Archive for

MATH 224, Linear Algebra 2, Spring 2022

Bilkent University, Laurence Barker, 10 June 2022.

Source file: arch224spr22.tex

page 2: Course specification.

page 4: Homeworks, with solutions.
page 17: Quizzes, with solutions.
page 19: Midterm 1, with solutions.
page 22: Midterm 2, with solutions.
page 24: Makeup 2.
page 25: Final, with solutions.

1
 Course specification
MATH 224, Linear Algebra 2, Spring 2022
Laurence Barker, Bilkent University. Version: 14 April 2022.

Course Aims: To aquire theoretical and practical skill at some techniques of linear algebra
that advance beyond a first course such as MATH 223 Linear Algebra 1.
Course Description: This is a second course in linear algebra, more advanced and more
theoretically inclined than MATH 223. To obtain a satsfactory grade, it will be necessary
to solve most of the routine problems and to answer brief theoretical questions. To obtain a
high grade, it will also be necessary to solve some of the difficult problems and to give clear
explanations.
Course Requirements: Knowledge of material in a first course in linear algebra, notably,
the technique of diagonalization and the material behind it, such as the theory of coordinate
transformations, the theory of eigenvalues and eigenvectors.
Instructor: Laurence Barker, Office SAZ 129, barker at fen dot bilkent dot edu dot tr.
Assistant: Serkan Sonel, serkan dot sonel at bilkent dot edu dot tr.
Recommanded textbooks:
• Stephen H. Friedberg, Arnold J. Insel, Lawrence E. Spence, “Linear Algebra”, 5th edition,
(Pearson, 2018).
• Bill Jacob, “Linear Algebra”, (W. H. Freeman, New York, 1990).
• Tom M. Apostol, “Linear Algebra”, (Wiley, New York, 1997).
Supplementary material on my homepage: [To be supplied later.]
• Homeworks, with solutions: homework224spr22.pdf .
• Quizzes, with solutions: quiz224spr22.pdf .
• Course notes on selected topics: PDF files with names of the form: LinAlgTheoNotes... .

Classes: Tuesdays 10:30 - 12:20, Thursdays 15:30 - 16:20, room SAZ-04.

Office Hours: Thursdays 16:30 - 17:20, in my office, room SA-129.

Office hours is for all the students on the course, not just the proficient. If you are having
difficulty with the course, then it is best to come to see me for advice.

Weekly Syllabus
The format below is, Week number; Monday date; Subtopics.
1: Mon 30 Jan - Fri 4 Feb: Review of diagonalization over the real and complex numbers.
Rings, fields, finite fields.
2: Mon 7 Feb - Fri 11 Feb: Vector spaces over arbitrary fields. Steinitz Exchange Lemma
and applications. Error-correcting codes.

2
3: Mon 14 Feb - Fri 18 Feb: Error-correcting codes. Unique factorization of polynomial
rings. Annihilating polynomials of a matrix and of a linear map.
4: Mon 21 Feb - Fri 25 Feb: Minimal polynomial of a linear map. Generalized eigenvectors.
5: Mon 28 Feb - Fri 4 Mar: Generalized eigenvectors. Jordan normal form.
6: Mon 7 Mar - Wed 9 Mar: Cayley–Hamilton Theorem and applications.
7: Mon 14 Mar - Fri 18 Mar: Inner product spaces. Cauchy–Schwartz inequality. Triangle
inequality. Orthonormal bases. Pearson correlation coefficient.
8: Mon 21 Mar - Fri 25 Mar: Orthonormal bases. Gram–Schmidt orthogonalization
9: Mon 28 Mar - Fri 1 Apr: Quadratic forms. Bilinear forms. Sylvester’s law of inertia.
10: Mon 4 Apr - Fri 8 Apr: Quadratic forms. Bilinear forms. Sylvester’s law of inertia.
11: Mon 11 Apr - Fri 15 Apr: Symmetric, Hermitian, orthogonal and unitary operators,
their diagonmalizability and their spectral theorems.
12: Mon 18 Apr - Fri 22 Apr: Symmetric, Hermitian, orthogonal and unitary operators,
their diagonmalizability and their spectral theorems.
13: Mon 25 Apr - Fri 29 Apr: Skew-symmetric bilinear forms. Groups preserving bilinear
forms.
14: Thur 5 May - Fri 6 May: Rational canonical form.
15: Mon 9 May - Fri 13 May: Review for Final.

Assessment: The method of assessment is by curve. Grades F, FX, FZ are to be for candidates
judged to be thoroughly incompetent in the routine course material.
• Quizzes 10% (marks for worst 2 quizzes to be dropped).
• Midterm I, 30%, face-to-face exam, Thursday 17 March, 15:30 - 17:20, rooms SA-Z04, B-108.
• Midterm II, 30%, face-to-face exam, Thursday 21 April, 15:30 - 17:20, rooms SA-Z04, B-108.
• Final, 30%, face-to-face exam, Friday, 20 May.
FZ criteria: less than 30% of the sum of the Midterm 1 and Midterm 2 marks.

Masks: Masks must be worn at all times in class.

Speaking: All speaking in class must address everyone in the room. Questions and comments
are welcome: they make learning easier, and they make teaching easier too.
Class Announcements: All students, including any absentees from a class, will be deemed
responsible for awareness of class announcements.

3
 Homework
MATH 224, Linear Algebra 2, Spring 2022, Laurence Barker
version: 15 April 2022

Homework 1: page 5
Solutions 1: page 6
Homework 2: page 8
Solutions 2: page 9
Homework 3: page 12
Solutions 3: page 14

4
 Homework 1
Solutions to be discussed in class on or after Tuesday 8 March.
Question 1.1: The modulus of a quaternion q = t + ix + jy + kz, with t, x, y, z ∈ R, is
defined by the Pythagorean formula
√
|q| = t2 + x2 + y 2 + z 2 .
As elements of Mat2 (C), write
[ ] [ ] [ ] [ ]
1 0 −i 0 0 −1 0 i
I= , I= , J = , K= .
0 1 0 i 1 0 i 0
(a) Calculate the determinant of the matrix tI + xI + yJ + zK.
(b) Using part (a), show that, given q, q ′ ∈ H, then |qq ′ | = |q|.|q ′ |.
(c) Let q be a non-zero quaternion. Again writing q = t + ix + jy + kz, give a formula for q −1 .
Question 1.2: A field F is said to be finite provided the underlying set F is finite, in other
words, provided F has only finitely many elements. Show that, if F is finite, then |F | is a
power of a prime, in other words, |F | = pm for some prime p and some positive integer m.
(Hint: First show that F contains a copy of the ring Z/n for some positive integer n. Then
make use of the fact that every finite-dimensional vector space has a basis.)
Question 1.3: Suppose F is finite. Consider a system of F -linear equations Ax = y. Thus, x
and y are column vectors over F and A is a matrix over F . Show that, for each y, the number
|{x : Ax = y}| is zero or a power of a prime.
Question 1.4: Prove the following two propositions concerning the polynomial ring F [t],
where F is a field:
(a) Given nonzero a(t), b(t) ∈ F [t], then there exist unique q(t), r(t) ∈ F [t] such that a(t) =
q(t)b(t) + r(t) and deg(r(t)) < deg(b(t)).
(b) Let a(t), b(t) ∈ F [t]. Let c(t) be a polynomial of maximal degree such that c(t) divides a(t)
and b(t). Then there exist u(t), v(t) ∈ F [t] such that c(t) = u(t)a(t) + v(t)b(t). Furthermore,
there exists a unique such monic c(t).
Question 1.5: Consider the linear coding scheme with Hamming matrix
[ ]
1 0 0 1 0
H= .
0 1 1 0 1
(a) Write down the generating matrix G for the coding scheme.
(b) Write down a decoding table, including the column of syndromes.
(c) Encode the message words 100, 110, 111.
(d) For the received words 00011, 00111, 01111, 11111, write down the syndromes, then write
down the decoded words.
(e) What is the rate of the code?
(f ) If a single codeword is transmitted, what is the maximum number of errors of transmission
(the maximum number of inversions of binary digits) such that any error can be detected?
What is the maximum number of errors of transmission (the maximum number of inversions
of binary digits) such that any error can be corrected?

