Linear matrix inequality

(Difference between revisions)
 Revision as of 21:17, 4 December 2011 (edit) (→Convexity of the LMI constraint)← Previous diff Revision as of 13:49, 21 September 2016 (edit) (undo)Next diff → Line 1: Line 1: In convex optimization, a '''linear matrix inequality (LMI)''' is an expression of the form In convex optimization, a '''linear matrix inequality (LMI)''' is an expression of the form - : $LMI(y):=A_0+y_1A_1+y_2A_2+\dots+y_m A_m\succeq0\,$ + : $LMI(y):=A_0+y_1A_1+y_2A_2+\ldots+y_m A_m\succeq0\,$ where where * $y=[y_i\,,~i\!=\!1\ldots m]$ is a real vector, * $y=[y_i\,,~i\!=\!1\ldots m]$ is a real vector,

Revision as of 13:49, 21 September 2016

In convex optimization, a linear matrix inequality (LMI) is an expression of the form

$LaTeX: LMI(y):=A_0+y_1A_1+y_2A_2+\ldots+y_m A_m\succeq0\,$

where

• $LaTeX: y=[y_i\,,~i\!=\!1\ldots m]$ is a real vector,
• $LaTeX: A_0\,, A_1\,, A_2\,,\ldots\,A_m$ are symmetric matrices in the subspace of $LaTeX: n\times n$ symmetric matrices $LaTeX: \mathbb{S}^n$,
• $LaTeX: B\succeq0$ is a generalized inequality meaning $LaTeX: B$ is a positive semidefinite matrix belonging to the positive semidefinite cone $LaTeX: \mathbb{S}_+$ in the subspace of symmetric matrices $LaTeX: \mathbb{S}$.

This linear matrix inequality specifies a convex constraint on y.

Convexity of the LMI constraint

$LaTeX: LMI(y)\succeq 0$ is a convex constraint on y which means membership to a dual (convex) cone as we now explain: (Dattorro, Example 2.13.5.1.1)

Consider a peculiar vertex-description for a convex cone defined over the positive semidefinite cone

(instead of the more common nonnegative orthant, $LaTeX: x\succeq0$):

for $LaTeX: X\!\in\mathbb{S}^n$ given $LaTeX: \,A_j\!\in\mathbb{S}^n$, $LaTeX: \,j\!=\!1\ldots m$

$LaTeX: \begin{array}{ll}\mathcal{K} \!\!&=\left\{\left[\begin{array}{c}\langle A_1\,,\,X^{}\rangle\\\vdots\\\langle A_m\;,\,X^{}\rangle\end{array}\right]|~X\!\succeq_{\!}0\right\}\subseteq_{}\mathbb{R}^m\\\\ &=\left\{\left[\begin{array}{c}\textrm{svec}(A_1)^T\\\vdots\\\textrm{svec}(A_m)^T\end{array}\right]\textrm{svec}X~|~X\!\succeq_{\!}0\right\}\\\\ &:=\;\{A\,\textrm{svec}X~|~X\!\succeq_{\!}0_{}\} \end{array}$

where

• $LaTeX: A\!\in_{}\!\mathbb{R}^{m\times n(n+1)/2}$,
• symmetric vectorization svec is a stacking of columns defined in (Dattorro, ch.2.2.2.1),
• $LaTeX: A_0=\mathbf{0}$ is assumed without loss of generality.

$LaTeX: \mathcal{K}$ is a convex cone because

$LaTeX: A\,\textrm{svec}{X_{{\rm p}_1}}_{\,},_{_{}}A\,\textrm{svec}{X_{{\rm p}_2}}\!\in\mathcal{K}~\Rightarrow~ A(\zeta_{\,}\textrm{svec}{X_{{\rm p}_1\!}}+_{}\xi_{\,}\textrm{svec}{X_{{\rm p}_2}})\in_{}\mathcal{K} \textrm{~~for\,all~\,}\zeta_{\,},\xi\geq0$

since a nonnegatively weighted sum of positive semidefinite matrices must be positive semidefinite.

