ABCExceptions

2 Reduction to Diophantine equations

We will work with a variant of \(S_{\alpha ,\beta ,\gamma }(X)\)

Definition 2.1
#

Let \(S^*_{\alpha ,\beta ,\gamma }(X)\) to be the number of \((a,b,c)\in \mathbb {N}^3\) with \(\gcd (a,b,c)=1\) and

\begin{align*} c\in [X/2,X],\quad a+b=c,\quad \operatorname{rad}(a)\sim X^{\alpha },\quad \operatorname{rad}(b)\sim X^{\beta },\quad \operatorname{rad}(c)\sim X^{\gamma }. \end{align*}

We begin by noting that by the pigeonhole principle,

We have

\begin{equation} \label{eq:step1} N_\lambda (X)\ll (\log X)^4 \max _{\substack { \alpha ,\beta ,\gamma {\gt}0\\ \alpha +\beta +\gamma \leq \lambda }} S^*_{\alpha ,\beta ,\gamma }(X). \end{equation}
2.0.1

Proof

There exists \(\varepsilon {\gt}0\) such that for all \(\mathbf{c}\in \mathbb {Z}^3\) and \(\mathbf{X},\mathbf{Y},\mathbf{Z}\in \mathbb {R}_{{\gt}0}^{d}\). we have

\begin{align} B_d({\bf c, X,Y,Z}) \ll X^{0.66-\varepsilon }. \end{align}
Proof
Proof of Theorem 1.4

The following result allows us to bound \(S^*_{\alpha ,\beta ,\gamma }(X)\) in terms of the number of solutions to certain monomial Diophantine equations. In order to state it, we need to introduce the quantity \(B_d\)

Definition 2.4
#

For \(\mathbf{c}\in \mathbb {Z}^3\) and \(\mathbf{X},\mathbf{Y},\mathbf{Z}\in \mathbb {R}_{{\gt}0}^{d}\). we have

\begin{equation} \label{eq:Bk} B_d(\mathbf{c},\mathbf{X},\mathbf{Y},\mathbf{Z}) := \# \left\{ (\mathbf{x},\mathbf{y},\mathbf{z})\in \mathbb {N}^{3d}\; :\; \begin{array}{l} x_i\sim X_i,\, y_i\sim Y_i,\, z_i\sim Z_i \\ c_1\prod _{j\leq d}x_j^j+c_2\prod _{j\leq d}y_j^j=c_3\prod _{j\leq d}z_j^j\\ \gcd (c_1\prod _{j\leq d}x_j,c_2\prod _{j\leq d}y_j , c_3\prod _{j\leq d}z_j)=1 \end{array} \right\} . \end{equation}
2.0.3

Lemma 2.5

Let \(\varepsilon \in (0,1/2)\), and let \(2\leq n\leq X\) be an integer. Then there exists a factorisation

\begin{align*} n=c\prod _{j\leq \frac{5}{2}\varepsilon ^{-2}}x_j^{j}, \end{align*}

for positive integers \(x_j,c\) such that \(c\leq X^{\varepsilon /2}\), the \(x_j\) are pairwise coprime, and

\begin{align*} X^{-\varepsilon }\prod _{j\leq \frac{5}{2}\varepsilon ^{-2}}x_j \leq \operatorname{rad}(n) \leq X^{\varepsilon }\prod _{j\leq \frac{5}{2}\varepsilon ^{-2}}x_j. \end{align*}
Proof

Fix \(2\le n\le X\) and let \(K=2\lceil \varepsilon ^{-1}\rceil \), \(M=\lfloor \frac{5}{2}\varepsilon ^{-2}\rfloor \). Define

\begin{align*} y_j:=\prod _{p^j\| n} p. \end{align*}

For \(j\leq M\), we set

\begin{alignat*}{2} x_j:=\begin{cases} y_j & \text{for $j\neq K$},\\ y_j\prod _{m{\gt}M}y_m^{\lfloor m/K\rfloor } & \text{for $j=K$}, \end{cases}\qquad \text{and}\qquad c\, :=\prod _{m{\gt}M}y_m^{m-K\lfloor m/K\rfloor }. \end{alignat*}

All the \(x_j\) are pairwise coprime, since the \(y_j\) are pairwise coprime.

Note that by definition \(c\prod _{j\leq M}x_j^j=\prod _{m\geq 1}y_m^m=n\leq X\). In particular,

\begin{align*} \prod _{m\geq M}y_m\leq \Big(\prod _{m\geq M}y_m^{m}\Big)^{1/M}\leq X^{1/M}. \end{align*}

Then, since \(m-K\lfloor m/K\rfloor \le K\), it follows from the definition of \(c\) that

\begin{align*} c\leq \prod _{m\geq M}y_m^K\leq X^{K/M}\leq X^{\varepsilon /2}. \end{align*}

Thus

\[ \operatorname{rad}(n)\leq \operatorname{rad}(c)\prod _{j\leq M}\operatorname{rad}(x_j)\leq X^{\varepsilon /2} \prod _{j\leq M} x_j. \]

On the other hand, we have

\begin{align*} x_K = y_K\prod _{m{\gt} M}y_m^{\lfloor m/K\rfloor }\leq \Big(y_K^K\cdot \prod _{m{\gt} M}y_m^m\Big)^{1/K}\leq n^{1/K}\leq X^{\varepsilon /2}. \end{align*}

Recalling that the \(y_j\) are squarefree and pairwise coprime for \(j\neq K\), gives the lower bound

\[ \operatorname{rad}(n)=\prod _{m\geq 1}y_m \geq \prod _{\substack {j\leq M\\ j\neq K}}x_j\geq X^{-\varepsilon /2}\prod _{j\leq M}x_j, \]

as claimed.

Proposition 2.6

Let \(\alpha ,\beta ,\gamma \in (0,1]\) be fixed and let \(X\geq 2\). For any \(\varepsilon {\gt}0\) there exists an integer \(d=d(\varepsilon )\geq 1\) such that the following holds. There exist \(X_1,\ldots , X_d,Y_1,\ldots , Y_d, Z_1,\ldots , Z_d\geq 1\) satisfying

\begin{align} \label{eq:xiyizi_1} X^{\alpha -\varepsilon }\ll _{\varepsilon } \prod _{j=1}^{d}X_j\leq 2 X^{\alpha +\varepsilon },\quad X^{\beta -\varepsilon }\ll _{\varepsilon }\prod _{j=1}^{d}Y_j\leq 2 X^{\beta +\varepsilon },\quad X^{\gamma -\varepsilon }\ll _{\varepsilon }\prod _{j=1}^{d}Z_j\leq 2 X^{\gamma +\varepsilon } \end{align}

and

\begin{align} \label{eq:xiyizi_2} \prod _{j=1}^d X_j^j \leq X, \quad \prod _{j=1}^d Y_j^j\leq X,\quad X^{1-\varepsilon ^2}\ll _{\varepsilon } \prod _{j=1}^d Z_j^j\leq X \end{align}

and pairwise coprime integers \(1\leq c_1,c_2,c_3\leq X^{\varepsilon }\), such that

\begin{align*} S^*_{\alpha ,\beta ,\gamma }(X) \ll _{\varepsilon } X^{\varepsilon }B_d(\mathbf{c},\mathbf{X},\mathbf{Y},\mathbf{Z}). \end{align*}
Proof of Proposition 2.6

We may assume that \(X\) is large enough in terms of \(\varepsilon \), since otherwise the claim is trivial. Let \((a,b,c)\) be a triple counted by \(S^*_{\alpha ,\beta ,\gamma }(X)\). Apply Lemma 2.5 (with \(\varepsilon ^2/2\) in place of \(\varepsilon \)) to each of \(a,b,c\) to obtain factorisations of the form

\begin{align*} a=c_1\prod _{j\leq d}x_j^j,\quad b=c_2\prod _{j\leq d}y_j^j,\quad c=c_3\prod _{j\leq d}z_j^j, \end{align*}

where \(d=\lfloor 10\varepsilon ^{-4}\rfloor \) and \(1\leq c_1,c_2,c_3\leq X^{\varepsilon ^2/4}\). Since \((a,b,c)\) is counted by \(S^*_{\alpha ,\beta ,\gamma }(X)\), we have \(\gcd (a,b,c)=1\) and \(a+b=c\), so \(a,b,c\) are pairwise coprime. Hence, all the \(3d+3\) numbers \(x_i,y_i,z_i,c_1,c_2,c_3\) are pairwise coprime. Note also that by the properties of the factorisation given by Lemma 2.5, we have

\begin{align*} & X^{-\varepsilon /2}\prod _{j\leq d}x_j\leq \operatorname{rad}(a)\leq X^{\varepsilon /2}\prod _{j\leq d}x_j,\quad X^{-\varepsilon /2}\prod _{j\leq d}y_j\leq \operatorname{rad}(b)\leq X^{\varepsilon /2}\prod _{j\leq d}y_j,\\ & X^{-\varepsilon /2}\prod _{j\leq d}z_j\leq \operatorname{rad}(c)\leq X^{\varepsilon /2}\prod _{j\leq d}z_j. \end{align*}

Since \(\operatorname{rad}(a)\sim X^{\alpha },\operatorname{rad}(b)\sim X^{\beta },\operatorname{rad}(c)\sim X^{\gamma }\) for all triples under consideration, this implies

\begin{align*} X^{\alpha -\varepsilon }\leq \prod _{j\leq d}x_j\leq X^{\alpha +\varepsilon },\quad X^{\beta -\varepsilon }\leq \prod _{j\leq d}y_j\leq X^{\beta +\varepsilon },\quad X^{\gamma -\varepsilon }\leq \prod _{j\leq d}z_j\leq X^{\gamma +\varepsilon }. \end{align*}

By dyadic decomposition, we can now find some \(X_i, Y_i,Z_i\) such that ?? and ?? hold, and such that

\[ S^*_{\alpha ,\beta ,\gamma }(X)\ll _{\varepsilon } (\log X)^{3d} \sum _{\substack {\mathbf{c}\in \mathbb {N}^3\\ c_1,c_2,c_3 \leq X^{\varepsilon /4}}}B_d(\mathbf{c},\mathbf{X},\mathbf{Y},\mathbf{Z}). \]

Now the claim follows from the pigeonhole principle.