Skip to main content
Joseph Li

LP Duality Practice

To prepare for my linear programming midterm, I looked at these problems from Bertsismas and Tsitsiklas' Introduction to Linear Optimization.

Problem 4.27 #

For a given matrix A\mathbf{A}, show that the following are equivalent: (a)\text{(a)} Every vector x\mathbf{x} such that Ax0\mathbf{A}\mathbf{x}\geq\mathbf{0} and x0\mathbf{x}\geq\mathbf{0} must satisfy x1=0x_1=0. (b)\text{(b)} There exists some p\mathbf{p} such that Ap0\mathbf{A}^\top \mathbf{p} \leq \mathbf{0}, p0\mathbf{p}\geq\mathbf{0}, and A1p<0\mathbf{A}_1^\top\mathbf{p}<0 where A1\mathbf{A}_1 is the first column of A\mathbf{A}.

Solution #

Consider the following primal-dual LP pair: (Dual) Minimize0psubject toA1p1<0Aip0i1p0\begin{align*} \text{Minimize}\qquad\qquad &\mathbf{0}^\top \mathbf{p} \newline \text{subject to}\qquad \mathbf{A}_1^\top \mathbf{p} &\leq -1 < 0 \newline \mathbf{A}_i^\top \mathbf{p} &\leq \mathbf{0}\qquad \forall i \neq 1 \newline \mathbf{p} &\geq 0\end{align*} (Primal) Maximizey1subject toAy0y0\begin{align*} \text{Maximize}\qquad\qquad &-y_1 \newline \text{subject to}\qquad \mathbf{A}\mathbf{y} &\leq \mathbf{0} \newline \mathbf{y} &\leq \mathbf{0}\end{align*}

(b) is true    (a) is true(b)\text{ is true} \implies (a)\text{ is true} #

If (b)(b) is true, then the dual is feasible with an optimal value of the objective function equal to zero. Then for any x0\mathbf{x}\leq\mathbf{0} satisfying Ax0\mathbf{A}\mathbf{x}\geq\mathbf{0} and x0\mathbf{x}\geq\mathbf{0}, we have y=x\mathbf{y}=-\mathbf{x} satisfying Ay0\mathbf{A}\mathbf{y}\leq\mathbf{0} and y0\mathbf{y}\leq\mathbf{0} where x1=y10x_1=-y_1\leq0 and y10y_1\leq0 (by weak duality). Therefore, y1=0y_1=0 and thus x1=0x_1=0.

(b) is false    (a) is false(b)\text{ is false} \implies (a)\text{ is false} #

If (b)(b) is false, then the dual is infeasible. This implies that the primal must either be infeasible or unbounded. Clearly, the primal is feasible with y=0\mathbf{y}=\mathbf{0}, so the primal must be unbounded. Then for any MRM\in\mathbb{R}, there exists y\mathbf{y} such that Ay0\mathbf{A}\mathbf{y}\leq\mathbf{0}, y0\mathbf{y}\leq\mathbf{0}, and y1>M-y_1>M. Choosing M=0M=0 and setting x=y\mathbf{x}=-\mathbf{y} yields Ax0\mathbf{A}\mathbf{x}\geq\mathbf{0}, x0\mathbf{x}\geq\mathbf{0}, and x10x_1 \neq 0, which shows (a)(a) is false.

Problem 4.28 #

Let a\mathbf{a} and a1,,am\mathbf{a}_1,\dots,\mathbf{a}_m be vectors in Rn\mathbb{R}^n. Show the following are equivalent. (a)\text{(a)} For all x0\mathbf{x}\geq\mathbf{0}, we have axmaxiaix\mathbf{a}^\top \mathbf{x}\leq \max_i \mathbf{a}^\top_i \mathbf{x}. (b)\text{(b)} There exist nonnegative coefficients λi\lambda_i that sum to 11 such that ai=1mλiai\mathbf{a}\leq \sum^{m} _{i=1} \lambda_i \mathbf{a}_i.

Solution #

Consider the primal-dual pair: (Primal) Maximizeax+ysubject toaix+y0i=1,,mx0\begin{align*}\text{Maximize}\qquad\qquad\mathbf{a}^\top\mathbf{x} + y& \newline \text{subject to}\qquad\qquad \mathbf{a}_i^\top \mathbf{x} + y &\leq 0 \qquad i=1,\dots,m \newline \mathbf{x}\geq\mathbf{0}\end{align*}

(Dual) Minimize0Λsubject to(a1am)Λa1Λ=1Λ0\begin{align*}\text{Minimize} \qquad\qquad\qquad\qquad \mathbf{0}^\top\Lambda& \newline \text{subject to}\qquad \begin{pmatrix} \mathbf{a}_1 & \dots & \mathbf{a}_m \end{pmatrix} \Lambda &\geq \mathbf{a} \newline \mathbf{1}^\top\Lambda &= 1 \newline \Lambda &\geq 0\end{align*}

(b) is true    (a) is true(b)\text{ is true} \implies (a)\text{ is true} #

If (b)(b) is true, then the dual is feasible with optimal solution 0. Then the primal is feasible. Let x0\mathbf{x}\geq\mathbf{0}. Then choose y=maxiaix\mathbf{y} = -\max_i \mathbf{a}_i^\top \mathbf{x}. We get axmaxiaix0\mathbf{a}^\top \mathbf{x} - \max_i \mathbf{a}_i^\top x \leq \mathbf{0}, so (a)(a) is true.

(b) is false    (a) is false(b)\text{ is false} \implies (a)\text{ is false} #

If (b)(b) is false, then the dual is infeasible, so primal is unbounded or infeasible. The primal is clearly feasible with x=0\mathbf{x}=\mathbf{0} and y=0y=0, so it must be unbounded. Then there exists x0\mathbf{x}\geq\mathbf{0}, yRy\in\mathbb{R}, such that aixy\mathbf{a}_i^\top \mathbf{x} \leq -y for all i=1,,mi=1,\dots,m, and ax+y>0\mathbf{a}^\top \mathbf{x}+y > 0. That is, ax>y>=aiTx\mathbf{a}^\top \mathbf{x}>-y>=a_i^{T}x for all i=1,,mi=1,\dots,m, so ax>maxiaix\mathbf{a}^\top\mathbf{x} > \max_i \mathbf{a}_i^\top \mathbf{x}.