Now consider the (closed convex) dual cone:

$LaTeX: \begin{array}{rl}\mathcal{K}^* \!\!\!&=_{}\left\{_{}y~|~\langle z\,,\,y_{}\rangle\geq_{}0\,~\textrm{for\,all}~\,z\!\in_{_{}\!}\mathcal{K}_{}\right\}\subseteq_{}\mathbb{R}^m\\ &=_{}\left\{_{}y~|~\langle z\,,\,y_{}\rangle\geq_{}0\,~\textrm{for\,all}~\,z_{\!}=_{\!}A\,\textrm{svec}X\,,~X\succeq0_{}\right\}\\ &=_{}\left\{_{}y~|~\langle A\,\textrm{svec}X\,,~y_{}\rangle\geq_{}0\,~\textrm{for\,all}~\,X\!\succeq_{_{}\!}0_{}\right\}\\ &=\left\{y~|~\langle\textrm{svec}X\,,\,A^{T\!}y\rangle\geq_{}0\;~\textrm{for\,all}~\,X\!\succeq_{\!}0\right\}\\ &=\left\{y~|~\textrm{svec}^{-1}(A^{T\!}y)\succeq_{}0\right\} \end{array}$

that follows from Fejer's dual generalized inequalities for the positive semidefinite cone:

• $LaTeX: Y\succeq0~\Leftrightarrow~\langle Y\,,\,X\rangle\geq0\;~\textrm{for\,all}~\,X\succeq0$

This leads directly to an equally peculiar halfspace-description

$LaTeX: \mathcal{K}^*=\{y\!\in\mathbb{R}^m~|\,\sum\limits_{j=1}^my_jA_j\succeq_{}0_{}\}$

The summation inequality with respect to the positive semidefinite cone is known as a linear matrix inequality.

LMI Geometry

Although matrix $LaTeX: \,A\,$ is finite-dimensional, $LaTeX: \mathcal{K}$ is generally not a polyhedral cone (unless $LaTeX: \,m\,$ equals 1 or 2) simply because $LaTeX: \,X\!\in\mathbb{S}_+^n\,.$

Relative interior of $LaTeX: \mathcal{K}$ may always be expressed $LaTeX: \textrm{rel\,int}\,\mathcal{K}=\{A\,\textrm{svec}X~|~X\!\succ_{\!}0_{}\}.$

Provided the $LaTeX: \,A_j$ matrices are linearly independent, then $LaTeX: \textrm{rel\,int}\,\mathcal{K}=\textrm{int}\,\mathcal{K}$

meaning, cone $LaTeX: \mathcal{K}$ interior is nonempty; implying, dual cone $LaTeX: \mathcal{K}^*$ is pointed (Dattorro, ch.2).

If matrix $LaTeX: \,A\,$ has no nullspace, then $LaTeX: \,A\,\textrm{svec}X\,$ is an isomorphism in $LaTeX: \,X\,$ between the positive semidefinite cone $LaTeX: \mathbb{S}_+^n$ and range $LaTeX: \,\mathcal{R}(A)\,$ of matrix $LaTeX: \,A.$

That is sufficient for convex cone $LaTeX: \,\mathcal{K}\,$ to be closed, and necessary to have relative boundary $LaTeX: \textrm{rel}\,\partial^{}\mathcal{K}=\{A\,\textrm{svec}X~|~X\!\succeq_{\!}0\,,~X\!\not\succ_{\!}0_{}\}.$

Relative interior of the dual cone may always be expressed $LaTeX: \textrm{rel\,int}\,\mathcal{K}^*=\{y\!\in_{}\!\mathbb{R}^m~|\,\sum\limits_{j=1}^my_jA_j\succ_{}0_{}\}.$

When the $LaTeX: A_j$ matrices are linearly independent, function $LaTeX: \,g(y)_{\!}:=_{_{}\!}\sum y_jA_j\,$ is a linear bijection on $LaTeX: \mathbb{R}^m.$

Inverse image of the positive semidefinite cone under $LaTeX: \,g(y)\,$ must therefore have dimension equal to $LaTeX: \dim\!\left(\mathcal{R}(A^{\rm T})_{}\!\cap\mbox{svec}\,\mathbb{S}_+^{_{}n}\right)$

and relative boundary $LaTeX: \textrm{rel\,}\partial^{}\mathcal{K}^*=\{y\!\in_{}\!\mathbb{R}^m~|\,\sum\limits_{j=1}^my_jA_j\succeq_{}0\,,~\sum\limits_{j=1}^my_jA_j\not\succ_{}0_{}\}.$

When this dimension is $LaTeX: \,m\,$, the dual cone interior is nonempty $LaTeX: \textrm{rel\,int}\,\mathcal{K}^*=\textrm{int}\,\mathcal{K}^*$

and closure of convex cone $LaTeX: \mathcal{K}$ is pointed.

Applications

There are efficient numerical methods to determine whether an LMI is feasible (i.e., whether there exists a vector $LaTeX: y$ such that $LaTeX: LMI(y)\succeq0$ ), or to solve a convex optimization problem with LMI constraints. Many optimization problems in control theory, system identification, and signal processing can be formulated using LMIs. The prototypical primal and dual semidefinite program are optimizations of a real linear function respectively subject to the primal and dual convex cones governing this LMI.