Square root theorem
Statement and proof
\(\mc{M}^{n \times n}\) denotes the set of \(n \times n\) matrices. \(\mc{P}^{n \times n} \subset \mc{P}^{n \times n} \subset \mc{M}^{n \times n}\) denotes the set of symmetric positive semidefinite real valued \(n \times n\) matrices.
Definition 1 Positive semidefinite matrices
A matrix \(\bsf{C} \in \mc{M}^{n \times n}\) is positive semidefinite if \[ \bsf{v} \cdot \bsf{C} \bsf{v} \geq 0 \] for all \(\bsf{v} \in \mbb{R}^n\).
Example 1 Check that the following matrix is positive semidefinite: \[ \begin{bmatrix} 1 & 1 & 1 \\ 1 & 1 & 1 \\ 1 & 1 & 1 \end{bmatrix} \]
Definition 2 Square root of a matrix
The matrix \(\bsf{U}\) is called the square root of the matrix \(\bsf{C}\) if \[ \bsf{U} \bsf{U} = \bsf{U}^{2} = \bsf{C}. \]
Theorem 1 Equivalent characterizations of positive semidefinite matrices
Let \(\bsf{C} \in \mc{P}^{n \times n}\). Then the following are true:
1. \(\bsf{C}\) is symmetric. 1. All the eigenvalues of \(\bsf{C}\) are nonnegative.
1. \(\bsf{C}\) has a positive definite square root.
1. There exists a matrix \(\bsf{F} \in \mc{M}^{n \times n}\) such that \(\bsf{C} = \bsf{F}^{\msf{T}} \bsf{F}\).
Proof 1. Proof of equivalent characterizations
To be filled.
Now, we can state the square root theorem.
Theorem 2 Square root theorem
Let \(\bsf{C} \in \mc{P}^{n \times n}\). Then \(\bsf{C}\) has a unique positive definite square root \(\bsf{U}\).
Proof 2. Proof of the square-root theorem
To be filled.
Example 2
Spectral and square root theorems
Let \(\bsf{C} \in \mc{P}^{n \times n}\). Since \(\mc{P}^{n \times n} \subset \mc{S}^{n \times n}\), the spectral theorem can be invoked and \(\bsf{C}\) can be written in terms of its spectral decomposition \[ \bsf{C} = \sum_{i = 1}^{n} \lambda_{i}^{2} \, {^{i}\bsf{N}} \otimes {^{i}\bsf{N}} \] where \(\lambda_{i}^{2}\) is the \(i\)-th eigenvalues of \(\bsf{C}\) and \(^{n}\bsf{N}\) is the \(n\)-th eigenvector of \(\bsf{C}\). Next, since \(\bsf{C}\) is positive semidefinite, the square root theorem can be invoked. \(^{n}\bsf{N}\) form an orthoromal basis, the square root is given simply by taking the square root of the scalar coefficients: \[ \sqrt{\bsf{C}} = \sum_{i = 1}^{n} \lambda_{i} \, {^{i}\bsf{N}} \otimes {^{i}\bsf{N}} \] Finally, the square root is unique, so we can designate it with a unique name: \[ \bsf{U} = \sum_{i = 1}^{n} \lambda_{i} \, {^{i}\bsf{N}} \otimes {^{i}\bsf{N}} \]
Note, the square root \(\bsf{U}\) is written as the sum of the scalars \(\lambda_{i}\) multiplying the dyad product of orthonormal vectors. This implies that this is the spectral decomposition of \(\bsf{U}\) and therefore, \(\lambda_{i}\) are the eigenvalues of \(\bsf{U}\) and \({^{i}\bsf{N}}\) are the eigenvectors of \(\bsf{U}\). Note that \(\bsf{C}\) and \(\bsf{U}\) share the same eigenvectors but not the same eigenvalues, although they are related.