How to mathematically compute the transition probabilities in the secretary problem

120 Views Asked by At

My model for the secretary problem is as in (1) in this Mathstackexchange question. More precisely, I look at the unit interval $I=[0,1]$, and $\Omega = I^n \smallsetminus D$ where $D \subset I^n$ is the set of points with at least two identical coordinates (e.g. $(\frac{1}{n},\frac{1}{n},\frac{1}{n},\frac{4}{n},\dots,1)$ would lie in $D$). On $\Omega$ there the Lebesgue probability.

Definition: Given a point $a =(a_1,\dots, a_n) \in \Omega$, an index $k \in \{1,\dots, n\}$ is said to be a maximum if $\max \{ a_i: i=1,\dots, k-1\} < a_k$. By way of definition, $1$ is always a maximum and further, we also put $n+1$ as a maximum.

Given a point $a \in \Omega$, we let $X_1(a),\dots, X_n(a)$ be its set of maxima arranged in increasing order. In particular, $X_1(a) =1$ for every $a \in \Omega$ and if we happen to have $m < n$ with $X_m(a) = k$ and $a_k = \max\{a_i: i =1,\dots, n\}$, then we set $X_{m+1}(a) = \dots = X_n(a) = n+1$. It is easy to see that the $X_i's$ are measurable functions. Indeed, the inverse image of an index $k$ is always the union of simplices of the form $\{a = (a_1, \dots, a_n)\in \Omega: a_{\sigma(1)} < \dots < a_{\sigma(n)}\}$ where $\sigma$ is some permutation.

Question: Given $i \in \{1,\dots, n-1\}$ and $i \leq k < l\leq n$, how do I prove the conditional probability formula \begin{equation}\label{1}\tag{1} P[X_{i+1} =l| X_{i}=k] = \frac{k}{l(l-1)}? \end{equation} Note, if $i=1$ I only consider the case $k=1$. I saw this formula in the book of Billingsley as well as the book of Dynkin-Yushkevich.

Attempt: For example, if I want to compute $P[X_2=k| X_1=1]$, I compute the volume of the union of simplices $$ \left\lbrace a =(a_1,\dots, a_n) \in \Omega: a_{\sigma(2)} < \dots < a_{\sigma(k-1)} < a_1 < a_k\right\rbrace $$ where $\sigma$ varies over all permutations over $k-2$ symbols. Since each simplex has volume $1/k!$, I get the formula $$ P[X_2 = k| X_1 = 1] = \frac{1}{k(k-1)}. $$ I can do something similar for $P[X_3 = l| X_2 = k]$ by counting the number for simplices occurring in the set $$ \left\lbrace (a_1, \dots, a_n) \in \Omega: \max\{a_2,\dots, a_{k-1}\} < a_1 < a_k < a_l \text{ and } \max\{a_{k+1},\dots, a_{l-1}\} < a_k. \right\rbrace$$ But if I try to go to $P[X_4=m|X_3=l]$, I become incapacitated.

Can someone please help me see the truth of formula \eqref{1}? I am weak in probability and combinatorics.

1

There are 1 best solutions below

0
On BEST ANSWER

Ok, so consider $1<k<l\leq n$ and $1<i \leq k$ and the conditional probability $$ P[X_{i+1}=l | X_i =k] := \frac{P[X_{i+1}=l,\ X_i = k]}{P[X_i=k]}. $$ First note that each of the sets in the above fraction is a subset of $\Omega = I^n \smallsetminus D$ where the last coordinates $a_{l+1},\dots,a_n$ are free. Thus, in light of the product structure of Lebesgue probability, it suffices to work in the setup where $n=l$.

Then note that the set $\{a \in \Omega: X_i(a)=k\}$ is, upto a set of measure zero, of the form \begin{equation} \bigcup_{\sigma \in \mathcal{F}} (\Delta_\sigma\times I^{l-k}) = \left(\bigcup_{\sigma \in \mathcal{F}}\Delta_\sigma\right) \times I^{k-l} \end{equation} where $\mathcal{F}$ is some set of permutations of $\{1,\dots,k\}$ and $$ \Delta_\sigma = \left\lbrace (a_1,\dots,a_k) \in I^k: a_{\sigma(1)} < \dots < a_{\sigma(k)} (= a_k)\right\rbrace. $$ Let $\pi: \Omega \to I^k$ denote the projection to the first $k$ coordinates. The set $$S:= \{a \in \Omega: X_{i+1}(a) = l,\ X_i(a) = k\} \text{ projects onto } T:= \bigcup_{\sigma \in \mathcal{F}} \Delta_\sigma.$$ Thus, we have $$ S = \bigcup_{\sigma \in \mathcal{F}} \left((\pi^{-1} \Delta_\sigma) \cap S\right) $$ For a fixed $\sigma \in \mathcal{F}$, a point $(a_1,\dots,a_k,\dots,a_l)$ lies in $(\pi^{-1} \Delta_\sigma) \cap S$ only if \begin{equation}\label{2}\tag{2} \max \{a_{k+1},\dots,a_{l-1}\} < a_k < a_l. \end{equation} Since we also have that $$ \max \{a_1, \dots, a_{k-1}\} < a_k \text{ (by the defining property of } \mathcal{F}), $$ the condition \eqref{2} gives us \begin{equation}\label{3}\tag{3} k\times(k+1) \times \dots \times (l-2) \end{equation} different configurations for $\{a_{k+1},\dots, a_{l-1}\}$. In particular, this demonstrates that $ (\pi^{-1}\Delta_\sigma) \cap S$ is the union of \eqref{3} many simplices of the form $$ \Phi_\tau := \left\lbrace (a_1,\dots, a_l) \in I^l : a_{\tau(1)}< \dots < a_{\tau(l)}\right\rbrace \text{ where } \tau \text{ is some permutation of } \{1,\dots, l\}. $$ And, more importantly, this estimate \eqref{3} is independent of $\sigma \in \mathcal{F}$. Thus we can compute \begin{equation} \begin{split} \frac{P[X_{i+1}=l,\ X_i = k]}{P[X_i=k]} &= \frac{1}{P[X_i=k]}\sum_{\sigma \in \mathcal{F}} P[(\pi^{-1}\Delta_\sigma) \cap S] \\&= \frac{1}{P[X_i=k]} \left( \# \mathcal{F} \cdot \frac{k\cdots (l-2)}{l!}\right) \\ &= \frac{1}{\sum_{\sigma \in \mathcal{F}} P[\Delta_{\sigma} \times I^{l-k}]} \left( \# \mathcal{F} \cdot \frac{k\cdots (l-2)}{l!}\right) \\ &= \frac{1}{\# \mathcal{F} \frac{1}{k!}} \left( \# \mathcal{F} \cdot \frac{k\cdots (l-2)}{l!}\right) \\ &= \frac{k}{l(l-1)}. \end{split} \end{equation}