next up previous
Next: Implementation Details Up: The Decomposition Algorithm Previous: Shrinking

Termination Criterion

Given what has been said in the section on shrinking, if we can always have (29) and (30) during the resolution of a subproblem, a reasonable termination criterion is to verify that the $\lambda$ estimated by (31) and (32) verifies the conditions (24)-(28) with a given precision $\epsilon_{end}$.

Thus, we simply verify that

\textrm{for} \ i \ \textrm{such that} \ 0 <...
...r}) + \delta_i \lambda^{eq} \leq \epsilon_{end} \\

with $\delta_i = 1$ for $1 \leq i \leq l$, $\delta_i = -1$ for $(l+1) \leq i \leq 2l$ and

\begin{displaymath}\boldsymbol{\beta} = \left( \begin{array}{c}
\boldsymbol{\alpha} \\
- \boldsymbol{\alpha}^{\star}

Journal of Machine Learning Research