Linear Algebra › Section E · Eigenvalues and eigenvectors › Topic E.4

Eigenvectors and linear transformations

A linear transformation does not come with one fixed matrix: its matrix depends on the coordinates you measure it in. Choosing a basis of eigenvectors gives the simplest matrix of all, and explains what the factorisation A = PDP⁻¹ really says.

One transformation, many matrices

"The matrix of a linear transformation" built the standard matrix of a map from \(\mathbb{R}^n\) to \(\mathbb{R}^m\) by recording what happens to the standard basis vectors. Nothing forces that choice. Any basis can serve as the measuring grid, and the same transformation then has a different matrix. The question of this topic is which grid makes the matrix simplest. For a diagonalisable matrix the answer is a basis of eigenvectors, and the resulting matrix is diagonal. The same idea also lets us use matrices for transformations of spaces such as polynomials, which have no standard columns at all.

The matrix relative to two bases

Let \(T\) be a linear transformation from an \(n\)-dimensional space \(V\) with basis \(\mathcal{B} = \{\mathbf{b}_1, \dots, \mathbf{b}_n\}\) to an \(m\)-dimensional space \(W\) with basis \(\mathcal{C}\). The matrix of \(T\) relative to \(\mathcal{B}\) and \(\mathcal{C}\) is the \(m \times n\) matrix

M = \big[\, [T(\mathbf{b}_1)]_{\mathcal{C}} \;\; \cdots \;\; [T(\mathbf{b}_n)]_{\mathcal{C}} \,\big]

Its column \(j\) is the \(\mathcal{C}\)-coordinate vector of the image of \(\mathbf{b}_j\). It does the job of \(T\) in coordinates: \([T(\mathbf{x})]_{\mathcal{C}} = M[\mathbf{x}]_{\mathcal{B}}\). That follows from linearity, since \(\mathbf{x} = c_1\mathbf{b}_1 + \dots + c_n\mathbf{b}_n\) gives \(T(\mathbf{x}) = c_1T(\mathbf{b}_1) + \dots + c_nT(\mathbf{b}_n)\), and coordinates are linear.

Worked example: evaluating a polynomial

Let \(T\) send a polynomial \(p\) of degree at most 2 to the pair \((p(0), p(1))\) in \(\mathbb{R}^2\). Use the basis \(\{1, t, t^2\}\) of \(\mathbb{P}_2\) and the standard basis of \(\mathbb{R}^2\). 1. Apply \(T\) to each basis vector: \(T(1) = (1, 1)\), \(T(t) = (0, 1)\) and \(T(t^2) = (0, 1)\). 2. These are the columns:

M = \begin{bmatrix} 1 & 0 & 0 \\ 1 & 1 & 1 \end{bmatrix}

3. Test it on \(p = 4 - 3t + 5t^2\), whose coordinate vector is (4, -3, 5): \(M(4, -3, 5) = (4, 6)\). 4. Check directly: \(p(0) = 4\) and \(p(1) = 4 - 3 + 5 = 6\).

Transformations from a space to itself

When \(T\) maps \(V\) to itself, it is natural to use the same basis \(\mathcal{B}\) on both sides. The result is the \(\mathcal{B}\)-matrix of \(T\), written \([T]_{\mathcal{B}}\), whose column \(j\) is \([T(\mathbf{b}_j)]_{\mathcal{B}}\), with \([T(\mathbf{x})]_{\mathcal{B}} = [T]_{\mathcal{B}}[\mathbf{x}]_{\mathcal{B}}\). Differentiation on \(\mathbb{P}_2\) with basis \(\{1, t, t^2\}\) is a classic case. It sends 1 to 0, \(t\) to 1 and \(t^2\) to \(2t\), so

[D]_{\mathcal{B}} = \begin{bmatrix} 0 & 1 & 0 \\ 0 & 0 & 2 \\ 0 & 0 & 0 \end{bmatrix}

Multiplying it by (3, 5, 4), the coordinates of \(3 + 5t + 4t^2\), gives (5, 8, 0), the coordinates of the derivative \(5 + 8t\).

Eigenvectors of a transformation

Eigenvectors make sense for any linear \(T\) from \(V\) to itself: a non-zero \(\mathbf{x}\) with \(T(\mathbf{x}) = \lambda\mathbf{x}\). In a space of functions they are sometimes called eigenfunctions. Take \(T(p) = p + tp'\) on \(\mathbb{P}_2\). Then \(T(1) = 1\), \(T(t) = 2t\) and \(T(t^2) = 3t^2\), so each basis polynomial is an eigenvector, with eigenvalues 1, 2 and 3, and the \(\mathcal{B}\)-matrix is diagonal. Differentiation is the opposite case. Its matrix above is triangular with zero diagonal, so 0 is its only eigenvalue, and the eigenvectors are the non-zero constants. Three dimensions but a one-dimensional eigenspace: no basis makes \(D\) diagonal.

