A Taste of Analytic Number Theory

Ayan Nath∗
ayan.nmath

This article, aimed at olympiad contestants, focuses on solving olympiad

number theory problems using analytic techniques and making contestants
familiar with common techniques and results in this topic. We start with
the prime number theorem, give an elementary proof of the weak version
and establish a few well known estimates for the two Chebyshev functions.
We also show Mertens’ first theorem on the fly and discuss Mertens’ second
theorem. Asymptotic density, equidistribution theorem are also added.

§1 The Prime Number Theorem

§1.1 Introduction
Primes are the building blocks of the integers, just as molecules and atoms are the
building blocks of nature, hence it makes great sense to study about primes, and in
particular, distribution of primes. I think you know that there are infinitely many primes
and most likely you already know Euclid’s proof to it, but it doesn’t tell us anything
significant about the distribution of primes. This is exactly what our objective is, i.e. try
to understand the distribution of primes. It is natural to define the function

π(x) = No. of primes at most x.

We would like to find a “formula” for π(x) in terms of x, it turns out finding an exact
formula is not really possible due to the raggedy nature of primes. Instead we try to
estimate it. The well known estimate which we call the Prime number theorem (PNT)
asserts that:

Theorem 1.1 Prime Number Theorem (PNT) —

x
π(x) ∼
log x

here we write f (x) ∼ g(x) if limx→∞ fg(x)(x)

= 1, the way to think about this is, f (x)
is approximated by g(x), the larger the x, the better the approximation. We do not
prove this here, but instead we establish a weaker estimate that there exist positive real
numbers a and b such that
ax bx
< π(x) < .
log x log x
Before we move on to the proof, it will help to get comfortable with big-O and little-o
notation:
∗
AoPS user : https://artofproblemsolving.com/community/user/362567

1
Ayan Nath A Taste of Analytic Number Theory

Digression 1.2. If f and g are two functions then we say that f (x) is O(g(x)) if and only
if there exist some constant C so that |f (x)| < Cg(x) for all large x. And we say that f (x)
is o(g(x)) if and only if limx→∞ fg(x)
(x)
= 0 where g(x) should be non-zero for all large enough
x. For example, sin x + e91 is O(1), x + (log x)10 is O(x), sin x + log x is o(x0.001 ). If f (x)
is O(g(x)) then this is often expressed as f (x) = O(g(x)) and similarly f (x) = o(g(x)) if
f (x) is o(g(x)), but you should remember that this is an abuse of notation. For example,
we write bxc = x + O(1), log x = o(x0.001 ), n! = O(nn ) in this article. It will help to get
comfortable with these notations, it lets us to not compute stuff we don’t care about. I
suggest reading this : https://en.wikipedia.org/wiki/Big_O_notation

§1.2 Proving weak PNT

Let me define some functions, bear with me for now, I will explain the motivation in a
moment. Define the von Mangoldt function Λ : N → R as
(
log p, n = pk for some prime p and positive integer k
Λ(n) =
0, otherwise

one can also think of this as weighting all the prime powers pk with log p. You can easily
see that X
Λ(d) = log n.
d|n

Whenever we are trying to find bounds it is a common theme to look at the “big picture”
at once, also called “global” methods, or in simple terms, double counting. Since we are
looking for bounds related to primes, it is somewhat motivated to “sum” everything up
and try to double count:
X XX
log n = Λ(d)
n≤x n≤x d|n
X X
= Λ(d) 1
d≤x n≤x,d|n
X jxk
= Λ(d)
d
d≤x
X x 
= Λ(d) + O(1)
d
d≤x
X Λ(n) X
=x + O(1) Λ(n)
n
n≤x n≤x

Now where did we double count? We double counted when we swapped the summations.
The left hand side is very easy to estimate accurately (Unimportant: Those who know
integration be like - lol just integrate). For now let us focus on the RHS, notice how
the RHS is related to primes while the LHS is not, clearly the above equation has some
information
P about primes encoded via the von Mangoldt function. We would want to
estimate n≤x Λ(n) to get rid of the awkward O(1) multiple. It now makes sense to
define X
ψ(x) = Λ(n).
n≤x

This is called the Second Chebyshev Function, we take the domain as R instead of N
to avoid writing floors whenever we have a non-integral input, we do this with almost all

2
Ayan Nath A Taste of Analytic Number Theory

the discussed functions in this article. In what follows, p always denotes a prime number.
Observe the following rough calculation
  X
X X   X log x
ψ(x) = Λ(n) = log p logp x = log p ≈ log x = π(x) log x.
log p
n≤x p≤x p≤x p≤x

This suggests that ψ(x) ∼ x (which is indeed true). We are deliberately ambiguous about
what ≈ means. We just need a rough estimate for ψ(x), even ψ(x) = O(x) should do.
For now assume this is true, we will get that
X X Λ(n)
log n = x + O(x).
n
n≤x n≤x

We would now want an estimate for the LHS.

Lemma 1.3 Weak Stirling’s —

P
P Approximation n≤x log n > x log x − x for all x ∈ N
and in particular, n≤x log n = x log x + O(x) for all x ∈ R.

Proof. The most common and natural way to prove this would be direct integration but
we won’t do that here. Look at the expansion of ex where x > 0 is an integer:

x x2 x3
ex = 1 + + + + ···
1! 2! 3!
xx
Clearly x! is a term in the expansion, therefore

xx X
ex > =⇒ x > x log x − log x! =⇒ log n > x log x − x.
x!
n≤x
P
The lemma is proved now because n≤x log n < x log x is trivial.

Using the above lemma we have that

X Λ(n) X Λ(n)
x log x + O(x) = x + O(x) =⇒ = log x + O(1).
n n
n≤x n≤x

The above relation is clearly useful since the LHS encodes primes in it, and gives us an
estimate of a “global” sum involving primes. What is left is to prove that ψ(x) = O(x).
The main idea is to write
X X X
ψ(x) = log p + log p + log p + · · ·
p≤x p2 ≤x p3 ≤x
X X X
= log p + log p + log p + · · · (?)
√ √
3
p≤x p≤ x p≤ x

For brevity let us define X

θ(x) = log p.
p≤x

3
Ayan Nath A Taste of Analytic Number Theory

This is called the First Chebyshev Function, again the domain here is R. Another
way to motivate the first Chebyshev function is :
X X  
ψ(x) = Λ(n) = log p logp x
n≤x p≤x
 
X log x
= log p + O(1)
log p
p≤x
X
= π(x) log x + O(1) log p .
p≤x
| {z }
θ(x)

The above relation also suggests that θ(x) ∼ x (which is indeed true, but we don’t need
that here). We can write (?) concisely as

ψ(x) = θ(x) + θ(x1/2 ) + θ(x1/3 ) + · · ·

This sum is not infinite, the terms become zero eventually. Note that θ(x) is the “biggest”
term in the RHS and the rest of them are “small”, let us try to prove θ(x) = O(x),
the fact that ψ(x) − θ(x) is “small” (compared to O(x) of course) will be automatically
implied.

