{
  "id": "jordan-normal-form",
  "title": "Jordan Normal Form",
  "lang": "en",
  "source_lang": "en",
  "summary": "Jordan normal form classifies square matrices up to similarity using eigenvalues and the sizes of their Jordan blocks.",
  "category": "mathematics",
  "subcategory": "algebra",
  "type": "concept",
  "aliases": [
    "Jordan canonical form",
    "Jordan form"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/jordan-normal-form/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/jordan-normal-form/",
      "json": "https://aiwiki.page/zh/jordan-normal-form.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "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.\n\nJordan 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."
    },
    {
      "heading": "Definition and classification theorem",
      "text": "For a scalar \\(\\lambda\\) and a positive integer \\(m\\), the Jordan block of size \\(m\\) is\n\n\\[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}.\\]\n\nA block of size one is simply the scalar matrix \\((\\lambda)\\). Writing \\(J_m(\\lambda)=\\lambda I_m+N_m\\) separates its scalar part from a nilpotent matrix: \\(N_m^m=0\\), while \\(N_m^{m-1}\\ne0\\). Here \\(I_m\\) denotes the identity matrix of order \\(m\\).\n\nLet \\(A\\) be an \\(n\\times n\\) matrix over a field \\(F\\). The Jordan normal form theorem states that if the characteristic polynomial of \\(A\\) splits completely over \\(F\\), then there is an invertible matrix \\(P\\) such that\n\n\\[P^{-1}AP=J =\\operatorname{diag}\\bigl( J_{m_1}(\\lambda_1),\\ldots,J_{m_r}(\\lambda_r) \\bigr).\\]\n\nThe sizes satisfy \\(m_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 \\(F\\) implies that the characteristic polynomial splits over \\(F\\).\n\nThe matrix \\(J\\) is unique up to a permutation of its blocks. The change-of-basis matrix \\(P\\), 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.\n\nThe columns of \\(P\\) form a basis in which the linear transformation represented by \\(A\\) 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.\n\nSome 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."
    },
    {
      "heading": "Generalized eigenvectors and the decomposition of the space",
      "text": "An ordinary eigenvector for an eigenvalue \\(\\lambda\\) is a nonzero vector \\(v\\) satisfying \\((A-\\lambda I)v=0\\). A generalized eigenvector satisfies\n\n\\[(A-\\lambda I)^k v=0\\]\n\nfor some positive integer \\(k\\). If \\(k\\) is the smallest such integer, the vector has order \\(k\\). 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.\n\nFor a fixed eigenvalue, the generalized eigenspace is\n\n\\[G_\\lambda=\\ker(A-\\lambda I)^n.\\]\n\nThe exponent \\(n\\) is always sufficient for an \\(n\\)-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 \\(\\lambda\\). It is invariant under \\(A\\), and the restriction of \\(A-\\lambda I\\) to it is nilpotent.\n\nIf the characteristic polynomial splits, the ambient vector space has the decomposition\n\n\\[V=\\bigoplus_{\\lambda}G_\\lambda,\\]\n\nwhere 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.\n\nInside \\(G_\\lambda\\), a Jordan chain of length \\(m\\) is a sequence \\(v_1,\\ldots,v_m\\) satisfying\n\n\\[(A-\\lambda I)v_1=0,\\qquad (A-\\lambda I)v_j=v_{j-1}\\quad(2\\le j\\le m),\\]\n\nwith \\(v_1\\ne0\\). These vectors are linearly independent. Indeed, applying a sufficiently high power of \\(A-\\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,\n\n\\[Av_1=\\lambda v_1,\\qquad Av_j=\\lambda v_j+v_{j-1},\\]\n\nwhich is precisely the action of an upper-diagonal Jordan block.\n\nThe 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\n\n\\[\\{0\\}\\subseteq\\ker N\\subseteq\\ker N^2\\subseteq\\cdots\\]\n\norganize the construction. Compatible choices of chain tops, followed by repeated application of \\(N\\), 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."
    },
    {
      "heading": "Block sizes, multiplicities, and complete invariants",
      "text": "The Jordan blocks encode more information than either the eigenvalues or their multiplicities alone. For an eigenvalue \\(\\lambda\\), its algebraic multiplicity is its multiplicity as a root of the characteristic polynomial. Its geometric multiplicity is the dimension of \\(\\ker(A-\\lambda I)\\). In Jordan form these become, respectively, the sum of the sizes of the blocks associated with \\(\\lambda\\) and the number of those blocks. Consequently,\n\n\\[1\\le \\dim\\ker(A-\\lambda I) \\le \\operatorname{mult}_\\lambda(\\chi_A).\\]\n\nEquality on the right holds exactly when every block for that eigenvalue has size one.\n\nThe minimal polynomial \\(m_A(t)\\) is the monic polynomial of least degree such that \\(m_A(A)=0\\). If \\(s_\\lambda\\) is the size of the largest block belonging to \\(\\lambda\\), then\n\n\\[\\chi_A(t)=\\prod_\\lambda(t-\\lambda)^{a_\\lambda}, \\qquad m_A(t)=\\prod_\\lambda(t-\\lambda)^{s_\\lambda},\\]\n\nwhere \\(a_\\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.\n\nAll the block sizes can be recovered from kernel dimensions. For a fixed \\(\\lambda\\), set\n\n\\[d_0=0,\\qquad d_k=\\dim\\ker(A-\\lambda I)^k.\\]\n\nIf the block sizes are \\(s_1,\\ldots,s_g\\), then\n\n\\[d_k=\\sum_{i=1}^{g}\\min(k,s_i).\\]\n\nEach block contributes one new kernel dimension at every step until its full size is reached. It follows that \\(d_k-d_{k-1}\\) counts blocks of size at least \\(k\\), and\n\n\\[b_k=2d_k-d_{k-1}-d_{k+1}\\]\n\ncounts blocks of size exactly \\(k\\). These identities also prove uniqueness of the Jordan block sizes without relying on any particular construction of a Jordan basis.\n\nFor example, suppose the kernel dimensions for one eigenvalue are\n\n\\[d_1=3,\\quad d_2=5,\\quad d_3=6,\\quad d_4=7,\\quad d_5=7.\\]\n\nThe successive increases are \\(3,2,1,1,0\\). 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 \\(4,2,1\\). Their total is seven, and the largest size is four.\n\nThe same information is available from ranks, since the rank–nullity theorem gives \\(d_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 \\(3,3,1,1\\) and blocks of sizes \\(3,2,2,1\\), all with eigenvalue zero, have characteristic polynomial \\(t^8\\) and minimal polynomial \\(t^3\\), but are not similar."
    },
    {
      "heading": "Exact construction and a worked example",
      "text": "An exact computation usually begins by factoring the characteristic polynomial over the intended field. For each eigenvalue \\(\\lambda\\), the kernels of successive powers of \\(A-\\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 \\(J\\), and constructing a particular invertible matrix \\(P\\) satisfying \\(AP=PJ\\).\n\nConsider\n\n\\[A= \\begin{pmatrix} 2&1&0&0\\\\ 0&2&0&0\\\\ 0&1&2&0\\\\ 0&0&0&3 \\end{pmatrix}.\\]\n\nBecause \\(A\\) is block triangular, its characteristic polynomial is\n\n\\[\\chi_A(t)=(t-2)^3(t-3).\\]\n\nOn the subspace spanned by the first three standard coordinate vectors, \\(N=A-2I\\) satisfies\n\n\\[Ne_2=e_1+e_3,\\qquad Ne_1=Ne_3=0.\\]\n\nThus \\(N\\) has rank one on this subspace and its square is zero. The kernel dimensions associated with the eigenvalue \\(2\\) are \\(d_1=2\\) and \\(d_2=3\\); the fourth coordinate contributes no kernel because \\(A-2I\\) acts there as multiplication by one. Hence the blocks for \\(2\\) have sizes two and one.\n\nA length-two chain is\n\n\\[v_1=e_1+e_3,\\qquad v_2=e_2,\\]\n\nbecause \\((A-2I)v_2=v_1\\) and \\((A-2I)v_1=0\\). An independent eigenvector for the remaining size-one block is \\(v_3=e_3\\). The fourth standard vector is an eigenvector for \\(3\\). Therefore\n\n\\[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}\\]\n\ngives\n\n\\[P^{-1}AP= \\begin{pmatrix} 2&1&0&0\\\\ 0&2&0&0\\\\ 0&0&2&0\\\\ 0&0&0&3 \\end{pmatrix}.\\]\n\nThe minimal polynomial is \\((t-2)^2(t-3)\\), and \\(A\\) is not diagonalizable despite having a two-dimensional ordinary eigenspace for its repeated eigenvalue.\n\nFor 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.\n\nDirect verification is valuable because it avoids explicitly computing an inverse: checking that \\(P\\) is nonsingular and that \\(AP=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."
    },
    {
      "heading": "Matrix powers, analytic functions, and resolvents",
      "text": "Jordan form reduces evaluating many matrix functions to evaluating scalar functions and their derivatives. Let \\(J=\\lambda I+N\\) be a block of size \\(m\\). Since \\(N^m=0\\), the binomial theorem gives, for a nonnegative integer \\(k\\),\n\n\\[J^k= \\sum_{r=0}^{\\min(k,m-1)} \\binom{k}{r}\\lambda^{k-r}N^r.\\]\n\nWhen \\(\\lambda\\ne0\\), entries on successive superdiagonals involve polynomial factors in \\(k\\), multiplied by powers of \\(\\lambda\\). When \\(\\lambda=0\\), the block is nilpotent and its \\(k\\)th power vanishes for every \\(k\\ge m\\). These are fundamentally different long-term behaviors.\n\nMore generally, if \\(f\\) is an analytic function on neighborhoods of the eigenvalues, then\n\n\\[f(J_m(\\lambda)) = \\sum_{r=0}^{m-1} \\frac{f^{(r)}(\\lambda)}{r!}N_m^r.\\]\n\nThe scalar Taylor expansion terminates after substitution because all higher powers of \\(N_m\\) are zero. For a polynomial, the identity holds without analytic assumptions. For a general matrix,\n\n\\[f(A)=P f(J)P^{-1}.\\]\n\nAlthough \\(P\\) is not unique, this definition is independent of the chosen Jordan basis.\n\nFor example, the matrix exponential satisfies\n\n\\[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).\\]\n\nFor a block of size three this is\n\n\\[e^{tJ_3(\\lambda)} =e^{\\lambda t} \\begin{pmatrix} 1&t&t^2/2\\\\ 0&1&t\\\\ 0&0&1 \\end{pmatrix}.\\]\n\nThe polynomial factors arise from generalized eigenvectors rather than from additional eigenvalues. For an invertible block, applying the same principle to \\(f(z)=1/z\\) produces a finite expansion for its inverse.\n\nJordan blocks also describe singularities of the resolvent. For \\(z\\ne\\lambda\\),\n\n\\[(zI-J_m(\\lambda))^{-1} = \\sum_{r=0}^{m-1} \\frac{N_m^r}{(z-\\lambda)^{r+1}}.\\]\n\nThus the resolvent has a pole of order \\(m\\) at the eigenvalue of a single block. For a full matrix, the pole order at \\(\\lambda\\) is the size of its largest block, not its algebraic multiplicity.\n\nThese formulas connect matrix functions with polynomial interpolation. To reproduce \\(f(A)\\), a polynomial need only match \\(f\\) and its derivatives through order \\(s_\\lambda-1\\) at each eigenvalue, where \\(s_\\lambda\\) is the largest relevant block size. The resulting Hermite interpolation conditions determine a unique polynomial of degree less than \\(\\deg m_A\\). Consequently, the minimal polynomial specifies exactly how much scalar derivative information is required."
