Maths Olympiad Prep

Library / /436 of 462

Algebra Difficulty 7.3 National Olympiad, round 2 Prove it Ireland

Suppose a doubly infinite sequence of real numbers
,a2,a1,a0,a1,a2, \dots, a_{-2}, a_{-1}, a_0, a_1, a_2, \dots
has the property that
an+3=an+an+1+an+23,for all integers n. a_{n+3} = \frac{a_n + a_{n+1} + a_{n+2}}{3}, \quad \text{for all integers } n.
Show that if this sequence is bounded (i.e., if there exists a number RR such that anR|a_n| \le R for all nn), then ana_n has the same value for all nn.

Solutions — 3

Solution 1

Assume that the sequence is not constant. Define a function D:Z[0,)D: \mathbb{Z} \to [0, \infty) by the equation
D(n)=an+1an+an+2an+1+an+2an, D(n) = |a_{n+1} - a_n| + |a_{n+2} - a_{n+1}| + |a_{n+2} - a_n|,
or equivalently,
D(n)=2(xnyn),(3) D(n) = 2(x_n - y_n), \quad (3)
where xnx_n and yny_n are the largest and the smallest, respectively, of the three numbers an,an+1,an+2a_n, a_{n+1}, a_{n+2}.
Note that if D(n)=0D(n) = 0 for any one value of nn, then the original sequence would be constant. Thus by our assumption, we must have D(n)>0D(n) > 0 for all nn. It is also trivial that (an)(a_n) is unbounded if the function DD is unbounded. We define [x,y][x, y] to be the closed interval of values between the numbers xx and yy. Unlike the usual meaning, we do not insist that xyx \le y, so [x,y]=[y,x][x, y] = [y, x] is always an interval of length yx|y - x|.

Claim 1: DD is monotonically decreasing.
By translation invariance of the indices, it suffices to show that D(0)D(1)D(0) \ge D(1). Bearing in mind that a3a_3 is an arithmetic average of a1,a2a_1, a_2, and a0a_0, we see that if a0[a1,a2]a_0 \in [a_1, a_2], then a3[a1,a2]a_3 \in [a_1, a_2]. Alternatively, if a0[a1,a2]a_0 \notin [a_1, a_2], then either a3[a1,a2]a_3 \in [a_1, a_2] or both a0a_0 and a3a_3 are on the same side of [a1,a2][a_1, a_2], with a3a_3 being strictly closer than a0a_0 to [a1,a2][a_1, a_2].
By the above considerations and (3), we deduce the following. If a0[a1,a2]a_0 \in [a_1, a_2], then D(1)=D(0)=2a2a1D(1) = D(0) = 2|a_2 - a_1|. If instead a0[a1,a2]a_0 \notin [a_1, a_2], then either a3[a1,a2]a_3 \in [a_1, a_2], in which case D(0)D(1)=2dist(a0,[a1,a2])>0D(0) - D(1) = 2 \text{dist}(a_0, [a_1, a_2]) > 0, or a3[a1,a2]a_3 \notin [a_1, a_2], in which case D(0)D(1)=2a0a3>0D(0) - D(1) = 2|a_0 - a_3| > 0. This shows that D(0)D(1)D(0) \ge D(1) with equality iff a0[a1,a2]a_0 \in [a_1, a_2].

Claim 2: We cannot have an[an+1,an+2]a_n \in [a_{n+1}, a_{n+2}] for two consecutive values of nn.
As in Claim 1, we reduce to the statement that if a0[a1,a2]a_0 \in [a_1, a_2], then a1[a2,a3]a_1 \notin [a_2, a_3]. Suppose a0[a1,a2]a_0 \in [a_1, a_2]. As noted above, we then have a3[a1,a2]a_3 \in [a_1, a_2], which implies that a1[a2,a3]a_1 \notin [a_2, a_3] unless a1=a3a_1 = a_3. But if a1=a3a_1 = a_3, then our assumption that a0[a1,a2]a_0 \in [a_1, a_2] forces a0=a1=a2a_0 = a_1 = a_2, contradicting the initial assumption that (an)(a_n) is non-constant. Claim 2 now follows.
Claim 3: D(n+3)<(2/3)D(n)D(n+3) < (2/3)D(n), nNn \in \mathbb{N}.
In view of Claims 1 and 2, it suffices to show that D(n+2)<2D(n)/3D(n+2) < 2D(n)/3 whenever an[an+1,an+2]a_n \notin [a_{n+1}, a_{n+2}]. As usual, we may assume that n=0n=0. The function DD is invariant under translations of the sequence, and under multiplication by 1-1. So we may assume that a0=0a_0 = 0, and that a:=a1a := a_1 and b:=a2b := a_2 are both positive. Since
a3=a+b3anda4=4a+4b9, a_3 = \frac{a+b}{3} \quad \text{and} \quad a_4 = \frac{4a+4b}{9},
it follows that
D(0)=a+ba+b, D(1)=ba+a2b3+b2a3, D(2)=a2b3+a+b9+4a5b9. \begin{align} D(0) &= a + |b-a| + b, \notag \ D(1) &= |b-a| + \frac{|a-2b|}{3} + \frac{|b-2a|}{3}, \tag{4} \ D(2) &= \frac{|a-2b|}{3} + \frac{|a+b|}{9} + \frac{|4a-5b|}{9}. \notag \end{align}
If bab \ge a, then by (4), we have D(0)=2bD(0) = 2b and
D(2)=2ba3+a+b9+5b4a9=12b6a9<4b3, D(2) = \frac{2b-a}{3} + \frac{a+b}{9} + \frac{5b-4a}{9} = \frac{12b-6a}{9} < \frac{4b}{3},
and so D(2)<2D(0)/3D(2) < 2D(0)/3, as required.
If instead a/2b<aa/2 \le b < a, then D(0)=2aD(0) = 2a and
D(1)=(ab)+2ba3+2ab3=4a2b3<4a3. D(1) = (a-b) + \frac{2b-a}{3} + \frac{2a-b}{3} = \frac{4a-2b}{3} < \frac{4a}{3}.
By Claim 1, we conclude that D(2)D(1)<2D(0)/3D(2) \le D(1) < 2D(0)/3, as required.
Finally, if b<a/2b < a/2, then D(0)=2aD(0) = 2a and
D(2)=a2b3+a+b9+4a5b9=8a10b9<8a9, D(2) = \frac{a-2b}{3} + \frac{a+b}{9} + \frac{4a-5b}{9} = \frac{8a-10b}{9} < \frac{8a}{9},
and so D(2)<4D(0)/92D(0)/3D(2) < 4D(0)/9 \le 2D(0)/3, as required. We have proved Claim 3.
With Claim 3 in hand, the unboundedness of D(n)D(n) for negative nn follows easily.

Solution 2

Let bn=anb_n = a_{-n}, so we have bn+3=bn+2bn+1+3bnb_{n+3} = -b_{n+2} - b_{n+1} + 3b_n for all nZn \in \mathbb{Z}, and in particular for n0n \ge 0. The unique solution is an=βrn+γsn+δtna_n = \beta r^n + \gamma s^n + \delta t^n for n0n \ge 0, where r=1r = 1 and s=tˉ=1+i2s = \bar{t} = -1 + i\sqrt{2} are the three roots of the equation x3+x2+x3=0x^3 + x^2 + x - 3 = 0, and β,γ,δC\beta, \gamma, \delta \in \mathbb{C} depend on b0,b1b_0, b_1, and b2b_2. Since every ana_n is real, β\beta is real and δ=γˉ\delta = \bar{\gamma}. Thus bn=β+2Re(γsn)b_n = \beta + 2 \operatorname{Re}(\gamma s^n), n0n \ge 0. We assume that (an)(a_n) is non-constant, and so γ0\gamma \ne 0.
Let arg:C{0}(π,π]\arg : \mathbb{C} \setminus \{0\} \to (-\pi, \pi] be the argument function. Call nZn \in \mathbb{Z} good if arg(γsn)<3π/8|\arg(\gamma s^n)| < 3\pi/8. Writing c:=cos(3π/8)>0c := \cos(3\pi/8) > 0, we have Re(γsn)>cγsn\operatorname{Re}(\gamma s^n) > c|\gamma| \cdot |s|^n whenever nn is good.
Now arg(s)=πtan12\arg(s) = \pi - \tan^{-1} \sqrt{2}, so
arg(s)(πtan13,πtan11)=(2π/3,3π/4). \arg(s) \in (\pi - \tan^{-1}\sqrt{3}, \pi - \tan^{-1}1) = (2\pi/3, 3\pi/4).
Thus the number γsn/γsn\gamma s^n / |\gamma s^n| "hops more than a full circuit around the unit circle" as nn ranges over a set of the form
S(n0):={n0,n0+1,n0+2,n0+3},n0Z, S(n_0) := \{n_0, n_0 + 1, n_0 + 2, n_0 + 3\}, \quad n_0 \in \mathbb{Z},
and every S(n0)S(n_0) contains a good number nn (because 0<arg(s)<2(3π/8)0 < \arg(s) < 2(3\pi/8)). We can pick arbitrarily large good numbers nNn \in \mathbb{N}, and for each good nn we have bn>β+2cγsnb_n > \beta + 2c|\gamma| \cdot |s|^n. But if (an)n=(a_n)_{n=-\infty}^{\infty} were bounded, and hence (bn)n=0(b_n)_{n=0}^{\infty} bounded, then sn|s|^n would be bounded for all nNn \in \mathbb{N}. This is impossible, since s>1|s| > 1 and so if snK|s|^n \le K, then nlogslogKn \log|s| \le \log K, and nlogK/logsn \le \log K / \log|s| (where log indicates a logarithm to any preferred base).

Solution 3

This solution is similar to Solution 2, except that we take the more obvious approach of using the given recurrence relation directly. The solution is an=β0rn+γ0sn+δ0tna_n = \beta_0 r^n + \gamma_0 s^n + \delta_0 t^n for n0n \ge 0, where r=1r = 1 and s=tˉ=(1+i2)/3s = \bar{t} = (-1 + i\sqrt{2})/3 are the three roots of the equation x3(x2+x+1)/3=0x^3 - (x^2+x+1)/3 = 0, and β0,γ0,δ0C\beta_0, \gamma_0, \delta_0 \in \mathbb{C} depend on a0,a1a_0, a_1, and a2a_2. Since every ana_n is real, β0\beta_0 is real and δ0=γˉ0\delta_0 = \bar{\gamma}_0. Thus an=β0+2Re(γ0sn)a_n = \beta_0 + 2 \operatorname{Re}(\gamma_0 s^n) for n0n \ge 0; we call this Formula 1, and we need to show that it is also valid for n<0n < 0. Assume that (an)(a_n) is non-constant, and so γ00\gamma_0 \ne 0.
Suppose that kNk \in \mathbb{N}. The sequence (bn)n=0(b_n)_{n=0}^{\infty}, where bn=ankb_n = a_{n-k}, satisfies the same recurrence relation as (an)(a_n), so it can be written in the same form: an=βk+2Re(γksn+k)a_n = \beta_k + 2 \operatorname{Re}(\gamma_k s^{n+k}) for nkn \ge -k; we call this Formula 2. To show that Formula 1 extends to n=kn = -k, we need to show that βk=β0\beta_k = \beta_0 and Re(γksn+k)=Re(γ0sn)\operatorname{Re}(\gamma_k s^{n+k}) = \operatorname{Re}(\gamma_0 s^n).
Let dk:=βkβ0d_k := \beta_k - \beta_0. Comparing Formulae 1 and 2, we see that
2Re(γ0sn)=2Re(γksn+k)+dk,n0.(5) 2 \operatorname{Re}(\gamma_0 s^n) = 2 \operatorname{Re}(\gamma_k s^{n+k}) + d_k, \quad n \ge 0. \quad (5)
Taking half the difference of two such equations (for n=0n = 0 and n=mn = m), we get
Re(γ0(1sm))=Re(γksk(1sm)) \mathrm{Re}(\gamma_0(1 - s^m)) = \mathrm{Re}(\gamma_k s^k (1 - s^m))
and so
Re((γ0γksk)(1sm))=0. \mathrm{Re}((\gamma_0 - \gamma_k s^k)(1 - s^m)) = 0.
Now arg(1sm)\arg(1 - s^m) takes on different values for m=1,2m = 1, 2, and these values do not differ by π\pi, so it follows that γ0=γksk\gamma_0 = \gamma_k s^k. Now (5) tells us that dk=0d_k = 0, i.e. βk=β0\beta_k = \beta_0, and so Formula 1 is valid for n=kn = -k.
Since kNk \in \mathbb{N} is arbitrary, our solution for n0n \ge 0 is also valid for n<0n < 0. The value of s/ss/|s| is the same as in Solution 2. Thus as before we see that there are infinitely many good values of n<0n < 0 for which Re(γsn)>cγsn\mathrm{Re}(\gamma s^n) > c|\gamma| \cdot |s|^n. But s<1|s| < 1, so as in Solution 2, boundedness is impossible.

Want a route through all this instead of an archive? The track puts 2,000 problems in a working order, from AMC 10 level to the IMO shortlist.

Source: MathNet, licensed CC-BY-4.0. Statement reproduced verbatim; metadata (topic, difficulty) added by this project.