Jordan Normal Form

⚡ Key Takeaways

Jordan Normal Form (JNF): every square matrix over an algebraically closed field is similar to a block-diagonal matrix built from "Jordan blocks" — nearly-diagonal pieces with eigenvalues on the diagonal and 1's on the superdiagonal. It is the closest a non-diagonalizable matrix can get to being diagonal. Applies wherever linear operators must be classified up to similarity: ODE systems, matrix functions, control theory, dynamical systems. Established by Camille Jordan (1870).

💡 The Core Insight

Diagonalization fails when an operator does not have "enough" eigenvectors. The fix is to enlarge the notion of eigenvector to generalized eigenvectors — vectors $v$ killed by some power $(A-\lambda I)^k$ rather than by $(A-\lambda I)$ itself. These organize into Jordan chains, where each vector maps (under $A-\lambda I$ ) to the previous one. Each chain becomes one Jordan block. The deep statement is that the whole space decomposes into independent chains, so any operator is, up to a change of basis, a disjoint union of these elementary "shift-plus-scalar" pieces.

DefinitionJordan Block

A Jordan block of size $k$ with eigenvalue $\lambda$ is the $k \times k$ matrix

1

\quad J_k(\lambda) = \begin{pmatrix} \lambda & 1 & & \\ & \lambda & 1 & \\ & & \ddots & \ddots \\ & & & \lambda & 1\\ & & & & \lambda \end{pmatrix}

Derivation hint: place $\lambda$ on the diagonal, $1$ on the superdiagonal, $0$ everywhere else; $J_1(\lambda)=(\lambda)$ is the smallest case. Here $\lambda$ is the eigenvalue, $k$ is the block size, the $1$ 's sit on the entries $(i,i+1)$ , and all other entries are $0$ .

DefinitionJordan Normal Form

A matrix $J$ is in Jordan Normal Form if it is block-diagonal with Jordan blocks:

2

\quad J = \bigoplus_{i=1}^{m} J_{k_i}(\lambda_i) = \begin{pmatrix} J_{k_1}(\lambda_1) & & \\ & \ddots & \\ & & J_{k_m}(\lambda_m) \end{pmatrix}

Derivation hint: arrange the individual blocks $J_{k_i}(\lambda_i)$ along the diagonal of a larger matrix, zeros off the blocks. The $\lambda_i$ need not be distinct (an eigenvalue may own several blocks), and $\sum_i k_i = n$ , the matrix dimension.

TheoremExistence and Uniqueness of JNF

Let $V$ be a finite-dimensional vector space over an algebraically closed field $\mathbb{F}$ , and let $A: V \to V$ be linear. Then there is a basis of $V$ in which $A$ is represented by a matrix $J$ in Jordan Normal Form. This $J$ is unique up to the order of the blocks. Equivalently, every square matrix over $\mathbb{F}$ is similar to such a $J$ :

3

\quad P^{-1} A P = J

Derivation hint: $P$ is the matrix whose columns are the Jordan-chain basis vectors; reading off how $A$ acts on those basis vectors reproduces equation [2].

TheoremBlock-Counting Formula

For eigenvalue $\lambda$ , the number of Jordan blocks of size exactly $j$ is determined by ranks of powers of $N = A - \lambda I$ :

4

\quad b_j(\lambda) = \operatorname{rank}(N^{j-1}) - 2\operatorname{rank}(N^{j}) + \operatorname{rank}(N^{j+1})

Derivation hint: $\dim\ker(N^j)-\dim\ker(N^{j-1})$ counts blocks of size $\ge j$ ; differencing that quantity again isolates blocks of size exactly $j$ . Use $\operatorname{rank}=n-\dim\ker$ . The number of blocks for $\lambda$ (total) equals $\dim\ker(A-\lambda I)$ , the geometric multiplicity; the total size of all $\lambda$ -blocks equals the algebraic multiplicity.

ProofDecomposition into Jordan chains

We sketch the standard constructive proof over $\mathbb{C}$ .

Step 1 — Reduce to a single eigenvalue. By the Primary Decomposition Theorem, if the characteristic polynomial factors as $\prod_i (x-\lambda_i)^{m_i}$ , then

5

\quad V = \bigoplus_i \ker\big((A-\lambda_i I)^{m_i}\big)

Derivation hint: the factors $(x-\lambda_i)^{m_i}$ are pairwise coprime, so Bézout gives idempotent projections summing to the identity; each projects onto a generalized eigenspace, which is $A$ -invariant. Each summand $V_i = \ker((A-\lambda_i I)^{m_i})$ is $A$ -invariant, so we may work in one $V_i$ at a time.

Step 2 — Reduce to a nilpotent operator. On $V_i$ set $N = A - \lambda_i I$ . Then $N$ is nilpotent: $N^{m_i} = 0$ on $V_i$ . It suffices to find a Jordan basis for $N$ , since adding back $\lambda_i I$ turns each nilpotent Jordan block $J_k(0)$ into $J_k(\lambda_i)$ .

