2
$\begingroup$
  • We consider infinite binary sequences $x \in[0,1]$ via their binary expansion.
  • $O: \{0,1\}^* \rightarrow \{0,1\}^*$ maps finite binary strings to finite binary strings.
  • for a string $m$ we define: $$ s_O(x,m) = \left\{ \begin{array}{ll} 0 \mbox{ if O(m) is not a prefix of x }\\ |O(m)|- |m| \mbox{ otherwise} \end{array} \right. $$
  • The question: is there such an $O$ such that: $\mathbb{P}_{x \mbox{ uniform in [0,1]}} (\mbox{sup}_m (s_O(m,x)) = + \infty) > 0$

I ve a ton of questions on $\mbox{sup}_m (s_O(m,x))$ , this one has to come first.

edited: $|O(m)|- \mbox{ln}_2(m) \rightarrow |O(m)|- \mbox{ln}_2(|m|) \rightarrow |O(m)|- \mbox |m|$ sorry

Thank you Brian for the parantheses editing.

$\endgroup$

2 Answers 2

2
$\begingroup$

Yes, such an $O$ exists. In fact, one can construct $O$ so that $\mathbb{P}\!\left(\sup_m s_O(x,m)=+\infty\right)=1$.

Here is one explicit construction.

For each integer $j\ge 1$, define $k_j:=\lceil \log_2 j\rceil$. Then $k_j\to\infty$ as $j\to\infty$, but it grows slowly enough that $\sum_{j=1}^\infty 2^{-k_j}=\infty$. Next choose integers $n_j$ so that the blocks of coordinates we use are disjoint. For example, set $n_1:=0$ and $n_{j+1}:=n_j+k_j+1$.Then the intervals $$I_j:=\{n_j+1,\dots,n_j+k_j\}$$ are pairwise disjoint. Now define the map $O:{0,1}^*\to{0,1}^*$ as follows. If a finite binary string $m$ has length $|m|=n_j$ for some $j$, then let $O(m):=m\,0^{k_j}$, that is, append $k_j$ zeros to the end of $m$. For all other strings $m$, define $O(m):=m$.

We now show that this choice works. Let $x=x_1x_2x_3\cdots$ be a random infinite binary sequence, with independent fair bits. For each $j$, consider the event $E_j:=\{x_{n_j+1}=\cdots=x_{n_j+k_j}=0\}$. This is the event that the block of bits of $x$ on the interval $I_j$ consists entirely of zeros. Since the intervals $I_j$ are pairwise disjoint, the events $E_j$ are independent. Also, because each bit is $0$ with probability $1/2$, we have $\mathbb{P}(E_j)=2^{-k_j}$.

Now suppose that $E_j$ occurs. Let $m$ be the prefix of $x$ of length $n_j$, namely $m=x_1x_2\cdots x_{n_j}$. By construction, $O(m)=m\,0^{k_j}$. Because $E_j$ says that the next $k_j$ bits of $x$ are all zero, it follows that $O(m)$ is a prefix of $x$.

Therefore, from the definition of $s_O(x,m)$, we get: $$s_O(x,m)=|O(m)|-|m|=(n_j+k_j)-n_j=k_j$$ So whenever $E_j$ occurs, there exists some $m$ such that $s_O(x,m)=k_j$. Hence $$\sup_m s_O(x,m)\ge k_j$$ on the event $E_j$. Now $\sum_{j=1}^\infty \mathbb{P}(E_j)=\sum_{j=1}^\infty 2^{-k_j}=\infty$, and the events $E_j$ are independent. Therefore, by the second Borel--Cantelli lemma, with probability $1$, infinitely many of the events $E_j$ occur. But whenever $E_j$ occurs, we have $\sup_m s_O(x,m)\ge k_j$, and since $k_j\to\infty$, these lower bounds become arbitrarily large. Therefore, with probability $1$, $\sup_m s_O(x,m)=+\infty$. So the answer to your question is yes, and in fact one can achieve the stronger statement $$\boxed{\mathbb{P}\!\left(\sup_m s_O(x,m)=+\infty\right)=1.}$$

EDIT: Only one part needs to be corrected. The construction of $O$ stays the same, and the Borel--Cantelli argument stays the same. The only point that needs to be replaced is the step where one justifies that the modified score still goes to $+\infty$ after subtracting the penalty term $\ln|m|$. The correct statement to insert is the following. On the event $E_j$, if $m$ is the prefix of $x$ of length $n_j$, then: $$s_O(x,m)=k_j-\ln n_j$$

Therefore it is enough to prove that $k_j-\ln n_j\to+\infty$. Now $n_j=1+\sum_{i=1}^{j-1}(k_i+1)$, and since $k_i=\lceil \log_2 i\rceil\le \log_2 i+1$, we get $n_j\le 1+\sum_{i=1}^{j-1}(\log_2 i+2)$. In particular, there exists a constant $C>0$ such that for all sufficiently large $j$, $n_j\le C_j\log j$. Taking natural logarithms gives $\ln n_j\le \ln C+\ln j+\ln\ln j$. On the other hand, $k_j=\lceil \log_2 j\rceil\ge \log_2 j=\dfrac{\ln j}{\ln 2}$. Hence: $$k_j-\ln n_j \ge \dfrac{\ln j}{\ln 2}-(\ln C+\ln j+\ln\ln j)$$ Equivalently, $$k_j-\ln n_j\ge\left(\dfrac{1}{\ln 2}-1\right)\ln j-\ln\ln j-\ln C$$ Since $\dfrac{1}{\ln 2}-1>0$ and $\ln\ln j=o(\ln j)$, the right-hand side tends to $+\infty$. Therefore $k_j-\ln n_j\to+\infty$.

$\endgroup$
4
  • $\begingroup$ I read 3 times, refreshed me on borel-cantelli, and it seems on point, thank you, i will reread tomorrow to be sure. This is a bad news for what i want to archieve even if intuitively i felt it, cause i need a function i can integrate giving finite value. I would like to ask a related question: Does the result hold if we add the size of m as cost of its description, so in the definition of s_o it would be |O(m)|-|m|- ln(|m|) instead. Can i add an edit on the original question adding the new one or should i create a new question ? $\endgroup$ Commented 8 hours ago
  • $\begingroup$ Yes. The same kind of construction still works, and in fact you still get probability $1$. $\endgroup$ Commented 8 hours ago
  • $\begingroup$ It s not obvious to me in your construction it s true if: lim ln(n) - ln(sum k< n ln(k)) -> infty if im not mistaken, idk if it's true $\endgroup$ Commented 7 hours ago
  • $\begingroup$ you were right to point that out, I made an edit to the post. $\endgroup$ Commented 2 hours ago
1
$\begingroup$

We can get probability $1$, and, moreover, $\sup$ equal to $1$ for all $x$.

For number $k$, let $f(k)$ be a string - $k$ written in binary. For string $s$, let $g(s)$ be number we get interpreting $s$ as binary number, and $s[i:j]$ be substring of $s$ that starts at position $i$ and ends at position $j - 1$ (including, zero-based indexing). For strings $p$ and $q$ let $p + q$ be concatenation of $p$ and $q$.

Let $O(m) = (f(|m|))[1:] + m$. For any $x$ and $k$, let $t = g("1" + x[:k])$ and $m = x[k : k + t]$. Then $f(|m|) = f(t) = "1" + x[:k]$ and $O(m) = ("1" + x[:k])[1:] + x[k:k+t] = x[:k] + x[k:k+t] = x[:k+t]$.

Thus, $s_O(m, x) = (k + t) - t = k$.

So, $\forall x, k \exists m: s_O(m, x) = k$, and thus $\forall x: \sup_m s_O(m, x) = \infty$.

$\endgroup$

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.