Differential Equations › Section E · Systems of differential equations › Topic E.7

Repeated eigenvalues

When the two eigenvalues of a 2 × 2 matrix coincide, there may be only one eigenvector, and the method of the last two topics runs one solution short. A generalised eigenvector supplies the missing solution and explains the bent trajectories of a degenerate node.

When the roots coincide

The characteristic equation \(\lambda^2 - \tau\lambda + \Delta = 0\), with \(\tau = \operatorname{tr} A\) and \(\Delta = \det A\), has a repeated root exactly when \(\tau^2 = 4\Delta\), and then \(\lambda = \frac{\tau}{2}\) is an eigenvalue of multiplicity 2. As "Review: matrices, vectors, and eigenvalues" warned, such an eigenvalue can have two independent eigenvectors or only one. For a 2 × 2 matrix the first case is rare and easy, and the second needs a new idea. Either way, a fundamental set still needs two independent solutions.

The easy case: A = λI

If a 2 × 2 matrix has a double eigenvalue \(\lambda\) with two independent eigenvectors, then \(A\) multiplies every vector of the plane by \(\lambda\), so \(A = \lambda I\). The system then splits into \(x_1' = \lambda x_1\) and \(x_2' = \lambda x_2\), and the general solution is \(\mathbf{x} = e^{\lambda t}\mathbf{c}\) for any constant vector \(\mathbf{c}\), for instance \(c_1e^{\lambda t}(1, 0) + c_2e^{\lambda t}(0, 1)\). Every trajectory is a straight ray through the origin, pointing in every possible direction. This picture is called a star node, or proper node: asymptotically stable if \(\lambda < 0\) and unstable if \(\lambda > 0\).

Why t times the first solution fails

Now suppose there is only one eigenvector direction, \(\mathbf{v}\). One solution is \(e^{\lambda t}\mathbf{v}\). Inspired by "Repeated roots", where \(te^{rt}\) supplied the second solution, try \(\mathbf{x} = te^{\lambda t}\mathbf{v}\). Its derivative is \(e^{\lambda t}\mathbf{v} + \lambda te^{\lambda t}\mathbf{v}\), while \(A\mathbf{x} = \lambda te^{\lambda t}\mathbf{v}\). The two sides differ by \(e^{\lambda t}\mathbf{v}\), which is never zero because eigenvectors are non-zero. So \(te^{\lambda t}\mathbf{v}\) on its own is not a solution. The leftover term suggests adding a correction of the form \(e^{\lambda t}\mathbf{w}\).

The second solution

Try \(\mathbf{x}_2 = e^{\lambda t}(t\mathbf{v} + \mathbf{w})\) for an unknown constant vector \(\mathbf{w}\). Then \(\mathbf{x}_2' = e^{\lambda t}(\lambda t\mathbf{v} + \mathbf{v} + \lambda\mathbf{w})\) and \(A\mathbf{x}_2 = e^{\lambda t}(\lambda t\mathbf{v} + A\mathbf{w})\). These agree exactly when \(A\mathbf{w} = \lambda\mathbf{w} + \mathbf{v}\), that is,

(A - \lambda I)\mathbf{w} = \mathbf{v}

Such a \(\mathbf{w}\) is called a generalised eigenvector. Its system has the same coefficient matrix as the eigenvector system, but the right side is \(\mathbf{v}\) instead of \(\mathbf{0}\). The general solution is

\mathbf{x} = c_1e^{\lambda t}\mathbf{v} + c_2e^{\lambda t}(t\mathbf{v} + \mathbf{w})

Worked example: an unstable degenerate node

Solve \(\mathbf{x}' = A\mathbf{x}\), \(\mathbf{x}(0) = (3, 1)\), for

A = \begin{bmatrix} 5 & -1 \\ 1 & 3 \end{bmatrix}

1. Trace 8 and determinant \(15 + 1 = 16\) give \((\lambda - 4)^2 = 0\). 2. \(A - 4I\) has rows (1, -1) and (1, -1), so \(\mathbf{v} = (1, 1)\), and there is only one eigenvector direction. 3. Solve \((A - 4I)\mathbf{w} = (1, 1)\): \(w_1 - w_2 = 1\), so take \(\mathbf{w} = (1, 0)\). 4. At \(t = 0\): \(c_1(1, 1) + c_2(1, 0) = (3, 1)\), so \(c_1 = 1\) and \(c_2 = 2\). 5. Collecting terms, \(\mathbf{x} = e^{4t}(3 + 2t, 1 + 2t)\). 6. Check: \(x_1' = e^{4t}(14 + 8t)\) and \(5x_1 - x_2 = e^{4t}(14 + 8t)\).

Choosing w

The equation \((A - \lambda I)\mathbf{w} = \mathbf{v}\) always has infinitely many solutions, because adding any multiple of \(\mathbf{v}\) to one solution gives another. In the example, \(\mathbf{w} = (2, 1)\) works as well as (1, 0). Choose the simplest, usually by setting one entry to 0. A different choice only shifts part of \(c_2\mathbf{w}\) into the \(c_1\mathbf{v}\) term, so after the initial condition is applied the final solution is identical. Do not confuse this system with the eigenvector system: if you solve \((A - \lambda I)\mathbf{w} = \mathbf{0}\), you simply find \(\mathbf{v}\) again.

