Diagonalization#
An \(n \times n\) matrix \(A\) is diagonalizable if and only if \(A\) has \(n\) linearly independent eigenvectors. If a matrix \(A\) is real and symmetric then it is diagonalizable, the eigenvalues are real numbers and the eigenvectors (for distinct eigenvalues) are orthogonal.
Eigenvalues and Eigenvectors#
An eigenvalue of a matrix \(A\) is a number \(\lambda\) such that
for some nonzero vector \(\boldsymbol{v}\). The vector \(\boldsymbol{v}\) is called an eigenvector for the eigenvalue \(\lambda\). Eigenvalues of a real matrix may be real or complex numbers.
If \(\lambda\) is an eigenvalue of \(A\) with eigenvector \(\boldsymbol{v}\) then \((A - \lambda I)\boldsymbol{v} = \boldsymbol{0}\) which implies that \(A - \lambda I\) is not invertible and therefore \(\det(A - \lambda I) = 0\). This suggests that to find eigenvalues and eigenvectors of \(A\) we should:
Find \(\lambda\) such that \(\det(A - \lambda I) = 0\).
Find solutions of the linear system \((A - \lambda I)\boldsymbol{v} = \boldsymbol{0}\).
This works when \(A\) is a small matrix and we have done this in previous linear algebra courses. However, this is impractical when \(A\) is a large matrix. For example, if \(A\) is \(5 \times 5\), then \(\det(A - \lambda I) = 0\) is a polynomial equation of degree 5 and there is no formula for the roots. We’ll see better algorithms for computing eigenvalues in later sections.
Let \(A\) be an \(n \times n\) matrix. The characteristic polynomial of \(A\) is
where \(I\) is the identity matrix. The polynomial \(c_A(x)\) has degree \(n\) and the roots of \(c_A(x)\) are the eigenvalues of \(A\).
Diagonalization#
A matrix \(A\) is diagonalizable if there exists an invertible matrix \(P\) and a diagonal matrix \(D\) such that \(A = PD P^{-1}\).
If \(A\) is diagonalizable with \(A = PDP^{-1}\) then the diagonal entries of \(D\) are eigenvalues of \(A\) and the columns of \(P\) are the corresponding eigenvectors.
Proof
Let \(\boldsymbol{v}_1,\dots, \boldsymbol{v}_n\) be the columns of \(P\) and let \(\lambda_1,\dots,\lambda_n\) be the diagonal entries of \(D\)
Matrix multiplication \(AP = PD\) yields the equation
Therefore \(A \boldsymbol{v}_i = \lambda_i \boldsymbol{v}_i\) for each \(i=1,\dots,n\).
If \(A\) has distinct eigenvalues then \(A\) is diagonalizable.
Proof
Let \(\lambda_1,\dots,\lambda_n\) be the distinct eigenvalues of \(A\). That is, \(\lambda_i \not= \lambda_j\) for \(i \ne j\). Each \(\lambda_i\) has a corresponding eigenvector \(\boldsymbol{v}_i\). Construct \(P\) with eigenvectors in the columns
and construct \(D\) with the eigenvalues in the diagonal entries
Then \(A = PDP^{-1}\) by construction.
By the Fundamental Theorem of Algebra, the characteristic polynomial of a matrix \(A\) factors as a product
where \(\lambda_1, \dots, \lambda_k\) are the distinct eigenvalues of \(A\). The algebraic multiplicity of \(\lambda_i\) is the power \(m_i\) in the factored characteristic polynomial. In other words, the algebraic multiplicity of \(\lambda_i\) is the number of times \(\lambda_i\) occurs as a root of the characteristic polynomial \(c_A(x)\).
Let \(\lambda\) be an eigenvalue of \(A\). The geometric multiplicity of \(\lambda\) is the number of linearly independent eigenvectors corresponding to \(\lambda\). In other words, the geometric multiplicity is the dimension of the eigenspace for \(\lambda\)
Not every matrix is diagonalizable. For example, consider the matrix
Then \(c_A(x) = (x - 3)^2\) and there is only one eigenvalue \(\lambda = 3\) and it has algebraic multiplicity 2. Solving the equation \((A - 3I)\boldsymbol{v} = \boldsymbol{0}\) yields only one independent solution
and so \(\lambda = 3\) has geometric multiplicity 1. Therefore \(A\) does not have enough eigenvectors to be diagonalizable.
A matrix \(A\) is diagonalizable if and only if, for all eiganvalues \(\lambda\), the algebraic multiplicity of \(\lambda\) equals the geometric multiplicity of \(\lambda\).
Spectral Theorem#
All eigenvalues of a symmetric matrix are real numbers.
Proof
Let \(\lambda\) be an eigenvalue of a symmetric matrix \(A\) with eigenvector \(\boldsymbol{v}\). Compute the complex inner product \(\langle \boldsymbol{v} , A \boldsymbol{v} \rangle = \boldsymbol{v}^T \, \overline{A \boldsymbol{v}}\) in two different ways. First, compute
Since \(A\) is real and symmetric, we have \(\overline{A}^T = A\) and we compute
Since \(\| \boldsymbol{v} \| \not= 0\) we have \(\lambda = \overline{\lambda}\) and therefore \(\lambda\) is a real number.
Let \(A\) be a symmetric matrix and let \(\lambda_1\) and \(\lambda_2\) be distinct real eigenvalues of \(A\) with eigenvectors \(\boldsymbol{v}_1\) and \(\boldsymbol{v}_2\) respectively. Then \(\boldsymbol{v}_1\) and \(\boldsymbol{v}_2\) are orthogonal.
Proof
Since \(\lambda_1\) and \(\lambda_2\) are real eigenvalues we may assume \(\boldsymbol{v}_1\) and \(\boldsymbol{v}_2\) are real vectors. Compute \(\langle A \boldsymbol{v}_1 , \boldsymbol{v}_2 \rangle\) in two different ways. First, compute
Now compute
Therefore
since \(\lambda_1 - \lambda_2 \not = 0\) because the eigenvalues are distinct.
Let \(A\) be a symmetric matrix. Then there exists an orthogonal matrix \(P\) and diagonal matrix \(D\) such that \(A = PDP^T\). In other words, \(A\) is orthogonally diagonalizable.
Exercises#
Determine whether the statement is True or False.
Let \(\boldsymbol{v}_1 , \boldsymbol{v}_2 \in \mathbb{R}^2\) be linearly independent vectors. Let \(\lambda_1 , \lambda_2\) be real numbers. There exists a unique \(2 \times 2\) matrix \(A\) with eigenvalues \(\lambda_1 , \lambda_2\) and corresponding eigenvectors \(\boldsymbol{v}_1 , \boldsymbol{v}_2\).
Suppose \(A\) and \(B\) are symmetric \(n \times n\) matrices. The eigenvectors of \(AB\) corresponding to distinct eigenvalues are orthogonal.
Solution
True
False
Let \(A\) be a \(m \times n\) matrix. Determine whether the statement is True or False.
If \(\lambda\) is an eigenvalue of \(AA^T\) then \(\lambda\) is a real number.
If \(\boldsymbol{v}_1 , \boldsymbol{v}_2\) are eigenvectors of \(AA^T\) for distinct eigenvalues \(\lambda_1 , \lambda_2\) then \(\langle \boldsymbol{v}_1 , \boldsymbol{v}_1 \rangle = 0\).
Solution
True
True
Let \(U \subset \mathbb{R}^n\) be a subspace with \(\mathrm{dim}(U) = m\) such that \(0 < m < n\), and let \(P\) be the orthogonal projection matrix onto \(U\). Determine the characteristic polynomial of \(P\).
Solution
Let \(\boldsymbol{u} \in \mathbb{R}^n\) be a nonzero vector and let
be the corresponding elementary reflector. Determine the characteristic polynomial of \(H\).
Solution
Let \(\lambda\) be a eigenvalue of an invertible matrix \(A\). Determine whether the statement is True or False.
\(\lambda^{-1}\) is an eigenvalue of \(A^{-1}\)
\(\lambda\) is an eigenvalue of \(A^T\)
\(\lambda^2\) is an eigenvalue of \(AA^T\)
\(\lambda\) is an eigenvalue of \(BAB^{-1}\) for any invertible matrix \(B\)
\(\lambda \ne 0\)
Solution
True
True
False
True
True
Suppose \( A \) is a symmetric \( 3 \times 3 \) matrix with distinct eigenvalues \( \lambda_1 , \lambda_2 , \lambda_3 \) and eigenvectors
Find an eigenvector \(\boldsymbol{v}_3\) for eigenvalue \(\lambda_3\).