1: \begin{abstract}   

2:     

3:     {\noindent In batch learning, stability together with existence

4:       and uniqueness of the solution corresponds to well-posedness of

5:       Empirical Risk Minimization (ERM) methods; recently, it was

6:       proved that {\it $\cv$ stability} is necessary and sufficient

7:       for generalization and consistency of ERM (\cite{PRMN04}). In this

8:       note, we introduce {\it $\cvon$ stability}, which plays a similar role in 

9:      online learning. We show that stochastic gradient descent (SDG) with 

10:       the usual hypotheses is $\cvon$ stable and we then discuss the 

11:       implications of $\cvon$ stability for convergence of SGD.}

12: 

13: \end{abstract}