New approach to Farkas' theorem of the alternative (Q2313263)

The famous Farkas lemma says that, if \(a_1,\dots,a_m,b\in \mathbb{R}^n\) satisfy the condition (F): \(\langle a_i,x\rangle\le 0\), \(i=1,\dots,m\, \Rightarrow \langle b,x\rangle\le 0\) for all \(x\in \mathbb{R}^n\), then \(b\in \operatorname{cone}(a_1,\dots,a_m)\), i.e., \(b=\sum_{i=1}^m\lambda_i a_i\) for some nonnegative numbers \(\lambda_i\), \(i=1,\dots,m\). The authors propose a new proof of this theorem. The basic idea is to show that the functional \(\Phi(x)=\langle b,x\rangle+\sum_{i=1}^m\big(\langle a_i,x\rangle_{-}\big)^2\), \(x\in\mathbb{R}^n\), attains its minimum at some point \(x_0\in\mathbb{R}^n\), provided that \(a_1,\dots,a_m\) and \(b\) satisfy Farkas' condition (F). Then the gradient of \(\Phi\) at \(x_0 \) is 0, \(\operatorname{grad}\Phi(x_0)=b-\sum_{i=1}^m\alpha_ia_i=0\), where \(\alpha_i=-2\langle a_i,x_0\rangle_{-}\geq 0\), that is, \(b\in \operatorname{cone}(a,\dots,a_m)\). Here, for \(\alpha\in\mathbb{R}\), \(\alpha_{-}=\alpha\) if \(\alpha<0\) and \(= 0\) otherwise. The paper also contains some comments on various proofs of the Farkas lemma.