The Jordan normal form is a block-diagonal representation of a square matrix that describes its structure under a change of basis. Each block has a single eigenvalue on its diagonal, ones immediately above the diagonal, and zeros elsewhere. Every square matrix over the complex numbers has such a representation, and the blocks are uniquely determined apart from their order. More generally, the representation exists over any field in which the matrix’s characteristic polynomial splits into linear factors.
Jordan normal form is a central classification theorem in linear algebra. It extends diagonalization to matrices that do not possess a basis of ordinary eigenvectors, recording precisely how diagonalization fails. Its blocks determine matrix powers, matrix functions, and the polynomial factors that accompany exponential solutions of linear differential equations. Although indispensable for exact algebraic reasoning, it is highly sensitive to perturbations and is therefore usually replaced by more stable decompositions in numerical computation.
Definition and classification theorem
For a scalar and a positive integer , the Jordan block of size is
A block of size one is simply the scalar matrix . Writing separates its scalar part from a nilpotent matrix: , while . Here denotes the identity matrix of order .
Let be an matrix over a field . The Jordan normal form theorem states that if the characteristic polynomial of splits completely over , then there is an invertible matrix such that
The sizes satisfy . The same eigenvalue may occur in several blocks, and the block sizes associated with it need not be equal. Conversely, the existence of such a representation over implies that the characteristic polynomial splits over .
The matrix is unique up to a permutation of its blocks. The change-of-basis matrix , however, is generally far from unique. Thus the theorem classifies matrices under similarity, not under equality: two matrices over a common field, with split characteristic polynomials, are similar exactly when they have the same eigenvalues and the same multiset of block sizes for each eigenvalue. Neither the order of the blocks nor the choice of generalized eigenvectors adds a new invariant.
The columns of form a basis in which the linear transformation represented by acts in Jordan form. This interpretation distinguishes similarity from other matrix transformations. Similarity uses the same change of coordinates in the domain and codomain of an endomorphism; independent changes in the two spaces lead instead to matrix equivalence, whose classification is governed by rank.
Some texts place the ones immediately below, rather than above, the diagonal. These conventions give similar matrices, related by reversing the basis order within each block. With either convention, diagonalizability means exactly that every Jordan block has size one. A matrix with a block of size greater than one is often called defective.
Generalized eigenvectors and the decomposition of the space
An ordinary eigenvector for an eigenvalue is a nonzero vector satisfying . A generalized eigenvector satisfies
for some positive integer . If is the smallest such integer, the vector has order . Ordinary eigenvectors are therefore generalized eigenvectors of order one. Higher-order generalized eigenvectors supply the vectors missing when ordinary eigenvectors do not span the whole space.
For a fixed eigenvalue, the generalized eigenspace is
The exponent is always sufficient for an -dimensional space, although the sequence of kernels may stabilize much earlier. Unlike the ordinary eigenspace, the generalized eigenspace has dimension equal to the algebraic multiplicity of . It is invariant under , and the restriction of to it is nilpotent.
If the characteristic polynomial splits, the ambient vector space has the decomposition
where the sum runs over distinct eigenvalues. This is a direct sum: every vector has a unique expression as a sum of components from these generalized eigenspaces. One way to establish the decomposition uses the fact that powers of distinct linear factors are relatively prime. Polynomial identities obtained by the Euclidean algorithm then produce projections onto the individual components.
Inside , a Jordan chain of length is a sequence satisfying
with . These vectors are linearly independent. Indeed, applying a sufficiently high power of to a purported linear relation first forces the coefficient of the last vector to vanish; repeating the argument eliminates the remaining coefficients. In this ordered basis,
which is precisely the action of an upper-diagonal Jordan block.
The main existence step is consequently a theorem about nilpotent transformations: every nilpotent transformation on a finite-dimensional space admits a basis consisting of chains. The nested kernels
organize the construction. Compatible choices of chain tops, followed by repeated application of , furnish the required basis. Applying this nilpotent result separately to each generalized eigenspace proves the Jordan theorem. The nilpotent case does not itself require an algebraically closed field; all its eigenvalues are already zero.
Block sizes, multiplicities, and complete invariants
The Jordan blocks encode more information than either the eigenvalues or their multiplicities alone. For an eigenvalue , its algebraic multiplicity is its multiplicity as a root of the characteristic polynomial. Its geometric multiplicity is the dimension of . In Jordan form these become, respectively, the sum of the sizes of the blocks associated with and the number of those blocks. Consequently,
Equality on the right holds exactly when every block for that eigenvalue has size one.
The minimal polynomial is the monic polynomial of least degree such that . If is the size of the largest block belonging to , then
where is the algebraic multiplicity. Thus the characteristic polynomial measures the total dimension belonging to an eigenvalue, while the minimal polynomial measures the length of its longest chain. A matrix with split characteristic polynomial is diagonalizable exactly when its minimal polynomial has no repeated factor.
All the block sizes can be recovered from kernel dimensions. For a fixed , set
If the block sizes are , then
Each block contributes one new kernel dimension at every step until its full size is reached. It follows that counts blocks of size at least , and
counts blocks of size exactly . These identities also prove uniqueness of the Jordan block sizes without relying on any particular construction of a Jordan basis.
For example, suppose the kernel dimensions for one eigenvalue are
The successive increases are . There are three blocks altogether, two of size at least two, and one each of size at least three and four. The block sizes must therefore be . Their total is seven, and the largest size is four.
The same information is available from ranks, since the rank–nullity theorem gives . By contrast, the trace, determinant, and characteristic polynomial cannot distinguish different block structures with the same eigenvalues. Even the characteristic and minimal polynomials together need not suffice: blocks of sizes and blocks of sizes , all with eigenvalue zero, have characteristic polynomial and minimal polynomial , but are not similar.
Exact construction and a worked example
An exact computation usually begins by factoring the characteristic polynomial over the intended field. For each eigenvalue , the kernels of successive powers of determine the block sizes. Only after that discrete structure is known is it necessary to choose chain vectors. This separates two tasks that are sometimes conflated: determining the canonical matrix , and constructing a particular invertible matrix satisfying .
Consider
Because is block triangular, its characteristic polynomial is
On the subspace spanned by the first three standard coordinate vectors, satisfies
Thus has rank one on this subspace and its square is zero. The kernel dimensions associated with the eigenvalue are and ; the fourth coordinate contributes no kernel because acts there as multiplication by one. Hence the blocks for have sizes two and one.
A length-two chain is
because and . An independent eigenvector for the remaining size-one block is . The fourth standard vector is an eigenvector for . Therefore
gives
The minimal polynomial is , and is not diagonalizable despite having a two-dimensional ordinary eigenspace for its repeated eigenvalue.
For larger problems, arbitrary choices of generalized eigenvectors can produce dependent chains. Selecting chain tops therefore requires compatibility with both previously selected chains and the kernel filtration. One approach starts with the longest required chains and then supplies shorter ones to fill the remaining dimensions. Another organizes the choices through quotient spaces derived from the kernels. In either approach, solving the chain equations is a sequence of exact linear systems, typically handled by Gaussian elimination.
Direct verification is valuable because it avoids explicitly computing an inverse: checking that is nonsingular and that proves the result. Exact symbolic computation may nevertheless become costly when eigenvalues lie in complicated algebraic extensions, even though the underlying linear-algebraic steps are finite.
Matrix powers, analytic functions, and resolvents
Jordan form reduces evaluating many matrix functions to evaluating scalar functions and their derivatives. Let be a block of size . Since , the binomial theorem gives, for a nonnegative integer ,
When , entries on successive superdiagonals involve polynomial factors in , multiplied by powers of . When , the block is nilpotent and its th power vanishes for every . These are fundamentally different long-term behaviors.
More generally, if is an analytic function on neighborhoods of the eigenvalues, then
The scalar Taylor expansion terminates after substitution because all higher powers of are zero. For a polynomial, the identity holds without analytic assumptions. For a general matrix,
Although is not unique, this definition is independent of the chosen Jordan basis.
For example, the matrix exponential satisfies
For a block of size three this is
The polynomial factors arise from generalized eigenvectors rather than from additional eigenvalues. For an invertible block, applying the same principle to produces a finite expansion for its inverse.
Jordan blocks also describe singularities of the resolvent. For ,
Thus the resolvent has a pole of order at the eigenvalue of a single block. For a full matrix, the pole order at is the size of its largest block, not its algebraic multiplicity.
These formulas connect matrix functions with polynomial interpolation. To reproduce , a polynomial need only match and its derivatives through order at each eigenvalue, where is the largest relevant block size. The resulting Hermite interpolation conditions determine a unique polynomial of degree less than . Consequently, the minimal polynomial specifies exactly how much scalar derivative information is required.
Linear differential equations and discrete dynamics
For a constant matrix , the homogeneous ordinary differential equation
has solution . The substitution transforms the equation into , so each Jordan block can be solved independently. A block of size associated with contributes terms involving
The maximal polynomial degree attached to an eigenvalue is therefore one less than its largest block size.
For a concrete example,
produces
Both coordinates tend to zero, but the coupling produces a polynomial factor absent from a diagonal system. For a larger off-diagonal coupling, some initial conditions can exhibit substantial transient amplification even though every solution eventually decays. Jordan structure identifies the polynomial mechanism, while the geometry of the coordinate transformation also affects amplification measured in a fixed norm.
For continuous-time linear systems, all solutions tend to zero precisely when every eigenvalue has strictly negative real part. Polynomial factors cannot overcome exponential decay. The equilibrium at the origin is stable in the sense that sufficiently small initial states remain small for all precisely when all eigenvalues have nonpositive real part and every block associated with an eigenvalue on the imaginary axis has size one. Such an eigenvalue is called semisimple. A larger block on the imaginary axis creates polynomial growth.
For a discrete-time dynamical system,
the solution is . All solutions tend to zero precisely when every eigenvalue has modulus less than one. The powers remain bounded precisely when all eigenvalues have modulus at most one and those on the unit circle are semisimple. These conditions apply directly to linear recurrence relations after conversion to a first-order vector equation.
Jordan blocks also clarify why boundedness assumptions constrain the spectrum. Powers of a finite stochastic transition matrix remain bounded. Consequently, any eigenvalue on the unit circle must have only size-one Jordan blocks; otherwise the block-power formula would force polynomial growth. Blocks associated with eigenvalues strictly inside the unit circle may be larger, producing decaying polynomial transients. This distinction matters in the analysis of finite Markov chains, where convergence rates depend not only on eigenvalue moduli but also on nontrivial block structure and the chosen coordinates.
Ground fields, real forms, and rational canonical form
The existence of an ordinary Jordan form depends on the field of scalars. Over the complex numbers, the fundamental theorem of algebra ensures that every characteristic polynomial splits. Over the real numbers, a matrix can have nonreal eigenvalues and therefore no real Jordan form consisting only of scalar-eigenvalue blocks. For example,
has characteristic polynomial . It is diagonalizable over the complex numbers, with eigenvalues and , but has no real eigenvector.
A real Jordan form accommodates conjugate pairs using two-dimensional diagonal blocks. If , with , is an eigenvalue, define
A pair of conjugate complex Jordan blocks of size corresponds to a real block of size , with on its block diagonal and on its first block superdiagonal:
Conventions differing in the signs of the off-diagonal entries of are equivalent by a real change of basis.
Over an arbitrary field, the general replacement is the rational canonical form, also called Frobenius normal form. It uses companion matrices of monic polynomials rather than requiring eigenvalues to belong to the field. Its invariant factors satisfy a divisibility chain, and their product is the characteristic polynomial; the largest invariant factor is the minimal polynomial. This makes it particularly useful for rational matrices whose eigenvalues are irrational or otherwise require a field extension.
The conceptual foundation is the structure theorem for finitely generated modules over a principal ideal domain. A transformation makes into a module over by defining
Because the vector space is finite-dimensional, this module is finitely generated and torsion. Decomposing it into cyclic modules yields rational canonical form. Decomposing further into primary cyclic modules produces factors of the form , where is irreducible.
When every relevant irreducible polynomial is linear, these elementary divisors become , and each corresponds to a Jordan block of size . Thus Jordan form is the split-polynomial version of a more general classification rather than a separate phenomenon. The form is named after Camille Jordan, whose 1870 Traité des substitutions et des équations algébriques is a central historical source for the canonical-form tradition. Its modern module-theoretic interpretation places the result within abstract algebra and explains why related canonical forms work over fields where eigenvalues are unavailable.
Similarity, commuting transformations, and spectral structure
Jordan form does more than provide a convenient representative: it specifies exactly when two finite-dimensional transformations are equivalent after coordinates are changed. Eigenvalues alone do not supply this classification. For instance,
have identical characteristic polynomials but different eigenspace dimensions, so they cannot be similar. The block partition at each eigenvalue is the additional information required.
The same structure controls matrices that commute with . The set
is called its centralizer or commutant. If is a single Jordan block, every commuting matrix is a polynomial in of degree less than the block size. In the standard Jordan basis these are upper-triangular matrices constant along each superdiagonal. Multiple blocks allow additional maps between chains, so the centralizer generally becomes larger.
For blocks of sizes and with the same eigenvalue, the space of rectangular matrices satisfying
has dimension . If the eigenvalues are distinct, the corresponding intertwining space is zero. Hence, if are the block sizes for ,
This formula measures how many linear transformations preserve the action of , and it is useful in studying similarity classes and their symmetries.
Another consequence is the additive decomposition
where is diagonalizable and is nilpotent. In Jordan coordinates, replaces each block by , while retains its superdiagonal ones. When the characteristic polynomial splits, both parts can be expressed as polynomials in . This is the matrix form of the Jordan–Chevalley decomposition, separating the eigenvalue data from the nontrivial chain structure.
Jordan diagonalization must also be distinguished from orthogonal or unitary diagonalization. The Jordan theorem permits any invertible change of basis; it imposes no preservation of lengths or angles. By contrast, the spectral theorem states that a complex normal matrix admits diagonalization by a unitary matrix. Such a matrix has no nontrivial Jordan blocks. Every complex matrix has a Jordan form, but only the normal matrices have an orthonormal eigenbasis. This distinction explains why spectral decompositions preserve geometric information that an arbitrary Jordan basis may distort.
Perturbations and numerical computation
Jordan block structure is an exact algebraic invariant but not a continuous function of the entries of a matrix. An arbitrarily small perturbation can split a repeated eigenvalue and replace a nontrivial block by several size-one blocks. For example, consider
At , this is a size-two nilpotent Jordan block. For every nonzero , its complex eigenvalues are the distinct numbers and , so it is diagonalizable. Its Jordan structure therefore changes discontinuously at zero.
The sensitivity becomes more pronounced for larger blocks. Starting with the nilpotent block , add a perturbation in its lower-left corner. The resulting characteristic polynomial is
Its eigenvalues have modulus . Thus a perturbation of size can cause eigenvalue displacements on the much larger scale . This example explains why multiple eigenvalues associated with long chains are especially sensitive; it is a worst-case perturbation mechanism, not a claim that every perturbation produces the same displacement.
The Jordan basis can also become poorly conditioned near a defective matrix. In the two-dimensional example, eigenvectors for the two perturbed eigenvalues can be chosen as
As approaches zero, these vectors become nearly parallel. The matrix formed from them then has a growing condition number, amplifying the effects of floating-point arithmetic in coordinate transformations.
These difficulties are not simply shortcomings of a particular algorithm. Determining exact block sizes requires deciding whether certain ranks drop exactly, or whether particular quantities are exactly zero. Rounded data generally do not contain that information. Tolerances can identify a plausible nearby structure, but different tolerance choices may imply different blocks. Exact symbolic inputs and numerical approximations therefore call for different computational interpretations.
For routine numerical linear algebra, the Schur decomposition
is usually preferable: is unitary and is upper triangular over the complex numbers. Its real counterpart uses an upper quasi-triangular matrix with one- and two-dimensional diagonal blocks. Unlike an arbitrary Jordan change of basis, a unitary transformation has condition number one in the Euclidean operator norm, making Schur form a practical foundation for computations designed for numerical stability.
The QR algorithm computes Schur forms without requiring decisions about exact Jordan chains. Questions about near-defectiveness are instead studied through perturbation theory, eigenvalue conditioning, and the pseudospectrum, which describes how the spectrum may move under bounded perturbations. Jordan form remains appropriate for exact classification and proofs; Schur form and related decompositions serve computations in which the data and arithmetic are inexact.