3
$\begingroup$

I am looking for approximations, or a closed form, if available, for the sum
$$S(n,a,b)=\sum_{1\leq x,y,\leq n} \frac{x^a}{\mathrm{lcm}(x,y)^{b}},$$ where $\mathrm{lcm}(x,y)$ is the least common multiple of integers $x,y$ and $a,b$ are positive quantities. I am in particular interested in $a=b=1.$ For this case numerical evidence suggests $$ S(n,1,1)=O( n \log n) $$ may hold. In particular, I am wondering whether by using the technique in the answer to this question here, one might obtain (as $n \rightarrow \infty$), by letting $a,b\downarrow 1,$ an estimate in terms of zeta functions. In that question the upper bound $$ S(n,0,b)\leq\frac{\zeta(b)^3}{\zeta(2b)},\quad b>1 $$ is derived by letting $n\rightarrow \infty.$ Any pointers,comments welcome.

$\endgroup$

2 Answers 2

6
$\begingroup$

The original sum can be written as $$T(\alpha,\beta,\gamma,n)=\sum_{x,y\le n}x^\alpha y^\beta(x,y)^\gamma,$$ where $(x,y)=\mathrm{gcd}(x,y)$. One can find asymptotic formula for this sum using standart approach. Let $d=(x,y)$. Then $$T(\alpha,\beta,\gamma,n)=\sum_{d\le n}d^\gamma\sum_{{x,y\le n\atop (x,y)=d}}x^\alpha y^\beta=\sum_{d\le n}d^{\alpha+\beta+\gamma}\sum_{{x,y\le n/d\atop (x,y)=1}}x^\alpha y^\beta.$$ The condition $(x,y)=1$ can be removed using Möbius function: $$T(\alpha,\beta,\gamma,n)=\sum_{d\le n}d^{\alpha+\beta+\gamma}\sum_{\delta\le n/d}\mu(\delta)\sum_{{x,y\le n/d\atop \delta\mid(x,y)}}x^\alpha y^\beta=\sum_{d\le n}d^{\alpha+\beta+\gamma}\sum_{\delta\le n/d}\mu(\delta)\delta^{\alpha+\beta}\sum_{x,y\le n/(d\delta)}x^\alpha y^\beta.$$ The last sum (for $\alpha,\beta>-1$) is $\sim\frac{n^{\alpha+\beta+2}}{(\alpha+1)(\beta+1)(d\delta)^{\alpha+\beta+2}},$ so (for $\gamma>1$) $$T(\alpha,\beta,\gamma,n)\sim \frac{n^{\alpha+\beta+\gamma+1}}{\zeta(2)(\alpha+1)(\beta+1)}.$$

The special case $\gamma=1$, $\alpha=0$, $\beta=−1$ is more tricky. We can write the given sum as $$T(n)=\sum_{x,y\le n}\frac{(x,y)}{y}=T_1(n)+T_2(n),$$ where for some $U>1$ $$T_1(n)=\sum_{{x,y\le n \atop (x,y)\le U}}\frac{(x,y)}{y},\quad T_2(n)=\sum_{{x,y\le n \atop (x,y)> U}}\frac{(x,y)}{y}.$$ The second sum will be in the error term ($y=dy_1$, $x=dx_1$): $$T_2(n)=\sum_{d> U}d\sum_{{x,y\le n \atop (x,y)=d}}\frac{1}{y}\ll \sum_{d> U}\sum_{x_1,y_1\le n/d }\frac{1}{y_1}\ll \sum_{d> U}\frac{n}{d}\log\frac{n}{d}\ll n\log^2 \frac{n}{U}.$$ Here it is clear that for $U=n\log^{-2}n$ we get error term $O(R(n))$ with $R(n)=n\log^2\log n.$

The first sum gives main term: $$T_1(n)=\sum_{y\le n}\frac{1}{y}\sum_{{d\le U \atop d\mid y}}d\sum_{{x\le n \atop (x,y)=d}}1=\sum_{y\le n}\frac{1}{y}\sum_{{d\le U \atop d\mid y}}d\sum_{{x_1\le n/d \atop (x_1,y/d)=1}}1.$$ The last sum is known: $$\sum_{{x_1\le n/d \atop (x_1,y/d)=1}}1=\frac{\varphi(y/d)}{y/d}\frac nd+O(\tau(y/d)).$$ Hence $$T_1(n)=n\sum_{y\le n}\frac{1}{y}\sum_{{d\le U \atop d\mid y}}\frac{\varphi(y/d)}{y/d}+O(R_1(n)),$$ where $$R_1(n)=\sum_{d\le U}d\sum_{{y\le n \atop d\mid y}}\frac{\tau(y/d)}{y}=\sum_{d\le U}\sum_{y_1\le n/d }\frac{\tau(y_1)}{y_1}\ll U\log^2n\ll R(n).$$ So $$T_1(n)=n\sum_{d\le U}\frac{1}{d}\sum_{y_1\le n/d}\frac{\varphi(y_1)}{y_1^2}+O(R(n))=n\sum_{d\le U}\frac{1}{d}\left(\frac{1}{\zeta(2)}\left(\log (n/d)+\gamma-\frac{\zeta'(2)}{\zeta(2)}\right)+O\left(\frac{\log n}{n/d}\right)\right)+O(R(n))=\frac{n}{\zeta(2)}\sum_{d\le U}\frac{1}{d}\left(\log (n/d)+\gamma-\frac{\zeta'(2)}{\zeta(2)}\right)+O(R(n))=\frac{n}{\zeta(2)}\left(\left(\log n+\gamma-\frac{\zeta'(2)}{\zeta(2)}\right)\left(\log U+\gamma+O(1/U)\right)-\sum_{d\le U}\frac{\log d}{d}\right)+O(R(n)).$$ We also know that $$\sum_{d\le U}\frac{\log d}{d}=\frac{\log^2 U}{2}+\gamma_1+O(U^{-1}\log U).$$ Collecting all together we'll have $$T(n)=\frac{n}{\zeta(2)}\left(\frac{\log^2n}{2}+\log n\left(2\gamma-\frac{\zeta'(2)}{\zeta(2)}\right)\right)+O(n\log^2\log n).$$

$\endgroup$
3
  • $\begingroup$ Thanks. Your answer does not have a $\log n$ term, unlike Henri Cohen’s estimate, however, so the two seem inconsistent, if I let $LCM(x,y)=xy/(x,y)$. $\endgroup$
    – kodlu
    Apr 4, 2020 at 22:54
  • 1
    $\begingroup$ @kodlu My answer will have another form if $\alpha=-1$ or $\beta=-1$ or $\alpha<-1$,... $\endgroup$ Apr 5, 2020 at 1:43
  • $\begingroup$ Great! Since I'm specifically interested in sum of $x/LCM(x,y)=GCD(x,y)/y,$ that case would be $\gamma=1,$ $\alpha=0,\beta=-1.$ When you have a chance, can you address that case please? $\endgroup$
    – kodlu
    Apr 5, 2020 at 3:15
4
$\begingroup$

Partial answer: elementary arithmetic transformations show that $$S(n,1,1)=\sum_{1\le y\le n}\dfrac{1}{y}\sum_{d\mid y}\phi(d)\lfloor n/d\rfloor$$ which allows for much faster computation since it is essentially a single sum. I didn't push the analysis further, but my guess is that $S(n,1,1)$ is asymptotic to $Cn\log(n)^2$ (with a log squared), perhaps with $C=3/\pi^2=1/(2\zeta(2))$.

Complete answer: I was really lazy. From the expression above, it is immediate to show that $$S(n,1,1)=\sum_{1\le d\le n}\dfrac{\phi(d)}{d}\log(n/d)+O(n\log(n))\;,$$ and the main term is indeed asymptotic to $Cn\log(n)^2$ with $C=3/\pi^2$ if I am not mistaken.

$\endgroup$
2
  • $\begingroup$ Thanks. In Alexey Ustinov’s answer, there is no $\log n$ term, so it is a bit puzzling to me. $\endgroup$
    – kodlu
    Apr 4, 2020 at 22:55
  • 1
    $\begingroup$ The two results are perfectly consistent (I write $\log(n)^2$ instead of $\log^2(n)$), but Alexey's answer is more precise. $\endgroup$ Apr 5, 2020 at 9:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.