Linear Algebra › Section E · Eigenvalues and eigenvectors › Topic E.1
Eigenvectors and eigenvalues
A matrix pushes most vectors off their own line, but a few special directions are only stretched, shrunk or flipped. Those directions and their stretch factors are the key to understanding what a matrix does when it is applied again and again.
Directions a matrix leaves alone
Take the matrix
and feed it a few vectors. It sends (1, 1) to (7, 4), which points in a new direction. But it sends (2, 1) to (12, 6), exactly 6 times (2, 1), and it sends (1, -2) to (1, -2) itself. Along the line through (2, 1) the matrix acts like the number 6, and along the line through (1, -2) it acts like the number 1. Every vector in the plane is a combination of these two directions, so two numbers summarise the whole matrix.
The definition
Let \(A\) be an \(n \times n\) matrix. An eigenvector of \(A\) is a non-zero vector \(\mathbf{x}\) with \(A\mathbf{x} = \lambda\mathbf{x}\) for some scalar \(\lambda\); that scalar is the eigenvalue belonging to \(\mathbf{x}\). The vector must be non-zero, because \(A\mathbf{0} = \lambda\mathbf{0}\) holds for every \(\lambda\) and tells us nothing, but the eigenvalue may be 0. Only square matrices have eigenvectors, since \(A\mathbf{x}\) must live in the same space as \(\mathbf{x}\). If \(\lambda > 1\) the vector is stretched, if \(0 < \lambda < 1\) it is shrunk, and if \(\lambda < 0\) its direction is reversed. "Eigen" is German for "own".
Testing a vector
To decide whether a given vector is an eigenvector, compute \(A\mathbf{x}\) and ask whether it is a scalar multiple of \(\mathbf{x}\). For the matrix of the first card, \(A(1, 1) = (7, 4)\). A multiple of (1, 1) has equal entries, so (1, 1) is not an eigenvector. Any non-zero multiple of an eigenvector is again an eigenvector with the same eigenvalue: \(A(4, 2) = (24, 12) = 6(4, 2)\). So it is really a direction, not one particular vector, that the matrix singles out.
Testing a scalar
The equation \(A\mathbf{x} = \lambda\mathbf{x}\) is the same as
so \(\lambda\) is an eigenvalue exactly when this homogeneous system has a non-trivial solution, that is, when \(A - \lambda I\) is not invertible. In practice, check whether \(A - \lambda I\) has a free variable when row reduced, or whether its determinant is zero. The identity matrix matters: \(\lambda I\) subtracts \(\lambda\) from the diagonal entries only. One warning: row reducing \(A\) itself does not reveal its eigenvalues, because row operations change them.
Eigenspaces
Once \(\lambda\) is known to be an eigenvalue, the set of all solutions of \((A - \lambda I)\mathbf{x} = \mathbf{0}\) is the null space of \(A - \lambda I\). Like every null space it is a subspace of \(\mathbb{R}^n\), called the eigenspace of \(A\) for \(\lambda\). It holds the zero vector together with every eigenvector for \(\lambda\). It may be a line through the origin, a plane or something larger, and on it \(A\) acts simply as multiplication by \(\lambda\). A basis comes from the null-space method of "Subspaces, column space, and null space": row reduce and write the general solution in terms of the free variables.
Worked example: testing two candidate eigenvalues
Is 2 an eigenvalue of the matrix \(B\) below? Is 5?
1. \(B - 2I\) has rows (-1, 4) and (2, 1). Neither column is a multiple of the other, so only the zero vector solves \((B - 2I)\mathbf{x} = \mathbf{0}\), and 2 is not an eigenvalue (its determinant, -9, agrees). 2. \(B - 5I\) has rows (-4, 4) and (2, -2). Its second column is -1 times the first, so it is not invertible and 5 is an eigenvalue. 3. Row reduce: the rows become (1, -1) and (0, 0), so \(x_1 = x_2\) and the eigenspace is the line spanned by (1, 1). 4. Check: \(B(1, 1) = (5, 5) = 5(1, 1)\).
Worked example: a two-dimensional eigenspace
The matrix \(M\) below has 2 as an eigenvalue. Find a basis for its eigenspace.
1. Subtract 2 from each diagonal entry: every row of \(M - 2I\) is (1, -1, 2). 2. Row reduce: two rows become zero, leaving \(x_1 - x_2 + 2x_3 = 0\) with \(x_2\) and \(x_3\) free. 3. The general solution is \(x_2(1, 1, 0) + x_3(-2, 0, 1)\), so a basis is \(\{(1, 1, 0), (-2, 0, 1)\}\) and the eigenspace is a plane. 4. Check: \(M(1, 1, 0) = (2, 2, 0)\) and \(M(-2, 0, 1) = (-4, 0, 2)\), each twice the original vector.
Triangular matrices and the eigenvalue zero
For triangular matrices no work is needed: the eigenvalues are the main diagonal entries. The reason is that \(A - \lambda I\) is still triangular, with diagonal entries \(a_{ii} - \lambda\), and a triangular matrix is invertible only when none of its diagonal entries is zero. So a lower triangular matrix with diagonal 4, 2, 4 has eigenvalues 4 and 2 only. Separately, \(\lambda = 0\) is an eigenvalue exactly when \(A\mathbf{x} = \mathbf{0}\) has a non-trivial solution, that is, when \(A\) is not invertible. This adds a line to "The invertible matrix theorem": \(A\) is invertible if and only if 0 is not an eigenvalue of \(A\).
Different eigenvalues, independent eigenvectors
Eigenvectors \(\mathbf{v}_1, \dots, \mathbf{v}_r\) for distinct eigenvalues \(\lambda_1, \dots, \lambda_r\) are linearly independent. For two vectors, suppose \(\mathbf{v}_2 = c\mathbf{v}_1\). Then \(A\mathbf{v}_2 = c\lambda_1\mathbf{v}_1 = \lambda_1\mathbf{v}_2\), but also \(A\mathbf{v}_2 = \lambda_2\mathbf{v}_2\), so \((\lambda_1 - \lambda_2)\mathbf{v}_2 = \mathbf{0}\), impossible when \(\lambda_1 \ne \lambda_2\) and \(\mathbf{v}_2 \ne \mathbf{0}\). Hence an \(n \times n\) matrix has at most \(n\) distinct eigenvalues. Eigenvectors also explain repeated processes: if \(\mathbf{x}_{k+1} = A\mathbf{x}_k\) and \(\mathbf{x}_0\) is an eigenvector, then \(\mathbf{x}_k = \lambda^k\mathbf{x}_0\), keeping its direction at every step.
Eigenvectors you can see
A reflection of the plane across a line fixes every vector on the line (eigenvalue 1) and flips every vector perpendicular to it (eigenvalue -1). A projection onto a line keeps vectors on the line (eigenvalue 1) and sends perpendicular vectors to zero (eigenvalue 0). A horizontal shear, with rows (1, k) and (0, 1) for some \(k \ne 0\), has only the eigenvalue 1 and only one line of eigenvectors, the horizontal axis. A rotation by 90° moves every non-zero vector off its line, so it has no real eigenvectors; "Complex eigenvalues" shows what replaces them.
Common mistakes to avoid
First, calling the zero vector an eigenvector: it satisfies the equation for every \(\lambda\), so it is excluded, even though 0 is allowed as an eigenvalue. Second, subtracting \(\lambda\) from every entry of \(A\) instead of only the diagonal. Third, row reducing \(A\) and reading eigenvalues off the echelon form; only a matrix that is triangular to begin with can be read directly. Fourth, stopping at one eigenvector when the eigenspace is larger: count the free variables of \(A - \lambda I\) to find its dimension.
Key takeaways
An eigenvector of a square matrix is a non-zero vector with \(A\mathbf{x} = \lambda\mathbf{x}\), and \(\lambda\) is its eigenvalue. Test a vector by computing \(A\mathbf{x}\); test a scalar by asking whether \(A - \lambda I\) fails to be invertible. The eigenspace for \(\lambda\) is the null space of \(A - \lambda I\). Triangular matrices have their diagonal entries as eigenvalues, and 0 is an eigenvalue exactly when \(A\) is not invertible. Eigenvectors for distinct eigenvalues are linearly independent.
Flashcards
Flashcard
A non-zero vector x with Ax = λx
Test your understanding
Question 1 of 5
Let A have rows (2, 1) and (1, 2). Which vector is an eigenvector of A, with its eigenvalue stated correctly?
Question 2 of 5
Which statement about the number 0 and eigenvalues is correct?
Question 3 of 5
A lower triangular matrix has rows (3, 0, 0), (7, -2, 0) and (1, 4, 5). What are its eigenvalues?
Question 4 of 5
A is 3 × 3, and A - 2I row reduces to a matrix with exactly one pivot. What is the dimension of the eigenspace of A for λ = 2?
Question 5 of 5
In x(k+1) = A x(k), the starting vector x(0) = (4, 8) is an eigenvector of A with eigenvalue 0.5. What is x(3)?
