$$ \newcommand{\qed}{\tag*{$\square$}} \newcommand{\span}{\operatorname{span}} \newcommand{\dim}{\operatorname{dim}} \newcommand{\rank}{\operatorname{rank}} \newcommand{\norm}[1]{\|#1\|} \newcommand{\grad}{\nabla} \newcommand{\prox}[1]{\operatorname{prox}_{#1}} \newcommand{\inner}[2]{\langle{#1}, {#2}\rangle} \newcommand{\mat}[1]{\mathcal{M}[#1]} \newcommand{\null}[1]{\operatorname{null} \left(#1\right)} \newcommand{\range}[1]{\operatorname{range} \left(#1\right)} \newcommand{\rowvec}[1]{\begin{bmatrix} #1 \end{bmatrix}^T} \newcommand{\Reals}{\mathbf{R}} \newcommand{\RR}{\mathbf{R}} \newcommand{\Complex}{\mathbf{C}} \newcommand{\Field}{\mathbf{F}} \newcommand{\Pb}{\operatorname{Pr}} \newcommand{\E}[1]{\operatorname{E}[#1]} \newcommand{\Var}[1]{\operatorname{Var}[#1]} \newcommand{\argmin}[2]{\underset{#1}{\operatorname{argmin}} {#2}} \newcommand{\optmin}[3]{ \begin{align*} & \underset{#1}{\text{minimize}} & & #2 \\ & \text{subject to} & & #3 \end{align*} } \newcommand{\optmax}[3]{ \begin{align*} & \underset{#1}{\text{maximize}} & & #2 \\ & \text{subject to} & & #3 \end{align*} } \newcommand{\optfind}[2]{ \begin{align*} & {\text{find}} & & #1 \\ & \text{subject to} & & #2 \end{align*} } $$
The Loewner order is a partial order on the set of positive semidefinite symmetric matrices. For two positive semidefinite matrices and , we write to denote that is positive semidefinite (and symmetric) and to denote that is positive definite (and symmetric). The set of positive semidefinite symmetric matrices is denoted , and the set of positive definite symmetric matrices is denoted .
Below are some facts about this partial order.
For the following fact, let , with eigenvalues .
(1): . This holds because .
(2): For and , if and only if . This is clear, for for some such that (which is possible because C is full rank).
Let and . Then
(3): if and only if . Furthermore, if , if and only if . We prove the former statement below; the latter is proved in an analogous fashion by starting with .
where the last inequality follows because is full rank.
Let and . Then
(4): if and only if . One way to prove this is to use (2):
where the second equivalence used the fact that for any two matrices and , the eigenvalues of equal those of .
Since for every one can associate an ellipsoid , it follows from (4) that if and only if .
Let be partitioned as
where . If is invertible, then the matrix
is called the Schur complement of in . The following facts are useful:
(5): if and only if and .
(6): If , then if and only if .
These facts can be derived by considering the minimization of the following quadratic, with variable :
where
If , then the optimal point is and the optimal value is .
To see why (5) is true, first observe that implies by the definition of positive definite. So implies , and implies the optimal value of the quadratic minimization problem is ; since , the optimal value must be positive (no matter the value of the paremeter ). Hence . Conversely, if and , then is strictly greater than for all , that is, .
To see why (6) is true, first assume that . Because and the minimum of the quadratic in question is , it must be the case that , that is, . Conversely, if , then for all , that is, .
Boyd, S., & Vandenberghe, L. (2004). Convex optimization. Cambridge university press.