Changing the basis of Rⁿ

Now take \(T(\mathbf{x}) = A\mathbf{x}\) on \(\mathbb{R}^n\) and a basis \(\mathcal{B}\) whose vectors form the columns of an invertible \(P\). From "Coordinate systems", \(\mathbf{x} = P[\mathbf{x}]_{\mathcal{B}}\), so \([\mathbf{x}]_{\mathcal{B}} = P^{-1}\mathbf{x}\). Follow a vector through: \(A\mathbf{x} = AP[\mathbf{x}]_{\mathcal{B}}\), and converting the result back to \(\mathcal{B}\)-coordinates gives

[A\mathbf{x}]_{\mathcal{B}} = P^{-1}AP\,[\mathbf{x}]_{\mathcal{B}}

So the \(\mathcal{B}\)-matrix of \(\mathbf{x} \mapsto A\mathbf{x}\) is \(P^{-1}AP\). Read from right to left: convert \(\mathcal{B}\)-coordinates to standard coordinates, apply \(A\), then convert back.

The diagonal matrix representation

Put the two ideas together. If \(A = PDP^{-1}\) with \(D\) diagonal, then \(P^{-1}AP = D\), so \(D\) is the \(\mathcal{B}\)-matrix of \(\mathbf{x} \mapsto A\mathbf{x}\) when \(\mathcal{B}\) is the basis formed by the columns of \(P\). This is what diagonalisation means geometrically. In standard coordinates the map may shear and mix every coordinate. In eigenvector coordinates it does something very plain: it multiplies the first coordinate by \(\lambda_1\), the second by \(\lambda_2\), and so on. The factorisation \(PDP^{-1}\) reads as change to eigenvector coordinates, scale each one, change back.

Worked example: two bases for one map

Let \(A\) be the matrix below. Find its \(\mathcal{B}\)-matrix for \(\mathcal{B} = \{(1, 0), (1, 1)\}\) and then for a basis of eigenvectors.

A = \begin{bmatrix} 3 & 1 \\ -2 & 0 \end{bmatrix}

1. With \(P\) having columns (1, 0) and (1, 1), \(P^{-1}\) has rows (1, -1) and (0, 1), and \(AP\) has rows (3, 4) and (-2, -2). 2. So \(P^{-1}AP\) has rows (5, 6) and (-2, -2): no simpler than \(A\). 3. Trace 3 and determinant 2 give eigenvalues 1 and 2, with eigenvectors (1, -2) and (1, -1). 4. In that basis the matrix is diagonal, with entries 1 and 2. 5. Check: all three matrices have trace 3 and determinant 2, as similar matrices must.

Similar matrices, same transformation

Every \(\mathcal{B}\)-matrix of \(\mathbf{x} \mapsto A\mathbf{x}\) has the form \(P^{-1}AP\), so all of them are similar to \(A\) and to each other. That is why similar matrices share eigenvalues: the eigenvalues belong to the transformation, not to the grid. A projection shows this. Projection of \(\mathbb{R}^2\) onto the line \(y = x\) has standard matrix \(\frac{1}{2}\) times the all-ones 2 × 2 matrix. In the basis \(\{(1, 1), (1, -1)\}\) its matrix is diagonal with entries 1 and 0: keep the first coordinate, delete the second. Projecting twice changes nothing more, and in this form it is plain to see that the matrix squared equals itself.

Common mistakes to avoid

First, putting the images \(T(\mathbf{b}_j)\) in as columns instead of their coordinate vectors relative to the output basis. Second, writing \(PAP^{-1}\) for the \(\mathcal{B}\)-matrix; with \(\mathbf{x} = P[\mathbf{x}]_{\mathcal{B}}\), it is \(P^{-1}AP\). Third, forgetting that a matrix relative to bases of spaces of different dimensions is not square: \(n\) columns for the input dimension, \(m\) rows for the output. Fourth, thinking a change of basis changes the eigenvalues; only the eigenvectors' coordinates change.

Key takeaways

Relative to bases \(\mathcal{B}\) and \(\mathcal{C}\), a linear transformation has the matrix whose columns are \([T(\mathbf{b}_j)]_{\mathcal{C}}\), so that \([T(\mathbf{x})]_{\mathcal{C}} = M[\mathbf{x}]_{\mathcal{B}}\). For a map of a space to itself, one basis gives the \(\mathcal{B}\)-matrix. For \(\mathbf{x} \mapsto A\mathbf{x}\) and a basis forming the columns of \(P\), the \(\mathcal{B}\)-matrix is \(P^{-1}AP\). If \(A = PDP^{-1}\), then \(D\) is the matrix of the same map in eigenvector coordinates. All matrices of one transformation are similar and share its eigenvalues.

Flashcards

Flashcard

Matrix relative to B and C

Columns [T(b₁)]_C, …, [T(bₙ)]_C

Coordinates of the images, not the images.

Test your understanding

Question 1 of 5

T sends a polynomial p in P₂ to (p(0), p(2)) in R². Using the basis {1, t, t²} and the standard basis of R², what is the third column of the matrix of T?

Question 2 of 5

Using the basis {1, t, t²}, multiply the matrix of differentiation on P₂ by the coordinates (3, 2, −1) of 3 + 2t − t². What do you get?

Question 3 of 5

A = PDP⁻¹ with D diagonal, and B is the basis formed by the columns of P. What is the B-matrix of x ↦ Ax?

Question 4 of 5

A has rows (1, 1) and (0, 2), and B = {(1, 0), (1, 1)}. What is the B-matrix of x ↦ Ax?

Question 5 of 5

Two matrices represent the same linear transformation of R³ relative to two different bases. What must they have in common?

← Previous topicNext topic →