How unexpected is the Prime Number Theorem? Let p1 , · · · , pn be the first n primes. Let x be chosen randomly from among the integers greater than pn . The probability that x is divisible by pi is 1/pi , so the probability that x is divisible by none of p1 , · · · , pn is 1 1 ··· 1 − . 1− p1 pn If we make the unjustified assumption that the property of being divisible by none of p1 , · · · , pn is held randomly by numbers greater than pn with probability (1 − 1/p1 ) · · · (1 − 1/pn ), then the probability that pn+1 − pn = r is r−1 1 1 1 1 ··· 1 − 1− ··· 1 − . 1− 1− p1 pn p1 pn So the expected value of pn+1 − pn is r−1 ∞ X 1 1 1 1 r 1− 1− ··· 1 − 1− ··· 1 − p1 pn p1 pn r=1 1 = 1/ 1 − p1 ··· 1 1− pn . Accordingly, we define the series of problimes by q1 = 2, qn+1 1 = qn + 1/ 1 − q1 ··· 1 1− qn . The interesting question that arises is, what is the asymptotic behaviour of the qn ? In particular, is it true that qn ∼ n loge n? Typeset by AMS-TEX 1 2 If this is true, then perhaps the prime number theorem is a little less surprising, since it is a cosequence of the prime number theorem that pn ∼ n loge n. I have been able to prove that (1) qn /n is unbounded, (2) qn /n1+ǫ → 0 as n → ∞ for any fixed ǫ > 0, but I have not been able to prove that qn /n is eventually monotonic increasing. It is clear that the problimes soon become non-integral. It may be more attractive to study the various integral sequences defined by q1 = 2, qn+1 1 1 = qn + F 1/ 1 − ··· 1 − , q1 qn where F (x) = [x], the greatest integer not greater than x, F (x) =< x >, the closest integer to x, or F (x) = {x}, the least integer not less than x. The first few terms of each of these sequences are given in the following table, together with the primes for comparison: n pn qn∗ qn , [] qn , <> qn , {} 1 2 2.0 2 2 2 2 3 4.0 4 4 4 3 5 6.7 6 7 7 4 7 9.8 9 10 11 5 11 13.3 12 13 15 6 13 17.1 15 17 19 7 17 21.1 19 21 23 8 19 25.3 23 25 28 9 23 29.7 27 29 33 10 29 34.2 31 34 38 n 12 pn 37 qn 43.7 qn , [] 40 qn , <> 44 qn , {} 48 13 41 48.6 45 49 53 14 43 53.6 50 54 58 15 47 58.7 55 59 63 16 53 63.9 60 64 68 17 59 69.2 65 69 73 18 61 74.6 70 74 79 19 67 80.0 75 79 85 20 71 85.5 80 84 91 21 73 91.1 86 90 97 11 31 38.9 35 39 43 3 ∗ to 1 decimal place P. Erdös (written communication) believes that he can prove by Tauberian arguments that (qn+1 − qn )/ loge n → 1, and hence that qn /n loge n → 1. I am indebted to the referee and to Richard K. Guy for their helpful comments and suggestions.