{
  "id": "moore-penrose-inverse",
  "title": "Moore–Penrose Inverse",
  "lang": "en",
  "source_lang": "en",
  "summary": "The Moore–Penrose inverse extends matrix inversion to rectangular and singular matrices, selecting unique minimum-norm least-squares solutions.",
  "category": "mathematics",
  "subcategory": "algebra",
  "type": "concept",
  "aliases": [
    "Moore–Penrose pseudoinverse",
    "pseudoinverse",
    "MP inverse"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/moore-penrose-inverse/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/moore-penrose-inverse/",
      "json": "https://aiwiki.page/zh/moore-penrose-inverse.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The Moore–Penrose inverse of a real or complex matrix \\(A\\) is the unique matrix \\(A^+\\) satisfying four algebraic conditions that extend the defining properties of an ordinary inverse. It exists for every finite-dimensional matrix, including rectangular matrices and singular square matrices. When \\(A\\) is invertible, \\(A^+\\) is exactly its inverse. Otherwise, it reverses the action of \\(A\\) on the directions where reversal is possible and assigns zero to the remaining orthogonal directions.\n\nIts central role is in solving linear systems that have either no exact solution or more than one. For every right-hand side \\(b\\), the vector \\(A^+b\\) minimizes the residual \\(\\|Ax-b\\|_2\\); among all vectors attaining that minimum, it has the smallest Euclidean norm. This makes the construction fundamental to least squares, numerical computation, statistical modeling, and inverse problems. Its geometry is described by orthogonal projections, while its most direct computational characterization uses the singular value decomposition.\n\nThe name commemorates E. H. Moore, who presented an early form of the generalized inverse in 1920, and Roger Penrose, who characterized it through four matrix equations in 1955. Related work by Arne Bjerhammar appeared in 1951. Although “pseudoinverse” sometimes refers to a broader class of generalized inverses, it commonly means the Moore–Penrose inverse."
    },
    {
      "heading": "Defining equations and geometric meaning",
      "text": "Let \\(A\\in\\mathbb C^{m\\times n}\\), and let \\(A^*\\) denote its conjugate transpose. The Moore–Penrose inverse is the matrix \\(A^+\\in\\mathbb C^{n\\times m}\\) satisfying the four Penrose equations:\n\n\\[AA^+A=A,\\qquad A^+AA^+=A^+,\\]\n\n\\[(AA^+)^*=AA^+,\\qquad (A^+A)^*=A^+A.\\]\n\nFor real matrices, the conjugate transpose reduces to the ordinary transpose. The first equation says that \\(A^+\\) reverses \\(A\\) sufficiently to reproduce every output of \\(A\\). The second excludes unnecessary action outside the part needed for this reversal. The last two impose orthogonality rather than merely algebraic projection.\n\nTo interpret these requirements, regard \\(A\\) as a linear map from \\(\\mathbb C^n\\) to \\(\\mathbb C^m\\), both equipped with their standard inner products. Its domain and codomain have orthogonal decompositions\n\n\\[\\mathbb C^n=\\mathcal R(A^*)\\oplus\\mathcal N(A), \\qquad \\mathbb C^m=\\mathcal R(A)\\oplus\\mathcal N(A^*),\\]\n\nwhere \\(\\mathcal R\\) denotes the range and \\(\\mathcal N\\) the null space. The restriction of \\(A\\) to \\(\\mathcal R(A^*)\\) is a bijection onto \\(\\mathcal R(A)\\). The pseudoinverse is its inverse on that range and is zero on \\(\\mathcal N(A^*)\\). This construction establishes both existence and uniqueness without requiring that \\(A\\) be square.\n\nThe two products\n\n\\[P=AA^+,\\qquad Q=A^+A\\]\n\nare orthogonal projections. Specifically, \\(P\\) projects onto the column space \\(\\mathcal R(A)\\), while \\(Q\\) projects onto \\(\\mathcal R(A^*)\\), the conjugate-transposed row space. They are self-adjoint and idempotent:\n\n\\[P^*=P,\\quad P^2=P,\\qquad Q^*=Q,\\quad Q^2=Q.\\]\n\nConsequently, \\(I-P\\) projects onto \\(\\mathcal N(A^*)\\), and \\(I-Q\\) projects onto \\(\\mathcal N(A)\\). These complementary projections separate attainable data from residual error, and identifiable input directions from directions annihilated by \\(A\\).\n\nThe orthogonality requirements distinguish the Moore–Penrose inverse from a general generalized inverse. A matrix \\(G\\) satisfying only \\(AGA=A\\) can reproduce outputs of \\(A\\), but need not produce orthogonal projections or minimum-norm solutions. Many such matrices may exist. The Moore–Penrose inverse is canonical relative to the specified inner products, not independent of geometry: changing the inner products generally changes which solution has minimum norm."
    },
    {
      "heading": "Singular values and explicit constructions",
      "text": "The singular value decomposition gives a direct formula valid for every matrix. Write\n\n\\[A=U\\Sigma V^*,\\]\n\nwhere \\(U\\in\\mathbb C^{m\\times m}\\) and \\(V\\in\\mathbb C^{n\\times n}\\) are unitary matrices, and \\(\\Sigma\\) is rectangular diagonal. If \\(A\\) has rank \\(r\\), its positive singular values can be ordered as\n\n\\[\\sigma_1\\geq \\sigma_2\\geq\\cdots\\geq\\sigma_r>0.\\]\n\nForm \\(\\Sigma^+\\in\\mathbb R^{n\\times m}\\) by replacing each positive singular value with its reciprocal and leaving every zero entry zero. Then\n\n\\[A^+=V\\Sigma^+U^*.\\]\n\nThus pseudoinversion reverses the input and output singular directions while reciprocating only the nonzero scale factors.\n\nIn terms of left and right singular vectors,\n\n\\[A=\\sum_{i=1}^{r}\\sigma_i u_i v_i^*, \\qquad A^+=\\sum_{i=1}^{r}\\frac{1}{\\sigma_i}v_i u_i^*.\\]\n\nThe associated projections are\n\n\\[AA^+=\\sum_{i=1}^{r}u_i u_i^*, \\qquad A^+A=\\sum_{i=1}^{r}v_i v_i^*.\\]\n\nThese formulas show that \\(\\operatorname{rank}(A^+)=r\\), that the nonzero singular values of \\(A^+\\) are the reciprocals of those of \\(A\\), and that the pseudoinverse of the zero matrix is the zero matrix of transposed dimensions.\n\nImportant special cases admit shorter expressions. If \\(A\\) has full column rank, then \\(m\\geq n\\) and\n\n\\[A^+=(A^*A)^{-1}A^*.\\]\n\nIf it has full row rank, then \\(m\\leq n\\) and\n\n\\[A^+=A^*(AA^*)^{-1}.\\]\n\nThe matrices \\(A^*A\\) and \\(AA^*\\) are Gram matrices; the relevant one is invertible precisely under the stated rank condition. These expressions are algebraically useful, although explicitly forming Gram matrices can be numerically undesirable.\n\nA rank factorization provides another exact construction. If\n\n\\[A=BC,\\]\n\nwhere \\(B\\in\\mathbb C^{m\\times r}\\) has full column rank and \\(C\\in\\mathbb C^{r\\times n}\\) has full row rank, then\n\n\\[A^+ =C^*(CC^*)^{-1}(B^*B)^{-1}B^*.\\]\n\nOnly matrices of order \\(r\\) need to be inverted in this formula. For a nonzero rank-one matrix \\(A=uv^*\\), it reduces to\n\n\\[A^+=\\frac{vu^*}{(u^*u)(v^*v)}.\\]\n\nFor a Hermitian matrix, the spectral theorem supplies an analogous construction: diagonalize the matrix, reciprocate each nonzero eigenvalue, and retain zero eigenvalues as zero. Negative eigenvalues remain negative after reciprocation."
    },
    {
      "heading": "Least-squares solutions and minimum norm",
      "text": "For \\(A\\in\\mathbb C^{m\\times n}\\) and \\(b\\in\\mathbb C^m\\), consider\n\n\\[\\min_{x\\in\\mathbb C^n}\\|Ax-b\\|_2.\\]\n\nThe pseudoinverse selects\n\n\\[x_0=A^+b.\\]\n\nIts fitted output is\n\n\\[Ax_0=AA^+b,\\]\n\nthe orthogonal projection of \\(b\\) onto \\(\\mathcal R(A)\\). To see why this minimizes the residual, decompose\n\n\\[b=AA^+b+(I-AA^+)b.\\]\n\nFor every \\(x\\), the vectors \\(Ax-AA^+b\\) and \\((I-AA^+)b\\) are orthogonal. The Pythagorean identity therefore gives\n\n\\[\\|Ax-b\\|_2^2 = \\|Ax-AA^+b\\|_2^2 + \\|(I-AA^+)b\\|_2^2.\\]\n\nThe second term is unavoidable; the first vanishes exactly when \\(Ax=AA^+b\\).\n\nAll least-squares minimizers have the form\n\n\\[x=A^+b+(I-A^+A)z,\\qquad z\\in\\mathbb C^n.\\]\n\nThe second term ranges over \\(\\mathcal N(A)\\), so it changes the input without changing the fitted output. Since \\(A^+b\\in\\mathcal R(A^*)\\), it is orthogonal to every such addition. Hence\n\n\\[\\|x\\|_2^2 = \\|A^+b\\|_2^2+\\|(I-A^+A)z\\|_2^2.\\]\n\nThis proves that \\(A^+b\\) is the unique minimum-norm least-squares solution. The two minimizations are hierarchical: residual norm is minimized first, and solution norm only among residual minimizers.\n\nWhen \\(b\\in\\mathcal R(A)\\), the residual is zero, and the formula describes all exact solutions of \\(Ax=b\\). For example,\n\n\\[A=\\begin{pmatrix}1&1\\end{pmatrix}, \\qquad A^+=\\frac12\\begin{pmatrix}1\\\\1\\end{pmatrix}.\\]\n\nThe equation \\(x_1+x_2=b\\) has infinitely many solutions, but the pseudoinverse chooses \\(x_1=x_2=b/2\\). Conversely, for\n\n\\[A=\\begin{pmatrix}1\\\\1\\end{pmatrix},\\]\n\nthe pseudoinverse is \\(\\tfrac12(1\\ \\ 1)\\). Applied to \\(b=(b_1,b_2)^{\\mathsf T}\\), it returns the arithmetic mean, the scalar minimizing the sum of squared discrepancies from the two observations.\n\nEvery least-squares minimizer satisfies the normal equations\n\n\\[A^*Ax=A^*b.\\]\n\nIf \\(A\\) is rank deficient, these equations do not determine a unique vector. The pseudoinverse adds the condition \\(x\\in\\mathcal R(A^*)\\), excluding null-space components. For several right-hand sides collected in a matrix \\(B\\), the same argument applies column by column: \\(A^+B\\) minimizes \\(\\|AX-B\\|_F\\) and has minimum Frobenius norm among all minimizers."
    },
    {
      "heading": "Algebraic identities and their limitations",
      "text": "Several familiar inverse identities survive pseudoinversion. For every finite matrix,\n\n\\[(A^+)^+=A,\\qquad (A^*)^+=(A^+)^*.\\]\n\nFor a nonzero scalar \\(\\alpha\\),\n\n\\[(\\alpha A)^+=\\alpha^{-1}A^+.\\]\n\nIf \\(U\\) and \\(V\\) are unitary matrices of suitable dimensions, then\n\n\\[(UAV)^+=V^*A^+U^*.\\]\n\nThus pseudoinversion behaves naturally under orthonormal changes of coordinates. This property follows either from the Penrose equations or from the singular value decomposition.\n\nThese identities do not extend to arbitrary invertible coordinate transformations. Nonunitary transformations distort lengths and angles, changing the minimum-norm criterion. In general,\n\n\\[(SAT)^+\\ne T^{-1}A^+S^{-1}.\\]\n\nFor example, let \\(A=(1\\ \\ 1)\\) and \\(T=\\operatorname{diag}(2,1)\\). Then\n\n\\[(AT)^+=\\frac15\\begin{pmatrix}2\\\\1\\end{pmatrix},\\]\n\nwhereas\n\n\\[T^{-1}A^+=\\begin{pmatrix}1/4\\\\1/2\\end{pmatrix}.\\]\n\nBoth are right inverses of \\(AT\\), but only the first selects solutions of minimum standard Euclidean norm.\n\nThe ordinary reverse-order law also fails in general:\n\n\\[(AB)^+\\ne B^+A^+.\\]\n\nA small example is\n\n\\[A=\\begin{pmatrix}1&1\\end{pmatrix}, \\qquad B=\\begin{pmatrix}1\\\\0\\end{pmatrix}.\\]\n\nHere \\(AB=(1)\\), so \\((AB)^+=(1)\\), while\n\n\\[B^+A^+ = \\begin{pmatrix}1&0\\end{pmatrix} \\frac12\\begin{pmatrix}1\\\\1\\end{pmatrix} =\\frac12.\\]\n\nReverse order does hold under additional hypotheses. In particular, it holds when \\(A\\) has full column rank and \\(B\\) has full row rank, as follows from the rank-factorization formula. It also holds when the relevant factor is unitary.\n\nUseful Gram-matrix identities valid without full-rank assumptions are\n\n\\[A^+=(A^*A)^+A^*=A^*(AA^*)^+,\\]\n\nand\n\n\\[(A^*A)^+=A^+(A^+)^*.\\]\n\nFor an orthogonal projection \\(P\\), one has \\(P^+=P\\); for a unitary matrix, the pseudoinverse is its adjoint. Block-diagonal matrices are pseudoinverted block by block. By contrast, there is no general additive rule: \\((A+B)^+\\) is not usually \\(A^++B^+\\). These distinctions are important because manipulations valid for invertible matrices can silently fail when intermediate factors lose rank."
    },
    {
      "heading": "Numerical computation and numerical rank",
      "text": "The singular value decomposition is the standard general-purpose basis for computing a pseudoinverse. Its central advantage is that it exposes the directions in which a matrix is nearly singular. In exact arithmetic, every positive singular value is reciprocated. In floating-point arithmetic, however, computed singular values near the scale of rounding error cannot reliably distinguish exact dependence from very weak independence. Software therefore commonly applies a threshold \\(\\tau\\), retaining \\(\\sigma_i^{-1}\\) only when \\(\\sigma_i>\\tau\\).\n\nA typical scale-dependent threshold has the form\n\n\\[\\tau=c\\,\\max(m,n)\\,\\varepsilon\\,\\sigma_1,\\]\n\nwhere \\(\\varepsilon\\) is machine precision and \\(c\\) is an implementation-dependent factor. The resulting matrix is the exact pseudoinverse of a truncated approximation to \\(A\\), not necessarily the exact pseudoinverse of the original mathematical matrix. This distinction becomes significant for poorly conditioned data. A threshold based only on arithmetic precision addresses rounding error; a threshold chosen to suppress measurement noise expresses an additional modeling decision.\n\nWhen full column rank is known, a thin QR decomposition\n\n\\[A=QR,\\qquad Q^*Q=I,\\]\n\ngives\n\n\\[A^+=R^{-1}Q^*.\\]\n\nFor a single right-hand side, the computation can instead form \\(Q^*b\\) and solve the triangular system \\(Rx=Q^*b\\), without constructing \\(A^+\\). Rank-revealing QR and complete orthogonal decompositions can handle rank deficiency, although rank decisions remain necessary. The SVD is generally more informative about singular values, while QR-based approaches can be less expensive.\n\nThe algebraic formula \\((A^*A)^{-1}A^*\\) is not usually the preferred numerical construction. For a full-column-rank matrix, the spectral condition number satisfies\n\n\\[\\kappa_2(A^*A)=\\kappa_2(A)^2.\\]\n\nForming the normal equations therefore squares the condition number and may lose information about small singular directions before the inverse is computed. Orthogonal factorizations avoid this particular amplification, although no algorithm can eliminate sensitivity inherent in the underlying problem.\n\nFor a dense \\(m\\times n\\) matrix with \\(m\\geq n\\), standard SVD computation has arithmetic cost on the order of \\(mn^2\\). Explicitly storing the pseudoinverse requires \\(mn\\) entries and may destroy sparsity. Large-scale methods therefore often compute \\(A^+b\\) indirectly. Krylov-subspace least-squares methods such as LSQR, in exact arithmetic and with a zero initial vector, select the minimum-norm least-squares solution upon convergence. Early stopping, preconditioning, and finite precision can alter that behavior and must be distinguished from exact pseudoinversion."
