# soft thresholding derivation with two terms

I am trying to minimize the following function $$argmin_x||x-y||^2_2 + ||x-z||_2^2 + \lambda || \frac{x- w}{c}||_1$$

I have been doing it for while but i am not sure how to do it. I have already been through few of the examples for simple cases such as this one (Derivation of soft thresholding operator).

Taking the derivative w.r.t $x$ is given by

$$\frac{x{-w}}{\left|c\right|\left|x{-w}\right|}+4x{-2z}{-2y}$$, i would like to know that how i would be able to go from here to the soft thresholding function described in the above link.

Any help regarding this would be quite useful

• Why not just perform a variable substitution: $\tilde{x} = (x - w)/c$ so that your objective becomes $$\|c\tilde{x}+w-y\|_2^2+\|c\tilde{x}+w-z\|_2^2+\lambda\|\tilde{x}\|_1?$$ – Michael Grant Oct 27 '15 at 14:44
• Do @MichaelGrant's variable substitution, and then complete the square in $c\tilde{x}$ – dohmatob Oct 28 '15 at 8:09

After applying @MichaelGrant 's variable substitution $u=(x-w)/c$ you end up with

$$\arg \min_u \Vert cu+w-y\Vert^2 + \Vert cu+w-z\Vert^2 + \lambda \Vert u \Vert_1$$

which basically looks like the standard soft-thresholding $$\arg \min_u (1/2)\Vert u-p\Vert^2 + \lambda \Vert u\Vert_1$$

after some manipulation of the two other terms i.e.

$$\Vert cu+w-y\Vert^2 + \Vert cu+w-z\Vert^2 = 2 c^2 \left\Vert u-\frac{2w-y-z}{c} \right\Vert^2 + f(w,y,z)$$

and it becomes

$$\begin{eqnarray} \arg \min_u 2 c^2 \left\Vert u-\frac{2w-y-z}{c} \right\Vert^2 + \lambda \Vert u\Vert_1 & = & 4c^2 \arg \min_u \frac{1}{2}\left\Vert u-\frac{2w-y-z}{c} \right\Vert^2 + \frac{\lambda}{4c^2} \Vert u\Vert_1 \\ & = & 4c^2 S_{\lambda / 4c^2} \left( \frac{2w-y-z}{c} \right) \end{eqnarray}$$

where $S_\rho$ is the soft-thresholding function with parameter $\rho$.