1: \begin{abstract}

2:         We consider the  problem of demixing a sequence of 

3:         source signals from the sum of noisy bilinear measurements. It is a generalized mathematical model for blind demixing with blind deconvolution, which is prevalent across the areas of dictionary learning, image processing, and communications. However, state-of-the-art convex methods for  blind demixing via semidefinite programming are computationally infeasible for large-scale problems. Although the existing nonconvex algorithms are able to address the scaling issue, they normally  require proper regularization  to establish optimality guarantees. The additional regularization  yields tedious

4: algorithmic parameters and pessimistic convergence rates with conservative step sizes.

5: To address the limitations of existing methods, we thus develop a provable nonconvex demixing procedure via Wirtinger flow, much like vanilla gradient descent, to harness the  benefits of {{regularization free, fast convergence rate with aggressive step size and computational optimality guarantees}}. This is achieved by exploiting the benign geometry of the blind demixing problem, thereby revealing that Wirtinger flow enforces the regularization-free iterates in the region of strong convexity and qualified level of smoothness where the step size can be chosen aggressively. 

6: 

7: \begin{IEEEkeywords}

8: Blind demixing, blind deconvolution, bilinear measurements, nonconvex optimization, Wirtinger flow, regularization-free, statistical and computational guarantees.    

9: \end{IEEEkeywords}

10: \end{abstract}

11: