1: \begin{abstract}

2:       The decomposition of tensors into simple rank-1 terms is key in a variety

3:       of applications in signal processing, data analysis and machine

4:       learning. While this canonical polyadic decomposition (CPD) is unique

5:       under mild conditions, including prior knowledge such as nonnegativity can

6:       facilitate interpretation of the components. Inspired by the effectiveness

7:       and efficiency of Gauss--Newton (GN) for unconstrained CPD, we derive a

8:       proximal, semismooth GN type algorithm for nonnegative tensor

9:       factorization. If the algorithm converges to the global optimum, we show

10:       that $Q$-quadratic convergence can be obtained in the exact case. Global

11:       convergence is achieved via backtracking on the forward-backward envelope

12:       function. The $Q$-quadratic convergence is verified experimentally, and we

13:       illustrate that using the GN step significantly reduces number of

14:       (expensive) gradient computations compared to proximal gradient descent.

15: 	\end{abstract}

16: