ODE: Stability via Eigenvalue Analysis

Understanding the long-term behavior of a system near its equilibrium points is a cornerstone of engineering analysis. Whether you're predicting the damping of a suspension bridge, the voltage regulation of a power grid, or the concentration of reactants in a chemical reactor, classifying stability is essential. Eigenvalue analysis provides a powerful, systematic method to determine if small perturbations will die out, grow uncontrollably, or oscillate, allowing engineers to design for safety and performance.

The Fundamental Link: Eigenvalues and Trajectories

For a linear system of ordinary differential equations (ODEs) expressed in matrix form, $\dot{\mathbf{x}}=A\mathbf{x}$, the equilibrium point is at the origin. The behavior of trajectories near this point is entirely governed by the eigenvalues and eigenvectors of the coefficient matrix $A$. An eigenvalue $\lambda$ is a scalar that satisfies the equation $A\mathbf{v}=\lambda \mathbf{v}$ for a non-zero eigenvector $\mathbf{v}$.

The general solution to the system is built from terms of the form $e^{\lambda t}\mathbf{v}$. Therefore, the exponential function $e^{\lambda t}$ dictates the growth or decay of solution components over time. The real part of $\lambda$, denoted $\text{Re}(\lambda )$, is the critical factor:

• If $\text{Re}(\lambda )<0$, the term $e^{\lambda t}$ decays exponentially to zero.
• If $\text{Re}(\lambda )>0$, the term $e^{\lambda t}$ grows exponentially without bound.
• If $\text{Re}(\lambda )=0$, the term $e^{\lambda t}$ has constant magnitude, leading to sustained oscillation.

This simple observation forms the bedrock of all stability classification for linear systems.

Classifying Real Eigenvalues: Nodes and Saddles

When the eigenvalues of $A$ are real and distinct, the solutions are composed of pure exponential functions. The phase portrait—a plot of system trajectories—takes on characteristic shapes.

A stable node occurs when both eigenvalues are real, distinct, and negative (e.g., ${\lambda }_1<{\lambda }_2<0$). All exponential terms decay, so every trajectory is pulled directly into the equilibrium point along the eigenvectors. Think of a ball bearing rolling to the bottom of a smooth, steep bowl without oscillation.

An unstable node occurs when both eigenvalues are real, distinct, and positive (e.g., ${\lambda }_1>{\lambda }_2>0$). All exponential terms grow, so trajectories are repelled directly away from the equilibrium. This is like reversing time for the stable node; trajectories emerge from the origin along the eigenvector directions.

A saddle point occurs when the eigenvalues are real and of opposite signs (e.g., ${\lambda }_1<0<{\lambda }_2$). This is the critical case of conditional instability. The negative eigenvalue attracts trajectories along its corresponding eigenvector (the stable manifold), while the positive eigenvalue repels them along its eigenvector (the unstable manifold). Most trajectories are hyperbolic curves that approach the saddle along the stable direction before being flung out along the unstable direction. In engineering, a saddle point represents an unstable equilibrium; even tiny deviations not perfectly aligned with the stable manifold will lead to runaway behavior.

Classifying Complex Eigenvalues: Spirals and Centers

When eigenvalues are complex conjugates, $\lambda =\alpha \pm i\beta$, the solutions involve terms $e^{\alpha t}(\cos (\beta t)+i\sin (\beta t))$. The imaginary part $\beta$ causes rotation, while the real part $\alpha$ controls the radial growth or decay.

A stable spiral (or spiral sink) occurs when the real part is negative ($\alpha <0$) and the imaginary part is non-zero ($\beta \ne 0$). Trajectories spiral inward toward the equilibrium point. This is a very common outcome in damped mechanical or electrical oscillatory systems, like a spring-mass system with friction.

An unstable spiral (or spiral source) occurs when the real part is positive ($\alpha >0$) and the imaginary part is non-zero ($\beta \ne 0$). Trajectories spiral outward away from the equilibrium. This represents an oscillatory system with negative damping, where energy is added, causing amplitudes to grow.

A center occurs when the eigenvalues are purely imaginary ($\alpha =0,\beta \ne 0$). The magnitude $e^{\alpha t}$ is constant, so trajectories form closed elliptical orbits around the equilibrium. This represents a conservative system with no dissipation, like an ideal frictionless spring-mass system or an LC circuit. While not asymptotically stable, a center is Lyapunov stable—small perturbations lead to small, bounded oscillations, not runaway growth.

