You are on page 1of 22

PRINCIPLES OF MATHEMATICAL ANALYSIS.

WALTER RUDIN

Disclaimer: these solutions were typed at warp speed, often with little or
no preparation. And very little proofreading. Consequently, there are bound to be
loads of mistakes. Consider it part of the challenge of the course to find these errors.
(And when you do find some, please let me know so I can fix them!)

7. Sequences and Series of Functions


1. Prove that every uniformly convergent sequence of bounded functions is uni-
formly bounded.
{fn } is uniformly
¯ Cauchy, so |f¯n (x) − fm (x)| < 1, for n, m bigger than some
¯ ¯
large N . Then ¯|fn (x)| − |fN (x)|¯ ≤ |fn (x) − fN (x)| < 1, so
|fn (x)| < |fN (x)| + 1.
Let M be a bound for fN . Then
|fn (x)| < M + 1,
so {fn }∞
n=N is uniformly bounded by M + 1. Define

K := max{sup |f1 |, sup |f2 |, . . . , sup |fN −1 |, M + 1}.


Then {fn }∞
n=1 is uniformly bounded by K.

2. Show that {fn }, {gn } converge uniformly on E implies {fn + gn } converges


uniformly on E. If, in addition, {fn }, {gn } are sequences of bounded functions,
prove that {fn gn } converges uniformly.

sup{|(fn + gn )(x) − (f + g)(x)|} = sup{|fn (x) − f (x) + gn (x) − g(x)|}


x∈E x∈E
≤ sup{|fn (x) − f (x)| + |gn (x) − g(x)|}
x∈E
≤ sup |fn (x) − f (x)| + sup |gn (x) − g(x)|
x∈E x∈E
n→∞
−−−−−→ 0 + 0.
When {fn }, {gn } are bounded sequences,
sup{|(fn gn )(x) − (f g)(x)|}
x∈E

March 23, 2006. Solutions by Erin P. J. Pearse.

1
Principles of Mathematical Analysis — Walter Rudin
= sup{|fn gn (x) − f gn (x) + f gn (x) − f g(x)|}
x∈E
≤ sup{|fn (x) − f (x)||gn (x)| + |f (x)||gn (x) − g(x)|}
x∈E
≤ sup |fn (x) − f (x)||gn (x)| + sup |f (x)||gn (x) − g(x)|
x∈E x∈E
≤ sup |fn (x) − f (x)| · Mg + sup Mf · |gn (x) − g(x)|.
x∈E x∈E
We can introduce the uniform bounds Mf and Mg by problem 1 and the
additional hypothesis. Then it is clear that the last line goes to 0 as n → ∞.

3. Construct sequences {fn }, {gn } which converge uniformly on a set E, but such
that {fn gn } does not converge uniformly.
Work on I = (0, 1). Let fn (x) = x1 + nx and gn (x) = − 1−x 1
− 1−x
n
so that
gn is the horizontal reflection of fn , translated 1 to the right. (Graph them to
see it.)
1 x n→∞ 1
fn (x) = x
+ n
−−−−−→ x
= f (x) sup |fn − f | = sup nx = 1
n
→0
0≤x≤1 x∈I
1 1−x n→∞ 1
gn (x) = − 1−x − n
−−−−−→ − 1−x = g(x) sup |gn − g| = sup 1−x
n
= 1
n
→0
0≤x≤1 x∈I

(n + (x − 1)2 )(n + x2 ) n→∞ 1


fn gn (x) = − 2
−−−−−→ = f g(x)
n x(x − 1) x − x2
¯ 2 ¯
¯ x (x − 1)2 + n + 2nx(x − 1) ¯
sup |fn gn − f g| = sup ¯¯ ¯ = ∞, for any n.
0≤x≤1 x∈I n 2 x(x − 1) ¯
To see the sup is infinite, check x = 0, 1.

4. Consider the sum ∞


X 1
f (x) = .
n=1
1 + n2 x
For what x does the series converge absolutely? On what intervals does it
converge uniformly? On what intervals does it fail to converge uniformly? Is
f bounded?
For xk = − k12 , we get

X 1
f (xk ) = ¡ n ¢2 .
n=1 1− k
The k term of the sum is undefined, so f is undefined at xk = − k12 , k =
th

1, 2, . . .
To be completed.

2
Solutions by Erin P. J. Pearse

5. Define a sequence of functions by


 ¡ ¢
 1
0 x < n+1
¡ 1
,
¢
fn (x) = sin2 πx < x ≤ 1
,

0 ¡ n+1
1
¢ n

n
<x .
Show that the series {fPn } converges to a continuous function, but not uni-
formly. Use the series fn to show that absolute convergence, even for all x,
does not imply uniform convergence.
For x ≤ 0, fn (x) = 0 for every n, so lim fn (x) = 0. For x > 0,
£ ¤
n > N := x1 =⇒ n1 < x =⇒ fn (x) = 0.
pw
Thus fn (x) −−−→ f (x) := 0 for x ∈ R.
To see that the convergence is not uniform, consider
¡ ¢¡ ¢¡ ¢
fn0 (x) = 2 sin πx cos πx − xπ2 .
fn0 (x) = 0 when
• sin πx = 0, in which case x = k1 for some k ∈ Z, or
• cos πx = 0, in which case x = 2k+1
2
for some k ∈ Z.
For each fn , only a few of these values occur where fn is not defined to be 0,
so checking these values of x,
¡ ¢
fn n1 = sin(nπ) = 0
¡ 1 ¢
fn n+1 = sin((n + 1)π) = 0
¡ 2 ¢ ¡ ¢
fn 2n+1 = sin 2n+1 2
π = 1.
So Mn = supxP {|fn (x) − f (x)|} = 1 for each n, and clearly Mn → 1 6= 0.
The series ∞ n=1 fn (x) converges absolutely for all x ∈ R: for any fixed x,
there is only one
P∞ nonzero term in the sum.
The series n=1 fn (x) does not converge uniformly: check partial sums.
¯ ¯
¯X N N
X +1 ¯
¯ ¯
sup ¯ fn (x) − fn (x)¯ = sup |fN +1 (x)| = 1,
x ¯ ¯ x
n=1 n=1

so the sequence of partial sums is not Cauchy (in the topology of uniform
convergence), hence cannot converge.

6. Prove that the series ∞


X x2 + n
(−1)n
n=1
n2
converges uniformly in every bounded interval, but does not converge abso-
lutely for any value of x.
To see uniform convergence on a bounded interval,
n¯X∞ ¯o
¯ n x2 +n ¯
sup {|fN (x) − f (x)|} = sup ¯ (−1) n2 ¯
a<x<b a<x<b n=N +1

3
Principles of Mathematical Analysis — Walter Rudin
n¯X∞ ¯o
¯ 2 ¯
= ¯ (−1)n c n+n
2 ¯ (7.1)
n=N +1
where c := max{|a|, |b|}. We will use the alternating series test to show
P∞ n c2 +n
n=1 (−1) n2
converges. From this, it will follow that (7.1) goes to 0 as
N → ∞.
2
(i) For all x ∈ R, x n+n
2 > 0. So the sum alternates.
x2 +n 1
(ii) For fixed x, lim n2 = lim 2n = 0. (L’Hôp)
(iii) To check monotonicity, prove
x2 + n + 1 x2 + n

(n + 1)2 n2
by cross-multiplying.

x
7. For n = 1, 2, 3, . . . , and x ∈ R, put fn (x) = 1+nx2 . Show that {fn } converges

uniformly to a function f , and that the equation f 0 (x) = limn→∞ fn0 (x) is
correct if x 6= 0 but false if x = 0.

8. If (
0 (x ≤ 0),
I(x) =
1 (x > 0),
P
if {xn } is a sequence of distinct points in (a, b), and if |cn | converges, then
prove that the series
X∞
f (x) = cn I(x − xn ) (a ≤ x ≤ b)
n=1
converges uniformly, and that f is continuous for every x 6= xn .

4
Solutions by Erin P. J. Pearse

unif
9. {fn } are continuous and fn −−−→ f on E. Show lim fn (xn ) = f (x) for every
sequence {xn } ⊆ E with xn → x.
Pick N1 such that
n ≥ N1 =⇒ sup |fn (x) − f (x)| < ε/2.
x∈E

Then surely |fn (xn ) − f (xn )| < ε/2 holds for each n ≥ N1 . By Thm. 7.12, f
is continuous, so pick N2 such that
n ≥ N2 =⇒ |f (xn ) − f (x)| < ε/2.
Then we are done because
|fn (xn ) − f (x)| ≤ |fn (xn ) − f (xn )| + |f (xn ) − f (x)|.

11. {fn }, {gn } are defined on E with:


P unif
(a) { N n=1 fn } uniformly bounded,
P (b) gn −−−→ g on E, and (c) gn (x) ≤
gn−1 (x) ∀x ∈ E, ∀n. Prove fn gn converges uniformly on E.
Define a := supx∈E |fn (x)| and bn := supx∈E |gn (x)|. Then
¯ ¯
¯X ∞ ¯ X ∞ ∞
X
¯ ¯
sup ¯ fn (x)gn (x)¯ ≤ sup |fn (x)gn (x)| ≤ an bn → 0,
x∈E ¯ ¯ x∈E
N +1 N +1 N +1

by Thm. 3.42.

13. {fn } is monotonically increasing on R, and 0 ≤ fn (x) ≤ 1


(a) Show ∃f, {nk } such that f (x) = limk→∞ fnk (x), ∀x ∈ R.
By Thm. 7.23, we can find a subsequence {fni } such that {fni (r)} con-
verges for every rational r. Thus we may define f (x) := supr≤x f (r),
where the supremum is taken over r ∈ Q. It is clear that f is monotone,
because
x < y =⇒ {r ≤ x} ⊆ {r ≤ y}
and the supremum can only increase on a larger set. Thus, f has at most
a countable set of discontinuities, by Thm. 4.30, pick x such that f is
continuous at x.
We want to show fni (x) → f (x). Fix ε > 0. Since f is continuous at x,
choose δ such that |x − y| < δ =⇒ |f (x) − f (y)| < ε/3. Now pick a
rational number r ∈ [x − 3δ , x]. Then
|fni (x) − f (x)| ≤ |fni (x) − fni (r)| + |fni (r) − f (r)| + |f (r) − f (x)|. (7.2)
On the RHS of (7.2): the last term is less than ε/3 by the choice of δ;
and the middle term is less than ε/3 whenever i ≥ N1 for some large N1 ,
because the subsequence converges on the rationals. It remains to show
the first term gets small.

5
Principles of Mathematical Analysis — Walter Rudin

Pick some rational s ∈ [x, x + δ/3]. Then r ≤ x ≤ s and the continuity


of f at x shows

|r − s| < δ =⇒ |f (r) − f (s)| < ε/3.

Also, since the fni are monotone,

fni (r) ≤ fni (x) ≤ fni (s). (7.3)

With

|fni (r) − fni (s)| ≤ |fni (r) − f (r)| + |f (r) − f (s)| + |f (s) − fni (s)|,

some large N2 , i ≥ N2 gives |fni (r) − fni (s)| < ε. By (7.3), this shows
|fni (r) − fni (x)| < ε.

(b) If f is continuous, show fnk → f uniformly on compact sets.


Let K be compact. Fix ε > 0. Since f is uniformly continuous on
K, pick δ such that |x − y| < δ =⇒ |f (x) − f (y)| < ε/3. Since K
S
is compact, we can find {x1 , . . . xJ } such that K ⊆ Jj=1 Bδ (xj ), where
Bδ (xj ) := (xj − δ, xj + δ).

pw
16. {fn } is equicontinuous on a compact set K, fn −−−→ f on K. Prove {fn }
converges uniformly on K.
Define f by f (x) := lim fn (x). Fix ε. From equicontinuity, find δ such that

|x − y| < δ =⇒ |fn (x) − fn (y)| < ε/3, ∀n, x, y.

Letting n → ∞, this gives

|x − y| < δ =⇒ |f (x) − f (y)| < ε/3, ∀x, y.

Since K is compact, we can choose a finite set {x1 , . . . , xJ } such that


[J
K⊆ Bδ (xj )
j=1

where Bδ (xj ) := (xj − δ, xj + δ). For each xj , we know fn (xj ) → f (xj ), so


pick N big enough that

n ≥ N =⇒ |fn (xj ) − f (xj )| < ε/3, ∀j = 1, . . . , J.

For any x ∈ K, x ∈ Bδ (xj ) for some j. Thus for all x ∈ K,

|fn (x) − f (x)| ≤ |fn (x) − fn (xj )| + |fn (xj ) − f (xj )| + |f (xj ) − f (x)|.

6
Solutions by Erin P. J. Pearse
Rx
18. Let {fn } be uniformly bounded and Fn (x) := a f (t) dt for x ∈ [a, b]. Prove
∃{Fnk } which converges uniformly on [a, b]
We need to show {Fn } is equicontinuous. Then by Thm. 6.20, each Fn is
continuous; and by Thm. 7.25(b), we’re done. So fix ε > 0, let x < y, and let
M be the uniform bound on the {fn }.
¯Z y ¯ Z y
¯ ¯
|Fn (x) − Fn (y)| = ¯¯ fn (t) dt¯¯ ≤ |fn (t)| dt ≤ M (y − x)
x x
Then pick any δ < ε/M and
|x − y| < δ =⇒ |Fn (x) − Fn (y)| ∀n,
so {Fn } is equicontinuous.

7
Principles of Mathematical Analysis — Walter Rudin
R1
20. f is continuous on [0, 1] and 0 f (x)xn dx = 0, n = 0, 1, 2, . . . . Prove that
f (x) ≡ 0.
Let g be any polynomial. Then g(x) = a0 + a1 x + a2 x2 + · · · + aK xK . By
linearity of the integral,
Z 1 K
X Z 1 K
X
k
f (x)g(x) dx = ak f (x)x dx = 0 = 0.
0 k=0 0 k=0

By the Weierstrass theorem, let {fn }∞n=1 be a sequence of polynomials which


converge uniformly to f on [0, 1]. Then
Z 1 Z 1
2
f (x) dx = lim f (x)fn (x) dx = lim 0 = 0.
0 n→∞ 0 n→∞

Then by Chap. 6, Exercise 2, f 2 (x) ≡ 0. Thus f (x) ≡ 0.

21. Let K be the unit circle in C and define


( N
)
X
iθ inθ ..
A := f (e ) = cn e . cn ∈ C, θ ∈ R .
n=0

To see that A separates points and vanishes at no point, note that A con-
tains the identity function f (eiθ ) = eiθ .
To see that there are continuous functions on K that are not in the uniform
closure of A, note that
Z 2π
f (eiθ )eiθ dθ = 0 ∀f ∈ A, (7.4)
0

and hence for g = lim gn (uniform limit) with gn ∈ A,


Z 2π Z 2π
iθ iθ
g(e )e dθ = lim gn (eiθ )eiθ dθ = 0,
0 n→∞ 0

by Thm. 7.16. Thus, all functions in the closure of A satisfy (7.4). However,
if we choose an h which is not in A, like

h(eiθ ) = e−iθ ,

then h is clearly continuous on K, and


Z 2π Z 2π
iθ iθ
h(e )e dθ = 1 dθ = 2π.
0 0

Thus h is not in the uniform closure of A.

8
Solutions by Erin P. J. Pearse

22. Assume f ∈ R(α) on [a, b] and prove that there are polynomials Pn such that
Z b
lim |f − Pn |2 dα = 0.
n→∞ a
n→∞
We need to find {Pn } such that kf −Pn k2 −−−−−→ 0. Fix ε > 0. By Chap. 6,
Exercise 12, we can find g ∈ C[a, b] such that kf − gk2 < ε/2. Note that
Z b Z b
2
|g − P | ≤ sup |g − P |2 = sup |g − P |2 (b − a).
a a
Then by the Weierstrass theorem, we can find a polynomial P such that
kg − P k2 ≤ sup |g − P |(b − a) < ε/2.
By Chap. 6, Exercise 11, this gives kf − P k2 < ε.

23. Put P0 = 0 and define Pn+1 (x) := Pn (x) + (x2 − Pn2 (x)) /2 for n = 0, 1, 2, . . . .
Prove that limn→∞ Pn (x) = |x| uniformly on [−1, 1].
Note that if Pn is even, the definition will force Pn+1 to be even, also. Now
2
P1 = x2 , so assume 0 ≤ Pn−1 ≤ 1 for |x| ≤ 1. Then
2
µ ¶
x2 Pn−1 x2 Pn−1
Pn = Pn−1 + − = + Pn−1 1 − .
2 2 2 2
By elementary calculus,
0≤y≤1 =⇒ f (y) = y(1 − y2 ) takes values in [0, 21 ]. (7.5)
x2
Since |x| ≤ 1 also implies 2
∈ [0, 1], this gives 0 ≤ Pn ≤ 1. Then
|x| + Pn (x)
0≤ ≤1
2
|x| + Pn (x)
0≤1− ≤ 1. (7.6)
2
To see Pn (x) ≤ |x|, consider that for x ≥ 0, the inequality
x − P1 (x) = x − x2 /2 ≥ 0
holds in virtue of the positivity of f in (7.5). Then by the symmetry of even
functions, this is true for |x| ≤ 1. Now suppose Pn−1 ≤ |x|, i.e., |x|−Pn−1 (x) ≥
0. Then the given identity, and (7.6), give
¡ ¢
|x| − Pn = (|x| − Pn−1 ) 1 − 12 (|x| + Pn−1 ) ≥ 0.
We have established
0 ≤ Pn (x) ≤ Pn+1 (x) ≤ |x| for − 1 ≤ x ≤ 1. (7.7)
Now, for n = 1, it is clear that
³ ´1
|x|2 |x|
|x| − P1 (x) = |x| − 2
≤ |x| 1 − 2
.

9
Principles of Mathematical Analysis — Walter Rudin
³ ´n−1
|x|
So suppose |x| − Pn−1 (x) ≤ |x| 1 − 2
. Then, multiplying each side by
³ ´
1 − |x|
2
− Pn−12 (x) and using the identity, we get
³ ´ ³ ´n−1 ³ ´
|x| Pn−1 (x) |x| |x| Pn−1 (x)
(|x| − Pn−1 (x)) 1 − 2 − 2 ≤ |x| 1 − 2 1− 2 − 2
³ ´n−1 ³ ´
|x| |x|
|x| − Pn (x) ≤ |x| 1 − 2 1− 2
³ ´n
|x| − Pn (x) ≤ |x| 1 − |x|2
.
¡ ¢n ¡ ¢n−1 ³ ´
Now consider gn (x) = x 1 − x2 on [0, 1]. gn (x) = 1 − x2 1 − (n+1)x
2
2
shows that gn has extrema at 2 and n+1 ; only the latter is in [0, 1]. Then for
³ ´n
|x|
fn (x) = |x| 1 − 2 , fn has extrema
¡ 2 ¢ 2
¡ n ¢n 2
fn ± n+1 = n+1 n+1
< n+1 .
We have established
³ ´n
2 |x|2
|x| − Pn (x) ≤ |x| 1 − < n+1 .x
(7.8)
¯ ¯
¯ ¯ 2 n→∞ unif
Now sup ¯|x| − Pn (x)¯ < n+1 −−−−−→ 0 and Pn −−−→ |x|.

10
Solutions by Erin P. J. Pearse

8. Some Special Functions


7. If 0 < x < π2 , prove that π2 < sinx x < 1.
To see the first inequality, suppose ∃x0 ∈ (0, π2 ) with π2 ≥ sinxx0 0 . Since
limx→0 sinx x = 1, IVT gives an y with π2 = siny y . Define g(x) = sin x − π2 x so
g(y) = 0. Then g 0 (x) = cos(x) − π2 and g 00 (x) = − sin x < 0 for x in the
interval. By IVT again, there is a point z ∈ (0, y) with g 0 (z) = 0, so g 0 (x) < 0
<
for x ∈ (y, π2 ). Then g(y) = 0 implies that g( π2 ) < 0. .
To see the second inequality, put f (x) = x − sin x. Then f 0 (x) = 1 − cos x
shows f 0 (0) = 0 and f 0 (x) > 0 for 0 < x < π2 .

8. For n = 0, 1, 2, . . . and x ∈ R, prove | sin nx| ≤ n| sin x|.


This is clearly true (with equality) for n = 0, 1, so induct on n; suppose
| sin nx| ≤ n| sin x|. We could use this assumption, to say
| sin(n + 1)x| = | sin(nx + x)| ≤ | sin nx| + | sin x| = (n + 1)| sin x|,
if we could prove the central inequality. From the definitions of C(x) and S(x)
given in (46),
sin(x + y) = sin x cos y + cos x sin y.
Applying this,
| sin(nx + x)| ≤ | sin nx cos x| + | cos nx sin x|
≤ | sin nx| + | sin x|,
since | cos x| ≤ 1 for all x.

9. (a) Put sN = 1 + 12 + · · · + N1 . Prove that γ = limN →∞ (sN − log N ) exists.


Note that
X N Z ∞µ ¶
1 1 1
γ = lim − log N = − dx.
N →∞
k=1
k 1 [x] x
The following sum telescopes:
XN · ¸
1
− (log(k + 1) − log k) = sN − log(N + 1),
k=1
k
so rewrite the summand as
µ ¶
1 1
ak := − log 1 + .
k k
We will bound ak by 12 x−2 . Note that for
x2
f (x) = − x + log(1 + x),
2

11
Principles of Mathematical Analysis — Walter Rudin
1 1
we have f 0 (x) = x − 1 + 1+x and f 00 (x) = 1 = (1+x) 0
2 . This gives f (0) = 0

and f 00 (x) > 0 for all x > 0, so that f is always positive for x > 0 and
x2
0 ≤ x − log(1 + x) ≤ .
2
P 1
In particular, this is true for x = k1 . Since k2
converges, this shows
N
X ³ ´
lim ak = lim sN − log(N + 1)
N →∞ N →∞
k=1

exists. Finally,
µ ¶
1 N →∞
log(N + 1) − log N = log 1 + −−−−−→ log 1 = 0
N
shows that limN →∞ (sN − log N ) = 0.

(b) Roughly how large must m be such that N = 10m satisfies sN > 100?
From the above, and problem #13, we have
π2 π2
0 ≤ sN − log N ≤ =⇒ log N ≤ sN ≤ log N + .
6 6
m
Then m ≈ 43.43 will ensure log 10 > 100.

12. Suppose that f is periodic with f (x) = f (x + 2π), and for δ ∈ (0, π),
(
1, |x| ≤ δ,
f (x) =
0, δ < |x| < π.

(a) Compute the Fourier coefficients of f .


Z δ
1 −1 ¡ −inδ ¢ sin nδ
cn = e−inx dx = e − einδ = .
2π −δ 2πin πn

(b) Conclude that



X sin nδ π−δ
= , (0 < δ < π).
n=1
n 2

X X sin nδ
=
n∈Z n∈Z
πn

X 1 X sin nδ
2 cn = .
n=1
π n∈Z πn

12
Solutions by Erin P. J. Pearse

(c) Deduce from Parseval’s theorem that


X∞
sin2 nδ π−δ
2
= .
n=1
nδ 2
Parseval’s theorem gives equality between
Z π Z δ
1 2 1 δ
|f (x)| dx = 1 dx =
2π −π 2π −δ π
and
X X sin2 nδ δ X sin2 nδ
2
|cn | = = 2 .
n∈Z n∈Z
π 2 n2 π n∈Z n2 δ
Then µ ¶

X
2 1 δ
|cn | = − c0 =
n=1
2 π
δ
Note that c0 = π
also.

(d) Let δ → 0 and prove that


Z ∞µ ¶2
sin x π
dx = .
0 x 2

(e) Put δ = π/2 in (c). What do you get?


∞ ∞
2 X sin2 ( π2 n) π X 1 π2
= =⇒ = .
π n=0 n2 4 k=0
(2k + 1)2 8

13. Put f (x) = x for 0 ≤ x < 2π, and apply Parseval’s Theorem to conclude
X∞
1 π2
2
= .
n=1
n 6
Apply it to the 2π-periodic function f (x) = x on (−π, π) instead. Integra-
tion by parts gives
Z π
1 (−1)n
cn = xeinx dx = ,
2π −π in
so with Parseval’s theorem,
X 1 X Z π
2 1 2 π2
2
= |c n | = |f (x)| dx = .
n∈Z
n n∈Z
2π −π 3
Since we have |cn | = |c−n |, we find the desired series by subtracting
Z π
1
c0 = x dx = 12 [x2 ]π−π = 0
2π −π

13
Principles of Mathematical Analysis — Walter Rudin

from each side, and dividing the remainder by 2.



à ! µ ¶
X 1 1 X 1 π2 π2
2
2
= |cn | − c0 = − 0 = .
n=1
n 2 n∈Z
2 3 6

9. Functions of Several Variables


5. Prove that to every A ∈ L(Rn , R) there corresponds a unique y ∈ Rn such
that Ax = x · y.
For N = {x ∈ Rn ... Ax = 0}, N is a closed subspace by Problem 4. If
N = Rn , then Ax = 0 · x ∀x, and we’re done. So suppose N 6= Rn . Then
there is a nonzero vector
x0 ∈ N ⊥ = {u ∈ Rn ... u · v = 0, ∀v ∈ N }.
We check Ax = yA · x for yA defined by
Ax0
yA = x0 .
kx0 k2
First, if x ∈ N , then Ax = 0 = yA · x. Next, if x = αx0 , then
Ax0
Ax = A(αx0 ) = αAx0 = x0 · αx0 = yA · αx0 .
kx0 k2
Since the functions A(x) and yA · x are linear functions of x and agree on N
and x0 , they must agree on the space spanned by N and x0 . But N and x0
span Rn , since every element y ∈ Rn can be written
µ ¶
Ay Ay
y= y− x0 + x0 .
Ax0 Ax0
Thus Ax = yA · x for all x ∈ Rn . If Ax = y 0 · x also, then
ky 0 − yA k2 = A(y 0 − yA ) − A(y 0 − yA ) = 0.
So y 0 = yA is unique. Equality of the norms comes from the inequalities:
kAk = sup |Ax| = sup |yA · x| ≤ sup kyA kkxk = kyA k, and
kxk≤1 kxk≤1 kxk≤1
¯ µ ¶¯
¯ y ¯
kAk = sup |Ax| ≥ ¯¯A
A ¯ = yA · yA = kyA k.
kxk≤1 kyA k ¯ kyA k

6. If f (0, 0) = 0 and
x1 x2
f (x1 , x2 ) = , for (x1 , x2 ) 6= 0,
x21
+ x22
prove that D1 f (x1 , x2 ) and D2 f (x1 , x2 ) exist for every point (x1 , x2 ) ∈ R2 ,
although f is not continuous at (0, 0).

14
Solutions by Erin P. J. Pearse

First, the derivatives are


x2 (x22 − x21 ) x1 (x21 − x22 )
D1 f (x1 , x2 ) = , and D2 f (x1 , x2 ) = ,
(x21 + x22 )2 (x21 + x22 )2
and so clearly exist wherever (x, y) 6= (0, 0). To check the origin, consider that
along the axes,
0 0
D1 f (x1 , 0) = 4 = 0, and D2 f (0, x2 ) = 4 = 0.
x1 x2
However, for f to be continuous at (0, 0), we must have
f (0, 0) = lim f (x, y),
(x,y)→(0,0)

no matter how (x, y) → (0, 0)! Define γ(t) : R → R2 by


γ(t) = (x(t), y(t)) = (t, t),
so that we approach the origin along the diagonal. Then
t2 1 1
lim f (x(t), y(t)) = lim f (t, t) = lim 2
= lim = 6= 0.
t→0 t→0 t→0 2t t→0 2 2

7. Suppose that f is a R-valued defined in an open set E ⊆ Rn , and that the


partial derivatives D1 f, . . . , Dn f are bounded in E. Prove that f is continuous
in E.
Take m = 1 as in 9.21. Pick x ∈ E and consider the ball of radius r
B(x, r) ⊆ E. Choose h ∈ Rn such that khk < r. Define vk ∈ Rn by
vk = (h1 , . . . , hk , 0, . . . , 0) for k = 1, . . . , n.
Then as in (42),
n
X
f (x + h) − f (x) = [f (x + vj ) − f (x + vj−1 )] .
j=1

Since |vk | < r for 1 ≤ k ≤ n, and since B = B(x, r) is convex, the segments
with endpoints x + vj−1 and x + vj lie in B. Then vj = vj−1 + hj ej , where
hj ej = (0, . . . , 0, hj , 0, . . . , 0).
Then the Mean Value Theorem applies to the partials and shows that the j th
summand is
hj (Dj f )(x + vj−1 + tj hj ej )
for some tj ∈ (0, 1). By hypothesis, we have M such that
|(Dj f )(x)| ≤ M j = 1, . . . , n, ∀x ∈ E.
Applying this to the absolute value of the above difference,
n
X
|f (x + h) − f (x)| ≤ |hj | · M ≤ nM max{|hj |} ≤ nM khk.
j=1

15
Principles of Mathematical Analysis — Walter Rudin

8. Suppose that f is a differentiable real function in an open set E ⊆ Rn , and


that f has a local maximum at a point x ∈ E. Prove that f 0 (x) = 0.
Let {ej } be the standard basis vectors of Rn , and define
ϕj (t) = f (x0 + tej ), for t ∈ R.
Then ϕj : R → R is differentiable, so by Thm. 5.8,
ϕ0j (0) = (Dj f )(x0 ) = 0.
But the partial derivatives (Dj f )(x0 ) are precisely the columns of the matrix
(Df )(x0 ) = f 0 (x).

9. If f is a differentiable mapping of a connected open set E ⊆ Rn into Rm , and


in f 0 (x) = 0 for every x ∈ E, prove f is constant in E.
First, suppose m = 1. Pick some x0 ∈ E and define
C = {x ∈ E ... f (x) = f (x0 )},
D = {x ∈ E ... f (x) 6= f (x0 )} = E \ C.
Clearly, C ∪ D = E and C ∩ D = ∅. If we can show each of C, D is open,
then D must be empty, or else we’d have a disconnection of the connected set
E.
E is open, so for any x ∈ C, pick ε > 0 such that B(x, ε) ⊆ E. Then
for any y ∈ B(x, ε), the segment [x, y] ⊆ B(x, ε). Since f 0 (x) = 0, we have
f (x) = f (y) by Thm. 5.11(b). Also, we chose x ∈ C, so f (x) = f (y) = f (x0 ).
This puts B(x, ε) ⊆ C and shows C is open.
Now suppose we have some x ∈ E \ C. Since E is open, we can again choose
ε > 0 such that B(x, ε) ⊆ E. By the exact same argument, we get
y ∈ B(x, ε) =⇒ f (x) = f (y) 6= f (x0 ),
so that B(x, ε) ⊆ D and D is open.
Finally, if m > 1, then apply this argument to each component of f .

10. Suppose f is a differentiable mapping of R into R3 such that |f (t)| = 1 for


every t. Prove f 0 (t) · f (t) = 0. Interpret this geometrically.
Since kf (t)k = 1, we have
(f1 (t))2 + (f2 (t))2 + (f3 (t))2 = 1.
Differentiating both sides,
2f1 (t)f10 (t) + 2f2 (t)f20 (t) + 2f3 (t)f30 (t) = 0
f (t) · f 0 (t) = 0.
This means that for any curve on the unit sphere, the tangent at p ∈ S 1 is
orthogonal to p, i.e., the surface of a sphere is orthogonal to the radius at any
point.

16
Solutions by Erin P. J. Pearse

14. Define f (0, 0) = 0 and f (x, y) = x3 /(x2 + y 2 ) for (x, y) 6= (0, 0).

2 1
0
-1
-2
2
1
0
-1
-2
1 2
-1 0
-2

(a) Prove that D1 f and D2 f are bounded functions in R2 so that f is con-


tinuous.
We have the partial derivatives
x4 + 3x2 y 2
D1 f = , and
x4 + 2x2 y 2 + y 4
−2x3 y
D2 f = 4 .
x + 2x2 y 2 + y 4
Boundedness at ±∞:
3(x4 + 3x2 y 2 )
|D1 f | ≤ = 3, as |x| → ∞, and
x4 + 2x2 y 2
|2x3 | |y|→∞
|D2 f | = 4 2 3
−−−−−→ 0.
|x /y + 2x y + y |
Boundedness away from ∞: only need to check zeroes of the denomina-
tors, and there is only (0, 0).
f (x, 0) − f (0, 0)
D1 f (0, 0) = lim lim 1 = 1, and
x→0 x x→0
f (0, y) − f (0, 0)
D2 f (0, 0) = lim lim 0 = 0.
y→0 y y→0

Thus f is continuous by Prob. 7.

(b) Let u ∈ R2 , |u| = 1. Show (Du f )(0, 0) exists and |(Du f )(0, 0)| ≤ 1.
The directional derivative is
f [(0, 0) + tu] − f (0, 0) 1 t 3 x3
Du f (0, 0) = lim = lim · 2 2
t→0 t t→0 t t (x + y 2 )

x3
= 2 ,
x + y2

17
Principles of Mathematical Analysis — Walter Rudin

for u = (x, y). Since |u| = x2 + y 2 = 1, this implies |x| ≤ 1. Hence we


have
Du f (0, 0) = x3 and |Du f (0, 0)| = |x3 | ≤ 1.

(c) Let γ be a differentiable mapping of R into R2 with γ(0) = (0, 0) and


|γ 0 (0)| > 0. Put g(t) = f (γ(t)) and prove that g is differentiable for every
t ∈ R.
Put γ(t) = (x(t), y(t)) so that
x(t)3
g(t) = , for t 6= 0.
x(t)2 + y(t)2
Since γ is differentiable for t 6= 0, we have x(t), y(t) differentiable for t 6= 0,
and hence g(t) is differentiable for t 6= 0, by the chain rule, Thm. 9.15.1
It remains to check that g is differentiable at the origin.
g(h) − g(0)
g 0 (0) = lim
h→0 h
1 x(h)3
= lim ·
h→0 h x(h)2 + y(h)2

x(h)3 h2
= lim ·
h→0 h3 x(h)2 + y(h)2
µ ¶3 µ ¶
x(h) h2
= lim lim
h→0 h h→0 x(h)2 + y(h)2

3 1
= (x0 (0)) · 0 .
|γ (0)|2
The last two lines are easiest to see by working backwards.
Also show γ ∈ C 0 =⇒ g ∈ C 0 .
Note that g 0 (t) = f 0 (γ(t))γ 0 (t). Since the additional hypothesis is that
γ 0 (t) is continuous (which is equivalent to saying x(t), y(t) are continuous,
by Thm. 4.10), we just need that f 0 (γ(t)) is continuous. Since the chain
rule gives
x(t)2 (x(t)2 x0 (t) + 3y(t)2 x0 (t) − 2x(t)y(t)y 0 (t))
g 0 (t) = ,
(x(t)2 + y(t)2 )2
it is clear that g 0 (t) is continuous whenever x, y are not simultaneously 0.
For t = 0 (or other t0 s.t. x(t0 ) = y(t0 ) = 0), replace x(t) by x0 (0)t + o(t)
(using Taylor’s Thm.) and similarly replace y(t) by y 0 (0)t + o(t). Thus
(x0 (0)t+o(t))2 ((x0 (0)t+o(t))2 x0 (t)+3(y 0 (0)t+o(t))2 x0 (t)−2(x0 (0)t+o(t))(y 0 (0)t+o(t))y 0 (t))
g 0 (t) ≈ ((x0 (0)t+o(t))2 +(y 0 (0)t+o(t))2 )2
,

1By (a), the partials of f exist and are continuous in an open nbd of (0, 0). Hence, f is continuously
differentiable at (0, 0) by Thm. 9.21.

18
Solutions by Erin P. J. Pearse

t→0
so that g 0 (t) −−−−→ g 0 (0).

(d) In spite of this, f is not differentiable at (0, 0).


Let u = (x, y) be a unit vector. By Rudin’s (40),
Du = D1 f (0, 0)x + D2 f (0, 0)y = 1 · x + 0 · y = x.
But from (b) we have Du = x3 6= x for any x other than −1, 0, 1.

(Bonus problem) This problem isn’t in Rudin.


Define f : R2 → R by
(
0, (x, y) = (0, 0),
f (x, y) = x2 y
x4 +y 2
, otherwise.
Let γa be any straight line through the origin with slope a and γa (0) = (0, 0). This
means that (up to parametrization) γa (t) = (t, at) for some a ∈ R, or γ∞ (t) = (0, t).
Show that
lim f (γa (t)) = f (0, 0) = 0,
t→0
but that f is not continuous at the origin.

0.5
0.25
0
-1 -0.25
-0.5
-0.5 1
0
0
0.5
-1
1

Continuity of the restriction follows by a basic computation:


at3 at
lim f (γa (t)) = lim 4 2
= lim 2 = 0, and
t→0 t→0 t + (at) t→0 t + a2

0
lim f (γ∞ (t)) = lim = 0,
t→0 t→0 0 + t2

To see the other part, approach the origin along a parabolic curve: let γ(t) =
(x(t), y(t)) = (t, t2 ). Then
t2 t2 t4 1 1
lim f (γ(t)) = lim 4 2 2
= lim 4
= lim = 6= 0.
t→0 t→0 t + (t ) t→0 2t t→0 2 2

19
Principles of Mathematical Analysis — Walter Rudin

4x6 y 2
15. Define f (x, y) = x2 + y 2 − 2x2 y − (x4 +y 2 )2
.

