Logarithmic exponential sums with multiplicative coefficients

From (?), Proposition 3.1, and Proposition 5.1, we get

Since \log\log x\ll\log q in this range, this proves (1.1) at a/q, and (?) returns us to \alpha. If f is completely multiplicative, the second part of Proposition 5.1 gives us the same result with 4/\sqrt{6} replaced by 1.

The constant 4/\sqrt6 is best possible by [GranvilleLamzouri2026, Example 1.7]. To see that the constant 1 is also best possible, we fix a primitive character \chi\pmod{q_0} and take f=\overline{\chi} and \alpha=a/q_0, where (a,q_0)=1. Then

By [GranvilleLamzouri2026, Corollary 7.1], there is a primitive character \chi\pmod\ell, with \ell\mid q, such that

where f=\kappa*\chi. The error term is clearly o((\log x)/\sqrt q) since q<\log x.

For every n with \chi(n)\neq0, writing n=dm, where all prime factors of d divide q but not \ell, and (m,q)=1, we obtain

where we replace m\le x/d by m\le x in the inner sum of the first line, at a cost \sum_{x/d<m\leqslant x}m^{-1} \ll \log d, and then extend the d-sum to infinity. A crude estimate gives the error

while the contribution from d>x is o(1) by Rankin’s trick.

Moreover, by Möbius inversion,

Thus, using |\tau(\chi)|=\sqrt\ell, it remains to show that

We verify this prime by prime. Write p^e\Vert q and p^j\Vert\ell, where 0\leqslant j\leqslant e. If j\geqslant1, then \chi(p)=0 and \kappa(p^{e-j})=f(p)^{e-j}. Hence the corresponding local factor on the left-hand side of (?) is at most p^{j/2-e}\leqslant p^{-e/2}. If j=0, put z:=f(p)\overline{\chi(p)}. Since f=\kappa*\chi and f is completely multiplicative, we have

The corresponding local factor on the left-hand side of (?) is therefore at most

where we used |1-z|\leqslant\sqrt p |1-z/p|, since |z|\leqslant1 and p|1-z/p|^2-|1-z|^2 =(p-1)\left(1-{|z|^2}/{p}\right)\geqslant0. Multiplying over the local factor at p^e\Vert q proves (?).

Substituting (?) into (Proof) we get

The two O(\log^2 q) errors above contribute, after substitution into (Proof), at most

where the second term follows from (?) and the first term comes from |\kappa(a)| \le 2^{\omega(a)}. This proves the lemma.

The contribution of L_0 is immediate from Mertens’ theorem:

We now turn to L_1. After removing the terms divisible by p^2 for some p>y, as in the previous section, every remaining integer counted by L_1 can be written uniquely as n=pm, where p>y and P^+(m)\leqslant y. Hence

We first consider m\leqslant\sqrt q. We may replace p\leqslant x/m by p\leqslant x/\sqrt q, at a cost

Since \sqrt q<y, the condition P^+(m)\leqslant y is automatic. Thus, after grouping the prime variable by its residue class modulo q, the contribution from m\leqslant\sqrt q is

by Cauchy–Schwarz, where

For \|A\|_2, let I=(P,2P] be a dyadic interval. We write

Since P\geqslant y=q^2, \log (P/q)\asymp\log P, then by the Brun–Titchmarsh inequality and partial summation formula, we get

Hence by Minkowski’s inequality, we have

For \|V\|_2, let c_q(h):= \sum_{u\in(\mathbb{Z}/q\mathbb{Z})^\times}\mathrm e(hu/q) denote the Ramanujan sum. For m\neq n, the standard bound |c_q(h)|\le(q,h) gives |c_q(m-n)|\le(q,m-n)\le |m-n|<\sqrt q. Therefore

Combining this with the bound for \|A\|_2, the contribution from m\leqslant\sqrt q is \|A\|_2\|V\|_2\ll\log\log x.

We now consider m>\sqrt q. Split the prime variable into dyadic intervals I=(P,2P]. For p\in I, we replace m\leqslant x/p by \sqrt q<m\leqslant x/(2P). The omitted boundary terms are bounded by

For a fixed I, grouping both variables by their residue classes modulo q, the contribution from the interval I is

where

and

Now, summing over the dyadic intervals I and using the same estimate for \|A_I\|_2 as above, we obtain the contribution from m>\sqrt{q} is

where we used \sum_I\sum_{p\in I}\frac1p\ll\log\log x and (?). Together with the boundary terms and the range m\leqslant\sqrt q, this proves the proposition.

This follows directly from the identity we stated above by writting k:=\omega_y(n).

By the Chinese remainder theorem, it is enough to estimate the corresponding sums prime by prime. Let q=p^e, and suppose that \chi is induced by a primitive character \chi^\ast\pmod{p^j}, where 0\leqslant j\leqslant e. If j\geqslant1, then

Here we used the standard formula |\tau(\chi^\ast)|=p^{j/2} for primitive characters.

If j=0, then \chi is principal modulo p^e, and \tau(\overline{\chi},ap^v)=c_{p^e}(ap^v). Hence

Returning to general q, we get

This proves (4.2).

Suppose now that f is completely multiplicative. By the Chinese remainder theorem, the sum in (4.2) factors prime by prime, up to unit factors of modulus 1. For p^e\Vert q, the nonprincipal case contributes at most p^{j/2}\leqslant p^{e/2}. In the principal case, after absorbing the unit factor into f(p), the local sum contributes, for some |z|\le1,

by the same inequality |1-z|\leqslant\sqrt p |1-z/p| we used in Lemma 2.1. Multiplying over p^e\Vert q proves (4.2).

Finally, we use Rankin’s trick to bound

The same calculation at prime power p^e gives

and therefore (4.2) follows by

Split the prime variables into dyadic intervals I=(P,2P] and J=(R,2R], we define

The choice of X separates the condition dcpr\leqslant x on each dyadic block.

If d>X, then {x}/{4}<n=dcpr\leqslant x. After summing over c,I,J, taking absolute values and applying Lemma 4.1, the total contribution of all such terms to L_{\ge2} is at most

For d\leqslant X, we have dcpr\leqslant x for every p\in I and r\in J. After temporarily including the diagonal p=r when I=J, whose contribution is acceptable, we have, by plugging (?) into ?,

where

By character orthogonality and the same Brun–Titchmarsh estimate used in Section ?,

Then by (4.2), (?), and Cauchy–Schwarz, we have

We now fix 0<\delta<1 and put z:=\exp((\log x)^\delta). We first consider the part of L_{\geqslant2} for which at least one of the two marked primes lies between y and z. By (?) and (?), and extending the sums, their total contribution is

Here

It remains to consider the part of L_{\geqslant2} for which both marked primes exceed z. Since p\in(P,2P] and r\in(R,2R], we have P,R\geqslant z/2. Since q\leqslant(\log x)^2\leqslant(\log P)^{2/\delta}, we can apply the Siegel–Walfisz theorem to get

Suppose X=x/(4cPR)<\log x, then \frac{x}{4\log x}<cpr\leqslant x. Using (?) and Lemma 4.1, the total contribution of these blocks is

by (?), which is negligible.

We are left with X\geqslant\log x. For a general multiplicative function, (4.2) gives

If f is completely multiplicative, then (4.2) and (4.2) give, for \varepsilon=1/8 (say),

Hence, by (?) and Cauchy–Schwarz,

in the multiplicative case. If f is completely multiplicative, the same estimate holds with 4/\sqrt6 replaced by 1.

To recombine the prime sums, we may drop the conditions p\nmid c and r\nmid c. The corresponding loss is

which is again negligible to L_{\geqslant2}. For the remaining terms, put m=cpr and write

For each fixed m, the marked primes can be chosen in K_z(K_z-1) ordered ways, while the two-prime Ramaré weight gives

Thus, in the multiplicative case, the contribution from p,r>z and X\geqslant\log x is at most

For completely multiplicative f, the same argument gives (1+o_x(1))\frac{\log x}{\sqrt q}.

Combining this with the contribution from the range in which at least one marked prime lies in (y,z], we obtain

for general multiplicative functions, and 4/\sqrt{6} replaced by 1 for completely multiplicative functions. We finally let x\to\infty first with \delta fixed, and then let \delta\to0. This proves the proposition.