<div dir="ltr">Hi,<div><br></div><div>Suppose I would like to maximize</div><div><br></div><div>max_{X} [ Tr(C X) ]_{+}</div><div><br></div><div>where [z]_{+} = max(z, 0), the hinge loss function. I can replace the hinge loss with a slack variable and get</div>
<div><br></div><div>max_{X} \xi</div><div>s.t.</div><div>\xi >= Tr(C X)<br></div><div><br></div><div>But turning the inequality constraint into an equality constraint would require another slack variable, i.e.</div><div>
<br></div><div>max_{X} \xi</div><div>s.t.</div><div>\xi - \eta - Tr(C X) = 0</div><div><br></div><div>Is that the right approach? The fact that I now have two slack variables, and use only one in the cost function strikes me as potentially problematic.</div>
<div><br></div><div>Thanks already,</div><div><br></div><div>nick</div><div><br></div></div>