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

IMC / 2009 / Problems / Day 1, P4

IMC 2009 · Day 1 · P4

killer

Let p(z)=a0+a1z+a2z2++anznp(z) = a_0 + a_1 z + a_2 z^2 + \dots + a_n z^n be a complex polynomial. Suppose that 1=c0c1cn01 = c_0 \ge c_1 \ge \dots \ge c_n \ge 0 is a sequence of real numbers which is convex (i.e. 2ckck1+ck+12 c_k \le c_{k-1} + c_{k+1} for every k=1,2,,n1k = 1, 2, \dots, n-1), and consider the polynomial q(z)=c0a0+c1a1z+c2a2z2++cnanzn.q(z) = c_0 a_0 + c_1 a_1 z + c_2 a_2 z^2 + \dots + c_n a_n z^n. Prove that maxz1q(z)maxz1p(z).\max_{|z| \le 1} \bigl| q(z) \bigr| \le \max_{|z| \le 1} \bigl| p(z) \bigr|.

Solution (official)

The polynomials pp and qq are regular on the complex plane, so by the Maximum Principle, maxz1q(z)=maxz=1q(z)\max_{|z| \le 1} |q(z)| = \max_{|z| = 1} |q(z)|, and similarly for pp. Let us denote Mf=maxz=1f(z)M_f = \max_{|z| = 1} |f(z)| for any regular function ff. Thus it suffices to prove that MqMpM_q \le M_p.

First, note that we can assume cn=0c_n = 0. Indeed, for cn=1c_n = 1, we get p=qp = q and the statement is trivial; otherwise, q(z)=cnp(z)+(1cn)r(z)q(z) = c_n p(z) + (1 - c_n) r(z), where r(z)=j=0ncjcn1cnajzjr(z) = \sum_{j=0}^{n} \frac{c_j - c_n}{1 - c_n} a_j z^j. The sequence cj=cjcn1cnc_j' = \frac{c_j - c_n}{1 - c_n} also satisfies the prescribed conditions (it is a positive linear transform of the sequence cjc_j with c0=1c_0' = 1), but cn=0c_n' = 0 too, so we get MrMpM_r \le M_p. This is enough: Mq=q(z0)cnp(z0)+(1cn)r(z0)cnMp+(1cn)MrMpM_q = |q(z_0)| \le c_n |p(z_0)| + (1 - c_n) |r(z_0)| \le c_n M_p + (1 - c_n) M_r \le M_p.

Using the Cauchy formulas, we can express the coefficients aja_j of pp from its values taken over the positively oriented circle S={z=1}S = \{|z| = 1\}: aj=12πiSp(z)zj+1dz=12πSp(z)zjdza_j = \frac{1}{2\pi i} \int_S \frac{p(z)}{z^{j+1}}\,dz = \frac{1}{2\pi} \int_S \frac{p(z)}{z^j}\,|dz| for 0jn0 \le j \le n, otherwise Sp(z)zjdz=0.\int_S \frac{p(z)}{z^j}\,|dz| = 0. Let us use these identities to get a new formula for qq, using only the values of pp over SS: 2πq(w)=j=0ncj(Sp(z)zjdz)wj.2\pi \cdot q(w) = \sum_{j=0}^{n} c_j \left( \int_S p(z) z^{-j}\,|dz| \right) w^j. We can exchange the order of the summation and the integration (sufficient conditions to do this obviously apply): 2πq(w)=S(j=0ncj(w/z)j)p(z)dz.2\pi \cdot q(w) = \int_S \left( \sum_{j=0}^{n} c_j (w/z)^j \right) p(z)\,|dz|. It would be nice if the integration kernel (the sum between the brackets) was real. But this is easily arranged — for nj1-n \le j \le -1, we can add the conjugate expressions, because by the above remarks, they are zero anyway: 2πq(w)=j=0ncj(Sp(z)zjdz)wj=j=nncj(Sp(z)zjdz)wj,2\pi \cdot q(w) = \sum_{j=0}^{n} c_j \left( \int_S p(z) z^{-j}\,|dz| \right) w^j = \sum_{j=-n}^{n} c_{|j|} \left( \int_S p(z) z^{-j}\,|dz| \right) w^j, 2πq(w)=S(j=nncj(w/z)j)p(z)dz=SK(w/z)p(z)dz,2\pi \cdot q(w) = \int_S \left( \sum_{j=-n}^{n} c_{|j|} (w/z)^j \right) p(z)\,|dz| = \int_S K(w/z)\, p(z)\,|dz|, where K(u)=j=nncjuj=c0+2j=1ncj(uj)K(u) = \sum_{j=-n}^{n} c_{|j|} u^j = c_0 + 2 \sum_{j=1}^{n} c_j\, \Re(u^j) for uSu \in S.

Let us examine K(u)K(u). It is a real-valued function. Again from the Cauchy formulas, SK(u)du=2πc0=2π\int_S K(u)\,|du| = 2\pi c_0 = 2\pi.

If SK(u)du=2π\int_S |K(u)|\,|du| = 2\pi still holds (taking the absolute value does not increase the integral), then for every ww: 2πq(w)=SK(w/z)p(z)dzSK(w/z)p(z)dzMpSK(u)du=2πMp;2\pi |q(w)| = \left| \int_S K(w/z)\, p(z)\,|dz| \right| \le \int_S |K(w/z)| \cdot |p(z)|\,|dz| \le M_p \int_S |K(u)|\,|du| = 2\pi M_p; this would conclude the proof. So it suffices to prove that SK(u)du=SK(u)du\int_S |K(u)|\,|du| = \int_S K(u)\,|du|, which is to say, KK is non-negative.

Now let us decompose KK into a sum using the given conditions for the numbers cjc_j (including cn=0c_n = 0). Let dk=ck12ck+ck+1d_k = c_{k-1} - 2 c_k + c_{k+1} for k=1,,nk = 1, \dots, n (setting cn+1=0c_{n+1} = 0); we know that dk0d_k \ge 0. Let Fk(u)=j=k+1k1(kj)ujF_k(u) = \sum\limits_{j=-k+1}^{k-1} (k - |j|) u^j. Then K(u)=k=1ndkFk(u)K(u) = \sum\limits_{k=1}^{n} d_k F_k(u) by easy induction (or see Figure for a graphical illustration).

So it suffices to prove that Fk(u)0F_k(u) \ge 0 for uSu \in S. This is reasonably well-known (as FkF_k is the Fejér kernel), and also very easy: Fk(u)=(1+u+u2++uk1)(1+u1+u2++u(k1))==(1+u+u2++uk1)(1+u+u2++uk1)=1+u+u2++uk120\begin{align*} F_k(u) &= (1 + u + u^2 + \dots + u^{k-1}) (1 + u^{-1} + u^{-2} + \dots + u^{-(k-1)}) = \\ &= (1 + u + u^2 + \dots + u^{k-1}) \overline{(1 + u + u^2 + \dots + u^{k-1})} = \left| 1 + u + u^2 + \dots + u^{k-1} \right|^2 \ge 0 \end{align*} This completes the proof.

How the field did

contestants scored
334
average (of 10)
0.07
solved (≥ 80%)
0.0%
near-0 (≤ 10%)
98.2%
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 2017 · Day 1 · P5killeravg 0.2/10 · solved 2% · near-0 97% · disc 0.42
IMC 2007 · Day 1 · P6killeravg 0.6/10 · solved 3% · near-0 91% · disc 0.37
IMC 2000 · Day 2 · P9easyavg 5.5/10 · solved 53% · near-0 32% · disc 0.65