The following question was asked in a comment at Almost sure convergence vs convergence of probability density functions :

Suppose that $(X_n)$ is a sequence of random variables (r.v.'s) converging almost surely (a.s.) to a r.v. $X$. Suppose that $X_n$ and $X$ have pdf's $p_n$ and $p$, respectively. Does it then necessarily follow that $\int fp_n\to\int fp$ for all $f\in L^\infty[0,1]$ (and $n\to\infty$)?

Here this question will be answered.

  • $\begingroup$ Remark: Convergence does hold if f is Riemann integrable, i.e., continuous a.e. $\endgroup$ Feb 9, 2021 at 16:28

1 Answer 1


The answer is no. Indeed, let $C$ be any countable subset of $(0,1)$ which is dense in $[0,1]$. Write $C=\{c_1,c_2,\dots\}$, where the $c_n$'s are pairwise distinct. Let $$U:=(0,1)\cap\bigcup_{n=1}^\infty(c_n+4^{-n}B)\quad\text{and}\quad F:=[0,1]\setminus U,$$ where $B:=(-1,1)$. Then $U\supset S$ and hence $U$ is dense in $[0,1]$. The Lebesgue measure $|U|$ of $U$ is $\le2/3$ and hence $|F|\ge1/3>0$. Let now $X$ be any r.v. uniformly distributed on $F$, so that the pdf of $X$ is $$p=1_F/|F|.$$

Also, the set $U$ is open and therefore $$U=\bigcup_{k=1}^\infty I_k,$$ where the $I_k$'s are some pairwise disjoint open subintervals of $(0,1)$. Let $$U_n:=\bigcup_{k=1}^n I_k$$ and then let $E_n$ denote the set of all endpoints of the intervals $I_1,\dots,I_n$. For any $x\in[0,1]$ and any $A\subseteq[0,1]$, let $d(x,A):=\inf\{|y-x|\colon y\in A\}$.

Since $U$ is dense in $[0,1]$, for all $x\in F$ we have $d(x,U)=0$ and hence $d(x,U_n)\to0$, so that $d(x,E_n)=d(x,U_n)\to0$. Letting now $$Y_n:=y_n(X),$$ where $$y_n(x):=\min\{y\in E_n\colon |x-y|=d(x,E_n)\},$$ we see that $|X-Y_n|=d(X,E_n)\to 0$ a.s. and hence $Y_n\to X$ a.s.

Finally, define $X_n$ as follows. For any $y\in E_n$ and any $t\in(0,1)$, let

(i) $x_n(y,t):=y+|I_k|t/n$ if $y$ is the left endpoint of the interval $I_k$ for some $k\in[n]:=\{1,\dots,n\}$, and

(ii) $x_n(y,t):=y-|I_k|t/n$ if $y$ is the right endpoint of the interval $I_k$ for some $k\in[n]$ but $y$ is not the left endpoint of the interval $I_l$ for any $l\in[n]$.

Then let $X_n:=x_n(Y_n,T)$, where $T$ is a r.v. uniformly distributed on $(0,1)$ and independent of $X$ (and hence of $Y_n$). Then $X_n$ is an absolutely continuous r.v. with all values in $U_n\subseteq U$; as before, let $p_n$ denote the pdf of $X_n$. Also, $|X_n-Y_n|\le1/n$. Therefore and because $Y_n\to X$ a.s., we have $X_n\to X$ a.s. However, $1_U\in L^\infty[0,1]$ but $$\int 1_U\,p_n=P(X_n\in U)=1\not\to0=P(X\in U)=\int 1_U\,p.$$

  • $\begingroup$ Once $U$ and $X$ are defined, you can just define the conditional distribution of $X_n$ given $X$ to be uniformly distributed on the open set $(X-1/n,X+1/n) \cap U$. That will shorten the proof. $\endgroup$ Feb 9, 2021 at 3:50
  • $\begingroup$ @YuvalPeres : Thank you for your comment. $\endgroup$ Feb 9, 2021 at 16:20
  • $\begingroup$ Thank you very much, @IosifPinelis! I guess this shows that the convergence you gave is the strongest possible. $\endgroup$
    – Nate River
    Feb 14, 2021 at 12:35
  • $\begingroup$ @NateRiver : I am glad you found this of help. So, to have a closure, are you satisfied with this answer? $\endgroup$ Feb 14, 2021 at 16:21
  • $\begingroup$ Yes, is there anything I should do to mark this? I think I have accepted your answer in my original post. $\endgroup$
    – Nate River
    Feb 19, 2021 at 23:46

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.