OLD SOLUTIONS

JAS SINGH, SAM QUNELL, JACOB SWENBERG

Feel free to add to, revise, or comment on this document as you see fit. Solutions will be (hopefully)
correct, but for sake of time, will probably not be typed up the detail required for the actual exam.
Once more problems are added here, I will put some thoughts about common problem types that
one should be prepared to deal with (i.e. "show this is an inner product", "use Arzela Ascoli but not
really", "know how to compute exp(A)" , "iterate a function and show it converges", "do infinite
dimensional vector spaces behave well? (no)")

1. S17

Problem 1 (1). Let M be an n × n real matrix with transpose M T . Prove that M and
M M T have the same image.

Via the Friedholm alternative, we have im(M ) = ker(M T )⊥ and im(M M T ) = ker((M M T )T )⊥ =
ker(M M T )⊥ . Since V = ker(M T )⊕ker(M T )⊥ , and likewise for ker(M M T ), it is sufficient to prove
that ker(M T ) = ker(M M T ). Fix v ∈ ker(M T ). Then M M T ∗ v − M ∗ 0 = 0, so v ∈ ker(M M T ).
Now, if v ∈ ker(M M T ), then M M T v = 0, so hM T v, M T vi = hv, M M T vi = hv, 0i = 0, and by
property of inner products, this implies that M T v = 0, and so v ∈ ker(M T ). Thus, ker(M T ) =
ker(M M T ), and so im(M ) = im(M M T ).
 
1 0 a b
0 1 0 0 
Problem 2 (2). Let a, b, c, d ∈ R and M  0 c 3 −2.


0 d 2 −1
(1) Determine conditions on a, b, c, d s.t. there is only one Jordan block for each eigenvalue
of M in the Jordan form of M .
(2) Find the Jordan form of M when a = c = d = 2 and b = −2.

We first compute the characteristic polynomial of M . We see this is (x − 1)4 . Therefore, the Jordan
canonical form of M has only ones on the diagonal,
 and so has
a single Jordan block for eigenvalue
0 0 a b
0 0 0 0 
1 iff nullity(M − I) = 1. We compute M −  0 c 2 −2. This needs to have exactly three


0 d 2 −2
linearly independent rows. We require c 6= d and a 6= −b.

If a = c = d = 2 and b = −2, nullity(M − I) = 2, so there are exactly 2 blocks in the Jor-

dan form. The size of the largest block corresponding to eigenvalue 1 is equal to the multiplicity of
x − 1 in the minimal polynomial of M . We compute that (X − 1)2 = 0, and since (X − 1) 6= 0, we
have that there are 2 blocks of size 2 in the J.C.F.
1
 Problem 3 (3). Let M be an n × n real matrix. Suppose M is orthogonal and symmetric
(1) Prove that if M is positive definite, then M is the identity.
(2) Does the answer change if M is only positive semidefinite?

M being real and symmetric implies by spectral theorem that it is diagonalizable with real eigen-
values. M being positive definite implies that each eigenvalue is strictly greater than zero. M being
orthogonal implies that the eigenvalues of M must have magnitude 1. This means that the only
eigenvalue of M is 1, and by diagonalizability, M = 1.

The answer does not change if M is only positive semidefinite. The condition that M is orthogonal
prohibits that M has a zero eigenvalue, so it is still true that M can only have eigenvalue 1.

Problem 4 (4). Let F (x) be the determinant of the given matrix (not retyping). Compute
dF/dx(0).

Note that F (x) is a polynomial in x, and therefore, if F (x) = ni=0 ai xi , dF/dx(0) = a1 i.e. the
P
P Q5
coefficient of x. Note also Det(M ) = σ∈S5 i=1 sign(σ) ∗ Mi,σ(i) . We will find all σ such that
Q5
i=1 Mi,σ(i) = x. For any given σ, exactly one of the Mi,σ(i) can be x, and the rest must be 1. One
can check that there is only one such σ, (not sure how to write this out concisely), where we take
a2,1 ∗ a3,2 ∗ a4,3 ∗ a1,4 ∗ a5,5 = x. The sign of this permutation is -1, so our derivative is −1.

Problem 5 (5). Suppose V is a finite dimensional vector space over C and T : V → V is a

linear transformation. Let F (X) ∈ C[X] be a polynomial. Show that F (T ) is an invertible
transformation iff F (X) and the minimal polynomial of T have no common factors.

If the minimal polynomial of T and F (X) indeed have a common factor, say (x − λ), then let v
be an eigenvector of T corresponding to λ. F (X) = q(X) ∗ (X − λ) for some polynomial q(X), so
F (T )v = q(T )((T − λ)v) = 0, so F (T ) is not invertible.

Qn that F (T
Now, suppose
n
) has no common factor with the minimal polynomial of T , and write
F (X) = i=1 (x − λi ) where the λi are distinct. Suppose
i
QF (T )v = 0 for some vector v (not
necessarily nonzero). In particular, (T − λ1 I)o(T − λ1 I)n1 −1 ni=2 (T − λi )ni v) = 0. Since λ1 is not
an eigenvalue of T (by assumption on no common factors), that means that the inner operator on
v must return zero. We repeat this argument on the polynomial of T of smaller degree, until we
eventually obtain that v = 0.

Problem 6 (6). (1) Let V denote the vector space of real n × n matrices. Prove that
hA, Bi − trace(AT B) is an inner product on V .
(2) For n = 2, find an orthonormal basis of V .

Note that for any matrix A, tr(A) = tr(AT ). Then hA, Bi = tr(AT B) = tr((AT B)T ) = tr(B T A) =
hB, Ai, and so we have symmetry.

For any matrix A, hA, PAi = tr(AT A). We P see that the element in the iProw and i column of
A A is by definition j=1 (AT )i,j Aj, i = nj=1 A2j,i ≥ 0. Then tr(AT A) = i,j A2i,j ≥ 0. We also
T n

see that this equals zero iff each Ai,j = 0 i.e. A = 0.

2
Fix matrices A, B, C and λ ∈ C. Then hA + λC, Bi = tr((A + λC)T B) = tr((AT + λC T )B) =
tr(AT B) + λ(C T B) = λhA, Bi + hC, Bi, and so we have linearity.

I will show that the standard basis of V is in fact orthonormal, i.e., bi,j is 1 in the i, j position
and zero elsewhere. One checks quickly that tr((bTp,q ∗ bi,j ) = tr(bq,p bi,j ). From here, one sees
that bi,j bj,i is always diagonal with one in one place and zero in the other, so each of these basis
elements is unit length. Furthermore, one also can confirm that these basis elements are pairwise
orthonormal.

Problem 7 (7). Prove existence and R x uniqueness of a non-negative continuous function f :

[0, 1] → [0, 1] satisfying f (x) = 1 − [ 0 tf (t)dt]2 .

Note that the set of continuous functionsR from [0, 1] to [0, 1] is complete with the supnorm metric.
x
Hence, if I can show that g(f (x)) = 1 − [ 0 tf (t)dt]2 is a contraction mapping, then by contraction
mapping principle, g has a unique fixed point, and note that f is a fixed point of g iff the given
condition is true.

R x first that g(f (x)) is always continuous since it is a composition of continuous

Note Rx functions.
Rx Since
[ 0 tf (t)dt]2 ≥ 0, g(f (x)) is always ≤ 1. Furthermore, since f (t) ≤ 1 always, 0 tf (t) ≤ 0 t = x2 /2.
For x ∈ [0, 1], 1 − x2 /2 ∈ [0, 1], and so g is indeed a mapping from this space to itself.
Rx 2
Rx 2
Rx
R x p(x), q(x) in this space. Then g(p) − g(q) = [ 0 tq(t)dt] − [ 0 tp(t)dt] = ( 0 t(q − p)dt) ∗
Fix
( 0 t(q + p)dt. If d = supx∈[0,1] |p(x) − q(x)|, the absolute value of the first term of this product is
Rx
less than 0 t ∗ d = x2 d/2. Since both f and g take values in [0, R x1], we see |f + g| ≤ 2, and so the
absolute value of the second term in this product is less than 0 t ∗ 2dt = x2 . We now have, for
any x, |p(x) − q(x)| ≤ x4 d/2, and since x ∈ [0, 1], this is always less than or equal to d/2. Then
supx∈[0,1] |p(x) − q(x)| ≤ d/2 as well, and so this is indeed a contraction mapping.

R1
Problem 8 (8). Show that there is a constant C so that | f (0)+f
2
(1)
− 0 f (x)dx| ≤
R 1 00 2
C 0 |f (x)|dx for every C function f : R → R.

R1
Consider the integral 0 f (t). We may integrate by parts taking u = f and v = x − 1/2 since f is
R1 R1
differentiable. Then 0 f (x)dx = (f (x) ∗ (x − 1/2)|10 − 0 (x − 1/2)f 0 (x)dx. The first term evaluates
R1 R1
to f (0)+f
2
(1)
, and so | f (0)+f
2
(1)
− 0 f (x)dx| ≤ | 0 f 0 (x) ∗ (x − 1/2)|. We may integrate by parts
on this right integral, with u = f 0 (x) and v = (x − 1/2)2 /2 since f 0 is differentiable. We compute
R1 0 0 2 1
R 1 00 2 f 0 (1)−f 0 (0)
0 fR(x)(x − 1/2)dx = f (x)(x − 1/2) /2|0 − 0 f (x)(x − 1/2) /2dx. The first term is 8 =
1
1/8 0 f 00 (x)dx by fundamental theorem of calculus. The second term, since (x − 1/2)2 ≤ 1/4 for
x ∈ 0, 1, we see that we may take C = 1/8 + 1/8 = 1/4.

Problem 9 (9). Let (X, d) be a bounded metric space and let C(X) be the space of bounded
continuous real functions on X endowed with the supremum norm. Suppose C(X) is sepa-
rable.
(1) Show that for every ε > 0, there is a countable set Zε ⊂ X so that for all x ∈ X,
there exists a z ∈ Zε where for any y in X, |d(x, y) − d(z, y)| < ε.
(2) Show that X is separable.

3
For part a, observe that for any x ∈ X, the function fx = d(x, y) is a bounded (since X is bounded)
continuous function of y. Since X is separable, let S be a countable set of functions in C(X) where
for any f ∈ C(X), there exists a g ∈ S such that sup|f (x) − g(x)| < ε/2. Note that since this is true
if we take f to be fx , each fx is within ε/2 of some g ∈ S. Let S2 be the subset of S of functions
containing at least one fx within the ε/2 ball around it, and note that this set is also ε/2 dense with
the fx . For each g ∈ S2 , pick some fy , and let the set of these be S3 . I will show that we may take
Zε = {y|fy ∈ S3 } works. Note first that this set is countable since S3 is countable. For any fx , fx
has supnorm distance ε/2 or less from some S2 member g. Let fy be the S3 member corresponding
to g. Then by triangle inequality, |fx (z) − fy (z)| < ε/2 + ε/2 = ε for all z, as desired.

Given that part a is true, note that if x, z ∈ X, then |d(x, y) − d(z, y)| ≤ d(x, z) by reverse triangle
inequality, and since |d(x, z) − d(z, z)| = d(x, z), we see that supy∈X |d(x, y) − d(z, y)| = d(x, z). For
any positive natural number n, part a gives that there exists a countable set Z1/n where for any
x ∈ X, there is a z ∈ Z1/n where d(x, z) ≤ ε. Take the union of the Z1/n , say Z, and note that since
this is a countable union of countable sets, it is countable. Now fix ε > 0 and x ∈ X. Then pick
1/n < ε. There exists z ∈ Z1/n , and hence z ∈ Z, such that d(x, z) ≤ ε. Thus, Z is a countable
dense subset of X, and so X is separable.

Problem 10 (10). Let K ⊂ Rn be compact. Suppose that for every ε > 0 and every pair
a, b ∈ K there is an integer n ≥ 1 and a sequence of points x0 ...xn ∈ K so that x0 = 0,
xn = b, and ||xk − xk−1 || < ε for every 1 ≤ k ≤ n
(1) Show K is connected
(2) Show by example that K may not be path connected.

Suppose towards contradiction that K is the disjoint union of two nonempty sets that are both
closed relative to K, call these A and B. Fix a point a ∈ A and b ∈ B. For every 1/m with m a
positive natural number, there exists a sequence with x0 = a, xn = b, and ||xk − xk−1 || < 1/m. But
this implies that there exists an am ∈ A and bm ∈ B with ||am − bm || < ε (there must be at least
one set change in the sequence since x0 ∈ A and xn ∈ B). Since K is compact, some subsequence
of the am and bm , say the amk and bmk must converge. Since A and B are relatively closed in K,
the amk converge to some a∗ ∈ A and the bmk converge to some b∗ in B. But the condition on
the am − bm implies that a∗ − b∗ = 0, and so a∗ = b∗ . This contradicts that A and B are disjoint.
Therefore, K must be connected.

To show K is not necessarily path connected, consider the set of points in R2 given by {(x, sin(1/x)|x ∈
(0, 1]} ∪ {(0, 0)} i.e. the topologist’s sine curve. It is known that this is not path connected. For
any a, b ∈ K, if the x component of both a and b is nonzero, then both are in the path connected
component of K, and so the given condition is immediately satisfied. If WLOG a = (0, 0) and
b = (x, sin(1/x)) for some x ∈ (0, 1], there will always be some zero of sin(1/x) with x < ε since
sin(1/x) has infinitely many periods approaching zero. Let such a zero be x1 . Then x1 is in the
path connected component of K and so we may find a chain connecting it to b via the argument
above.

