Linear Algebra › Section E · Eigenvalues and eigenvectors › Topic E.3
Diagonalization
Many matrices are diagonal matrices in disguise: change to a basis of eigenvectors and all that is left is a list of stretch factors. Writing a matrix this way makes its powers, and so its long-run behaviour, easy to compute.
Why diagonal matrices are easy
Computing \(A^{10}\) for a general matrix takes nine matrix products. For a diagonal matrix it takes no real work: \(D^k\) is diagonal, with each diagonal entry raised to the power \(k\). Now suppose \(A = PDP^{-1}\) for some invertible \(P\) and diagonal \(D\). Then
because the inner \(P^{-1}P\) collapses to \(I\). Repeating the argument gives \(A^k = PD^kP^{-1}\) for every positive integer \(k\). The whole cost of a high power becomes one power of each diagonal entry plus two matrix products.
The diagonalisation theorem
A square matrix \(A\) is diagonalisable if \(A = PDP^{-1}\) for some invertible \(P\) and diagonal \(D\); in the language of "The characteristic equation", \(A\) is similar to a diagonal matrix. The diagonalisation theorem says exactly when this happens: an \(n \times n\) matrix is diagonalisable if and only if it has \(n\) linearly independent eigenvectors. When it does, the columns of \(P\) are those eigenvectors, and the diagonal entries of \(D\) are their eigenvalues, listed in the same order. Diagonalising a matrix therefore means finding enough eigenvectors and arranging them carefully.
Why the theorem works
Write \(P\) with columns \(\mathbf{v}_1, \dots, \mathbf{v}_n\) and let \(D\) have diagonal entries \(\lambda_1, \dots, \lambda_n\). Multiplying column by column, \(AP\) has columns \(A\mathbf{v}_1, \dots, A\mathbf{v}_n\), while \(PD\) has columns \(\lambda_1\mathbf{v}_1, \dots, \lambda_n\mathbf{v}_n\). So \(AP = PD\) says precisely that each \(\mathbf{v}_i\) is an eigenvector with eigenvalue \(\lambda_i\). Multiplying \(AP = PD\) on the right by \(P^{-1}\) gives \(A = PDP^{-1}\), and that step needs \(P\) to be invertible, which by "The invertible matrix theorem" means its \(n\) columns are linearly independent. In the eigenvector basis, \(A\) simply multiplies each coordinate by its own eigenvalue.
Worked example: a 2 × 2 matrix and its powers
Diagonalise \(A\) and find a formula for \(A^k\).
1. Trace 3 and determinant -4 give \(\lambda^2 - 3\lambda - 4 = (\lambda - 4)(\lambda + 1)\), so \(\lambda = 4\) and \(-1\). 2. \(A - 4I\) has rows (-3, 2) and (3, -2), giving eigenvector (2, 3); \(A + I\) has rows (2, 2) and (3, 3), giving (1, -1). 3. So \(P\) has columns (2, 3) and (1, -1), and \(D\) has diagonal 4, -1. 4. \(\det P = -5\), and \(A^k = PD^kP^{-1}\) works out to
5. Check: \(k = 1\) returns \(A\), and \(k = 3\) gives rows (25, 26) and (39, 38), matching \(A \cdot A \cdot A\).
Distinct eigenvalues are enough
"Eigenvectors and eigenvalues" showed that eigenvectors belonging to distinct eigenvalues are linearly independent. So an \(n \times n\) matrix with \(n\) distinct eigenvalues automatically has \(n\) independent eigenvectors, one for each eigenvalue, and is diagonalisable. This is the quickest test there is: a 3 × 3 matrix with eigenvalues 1, 2 and 7 is diagonalisable before you find a single eigenvector. The condition is sufficient but not necessary. The identity matrix has only the eigenvalue 1, repeated \(n\) times, and it is already diagonal. With repeated eigenvalues, you have to look at the eigenspaces.
Repeated eigenvalues
Suppose all the eigenvalues of \(A\) are real. Then \(A\) is diagonalisable exactly when, for every eigenvalue, the geometric multiplicity (the eigenspace dimension) equals the algebraic multiplicity. Equivalently, the eigenspace dimensions add up to \(n\), and putting the bases of all the eigenspaces together gives the \(n\) independent columns of \(P\). A matrix that falls short is called defective. The matrix with rows (5, 1) and (0, 5) is the standard example: 5 has algebraic multiplicity 2, but \(A - 5I\) has one free variable, so there is only one line of eigenvectors and no invertible \(P\) can be built.
Worked example: a repeated eigenvalue
The matrix below has characteristic polynomial \(-(\lambda - 3)^2(\lambda - 1)\). Is it diagonalisable?
1. For \(\lambda = 3\), \(A - 3I\) reduces to the single equation \(x_1 - x_2 - x_3 = 0\), with two free variables, so a basis is (1, 1, 0) and (1, 0, 1). 2. For \(\lambda = 1\), \(A - I\) reduces to \(x_1 = 0\) and \(x_2 = x_3\), giving (0, 1, 1). 3. Dimensions \(2 + 1 = 3\), so \(A\) is diagonalisable: \(P\) has columns (1, 1, 0), (1, 0, 1), (0, 1, 1) and \(D\) has diagonal 3, 3, 1. 4. Check \(AP = PD\) instead of inverting \(P\): both have rows (3, 3, 0), (3, 0, 1) and (0, 3, 1).
Powers and the long run
Diagonalisation is the engine behind repeated processes. If \(\mathbf{x}_{k+1} = A\mathbf{x}_k\), then \(\mathbf{x}_k = A^k\mathbf{x}_0 = PD^kP^{-1}\mathbf{x}_0\), and only the powers \(\lambda_i^k\) change with \(k\). In the 2 × 2 example, \(4^k\) swamps \((-1)^k\), so for large \(k\) the matrix \(A^k\) is close to \(4^k\) times a fixed matrix, and \(A^k\mathbf{x}_0\) lines up with the eigenvector (2, 3) unless \(\mathbf{x}_0\) has no component along it. "Discrete dynamical systems" develops this idea. Note that diagonalisability and invertibility are separate questions: an eigenvalue 0 on the diagonal of \(D\) is perfectly allowed.
Beyond diagonalisation
Some families are always diagonalisable. A real symmetric matrix always is, with real eigenvalues and an orthogonal \(P\), a result developed in "Diagonalizing symmetric matrices". Defective matrices cannot be diagonalised, but every square matrix is similar to a Jordan form, which is nearly diagonal, with eigenvalues on the diagonal and some 1s just above it; it is named after Camille Jordan, who published it in 1870. In numerical work, the Schur form, an upper triangular matrix reached by an orthogonal change of basis, is often preferred because it can be computed stably.
Common mistakes to avoid
First, listing the eigenvalues in \(D\) in a different order from the eigenvectors in \(P\): column \(i\) of \(P\) must belong to entry \(i\) of \(D\). Second, writing \(P^{-1}DP\) instead of \(PDP^{-1}\). Third, assuming a repeated eigenvalue rules out diagonalisation; it does only when its eigenspace is too small. Fourth, assuming any \(n\) eigenvectors will do: they must be linearly independent, so take a basis of each eigenspace. Fifth, confusing diagonalisable with invertible.
Key takeaways
An \(n \times n\) matrix is diagonalisable, \(A = PDP^{-1}\), exactly when it has \(n\) linearly independent eigenvectors; they form the columns of \(P\), and \(D\) holds their eigenvalues in the same order. The identity \(AP = PD\) is the reason, and the quickest check. \(n\) distinct eigenvalues guarantee diagonalisability; with repeated ones, each eigenspace dimension must match its algebraic multiplicity, or the matrix is defective. Powers follow from \(A^k = PD^kP^{-1}\).
Flashcards
Flashcard
A = PDP⁻¹ with P invertible and D diagonal
Test your understanding
Question 1 of 5
A 4 × 4 matrix has eigenvalues 1, 3, −2 and 5. What can you conclude?
Question 2 of 5
A = PDP⁻¹, where D has diagonal entries 2 and −1. Which statement about A⁴ is correct?
Question 3 of 5
A has eigenvector (1, 2) for λ = 3 and eigenvector (1, −1) for λ = −1. P has first column (1, −1) and second column (1, 2). Which D gives A = PDP⁻¹?
Question 4 of 5
Which of these matrices is not diagonalisable?
Question 5 of 5
A 3 × 3 matrix A has characteristic polynomial −(λ − 2)²(λ − 5), and A − 2I has rank 2. Is A diagonalisable?
