Expected Value of an Exploding Die

The rule is simple to state: roll the die, and if it lands on its highest face, roll again and add — repeating for as long as the highest face keeps coming up. The result has no upper bound, which is the appeal. The average is nonetheless a small finite number, and working out why is a genuinely satisfying piece of arithmetic.

Derive it in one line with a self-referential argument. Let $E$ be the answer. Every roll contributes its own face value, averaging $(d+1)/2$, and with probability $1/d$ it also triggers a fresh roll whose average is — by definition — $E$ again, because the new roll knows nothing about being a continuation. So $E = (d+1)/2 + E/d$. Solve: $E(1 - 1/d) = (d+1)/2$, giving $E = d(d+1)/(2(d-1))$. For a six-sided die that is $42/10 = 4.2$, against 3.5 without the rule — the explosions are worth exactly 0.7.

The same answer falls out of summing a geometric series, and it is worth seeing that way too: the first roll averages 3.5, the chance of a second is 1/6 and it contributes 3.5/6, the third contributes 3.5/36, and so on. The total is $3.5 \times (1 + 1/6 + 1/36 + \cdots) = 3.5 \times 6/5 = 4.2$. The series converges because each further explosion is $d$ times rarer than the last, and it is exactly this convergence that keeps an unbounded outcome from having an unbounded average.

The one case where it fails is instructive. A two-faced die explodes half the time and the series still converges, to 3. A one-faced die explodes every time, the series is $1 + 1 + 1 + \cdots$, and the expected value is genuinely infinite — the only input this formula honestly cannot answer.

Two things the average will not tell you. The distribution is highly skewed: most rolls are ordinary, a minority are large, and the average sits above the median, so "4.2" describes a shape that hardly ever produces 4.2. And the number of explosions in a single roll averages $1/(d-1)$, so on a six-sided die roughly one roll in five explodes at all and one in thirty explodes twice.

This is the page where the gambler's fallacy is most tempting, so it gets said plainly. Explosions feel streaky and they are not. Each roll is independent of every roll before it. A die that has just exploded three times running is exactly as likely to explode again as one that has never exploded, and a die that has gone a hundred rolls without exploding is owed nothing whatsoever. There is no mechanism by which a die could know, and no bookkeeping anywhere that records what it has done. What people mistake for a streak is the ordinary clustering that independent trials produce all the time — genuinely random sequences look far lumpier than most people expect, and a sequence with no clusters in it would be the suspicious one.