1: \begin{abstract}

2: 

3: In the present paper, we propose a novel sparse signal recovery algorithm called the Trainable Iterative Soft Thresholding Algorithm (TISTA).

4: The proposed algorithm consists of two estimation units: a linear estimation unit  

5: and a minimum mean squared error (MMSE) estimator-based shrinkage unit.

6: The estimated error variance required in the MMSE shrinkage 

7: unit is precisely estimated from a tentative estimate of

8: the original signal.

9: The remarkable feature of the proposed scheme is 

10: that TISTA includes adjustable variables that control step size

11: and the error variance for the MMSE shrinkage.

12: The variables are adjusted by standard deep learning techniques. 

13: The number of trainable variables of TISTA is { nearly} equal to the number of iteration rounds 

14: and is much smaller than that of known learnable sparse signal recovery algorithms.

15: This feature leads to highly stable and fast training processes of TISTA.

16: Computer experiments show that TISTA is applicable to various classes of sensing matrices 

17: such as  Gaussian matrices, binary matrices, and matrices with large condition numbers.

18: Numerical results also demonstrate that, in many cases, TISTA provides significantly faster convergence 

19: than AMP and the Learned {ISTA} and { also outperforms OAMP in the NMSE performance.}

20: \end{abstract}