Ashwin Sah Analysis 18.

A34

Problem 1. Prove that there is a constant C such that, if p(x) is a polynomial of

degree 1999, then Z 1
|p(0)| ≤ C |p(x)| dx.
−1

Solution 1:
Consider p(x) = a1999 x1999 + a1998 x1998 + · · · + a0 , where not all the coefficients are
zero. We wish to prove a bound of the form

|a0 |
f (a0 , . . . , a1999 ) = R 1 ≤ C.
−1
|a1999 x1999 + · · · + a0 | dx

Note that as a function f : R2000 \{⃗0} → R, this is continuous. To check this, we

note that ⃗a 7→ |a1999 x1999 + · · · + a0 | is continuous for each x ∈ [−1, 1], and in fact
⃗a and ⃗b map to values differing by at most 1999
P
i=0 |ai − bi |. Thus the integral in
question differs by at most twice this, which goes to zero as ⃗b → ⃗a. Additionally,
this map is nonzero when ⃗a is nonzero, since the integral of a nonzero continuous
function is nonzero. Thus inverting it remains well-defined and hence continuous.
We want a uniform upper bound. Notice that f is scale-invariant, and thus the
set of values attained by f is the same as the set of value obtained by f on
S 1999 ⊆ R2000 . Now, the sphere is compact, and the image of a compact set is
compact, so f (R2000 \{⃗0}) = f (S 1999 ) ⊆ R is compact. Compact subsets of R are
bounded, by the Heine-Borel theorem, and if we let C be an upper bound on the
magnitudes of elements in the compact set f (S 1999 ), then the above inequality clearly
holds. We are done. □ R1
Note: by attempting to prove the stronger inequality supx∈[−1,1] |p(x)| ≤ C −1 |p(x)| dx
for some C, we can actually find that the problem is a direct consequence of the
equivalence of all norms on Rn .

Solution 2:
These compactness-style arguments tell us a number exists, but don’t tell us which
number. It is also worth seeing a more constructive argument with explicit bound-
ing. This approach has its merits. Basically, we want a lower bound on the value of
|p(x)| over some interval of some fixed size. How do we lower-bound the magnitude
of a polynomial? See how far the value is from the roots!
We are given some polynomial p(x), which factors as p(x) = 1999
Q
i=1 (x − ri ), where ri

Page 1 of 6
Ashwin Sah Analysis 18.A34

are in the complex plane. Now

1999
|p(x)| Y |x − ri |
= . (**)
|p(0)| i=1
|ri |

Here’s the idea: Fix some ϵ > 0 (we will choose it later). Then we look at all x
which are distance at least than ϵ from each of r1 , . . . , r1999 . For technical reasons
we only look at x in the interval − 12 , 12 . Then

|x − ri |
≥ϵ
|ri |

if |ri | ≤ 1 and
|x − ri | x x 1
= 1− ≥1− ≥
|ri | ri ri 2
if |ri | ≥ 1, using that |x| ≤ 12 . Thus, returning to (**), we have
1999    1999
|p(x)| Y 1 1
≥ min ϵ, = min ϵ,
|p(0)| i=1
2 2

for these x. Finally, we choose ϵ so that there are “a lot” of x which satisfy the
imposed conditions: |x − ri | ≥ ϵ for i ∈ [1999] and |x| ≤ 21 . Well, choose ϵ = 4000 1
.
Each “bad disk” of the form |z − ri | ≤ ϵ can cover at most a length 2ϵ interval of
the real numbers, for a total of 3998ϵ (in length). The interval − 12 , 12 has length 1.
1
Thus at least a length of 1−3998ϵ = 2000 is uncovered (there is technically something
nontrivial in this statement: alternatively, there are at most 2000 intervals in the
R 1 |p(x)|
1
complement, and one of them has length at least 2000 2 ), so the integral −1 |p(0)|
dx
is at least  1999
1 1
· .
2000 4000
In general, for degree n polynomials, this gives C(n) = exp(O(n log n)) or something.
Can you do better? □

Page 2 of 6
Ashwin Sah Analysis 18.A34

Problem 2. Let k be an integer greater than 1. Suppose a0 > 0, and define

1
an+1 = an + √
k a
n

for n > 0. Evaluate

ak+1
n
lim .
n→∞ nk

Solution 1:
We’ll first give a solution that uses some important bounding ideas, but there is a
second more natural solution that we will get to. First, what should the answer be?
k
The question implies that an = ck n k+1 is a good model of the growth. Then
k k 1
ck (n + 1) k+1 − ck n k+1 ≈ 1 1
.
ck n k+1
k

1
ck ·k − k+1
The left difference looks like a derivative so is around k+1 n , so we would expect
k+1
k
ck k = k+1 , for a final answer of k+1


k k
. Of course, this all assumes that the limit
k+1
even exists. Okay, so maybe the natural thing to look at is an k , which we expect
to be around k+1
k
n (this expression has no fractional exponents in sight). Well, if
k+1 k+1
k+1
we can show that an+1
k
− an k is “near” k
, then we should be morally done.
Well,
! k+1
k  
k+1 k+1 k+1 k+1 1 k+1 − k+1
an+1 − an
k k
− = an k 1+ k+1 −1− an k

k an k k

looks very promising due to the Taylor approximation (1 + x)p ≈ 1 + px + p(p−1)

2
x2 +
3 p
O(x ). In fact, for each p there is ϵp so that x < ϵp implies
  ≥ 1 + px (true at
(1 + x)
p(p−1)
x = 0 and take derivatives) and (1 + x)p ≤ 1 + px + 2
+1 x2 , true by looking
k+1
at Taylor series. Now applying this for the constant p = k we find an estimate
! k+1 !
k k+1 1
·
 
k+1 1 k+1 − k+1 k+1
k k − k+1 ·2
ank
1+ k+1 −1− an k ≤ an k
+ 1 an k

ank k 2
− k+1
≤ C k an k
.