Handling Degenerate and Special Cases

Real-world systems don't always yield distinct eigenvalues. Degenerate cases require careful analysis. One common case is a repeated real eigenvalue (${\lambda }_1={\lambda }_2$). The stability is still determined by the sign of this repeated eigenvalue (negative for stability, positive for instability), but the geometric phase portrait can be a proper node or an improper node, depending on whether one or two linearly independent eigenvectors exist. A proper node has trajectories that approach from all directions, while an improper node has a preferred direction with a "sheared" appearance.

Another special case is when one eigenvalue is exactly zero. This indicates the system matrix $A$ is singular, and the equilibrium is not isolated—there is an entire line (or plane) of equilibrium points. Stability analysis for the full system becomes more nuanced, often requiring examination of the nonlinear system or the use of center manifold theory.

Constructing the Complete Classification Chart

The power of eigenvalue analysis is that it condenses all possible stability behaviors into a simple visual chart based on the location of the eigenvalues in the complex plane. You can create this chart by considering the two eigenvalues, ${\lambda }_1$ and ${\lambda }_2$.

1. Trace-Determinant Plane: For a 2x2 matrix $A=\left[\begin{array}{cc}a& b\\ c& d\end{array}\right]$, the eigenvalues are determined by ${\lambda }^2-\tau \lambda +\Delta =0$, where $\tau =a+d$ (the trace) and $\Delta =ad-bc$ (the determinant). The discriminant is $D={\tau }^2-4\Delta$.
2. Plot Stability Regions: In the $\tau$-$\Delta$ plane:

• $\Delta <0$ always gives real eigenvalues of opposite signs: Saddle Point (unstable).
• $\Delta >0$ and $\tau <0$ and $D>0$: Stable Node.
• $\Delta >0$ and $\tau >0$ and $D>0$: Unstable Node.
• $\Delta >0$ and $\tau <0$ and $D<0$: Stable Spiral.
• $\Delta >0$ and $\tau >0$ and $D<0$: Unstable Spiral.
• $\Delta >0$ and $\tau =0$: Center (purely imaginary eigenvalues).
• The parabola ${\tau }^2=4\Delta$ (where $D=0$) is the boundary between nodes and spirals, representing the degenerate repeated eigenvalue cases.

1. The Vertical Axis ($\tau =0$) separates systems that dissipate energy (stable, $\tau <0$) from those that add energy (unstable, $\tau >0$).

This chart allows you to classify stability at a glance by simply computing the trace and determinant of your system matrix, a crucial skill for rapid engineering assessment.

Common Pitfalls

1. Misapplying Linear Analysis to Nonlinear Systems: Eigenvalue analysis is strictly for the linearized system near an equilibrium. For a nonlinear system $\dot{\mathbf{x}}=f(\mathbf{x})$, you must first find the equilibria, compute the Jacobian matrix $J=Df$ at that point, and then analyze its eigenvalues. The results only predict local stability near that specific point.
2. Overlooking the Saddle Point's Partial Stability: A saddle point is unstable overall. A common exam trap is to see the negative eigenvalue and incorrectly label the equilibrium as "stable." Remember, stability requires all eigenvalues to have negative real parts.
3. Confusing a Center with a Spiral: With numerical or approximate calculations, eigenvalues like $-0.001\pm 3i$ are very close to being a center but are technically a stable spiral. In practice, this distinction is critical: the spiral is asymptotically stable, while the center is only neutrally stable. Small modeling errors can shift a predicted center into a very slow spiral in reality.
4. Ignoring Eigenvectors for Saddles: While eigenvalues tell you a saddle is unstable, the eigenvectors tell you the directions of the stable and unstable manifolds. In control engineering, knowing the stable manifold direction might be used to design a controller that steers the system onto that safe path.

Summary

• The sign of the real part of the eigenvalues of the system matrix (or its Jacobian) dictates the local stability of an equilibrium point.
• Real eigenvalues lead to non-oscillatory behaviors: both negative for a stable node, both positive for an unstable node, and opposite signs for a saddle point.
• Complex eigenvalues lead to oscillatory behaviors: negative real part for a stable spiral, positive real part for an unstable spiral, and zero real part for a center.
• Degenerate cases like repeated real eigenvalues require checking the geometric multiplicity of eigenvectors but do not change the fundamental stability condition based on the sign.
• A complete classification chart in the trace-determinant plane provides an immediate graphical method for stability analysis of any two-dimensional linear system.