Problem 11 (11). Let fn : [0, 1] → R be Ra sequence of continuous functions satisfying

x
|fn (x)| ≤ 1 + 1+nn2 x2 , and define Fn (x) = 0 fn (t)dt. Show that there is a subsequence
nk → ∞ so that Fnk converges for every x ∈ [0, 1].

4
 Rx Rx
We first show that the Fn are equibounded. Note that |Fn (x)| ≤ 0 |fn (t)| ≤ 0 1 + 1+nn2 t2 =
x + arctan(nx) ≤ 1 + π/2 for any x and n.
I will now show that the Fn are equicontinuous on [1/2, 1]. By Arzela Ascoli, this implies that
some subsequence,
Ry the Fnk converges for all x ∈ [1/2, 1]. If 1/2 ≤ x < y ≤ 1, then |Fn (y) −
Fn (x)| ≤ x 1 + 1+nn2 t2 = (y − x) + arctan(ny) − arctan(nx). arctan is a continuous function
on [1/2, 1], and hence, is uniformly continuous. The sum (y − x) + arctan(ny) − arctan(nx) will
also be uniformly continuous on [1/2, 1], and so there exists δ > 0 s.t. |x − y| < δ implies that
(y − x) + arctan(ny) − arctan(nx) < ε. This shows that |Fn (y) − Fn (x)| < ε for such x, y, so the
Fn are equicontinuous on [1/2, 1]. By Arzela Ascoli, some subsequence, say the F2,i converges on
[1/2, 1].
Now suppose that some subsequence of our original Fn , say the Fj,i converges on [1/j, 1]. Then
[1/(j + 1)/1] is also a compact interval, and so this subsequence is also equicontinuous here. There-
fore, some subsequence of the Fj,i , say the Fj+1,i , converges on [1/(j + 1), 1].

We have shown that we can iteratively define subsequences that converge on [1/j, 1] for any j > 1.
To complete the proof, define our sequence to be F2,1 , F3,2 , F4,3 .... Note that this subsequence con-
verges at zero since for any n, Fn (0) = 0. Now fix any x ∈ (0, 1]. There exists some n > 1 natural
number where 1/n < x. Then for all i > n, the Fi,i−1 were elements of the sequence Fn,j , and so
converge on [1/n, 1], and in particular, at x.

Problem 12 (12). Show that for each t ∈ R fixed, that the function F (y, t) = y 4 + ty 2 + t2 y
defined for all y ∈ R acheives its global minimum at a single point y0 (t)

For fixed t, define f (y) = F (y, t), and note that this is continuous. Observe that since this function
f , as a polynomial of y, is quartic, limy→∞ F (y, t) = limy→−∞ F (y, t) = ∞. Therefore, there exists
a positive M such that |y| > M implies that F (y, t) > 10. Consider f where y is restricted to the
bounded set [−M, M ] and t is fixed. f is a continuous function of y on a compact set, and so by
extreme value theorem, it acheives a local minimum for some y ∈ [−M, M ]. Furthermore, since
F (0, t) = 0, we see that the minimum acheived on this set will be less than any value F takes for
|y| > M . Equivalently, some local minimum of f for y ∈ [−M, M ] is a global minimum of F (y, t). It
remains to show that there is exactly one local minimum of f . We compute f 0 (y) = 4y 3 + 2yt + t2 ,
and f 00 (y) = 12y 2 +2t. If t = 0, then f (y) = y 4 , and so there is exactly one global minimum at y = 0.
If t > 0, then f 00 (y) > 0, so f cannot obtain a minimum at more than one point. We note that
f 0 (y) must have at least one real root since it is cubic, so there is indeed exactly one local minimum,
which is therefore a global minimum. Now suppose t < 0, and suppose towards contradiction that
y1 and y2 are both global minima. Then the polynomial g(y) = F (y, t) − F (y1 , t) has roots at
both y2 and y1 . Neither of these can be a simple root since g(y) ≥ 0 always, and so (y1 − y)2 and
(y2 − y)2 both divide g. But since g is monic and degree 4, we see g = (y1 − y)2 (y2 − y)2 . Then
F (y, t) = (y1 − y)2 (y2 − y)2 + F (y1 , t). Equating the coefficients of y 3 , we must have 2y1 + 2y2 = 0,
and so y1 = −y2 . Equating the coefficients of y, we see 2y12 y2 + 2y22 y1 = t2 , but the left hand side is
also zero. This is a contradiction since t < 0. Thus, f obtains a unique local minimum in [−M, M ],
which is a global minimum.

2. F17

Problem 13 (1). Let V = {f (X) = a0 + a1 X + a2 X 2 + a3 X 3 |a1 , ...a3 ∈ C} be the complex

vector space of polynomials in the variable X of degree at most 3.

5
 R1
(1) Show that V is an inner product space with hf, gi = −1 f (t)g(t)dt
(2) Find an orthonormal basis of V
R1
First, fix f ∈ V . Then hf, f i = −1 f (t)|f (t)|dt. The integrand is nonnegative and is continuous.
Therefore, this integral is always nonnegative, and is zero iff f f = 0 everywhere on [−1, 1] iff f = 0.
R1 R1 R1
For symmetry, fix g ∈ V as well. Then hf, gi = −1 f g = −1 f g = −1 gf = hg, f i. Finally, fix
R1 R1 R1
h ∈ V and a ∈ C. Then haf + h, gi = −1 (af + h)g = a −1 f g + −1 hg = ahf, gi + hh, gi, all by
linearity of the integral. Thus, this defines an inner product.

I won’t type out the full Gram Schmidt calculation. Start with a basis for Rthis space, i.e. 1, t, t2 , t3 .
1
We first normalize basis vector 1 with respect to this product. h1, 1i = −1 1 = 2. So if we take
√ √ √
e1 = 1/ 2, this vector will be unit length. For our next vector, take b2 = t − ht, 1/ 2i ∗ 1/ 2. We
note that the inner product in this expression is aR symmetric integral of an odd
p function, and so
1 2
is zero. So b2 = t. To normalize, we see ht, ti = −1 t = 2/3, so take e2 = 3/2t, and this will
√
be unit length and orthogonal to 1/ 2. The remaining vectors can be computed in the same way,
and calculation is made simpler by using the odd and even functions.
p These polynomials are the
Legendre polynomials, and when normalized, the next one is 45/8(t2 − 1/3).

Problem 14 (2). Let n ≥ 1 be an integer, and A, B be n × n matrices.

(1) Show that if A is invertible, AB and BA have the same characteristic polynomial.
(2) Prove or disprove: the same is true even if A is not invertible.

By definition, the characteristic polynomial of AB is det(AB −xI). Since det(A) = det(A−1 )−1 , this
is det(A−1 )det(AB − xI)det(A). By multiplicativity of determinant, this is det(A−1 (ABx I)A) =
det(BA − xI), which is exactly the characteristic polynomial of BA. (Note, we have shown that
similar matrices have the same characteristic polynomial.)

This is true even if A is not invertible. First, we will show that the invertible matrices are dense
in Mn (C). Consider any ε > 0, and the matrix A − εI. This matrix is non-invertible iff. ε is an
eigenvalue of A. A has finitely many eigenvalues, so if ε > 0 and is less than the smallest positive
eigenvalue of A, then A−εI will always be invertible. Therefore, |A−(A−εI)| = |εI| = ε, and so in-
deed, A is a limit of invertible matrices. Let An be a sequence of invertible matrices converging to A,
and let B be fixed. Consider the function f : Mn (C) → C[x], f (M ) = det(M B−xI)−det(BM −xI).
This is continuous since it is a composition of continuous functions (in particular, determinant is
continuous). For each An , An is invertible, and we have shown f (An ) = 0. Then since An have
limit A and since f is continuous, f (A) = 0 as well, so det(AB − xI) = det(BA − xI), and this is
exactly what we wanted.

Problem 15 (3). Solve the following linear system of differential equations (not going to
retype)
 
6 −1
If A = , then the solution is exactly x(t) = exp(A) ∗ x(0), where x(0) is the initial
2 3
conditions at t = 0 (I believe this can be stated without proof, since this is relatively common use
of exponentials of matrices. If this seems strange to you, compare to the case where you only have
one variable. x0 = ax → x = eat ∗ x(0)). So, it remains to compute exp(A). It is easiest to compute
6
the Jordan canonical form. The characteristic polynomial of this matrix   (x − 4)(x − 5). Since
is
4 0
this has two distinct roots, the Jordan canonical form is exactly D = . We also compute the
0 5
 
−1 2 −1
matrix P such that P AP = D. To do so, we need only find the eigenvectors. A − 4I= .
2 −1
 
1 −1
Then (A − 4I)v = 0 iff v is a scalar multiple of (1, 2). Likewise, (A − 5I)v = 0 iff v = 0 iff
2 −2
   
1 1 −1 1
v is a scalar multiple of (1, 1). If P = , we see P −1 = . We also see that indeed
2 1 2 −1
P −1 AP = D, and so A = P DP −1 . Then exp(A) = exp(P DP −1 )= P exp(D)P −1 . But exp(D) is
e4 0 2e5 − e4 −e5 + e4
 
exactly 5 . Then P exp(D) ∗ P −1 =
0 e 2e5 − 2e4 2e4 − e5

Problem 16 (4). Let V be a vector space over R and let V ∗ be the dual space of V . Let
B = {ei } be a basis of V . For each i, define e# ∗ #
i ∈ V by ei (ej ) = δi,j .
(1) Show that these e#
i are linearly independent in V .
∗

(2) Give necessary and sufficient conditions (with proof) on V for these vectors to form
a basis of V ∗ .

Linear independence is defined with finite support, so, after reordering, suppose a1 e# #
1 + a2 e 2 +
...an e#
n = 0 i.e. is the zero operator. In particular, for each ei with i ∈ {1, 2...n}, we have
0(ei ) = (a1 e# # #
1 + a2 e2 + ...an en )(ei ) = ai . Thus, all coefficients are zero.

We will show that this is a basis iff V is finite dimensional. First, suppose V is indeed finite di-
mensional. It remains to show that this basis spans V ∗ . Fix an operator L ∈ V ∗ . Write ai = L(ei )
for each of the (finitely many) ei . Then L and a1 e# # #
1 + a2 e2 + ...an en agree on all vectors of V by
linearity, and are thus the same operator. Therefore, L is in the span of B.

If V is infinite dimensional, define L(ei ) = 1 for all basis vectors ei (this is well-defined, since
sums of V elements are only defined on finite support). Let e# # #
1 , e2 , ...en be a finite subset of B that
# #
is reindexed. Then L(en+1 ) = 1, but (a1 e1 + ...an en )(en+1 ) = 0. Thus, L is not a (finite) linear
combination of B, and so is not in the span of B.

Problem 17 (5). Let V and W be two infinite dimensional vecto spaces over C. Let
LinF (V, W ) be the C vector space of C linear maps from V to W
(1) Is X = {f ∈ LinF (V, W )|f has finite rank} a subspace?
(2) What about Y = {f ∈ LinF (V, W )|Ker(f ) has finite dimension}?
(3) What is X ∩ Y ?

(1) Yes. Fix f, g ∈ X and a ∈ C. Then (af + g)v = a(f v) + gv ∈ im(f ) + im(g) ⊂
span(im(f ), im(g)). Since im(f ) and im(g) are finite dimensional, their combined span
is as well. So im(af + g) ⊂ span(im(f ), im(g)), and hence, is finite dimensional. So
af + g ∈ X.
(2) No. Fix f ∈ Y (if Y is a subspace, it must be nonempty). Then −f ∈ Y as well (where
f (v) = w ⇐⇒ −f (v) = −w). This is because f (v) = 0 ⇐⇒ −f (v) = −0 = 0. Then
f + (−f ) = 0, which has kernel infinite dimensional.
7
 One could also simply say that 0 ∈
/ Y implies Y is not a subspace immediately.

Interesting thought: I don’t think it is even guaranteed that Y is nonempty if V has a

continuum basis (like the set ert are linearly independent for smooth functions) and W has
countably infinite basis like R[x], it seems that you could not produce any operator with
finite dimensional kernel, since the image will always be countable dimension.

(3) This is the empty set. If Ker(f ) is finite dimensional, extend a basis of Ker(f ) to a
basis for V (technically requires axiom of choice, but then again, so does the fact that V
has a basis at all). The images of each element that was added in the extension must be
linearly independent, or else we have contradicted that these elements were not in the kernel.
Therefore, the image will be infinite rank.
Basically an infinite version of rank-nullity theorem, although we technically haven’t
proven this for the infinite dimensional case.

Problem 18 (6). For each of F = R, C, F3 , is it true that every symmetric matrix A ∈

M2×2 (F ) is diagonalizable?

We use that if A is diagonalizable, then the diagonal entries of D will be the eigenvalues of A with
algebraic multiplicity of λ corresponding to the number of times λ appears on the diagonal of D.

(1) Yes. This matrix A is in fact self adjoint, and so is diagonalizable over C. Note that the
eigenvalues of A are strictly real since it is self adjoint. If A has only one distinct eigenvalue,
this implies that A = λ∗I for some real λ, since this is its Jordan canonical form. A is similar
to itself. If A has two distinct (real) eigenvalues, then note the characteristic polynomial of
A = (x − λ1 )(x − λ2 ). This will also be the characteristic polynomial of A when viewed as
an operator on R2 , and so A has a real eigenvector for each of the eigenvalues λ1 and λ2 .
These are linearly independent since they correspond to different eigenvalues. Thus, they
form a basis of R2 , so A is diagonalizable.

One might note that I couldn’t find a quick argument for why diagonalizability over C
implies the same over R. If
 anyone wants to write that up, be my guest.
1 i
(2) No. Consider A = . The characteristic polynomial of this matrix is (λ − 2)2 , but
i 3
nullity(A − 2I) = 1, so A cannot have an orthonormal basis of eigenvectors. (One should
guess that this is false since symmetry is not self-adjointness in C. To construct such a
matrix, one should start by guessing that the off-diagonal elements are i, and hope that the
first entry = 1 works. Solve for what the last entry should be so that the matrix has one
eigenvalue).  
1 1
(3) No. Consider A = . The characteristic polynomial of this matrix is λ2 + 1 (since
1 2
3 = 0 in this field). But this equation has no roots in F3 (check each element separately).
Therefore, A has no eigenvalues at all, and hence, cannot be diagonalizable. (One might
also guess that this is false since this field is not algebraically closed. There are only 18
matrices worth checking, and this seemed like the first one to try.)
8
 P∞
Problem 19 (7). Let an be non increasing and positive, where n=1 an < ∞. Show that
limn→∞ nan = 0
P∞
Pm ε > 0. Since n=1 an < ∞, there exists a natural number N such that n, m > N implies that
Fix
a < ε (since each ai is positive). Since the ai are nonincreasing, we have that (m−n+1)∗am =
Pnm i PM
n aM ≤ N ai < ε. Therefore, ε ≥ limsupm (m − n + 1)am = limsupm mam − lim(1 − n)am .
The rightmost limit is zero since 1 − n is fixed and am decreases to zero (since the infintie sum
converges). Since limsup mam < ε for any epsilon, we conclude that limsupmam = 0. But since
mam is strictly positive, we must have mam → 0.

Problem 20 (8). Let a < b be real numbers with f : [a.b] a function s.t. L(x) = limy→x f (y)
exists for all x ∈ [a, b].
(1) Show L is continuous on [a, b].
(2) Show that {x ∈ [a, b] : f (x) 6= L(x)} is countable
(3) Show f is Riemann integrable.

(1) Fix x ∈ [a, b]. Then limz→x L(z) = limz→x limy→z f (y) = limy→x f (y) = L(x) by definition
of L.
(2) I will show that Sn = {x ∈ [a, b] : |f (x) − L(x)| > 1/n} is finite, since the given set will then
be a countable union of finite sets, and hence countable. Suppose instead that there are
infinitely many x ∈ [a, b] such that |f (x) − L(x)| > 1/n, and let xi be an infinite sequence
of distinct such elements. Since [a, b] is compact, there exists a convergent subsequence,
which we call the yi , that converges to y. Since L is continuous, L(y) = limi L(yi ). Since
|f (yi ) − L(yi )| > 1/n for each yi , we have lim(L(yi )) > lim(f (yi ) + 1/n) = L(y) + 1/n by
definition of L. This is a contradiction, since 1/n is not less than zero. Thus, Sn must be
finite.
(3) The set from part b is precisely the set of discontinuities of f , and so this is countable. If
f is unbounded, let an be a sequence of points in [a, b] such that f (an ) ≥ n. Since [a, b]
is compact, there exists a convergent subsequence, converging to some x. By definition,
L(x) = limy→x f (y). But since the ank become arbitrarily close to x, and f (ank ) does not
converge, L(x) cannot be defined. Thus f must be bounded. Since f is bounded and has
countably many discontinuities, by Riemann-Lebesgue criterion, f is integrable on [a, b].

Problem 21 (9). Let (X, ρ) be a complete metric space and f : X → X a function. If f n is

n (x),f n (y))
the n − th iterate of f , denote cn = supx,y∈X,x6=y ρ(f ρ(x,y) . If ∞
P
n=1 cn < ∞, show f has
a unique fixed point in X.

For any x, y ∈ X with x 6= y and any natural number n, the given expression says that ρ(f n (x), f n (y)) ≤
ρ(x, y) ∗ cn . Fix x ∈ X and consider the sequence x, f (x), f 2 (x).... Then we see, if i < j, that
ρ(f i (x), f j (x)) = ρ(f i (x), f i (f j−i (x))) ≤ ci ∗ ρ(x, f j−i (x)). Likewise, we use triangle inequality to
≤ j−i
Pj−i
show that ρ(x, f j−i (x)) n n−1 (x)) ≤
P
P∞ n=1 ρ(f (x), f n=1 cn−1 ρ(f (x), x) (we say c0 = 1). Now fix
ε > 0 and write c = n=1 cn . There exists N ∈ N s.t. i > N implies that ci < ε/(c ∗ ρ(f (x), x)).
For any N < i < j, we have that ρ(f i (x), f j (x)) ≤ ci ρ(f (x), x) j−i
P
n=1 cn ≤ ci ρ(f (x), x) ∗ c ≤ ε. This
sequence is thus Cauchy, so converges to some y.

9
I will now show that this y is a fixed point of f , and that it is unique. Fix ε > 0, and choose N so
that i > N implies that ρ(f i (x), y) < min ε/2, ε/(2c1 ) (if c1 = 0 then the proof was trivial). Then
ρ(f (y), y) ≤ ρ(f (y), f i+1 (x)) + ρ(f i+1 (x), y) ≤ c1 ∗ ρ(y, f i (x)) + ρ(f i+1 (x), y) ≤ ε/2, ε/2. Since this
is true for any ε, we have ρ(y, f (y)) = 0, and so y = f (y).

For uniqueness, suppose that z is another fixed point of f . Then for any positive integer n,
ρ(y, z) = ρ(f n (y), f n (z)) ≤ cn ρ(y, z). Since the cn must converge to zero, this will be false un-
less ρ(y, z) = 0 i.e. y = z.

Problem 22 (10). Let a < b be real numbers and f : [a, b] → R a continuous function s.t.
Rb n
a f (x)x dx = 0 for each integer n ≥ 0. Prove f = 0

By Stone-Weierstrauss theorem, the set of polynomials with real coefficients is P dense in the set
n i
of continuous functions on [a, b] For any p(x) a real polynomial, write p(x) = i=1 ai x . Then
Rb Pn Rb i
a f (x)p(x) = i=0 ai a f (x)x = 0. f is a continuous function on a compact set, so it is bounded
by some M > 0. Fix an ε > 0, and choose a p(x) such that sup|p − f | < ε/(M (b − a)). Then
Rb 2 Rb Rb Rb Rb R 2
a f = a f p + a f (f − p) = a f (f − p) ≤ a M ∗ ε/(M (b − a)) = ε. Therefore, f = 0. Since
f 2 is continuous and positive, this implies that f 2 = 0, and so f = 0.

Problem 23 (11). Prove Young’s inequality: If p, q ∈ (1, ∞) and 1/p + 1/q = 1, then for
any a, b ≥ 0, ab ≤ ap /p + bq /q

Rewriting, we must show ap /p + bq /q − ab ≥ 0. Fix b ≥ 0, and consider the continuous func-

tion f (a) = ap /p + bq /q − ab. Since p > 1, lima→∞ f (a) = ∞, so for some M > 0, a ≥ M
implies that f (a) > 1. Consider the compact interval [0, M ], and since f is continuous on this
compact set, it obtains a minimum on this set. I will show that for a local minimum inside this
set, f (x) = 0. Since f (0), f (M ) > 0, 0 is the global minimum for f (a). Since this is true
for any b, we have ap /p + bq /q − ab ≥ 0 always, which is what we wanted. It remains to com-
pute the local minima of f inside [0, M ]. The derivative of f is ap−1 − b. This is zero when-
ever ap−1 = b. Furthermore, f 00 (a) ≥ 0, so any critical point is a local minimum. We compute
f (b1/(p−1) ) = bp/(p−1) /p + bq /q − b1/(p−1) b = bq /p + bq /q − bq = 0, and hence, the global minimum
of f is zero.

Note: Most online sources give a slicker proof using that log is concave and increasing. I don’t
see how anyone would think of that on the exam, although the condition on p and q may suggest
concavity to some people. I am sure there are other solutions.

Problem 24 (12). Let X be a compact metric space and C(X) be the space of continuous
real-valued functions on X endowed with the supremum norm. Let F ⊂ C(X) be non-empty.
Prove that F is compact iff F is closed, bounded, and equicontinuous.

If F is compact, then F is immediately closed and bounded. It remains to show it is equicontinuous.

Fix ε > 0. Let Sδ be the set of f ∈ F such that |x − y| ≤ δ implies |f (x) − f (y)| ≤ ε. Since each
f is continuous on a compact set, it is uniformly continuous, and so the set of Sδ forms a cover of
F . Note that if δ1 < δ2 that Sδ1 ⊃ Sδ2 . So for our finite subcover, some Sδ is in fact a cover of
F , implying that every f ∈ F satisfies |x−y| < δ implies |f (x)−f (y)| ≤ ε. F is thus equicontinuous.

10
For the reverse direction, we will show F is closed and totally bounded, since this implies com-
pactness in a metric space. We have already assumed F is closed. Fix ε > 0. We will construct
a finite epsilon net of F . Since F is equicontinuous, there exists δ > 0 where |x − y| < δ im-
plies that |f (x) − f (y)| < ε/3 for all x, y ∈ X and f ∈ F . Since X is compact, there exists a
finite δ net of X, say the ai . For all f ∈ F , consider the vector in Rn of (f (a1 ), f (a2 ), f (a3 )...).
F is equibounded, so each component of these vectors are bounded, and hence each vector is
bounded in magnitude. These vectors form a bounded subset of Rn , and so by compactness of
this set, there exists a finite ε/3 net of these vectors. Equivalently, there exists a finite set of fi
where for any g ∈ F , |g(aj ) − fi (aj )| ≤ ε/3 for each aj and fi . Fix x ∈ X. Then there exists
an aj where |x − aj | < δ. There also exists an fi where |fi (aj ) − g(aj )| < ε/3 for all j. So
|g(x) − fi (x)| ≤ ||g(x) − g(aj )| + |g(aj ) − fi (aj )| + |fi (aj ) − fi (x)|. Since |aj − x| < δ, the first and
last term are less than ε/3, and the middle is less than ε/3 by choice of fi and aj . This implies that
the fi are a finite ε net of F , and so F is totally bounded, and hence also compact.

3. S18

Problem 25 (1). ] Prove that et , e2t ...ent are linearly independent in the space of continuous
functions on the interval [1, 2].

We argue by induction on n. The base case n = 1 is vacuous. Suppose that the first n of these
functions are linearly independent. Suppose that a1 et + a2 e2t + ...an+1 e(n+1)t = 0 for some ai and
for all t in this interval. Taking derivatives of both sides, we obtain a1 et + 2a2 e2t + ...nan ent + (n +
1)an+1 e(n+1)t = 0. Subtracting n+1 of the first equation from the second, we have (1−(n+1))a1 et +
(2 − (n + 1))a2 e2t + ...(n − (n + 1))ent = 0. Simplifying, −na1 et + (−n + 1)a2 e2t + ... − 1an ent = 0.
By induction, the coefficients bi = (i − n − 1)ai are all zero. Since i > n + 1 for these i, this implies
all ai are zero. Our original equation now reads an+1 e(n+1)t = 0 for all t, so an+1 = 0. This set is
thus linearly independent.

Problem 26 (2). Let A and B be two real 5 × 5 real matrices such that A2 = A, B 2 = B,
amd 1 − (A + B) is invertible. Prove rank(A) = rank(B).

Observe that A(1 − A − B) = A − A2 − AB = −AB. Likewise, (1 − A − B)B = B − B 2 − AB =

−AB = A(1 − A − B). Therefore, rank(A(1 − A − B)) = rank((1 − A − B)B). Since (1 − A − B) is
invertible, rank(A) = rank(A(1 − A − B)) = rank((1 − A − B)B) = rank(B). (To see this last step,
one can use Sylvester’s rank inequality, or just note that the rank of (1 − A − B)|W = dim(W )).

Problem 27 (3). Let A = (ai,j ) be a complex n × n matrix. Suppose eA = 1 + A + A2 .

Prove or disprove: A = 0
 
0 1
This is false. Every Mn (C) contains a nonzero nilpotent matrix A where A2
= 0 (i.e. and
0 0
its extensions). For this matrix, eA = 1 + A + A2 /2! + ... = 1 + A. Likewise, 1 + A + A2 = 1 + A.

For an alternative solution, we show that the real equation f (x) = ex − x2 − x − 1 = 0 has a positive
solution. Note that this function on the left side is continuous since it is the difference of continuous
functions. Then f (1) < 0, f (10) > 0, and so by I.V.T., there exists a c ∈ (1, 10) s.t. f (c) = 0. Now,
consider the diagonal matrix D where each diagonal entry is c. One sees that eD = 1 + D + D2 .
11
 Problem 28 (4). Consider the following matrices (I’m not retyping these, go look at the
problem sheet). Which pairs are similar over R?

Two real matrices are similar over R iff they have the same real Jordan canonical form. Each of
these matrices is triangular, so we read off that the characteristic polynomial of each matrix is
(x − 1)3 . There are thus three possibilities for the real canonical form: 3 blocks of size 1, one block
of size 2 and one block of size 1, or 1 block of size 3. Therefore, the Jordan form for each matrix is
determined by the total number of blocks.
Note that the number of blocks in the J.C.F. for eigenvalue λ is equal to nullity(X − λI). For
matrices A through E, we see that this nullity is 2, so these are all similar. For matrix F , this
nullity is 1, so this is not similar to any of the other matrices.

Problem 29 (5). Let A, B be two positive definite 2 × 2 matrices. Prove or disprove:

(1) A+B is positive definite
(2) AB+BA is positive definite

The first claim is true. Fix any nonzero v ∈ C2 . Then hx, (A+B)xi = hx, Axi+hx, Bxi by linearity,
and each term is strictly greater than zero
 by positive
 definiteness of A, B.
2 0
The second claim is false. Consider A = . This is positive definite since it is symmetric and
0 1
 √ 
1/8
√ 1/ 2
both eigenvalues (2 and 1) are positive. Let B = . The characteristic polynomial
1/ 2 9/2
of this matrix is λ2 − 37λ/8 + 1/16. By Descartes’ rule of signs, this polynomial has only positive
roots, and so B is symmetric
√  with positive eigenvalues, and so is positive definite. We compute
1/2
√ 3/ 2
AB + BA = . The characteristic polynomial is λ2 − 19λ/2, and so zero is an
3/ 2 9
eigenvalue of this matrix, and so this matrix is not positive definite.

The example seems rather strange, but the idea was that if we can force the constant coefficient of
the characteristic polynomial to be zero or less, then we are done. To do so, if we let B be a matrix
that is "almost" not positive definite, and scale its rows and columns, we ought to be able to make
it no longer positive definite.

Problem 30 (6). Compute the determinant of the following matrix: (not retyped for sake
of time)

Note that subtracting a multiple of one row from another does not change the determinant (this is
equivalent to left multiplying by a particular elementary matrixwith determinant one). We subtract

1 2 4 8 16
0
 0 0 0 −31 

multiples of the first row from each remaining row to obtain 0  0 0 −31 −62  . We
0 0 −31 −62 −124
0 −31 −62 −124 −248
compute the determinant by taking a column sum along the first column. The determinant is
1 ∗ det(A1,1 ) + 0 ∗ det(A2,1 ) + 0 ∗ .... = 1 ∗ det(A1,1 ). The determinant of A1,1 can be computed by
taking repeated column sums along the first column, yielding that det(A) = 1 ∗ (−1 ∗ −31) ∗ (−31) ∗
(−1 ∗ −31) ∗ −31 = 314 .
12
 P∞ sin(πn/p)
Problem 31 (7). Prove that for each p ∈ N, the infinite series n=1 n converges.

Let an = sin(πn/p) and bn =P1/n. By theorem, if an has bounded partial sums, bn is non-
increasing, and bn → 0, then ∞ n=1 an bn converges. 1/n is indeed nonincreasing and converges
to zero, so it remains to show that an = sin(πn/p) has bounded partial sums. Observe that
P2p
sin(πn/p) = − sin(π(−n)/p) = − sin(π(2p − n)/p). Consider the sum n=1 sin(πn/p). For
each n ∈ [1, 2p]\{p, 2p}, we see that 2p − n ∈ [1, 2p]\{p, 2p} and is distinct from n. Therefore,
P2p P
n=1 sin(πn/p) = sin(π) + sin(2π) + n=1,n6=p,n6=2p sin(π ∗ n/p) = 0. Likewise, we compute that for
P2pk
any positive integer k, n=1 sin(πn/p) = 0 since the summand is periodic in n with period 2p. Now,
for any positive integer m, by Euclidean
P2p Pm m = 2p ∗ q + r for integers
algorithm, write Pm q and r and r <
2p. Then m
P
n=1 sin(2π∗n/p) = n=1 sin(2π∗n/p)+ n=2p+1 sin(2π∗n/p) = n=2p+1 sin(2π∗n/p).
Pm Pm
Since | sin(2π ∗ n/p)| ≤ 1, we see | n=1 sin(2π ∗ n/p)| ≤ n=2p+1 | sin(2π ∗ n/p)| = r < 2p. There-
Pevery partial sum of the an is bounded in [−2p, 2p]. The conditions of our theorem are satisfied,
fore,
so ∞ n=1 an bn converges.

Problem 32 (8). This problem is notoriously hard and I don’t have the bravery to try to
understand and type it up. Sylvester posted his solution on CCLE, although it is unlikely
anyone could write it up in full detail for the exam. It just seems like an unreasonable and
uninstructive problem.

Problem 33 (9). Let f : R → R be non-decreasing (but not necessarily continuous).

Prove f is Riemann integrable on any finite interval (a, b) ⊂ R. DO NOT PROVE US-
ING LEBESGUE THEORY UNLESS YOU ALSO PROVE LEBESGUE THEORY.

To prove using Lebesgue theory, one would first have to demonstrate that f has countably many
discontinuities (assign a distinct rational number to each jump, or use the trick Sylvester showed
on CCLE), and then prove that this implies that we can find partitions that minimize the contribu-
tion of each jump (create an interval around each jump of size less than 2−j if this is jump number j).

A direct proof is more straightforward. Fix such (a, b) and let Pn be the partition of (a, b) into
n uniform segments, i.e., x0 = a, x1 = a + (b − a)/n, x2 = a + (b − a) ∗ 2/n, xn = b. Then
Pn−1
the upper sum U (f, Pn ) = i=0 (xi+1 − xi ) ∗ supx∈[xi ,xi+1 ] f (x). Since f is nondecreasing, this
supremum is always f (xi+1 ), and by choice of partition, xi+1 − xi is always (b − a)/n. Then
U (f, Pn ) = n−1
P Pn−1
i=1 f (xi+1 )∗(b−a)/n. Likewise, we compute L(f, Pn ) = i=1 f (xi )∗(b−a)/n. There-
fore, U (f, Pn ) − L(f, Pn ) = (f (b) − f (a)) ∗ (b − a)/n. The limit of the sequence U (f, Pn ) − L(f, Pn )
as n → ∞ is therefore zero, and so f is Riemann integrable.

Problem 34 (10). Let n ∈ N and U ⊂ Rn be nonempty, open, and connected. Suppose

f : U → R is such that all first partial derivatives of f exist and vanish at each point of U .
Prove f is constant.

I will first show f is constant on each open ball B contained in U . Fix any point x ∈ U , and
since U is open, let B(x, r) be the ball of radius r around x, and suppose this ball is contained
in U . Fix y ∈ B(x, r). Write x = (x1 , x2 , ...xn ) and y = (x1 + h1 , x2 + h2 , ...xn + hn ). Then
f (y) − f (x) = (f (y) − f (x1 + h1 , x2 + h2 , ...xn )) + (f (x1 + h1 , ...xn ) − f (x1 + h1 , ...xn−1 , xn )) +
(f (x1 + h1 , x2 ...xn ) − f (x)), where for any term in this sum, the arguments of f differ in only one
13
coordinate. Note that each (x1 + h1 , x2 + ....., xi + hi , xi+1 , xi+2 ..., xn ) is contained in B(x, r) since
their distance from x must be less than d(x, y). By fundamental theorem of calculus, one com-
putes that f (x1 + h1 , x2 + ....., xi + hi , xi+1 , xi+2 ..., xn ) − f (x1 + h1 , x2 + ....., xi , xi+1 , xi+2 ..., xn ) =
R hi R hi
0 ∂i f (x1 + h1 , x2 + ....., xi + s, xi+1 , xi+2 ..., xn )ds = 0 0 = 0 since the partial derivatives vanish
everywhere in U . This means that each term from our sum is zero, and so f (y) = f (x). Therefore,
f is constant on any open ball contained in U .

We now use an interesting characterization of connected sets (an exercise in the week two anal-
ysis notes). Let C be the open cover of U consisting of every open ball contained in U . Since U
is connected, for any Ba and Bb , there exists a sequence of balls B1 = Ba , B2 ...Bn = Bb where
Bi ∩ Bi+1 6= ∅. Now, fix any x, y ∈ U , and pick a Bx in C containing x and a By containing y in
C. There exists a chain between them as described above. Then since Bx ∩ B2 6= ∅, and since f is
constant on each of Bx and B2 , we must have that f (c) = f (x) for each c ∈ B2 . Continuing in this
way, f (ci ) = f (x) for each ci ∈ Bi . In particular, since y ∈ Bn , f (y) = f (x), as desired.

There is probably a way to do this with path connectedness that I am not seeing. Since U is
nonempty, open, and connected, it is path connected. Every directional derivative vanishes since
each partial one does. Is there a quick way we can integrate along a path connecting x and y?

Problem 35 (11). Let (X, ρ) be a compact metric space and let f : X → X be an isometry.
Prove f (X) = X.

Suppose towards contradiction that f (X) is a proper subset of X, and pick x0 ∈ X\f (X). f is
an isometry, and so maps open sets to open sets (and hence, closed sets to closed sets), and so
we see that f (X) is clopen. This implies that there exists an r > 0 such that B(x, r) ⊂ X\f (X)
since X\f (X) is open. One sees that f (B(x, r)) = B(f (x), r). Note that we may repeat these
arguments to show that, if f i is f applied i times, that f i (B(x, r)) = B(f i (x), r). Since X is
compact, some subsequence of the f i (x) must converge. But since each f i (x) ∈ f i (X)\f i+1 (X), the
balls B(f i (x), r) are actually disjoint. This is a contradiction, since the f i (x) are always separated
by distance at least r.

Problem 36 (12). Let F be a family of real-valued functions on a compact metric space

taking values in [−1, 1]. Prove that if F is equicontinuous then the function
g(x) = sup{f (x) : f ∈ F } is continuous.

Pick ε > 0 and x in our metric space. By equicontinuity of F and compactness of our metric space,
there exists δ > 0 such that |x − y| < δ implies that |f (x) − f (y)| < ε/2 for any x, y in our space and
f ∈ F . Now suppose |x − y| < δ. By definition of supremum, there exists a function fa ∈ F such
that g(x) ≤ fa (x) + ε/2. By equicontinuity, fa (x) + ε/2 ≤ fa (y) + ε/2 + ε/2 ≤ g(y) + ε. Therefore,
g(x) − g(y) ≤ ε.

Likewise, there exists fb ∈ F such that g(y) ≤ fb (y)+ε/2. We likewise compute that g(y) ≤ g(x)+ε.
Thus, we have |g(x) − g(y)| ≤ ε, so g is continuous.
14
 4. F18

Problem 37 (1). Let {an }n≥1 be a sequence of nonnegative numbers such that
X
an diverges.
n≥1
Show that X an
diverges.
2an + 1
n≥1

Suppose that there exists some positive integer N such that an < 1 for all n ≥ N . Then for n ≥ N ,
we have
an 1
> an .
2an + 1 3
P∞
Since the series
PM n=N an diverges and an ≥ 0 for all n, we have that the set of partial sums
n=N an is unbounded. Since
M M
X an 1 X
> an ,
2an + 1 3
n=N n=N
P∞
Then the partial sums M an an
P
P∞ n=N 2an +1 are unbounded, so the series n=N 2an +1 diverges, so the series
an
n=1 2an +1 diverges.
Suppose instead that there are infinitely many n such that an ≥ 1. For such n,
an 1 1/2 1 1/2 1
= − ≥ − = .
2an + 1 2 2an + 1 2 3 3
Since an /(2an + 1) ≥ 1/3 for infinitely many n, the limit of an /(2an + 1) cannot be zero, so the
series diverges again.

Problem 38 (2). Let A be a connected subset of Rn such that the complement of A is the
union of two separated sets B and C, that is
Rn \ A = B ∪ C with B ∩ C = B ∩ C = ∅.
Show that A ∪ B is a connected subset of Rn .

Let D be a nonempty subset of A ∪ B that is relatively open and closed in A ∪ B. Then D ∩ A is

relatively open and closed in A, and since A is connected, we either have A ⊆ D or D ∩ A = ∅.
Suppose first that D ∩ A = ∅. Then D ⊆ B, and (A ∪ B) \ D is also relatively open and closed in
A ∪ B, and A ⊆ (A ∪ B) \ D. So we can replace D with (A ∪ B) \ D and get A ⊆ D.
We claim that D ∪ C is open and closed in Rn . First, D ∪ C is closed: let x ∈ C ∪ D. If x 6∈ C
and x 6∈ D, then there exists ε > 0 such that the ε-ball around x doesn’t meet either D or C,
contradicting x being in the closure of D ∪ C. So x ∈ D ∪ C. If x ∈ C, then x 6∈ B. So x ∈ A ⊆ D
or x ∈ C, so x ∈ D ∪ C. If x 6∈ C, then x ∈ D. If x ∈ A ∪ B, then x ∈ D because D is relatively
closed in A ∪ B. Otherwise, x ∈ C. In any case, x ∈ D ∪ C. This shows that D ∪ C is closed in Rn .
Now we show that D ∪ C is open in Rn . Let x ∈ D ∪ C. Case 1: x ∈ D. Since D is relatively open
in A ∪ B, there exists ε1 > 0 such that |y − x| < ε1 and y ∈ A ∪ B implies y ∈ D. Then for any
y ∈ Rn with |y − x| < ε1 , either y ∈ D or y 6∈ A ∪ B, in which case y ∈ C. So D ∪ C contains the
ε1 -ball centered at x. Case 2: x ∈ C. Then x 6∈ B, so there exists ε2 > 0 such that |y − x| < ε2
implies y 6∈ B. Then y ∈ A ∪ C ⊆ D ∪ C. So D ∪ C contains an ε2 -ball centered at x. Since D ∪ C
contains an open ball around each point, we have D ∪ C is open in Rn .
15
Since Rn is connected, we must have D ∪ C = Rn . Then since C is disjoint from A ∪ B,
D ∩ (A ∪ B) = (D ∪ C) ∩ (A ∪ B) = A ∪ B.
So A ∪ B = D. We have shown that the only nonempty subset of A ∪ B that is both relatively open
and closed in A ∪ B is all of A ∪ B, so A ∪ B is connected.

Problem 39 (3). Let f : [0, 1] → R and g : [0, 1] → [0, 1] be two Riemann integrable
functions. Assume that
|g(x) − g(y)| ≥ α|x − y| for any x, y ∈ [0, 1]
and some fixed α ∈ (0, 1). Show that f ◦ g is Riemann integrable.

We rely on the Riemann-Lebesgue theorem, which says that a bounded function on [a, b] is Riemann
integrable if and only if its set of discontinuities has measure zero. By measure zero, we mean for
all ε > 0, the set of discontinuities can be covered by countably many open intervals whose lengths
sum to less than ε. Let `(I) denote the length of an interval I, so that `((a, b)) = b − a.
Note that if g is continuous at x and f is continuous at g(x), then f ◦ g is continuous at x. Let
Df ◦g , Df , and Dg be the sets of discontinuities of f ◦ g, f , and g respectively. So x ∈ Df ◦g if and
only if x ∈ Dg or g(x) ∈ Df , so
Df ◦g = Dg ∪ g −1 (Df ).
P∞ ε > 0. Since Dg has measure
Let S zero, there exist open intervals I1 , I2 , · · · ⊆ R such that
k=1 `(I k ) < ε/2 and D
Pg ⊆ I
k k . Since Df hasS measure zero, there exist open intervals
I10 , I20 , · · · ⊆ R such that ∞k=1 `(Ik
0 ) < αε/2 and D ⊆
g
0
k Ik .
Consider an open interval (a, b). If x, y ∈ g −1 ((a, b)), then
1 b−a
|x − y| ≤ |g(x) − g(y)| < .
α α
It follows that g −1 ((a, b)) is contained in an open interval of length (b − a)/α. For each Ik0 , let Jk
be an open interval containing g −1 (Ik0 ) such that `(Jk ) = `(Ik0 )/α. For all x ∈ g −1 (Df ), we have
g(x) ∈ Df , so g(x) ∈ Ik0 for some k, so x ∈ g −1 (Ik0 ), so x ∈ Jk . This shows that
[
g −1 (Df ) ⊆ Jk ,
k
and
∞ ∞
X 1X 0
`(Jk ) = `(Ik ) < ε/2.
α
k=1 k=1
Then [ [
Df ◦g ⊆ Ik ∪ Jk ,
k k
and X X
`(Ik ) + `(Jk ) < ε/2 + ε/2 = ε.
k k
We have covered Df with countably many open intervals whose lengths sum to less than ε. Since
ε > 0 was arbitrary, Df ◦g has measure zero. Also, f is bounded, so f ◦ g is bounded. It follows
from Lebesgue’s criterion for Riemann integrability that f ◦ g is Riemann integrable.

Problem 40 (4). Let f : [0, 1] → R be a continuous function on the closed interval [0, 1]
and differentiable on the open interval (0, 1). Assume that f (0) = 0 and f 0 is a decreasing

16
 function on (0, 1). Show that
f (x)
g(x) =
x
is a decreasing function on (0, 1).

Suppose for the sake of contradiction that g is not decreasing on (0, 1). Then for some 0 < x < y < 1,
we have g(y) ≥ g(x). Since g is differentiable on (0, 1), by the Mean Value Theorem, there exists
c ∈ (x, y) such that
f (y) − f (x)
g 0 (c) = ≥ 0.
x−y
Note that
cf 0 (c) − f (c)
g 0 (c) = .
c2
So cf 0 (c) ≥ f (c). Since f is continuous on [0, c] and differentiable on (0, c), the mean value theorem
again says that there exists d ∈ (0, c) such that
f (c) − f (0) f (c)
f 0 (d) = = ≤ f 0 (c).
c−0 c
So d < c and f 0 (d) ≤ f 0 (c). But this contradicts f 0 decreasing on (0, 1). So the original assumption
was false, and g is decreasing on (0, 1).

Problem 41 (5). Let B := {x ∈ Rn : |x| ≤ 1} and let g : ∂B → R be a 1-Lipschitz function.

(a) Show that the function f : B → R given by
f (x) := inf [g(y) + |x − y|]
y∈∂B
is 1-Lipschitz.
(b) Show that the set M (g) := {h : B → R | h is 1-Lipschitz and h|∂B = g} is compact
in the space of continuous functions on B endowed with the supremum norm.

(a) Let x1 , x2 ∈ B. The function y 7→ g(y) + |xk − y| is continuous on the set ∂B. Also, ∂B is
closed and bounded, so is compact. So for k = 1, 2, there exists some yk ∈ ∂B such that
f (xk ) = inf [g(y) + |xk − y|] = g(yk ) + |xk − yk |.
y∈∂B

By definition of infimum, we have

f (x2 ) = g(y2 ) + |x2 − y2 | ≤ g(y1 ) + |x2 − y1 |.
Then by the reverse triangle inequality,
f (x2 ) − f (x1 ) ≤ g(y1 ) + |x2 − y1 | − g(y1 ) − |x1 − y1 | ≤ |x2 − x1 |.
By interchanging x2 and x1 ,
f (x1 ) − f (x2 ) ≤ |x2 − x1 |.
So
|f (x1 ) − f (x2 )| ≤ |x2 − x1 |
and f is 1-Lipschitz.
17
 (b) The Arzelà-Ascoli Theorem states that for a compact metric space K and a subset F ⊆
C(K, R) of the continuous functions from K → R endowed with the supremum norm, we
have F compact if and only if F is closed, uniformly bounded, and equicontinuous. In
this case, K = B is compact because B is closed and bounded in Rn . Furthermore, every
function in M (g) is 1-Lipschitz, so continuous, so M (g) ⊆ C(B, R). So it suffices to show
that M (g) is equicontinuous, and closed and bounded with respect to the supremum norm.
Let ε > 0. Suppose |x − y| < ε for some x, y ∈ B. Let h ∈ M (g). Since h is 1-Lipschitz,
we have
|h(x) − h(y)| ≤ |x − y| < ε.
This shows that M (g) is equicontinuous.
Let h1 , h2 , · · · ∈ M (g) be a sequence with hn converging uniformly to some continuous
function h : B → R (equivalently, converging in the supremum norm). In particular, hn → h
pointwise. Since all hn are 1-Lipschitz, for any x, y ∈ B, we have
|h(x) − h(h)| = lim |hn (x) − hn (y)| ≤ lim |x − y| = |x − y|.
n n

So h is 1-Lipschitz. Let x ∈ ∂B. Then hn (x) = g(x) for all n. So h(x) = limn hn (x) =
limn g(x) = g(x). So h|∂B = g. So h ∈ M (g). Since M (g) contains all of its limit points,
M (g) is closed.
For all x, y ∈ B, we have |x − y| ≤ |x| + |y| ≤ 2. Fix some x0 ∈ ∂B. Then for all x ∈ B
and all h ∈ M (g), we have
|h(x)| ≤ |h(x) − h(x0 )| + |h(x0 )| ≤ |x − x0 | + |g(x0 )| ≤ 2 + |g(x0 )|.
So M (g) is uniformly bounded by 2 + |g(x0 )|. So M (g) is closed and bounded with respect
to the supremum norm, and equicontinuous, so is compact with respect to the supremum
norm.

Problem 42 (6). For x ∈ (0, ∞), let

∞
1 − e−tx
Z
F (x) = dt.
0 t3/2
Show that F : (0, ∞) → (0, ∞) is well-defined, bijective, of class C 1 , and that its inverse is
of class C 1 .

By using u-substitution, we let u(t) = tx for x > 0 fixed and t > 0. Then for ε > 0 and M > ε,

1 − e−tx 1 − e−u
Z M Z Mx
√
3 dt = x du
ε t2 εx u3/2
Let
M
1 − e−u
Z
C(ε, M ) := du.
ε u3/2
The integrand is nonnegative for u > 0. To show that C(ε, M ) converges as ε → 0 and M → ∞, it
suffices to show that C(ε, M ) is bounded above by some constant, since C(ε, M ) is monotonically
increasing as both ε decreases and M increases. For u > 0, by the mean value theorem, there exists
some v > 0 such that (1 − e−u )/u = e−v ≤ 1. So 1 − e−u ≤ u. Then
1 − e−u
Z 1 Z 1
3/2
du ≤ u−1/2 du = [2u1/2 ]1ε ≤ 2.
ε u ε
18
For u ≥ 1, we have 1 − e−u ≤ 1 − e−1 . So
1 − e−u
Z M Z M
−1
du ≤ (1 − e ) u−3/2 du = (1 − e−1 )(2 − 2M −1/2 ) ≤ 2 − 2e−1 ≤ 2.
1 u3/2 1

This shows that

M
1 − e−u
Z
du ≤ 4.
ε u3/2
So C(ε, M ) converges to some constant C > 0 as ε → 0 and M → ∞. Then for all x > 0, we have
C(εx, M x) → C as well. So
Z ∞
1 − e−tx √
F (x) = 3 dt = C x
0 t2
Is well-defined. This function is bijective (0, ∞) → (0, ∞), with inverse G(y) = (y/C)2 . We check
that G and F are mutually inverse: for x > 0 and y > 0,
√ 2 p
G(F (x)) = (F (x)/C)2 = x = x, F (G(y)) = C G(y) = C(y/C) = y.
We have F 0 (x) = 2√ C
x
is continuous on (0, ∞), so F is C 1 , and G0 (x) = 2y/C 2 is also continuous,
so G is C 1 as well.

Problem 43 (7). Let T : Rn → Rn be a linear transformation with the property that

T (T (x)) = T (x) for all x ∈ Rn .
Show that there exists 1 ≤ m ≤ n and a basis of Rn such that in this basis the entries of T
satisfy (
1 if i = j and 1 ≤ i ≤ m,
Tij =
0 otherwise.

Note: this problem is technically false, since the conclusion would imply T11 = 1 always, but T
could be the zero map.
Let m = rank(T ), and let v1 , . . . , vm be a basis of im(T ) ⊆ Rn . So there exists x1 , . . . , xm ∈ Rn
such that T (xk ) = vk for all k. By the Rank-Nullity Theorem, dim(ker(T )) = n − rank(T ) = n − m.
So let vm+1 , . . . , vn be a basis of ker(T ). We claim that v1 , . . . , vn is a basis for Rn . It suffices to
show that these n vectors span Rn , since Rn has dimension n. Let x ∈ Rn . Let x0 = x − T (x).
Then
T (x0 ) = T (x) − T (T (x)) = 0.
So x0 ∈ ker(T ) = span{vm+1 , . . . , vn }, and T (x) ∈ im(T ) = span{v1 , . . . , vm }, so x = T (x) + x0 ∈
span{v1 , . . . , vn }. So v1 , . . . , vn is a basis for Rn .
In this basis, for 1 ≤ i ≤ m, we have
T (vi ) = T (T (xi )) = T (xi ) = vi .
So Tij = 1 if i = j and 1 ≤ i ≤ m, and Tij = 0 if 1 ≤ j ≤ m and i 6= j. Also, for j > m, we have
vj ∈ ker(T ), so T (vj ) = 0, so Tij = 0 if j > m.

Problem 44 (8). Let X be an n × n symmetric (real) matrix and z ∈ C with Im(z) > 0.
Define
G = (X − z)−1 .

19
 Show that
X Im(Gii )
|Gij |2 =
Im(z)
1≤j≤n

By the Spectral Theorem, since X is real and symmetric, X is diagonalizable with real eigenvalues.
So z is not an eigenvalue of X, so X − z has trivial kernel, so is invertible. So G is well-defined. We
have
I = G(X − z) = (X − z)G.
So
GX = G(X − z) + zG = (X − z)G + zG = XG.
So G and X commute. Let G∗ denote the conjugate transpose (adjoint) of G. Then taking adjoints
of the above equation, since X is real and symmetric, X ∗ = X, so
G∗ X = G∗ X ∗ = (XG)∗ = (GX)∗ = X ∗ G∗ = XG∗ .
So X also commutes with G∗ . Also,
GX = I + zG,
and
G∗ X = (XG)∗ = (GX)∗ = I + zG∗ .
So
X X
z |Gij |2 = z Gij Gij
1≤j≤n 1≤j≤n
X
=z Gij (G∗ )ji
1≤j≤n
∗
= (zGG )ii
= ((GX − I)G∗ )ii
= (GXG∗ − G∗ )ii
= (GG∗ X − G)ii + Gii − G∗ii
= (G(G∗ X − I))ii + Gii − Gii
= z(GG∗ )ii + 2 Im(Gii )
X
= 2 Im(Gii ) + z |Gij |2 .
1≤j≤n

Then
X X X
2 Im(z) |Gij |2 = z |Gij |2 − z |Gij |2
1≤j≤n 1≤j≤n 1≤j≤n
= 2 Im(Gii ),
so
X 2 Im(Gii ) Im(Gii )
|Gij |2 = = .
2 Im(z) Im(z)
1≤j≤n

Problem 45 (9). Let f, g : Rn → R be linearly independent elements of the vector space

(over R) of linear mappings from Rn to R. Show that for any v ∈ Rn , there exists v1 and v2

20
 such that
v = v1 + v2 , f (v) = f (v1 ), and g(v) = g(v2 ).

Since f and g are linearly independent, neither can be the zero map, so rank(f ) = rank(g) = 1. By
the Rank-Nullity theorem, we have dim ker f = dim ker g = n − rank(f ) = n − 1.
Let V = ker f ∩ ker g. Suppose for the sake of contradiction that dim V > n − 2. Since V ⊆ ker f ,
we have dim(V ) ≤ n − 1, so dim(V ) = n − 1. But since dim ker f = n − 1, we have V = ker f .
Similarly, V = ker g, so ker f = ker g. Let v ∈ Rn be such that f (v) 6= 0. Let λ = g(v)/f (v), so
g(v) = λf (v). Now let w ∈ Rn . Then
 
f (w)
f w− v = f (w) − f (w) = 0.
f (v)
f (w)
So w − ∈ ker f = ker g, so
f (v) v
 
f (w) f (w) f (w)
0=g w− v = g(w) − g(v) = g(w) − λf (v) = g(w) − λf (w).
f (v) f (v) f (v)
So g(w) = λf (w) for all w ∈ Rn . But this contradicts f and g being linearly independent. So our
assumption was wrong, and dim V ≤ n − 2.
Furthermore, the inclusion-exclusion formula gives
dim(ker f + ker g) = dim ker f + dim ker g − dim(ker f ∩ ker g) ≥ n − 1 + n − 1 − (n − 2) = n.
So ker f + ker g = Rn , so for all v ∈ Rn , there exists v1 ∈ ker g and v2 ∈ ker f such that v = v1 + v2 ,
and
f (v) = f (v1 ) + f (v2 ) = f (v1 ),
and
g(v) = g(v1 ) + g(v2 ) = g(v2 ).
 
1 0 0
Problem 46 (10). Let A :=  12 1
2 0 . Calculate limn→∞ An .
1 1 1
3 3 3

Since A is lower triangular, we can read the eigenvalues off the diagonal as 1, 1/2 and 1/3. Since
there are 3 distinct eigenvalues, A is diagonalizable. Let v1 , v2 , v3 be eigenvectors for 1, 1/2, and
1/3, respectively. Then the matrix P with columns vk is invertible, and A = P DP −1 , where
 
1 0 0
D = 0 1/2 0  .
0 0 1/3
Specifically, we calculate that v1 can be (1, 1, 1), v2 can be (0, 1, 2), and v3 can be (0, 0, 1). Note
that  
1 0 0
Dn = 0 2−n 0 
0 0 3−n
Since matrix multiplication is continuous,
 
1 0 0
lim An = lim P Dn P −1 = P (lim Dn )P −1 = P 0 0 0 P −1 .
n n n
0 0 0
21
Then
        
1 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 0
lim An = 1 1 0 0 0 0 −1 1 0 = 1 1 0 0 0 0 = 1 0 0 .
n
1 2 1 0 0 0 1 −2 1 1 2 1 0 0 0 1 0 0

Problem 47 (11). Let V be the space of all 3 × 3 real matrices that are skew-symmetric,
i.e. At = −A, where At denotes the transpose of A. Prove that the expression
1
hA, Bi = Tr(ABt )
2
defines an inner product on V . Exhibit an orthonormal basis of V with respect to this inner
product.

The trace of a matrix’s transpose is the same as the original trace, so

1 1 1
hB, Ai = Tr(BAt ) = Tr((AB t )t ) = Tr(AB t ) = hA, Bi.
2 2 2
So the form is symmetric.
Now for any A1 , A2 , B ∈ V and c1 , c2 ∈ R, by linearity of matrix multiplication and trace,
1
hc1 A1 + c2 A2 , Bi = Tr((c1 A1 + c2 A2 )B t )
2
1
= Tr(c1 A1 B t + c2 A2 B t )
2
1 1
= c1 Tr(A1 B t ) + c2 Tr(A2 B t )
2 2
= c1 hA1 , Bi + c2 hA2 , Bi.
So the form is linear in the first coordinate (so linear in the second coordinate by symmetry). For
any A ∈ V ,
1
hA, Ai = Tr(AAt )
2
n
1X
= (AAt )ii
2
i=1
n n
1 XX
= Aij (At )ji
2
i=1 j=1
n X n
1 X
= A2ij
2
i=1 j=1
≥ 0,
with equality if and only if Aij = 0 for all i, j, if and only if A = 0. This completes the proof that
the form is an inner product.
For an orthonormal basis of V , let
     
0 1 0 0 0 1 0 0 0
A1 := −1 0 0 , A2 :=  0 0 0 , A3 := 0 0 1 .
0 0 0 −1 0 0 0 −1 0
Each of Ak is visibly skew-symmetric. It suffices to show that they are orthonormal (which implies
they are linearly independent) and that they span V . Note that Ak Atk is diagonal, with two 1s and
22
a zero on the diagonal, so hAk , Ak i = 21 Tr(Ak Atk ) = 12 (2) = 1. It is straightforward to calculate
that the Ak are mutually orthogonal. To see that they span V , suppose A ∈ V . Then for i = 1, 2, 3,
we have Aii = −Atii = −Aii , so Aii = 0. For j 6= i, Aij = −Aji . So
 
0 A12 A13
A = −A12 0 A23  = A12 A1 + A13 A2 = A23 A3 .
−A13 −A23 0
This completes the problem.

Problem 48 (12). Let V be a finite-dimensional vector space. Let T : V → V be a linear

transformation such that T (W ) ⊆ W for every subspace W of V with dim(W ) = dim(V ) − 1.
Prove that T is a scalar multiple of the identity.

Let n = dim(V ). First, we claim that every nonzero vector of V is an eigenvector. Suppose that
some v 6= 0 is not an eigenvector. Then v, T (v) are linearly independent. Let v1 = v, v2 = T (v),
and extend to a basis v1 , . . . , vn of V . Let W = span{v1 , v3 , . . . , vn }. Then dim(W ) = n − 1, so
T (W ) ⊆ W . But T (v) = v2 6∈ W , a contradiction. It follows that for every nonzero vector v, there
exists a scalar λ such that T (v) = λv. A priori, the scalar λ depends on v, but we will prove it does
not. Let vP1 , . . . , vn be a basis of V . Then for each k, there exists a scalar λk such that T (vk ) = λk .
Consider k vk . On one hand, there exists some scalar λ such that
!
X X X
T vk = λ vk = λvk .
k k k
On the other hand, !
X X X
T vk = T (vk ) = λ k vk .
k k k
Then X
0= (λ − λk )vk .
k
Since the vk are linearly independent, we must have λ − λk = 0. So P λk = λ for all k. Then
T (vk ) = λvk for all k. For any v ∈ V , there are scalars ak such that v = k ak vk . Then
!
X X
T (v) = T ak vk = ak λvk = λv.
k k
So T = λid.

5. S19

Problem 49 (1). Consider C([0, 1]), and let X be the subset of C([0, 1]) of all 1 lipschitz
functions f such that f (0) = 0. Show X is connected and complete.

For completeness, it is sufficient to show this space is closed, since the given space is already com-
plete. If fn ∈ X converge uniformly to some f ∈ C([0, 1]), we see that f (0) = 0. Likewise, for any
x, y ∈ [0, 1], we have |f (x) − f (y)| ≤ |f (x) − fn (x)| + |fn (x) − fn (y)| + |fn (y) − f (y)| by triangle
inequality. For n large enough, the first and last term are both less than ε/2 by uniform convergence.
The middle term is less than or equal to |x − y| for all n. Then |f (x) − f (y)| ≤ |x − y| + ε for all
ε > 0, and so |f (x) − f (y)| ≤ |x − y|. Thus, f ∈ X, so X is closed, and so complete.

23
For connectedness, we will show convexity. For any t ∈ [0, 1], and f, g ∈ X, one checks quickly
that tf + (1 − t)g ∈ X.

Problem 50 (2). Let a, b ∈ R be two real numbers and consider a sequence an defined
recursively:
1
a0 = a, a1 = b, an = 2 (an−1 + an−2 ) when n ≥ 2. Prove that limn→∞ exists and
compute its value.

First, observe that both a0 and a1 can be written in the form c ∗ a + (1 − c) ∗ b for some c ∈ [0, 1].
Suppose that this is true for some an−2 and an−1 . Then an = 12 (cn−1 a + cn−2 a + (1 − cn−1 )b +
(1 − cn−2 )b) = da + (1 − d)b where d = cn−2 +c 2
n−1
, and note d ∈ [0, 1] as well. So, write each
ai as ci ∗ a + (1 − ci ) ∗ b for some ci ∈ [0, 1]. I will show that the ai have limit a+2b 3 . Note
a+2b 1 1
ai − 3 = (ci − 3 )a + ( 3 − ci )b. It is sufficient then to show that ci → 1/3. I will show that for
i−1 +1 i−1 −1
i even and at least 2, ci = 23∗2i−1 , and for i odd and at least 3, ci = 23∗2i−1 . From here, it is easy
to see that limci = 1/3. It is true for c2 = 1/2 and c3 = 1/4. If n is even and at least 4, then
n−1 +1
cn = (cn−1 + cn−2 )/2 = 23∗2n−1 , as desired. Similar can be shown for n odd and at least 5.

Problem 51 (3). Let a, b ∈ R with a ≤ b and let f : [a, b] → R be a Riemann integrable

function. Assume there is δ > 0 such that f (x) ≥ δ for all x ∈ [a.b]. Show that 1/f is
Riemann integrable.

Fix ε > 0. Since f is Riemann integrable, let Pn be a partition P of [a, b] such that U (f, Pn ) −
L(f, Pn ) < ε/δ 2 . On Pn , we have that U (1/f, Pn )−L(1/f, Pn ) = [xi ,xi+1 ]∈Pn (xi+1 −xi )supx,y∈[xi ,xi+1 ] |1/f (x)−
1/f (y)| = [xi ,xi+1 ]∈Pn (xi+1 −xi )supx,y∈[xi ,xi+1 ] | ff(y)−f (x) f (y)−f (x)
P P
(y)∗f (x) | ≤ [xi ,xi+1 ]∈Pn (xi+1 −xi )supx,y∈[xi ,xi+1 ] | δ2
|<
2
δ ∗ (U (f, Pn ) − L(f, Pn )) = ε.

Problem 52 (4). Let a, b ∈ R with a < b and let f, g : [a, b] → R be continuous functions
differentiable on (a, b). Show there exists ζ ∈ (a, b) such that (f (b) − f (a))g 0 (ζ) = (g(b) −
g(a))f 0 (ζ)

Consider the function h(x) = g(x) ∗ (f (b) − f (a)) − f (x) ∗ (g(b) − g(a)). As this is a difference of
functions that are continuous on [a, b] and differentiable on (a, b), it has both of these properties as
well. We see that h(b) = g(a)f (b)−g(b)f (a). Likewise, h(a) = g(a)f (b)−g(b)f (a) = h(b). By Rolle’s
theorem, there exists a point ζ ∈ (a, b) such that h0 (ζ) = g 0 (ζ)(f (b) − f (a)) − f 0 (ζ)(g(b) − g(a)) = 0.
This is equivalent to the given condition.

Problem 53 (5). Show that each metric space can be embedded isometrically into a Banach
space.

Let Cb (D) be the space of bounded continuous functions on X into R with the standard supremum
norm, and note that this space is complete. If D is empty we are trivially done, so if not, fix x0 ∈ D,
and define S(x) = fx (y) = d(x, y) − d(x0 , y). This is continuous since the metric is, and is bounded
by reverse triangle inequality. For isometry, (S(x) − S(z))(y) = d(x, y) − d(z, y). By reverse triangle
inequality, this is less than or equal to d(x, z), and taking y = z gives that |S(x) − S(y)| ≥ d(x, z),
and so |S(x) − S(z)| = d(x, z), as desired.

24
Note: This is in the week 2 notes. I have no idea how someone would just come up with this
on their own, and I am 90 percent positive it is only in the notes because Sylvester saw this ques-
tion and got worried that there would be more. It seems like the sort of question that a professor just
thought was a neat fact and wasn’t actually a suitable question. I wouldn’t worry about whether you
could solve this from scratch, but rather, keep in mind that it could be similar to future questions.

x −x/n
Problem 54 (6). For n ≥ 1, letR fn : R+ → R, fn (x) = n2
e . Show that the fn converge
∞
uniformly to zero, but limn→∞ 0 fn (x) = 1.

Note each fn is nonnegative. For each fn , we compute fn0 (x) = n12 e−x/n + −x n3
e−x/n = n−x
n3
e−x/n .
00 −1 −x/n n−x −x/n
Therefore, fn has a critical point iff x = n. We also check that fn (x) = n3 e − n4 e =
x−2n −x/n
n4
e , which is negative when x = n, and so this is a local maximum. We check fn (0) = 0
and fn (kn) for integer k is nekk . The root test shows that this sequence converges as k → ∞, and
since fn0 (x) is negative for x > n, we see that limx→∞ fn (x) = 0. Therefore, the global maximum of
1
fn (x) is fn (n) = ne .

Now pick ε > 0. For all n satisfying 1/n < e ∗ ε, and all x positive, we have |fn (x) − 0(x)| ≤ ε, and
so we have uniform convergence. R∞
We
R∞ calculate
−u
the given integrals via integration
−u
by parts. First, a u substitution gives
−u
that
x
R0x fn−u
(x) =
0 u∗e du. Taking u2 = u, dv = e , this expression is equalR to limx→∞ − ue |0 − 0 −e =
∞
limx→∞ − xe−x + 1 = 1 by our above calculation. So, limn→∞ 0 fn (x) = limn→∞ 1 = 1.
 
1 0 3
Problem 55 (7). For A = 3 −5 3, write A−1 as a polynomial in A with real coeffi-
1 −1 2
cients.

We compute the characteristic polynomial of A to be x3 + 2x2 + 13x − 1 (I’m not typing out this
calculation). By Cayley-Hamilton theorem, A3 + 2A2 + 13A − 1 = 0. Applying A−1 to both sides,
A2 + 2A + 13 = A−1 .

Problem 56 (8). Let V be the vector space of all 2 × 2 matrices with real entries. Let W
be the subspace generaged by all matrices of the form AB − BA for A, B ∈ V . What is the
dimension of W ? Justify your answer

We note that for any matrix of the form AB − BA, tr(AB − BA) = tr(AB) − tr(BA) = tr(AB) −
tr(AB) = 0. Any matrix in the span of such matrices must also be traceless, and so this subspace
is a supspace
 of the kernel
  of the trace operator. We know that the trace operator has kernel 3,
1 0 0 1 0 0
since , , are all in this space, are linearly independent, but I is not in this
0 −1 0 0 1 0
space, so 3 ≤ dim(ker(tr)) < 4 i.e. dim(ker(tr)) = 3. In particular, dim(W ) ≤ 3. Wewill exhibit 
0 1
three linearly independent matrices in W , and so dim(W ) = 3. First, taking A = and
0 0
     
1 0 0 0 0 0
B = gives that A ∈ W . Now, taking C = and D = gives that D ∈ W .
0 0 0 1 1 0
 
1 −1
Finally, taking M = A + D and N = B + D gives that ∈ W . We see that this matrix,
1 −1
and A, D are linearly independent, so dim(W ) = 3.
25
 Problem 57 (9). Let V be the vector space over C of all complex polynomials of degree at
most 10. Let D : V → V be the differentiation operator. find all eigenvalues and eigenvectors
of eD on V .
P10
Observe that D is nilpotent with D11 = 0 but D10 6= 0. We note that eD = i
i=0 D /i!. We
D
see that e (c) = c where c is constant (and hence, eigenvalue 1). I will show that in fact the
nonzero constant functions (a subspace of degree 1 over C) are the only eigenvectors. Let f be any
polynomial in V with degree m more than 1. Then eD (f ) = f + m−1 i
P
i=1 D (f )/i!, and note that
this sum term will be a polynomial of degree exactly m − 1. The leading coefficient of eD (f ) is the
leading coefficient of f , and hence, if f is an eigenvector, it must have eigenvalue 1. This implies
Pm−1
that i=1 Di (f )/i! = 0, which is impossible, since each Di (f ) has distinct degree.

 A be
Problem 58 (10). Let  an n × n complex diagonalizable matrix and I the n × n identity
A I
matrix. Show M = is not diagonalizable.
0 A

We use that a matrix is diagonalizable over  Cn iff itsn−1

minimal
 polynomial is squarefree. For any
A nA
positive integer n, we compute that M n = . Then, for any polynomial ni=0 ai ∗ xi
P
0 A n
Pn
ai ∗ i ∗ An−1
 
p(A) i=0
with ai ∈ C, we have that p(M ) = . We have p(M ) is zero iff p(A) is
0 p(A)
zero and the upper right component is zero. But p(A) is zero iff the characteristic polynomial of A,
mA (x), divides p(x). We also note that the upper right corner is exactly equal to p0 (A). This is zero
exactly when mA (x)|p0 (A). Since mA is squarefree, mA divides p and p0 if and only if m2 |p. This
implies that the minimal polynomial of M must be divisible by m2A , and so cannot be squarefree.
M is not diagonalizable.

Problem 59 (11). Let A be an n × n complex matrix. Prove that rank(A) = rank(A2 ) iff
limλ→0 (A + λI)−1 A exists.

Problem 60 (12). LetPA = (ai,j ) be an n × n real matrix whose diagonal entries are all at
least 1, and such that i6=j a2i,j < 1. Prove A−1 exists.

Write A = D + B where D is the diagonal component of A and B is the off diagonal component of
B. We note that D is positive definite since its eigenvalues are along the diagonal and all positive.
We note also thatqthe condition on the remaining terms implies that ||B|| < 1, since (B T Bv)i =
T T
P P 2 P 2
j (B B)i,j vj ≤ j bi,j |v| by Cauchy Schwarz, and therefore, ||B B|| < |v| i,j bi,j < |v|. This
implies that |hBv, Bvi| = | < B T Bv, vi ≤ |v| ∗ |B T Bv| < |v|2 , again by Cauchy-Schwarz.

Now, suppose Av = Dv + Bv = 0 for some v. Then Dv = −Bv. But if v 6= 0, we have |Dv| > |v|,
but | − Bv| < |v|, a contradiction. Thus, v = 0, and A is invertible.

Note: This was an old Berkeley problem (week 4 Thursday), but I think it is the first time it
has appeared on a test for UCLA. No idea what to think about that. Maybe keep in mind that old
Berkeley probs could appear?
26
 6. F19

Problem 61 (1). Let A be an invertible n × n matrix with real entries and let e1 be the
first standard basis vector. For each λ ∈ R, define Aλ (x) = Ax + λhe1 , xie1 . Show that Aλ
is invertible iff 1 + λhe1 , A−1 e1 i =
6 0.

Consider the set of vectors A−1 ei . Since A−1 has full rank and the ei are a basis, the A−1 ei are a
basis as well. For each A−1 ei , we see that Aλ (A−1 ei ) = ei + λhe1 , A−1 ei ie1 . From here, we see that
the Aλ (A−1 ei ) for i ≥ 2 are linearly independent, and Aλ (A−1 e1 ) is in their span iff Aλ (A−1 ei ) = 0.
Therefore, Aλ is invertible iff Aλ (A−1 e1 ) 6= 0 iff (1 + λhe1 , A−1 e1 i)e1 6= 0 iff the given condition is
true.

Problem 62 (2). Find a real symmetric matrix A so that A2 + A = (not typing it out)
 
1 0 0 1
0 3/2 1/2 0
We can see that 
0 1/2 3/2 0 works.


1 0 0 1

To find this matrix, the real symmetry condition suggests looking at eigenvalues and eigenvectors
of A. Such an eigenvector would also be an eigenvector of A2 + A with eigenvalue λ2 + λ. Finding
eigenvectors of A2 + A isn’t difficult. From here, you can compute what Aei for each ei is, and this
gives the above matrix for a certain choice of the eigenvalues.

Problem 63 (4). Let V be a vector space and 1 ≤ n < dim(V ) be an integer. Let {Vi } be
a collection of n dimensional subspaces of V with the property that dim(Vi ∩ Vj ) = n − 1 for
every i 6= j. Show that at least one of the following holds:
All Vi share a common n − 1 dimensional subspace.
There is an n + 1 dimensional subspace of V containing all Vi .

Problem 64 (5). Show that an x × n matrix A with real entries obeys limk→∞ ||Ak || = 0 iff
all (possibly complex) eigenvalues have modulus strictly less than 1.

Note that the operator norm of any real matrix A, when viewed as an operator of a real vector space,
is the same as the operator norm when viewed as an operator over a complex space. This is because
the singular values remain the same, and the operator norm is the largest singular value modulus.
Therefore, it is sufficient to show this equivalence holds when A is viewed as an operator over a
complex vector space. If A has any eigenvalue of modulus ≥ 1, then fix a unit length eigenvector v
corresponding to this λ. Then Ak v = λk v, which has norm at least one. The operator norm is the
supremum of such vector norms, and so must be at least one for all k. This implies that the limit
cannot go to zero.

Now, if all eigenvalues have modulus less than one, consider the Jordan canonical form of A. One
checks that its entries are vanishing, and so the operator norm of J k → 0. Then ||Ak || ≤ ||P || ∗
||J k || ∗ ||P −1 || → 0, where P is the invertible matrix satisfying P JP −1 = A.
27
 Problem 65 (6). Let Mn be the vector space of n × n matrices with real entries. Given
B ∈ Mn , we define a linear transformation Lb : Mn → Mn by LB (A) = B T AB.
Prove LB is invertible iff B is.

Find rank(LB ) when B is diagonal.

Find rank(LB ) in general.

For the first part, if B is invertible, then so is B −T , and we see that LB (A)−1 = (B T )−1 AB −1 . If
B is not invertible, then B T is not invertible. This implies that B T v = 0 for some nonzero v. Take
A to be the matrix each of whose columns is v. Then LB (A) = 0, so LB is not invertible.

Let Ei,j be the standard basis of Mn i.e. Ei,j is only nonzero in row i column j, where it is one.
Then rank(LB ) is precisely the number of linearly independent LB (Ei,j ). Suppose that Bi,i = λi .
Then we compute that LB (Ei,j ) = λi ∗ λj ∗ Ei,j . The set of nonzero LB (Ei,j ) is linearly independent
since the Ei,j are linearly independent, and so LB (Ei,j ) is nonzero iff λi and λj are both nonzero.
Note that rank(B) is exactly the number of nonzero λi , and so there are rank(B)2 eigenmatrices
not corresponding to the eigenvalue zero. Thus, rank(LB ) = rank(B)2 .

Problem 66 (7). Show that the equation x = cos(x) has exactly one solution on [0, 1]

x = cos(x) iff f (x) = x − cos(x) = 0, so we show that f (x) has exactly one root on [0, 1]. Note
that f is continuous since it is a difference of continuous functions. f (0) = 0 − cos(0) = −1.
f (1) = 1 − cos(1). Since 1 ∈ [0, π/2], we have that 0 < cos(1) < 1, and so f (1) > 0. Then
0 ∈ (f (0), f (1)). By I.V.T., there exists a c ∈ (0, 1) such that f (c) = 0. For uniqueness, observe
that f 0 (x) = 1 + sin(x). For x ∈ [0, 1], sin(x) > −1, and so f 0 (x) is positive. This means that f (x)
is strictly increasing on [0, 1], and hence injective.
P h
Problem 67 (8). Show that sup0<h≤1 n∈Z 1+n2 h2 <∞

Write 1+nh2 h2 = h ∗ f (nh) where f (x) = 1+x 1

2 . Note that f is an even function, and therefore,

sum is then equal to h ∗ f (0) + 2 ∗ ∞

P
h ∗ f (nh) = h ∗ f ((−n) ∗ h). The given
P n=1 h ∗ f (nh). It
is sufficient to show that sup0<h≤1 n∈Z+ h ∗ f (nh) < ∞ since h ∗ f (0) = h < 1. Fix N > 0.
Then N 1
P
n=1 h ∗ f (nh) = L(f, Pn ) where f = 1+x2 and Pn partitions [0, N h] into N uniformly sized
R Nh
blocks, where we have used that f is decreasing on the positive reals. But L(f, Pn ) ≤ 0 f (x)dx =
arctan(N h) − arctan(0) < π/2. Since N
P
n=1 h ∗ f (nh) is increasing in N and is bounded above
by π/2, by M CT , it converges to some positive real less than or equal to π/2. Since this was
independent of h, the supremum over all such h of this sum will also be no greater than π/2.

Problem 68 (9). (1) Show that the relation (2 + x + y) = z 2 + ex + ey determines z as

a smooth function of (x, y) in some neighborhood of the origin.
(2) Show that (0, 0) is a critical point of z(x, y) and determine its nature (minimum,
maximum, etc.)

28
 Problem 69 (11). Let X denote the set of non-decreasing functions f : [0, 1] → [0, 1] with
the supnorm metric.
Prove (X, d) is complete.
Prove it is not compact.

