This is a optimization problem.

$$ \arg\min\limits_{x}|X-B|_2^2+\lambda|X|_1 $$

Apparently,the lasso term isn't derivable.

But if we discuss it by category, we can get the following results.

For the i-th demension,we have

$$ x_i^*=\begin{cases}B_i+\lambda/2,&B_i<-\lambda/2\\0,&|B_i|<\lambda/2\\B_i-\lambda/2,&B_i>\lambda/2 \end{cases} $$


This form is called "Soft Thresholding",and it is a closed form solution.

In practice, we use subgradient descent algorithm or near proximal algorithm to solve this problem and get an approximate solution.

I also see the soft thresholding form in the process by using above two algorithms.But these two methods are iterative processes step by step and then get an approximate solution.

I don't know why since there is a closed form solution we still need two iterative algorithms to find an approximate solution

I don't know either what is the relationship between subgradient descent algorithm and near proximal algorithm——they all get the soft thresholding form


1 Answer 1


The closed-solution needs some conditions.For example,the $x$ is a vector or the coefficient matrix is ones matrix.In other condition,you must seek an approximate solution.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .