Triangular when diagonal is impossible
Diagonalization needs an eigenbasis and may fail. Triangularizability asks for less: a basis in which T is upper-triangular. Over the complex numbers this always succeeds, because the characteristic polynomial always has a root, so T always has at least one eigenvector to start the construction.
Upper-triangular has a clean meaning in invariant-subspace language. T is upper-triangular in a basis (b_1, ..., b_n) exactly when each initial chunk span{b_1, ..., b_k} is an invariant subspace. Triangularizing is nothing but building a nested tower of invariant subspaces.
The flag: a nested tower of invariant subspaces
A flag is a chain {0} = V_0 ⊂ V_1 ⊂ ... ⊂ V_n = V with dim V_k = k. When every V_k is T-invariant we call it a T-invariant flag. The whole content of triangularization is: every complex operator admits a T-invariant flag, and you can even choose it so consecutive basis vectors are orthonormal.
- Find an eigenvector v_1 of T (guaranteed over C); set V_1 = span{v_1}, invariant.
- Pass to the quotient V / V_1; the induced operator there also has an eigenvector, lifting to extend V_1 to an invariant V_2.
- Repeat, growing the invariant subspace one dimension at a time, until you reach V_n = V.
- Orthonormalize the resulting basis with Gram-Schmidt to get the UNITARY version — the Schur form.
Schur decomposition: for any complex matrix A there is a
UNITARY U (U* U = I) and an UPPER-TRIANGULAR matrix R with
A = U R U* ( U* = conjugate transpose of U )
R = [ l1 * * ;
0 l2 * ;
0 0 l3 ]
The diagonal entries l1, l2, l3 are exactly the eigenvalues of A
(counted with multiplicity). The columns of U are an orthonormal
basis realizing a T-invariant flag: span{u1} subset span{u1,u2} ...Reading off the spectrum
Once you have the Schur decomposition A = U R U*, the eigenvalues are simply the diagonal of R. The unitary U means the change of basis is numerically stable — no inversion of an ill-conditioned matrix — which is why Schur form, not Jordan form, is what real eigenvalue algorithms actually compute.
The payoff: from one operator to a commuting family
The flag idea scales beautifully. If several operators commute, they share a common invariant subspace at every rung, so they can be brought to upper-triangular form in the same basis at once. This is simultaneous triangularization, and it is the linear-algebra skeleton of Lie theory.
Step back and see the arc of the whole track. We began with a single eigenline, generalized it to invariant subspaces, discovered that complements can fail, repaired the gap with generalized eigenvectors and the primary decomposition, and arrived at a universal structure theorem: over C every operator triangularizes, every commuting family triangularizes together, and the diagonal always shows you the spectrum. Even when an operator refuses to be diagonal, the block-triangular form tames it — that is the bridge from raw eigentheory to the canonical forms ahead.