#effective-step-size

Plain SGD works well with $\eta = 0.1$. You switch to momentum, $\mathbf{v}_t = \beta\mathbf{v}_{t-1} + \mathbf{g}_t$, $\mathbf{x}_t = \mathbf{x}_{t-1} - \eta\mathbf{v}_t$, with $\beta = 0.9$. Where the gradient is roughly constant, which $\eta$ gives the same step size as before?