= Replacement breakdown point
{title2=$\varepsilon^*$}
For a fixed sample $y$ of size $n$, let $b_n(m;y)$ be its <finite-sample maximum bias> after at most $m$ arbitrary replacements, with $m\in\{0,\ldots,n\}$. The largest-bounded convention used here is
$$
\varepsilon_{\mathrm{bounded}}^*(T_n;y)=\frac1n\max\{m:b_n(m;y)<\infty\}.
$$
When breakdown is possible, the first-unbounded convention is
$$
\varepsilon_{\mathrm{unbounded}}^*(T_n;y)=\frac1n\min\{m:b_n(m;y)=\infty\}.
$$
They differ by $1/n$. For five observations, the <sample median> remains bounded under any two replacements, while three replacements can make it unbounded. These conventions therefore give $2/5$ and $3/5$, respectively. Their common large-sample limit does not remove the need to specify the finite-sample convention. The <finite-sample maximum bias> article displays the first-unbounded form explicitly.
Back to article page