5
 Solutions 1

1.1: Part (a). We have

t − ix −y + iz
det(tI + xI + yJ + zK) = = |t + ix|2 + |y + iz|2 = |q|2 .
y + iz t + ix

Part (b). The required equality follows from part (a) because det(AB) = det(A) det(B)
for 2 × 2 matrices A and B over C.
Part (c). By the formula for the inverse of a 2 × 2 matrix with non-zero determinant, if
q ̸= 0, then
[ ]−1 [ ]
t − ix −y + iz 2 2 2 2 −1 t + ix y + iz
= (t + x + y + z )
y + iz t + ix −y + iz t − ix

and hence q −1 = (t − ix − jy − kz)/(t2 + x2 + y 2 = z 2 ).

1.2: Let K be the subring of F generated by the unity element 1F . Let p be the smallest
positive integer such that p1F = 0. Then K is a copy of the ring Z/p of modulo p integers.
For x, y ∈ K, if xy = 0 then x = 0 or y = 0. Therefore, p is prime and K is a field.
We can regard F as a vector space over K with the same addition operation and with
scalar multiplication operation K × F → F coming from the multiplication operation of F .
Letting m be the dimension of F as a K-vector space, then |F | = pm .
1.3: Let S = {x : Ax = y}. We may assume that S ̸= ∅. Let x0 ∈ S. For a column vector x
with the appropriate number of coordinates, we have x ∈ S if and only if x − x0 belongs to the
null space of A. So |S| = |Null(A)| = |F |n where n = dim(Null(A)). In Question 2, we showed
that |F | is a power of a prime. Therefore, |S| is a power of a prime.
1.4: Part (a). Let m = deg(a) and n = deg(b). We may assume that m ≥ n. Write
a(t) = an tn + ... + a1 t + a0 and b(t) = bm tm + ... + b1 t + b0 . Consder a polynomial q(t) =
qm−n tm−n + q1 t + q0 . The coefficient of tm in a − qb is am − qm−n bn . Putting qm−n = am /bn ,
we ensure that a − qb has degree less than m. Inductively, for m − n ≥ s ≥ 1, having solved for
qm−n , ..., qs so as to ensure that a − qb has degree less than n + s, we can consider coefficients
of tn+s−1 to solve for qs−1 so as to ensure that a − qb has degree less than n + s − 1. The case
s = 0 is the required conclusion.
Part (b). We apply the Euclidian algorithm. Again, we may assume that deg(a) ≥ deg(b).
Put r0 = a and r1 = b. Inductively, having defined r0 , ..., ri , we terminate the process when
ri = 0, otherwise, we define polynomials qi and ri−1 such that

ri+1 = qi ri + ri−1

and deg(ri ) > deg(ri−1 ). Since the degrees deg(ri ) are strictly decreasing, the process must
terminate. Let k be such that rk+1 = 0. Let c = rk . If c divides ri−1 and ri , then c divides ri+1
and ri . So, by an inductive argument, c divides a and b. On the other hand, another inductive
argument shows that each ri has the form ui a + vi b for polynomials ui and vi . In particular,
c = ua + vb for some u and v.
Finally, if d is a polynomial dividing a and b, then d must divide c because c = ua + vb.
In particular, deg(d) ≤ deg(c) and if deg(d) = deg(c) then d must be an F -multiple of c. The
required conclusion follows.

6
  
1 0 0
0 1 0
 
1.5: Part (a). We have G = 
0 0 1.
1 0 0
0 1 1
Part (b). The decoding table is as follows.
000 001 010 011 100 101 110 111 syndrome
00000 00101 01001 01100 10010 10111 11011 11110 00
00001 00100 01000 01101 10011 10110 11010 11111 01
00010 00111 01011 01110 10000 10101 11001 11100 10
00011 00110 01010 01111 10001 10100 11000 11101 11
Part (c). The encodings of 100, 110, 111 are 10010, 11011, 11110, respectively.
Part (d). The received words 00011, 00111, 01111, 11111 have syndromes 11, 10, 11, 01
and decodings 000, 001, 011, 111, respectively
Part (e). The rate is 3/5.
Part (f). The minimum weight of a nonzero codeword is 2. So up to 1 error of transmission
can always be detected, and up to 0 errors of transmission can always be corrected.

7
 Homework 2
Solutions to be discussed in class on or after Tuesday 15 March.
 
−1 2 −1
Question 2.1: Consider the matrix A =  0 −1 1 .
0 0 2
Express A in the form A = P EP −1 where E has Jordan normal form.
 
−2 1 0

Question 2.2: Repeat Question 1 with A = −1 0 0.
−1 −1 1
Question 2.3: Let A be an n × n matrix over an algebraically closed field F .
(a) Show that, when n = 3, the Jordan normal form of A is determined by the characteristic
polynomial and the minimal polynomial of A.
(b) In the case n = 4, give a counter-example to the conclusion of part (a).
Question 2.4: We write F2 for the field with order 2, and we write Matn (F2 ) to denote the
ring of n × n matrices over F2 . Let ∼ denote the equivalence relation on Matn (F2 ) such that,
given A, B ∈ Matn (F2 ), then A ∼ B when A is similar to B, in other words, A = P BP −1 for
some invertible P ∈ Matn (F2 ).
(a) When n = 2, how many equivalence classes of ∼ include a matrix that has Jordan normal
form?
(b) When n = 3, how many equivalence classes of ∼ include a matrix that has Jordan normal
form?
(c) When n = 4, how many equivalence classes of ∼ include a matrix that has Jordan normal
form?

8
 Solutions 2

2.1: The eigenvalues are −1 and 2 with multiplicities 2 and 1, respectively.

Let f1 be an eigenvector with eigenvalue −1. Write f1 = (x, y, z). Then

−x + 2y − z = −x , −y + z = −y , 2z = −z .

We have y = z = 0. So we can put x = 1 and f1 = (1, 0, 0).

Let f2 be a generalized eigenvector such that Af2 = −f2 + f1 . Write f2 = (x, y, z). Then

−x + 2y − z = −x + 1 , −y + z = −y , 2z = −z .

Hence, z = 0 and y = 1/2. We can put x = 0. Then f2 = (0, 1/2, 0).

Let f3 be an eigenvector with eigenvalue 2. Write f2 = (x, y, z). Then

−x + 2y − z = 2x , −y + z = 2y , 2z = 2z .

We can put z = 1. Then y = 1/3 and, since 3x = 2y − z = −1/3, we have x = −1/9. In

conclusion, we can put
   
1 0 −1/9 −1 1 0
P = 0 1/2 1/3  , E =  0 −1 0 .
0 0 1 0 0 2
 
1 0 1/9
Comment: As an aid to checking the answer, we mention that P −1 
= 0 2 −2/3.
0 0 1
2.2: The characteristic polynomial of A is

λ + 2 −1 0
1 λ 0 = (λ + 2)λ(λ − 1) + λ − 1 = (λ2 + 2λ + 1)(λ − 1) = (λ + 1)2 (λ − 1) .
1 1 λ−1

So the eigenvalues are −1 and 1 with multiplicities 2 and 1, respectively.

Let f1 be an eigenvector with eigenvalue −1. Write f1 = (x, y, z). Then

−2x + y = −x , −x = −y , −x − y + z = z .

We can put x = y = z = 1. Thus, f1 = (1, 1, 1).

Let f2 be a generalized eigenvector such that Af2 = −f2 + f1 . Write f2 = (x, y, z). Then

−2x + y = −x + 1 , −x = −y + 1 , −x − y + z = z + 1 .

We can put y = 1, whereupon x = 0 and z = 1. Thus, f2 = (0, 1, 1).

Putting f3 = (0, 0, 1), then f3 is plainly an eigenvector with eigenvalue 1. We have
   
1 0 0 −1 1 0
P = 1 1 0 , E =  0 −1 0 .
1 1 1 0 0 1

9
  
1 0 0
Comment: As an aid to checking the answer, we mention that P −1 = −1 1 0.
0 −1 1
2.3: Part (a). Write cA (t) and mA (t), respectively, for the characteristic polynomial and the
minimal polynomial of A in F [t]. We run through all the possible cases.
If A has 3 mutually distinct eigenvalues, λ1 , λ2 , λ3 , then A is diagonalizable and

cA (t) = mA (t) = (t − λ1 )(t − λ2 )(t − λ3 ) .

Now suppose A has distinct eigenvalues λ1 and λ2 with multiplicities 1 and 2, respectively.
Then cA (t) = (t − λ1 )(t − λ2 )2 . Up to renumbering of basis elements, the two possible Jordan
normal forms for A are    
λ1 0 0 λ1 0 0
 0 λ2 0  ,  0 λ2 1 
0 0 λ2 0 0 λ2
whose minimal polynomials are (t − λ1 )(t − λ2 ) and cA (t), respectively. Finally, suppose A has
a unique eigenvalue λ. Then cA (t) = (t − λ)3 . Up to renumbering, the three possible Jordan
normal forms for A are
     
λ 0 0 λ 1 0 λ 1 0
0 λ 0 , 0 λ 0 , 0 λ 1
0 0 λ 0 0 λ 0 0 λ

whose minimal polynomials are (t − λ), (t − λ)2 , cA (t), respectively. Evidently, no two of the
six possible cases have the same polynomials cA (t) and mA (t).
Part (b). Let λ ∈ F . The two distinct matrices
   
λ 1 0 0 λ 1 0 0
0 λ 0 0 0 λ 0 0
   
0 0 λ 0 , 0 0 λ 1
0 0 0 λ 0 0 0 λ

are both in Jordan normal form, they both have characteristic polynomial (t − λ)4 ∈ F [t] and
minimal polynomial (t − λ)2 .
2.4: In all three parts, we are to count the equivalence classes of ∼ that include a matrix
having Jordan normal form.
Part (a). The answer is 5, since there are 5 possible cases: 2 distinct eigenvalues; unique
eigenvalue 0 and diagonalizable; unique eigenvalue 0 and non-diagonalizable; unique eigenvalue
1 and diagonalizable; unique eigenvalue 2 and non-diagonalizable.
Part (b). We can classify the equivalence classes according to Jordan normal form, bearing
in mind that renumberings of the Jordan basis yield the same Jordan normal form up to
permutation of the Jordan block matrices.
There are no cases with 3 mutually distinct eigenvalues. In the the case of 2 distinct
eigenvalues, there are 2 choices for which of the eigenvalues 0 or 1 has multiplicity 2 and then,
as we saw in Question 2.3, there are 2 choices for the Jordan normal form. That yields 4
equivalence classes. Meanwhile, in the case of a unique eigenvalue, there are 2 choices for the
eigenvalue, then 3 choices for the Jordan normal form. That yields 6 equivalence classes. So
the total number of equivalence classes is 4 + 6 = 10.

10
 Part (c). Let (m1 , m2 , ...) be the dimensions of the Jordan blocks with eigenvalue 0, ar-
ranged such that m1 ≥ m2 ≥ .... Let (n1 , n2 , ...) be defined similarly for eigenvalue 1. Let
m = m1 + m2 + ... and n = n1 + n2 + ..., which are the multiplicities of the eigenval-
ues 0 and 1. The number of equivalence classes is the number of tuples having the form
(m1 , m2 , ...; n1 , n2 , ...) such that m + n = 4.
When m = 0, the possible tuples are (; 1, 1, 1, 1), (; 2, 1, 1), (; 2, 2), (; 3, 1), (; 4).
When m = 1, the possilities are (1; 1, 1, 1), (1; 2, 1), (1; 3).
When m = 2, they are (1, 1; 1, 1), (1, 1; 2), (2; 1, 1), (2; 2).
When m = 3, the number of possibilities is the same as for m = 1.
When m = 4, the number of possibilities is the same as for m = 0.
So the number of equivalence classes is 5 + 3 + 4 + 3 + 5 = 20.
Comment: Consider a field F . Recall that, when F is algebraically closed, every square matrix
over F is similar to a matrix having Jordan normal form. For an arbitrary field F and a square
matrix A over F , we may have to extend to a larger field in order to express A in the form
A = P EP −1 where E has Jordan normal form.
Consider the matrix  
0 1 0
A = 0 0 1 ∈ Mat3 (F2 ) .
1 0 0
We have A3 = I. But there does not exist a non-identity matrix E ∈ Mat3 (F2 ) such that
E 3 = I. So A cannot be expressed in Jordan normal form over F2 . However, now let us view
A as a matrix over the field F4 with order 4. The field F4 = {0, 1, ω, ω 2 } has the following
addition and multiplication tables.
y y
x+y 0 1 ω ω2 xy 0 1 ω ω2
x 0 0 1 ω ω2 x 0 0 0 0 0
1 1 ω ω2 0 1 0 1 ω ω2
ω ω ω2 0 1 ω 0 ω ω2 1
ω2 ω2 ω 1 0 ω2 0 ω2 1 ω
It is not hard to check that, as a product of three matrices over F4 , we have
   −1
1 1 1 1 0 0 1 1 1
A = 1 ω ω 2  0 ω 0  1 ω ω 2  .
1 ω2 ω 0 0 ω2 1 ω2 ω

Thus, in fact, A is diagonalizable over F4 .

11
 Homework 3
Solutions to be discussed in class on Thursday 14 April and Tuesday 19th April.
Question 3.1: Let π be the linear map R2 ← R2 such that π(x, y) = (x + y, x + y)/2. Find:
(a) The matrix representing π with respect to the standard basis {(1, 0), (0, 1)} of R2 .
(b) Evaluate the trace tr(π) and the determinant det(π).
Question 3.2: Let e1 = (1, 1, 1) and e2 = (1, 2, 2) and e3 = (1, 2, 3) in R3 . Apply the
Gram–Schmidt process to the basis {e1 , e2 , e3 } to obtain an orthonormal basis for R3 .
Question 3.3: Let

U = {(x1 , x2 , x3 , x4 ) ∈ C4 : x1 + x3 = x2 + x4 = 0}

as a subspace of the inner product space C4 .

(a) Find an orthonormal basis for U .
(b) Find an orthonormal basis for the orthogonal complement U ⊥ of U in C4 .
(c) Find the matrix A representing the orthogonal projection U ← C4 with respect to the
standard basis of C4 .
(d) Find the matrix B representing the orthogonal projection U ⊥ ← C4 with respect to the
standard basis of C4 .
Question 3.4: Let V and W be subspaces of a finite-dimensional inner product space U over
C. We write V ⊥ for the orthogonal complement of V in U . Show that:
(a) We have (V ⊥ )⊥ = V .
(b) We have (V + W )⊥ = V ⊥ ∩ W ⊥ .
(c) We have (V ∩ W )⊥ = V ⊥ + W ⊥ .
(d) The inner ∑ product space ℓ2 (C) consists of the sequences (x1 , x2 , ...) of complex num-
bers such that n |xn |2 < ∞. Let V be the subspace of ℓ2 (C) consisting of those sequences
(x1 , x2 , ...) such that only finitely many of the xi are non-zero. Show that, if we drop the
assumption that U is finite-dimensional and put U = ℓ2 (C), then V is a counter-example to
part (a).
(e) Again dropping the assumption that U is finite-dimensional, give a counter-example to
part (c). (Hint: Put U and V as in part (d) and find a suitable W .)
Question 3.5: Let n be a positive integer. The n × n discrete Fourier transform matrix
F is the matrix with rows and columns indexed by the ring of modulo n integere Z/n, with
√
(s, t)-entry ζ st / n, where ζ = e2πi/n .
(a) Why is F diagonalizable?
(b) Evaluate F 2 . Hence show that every eigenvalue of F belongs to the set {1, i, −1, −i}.
(c) Suppose n is odd. Evaluate dimC (E1 ⊕ E−1 ) and dimC (Ei ⊕ E−i ), where Eλ denotes the
eigenspace for eigenvalue λ.
Question 3.6: Let ρ be an orthogonal operator on a finite-dimensional real inner product

12
space V . Show that, with respect to some basis of V , the matrix representing ρ has the form
 
A1 0 ... 0
0 A2 ... 0 
 
 .. .. .. 
 . . . 
0 0 ... Am
[ ]
cos(θ) − sin(θ)
where each As is either a 1 × 1 matrix [1] or [−1] or else a 2 × 2 matrix .
sin(θ) cos(θ)
(Hint: Extend to the complex numbers and adapt the trick we used to show that real symmetric
matrics are diagonalizable.)

13
 Solutions 3
[ ]
1 1 1
3.1: Part (a). The required matrix is P = .
2 1 1
Part (b). We have tr(π) = tr(P ) = 1 and det(π) = det(P ) = 0.
Comment: Many invariants of a linear map, such as trace, determinant, characteristic poly-
nomial, minimal polynomial, are features of any matrix representing the linear map. Given a
linear map, then the representing matrix depends on the choice of basis. But, when we change
the basis, the invariants do not change. That is why they are called invariants. (I emphasize
this fundamental point because grasping it is the beginning of a comprehension of the theory
of linear algebra.)
3.2: We calculate the orthogonal basis {f1′ , f2′ , f3′ } given by
⟨f1′ | e2 ⟩ ′ ⟨f1′ | e3 ⟩ ′ ⟨f2′ | e3 ⟩ ′
f1′ = e1 , f2′ = e2 − f , f3′ = e3 − f − f .
∥ f1′ ∥2 1 ∥ f1′ ∥2 1 ∥ f2′ ∥2 2

and then the orthonormal basis {f1 , f2 , f3 } given by fi = fi′ / ∥ fi′ ∥. We have f1′ = (1, 1, 1).
Since
⟨f1′ | e2 ⟩ 1+2+2 5
′ = =
∥ f1 ∥ 2 1+1+1 3
5 1
we have f2′ = (1, 2, 2) − (1, 1, 1) = (−2, 1, 1). Since
3 3
⟨f1′ | e3 ⟩ 1+2+3 ⟨f2′ | e3 ⟩ (−2 + 2 + 3)/3 3
′ = =2, ′ = =
∥ f1 ∥ 2 3 ∥ f2 ∥ 2 (4 + 1 + 1)/9 2
1 1
we have f3′ = (1, 2, 3) − 2(1, 1, 1) − (−2, 1, 1) = (0, −1, 1). Therefore,
2 2
√ √ √
f1 = (1, 1, 1)/ 3 , f2 = (−2, 1, 1)/ 6 , f3 = (0, −1, 1)/ 2 .

3.3: We define
√ √ √ √
f1 = (1, 0, −1, 0)/ 2 , f2 = (0, 1, 0, −1)/ 2 , f3 = (1, 0, 1, 0)/ 2 , f4 = (0, 1, 0, 1)/ 2 .

Note that U = {x ∈ C4 : ⟨f1 | x⟩ = ⟨f2 | x⟩ = 0}.

Part (a). An orthonormal basis for U is {f1 , f2 }.
Part (b). An orthonormal basis for U ⊥ is {f3 , f4 }.
Part (c). Given x = (x1 , x2 , x3 , x4 ) ∈ C4 , then the orthogonal projection of x to U is

⟨f1 | x⟩f1 + ⟨f2 | x⟩f2 = (x1 − x3 , x2 − x4 , −x1 + x3 , −x2 + x4 )/2 .

 
1 0 −1 0
 0 1 0 −1
Therefore, A = 
−1
.
0 1 0
0 −1 0 1
Part (d). The orthogonal projection of x to U ⊥ is

⟨f3 | x⟩f3 + ⟨f4 | x⟩f4 = (x1 + x3 , x2 + x4 , x1 + x3 , x2 + x4 )/2 .

14
  
1 0 1 0
 0 1 0 1
Therefore, B = 
 1
.
0 1 0
0 1 0 1
3.4: Part (a). As a preliminary, recall that dim(U ) = dim(V ) + dim(V ⊥ ). (This is proved
by extending a basis for V to a basis for U , then applying the Gram–Schmidt process, which
yields an orthonormal basis for V that extends to a orthonormal basis for U .) Hence dim(V ) =
dim((V ⊥ )⊥ ). Plainly, V ≤ (V ⊥ )⊥ . The required conclusion follows.
Part (b). Since V + W ≥ V , we have (V + W )⊥ ≤ V ⊥ . Similarly, (V + W )⊥ ≤ W ⊥ . Hence
(V + W )⊥ ≤ V ⊥ ∩ W ⊥ . For the reverse inequality, let z ∈ V ⊥ ∩ W ⊥ . Given v ∈ V and w ∈ W ,
then ⟨v ⊥ z⟩ = ⟨w | z⟩, hence ⟨v + w | z⟩ = 0 and z ∈ (V + W )⊥ .
Part (c). Using parts (a) and (b), we have

(V ∩ W )⊥ = (V ⊥⊥ ∩ W ⊥⊥ )⊥ = (V ⊥ + W ⊥ )⊥⊥ = V ⊥ + W ⊥ .

Part (d). We have V ⊥ = {0}, hence V ⊥⊥ = U ̸= V .