Let fn (x) be a Cauchy sequence of functions in X. Note that this implies that fn (x) is a Cauchy
sequence of real numbers for each x ∈ [0, 1], and by completeness of R, fn (x) converges pointwise
to some function f . It remains to show that f ∈ X. We see that for any x, f (x) = limn fn (x)
by pointwise convergence, and so f (x) ∈ [0, 1]. Now, fix 0 ≤ x < y ≤ 1. Then f (y) − f (x) =
limn fn (y) − fn (x). Since fn (y) − fn (x) is nonnegative, so to is the limit, so we have f is nonde-
creasing.

Since compactness implies separability in metric spaces, it is sufficient to show that X is not
separable. Consider the functions fr (x) = 0 if x ≤ r and 1 if x > r, where r ∈ [0, 1]. Then
sup|fr1 (x) − fr2 (x)| ≤ 1, and if r1 < r2 , then fr1 (r1 ) − fr2 (r1 ) = −1, and so sup|fr1 (x) − fr2 (x)| = 1
for all distinct r1 , r2 . Note that there are uncountably many fr since [0, 1] is uncountable. There
can be no countable dense subset of X, since this would imply that the balls B(fr1 , 1/2) all contain
a distinct element of our dense subset, which contradicts the countability of such a set. X is not
separable, and hence, not compact.

Problem 70 (12). Let l∞ (Z) denote the space of bounded functions x : Z → R together with
the metric d(x, y) = supn∈Z |x(n) − y(n)|. Show that a function f : l∞ (Z) → R is continuous
iff its restriction to any compact subset of l∞ (Z) is continuous.

For the forward direction, assume f is continuous, and let K be a compact subset of L∞ (Z). For
any point x ∈ K, f is continuous at x since any sequence of K elements xn converging to x is a
sequence of L∞ (Z) elements converging to x. By continuity of f in l∞ (Z), f (xn ) → f (x), and so f
is continuous on K.

For the reverse direction, assume f is continuous on any compact K ⊂ l∞ (Z). Let xn be a se-
quence of l∞ (Z) elements converging to x ∈ l∞ (Z). I will show that K = {xn } ∪ {x} is compact.
For any infinite sequence yn in K, either the yn have only finitely many distinct values, in which case
some subsequence is constant, or the yn have infinitely many distinct values, in which case some sub-
sequence must converge to x. K is sequentially compact, and hence, compact. By assumption, f is
continuous on K, and so limn→∞ f (xn ) = f (x). f is thus sequentially continuous, and so continuous.

Note that the ambient space did not matter at all. The conclusion is actually generally true for
metric spaces.

7. S20
8. F20

Problem 71 (1). Let M be an n × n matrix with rational entries such that M 2 = 2I.
Prove n is even.
Give an example of such matrix M for n = 2.
√
The minimal polynomial of M must divide x2 − 2. But M 6= ± 2I, and so x2 − 2 is the minimal
polynomial of M . Note also that the characteristic polynomial must divide some power of the
29
minimal polynomial since λ is a (complex) root of the minimal polynomial iff it is a root of the
characteristic polynomial. This implies that the characteristic polynomial is exactly (x2 − 2)i for
some positive integer i, and hence, n must be even since this polynomial is degree 2i.

 
1 1
We see that the matrix works. (To find this, just write out the general expression for a
1 −1
2 by 2 matrix and solve for possible entries)

Problem 72 (2). Let A be an orthogonal n × n matrix.

Prove A3 is orthogonal
Prove A + 12 I is invertible.

A3 is orthogonal iff (A3 )T A3 = I. But (A3 )T = (AT )3 . We see that AT ∗ AT ∗ AT ∗ A ∗ A ∗ A = I

by orthogonality of A.

A + 21 I is invertible iff 1/2 is not an eigenvlaue of A. I will show that orthogonal matrices can
only have eigenvalues of modulus 1. Fix an eigenvector of A, v, corresponding to eigenvalue λ.
Then λhv, vi = hAv, vi = hv, AT vi = hv, A−1 vi. Since A−1 Av = v, we have that v is an eigen-
vector of A−1 with eigenvalue 1/λ. (λ must be nonzero or else A is not invertible at all). Then
hv, A−1 vi = λ1 hv, vi. We have that λ2 = 1, as desired.

Problem 73 (3). Let M be a complex 4 × 4 matrix such that M 6 = M 4 = 2M 3 − M 2 .

Describe all possible J.C.F. of M .

The condition M 4 − 2M 3 + M 2 = 0 gives that the minimal polynomial of M divides x2 (x − 1)2 , and
so the only possible eigenvalues of M are 0 and 1. Note also that the power of x − λ in the minimal
polynomial is equal to the size of the largest Jordan Block for that eigenvalue. In particular, there
can be no Jordan Blocks of size 3 or more for either possible eigenvalue. I will now show that there
can be no block for eigenvalue 1 of size greater than 1. Let B be such a block. Note that (B n )i,j = n
for all n, and so M 6 6= M 4 . All blocks for eigenvalue zero are nilpotent with order no greater than
4, and so B 6 = B 4 for such blocks. This leaves us with only a few possible matrices, as described
below.
(1) 4 zero’s on the diagonal, 4 blocks
(2) 4 zeros on the diagonal, 1 block of size 2, 2 blocks of size 1
(3) 4 zero’s on the diagonal, 2 blocks of size 2
(4) 3 zeros on the diagonal and one 1, all blocks of size 1.
(5) 3 zeros on the diagonal, one block of size 2 for the zeros
(6) 2 zeros and 2 ones on the diagonal, all blocks size 1.
(7) 2 zeros and 2 ones on the diagonal, block of zeros is of size 2.
(8) 1 zero on the diagonal and 3 ones, all blocks size 1.
(9) 4 ones on the diagonal, all blocks size 1.

Problem 74 (4). Let A be a 2 × 2 real matrix

  eigenvalues 2 and −1. Consider the set
with
A B
X of 2 × 2 real matrices C such that C = is diagonalizable over C. Prove X is a 2
0 A
dimensional subspace of Mn (R)

30
A matrix of this form is diagonalizable iff its minimal polynomial is squarefree. We compute that
the characteristic polynomial of C is µA (x)2 , where µA (x) is the characteristic polynomial of A.
But since A is a 2 × 2 matrix with eigenvalues 2 and −1, µA (x) is exactly (x − 2)(x + 1). We see
that C − 2I and C + I are nonzero since A − 2I and A + I are nonzero,
 and so C is diagonalizable

0 AB + BA − B
iff (C − 2I)(C + I) = 0. We compute that (C − 2I)(C + 2I) = , and so C
0 0
is diagonalizable iff B = AB + BA. Write A = P DP −1 where D is the diagonal matrix with 2
and −1 on the diagonal. Write F = P −1 BP . We see that B = AB + BA iff F = DF + F D, and
each  of such F C yields a unique
choice   such  B,and each such B yields a unique such C. Write
w x w x 4w x
F = . Then we must have = . We see that w = z = 0 is required,
y z y z y −2z
and any choice of x and y are valid. This is then indeed a dimension 2 subspace.

Problem 75 (5). Let K = Fp be a finite field with p elements where p is prime. Let V = K 9
be a vector space, and let W ⊂ V be a subspace of V such that dim(W ) = 5. Compute the
number of subspaces U ⊂ V such that dim(U ) = 6 and dim(W ∩ U ) = 3.

Problem 76 (6). Let v1 ,....vk ∈ Rn satisfy hvi , vj i < 0 for all 1 ≤ i < j ≤ k. Prove k ≤ n+1.

We argue by induction on n. The case for n = 1 is trivial. Suppose that the claim is true for all
n ≤ N for some fixed N . Suppose towards contradiction that in RN +1 , there exists N + 2 vectors
v1 through v2 such that hvi , vj i < 0 for i < j. Consider the N + 1 vectors wi = vi − hvi , vN +2 ivN +2 .
We see that these vectors inhabit vN ⊥ , which is an N dimensional subspace of RN +1 . However,
+2
for i < j, we see hwi , wj i = hvi , vj i − 2hvi , vN +2 ihvN +2 , vj i + hvi , vN +2 ihvj , vN +2 ihvN +2 , vN +2 i. The
first two terms are negative by assumption, and the last term can be made to have arbitrarily small
absolute value by rescaling vN +2 . If vN +2 is rescaled such that the (finitely many) hwi , wj i < 0,
then we have a contradiction, because vN ⊥ N
+2 is isomorphic to R , but we have found N + 1 vectors
with mutually negative inner product in this space.

Problem 77 (9). Let f (x) be a real and continuous function on [0, 1]. Show that limn→∞ (n+
R1
1) 0 xn f (x)dx = f (1).

Fix ε > 0. Since f is continuous on the compact set [0, 1], it is bounded. In particular, there exists
positive M such that f (x) < M for all x and f (x) > −M for all x. Write ζn = (ε/M )1/(n+1) . For
Rζ R1
any n, observe that the given integral is equal to (n + 1) 0 n xn f (x) + (n + 1) ζn xn f (x). The first
integral is less than ε and greater than −ε by using the boundedness of f . Since xn isR positive on
1
(ζn , 1) and xn and f are both continuous, by Mean Value theorem for integrals, (n + 1) ζn xn f (x) =
R1 n
(n+1)f (cn ) ζn x = f (cn )∗(1−ε/M ) for some cn ∈ (ζn , 1). Taking limits, since cn → 1 and since f
is continuous, we have that the given limit is in the interval (f (1)∗(1−ε/M )−ε, f (1)∗(1−ε/M )+ε)
for any ε > 0. This implies that the given limit is indeed f (1).

Problem 78 (10). Let fn be a uniformly bounded equicontinuous sequence of real valued

functions on a compact metric space X with distance function d. Define the functions gn
from X → R by gn = max{f1 (x), ...fn (x)}. Show that gn converges uniformly.

Note that for any x, gn+1 (x) ≥ gn (x). Note also that the gn converge pointwise at any x since the
sequence gn (x) is a bounded (by uniform boundedness of fn ) monotone sequence of real numbers.
31
By Dini’s theorem, since X is compact, it is sufficient to show that the limit function g is continuous.

First, we will show that g(x) = sup(f1 (x), f2 (x)...). The supremum is defined since fn are uni-
formly bounded. Note that g(x) ≥ fn (x) for all n, and so g(x) ≥ sup(f1 (x), f2 (x)...). Since each
gn (x) ≤ sup(f1 (x)...), we have that the limit g(x) ≤ sup(f1 (x)...) as well.

We now show continuity of g. Fix x, y ∈ X. Then g(x) − g(y) = supi (fi (x) − g(y)) ≤ supi (fi (x) −
fi (y)). Fix ε > 0. By equicontinuity of the fn , there exists δ > 0 such that |fn (x) − fn (y)| ≤ ε
when |x − y| ≤ δ. Then for this delta, we have |g(x) − g(y)| ≤ ε as well, implying continuity of g.

Problem 79 (11). For each n ∈ N, let fn : R → R, and assume that these functions are
uniformly bounded. Let X be a countable subset of R. Show that the sequence fn has a
subsequence that converges pointwise for all x ∈ X.

Enumerate the X elements as x1 , x2 .... Consider the sequence f1 (x1 ), f2 (x1 ), f3 (x1 )... This is a
bounded sequence of real numbers since the fn are uniformly bounded. Therefore, by compactness,
some subsequence of the f1 (x1 ), f2 (x1 )... converges, i.e., some subsequence of the fn converge
pointwise at x1 . Relabel this subsequence as f1,1 , f1,2 , f1,3 .... Now suppose that we have a sequence
gn that converges for xi for all 1 ≤ i ≤ k for some k. Then we may apply the above argument to
find a subsequence of the gn that converge at xk+1 as well. Applying this argument to our sequence
f1,1 , f1,2 ... repeatedly, we obtain nested subsequences, where fk,1 , fk,2 ... converges for all xi with
i ≤ k.
Now, define fk = fk,k . Fix xi ∈ X. For k ≥ i, we see that the fk are a subsequence of fi,1 , fi,2 ...,
and this sequence was defined to converge at xi . Thus, the fk are a subsequence of the fn that
converge at all x ∈ X4.

Problem 80 (12). Let X be an open convex subset of Rn . Let f : X → R be a differentiable

function.
Show that for any x, y ∈ X, there is a point z lying on the line segment from x to y for
which f (y) − f (x) = ∇f (z) ∗ (y − x).

Use part a to show that if the partial derivatives of f are bounded, then f is uniformly
continuous on X.

Fix x, y ∈ X. Define g(t) : [0, 1] → R where g(t) = f ((1−t)∗x+t∗y). Note that g is continuous since
it is a composition of continuous functions, and since X is convex, (1 − t) ∗ x + t ∗ y ∈ X for all t in
this range. By Mean Value Theorem, there exists a t∗ ∈ [0, 1] such that g 0 (t∗ ) ∗ (1 − 0) = f (1) − f (0).
Now, observe that f (1) − f (0) = f (y) − f (x). By chain rule, g 0 (t∗ ) = ∇f ((1 − t∗ )x + t∗ y) ∗ (y − x).
Taking z = (1 − t∗ )x + t∗ y works.

Suppose that the partial derivatives of f are bounded. Then |∇f | ≤ M for some positive M .
Fix ε > 0. Then if |y − x| ≤ ε/M , we have that |f (y) − f (x)| ≤ |∇f | ∗ |(y − x)| (Cauchy Shchwarz)
< ε, as desired.

32