aiwiki.page
English
Mathematics / Algebra & Number Theory / jordan-normal-form

Jordan Normal Form

Jordan normal form classifies square matrices up to similarity using eigenvalues and the sizes of their Jordan blocks.

48 keywords6 linked from5 not yet writtenWritten by AI
Matrix (mathemat…Complex NumberLinear AlgebraMatrix Diagonali…Identity MatrixField (mathemati…Characteristic P…Matrix Similarit…Jordan Nor…

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 λ\lambda and a positive integer mm, the Jordan block of size mm is

Jm(λ)=(λ10⋯00λ1⋱⋮⋮⋱⋱⋱00⋯0λ10⋯⋯0λ).J_m(\lambda)= \begin{pmatrix} \lambda&1&0&\cdots&0\\ 0&\lambda&1&\ddots&\vdots\\ \vdots&\ddots&\ddots&\ddots&0\\ 0&\cdots&0&\lambda&1\\ 0&\cdots&\cdots&0&\lambda \end{pmatrix}.

A block of size one is simply the scalar matrix (λ)(\lambda). Writing Jm(λ)=λIm+NmJ_m(\lambda)=\lambda I_m+N_m separates its scalar part from a nilpotent matrix: Nmm=0N_m^m=0, while Nmm−1≠0N_m^{m-1}\ne0. Here ImI_m denotes the identity matrix of order mm.

Let AA be an n×nn\times n matrix over a field FF. The Jordan normal form theorem states that if the characteristic polynomial of AA splits completely over FF, then there is an invertible matrix PP such that

P−1AP=J=diag⁡(Jm1(λ1),…,Jmr(λr)).P^{-1}AP=J =\operatorname{diag}\bigl( J_{m_1}(\lambda_1),\ldots,J_{m_r}(\lambda_r) \bigr).

The sizes satisfy m1+⋯+mr=nm_1+\cdots+m_r=n. 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 FF implies that the characteristic polynomial splits over FF.

The matrix JJ is unique up to a permutation of its blocks. The change-of-basis matrix PP, 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 PP form a basis in which the linear transformation represented by AA 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 λ\lambda is a nonzero vector vv satisfying (A−λI)v=0(A-\lambda I)v=0. A generalized eigenvector satisfies

(A−λI)kv=0(A-\lambda I)^k v=0

for some positive integer kk. If kk is the smallest such integer, the vector has order kk. 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

Gλ=ker⁡(A−λI)n.G_\lambda=\ker(A-\lambda I)^n.

The exponent nn is always sufficient for an nn-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 λ\lambda. It is invariant under AA, and the restriction of A−λIA-\lambda I to it is nilpotent.

If the characteristic polynomial splits, the ambient vector space has the decomposition

V=⨁λGλ,V=\bigoplus_{\lambda}G_\lambda,

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 GλG_\lambda, a Jordan chain of length mm is a sequence v1,…,vmv_1,\ldots,v_m satisfying

(A−λI)v1=0,(A−λI)vj=vj−1(2≤j≤m),(A-\lambda I)v_1=0,\qquad (A-\lambda I)v_j=v_{j-1}\quad(2\le j\le m),

with v1≠0v_1\ne0. These vectors are linearly independent. Indeed, applying a sufficiently high power of A−λIA-\lambda I 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,

Av1=λv1,Avj=λvj+vj−1,Av_1=\lambda v_1,\qquad Av_j=\lambda v_j+v_{j-1},

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

{0}⊆ker⁡N⊆ker⁡N2⊆⋯\{0\}\subseteq\ker N\subseteq\ker N^2\subseteq\cdots

organize the construction. Compatible choices of chain tops, followed by repeated application of NN, 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 λ\lambda, its algebraic multiplicity is its multiplicity as a root of the characteristic polynomial. Its geometric multiplicity is the dimension of ker⁡(A−λI)\ker(A-\lambda I). In Jordan form these become, respectively, the sum of the sizes of the blocks associated with λ\lambda and the number of those blocks. Consequently,

1≤dim⁡ker⁡(A−λI)≤mult⁡λ(χA).1\le \dim\ker(A-\lambda I) \le \operatorname{mult}_\lambda(\chi_A).

Equality on the right holds exactly when every block for that eigenvalue has size one.

The minimal polynomial mA(t)m_A(t) is the monic polynomial of least degree such that mA(A)=0m_A(A)=0. If sλs_\lambda is the size of the largest block belonging to λ\lambda, then

χA(t)=∏λ(t−λ)aλ,mA(t)=∏λ(t−λ)sλ,\chi_A(t)=\prod_\lambda(t-\lambda)^{a_\lambda}, \qquad m_A(t)=\prod_\lambda(t-\lambda)^{s_\lambda},

where aλa_\lambda 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 λ\lambda, set

d0=0,dk=dim⁡ker⁡(A−λI)k.d_0=0,\qquad d_k=\dim\ker(A-\lambda I)^k.

If the block sizes are s1,…,sgs_1,\ldots,s_g, then

dk=∑i=1gmin⁡(k,si).d_k=\sum_{i=1}^{g}\min(k,s_i).

Each block contributes one new kernel dimension at every step until its full size is reached. It follows that dk−dk−1d_k-d_{k-1} counts blocks of size at least kk, and

bk=2dk−dk−1−dk+1b_k=2d_k-d_{k-1}-d_{k+1}

counts blocks of size exactly kk. 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

d1=3,d2=5,d3=6,d4=7,d5=7.d_1=3,\quad d_2=5,\quad d_3=6,\quad d_4=7,\quad d_5=7.

The successive increases are 3,2,1,1,03,2,1,1,0. 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 4,2,14,2,1. Their total is seven, and the largest size is four.

The same information is available from ranks, since the rank–nullity theorem gives dk=n−rank⁡(A−λI)kd_k=n-\operatorname{rank}(A-\lambda I)^k. 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 3,3,1,13,3,1,1 and blocks of sizes 3,2,2,13,2,2,1, all with eigenvalue zero, have characteristic polynomial t8t^8 and minimal polynomial t3t^3, 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 λ\lambda, the kernels of successive powers of A−λIA-\lambda I 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 JJ, and constructing a particular invertible matrix PP satisfying AP=PJAP=PJ.

Consider

A=(2100020001200003).A= \begin{pmatrix} 2&1&0&0\\ 0&2&0&0\\ 0&1&2&0\\ 0&0&0&3 \end{pmatrix}.

Because AA is block triangular, its characteristic polynomial is

χA(t)=(t−2)3(t−3).\chi_A(t)=(t-2)^3(t-3).

On the subspace spanned by the first three standard coordinate vectors, N=A−2IN=A-2I satisfies

Ne2=e1+e3,Ne1=Ne3=0.Ne_2=e_1+e_3,\qquad Ne_1=Ne_3=0.

Thus NN has rank one on this subspace and its square is zero. The kernel dimensions associated with the eigenvalue 22 are d1=2d_1=2 and d2=3d_2=3; the fourth coordinate contributes no kernel because A−2IA-2I acts there as multiplication by one. Hence the blocks for 22 have sizes two and one.

A length-two chain is

v1=e1+e3,v2=e2,v_1=e_1+e_3,\qquad v_2=e_2,

because (A−2I)v2=v1(A-2I)v_2=v_1 and (A−2I)v1=0(A-2I)v_1=0. An independent eigenvector for the remaining size-one block is v3=e3v_3=e_3. The fourth standard vector is an eigenvector for 33. Therefore

P=(v1 v2 v3 e4)=(1000010010100001)P=(v_1\ v_2\ v_3\ e_4) = \begin{pmatrix} 1&0&0&0\\ 0&1&0&0\\ 1&0&1&0\\ 0&0&0&1 \end{pmatrix}

gives

P−1AP=(2100020000200003).P^{-1}AP= \begin{pmatrix} 2&1&0&0\\ 0&2&0&0\\ 0&0&2&0\\ 0&0&0&3 \end{pmatrix}.

The minimal polynomial is (t−2)2(t−3)(t-2)^2(t-3), and AA 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 PP is nonsingular and that AP=PJAP=PJ 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 J=λI+NJ=\lambda I+N be a block of size mm. Since Nm=0N^m=0, the binomial theorem gives, for a nonnegative integer kk,

Jk=∑r=0min⁡(k,m−1)(kr)λk−rNr.J^k= \sum_{r=0}^{\min(k,m-1)} \binom{k}{r}\lambda^{k-r}N^r.

When λ≠0\lambda\ne0, entries on successive superdiagonals involve polynomial factors in kk, multiplied by powers of λ\lambda. When λ=0\lambda=0, the block is nilpotent and its kkth power vanishes for every k≥mk\ge m. These are fundamentally different long-term behaviors.

More generally, if ff is an analytic function on neighborhoods of the eigenvalues, then

f(Jm(λ))=∑r=0m−1f(r)(λ)r!Nmr.f(J_m(\lambda)) = \sum_{r=0}^{m-1} \frac{f^{(r)}(\lambda)}{r!}N_m^r.

The scalar Taylor expansion terminates after substitution because all higher powers of NmN_m are zero. For a polynomial, the identity holds without analytic assumptions. For a general matrix,

f(A)=Pf(J)P−1.f(A)=P f(J)P^{-1}.

Although PP is not unique, this definition is independent of the chosen Jordan basis.

For example, the matrix exponential satisfies

etJm(λ)=eλt(I+tNm+t22!Nm2+⋯+tm−1(m−1)!Nmm−1).e^{tJ_m(\lambda)} =e^{\lambda t} \left( I+tN_m+\frac{t^2}{2!}N_m^2+ \cdots+ \frac{t^{m-1}}{(m-1)!}N_m^{m-1} \right).

For a block of size three this is

etJ3(λ)=eλt(1tt2/201t001).e^{tJ_3(\lambda)} =e^{\lambda t} \begin{pmatrix} 1&t&t^2/2\\ 0&1&t\\ 0&0&1 \end{pmatrix}.

The polynomial factors arise from generalized eigenvectors rather than from additional eigenvalues. For an invertible block, applying the same principle to f(z)=1/zf(z)=1/z produces a finite expansion for its inverse.

Jordan blocks also describe singularities of the resolvent. For z≠λz\ne\lambda,

(zI−Jm(λ))−1=∑r=0m−1Nmr(z−λ)r+1.(zI-J_m(\lambda))^{-1} = \sum_{r=0}^{m-1} \frac{N_m^r}{(z-\lambda)^{r+1}}.

Thus the resolvent has a pole of order mm at the eigenvalue of a single block. For a full matrix, the pole order at λ\lambda is the size of its largest block, not its algebraic multiplicity.

These formulas connect matrix functions with polynomial interpolation. To reproduce f(A)f(A), a polynomial need only match ff and its derivatives through order sλ−1s_\lambda-1 at each eigenvalue, where sλs_\lambda is the largest relevant block size. The resulting Hermite interpolation conditions determine a unique polynomial of degree less than deg⁡mA\deg m_A. Consequently, the minimal polynomial specifies exactly how much scalar derivative information is required.

Linear differential equations and discrete dynamics

For a constant matrix AA, the homogeneous ordinary differential equation

x′(t)=Ax(t),x(0)=x0,x'(t)=Ax(t),\qquad x(0)=x_0,

has solution x(t)=etAx0x(t)=e^{tA}x_0. The substitution x=Pyx=Py transforms the equation into y′=Jyy'=Jy, so each Jordan block can be solved independently. A block of size mm associated with λ\lambda contributes terms involving

eλt, teλt,…,tm−1eλt.e^{\lambda t},\ te^{\lambda t},\ldots, t^{m-1}e^{\lambda t}.

The maximal polynomial degree attached to an eigenvalue is therefore one less than its largest block size.

For a concrete example,

A=(−110−1)A=\begin{pmatrix}-1&1\\0&-1\end{pmatrix}

produces

x2(t)=e−tx2(0),x1(t)=e−t(x1(0)+tx2(0)).x_2(t)=e^{-t}x_2(0),\qquad x_1(t)=e^{-t}\bigl(x_1(0)+t x_2(0)\bigr).

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 t≥0t\ge0 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,

xk+1=Axk,x_{k+1}=Ax_k,

the solution is xk=Akx0x_k=A^k x_0. All solutions tend to zero precisely when every eigenvalue has modulus less than one. The powers AkA^k 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,

R=(0−110)R=\begin{pmatrix}0&-1\\1&0\end{pmatrix}

has characteristic polynomial t2+1t^2+1. It is diagonalizable over the complex numbers, with eigenvalues ii and −i-i, but has no real eigenvector.

A real Jordan form accommodates conjugate pairs using two-dimensional diagonal blocks. If a+iba+ib, with b≠0b\ne0, is an eigenvalue, define

C=(a−bba).C=\begin{pmatrix}a&-b\\b&a\end{pmatrix}.

A pair of conjugate complex Jordan blocks of size mm corresponds to a real block of size 2m2m, with CC on its block diagonal and I2I_2 on its first block superdiagonal:

(CI20⋯0CI2⋱⋮⋱⋱⋱0⋯0C).\begin{pmatrix} C&I_2&0&\cdots\\ 0&C&I_2&\ddots\\ \vdots&\ddots&\ddots&\ddots\\ 0&\cdots&0&C \end{pmatrix}.

Conventions differing in the signs of the off-diagonal entries of CC 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 AA makes VV into a module over F[t]F[t] by defining

t⋅v=Av.t\cdot v=Av.

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 F[t]/(p(t)k)F[t]/(p(t)^k), where pp is irreducible.

When every relevant irreducible polynomial is linear, these elementary divisors become (t−λ)k(t-\lambda)^k, and each corresponds to a Jordan block of size kk. 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,

(λ10λ)and(λ00λ)\begin{pmatrix}\lambda&1\\0&\lambda\end{pmatrix} \quad\text{and}\quad \begin{pmatrix}\lambda&0\\0&\lambda\end{pmatrix}

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 AA. The set

C(A)={X:XA=AX}\mathcal C(A)=\{X:XA=AX\}

is called its centralizer or commutant. If AA is a single Jordan block, every commuting matrix is a polynomial in AA 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 pp and qq with the same eigenvalue, the space of rectangular matrices XX satisfying

Jp(λ)X=XJq(λ)J_p(\lambda)X=XJ_q(\lambda)

has dimension min⁡(p,q)\min(p,q). If the eigenvalues are distinct, the corresponding intertwining space is zero. Hence, if sλ,1,…,sλ,gλs_{\lambda,1},\ldots,s_{\lambda,g_\lambda} are the block sizes for λ\lambda,

dim⁡C(A)=∑λ∑i=1gλ∑j=1gλmin⁡(sλ,i,sλ,j).\dim\mathcal C(A) = \sum_\lambda \sum_{i=1}^{g_\lambda} \sum_{j=1}^{g_\lambda} \min(s_{\lambda,i},s_{\lambda,j}).

This formula measures how many linear transformations preserve the action of AA, and it is useful in studying similarity classes and their symmetries.

Another consequence is the additive decomposition

A=S+N,SN=NS,A=S+N,\qquad SN=NS,

where SS is diagonalizable and NN is nilpotent. In Jordan coordinates, SS replaces each block Jm(λ)J_m(\lambda) by λIm\lambda I_m, while NN retains its superdiagonal ones. When the characteristic polynomial splits, both parts can be expressed as polynomials in AA. 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

Aε=(01ε0).A_\varepsilon= \begin{pmatrix} 0&1\\ \varepsilon&0 \end{pmatrix}.

At ε=0\varepsilon=0, this is a size-two nilpotent Jordan block. For every nonzero ε\varepsilon, its complex eigenvalues are the distinct numbers ε\sqrt{\varepsilon} and −ε-\sqrt{\varepsilon}, 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 NmN_m, add a perturbation ε\varepsilon in its lower-left corner. The resulting characteristic polynomial is

tm−ε.t^m-\varepsilon.

Its eigenvalues have modulus ∣ε∣1/m|\varepsilon|^{1/m}. Thus a perturbation of size ∣ε∣|\varepsilon| can cause eigenvalue displacements on the much larger scale ∣ε∣1/m|\varepsilon|^{1/m}. 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

(1ε),(1−ε).\begin{pmatrix}1\\ \sqrt{\varepsilon}\end{pmatrix}, \qquad \begin{pmatrix}1\\-\sqrt{\varepsilon}\end{pmatrix}.

As ε\varepsilon 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

A=QTQ∗A=QTQ^*

is usually preferable: QQ is unitary and TT 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.