Step 3 — Jordan basis for a nilpotent $N$ . Consider the descending chain of kernels:

6

\quad \{0\} \subsetneq \ker N \subsetneq \ker N^2 \subsetneq \cdots \subsetneq \ker N^p = V_i

Derivation hint: $\ker N^{j}\subseteq \ker N^{j+1}$ always; for nilpotent $N$ the chain strictly grows until it fills $V_i$ at the index of nilpotency $p$ . Choose a basis of $V_i$ adapted to this filtration from the top down. Take vectors completing $\ker N^{p-1}$ to $\ker N^{p}$ ; apply $N$ repeatedly to each to generate a Jordan chain

7

\quad v,\; Nv,\; N^2 v,\;\dots,\; N^{k-1}v \quad (N^k v = 0)

Derivation hint: $N$ maps $\ker N^{j}$ into $\ker N^{j-1}$ , so each chain descends one level per application; chains are independent because their leading vectors are chosen linearly independent modulo lower kernels.

Step 4 — Read off the block. On the chain in [7], $N$ sends each basis vector to the next, so in the basis $(N^{k-1}v,\dots,Nv,v)$ the operator $N$ is exactly $J_k(0)$ . Collecting all chains across all $V_i$ gives the full basis $P$ and equation [2]. $\qquad\blacksquare$

Uniqueness follows from [4]: the block sizes are forced by rank invariants of $A$ , which are unchanged by similarity.

∎

CJ

Camille Jordan

1838–1922

Marie Ennemond Camille Jordan was a French mathematician and engineer whose 1870 Traité des substitutions et des équations algébriques introduced the canonical form now bearing his name, in the context of studying linear substitutions over finite fields and the solvability of equations. (He is also responsible for the Jordan curve theorem and Jordan measure — a different, frequently confused Jordan from Wilhelm Jordan of Gauss–Jordan elimination.) His treatise was a foundational text in group theory and bridged Galois's ideas with concrete matrix computation.

🧠 Intuition

Think of a Jordan block $J_k(\lambda)$ as a scalar $\lambda$ plus a one-step shift. Write $J_k(\lambda) = \lambda I + N$ where $N$ is the "shift" sending basis vector $e_{i+1}\mapsto e_i$ and $e_1 \mapsto 0$ . Applying $N$ once "moves you down the chain"; applying it $k$ times annihilates everything. So a Jordan block is the algebraic incarnation of a finite conveyor belt: stuff enters at the top, slides down one notch each time $N$ acts, and falls off the bottom after $k$ steps. Diagonalizability is the special case where every belt has length 1 — every generalized eigenvector is already a genuine eigenvector. A nontrivial Jordan block of size $\ge 2$ is precisely the obstruction to diagonalizing: it signals an eigenvalue whose geometric multiplicity (number of belts) is less than its algebraic multiplicity (total belt length).

Figure

A Jordan chain visualized as a conveyor belt under N = A − λI

ExampleA 3×3 with a defective eigenvalue

Given.

A = \begin{pmatrix} 2 & 1 & 0 \\ 0 & 2 & 0 \\ 0 & 0 & 3 \end{pmatrix}

Step 1 — Characteristic polynomial. $\det(A-xI) = (2-x)^2(3-x)$ , so eigenvalues are $\lambda=2$ (algebraic multiplicity 2) and $\lambda=3$ (algebraic multiplicity 1).

Step 2 — Geometric multiplicity of $\lambda=2$ . Compute $N=A-2I=\begin{pmatrix}0&1&0\\0&0&0\\0&0&1\end{pmatrix}$ , $\operatorname{rank}(N)=2$ , so $\dim\ker N = 3-2 = 1$ . Only one eigenvector for $\lambda=2$ , yet algebraic multiplicity is 2 → one Jordan block of size 2.

Step 3 — Block counting via [4]. For $\lambda=3$ : geometric mult $=1$ , alg mult $=1$ → one size-1 block.

Step 4 — Assemble.

\boxed{\,J = \begin{pmatrix} 2 & 1 & 0 \\ 0 & 2 & 0 \\ 0 & 0 & 3 \end{pmatrix} = J_2(2)\oplus J_1(3)\,}

(Here $A$ is already in Jordan form.)

ExampleComputing

e^{At}

via JNF

Given. $A = \begin{pmatrix} 5 & 1 \\ -1 & 3 \end{pmatrix}$ , find $e^{At}$ .

Step 1 — Eigenvalues. $\det(A-xI)=(5-x)(3-x)+1=x^2-8x+16=(x-4)^2$ . Single eigenvalue $\lambda=4$ , alg mult 2.

Step 2 — Geometric multiplicity. $N=A-4I=\begin{pmatrix}1&1\\-1&-1\end{pmatrix}$ , $\operatorname{rank}=1$ , $\dim\ker N=1$ . Defective → one $J_2(4)$ block.

