A diagonal matrix is the simplest kind of matrix. Here are some facts that justify such a strong statement: For diagonal matrices, (1) the eigenvalues are the entries along the diagonal, (2) the standard basis vectors are eigenvectors, (3) the determinant is the product of diagonal entries, (4) the rank is the number of non-zero entries on the diagonal, and (5) all diagonal matrices commute with each other (that is, , if and are diagonal). In some statistical applications, diagonal matrices correspond to uncorrelated variables, which is the easiest situation to study. A diagonal adjacency matrix in graph theory corresponds to a completely disconnected graph.
If is a linear transformation, it is desirable to choose a basis of for which the matrix is diagonal. This is not always possible (hence the reason for most of Chapter 6), but when is it possible, we will call diagonalizable.
Theorem 4.52 shows how the matrix of a linear transformation changes when we change the basis. We would like to think of the resulting matrices as being closely related to each other in some way; that is the purpose of the following terminology.
Let . We say that is similar to if there exists an invertible matrix such that
A matrix is called diagonalizable if there exists a such that is a diagonal matrix.
[Aside: In fact, any invertible matrix can be thought of as a change of basis matrix for an appropriate choice of bases; so if two matrices are similar to each other, then they can always be visualized as representing the same linear transformation, but with a different choice of basis.]
An important special case of Theorem 4.52 is when one of the bases consists of eigenvectors, as has been the case in several of the examples we have already seen. We summarize this case as follows, and omit the proof (compare with Theorem 4.11):
Let be a square matrix. Let be the standard basis of .
If is a basis of eigenvectors of , and , then is diagonal.
If is diagonal, then is a basis of eigenvectors of .
Note that are the column vectors of the matrix . The matrix in part (i) is called a diagonalization of .
Let , and prove that is a basis of eigenvectors, and hence find a diagonalization of .
Solution: One checks that these basis vectors are eigenvectors as follows:
So let be the standard basis, and then we have
By Theorem 4.54, a diagonalization is
Let . Prove is not diagonalizable ().
Using the basis , and from Exercise 4.51,
Verify that consists of eigenvectors of .
Verify, by matrix multiplication, that is a diagonal matrix.
Verify, by matrix multiplication, that .
[End of Exercise]
Prove that similar matrices always have the same eigenvalues.
[End of Exercise]