Some problems related to Achim Klenke: Probability theory 3.1.2

63 Views Asked by At

The originally problem is as follows:

Give an example for two different probability generating functions that coincide at countably many points $x_i ∈ (0, 1)$, $i ∈ \mathbb N$.

In other words, find $\{a_n\}_{n \in \mathbb N},\{b_n\}_{n \in \mathbb N}$, two different sequences of non negative real numbers which sum to $1$, but such that: $$\sum_{n \in \mathbb N} a_nx^n = \sum_{n \in \mathbb N} b_nx^n $$ For countably infinite many points $x\in (0,1)$.

My thoughts are as follows, if we let the LHS be $f(x)$ and the RHS be $g(x)$ then $f,g$ are holomorphic with radius of convergence $1$, and hence so is $f-g$. If the zero set of $f-g$ (i.e. where $f=g$) has a limit point in the interior of the unit disc, then $f=g$ identically, so we must have that the zero set of $f-g$ (in the unit interval) is a sequence approaching $1$. Furthermore, let $t_n$ be the sequence of zeros of $f-g$, we may cite the Blaschke condition to rule out sequences such as the harmonic series (i.e. $t_n = 1-\frac 1n$ since we need the sum of $1-t_n$ to be finite).

Also, suppose $h(x): = \sum_{n \in \mathbb N} c_nx^n$ is analytic such that the sum of its coefficients is absolutely convergent, and such that $h(x_i) = 0$ for some sequence $x_i$ tending to 1, then we can normalize $c_n$ such that $\sum_{n \in \mathbb N}c_n^+ = \sum_{n \in \mathbb N}c_n^-=1$ (the superscript $+$ denotes $\min(\cdot,0)$, similarly for $-$), then letting $f=\sum_{n \in \mathbb N}c_n^+x^n, g=\sum_{n \in \mathbb N}c_n^-x^n$ gives us the desired coefficients.

To this end, I think $h(x)=\sin\left(\frac \pi {\ln 2}\ln(1-x)\right)$ works but I have no proof of this (the coefficients are hard to compute explicitly).

My question is as follows:

  1. I know that the product of Blaschke factors (which satisfy the Blaschke condition) defines a bounded holomorphic function on the unit disc, is it true that such a function also has absolutely convergent coefficients?
  2. can the coefficients of $h(x)=\sin\left(\frac \pi {\ln 2}\ln(1-x)\right)$ be explicitly found/proven to be absolutely convergent in sum or is it just too much effort? If so is there a example with simpler example that's easier to work with?
  3. If such an example cannot be written down exactly. Is it possible to at least prove the existence of it? Or is there no such example possible?
1

There are 1 best solutions below

0
On BEST ANSWER

Context: This problem was originally part of homework of a graduate probability class which was not meant to be handed in. I deleted the question originally because this particular problem was made into extra credit for the first person who solved it. I am now undeleting it since I think this is a fairly interesting problem and the extra credit has been claimed, so I think it might be good to have a record of this solution for future reference. The idea of the following solution is from my classmate, the explicit example is my original work.

We wish to define $f,g:[0,1] \to [0,1]$ with all positive coefficients such that $f>g$ for infinitely many points and $g>f$ for infinitely many points, and such that $f(1)=g(1)=1$. To do so, we recursively define two sequences $\{(n_k,m_k)\}_{k \in \mathbb N}$ and $\{t_k\}_{k \in \mathbb N}$, we also now define $f_n(x): = \sum_{i=0}^n2^{-i-1}x^{n_i}$ and $g_n(x): = \sum_{i=0}^n2^{-i-1}x^{m_i}$. Let $t_0=0=n_0 = m_0-1$, then $f_0 = \frac 12$ and $g_0 = \frac t2$ and $f_0(t_0) = \frac 12 > 0 = g_0(t_0)$. We intend to make these defined such that $f_N(t_k)>g_N(t_k)$ for $k<N$ even and $g_N(t_k)>f_N(t_k)$ for $k<N$ odd, and the $n_k$s are the powers of $t$ in the power series expansion of $f$ with nonzero coefficient $2^{-k-1}$, and $m_k$ for $g$.\\ Now suppose $f_k,g_k,t_k,n_k,m_k$ are defined, we split into two cases, if $k$ even, then $\epsilon:=f_k(t_k)-g_k(t_k)>0$, and if odd, we have $g_k(t_k)-f_k(t_k)>0$ in which case we repeat the even construction, replacing $f \to g$. So we fix some $m_{k+1}$ such that $2^{-k-1}t_k^{m_{k+1}} < \epsilon$, so that we have: \begin{align*} f_k(t_k)>g_k(t_k)+2^{-k-1}t_k^{m_{k+1}} & \geq g_k(t_k)+2^{-k-2}t_k^{m_{k+1}}(\sum_{i \in \mathbb N}2^{-i}) \\ & \geq g_k(t_k)+2^{-k-2}t_k^{m_{k+1}}(\sum_{i \in \mathbb N}2^{-i}t_k^i)\\ & \geq g_k(t_k)+2^{-k-2}t_k^{m_{k+1}}+2^{-k-3}t_k^{m_{k+1}+1}+ \dots \end{align*} Hence we have: $$f_k(t_k)>g_k(t_k)+2^{-k-1}t_k^{m_{k+1}} > g_k(t_k)+2^{-k-2}t_k^{m_{k+1}}=g_{k+1}(t_k) $$ Now, if we evaluate at $1$: $$f_k(1) = \sum_{i=0}^k2^{-i-1} <\sum_{i=0}^k2^{-i-1}+2^{-k-2} = g_k(1)+2^{-k-2} = g_{k+1}(1)$$ Hence there exists $t_k<t_{k+1}<1$ such that $g_{k+1}(t_{k+1}) > f_k(t_{k+1})$. Now we repeat the process at $t_{k+1}$, choose $2^{-k-1}t_{k+1}^{n_{k+1}} < g_{k+1}(t_{k+1})-f_k(t_{k+1})$ so that we have: \begin{align*} g_{k+1}(t_{k+1})>f_k(t_{k+1})+2^{-k-1}t_{k+1}^{n_{k+1}} & \geq f_k(t_{k+1})+2^{-k-2}t_{k+1}^{n_{k+1}}(\sum_{i \in \mathbb N}2^{-i}) \\ & \geq f_k(t_{k+1})+2^{-k-2}t_{k+1}^{n_{k+1}}(\sum_{i \in \mathbb N}2^{-i}t_{k+1}^i)\\ & \geq f_k(t_{k+1})+2^{-k-2}t_{k+1}^{n_{k+1}}+2^{-k-3}t_{k+1}^{n_{k+1}+1}+ \dots \end{align*} Hence we have: $$g_{k+1}(t_{k+1}) > f_k(t_{k+1})+2^{-k-2}t_{k+1}^{n_{k+1}}=f_{k+1}(t_{k+1})$$ And we also have: $$g_{k+1}(1) = \sum_{i=0}^{k+1} 2^{-i-1}=f_{k+1}(1)$$ We define $f,g$ as the limits of $f_n,g_n$ respectively, both have coefficients which are absolutely convergent to $1$ since the coefficients are just dyadics. \\ Claim that $f(t_k) > g(t_k)$ for all even $k$ and the inequality reverses for all odd $k$. Indeed, fix $k$ even, by construction we have: $$f_k(t_k) > g_k(t_k) + 2^{-k-1}t_k^{m+1} \geq g_k(t_k)+2^{-k-2}t_k^{m_{k+1}} +2^{-k-3}t_k^{m_{k+2}}+2^{-k-3}t_k^{m_{k+2}}+ \dots=g(t_k)$$ And also: $$f_k(t_k) < f_k(t_k) + 2^{-k-2}t_k^{n_k+1} + \dots=f(t_k)$$ The proof for the odd case follows similarly by the exact same computation (but switch $f \to g$). Now we may conclude by IVT that there exists $\{s_n\}_{n \in \mathbb N}$ with $s_n \in (t_n, t_{n+1})$ with $f(s_n) = g(s_n)$ for all $n \in \mathbb N$. \\ In summary, we constructed: $$f: = \sum_{i \in \mathbb N} 2^{-i-1}t^{n_i}, g: = \sum_{i \in \mathbb N} 2^{-i-1}t^{m_i}$$ which clearly satisfy: $$f(1)=g(1) = \sum_{i \in \mathbb N} 2 ^{-i-1} = 1 $$ And which are equal on a countably infinite set $\{s_n\}_{n \in \mathbb N}$.