Jordan Normal Form: What Matrices Look Like When They Cannot Be Diagonalised

A linear transformation acts on a vector space. Sometimes it acts simply: there is a basis in which the transformation is just multiplication of each coordinate by a fixed eigenvalue. The matrix in that basis is diagonal, and the transformation is said to be diagonalisable. From a diagonal matrix you can read off everything — eigenvalues, powers, the matrix exponential — almost without computation.

But not every matrix can be diagonalised. The reason is structural: a matrix can fail to have enough independent eigenvectors to make up a basis. The simplest example is

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

whose only eigenvalue is $2$ (with algebraic multiplicity $2$), but whose eigenspace is one-dimensional: any eigenvector is a scalar multiple of $(1, 0)$. There is no second independent eigenvector, no diagonalising basis, no clean diagonal form.

What can be done with such a matrix? Camille Jordan answered the question in 1870. He showed that every square matrix over the complex numbers (or any algebraically closed field) is similar to a particularly simple, near-diagonal form called the Jordan canonical form, or Jordan normal form. The form is unique up to reordering of its blocks. From it, you can read off everything about the matrix’s linear-algebraic structure, just as you can from a diagonal matrix — and most of the practical questions you could ask reduce to questions about this canonical form. This article is about what Jordan form looks like, why it exists, and what it is good for.

The structure

A Jordan block $J_n(\lambda)$ is an $n \times n$ matrix with the eigenvalue $\lambda$ repeated on the diagonal and $1$s immediately above the diagonal:

$J_n(\lambda) = \begin{pmatrix} \lambda & 1 & 0 & \cdots & 0 \\ 0 & \lambda & 1 & \cdots & 0 \\ 0 & 0 & \lambda & \ddots & 0 \\ \vdots & \vdots & \vdots & \ddots & 1 \\ 0 & 0 & 0 & \cdots & \lambda \end{pmatrix}.$

A diagonal matrix is the special case where every Jordan block has size $1$. The $1$s above the diagonal are the obstruction to diagonalisability — they represent the failure of the eigenspace to fill out the full algebraic multiplicity of the eigenvalue.

The Jordan normal form theorem states: for any square matrix $A$ over an algebraically closed field, there exists an invertible matrix $P$ such that

$P^{-1} A P = J,$

where $J$ is block-diagonal with each diagonal block a Jordan block. The blocks are determined uniquely (up to ordering) by $A$, and they encode the eigenvalues of $A$ and the way the eigenvalues’ eigenspaces fall short of being independent bases.

The picture shows a typical Jordan form. Three blocks sit on the diagonal: a $3 \times 3$ block for eigenvalue $\lambda_1$ (with $1$s on the superdiagonal), a $2 \times 2$ block for $\lambda_2$, and a $1 \times 1$ block for $\lambda_3$. Everything outside the blocks is zero. The matrix is block-diagonal in a strong sense: it acts independently on three invariant subspaces, one for each block.

How blocks encode geometry

The $1$s above the diagonal are not just a typesetting choice; they have a precise geometric meaning. Apply $J_n(\lambda)$ to the standard basis vectors $e_1, e_2, \dots, e_n$:

$J_n(\lambda) \cdot e_1 = \lambda e_1, \qquad J_n(\lambda) \cdot e_k = \lambda e_k + e_{k-1} \;\; \text{for } k \geq 2.$

The first vector $e_1$ is an honest eigenvector. The subsequent vectors $e_2, \dots, e_n$ are generalised eigenvectors: they get sent to a scalar multiple of themselves plus the previous basis vector. Each generalised eigenvector “remembers” a piece of the eigenspace that the matrix has not yet been able to align. A Jordan block of size $n$ corresponds to a chain of $n$ generalised eigenvectors, exactly one of which is a true eigenvector and the others “hanging off” of it.

The diagonalisable case is when every block has size $1$ — there are no chains, just standalone eigenvectors. The most non-diagonalisable case is a single Jordan block of full size $n$ — one eigenvector at the bottom of a chain of $n$ generalised eigenvectors. The dimension of the eigenspace for $\lambda$ equals the number of Jordan blocks for $\lambda$, while the algebraic multiplicity equals the sum of sizes of those blocks.

Computing functions of matrices

The practical value of Jordan form becomes obvious when you need to compute a function $f$ of a matrix. The matrix exponential

$e^A = I + A + \tfrac{1}{2!}A^2 + \tfrac{1}{3!}A^3 + \cdots$

shows up in the general solution of linear ODE systems $\mathbf{y}' = A\mathbf{y}$, whose closed-form solution is $\mathbf{y}(t) = e^{At}\mathbf{y}(0)$. Computing $e^A$ from a Taylor series is hopeless for any matrix of nontrivial size.

Jordan form solves the problem. If $A = PJP^{-1}$, then $A^k = PJ^kP^{-1}$ for any $k$, and therefore

$e^A = P\,e^J\,P^{-1}.$

The exponential of a block-diagonal matrix is block-diagonal with the exponential of each block, so the problem reduces to computing $e^{J_n(\lambda)}$ for individual Jordan blocks. For a $2 \times 2$ block:

$e^{tJ_2(\lambda)} = e^{\lambda t}\begin{pmatrix} 1 & t \\ 0 & 1 \end{pmatrix}.$

For larger blocks, the formula generalises with higher powers of $t$ appearing on successive superdiagonals. The whole calculation, intractable on $A$, becomes routine on $J$. This is the standard route used in numerical analysis, control theory, and the analytical solution of linear ODE systems.

Theoretical consequences

Jordan form is also the key to several structural results in linear algebra.