    },
    {
      "heading": "Linear differential equations and discrete dynamics",
      "text": "For a constant matrix \\(A\\), the homogeneous ordinary differential equation\n\n\\[x'(t)=Ax(t),\\qquad x(0)=x_0,\\]\n\nhas solution \\(x(t)=e^{tA}x_0\\). The substitution \\(x=Py\\) transforms the equation into \\(y'=Jy\\), so each Jordan block can be solved independently. A block of size \\(m\\) associated with \\(\\lambda\\) contributes terms involving\n\n\\[e^{\\lambda t},\\ te^{\\lambda t},\\ldots, t^{m-1}e^{\\lambda t}.\\]\n\nThe maximal polynomial degree attached to an eigenvalue is therefore one less than its largest block size.\n\nFor a concrete example,\n\n\\[A=\\begin{pmatrix}-1&1\\\\0&-1\\end{pmatrix}\\]\n\nproduces\n\n\\[x_2(t)=e^{-t}x_2(0),\\qquad x_1(t)=e^{-t}\\bigl(x_1(0)+t x_2(0)\\bigr).\\]\n\nBoth 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.\n\nFor 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\\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.\n\nFor a discrete-time dynamical system,\n\n\\[x_{k+1}=Ax_k,\\]\n\nthe solution is \\(x_k=A^k x_0\\). All solutions tend to zero precisely when every eigenvalue has modulus less than one. The powers \\(A^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.\n\nJordan 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."
    },
    {
      "heading": "Ground fields, real forms, and rational canonical form",
      "text": "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,\n\n\\[R=\\begin{pmatrix}0&-1\\\\1&0\\end{pmatrix}\\]\n\nhas characteristic polynomial \\(t^2+1\\). It is diagonalizable over the complex numbers, with eigenvalues \\(i\\) and \\(-i\\), but has no real eigenvector.\n\nA real Jordan form accommodates conjugate pairs using two-dimensional diagonal blocks. If \\(a+ib\\), with \\(b\\ne0\\), is an eigenvalue, define\n\n\\[C=\\begin{pmatrix}a&-b\\\\b&a\\end{pmatrix}.\\]\n\nA pair of conjugate complex Jordan blocks of size \\(m\\) corresponds to a real block of size \\(2m\\), with \\(C\\) on its block diagonal and \\(I_2\\) on its first block superdiagonal:\n\n\\[\\begin{pmatrix} C&I_2&0&\\cdots\\\\ 0&C&I_2&\\ddots\\\\ \\vdots&\\ddots&\\ddots&\\ddots\\\\ 0&\\cdots&0&C \\end{pmatrix}.\\]\n\nConventions differing in the signs of the off-diagonal entries of \\(C\\) are equivalent by a real change of basis.\n\nOver 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.\n\nThe conceptual foundation is the structure theorem for finitely generated modules over a principal ideal domain. A transformation \\(A\\) makes \\(V\\) into a module over \\(F[t]\\) by defining\n\n\\[t\\cdot v=Av.\\]\n\nBecause 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)\\), where \\(p\\) is irreducible.\n\nWhen every relevant irreducible polynomial is linear, these elementary divisors become \\((t-\\lambda)^k\\), and each corresponds to a Jordan block of size \\(k\\). 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."
    },
    {
      "heading": "Similarity, commuting transformations, and spectral structure",
      "text": "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,\n\n\\[\\begin{pmatrix}\\lambda&1\\\\0&\\lambda\\end{pmatrix} \\quad\\text{and}\\quad \\begin{pmatrix}\\lambda&0\\\\0&\\lambda\\end{pmatrix}\\]\n\nhave identical characteristic polynomials but different eigenspace dimensions, so they cannot be similar. The block partition at each eigenvalue is the additional information required.\n\nThe same structure controls matrices that commute with \\(A\\). The set\n\n\\[\\mathcal C(A)=\\{X:XA=AX\\}\\]\n\nis called its centralizer or commutant. If \\(A\\) is a single Jordan block, every commuting matrix is a polynomial in \\(A\\) 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.\n\nFor blocks of sizes \\(p\\) and \\(q\\) with the same eigenvalue, the space of rectangular matrices \\(X\\) satisfying\n\n\\[J_p(\\lambda)X=XJ_q(\\lambda)\\]\n\nhas dimension \\(\\min(p,q)\\). If the eigenvalues are distinct, the corresponding intertwining space is zero. Hence, if \\(s_{\\lambda,1},\\ldots,s_{\\lambda,g_\\lambda}\\) are the block sizes for \\(\\lambda\\),\n\n\\[\\dim\\mathcal C(A) = \\sum_\\lambda \\sum_{i=1}^{g_\\lambda} \\sum_{j=1}^{g_\\lambda} \\min(s_{\\lambda,i},s_{\\lambda,j}).\\]\n\nThis formula measures how many linear transformations preserve the action of \\(A\\), and it is useful in studying similarity classes and their symmetries.\n\nAnother consequence is the additive decomposition\n\n\\[A=S+N,\\qquad SN=NS,\\]\n\nwhere \\(S\\) is diagonalizable and \\(N\\) is nilpotent. In Jordan coordinates, \\(S\\) replaces each block \\(J_m(\\lambda)\\) by \\(\\lambda I_m\\), while \\(N\\) retains its superdiagonal ones. When the characteristic polynomial splits, both parts can be expressed as polynomials in \\(A\\). This is the matrix form of the Jordan–Chevalley decomposition, separating the eigenvalue data from the nontrivial chain structure.\n\nJordan 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."
    },
    {
      "heading": "Perturbations and numerical computation",
      "text": "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\n\n\\[A_\\varepsilon= \\begin{pmatrix} 0&1\\\\ \\varepsilon&0 \\end{pmatrix}.\\]\n\nAt \\(\\varepsilon=0\\), this is a size-two nilpotent Jordan block. For every nonzero \\(\\varepsilon\\), its complex eigenvalues are the distinct numbers \\(\\sqrt{\\varepsilon}\\) and \\(-\\sqrt{\\varepsilon}\\), so it is diagonalizable. Its Jordan structure therefore changes discontinuously at zero.\n\nThe sensitivity becomes more pronounced for larger blocks. Starting with the nilpotent block \\(N_m\\), add a perturbation \\(\\varepsilon\\) in its lower-left corner. The resulting characteristic polynomial is\n\n\\[t^m-\\varepsilon.\\]\n\nIts eigenvalues have modulus \\(|\\varepsilon|^{1/m}\\). Thus a perturbation of size \\(|\\varepsilon|\\) can cause eigenvalue displacements on the much larger scale \\(|\\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.\n\nThe 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\n\n\\[\\begin{pmatrix}1\\\\ \\sqrt{\\varepsilon}\\end{pmatrix}, \\qquad \\begin{pmatrix}1\\\\-\\sqrt{\\varepsilon}\\end{pmatrix}.\\]\n\nAs \\(\\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.\n\nThese 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.\n\nFor routine numerical linear algebra, the Schur decomposition\n\n\\[A=QTQ^*\\]\n\nis usually preferable: \\(Q\\) is unitary and \\(T\\) 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.\n\nThe 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."
    }
  ],
  "links": [
    "matrix",
    "complex-number",
    "linear-algebra",
    "matrix-diagonalization",
    "identity-matrix",
    "field-mathematics",
    "characteristic-polynomial",
    "matrix-similarity",
    "basis-linear-algebra",
    "linear-map",
    "eigenvalues-and-eigenvectors",
    "eigenspace",
    "vector-space",
    "direct-sum",
    "minimal-polynomial",
    "matrix-rank",
    "rank-nullity-theorem",
    "matrix-trace",
    "determinant",
    "system-of-linear-equations",
    "gaussian-elimination",
    "binomial-theorem",
    "analytic-function",
    "taylor-series",
    "matrix-exponential",
    "resolvent-operator",
    "ordinary-differential-equation",
    "dynamical-system",
    "recurrence-relation",
    "transition-matrix",
    "markov-chain",
    "fundamental-theorem-of-algebra",
    "real-number",
    "module-mathematics",
    "abstract-algebra",
    "spectral-theorem",
    "unitary-matrix",
    "condition-number",
    "floating-point-arithmetic",
    "numerical-linear-algebra",
    "schur-decomposition",
    "numerical-stability",
    "qr-algorithm"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}