1: \begin{abstract}

2:   It is well-known that the reparameterisation gradient estimator, which exhibits low variance in practice, is biased for non-differentiable models. This may compromise correctness of gradient-based optimisation methods such as stochastic gradient descent (SGD).

3:   We introduce a simple syntactic framework to define non-differentiable functions piecewisely and present a systematic approach to obtain smoothings for which the reparameterisation gradient estimator is unbiased.

4:   Our main contribution is a novel variant of SGD, \emph{Diagonalisation Stochastic Gradient Descent}, which progressively enhances the accuracy of the smoothed approximation \emph{during} optimisation, and we prove convergence to stationary points of the \emph{unsmoothed} (original) objective.

5:   Our empirical evaluation reveals benefits over the state of the art: our approach is simple, fast, stable and attains orders of magnitude reduction in work-normalised variance.

6: \end{abstract}

7: