1: \begin{abstract}

2: Anderson acceleration (AA) is a popular method for accelerating fixed-point iterations, but may suffer from instability and stagnation.

3: We propose a globalization method for AA to improve stability and achieve unified global and local convergence.

4: Unlike existing AA globalization approaches that rely on safeguarding operations and might hinder fast local convergence, we adopt a nonmonotone trust-region framework and introduce an adaptive quadratic regularization together with a tailored acceptance mechanism. 

5: We prove global convergence and show that our algorithm attains the same local convergence as AA under appropriate assumptions. The effectiveness of our method is demonstrated in several numerical experiments.

6: \end{abstract}

7: