Gluing the pieces into the whole space
The primary decomposition theorem is the harvest of the last guide. Over a field where the characteristic polynomial factors fully (always true over C), the space splits as V = G(lambda_1, T) (+) G(lambda_2, T) (+) ... (+) G(lambda_r, T), one generalized eigenspace per distinct eigenvalue, each one T-invariant.
This is the invariant direct-sum decomposition we wanted from the start of the track: V is broken into invariant blocks, and on each block T - lambda_i*I is nilpotent. The hard obstruction of Guide 2 — missing complements — is dodged because these particular complements always exist. The blocks may not be eigenlines, but they are honest invariant summands.
A = [ 3 1 0 0 ;
0 3 0 0 ;
0 0 3 0 ;
0 0 0 7 ] eigenvalues: 3 (mult 3), 7 (mult 1)
G(3, A) = span{ e1, e2, e3 } (dim 3 = algebraic mult of 3)
G(7, A) = span{ e4 } (dim 1 = algebraic mult of 7)
Primary decomposition: R^4 = G(3,A) (+) G(7,A).
On G(3,A): A - 3I is nilpotent.
On G(7,A): A - 7I is zero (e4 is a genuine eigenvector).Splitting the operator: semisimple plus nilpotent
The decomposition of the space induces a decomposition of the operator. Define S to act on each block as multiplication by lambda_i, and let N = T - S. Then S is a semisimple operator (diagonalizable over C), N is nilpotent, and crucially S and N commute. This is the Jordan-Chevalley decomposition: T = S + N, uniquely.
- Run the primary decomposition to find each G(lambda_i, T) and its eigenvalue lambda_i.
- Define S = lambda_i * I on the block G(lambda_i, T); this is the semisimple part.
- Set N = T - S; on each block this is T - lambda_i*I, which is nilpotent there.
- Verify S N = N S — the commuting property that makes the split unique and useful.
What this buys you
Diagonalization is now revealed as the special case where N = 0 — where the operator is purely semisimple. Every other operator over C is semisimple-plus-a-genuine-nilpotent, and that nilpotent tail is precisely the obstruction we first met as the shear in Guide 2.