Maths Olympiad Prep

Library / /267 of 397

, 2023

Algebra Difficulty 6.3 National Olympiad Prove it Taiwan

Given a positive integer k>1k > 1 and nn distinct points in kk-dimensional coordinates a(1)=(a1(1),,ak(1))a^{(1)} = (a_1^{(1)}, \dots, a_k^{(1)}), ,a(n)=(a1(n),,ak(n))\dots, a^{(n)} = (a_1^{(n)}, \dots, a_k^{(n)}), we define the "score" of a(i)a^{(i)} as follows:
j=1k#{i1in such that πj(a(i))=πj(a(i))}, \prod_{j=1}^{k} \#\{i' \mid 1 \le i' \le n \text{ such that } \pi_j(a^{(i')}) = \pi_j(a^{(i)})\},
where #S\#S denotes the number of elements in the set SS, and πj:RkRk1\pi_j : \mathbb{R}^k \to \mathbb{R}^{k-1} is the projection map that sends a point in kk-dimensional coordinates to the point obtained by deleting its jj-th coordinate. Find the largest real number tt such that the tt-th power mean of the "scores" of all a(i)a^{(i)} is n\le n.

Note: The tt-th power mean of positive real numbers x1,,xnx_1, \dots, x_n is defined as

( x 1 t +\text{( x 1 t +} + x_n^t}{n} )1t(if\left. \right)^{\frac{1}{t}} \quad (\text{if} } t \neq 0); x1n\quad \sqrt[n]{x_1 \cdots} x_n} (if\quad (\text{if} } t = 0).

For every positive integer kk greater than 1, find the largest real number tt such that the following hold: Given nn distinct points a(1)=(a1(1),,ak(1))a^{(1)} = (a_1^{(1)}, \cdots, a_k^{(1)}), \cdots, a(n)=(a1(n),,ak(n))a^{(n)} = (a_1^{(n)}, \cdots, a_k^{(n)}) in Rk\mathbb{R}^k, we define the score of the tuple a(i)a^{(i)} as
j=1k#{1in such that πj(a(i))=πj(a(i))} \prod_{j=1}^{k} \#\{1 \le i' \le n \text{ such that } \pi_j(a^{(i')}) = \pi_j(a^{(i)})\}
where #S\#S is the number of elements in set SS, and πj\pi_j is the projection RkRk1\mathbb{R}^k \to \mathbb{R}^{k-1} omitting the jj-th coordinate. Then the tt-th power mean of the scores of all a(i)a^{(i)}'s is at most nn.

Note: The tt-th power mean of positive real numbers x1,,xnx_1, \cdots, x_n is defined as
(x1t++xntn)1t \left( \frac{x_1^t + \cdots + x_n^t}{n} \right)^{\frac{1}{t}}

Solution

The answer is 1k1\frac{1}{k-1}. To see that t1k1t \le \frac{1}{k-1}, consider the sequences where all but one of the components are equal to 1, and the remaining one component is 1,2,,m1, 2, \dots, m. Then n=km+1n = km + 1 and the score of the zero sequence is mkm^k. In order for the statement to hold, we must have
(mktkm+1)1tkm+1, \left( \frac{m^{kt}}{km + 1} \right)^{\frac{1}{t}} \le km + 1,
i.e., mk(km+1)1+1/tm^k \le (km + 1)^{1+1/t}. By choosing mm to be sufficiently large, we get
k1+1t, or t1k1. k \le 1 + \frac{1}{t}, \text{ or } t \le \frac{1}{k-1}.

Now we will prove by induction on kk that t=1k1t = \frac{1}{k-1} works. We first assume that k3k \ge 3 and the statement holds for k1k-1. For every real number rr, let crc_r be the number of sequences a(i)a^{(i)} with a1(i)=ra_1^{(i)} = r. For each rr with cr0c_r \ne 0, we gather the sequences with a1(i)=ra_1^{(i)} = r and compute their scores when we remove the first component. Let the scores be s1,,scrs'_1, \cdots, s'_{c_r}. Moreover, for each truncated sequence, we see how many sequences from the original list would have produced the same truncated sequence. Let the number of such sequences be d1,,dcrd_1, \cdots, d_{c_r}. Then clearly d1++dcrnd_1 + \cdots + d_{c_r} \le n as all sequences are distinct. Note that the scores of the original sequences we gather are precisely d1s1,,dcrscrd_1 s'_1, \cdots, d_{c_r} s'_{c_r}. By the inductive hypothesis, we have
s11k2++scr1k2=crk1k2. s_1^{\frac{1}{k-2}} + \cdots + s_{c_r}^{\frac{1}{k-2}} = c_r^{\frac{k-1}{k-2}}.
By Hölder's inequality, we have
((d1s1)1k1++(dcrscr)1k1)k1(d1++dcr)(s11k2++scr1k2)k2ncrk1. ((d_1 s'_1)^{\frac{1}{k-1}} + \cdots + (d_{c_r} s'_{c_r})^{\frac{1}{k-1}})^{k-1} \le (d_1 + \cdots + d_{c_r}) \cdot \left(s_1^{\frac{1}{k-2}} + \cdots + s_{c_r}^{\frac{1}{k-2}}\right)^{k-2} \le n c_r^{k-1}.
By taking the (k1)(k-1)-th root of both sides and adding it up for all rRr \in \mathbb{R} with cr0c_r \ne 0, we get that the 1k1\frac{1}{k-1}-th power mean is bounded by
(cr0n1k1crn)k1=n \left( \frac{\sum_{c_r \ne 0} n^{\frac{1}{k-1}} c_r}{n} \right)^{k-1} = n
as
cr0cr=n. \sum_{c_r \ne 0} c_r = n.
For the base case k=2k=2, we can follow the same proof except that we clearly have
s1==scr=n s'_1 = \cdots = s'_{c_r} = n
and so
(d1s1++dcrscr)=n(d1++dcr)n2 (d_1 s'_1 + \cdots + d_{c_r} s'_{c_r}) = n(d_1 + \cdots + d_{c_r}) \le n^2
and the rest of the proof follows verbatim.

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 translated into English from zh; metadata (topic, difficulty) added by this project.