Part (e). Let U and V be as suggested in the question. Let x = (x1 , x2 , ...) be any element of
ℓ (C) such that infinitely many of the xi are non-zero. (For instance, we can put xi = z i where
2

z is any complex number satisfying 0 < |z| < 1.) Let W = spanC {x}. Then V ∩ W = {0},
hence (V ∩ W )⊥ = U . Meanwhile, w ̸∈ W ⊥ = V ⊥ ∩ W ⊥ . Therefore, (V ∩ W )⊥ > V ⊥ ∩ W ⊥ .
Comment: For each positive integer m, let em be the element of ℓ2 (C) such that the m-th term
of em is 1 and all the other terms are 0. In functional analysis, we write
∞
∑
(x1 , x2 , ...) = xm em
m=1

and we call the infinite set {e1 , e2 , ...} a topological basis for ℓ2 (C).
However, a vector space S over a field F comes equipped with an addition operation
S × S ∋ (s, t) 7→ s + t ∈ S. By the associativity of that operation, we can form finite sums
s1 + ... + sn for s1 , ..., sn in X. By the commutativity of the addition operation, the sum
s1 + ... + sn does not depend on the ordering of the terms sm . In linear algebra, we define
a basis for S to be a∑subset {ei : i ∈ I} ⊆ S such that every element of X can be uniquely
expressed as a sum i xi ei such that xi ∈ F and only finitely many of the indices xi are
nonzero. Sometimes, a basis in that sense is called an algebraic basis. Using Zorn’s Lemma,
it can be proved that every vector space has an algebraic basis.
In particular, ℓ2 (C) must have an algebraic basis. But it would be very difficult, perhaps
impossible, to explicitly specify any algebraic basis for ℓ2 (C). The above infinite sum only
makes sense because ℓ2 (C) has more structure than just that of a vector space. Since ℓ2 (C) is
an inner product space, we can view ℓ2 (C) as a metric space where the distance between two
vectors u and v is the norm ∥ u − v ∥. Via the metric space structure, we can form limits. To
give sense to the above infinite sum, we define it to be the limit
∞
∑ ∑
n
xm em = lim x m em .
n→∞
m=1 m=1

But span{e1 , e2 , ...} is not the whole of ℓ2 (C). In fact, span{e1 , e2 , ...} is precisely the subspace
V defined in the question. Thus, the topological basis {e1 , e2 , ...} for ℓ2 (C) is an algebraic basis
for V and it is not an algebraic basis for ℓ2 (C).

15
3.5: Part (a). The matrix F is unitary, hence diagonalizable.
Part (b). Given r, t ∈ Z/n, then the (r, t) entry of F 2 is
{
1 ∑ rs st 1 ∑ r+t 1 if r = −t,
ζ ζ = ζ s=
n n 0 otherwise.
s∈Z/n s∈Z/n

It follows that F 4 is the identity matrix. So the minimal polynomial of F must divide X 4 ∈
C[X]. The required conclusion follows.
Part (c). Since F is diagonalizable, part (b) yields Cn = E1 ⊕ E−1 ⊕ Ei ⊕ E−i . So, letting
a = dim(E1 ⊕ E−1 ) and b = dim(Ei ⊕ E−i ), we have a + b = n. On the other hand, E1 ⊕ E−1
is the 1-eigenspace of F 2 , while Ei ⊕ E−i is the −1-eigenspace of F 2 . The evaluation of F 2 in
part (b) implies that a − b = tr(F 2 ) = 1. Therefore,

dim(E1 ⊕ E−1 ) = (n + 1)/2 , dim(Ei ⊕ E−i ) = (n − 1)/2 .

3.6: We C-linearly extend ρ to a unitary operator ϕC on the C-vector space CV = V ⊕ iV .

Thus, any R-basis E for V becomes a C-basis for CV , and the matrix representing ρ with
respect to E is also the matrix representing ρC with respect to E.
Let z be an eigenvector of ρC . Since ρC is unitary, the corresponding eigenvalue has the form
e with θ ∈ R. Write z = x + iy where x, y ∈ V . Suppose dim(span{x, y}) = 1. Then we can
iθ

put z = x or z = y whence, perforce, z ∈ V and eiθ = ±1. We have V = span{z} ⊕ span({z}⊥ )

and ρ restricts to an orthogonal operator on span{z}⊥ . The required conclusion follows, in
this case, by an inductive argument on dim(V ).
So we may assume that dim(span{x, y}) = 2. We have ρ(x+iy) = eiθ (x+iy) and, applying
complex conjugation, ρ(x − iy) = e−iθ (x − iy). With respect to the basis {x, y} of span{x, y},
the matrix representing ρ is the 2 × 2 rotation matrix specified in the question. We have
V = span{x, y} ⊕ span({x, y}⊥ ) and ρ restricts to an orthogonal operator on span({x, y}⊥ ).
Again, the required conclusion follows by an inductive argument on dim(V ).

16
 Quizzes, with solutions
MATH 224, Linear Algebra 2, Spring 2022, Laurence Barker
version: 19 May 2022

[ ]
1 1/2
Quiz 1: Diagonalize . That is, express the specified matrix in the form P DP −1
−2 1
with D diagonal.
Solution: Write A for the specified matrix. The charactreristic polynomial of A is

det(λI − A) = λ2 − 2λ + 2
√ [ ]
2± 4−8 x
which has roots = 1 ± i. The eigenvalue 1+i has associated eigenvectors where
2 y
[ ] [ ]
x 1
x + y/2 = (1 + i)x and 2x + y = (1 + i)y. We can put = . Similarly (or by taking
y 2i
[ ]
1
complex conjugates), we see that the eigenvalue 1 − i has an associated eigenvector . So
−2i
[ ] [ ]
1 1 1+i 0
A = P DP −1 , where P = and D = . ⊓⊔
2i −2i 0 1−i
Comment: The example illustrates fact that, to diagonalize a real square matrix, we may need
to extend to the complex numbers.

Quiz 2: The linear code over F2 with Hamming matrix

[ ]
1 1 0 1 0
H=
0 1 1 0 1

has the following decoding table:

000 001 010 011 100 101 110 111 syndrome
00000 00101 01011 01110 10010 10111 11001 11100 00
00001 00100 01010 01111 10011 10110 11000 11101 01
00010 00111 01001 01100 10000 10101 11011 11110 10
01000 01101 00011 00110 11010 11111 10001 10100 11
(a) For the message words 011, 001, 000, write down the corresponding codewords.
(b) For the received words 00110, 01011, 11010, write down the corresponding syndromes and
the corresponding decoded words.
Solution: Part (a). Respectively, those message words have codewords 01110, 00101, 00000.
Part (b). Respectively, those received words have syndromes 11, 00, 11 and decodings 011,
010, 100.
] [
1 1
Quiz 3: Express the matrix A = as A = P EP −1 with E in Jordan normal form.
−4 5

17
Solution: The characteristic polynomial of A is
t − 1 −1
det(tI − A) = = (t − 1)(t − 5) + 4 = t2 − 6t + 9 = (t − 3)2 .
4 t−5
Solving for an eigenvector f1 = (x, y) with eigenvalue 3, we have
[ ][ ] [ ]
1 1 x x
=3 .
−4 5 y y
In other words, x + y = 3x and −4x + 5y = 3y. Thus, y = 2x. We can put f1 = (1, 2). Solving
for f2 = (x, y) such that Af2 = 3f2 + f1 , we have
[ ][ ] [ ] [ ]
1 1 x x 1
=3 + .
−4 5 y y 2
In other words, x + y = 3x + 1 and −4x + 5y = 3y + 2. Thus, y = 2x + 1. We can put
f2 = (0, 1). The Jordan basis {f1 , f2 } yields
[ ] [ ]
1 0 3 1
P = , E= .
2 1 0 3
Comment: By the uniqueness property of Jordan normal form, there is no other solution for
E. Of course, there are infinitely many solutions for P .

Quiz 4: Let e1 = (1, 1) and e2 = (0, 1). Apply the Gram–Schmidt process to the basis {e1 , e2 }
of R2 to obtain an orthonormal basis {f1 , f2 } for R2 . (Recall, the orthonormality condition is
⟨f1 | f1 ⟩ = ⟨f2 | f2 ⟩ = 1 and ⟨f1 | f2 ⟩ = 0.)
Solution: We construct an orthogonal basis {f1′ , f2′ } where f1′ = e1 and f2′ = e2 − ⟨f1 | e2 ⟩f1 .
Then we construct an orthonormal basis {f1 , f2 } where fi = fi′ / ∥ fi′ ∥. We have ∥ f1 ∥2 = 2, so
√ √
f1 = (1/ 2, 1/ 2) .
√ √ √
Hence f2′ = (0, 1) − (1/ 2, 1/ 2)/ 2 = (−1/2, 1/2). Since ∥ f2′ ∥2 = 1/2, we have
√ √
f2 = (−1/ 2, 1/ 2) .

Quiz 5: Which of the following form a subspace of Matn (R) and, in those cases, what is the
dimension?
(a) The set of symmetric matrices on Rn ?
(b) The set of orthogonal matrices on Rn ?
Solution: Part (a), obviously, yes. The dimension is n(n + 1)/2. Letting ϵi,j be the matrix in
Matn (R) with (i, j) entry 1 and all other entries 0, then the subspace of symmetric matrices
has a basis consisting of the matrices ϵi,j + ϵj,i .
Part (b), obviously, no.

Quiz 6: What is the set of real numbers t such that t is the trace of an orthogonal n × n
matrix over R?
Solution: When n = 1, the set is {−1, 1}. Otherwise, the set is the closed interval [−n, n].
Indeed, this is because the real eigenvalues of an orthogonal matrix are 1 and −1, while the
non-real eigenvalues can be arranged in pairs {z, z −1 } where z and hence z −1 are complex
numbers with modulus unity.

18
MATH 224: Linear Algebra 2. Spring 2022. Midterm 1
LJB, 17 March 2022, Bilkent University.
Time allowed: 110 minutes hours. Please put your name on EVERY sheet of your manuscript.
The use of telephones, calculators or other electronic devices is prohibited. The use of red pens
or very faint pencils is prohibited too. You may take the question sheet home.

1: 35 marks. Consider the linear coding scheme over the field {0, 1} with Hamming matrix
[ ]
1 0 1 1 0
H= .
1 1 1 0 1

(a) What is the generating matrix G for the coding scheme?

(b) Encode the message words 000, 001, 011.
(c) Write down a decoding table, including the column of syndromes, ensuring that the received
words 00001, 00010, 00100 all have decoding 000
(d) Using that decoding table, for the received words 10000, 11000, 11001, write down the
syndromes and the decoded words.
(e) How many possible decoding tables are there, assuming that the received words with
decoding 000 all have weight 0 and 1?
 
0 1 0
2: 35 marks. Let A = −2 2 1 as a matrix over C.
0 −2 4
(a) Show that A has unique eigenvalue 2.
(b) Find P and E, with E in Jordan normal form, such that A = P EP −1 .
(c) What is the minumal polynomial of A?

3: 10 marks. Let F be a field, let m and n be positive integers with m ≤ n, and let M
and N be F -vector spaces with dimensions m = dim(M ) and n = dim(N ) such that M is a
subspace of N . Let E be the F -vector space consisting of the linear maps α : N → N such
that α(M ) ≤ M . Evaluate dim(E).

4: 20 marks. Let F be a field, n a positive integer, V an F -vector space with dimension

n = dim(V ). Let θ : V → V be a linear map such that θ2 = idV .
(a) Suppose 1F + 1F ̸= 0F . Let E1 and E−1 denote the 1-eigenspace and the −1-eigenspace
of θ, respectively. Show that dim(E1 ) + dim(E−1 ) = n.
(b) Now suppose that 1F + 1F = 0F . Let E1 denote the 1-eigenspace of θ. Show that
dim(E1 ) ≥ n/2.

19
 Solutions to Midterm 1
 
1 0 0
0 1 0
 
1: We have G = 0 0 1.

1 0 1
1 1 1
Part (b). Respectively, 000, 001, 011 have encodings 00000, 00111, 01110.
Part (c). The decoding table is as follows.
000 001 010 011 100 101 110 111 syndrome
00000 00111 01001 01110 10011 10100 11010 11101 00
00001 00110 01000 01111 10010 10101 11011 11100 01
00010 00101 01011 01100 10001 10110 11000 11111 10
00100 00011 01101 01010 10111 10000 11110 11001 11
Part (d). Respectively, 10000, 11000, 11001 have syndromes 11, 10, 11 and decodings 101,
110, 111.
Part (e). The received word with syndrome 00 and decoding 000 must be 00000. The
received word with syndrome 01 and decoding 000 must be 00001 or 01000. The received word
with syndrome 01 and decoding 000 must be 00010. The recevied word with syndrome 11 and
decoding 000 must be 00100 or 10000. So the number of possible decoding tables is 1.2.1.2 = 4.
2: Part (a). The characteristic polynomial of A in C[t] is
t −1 0
2 t − 2 −1 = t((t − 1)(t − 4) + 2) + 2(t − 4)
0 2 t−4

= t(t2 − 6t + 10) + 2t − 8 = t3 − 6t2 + 12t − 8 = (t − 2)3 .

So the unique eigenvalue is 2.
Part (b). Let f1 be an eigenvector of A. Say f1 = (x, y, z). Then
y = 2x , −2x + 2y + z = 2y , −2y + 4z = 2z .
We have z = y. Putting x = 1, then f1 = (1, 2, 2). The argument shows that every eigenvector
of A is a scalar multiple of f1 . So there is a Jordan basis {f1 , f2 , f3 } of A such that Af2 = 2f2 +f1
and Af3 = 2f3 + f2 .
Writing f2 = (x, y, z), then
y = 2x + 1 , −2x + 2y + z = 2y + 2 , −2y + 4z = 2z + 2 .
Hence 2x = y − 1 and z = y + 1. Putting y = 1, then f2 = (0, 1, 2).
Now writing f3 = (x, y, z), then
y = 2x , −2x + 2y + z = 2y + 1 , −2y + 4z = 2z + 2 .
Again, z = y + 1. Putting z = 1, then f3 = (0, 0, 1). From the coordinates of the vectors fi ,
we arrive at    
1 0 0 2 1 0
P = 2 1 0 , E = 0 2 1 .
2 2 1 0 0 2