Step 3 — Jordan basis. Eigenvector $v_1$ : solve $Nv_1=0 \Rightarrow v_1=(1,-1)^T$ . Generalized eigenvector $v_2$ with $Nv_2=v_1$ : $\begin{pmatrix}1&1\\-1&-1\end{pmatrix}v_2=(1,-1)^T \Rightarrow v_2=(1,0)^T$ . So $P=\begin{pmatrix}1&1\\-1&0\end{pmatrix}$ and $J=\begin{pmatrix}4&1\\0&4\end{pmatrix}$ .

Step 4 — Exponentiate the block. For $J_2(\lambda)=\lambda I+N_0$ with $N_0^2=0$ :

e^{Jt}=e^{4t}\begin{pmatrix}1&t\\0&1\end{pmatrix}.

Step 5 — Transform back. $e^{At}=P\,e^{Jt}P^{-1}$ with $P^{-1}=\begin{pmatrix}0&-1\\1&1\end{pmatrix}$ :

\boxed{\,e^{At}=e^{4t}\begin{pmatrix}1+t & t \\ -t & 1-t\end{pmatrix}\,}

(units: dimensionless matrix, $t$ in the same time units as the underlying ODE $\dot{x}=Ax$ ).

⚠️ Common mistake

JNF over $\mathbb{R}$ may not exist as written. The existence theorem requires an algebraically closed field. A real matrix with complex eigenvalues (e.g. a rotation) has no real Jordan form with real $\lambda$ 's on the diagonal; one uses the real Jordan form with $2\times2$ rotation-scaling blocks instead. Also: the Jordan form is not a continuous function of the matrix entries — an arbitrarily small perturbation can split a Jordan block into distinct eigenvalues, which is exactly why JNF is numerically ill-conditioned and rarely computed in floating-point practice.

Counterexample

A common error: "algebraic multiplicity equals the size of the largest Jordan block." Consider $\lambda$ with algebraic multiplicity 4 but Jordan structure $J_2(\lambda)\oplus J_1(\lambda)\oplus J_1(\lambda)$ . The algebraic multiplicity is $2+1+1=4$ (sum of all block sizes), the geometric multiplicity is $3$ (number of blocks), and the largest block has size $2$ — which is the multiplicity of $(x-\lambda)$ in the minimal polynomial, not the characteristic polynomial. Conflating these three numbers is the single most frequent JNF mistake.

connection

JNF sits at the heart of the structure theory of linear operators. It is the special case (over an algebraically closed field) of the Rational Canonical Form, which works over any field by using companion blocks of invariant factors. The block sizes are governed by the Minimal Polynomial (largest block per eigenvalue) and the Characteristic Polynomial (total block size per eigenvalue). The diagonalizable case is exactly when JNF reduces to the Spectral Theorem situation of distinct eigenspaces. Abstractly, JNF is the manifestation of the Structure Theorem for Finitely Generated Modules over a PID applied to $V$ as a module over $\mathbb{F}[x]$ , where the Jordan blocks correspond to primary cyclic summands $\mathbb{F}[x]/(x-\lambda)^k$ .

🔬 Where this appears

⏳ History

- 1858 — Cayley and Sylvester develop the theory of matrices; the Cayley–Hamilton theorem hints at minimal/characteristic polynomial structure.

- 1868 — Karl Weierstrass introduces elementary divisors and a canonical form for matrix pencils, anticipating the invariant-factor viewpoint.

- 1870 — Camille Jordan publishes Traité des substitutions, defining the canonical form over finite fields in connection with Galois theory.

- 1878 — Georg Frobenius unifies and clarifies the invariant-factor and elementary-divisor theories, giving the rational canonical form.

- 1880s — The module-over-PID interpretation crystallizes through the work on finitely generated abelian groups (Frobenius, Stickelberger).

- 1930s — Jordan–Chevalley decomposition formalized in the theory of algebraic groups and Lie algebras.

- 1965 — Numerical analysts (Wilkinson) document the extreme ill-conditioning of JNF computation, steering applied practice toward the Schur form.

- 20th–21st c. — Exceptional points in non-Hermitian physics revive interest in Jordan blocks as physical, observable phenomena rather than mere algebraic curiosities.

open-problem

The chief practical limitation of Jordan Normal Form is numerical instability: because Jordan structure is destroyed by generic perturbations, no backward-stable algorithm computes it from inexact data. This raises the still-active question of how to define a meaningful numerical Jordan form — work by Golub, Wilkinson, and others on staircase algorithms and the Kronecker canonical form addresses it partially, but a fully robust, parameter-free method remains elusive. Related open territory includes the structured perturbation theory of exceptional points (how Jordan blocks deform under physically constrained perturbations) and effective bounds on the conditioning of the similarity matrix $P$ , which can be exponentially large in the block size and is central to controlling errors in matrix-function computation.

Formal Statement

Why It Holds / Derivation

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also