Diagonalization is the most pleasant outcome in matrix theory: choose a basis of eigenvectors and the matrix becomes a diagonal of scalars. But many important matrices are not diagonalizable, even over $\mathbb{C}$. The right response is not to abandon structure, but to ask what structure is still forced. Invariant subspaces, minimal polynomials, and Jordan form answer that question. They explain precisely what breaks when diagonalization fails, and what remains computable and stable.
This article develops the logic behind Jordan form, shows how invariant subspaces organize the theory, and emphasizes the “survivors”: quantities and decompositions that still behave predictably when eigenvectors are insufficient.
Featured Gaming CPUTop Pick for High-FPS GamingAMD Ryzen 7 7800X3D 8-Core, 16-Thread Desktop Processor
AMD Ryzen 7 7800X3D 8-Core, 16-Thread Desktop Processor
A strong centerpiece for gaming-focused AM5 builds. This card works well in CPU roundups, build guides, and upgrade pages aimed at high-FPS gaming.
- 8 cores / 16 threads
- 4.2 GHz base clock
- 96 MB L3 cache
- AM5 socket
- Integrated Radeon Graphics
Why it stands out
- Excellent gaming performance
- Strong AM5 upgrade path
- Easy fit for buyer guides and build pages
Things to know
- Needs AM5 and DDR5
- Value moves with live deal pricing
Invariant subspaces as the real objects
A subspace $W\subseteq \mathbb{F}^n$ (with $\mathbb{F}=\mathbb{R}$ or $\mathbb{C}$) is invariant under a matrix $A$ if $AW\subseteq W$. Invariance is the basis‑free way to say “the action of $A$ closes on this subspace.” If you restrict $A$ \to $W$, you get a smaller linear operator $A|_W$ whose matrix depends on a basis of $W$, but whose algebraic properties are intrinsic.
Diagonalization is the special case where $\mathbb{F}^n$ splits as a direct sum of one‑dimensional invariant subspaces spanned by eigenvectors. When diagonalization fails, the decomposition into invariant subspaces is still the correct target, but the invariant pieces can be larger than one dimension.
A guiding slogan that is accurate without being misleading:
- eigenvectors identify invariant lines,
- generalized eigenvectors identify invariant chains,
- Jordan blocks are the matrices of those chains.
Eigenvalues, eigenspaces, and the first obstruction
Fix an eigenvalue $\lambda$. The eigenspace is $\ker(A-\lambda I)$. Its dimension is the geometric multiplicity. The algebraic multiplicity is the power of $(t-\lambda)$ in the characteristic polynomial $\chi_A(t)$.
Diagonalization over $\mathbb{C}$ happens exactly when, for every eigenvalue, geometric and algebraic multiplicities agree. When they do not, you run out of eigenvectors. That shortage is the first obstruction.
The remedy is to enlarge the eigenspace \to a generalized eigenspace.
Generalized eigenspaces and primary decomposition
For a complex matrix $A$ with eigenvalue $\lambda$, define the generalized eigenspace
for $k$ large enough that the kernel stabilizes (it stabilizes by finite dimensionality). Vectors in $G_\lambda$ are those annihilated by some power of $A-\lambda I$. The map $A-\lambda I$ acts nilpotently on $G_\lambda$.
A deep but standard fact is the primary decomposition:
a direct sum over distinct eigenvalues. Each $G_\lambda$ is $A$-invariant, and $A$ restricted \to $G_\lambda$ has the single eigenvalue $\lambda$.
This decomposition is already a major structural win. It tells you that the study of a general matrix reduces to studying matrices with one eigenvalue, which are “a scalar plus a nilpotent.”
On $G_\lambda$,
where $N = A-\lambda I$ is nilpotent. The classification problem becomes: how can a nilpotent operator look, up to change of basis?
Nilpotent operators and Jordan chains
A nilpotent operator $N$ satisfies $N^p=0$ for some $p$. Nilpotency forces a filtration of invariant subspaces:
Each inclusion is invariant under $N$ and hence under $A$ on a generalized eigenspace.
Jordan chains arise from choosing vectors that sit just outside one kernel and then pushing them down by $N$. A length‑$\ell$ Jordan chain for eigenvalue $\lambda$ is a sequence $v_1,\dots,v_\ell$ such that
So $v_1$ is an eigenvector, $v_2$ maps \to $v_1$, $v_3$ maps \to $v_2$, and so on. The span of the chain is invariant, and in the chain basis the restriction of $A$ becomes a Jordan block:
The ones above the diagonal record the failure of diagonalization. If those ones are absent, the block is diagonal and the chain length is $1$.
Jordan form: what it is and what it means
Jordan form states that over $\mathbb{C}$, any matrix $A$ is similar \to a block diagonal matrix whose blocks are Jordan blocks $J_{\ell}(\lambda)$ for its eigenvalues. Similarity means $A = SJS^{-1}$ for an invertible $S$.
Jordan form simultaneously answers two questions.
- Which invariant subspaces are forced by the spectrum: the generalized eigenspaces $G_\lambda$.
- How the operator behaves inside each $G_\lambda$: a nilpotent part decomposed into Jordan chains.
The decomposition is not merely symbolic. It encodes the exact sizes of the chains, which control powers of $A$, matrix functions, and the dimensions of the kernels $\ker(A-\lambda I)^k$.
A concise way to remember the relationship between kernels and blocks:
- A Jordan block of size $\ell$ contributes one dimension \to $\ker(A-\lambda I)^k$ for each $k\ge 1$, until $k$ reaches $\ell$.
- So the sequence $\dim\ker(A-\lambda I)^k$ determines the multiset of Jordan block sizes.
Minimal polynomials: the invariant you can trust
Jordan form is not unique when eigenvalues repeat, because blocks can be permuted. But there are cleaner invariants that capture most of what you need without reconstructing $J$ explicitly.
The minimal polynomial $m_A(t)$ is the monic polynomial of least degree with $m_A(A)=0$. It divides the characteristic polynomial. Its factorization reveals chain lengths: for each eigenvalue $\lambda$, the exponent of $(t-\lambda)$ in $m_A$ is the size of the largest Jordan block for $\lambda$.
This has immediate consequences.
- $A$ is diagonalizable over $\mathbb{C}$ exactly when $m_A$ has no repeated linear factor, meaning each $(t-\lambda)$ appears to the first power.
- The degree of $m_A$ bounds the complexity of polynomials in $A$: every polynomial in $A$ reduces modulo $m_A$.
Minimal polynomials are also practical because they can be inferred from Krylov subspaces. The Krylov sequence $\{v, Av, A^2v,\dots\}$ spans an invariant subspace, and the first linear dependence gives a polynomial that annihilates that subspace. In numerical methods, this is the backbone of iterative solvers and eigenvalue techniques.
What survives similarity: trace, determinant, and more
Jordan form may look complicated, but similarity preserves several quantities that remain easy to compute.
| Similarity invariant | How to compute | What it tells you |
|—|—|—|
| trace $\operatorname{tr}(A)$ | sum of diagonal entries | sum of eigenvalues with multiplicity |
| determinant $\det(A)$ | product of pivots or eigenvalues | product of eigenvalues with multiplicity |
| characteristic polynomial $\chi_A(t)$ | $\det(tI-A)$ | eigenvalues and algebraic multiplicities |
| minimal polynomial $m_A(t)$ | smallest annihilating polynomial | largest Jordan block sizes |
| rank of $A-\lambda I$ | row reduction | geometric multiplicity via nullity |
These invariants are not substitutes for Jordan form, but they tell you what aspects of structure cannot change under any basis choice.
Powers and matrix functions: nilpotent terms you cannot ignore
On a generalized eigenspace $G_\lambda$, $A=\lambda I + N$ with nilpotent $N$. This gives explicit formulas for powers:
where $p$ is a nilpotency index with $N^p=0$. The binomial sum terminates because $N^j=0$ for large $j$.
This formula explains a qualitative distinction that diagonalization hides. When $N\ne 0$, powers of $A$ include polynomial factors in $k$ multiplying $\lambda^k$. Even if $|\lambda|<1$, the polynomial factor can delay decay; even if $|\lambda|=1$, the polynomial factor can produce growth in norm. The nilpotent part is the source of these polynomial terms.
Matrix functions behave similarly. For an analytic function $f$, one can define $f(J_\ell(\lambda))$ explicitly: it is an upper triangular Toeplitz matrix whose diagonals involve derivatives $f^{(j)}(\lambda)$. The presence of derivatives is another way to see that Jordan structure matters: nontrivial blocks force higher‑order information about the function at the eigenvalue.
A small example where diagonalization fails
Consider
The characteristic polynomial is $(t-1)^2$. The eigenspace $\ker(A-I)$ is one‑dimensional, spanned by $(1,0)$. So there is only one eigenvector, not enough to diagonalize.
But $A-I = \begin{pmatrix}0&1\\0&0\end{pmatrix}$ is nilpotent with $(A-I)^2=0$. Every vector is in the generalized eigenspace $G_1=\ker(A-I)^2$, and $A$ is already a Jordan block $J_2(1)$.
Compute powers:
The off‑diagonal entry grows linearly with $k$, an effect entirely driven by the nilpotent part. This is the simplest illustration of why Jordan structure changes long‑term behavior of repeated application, even when the eigenvalue is exactly $1$.
What Jordan form is not: orthogonality and numerical fragility
Jordan form is a classification under similarity, not an orthogonal decomposition. The change of basis matrix $S$ in $A=SJS^{-1}$ can be ill‑conditioned. That matters numerically: computing Jordan form from floating‑point data is often unstable, and tiny perturbations can change the Jordan structure dramatically when eigenvalues are close.
This does not make the theory useless. It changes how you use it.
- Use Jordan form as an exact algebraic classification when matrices are exact or symbolic.
- In numerical settings, prefer invariant subspaces and Schur form; they retain triangular structure with unitary changes of basis.
- Use minimal polynomials, ranks of $(A-\lambda I)^k$, and Krylov behavior to reason about structure without forcing an unstable canonical form.
The lesson is not to avoid Jordan ideas, but to separate algebraic truth from computational representation.
The core message
When diagonalization fails, linear algebra does not become chaotic. It becomes layered.
- The space decomposes into generalized eigenspaces, each tied \to a single eigenvalue.
- Inside each piece, the operator is a scalar plus a nilpotent map.
- Jordan chains describe the nilpotent structure, and Jordan blocks are their coordinate matrices.
- Minimal polynomials record the largest chain lengths and give a robust invariant.
- Many key quantities remain similarity invariants and are computable without finding a Jordan basis.
Invariant subspaces are the real objects: they tell you where the action lives, which pieces communicate, and how to restrict the problem. Jordan form is the sharpened statement of that structure. Understanding it equips you to reason about matrices beyond the comfortable diagonal world, while still staying within the disciplined geometry of linear maps.
Books by Drew Higgins
Leave a Reply