Linear Algebra › Section E · Eigenvalues and eigenvectors › Topic E.6
Discrete dynamical systems
Populations, markets and networks are often modelled one time step at a time, with a matrix carrying this year's state to next year's. Eigenvalues and eigenvectors predict where such a system ends up: whether it grows, dies out or settles, and in what proportions.
Systems that move in steps
A discrete dynamical system is a rule
where \(\mathbf{x}_k\) is the state at step \(k\), such as the numbers of juvenile and adult animals in year \(k\), and the square matrix \(A\) describes how one step leads to the next. Given a starting state \(\mathbf{x}_0\), the states are \(\mathbf{x}_k = A^k\mathbf{x}_0\). The systems of "Difference equations" are of this type. Computing \(A^k\) step by step answers "what happens in year 10?", but it hides the question that matters most: what happens in the long run, and why. Eigenvectors answer that directly.
The eigenvector decomposition
Suppose \(A\) is diagonalisable, with eigenvectors \(\mathbf{v}_1, \dots, \mathbf{v}_n\) forming a basis and eigenvalues \(\lambda_1, \dots, \lambda_n\). Write the starting state as \(\mathbf{x}_0 = c_1\mathbf{v}_1 + \dots + c_n\mathbf{v}_n\). Each step multiplies each eigenvector by its own eigenvalue, so
This one formula holds the whole future of the system. The weights \(c_i\) come from solving a linear system, \(P\mathbf{c} = \mathbf{x}_0\), where \(P\) has the eigenvectors as columns. After that, every state is just powers of numbers.
Worked example: a two-stage population
A species has juveniles \(j_k\) and adults \(a_k\) each year. Each juvenile produces 0.5 young, each adult 1.1, and 60% of juveniles survive to become adults; adults then die. So \(\mathbf{x}_{k+1} = L\mathbf{x}_k\) with
1. Trace 0.5 and determinant -0.66 give \(\lambda^2 - 0.5\lambda - 0.66 = 0\), with roots 1.1 and -0.6. 2. Eigenvectors: (11, 6) for 1.1, from \(0.6x_1 = 1.1x_2\), and (1, -1) for -0.6. 3. For \(\mathbf{x}_0 = (120, 50)\), solve to get \(\mathbf{x}_0 = 10(11, 6) + 10(1, -1)\). 4. So
5. Check \(k = 1\): the formula gives (115, 72), and so does \(L(120, 50)\).
Reading the long run
Order the eigenvalues so that \(|\lambda_1|\) is strictly the largest; \(\lambda_1\) is then the dominant eigenvalue. Dividing the decomposition by \(\lambda_1^k\) shows that every other term shrinks relative to the first, so for large \(k\), \(\mathbf{x}_k \approx c_1\lambda_1^k\mathbf{v}_1\) as long as \(c_1 \ne 0\). Two predictions follow. The state grows or shrinks by a factor of about \(\lambda_1\) each step, and its proportions approach those of \(\mathbf{v}_1\). In the population example, \((-0.6)^k\) fades, so the population grows about 10% a year, with juveniles making up \(\frac{11}{17} \approx 65\%\) of it. If \(\lambda_1 = 1\), the state settles to a steady state, as in "Markov chains".
Age-structured population models
The matrix in the worked example is a small Leslie matrix, named after the ecologist Patrick Leslie, who introduced this kind of model in 1945. The population is split into age classes. The first row holds the birth rate of each class, and the entries just below the diagonal hold the fraction of each class that survives into the next. The dominant eigenvalue is the long-run growth factor per time step: above 1 the population grows, below 1 it declines towards extinction. Its eigenvector, scaled to add up to 1, is the stable age distribution that the population approaches whatever mix it starts with, provided \(c_1 \ne 0\).
Pictures in the plane: attractors, repellers, saddles
For a 2 × 2 system, plot the points \(\mathbf{x}_0, \mathbf{x}_1, \mathbf{x}_2, \dots\); the path they trace is a trajectory. With real eigenvalues, the origin falls into one of three types. If both \(|\lambda| < 1\), every trajectory heads into the origin, which is an attractor. If both \(|\lambda| > 1\), every trajectory except the one starting at \(\mathbf{0}\) moves away, and the origin is a repeller. If one \(|\lambda|\) is above 1 and the other below, the origin is a saddle point: trajectories starting on the line of the smaller eigenvalue's eigenvector move in towards it, while all others eventually move away along the dominant eigenvector.
Worked example: a saddle point
Classify the origin for the matrix \(A\) below and describe the trajectory from \(\mathbf{x}_0 = (5, 0)\).
