Digit variance and Dedekind sums (Q1363093)

Given positive integers \(z,n,b\) with \(1\leq z\leq n-1\), \((z,n)=1\), \(n\geq 2\), \(b\geq 2\), \((b,n)=1\), the expansion of the rational \(z/n\) in base \(b\) has the form \(z/n= \sum^\infty_{j =1} c_jb^{-j}\), where each coefficient \(c_j\) is one of the digits \(0,1, \dots, b-1\), and \(c_j \neq b-1\) for infinitely many \(j\). The sequence of digits \((c_1, \dots, c_t)\) is a period with period length \(t\) equal to the order of \(b\bmod n\), and there is no shorter one. The author views this period as a sort of random sequence that takes values in \(\{0,1, \dots, b-1\}\). The mean value \(m=t^{-1} \sum^t_{j=1} c_j\) is easily shown to be \((b-1)/2\), but the variance \(\sigma^2 =t^{-1} \sum^t_{j=1} (c_j-m)^2\) is not easy to determine. In the special case where \(n\) is an odd prime \(p\) and \(b\) is a primitive root mod \(p\) the author proves that \[ \sigma^2 ={b^2-1 \over 12}- {S (b,p) \over p-1}, \] where \(S(p,b)= (b-1)(b-2)/6- 2bs(p,b)\) and \(s(p,b)\) is the classical Dedekind sum, \[ s(p,b)= \sum^{b-1}_{k=1} {k\over b} \left({kp \over b}- \left[{kp \over b} \right]- {1\over 2} \right). \] The reciprocity law for Dedekind sums is used to establish this result and others concerning upper and lower bounds for \(\sigma^2\).