Page 3 of 6
Ashwin Sah Analysis 18.A34

This is really good. Now let’s justify that an → ∞ as n → ∞. Indeed, notice

that a0 ≤ a1 ≤ · · · so that 11 ≤ 11 = D. Then an ≤ a0 + Dn, so an+1 ≥
ank+1 a0k+1
1 1
an + 1 . Since k+1
< 1, hashing out this series we see that it diverges, so
(a0 +Dn) k+1
an → ∞. We can also just note that an is increasing so if not unbounded it tends
toward a limit, implying that an+1 − an = 11 eventually approaches 0. But this
ank+1
makes no sense, since this forces an to get arbitrarily large.
Okay, so formally, we have proved that
 k+1 k+1  k+1
lim an+1 − an
k k
= .
n→∞ k
I claim that this means k+1
an k k+1
lim = ,
n→∞ n k
which is what we want.
In fact, more is true:
Lemma. (Cesàro’s Lemma) If limn→∞ an = c then limn→∞ n1 ni=1 ai = c.
P
Proof:
Fix ϵ > 0. Then for n > N we have |an − c| < ϵ, hence n1 ni=1 ai ≥ n1 (a1 + · · · +
P
aN + (c − ϵ)(n − N )) and similarly it is less than n1 (a1 + · · · + aN + (c + ϵ)(n − N )).
Thus we see that the ratio eventually is in the range (c − 2ϵ, c + 2ϵ) for n sufficiently
large. That is, the limit is c. □
We are done! □

Solution 2: Since we have time, let’s look at the second approach. First, fix k
k
and assume without loss of generality that a0 ≥ k k+1 (we’ll see why we need this in
a bit). Why can we assume this? Well, we can always shift the starting index of the
sequence a little bit, and we know that an → ∞ through an elementary argument
(see above).
n+1 −an
Now the idea is to look at a(n+1)−n = √1
k a
n
as a discretization of some differential
equation (this is basically Euler’s method, applied to infinity). In particular, the
dy 1
equation is dx = √k y with boundary condition f (0) = a0 > 0. It is well-known how

1 k+1 k+1
dy k
to solve these: we see y k dx = 1, and integrating dx gives k+1
y k = x + a0 k . Then
k k+1
y(x) = k+1



k
(x + d) k+1 for d = a0 k .

Page 4 of 6
Ashwin Sah Analysis 18.A34

Two things to notice: y(0) = a0 and y ′ (0) isR decreasing, i.e., y is concave. Thus it
x+1
is not hard to check that y(x + 1) − y(x) = x y ′ (t)dt ≤ y ′ (x), so that y(x + 1) ≤
y(x) + √k
1
.
y(x)
1
Thus the sequence bn = y(n) satisfies bn+1 ≤ bn + √ k
bn
for n ≥ 0 and b0 = a0 . It’s
clear that b0 ≤ a0 , and we can induct to show that bn ≤ an for all n ≥ 0: since
1 1
bn ≤ an we have bn+1 ≤ bn + √
k
b
≤ an + √
k a
n
= an+1 . We know the middle inequality
n
k+1
since f (x) = x + 11 is increasing for x > k k by taking derivatives (this is where
xk
the strange constant comes in!).
k k k
Now bn = y(n) = k+1

k+1
k
(n + d) k+1 , so an ≥ (n + d) k+1 . This is a pretty explicit
bound, and you can also prove it directly by induction (again only if we start with
k
a0 ≥ k k+1 ), but this tells you what the correct lower bound to induct on is, and
simplifies the inequalities involved in the induction step by allowing us to express
the difference bn+1 − bn as an integral that then has nice bounding properties.
Anyways, this expression gives the right constant that we want: we just need an
upper bound. How do we do this? Find a way to invert the inequality! We have
1 1
an+1 = an + √
k a
≤ a n + √
k
,
n bn

so expanding gives
n−1
X 1
an ≤ a0 + √
k
.
i=0
bn
If you think about this a little, it’s clear that this gives the “right expression” for
an upper bound. More precisely, this is
n−1 n−1
X 1 1 X 1
an ≤ a0 + 1 1
≤ a0 + s + 1 1
k+1 k+1 k+1

k+1

i=0 k
(n + d) k+1 k i=1 i k+1

1
where s is some inconsequential constant like . Finally, remember some
1 1
k+1
( ) k
k+1 d k+1

tricks
R n 1 from Evan’s lecture: this series looks like a Riemann sum for the integral
1 1 dx, and it is not hard to upper bound the right by
x k+1
 Z n 
1 1
− k+1
an ≤ a0 + s +
1 1+ x dx .
k+1 k+1 1
k

Page 5 of 6
Ashwin Sah Analysis 18.A34

Finally,
   k
 k+1
1 k + 1 k+1
k 1 ′ k+1 k
an ≤ a0 + s +
1 · n − = a0 + s + n k+1 .
k+1 k+1 k k k
k

The upper and lower bounds are of the same order and trivially give
k
ak+1

k+1
lim n k = ,
n→∞ n k

as claimed. □

Page 6 of 6