On a question of Erdős and Graham (Q5915541)

\textit{P. Erdős} and \textit{R. L. Graham} [Old and new problems and results in combinatorial number theory, Monographie No. 28. Genève: L'Enseignement Mathématique, Université de Genève (1980; Zbl 0434.10001), page 103] conjectured that for any \(\varepsilon >0\) there exists \(k(\varepsilon)\) such that for any \(c\) and any prime \(p\) there exist \( k \leq k(\varepsilon)\) pairwise distinct integers \(1 \leq x_i \leq p^{\varepsilon}\), \(i=1, \ldots , k\) such that \(\sum_{i=1}^k \frac{1}{x_i} \equiv c \bmod p.\) \textit{E. S. Croot} [Mathematika 46, 359--372 (1999; Zbl 1033.11014)] proved that \(k(\varepsilon,p)= (\log p)^{3+\varepsilon}\) is admissible, even for prime values of \(x_i\). Using results of Karatsuba in a version of \textit{J. Friedlander} and \textit{H. Iwaniec} [Lond. Math. Soc. Lect. Note Ser. 247, 85-93 (1997; Zbl 0910.11036)] the present author shows that \(k(\varepsilon)=4 \varepsilon^{-3}+O(\varepsilon^{-2})\) is admissible, where the \(x_i\) are products of two primes, from a suitable dyadic interval \((X,2X)\).