This little result came up when proving the convergence of a stochastic gradient algorithm and I want to write it down to remember it after discussions with Matt Johnson and Alex Tank.
Let \(a_1, a_2, \ldots\) be a positive sequence of numbers. If \(\sum_{n=1}^\infty \frac{1}{n} a_n < \infty\), then \(\liminf_{n\rightarrow \infty} a_n = 0\).
The proof is by contrapositive. Assume that \(\liminf_{n \rightarrow \infty} = c \neq 0\). Then by the \(\epsilon\) characterization of the limit inferior we have that \(\forall \epsilon > 0\), \(\exists N\) such that for \(n > N\), \(a_n > c - \epsilon\). This implies that
$$
\begin{aligned}
\sum_{n=1}^\infty \frac{1}{n} a_n & \geq \sum_{n'=N}^\infty \frac{1}{n'}a_{n'} \\
& > (c - \epsilon) \sum_{n'=N}^\infty \frac{1}{n'} = \infty.
\end{aligned}
$$
But we assumed that the original sum converged, implying that \(\liminf_{n\rightarrow \infty} a_n = 0\).