1: \begin{abstract}

2: In this paper, we propose a time-frequency analysis method to obtain instantaneous frequencies and the corresponding decomposition by solving an optimization problem.  In this optimization problem,

3: the basis to decompose the signal is not known {\it a priori}. Instead, it is adapted to the signal

4: and is determined as part of the optimization problem.

5: In this sense, this optimization problem can be seen as a dictionary learning problem.

6: This dictionary learning problem is solved by using the Augmented Lagrangian Multiplier method (ALM) iteratively.

7:  We further accelerate the convergence of the ALM method in each iteration by using the fast wavelet transform.

8: We apply our method to decompose several signals, including signals with poor scale separation, signals with outliers and polluted by noise and a real signal. The results show that this method can give accurate recovery

9: of both the instantaneous frequencies and the intrinsic mode functions.

10: \end{abstract}