Joeyonng
  • Notebook
  • Pages
  • About
  • Backyard
  1. Linear Algebra
  2. 14  Normal and Hermitian Matrices
  • Welcome
  • Notations and Facts
  • Linear Algebra
    • 1  Fields and Spaces
    • 2  Vectors and Matrices
    • 3  Span and Linear Independence
    • 4  Basis and Dimension
    • 5  Linear Map and Rank
    • 6  Inner Product and Norm
    • 7  Orthogonality and Unitary Matrix
    • 8  Complementary Subspaces and Projection
    • 9  Orthogonal Complement and Decomposition
    • 10  SVD and Pseudoinverse
    • 11  Orthogonal and Affine Projection
    • 12  Determinants and Eigensystems
    • 13  Similarity and Diagonalization
    • 14  Normal and Hermitian Matrices
    • 15  Positive Definite Matrices
  • Calculus
    • 16  Derivatives
    • 17  Chain rule
  • Probability and Statistics
    • 18  Probability
    • 19  Random Variables
    • 20  Expectation
    • 21  Common Distributions
    • 22  Moment Generating Function
    • 23  Concentration Inequalities I
    • 24  Convergence
    • 25  Limit Theorems
    • 26  Maximum Likelihood Estimation
    • 27  Bayesian Estimation
    • 28  Expectation-maximization
    • 29  Concentration Inequalities II
  • Learning Theory
    • 30  Statistical Learning
    • 31  Bayesian Classifier
    • 32  Effective Class Size
    • 33  Empirical Risk Minimization
    • 34  Uniform Convergence
    • 35  PAC Learning
    • 36  Rademacher Complexity
  • Machine Learning
    • 37  Linear Discriminant
    • 38  Perceptron
    • 39  Logistic Regression
    • 40  Multi-layer Perceptron
    • 41  Boosting
    • 42  Support Vector Machine
    • 43  Decision Tree
    • 44  Principle Component Analysis
  • Deep Learning
    • 45  Transformer

Table of contents

  • Normal Matrices
    • Orthogonal diagonalization
  • Symmetric (Hermitian) matrices
  • Rayleigh quotient
  1. Linear Algebra
  2. 14  Normal and Hermitian Matrices

14  Normal and Hermitian Matrices

Normal Matrices

Definition 14.1 The square matrix \mathbf{A} \in \mathbb{C}^{n \times n} is a normal matrix if and only if

\mathbf{A}^{H} \mathbf{A} = \mathbf{A} \mathbf{A}^{H}.

Theorem 14.1 Diagonal matrices are normal.

Proof

Consider \mathbf{D} \in \mathbb{C}^{n \times n} as a diagonal matrix with d_{1}, \dots, d_{n} in its diagonal.

\mathbf{D}^{H} \mathbf{D} = \sum_{i=1}^{n} d_{i}^{2} = \mathbf{D} \mathbf{D}^{H}.

Theorem 14.2 Unitary (orthogonal) matrices are normal.

Proof

According to the property of unitary matrices, a matrix \mathbf{U} is unitary if and only if

\mathbf{U}^{-1} \mathbf{U} = \mathbf{U} \mathbf{U}^{-1} = \mathbf{I}.

Thus, unitary matrices are normal.

Theorem 14.3 Unitary similarity preserves normality. That is, if \mathbf{A} is a normal matrix and is unitarily similar to \mathbf{B}

\mathbf{U}^{-1} \mathbf{A} \mathbf{U} = \mathbf{B}

\mathbf{U}^{H} \mathbf{A} \mathbf{U} = \mathbf{B},

then \mathbf{B} is also a normal matrix.

Proof

The goal is to prove

\mathbf{B}^{H} \mathbf{B} = \mathbf{B} \mathbf{B}^{H}.

First we expand \mathbf{B}^{H} \mathbf{B} to have

\begin{aligned} \mathbf{B}^{H} \mathbf{B} & = (\mathbf{U}^{H} \mathbf{A} \mathbf{U})^{H} (\mathbf{U}^{H} \mathbf{A} \mathbf{U}) \\ & = \mathbf{U} \mathbf{A}^{H} \mathbf{U}^{H} \mathbf{U}^{H} \mathbf{A} \mathbf{U} \\ & = \mathbf{U}^{H} \mathbf{A}^{H} \mathbf{U} \mathbf{U}^{H} \mathbf{A} \mathbf{U} & [TODO] \\ & = \mathbf{U}^{H} \mathbf{A}^{H} \mathbf{I} \mathbf{A} \mathbf{U} & [\mathbf{U} \mathbf{U}^{H} = \mathbf{U} \mathbf{U}^{-1} = \mathbf{I}] \\ & = \mathbf{U}^{H} \mathbf{A}^{H} \mathbf{A} \mathbf{U}. \end{aligned}

Since \mathbf{A} is a normal matrix,

\mathbf{A} \mathbf{A}^{H} = \mathbf{A}^{H} \mathbf{A}.

Continue from the derivation above,

\begin{aligned} \mathbf{B}^{H} \mathbf{B} & = \mathbf{U}^{H} \mathbf{A} \mathbf{A}^{H} \mathbf{U} \\ & = \mathbf{U}^{H} \mathbf{A}^{H} \mathbf{A} \mathbf{U} \\ & = \mathbf{U}^{H} \mathbf{A}^{H} \mathbf{I} \mathbf{A} \mathbf{U} \\ & = \mathbf{U}^{H} \mathbf{A}^{H} \mathbf{U} \mathbf{U}^{H} \mathbf{A} \mathbf{U} \\ & = \mathbf{U} \mathbf{A}^{H} \mathbf{U}^{H} \mathbf{U}^{H} \mathbf{A} \mathbf{U} & [TODO] \\ & = (\mathbf{U}^{H} \mathbf{A}^{H} \mathbf{U})^{H} (\mathbf{U}^{H} \mathbf{A} \mathbf{U}) \\ & = \mathbf{B} \mathbf{B}^{H}. \end{aligned}

Orthogonal diagonalization

Theorem 14.4 A matrix \mathbf{A} \in \mathbb{C}^{n \times n} is normal if and only if \mathbf{A} is orthogonally (unitarily) similar to a diagonal matrix

\mathbf{A}^{H} \mathbf{A} = \mathbf{A} \mathbf{A}^{H} \iff \mathbf{A} = \mathbf{U} \mathbf{\Lambda} \mathbf{U}^{-1}.

where \mathbf{U} is an orthogonal (unitary) matrix and \mathbf{\Lambda} is a diagonal matrix.

Proof

We first prove that

\mathbf{A}^{H} \mathbf{A} = \mathbf{A} \mathbf{A}^{H} \Longrightarrow \mathbf{A} = \mathbf{U} \mathbf{\Lambda} \mathbf{U}^{-1}

TODO

Then we prove that

\mathbf{A} = \mathbf{U} \mathbf{\Lambda} \mathbf{U}^{-1} \Longrightarrow \mathbf{A}^{H} \mathbf{A} = \mathbf{A} \mathbf{A}^{H}.

According to the property of unitary matrices, unitary similarity preserves normality.

Also according to property of unitary matrices, all diagonal matrices are normal.

Thus, \mathbf{A} is normal because \mathbf{A} is unitarily similar to a diagonal matrix, which is always normal.

Since the columns of \mathbf{U} are eigenvectors of \mathbf{A} and are orthonormal to each other the columns of \mathbf{U} must be a complete orthonormal set of eigenvectors for \mathbf{A}, and the diagonal entries of \mathbf{\Lambda} are the associated eigenvalues.

Symmetric (Hermitian) matrices

Definition 14.2 A square matrix \mathbf{A} \in \mathbb{C}^{n \times n} is symmetric (Hermitian) if and only if

\mathbf{A}^{H} = \mathbf{A},

which implies a symmetric (Hermitian) matrix is a normal matrix.

Theorem 14.5 All eigenvalues of symmetric (Hermitian) matrices are real.

Proof

We will prove it for the Hermitian matrix and symmetric matrix is a special case of the hermitian matrix.

Suppose (\lambda, \mathbf{v}) is a eigenpair for the Hermitian matrix \mathbf{A}.

\mathbf{A} \mathbf{v} = \lambda \mathbf{v}.

Multiplying both sides by \mathbf{v}^{H} on the left to get,

\begin{aligned} \mathbf{A} \mathbf{v} & = \lambda \mathbf{v} \\ \mathbf{v}^{H} \mathbf{A} \mathbf{v} & = \lambda \mathbf{v}^{H} \mathbf{v} \\ \mathbf{v}^{H} \mathbf{A} \mathbf{v} & = \lambda \lVert \mathbf{v} \rVert^{2}. \\ \end{aligned}

Alternatively we can take the transpose conjugate of both sides, and then multiply both sides by \mathbf{v} on the right to get,

\begin{aligned} \mathbf{A} \mathbf{v} & = \lambda \mathbf{v} \\ (\mathbf{A} \mathbf{v})^{H} & = (\lambda \mathbf{v})^{H} \\ \mathbf{v}^{H} \mathbf{A}^{H} & = \lambda^{*} \mathbf{v}^{H} \\ \mathbf{v}^{H} \mathbf{A}^{H} \mathbf{v} & = \lambda^{*} \mathbf{v}^{H} \mathbf{v} \\ \mathbf{v}^{H} \mathbf{A}^{H} \mathbf{v} & = \lambda^{*} \lVert \mathbf{v} \rVert^{2} \\ \end{aligned}

Since \mathbf{A} is a hermitian matrix

\begin{aligned} \mathbf{v}^{H} \mathbf{A} \mathbf{v} & = \mathbf{v}^{H} \mathbf{A}^{H} \mathbf{v} \\ \lambda \lVert \mathbf{v} \rVert^{2} & = \lambda^{*} \lVert \mathbf{v} \rVert^{2} \\ \lambda & = \lambda^{*}, \end{aligned}

which means \lambda is a real number.

