When the criterion fails: defective operators
When some eigenvalue has geometric multiplicity strictly less than its algebraic multiplicity, the criterion fails and the operator is defective. There simply aren't enough eigenvectors to fill a basis. The smallest example is the shear [4, 1; 0, 4] from Guide 1: a double eigenvalue with a one-dimensional eigenspace, missing exactly one direction.
Defectiveness is not a flaw in your computation — it's a real structural property. It is the precise obstruction that the rest of the curriculum exists to handle: invariant subspaces, the Jordan form, and the matrix exponential all grow from this single gap.
Patching the gap: generalized eigenspaces
The repair is to enlarge each eigenspace. The generalized eigenspace for lambda is the null space of (A - lambda I)^k for k large enough — vectors killed not by one application of (A - lambda I) but by some power of it. Crucially, the dimension of the generalized eigenspace equals the algebraic multiplicity, so the generalized eigenspaces always sum to the whole space, even when ordinary eigenspaces don't.
The payoff: powers, dynamics, and PageRank
Here's why all of this was worth it. When A diagonalizes, computing matrix powers collapses from O(k) matrix multiplications to one diagonal power: A^k = P D^k P^-1, and D^k just raises each eigenvalue to the k. That single identity solves linear recurrences (Fibonacci in closed form), evolves discrete dynamical systems x_{k+1} = A x_k, and computes steady states.
Why the largest eigenvalue wins (power iteration intuition):
write any start vector in the eigenbasis: x = c_1 v_1 + ... + c_n v_n
apply A repeatedly:
A^k x = c_1 lambda_1^k v_1 + ... + c_n lambda_n^k v_k
if |lambda_1| > |lambda_2| >= ... (lambda_1 dominant), factor it out:
A^k x = lambda_1^k ( c_1 v_1 + c_2 (lambda_2/lambda_1)^k v_2 + ... )
\__ ratios (lambda_j/lambda_1)^k -> 0 as k grows __/
=> direction of A^k x -> the dominant eigenvector v_1.
This IS the engine of PageRank: the steady-state web-rank
is the dominant eigenvector of the link matrix.That asymptotics is the dominant eigenvalue at work, and its convergence speed is set by the ratio |lambda_2 / lambda_1| — the spectral gap. To predict behavior without solving the characteristic polynomial, eigenvalue estimates like Gershgorin's disks pen the spectrum into regions read straight off the matrix entries. From one operator's spectrum, you now get its long-run dynamics, its stability, and a ranking of the entire web — the destination this whole track was climbing toward.