1. Trace 1.7 and determinant 0.66 give \((\lambda - 1.1)(\lambda - 0.6) = 0\). 2. Eigenvectors: (1, 1) for 1.1 and (3, -2) for 0.6. 3. One eigenvalue is above 1 in size and the other below, so the origin is a saddle point. 4. Solving \(c_1(1, 1) + c_2(3, -2) = (5, 0)\) gives \(c_1 = 2\) and \(c_2 = 1\), so
5. Check: \(\mathbf{x}_1 = (4, 1) = A(5, 0)\). The second term dies away, and the trajectory bends towards the line through (1, 1) while growing. Only starts on the line through (3, -2) approach the origin.
Changing variables to decouple
Why does the eigenvector picture work so cleanly? Let \(P\) have the eigenvectors as columns and define new coordinates \(\mathbf{y}_k = P^{-1}\mathbf{x}_k\). Since \(A = PDP^{-1}\),
With \(D\) diagonal, each new coordinate evolves on its own: the first is multiplied by \(\lambda_1\) every step, the second by \(\lambda_2\), and so on. The original variables affect each other, which is what makes the system hard to read, but in eigenvector coordinates the system decouples into separate one-number problems. This is "Eigenvectors and linear transformations" applied to a moving state.
Complex eigenvalues: spirals
When a real 2 × 2 matrix has complex eigenvalues \(a \pm bi\), "Complex eigenvalues" showed that \(A = PCP^{-1}\), where \(C\) rotates through an angle and scales by \(r = \sqrt{a^2 + b^2}\). The trajectories therefore wind around the origin. If \(r < 1\) they spiral inwards and the origin is an attractor; if \(r > 1\) they spiral outwards; if \(r = 1\) they circle forever on a closed curve. For eigenvalues \(0.5 \pm 0.5i\), \(r = \sqrt{0.5} \approx 0.71\), so every state spirals in, and its size measured in the coordinates given by \(P\) shrinks by about 29% each step.
Ranking web pages
The PageRank method behind Google's early search engine is a dynamical system on a huge scale. Picture a surfer who follows a random link on each page, and now and then jumps to a random page instead. The probabilities of being on each page form a state vector, and one click multiplies it by a stochastic matrix built from the web's links. The long-run state is the eigenvector with eigenvalue 1, and its entries rank the pages. With billions of pages no one solves for it directly: repeating the multiplication from a starting guess until the state stops changing is the power method of "Iterative and numerical eigenvalue methods".
Common mistakes to avoid
First, judging growth by the eigenvalues themselves rather than their sizes: an eigenvalue of -1.2 still makes its term grow, while flipping sign each step. Second, forgetting the condition \(c_1 \ne 0\): a start with no component along the dominant eigenvector does not follow it. Third, reading the stable proportions wrongly: the eigenvector (11, 6) means 11 parts in 17, not 11 parts in 6. Fourth, judging complex eigenvalues by the real part alone; for spirals, the modulus decides.
Key takeaways
For \(\mathbf{x}_{k+1} = A\mathbf{x}_k\) with a basis of eigenvectors, \(\mathbf{x}_0 = \sum c_i\mathbf{v}_i\) gives \(\mathbf{x}_k = \sum c_i\lambda_i^k\mathbf{v}_i\). A strictly dominant eigenvalue sets the long-run growth factor, and its eigenvector sets the long-run proportions. In the plane, the origin is an attractor if both \(|\lambda| < 1\), a repeller if both exceed 1, and a saddle point if one of each. Complex eigenvalues make trajectories spiral, inwards when \(r < 1\) and outwards when \(r > 1\).
Flashcards
Flashcard
A discrete dynamical system
Test your understanding
Question 1 of 5
A 2 × 2 matrix A has eigenvalues 0.9 and 0.5. For xₖ₊₁ = Axₖ, what kind of point is the origin?
Question 2 of 5
x₀ = 2v₁ + 5v₂, where v₁ and v₂ are eigenvectors of A with eigenvalues 3 and 0.5. What is x₂ for xₖ₊₁ = Axₖ?
Question 3 of 5
A two-class Leslie matrix (juveniles, adults) has dominant eigenvalue 1.05 with eigenvector (3, 2). What happens in the long run?
Question 4 of 5
A has eigenvector (1, 1) for λ = 1.1 and eigenvector (3, −2) for λ = 0.6. Which starting vector gives a trajectory that approaches the origin?
Question 5 of 5
A real 2 × 2 matrix has eigenvalues 0.5 ± 0.5i. How do the trajectories of xₖ₊₁ = Axₖ behave?
