Skip to main content
1 of 2
Iosif Pinelis
  • 127.9k
  • 8
  • 107
  • 229

$\newcommand{\de}{\delta}$You have the bounds \begin{equation} P(|X_i-m_i|\ge \de_i m_i]\le2e^{-\de_i^2 m_i/3} \end{equation} for $i=1,2$ and $\de_i>0$, where \begin{equation} m_i:=EX_i. \end{equation} Clearly, these bounds are useful only if $m_i>0$ for $i=1,2$, which will be henceforth assumed. Assume also that $0<\de_i<1$ for $i=1,2$.

Then on the event \begin{equation} \begin{aligned} A&:=\{|X_1-m_1|<\de_1 m_1,|X_1-m_1|<\de_2 m_2\} \\ &=\{m_1(1-\de_1)<X_1<m_1(1+\de_1),m_2(1-\de_2)<X_2<m_2(1+\de_2)\} \end{aligned} \end{equation} we have $X_1>0$ and $X_2>0$, which implies that \begin{equation} Y = \frac{X_1-X_2}{X_1+X_2} \end{equation} is increasing in $X_1$ and decreasing in $X_2$, so that the event
\begin{equation} B:=\Big\{\frac{m_1(1-\de_1)-m_2(1+\de_2)}{m_1(1+\de_1)+m_2(1-\de_2)} <Y<\frac{m_1(1+\de_1)-m_2(1-\de_2)}{m_1(1-\de_1)+m_2(1+\de_2)}\Big\} \end{equation} occurs. That is, $A\subseteq B$.

So, by the union bound, we get the concentration result: \begin{equation} 1-P(B)[\le1-P(A)]\le 2e^{-\de_1^2 m_1/3}+2e^{-\de_2^2 m_2/3}. \end{equation} One can now (quasi-)optimize the choices of $\de_1$ and $\de_2$, depending on one's objectives.

Iosif Pinelis
  • 127.9k
  • 8
  • 107
  • 229