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

IMC / 2008 / Problems / Day 2, P12

IMC 2008 · Day 2 · P12

killer

Let H\mathcal{H} be an infinite-dimensional real Hilbert space, let d>0d > 0, and suppose that SS is a set of points (not necessarily countable) in H\mathcal{H} such that the distance between any two distinct points in SS is equal to dd. Show that there is a point y∈Hy \in \mathcal{H} such that {2d(x−y):x∈S}\left\{ \frac{\sqrt{2}}{d} (x - y) : x \in S \right\} is an orthonormal system of vectors in H\mathcal{H}.

Solution (official)

It is clear that, if BB is an orthonormal system in a Hilbert space H\mathcal{H}, then {(d/2)e:e∈B}\{ (d/\sqrt{2}) e : e \in B \} is a set of points in H\mathcal{H}, any two of which are at distance dd apart. We need to show that every set SS of equidistant points is a translate of such a set.

We begin by noting that if x1,x2,x3,x4∈Sx_1, x_2, x_3, x_4 \in S are four distinct points, then ⟨x2−x1,x2−x1⟩=d2,⟨x2−x1,x3−x1⟩=12(∥x2−x1∥2+∥x3−x1∥2−∥x2−x3∥2)=d22,⟨x2−x1,x4−x3⟩=⟨x2−x1,x4−x1⟩−⟨x2−x1,x3−x1⟩=d22−d22=0.\begin{align*} \langle x_2 - x_1, x_2 - x_1 \rangle &= d^2, \\ \langle x_2 - x_1, x_3 - x_1 \rangle &= \tfrac{1}{2} \bigl( \|x_2 - x_1\|^2 + \|x_3 - x_1\|^2 - \|x_2 - x_3\|^2 \bigr) = \tfrac{d^2}{2}, \\ \langle x_2 - x_1, x_4 - x_3 \rangle &= \langle x_2 - x_1, x_4 - x_1 \rangle - \langle x_2 - x_1, x_3 - x_1 \rangle = \tfrac{d^2}{2} - \tfrac{d^2}{2} = 0. \end{align*} This shows that scalar products among vectors which are finite linear combinations of the form λ1x1+λ2x2+⋯+λnxn,\lambda_1 x_1 + \lambda_2 x_2 + \dots + \lambda_n x_n, where x1,x2,…,xnx_1, x_2, \dots, x_n are distinct points in SS and λ1,λ2,…,λn\lambda_1, \lambda_2, \dots, \lambda_n are integers with λ1+λ2+⋯+λn=0\lambda_1 + \lambda_2 + \dots + \lambda_n = 0, are universal across all such sets SS in all Hilbert spaces H\mathcal{H}; in particular, we may conveniently evaluate them using examples of our choosing, such as the canonical example above in Rn\mathbb{R}^n. In fact this property trivially follows also when coefficients λi\lambda_i are rational, and hence by continuity any real numbers with sum 0.

If S={x1,x2,…,xn}S = \{x_1, x_2, \dots, x_n\} is a finite set, we form x=1n(x1+x2+⋯+xn)x = \frac{1}{n}(x_1 + x_2 + \dots + x_n), pick a non-zero vector z∈[span⁡(x1−x,x2−x,…,xn−x)]⊥z \in [\operatorname{span}(x_1 - x, x_2 - x, \dots, x_n - x)]^{\perp} and seek yy in the form y=x+λzy = x + \lambda z for a suitable λ∈R\lambda \in \mathbb{R}. We find that ⟨x1−y,x2−y⟩=⟨x1−x−λz,x2−x−λz⟩=⟨x1−x,x2−x⟩+λ2∥z∥2.\langle x_1 - y, x_2 - y \rangle = \langle x_1 - x - \lambda z, x_2 - x - \lambda z \rangle = \langle x_1 - x, x_2 - x \rangle + \lambda^2 \|z\|^2. ⟨x1−x,x2−x⟩\langle x_1 - x, x_2 - x \rangle may be computed by our remark above as ⟨x1−x,x2−x⟩=d22⟨(n−1n,−1n,−1n,…,−1n) ⁣⊤,(−1n,n−1n,−1n,…,−1n) ⁣⊤⟩Rn=d22(−2(n−1)n2+n−2n2)=−d22n.\langle x_1 - x, x_2 - x \rangle = \frac{d^2}{2} \left\langle \left( \tfrac{n-1}{n}, -\tfrac{1}{n}, -\tfrac{1}{n}, \dots, -\tfrac{1}{n} \right)^{\!\top}, \left( -\tfrac{1}{n}, \tfrac{n-1}{n}, -\tfrac{1}{n}, \dots, -\tfrac{1}{n} \right)^{\!\top} \right\rangle_{\mathbb{R}^n} = \frac{d^2}{2} \left( -\frac{2(n-1)}{n^2} + \frac{n-2}{n^2} \right) = -\frac{d^2}{2n}. So the choice λ=d2n ∥z∥\lambda = \dfrac{d}{\sqrt{2n}\, \|z\|} will make all vectors 2d(xi−y)\frac{\sqrt{2}}{d}(x_i - y) orthogonal to each other; it is easily checked as above that they will also be of length one.

