From one scalar equation to a system, and back to the same trick
The previous guide showed you how any higher-order linear equation can be flattened into a first-order system by naming each derivative as a fresh unknown. The reward of that matrix form x' = A x is that it looks exactly like the simplest differential equation you have ever solved. Remember the scalar case y' = a y from the very first rung: its solution is y = c * e^(at), a single exponential whose rate is the coefficient a. The bold hope of this guide is that the vector system x' = A x — many coupled unknowns at once — solves the same way, with the matrix A playing the role of the rate a.
But coupling is the whole obstacle. In x' = A x each unknown's rate of change depends on the *others*, so you cannot just integrate one equation at a time — pull on one thread and the others move. The single idea behind this entire rung is to find special directions in which the coupling *disappears*: directions along which the matrix A behaves like an ordinary number. Move along such a direction and the vector system briefly forgets it is a system; it acts like one scalar equation y' = lambda y, whose answer we already know cold.
The guess that turns calculus into algebra
Here is the leap. Guess a solution of the special shape x(t) = e^(lambda t) v, where lambda is an unknown number and v is an unknown *fixed* vector — a single direction that never rotates, only stretches or shrinks in time by the scalar e^(lambda t). This mirrors the scalar guess y = e^(rt) you trusted for the characteristic equation, only now the constant out front is a whole vector. Plug it in and watch the differential equation evaporate.
Guess: x(t) = e^(lambda t) v (v is a fixed, nonzero vector)
Left: x' = lambda e^(lambda t) v (only the scalar e^(lambda t) gets differentiated)
Right: A x = A ( e^(lambda t) v ) = e^(lambda t) (A v)
Set equal and cancel the never-zero scalar e^(lambda t):
A v = lambda v <-- the eigenvalue problem
rearranged: (A - lambda I) v = 0, v != 0Read what just happened. The differentiation only touched the scalar e^(lambda t), bringing down a factor lambda; the matrix A only touched the fixed vector v. After cancelling the exponential — which is never zero, so the cancellation is always legal — nothing remains but A v = lambda v. That is the eigenvalue problem: find a scalar lambda (an eigenvalue) and a nonzero vector v (an eigenvector) so that A merely scales v rather than tilting it. Each such pair (lambda, v) hands you one solution e^(lambda t) v of the system, for free. The entire eigenvalue method is this single exchange: a calculus problem for a linear-algebra one.
How to actually find lambda and v
Rewrite A v = lambda v as (A - lambda I) v = 0, where I is the identity matrix. This asks: for which lambda does the matrix (A - lambda I) crush some nonzero vector v down to the zero vector? A matrix only collapses a nonzero vector when it is *singular* — when its determinant is zero. So lambda must satisfy det(A - lambda I) = 0. For a 2-by-2 system this is just a quadratic in lambda, the system's own characteristic equation; its roots are the eigenvalues. Once you have an eigenvalue, substitute it back into (A - lambda I) v = 0 and solve that ordinary algebra system for the direction v.
- Form the matrix A from your first-order system x' = A x (the conversion from the previous guide).
- Solve the characteristic equation det(A - lambda I) = 0 for the eigenvalues lambda.
- For each lambda, solve (A - lambda I) v = 0 to find an eigenvector v (a direction, not a unique vector — any nonzero multiple works).
- Assemble each pair into a building-block solution e^(lambda t) v.
- Superpose: the general solution is c1 e^(lambda1 t) v1 + c2 e^(lambda2 t) v2 + ... , with constants fixed by the initial condition.
Three cases, because eigenvalues come in three flavours
The recipe above is perfect when everything is generic, but the characteristic equation is a polynomial, and polynomials have exactly the same three kinds of roots you met for second-order scalar equations. That parallel is not a coincidence — it is the same algebra wearing a matrix costume. So the eigenvalue method splits into three cases, and the remaining three guides of this rung are one apiece. Case one: real and distinct eigenvalues. Each lambda is a different real number, you get one honest eigenvector per eigenvalue, and the building blocks e^(lambda1 t) v1, e^(lambda2 t) v2 are automatically independent. This is the real-distinct case, the cleanest of all — every trajectory is governed by a pair of pure exponentials along fixed straight lines.
Case two: complex eigenvalues. When the characteristic quadratic has a negative discriminant, the eigenvalues come as a conjugate pair lambda = mu +/- i*omega, and so do the eigenvectors. A complex exponential e^((mu + i omega) t) is, by Euler's formula, a real exponential e^(mu t) wrapped around a rotation cos(omega t) + i sin(omega t). Taking real and imaginary parts of one complex solution hands you two *real* solutions that spiral: the real part mu sets growth or decay, the imaginary part omega sets the rate of rotation. This is the complex case, and it is exactly why a phase portrait can swirl — the next guide on rotation lives here.
Case three: repeated eigenvalues. A double root lambda is the troublemaker. Sometimes it still supplies two independent eigenvectors and behaves nicely, but often it is *defective* — algebraically a double root, yet it yields only one eigenvector direction. Then e^(lambda t) v alone cannot span a two-dimensional solution space, and you are genuinely one solution short. The cure mirrors the lone x*e^(rt) partner from the repeated-root scalar case: you build a second solution of the form e^(lambda t) (t*v + w), where w is a generalized eigenvector solving (A - lambda I) w = v. This is the repeated-eigenvalue case, the subtlest of the three and the subject of the rung's final guide.
Why the building blocks are enough — and where the method's limits lie
Why are n eigen-solutions enough to capture *every* solution? Because for an n-by-n system the solution space is exactly n-dimensional, and n independent solutions form a basis — a fundamental set of vector solutions. When the eigenvectors v1, ..., vn are independent (which is automatic for distinct eigenvalues, and arrangeable in the well-behaved cases), the combination c1 e^(lambda1 t) v1 + ... + cn e^(lambda n t) vn already reaches every solution. Fixing the constants from a single initial condition x(0) is then just solving one linear system: write x(0) as a combination of the eigenvectors. That is the whole solution, start to finish.
Step back and admire the bridge you just crossed. A differential equation about change over time has been answered entirely by static algebra about directions and stretches — A v = lambda v. The eigenvalues decide the *fate* of every trajectory (grow, decay, oscillate), and the eigenvectors decide its *geometry* (the lines and spirals it follows). This is the exact spot where linear algebra and differential equations fuse into one subject. The three remaining guides simply zoom into the three cases — straight-line solutions, rotation, and generalized eigenvectors — so that by the end you can look at any 2-by-2 matrix and sketch the flow it generates by eye.