What repeated application reveals
Let be a finite-dimensional vector space over a field and let be linear. Repeated application creates two sequences:
The kernel grows because once a vector maps to zero, every later power still maps it to zero. The image shrinks because . Finite dimension forces the kernel chain eventually to stop strictly growing. More precisely, if , some works when ; for the zero space take .
Lemma 11.1.1
(Stabilization persists)
If for some , then and for every .
Proof
(One step at a time)
If , then , hence . Thus , and repeating the same reasoning gives every later power. For the images, rank–nullity gives for , and ; equal finite dimensions force equality.
Definition 11.1.2
(Stable kernel and stable image)
For any at which the kernel chain has stabilized, is the stable kernel and the stable image of . By the lemma, neither depends on which such is chosen.
The decomposition and its proof
Theorem 11.1.3
(Fitting decomposition)
Fix a positive at which the kernel chain stabilizes. Then
Both summands are -invariant, is nilpotent on and invertible on , and every linear map commuting with preserves both summands.
Proof
(Intersection, dimension count and invariance)
To prove the intersection is zero, take in . Then , so stabilization gives , and thus . Rank–nullity gives , so the direct sum is all of .
Both summands are -invariant. On , the restriction of is nilpotent: its th power is zero. On , the restriction is onto because ; in a finite-dimensional space it is therefore invertible. If , then : for we get , and implies .
Proof X-ray
Where finite dimension is spent
Stabilization is used once, in the form , to show that the two summands meet only in zero. Finite dimension is used three times: to guarantee that the kernel chain stabilizes at all, to pass from a zero intersection to a spanning sum through rank–nullity, and to upgrade the surjective restriction on to an invertible one. Each of these steps can fail in infinite dimension, which is why the multiplication operator below defeats the theorem even though its kernel chain is constant.
This is a finite-dimensional form of Fitting's decomposition. Treating as an -module with acting as , it separates the part killed by a power of from the part where multiplication by is invertible. The proof does not assume an eigenbasis or an algebraically closed field.
Worked example
Worked example 11.1.4
(A nilpotent block beside an identity)
Let be . Then and . Thus
Every vector splits as . The first restriction is nonzero but has square zero; the second is the identity.
Common mistake
(Using instead of the stable kernel)
In the previous example and . These two subspaces intersect in the line and their sum misses , which is not killed by but is killed by . The decomposition holds only once is large enough for the kernel chain to have stopped growing.
Worked example 11.1.5
(A projection: stabilization at the first step)
Over any field let on . Then , so and already works. Here and , and indeed . The restriction to is the identity and the restriction to is zero.
Worked example 11.1.6
(An invertible part without eigenvectors)
Over let . Then , so and . The stable image is the plane , on which is the rotation through a right angle. It is invertible, yet it has no real eigenvector, since there. The decomposition is obtained without any eigenbasis.
Why finite dimension matters
The theorem can fail for an operator whose kernel chain is constant from the start.
Counterexample mode
Multiplication by x on polynomials
On the infinite-dimensional space , let multiply by . Every is zero, so the kernel chain is stable from , while every is a proper subspace for . The sum is not all of , and restricted to is injective but not onto. Stabilization of the kernel alone does not imply the direct-sum result without the finite-dimensional rank–nullity argument.
Exercises with explanations
Quick check
(For , find and , and explain why is too early.)
Solution · Answer
They are and respectively. Their sum omits vectors with nonzero second coordinate, so does not give the decomposition.
Quick check
(If , identify the two summands for .)
Solution · Answer
and ; the invertible restriction on the zero space is its unique identity map.
Quick check
(If is invertible, identify the two summands.)
Solution · Answer
Every is invertible, so its kernel is zero and its image is .
Quick check
(Let commute with and let . Show that belongs to .)
Solution · Answer
. Likewise, implies , so preserves the stable kernel as well.