0 -1 -2
1
2
20
10
0
-10
-2
-1
0
1
2

(a) Prove that 4x4 y 2 ≤ (x4 + y 2 )2 , so that f ∈ C 0 .

(x4 + y 2 )2 − 4x4 y 2 = x8 + 2x4 y 2 + y 4 − 4x4 y 2


= (x4 − y 2 )2 ≥ 0

Now f (x, y) = x2 + y 2 − 2x2 y − x2 ϕ(x, y), where 0 ≤ ϕ(x, y) ≤ 1. This


gives f continuous at (0, 0).

(b) For 0 ≤ θ < π, t ∈ R, let gθ (t) = f (t cos θ, t sin θ). Show that gθ (0) = 0,
gθ0 (0) = 0, gθ00 (0) = 2. Each gθ thus has a strict local minimum at t = 0,
i.e., on each line through (0, 0), f has a strict local minimum at (0, 0).
Since sin θ, cos θ are bounded, gθ (0) = f (0, 0) = 0 by (a). By some
nightmarishly tortuous (but elementary) calculation, get the other two
results.

(c) Show that (0, 0) is not a local minimum of f , since f (x, x2 ) = −x4 .
Substituting in y = x2 :

4x10
f (x, x2 ) = x2 + x4 − 2x4 −
(2x4 )2
= x2 − x4 − x2
= −x4 .

¡ ¢
16. Define f (t) = t + 2t2 sin 1t , f (0) = 0. Then f 0 (0) = 1, f 0 bounded in (−1, 1),
but f is not injective in any nbd of 0.

20
Solutions by Erin P. J. Pearse

0.2

0.1

-0.2 -0.1 0.1 0.2

-0.1

-0.2

¡1¢ ¡1¢
First note that f 0 (t) = 1 − 2 cos t
+ 4t sin t
for t 6= 0. Since sin t, cos t
are bounded by 1, we have
|f 0 (t)| ≤ 1 + 4 + 2 = 7, for t ∈ (−1, 1) \ {0}.
At t = 0, we have that
f (t) − f (0) t + 2t2 sin 1t
= = 1 + 2t sin 1t
t t
shows f (0) = 1 by the Sandwich theorem applied to t + 2t2 and t − 2t2 . So
0

f 0 is bounded in all of (−1, 1).


1
Define sn = 2nπ for n = 1, 2, . . . , so that
sin 2nπ
f 0 (sn ) = 1 − 2 cos(2nπ) + 2 = 1 − 2 = −1.

2
Then define tn = (4n+3)π
for n = 1, 2, . . . , so that

f 0 (tn ) = 1 − 2 cos (4n+3)π


2
+ 8
(4n+3)π
sin (4n+3)π
2
=1+ 8
(4n+3)π
> 0.
Since f 0 changes sign between each successive sn and tn , and since sn , tn → 0,
f fails to be injective in every neighbourhood of 0.

17. Let f = (f1 , f2 ) : R2 → R2 be given by


f1 (x, y) = ex cos y, f2 (x, y) = ex sin y.

(a) What is the range of f ?



Note that if we identify (0, 1) = −1 = i, then
f (z) = f (x + iy) = ez = ex+iy = ex cos y + iex sin y.
But you only really need to see that (ex cos y, ex sin y) is polar coordinates
for a point with radius ex and argument y, to see that the range is any
point of R2 with radius r > 0, i.e., R2 \ (0, 0).

21
Principles of Mathematical Analysis — Walter Rudin

(b) Show that the Jacobian of f is nonzero. Thus, every point of R2 has a
neighbourhood in which f is injective. However, f is not injective.
The Jacobian is the determinant of partials:
¯ x ¯
¯ e cos y −ex sin y ¯
¯ x ¯ 2x 2 2 2x
¯ e sin y ex cos x ¯ = e (cos x + sin x) = e 6= 0.
But f is not injective, since f (x, y) = f (x, y + 2nπ), for n ∈ Z.

(c) Put a = (0, π3 ) and b = f (a). Let g be the continuous inverse of f ,


defined in a neighbourhood of b, such that g(b) = a. Find an explicit
formula for g, compute f 0 (a) and g 0 (b), and verify Rudin’s (52).
For u = ex cos y, v = ex sin y, one verifies that
x = 12 log(u2 + v 2 ) = log r, for r = ex , and
y = tan−1 v
u
= arg θ, for θ = uv .
For the derivatives,
" √ #
1 3

f 0 (a) = √
2 2
,
3 1
2 2
and
" # " √ #
u v 1 3
g 0 (u, v) = 1
u2 +v 2
=⇒ g 0 (b) = 2

2
,
−v u − 23 1
2
Finally,
³ √ ´
f 0 (g(u, v)) = f 0 log u2 + v 2 , tan−1 uv
" ¡ ¢ ¡ ¢ #
√ cos tan−1 uv − sin tan−1 uv
= u2 + v 2 ¡ ¢ ¡ ¢
sin tan−1 uv cos tan−1 uv
" #
√ √ u − √ v
2
u +v 2 2
u +v 2
= u2 + v 2 v

u2 +v 2
− √u2u+v2
· ¸
u −v
=
v u
−1
= [g 0 (u, v)]

(d) What are the images under f of lines parallel to the coordinate axes?
Lines parallel to the x-axis are mapped to straight lines through the
origin, parameterized exponentially. Lines parallel to the y-axis are cir-
cles about the origin of radius ex , parameterized with constant speed.
Diagonal lines γ(t) = (at + b, btc ) will get mapped to (et cos t, et sin t),
counterclockwise logarithmic spirals emanating from the origin.

22

You might also like