KKT optimisation - condition of inequality constraint being zero

433 Views Asked by At

For example, given the following:

Minimise $$ f(x_1, x_2) $$

Subject to $$ h(x_1, x_2) = 0 $$

$$ g(x_1, x_2) \leq 0 $$

The KKT conditions are written out as

$$ l(x, \lambda, \mu) = f(x_1, x_2) + \lambda h(x_1, x_2) + \mu g(x_1, x_2) $$

And one of the conditions is that

$$ \mu \cdot g(x_1, x_2) = 0 $$

I'm not sure why this condition is necessary though, I think that it has something to do with the constraint being active / inactive and/or there being slack? I'm not sure what the reasoning is though.

1

There are 1 best solutions below

0
On

The condition $\mu \cdot g(x_1,x_2) = 0$ is a complementary slackness condition. It says that either the dual multiplier $\mu$ is 0, or there is no slack in the $g(x_1,x_2) \le 0$ constraint.

The dual values give the change in the optimal objective function value if the right-hand side of the constraint changes. So the intuition behind the complementary slackness condition is that if there is slack in the constraint ($g(x_1,x_2) \lneqq 0$), then changing the RHS would have no effect on the optimal objective function value, hence $\mu=0$. And, if $\mu \ne 0$, then a change in RHS would lead to a change in the objective function, so the constraint must be tight ($g(x_1,x_2) = 0$).