
1
1 1
ellipse with semi-axes √ √ ,√ √ ,
0.5 2+ 2 2− 2
“ √ ”T “ √ ”T
(iii) -1.5 -1 -0.5 0.5 1 1.5 and principal axes 1+ 2, 1 , 1− 2, 1 .
-0.5
-1
-1.5
(c) If K is symmetric and positive semi-definite it is a parabola; if K is indefinite, a hy-

perbola; if negative (semi-)definite, the empty set. If K is not symmetric, replace K by
1 T
2 (K + K ), and then apply the preceding classification.
♦ 8.4.24. Let K be a positive definite 3×3 matrix. (a) Prove that the quadratic equation x T K x = 1
defines an ellipsoid in R 3 . What are its principal axes and semi-axes? (b) Describe the
surface defined by the quadratic equation 11 x2 − 8 x y + 20 y 2 − 10 x z + 8 y z + 11 z 2 = 1.
Solution: (a) Same method as in Exercise 8.4.23. Its principal axes are the eigenvectors of K,
and the semi-axes are the reciprocals of the square roots of the eigenvalues. (b) Ellipsoid with
principal axes: ( 1, 0, 1 )T , ( −1, −1, 1 )T , ( −1, 2, 1 )T and semi-axes √1 , √1 , √1 .
6 12 24
T
8.4.25. Prove that A = A has a repeated eigenvalue if and only if it commutes, A J = J A,
with a nonzero skew-symmetric matrix: J T = − J =
6 O. Hint: First prove this when A is a
diagonal matrix.
Solution: If Λ = diag (λ1 , . . . , λn ), then the (i, j) entry of Λ M is di mij , whereas the (i, j) entry
of M Λ is dk mik . These are equal if and only if either mik = 0 or di = dk . Thus, Λ M = M Λ
with M having one or more non-zero off-diagonal entries, which includes the case of non-zero
skew-symmetric matrices, if and only if Λ has one or more repeated diagonal entries. Next,
suppose A = Q Λ QT is symmetric with diagonal form Λ. If A J = J A, then Λ M = M Λ
where M = QT J Q is also nonzero, skew-symmetric, and hence A has repeated eigenvalues.
Conversely, if λi = λj , choose M such that mij = 1 = − mji , and then A commutes with
J = Q M QT . Q.E.D.
♦ 8.4.26. (a) Prove that every positive definite matrix K has a unique positive definite square
root, i.e., a matrix B > 0 satisfying B 2 = K. (b) Find the positive definite square roots of
the following matrices: 0 1 0 1
! ! 2 0 0 6 −4 1
2 1 3 −1
(i) , (ii) , (iii) B C
@ 0 5 0 A, (iv ) B
@ −4 6 −1 C A.
1 2 −1 1
0 0 9 1 −1 11
Solution: √ √
(a) Set B = Q Λ QT , where Λ is the diagonal matrix with the square roots of the eigen-
values of A along the diagonal. Uniqueness follows from the fact that the eigenvectors
and eigenvalues are uniquely determined. (Permuting them does not change the final
form of B.)
evv 9/9/04 445 °

0 1
√ √ √ √ !
1B 3 + 1 3 − 1C 1 −1
2 2√ 1− 2 ;
(b) (i) @ √ √ A; (ii) √ q √
2 3−1 3+1 (2 − 2) 2 + 2 1− 2 1
0 1
0√ 1 B 1+ √1 + √1 −1 +√1 − √1 −1 + √2 C
2 2 3 2 3 3
B √0 0
C
B
B
C
C
(iii) @ 0 5 0 A; (iv ) B
B − 1 + √ − √1
1 1
1+ √ + √ 1
1− √2 C.
C
B 2 3 2 3 3 C
0 0 3 @ A
− 1 + √2 1 − √2 1+ √4
3 3 3
8.4.27. Find all positive definite orthogonal matrices.

Solution: Only the identity matrix is both orthogonal and positive definite. Indeed, if K =
K T > 0 is orthogonal, then K 2 = I , and so its eigenvalues are all ± 1. Positive definiteness
implies that all the eigenvalues must be +1, and hence it diagonal form is Λ = I . But then
K = Q I QT = I also.
♦ 8.4.28. The Polar Decomposition: Prove that every invertible matrix A has a polar decomposi-
tion, written A = Q B, into the product of an orthogonal matrix Q and a positive definite
matrix B > 0. Show that if det A > 0, then Q is a proper orthogonal matrix. Hint: Look
at the Gram matrix K = AT A and use Exercise 8.4.26. Remark : In mechanics, if A repre-
sents the deformation of a body, then Q represents a rotation, while B represents a stretch-
ing along the orthogonal eigendirections of K. Thus, any linear deformation of an elastic
body can be decomposed into a pure stretching transformation followed by a rotation.
√
Solution: If A = Q B, then K = AT A = B T QT Q B = B 2 , and hence B = K is the posi-
tive definite square root of K. Moreover, Q = A B −1 then satisfies QT Q = B −T AT A B −1 =
B −1 K B −1 = I since B 2 = K. Finally, det A = det Q det B and det B > 0 since B > 0. So if
det A > 0, then det Q = +1 > 0. Q.E.D.
8.4.29. Find the polar decompositions A = Q B, as defined in Exercise 8.4.28, of the following
matrices: 0 1 0 1
! ! ! 0 −3 8 1 0 1
0 1 2 −3 1 2 B C
(a) , (b) , (c) , (d) @ 1 0 0 A, (e) @ 1 −2 0 C
B
A.
2 0 1 6 0 1
0 4 6 1 1 0
0 1 0 10 1
! ! ! 2
− √1 √
0 1 0 1 2 0 B2 −3 C B √5 5 CB 5
C 0C
Solution: (a) = , (b) = B
2 0 1 0 0 1
@ A @ 1 2 A@ √ A,
1 6 √
5
√
5
0 3 5
0 10 1 0 1 0 10 1
1 2
!
B √1 √1 CB √1 √1 C 0 −3 8 0 − 53 45 1 0 0
B
(c) = B
@
2 2 CB 2
A@ 1
2C
A, (d) @1 0 0C B
A = @1
CB
0 0A@0 5 0CA,
0 1 − √1 √1 √ √3 4 3
0 4 6 0 0 0 10
0 1 02 2 2 2 10 15 5
1 0 1 .3897 .0323 .9204 1.6674 −.2604 .3897
B
(e) @1 −2 0C B
A = @ .5127 −.8378 −.1877 C B
A@ −.2604 2.2206 .0323 C
A.
1 1 0 .7650 .5450 −.3430 .3897 .0323 .9204
♥ 8.4.30. The Spectral Decomposition: (i) Let A be a symmetric matrix with eigenvalues λ 1 , . . . , λn
and corresponding orthonormal eigenvectors u1 , . . . , un . Let Pk = uk uT
k be the orthogo-
nal projection matrix onto the eigenline spanned by uk , as defined in Exercise 5.5.8. Prove
that the spectral factorization (8.32) can be rewritten as
A = λ 1 P 1 + λ 2 P 2 + · · · + λ n P n = λ 1 u1 uT T T
1 + λ 2 u2 u2 + · · · + λ n un un , (8.34)
expressing A as a linear combination of projection matrices. (ii) Write out the spectral
decomposition (8.34) for the matrices in Exercise 8.4.13.
Solution:
evv 9/9/04 446 °

(i) This follows immediately from the spectral factorization. The rows of Λ Q T are
λ 1 uT T
1 , . . . , λn un . The result then follows from the alternative version of matrix multi-
plication given in Exercise
0 11.2.32.
0 1
! 2 1 4 2
(ii) (a)
−3 4 5C= 5B
A − 5@
@
B
5 5 −5 C A.
4 3 4 2 2 1
5
0 5 − 5√ 1 5 0 1
√ √ √
! 3−2√2 1− √2 3+2√2 1+ √2
2 −1 √ B 4−2 2 4−2 2 C
√ B C
(b) = (3 + 2 )B C + (3 − 2 )B 4−2√ 2 4−2 2 C.
−1 4 @ 1−√2 1 A @ 1+ 2 1√ A
√ √ √
0 4−21 20 4−2 2 1 4−2 2 4−2 2
0 1 1 1 1 1 1
1 1 0 B6
B 3 6C C
B
B 2 0 −2 C C
B C
(c) @ 1 2 1 A = 3B
B3
1 2 1 C+B
C B 0 0 0 C.
C
@ 3 3 A @ A
0 1 1 1 1 1 1 1
6 03 6 − 2 1 0 02 1 0 1
0 1 2 1 1 1 1 1
3 −1 −1 B 3 − 3 − 3C B 0 0 0 C B 3 3 3 C
B B C B C B C
(d) @ −1 2 0CA = 4BB − 3
1 1 1 C + 2B
C B 0 1
− 1 C+B
C B
1 1 1 C.
C
@ 6 6 A @ 2 2 A @ 3 3 3 A
−1 0 2 1 1
− 13 1
6
1
6 0 − 2 2
1
3
1
3
1
3
♦ 8.4.31. The Spectral Theorem for Hermitian Matrices. Prove that any complex Hermitian ma-
trix can be factored as H = U Λ U † where U is a unitary matrix and Λ is a real diagonal
matrix. Hint: See Exercises 5.3.25, 8.4.7.
Solution:
8.4.32. Find the spectral factorization, as in Exercise 8.4.31, of the following Hermitian matri-
0 1
! ! −1 5i −4
3 2i 6 1 − 2i
ces: (a) , (b) , (c) B @ −5 i −1 4 i CA.
−2 i 6 1 + 2i 2
−4 −4 i 8
Solution:
0 1 0 10 10 1
√i √2 C − √i 2i
√
B 3 2i C B
B 5 5 CB 2 0 C B B 5 5CC,
(a) @ A= @ A @ A@ A
2i √1 √2 √1
− 2i 6 −√ 0 7
0 5 5 10 5 5
10 1
! 1−2
√ i −1+2
√ i 1+2
√ i √1 C
6 1 − 2i B 6 30 C B 7 0 CB 6
(b) =@B √ C
A@ A@ B √6 CA,
1 + 2i 2 √1 √5 − 1+2 i √5
6 6
0 1 √
30 6
0 1 0 1
1 i 0 1
0 1 B − √6 − √2 √1 C − √1 − √i √2 C
−1 5i −4 B 3 CB 12 0 0 CBB 6 6 6C
B C B i 1 i CB CB C
(c) @ −5 i −1 4 i A = B √ B C B CB √i 1
0 C
6
√
2
− √ CB 0 −6
3 C@ 0C B 2
√
2 C.
B AB C
−4 −4 i 8 @
2 1
A @
1 i
A
√ 0 √ 0 0 0 √ √ √1
6 3 3 3 3
8.4.33. Find the minimum and maximum values of the quadratic form 5 x2 + 4 x y + 5 y 2 where
x, y are subject to the constraint x2 + y 2 = 1.
Solution: Maximum: 7; minimum: 3.
8.4.34. Find the minimum and maximum values of the quadratic form 2 x2 + x y + 2 x z + 2 y 2 +
2 z 2 where x, y, z are required to satisfy x2 + y 2 + z 2 = 1.
evv 9/9/04 447 °

√ √
4+ 5 4− 5
Solution: Maximum: 2 ; minimum: 2 .
8.4.35. Write down and solve a optimization principle characterizing the largest and smallest
eigenvalue of the following positive definite matrices:
0 1 0 1
! ! 6 −4 1 4 −1 −2
2 −1 4 1
(a) , (b) , (c) B @ −4 6 −1 C
A, (d) B
@ −1 4 −1 CA.
−1 3 1 4
1 −1 11 −2 −1 4
Solution: √
(a) 5+2 5 = max{ 2 x2 − 2 x y + 3 y 2 | x2 + y 2 = 1 },
√
5− 5
2 = min{ 2 x2 − 2 x y + 3 y 2 | x2 + y 2 = 1 };
(b) 5 = max{ 4 x2 + 2 x y + 4 y 2 | x2 + y 2 = 1 },
3 = min{ 4 x2 + 2 x y + 4 y 2 | x2 + y 2 = 1 };
(c) 12 = max{ 6 x2 − 8 x y + 2 x z + 6 y 2 − 2 y z + 11 z 2 | x2 + y 2 + z 2 = 1 },
2 = min{ 6 x2 − 8 x y + 2 x z + 6 y 2 − 2 y z + 11 z 2 | x2 + y 2 + z 2 = 1 };
max{ 4 x2 − 2 x y − 4 x z + 4 y 2 − 2 y z + 4 z 2 | x2 + y 2 + z 2 = 1 },
(d) 6 = √
3 − 3 = min{ 4 x2 − 2 x y − 4 x z + 4 y 2 − 2 y z + 4 z 2 | x2 + y 2 + z 2 = 1 }.
8.4.36. Write down a maximization principle that characterizes the middle eigenvalue of the
matrices in parts (c–d) of Exercise 8.4.35.
√(c) 12 = max{
Solution: 6 x2 − 8 x y + 2 x z + 6 y 2 − 2 y z + 11 z 2 | x2 + y 2 + z 2 = 1, x − y + 2 z = 0 };
(d) 3 + 3 = max{ 4 x2 − 2 x y − 4 x z + 4 y 2 − 2 y z + 4 z 2 | x2 + y 2 + z 2 = 1, x − z = 0 },
8.4.37. What is the minimum and maximum values of the following rational functions:
3 x2 − 2 y 2 x2 − 3 x y + y 2 3 x2 + x y + 5 y 2 2 x2 + x y + 3 x z + 2 y 2 + 2 z 2
(a) , (b) , (c) , (d) .
x2 + y 2 x2 + y 2 x2 + y 2 x2 + y 2 + z 2
Solution:
(a) Maximum: 3; minimum: −2.
(b) Maximum: 52 ; minimum: − 21 .
√ √
8+ 5 8− 5
(c) Maximum: 2√ = 5.11803; minimum: 2 √ = 2.88197.
4+ 10
(d) Maximum: 2 = 3.58114; minimum: 4−2 10 = .41886.
n−1
X
8.4.38. Find the minimum and maximum values of q(x) = 2 xi xi+1 for k x k2 = 1. Hint: See
Exercise 8.2.46. i=1
π π
Solution: Maximum: 2 cos ; minimum − 2 cos .
n+1 n+1
8.4.39. Suppose K > 0. What is the maximum value of q(x) = xT K x when x is constrained
to a sphere of radius k x k = r?
Solution: Maximum: r 2 λ1 ; minimum r 2 λn , where λ1 , λn are, respectively, the maximum and
minimum eigenvalues of K.
8.4.40. Let K >n 0. Prove˛ the product
o
formula
n ˛ o
max xT K x ˛˛ k x k = 1 min xT K −1 x ˛
˛ kxk = 1 = 1.
Solution: max{ xT K x | k x k = 1 } = λ1 is the largest eigenvalue of K. On the other hand, K −1

is positive definite, cf. Exercise 3.4.9, and hence min{ xT K −1 x | k x k = 1 } = µn is its smallest
eigenvalue. But the eigenvalues of K −1 are the reciprocals of the eigenvalues of K, and hence
its smallest eigenvalue is µn = 1/λ1 , and so the product is λ1 µn = 1. Q.E.D.
evv 9/9/04 448 °

♦ 8.4.41. Write out the details in the proof of Theorem 8.30.
Solution: According to the discussion preceding the statement of the Theorem 8.30,
n ˛ o
˛
λj = max yT Λ y ˛ k y k = 1, y · e1 = · · · = y · ej−1 = 0 .
Moreover, using (8.33), setting x = Qy and using the fact that Q is an orthogonal matrix and
so (Q v) · (Q w) = v · w for any v, w ∈ R n , we have
xT A x = yT Λ y, k x k = k y k, y · e i = x · vi ,
where vi = Q ei is the ith eigenvector of A. Therefore, by the preceding formula,
n ˛ o
λj = max xT A x ˛˛ k x k = 1, x · v1 = · · · = x · vj−1 = 0 . Q.E .D.
♦ 8.4.42. Reformulate Theorem 8.30 as a minimum principle for intermediate eigenvalues.

Solution: Let A be a symmetric matrix with eigenvalues λ1 ≥ λ2 ≥ · · · ≥ λn and corresponding
orthogonal eigenvectors v1 , . . . , vn . Then the minimal value of the quadratic form xT A x over
all unit vectors which are orthogonal to the last n − j eigenvectors is the j th eigenvalue:
n ˛ o
λj = min xT A x ˛˛ k x k = 1, x · vj+1 = · · · = x · vn = 0 .
8.4.43. Under the set-up of Theorem 8.30, explain why

8 ˛ 9
˛
< vT K v ˛
˛
=
λj = max : ˛ v 6= 0, v · v1 = · · · = v · vj−1 = 0 ; .
k v k2 ˛
vT K v v
Solution: Note that = uT K u, where u = is a unit vector. Moreover, if v is or-
k v k2 kvk
thogonal to an eigenvector vi , so is u. Therefore,
8 ˛ 9
˛
< vT K v ˛
˛
=
max : ˛ v 6= 0, v · v1 = · · · = v · vj−1 = 0 ;
k v k2 ˛
n ˛ o
= max uT K u ˛˛ k u k = 1, u · v1 = · · · = u · vj−1 = 0 = λj
by Theorem 8.30.
♥ 8.4.44. (a) Let K, M be positive definite n × n matrices and λ1 ≥ · · · ≥ λn their generalized
eigenvalues, as in Exercise 8.4.9. Prove that that the largest generalized eigenvalue can be
characterized by the maximum principle λ1 = max{ xT K x | xT M x = 8
1 }. Hint:˛ Use Ex-9
< xT K x ˛˛ =
ercise 8.4.26. (b) Prove the alternative maximum principle λ1 = max : T ˛ x 6= 0 .
˛ ;
x Mx ˛
(c) How would you characterize the smallest generalized eigenvalue? (d) An intermediate
generalized eigenvalue?
Solution: √
(a) Let R = M be the positive definite square root of M , and set Kc = R−1 K R−1 . Then
f y, xT M x = yT y = k y k2 , where y = R x. Thus,
xT K x = y T K
n ˛ o n ˛ o
˛ ˛
max xT K x ˛ xT M x = 1 = max yT K
fy ˛ k y k2 = 1 e ,
=λ 1
f But K
the largest eigenvalue of K. f y = λ y implies K x = λ x, and so the eigenvalues of
f
K and K coincide. Q.E.D.
x T
(b) Write y = √ so that y M y = 1. Then, by part (a),
xT 8
Mx ˛ 9
< xT K x ˛˛ = n ˛ o
˛ x 6= 0 T ˛ T
max : T ˛ ;
= max y K y ˛ y M y = 1 = λ1 .
x Mx ˛
evv 9/9/04 449 °

(c) λn = min{ xT K x | xT M x = 1 }.
(d) λj = max{ xT K x | xT M x = 1, xT M v1 = · · · = xT M vj−1 = 0 } where v1 , . . . , vn are
the generalized eigenvectors.
8.4.45. Use Exercise 8.4.44 to find the minimum and maximum of the rational functions
3 x2 + 2 y 2 x2 − x y + 2 y 2 2 x2 + 3 y 2 + z 2 2 x2 + 6 x y + 11 y 2 + 6 y z + 2 z 2
(a) , (b) , (c) , (d) .
4 x2 + 5 y 2 2 x2 − x y + y 2 x2 + 3 y 2 + 2 z 2 x2 + 2 x y + 3 y 2 + 2 y z + z 2
√ √
Solution: (a) max = 43 , min = 52 ; (b) max = 9+47 2 , min = 9−4 2
7 ;
(c) max = 2, min = 12 ; (d) max = 4, min = 1.
8.4.46. Let A be a complete square matrix, not necessarily symmetric, with all positive eigen-
values. Is the quadratic form q(x) = xT A x > 0 for all x 6= 0?
!
1 b
Solution: No. For example, if A = , then q(x) > 0 for x 6= 0 if and only if | b | < 4.
0 4
SVD 8.5. Singular Values.
! !
1 1 0 1
8.5.1. Find the singular values of the following matrices: (a) , (b) ,
0 2 −1 0
0 1
! ! ! 1 −1 0
1 −2 2 0 0 2 1 0 −1 B
(c) , (d) , (e) , (f ) @ −1 2 −1 CA.
−3 6 0 3 0 0 −1 1 1
0 −1 1
√ √ √ √
Solution: (a) 3 ± 5, (b) 1, 1; (c) 5 2; (d) 3, 2; (e) 7, 2; (f ) 3, 1.
8.5.2. Write out the singular value decomposition (8.40) of the matrices in Exercise 8.5.1.
Solution: 0 √ √ 1 0 √ 1
0q 1
!
B √−1+ √
5 √−1− 5
√ C
√ √−2+ √
5 √ 1
√
1 1 3+ 5 0 CB C
(a) =B
B
10−2 5 10+2 5 CB
C@ q √ AB
B 10−4 5
√
10−4 5 C
C,
0 2 @
√ 2 √ √ 2 √ A @
√−2− √5 √ 1 √ A
0 3− 5
! 10−2 5 ! 10+2
! 5 ! 10+4 5 10+4 5
0 1 1 0 1 0 0 1
(b) = ,
−1 0 0 −1 0 1 1 0
0 1
!
1 −2 − √1 “ √ ”“ 1 ”
(c) =@ B 10 C
A 5 2 − √5 √2 ,
−3 6 √3 5
! 10! ! !
2 0 0 0 1 3 0 0 10
(d) = ,
0 3 0 1 0 0 2 1 00
0 10 10 1
!
− √2 √1 √ − √4 − √3 √1 √3 C
2 1 0 −1 B 5 5 CB 7 0 CB 35 35 35 35 C
(e) =B
@ A@
C
√ AB@ A,
0 −1 1 1 √1 √2 √2 − √1 √2 √1
5 5 1
0 2 10 10 10 10
0
0 1 B √1 − √1 C0 10 √ 1
1 −1 0 B √6 2C
CB 3 √1 − √2 √1 C
B 0 CB
(f ) B
@ −1 2 −1 C
A=B
B − √2 0CC@
BA@
6 3 6C
A.
B 3 C 0 √1 1
0 −1 1 @ A 1 − 0 √
√1 √1 2 2
6 2
evv 9/9/04 450 °

!
1 1
8.5.3. (a) Construct the singular value decomposition of the shear matrix A = .
0 1
(b) Explain how a shear can be realized as a combination of a rotation, a stretch, followed
by a second rotation.
Solution: 0 √ √ 1 0 √ 1
0q √ 1
!
B √1+ 5
√ √1− 5
√ C 3 1 √−1+ √
5 √ 2
√
1 1 B 10+2 5 10−2 5 CB 2 +2 5 0 CB
B 10−2 5 10−2 5
C
C
(a) = B C@ q √ AB √ C;
0 1 @
√ 2 √ √ 2 √ A
0 3
− 1 @
√−1− √5 √ 2 √ A
10+2 5 10−2 5 2 2 5 10+2 5 10+2 5
(b) The first and last matrices are proper orthogonal, and so represent rotations, while the mid-
dle matrix is a stretch along the coordinate directions, in proportion to the singular values. Ma-
trix multiplication corresponds to composition of the corresponding linear transformations.
8.5.4. Find the condition number of the!following matrices. Which
! would you characterize!
2 −1 −.999 .341 1 2
as ill-conditioned? (a) , (b) , (c) ,
−3 1 −1.001 .388 1.001 1.9997
0 1
0 1 0 1 5 7 6 5
−1 3 4 72 96 103 B
B C B C B 7 10 8 7CC
(d) @ 2 10 6 A, (e) @ 42 55 59 A, (f ) B
@6
C.
8 10 9A
1 2 −3 67 95 102
5 7 9 10
Solution: √
(a) The eigenvalues of K = AT A are 15 2 ± 221
2 = 14.933, .0669656. The square roots
of these eigenvalues give us the singular values of A. i.e., 3.8643, .2588. The condition
number is 3.86433 / .25878 = 14.9330.
(b) The singular values are 1.50528, .030739, and so the condition number is 1.50528 / .030739 =
48.9697.
(c) The singular values are 3.1624, .0007273, and so the condition number is 3.1624 / .0007273 =
4348.17; slightly ill-conditioned.
(d) The singular values are 12.6557, 4.34391, .98226, so he condition number is 12.6557 / .98226 =
12.88418.
(e) The singular values are 239.138, 3.17545, .00131688, so the condition number is 239.138 / .00131688 =
181594; ill-conditioned.
(f ) The singular values are 30.2887, 3.85806, .843107, .01015, so the condition number is
30.2887 / .01015 = 2984.09; slightly ill-conditioned.
♠ 8.5.5. Solve the following systems of equations using Gaussian Elimination with three-digit
rounding arithmetic. Is your answer a reasonable approximation to the exact solution?
Compare the accuracy of your answers with the condition number of the coefficient matrix,
and discuss the implications of ill-conditioning.
97 x + 175 y + 83 z = 1, 3.001 x + 2.999 y + 5 z = 1,
1000 x + 999 y = 1,
(a) (b) 44 x + 78 y + 37 z = 1, (c) − x + 1.002 y − 2.999 z = 2
1001 x + 1002 y = −1,
52 x + 97 y + 46 z = 1. 2.002 x + 4 y + 2 z = 1.002.
Solution:
(a) x = 1, y = −1.
(b) x = −1, y = −109, z = 231, singular values 26.6, 1.66, .0023, condition number: 1.17 ×
105 .
(c)
♠ 8.5.6. (a) Compute the singular values and condition numbers of the 2 × 2, 3 × 3, and 4 × 4
Hilbert matrices. (b) What is the smallest Hilbert matrix with condition number larger
than 106 ?
evv 9/9/04 451 °

Solution:
(a) The 2×2 Hilbert matrix has singular values 1.2676, .0657 and condition number 19.2815.
The 3 × 3 Hilbert matrix has singular values 1.4083, .1223, .0027 and condition number
524.057. The 4 × 4 Hilbert matrix has singular values 1.5002, .1691, .006738, .0000967
and condition number 15513.7.
(b) An 5 × 5 Hilbert matrix has condition number 4.7661 × 105 . An 6 × 6 Hilbert matrix has
condition number 1.48294 × 107 .
8.5.7. (a) What are the singular values of a 1 × n matrix? (b) Write down its singular value
decomposition. (c) Write down its pseudoinverse.
Solution: Let A = v ∈ R n be the matrix (column vector) in question. (a) It has one singular
v “ ” vT
value: k v k; (b) P = , Σ = k v k , Q = (1); (c) v+ = .
kvk k v k2
8.5.8. Answer Exercise 8.5.7 for an m × 1 matrix.
Solution: Let A = vT , where v ∈ R n , be the matrix (row vector) in question. (a) It has one
“ ” v v
singular value: k v k; (b) P = (1), Σ = k v k , Q = ; (c) v+ = .
kvk k v k2
8.5.9. True or false: Every matrix has at least one singular value.
Solution: True, with one exception — the zero matrix.
8.5.10. Explain why the singular values of A√are the same as the nonzero eigenvalues of the
positive definite square root matrix S = AT A , defined in Exercise 8.4.26.
Solution: Since S 2 = K = AT A, the eigenvalues λ of K are the squares, λ = σ 2 of√ the eigen-
values σ of S. Moreover, since S > 0, its eigenvalues are all non-negative, so σ = + λ, and, by
definition, the nonzero σ > 0 are the singular values of A.
8.5.11. True or false: The singular values of AT are the same as the singular values of A.
Solution: True. If A = P Σ QT is the singular value decomposition of A, then the transposed
equation AT = Q Σ P T gives the singular value decomposition of AT , and so the diagonal en-
tries of Σ are also the singular values of AT .
♦ 8.5.12. Prove that if A is square, nonsingular, then the singular values of A−1 are the recipro-
cals of the singular values of A. How are their condition numbers related?
Solution: Since A is nonsingular, so is K = AT A, and hence all its eigenvalues are nonzero.
Thus, Q, whose columns are the orthonormal eigenvector basis of K, is a square orthogonal
matrix, as is P . Therefore, the singular value decomposition of the inverse matrix is A −1 =
Q−T Σ−1 P −1 = Q Σ−1 P T . The diagonal entries of Σ−1 , which are the singular values of A−1 ,
are the reciprocals of the diagonal entries of Σ. Finally, κ(A−1 ) = σn /σ1 = 1/κ(A). Q.E.D.
♦ 8.5.13. Let A be any matrix. Prove that any vector x that minimizes the Euclidean least squares
error k A x − b k must satisfy the normal equations AT A x = AT b, even when ker A 6= {0}.
Solution:
♦ 8.5.14. (a) Let A be a nonsingular square matrix. Prove that the product of the singular val-
ues of A equals the absolute value of its determinant: σ1 σ2 · · · σn = | det A |. (b) Does
their sum equal the absolute value of the trace: σ1 + · · · + σn = | tr A |? (c) Show that
if | det A | < 10− k , then its minimal singular value satisfies σn < 10− k/n . (d) True or
false: A matrix whose determinant is very small is ill-conditioned. (e) Construct an ill-
evv 9/9/04 452 °

conditioned matrix with det A = 1.
Solution:
(a) Since all matrices in its singular value decomposition (8.40) are square, we can take the
determinant as follows:
det A = det P det Σ det QT = ±1 det Σ = ± σ1 σ2 · · · σn ,
since the determinant of an orthogonal matrix is ±1. The result follows upon taking ab-
solute values of this equation and using the fact that the product of the singular values
is non-negative.
(b) No — even simple nondiagonal examples show this is false.
(c) Numbering the singular values in decreasing order, so σk ≥ σn for all k, we conclude
10− k > | det A | = σ1 σ2 · · · σn ≥ σn
n
, and the result follows by taking the nth root.
(d) Not necessarily, since all the singular values could be very small but equal, and so the
condition number would be 1.
(e) The diagonal matrix with entries 10k and 10−k for k À 0, or more generally, any 2 × 2
matrix with singular values 10k and 10−k , has condition number 102 k .
8.5.15. True or false: If det A > 1, then A is not ill-conditioned.
Solution: False. For example, the diagonal matrix with entries 2 · 10k and 10−k for k À 0 has
determinant 2 but condition number 2 · 102 k .
8.5.16. True or false: If A is a symmetric matrix, then its singular values are the same as its
eigenvalues.
Solution: False — the singular values are the absolute values of the nonzero eigenvalues.
8.5.17. True or false: If U is an upper triangular matrix whose diagonal entries are all posi-
tive, then its singular values are the same as its diagonal entries.
!
1 1 √
Solution: False. For example, U = has singular values 3 ± 5.
0 2
8.5.18. True or false: The singular values of A2 are the squares σi2 of the singular values of A.
Solution: False, unless A is symmetric or, more generally,

! r
normal, meaning that A T A = A AT .
√
1 1
For example, the singular values of A = are 32 ± 25 , while the singular values of
0 1
! q
1 2 √
A2 = are 3 ± 2 2 .
0 1
8.5.19. True or false: If B = S −1 A S are similar matrices, then A amd B have the same singu-
lar values.
Solution: False. This is only true if S is an orthogonal matrix.
♦ 8.5.20. A complex matrix A is called normal if it commutes with its Hermitian transpose
A† = AT , as defined in so A† A = A A† .(a) Show that a real matrix is normal if it com-
mutes with its transpose: AT A = A AT . (b) Show that every real symmetric matrix is
normal. (c) Find a normal matrix which is real but not symmetric. (d) Show that the
eigenvectors of a normal matrix form an orthogonal basis of C n under the Hermitian dot
product. (e) Show that the converse is true: a matrix has an orthogonal eigenvector basis
of C n if and only if it is normal. (f ) Prove that if A is normal, the singular values of An
are σin where σi are the singular values of A. Show that this result is not true if A is not
normal.
evv 9/9/04 453 °

Solution:
(a) If A is real, A† = AT , and so the result follows immediately.
T
(b) If A = A! then AT A = A2 = A AT . !
0 1 a b
(c) , or more generally, where b 6= 0, cf. Exercise 1.6.7.
−1 0 −b a
(d)
(e) Let U = ( u1 u2 . . . un ) be the correspponding unitary matrix, with U −1 = U † . Then
A U = U Λ, where Λ is the diagonal eigenvalue matrix, and so A = U Λ U † = U Λ U T .
Then A AT = U ΛU T U Λ U T
♥ 8.5.21. Let A be a nonsingular 2×2 matrix with singular value decomposition A = P Σ Q T and
singular values σ1 ≥ σ2 > 0. (a) Prove that the image of the unit (Euclidean) circle under
the linear transformation defined by A is an ellipse, E = { A x | k x k = 1 }, whose princi-
pal axes are the columns p1 , p2 of P , and whose corresponding semi-axes are the singular
values σ1 , σ2 . (b) Show that if A is symmetric, then the ellipse’s principal axes are the
eigenvectors of A and the semi-axes are the absolute values of its eigenvalues. (c) Prove
that the area of E equals π | det A
!
|. (d) Find the !principal axes, semi-axes,
!
and area of the
0 1 2 1 5 −4
ellipses defined by: (i) , (ii) , (iii) . (e) What happens
1 1 −1 2 0 −3
if A is singular?
Solution:
(a) If k x k = 1, then y = A x satisfies the equation y T B y = 1, where B = A−T A−1 =
P T Σ−2 P . Thus, by Exercise 8.4.23, the principal axes of the ellipse are the columns of
P , and the semi-axes are the reciprocals of the square roots of the diagonal entries of
Σ−2 , which are precisely the singular values σi .
(b) If A is symmetric, P = Q is the orthogonal eigenvector matrix, and so the columns of
P coincide with the eigenvectors of A. Moreover, the singular values σi = | λi | are the
absolute values of its eigenvalues.
(c) From elementary geometry, the area of an ellipse equals π times the product of its semi-
axes. Thus, areaE = π σ1 σ2 ! = π | det A | !
using Exercise 8.5.14.
r r
√ √
2√ 2√ 3 5 3 5
(d) (i) principal axes: , ; semi-axes: 2 + 2 , 2 − 2 ; area: π.
1+ 5 1− 5
! !
1 2
(ii) principal axes: , (but any orthogonal basis of R 2 will also do); semi-axes:
2 −1
√ √
5, 5 (circle); area:
!
5 π.!
3 1 √ √
(iii) principal axes: , ; semi-axes: 3 5 , 5 ; area: 15 π.
1 −3
(iv ) If A = O, then E = {0} is a point. Otherwise, rank A = 1 and its singular value decom-
position is A = σ1 p1 qT 1 where A q1 = σ1 p1 . Then E is a line segment in the direction
of p1 whose length is 2 σ1 .
!
2 −3
8.5.22. Let A = . Write down the equation for the ellipse E = { A x | k x k = 1 }
−3 10
and draw a picture. What are its principal axes? Its semi-axes? Its area?
“ ”2 “ ”2
Solution: 11 10 3
u + 11 v + 11 3 2
u + 11 v = 1 or 109 u2 + 72 u v + 13 v 2 = 121. Since A is
symmetric, the semi-axes are the eigenvalues, which ! are the
! same as the singular values, namely
−1 3
11, 1; the principal axes are the eigenvectors, , ; the area is 11 π. ell1
3 1
evv 9/9/04 454 °

0 1
−46 1
B
8.5.23. Let A = @ −4 6 −1 CA, and let E = { y = A x | k x k = 1 } be the image of the unit
1 −1 11
Euclidean sphere under the linear map induced by A. (a) Explain why E is an ellipsoid
and write down its equation. (b) What are its principal axes and their lengths — the semi-
axes of the ellipsoid? (c) What is the volume of the solid ellipsoidal domain enclosed by E?
Solution:
(a) In view of the singular value decomposition of A = P Σ QT , the set E is obtained by
first rotating the unit sphere according to QT , which doesn’t change it, then stretching
it along the coordinate axes into an ellipsoid according to Σ, and then rotating it ac-
cording to P , which aligns its principal axes with the columns of P . The equation is
(65 u + 43 v − 2 w)2 + (43 u + 65 v + 2 w)2 + (− 2 u + 2 v + 20 w)2 = 2162 , or

2 2 2
1013 u + 1862 u v + 1013 v − 28 u w + 28 v w + 68 w = 7776.
(b) The semi-axes are the eigenvalues: 12, 9, 2; the principal axes are the eigenvectors:
( 1, −1, 2 )T , ( −1, 1, 1 )T , ( 1, 1, 0 )T .
(c) Since the unit sphere has volume 34 π, the volume of E is 43 π det A = 288 π.
♦ 8.5.24. Optimization Principles for Singular Values: Let A be any nonzero m × n matrix.
Prove that (a) σ1 = max{ k A u k | k u k = 1 }. (b) Is the minimum the smallest singular
value? (c) Can you design an optimization principle for the intermediate singular values?
Solution:
(a) k A u k2 = (A u)T A u = uT K u, where K = AT A. According to Theorem 8.28,
max{ k A u k2 = uT K u | k u k = q
1 } is the largest eigenvalue λ1 of K = AT A, hence
the maximum value of k A u k is λ1 = σ1 . Q.E.D.
(b) This is true if rank A = n by the same reasoning, but false if ker A 6= {0}, since then the
minimum is 0, but, according to our definition, singular values are always nonzero.
(c) The k th singular value σk is obtained by maximizing k A u k over all unit vectors which
are orthogonal to the first k − 1 singular vectors.
♦ 8.5.25. Let A be a square matrix. Prove that its maximal eigenvalue is smaller than its maxi-
mal singular value: max | λi | ≤ max σi . Hint: Use Exercise 8.5.24.
Solution: Let λ1 be the maximal eigenvalue, and let u1 be a corresponding unit eigenvector. By
Exercise 8.5.24, σ1 ≥ k A u1 k = | λ1 |.
max{ k A u k | k u k = 1 }
8.5.26. Let A be a nonsingular square matrix. Prove that κ(A) = .
min{ k A u k | k u k = 1 }
Solution: By Exercise 8.5.24, the numerator is the largest singular value, while the denominator
is the smallest, and so the ratio is the condition number.
! !
1 −1 1 −2
8.5.27. Find the pseudoinverse of the following matrices: (a) , (b) ,
−3 3 2 1
0 1 0 1 0 1 0 1
2 0 0 0 1 ! 1 3 1 2 0
1 −1 1
(c) B
@0 −1 C
A, (d)
B
@0 −1 0CA, (e) , (f ) B C B
@ 2 6 A, (g) @ 0 1 1CA.
−2 2 −2
0 0 0 0 0 3 9 1 1 −1
0 1 0 1 0 1 0 1
1 3 1 2 1
0 0A 0 0 0
B − 20 C B 5C
Solution: (a) @ 20 A, (b) @ 5 A, (c) @2 , (d) B
@0 −1 0CA,
− 1 3
− 2 1 0 −1 0 1 0 0
20 20 5 5
evv 9/9/04 455 °

0 1 0 1
1 2 0 1 1
B 15 − 15 C 1 1 3 B 9 − 19 2
9 C
B C B 140 B C
(e) B 1 2 C, (f ) @ 70 140 C
A, (g) B 5 2 1 C.
B
@
− 15 15 C
A 3 3 9
B 18
@ 9 18 C
A
1 2 140 70 140 1 4 7
15 − 15 18 9 − 18
8.5.28. Use the pseudoinverse to find the least squares solution of minimal norm to the follow-
x − 3 y = 2,
x + y = 1, x + y + z = 5,
ing linear systems: (a) (b) 2 x + y = −1, (c)
3 x + 3 y = −2. 2 x − y + z = 2.
x + y = 0.
Solution: 0 1 0 1
! 3 1 ! 1
1 1 + C
B 20
20 A, ? + 1 B − 4C
(a) A = , A = @ x =A =@ A;
3 3 3 1 −2 1
0 1 20 20
0 1 0
− 4
1
1 −3 1 1 1 2 !
0
(b) A = B
@2
C + B
1 A, A = @ 6 3 6 C
A,
? + B
x = A @ −1 A =C
7 ;
3 1 1 − 11
1 1 − 11 11 11 0
0 1 0 1
1 2 9
! B7 7C ! B 7C
1 1 1 B C 5 B C
(c) A = , A+ = B
B
4
− 5 C,
C x? = A+ =B 15 C
B 7 C.
2 −1 1 @ 7 14 A 2 @ A
2 1 11
7 14 7
♥ 8.5.29. Prove that the pseudoinverse satisfies the following identities: (a) (A + )+ = A,
(b) A A+ A = A, (c) A+ A A+ = A+ , (d) (A A+ )T = A A+ , (e) (A+ A)T = A+ A.
Solution: We repeatedly use the fact that the columns of P, Q are orthonormal, and so
P T P = I , QT Q = I .
(a) Since A+ = Q Σ−1 P T is the singular value decomposition of A+ , we have (A+ )+ =

P (Σ−1 )−1 QT = P Σ QT = A.
(b) A A+ A = (P Σ QT )(Q Σ−1 P T )(P Σ QT ) = P Σ Σ−1 Σ QT = P Σ QT = A.
(c) A+ A A+ = (Q Σ−1 P T )(P Σ QT )(Q Σ−1 P T ) = Q Σ−1 Σ Σ−1 P T = Q Σ−1 P T = A+ .
Or, you can use the fact that (A+ )+ = A.
(d) (A A ) = (Q Σ−1 P T )T (P Σ QT )T = P (Σ−1 )T QT QΣT P T = P (Σ−1 )T ΣT P T =
+ T
P P T = P Σ−1 Σ P T = (P Σ QT )(Q Σ−1 P T ) = A A+ .

(e) This follows from part (d) since (A+ )+ = A.
8.5.30. Suppose b ∈ rng A and ker A = {0}. Prove that x = A+ b is the unique solution to the
linear system A x = b. What if ker A 6= {0}?
Solution:
Jordan 8.6. Incomplete Matrices and the Jordan Canonical Form.
8.6.1. For each of the following Jordan matrices, identify the Jordan blocks. Write down the
eigenvalues, the eigenvectors, and the Jordan basis. Clearly identify the Jordan chains.
evv 9/9/04 456 °

0 1
0 1 0 1 4 0 0 0
! ! 1 0 0 0 1 0 B
2 1 −3 0 B0 3 1 0C
(a) , (b) , (c) B
@0 1 1CA, (d)
B
@0 0 1CA, (e) B
C
C.
0 2 0 6 @0 0 3 0A
0 0 1 0 0 0
0 0 0 2
Solution:
(a) One 2 × 2 Jordan block; eigenvalue 2; eigenvector e1 .
(b) Two 1 × 1 Jordan blocks; eigenvalues −3, 6; eigenvectors e1 , e2 .
(c) One 1 × 1 and one 2 × 2 Jordan blocks; eigenvalue 1; eigenvectors e1 , e2 .
(d) One 3 × 3 Jordan block; eigenvalue 0; eigenvector e1 .
(e) An 1 × 1, 2 × 2 and 1 × 1 Jordan blocks; eigenvalues 4, 3, 2; eigenvectors e1 , e2 , e4 .
8.6.2. Write down all possible 4 × 4 Jordan matrices that only have an eigenvalue 2.
Solution:
0 1 0 1 0 1 0 1
2 0 0 0 2 1 0 0 2 0 0 0 2 0 0 0
B
B0 2 0 0CC
B
B0 2 0 0CC
B
B0 2 1 0CC
B
B0 2 0 0CC
B C, B C, B C, B C,
@0 0 2 0A @0 0 2 0A @0 0 2 0A @0 0 2 1A
0 0 0 2 0 0 0 2 0 0 0 2 0 0 0 2
0 1 0 1 0 1 0 1
2 1 0 0 2 1 0 0 2 0 0 0 2 1 0 0
B
B0 2 0 0CC
B
B0 2 1 0CC
B
B0 2 1 0CC
B
B0 2 1 0CC
B C, B C, B C, B C.
@0 0 2 1A @0 0 2 0A @0 0 2 1A @0 0 2 1A
0 0 0 2 0 0 0 2 0 0 0 2 0 0 0 2
8.6.3. Write down all possible 3×3 Jordan matrices that have eigenvalues 2 and 5 (and no others).
Solution:
0 1 0 1 0 1 0 1 0 1 0 1
2 0 0 2 0 0 5 0 0 2 0 0 5 0 0 5 0 0
B
@0 2 0CA,
B
@0 5 0CA,
B
@0 2 0CA,
B
@0 5 0CA,
B
@0 2 0CA,
B
@0 5 0CA,
0 0 5 0 0 2 0 0 2 0 0 5 0 0 5 0 0 2
0 1 0 1 0 1 0 1
2 1 0 5 0 0 2 0 0 5 1 0
B
@0 2 0CA,
B
@0 2 1CA,
B
@0 5 1CA,
B
@0 5 0CA.
0 0 5 0 0 2 0 0 5 0 0 2
8.6.4. Find Jordan bases and the Jordan canonical

0 form1 for the following matrices:
! ! 1 1 1
2 3 −1 −1
(a) , (b) , (c) B
@ 0 1 1 A,
C
0 2 4 −5
0 0 1 0 1
0 1 0 1 2 −1 1 2
−3 1 0 −1 1 1 B
0 2 0 1C
(d) B
@ 1 −3 −1 A,
C
(e) B
@ −2 −2 −2 A,
C
(f ) B
B
@0
C
C.
0 2 −1 A
0 1 −3 1 −1 −1
0 0 0 2
Solution: ! ! !
1 0 2 1
(a) Eigenvalue: 2. Jordan basis: v1 = , v2 = 1 . Jordan canonical form: .
0 3 0 2
! ! !
1 1 −3 1
(b) Eigenvalue: −3. Jordan basis: v1 = , v2 = 2 . Jordan canonical form: .
2 0 0 −3
0 1 0 1 0 1
1 0 0
(c) Eigenvalue: 1. Jordan basis: v1 =B C B C B C
@ 0 A , v2 = @ 1 A , v3 = @ −1 A.
0 1
0 0 1
1 1 0
B
Jordan canonical form: @0 1 1CA.
0 0 1
evv 9/9/04 457 °

0 1 0 1 0 1
1 0 1
B C B C B C
(d) Eigenvalue: −3. Jordan basis: v1 = @ 0 A , v2
= @ 1 A , v3
= @ 0 A.
0
11
0 0
−3 1 0
Jordan canonical form: B
@ 0 −3 1CA.
0 0 −3 0 1 0 1 0 1
−1 0 0
B C B C B
(e) Eigenvalues: −2, 0. Jordan basis: v1 = @ 0 A , v2 = @ −1 A , v3 = @ −1 C
A.
0 1
1 0 1
−2 1 0
@ 0 −2 0 A.
C
0 0 0
0 1 0 1 0 1 0 1
−2 1 −1 0
B B−1 C
1C B C
B0C B 2C B
0C
(f ) Eigenvalue: 2. Jordan basis: v1 = B
B
@ 1A
C
C, v2 = B C, v3 = B
@0A
C
B 1 C, v 4 =
B
B
C
C.
@ 2A @ 0A
0 0 0 − 21
0 1
2 0 0 0
B
0 2 1 0C
B
@0 0 2 1A
C
C.
0 0 0 2
8.6.5. Write down a formula for the inverse of a Jordan block matrix. Hint: Try some small
examples first to help figure out the pattern.
Solution:
0 −1 1
λ − λ−2 λ−3 − λ−4 ... − (−λ)n
B C
B 0
B λ−1 − λ−2 λ−3 ... − (−λ)n−1 C C
B C
−1
B 0
B 0 λ−1 − λ−2 ... − (−λ)n−2 C C
Jλ,n = B
B 0 0 0 λ−1 ... − (−λ) n−3 C.
C
B C
B . .. .. .. .. .. C
B . C
@ . . . . . . A
0 0 0 0 ... λ−1
8.6.6. True or false: If A is complete, every generalized eigenvector is an ordinary eigenvector.

Solution: True. All Jordan chains have length one, and so consist only of eigenvectors.
♦ 8.6.7. Suppose you know all eigenvalues of a matrix as well as their algebraic and geometric
multiplicities. Can you determine the matrix’s Jordan canonical form?
Solution: No in general. If an eigenvalue has multiplicity ≤ 3, then you can tell the size of its
Jordan blocks by the number of linearly independent eigenvectors it has: if it has 3 linearly in-
dependent eigenvectors, then there are 3 1 × 1 Jordan blocks; if it has 2 linearly independent
eigenvectors then there are Jordan blocks of sizes 1 × 1 and 2 × 2, while if it only has one lin-
early independent eigenvector, then it corresponds to a 3 × 3 Jordan block. But if the multiplic-
ity of the eigenvalue is 4, and there are only 2 linearly independent eigenvectors, then it could
have two 2 × 2 blocks, or a 1 × 1 and an 3 × 3 block. Distinguishing between the two cases is a
difficult computational problem.
8.6.8. True or false: If w1 , . . . , wj is a Jordan chain for a matrix A, so are the scalar multiples
c w1 , . . . , c wj for any c 6= 0.
Solution: True. If zj = c wj , then A zj = cA wj = c λwj + c wj−1 = λzj + zj−1 .
8.6.9. True or false: If A has Jordan canonical form J, then A2 has Jordan canonical J 2 .
evv 9/9/04 458 °

Solution: False. Indeed, the square of a Jordan matrix is not necessarily a Jordan matrix, e.g.,
0 1 0 1
1 1 0 2 1 2 1
B C B
@0 1 1A = @0 1 2CA.
0 0 1 0 0 1
♦ 8.6.10. (a) Give an example of a matrix A such that A2 has an eigenvector that is not an eigen-
vector of A. (b) Show that every eigenvalue of A2 is the square of an eigenvalue of A.
Solution: !
0 1
(a) Let A = . Then e2 is an eigenvector of A2 = O, but is not an eigenvector of A.
0 0
(b) Suppose A = S J S −1 where J is the Jordan canonical form of A. Then A2 = S J 2 S −1 .
Now, even though J 2 is not necessarily a Jordan matrix, cf. Exercise 8.6.9, since J is
upper triangular with the eigenvalues on the diagonal, J 2 is also upper triangular and
its diagonal entries, which are its eigenvalues and the eigenvalues of A2 , are the squares
of the diagonal entries of J. Q.E.D.
8.6.11. Let A and B be n × n matrices. According to Exercise 8.2.22, the matrix products A B
and B A have the same eigenvalues. Do they have the same Jordan form?
Solution: No. ! ! ! !
1 1 0 1 0 1 0 0
A simple example is A = ,B= , so A B = , whereas B A = .
0 0 0 0 0 0 0 0
♦ 8.6.12. Prove Lemma 8.49.
Solution: First, since Jλ,n is upper triangular, its eigenvalues are on the diagonal, and hence λ
is the only eigenvalue. Moreover, v = ( v1 , v2 , . . . , vn )T is an eigenvector if and only if
(Jλ,n − λ I )v = ( v2 , . . . , vn , 0 )T = 0. This requires v2 = · · · = vn = 0, and hence v must be a
scalar multiple of e1 . Q.E.D.
♦ 8.6.13. (a) Prove that a Jordan block matrix J0,n with zero diagonal entries is nilpotent, as in
Exercise 1.3.13. (b) Prove that a Jordan matrix is nilpotent if and only if all its diagonal
entries are zero. (c) Prove that a matrix is nilpotent if and only if its Jordan canonical
form is nilpotent. (d) Explain why a matrix is nilpotent if and only if its only eigenvalue
is 0.
Solution:
k
(a) Observe that J0,n is the matrix with 1’s along the (k + 1)st upper diagonal, i.e., in posi-
n
tions (i, k + i + 1). Thus, for k = n, all entries are all 0, and so J0,n = O.
(b) Since a Jordan matrix is upper triangular, the diagonal entries of J k are the k th powers
of diagonal entries of J, and hence J m = O requires that all its diagonal entries are
zero. Moreover, J k is a block matrix whose blocks are the k th powers of the original
Jordan blocks, and so J m = O where m is the maximal size Jordan block.
(c) If A = S J S −1 , then Ak = S J k S −1 and hence Ak = O if and only if J k = O.
(d) This follow from parts (c–d).
8.6.14. Let J be a Jordan matrix. (a) Prove that J k is a complete matrix for some k ≥ 1
if and only if either J is diagonal, or J is nilpotent with J k = O. (b) Suppose that A is
an incomplete matrix such that Ak is complete for some k ≥ 2. Prove that Ak = O. (A
simpler version of this problem appears in Exercise 8.3.8.)
Solution:
(a) Since J k is upper triangular, Exercise 8.3.12 says it is complete if and only if it is a di-
evv 9/9/04 459 °

agonal matrix, which is the case if and only if J is diagonal, or J k = O.
(b) Write A = S J S −1 in Jordan canonical form. Then Ak = S J k S −1 is complete if and
only if J k is complete, so either J is diagonal, and so A is complete, or J k = O and so
Ak = O.
♥ 8.6.15. The Schur Decomposition: In practical computations, there are severe numerical diffi-
culties in constructing the Jordan canonical form, [ 59 ]. A simpler version, due to I. Schur,
often suffices in applications. The result to be proved is that if A is any n × n matrix, then
there exists a unitary matrix Q, as defined in Exercise 5.3.25, such that Q† A Q = U is up-
per triangular. (a) Explain why A and U have the same eigenvalues. How are their eigen-
vectors related? (b) To establish the Schur Decomposition, we work by induction. First
show that if Q1 is an unitary
!
matrix that has a unit eigenvector of A as its first column,
λ c
then Q†1 A Q1 = where λ is the eigenvalue, and B has size (n − 1) × (n − 1). Then
0 B
use the induction hypothesis for B to complete the induction step. (c) Is the Schur Decom-
position unique? (d) Show that if A has all real eigenvalues, then Q will be an orthogonal
matrix. (e) Show that if A is symmetric, then U is diagonal. Which result do we recover?
(f ) Establish a Schur Decomposition for the following0 matrices: 1 0 1
! ! ! −1 −1 4 −3 1 0
2 3 1 −2 3 1 B
(i) , (ii) , (iii) , (iv ) @ 1 3 −2 A, (v ) @ 1 −3 −1 C
C B
A.
0 2 −2 1 −1 1
1 1 −1 0 1 −3
Solution:
(a)
♥ 8.6.16. The Cayley-Hamilton Theorem: Let pA (λ) = det(A − λ I ) be the characteristic polyno-
mial of A. (a) Prove that if D is a diagonal matrix, then† pD (D) = O. Hint: Leave pD (λ)
in factored form. (b) Prove that if A is complete, then pA (A) = O. (c) Prove that if J
is a Jordan block, then pJ (J) = O. (d) Prove that this also holds if J is a Jordan matrix.
(e) Prove that any matrix satisfies its own characteristic equation: pA (A) = O.
Solution:
n
Y
(a) If D = diag (d1 , . . . , dn ), then pD (λ) = (λ − di ). Now D − di I is a diagonal matrix
i=1
n
Y
with 0 in its ith diagonal position. The entries of the product pD (D) = (D − di I ) of
i=1
diagonal matrices is the product of the individual diagonal entries, but each such prod-
uct has at least one zero, and so the result is a diagonal matrix with all 0 diagonal en-
tries, i.e., the zero matrix.
(b) First, according to Exercise 8.2.31, similar matrices have the same characteristic poly-
nomials, and so if A = S D S −1 then pA (λ) = pD (λ). On the other hand, if p(λ) is any
polynomial, then p(S D S −1 ) = S −1 p(D) S. Therefore, if A is complete, we can diago-
nalize A = S D S −1 , and so, by part (a) and the preceding two facts,
pA (A) = pA (S D S −1 ) = S −1 pA (D) S = S −1 pD (D) S = O.
(c) The characteristic polynomial of the upper triangular Jordan block matrix J = J µ,n
with eigenvalue µ is pJ (λ) = (λ − µ)n . Thus, pJ (J) = (J − µ I )n = J0,n
n
= O by Exercise
8.6.13.
(d) The determinant of a (Jordan) block matrix is the product of the determinants of the
†
See Exercise 1.2.33 for the basics of matrix polynomials.
evv 9/9/04 460 °

individual blocks. Moreover, by part (c), substituting J into the product of the charac-
teristic polynomials for its Jordan blocks gives zero in each block, and so the product
matrix vanishes.
(e) Same argument as in part (b), using the fact that a matrix and its Jordan canonical
form have the same characteristic polynomial.
♦ 8.6.17. Prove that the n vectors constructed in the proof of Theorem 8.46 are linearly inde-
pendent and hence form a Jordan basis. Hint: Suppose that some linear combination van-
ishes. Apply B to the equation, and then use the fact that we started with a Jordan basis
for W = rng B.
Solution: Suppose some linear combination vanishes:
c1 w1 + · · · cn−r+k wn−r+k + d1 z1 + · · · dr−k zr−k = 0.
We multiply by B, to find c1 B w1 + · · · cn−r+k B wn−r+k = 0 since the zi ∈ ker B.
evv 9/9/04 461 °


CH 8 Sol

Uploaded by

Copyright:

Available Formats

CH 8 Sol

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

CH 8 Sol

Uploaded by

Copyright:

Available Formats

Solutions — Chapter 8

sds 8.1. Simple Dynamical Systems.

8.1.1. Solve the following initial value problems:

evv 9/9/04 415 °

evv 9/9/04 416 °

ev 8.2. Eigenvalues and Eigenvectors.

8.2.1. Find the eigenvalues0and eigenvectors

evv 9/9/04 417 °

Solution: The eigenvalues are ±1 with eigenvectors ( sin θ, cos θ ∓ 1 )T .

evv 9/9/04 418 °

evv 9/9/04 419 °

and so ωk is an eigenvector with corresponding eigenvalue e2 k π i /n .

evv 9/9/04 420 °

evv 9/9/04 421 °

8.2.30. An elementary reflection matrix has the form Q = I − 2 u uT , where u ∈ R n is a unit

evv 9/9/04 422 °

evv 9/9/04 423 °

evv 9/9/04 424 °

evv 9/9/04 425 °

8.2.43. Find all invariant subspaces, cf. Exercise 7.4.32, of a rotation in R 3 .

evv 9/9/04 426 °

evv 9/9/04 427 °

evv 9/9/04 428 °

diag 8.3. Eigenvector Bases and Diagonalization.

8.3.1. Which of the following are complete eigenvalues for the

evv 9/9/04 429 °

evv 9/9/04 430 °

evv 9/9/04 431 °

evv 9/9/04 432 °

evv 9/9/04 433 °

evv 9/9/04 434 °

evv 9/9/04 435 °

evsym 8.4. Eigenvalues of Symmetric Matrices.

evv 9/9/04 436 °

evv 9/9/04 437 °

evv 9/9/04 438 °

evv 9/9/04 439 °

evv 9/9/04 440 °

C ωk = (c0 + c1 e2 k π i /n + c2 e4 k π i /n + · · · + +cn−1 e2 (n−1) k π i /n ) ωk

and using Exercise 8.2.12(c). ! !

8.4.13. Write out the spectral decomposition of the0following matrices:

evv 9/9/04 441 °

λ2 = −1, v2 = ( −1, 2 )T , (d) λ1 = 2, v1 = ( 2, 1 )T , λ2 = 2, v2 = ( 1, 2 )T .

evv 9/9/04 442 °

evv 9/9/04 443 °

Solution: Moments of inertia: 4, 2, 1; principal directions: ( 1, 2, 1 )T , ( −1, 0, 1 )T , ( 1, −1, 1 )T .

(i) -1 -0.5 0.5 1 ellipse with semi-axes 1, 12 and principal axes ( 1, 0 )T , ( 0, 1 )T .

evv 9/9/04 444 °

(c) If K is symmetric and positive semi-definite it is a parabola; if K is indefinite, a hy-

evv 9/9/04 445 °

8.4.27. Find all positive definite orthogonal matrices.

evv 9/9/04 446 °

evv 9/9/04 447 °

Solution: max{ xT K x | k x k = 1 } = λ1 is the largest eigenvalue of K. On the other hand, K −1

evv 9/9/04 448 °

♦ 8.4.42. Reformulate Theorem 8.30 as a minimum principle for intermediate eigenvalues.

8.4.43. Under the set-up of Theorem 8.30, explain why

evv 9/9/04 449 °

SVD 8.5. Singular Values.