The degenerate node

With one eigenvector, the only straight-line trajectories lie on the line through \(\mathbf{v}\). Any other solution is \(e^{\lambda t}\big[(c_1 + c_2t)\mathbf{v} + c_2\mathbf{w}\big]\), and as \(t \to \pm\infty\) the \(c_2t\mathbf{v}\) part dominates. So trajectories meet the origin tangent to the eigenvector line and end up parallel to it far away, but the sign of \(t\) flips between the two ends, so each trajectory turns round. This picture is a degenerate, or improper, node: unstable when \(\lambda > 0\), asymptotically stable when \(\lambda < 0\). For the example, \(A(1, 0) = (5, 1)\) shows the turning direction: at (1, 0), trajectories head right and slightly upwards.

Worked example: a stable degenerate node

Solve \(\mathbf{x}' = A\mathbf{x}\), \(\mathbf{x}(0) = (0, 2)\), where \(A\) has rows (-1, 1) and (-1, -3). 1. Trace -4 and determinant \(3 + 1 = 4\) give \((\lambda + 2)^2 = 0\). 2. \(A + 2I\) has rows (1, 1) and (-1, -1), so \(\mathbf{v} = (1, -1)\). 3. Solve \(w_1 + w_2 = 1\): take \(\mathbf{w} = (1, 0)\). 4. At \(t = 0\): \(c_1(1, -1) + c_2(1, 0) = (0, 2)\), so \(c_1 = -2\) and \(c_2 = 2\). 5. So \(\mathbf{x} = e^{-2t}(2t, 2 - 2t)\). 6. Check: \(x_1' = e^{-2t}(2 - 4t)\) and \(-x_1 + x_2 = e^{-2t}(2 - 4t)\). Since \(\lambda = -2 < 0\), every trajectory approaches the origin.

Links and larger systems

Converting \(y'' + 4y' + 4y = 0\), a critically damped spring from "Mechanical vibrations", gives rows (0, 1) and (-4, -4) with double eigenvalue -2. A converted second-order equation is never \(\lambda I\), because its upper-right entry is 1, so the generalised eigenvector is always needed, and the first entry of the second solution produces the familiar \(te^{-2t}\). In larger systems, an eigenvalue of multiplicity 3 with one eigenvector needs a chain: \((A - \lambda I)\mathbf{w} = \mathbf{v}\), then \((A - \lambda I)\mathbf{u} = \mathbf{w}\), giving a third solution \(e^{\lambda t}\big(\frac{t^2}{2}\mathbf{v} + t\mathbf{w} + \mathbf{u}\big)\).

Common mistakes to avoid

First, using \(te^{\lambda t}\mathbf{v}\) alone as the second solution; it needs the \(e^{\lambda t}\mathbf{w}\) term. Second, solving \((A - \lambda I)\mathbf{w} = \mathbf{0}\) instead of \((A - \lambda I)\mathbf{w} = \mathbf{v}\). Third, forgetting \(\mathbf{w}\) when applying the initial condition: at \(t = 0\) the second solution equals \(\mathbf{w}\), not \(\mathbf{0}\). Fourth, using a different multiple of \(\mathbf{v}\) in the second solution from the one used to find \(\mathbf{w}\); the \(\mathbf{v}\) on the right of \((A - \lambda I)\mathbf{w} = \mathbf{v}\) must be the same vector that multiplies \(t\). Fifth, calling every repeated-eigenvalue picture a star node.

Key takeaways

A 2 × 2 matrix has a repeated eigenvalue \(\lambda = \frac{\tau}{2}\) when \(\tau^2 = 4\Delta\). If \(A = \lambda I\), every vector is an eigenvector, \(\mathbf{x} = e^{\lambda t}\mathbf{c}\), and the origin is a star node. Otherwise there is one eigenvector \(\mathbf{v}\), and the general solution is \(c_1e^{\lambda t}\mathbf{v} + c_2e^{\lambda t}(t\mathbf{v} + \mathbf{w})\) with \((A - \lambda I)\mathbf{w} = \mathbf{v}\). The origin is then a degenerate node, tangent to the eigenvector line and stable exactly when \(\lambda < 0\).

Flashcards

Flashcard

Repeated eigenvalue (2 × 2)

λ = τ/2, occurring when τ² = 4Δ

τ = tr A, Δ = det A.

Test your understanding

Question 1 of 5

A 2 × 2 matrix A has the double eigenvalue 3 but only one eigenvector direction, (1, 2). Which pair is a fundamental set for x′ = Ax?

Question 2 of 5

A has rows (2, 1) and (−1, 4), with double eigenvalue 3 and eigenvector v = (1, 1). Which w solves (A − 3I)w = v?

Question 3 of 5

Why does x = te^(λt)v fail to solve x′ = Ax when v is an eigenvector for λ?

Question 4 of 5

A 2 × 2 matrix has trace −6 and determinant 9 and is not a multiple of I. What is the origin of x′ = Ax?

Question 5 of 5

The matrix with rows (−2, 0) and (0, −2) has the double eigenvalue −2. What is the general solution of x′ = Ax?

← Previous topicNext topic →