Lower bounds on binomial and Poisson tails: an approach via tail conditional expectations
Extracted source text · arXiv 1609.06651
\documentclass[11pt]{article}
%\documentclass[12pt, oneside]{amsart} % use "amsart" instead of "article" for AMSLaTeX format
\usepackage{geometry} % See geometry.pdf to learn the layout options. There are lots.
%\geometry{a4paper} % ... or a4paper or a5paper or ...
%\geometry{landscape} % Activate for for rotated page geometry
\usepackage[parfill]{parskip} % Activate to begin paragraphs with an empty line rather than an indent
\usepackage{graphicx} % Use pdf, png, jpg, or eps§ with pdflatex; use eps in DVI mode
\usepackage{subfig} % TeX will automatically convert eps --> pdf in pdflatex
\usepackage{amssymb}
\usepackage{amsmath}
\usepackage{palatino}
\usepackage{graphicx}
\usepackage[english]{babel}
\usepackage{amscd}
\usepackage{amsthm}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{definition}{Definition}
\newtheorem{conjecture}[theorem]{Conjecture}
\newtheorem{example}[theorem]{Example}
\newtheorem{problem}[theorem]{Problem}
\newtheorem{question}[theorem]{Question}
\renewcommand{\labelitemi}{-}
\title{Lower bounds on binomial and Poisson tails: an approach via tail conditional expectations}
\author{Christos Pelekis\thanks{The Czech Academy of Sciences, Institute of Computer Science, Pod Vod\'{a}renskou v\v{e}\v{z}\'{\i} 2, 182 07 Prague, Czech Republic. E-mail: pelekis.chr@gmail.com} }
%\date{} % Activate to display a given date or no date
\begin{document}
\maketitle
\begin{abstract} We derive upper bounds on the tail conditional expectation of binomial and Poisson random variables.
Those upper bounds are subsequently employed to the problem of obtaining non-asymptotic lower bounds on the probability that the aforementioned random variables are
significantly larger than their expectation.
\end{abstract}
\noindent \emph{Keywords}: lower bounds; binomial tails; Poisson tails; tail conditional expectation
\noindent \emph{MSC (2010):} 60E15; 60E05
\section{Prologue and main result}
Given a positive integer $n$ and a real number $p\in (0,1)$, we denote by $\text{Bin}(n,p)$ a binomial
random variable of parameters $n$ and $p$. Given a positive real number $\mu$, we denote by $\text{Poi}(\mu)$ a Poisson random variable of mean $\mu$.
Throughout the text, given two random variables $X,Y$, the notation
$X\sim Y$ will indicate that they have the same distribution.
In this article we shall be interested in \emph{lower bounds} on
the "tail" probabilities $\mathbb{P}\left[X\geq k\right]$, where $X\sim \text{Bin}(n,p)$ and $k>np$, and $\mathbb{P}\left[Y\ge k\right]$, where $Y\sim \text{Poi}(\mu)$ and $k>\mu$.
The problem of estimating the tail of a binomial random variable is arguably one of the most basic problems in
probability and statistics.
It arises in a plethora of problems and, despite the fact that the binomial tails are asymptotically
well understood (see, for example, \cite{Bahadur, Carter, Fellerzero, Littlewood, Massart, McKay, Slud, Zubkov}),
several questions coming from a variety of disciplines (see \cite{Alon, Ash, Greenberg, Jerabek, Rigollet})
do not decrease the need for upper and lower
bounds on $\mathbb{P}\left[\text{Bin}(n,p)\geq k\right]$ that are valid for fixed $n,p,k$.
Sharp upper bounds on binomial tails are provided by the, so-called, Chernoff-Hoeffding bound (see \cite{Hoeffdingone})
by employing an "exponential moment" approach:
\begin{equation}\label{Hoebound}
\mathbb{P}\left[\text{Bin}(n,p)\geq k\right] \leq \text{exp}\left(-n D(k/n||p)\right), \; \text{for}\; k>np,
\end{equation}
where $D(k/n||p)$ is the Kullback-Leibler distance between two Bernoulli $0/1$ random variables of parameter $k/n$ and $p$, respectively.
Upper bounds that are sharper than the previous one can be obtained by employing
a "factorial moment" approach (see \cite{Schmidt}):
\begin{equation}\label{HoeboundSiegel}
\mathbb{P}\left[\text{Bin}(n,p)\geq k\right] \leq p^{\ell} \binom{n}{\ell}/\binom{k}{\ell}, \; \text{where}\;
\ell = \big\lceil \frac{k-np}{1-p} \big\rceil , \; k>np.
\end{equation}
Here and later, $\lceil x\rceil$ denotes the minimum integer that is larger than or equal to $x$.
Elementary, though quite tedious, calculations (see \cite{Schmidt} for details) show that the bound
given by (\ref{HoeboundSiegel}) is smaller than the bound given by (\ref{Hoebound}) but the two
bounds are quite close to each other.
A well-known theorem of Cram\'er implies that the bound of Eq.(\ref{Hoebound}) is asymptotically tight and therefore one can only hope for sharp lower bounds on the
binomial tails that are a portion of the
Chernoff-Hoeffding bound.
In that regard, one has the following result (see \cite[Lemma $4.7.2$]{Ash}):
\begin{equation}\label{Ashbound}
\mathbb{P}\left[\text{Bin}(n,p)\geq k\right]\geq \frac{1}{\sqrt{8k(1-k/n)}}\cdot \text{exp}\left(-n D(k/n||p)\right), \; \text{for}\; k>np.
\end{equation}
Briefly, this bound
follows from the observation
$\mathbb{P}\left[\text{Bin}(n,p)\geq k\right]\geq \binom{n}{k}p^k(1-p)^{n-k}$ combined with elaborate use of entropy estimates on binomial coefficients and appears
to be among the best known lower bounds.
Several lower bounds on the tail of binomial random variables
can be found in \cite{Diaconis, Doerr, Greenberg, Jerabek, PelekisRamon, Telgarsky}, among many others.
Let us remark that most lower bounds in the literature appear to be
obtained using ideas from asymptotic analysis and/or rely on estimates of the
normal approximation to the binomial distribution. In this article we provide yet another lower
bound on binomial tails which results in a proportion of the bound given by (\ref{HoeboundSiegel}).
We embark on a different
approach that employs properties of tail conditional expectations.
In the case of Poisson random variables, the Chernoff-Hoeffding bound provides an upper estimate on the probability that a $\text{Poi}(\mu)$ random variable is significantly larger than its mean and reads as follows (see \cite[page 97]{Mitzenmacher}):
\begin{equation}\label{Poi_upper_bound}
\mathbb{P}\left[\text{Poi}(\mu) \ge k\right] \le \frac{e^{\mu}(e\mu)^k}{k^k}, \; \text{for} \; k >\mu.
\end{equation}
Several upper bounds on Poisson tails can be found in \cite{Glynn, Teicher}, among others.
However, the problem of obtaining lower bounds on Poisson tails seems to be less investigated and we were not able to find a systematic treatment of this topic.
Let us proceed by introducing our main results.
Notice that the trivial estimates
\[ p^k \leq \mathbb{P}\left[\text{Bin}(n,p)\geq k\right] \leq \binom{n}{k}p^k\]
imply that there exists a constant, $C_{n,p,k}$, that depends on $n,p,k$ such that
\[ \mathbb{P}\left[\text{Bin}(n,p)\geq k\right] = C_{n,p,k} \cdot p^k . \]
It turns out that $C_{n,p,k}$ can be described in terms of tails conditional expectations of certain binomial
random variables. More precisely, we will see in the next section that
\[ C_{n,p,k} = \prod_{j=0}^{k-1} \frac{n-j}{\mathbb{E}\left[X_{n-j}| X_{n-j}\geq k-j\right]}, \; \text{where} \;
X_{n-j} \sim \text{Bin}(n-j,p) \]
and so the problem of obtaining lower bounds on the tail probability is reduced to the problem
of obtaining upper bounds on tail conditional expectations of binomial random variables.
The main novelty of our paper is that it appears to be the first report that investigates upper
bounds on tail conditional expectations as a tool to obtain lower bounds on tail probabilities.
Let us also remark that the problem of obtaining estimates on the tail conditional expectation (a.k.a. \emph{conditional value at risk}) of spesific distributions is of particular interest in actuarial sciences (see, for example, \cite{Landsman, WangZhao}).
Our bounds on the tail conditional expectation and the tail probability of a binomial random variable read as follows.
Here and later, given a real number $x$, we will denote by $\lfloor x \rfloor$ the maximum integer that
is less than or equal to $x$.\\
\begin{theorem}\label{mainthm} Let $X \sim \text{Bin}(n,p)$ and fix positive integer $k$ such that $np < k \leq n$.
Then
\[ \mathbb{E}\left[X | X\geq k\right] \leq k + \frac{(n-k)p}{k-np+p} . \]
Furthermore, if $k$ satisfies $np<k\le n-1$ we have
\[ \mathbb{P}\left[X_n \geq k\right] \geq \frac{p^{2(\ell+1)}}{2}\cdot \binom{n}{\ell +1}/\binom{k}{\ell +1} ,\]
where $\ell := \lfloor \frac{k-np}{1-p} \rfloor <k$.
\end{theorem}
Notice that when $k$ is such that $np<k< np + 1-p$, then $\ell =0$ and the lower bound of the previous result reduces to
\begin{equation}\label{eqq:1}
\mathbb{P}\left[X\ge np\right] \ge \frac{1}{2}\cdot p^2\cdot\frac{n}{k}\ge \frac{p}{2}\cdot \left(1- \frac{1-p}{k}\right)\ge \frac{p}{2}\cdot \left(1- \frac{1}{k}\right)
\end{equation}
and the later bound is larger than $\frac{1}{4}\cdot \frac{9}{10}$, when $p>1/2$ and $k\ge 10$. We refer the reader to \cite{Doerr, Greenberg, PelekisRamon} for sharper lower bounds on the the tail probability in (\ref{eqq:1}).
In other words, the second statement of Theorem \ref{mainthm} provides a lower bound on
binomial tails that is equal to a particular proportion of the upper bound given by (\ref{HoeboundSiegel}).
We prove Theorem \ref{mainthm} in Section \ref{sec:2}. Let us remark that comparisons between
the lower bound provided by Theorem \ref{mainthm} and the lower
bound given by (\ref{Ashbound}) require quite tedious calculations. However,
it is rather straightforward
to put the computer to work and see that the bound of
Theorem \ref{mainthm} sometimes performs better than the
bound given by (\ref{Ashbound}). We provide some pictorial comparisons between the
two bounds in Section \ref{compare}.
Finally, in Section \ref{sec:3}, we show that similar ideas apply to Poisson random variables.
In particular, we obtain the following. \\
\begin{theorem}\label{poibound}
Let $Y\sim \text{Poi}(\mu)$, where $\mu>0$, and fix a positive integer $k$ such that $k\ge\mu$.
Then
\[ \mathbb{E}\left[Y|Y\ge k\right] \le k +\frac{\mu}{k+1-\mu}. \]
Furthermore, if we set $\ell = \lfloor k-\mu \rfloor$ we have
\[ \mathbb{P}\left[Y \ge k\right] \ge \frac{1}{2}\cdot \left(\frac{\mu}{k+\mu} \right)^{\ell+1} . \]
\end{theorem}
Notice that when $k$ satisfies $\mu\le k < \mu +1$ then we have $k+\mu <2\mu+1$ as well as $\ell =0$ and the lower bound of the previous result reduces to
\begin{equation}\label{eqq:2}
\mathbb{P}\left[Y\ge \mu\right] \ge \frac{1}{2}\cdot \frac{\mu}{2\mu+1} , \; \text{for}\; Y\sim \text{Poi}(\mu),
\end{equation}
and the right hand side is larger than $1/5$, when $\mu\ge 2$.
\section{Proof of Theorem \ref{mainthm}}\label{sec:2}
In this section we prove Theorem \ref{mainthm}.
Let us begin by recalling the following well-known result regarding the median of a binomial random variable. \\
\begin{theorem}[see \cite{Kaas}]
\label{kaas} Let $X \sim \text{Bin}(n,p)$. Then
$\mathbb{P}\left[ X\geq \lfloor np\rfloor \right] \geq 1/2$.
\end{theorem}
A basic ingredient in the proof of Theorem \ref{mainthm} is the following relation
between the tails of a binomial random variable and its tail conditional expectation. \\
\begin{lemma}\label{basiclem} Fix a positive integer $n$ and a real number $p\in (0,1)$.
For $j=1,\ldots,n$ let $X_j$ be a $\text{Bin}(j,p)$ random variable.
If $k$ is a positive integer such that $1\leq k \leq n-1$. Then
\[ \mathbb{P}\left[X_n\geq k \right] = \frac{np}{\mathbb{E}\left[X_n | X_n\geq k\right]} \cdot
\mathbb{P}\left[X_{n-1}\geq k-1\right] . \]
\end{lemma}
\begin{proof}
Since $X_n = \sum_i B_i$ is a sum of $n$ Bernoulli $0/1$ random variables $B_i$ of mean $p$, the linearity of expectation yields
\begin{eqnarray*} \mathbb{E}\left[X_n | X_n\geq k\right] = \sum_{i=1}^{n} \mathbb{P}\left[B_i =1 | X_n\geq k\right] = \frac{n p\cdot \mathbb{P}\left[X_{n-1}\geq k-1\right]}{\mathbb{P}\left[X_{n}\geq k\right]}
\end{eqnarray*}
and the result follows.
\end{proof}
Notice that Lemma \ref{basiclem} can be iterated. This provides a way to express the tail
probability of a binomial random variable in terms of tail conditional expectations. \\
\begin{corollary}\label{cormain} Fix a positive integer $n$ and a real number $p\in (0,1)$.
For every $j\leq n$, let $X_j$ be a $\text{Bin}(j,p)$ random variable. If $\ell$ is any integer from the set
$\{0,1,\ldots,k-1\}$ then
\[ \mathbb{P}\left[X_n\geq k\right] = \mathbb{P}\left[X_{n-(\ell +1)} \geq k-(\ell+1)\right] \cdot \prod_{j=0}^{\ell} \frac{p(n-j)}{\mathbb{E}\left[X_{n-j} \big| X_{n-j}\geq k-j\right]}
.\]
\end{corollary}
\begin{proof} Immediate upon iterating Lemma \ref{basiclem} $\ell +1$ times.
\end{proof}
Hence lower bounds on the tail probability can be provided via
upper bounds on the tail conditional expectations
$\mathbb{E}\left[X_{n-j} \big| X_{n-j}\geq k-j\right]$, for $j=0,\ldots,k-1$.
The next result provides a recursion for tail conditional expectations of integer valued random variables
and is interesting on its own. \\
\begin{lemma}\label{secondlem}
Let $X$ be an integer valued random variable and fix a positive integer $k$ such that
$\mathbb{P}\left[X\geq k\right]>0$ and $\mathbb{P}\left[X\geq k+1\right]>0$ .
Then
\[ \mathbb{E}\left[X| X\geq k\right] = k \cdot \frac{\mathbb{P}\left[X=k\right]}{\mathbb{P}\left[X\geq k\right]} + \mathbb{E}\left[X| X\geq k+1\right] \cdot \left(1- \frac{\mathbb{P}\left[X=k\right]}{\mathbb{P}\left[X\geq k\right]}\right) .\]
\end{lemma}
\begin{proof}
The result follows upon writing
\begin{eqnarray*} \mathbb{E}\left[X| X\geq k\right] &=& k\cdot \frac{\mathbb{P}\left[X=k\right]}{\mathbb{P}\left[X\geq k\right]} + \sum_{j\ge k+1} j\cdot \frac{\mathbb{P}\left[X=j\right]}{\mathbb{P}\left[X\geq k+1\right]}
\cdot \frac{\mathbb{P}\left[X\geq k+1\right]}{\mathbb{P}\left[X\geq k\right]} \\
&=& k \cdot \frac{\mathbb{P}\left[X=k\right]}{\mathbb{P}\left[X\geq k\right]} + \mathbb{E}\left[X| X\geq k+1\right] \cdot \left(1- \frac{\mathbb{P}\left[X=k\right]}{\mathbb{P}\left[X\geq k\right]}\right) .
\end{eqnarray*}
\end{proof}
We will also need the following result which provides estimates on ratios of binomial tails. \\
\begin{lemma}\label{feller} Let $X_n \sim \text{Bin}(n,p)$, for $n=1,2,\ldots$. Then for every positive integer $k$ such that $np<k\leq n$ we have
\[ \mathbb{P}\left[X_n \geq k\right] \leq \frac{k(1-p)}{k-np}\cdot \mathbb{P}\left[X_n = k\right] . \]
In particular,
\[ \frac{\mathbb{P}\left[X_n \geq k+1\right]}{\mathbb{P}\left[X_n \geq k\right]} \leq \frac{p(n-k)}{k(1-p)}. \]
\end{lemma}
\begin{proof} The first statement is a well known result. See, for example, Feller \cite[p. $151$]{Feller} for a proof that employs comparison with a geometric series, or
Diaconis et al. \cite[Theorem $1$]{Diaconis} for a proof that employs the, so-called, Todhunter's formula.
The second statement is equivalent to the first and follows upon observing that
\[ \mathbb{P}\left[X_n \geq k+1\right] = \mathbb{P}\left[X_n \geq k\right] - \mathbb{P}\left[X_n = k\right] . \]
For the sake of completeness, let us provide yet another proof of this result that employs Lemma
\ref{basiclem}.
Notice that, if we apply Lemma \ref{basiclem} with $n$ replaced by $n+1$ and $k$ replaced by
$k+1$, we get
\begin{eqnarray*} 0\leq \mathbb{E}\left[\max\{0,X_{n+1}-(k+1) \} \right] &=& \mathbb{E}\left[X_{n+1} -(k+1) | X_{n+1}\geq k+1\right]
\cdot \mathbb{P}\left[X_{n+1}\geq k+1\right] \\
&=& (n+1)p \cdot \mathbb{P}\left[ X_{n}\geq k \right] - (k+1)\cdot \mathbb{P}\left[X_{n+1} \geq k+1 \right] \\
&=& p(n-k)\cdot \mathbb{P}\left[ X_{n}\geq k \right] - (k+1)(1-p) \cdot \mathbb{P}\left[X_{n} \geq k+1 \right].
\end{eqnarray*}
Hence we have shown that
\[ \frac{\mathbb{P}\left[X_n \geq k+1\right]}{\mathbb{P}\left[X_n \geq k\right]} \leq \frac{p(n-k)}{(k+1)(1-p)} \]
and the result follows.
\end{proof}
We can now proceed with an upper bound on the tail conditional expectation of a binomial random variable.
\begin{proof}[Proof of Theorem \ref{mainthm}(First statement)]
Clearly, $\mathbb{E}\left[X | X\geq n\right] = n$. Moreover, elementary calculations show that
\[ \mathbb{E}\left[X | X\geq n-1\right] = n-1 + \frac{p}{(n-1)(1-p)+p} \leq n-1 + \frac{p}{n-1-np+p} \]
and so the result holds true for $k\in \{n-1,n\}$.
For the remaining values, i.e. $k\leq n-2$, we prove the result by reverse induction on $k$.
More precisely, we prove that the inequality holds true for $k>np$ under the assumption that it
holds true for $k+1$. The assumption
that the inequality holds true for $k+1$ implies that
\[\mathbb{E}\left[X | X\geq k+1\right] \leq k+1 + \frac{(n-k-1)p}{k+1-np+p} . \]
Now notice that the previous inequality combined with Lemma \ref{secondlem} yield
\begin{eqnarray*} \mathbb{E}\left[X| X\geq k\right] &=& k \cdot \frac{\mathbb{P}\left[X=k\right]}{\mathbb{P}\left[X\geq k\right]} + \mathbb{E}\left[X| X\geq k+1\right] \cdot \frac{\mathbb{P}\left[X\geq k+1\right]}{\mathbb{P}\left[X\geq k\right]} \\
&\leq& k + \left(1 + \frac{(n-k-1)p}{k+1-np+p}\right)\cdot \frac{\mathbb{P}\left[X\geq k+1\right]}{\mathbb{P}\left[X\geq k\right]} \\
&\leq& k + \left(1 + \frac{(n-k-1)p}{k+1-np+p}\right) \cdot\frac{p(n-k)}{k(1-p)} ,
\end{eqnarray*}
where the last estimate follows from Lemma \ref{feller}.
Now notice that the result will follow once we show
\[ \left(1 + \frac{(n-k-1)p}{k+1-np+p}\right) \cdot\frac{1}{k(1-p)} \leq \frac{1}{k-np+p} , \]
or, equivalently, once we show
\[ \frac{k+1-kp}{k+1-np+p } \cdot \frac{1}{k(1-p)} \leq \frac{1}{k-np+p}. \]
Since we assume $k< n-1$, it is now easy to verify that the last inequality holds true and the proof is complete.
\end{proof}
We are now ready to prove the second statement of Theorem \ref{mainthm}. The proof combines the previous upper bound on
tail conditional expectations with Corollary \ref{cormain}.
\begin{proof}[Proof of Theorem \ref{mainthm} (Second statement)]
Notice that $\ell = \lfloor \frac{k-np}{1-p} \rfloor$ is the maximum
integer $j$ such that $k-j \geq (n-j)p$.
From Corollary \ref{cormain} we know that
\[ \mathbb{P}\left[X_n\geq k\right] = \mathbb{P}\left[X_{n-(\ell +1)} \geq k-(\ell+1)\right] \cdot \prod_{j=0}^{\ell} \frac{p(n-j)}{\mathbb{E}\left[X_{n-j} \big| X_{n-j}\geq k-j\right]} ,
\]
where $X_{n-j}\sim \text{Bin}(n-j,p)$, for $j=0,\ldots, \ell$.
Since $k-(\ell +1)$ is a positive integer such that $k-(\ell +1) < (n-(\ell +1))p$,
Theorem \ref{kaas} yields
\[ \mathbb{P}\left[X_{n-(\ell +1)} \geq k-(\ell+1)\right] \geq \frac{1}{2} . \]
Now notice that the first statement of Theorem \ref{mainthm} combined with the fact that $j\le \ell <k$ imply
\[ \mathbb{E}\left[X_{n-j} \big| X_{n-j}\geq k-j\right] \le k-j + \frac{n-k}{k-j-(n-j)p+p} \le k-j + \frac{n-j}{k-j-(n-j)p+p} \]
and therefore we have
\[ \frac{p(n-j)}{\mathbb{E}\left[X_{n-j} \big| X_{n-j}\geq k-j\right]} \geq \frac{(n-j)\cdot p\cdot (k-j-(n-j)p)+ (n-j)p^2 }{(k-j)\cdot (k-j-(n-j)p)+(n-j)p} ,\; \text{for}\; j=0,\ldots, \ell .\]
Since $k-j \geq (n-j)p$, for $j=0,\ldots,\ell$,
straightforward calculations show that
\[ \frac{(n-j)\cdot p\cdot (k-j-(n-j)p)+ (n-j)p^2 }{(k-j)\cdot (k-j-(n-j)p)+(n-j)p} \geq \frac{n-j}{k-j}\cdot p^2 . \]
Putting all the above together, we conclude that
\[ \mathbb{P}\left[X_n\geq k\right] \geq \frac{1}{2}\cdot p^{2(\ell+1)}\prod_{j=0}^{\ell} \frac{n-j}{k-j} \]
and the result follows.
\end{proof}
\section{Proof of Theorem \ref{poibound}}\label{sec:3}
In this section we prove Theorem \ref{poibound}. We begin with a result that expresses the tail conditional expectation of a Poisson random variable in terms of a ratio between consecutive tails. \\
\begin{lemma}\label{poilem:1}
Let $Y\sim\text{Poi}(\mu)$, where $\mu>0$, and fix a positive integer $k$.
Then
\[\mathbb{P}\left[Y\ge k\right] = \mu\cdot \frac{\mathbb{P}\left[Y\ge k-1\right]}{\mathbb{E}\left[Y| Y\ge k\right] } . \]
In particular, we have
\[ \frac{\mathbb{P}\left[Y\ge k\right]}{\mathbb{P}\left[Y\ge k-1\right]} \le \frac{\mu}{k} . \]
\end{lemma}
\begin{proof}
Notice that for all $i\ge 1$, we have
\[ i\cdot\mathbb{P}\left[Y=i\right] =\mu\cdot \mathbb{P}\left[Y=i-1\right] .\]
This implies that
\[ \mathbb{E}\left[Y| Y\ge k\right] = \sum_{i=k}^{\infty} i\cdot \frac{\mathbb{P}\left[Y= i\right]}{\mathbb{P}\left[Y\ge k\right]} = \mu\cdot\sum_{i=k}^{\infty} \frac{\mathbb{P}\left[Y= i-1\right]}{\mathbb{P}\left[Y\ge k\right]} =
\mu\cdot \frac{\mathbb{P}\left[Y\ge k-1\right]}{\mathbb{P}\left[Y\ge k\right] } \]
and the first statement follows. The second statement follows from the last equation, upon observing that $\mathbb{E}\left[Y| Y\ge k\right]\ge k$.
\end{proof}
We are now ready to prove the first statement of Theorem \ref{poibound}.
\begin{proof}[Proof of Theorem \ref{poibound} (First statement)]
Notice that
\begin{eqnarray*}
\mathbb{E}\left[Y| Y\ge k\right] &=& \sum_{i=k}^{\infty} i\cdot \frac{\mathbb{P}\left[Y= i\right]}{\mathbb{P}\left[Y\ge k\right]} \\
&=& k + \sum_{j=1}^{\infty} j \cdot \frac{\mathbb{P}\left[Y=k+j\right]}{\mathbb{P}\left[Y\ge k\right]} \\
&=& k + \sum_{j=1}^{\infty} \frac{\mathbb{P}\left[Y\ge k+j\right]}{\mathbb{P}\left[Y\ge k\right]} .
\end{eqnarray*}
Now notice that the second statement of Lemma \ref{poilem:1} implies
\[ \frac{\mathbb{P}\left[Y\ge k+j\right]}{\mathbb{P}\left[Y\ge k\right]} = \prod_{\ell = 0}^{j-1} \frac{\mathbb{P}\left[Y\ge k+\ell+1\right]}{\mathbb{P}\left[Y\ge k+ \ell\right]} \le \frac{\mu^j}{\prod_{\ell =1}^{j}(k+\ell)} .\]
Since $\prod_{\ell =1}^{j}(k+\ell) \ge (k+1)^j$ we conclude
\[ \sum_{j=1}^{\infty} \frac{\mathbb{P}\left[Y\ge k+j\right]}{\mathbb{P}\left[Y\ge k\right]} \le \sum_{j=1}^{\infty} \left(\frac{\mu}{k+1}\right)^j = \frac{1}{1-\frac{\mu}{k+1}}-1 =\frac{\mu}{k+1-\mu} . \]
and the result follows.
\end{proof}
Recall the following, well-known, result regarding the median of a Poisson random variable (see \cite{Choi}). \\
\begin{lemma}\label{medpoi}
Let $Y\sim\text{Poi}(\mu)$. Then $\mathbb{P}\left[Y \ge \mu -\ln 2\right] \ge 1/2$.
\end{lemma}
We now proceed with the proof of the second statement of Theorem \ref{poibound}.
\begin{proof}[Proof of Theorem \ref{poibound} (Second statement)]
Let $Y\sim \text{Poi}(\mu)$ and fix a positive integer $k>\mu$. Let $\ell := \lfloor k-\mu \rfloor$ and notice that $\ell$ is the maximum positive integer $j$ such that $k-j\ge \mu$.
If we iterate Lemma \ref{poilem:1} $\ell +1$ times, we obtain
\[ \mathbb{P}\left[Y\ge k\right] = \mathbb{P}\left[Y\ge k-(\ell +1)\right]\cdot \prod_{j=0}^{\ell} \frac{\mu}{\mathbb{E}\left[Y|Y\ge k-j\right]} . \]
Since $\ell$ is a positive integer such that $k-(\ell +1) < \mu$, Lemma \ref{medpoi} implies that
\[ \mathbb{P}\left[Y\ge k-(\ell +1)\right] \ge \frac{1}{2} . \]
The first statement of Theorem \ref{poibound} now yields
\[ \prod_{j=0}^{\ell} \frac{\mu}{\mathbb{E}\left[Y|Y\ge k-j\right]} \ge \prod_{j=0}^{\ell} \frac{\mu}{k-j+\frac{\mu}{k-j+1-\mu}} = \prod_{j=0}^{\ell} \frac{\mu (k-j+1-\mu)}{(k-j)(k-j+1-\mu)+\mu}. \]
Since $k-j\ge \mu$, for $j=0,1,\ldots,\ell$, we conclude
\[ \frac{\mu (k-j+1-\mu)}{(k-j)(k-j+1-\mu) + \mu} \ge \frac{\mu}{(k-j)+ \mu} .\]
Putting all the above together, we conclude
\[ \mathbb{P}\left[Y\ge k\right] \ge \frac{1}{2}\cdot \prod_{j=0}^{\ell} \frac{\mu}{(k-j)+ \mu} \]
and the result follows.
\end{proof}
\section{Comparisons}\label{compare}
In this section we perform pictorial comparisons between the bound given by Theorem \ref{mainthm}
and the bound given by (\ref{Ashbound}). In Figure \ref{fig:compare} we fix the values of
$n$ and $k$ and plot the function $f(p):= \frac{p^{2\ell}}{2}\binom{n}{\ell +1}/\binom{k}{\ell+1} -
\frac{1}{\sqrt{8k(1-k/n)}}\cdot \text{exp}\left(-n D(k/n||p)\right)$,
for $p\in (0,k/n)$, where $\ell$ is as in Theorem \ref{mainthm}. For moderate values of $p$ our
bound is almost equal to the bound given by (\ref{Ashbound}) while for large values of $k$ and $p$
our bound is sharper. In most cases, the two bounds are rather close to each other.
\begin{figure}[htb!]
\subfloat[][]{\includegraphics[scale=0.22]{forty_six.pdf}}
\subfloat[][]{\includegraphics[scale=0.22]{forty_twentysix.pdf}}
\subfloat[][]{\includegraphics[scale=0.22]{twenty_four.pdf}}\\
\subfloat[][]{\includegraphics[scale=0.22]{twenty_twelve.pdf}}
\subfloat[][]{\includegraphics[scale=0.22]{hundred_fourtyfive.pdf}}
\subfloat[][]{\includegraphics[scale=0.22]{hundret_seventy.pdf}}
\caption{Comparison between the bound of Theorem \ref{mainthm} and the bound given by
Eq.(\ref{Ashbound})}\label{fig:compare}
\end{figure}
\textbf{Acknowledgements}.
I am grateful to Henk Don for fruitful discussions and valuable comments.
\begin{thebibliography}{99}
\bibitem{Alon} N. Alon, J. Spencer, \textit{The Probabilistic Method}, 3rd Edition, J. Wiley and Sons,
New York, 2008.
\bibitem{Ash} R.B. Ash, \textit{Information Theory}, Dover, New York, (1990).
\bibitem{Bahadur} R.R. Bahadur, \textit{Some approximations to the binomial distribution function}, The Annals of
Mathematical Statistics \textbf{31} (1) (1960) 43--54.
\bibitem{Carter} A. Carter, D. Pollard, \textit{Tusn\'ady's inequality revisited}, The Annals of Statistics \textbf{32}
(6) (2004) 2731--2741.
\bibitem{Choi} K.P. Choi, \textit{On the medians of Gamma distributions and an equation of Ramanujan}, Proc. Amer. Math. Soc. \textbf{121} (1) (1994) 245--251.
\bibitem{Diaconis} P. Diaconis, S. Zabell, \textit{Closed form summation for classical distributions: variations on a theme of
de Moivre}, Statistical Science \textbf{6} (3) (1991) 284--302.
\bibitem{Doerr} B. Doerr, \textit{An Elementary Analysis of the Probability that a Binomial Random Variable Exceeds its Expectation}, (preprint), arXiv:1712.00519.
\bibitem{Fellerzero} W. Feller, \textit{On the normal approximation to the binomial distribution}, The Annals of
Mathematical Statistics \textbf{16} (4) (1945) 319--329.
\bibitem{Feller} W. Feller, \textit{An introduction to probability theory and its applications}, Volume 1, Wiley New York,
(1957).
\bibitem{Glynn} P.W. Glynn, \textit{Upper bounds on Poisson tail probabilities}, Operations Research Letters
\textbf{6} (1) (1987) 9--14.
\bibitem{Greenberg} S. Greenberg, M. Mohri, \textit{Tight lower bound on the probability of a binomial exceeding its expectation}, Statistics and Probability Letters \textbf{86} (2014) 91--98.
\bibitem{Hoeffdingone} W. Hoeffding, \textit{Probability inequalities for sums of bounded random variables},
Journal of the American Statistical Association \textbf{58} (1963) 13--30.
\bibitem{Jerabek} E. Je\v{r}\'abek, \textit{Dual weak pigeonhole principle, Boolean complexity, and derandomization},
Annals of Pure and Applied Logic \textbf{129}, Issue 1-3 (2004) 1--37.
\bibitem{Kaas} R. Kaas, J.M. Burhman, \textit{Mean, median, mode in the binomial distribution}, Statistica Neerlandica
\textbf{34} (1) (1980) 13--18.
\bibitem{Landsman} Z. Landsman, E.A. Valdez, \textit{Tail Conditional Expectations for Elliptical
Distributions}, North American Actuarial Journal \textbf{7} (2003) 55--71.
\bibitem{Littlewood} J.E. Littlewood, \textit{On the probability in the tail of a binomial distribution}, Advances in Applied Probability \textbf{1} (1) (1969) 43--72.
\bibitem{Massart} P. Massart, \textit{Tusnady's Lemma, 24 years later}, Annales de l'Institut Henri Poincare (B) Probability and Statistics \textbf{38} (6) (2002) 991--1007.
\bibitem{McKay} B.D. McKay, \textit{On Littlewood's estimate for the binomial distribution}, Advances in Applied Probability \textbf{21} (2) (1989) 475--478.
\bibitem{Mitzenmacher} M. Mitzenmacher, E. Upfal, \textit{Probability and Computing: Randomized Algorithms and Probabilistic Analysis}, Cambridge University Press.
\bibitem{PelekisRamon} C. Pelekis, J. Ramon, \textit{A lower bound on the probability that a binomial random variable is exceeding its mean}, Statistics and Probability Letters \textbf{119} (2016) 305--309.
\bibitem{Rigollet} P. Rigollet, X. Tong, \textit{Neyman-Pearson classification, convexity and stochastic
constraints}, Journal of Machine Learning Research \textbf{12} (2011) 2831--2855.
\bibitem{Schmidt} J.P. Schmidt, A. Siegel, A. Srinivasan, \textit{Chernoff-Hoeffding bounds for applications with limited
independence}, SIAM Journal on Discrete Mathematics \textbf{8} (2) (1995) 223--250.
\bibitem{Slud} E.V. Slud, \textit{Distribution inequalities for the binomial law}, The Annals of Probability \textbf{5} (3) (1977) 404--421.
\bibitem{Teicher} H. Teicher, \textit{An inequality on Poisson probabilities}, Ann. Math. Statist.
\textbf{26} (1) (1955) 147--149.
\bibitem{Telgarsky} M. Telgarsky, \textit{Central binomial tail bounds}, (preprint), arXiv:0911.2077.
\bibitem{WangZhao} C.S. Wang, Z. Zhao, \textit{Conditional Value-at-Risk: Semiparametric estimation and inference},
Journal of Econometrics \textbf{195} (2016) 86--103.
\bibitem{Zubkov} A.M. Zubkov, A.A. Serov, \textit{A complete proof of universal inequalities for the distribution of the binomial law}, Theory Probab. Appl. \textbf{57} (3) (2013) 539--544.
\end{thebibliography}
\end{document}