Let $p_1, ... ,p_k$ denote the probabilities of drawing bin $1, .. ,k$, where $\sum_{i = 1}^{k} p_i= 1$. My question is if we draw $n$ times, how can I show that the probability that all bins are drawn less than $C$ times is maximized by $p_1 = p_2 = ... = p_k = 1/k$ ?

I've tried using conditional probability to solve inductively such as assuming bin $i$ is drawn $<C$ times and trying to reason about the remaining $k-1$, but no luck so far. Computing explicitly seems to be really difficult.


Here is a proof using generating functions. Let $c\geq 2$ and $k\geq 2$ be fixed.

Let $X=(X_1(n),\ldots,X_k(n))$ be the $k$-tuple of occupancy numbers at time $n$, i.e. $X_i(n)$ = number of bins of type $i$ drawn at time $n$.
Clearly $X$ has the multinomial distribution with parameters $n$ and $p=(p_1,\ldots,p_k)$. Therefore
$$\mathbb{P}( \max X_i(n)\leq c-1)=n!\,[t^n] \prod _{i=1}^n q_c(p_it)$$

where $q_c(t):=\sum_{j=0}^{c-1}\frac{t^j}{j!}$ is the $c$th partial sum of $\exp(t)$.

The following lemma is the key.

When $p_i,p_j$ are each replaced by their arithmetic mean $a:=(p_i+p_j)/2$
$\mathbb{P}(\max X_i(n) \leq c-1 )$ will not decrease.
If $c\leq n \leq k(c-1)$ and all other $p_\ell$ are positive $\mathbb{P}(\max X_i(n) \leq c-1 )$ will strictly increase if $p_i\not = p_j$.

Proof Consider first the case of two factors. Let $x,y\in\mathbb{R}_+,x<y$. We show that $$ [t^m] q_c(xt)q_c(yt)\leq [t^m] \left(q_c(((x+y)/2)t\right)^2\;\;.$$It is easy to see that equality holds for $m\leq c-1$ and $m>2c-2$. Let $c\leq m \leq 2c-2$. We have $$m!\,[t^m] q_{c}(xt)q_c(yt)=(x+y)^m-\sum_{i=0}^{m-c} {m \choose i}\left(x^i y^{m-i}+y^i x^{m-i}\right) $$ For fixed sum $s=x+y$ the function $x \mapsto f(x):=\sum_{i=0}^{m-c} {m \choose i}\left(x^i y^{m-i}+y^i x^{m-i}\right)$ has the derivative $f^\prime(x)=(m-c+1){m \choose c} \left(x^{m-c+1}(s-x)^c-(s-x)^{m-c+1}x^c\right)$. Thus $x\mapsto f(x)$ is strictly decreasing (resp. increasing) on $[0,s/2]$ (resp. $[s/2,s]$), attaining its minimum at $x=s/2$. The rest is easy. End proof.

For any distribution $p=(p_1,\ldots,p_k)$ we have that $\mathbb{P}( \max X_i(n)\leq c-1)=1$ for $n\leq c-1$, and $\mathbb{P}( \max X_i(n)\leq c-1)=0$ for $n > k(c-1)$, so that only the cases $c\leq n \leq k(c-1)$ are of interest.

For $c\leq n \leq k(c-1)$ the uniform distribution $p_1=\ldots=p_k=\tfrac{1}{k}$ uniquely maximises $\mathbb{P}(\max X_i(n) \leq c-1 )$.

Proof: Since we are maximising a continuous function (multivariate polynomial) on a compact set (simplex) the maximum is attained. Let $p$ be a maximising distribution. No $p_i$ can be $0$ (else replacing the last $0$ strictly increases the probability). But then $p$ can only be maximising if $p$ is the uniform distribution. End proof.

Remark Lemma 2 shows that the probability $\mathbb{P}(\max X_i(n) \leq c-1 )$ is a Schur-concave function on the simplex $p_i\geq 0, \sum_{i=1}^k p_i=1$. Possibly the result can be found under this heading in the literature (but a brief Google-search didn't reveal anything).


Experts may have a better answer.

This is I believe very hard if not impossible to do exactly, but if you use poissonization, i.e., model $X_i$ as independent poisson arrivals into bin $i$ with rate $np_i$ then the new process $(X_1,\ldots, X_k)$ equals the original bin arrival process $(Y_1,\ldots,Y_k)$ in expectation.

There is also a property that the probability of a monotone event can only be at most a factor of $2$ different in the bin arrival model and this model (I saw this in Mitzenmacher and Upfal's Probability and Computing textbook). Your event

$$ A=\{ \omega: \textrm{Max } X_i \leq C\} $$ is monotone. Maybe you can show strong concentration with upper and lower bounds on $A$ in the poisson model.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.