Rayleigh quotient

Given a Hermitian matrix \mathbf{M} \in \mathbb{C}^{n \times n}, the Rayleigh quotient is a function R_{\mathbf{M}} (\mathbf{x}): \mathbb{C}^{n} \setminus \{ 0 \} \rightarrow \mathbb{R}

R_{\mathbf{M}} (\mathbf{x}) = \frac{ \mathbf{x}^{H} \mathbf{M} \mathbf{x} }{ \mathbf{x}^{H} \mathbf{x} }

that takes a nonzero vector \mathbf{x} and returns a real number.

Since the Hermitian matrix \mathbf{M} has all real eigenvalues, they can be ordered. Suppose \lambda_{1}, \dots, \lambda_{n} is the eigenvalues in descending orders.

Then, given a Hermitian matrix, its Rayleigh quotient is upper bounded and lower bounded by maximum and minimum eigenvalues of \mathbf{M} respectively,

\lambda_{1} \geq R_{\mathbf{M}} (\mathbf{x}) \geq \lambda_{n}.

That is,

\lambda_{1} = \max_{\mathbf{x} \neq 0} \frac{ \mathbf{x}^{H} \mathbf{M} \mathbf{x} }{ \mathbf{x}^{H} \mathbf{x} },

\lambda_{n} = \min_{\mathbf{x} \neq 0} \frac{ \mathbf{x}^{H} \mathbf{M} \mathbf{x} }{ \mathbf{x}^{H} \mathbf{x} }.

Proof

Since \mathbf{M} is a Hermitian matrix, we can expand it using unitary diagonalization:

\begin{aligned} \frac{ \mathbf{x}^{H} \mathbf{M} \mathbf{x} }{ \mathbf{x}^{H} \mathbf{x} } & = \frac{ \mathbf{x}^{H} \mathbf{U}^{H} \mathbf{\Lambda} \mathbf{U} \mathbf{x} }{ \mathbf{x}^{H} \mathbf{x} } \\ & = \frac{ \mathbf{y}^{H} \mathbf{\Lambda} \mathbf{y} }{ \mathbf{x}^{H} \mathbf{x} } & [\mathbf{y} = \mathbf{U} \mathbf{x}]. \end{aligned}

Since \mathbf{U} is a unitary matrix, according to the property of the unitary matrix,

\mathbf{y}^{H} \mathbf{y} = \mathbf{x}^{H} \mathbf{x}.

Thus,

\begin{aligned} \frac{ \mathbf{y}^{H} \mathbf{\Lambda} \mathbf{y} }{ \mathbf{x}^{H} \mathbf{x} } & = \frac{ \mathbf{y}^{H} \mathbf{\Lambda} \mathbf{y} }{ \mathbf{y}^{H} \mathbf{y} } \\ & = \frac{ \sum_{i=1}^{n} \lambda_{i} y_{i}^{2} }{ \sum_{i=1}^{n} y_{i}^{2} }. \end{aligned}

Since \lambda_{1} \geq \lambda_{i} \geq \lambda_{n}, \forall i = 1, \dots, n,

\lambda_{1} = \lambda_{1} \frac{ \sum_{i=1}^{n} y_{i}^{2} }{ \sum_{i=1}^{n} y_{i}^{2} } = \frac{ \sum_{i=1}^{n} \lambda_{1} y_{i}^{2} }{ \sum_{i=1}^{n} y_{i}^{2} } \geq \frac{ \sum_{i=1}^{n} \lambda_{i} y_{i}^{2} }{ \sum_{i=1}^{n} y_{i}^{2} },

\lambda_{n} = \lambda_{n} \frac{ \sum_{i=1}^{n} y_{i}^{2} }{ \sum_{i=1}^{n} y_{i}^{2} } = \frac{ \sum_{i=1}^{n} \lambda_{n} y_{i}^{2} }{ \sum_{i=1}^{n} y_{i}^{2} } \leq \frac{ \sum_{i=1}^{n} \lambda_{i} y_{i}^{2} }{ \sum_{i=1}^{n} y_{i}^{2} }.

13  Similarity and Diagonalization
15  Positive Definite Matrices