Regularization
Regularize version of Lasso and basis-pursuit denoise can be obtained by augmenting A with the weighted identity matrix, and augmenting b with a vector of all zero. For convenience, SPGL1 also supports direct regularization, which changes the Lasso formulation to
\mathop{\mathrm{minimize}}_{x}\quad {\textstyle\frac{1}{2}}\Vert Ax-b\Vert_2^2 + {\textstyle\frac{\mu}{2}}\Vert x\Vert_2^2\quad\mathrm{subject\ to}\quad \Vert x\Vert_p \leq \tau
and basis-pursuit denoise to
\mathop{\mathrm{minimize}}_{x}\quad \Vert x\Vert_p\quad \mathrm{subject\ to}\quad \left\Vert \left[\begin{array}{c}A\\ \sqrt{\mu}I\end{array}\right]x-\left[\begin{array}{c}b\\0\end{array}\right]\right\Vert_2\leq \sigma.
The \mu parameter can be specified in the options as options.mu
or as parameter mu
.