It is "well-known" that there are basically only two types of memoryless distributions: The exponential distribution (in the continuous case) and the geometric distribution (in the discrete case). I wish to show this result. That is, given any real, positive random variable, I want to deduce that there are only two possibilities: Either the random variable has an exponential distribution, or it has a geometric distribution.

So let a positive random variable $T$ be given. Let us suppose that it satisfies the memorylessness property: $ P(T > t + s ) = P(T>t) P (T> s)$ for all $t,s \geq 0$. (Let's also assume that $P(T>t)\neq0$ for all $t\geq 0$, as this seems to be a natural assumption.)

If I assume that $P(T=0)=0$, then it's OK, because I am able to deduce from that that then $T$ should have an exponential distribution. So I assume that $P(T=0)>0$. The goal now is to prove that $T$ has a geometric distribution. Let's put $p:=P(T=0)$. Then $p \in (0,1]$.

I then argue as follows. Because $T$ satisfies memorylessness, we have that $T-t \; | \; T>t$ has the same distribution as $T$, for every $t\geq 0$.
But then I get $p=P(T=0) = P(T-1 = 0 \; | \; T>1) = P(T =1 \; | \; T>1) = 0$, contradiction. Yet this doesn't seem to make much sense.

Where did I make a mistake in my reasoning?

New contributor
herbhofsterd is a new contributor to this site. Take care in asking for clarification, commenting, and answering. Check out our Code of Conduct.
  • $\begingroup$ From glancing at Wikipedia's definition of discrete memorylessness it seems that you need to be careful with which inequalities are strict ($>$) and which are non strict ($\ge$). $\endgroup$
    – angryavian
    May 14 at 1:29
  • $\begingroup$ Is there no "general" notion of 'memorylessness' then? $\endgroup$ May 14 at 2:43
  • $\begingroup$ I think I solved the issue. The point is, according to 'my' definition of memorylessness (where the condition holds for all $t,s\geq0$), the geometric distribution is not a memoryless random variable. The reason is that if $T$ is geometrically distributed with parameter $p>0$, then $P(T>t+s)=P(T>t)P(T>s)$ holds for $s,t\geq0$ iff the sum of the fractional parts of $s$ and $t$ is strictly less than $1$. (In particular, it holds for integers.) So we must 'relax' the definition of memorylessness for discrete rvs. Following 'my' definition of memorylessness, the exponential is indeed unique. $\endgroup$ May 14 at 4:31


Your Answer

herbhofsterd is a new contributor. Be nice, and check out our Code of Conduct.

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.