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

IMC / 2002 / Problems / Day 1, P6

IMC 2002 · Day 1 · P6

killer

For an n×nn \times n matrix MM with real entries let M=supxRn{0}Mx2x2\|M\| = \sup\limits_{x \in \mathbb{R}^n \setminus \{0\}} \dfrac{\|Mx\|_2}{\|x\|_2}, where 2\|\cdot\|_2 denotes the Euclidean norm on Rn\mathbb{R}^n. Assume that an n×nn \times n matrix AA with real entries satisfies AkAk112002k\|A^k - A^{k-1}\| \le \dfrac{1}{2002 k} for all positive integers kk. Prove that Ak2002\|A^k\| \le 2002 for all positive integers kk.

Solution (official)

Lemma 1. Let (an)n0(a_n)_{n \ge 0} be a sequence of non-negative numbers such that a2ka2k+1ak2a_{2k} - a_{2k+1} \le a_k^2, a2k+1a2k+2akak+1a_{2k+1} - a_{2k+2} \le a_k a_{k+1} for any k0k \ge 0 and lim supnan<1/4\limsup n a_n < 1/4. Then lim supann<1\limsup \sqrt[n]{a_n} < 1.

Proof. Let cl=supn2l(n+1)anc_l = \sup_{n \ge 2^l} (n+1) a_n for l0l \ge 0. We will show that cl+14cl2c_{l+1} \le 4 c_l^2. Indeed, for any integer n2l+1n \ge 2^{l+1} there exists an integer k2lk \ge 2^l such that n=2kn = 2k or n=2k+1n = 2k+1. In the first case there is a2ka2k+1ak2cl2(k+1)24cl2(2k+1)(2k+2)a_{2k} - a_{2k+1} \le a_k^2 \le \frac{c_l^2}{(k+1)^2} \le \frac{4 c_l^2}{(2k+1)(2k+2)}, whereas in the second case there is a2k+1a2k+2akak+1cl2(k+1)(k+2)4cl2(2k+2)(2k+3)a_{2k+1} - a_{2k+2} \le a_k a_{k+1} \le \frac{c_l^2}{(k+1)(k+2)} \le \frac{4 c_l^2}{(2k+2)(2k+3)}.

Hence a sequence (an4cl2n+1)n2l+1\left( a_n - \frac{4 c_l^2}{n+1} \right)_{n \ge 2^{l+1}} is non-decreasing and its terms are non-positive since it converges to zero. Therefore an4cl2n+1a_n \le \frac{4 c_l^2}{n+1} for n2l+1n \ge 2^{l+1}, meaning that cl+14cl2c_{l+1} \le 4 c_l^2. This implies that a sequence ((4cl)2l)l0((4 c_l)^{2^{-l}})_{l \ge 0} is non-increasing and therefore bounded from above by some number q(0,1)q \in (0,1) since all its terms except finitely many are less than 1. Hence clq2lc_l \le q^{2^l} for ll large enough. For any nn between 2l2^l and 2l+12^{l+1} there is ancln+1q2l(q)na_n \le \frac{c_l}{n+1} \le q^{2^l} \le (\sqrt{q})^n yielding lim supannq<1\limsup \sqrt[n]{a_n} \le \sqrt{q} < 1, yielding lim supannq<1\limsup \sqrt[n]{a_n} \le \sqrt{q} < 1,

which ends the proof.

Lemma 2. Let TT be a linear map from Rn\mathbb{R}^n into itself. Assume that lim supnTn+1Tn<1/4\limsup n \|T^{n+1} - T^n\| < 1/4. Then lim supTn+1Tn1/n<1\limsup \|T^{n+1} - T^n\|^{1/n} < 1. In particular TnT^n converges in the operator norm and TT is power bounded.

Proof. Put an=Tn+1Tna_n = \|T^{n+1} - T^n\|. Observe that Tk+m+1Tk+m=(Tk+m+2Tk+m+1)(Tk+1Tk)(Tm+1Tm)T^{k+m+1} - T^{k+m} = (T^{k+m+2} - T^{k+m+1}) - (T^{k+1} - T^k)(T^{m+1} - T^m) implying that ak+mak+m+1+akama_{k+m} \le a_{k+m+1} + a_k a_m. Therefore the sequence (am)m0(a_m)_{m \ge 0} satisfies assumptions of Lemma 1 and the assertion of Proposition 1 follows.

Remarks. 1. The theorem proved above holds in the case of an operator TT which maps a normed space XX into itself, XX does not have to be finite dimensional.

2. The constant 1/41/4 in Lemma 1 cannot be replaced by any greater number since a sequence an=14na_n = \frac{1}{4n} satisfies the inequality ak+mak+m+1akama_{k+m} - a_{k+m+1} \le a_k a_m for any positive integers kk and mm whereas it does not have exponential decay.

3. The constant 1/41/4 in Lemma 2 cannot be replaced by any number greater that 1/e1/e. Consider an operator (Tf)(x)=xf(x)(Tf)(x) = x f(x) on L2([0,1])L_2([0,1]). One can easily check that lim supTn+1Tn=1/e\limsup \|T^{n+1} - T^n\| = 1/e, whereas TnT^n does not converge in the operator norm. The question whether in general lim supnTn+1Tn<\limsup n \|T^{n+1} - T^n\| < \infty implies that TT is power bounded remains open.

Remark The problem was incorrectly stated during the competition: instead of the inequality AkAk112002k\|A^k - A^{k-1}\| \le \frac{1}{2002 k}, the inequality AkAk112002n\|A^k - A^{k-1}\| \le \frac{1}{2002 n} was assumed. If A=(1ε01)A = \begin{pmatrix} 1 & \varepsilon \\ 0 & 1 \end{pmatrix} then Ak=(1kε01)A^k = \begin{pmatrix} 1 & k\varepsilon \\ 0 & 1 \end{pmatrix}. Therefore AkAk1=(0ε00)A^k - A^{k-1} = \begin{pmatrix} 0 & \varepsilon \\ 0 & 0 \end{pmatrix}, so for sufficiently small ε\varepsilon the condition is satisfied although the sequence (Ak)\bigl( \|A^k\| \bigr) is clearly unbounded.

How the field did

contestants scored
182
average (of 20)
0.42
solved (≥ 80%)
1.6%
near-0 (≤ 10%)
97.8%
discrimination
0.21

Score distribution (field cohort)

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

Similar problems

IMC 2016 · Day 2 · P10killeravg 0.1/10 · solved 1% · near-0 99% · disc 0.16
IMC 2014 · Day 2 · P7easyavg 7.3/10 · solved 71% · near-0 21% · disc 0.46