Hyperbolic Summation for Fractional Sums

For an integer $n\geq 1$, let $\tau(n)$ be the number of positive divisors of $n$. A classical result of Dirichlet states that

where is $\gamma$ is the Euler’s constant and $\Delta(x)$ is an error term. Dirichlet proved that $\Delta(x)\ll x^{1/2}$. Improved estimated for $\Delta(x)$ have been given by numerous authors ([16, 14, 15, 7, 6]), with the current record being Huxley’s [5] estimate: for any $\varepsilon>0$,

By the definition of $\tau(n)$, we note that

Thus, we can consider (1.1) as an asymptotic formula for the fractional sum $\sum_{n\leq x}\left\lfloor x/n\right\rfloor$, where $\lfloor t\rfloor$ is the largest integer not exceeding $t$. With this viewpoint, Bordellès, Dai, Heyman, Pan and Shparlinski [2] investigated the more general sums

for arithmetic functions $f$ satisfying various growth conditions. Later, Wu [17] and Zhai [18] independently showed that if

for some $\alpha\in[0,1)$ and $\theta\geq 0$, then

It is possible to improve (1.3) for functions $f$ satisfying certain decomposition identities. For example, if $f=\Lambda$, the von Mangoldt function, Ma and Wu [11] proved that

for any $\varepsilon>0$. Here the “decomposition identity” needed is Vaughan’s identity. The exponent $35/71$ was then improved to $9/19$ by Liu, Wu and Yang [8]. If $f=\tau$, Ma and Sun [10] showed that

The exponent $11/23$ was then improved to $19/40$ and $9/19$, respectively, by Bordellès [1] and Liu, Wu and Yang [9], and finally to $5/11$ by the third author [13].

Motivated by recent results, it is important to study sums of the form

where $r$ is a fixed positive integer and $\tau_{r}$ is the number of ways of writing $n$ as a product of $r$ positive integers. In this paper, we consider the case $r=2$ and put

Supposing that $\Delta(x)\ll x^{\theta+\varepsilon}$ and that $f(n)\ll n^{\alpha}$ for some $\alpha\in[0,1)$, it is not hard to show (see Section 3) that

where

and

Note that both sums converge since $\alpha<1$. Huxley’s bound (1.2) then gives

It is known that $\Delta(x)\gg x^{1/4}$ for infinitely many values of $x$, and it is conjectured that $\Delta(x)\ll x^{1/4+\varepsilon}$ for all $x$. Taking $\theta=\frac{1}{4}$ in (1.4), one conjectures that

Using an estimate for three-dimensional exponential sums due to Robert and Sargos [12], we are able to prove this result unconditionally.

In this section, we state some lemmas that will be needed in the proof of Theorem 1.1. For the error term $\Delta(x)$ in (1.1), we have the following expression.

The function $\psi(x)$ is periodic with period $1$ and so can be expanded into a Fourier series. We need the following truncated version due to Vaaler.

Lastly, we need the following lemma on certain three-dimensional exponential sums.

As a consequence of the above three lemmas, we deduce the following estimate for the average of $\Delta$ over a sequence of monomials.

We note that this agrees with the estimate one obtains from the conjectured bound $\Delta(x)\ll x^{1/4+\varepsilon}$ so long as $D\geq\sqrt{X}$.

From this, we deduce the following special case needed for the proof of Theorem 1.1.

Let $N\in[1,x)$ be a parameter that will be chosen later. We split $T_{f}(x)$ into two parts:

where

Since $\tau(n)\ll n^{\varepsilon}$ for any $\varepsilon>0$, we have

If $d=\lfloor x/n\rfloor$, then $x/n-1<d\leq x/n$ and $x/(d+1)<n\leq x/d$, so

By (1.1), we obtain

where

Since $f(d)\ll d^{\alpha}$ with $\alpha<1$, we have

and similarly

Therefore

and so

To prove (1.4), we note that the estimate $\Delta(x)\ll x^{\theta+\varepsilon}$ immediately gives

and (1.4) follows by choosing $N=x^{\frac{1}{2-\theta}}$.

Finally, to prove Theorem 1.1, we note that

so that Proposition 2.5 gives

The second term dominates so long as $N\leq x^{2/3}$. In this case, combining the above estimate with (3.10), we have

Choosing $N=x^{4/7}$, we complete the proof of Theorem 1.1.

The authors would like to thank Professor Kui Liu for his suggestions to improve this paper.