Olympiad Maths Prep

Track / Stage 8 / 17 of 180 #1717 of 2000

Problem 1717

IMO Shortlist mid-range; USAMO P2/P5
Number theory Difficulty 8.0 Prove it USAMO · United States

Determine which integers n>1n > 1 have the property that there exists an infinite sequence a1,a2,a3,a_1, a_2, a_3, \dots of nonzero integers such that the equality
ak+2a2k++nank=0 a_k + 2a_{2k} + \dots + n a_{nk} = 0
holds for every positive integer kk.

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 solutions — 2

Solution 1

We will show that the sequence exists for all n3n \ge 3.

For n=2n = 2, the sequence cannot exist. If it existed, we would have ak=2a2ka_k = -2a_{2k} for all kk, from which a1=(2)ra2ra_1 = (-2)^r a_{2r} for all rr by induction. Then a1a_1 would have to be divisible by 2r2^r for all rr, which is impossible for a10a_1 \ne 0.

Now fix n3n \ge 3. We will show that the desired sequence exists. The construction is a repeated application of the Chinese Remainder Theorem, but the details require substantial care.

First we prove two lemmas.

Lemma 1. It is possible to partition the positive integers into subsets S1,S2,S3,S_1, S_2, S_3, \dots so that for every positive integer kk,
(i) the numbers (n1)k(n-1)k and nknk are in the same subset, and
(ii) the numbers k,2k,,(n2)kk, 2k, \dots, (n-2)k are all in strictly earlier subsets than (n1)k(n-1)k.

*Proof.* Define a function ff from the positive integers to the positive reals as follows. Let PP be the set of primes dividing nn. No element of PP divides n1n-1. For any positive integer kk, write its prime factorization k=p1e1p2e2prerk = p_1^{e_1} p_2^{e_2} \cdots p_r^{e_r}, and then define
f(k)=piPpieipiP(piei)logn(n1). f(k) = \prod_{p_i \notin P} p_i^{e_i} \cdot \prod_{p_i \in P} (p_i^{e_i})^{\log_n(n-1)}.
Notice that for every kk, we have
f((n1)k)=(n1)f(k)=f(nk),(4) f((n-1)k) = (n-1)f(k) = f(nk), \qquad (4)
whereas for each t=1,2,,n2t = 1, 2, \dots, n-2, we have
f(tk)tf(k)<f((n1)k).(5) f(tk) \le tf(k) < f((n-1)k). \qquad (5)
Also notice that for each kk, f(k)klogn(n1)f(k) \ge k^{\log_n(n-1)}, which implies that for any fixed CC, there can only be finitely many values of kk with f(k)<Cf(k) < C. Therefore, we may arrange the elements of the image of ff in the increasing order x1<x2<x3<x_1 < x_2 < x_3 < \dots. Now let Si=f1(xi)S_i = f^{-1}(x_i) for each ii. The sets SiS_i are a partition of the positive integers, and (4) and (5) ensure that they satisfy (i) and (ii), respectively. \square

Lemma 2. Let pp and qq be relatively prime positive integers and t1,t2,,trt_1, t_2, \dots, t_r arbitrary integers. Then it is possible to choose nonzero integers b1,b2,,br+1b_1, b_2, \dots, b_{r+1} such that
pbi+qbi+1=tifor i=1,2,,r.(6) pb_i + qb_{i+1} = t_i \quad \text{for } i = 1, 2, \dots, r. \qquad (6)
*Proof.* We use induction on rr. If r=1r=1, then since pp and qq are relatively prime, we can find c,dc, d such that pc+qd=1pc+qd=1. Then b1=ct1b_1 = ct_1 and b2=dt1b_2 = dt_1 satisfy (6). Now suppose we have b1,,brb_1, \dots, b_r satisfying (6) for i=1,2,,r1i=1, 2, \dots, r-1. If we choose any integer kk and replace each bib_i with bi=bi+(1)ipi1qrikb'_i = b_i + (-1)^i p^{i-1} q^{r-i} k, then (6) still holds for i=1,2,,r1i=1, 2, \dots, r-1, and pbr=pbr+(1)rpr1kpb'_r = pb_r + (-1)^r p^{r-1} k. Since pp and qq are relatively prime, we can choose kk so as to make pbrpb'_r congruent to trt_r modulo qq, and then we take br+1=(trpbr)/qb_{r+1} = (t_r - pb'_r)/q. Then the numbers b1,b2,,br,br+1b'_1, b'_2, \dots, b'_r, b_{r+1} satisfy (6) for i=1,2,,ri=1, 2, \dots, r.

This shows that we can find b1,b2,,br+1b_1, b_2, \dots, b_{r+1} satisfying (6), but they may not all be nonzero. However, once again, we can make the replacements bi=bi+(1)ipi1qr+1ikb'_i = b_i + (-1)^i p^{i-1} q^{r+1-i} k for any integer kk, and the new sequence still satisfies (6). By an appropriate choice of kk, we can ensure each bib'_i is nonzero. \square

Now both lemmas are proven, and we resume the main proof. We will construct terms of the sequence inductively, but not in the order a1,a2,a_1, a_2, \dots.

Suppose SS is any set of positive integers, and we have chosen nonzero integers aka_k for each kSk \in S. Say that there is a *conflict* in SS if there exists some kk such that k,2k,,nkk, 2k, \dots, nk are all in SS, and
ak+2a2k++nank0. a_k + 2a_{2k} + \dots + n a_{nk} \neq 0.
Let S1,S2,S_1, S_2, \dots be as given by Lemma 1. We will inductively define our sequence as follows:

a. *Step 1:* Choose nonzero values aka_k for all kS1k \in S_1 simultaneously, without creating a conflict in S1S_1.

b. *Step t>1t > 1:* Given the values of aka_k for kS1St1k \in S_1 \cup \cdots \cup S_{t-1} chosen at previous steps, choose nonzero integers aka_k for all kStk \in S_t simultaneously, without creating a conflict in S1StS_1 \cup \cdots \cup S_t.

If we can show that each step of this process can indeed be carried out, then it will eventually define aka_k for all positive integers kk, satisfying the required condition
ak+2a2k++nank=0(7) a_k + 2a_{2k} + \cdots + n a_{nk} = 0 \qquad (7)
for all kk (since no conflicts are created).

For Step 1, Lemma 1 implies we can choose aka_k arbitrarily for kS1k \in S_1 without creating any conflicts, since (n1)k,nkS1(n-1)k, nk \notin S_1 for all kk. Now for Step t>1t > 1, suppose the aka_k have been assigned already for all kS1S2St1k \in S_1 \cup S_2 \cup \cdots \cup S_{t-1}. We need to assign aka_k for kStk \in S_t without creating any new conflicts. This just requires that the new assignments satisfy (7) for all integers kk such that (n1)k(n-1)k and nknk are in StS_t. For any other value kk, either {k,2k,,nk}S1St\{k, 2k, \dots, nk\} \nsubseteq S_1 \cup \cdots \cup S_t so no conflict can be created, or else Lemma 1 implies {k,2k,,nk}S1St1\{k, 2k, \dots, nk\} \subseteq S_1 \cup \cdots \cup S_{t-1} so that the corresponding constraint (7) has been dealt with at an earlier step.

Thus for each kk such that (n1)k,nkSt(n-1)k, nk \in S_t, we have a constraint
(n1)a(n1)k+nank=Xk,(8) (n-1)a_{(n-1)k} + n a_{nk} = X_k, \qquad (8)
where Xk=(ak++(n2)a(n2)k)X_k = -(a_k + \cdots + (n-2)a_{(n-2)k}) is determined by the assignments made at previous steps. We just need to show that it is possible to choose aka_k for all kStk \in S_t such that all these constraints are satisfied.

Form a directed graph whose vertices are the elements of StS_t, with an edge leading from (n1)k(n-1)k to nknk whenever both numbers are in StS_t. Then every component of this graph is either a single vertex or a (directed) path. We wish to show that nonzero integer values can be assigned to elements of StS_t so that for each edge, the corresponding constraint (8) is satisfied. It suffices to show this for each component of the graph. If the component is a single vertex, any nonzero value works. Otherwise, it is a path k1,k2,,kr+1k_1, k_2, \dots, k_{r+1}, and Lemma 2 ensures that we can choose nonzero integer values for ak1,ak2,,akr+1a_{k_1}, a_{k_2}, \dots, a_{k_{r+1}} so as to satisfy (8) for each edge.

This shows that each step of our constructive process can indeed be performed successfully, eventually constructing every term of the sequence.

Solution 2

(By Dai Yang). We claim that such a sequence exists for all nn except n=2n = 2. If we assume that n=2n = 2, then
a1=2a2=4a4=8a8= a_1 = -2a_2 = 4a_4 = -8a_8 = \dots
which is impossible since a1a_1 cannot be divisible by arbitrarily large powers of 2. Thus no valid sequence exists for n=2n = 2.

For n3n \ge 3, we begin with the following lemma.

Lemma 3. If there exists a multiplicative function f(x)f(x) from N\mathbb{N} to Z\mathbb{Z} such that f(x)0f(x) \ne 0 for all xx and
f(1)+2f(2)++nf(n)=0, f(1) + 2f(2) + \cdots + n f(n) = 0,
then there exists a sequence {ai}\{a_i\} satisfying the desired conditions.

*Proof.* Assume that f(x)f(x) satisfies the hypothesis. Let ak=f(k)a_k = f(k) for all kk. Since f(x)f(x) is multiplicative, we have for all positive integers kk that
ak+2a2k++nank=f(k)+2f(2k)++nf(nk)=kf(1)+2kf(2)++nkf(n)=k(f(1)+2f(2)++nf(n))=0. \begin{aligned} a_k + 2a_{2k} + \dots + n a_{nk} &= f(k) + 2f(2k) + \dots + n f(nk) \\ &= k f(1) + 2k f(2) + \dots + n k f(n) \\ &= k(f(1) + 2f(2) + \dots + n f(n)) \\ &= 0. \end{aligned} \quad \square

We now use Lemma 3 to establish the following lemma.

Lemma 4. If there exist primes pp and qq satisfying n<q<pn\sqrt{n} < q < p \le n and p>n2p > \frac{n}{2}, then there exists a sequence {ai}\{a_i\} satisfying the desired conditions.

*Proof.* Let pp and qq be given. For a prime rr, define vr(k)v_r(k) to be the largest exponent vv such that rvkr^v \mid k. Let f(x)=avp(x)bvq(x)f(x) = a^{v_p(x)} b^{v_q(x)} for all positive integers xx, where aa and bb are nonzero integers. Clearly f(x)f(x) is a multiplicative function. We will find suitable values for aa and bb so that f(x)f(x) satisfies the conditions of Lemma 3.

Let the multiples of qq that are less than or equal to nn be q,2q,,mqq, 2q, \dots, m q for some mm. Since n<q\sqrt{n} < q, we have m<qm < q. Note also that the only multiple of pp that is less than or equal to nn is pp itself. Hence, for all \ell satisfying 1n1 \le \ell \le n, we have
f()={aif =pbif =q,2q,,mq1otherwise. f(\ell) = \begin{cases} a & \text{if } \ell = p \\ b & \text{if } \ell = q, 2q, \dots, m q \\ 1 & \text{otherwise.} \end{cases}
Therefore, for some constant dd, we have
f(1)+2f(2)++nf(n)=ap+bq+2bq++mbq+d=ap+bqm(m+1)2+d.(9) f(1) + 2f(2) + \dots + n f(n) = a p + b q + 2 b q + \dots + m b q + d = a p + b \frac{q m(m+1)}{2} + d. \quad (9)
Note that m<q<pm < q < p, so that gcd(p,qm(m+1)2)=1\text{gcd}(p, \frac{q m(m+1)}{2}) = 1. By Bézout's Identity, there exist integers aa' and bb' for which ap+bqm(m+1)2=1a' p + b' \frac{q m(m+1)}{2} = 1, hence after appropriate scaling, we obtain aa and bb so that
ap+bqm(m+1)2=d. a p + b \frac{q m(m+1)}{2} = -d.
Then (9) reduces to
f(1)+2f(2)++nf(n)=0. f(1) + 2f(2) + \dots + n f(n) = 0.
Note that we can also stipulate that aa and bb be nonzero. Therefore, f(x)f(x) satisfies the hypothesis of Lemma 3, so a valid sequence {ai}\{a_i\} exists. \square

It now suffices to list for each n3n \ge 3 either a function f(x)f(x) satisfying Lemma 3 or a pair of primes pp and qq satisfying Lemma 4.

* If n=3n = 3, let f(x)=(1)v3(x)f(x) = (-1)^{v_3(x)}. Note that f(1)+2f(2)+3f(3)=0f(1) + 2f(2) + 3f(3) = 0, as required by Lemma 3.
* If n=4n = 4, let f(x)=(1)v2(x)(1)v3(x)f(x) = (-1)^{v_2(x)}(-1)^{v_3(x)}.
* If n=5,6,7,8n = 5, 6, 7, 8, let p=5p = 5 and q=3q = 3.
* If n=9,10n = 9, 10, let p=7p = 7 and q=5q = 5.
* If 11n1611 \le n \le 16, let p=11p = 11 and q=7q = 7.
* If n>16n > 16, by Bertrand's Postulate, we can find a prime pp satisfying n2<pn\frac{n}{2} < p \le n, and a prime qq satisfying n4<qn2\frac{n}{4} < q \le \frac{n}{2}. For all n>16n > 16, we have n4>n\frac{n}{4} > \sqrt{n}, so pp and qq satisfy the conditions of Lemma 4.

This exhausts all values of n3n \ge 3, so we are done.

Source: MathNet, licensed CC-BY-4.0. Statement and solution reproduced as published; topic, difficulty and ordering added by this site.