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

IMC / 2020 / Problems / Day 1, P3

IMC 2020 · Day 1 · P3

killer

Let d2d \ge 2 be an integer. Prove that there exists a constant C(d)C(d) such that the following holds: For any convex polytope KRdK \subset \mathbb{R}^d, which is symmetric about the origin, and any ε(0,1)\varepsilon \in (0, 1), there exists a convex polytope LRdL \subset \mathbb{R}^d with at most C(d)ε1dC(d) \varepsilon^{1-d} vertices such that (1ε)KLK.(1 - \varepsilon) K \subseteq L \subseteq K. (For a real α\alpha, a set TRdT \subset \mathbb{R}^d with nonempty interior is a convex polytope with at most α\alpha vertices, if TT is a convex hull of a set XRdX \subset \mathbb{R}^d of at most α\alpha points, i.e., T={xXtxxtx0,xXtx=1}T = \{ \sum_{x \in X} t_x x \mid t_x \ge 0, \sum_{x \in X} t_x = 1 \}. For a real λ\lambda, put λK={λxxK}\lambda K = \{ \lambda x \mid x \in K \}. A set TRdT \subset \mathbb{R}^d is symmetric about the origin if (1)T=T(-1) T = T.)

Fedor Petrov, St. Petersburg State University

Solution 1 of 2 (official)

[in elementary terms] Let {p1,,pm}\{p_1, \dots, p_m\} be an inclusion-maximal collection of points on the boundary K\partial K of KK such that the homothetic copies Ki:=pi+ε2KK_i := p_i + \frac{\varepsilon}{2} K have disjoint interiors. We claim that the convex hull L:=conv{p1,,pm}L := \operatorname{conv}\{p_1, \dots, p_m\} satisfies all the conditions.

First, note that by convexity of KK we have aK+bK=(a+b)KaK + bK = (a + b)K for a,b>0a, b > 0. It follows that Ki(1+ε2)KK_i \subset (1 + \frac{\varepsilon}{2}) K. On the other hand, if kKk \in K, a>0a > 0 and and akKiak \in K_i, then piakε2K=ak+ε2K(a+ε2)K,p_i \in ak - \frac{\varepsilon}{2} K = ak + \frac{\varepsilon}{2} K \subset \Bigl( a + \frac{\varepsilon}{2} \Bigr) K, and since pip_i is a boundary point of KK, we get a+ε21a + \frac{\varepsilon}{2} \geqslant 1, a1ε2a \geqslant 1 - \frac{\varepsilon}{2}. It means that all KiK_i lie between (1ε2)K(1 - \frac{\varepsilon}{2}) K and (1+ε2)K(1 + \frac{\varepsilon}{2}) K. Since their interiors are disjoint, by the volume counting we obtain m(ε2)d(1+ε2)d(1ε2)d(3/2)dεm \left( \frac{\varepsilon}{2} \right)^d \leqslant \left( 1 + \frac{\varepsilon}{2} \right)^d - \left( 1 - \frac{\varepsilon}{2} \right)^d \leqslant (3/2)^d \varepsilon (since F(ε)=(1+ε2)d(1ε2)dF(\varepsilon) = (1 + \frac{\varepsilon}{2})^d - (1 - \frac{\varepsilon}{2})^d is a polynomial in ε\varepsilon without constant term with non-negative coefficients which sum up to (3/2)d(1/2)d(3/2)^d - (1/2)^d), therefore m3dε1dm \leqslant 3^d \varepsilon^{1-d}.

It is clear that LKL \subseteq K, so it remains to prove that (1ε)KL(1 - \varepsilon) K \subseteq L. Assume the contrary: there exists a point p(1ε)KLp \in (1 - \varepsilon) K \setminus L. Separate pp from LL by a hyperplane: Choose a linear functional \ell such that (p)>maxxL(x)=maxi(pi)\ell(p) > \max\limits_{x \in L} \ell(x) = \max_i \ell(p_i). Choose xKx \in K such that (x)=:a\ell(x) =: a is maximal possible. Note that by our construction x+ε2Kx + \frac{\varepsilon}{2} K has a common point with some KiK_i: there exists a point z(x+ε2K)(pi+ε2K)z \in (x + \frac{\varepsilon}{2} K) \cap (p_i + \frac{\varepsilon}{2} K). We have (pi)+ε2a(z)(x)ε2a,\ell(p_i) + \frac{\varepsilon}{2} a \geqslant \ell(z) \geqslant \ell(x) - \frac{\varepsilon}{2} a, and therefore (pi)a(1ε)\ell(p_i) \geqslant a (1 - \varepsilon). Since p(1ε)Kp \in (1 - \varepsilon) K, we obtain (p)a(1ε)\ell(p) \leqslant a (1 - \varepsilon). A contradiction.

Solution 2 of 2 (official)

[in the language of Banach spaces] Equip Rd\mathbb{R}^d with the norm \| \cdot \|, whose unit ball is KK, call this Banach space VV. Choose an inclusion maximal set XKX \subset \partial K whose pairwise distances are ε\ge \varepsilon. Put L=convXL = \operatorname{conv} X.

The inclusion LKL \subseteq K follows from the convexity of KK. If the inclusion (1ε)KL(1 - \varepsilon) K \subseteq L fails then the Hahn–Banach theorem provides a unit linear functional λV\lambda \in V^* such that max{λ(L)}=max{λX}1ε\max\{\lambda(L)\} = \max\{\lambda X\} \le 1 - \varepsilon. Then the point xKx \in K, where the maximum max{λ(K)}=1\max\{\lambda(K)\} = 1 is attained (thanks to the finite dimension and compactness) is in K\partial K and, as λ\lambda witnesses, at distance ε\ge \varepsilon from all other points of LL and XX, contradicting the inclusion-maximality of XX.

The upper bound for the cardinality X|X| is obtained by noting that the ε/2\varepsilon/2 balls centered at the points of XX are pairwise disjoint and lie in the difference of balls (1+ε/2)K(1ε/2)K(1 + \varepsilon/2) K \setminus (1 - \varepsilon/2) K, whose volume is ((1+ε/2)d(1ε/2)d)volK\bigl( (1 + \varepsilon/2)^d - (1 - \varepsilon/2)^d \bigr) \operatorname{vol} K, the volume of each of the small balls being εd/2dvolK\varepsilon^d / 2^d \operatorname{vol} K. Hence X(2+ε)d(2ε)dεd=O(ε1d).|X| \le \frac{(2 + \varepsilon)^d - (2 - \varepsilon)^d} {\varepsilon^d} = O(\varepsilon^{1-d}).

How the field did

contestants scored
453
average (of 10)
0.11
solved (≥ 80%)
0.7%
near-0 (≤ 10%)
98.5%
discrimination
0.22

Score distribution (field cohort)

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

Similar problems

IMC 2014 · Day 2 · P9killeravg 0.6/10 · solved 4% · near-0 89% · disc 0.40
IMC 2014 · Day 1 · P5very hardavg 1.0/10 · solved 5% · near-0 78% · disc 0.49
IMC 2001 · Day 2 · P9hardavg 2.3/10 · solved 16% · near-0 66% · disc 0.48
IMC 2004 · Day 1 · P4mediumavg 4.0/10 · solved 35% · near-0 52% · disc 0.52