20
 Part (c). In view of the Jordan normal form of A, the minimal polynomial of A is (t − 2)3 .
3: Extending a basis {e1 , ..., em } for M to a basis E = {e1 , ..., en } for N , then E is the
vector space consisting of those linear maps α : N → N such that, with respect to E, the
matrix representing α has (i, j)-entry 0 whenever i > m ≥ j. The number of other entries is
dim(E) = n2 + m2 − nm.
4: Part (a). By the Cayley–Hamilton Theorem, the only possible eigenvalues of θ are 1 and
−1. When we extend to a larger field containing F , the nullities of θ − idV and θ + idV do
not change, in other words, dim(E1 ) and dim(E−1 ) do not change. So we may assume that F
is algebraically closed. Let T be the matrix representing θ with respect to a Jordan basis for
θ. We are to show that T is diagonal. Supposing otherwise, we may assume that the top-left
Jordan block of T is not a 1 × 1 matrix. Then the (1, 2) entry of T 2 must be 2 or −2, which
is impossible because T 2 is the identity matrix.
Part (b). Arguing as before, we may assume that F is algebraically closed. Again let T
be a matrix representing θ with respect to a Jordan basis for θ. It suffices to show that every
Jordan block of T is a 1×1 matrix or a 2×2 matrix. Supposing otherwise, we may assume that
the top-left Jordan block of T is a square matrix with at least 3 columns. Then the (1, 3)-entry
of T 2 is 1, which is again impossible because T 2 is the identity matrix.
Alternative, faster version of the solution to 4: The following was used in the exam script of
Buruk Yavus. ∏ Recall that, when λi runs over the eigenvalues of a linear map and the minimal
polynomial is i (t − λi )mi , the largest Jordan block with eigenvalue λi is an mi × mi matrix.
We have θ2 − id = 0. So the minimal polynomial of θ must divide t2 − 1. In part (a), we have
t2 − 1 = (t − 1)(t + 1). So the minimal polynomial must be t − 1 or t + 1 or (t − 1)(t + 1). It
follows that all the Jordan blocks must be 1 × 1 matrices, in other words, θ is diagonalizable.
In part (b), we have t2 − 1 = (t − 1)2 . So now the minimal polynomial must be t − 1 or (t − 1)2 .
Hence all the Jordan blocks must be 1 × 1 or 2 × 2 matrices.

21
MATH 224: Linear Algebra 2. Spring 2022. Midterm 2
LJB, 21 April 2022, Bilkent University.
Time allowed: 110 minutes hours. Please put your name on EVERY sheet of your manuscript.
The use of telephones, calculators or other electronic devices is prohibited. The use of red pens
or very faint pencils is prohibited too. You may take the question sheet home.

1: (30 marks). Let e1 = (0, 1, 1) and e2 = (1, 0, 1) and e3 = (1, 1, 0). Let {f1 , f2 , f3 } be the
basis for R3 obtained from {e1 , e2 , e3 } by the Gram–Schmidt process. Evaluate f1 and f2 and
f3 .

2: (30 marks). Let U = spanR {(1, 2, 3)} as a subspace of R3 . Let π : U ← R3 and

π ′ : U ⊥ ← R3 be the orthogonal projections. Express, with respect to the standard basis of
R3 , the matrices representing π and π ′ .

3: (12 marks). Let n be a positive integer. Determine the sets of:

(a) complex numbers z such that z is the determinant of a Hermitian operator on Cn ,
(b) complex numbers z such that z is the trace of a Hermitian operator on Cn ,
(c) complex numbers z such that z is the determinant of a unitary operator on Cn ,
(d) complex numbers z such that z is the trace of a unitary operator on Cn .

4: Let n be a positive integer.

(a) (2 marks) Let {f1 , ..., fn } be an orthonormal
∑ basis for Rn and let x ∈ Rn . Let xi , for
1 ≤ i ≤ n, be real numbers such that x = i xi fi . State a formula for xi in terms of x and fi .
(b) (26 marks) Let f : Rn ← Rn be a distance-preserving function, in other words,

∥ f (x) − f (y) ∥ = ∥ x − y ∥

for all x, y ∈ Rn . Show that there exist a ∈ Rn and an orthogonal n × n matrix A such that
f (x) = Ax + a for all x ∈ Rn . (Warning: be careful not to make an invalid assumption of
linearity.)

22
 Solutions to Midterm 2

1: We put f1′ = e1 = (0, 1, 1) and

⟨f1′ | e2 ⟩ ′
f2′ = e2 − f .
∥ f1′ ∥2 1
Now ⟨f1′ | e2 ⟩ = 1 and ∥ f1′ ∥2 = 2, so f2′ = (1, 0, 1) − (0, 1, 1)/2 = (1, −1/2, 1/2). We have
⟨f1′ | e3 ⟩ ′ ⟨f2′ | e3 ⟩ ′
f3′ = e3 − f − f .
∥ f1′ ∥2 1 ∥ f2′ ∥2 2
Now ⟨f1′ | e3 ⟩ = 1, also ⟨f2′ | e3 ⟩ = 1/2 and ∥ f2′ ∥2 = 3/2. So
f3′ = (1, 0, 0) − (0, 1, 1)/2 − (1, −1/2, 1/2)/3
( )
= (6, 0, 0) − (0, 3, 3) − (2, −1, 1) = (2, 2, −2)/3 .
Up to scaling, the obtained orthogonal basis√is {(0, 1, 1), (2,
√−1, 1), (1, 1, −1)}.
√ Normalizing,
the associated orthonormal basis is {(0, 1, 1)/ 2, (2, −1, 1)/ 6, (1, 1, −1)/ 3}.
Comment: In the phrasing of the question, I omitted to specify whether the basis should be
orthonormal or just orthogonal. I deemed any basis consisting of multiples of (0, 1, 1) and
(2, −1, −1) and (1, 1, −1) to be correct.
(1, 2, 3)(x, y, z) x + 2y + 3z
2: We have π(x, y, z) = (1, 2, 3) = (1, 2, 3)
∥ (1, 2, 3) ∥2 1+4+9
= (x + 2y + 3z, 2x + 4y + 6z, 3x + 6y + 9z)/14 .
   
1 2 3 13 −2 −3
1  1 
So π and π are represented by 2 4 6 and −2 10 −6, respectively.
14 14
3 6 9 −3 −6 5
3: The eigenvalues of a Hermitian operator on Rn are arbitrary real numbers, so the answer to
(a) is R and the answer to (b) is R. The eigenvalues of a unitary operator on Cn are arbitrary
complex numbers with modulus unity. So the answer to (c) is {z ∈ C : |z| = 1} and the answer
to (d) is {z ∈ C : |z| ≤ n}.
4: Part (a). Obviously, xi = ⟨fi | x⟩.
Part (b). Define g : Rn ← Rn such that g(x) = f (x) − f (0) for x ∈ Rn . We have
∥ g(x) − g(y) ∥=∥ f (x) − f (y) ∥=∥ x − y ∥. Replacing f with g, we may assume that f (0) = 0,
and we are to show that f (x) = Ax for some orthogonal matrix A.
We have ∥ f (x) ∥=∥ f (x) − f (0) ∥=∥ x − 0 ∥=∥ x ∥. Thus, f preserves the norm. We have

∥ f (x) ∥2 + ∥ f (y) ∥2 −2⟨f (x) | f (y)⟩ =∥ f (x) − f (y) ∥2 =∥ x − y ∥2 =∥ x ∥2 + ∥ y ∥2 −2⟨x | y⟩ .

But ∥ f (x) ∥ = ∥ x ∥ and similarly for y, so ⟨f (x) | f (y)⟩ = ⟨x | y⟩. Thus, f preserves the inner
product.
Let {e1 , ..., en } be the standard basis for Rn . Let fi = f (ei ). Since
∑ f preserves the inner
product, {f1 , ..., fn } is an orthonormal basis for Rn . Write x = i ei . Again using the
∑ of inner products, and also applying part (a), xi = ⟨ei | x⟩ = ⟨fi | f (x)⟩, hence
preservation
f (x) = i xi fi . It is now clear that f is an operator and that, in fact, f is an orthogonal
operator.

