arXiv2026
We construct counterexamples, along dyadic lacunary sequences, to two problems of Erdős on the almost-everywhere behavior of sums $\sum_{j\le N}f(n_jx)$ on the circle. The main device is a dyadic spike block: rare positive spikes, each counted many times by a run of consecutive dilates, make the lacunary averages large, while a deterministic lower bound on every block prevents the rest of the series from cancelling them. We obtain a mean-zero $f\in\bigcap_{1\le q<\infty}L^q(\mathbb T)$ and a sequence $n_j=2^{m_j}$ with $n_{j+1}/n_j\ge2$ such that $\|f-S_Nf\|_2\ll(\log\log N)^{-1/2}$ and $\limsup_{N\to\infty}\frac1N\sum_{j\le N}f(n_jx)=+\infty$ for almost every $x$. This answers negatively a question of Erdős on whether the condition $\|f-S_Nf\|_2\ll(\log\log\log N)^{-C}$ suffices for almost-everywhere convergence, and shows that Matsuyama's condition $\|f-S_Nf\|_2\ll(\log\log N)^{-c}$, $c>1/2$, cannot be relaxed to $c=1/2$. A second choice of parameters gives, for every $2\le p<\infty$, functions $f\in L^p(\mathbb T)$ with $\limsup_{N\to\infty}\sum_{j\le N}f(n_jx)/(N(\log N)^{1/p-\varepsilon})=+\infty$ for every $\varepsilon>0$, almost everywhere. This is sharp up to $\varepsilon$, and the case $p=2$ answers negatively a second question of Erdős. Finally, a bounded variant in the spirit of Marstrand's disproof of Khintchine's conjecture gives sets of arbitrarily small measure whose lacunary visiting frequencies have limsup $1$ almost everywhere.