Let $p_n$ denote the $n$th prime number and $g_n=p_{n+1}-p_n$ the $n$th prime number gap. This is to ask for which values of $\alpha$ the series $S_\alpha$ converges or diverges, where $$S_\alpha=\sum_n\left(\frac{g_n}{p_n}\right)^\alpha.$$


  • The obvious lower bound $g_n\geqslant1$ shows that $S_\alpha$ diverges for every $\alpha\leqslant1$.
  • It is known (see the WP page) that $g_n\lt (p_n)^\theta$ for every large enough $n$, for every $\theta\gt\frac34$, hence $S_\alpha$ converges for every $\alpha\gt4$.
  • Various unproven results, such as Cramér's conjecture that $g_n=O\left((\log p_n)^2\right)$, would imply that $S_\alpha$ is finite if and only if $\alpha\gt1$.
  • An answer for $\alpha=2$ would solve (and in fact would be equivalent to a solution of) this other question.

Edit: @GregMartin's answer below yields naturally the more general result that the series $$S_{\alpha,\beta}=\sum_n\frac{g_n^\beta}{p_n^\alpha}$$ converges for every $$\alpha\gt\max\{1,\tfrac5{18}\beta+\tfrac{13}{18}\}.$$ For example, two convergent series are $$\sum_n\frac{g_n^2}{p_n^{4/3}},\qquad\sum_n\frac{g_n^4}{p_n^2}.$$ Actually, an asymptotic control $$ \sum_{n\colon p_n \le x} g_n^2 \leqslant x^{1+\gamma}, $$ for some $\gamma$ in $(0,1)$ (Heath-Brown's result used by @GregMartin being the case of every $\gamma\gt\frac5{18}$) would yield the convergence of $S_{\alpha,\beta}$ for every $(\alpha,\beta)$ such that $$\alpha-1\gt\gamma\cdot(\beta-1)_+.$$

  • $\begingroup$ what happens if you just replace $g_n$ by $\log p_n?$ $\endgroup$
    – Will Jagy
    Nov 28 '14 at 19:43
  • $\begingroup$ right, simple estimates in Rosser and Schoenfeld (1962) for the substitute series give convergence for your $\alpha > 1.$ So, compare partial sums for the two series up to $p_n,$ could be possible to resolve this. projecteuclid.org/euclid.ijm/1255631807 not suggesting it's easy $\endgroup$
    – Will Jagy
    Nov 28 '14 at 19:53
  • $\begingroup$ @WillJagy "what happens if you just replace g_n by log(p_n)" Then the modified series converges for every α>1, but how is this related to my question? // "simple estimates in Rosser and Schoenfeld (1962) for the substitute series give convergence for your α>1. " Really? Which ones? How? Please be much more specific, I am afraid you lost me. $\endgroup$
    – Did
    Nov 28 '14 at 21:31
  • $\begingroup$ I just meant (3.12) and (3.13), for $n \geq 6$ we get $n \log n < p_n < n (\log n + \log \log n).$ As far as relationship, the logarithm is the expected size of the gap, we can not expect to compare series uniformly, but perhaps something based on averages is worth exploring. I don't have much more, this was never my area. If I get something that works i will leave an answer. $\endgroup$
    – Will Jagy
    Nov 28 '14 at 23:11
  • 5
    $\begingroup$ @XuefengMeng Please avoid unnecessary edits to questions of others. The one you suggested, which was, rather mysteriously, approved by two users, was superfetatory since the case $\alpha=1$ is already addressed in the question. $\endgroup$
    – Did
    Jan 9 '15 at 10:40

The series does converge for every $\alpha>1$.

While our knowledge of individual prime gaps is still somewhat lacking, our knowledge of the gaps on average is rather better. Suppose we have any estimate of the form $$ \sum_{n\colon p_n \le x} g_n^2 \ll x^{2-\delta} \tag{$\ast$} $$ for some $\delta>0$. Then for any $\alpha>1$ we can argue, using Hölder's inequality: \begin{align*} \sum_{n\colon x/2<p_n\le x} \frac{g_n^\alpha}{p_n^\alpha} &\ll \frac1{x^\alpha} \sum_{n\colon x/2<p_n\le x} g_n^\alpha \\ &\le \frac1{x^\alpha} \bigg( \sum_{n\colon x/2<p_n\le x} g_n^2 \bigg)^{\alpha-1} \bigg( \sum_{n\colon x/2<p_n\le x} g_n \bigg)^{2-\alpha} \\ &\ll \frac1{x^\alpha} \big( x^{2-\delta} \big)^{\alpha-1} (x)^{2-\alpha} = x^{-\delta(\alpha-1)}. \end{align*} (We use the fact that the sum of the gaps themselves, of all primes between $a$ and $b$, is just $b-a$ up to one prime gap on each end.) Therefore $$ \sum_{n=1}^\infty \frac{g_n^\alpha}{p_n^\alpha} = \sum_{k=1}^\infty \sum_{n\colon 2^{k-1}<p_n\le 2^k} \frac{g_n^\alpha}{p_n^\alpha} \ll \sum_{k=1}^\infty 2^{-k\delta(\alpha-1)} \ll 1, $$ and so the series converges.

Fortunately, we do know $(\ast)$; indeed, Heath-Brown has established the rather strong version $$ \sum_{n\colon p_n \le x} g_n^2 \ll x^{23/18+\varepsilon}. $$

  • $\begingroup$ +1. Very nice, thanks a lot. (I guess that, like many number theorists, you use $x_n\ll y_n$ in the Vinogradov sense, that is, to mean that, in Landau's notation, $x_n\in O(y_n)$.) $\endgroup$
    – Did
    Nov 29 '14 at 12:55
  • $\begingroup$ that's right, about $\ll$ $\endgroup$ Nov 29 '14 at 20:40
  • $\begingroup$ Thanks to (someone) for the bounty! $\endgroup$ Dec 2 '14 at 21:14
  • 1
    $\begingroup$ Here's an alternative argument in proving $(*)$. Let $\theta$ be the exponent in the estimate $g_n \ll p_n^{\theta}$. Then $$\sum_{p_n\leq x}g_n^2 \ll \sum_{p_n\leq x} p_n^{\theta} g_n \ll x^{\theta} \sum_{p_n\leq x} g_n \ll x^{\theta+1} .$$ Any $\theta$ with $0<\theta<1$ will suffice for the purpose of proving $(*)$. $\endgroup$ Jun 17 '17 at 19:10

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.