1: \begin{abstract}

2: Our aim is to study the backward problem, i.e. recover the initial data from the terminal observation, of the 

3: subdiffusion with time dependent coefficients. First of all, by using the smoothing property of solution operators

4: and a perturbation argument of freezing the diffusion coefficients, we show a stability estimate in Sobolev spaces, under some smallness/largeness condition on the terminal time.

5: Moreover, in case of noisy observation, we apply a quasi-boundary value method to regularize the problem and then show the convergence of the regularization scheme.

6: Finally, to numerically reconstruct the initial data, we propose a completely discrete scheme by applying the finite element method in space and backward Euler 

7: convolution quadrature in time. 

8: An \textsl{a priori} error estimate is then established. The proof is heavily built on a 

9: perturbation argument dealing with time dependent coefficients

10: and some nonstandard error estimates for the direct problem. 

11: The error estimate gives a useful guide for balancing discretization parameters, regularization parameter 

12: and noise level. Some numerical experiments are presented to illustrate our theoretical results.

13: 

14: 

15: 

16:  

17: 

18: %{\bf AMS subject classifications 2010:}

19: 

20: \end{abstract}

21: