If $\lim P(A_n \le B_n + C_n) = 1$ and $C_n = o(1)$, then $\limsup (P(A_n > u) - P(B_n > u)) \le 0, \forall u \in \mathbb{R}^+$

45 Views Asked by At

The problem statement is in the title. $A_n, B_n$ are random sequences, $C_n$ is deterministic.

The statement itself is complicated because of various notations. But I think the proof relies on this basic question. I'm not familiar with the proof technique when there is a $o(1)$ in side the probability statement and working with $\limsup$.

I know that in the event ${A_n \le B_n + C_n}$, since $A_n > u$ implies $B_n + C_n > u$, the event ${A_n > u}$ is a subset of ${B_n + C_n > u}$. However I feel wrong to write $P(A_n > u) - P(B_n + C_n > u) \le 0$, because it replies on the event ${A_n \le B_n + C_n}$, which has probability goes to 1 as $n\to \infty$. I can't write $P(A_n > u) - P(B_n + C_n > u) \le 0$ with probability converges to 1 after all. That is my starting point, and sooner I'm not sure also how to pull the $C_n$ out and utilize $C_n = o(1)$. If I can manage to this, I still need some help on how to get in touch with the limsup in the target statement. Thank you for all the advice.

1

There are 1 best solutions below

0
On BEST ANSWER

In general this is false. Take the case of $A_n = B_n + c_n$ w.p. $1,$ where $c_n > 0$. Suppose that $B_n$ is uniform over $(-c_n/2, c_n/2)$. Then we have $$ P(A_n > 0) - P(B_n>0) = P(B_n > -c_n) - P(B_n > 0) = 1 - 1/2 = 1/2,$$ so the limit in question is positive at $u = 0$. This holds independently of the behaviour of $c_n,$ and in particular also works when $c_n \to 0$ with $n$.


Notice though that when $c_n \to 0,$ the above law on $B_n$ degenerates with $n$ in the sense that $B_n$ converges to a point mass at $0$. This style of counterexample continues to work if $B_n$ goes to some sort of discrete set almost surely. So it is possible that the bound does hold if we demand some smoothness from limiting behaviour of the $B_n$.

Indeed, suppose that $B_n$ satisfies that for any sequence $\delta_n > 0$ such that $\delta_n \to 0,$ and any $u$, it holds that $$ \lim {P}(B_n \in (u-\delta_n, u]) = 0. \tag{*}$$ Then I claim that the bound does hold. To see this, first observe that $$ P(A_n > u) = P(A_n > B_n + c_n, A_n > u) + P(A_n \le B_n + c_n, A_n > u),\\ P(B_n>u) = P(A_n > B_n + c_n, B_n>u) + P(A_n \le B_n + c_n, B_n > u).$$

Now, first consider $$ \zeta_n := P(A_n > B_n + c_n, A_n > u) - P(A_n > B_n + c_n, B_n > u).$$ Clearly we have $$ \zeta_n \le P(A_n > B_n + c_n), \zeta_n \ge -P(A_n > B_n +c _n),$$ and so $\zeta_n \to 0.$ So, we must have that $$\limsup P(A_n > u) - P(B_n > u)\\ = \limsup P( A_n \le B_n + c_n, A_n > u) - P(A_n \le B_n + c_n, B_n > u). $$ To make notation a little more succinct, set $D_n = \min(A_n, B_n + c_n).$ It suffices to analyse $P(D_n > u) - P(B_n > u)$ (why?).

Now, we can break this as \begin{align} &P(D_n > u) - P(B_n > u)\\ &=\, P(D_n > B_n, D_n > u) + P(D_n \le B_n, D_n > u) \\ &\quad - P(D_n > B_n, B_n > u) - P(D_n \le B_n, B_n > u).\end{align} But $D_n\le B_n, D_n > u \implies B_n > u,$ and so we find that for every $n,$ $$ \xi_n := P(D_n \le B_n, D_n > u) - P(D_n \le B_n, B_n > u) \le 0.$$

Further, observe that \begin{align} P(D_n > B_n, D_n > u) - P(D_n > B_n, B_n > u) &= P(D_n > B_n, D_n> u, B_n \le u) \\ &\le P(B_n \le u, B_n + c_n > u) \\ &\le P(B_n \in (u - |c_n|, u]) ,\end{align} which goes to $0$ by our condition $(*)$.

Putting everything together, we can write the bound $$ P(A_n > u) - P(B_n > u) \le P(A_n > B_n + c_n) + \xi_n + P(B_n \in (u-|c_n|, u]).$$ Here the first and last term go to $0$, and the second term is nonpositive, so the $\limsup$ of this term can be at most $0$.


Finally, just to make the point explicitly, note that the condition $(*)$ is also necessary at this level of generality, in the sense that if there is some $u_0$ and $\{\delta_n\}$ such that the limit in $(*)$ is nonzero, then a counterexample exists: taking $A_n = B_n + \delta_n,$ we have that \begin{align} P(A_n > u_0) - P(B_n > u_0) &= P(B_n > u_0 - \delta_n) -P(B_n > u_0) \\ &= P(B_n \in (u_0 - \delta_n, u_0]),\end{align} which converges to something positive. Of course, it would also be interesting to figure out what types of $A_n$ suffice for the result if $B_n$ has this property, but I'm not sure that has a very clean answer.