# Upper bounds for the sum of primes <= n

Let $s(n)$ denote the sum of primes less than or equal to n. Clearly, $s(n)$ is bounded from above by the sum of the first $n/2$ odd integers $+1$. $s(n)$ is also bounded by the sum of the first $n$ primes, which is asymptotically equivalent to $\frac{n^2}{2\log{n}}$. It should thus be possible to find estimates for $s(n)$ using the fact that for an $\epsilon > 0$ and $n$ large enough $s(n) < (1+\epsilon)\frac{n^2}{\log{n}}.$

I would like to know if there are any known sharp upper bounds for $s(n)$. That is, I am looking for a function $f(n)$ such that for every $n > N_0$ $$s(n) \leq f(n)$$

As a way of relaxing the question, $s(n)$ could be regarded as the sum of the primes in the interval $[c,n]$ given a constant $c$.

-
I would have thought the sum of the first $n$ primes was asymptotically equivalent to $(n^2 \log n)/2$. (The $n$th prime is near $n \log n$, so the sum of the integers up to the $n$th prime is near $(n \log n)^2/2$, but only a proportion $1/\log n$ of the summands are prime. –  Michael Lugo Apr 29 '11 at 14:28
Did you try already to use Weil's explicit formula. Appropiated choosen tex functions, e.g. $e^{x tanh(x)} e^{-x^2 \delta}$ should give you an upper bound. –  Marc Palm Apr 29 '11 at 14:33

By partial summation $$s(n) = n\pi(n)-\sum_{m=2}^{n-1}\pi(m)$$ so by the Prime Number Theorem $$s(n) = \frac{n^2}{\log n}-\sum_{m=2}^{n-1}\frac{m}{\log m}+O\left(\frac{n^2}{\log^2 n}\right).$$ The sum on the right is $$\sum_{m=2}^{n-1}\frac{m}{\log m} = \int_2^n \frac{x}{\log x}dx + O\left(\frac{n}{\log n}\right)$$ using the monotonicity properties of the integrand. Now the integral equals, by partial integration, $$\int_2^n \frac{x}{\log x}dx = \left[\frac{x^2}{2\log x}\right]_2^n + \int_2^n \frac{x}{2\log^2 x}dx = \frac{n^2}{2\log n} + O\left(\frac{n^2}{\log^2 n}\right).$$ Altogether we have $$s(n) = \frac{n^2}{2\log n} + O\left(\frac{n^2}{\log^2 n}\right).$$ This can be made more precise both numerically and theoretically.

-

It is not difficult to calculate upper bounds on $s(n)$ from bounds on the prime counting function $\pi(n)$. Just use integration by parts, $$s(n) = \int\_0^n x\\,d\pi(x) = n\pi(n) - \int\_0^n\pi(x)\\,dx.$$ I'm not sure what the currently best known bounds for $\pi(x)$ are but, checking Wikipedia, gives $$\frac{x}{\log x}\left(1+\frac{1}{\log x}\right) < \pi(x) < \frac{x}{\log x}\left(1+\frac{1}{\log x}+\frac{2.51}{(\log x)^2}\right)$$ with the left hand inequality holding for $x\ge599$ and the right hand holding for $x\ge355991$. So,

$$s(n)\le \frac{n^2}{\log n}\left(1+\frac{1}{\log n}+\frac{2.51}{(\log n)^2}\right)-\int^n\left(1+\frac{1}{\log x}\right)\frac{x\\,dx}{\log x}+c$$ (where $c$ is a constant which you can compute if you feel so inclined). Applying integration by parts,

$$s(n)\le\frac{n^2}{2\log n}\left(1+\frac{1}{\log n}+\frac{5.02}{(\log n)^2}\right)-\frac12\int^n\left(1+\frac{2}{\log x}\right)\frac{x\\,dx}{(\log x)^2}+c$$

Bounding $\log x\le\log n$ in the integral gives a bound

$$s(n)\le\frac{n^2}{2\log n}\left(1+\frac{1}{2\log n}+\frac{4.02}{(\log n)^2}\right)+c$$

You can also take $c=0$ if you only require the bound to hold for $n\ge N$ (some $N$), since the term I neglected in the integral by applying $\log x\le \log n$ grows withuot bound, and will eventually dominate any constant term. Obviously, if you know any better bounds for $\pi(n)$ then you will get improved bounds for $s(n)$. For example, the same Wikipedia article linked to above states that $\left\vert\pi(x)-{\rm Li}(x)\right\vert\le\frac{\sqrt{x}\log x}{8\pi}$ for $x\ge2657$ under the assumption that the Riemann hypothesis holds.

-
+1 nice answer! –  Marc Palm Apr 29 '11 at 15:22

There definitely are earlier references than our book. An asymptotic formula for

$\sum_{p \leq x} p^a$

is in T. Salát and S. Znám, On the sums of the prime powers, Acta Fac. Rer. Nat. Univ. Com. Math. 21 (1968), pp. 21-24. (Cited by Spearman & Williams--I actually have not seen this paper.) It probably goes back further than that. The natural place to look would be Landau's "Primzahlen"--I forget the exact title--but I was unable to find that sum in there.

Eric.

-
Welcome to MathOVerflow, Eric! –  Charles May 20 '11 at 4:04

The following paper gives the asymptotic expansion of the sum of the first $n$ prime numbers. Hence for sufficiently large $n$, the first few positive and negative terms of the asymptotic expansion will give best upper and lower bound on the sum of primes.

http://arxiv.org/pdf/1011.1667.pdf

$$\sum_{r \le n}p_r = \frac{n^2}{2}\Bigg[\ln n + \ln\ln n - \frac{3}{2} + \frac{\ln\ln n}{\ln n} - \frac{3}{\ln n}- \frac{\ln^2 \ln n}{2\ln^2 n}$$

$$+ \frac{7 \ln \ln n}{2\ln^2 n} - \frac{27}{4\ln^2 n} + o\Bigg(\frac{1}{\ln^2 n}\Bigg) \Bigg].$$

-

http://mathworld.wolfram.com/PrimeSums.html

It says that

$$s(p_n) \tilde \quad n^2 \log n /2.$$

where $p_n$ is the $n$-th prime.

Perhaps you want to look at the reference, and figure out if you can make the bound effective.

-
So Michael Lugo's heuristic actually can be made effective... –  Marc Palm Apr 29 '11 at 15:08
Not exactly: there is a difference between the sum of primes less than or equal to $n$, and the sum of the $n$ first primes. I must have missed something: how could $s(n) \sim n^2 \ln(n)/2$ have not be proven before 1996 by Bach and Shallit? I do not see what is wrong with the fact that, since, $p_n \sim n \ln(n)$ (prime number theorem), one has: $\sum_{k=1}^n p_k \sim \sum_{k=1}^n k \ln(k) \sim \int_1^n t\ln(t) \mathrm{d}t \sim n^2\ln(n)/2$ –  Bernikov Apr 29 '11 at 15:11
Sorry, I have missed that difference. I did not read carefully. –  Marc Palm Apr 29 '11 at 15:18
Fixed the mistake... –  Marc Palm Apr 29 '11 at 15:25