# Evaluating $\sum\limits_{n=1}^{\infty} \frac{1}{n\operatorname{ GPF}(n)}$, where $\operatorname{ GPF}(n)$ is the greatest prime factor

$\operatorname{ GPF}(n)=$Greatest prime factor of $n$, eg. $\operatorname{ GPF}(17)=17$, $\operatorname{ GPF}(18)=3$.

How to evaluate convergence/divergence/value of the sum

$$\sum_{n=1}^{\infty} \frac{1}{n\operatorname{ GPF}(n)}\,?$$

• On the numerical front: the convergence rate for your sum is ridiculously slow, even after the use of high precision and the use of a convergence acceleration method. The best estimate I have is $2.22$ (yes, just three digits, and I don't even trust the last $2$... :( ), after ten thousand or so terms. – J. M. is a poor mathematician Feb 5 '12 at 14:26
• I assume you're defining $GPF(1)=1$? E.g., MathWorld only defines the GPF for $n\ge 2$. Depending on the context in which this arose, the sum starting from 2 might be of more interest. – Ben Crowell Feb 5 '12 at 18:49
• @J.M.: I think it is pretty close to $2.25441837$. I calculated partial sums $R_p=1+\sum_{q\leq p}\frac{T_q}{q}$ of the desired sum $S$, with $T_p=\frac{1}{p}\prod_{q\leq p}\left(1-\frac{1}{q}\right)^{-1}$ calculated exactly with a successive ratio trick to accelerate convergence and checked against the asymptotic formula $T_p\approx e^\gamma\frac{\log p}{p}$ from Mertin's theorem. See my answer below for the details. – bgins Feb 5 '12 at 23:37
• +1 Cool question! Just an aside: Number theorists typically use the notation $P^+(n)$ for the largest prime factor of $n$, and $P^-(n)$ for the smallest prime factor of $n$, with the values $P^+(1) = 1$ and $P^-1(1) = \infty$. – JavaMan Feb 6 '12 at 1:13
• I have another (surprisingly good) repeated rough analytic approximation, 2.06527893367, using an integral approximation to Merten's $\sum\frac{\log p}{p}$, change of variable using $\pi(x)$, and the (rough) approximation $\pi(x)\approx\text{li}(x)$. – bgins Feb 6 '12 at 12:15

If $\{r_1,\dots,r_k,p\}$ is the set of primes $\le p$, then

$$\{n\in \mathbb{Z}: \operatorname{gpf}(n)=p\}=\{r_1^{e_1}\cdots r_k^{e_k}p^f:\text{each }e_i\ge 0,f\ge1\}.$$

Now we have the Euler product factorization

$$\sum_{e_1\,\ge0}\cdots\sum_{e_k\,\ge0}\sum_{f\ge1}\frac{1}{r_1^{e_1}\cdots r_k^{e_k} p^f}= \prod_{i=1}^k \left(1+\frac{1}{r_i}+\frac{1}{r_i^2}+\cdots\right)\cdot \left(\frac{1}{p}+\frac{1}{p^2}+\frac{1}{p^3}+\cdots\right)$$

$$\implies \bullet =\sum_p \frac{1}{p}\sum_{\operatorname{gpf}(n)=p}\frac{1}{n}=\sum_p \frac{1}{p}\left[\,\prod_{q\,< \,p}\left(1-\frac{1}{q}\right)^{-1}\right]\frac{1/p}{1-1/p}$$

$$=\sum_p \frac{1}{p(p-1)}\prod_{q<p}\left(1-\frac{1}{q}\right)^{-1}.$$

With Mertens' Theorem, we know that the $\prod$ above is $\le C\log p$ eventually, so after we ignore a finite sum corresponding to the $p$ before this "eventually" we may say that

$$\frac{\bullet}{C} + \zeta_P'(2)\le \sum_p \frac{\log p}{p}\left(\frac{1}{p-1}-\frac{1}{p}\right)\le \sum_p\frac{\log p}{p^2}=-\zeta_P'(2).$$

Here $\zeta_P$ is the prime zeta function. With monotonicity of the terms, this proves convergence.

• Nice, but aren't you forgetting the $n=1$ term? – bgins Feb 5 '12 at 18:01
• @bgins: Good catch. Since $\operatorname{gpf}(1)$ is undefined and an $n=1$ term is superfluous anyway I think I'll just leave it at $\Sigma_2^\infty$. – anon Feb 5 '12 at 19:40

Call the sum $S$, and assume that $\text{GPF}(1)$ is defined to be $1$. For $n\geq1$ and $p,q$ prime, $\text{GPF}(n)=p$ iff $p|n$ and $n$ is a product of primes $q\leq p$, so that, by the Euler product factorization (as noted by @anon), the sum of reciprocals of all such $n$ is $$T_p= \sum_{\text{GPF}(n)=p} \frac{1}{n}= \frac{1}{p}\; \prod_{q \leq p}\; \sum_{i=0}^\infty q^{-i} = \frac{1}{p}\; \prod_{q \leq p} \left( 1-\frac{1}{q} \right)^{-1}.$$ Also (as @anon noted) by Mertens' third theorem, $p\;T_p\rightarrow e^\gamma\log p$ so that $T_p\rightarrow e^\gamma\frac{\log p}{p} \rightarrow 0$ as $p\rightarrow\infty$, with $T_p-e^\gamma\frac{\log p}{p}$ in fact changing sign infinitely often (Robin 1983) and $$S = \sum_{n=1}^\infty \frac{1}{n\text{GPF}(n)} = 1 + \sum_{p} \frac{1}{p} T_p$$ must converge: $\sum\frac{T_p}{p}\sim \sum\frac{\log p}{p^2} <\sum\frac{\log n}{n^2}$ all converge by the limit comparison and integral tests since $\int_1^\infty\frac{logx}{x^2}dx= -\left|\frac{1+\log x}{x}\right|_1^\infty= 1$. Note that $$R_p= \sum_{\text{GPF}(n)\leq p}\frac{1}{n\text{GPF}(n)}= 1+ \sum_{q\leq p}\; \sum_{\text{GPF}(n)=q} \frac{1}{nq}= 1+ \sum_{q\leq p} \frac{1}{q} T_q$$ monotonically increases to $S$ as $p\rightarrow\infty$, and that for $p,q$ and $r$ prime, $$\sum_{\text{GPF}(n)>p}\frac{1}{n\text{GPF}(n)}= S-R_p= \sum_{q>p} \frac{1}{q} T_q= \sum_{q>p} \frac{1}{q^2} \prod_{r\leq q} \left( 1-\frac{1}{r} \right)^{-1}.$$ Furthermore, note that multiplying by $(1-\frac{1}{p})$ on the right above has the effect of removing the $r=p$ term from the product, thus eliminating all terms with $p\mid n$ from the summation on the left. Thus $$S-R_p+\frac{1}{p}T_p= \sum_{\text{GPF}(n)\geq p}\frac{1}{n\text{GPF}(n)}= S_p+\sum_{ \begin{matrix} \text{GPF}(n)>p \\p\nmid n \end{matrix} } \frac{1}{n\text{GPF}(n)}$$ $$=S_p+ \left(1-\frac{1}{p}\right) \sum_{q>p}\frac{1}{q}T_q =S_p+ \left(1-\frac{1}{p}\right) \left(S-R_p\right)$$ $$\implies\quad 0 < S - R_p = p S_p - T_p \rightarrow 0 \quad\text{as}\quad p\rightarrow\infty$$ where $S_p$ is the sum of terms from $S$ for which $n$ is divisible by the prime $p$: $$S_p= \sum_{p|n}\frac{1}{n\text{GPF}(n)}= \sum_{n=1}^{\infty}\frac{1}{pn\text{GPF}(pn)}. %ERROR: %= %\frac{1}{p^2}- %\frac{1}{p}+ %\sum_{n=1}^{\infty}\frac{1}{pn\text{GPF}(n)} %= %\frac{S}{p}-\frac{p-1}{p^2}.$$

To check this, we might want the first few values of $T_p$: $T_2=T_3=1$, $T_5=\frac{3}{4}$, $T_7=\frac{5}{8}$, $T_{11}=\frac{7}{16}$, $T_{13}=\frac{77}{192}$ & $T_{17}=\frac{1001}{3072}$, along with the following recursion, which follows from the formula above for $T_p$: if $p_k$ is the $k^{\text{th}}$ prime, then $$\frac{T_{p_{k+1}}}{T_{p_k}} = \frac{p_k}{p_{k+1}-1}.$$ Working in sage (online), we could calculate and plot a rough lower bound for $S$ as follows:

P = Primes()
p = P.first()
T = 1; R = 3/2 # R = R_2 = 1 + T_2/2
x = []; y = []
for k in range(1,10000):
q = p
p = P.next(p)
T = T * q/(p-1)
R = R + (T/p).n(digits=16)
x.append(log(p).n(digits=6))
y.append(R)
print 'p = %3d T =' % p, T.n(digits=12), 'R =', R.n(digits=12)
print 'T ~', (e^euler_gamma*log(p)/p).n(digits=12)
list_plot(zip(x,y))

p = 104729 T = 0.000196636242440 R = 2.25441837672
T ~ 0.000196580221393


where we plot the partial sums $R_p$ versus $\log p$ for $p\in\{p_2=3,\dots,p_{10000}=104729\}$. The last computed term (with $p=104729$) has contributed about $T_p/p \approx 0.000196636/104729 \approx 2\times 10^{-9}$ to $R_p \approx 2.2544184$, so the convergence seems adequate to offer the estimate as a fairly good lower bound. We also compare the computed (exact) value of $T_p=0.00019663624$ with the asymptotic estimate, $e^\gamma\frac{\log p}{p}=0.000196580221393$.

Notice that the plot above appears to repeatedly cross its "best fit" continuous approximation. This observation corresponds to the 1983 result of Guy Robin, and can be made more precise, yielding in the process another, presumably course analytic estimate for $S$ by way of the asymptotic Merten approximation $T_p \approx e^\gamma\frac{\log p}{p}$, whose corresponding sum can also be expressed analytically in terms of the prime zeta function: $$S=1+\sum_p\frac{1}{p}\prod_{q\leq p}\left(1-\frac{1}{q}\right)^{-1} =1+\sum_p\frac{T_p}{p} \quad\implies$$ $$S \approx 1-e^\gamma\zeta_{P}'(2) = 1+e^\gamma\sum_p\frac{\log p}{p^2} = 1+e^\gamma\sum_{k=1}^\infty\frac{\log p_k}{p_k^2}, \quad\text{where}$$ $$\zeta_P(s) = \sum_p p^{-s} \implies \zeta_{P}'(s) = -\sum_{p}\frac{log p}{p^s} \quad\text{since}\quad \frac{d}{ds}e^{-s\log p} = e^{-s\log p}(-\log p).$$ Called from sage, mpmath calculates $\zeta_{P}'(2)=-0.93754825431584377$, yielding the "Merten" estimate $S\approx 1-e^\gamma\zeta_{P}'(2)=2.669841336296809$:

import mpmath; mpmath.zeta(2,1,1)
S_Merten = 1 + e^euler_gamma * 0.93754825431584377
S_Merten.n(digits=16) # -> 2.669841336296809


This in turn can be analytically approximated by transforming the discrete variables $k$ and $p=p_k$ to continuous variables $t=\pi(x)$ and $x$, respectively, where $\pi(x)$ is the prime-counting function. Using the logarithmic integral approximation for $\pi(x)$, the fundamental theorem of calculus (FTOC) and integration by substitution, we have: $$t=\pi(x) \approx \text{li}(x)=\int_0^x\frac{dt}{\log t} \quad\implies\quad dt\approx\frac{dx}{\log x} \qquad\text{(FTOC)}$$ $$\frac{\log p_k}{p_k^2} \approx \int_{t=k-.5}^{t=k+.5} \frac{\log x}{x^2}dt \quad \implies \quad S \approx 1 + e^\gamma \int_{\text{li}^{-1}(.5)}^\infty \frac{\log x}{x^2} \frac{dx}{\log x} = 1 + \frac{e^\gamma}{{\text{li}^{-1}(.5)}}$$

x=var('x'); c=find_root(Ei(log(x))==.5, 1, 10)
S=1+e^euler_gamma/c; [c, S.n(digits=12)]


Sage reports $\text{li}^{-1}(.5)\approx 1.6719305730098757$ and $S\approx 1+0.598111e^\gamma$ $=2.06527893367$, which is surprisingly accurate given the inherent inaccuracy in the approximation $\pi(x)=\text{li}(x)$, $$\frac{\text{li}(x)-\pi(x)}{\pi(x)} \approx \log x \cdot \exp\left(-\frac{\sqrt{\log x}}{15}\right),$$ and given the initial negative bias in the Merten approximation for $T_p$. However, $T_p-e^\gamma\frac\log{p}{p}$ changes sign infinitely often (Robin 1983), as does $\pi(x)-\text{li}(x)$ (Littlewood 1914), which exhibits quite a delay before the first sign change.

$S-1$ also appears to be rather close to the number $1.25506$ offered (apparently without citation here) as a bound on the supremum of $\frac{\pi(x)\log(x)}{x}$ for $x>1$.

Robin, G. (1983). "Sur l’ordre maximum de la fonction somme des diviseurs". Séminaire Delange–Pisot–Poitou, Théorie des nombres (1981–1982). Progress in Mathematics 38: 233–244.

Littlewood, J. E. (1914), "Sur la distribution des nombres premiers", Comptes Rendus 158: 1869–1872

• Can P.I.E. (en.wikipedia.org/wiki/Principle_of_inclusion-exclusion) get us any further? – bgins Feb 5 '12 at 17:57
• @Ben-Crowell: I haven't had the time to review your interesting approach yet (may take 1 more day). I have certainly already vetted my own fair share of mistakes! However, everything in my post up to the words "must converge" is a condensation of anon's and julian-aguirre's posts (reconciling my aesthetics with their math). I would of course appreciate any corrections, but I think $T_p$ ~ $\frac{\log p}{p}$ and the partial sums $R_p$ of $S$ are pretty accurate. You even see the alternating sign of $R_p''$ in the graph of $R_p$ proved by Guy Robin in 1983 (see the link to Merten's Theorem 3). – bgins Feb 5 '12 at 23:21
• Yes, my previous calculation had mistakes in it. My corrected result is now consistent with, but poorer than, yours. The fact that your 2.25 is close to J.M.'s 2.22 is further encouragement to believe that you guys are right. – Ben Crowell Feb 5 '12 at 23:25
• Interesting! I'll run my numerics on your reformulation of the sum and report back. :) – J. M. is a poor mathematician Feb 6 '12 at 1:06
• @bgins: the difference between $\pi(x)$ and $\mathrm{li}(x)$ changes sign infinitely often, but if I remember right the first change comes when $x$ is in the range of $10^{34}$ or so, so your relative error may be negative for quite a while! – Steven Stadnicki Feb 6 '12 at 16:03

Let $p_k$ be the $k$-th prime number. $$\sum_{n\ge1}\frac{1}{n\operatorname{GPF}(n)}=\sum_{k\ge1}\frac{1}{p_k}\sum_{\operatorname{GPF}(n)=p_k}\frac{1}{n}.$$ If $\operatorname{GPF}(n)=p_k$, then $n=p_1^{i_1}\dots p_{k-1}^{i_{k-1}}\,p_k^{i_k}$ with $i_j\ge0$, $1\le j<k$ and $i_k\ge1$. It folows that \begin{align*} \sum_{\operatorname{GPF}(n)=p_k}\frac{1}{n}&=\Bigl(\sum_{i_1\ge0}p_1^{-i_1}\Bigr)\dots\Bigl(\sum_{i_{k-1}\ge0}p_{k-1}^{-i_{k-1}}\Bigr)\Bigl(\sum_{i_k\ge1}p_k^{-i_k}\Bigr)\\ &=\frac{1}{p_k}\,\prod_{i=1}^k\Bigl(1-\frac{1}{p_i}\Bigr)^{-1}. \end{align*} From Merten's theorem $$\prod_{i=1}^k\Bigl(1-\frac{1}{p_i}\Bigr)^{-1}\sim e^\gamma\,\log p_k\ ,$$ and the original series has te same character as $$\sum_{k\ge1}\frac{\log p_k}{p_k^2} ,$$ which is convergent.

• The geometric series for $p_k$ evaluates not to $p_k^{-1}$ but $(p_k-1)^{-1}$, which is a slightly more work for proving convergence. – anon Feb 5 '12 at 9:59
• @anon The first term of the geometric series for $p_k$ is $1/p_k$, not $1$, so that it evaluates to $p_k^{-1}(1-p_k^{-1})$. – Julián Aguirre Feb 5 '12 at 10:26
• Oh you subsumed part of it into the product and part of it stays outside the product, I see now. – anon Feb 5 '12 at 10:31

I've attempted to put a numerical lower bound on the sum. Let $S=\sum_{j=1}^\infty 1/jGPF(j)$, where I assume $GPF(1)=1$. Let $r_i$ be the $i$th prime. Then using the geometric series, we have

$S=1+\displaystyle\sum_{\{e_i\}} r_n^{-2}\left[\frac{1}{1-r_n^{-1}}\right]\prod r_i^{-e_i}$

where the sum is over sets of exponents $\{e_i\}$ such that the highest $i$ occurring is $k$, where $0 \le k < n$. The idea here is that most big contributions to $S$ come from terms with fairly small $k$. Let $m=\sum e_i$. Then

$S \ge 1+\displaystyle\sum_{n=1}^\infty r_n^{-2}\left[\frac{1}{1-r_n^{-1}}\right] \left[1+\sum_{m=1}^\infty \sum_{k=1}^{n-1} r_k^{-m} \sum_{\{e_i\}} 1\right]$ ,

where the 1 term before the sum over $m$ is to account for the $m=0$, $k=0$ term. Now let $B(m,k)$ be the number of partitions of $m$ into $k$ nonnegative terms, with order counted as significant, and with the final term being nonzero. Then $B(m,k)={{m+k-1} \choose k}\ge (1+(m-1)/k)^k$, so

$S \ge 1+\displaystyle\sum_{n=1}^\infty r_n^{-2}\left[\frac{1}{1-r_n^{-1}}\right] \left[1+\sum_{m=1}^\infty \sum_{k=1}^{n-1} r_k^{-m}\left(1+\frac{m-1}{k}\right)^k\right]$

The following code, written in the open-source programming language Yacas, is intended to evaluate this expression:

n_max := 100;
m_max := 100;
prec := 8; /* digits of precision */
n := 1;
rn := 1;
s := N(1,prec); /* take GPF(1)=1, so first term=1 */
While (n<=n_max) [
u := N(1,prec); /* 1=contribution from m=0, k=0 */
rn := NextPrime(rn); /* rn=nth prime */
m := 1;
While (m<=m_max) [
k := 1;
rk := 1;
While (k<n) [
rk := NextPrime(rk); /* rk=kth prime */
u := N(u+(1+(m-1)/k)^k*rk^-m,prec);
k := k+1;
];
m := m+1;
];
sn := N(u*rn^-2*(1/(1-1/rn)),prec);
s := N(s+sn,prec);
Write(n,sn,s); NewLine();
n := n+1;
];
Write(s); NewLine();


Summing up to maximum $n$ and $m$ values of 100 gives $S\ge 1.39$, which is in agreement with, but poorer than, the estimates by J.M. and bgins.

I've caught and corrected several errors in this after I initially posted it. Maybe there are more...