9
$\begingroup$

Is it possible to prove that the sum of two independent r.v.'s $X$ and $Y$ with convex support cannot be uniformly distributed on an interval $[a,b]$, with $a < b$?

(Let us rule out the trivial case where $X$ is degenerate and $Y$ is uniform.) Note that $X$ and $Y$ need not be identically distributed.

I have a strong intuition supporting a negative answer because commutativity gets $X+Y$ to "pile up" away from the tails, but I cannot find a way to pin this down to an impossibility proof.

Related questions on Math.SE:

1) Can the sum of two independent identical random variables be uniform? and Can the sum of two i.i.d. random variables be uniformly distributed? answer negatively under the additional assumption that $X$ and $Y$ are identically distributed.

2) in Can sum of two random variables be uniformly distributed @GEdgar answers positively under the assumption that the two independent r.v.'s $X$ and $Y$ have non-convex support.

$\endgroup$
  • $\begingroup$ Are you asking for a general proof or a specific example? $\endgroup$ – copper.hat Mar 24 '17 at 23:56
  • $\begingroup$ An example would be nice, and a general proof would be even better. $\endgroup$ – mlc Mar 25 '17 at 6:11
5
+50
$\begingroup$

Your intuition is correct : the answer is negative. Here is a proof.

Suppose, by contradiction, that $X$ and $Y$ are independent, non-degenerate variables with convex support, say $[c,d]$ and $[e,f]$, and $X+Y$ is uniform on $[c+e,d+f]$. Replacing $(X,Y)$ with $(X-c,Y-e)$, we may assume without loss that $c=e=0$. Interchanging $X$ and $Y$ if necessary, we may further assume that $d\leq f$. If $d<f$ we can also replace $(X,Y)$ with $((2f)X+(3f+d)Y,(4f)X+(3f-d)Y)$ (remember that a linear bijective transformation preserves rv independence), so we may further assume that $d= f$. Finally, replacing $(X,Y)$ with $(\frac{X}{d},\frac{Y}{d})$ we may assume that $d=1$.

At this point, we then have $X,Y$ independent with support exactly equal to $[0,1]$, and $X+Y$ distributed uniformly on $[0,2]$. Now, define the following events :

$$ \begin{array}{cccl} X_1= \lbrace X\in[0,\frac{1}{3}] \rbrace, & X_2= \lbrace X\in(\frac{1}{3},\frac{2}{3}] \rbrace, & X_3= \lbrace X\in(\frac{2}{3},1] \rbrace, \\ Y_1= \lbrace Y\in[0,\frac{1}{3}] \rbrace, & Y_2= \lbrace Y\in(\frac{1}{3},\frac{2}{3}] \rbrace, & Y_3= \lbrace Y\in(\frac{2}{3},1] \rbrace, & \end{array} $$

$$ P_k=\cup_{i+j=k}X_i\times Y_j, Q_l=\cup_{k=2}^l P_k, L(x)=\lbrace X+Y \leq x \rbrace $$

We then have the inclusions :

$$ L\big(\frac{l-1}{3}\big) \subseteq Q_l \subseteq L\big(\frac{l}{3}\big) $$

Putting $q_l=P(Q_l)$, we deduce

$$ \frac{l-1}{6}\leq q_l \leq \frac{l}{6} \tag{1} $$

Next, put $x_k=P(X_k), y_k=P(Y_k), p_k=P(P_k)$. It follows from (1) that the following number is nonnegative :

$$ s=(1-x_1)\big(q_2-\frac{1}{6}\big)+x_1\big(\frac{2}{3}-q_4\big)+x_1x_2y_2 \tag{2} $$

Clearly $q_2=x_1y_1$. And using $x_3=1-(x_1+x_2)$ and $y_3=1-(y_1+y_2)$, we find that

$$ q_4 =x_1y_1+x_2y_1+x_1y_2+x_1y_3+x_2y_2+x_3y_1=x_1+y_1+x_2y_2-x_1y_1 \tag{3} $$

Reinjecting this back into (2), we obtain

$$ s=\bigg(x_1-\frac{1}{2}\bigg)\bigg(\frac{1}{3}-x_1\bigg) \tag{4} $$

It follows that $x_1\in[\frac{1}{3},\frac{1}{2}]$. Using natural symmetries of the problem, it follows that $x_3,y_1$ and $y_3$ are also in $[\frac{1}{3},\frac{1}{2}]$.

Next, consider

$$ \begin{array}{clc} t_1 &=& (x_1x_3-x_2^2)\big(\frac{5}{6}-q_5\big)+x_3x_2\big(\frac{2}{3}-q_4\big)+x_3^2\big(q_3-\frac{1}{3}\big) \\ t_2 &=& (x_1x_3-x_2^2)\big(q_2-\frac{1}{6}\big)+x_1x_2\big(q_3-\frac{1}{3}\big)+ x_1^2\big(\frac{2}{3}-q_4\big) \\ t &=& t_1+t_2 \end{array} $$

Note that since $x_1x_3 \geq \frac{1}{9} \geq x_2^2$, the numbers $t_1,t_2$ and $t$ are all nonnegative. Simplifying $t$ as we did for $s$, we obtain :

$$ t=-\frac{x_2}{4}(1+x_2)(\frac{1}{3}-x_2)-\frac{3(x_3-x_1)^2}{4}\bigg(\frac{4}{9}-x_2\bigg) \tag{5} $$

In (5) above we have a nonnegative LHS and a nonpositive RHS. Note also that $x_2>0$ because the support of $X$ is $[0,1]$. This forces $x_2=\frac{1}{3}$ and $(x_3-x_1)^2=0$, so that $x_1=x_2=x_3=\frac{1}{3}$, and similarly $y_1=y_2=y_3=\frac{1}{3}$. But then $q_2=\frac{1}{9}$ contradicts (1). This finishes the proof.

UPDATE: Since (5) was questioned in the comments, I append a computer verification of it.

                  GP/PARI CALCULATOR Version 2.7.4 (released)
          i386 running darwin (x86-64/GMP-6.0.0 kernel) 64-bit version
compiled: Sep 27 2015, Apple LLVM version 6.0 (clang-600.0.57) (based on LLVM 3.5svn)
                            threading engine: single
                 (readline v6.3 enabled, extended help enabled)

                     Copyright (C) 2000-2015 The PARI Group

PARI/GP is free software, covered by the GNU General Public License, and comes 
WITHOUT ANY WARRANTY WHATSOEVER.

Type ? for help, \q to quit.
Type ?12 for how to get moral (and possibly technical) support.

parisize = 8000000, primelimit = 500000
? x3=1-(x1+x2)
%1 = -x1 + (-x2 + 1)
? y3=1-(y1+y2)
%2 = -y1 + (-y2 + 1)
? 
? x(k)=if((1<=k)&&(k<=3),eval(Str("x",k)),0)
%3 = (k)->if((1<=k)&&(k<=3),eval(Str("x",k)),0)
? y(k)=if((1<=k)&&(k<=3),eval(Str("y",k)),0)
%4 = (k)->if((1<=k)&&(k<=3),eval(Str("y",k)),0)
? 
? p(k)=sum(j=1,k,x(j)*y(k-j))
%5 = (k)->sum(j=1,k,x(j)*y(k-j))
? 
? q(l)=sum(k=2,l,p(k))
%6 = (l)->sum(k=2,l,p(k))
? 
? s_term=(1-x1)*(q(2)-(1/6))+x1*((2/3)-q(4))+x1*x2*y2
%7 = -x1^2 + 5/6*x1 - 1/6
? 
? check1=s_term-(x1-(1/3))*((1/2)-x1)
%8 = 0
? 
? t1_term=(x1*x3-(x2^2))*((5/6)-q(5))+x3*x2*((2/3)-q(4))+(x3^2)*(q(3)-(1/3))
%9 = x1^3 + (3*x2 - 13/6)*x1^2 + (3*x2^2 - 25/6*x2 + 3/2)*x1 + (x2^3 - 11/6*x2^2 + 4/3*x2 - 1/3)
? t2_term=(x1*x3-(x2^2))*(q(2)-(1/6))+x1*x2*(q(3)-(1/3))+(x1^2)*((2/3)-q(4))
%10 = -x1^3 + 5/6*x1^2 + (-1/6*x2 - 1/6)*x1 + 1/6*x2^2
? t_term=t1_term+t2_term
%11 = (3*x2 - 4/3)*x1^2 + (3*x2^2 - 13/3*x2 + 4/3)*x1 + (x2^3 - 5/3*x2^2 + 4/3*x2 - 1/3)
? 
? easy_part=(x2/4)*(x2+1)*((1/3)-x2)
%12 = -1/4*x2^3 - 1/6*x2^2 + 1/12*x2
? hard_part=(3/4)*((x3-x1)^2)*((4/9)-x2)
%13 = (-3*x2 + 4/3)*x1^2 + (-3*x2^2 + 13/3*x2 - 4/3)*x1 + (-3/4*x2^3 + 11/6*x2^2 - 17/12*x2 + 1/3)
? 
? check2=(t_term+(easy_part+hard_part))
%14 = 0
$\endgroup$
  • $\begingroup$ I have been unable to verify (5). I have been doing it by hand, and got a slightly different result. Everything else works out. (I fixed three minor typos.) This approach is laborious but compelling. $\endgroup$ – mlc Apr 2 '17 at 20:24
  • $\begingroup$ @mlc Thank you for your feedback. I have added a verification of (5) by computer, and I invite you to double-check it, for example at tutorialspoint.com/execute_pari_online.php $\endgroup$ – Ewan Delanoy Apr 3 '17 at 7:01
  • $\begingroup$ I did not mean to question the computation, simply to plainly state that I could not (yet) vouch for the correctness of the proof. I have accepted your answer because it is most useful, but this is not equivalent to asserting that the proof is complete. And now it is :-) $\endgroup$ – mlc Apr 3 '17 at 8:45

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.