Let now SS be an infinite set. Pick an infinite sequence T={x1,x2,…,xn,… }T = \{x_1, x_2, \dots, x_n, \dots\} of distinct points in SS. We claim that the sequence yn=1n(x1+x2+⋯+xn)y_n = \frac{1}{n} (x_1 + x_2 + \dots + x_n) is a Cauchy sequence in H\mathcal{H}. (This is the crucial observation.) Indeed, for m>nm > n, the norm ∥ym−yn∥\|y_m - y_n\| may be computed by the above remark as ∥ym−yn∥2=d22∥(1m−1n,…,1m−1n,1m,…,1m) ⁣⊤∥Rm2=d22(n(m−n)2m2n2+m−nm2)=d22⋅m−nmn=d22(1n−1m)→0,m,n→∞.\|y_m - y_n\|^2 = \frac{d^2}{2} \left\| \left( \tfrac{1}{m} - \tfrac{1}{n}, \dots, \tfrac{1}{m} - \tfrac{1}{n}, \tfrac{1}{m}, \dots, \tfrac{1}{m} \right)^{\!\top} \right\|^2_{\mathbb{R}^m} = \frac{d^2}{2} \left( \frac{n(m-n)^2}{m^2 n^2} + \frac{m-n}{m^2} \right) = \frac{d^2}{2} \cdot \frac{m-n}{mn} = \frac{d^2}{2} \left( \frac{1}{n} - \frac{1}{m} \right) \to 0, \quad m, n \to \infty. By completeness of H\mathcal{H}, it follows that there exists a limit y=lim⁡n→∞yn∈H.y = \lim_{n \to \infty} y_n \in \mathcal{H}. We claim that yy satisfies all conditions of the problem. For m>n>pm > n > p, with n,pn, p fixed, we compute ∥xn−ym∥2=d22∥(−1m,…,−1m,1−1m,−1m,…,−1m) ⁣⊤∥Rm2=d22(m−1m2+(m−1)2m2)=d22⋅m−1m→d22,m→∞,\|x_n - y_m\|^2 = \frac{d^2}{2} \left\| \left( -\tfrac{1}{m}, \dots, -\tfrac{1}{m}, 1 - \tfrac{1}{m}, -\tfrac{1}{m}, \dots, -\tfrac{1}{m} \right)^{\!\top} \right\|^2_{\mathbb{R}^m} = \frac{d^2}{2} \left( \frac{m-1}{m^2} + \frac{(m-1)^2}{m^2} \right) = \frac{d^2}{2} \cdot \frac{m-1}{m} \to \frac{d^2}{2}, \quad m \to \infty, showing that ∥xn−y∥=d/2\|x_n - y\| = d/\sqrt{2}, as well as ⟨xn−ym,xp−ym⟩=d22(m−2m2−2m(1−1m))=−d22m→0,m→∞,\langle x_n - y_m, x_p - y_m \rangle = \frac{d^2}{2} \left( \frac{m-2}{m^2} - \frac{2}{m} \left( 1 - \frac{1}{m} \right) \right) = -\frac{d^2}{2m} \to 0, \quad m \to \infty, showing that ⟨xn−y,xp−y⟩=0\langle x_n - y, x_p - y \rangle = 0, so that {2d(xn−y):n∈N}\left\{ \frac{\sqrt{2}}{d} (x_n - y) : n \in \mathbb{N} \right\} is indeed an orthonormal system of vectors.

This completes the proof in the case when T=ST = S, which we can always take if SS is countable. If it is not, let x′,x′′x', x'' be any two distinct points in S∖TS \setminus T. Then applying the above procedure to the set T′={x′,x′′,x1,x2,…,xn,… }T' = \{x', x'', x_1, x_2, \dots, x_n, \dots\} it follows that the limit y=lim⁡n→∞x′+x′′+x1+x2+⋯+xnn+2=lim⁡n→∞x1+x2+⋯+xnn=yy = \lim_{n \to \infty} \frac{x' + x'' + x_1 + x_2 + \dots + x_n}{n + 2} = \lim_{n \to \infty} \frac{x_1 + x_2 + \dots + x_n}{n} = y satisfies that {2d(x′−y),2d(x′′−y)}∪{2d(xn−y):n∈N}\left\{ \frac{\sqrt{2}}{d} (x' - y), \frac{\sqrt{2}}{d} (x'' - y) \right\} \cup \left\{ \frac{\sqrt{2}}{d} (x_n - y) : n \in \mathbb{N} \right\} is still an orthonormal system.

This it true for any distinct x′,x′′∈S∖Tx', x'' \in S \setminus T; it follows that the entire system {2d(x−y):x∈S}\left\{ \frac{\sqrt{2}}{d} (x - y) : x \in S \right\} is an orthonormal system of vectors in H\mathcal{H}, as required.

How the field did

contestants scored
255
average (of 20)
1.39
solved (≥ 80%)
3.9%
near-0 (≤ 10%)
90.2%
discrimination
0.45

Score distribution (field cohort)

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

Similar problems

IMC 2005 · Day 2 · P12killeravg 1.5/10 · solved 4% · near-0 72% · disc 0.33
IMC 2002 · Day 2 · P11killeravg 1.9/10 · solved 4% · near-0 20% · disc 0.36
linear algebra
IMC 2009 · Day 1 · P5killeravg 0.5/10 · solved 4% · near-0 92% · disc 0.28
IMC 2019 · Day 1 · P5killeravg 0.8/10 · solved 5% · near-0 91% · disc 0.35