Lemma 1.4 — θ(n) < 4n log 2, in particular θ(x) = O(x).

Proof. The main idea is to consider a number which is divisible by many consecutive
primes but the size of the number is not Q too large. A crude example is x!. Clearly the
primes less than x divide x!, therefore p≤x p ≤ x! =⇒ θ(x) ≤ log x! = x log x + O(x),
yes this is a stupidly
trivial bound, but the point is that we want to do something similar.
We consider 2n n . Note that this number is divisible by all primes in the interval
[n + 1, 2n]. Therefore, n<p≤2n p ≤ 2n 2n
Q


n . By binomial theorem, n is trivially bounded
above by (1 + 1)2n = 22n . Therefore, taking logarithms we obtain

θ(2n) − θ(n) ≤ 2n log 2.

So we have that

θ(2k ) − θ(2k−1 ) ≤ 2k log 2

θ(2k−1 ) − θ(2k−2 ) ≤ 2k−1 log 2
..
.
θ(2) − θ(1) < 2 log 2

Summing up, θ(2k ) ≤ 2k+1 log 2. Therefore for general n, it holds that

θ(n) ≤ θ(2dlog2 ne ) ≤ 2dlog2 ne+1 log 2 < 2log2 n+2 log 2 < 4n log 2.

The following implies that ψ(x) − θ(x) is “small”, I suggest you to try to prove this on
your own.

4
Ayan Nath A Taste of Analytic Number Theory

√
Lemma 1.5 — ψ(x) = θ(x) + O( x), and in particular ψ(x) = O(x).

Proof. We have that

ψ(x) = θ(x) + θ(x1/2 ) + θ(x1/3 ) + · · ·

This summation is not infinite, we can write it as

dlog2 xe
1/2
X log x
ψ(x) = θ(x) + θ(x )+ θ(x1/k ) = θ(x) + O(x1/2 ) + O(x1/3 )
log 2
k=3
√
= θ(x) + O( x)

This finally implies that ψ(x) = O(x) since θ(x) = O(x).

And we can now state the result we obtained initially as the proof is now complete:

Theorem 1.6 —
X Λ(n)
= log x + O(1).
n
n≤x

Recall the rough calculation we did for ψ(x), it implied that ψ(x) is roughly π(x) log x.
But we already got ψ(x) = O(x), so this should somehow imply estimates on π(x) right?
Yes, but with some work. Doing the calculations properly:
X X  
ψ(x) = Λ(n) = log p logp x
n≤x p≤x
 
X log x
= log p + O(1)
log p
p≤x

= π(x) log x + O(1)θ(x)

= π(x) log x + O(x).
 
x
Therefore we conclude that π(x) = O log x , this proves that there exist a positive
bx
constant b such that π(x) < log x ,
if you really want to you can get an explicit constant b
easily, but this is not terribly important.

Remark 1.7. By elementary means you can show that b = 1.6 works.

We are left to prove a lower bound. There is a completely elementary proof1 of the
lower bound but here I will discuss a proof which uses the theory developed till now and
also proves Mertens’ first theorem on the fly.
Consider Theorem 1.6, for majority of the terms n ≤ x, Λ(n) is zero. And if n is not
a prime but a power of a prime p, then the denominator of Λ(n) n becomes very large
compared to the numerator, for those who know about p-series convergence it is probably
immediate that the sum of terms when n is not a prime is bounded above by a constant,
or in other words, O(1).

1
https://math.stackexchange.com/a/1890792

5
Ayan Nath A Taste of Analytic Number Theory

P∞
Digression 1.8. Consider Sp = n=1 n1p , called the p−series, it is well known that Sp
converges for p > 1 and diverges for p ≤ 1. Proving this is not hard, see here: https:
//math.stackexchange.com/a/29466

Theorem 1.9 Mertens’ First Theorem —

X log p
= log x + O(1)
p
p≤x

Proof.
X Λ(n) X log p X log p
= +
n p pk
n≤x p≤x pk ≤x,k≥2
X log p X X log p
= +
p pk
p≤x p≤x 2≤k≤logp x

X log p ∞
X X 1
< + log p
p pk
p≤x p≤x k=2
X log p X log p
= +
p p2 − p
p≤x p≤x

P∞ log n
Note that n=2 n2 −n is convergent because

log n n0.1 1 1
2
< 2
= 1.9 0.9
< 1.8
n −n n −n n −n n
holds for all sufficiently large n. We are now done by Theorem 1.6.

Fix a large constant c. The above result implies that

X log p X log p X log p
= − = log c + O(1).
x p p x p
p≤x
c <p≤x p≤ c

Here we pick c very large so that the RHS is positive, say the RHS is bounded below by
δ > 0. We do some trivial bounding,
log xc X log p
(π(x) − π(x/c)) ≥ >δ
x/c x p
c <p≤x

Now just neglect π(x/c) we get

x
π(x) > δ ·
c(log x − log c)
ax
Thus we have established that there exists some a such that π(x) > log x . Again, with
some hard work you can get an explicit constant but that is not very important.

Question 1.10. Put together a logical write-up of the proof of weak PNT.

6
Ayan Nath A Taste of Analytic Number Theory

Example 1.11 (Generalisation of Bertrand’s Postulate)

Let ε > 0. Prove that there exist a prime between n and (1 + ε)n for all large n, in
particular there always exist a prime between n and 2n for n > 1.

Demonstration. Just use PNT for the first part. Proving that there always exist a
prime between n and 2
2n

2n for n > 1 is doable without the full power of PNT though.
Hint: Consider n .

Example 1.12
Fix 1 > ε > 0. For some natural n, let g(n) be the number of divisors of n in
√ √
( n, (1 + ε) n). Prove that g : N → Z≥0 is surjective.

Demonstration.
1. Check that n = pk for some prime p won’t work.

2. Take n = p2a q 2b for two primes p and q.

3. Set some arbitrary k. You want to ensure that pa q b < px q y < (1 + )pa q b has k
solutions in 0 ≤ x ≤ 2a, 0 ≤ y ≤ 2b.

4. Fix x. At most how many possibilities for y are there?

5. It would be nice to have something like this: all such divisors are given by pa q b ( qpt )i
for i = 1, 2, . . . , k. Why do we expect this? When can this happen?

6. Obviously we want p/q t to be very close to 1 and tk = b.

7. Finish using generalised Bertrand’s postulate.

§1.3 Asymptotics for primes

We define pn to be the nth prime number. It will be quite nice to to find a smooth
function f (n) such that pn ∼ f (n). It turns out this is quite easy using PNT, the reader
may try this on their own.

Theorem 1.13 — pn ∼ n log n

Proof. Obviously π(pn ) = n. Therefore

pn pn log pn
∼ n =⇒ ∼ .
log pn n log n log n
But see that
pn log pn log log pn
n∼ =⇒ log n ∼ log pn − log log pn =⇒ ∼1−
log pn log n log pn
Thus it follows that pn ∼ n log n.

