Unofficial archive — problems, solutions & results © IMC, reproduced with permission.

IMC / 2021 / Problems / Day 1, P3

IMC 2021 · Day 1 · P3

very hard

We say that a positive real number dd is good if there exists an infinite sequence a1,a2,a3,(0,d)a_1, a_2, a_3, \dots \in (0, d) such that for each nn, the points a1,,ana_1, \dots, a_n partition the interval [0,d][0, d] into segments of length at most 1/n1/n each. Find sup{dd is good}.\sup \{ d \mid d \text{ is good} \}. (proposed by Josef Tkadlec)

Solution (official)

Hint: To get an upper bound, use that some of the gaps after nn steps are still intact some steps later.

Let d=sup{dd is good}d^\star = \sup\{ d \mid d \text{ is good} \}. We will show that d=ln(2)0.693d^\star = \ln(2) \doteq 0.693.

1. dln2d^\star \le \ln 2:

Assume that some dd is good and let a1,a2,a_1, a_2, \dots be the witness sequence.

Fix an integer nn. By assumption, the prefix a1,,ana_1, \dots, a_n of the sequence splits the interval [0,d][0, d] into n+1n + 1 parts, each of length at most 1/n1/n.

Let 012n+10 \le \ell_1 \le \ell_2 \le \dots \le \ell_{n+1} be the lengths of these parts. Now for each k=1,,nk = 1, \dots, n after placing the next kk terms an+1,,an+ka_{n+1}, \dots, a_{n+k}, at least n+1kn + 1 - k of these initial parts remain intact. Hence n+1k1n+k\ell_{n+1-k} \le \frac{1}{n+k}. Hence d=1++n+11n+1n+1++12n.(2)\tag{2} d = \ell_1 + \dots + \ell_{n+1} \le \frac{1}{n} + \frac{1}{n+1} + \dots + \frac{1}{2n}. As nn \to \infty, the RHS tends to ln(2)\ln(2) showing that dln(2)d \le \ln(2).

Hence dln2d^\star \le \ln 2 as desired.

2. dln2d^\star \ge \ln 2:

Observe that ln2=ln2nlnn=i=1n(ln(n+i)ln(n+i1))=i=1nln(1+1n+i1).\ln 2 = \ln 2n - \ln n = \sum_{i=1}^{n} \bigl( \ln(n+i) - \ln(n+i-1) \bigr) = \sum_{i=1}^{n} \ln \left( 1 + \frac{1}{n+i-1} \right). Interpreting the summands as lengths, we think of the sum as the lengths of a partition of the segment [0,ln2][0, \ln 2] in nn parts. Moreover, the maximal length of the parts is ln(1+1/n)<1/n\ln(1 + 1/n) < 1/n.

Changing nn to n+1n + 1 in the sum keeps the values of the sum, removes the summand ln(1+1/n)\ln(1 + 1/n), and adds two summands ln(1+12n)+ln(1+12n+1)=ln(1+1n).\ln \left( 1 + \frac{1}{2n} \right) + \ln \left( 1 + \frac{1}{2n+1} \right) = \ln \left( 1 + \frac{1}{n} \right). This transformation may be realized by adding one partition point in the segment of length ln(1+1/n)\ln(1 + 1/n).

In total, we obtain a scheme to add partition points one by one, all the time keeping the assumption that once we have n1n - 1 partition points and nn partition segments, all the partition segments are smaller than 1/n1/n.

The first terms of the constructed sequence will be a1=ln32a_1 = \ln\frac32, a2=ln54a_2 = \ln\frac54, a3=ln74a_3 = \ln\frac74, a4=ln98a_4 = \ln\frac98, ….

Remark. This remark describes in fact the same solution from a different view and some ideas behind it. It could be erased after marking is finished.

Estimate (2) is quite natural. To prove that RHS tends to ln2\ln 2 we use some integral estimates by n2n+11xdx=ln(2n+1)lnn.\int_n^{2n+1} \frac{1}{x}\,dx = \ln(2n+1) - \ln n. Here we can observe that n2n1xdx=ln2\int_n^{2n} \frac{1}{x}\,dx = \ln 2 is independent of nn. This can help us with the construction since the above equality means I1=nn+11xdx=2n2n+11xdx+2n+12n+21xdx=I2+I3,I_1 = \int_n^{n+1} \frac{1}{x}\,dx = \int_{2n}^{2n+1} \frac{1}{x}\,dx + \int_{2n+1}^{2n+2} \frac{1}{x}\,dx = I_2 + I_3, so, interval of length I1I_1 can be splitted into two intervals of lengths I2I_2 and I3I_3. In fact, after placing the point ana_n in the construction for d=ln2d = \ln 2, the lengths of the n+1n + 1 intervals are n+1n+21x,n+2n+31x,,2n+12n+21x\int_{n+1}^{n+2} \frac{1}{x}, \quad \int_{n+2}^{n+3} \frac{1}{x}, \quad \dots, \quad \int_{2n+1}^{2n+2} \frac{1}{x} with total length d=n+12n+21x=ln2.d = \int_{n+1}^{2n+2} \frac{1}{x} = \ln 2.

How the field did

contestants scored
514
average (of 10)
1.27
solved (≥ 80%)
7.8%
near-0 (≤ 10%)
80.2%
discrimination
0.50

Score distribution (field cohort)

Computed on contestants with a meaningful total (field cohort); discrimination is the corrected item–total correlation.

Similar problems

IMC 2000 · Day 1 · P6very hardavg 1.3/10 · solved 5% · near-0 68% · disc 0.47
IMC 2017 · Day 2 · P9hardavg 2.2/10 · solved 15% · near-0 69% · disc 0.59
IMC 2020 · Day 2 · P8killeravg 0.1/10 · solved 0% · near-0 98% · disc 0.16
IMC 2010 · Day 2 · P6mediumavg 7.4/10 · solved 47% · near-0 3% · disc 0.49