Decomposing the identity by eigenvalue
A diagonalizable A splits the space as a direct sum of its eigenspaces. For each eigenvalue lambda there is a spectral projection P_lambda — a projection onto E_lambda along all the other eigenspaces. These projectors are the spectral skeleton of A: they are idempotent, they multiply to zero across distinct eigenvalues, and they sum to the identity.
If A is diagonalizable with distinct eigenvalues lambda_1, ..., lambda_k: P_i^2 = P_i (each is a projection) P_i P_j = 0 for i != j (orthogonal idempotents) P_1 + ... + P_k = I (resolution of the identity) Spectral decomposition of A: A = lambda_1 P_1 + ... + lambda_k P_k And then ANY power / polynomial is read off the same projectors: A^m = lambda_1^m P_1 + ... + lambda_k^m P_k f(A) = f(lambda_1) P_1 + ... + f(lambda_k) P_k
This is the operator-level upgrade of A = P D P^-1: instead of one change-of-basis, you get k standalone pieces, each carrying one eigenvalue. Functions of A — powers, exponentials, inverses — become functions applied to scalars, eigenvalue by eigenvalue.
Building the projectors concretely
You don't have to guess the spectral projections — they read straight off A = P D P^-1. Group the columns of P by eigenvalue; the projector P_lambda keeps the coordinates belonging to lambda's eigenspace and zeroes the rest, expressed back in the standard basis. Equivalently, P_lambda = P E_lambda P^-1, where E_lambda is the diagonal indicator that is 1 on lambda's slots and 0 elsewhere.
A = [3, 1; 0, 2] eigenvalues 3 and 2 (distinct -> diagonalizable) eigenvector for 3: v1 = (1, 0) eigenvector for 2: v2 = (1, -1) P = [1, 1; 0, -1] P^-1 = [1, 1; 0, -1] (P is its own inverse here) D = diag(3, 2) P_3 = P diag(1,0) P^-1 = [1, 1; 0, 0] (onto E_3 along E_2) P_2 = P diag(0,1) P^-1 = [0, -1; 0, 1] (onto E_2 along E_3) check: P_3 + P_2 = I, P_3 P_2 = 0, 3 P_3 + 2 P_2 = A OK
Two operators, one basis
Now ask: can two diagonalizable operators A and B be diagonalized by the same P? That's simultaneous diagonalization. The clean theorem: two diagonalizable operators are simultaneously diagonalizable iff they are commuting operators, AB = BA. Commuting is exactly the algebraic shadow of sharing an eigenbasis.
The intuition for one direction: if B commutes with A, then B maps each eigenspace of A into itself (an A-eigenspace is invariant under B). Restrict B to each eigenspace, diagonalize it there, and stitch the local eigenbases together — the result diagonalizes both at once.