This result is useful for ad-hoc calculations to get a feel about whether a statement or a
conjecture should be true. Let me state a result without proof just to summarise:

2
https://www.cut-the-knot.org/arithmetic/algebra/BertrandPostulate.shtml

7
Ayan Nath A Taste of Analytic Number Theory

Theorem 1.14 — The following are equivalent :

• π(x) ∼ x/ log x

• θ(x) ∼ x

• ψ(x) ∼ x

• pn ∼ n log n

Remark 1.15 (Rosser’s theorem). pn > n log n

If you are interested then you may trying proving them. Finally here’s a real olympiad
problem:

Example 1.16 (EMMO 2016 Sr, Anant Mudgal)

We call a sequence of positive integers {an }n∈N as a scouter if it is strictly increasing
and an < 9000n. We call an integer i ≥ 1 as divisor friendly if ai divides the least
common multiple of all previous terms of the sequence and call divisor-unfriendly
otherwise. Is it necessarily true that a scouter has infinitely many

(a) divisor friendly

(b) divisor unfriendly

indices?

Demonstration.

1. Do part (b).

2. Assume that an is divisor unfriendly for all n ≥ N.

3. Look at the prime factorisation of an , precisely, look at the exponents.

4. Conclude that there is a sequence of prime powers dn such that dn | an and di = dj

if and only if i = j.

5. Intuitively, dn should grow faster than 9000n.

6. Prove it by estimating the proportion of prime powers less than some large fixed
number M. (We will discuss this idea in detail in the following section)
√ √
7. You may need to split the summation into the intervals [1, 9000n] and ( 9000n, 9000n]
and bound them separately, there are a lot of other ways to do this though.

8. Conclude.

9. Bonus: strengthen the bound. You can relax the upper bound for an to δn log n
for some sufficiently small δ > 0.

8
Ayan Nath A Taste of Analytic Number Theory

§2 Density
§2.1 Asymptotic Density
Density of a subset S of N refers to the “proportion” of positive integers which are in S.
For example, what is the density of even numbers? Or in other words, what proportion of
positive integers are even numbers? Intuitively, the answer is 12 . This notion is captured
formally as:
Definition 2.1. Let S be a set of positive integers. The asymptotic density of S is
defined as
|S ∩ {1, 2, . . . , n}|
d(S) = lim
n→∞ n
if the limit exists.
This may feel like a mouthful, but what the definition says is that we find the proportion
for a finite n and then take the limit n → ∞. One can also think of this as the probability
that a positive integer chosen at random belongs to S. What is the density of the set of
prime numbers? Zero.
The following example showcases the power of density :

Example 2.2 (China TST 3 2015/3)

Prove that there exist infinitely many integers n such that n2 + 1 is square-free.

Demonstration. The main idea is to estimate the number of positive integers n ≤ N

such that n2 + 1 is square-free using truncated Inclusion Exclusion Principle for some
fixed N.
1. Define AN = {n2 + 1 | n ≤ N, n2 + 1 is square-free}. Which primes divide numbers
of the form n2 + 1?

2. Fix some odd prime p ≡ 1 (mod 4). At most how many multiples of p2 are in AN ?
To find this, first show that the congruence n2 + 1 ≡ 0 (mod p2 ) has at most 2
solutions modulo p2 .

3. What happens when p = 2? How many multiples of 4 are there in AN ?

4. Show that number of non-squarefree numbers in AN is at most

  X N   
X N X 1 N
2 2 ≤2 + 1 = 2N +O
p p2 p2 log N
p≤N, 4|p−1 p≤N p≤N

Notice that we are totally dropping p ≡ 1 (mod 4) for now.

5. Show that the proportion (density) of square-free numbers in AN is at least

X 1  
1 X 1
1−2 − O ∼ 1 − 2 .
p2 log N p2
p≤N p≤N

6. Conclude by proving that

X 1 1
2
< .
p
p 2

9
Ayan Nath A Taste of Analytic Number Theory

Remark 2.3. By tightly bounding, you can show that the density of n such that n2 + 1 is
square-free is at least 0.8924. This means that a positive integer of the form n2 + 1 picked
at random has at least 89.24% chances of being square-free!

Remark 2.4. Note that here we are not proving the existence of the limit for density, it
will be painful to prove the existence every time we want to talk about density. Often we
only care about bounds on the density rather than computing its exact value. To take care
of this issue we define
|S ∩ {1, 2, . . . , n}|
dupper (S) = lim sup
n→∞ n
called the Upper Density of the set S and

|S ∩ {1, 2, . . . , n}|
dlower (S) = lim inf
n→∞ n
called the Lower Density of the set S. So density of a set S exists if and only if
dupper (S) = dlower (S). If you want to be fully rigorous you can replace every word “density”
with whatever seems suitable from the above two in the rest of this article.

Digression 2.5. Define

∞
X 1
ζ(s) = s
,
n=1
n
2
called the Riemann Zeta Function. It is well known that ζ(2) = π6 (buzzword: “Basel
problem”), the reason I am introducing this is because a few contest problems require you
to bound the sum of reciprocals of squares of primes, or in general sum of reciprocals of
squares of some set of positive integers. You can also prove that
Y −1
1
ζ(s) = 1− ,
p
ps

this is known as the Euler Product Formula.

§2.2 Kronecker’s Theorem and Equidistribution Theorem

This section is a bit dense (no pun intended) so take your time.
You may already know what dense means, if not, here is the definition (note that
denseness is defined far more generally, here we only consider R) :
Definition 2.6. Let A be a subset of S ⊆ R, we say that A is dense in S if for every
x ∈ S and ε > 0, there exists an element a ∈ A such that a ∈ (x − ε, x + ε).
For our purposes S will mostly be an interval. Suppose S = [0, 1] and let A ⊆ S be
dense in S. One can think of this as - there are elements of A arbitrarily close to both
0 and 1 and for any a, b ∈ A, there is some c ∈ (a, b) which belongs to A. For example,
the set of rational numbers in [0, 1] is dense in [0, 1]. The way I like to think about this
(in case of intervals of R of course) is that between any two elements of A there exists
another element of A.

Theorem 2.7 Kronecker’s Theorem — Let k be an irrational number. The set of

fractional parts of the terms of the sequence {nk}∞
n=1 is dense in [0, 1].

Proof. This is not difficult. Left as an exercise. See the end of the article for a proof.

10
Ayan Nath A Taste of Analytic Number Theory

Actually, far more is true about the sequence {nk (mod 1)}∞ n=1 (here we write (mod 1)
to denote fractional parts, quite self-explanatory) where k is irrational, it isn’t only dense
in [0, 1] but “uniformly” dense in [0, 1]. “Uniformly” dense is exactly what you think
it means - distributed evenly. So we can say that equidistribution is nicer than simply
being dense. This is defined formally as :

Definition 2.8. Let {an }n≥1 be a sequence of real numbers in the interval [0, 1]. We say
that the sequence is equidistributed if

|{i | 1 ≤ i ≤ n, ai ∈ [a, b]}|

lim =b−a
n→∞ n
holds for all real numbers 0 ≤ a ≤ b ≤ 1.

Question 2.9. Digest the above definition. Prove that Equidistributed =⇒ Dense.

Theorem 2.10 Equidistribution Theorem — Let k be an irrational number. The

sequence of fractional parts of the terms of the sequence {nk}∞
n=1 is equidistributed
in [0, 1].

Demonstration.

1. By Kronecker’s theorem there exists N ∈ N such that {N k} < ε for some very
small ε > 0.

2. Consider the sequence T :

{N k}, {2N k}, {3N k}, . . .

3. Imagine a number line and consider the interval [0, 1]. Plot the sequence T term
by term on the number line. Observe that there will be continuous runs of terms
which belong to I = [a, b] separated by runs of terms which don’t belong to I.

4. About how long are the runs of terms of both types?

5. Fix some large M. What proportion of the first M terms are in I?

6. Do the same thing with

{ik}, {(i + N )k}, {(i + 2N )k}, . . .

for all 1 ≤ i < N − 1.

7. Sum up and conclude.

See the end of the article for a complete proof.

Example 2.11
Find the (asymptotic) density of positive integers n such that 7n begins with the
digits 42 in base-10.

Demonstration.

11
Ayan Nath A Taste of Analytic Number Theory

1. Prove that this is equivalent to having

43 42
log10 > {n log10 7} ≥ log10
10 10

2. Using Equidistribution theorem, {n log10 7} is equidistributed in [0, 1]. Using the

definition of equidistribution, conclude that the required density is
43 42 43
log10 − log10 = log10 .
10 10 42

§3 Mertens’ Second Theorem

Few readers may already that the sum of reciprocals of primes is divergent. It is true
that

Theorem 3.1 weak form of Mertens’ Second Theorem —

X1
= log log n + O(1)
p
p≤n

Question 3.2. For those who know integration, why do we expect the LHS to be asymtotic
to log log n?

See the last page for the proof of one direction.

§4 PNT for Arithmetic Progressions

This section unfortunately will be lack proofs because they are not in the scope of
olympiad mathematics, but these results are very nice, so I decided to include them. We
state a marvellous result without proof :

Theorem 4.1 PNT for Arithmetic Progressions — Let r and d be two relatively prime
positive integers. The number of primes less than x which are congruent to r modulo
d is asymptotic to
1 x
·
ϕ(d) log x
.

One can kind of intuitively see why this should be true - there are ϕ(d) invertible residues
modulo d, namely, those coprime to d. Almost all the other theorems and estimates
change the way you would expect, they are scaled down by ϕ(d).

Theorem 4.2 — If a and d are relatively prime positive integers then

1.
X log p 1
∼ · log x
p ϕ(d)
p≤x
p≡a (mod d)

12
Ayan Nath A Taste of Analytic Number Theory

2. X 1
log p ∼ ·x
ϕ(d)
pk ≤x for some k∈N
p≡a (mod d)

3. X 1
log p ∼ ·x
ϕ(d)
p≤x
p≡a (mod d)

4. X 1 1
∼ · log log x
p ϕ(d)
p≤x
p≡a (mod d)

1
Question 4.3. Why do we expect the scaling by ϕ(d) ?

For further references on this, see [4] or [5].

§5 Examples
I feel that this topic requires more examples than usual so plenty of examples follow
from here, if you get bored feel free to skip to the latter examples or next sections
:)

Example 5.1
Prove that the sequence {bpn νc}∞
n=1 has infinitely many prime divisors where ν is
some positive real number greater than 1.

Demonstration. The main idea here is to look at sum of reciprocals.

1. Show that if {an }∞ i=1 is P
a sequence which has finitely many prime divisors say qi
1
for i = 1, 2, . . . , k, then an is convergent. Use the crude bound:

∞ k  
X 1 Y 1 1
≤ 1 + + 2 + ··· .
a qi qi
i=1 n i=1

2. Conclude.

Example 5.2 (Mathlinks)

Find all polynomials p(x) ∈ Z[x] such that for all positive integers n, we have that
p(n) is a palindrome number.
A number q written in base 10 is called a Palindrome number, if q reads the same
from left to right, as it reads from right to left. For example : 121, −123321 are
Palindrome numbers, but 113 is not a Palindrome number.

Demonstration.
1. Suppose p is non-constant. Let d = deg p > 1.

13
Ayan Nath A Taste of Analytic Number Theory

2. It is pretty intuitive that there exist arbitrarily long runs of consecutive natural
numbers whose p−values start with the same fixed digit.

3. Verify it using p(x) ∼ axd where a is the coefficient of xd .

4. Finish with modulo 10.

Example 5.3 (Canada MO 2020/4)

Let S = {1, 4, 8, 9, 16, . . .} be the set of all perfect powers i.e. S = {nk | n, k ∈ Z, k ≥
2}. We arrange the elements of S into an increasing sequence {ai }∞ i=1 . Show that
there are infinite many positive integers n such that 9999 | an+1 − an .

Demonstration.
1. Find pairs of consecutive perfect squares whose difference is divisible by 9999.
Parametrize to get many such pairs.

2. You want to show that there are infinitely many such pairs of consecutive perfect
squares between which there is no perfect power.

3. Verify that these “pairs” of perfect squares are far denser than perfect odd powers
to conclude.

Example 5.4 (Putnam 2007)

Find all polynomials f with real coefficients such that if n is a positive integer which
is written in base 10 only with ones, then f (n) has the same property.

Demonstration.
n 10xn −1
1. Let f ( 10 9−1 ) = 9 for all n where xn is a sequence of positive integers.

2. Suppose f is non-constant. Let the degree of f be d ≥ 1 and f (x) = axd + o(xd ).

3. So it follows that
10n − 1 a · 10nd
 
f ∼ .
9 9d
4. Show that the sequence xn − nd is convergent. Let the limit be L then see that
a = 9d−1 · 10L .

5. Observe that xn − nd must be eventually constant.

6. Finish the problem.

Example 5.5 (USAMO 2014/6)

Prove that there is a constant c > 0 with the following property: If a, b, n are positive
integers such that gcd(a + i, b + j) > 1 for all i, j ∈ {0, 1, . . . , n}, then
n
min{a, b} > cn · n 2 .

Demonstration.

14
Ayan Nath A Taste of Analytic Number Theory

1. Make an (n + 1) × (n + 1) table with the i, jth entry being the smallest prime
divisor of gcd(a + i, b + j).

2. Try filling up the table with primes. Observe that the primes get large really quick.

3. Take some prime p. Get an upper bound on the number of times p can appear in
the table.

4. Fix some large C. Show that the maximum number of entries that are occupied by
primes at most C is something like
X  n + 1 2
.
p
p≤C

5. Do some bounding and conclude that there exist a constant c > 0 such that at least
50% of the primes in the table are larger than cn2 .

6. Thus there is some row/column with at least half of its primes larger than cn2 .

7. Conclude.

Example 5.6 (Iran 3rd round 2011)

Suppose that α is a real number and a1 < a2 < · · · is a strictly increasing sequence
of natural numbers such that for each natural number n we have an ≤ nα . We
call the prime number q golden if there exists a natural number m such that q|am .
Suppose that q1 < q2 < q3 < · · · are all the golden prime numbers of the sequence
{an }.

(a) Prove that if α = 1.5, then qn ≤ 1390n . Can you find a better bound for qn ?

(b) Prove that if α = 2.4, then qn ≤ 13902n . Can you find a better bound for qn ?

Demonstration. This problem is quite tricky. We only demonstrate part (a), part (b)
is similar.
1. Assume the contrary that there exist qn > 1390n , and take n to be minimal.
Suppose r is the minimal index such that qn | ar .

2. Get a lower bound on

r−1
X 1
S= 1
k=1 ak3
just by using an ≤ n1.5 .

3. Using the fact that all prime factors of the elements of the set {a1 , a2 , . . . , ar−1 }
belong to the set {q1 , q2 , . . . , qn−1 }, get an upper bound on S.

4. Combine and conclude.

5. Strengthen the bound.

6. Try considering r−1

P 1 Pr−1 1
i=1 ak or i=1 ak0.5
, what happens? You may also try taking
Pr−1 −s
i=1 ak for some unspecified s.

15
Ayan Nath A Taste of Analytic Number Theory

Example 5.7 (IMO 2008/3 improved)

Let ε > 0. Prove that there exist infinitely many n such that there is a prime divisor
of n2 + 1 which is larger than (1 − ε)n log n.

Demonstration. This is a pretty hard problem if you are not familiar with some
common methods and ideas in this subject.

1. The key idea is to analyse the product f (N ) = N 2

Q
n=1 (n + 1) (note that this is the
exact same “global” idea discussed in the beginning of the Section 1.2).

2. If p ≤ N is a 1 (mod 4) prime then show that

       
N N N N
νp (f (N )) ≤ 2 + 2 2 + 2 3 + ··· + 2 k
p p p p

where k = dlogp N e. Handle p > N and p = 2 separately.

3. Get an upper bound on f (N ) which looks something like

X X log p X N
log f (N ) ≤ 2 log N 1 + 2N + log p + log 2
p−1 2
p≤N p≤N p≤t
p≡1 (mod 4) p≡1 (mod 4) p≡1 (mod 4)

where t is the largest prime divisor of f (N ).

4. Bound f (N ) from below and conclude.

Remark 5.8. Let c > 0 be a sufficiently small constant. You can try showing that the
set of all positive integers n such that n2 + 1 has a prime divisor larger than cn log n has
positive density.

Remark 5.9. In fact, there exist infinitely many n such that n2 + 1 has a prime divisor
larger than n6/5 . The proof is non-elementary.

Example 5.10 (STEMS 2020 B3/C5, Arka Karmakar)

Let S(a) := {ai + aj | i, j ∈ N}. Find all tuples (a, b, f ) with a, b ∈ N and f ∈ R[x]
such that f (S(a)) ⊆ S(b). (Here f (S(a)) denotes the image of S(a) under f.)

Demonstration. This one is a pretty hard and the solution is a bit involved as well.

1. Guess the solutions. Note that the leading coefficient of f must be positive and
f ∈ Q[x] (Prove it). For now assume that f is non-constant.

2. Consider f (an + a) = bx1 + by1 and f (an + a2 ) = bx2 + by2 where x1 ≥ y1 and
x2 ≥ y2 . We expect x1 = x2 for all large n. Why?

3. Prove the above. Hint: For n → ∞ the ratio of f (an + a) and f (an + a2 ) converges
to 1, this is basically the main intuition rigorised. Make separate cases for b > 2
and b = 2 if needed.

4. Generalise/extend the above.

16
Ayan Nath A Taste of Analytic Number Theory

5. Fix k, pick huge n and let f (an + ai ) = btn + bA(n,i) for all i = 1, 2, 3, . . . , k. Here
we can let it to be tn since it is independent of i for small i.

6. It is not a bad guess that A(n, i) is an arithmetic progression for small i, you should
actually expect this to be true if you have already guessed the solutions.

7. To prove this, analyse νp for some prime p. Writing f (x) = xd (xg(x) + c) might be
helpful (you secretly know what d and c should be).

8. You would probably need this as a lemma: If p | b then p | a for all primes p (Prove
it)

9. Finish the problem.

§6 Problems
All the problems below don’t necessarily use the theory discussed in this article. Many
of the following problems are hard so don’t get demotivated.

§6.1 Exercises
If you are experienced then you may skip this section.

Exercise 6.1. Let A be a set of positive integers with positive asymptotic density. Prove
that sum of reciprocals of elements of A is divergent.

Exercise 6.2. If the density of A ⊂ N and B ⊂ N is zero then prove that density of
A ∪ B is zero.

Exercise 6.3. You are given a string of base-10 digits. Prove that you can append some
finite number of digits so that the resultant number becomes a power of 2.

Exercise 6.4. Define ω(n) to be the number of distinct prime divisors of n. Prove that
X
ω(n) = x log log x + O(x).
n≤x

Exercise 6.5. Prove that

Y 1

1− =0
p
p

where the product is over all primes p.

Exercise 6.6. Let r2 (n) be the number of ways n can be written as a sum of two perfect
squares. Prove that
r2 (1) + r2 (2) + · · · + r2 (n)
lim = π.
n→∞ n
Exercise 6.7 (Mathotsav). We say that a positive integer t is good if the density of
positive integers n such that n2 + t is square-free is at least 0.99.
6
(a) Prove that the density of square free numbers is π2
.

(b) Prove that infinitely many natural numbers are good.

(c) Prove that there exists a positive constant c and a natural number N , such that for
all n > N , the number of natural numbers less than n which are good is at least cn.

17
Ayan Nath A Taste of Analytic Number Theory

§6.2 Easy
Problem 6.8 (Iranian Our MO 2020). Consider two sequences xn = an + b, yn = cn + d
where a, b, c, d are natural numbers and gcd(a, b) = gcd(c, d) = 1, prove that there exist
infinite n such that xn , yn are both square-free.
Problem 6.9 (Iran 3rd round 2010/8). Prove that there are infinitely many natural
numbers of the form n2 + 1 such that they don’t have any divisor of the form k 2 + 1
except 1 and themselves.
Problem 6.10 (China TST 2005). Prove that for any n (n ≥ 2) pairwise distinct
3
fractions in the interval (0, 1), the sum of their denominators is no less than 13 n 2 .
Problem 6.11 (China TST 2004). Let u be a fixed positive integer. Prove that the
equation n! = uα − uβ has a finite number of solutions (n, α, β).
Problem 6.12 (IMO Shortlist 2011/A2). Determine all sequences (x1 , x2 , . . . , x2011 ) of
positive integers, such that for every positive integer n there exists an integer a with
2011
X
jxnj = an+1 + 1.
j=1

Problem 6.13 (China TST 2010, Miklos
Schweitzer, Paul Erdos). Given positive integers
n and k such that n ≥ 9k , prove that nk has at least k different prime divisors.
Problem 6.14 (IMO ShortList 2003/N4). Let b be an integer greater than 5. For each
positive integer n, consider the number

· · · 1} 22
xn = |11 {z · · · 2} 5,
| {z
n−1 n

written in base b.
Prove that the following condition holds if and only if b = 10: there exists a positive
integer M such that for any integer n greater than M , the number xn is a perfect square.
Problem 6.15 (Vesselin Dmitrov). Prove that the set of positive integers n such that
1
2 n(n + 1)(n + 2)(n2 + 1)

is square free has positive density.

Problem 6.16 (Miklos Schweitzer). Prove that the set of positive integers n such that
τ (n) | n has density 0.

§6.3 Medium
Problem 6.17 (ARMO 2012 Grade 11 Day 2). For a positive integer n define Sn =
1! + 2! + . . . + n!. Prove that there exists an integer n such that Sn has a prime divisor
greater than 102012 .
Problem 6.18 (AoPS). Prove that n! = m3 + 8 has only finitely many solutions in
positive integers.
Problem 6.19 (China TST 2 Day 1 P1). Let n be a positive integer. Let Dn be the set
of all divisors of n and let f (n) denote the smallest natural m such that the elements of
Dn are pairwise distinct in mod m. Show that there exists a natural N such that for all
n ≥ N , one has f (n) ≤ n0.01 .

18
Ayan Nath A Taste of Analytic Number Theory

Problem 6.20. (Own & Superguy, KöMaL A. 787) Let pn denote the nth prime number
and define an = bpn νc for all positive integers n where ν is a positive irrational number.
Is it possible that there exist only finitely many k such that 2a by p10
k


ak is divisible i for
all i = 1, 2, . . . , 2020?
Problem 6.21 (Superguy). Prove that the set of positive integers n such that n and
2n − 1 are relatively prime has positive lower density.
Problem 6.22 (Paul Erdos, Miklos Schweitzer). Let a1 < a2 < · · · < an be a sequence
of positive integers such that ai − aj | ai for all i ≤ j. Prove that there is a positive
constant c such that for any such sequence of length n, a1 > ncn .
Problem 6.23. (Tuymaada 2011, Senior Level) Let P (n) be a quadratic trinomial with
integer coefficients. For each positive integer n, the number P (n) has a proper divisor
dn , i.e., 1 < dn < P (n), such that the sequence d1 , d2 , d3 , . . . is increasing. Prove that
either P (n) is the product of two linear polynomials with integer coefficients or all the
values of P (n), for positive integers n, are divisible by the same integer m > 1.
Problem 6.24. (Turkey TST 2015/6) Prove that there are infinitely many positive
integers n such that (n!)n+2015 divides (n2 )!.
Problem 6.25 (China TST 2015). Let a1 , a2 , a3 , . . . be distinct positive integers, and
0 < c < 32 . Prove that : There exist infinitely many positive integers k, such that
lcm(ak , ak+1 ) > ck.

Remark 6.26. The bound cannot be improved to lcm(ak , ak+1 ) > k 1+δ for some δ > 0.

Problem 6.27 (USA TSTST 2017/6). A sequence of positive integers (an )n≥1 is of
Fibonacci type if it satisfies the recursive relation an+2 = an+1 + an for all n ≥ 1. Is it
possible to partition the set of positive integers into an infinite number of Fibonacci type
sequences?
Problem 6.28 (Tuymaada 2007/8). Prove that there exists a positive c such that for
every positive integer N among any N positive integers not exceeding 2N there are two
numbers whose greatest common divisor is greater than cN . (Bonus: Strengthen the
bound)

§6.4 Hard
Problem 6.29 (IMO 2015/N6). Let Z>0 denote the set of positive integers. Consider
a function f : Z>0 → Z>0 . For any m, n ∈ Z>0 we write f n (m) = f (f (. . . f (m) . . .)).
| {z }
n
Suppose that f has the following two properties:
f n (m)−m
(i) if m, n ∈ Z>0 , then n ∈ Z>0 ;
(ii) The set Z>0 \ {f (n) | n ∈ Z>0 } is finite.
Prove that the sequence f (1) − 1, f (2) − 2, f (3) − 3, . . . is periodic.
Problem 6.30 (China TST 2018 Day 2 Q2). Given a positive integer k, call n good if
among        
n n n n
, , ,...,
0 1 2 n
at least 0.99n of them are divisible by k. Show that exists some positive integer N such
that among 1, 2, . . . , N , there are at least 0.99N good numbers.

19
Ayan Nath A Taste of Analytic Number Theory

Problem 6.31 (Paul Erdos). For any δ > 0 prove that there are at least ( 23 − δ) logn n
2
primes between n and 2n for sufficiently large n. (Using the full power of PNT would be
cheating :P)

Problem 6.32 (IMO Shortlist 2019/N7). Prove that there is a constant c > 0 and
infinitely many positive integers n with the following property: there are infinitely many
positive integers that cannot be expressed as the sum of fewer than cn log(n) pairwise
coprime nth powers.

Problem 6.33 (XIII Brazilian Olympic Revenge 2014). Let a > 1 be a positive integer
and f ∈ Z[x] with positive leading coefficient. Let S be the set of integers n such that

n | af (n) − 1.
|S∩{1,...,n}|
Prove that S has density 0; that is, prove that limn→∞ n = 0.

Problem 6.34 (PRIMES 2020 M5). We say an integer n ≥ 2 is chaotic if for any monic
nonconstant polynomial f (x) with positive integer coefficients, the set

{f (1), f (2), . . . , f (n)}

contains fewer than 10deg f · logn n prime numbers. Are there finitely many chaotic integers?

Remark 6.35. There is a theorem by Nagell & Heilbronn which says that for any f ∈ Z[x],
the number of primes in {|f (1)|, |f (2)|, . . . , |f (n)|} is O(n/ log n) but unfortunately the proof
is beyond the scope of olympiad Mathematics.

Problem 6.36 (Marius Cavachi, AMM). Let a and b be integers greater than 1 such
that an − 1 | bn − 1 for every positive integer n. Prove that b is a natural power of a.

Remark 6.37. You can relax the condition to “for infinitely many positive integers n”
instead of “for every positive integer n” and the problem would still hold. However the
proof is non-elementary.

Problem 6.38 (Fedor Petrov). Does there exist c > 0 such that among any n positive
integers one may find 3 with least common multiple at least cn3 ?

20
Ayan Nath A Taste of Analytic Number Theory

§7 Solutions to selected examples

§7.1 Example 1.12
Let’s suppose we want g(n) = k. Choose a large enough natural t such that 2t < q <
2t (1 + ε) where q is a prime. Note that n = q 2k 22kt works because all such k divisors are
of the form q k+i 2t(k−i) for i = 1, 2, . . . , k. No other divisor works because for any fixed
power of q we can have only one power of 2 which may work.

§7.2 Example 1.15 (EMMO 2016 Sr, Anant Mudgal)

Part (b) is easy so we only solve part (a). Assume the contrary that all sufficiently large
indices are divisor friendly. We have that

an - lcm(a1 , a2 , . . . , an−1 )

for all n > K, say. Observe that there must exist some sequence of primes qn such that
ν (a )
if bn = qnqn n for n > K then qn divides none of the preceding terms ai for i > K. See
that all the bn ’s must be distinct. Obviously b1 , b2 , . . . , bn ≤ 9000n and bi are distinct
prime powers. Number of prime powers at most 9000n is less than
X X 1
S= logp 9000n = log 9000n
log p
p≤9000n p≤9000n
 
X 1 X 1 
≤ log 9000n  +
√ log p √ log p
p< 9000n 9000n≤p≤9000n
√ !
9000n 1 √
≤ log 9000n + √ · (π(9000n) − π( 9000n))
log 2 log 9000n
√  !
9000n 1 n
= log 9000n + √ ·O
log 2 log 9000n log n
   
n n
= log 9000n · O 2 =O ,
log n log n

in the last second step we used PNT. This is a contradiction for large enough n since
there are n distinct prime powers at most 9000n, namely b1 , b2 , . . . , bn . And we are done.

§7.3 Example 5.3 (Canada MO 2020/4)

Consider (9999n + 4999)2 and (9999n + 5000)2 , verify that their difference is divisible by
9999, call a pair of such perfect squares good. Fix
√ some large N. Check that the number
of such pairs less than N is bounded below by c N for some constant c > 0. All perfect
powers between such a pair must be odd perfect powers. Number of odd perfect powers
ab less than N is at most

S = N 1/3 + N 1/5 + N 1/7 + · · ·

where the number of summands is at most log2 N as a ≥ 2 except

√ for the trivial perfect
1/3
power 1. Therefore S = O(N log N ), which is less than c N for all large N. Thus
there exists infinitely many good pairs.

21
Ayan Nath A Taste of Analytic Number Theory

§7.4 Example 5.6 (Iran 3rd round 2011)

(Solution by Superguy) We are going to prove the bound qn ≤ 35n for part (a). Let’s
assume for the sake of contradiction that there exists n such that qn > 35n , here suppose
2
n is minimal.Then suppose r is the minimal index such that qn | ar then r > 35 3 n (♣).
So all of {a1 , a2 , . . . , ar−1 } have prime factors in set {q1 , q2 , . . . , qn−1 }. Call this set of
primes as P. We clearly have
r−1
r−1
XX 1
1
1 ≥ √ (1)
3
k=1 ak k=1
k
√
Clearly RHS in (1) is greater than 2 r − 2 which can be shown using easy integration or
induction. Consider the following claim:

Claim —
r−1
X 1 YhX 1
i
1 ≤ p− 3 m ≤ 5(3.27)n−2
3
k=1 ak p∈P m≥0

where |P | = n − 1.

k
Proof. Note that all of ak are of the form q1k1 · q2k2 · · · qn−1
n−1
where all ki are non-negative
1
which gives the left side inequality. For right side we have that the sum m≥0 p− 3 m is
P
maximum for p = 2 and next greatest value is achieved by p = 3 and the value of the
sum for p = 2 is less than 5 and for p = 3 the sum would be less than 3.27 Now observe
YhX 1
i
p− 3 m ≤ 5(3.27)n−2
p∈P m≥0

So we get the claim.

Now by our claim,(1),(♣) and one fact:

n n
ln(2 · 35 3 ) < ln(2 · 35 3 − 2) + 1 for all natural n,

we get that we should have

n · ln(35)
ln(2) − 1 + − ln(5) − (n − 2)(ln(3.27)) < 0 (2)
3
Now we are going to prove the opposite inequality. Taking the function in LHS as f (n)
we get that f (n) is increasing. Hence we just need to check for n = 1 which we get that
f (1) > 0. Thus we have proved the opposite inequality and thus the contradiction for
our initial assumption. For part b exact similar process can give a nice bound of some
qn < 300n . 

§7.5 Example 5.7 (IMO 2008/3 improved)

Define Y
f (N ) = (n2 + 1).
n≤N

Lets assume that the largest prime divisor of f (N ) is t. Let f (N ) = p pαp be the prime
Q

factorisation of f (N ), each prime p > N can divide n2 + 1 for at most two different
values of n, and so αp ≤ 2 in this case. See that α2 = bN/2c . For p ≤ x, if p | n2 + 1,

22
Ayan Nath A Taste of Analytic Number Theory

then n2 ≡ −1 mod p which has solutions if and only if p ≡ 1 (mod 4), and in that case
there will be at most 2dN/pe values of n for which p | n2 + 1. Similarly, if pk | n2 + 1,
then n2 ≡ −1 mod pk , and there are at most 2 solutions to this congruence and hence at
most 2dN/pk e values of n for which pk | n2 + 1. Combining, we find that for p ≤ N and
p ≡ 1 (mod 4)        
N N N N
αp ≤ 2 + 2 2 + 2 3 + ··· + 2 k
p p p p
 
where k = logp N . This gives that

2N

αp ≤ + 2 logp N + 1
p−1
1 1 1 1
since 1 + p + p2
+ ··· + pk
≤ 1−1/p . Thus,

2N
+2 logp (N )+2
Y Y
f (N ) ≤ 2N/2 p p−1 p2 ,
p≤N N <p≤t
p≡1 (mod 4) p≡1 (mod 4)

and so,
Y Y 2N Y
f (N ) ≤ 2N/2 N2 p p−1 p2 .
p≤N p≤N p≤t
p≡1 (mod 4) p≡1 (mod 4) p≡1 (mod 4)

Taking the logarithm

X X log p X N
log f (N ) ≤ 2 log N 1 + 2N + log p + log 2.
p−1 2
p≤N p≤N p≤t
p≡1 (mod 4) p≡1 (mod 4) p≡1 (mod 4)

By PNT for AP and with some computations we can see that the RHS is asymptotic to
N log N + t. Notice that
Y
f (N ) ≥ n2 = (N !)2 = N 2N + O(N ),
n≤N

combining, we get that

2N log N + O(N ) ≤ log f (N ) ≤ N log N + t + o(N log N )

now if t ≤ (1 − ε)N log N for all large N then the above is false for sufficiently large N,
which is what we wanted.

§7.6 Example 5.10 (STEMS 2020 B3/C5, Arka Karmakar)

Clearly b 6= 1. Note that the leading coefficient of f must be positive and f ∈ Q[x]. For
now assume that f is non-constant. Consider the following claims :

Claim 1 — For n, i ∈ N let f (an + ai ) = bt(n,i) + bm(n,i) where t(n, i) ≥ m(n, i). And
let i1 , i2 , i3 , . . . , ik ∈ N, then it follows that t(n, i1 ) = t(n, i2 ) = . . . = t(n, ik ) for all
large n.

