Stationarity in KKT condition

228 Views Asked by At

I was learning the KKT conditions theory and in the notes the author was trying to prove the necessity of KKT conditions with bare hands. But I found a phenomenon which is a bit weird and I seek for some suggestions here.


Suppose $\hat{x}$ solves $$\min_{x\in \mathbb R^m} F(x) \quad\text{subject to } \quad G(x)\le 0\,,$$ where $F,G$ are from $\mathbb R^m$ to $\mathbb R$, continuously differentiable.

Define $\mathcal{L}(x,\mu)=F(x)+\mu G(x)$, where $\mu\in \mathbb R$.

Consider the case that $G(\hat{x})<0$, i.e. $\hat{x} \in \{G(x)<0 \}$, which is open in $\mathbb R^m$.

Then $\nabla_x F(\hat{x})=0$. Recall that the stationarity condition in KKT is, there exists $\hat{\mu}$ such that $\nabla_x F(\hat{x})+\hat{\mu} \nabla_x G(\hat{x})=0$. Therefore we need to have that $\hat{\mu} \nabla_x G(\hat{x})=0$. If we choose $\hat{\mu}=0$, then we are done. But then $\mathcal{L}(x,\hat{\mu})$ reduces to $F(x)$.

It seems like introducing $\mathcal{L(x,\mathcal{\mu})}$ is somehow meaningless. What is the meaning of introducing KKT in this specific case? Thanks for posts and comments.

1

There are 1 best solutions below

0
On BEST ANSWER

In general, you don't know before you find them, which, if any, of the minimizers $\ \hat{x}\ $ will satisfy the condition $\ G(\hat{x})<0\ $, and if $\ \hat{x}\ $ is a local minimizer for which $\ G(\hat{x})=0\ $, then it will not necessarily be true that $\ \nabla F(\hat{x})=0\ $. Thus, if you want to be sure of finding all the minimizers, you will normally need to find all solutions of the full KKT conditions, which are \begin{align} \nabla F(\hat{x})+\hat{\mu}\nabla G(\hat{x})&=0\\ \hat{\mu}G(\hat{x})&=0\\ \hat{\mu}&\ge0 \end{align} for this problem. Note that the last condition is also necessary. If $\ \nabla F(\hat{x})+\hat{\mu}\nabla G(\hat{x})=0\ $, $\ \nabla F(\hat{x})\ne0\ $ and $\ \hat{\mu}<0\ $, then $\ \hat{x}\ $ cannot be a minimizer.