1

I would like to solve the following problem:

$$ \beta = \arg\min_{\beta;\|\beta\|_1 \leq1} \|X\beta-y\|_2^2 $$

which happens to be the constrained formulation of the Lasso, considering only parameter vectors that lie inside the $\ell 1$-unit-ball.

However, Matlab uses the penalized formulation, i.e.

$$ \beta = \arg \min_{\beta} \|X\beta-y\|_2^2 + \lambda\|\beta\|_1 $$

So my question is the following: is there a way to choose $\lambda$, or a way to transform the formulation, so that I can represent the first problem statement (with its constraint) using the second formulation? In other words, can I choose a value of $\lambda$ such that the parameter vector stays within the $\ell 1$-unit-ball?

cangrejo
  • 2,121
  • 13
  • 22
  • 1
    For the constraint $t$ in the $\|\beta\|_1 \leq t$ formulation, there is a always an equivalent $\lambda$ formulation. See [two](http://stats.stackexchange.com/questions/17781/derivation-of-closed-form-lasso-solution) [related](http://stats.stackexchange.com/questions/90648/kkt-versus-unconstrained-formulation-of-lasso-regression) questions. – Affine May 24 '15 at 03:04

0 Answers0