Olympiad Maths Prep

Track / Stage 8 / 65 of 180 #1765 of 2000

Problem 1765

IMO Shortlist mid-range; USAMO P2/P5
Number theory Difficulty 8.2 Prove it

Let PP be a polynomial of degree greater than or equal to 44 with integer coefficients. An integer xx is called PP-[i]representable[/i] if there exists integer numbers aa and bb such that x=P(a)P(b)x = P(a) - P(b). Prove that, if for all N0N \geq 0, more than half of the integers of the set {0,1,,N}\{0,1,\dots,N\} are PP-[i]representable[/i], then all the even integers are PP-[i]representable[/i] or all the odd integers are PP-[i]representable[/i].

This one wants a proof. Work it on paper, read the official solution, then mark yourself honestly — the ladder only means something if the record is true.

Official solution

1. **First Claim: The degree of P P is even.**

Proof: Suppose not, i.e., degP=n \deg P = n with n n odd. Then there exist real numbers u u and v v such that if x>v x > v , then P(x)>P(v) P(x) > P(v) , and if x<u x < u , then P(x)<P(u) P(x) < P(u) . This is possible by taking u u and v v such that P(v) P(v) is strictly greater than all local maxima and P(u) P(u) is strictly smaller than all local minima.

There exists a constant C C such that if Q(x)=P(x+1)P(x) Q(x) = |P(x+1) - P(x)| , then Q(x)>Cxn1 Q(x) > C|x|^{n-1} for all x>C0>0 |x| > C_0 > 0 , where all the roots of Q Q belong to (C0/2,C0/2) (-C_0/2, C_0/2) .

Now take a large M M such that M>u+v+C+1 M > |u| + |v| + C + 1 . Fix a large integer N N . For each pair of integers (x,y) (x, y) with xy x \neq y such that P(x)P(y)N |P(x) - P(y)| \leq N (denoted as () (\star) ), we classify them as:
- Good: if x,yM |x|, |y| \geq M
- Bad: if exactly one of x,y |x|, |y| is less than M M
- Whatever: if both have sizes smaller than M M

Clearly, the number of Whatever Whatever pairs is finite (for all N N ). So we shall not worry about them.

Now take a bad bad pair. Suppose WLOG that xM |x| \geq M . Then
N>P(x)P(y)P(x)P(x+ϵ)>Cxn1, N > |P(x) - P(y)| \geq |P(x) - P(x+\epsilon)| > C|x|^{n-1},
where ϵ{1,1} \epsilon \in \{-1, 1\} , and it's 1 1 if x<0 x < 0 and 1 -1 if x>0 x > 0 .

Now take a good pair. WLOG x>y |x| > |y| . So we know that
N>P(x)P(y)P(x)P(x+ϵ)>Cxn1. N > |P(x) - P(y)| \geq |P(x) - P(x+\epsilon)| > C|x|^{n-1}.

So we must conclude that if (x,y) (x, y) are a solution to () (\star) , then both x |x| and y |y| are smaller than C1N1n1 C_1 N^{\frac{1}{n-1}} (where C11n=C C_1^{1-n} = C ). So the total number of solutions must be at most 4C12N2n1 4C_1^2 N^{\frac{2}{n-1}} .

But we know that the number of solutions is at least N/2 N/2 . So sending N N to infinity, we must have that
2n11    n3, \frac{2}{n-1} \geq 1 \implies n \leq 3,
which is a contradiction. Hence, the degree of P P must be even.

2. **Second Claim: Given that the degree of P P is even, the majority of the solutions (x,y) (x, y) for () (\star) have xy<0 xy < 0 .**

Proof: The number of solutions where xy0 xy \geq 0 grows smaller than N N . In fact, we can prove just as we did in Claim 1 that it grows with degree 2n1 \frac{2}{n-1} . The proof uses the same arguments as in Claim 1, where we lower their difference by bringing one of them close to the other and then use the lower bound of degree n1 n-1 .

3. **Third Claim: There exists a unique rational number L L such that the polynomial P(x)P(Lx) P(x) - P(L-x) has degree at most n2 n-2 .**

Proof: Let P(x)=anxn++a0 P(x) = a_n x^n + \dots + a_0 . We can write the polynomial P(x)P(Ax) P(x) - P(A-x) explicitly as
k=0nakxkk=0naki=0kxiAki(ki), \sum_{k=0}^{n} a_k x^k - \sum_{k=0}^n a_k \sum_{i=0}^k x^i A^{k-i} \binom{k}{i},
which can be rewritten as
k=0nxk(aki=0nkai+kAi(i+kk)). \sum_{k=0}^n x^k \left( a_k - \sum_{i=0}^{n-k} a_{i+k} A^i \binom{i+k}{k} \right).
Then P(x)P(Ax) P(x) - P(A-x) has always degree at most n1 n-1 , as the coefficient of xn x^n is 1(1)n=0 1 - (-1)^n = 0 , since n n is even.

The coefficient of xn1 x^{n-1} is 2an1+An 2a_{n-1} + An , which is 0 0 if and only if A=2an1n A = -\frac{2a_{n-1}}{n} . This is our desired rational L L .

So we conclude that the polynomial P(x)P(Lx) P(x) - P(L-x) has degree at most n2 n-2 . In other words, there is a constant C C such that P(x)P(Lx)Cxn2 |P(x) - P(L-x)| \leq C|x|^{n-2} for every x>0.1 |x| > 0.1 .

4. **Fourth Claim (Naughty Lemma): We say that a pair (x,y) (x, y) is naughty naughty if it satisfies () (\star) , x,y>M |x|, |y| > M , xy<0 xy < 0 , and x+yL x + y \neq L .**

Proof: Suppose WLOG xy |x| \geq |y| . The intuition for proving this claim is that you can't directly come up with a bound for this difference, and as xy<0 xy < 0 , you use the fact that P(Ly) P(L-y) is somehow close to P(y) P(y) , to then be able to compare their sizes. In other words, we will use the triangular inequality to say that:
P(x)P(y)P(x)P(Ly)P(y)P(Ly). |P(x) - P(y)| \geq |P(x) - P(L-y)| - |P(y) - P(L-y)|.
We already have an upper bound for P(y)P(Ly) |P(y) - P(L-y)| , so we're left with finding a bound for P(x)P(Ly) |P(x) - P(L-y)| . Let t=1 t = 1 if L L is an integer, and let t=min{{L},1{L}} t = \min\{\{L\}, 1-\{L\}\} if L L is not an integer. So we can say that if Ly<x L-y < x , then P(x)P(Ly)P(x)P(xt) |P(x) - P(L-y)| \leq |P(x) - P(x-t)| . If Ly>x L-y > x , then P(x)P(Ly)P(x)P(x+t) |P(x) - P(L-y)| \leq |P(x) - P(x+t)| . In other words, since x,Ly>M |x|, |L-y| > M , we know that bringing them together will decrease the difference, as x(Ly)>0 x(L-y) > 0 . And, as x,y x, y are integers, we know that the difference between x x and Ly L-y is at least t t . Since t>0 t > 0 , we can find a constant C2 C_2 such that both P(x)P(x+t) |P(x) - P(x+t)| and P(x)P(xt) |P(x) - P(x-t)| are smaller than C2xn1 C_2 |x|^{n-1} . So we find that
N>P(x)P(y)P(x)P(Ly)P(y)P(Ly)C2xn1Cyn2C3xn1 N > |P(x) - P(y)| \geq |P(x) - P(L-y)| - |P(y) - P(L-y)| \geq C_2 |x|^{n-1} - C|y|^{n-2} \geq C_3 x^{n-1}
for some constant C3 C_3 and large x>M0 |x| > M_0 . (The choice of M M does not interfere with M0 M_0 , so we may suppose as well that M>M0 M > M_0 ). So this means that if a pair is naughty, then both (x,y) (x, y) have to be smaller than C4N1n1 C_4 N^{\frac{1}{n-1}} for some constant C4 C_4 , meaning that there are at most 4C42N2n1 4C_4^2 N^{\frac{2}{n-1}} naughty pairs.

5. **Conclusion: Combining everything we proved so far, we must conclude that there is a constant C5 C_5 such that the number of solutions (x,y) (x, y) for () (\star) with x+yL x + y \neq L is at most C5N2n1 C_5 N^{\frac{2}{n-1}} .**

Then take N N sufficiently large such that N/2.5<N/2C5N2n1 N/2.5 < N/2 - C_5 N^{\frac{2}{n-1}} .

So for N N this large, the number of solutions for (x,y) (x, y) with x+y=L x + y = L must be at least N/2.5 N/2.5 . Now let Q(x)=P(x)P(Lx) Q(x) = P(x) - P(L-x) . So this directly implies that L L must be an integer, and it also means that the polynomial Q Q covers N/2.5 N/2.5 integers up to N N . Let d d be the degree of Q Q . Clearly d>0 d > 0 . We know that there is a constant C6 C_6 such that Q(x)>C6xd |Q(x)| > C_6 |x|^d for every non-zero x x . This means that if Q(x)<N Q(x) < N , then x<C7N1d x < C_7 N^{\frac{1}{d}} for some constant C7 C_7 . Sending N N to infinity, we conclude that d1 d \leq 1 , so d=1 d = 1 .

Then Q(x)=ax+b Q(x) = ax + b for some integers a a and b b . (We may suppose WLOG a>0 a > 0 . If not, we may switch P P by P -P ). But we know that if 0<ax+b<N 0 < ax + b < N , then b/a<x<N/ab/a -b/a < x < N/a - b/a . So there are at most N/a+1 N/a + 1 solutions for () (\star) , which implies a2.5    a2 a \leq 2.5 \implies a \leq 2 . If a=1 a = 1 , then Q Q covers all integers. If a=2 a = 2 , then Q Q must cover all even or all odd integers, concluding the problem.

\blacksquare

Source: NuminaMath-1.5, licensed Apache-2.0. Statement and solution reproduced as published; topic, difficulty and ordering added by this site.