23
Ayan Nath A Taste of Analytic Number Theory

Proof. Let i > j be two positive integers. It is obvious that t(n, i) ≥ t(n, j) for all large
n. Observe that
f (an + ai ) bt(n,i) + bm(n,i)
1 = lim = lim
n→∞ f (an + aj ) n→∞ bt(n,j) + bm(n,j)

1 + bm(n,i)−t(n,i)
= lim
n→∞ bt(n,j)−t(n,i) + bm(n,j)−t(n,i)
1
≥ lim −(t(n,i)−t(n,j))
n→∞ b + b−(t(n,i)−m(n,j))
1
≥ lim
n→∞ 2b−(t(n,i)−t(n,j))
1
= lim bt(n,i)−t(n,j) .
2 n→∞
If b > 2 we get t(n, i) = t(n, j) for all sufficiently large n. So let b = 2 then either
t(n, i) = t(n, j) for all sufficiently large n or t(n, i) = t(n, j) + 1 for all sufficiently large
n. We assume the later. Then note that,

bt(n,i) + bm(n,i) 2 + 2m(n,i)−t(n,j) 2

1 = lim t(n,j) m(n,j)
= lim m(n,j)−t(n,j)
≥ lim m(n,j)−t(n,j)
n→∞ b +b n→∞ 1 + 2 n→∞ 1 + 2

which implies that m(n, j) = t(n, j) for all large n. Let i > j > 1. Hence we obtain

f (an + ai ) = 2t(n,j)+1 + 2m(n,i)

f (an + aj ) = 2t(n,j)+1
f (an + a) = 2t(n,1) + 2m(n,1)

for all large n. We now must have t(n, 1) = t(n, j). Now again using the same reasoning
as above we will get m(n, 1) = t(n, 1) which will mean f (an + aj ) = f (an + a) for all
large n. Contradiction! Hence the claim.

Let us introduce some notation : Let i ∈ N and define tn and A(n, i) such that

f (an + ai ) = btn + bA(n,i)

for all large n (here we are using claim 1 and tn is independent of i for small i). Let
f (x) = xd (xg(x) + c) where c 6= 0 and g ∈ Q[x].

Claim 2 — Let p be a prime such that p | b then p | a.

r
Proof. Assume that gcd(a, b) = 1. Let r be some positive integer. Notice that ac1 φ(b )+d1 +
r
ac2 φ(b )+d2 ≡ ad1 + ad2 (mod br ). Therefore if we take c1 , c2 → ∞ then using Claim 1, we
r r
get br | f (ac1 φ(b )+d1 + ac2 φ(b )+d2 ) =⇒ br | f (ad1 + ad2 ). Now taking r to be sufficiently
large we get f (ad1 +ad2 ) which means that f ≡ 0, this is a contradiction to our assumption
that f is non-constant.

Claim 3 — Let N ∈ N be a constant. Then it follows that {A(n, i)}N

i=1 forms an A.P.
for large enough n and a is a power of b.

Proof. Let p | gcd(a, b) be a prime. Now consider

(an + ai )d ((an + ai )g(an + ai ) + c) = btn + bA(n,i)

24
Ayan Nath A Taste of Analytic Number Theory

Taking νp of both sides and n → ∞,

idνp (a) + νp (c)
diνp (a) + νp (c) = A(n, i)νp (b) =⇒ A(n, i) =
νp (b)
Hence the claim. Notice that the above also gives us that νp (b)(A(n, i + 1) − A(n, i)) =
dνp (a), which means both a and b have the same set of prime divisors. Now if a prime q
divides both a and b then by the same reasoning we have that
idνq (a) + νq (c) idνp (a) + νp (c)
A(n, i) = =
νq (b) νp (b)
taking i = 1, 2 we get that
νq (a) νp (a)
= =⇒ a = br
νq (b) νp (b)
for some r ∈ N.
Finishing the problem is easy using the above claim.

§8 Proofs of selected Theorems

§8.1 Theorem 2.7 (Kronecker’s Theorem)
Pick any n ∈ N. By the pigeonhole principle, there are two multiples of k whose fractional
part lie within 1/n of each other (to see this, divide [0, 1] into n equal intervals). Taking
the difference, there is a multiple of k with fractional part less than 1/n. It follows that
every x ∈ [0, 1] is within 1/m of some {nk}, for any m. It is easy to see that we are done.

§8.2 Theorem 2.10 (Equidistribution Theorem)

(Proof by mathcool2009) Define I = [a, b] and let S denote the set of non-negative
integers n for which {nk} ∈ I. We want to show that limn→∞ snn = b − a, where
sn = |S ∩ {1, 2, · · · , n}|.
By Kronecker’s theorem, for any ε > 0 there is a positive integer N for which {N k} < .
Now take an arbitrary integer 0 ≤ i < N and consider the infinite sequence Ti given by
{ik}, {(i + N )k}, {(i + 2N )k}, {(i + 3N )k}, . . .
Assume ε is sufficiently small. Observe that there must be runs of terms in I separated
with
 b−a runs b−a
of terms
 not in I. It is easy to see that the runs of terms in I has length
j L ork Ll where mL = {N k}. Similarly, the runs of terms not in I has length either
1−(b−a) 1−(b−a)
L or L . This means that in the limit, between
     
b−a b−a 1 − (b − a)
u− = / +
L L L
and      
b−a b−a 1 − (b − a)
u+ = / +
L L L
of the terms of Ti are in I. (More precisely, let tn denote the number of terms of Ti that are
in I within the first n terms of Ti . Then lim inf n→∞ tnn ≥ u− and lim supn→∞ tnn ≤ u+ .)
Of course, as we sum over i, we see that we must have lim inf n→∞ snn ≥ u− and
lim supn→∞ snn ≤ u+ .
Finally, since ε was arbitrary, we can choose ε small enough such that u− and u+ both
approach b − a. Hence lim inf n→∞ snn = lim supn→∞ snn = b − a, as desired.

25
Ayan Nath A Taste of Analytic Number Theory

§8.3 Theorem 3.1 (Mertens’ Second Theorem)

Theorem weak form of Mertens’ Second Theorem —

X1
= log log n + O(1)
p
p≤n

We only show the following:

X1
≥ log log n + O(1)
p
p≤n

The proof is motivated from the Euler product formula for the ζ function.
Demonstration.

1. Look at Digression 2.5.

2. You would want to somehow “truncate” the Euler product formula so that you get
information about sum of reciprocals of primes till n.

3. Show that −1

Y 1 X1
log 1− = + O(1)
p p
p≤n p≤n

using the Taylor series of log(1 − x).

4. Get a lower bound on p≤n (1 − p1 )−1 .

Q

5. Conclude.

Readers interested in the complete proof may have a look at Abel Summation Formula3
and start with Mertens’ First Theorem.

§9 Acknowledgements
I am very thankful to Superguy4 for valuable suggestions and for being a problem
resource.

References
[1] https://artofproblemsolving.com

[2] https://math.stackexchange.com

[3] An Introduction to The Theory of Numbers, Ivan Niven, Herbert S. Zuckerman, Hugh
L. Montgomery, Wiley India Pvt. Ltd.

[4] Problems from the Book, Titu Andreescu, Gabriel Dospinescu, XYZ Press

[5] Straight from the Book, Titu Andreescu, Gabriel Dospinescu, XYZ Press

3
https://en.wikipedia.org/wiki/Abel%27s_summation_formula
4
AoPS user : https://artofproblemsolving.com/community/user/388865 v2.1

26