Performance limits of stochastic sub-gradient learning, Part I: Single-agent case