Spectral theorem: eigenvalue decomposition for symmetric matrices

Spectral theorem

We can decompose any symmetric matrix A \in {\bf S}^n with the symmetric eigenvalue decomposition (SED)

A = \sum\limits_{i=1}^n \lambda_i u_i u_i^T = U\Lambda U^T, \Lambda = {\bf diag}(\lambda_1, \lambda_2, cdots, \lambda_n).

where the matrix of U:=[u_1, \cdots, u_n] is orthogonal (that is, U^TU=UU^T=I_n), and contains the eigenvectors of A, while the diagonal matrix \Lambda contains the eigenvalues of A.

Proof: The proof is by induction on the size n of the matrix A. The result is trivial for n=1. Now let n>1 and assume the result is true for any matrix of size n-1.

Consider the function of A, t\rightarrow p(t) = \det (tI-A). From the basic properties of the determinant, it is a polynomial of degree n, called the characteristic polynomial of A. By the fundamental theorem of algebra, any polynomial of degree n has n (possibly not distinct) complex roots; these are called the eigenvalues of A. We denote these eigenvalues by \lambda_1, \cdots, \lambda_n.

If \lambda is an eigenvalue of A, that is, \det (\lambda I-A)=0, then \lambda I- A must be non-invertible (see here). This means that there exists a non-zero real vector u such that Au = \lambda u. We can always normalize u so that u^Tu=1. Thus, \lambda = u^TAu is real. That is, the eigenvalues of a symmetric matrix are always real.

Now consider the eigenvalue \lambda_1 and an associated eigenvector u_1. Using the Gram-Schmidt orthogonalization procedure, we can compute a n \times (n-1) matrix V_1 such that [u_1, V_1] is orthogonal. By induction, we can write the (n-1) \times (n-1) symmetric matrix V_1^T A V_1 as, Q_1^T A Q_1 where Q_1 is a (n-1) \times (n-1) matrix of eigenvectors, and \Lambda_1 = {\bf diag}(\lambda_2, \cdots, \lambda_n) are the n-1 eigenvalues of V_1^T A V_1. Finally, we define the n \times (n-1) matrix U_1:= V_1 Q_1. By construction the matrix U:= [u_1, U_1] is orthogonal.

We have

U^T A U=\left(\begin{array}{c} u_1^T \\ U_1^T \end{array}\right) A\left(\begin{array}{cc} u_1 & U_1 \end{array}\right)=\left(\begin{array}{cc} u_1^T A u_1 & u_1^T A U_1 \\ U_1^T A u_1 & U_1^T A U_1 \end{array}\right)=\left(\begin{array}{cc} \lambda_1 & 0 \\ 0 & \Lambda_1 \end{array}\right),

where we have exploited the fact that U_1^T A u_1 = \lambda_1 U_1^T u_1 = 0, and U_1^T A U_1 =\Lambda_1.

We have exhibited an orthogonal n \times n matrix U such that U^TAU is diagonal. This proves the theorem.

License

Hyper-Textbook: Optimization Models and Applications Copyright © by L. El Ghaoui. All Rights Reserved.

Share This Book