We study principal component analysis (PCA), where given a dataset in
$\mathbb{R}^d$ from a distribution, the task is to find a unit vector $v$ that
approximately maximizes the variance of the distribution after being projected
along $v$. Despite being a classical task, standard estima