23
MATH 224: Linear Algebra 2. Spring 2022. Makeup 2
LJB, 12 May 2022, Bilkent University.
Time allowed: 110 minutes hours. Please put your name on EVERY sheet of your manuscript.
The use of telephones, calculators or other electronic devices is prohibited. The use of red pens
or very faint pencils is prohibited too. You may take the question sheet home.

1: (30 marks). Let e1 = (0, 1, 2) and e2 = (1, 2, 1) and e3 = (2, 1, 0). Let {f1 , f2 , f3 } be the
basis for R3 obtained from {e1 , e2 , e3 } by the Gram–Schmidt process. Evaluate f1 and f2 and
f3 .

2: (30 marks). Let

U = {(a, b, c, d) ∈ R4 : a + b + c + d = b + 2c + 3d = 0} .

Express, with respect to the standard basis of R3 , the matrix representing the orthogonal
projection U ← R4 .

3: (20 marks). Let m and n be positive integers with m ≤ n. Let U be an m-dimensional

subspace of an n-dimensional inner product space V over C. Let E be the set of Hermitian
operators V ← V that restrict to operators U ← U . Regarding E as a vector space over R in
the obvious way, evaluate the dimension of E.

4: (20 marks). Let θ be a unitary operator on Cn .

(a) Under what conditions on θ do there exist only finitely many unitary operators ϕ on Cn
such that θ = ϕ2 ?
(b) Under those conditions, how many such ϕ exist?

24
MATH 224: Linear Algebra 2. Spring 2022. Final
LJB, 20 May 2022, Bilkent University.
Time allowed: 2 hours. Please put your name on EVERY sheet of your manuscript. The use
of telephones, calculators or other electronic devices is prohibited. The use of red pens or very
faint pencils is prohibited too. You may take the question sheet home.
[ ]
1 2
1: (20 marks). Let A = .
2 1
(a) Find the eigenvalues and eigenvectors of A.
(b) Find the rank and signature of the real quadratic form represented by A.

2: (40 marks). Let n be an integer. When A2 = B for n × n matrices A and B, we call A a

square root of B.
 
2 1 1 1
 1 2 1 1
(a) Find the multiplicity of 1 as an eigenvalue of H =  
 1 1 2 1 .
1 1 1 2
(b) Find all the other eigenvalues of H. (Hint: to avoid complicated calculations, consider the
trace of H.)
(c) Show that H has a square root in Mat4 (R). (Hint: to avoid complicated calculations, use
parts (a) and (b).)
(d) Show that every symmetric matrix in Matn (R) has a square root in Matn (C).
(e) Give an example of an integer n and a symmetric matrix B in Matn (R) such that B has
no square root in Matn (R).
(f ) Give an example of an integer n and a matrix C in Matn (R) such that C has no square
root in Matn (C).

3: (20 marks). For a positive integer n,

(a) How many equivalence classes of symmetric bilinear forms Rn × Rn → R are there?
(b) For how many of those equivalence classes are the bilinear forms non-degenerate?

4: (20 marks). Let F denote the field with order |F| = 4, let n be a positive integer, and let
Fn denote the standard F-vector space of dimension n.
(a) What is the number of n-tuples (f1 , ..., fn ) such that the set {f1 , ..., fn } is a basis for Fn ?
(b) How many linear maps Fn → Fn are there?
(c) How many invertible linear maps Fn → Fn are there?
(d) What is the number of linear maps α : Fn → Fn such that det(α) = 1?

25
 Solutions to Final

1: Part (a). The characteristic polynomial of A is

1−λ 2
= (λ − 1)2 − 4 = λ2 − 2λ + 3 = (λ − 3)(λ + 1) .
2 1−λ

Eigenvectors with eigenvalue 3 and −1 are (1, 1) and (1, −1).

Part (b). Let Q be the quadratic form represented by A with respect to the standard
basis of R2 . With respect to the basis {(1, 1), (1, −1)}, the matrix representing Q has diagonal
entries 3 and −1. Normalizing the basis, then the matrix representing Q has diagonal entries
1 and −1. So the rank of Q is 2 and the signature of Q is 0.
2: Part (a). The matrix H − 1 has rank 1 and nullity 3. So the multiplicity of the eigenvalue
1 is 3.
Part (b). By part (a), H has one other eigenvalue λ, with multiplicity 1. The trace of H
is 8 = 3 + λ, so λ = 5.
Part (c). Since H is symmetric, H is diagonalizable. By parts (a) and (b), there exists
invertible P ∈ Mat4 (R) such that H = P diag(1, 1, 1, 5)P −1 , where diag(a1 ,√ a2 , ...) denotes the
diagonal matrix with (i, i) entry ai . One square root of H is P diag(1, 1, 1, 5)P −1 .
Part (d). Let S be a symmetric matrix in Matn (R). Then S is diagonalizable over R, say
S = Q diag(d1 , ..., dn )Q−1 . Letting ci be a square root of di in C, then Q diag(c1 , ..., cn )Q−1 is
a square root of S in Matn (C).
Part (e). Put n = [1 and]let B be the 1×1 matrix with entry −1.
0 1
Part (f). Let C = . For a contradiction, suppose C has a square root A in Mat2 (C).
0 0
Then A4 = 0. By considering the minimal polynomial of A, we find that 0 is the unique
eigenvalue of A. By considering the Jordan normal form of A, we deduce that A2 = 0 ̸= C,
which is a contradiction, as required.
3: Part (a). Sylvester’s Law of Inertia tells us that each equivalence class is represented by
a unique diagonal matrix whose first a entries are 1, whose next b entries are −1 and whose
next c entries are 0, where a, b, c are natural numbers such that a + b + c = n. Putting
m = a + b = n − c, then 0 ≤ m ≤ n and, for each m, the number of possibilities for (a, b, c) is
m + 1. So the number of equivalence classes is

∑
n
(m + 1) = 1 + 2 + ... + n + (n + 1) = (n + 1)(n + 2)/2 .
m=0

Part (b). The non-degenerate cases are those where c = 0. The number of such cases is
n + 1.
4: Part (a). There are 4n − 1 choices for f1 , then 4n − 4 choices for f2 and so on, finally
4n − 4n−1 choices for fn . So the number of such n-tuples is

(4n − 1)(4n − 4)...(4n − 4n−1 ) = 4n(n−1)/2 (4n − 1)(4n−1 − 1)...(4 − 1) .

Part (b). Since Matn (F) has dimension n2 as a vector space over F, the specified number
2
is |Matn (F)| = 4n .

26
 Part (c). Fixing an ordered basis {e1 , ..., en } for Fn , each Each invertible map α on Fn is
uniquely determined by the ordered basis {α(e1 ), ..., α(en )}. So the number of such maps is
4n(n−1)/2 (4n − 1)(4n−1 − 1)...(4 − 1).
Part (d). Given a nonzero element λ of F and A ∈ Matn (F), then multiplying the top row
of A by λ produces a matrix with determinant λ det(A). So the number of invertible n × n
matrices with determinant λ is independent of λ. There are exactly 3 non-zero elements of F.
So the number of α as specified is the answer to (c) divided by 3. In other words, the number
of such α is 4n(n−1)/2 (4n − 1)(4n−1 − 1)...(42 − 1).

27