Similarity classification. Two complex square matrices are similar (i.e., conjugate via an invertible matrix) if and only if they have the same Jordan form, up to ordering of blocks. This converts the question “are these two matrices conjugate?” — which is hard in general — into the much easier question “do they have the same Jordan structure?”

Minimal polynomial. The minimal polynomial of a matrix in Jordan form is the product $\prod (\lambda - \lambda_i)^{n_i}$, where the product runs over distinct eigenvalues and $n_i$ is the size of the largest Jordan block for $\lambda_i$. The Cayley–Hamilton theorem follows immediately: the characteristic polynomial is divisible by the minimal polynomial, so $p_A(A) = 0$.

Invariant subspaces. Each Jordan block corresponds to an invariant subspace of the matrix’s action. The whole vector space decomposes into a direct sum of these invariant subspaces, called the generalised eigenspaces. This decomposition is the foundation of the spectral theory of matrices.

Stability of dynamical systems. For a linear discrete-time system $\mathbf{x}_{n+1} = A\mathbf{x}_n$, asymptotic behaviour is controlled by the Jordan structure: the system is stable iff all eigenvalues have absolute value at most $1$, and strictly stable (all trajectories go to $0$) iff all have absolute value strictly less than $1$, with one additional caveat in the borderline case where Jordan blocks of size larger than $1$ can produce polynomial growth.

What it does not do

Jordan form has two important caveats.

It only works over algebraically closed fields. A matrix with real entries may have no Jordan form over the reals if its characteristic polynomial has complex roots; you must pass to complex matrices. The analogous real Jordan form uses $2 \times 2$ rotation-like blocks to handle complex-conjugate pairs.

It is numerically unstable. Small perturbations in $A$ can drastically change which Jordan blocks appear. A matrix that is theoretically a single Jordan block of size $5$ may, after rounding errors, look like a diagonal matrix with five very close eigenvalues — or anything in between. For practical numerical computation, the Schur form (upper triangular, achievable by unitary similarity) is preferred. Jordan form remains the cleanest theoretical tool, but is rarely computed numerically.

A small canonical form with enormous consequences

Jordan form is the kind of result that, once you have seen it, you start using everywhere. Almost every theorem in elementary linear algebra has a clean proof using Jordan form, and most of the structural results that don’t use it explicitly can be derived from it. The intuition that “every matrix can be decomposed into blocks acting independently on invariant subspaces” runs through dozens of subsequent results, from the spectral theorem for self-adjoint operators to the modern theory of representations of Lie algebras.

Camille Jordan’s 1870 Traité was a long, dense, ambitious book. Among its many contributions, the canonical form for matrices was perhaps not the most celebrated at the time. But a century and a half later, it is the result by which Jordan is most remembered in linear algebra, and it remains the standard tool for understanding what a linear transformation really does — not just what its eigenvalues are, but how it shears and stretches its space along the threads of generalised eigenvectors that the eigenvalues hint at but do not fully describe.

The Jordan form is, in one phrase, the structure of a linear operator up to similarity. Everything else in elementary linear algebra is more or less a corollary.

Frequently asked

Why does the Jordan form even need to exist? Aren't most matrices diagonalisable?

Most matrices with distinct eigenvalues are diagonalisable, but a matrix with a repeated eigenvalue may fail to have enough independent eigenvectors. The simplest example is the 2×2 matrix [[2, 1], [0, 2]], whose only eigenvalue is 2 with multiplicity 2 but whose eigenspace is one-dimensional. No basis of eigenvectors exists, so the matrix cannot be diagonalised. Jordan normal form is the closest you can get: it adds 1s on the superdiagonal of each block whose eigenspace falls short.

Who proved the theorem and when?

Camille Jordan published the canonical form in his 1870 Traité des substitutions et des équations algébriques, although ideas going in this direction had been developed earlier by Cauchy and Weierstrass. Jordan's particular contribution was the systematic statement and proof of the canonical form for matrices over the complex numbers (or any algebraically closed field), and the recognition of its central importance for the theory of linear transformations.

What is a Jordan block?

A Jordan block J_n(λ) is an n×n matrix with the eigenvalue λ repeated on the diagonal and 1s on the superdiagonal (immediately above the diagonal), with all other entries zero. The 1s are what distinguish a Jordan block from a diagonal matrix: they represent how the action of the matrix slightly 'shears' along the eigenspace, the obstruction to full diagonalisation. A 2×2 Jordan block looks like [[λ, 1], [0, λ]] and a 3×3 one like [[λ, 1, 0], [0, λ, 1], [0, 0, λ]].

What is Jordan form actually used for?

Three main things. First, computing functions of matrices — the matrix exponential e^A for solving systems of linear ODEs is much easier in Jordan form. Second, classifying linear operators up to similarity: two matrices are similar if and only if they have the same Jordan form (up to ordering of blocks). Third, theoretical results that need the fine structure of a linear operator — minimal polynomials, invariant subspaces, the Cayley–Hamilton theorem, and many control-theory problems all use Jordan form as a basic tool.

Continue reading

Blog

Eigenvalues and Eigenvectors: The Directions a Transformation Can't Turn

Most vectors get rotated when you apply a matrix. A special few only get stretched. Those directions — the eigenvectors — reveal the hidden skeleton of a linear transformation.

Read →Blog

The Cayley–Hamilton Theorem: Every Matrix Solves Its Own Equation

Compute the characteristic polynomial of a square matrix, then substitute the matrix itself into the polynomial. The result is zero — every single time. A strange, useful algebraic identity.

Read →Blog

Determinants: The Number That Measures How Much a Matrix Stretches Space

A determinant looks like an arbitrary recipe of multiply-and-subtract. Geometrically it is something simple and beautiful: the factor by which a transformation scales area or volume.

Read →