    },
    {
      "heading": "Perturbation, discontinuity, and regularization",
      "text": "Pseudoinversion is sensitive to small nonzero singular values because\n\n\\[\\|A^+\\|_2=\\frac1{\\sigma_r}\\]\n\nfor a nonzero rank-\\(r\\) matrix. If the matrix is fixed but the right-hand side changes by \\(\\delta b\\), then\n\n\\[\\delta x=A^+\\delta b,\\qquad \\|\\delta x\\|_2\\leq\\|A^+\\|_2\\|\\delta b\\|_2.\\]\n\nMore specifically, noise along \\(u_i\\) produces an input error along \\(v_i\\), amplified by \\(1/\\sigma_i\\). The difficulty is therefore directional, not simply a matter of the overall magnitude of the data.\n\nUnlike ordinary inversion on the set of nonsingular matrices, pseudoinversion is not continuous across rank changes. Consider\n\n\\[A(t)= \\begin{pmatrix} 1&0\\\\ 0&t \\end{pmatrix}.\\]\n\nFor \\(t\\ne0\\),\n\n\\[A(t)^+= \\begin{pmatrix} 1&0\\\\ 0&t^{-1} \\end{pmatrix},\\]\n\nbut\n\n\\[A(0)^+= \\begin{pmatrix} 1&0\\\\ 0&0 \\end{pmatrix}.\\]\n\nAlthough \\(A(t)\\to A(0)\\), the pseudoinverses diverge. Restricted to matrices of a fixed rank, the pseudoinverse varies continuously and smoothly. A deficient-rank matrix nevertheless has higher-rank matrices arbitrarily close to it, so unrestricted perturbations can cause singular behavior.\n\nIn an inverse problem, a small singular value may encode a genuine but weakly observed direction. Exact pseudoinversion can amplify noise in that direction to an unacceptable degree. Regularization modifies the inverse action rather than attempting to recover every nonzero singular component. Truncated SVD uses\n\n\\[x_\\tau=\\sum_{\\sigma_i>\\tau} \\frac{u_i^*b}{\\sigma_i}v_i.\\]\n\nThis eliminates poorly resolved directions entirely. The threshold controls a trade-off between suppressing noise and discarding signal; it is not determined by the Penrose equations.\n\nA smooth alternative is quadratic, or Tikhonov, regularization:\n\n\\[x_\\lambda = \\operatorname*{arg\\,min}_x \\left(\\|Ax-b\\|_2^2+\\lambda\\|x\\|_2^2\\right) = (A^*A+\\lambda I)^{-1}A^*b, \\qquad \\lambda>0.\\]\n\nIts singular-direction factors are\n\n\\[\\frac{\\sigma_i}{\\sigma_i^2+\\lambda},\\]\n\nrather than \\(1/\\sigma_i\\). For fixed finite-dimensional \\(A\\),\n\n\\[\\lim_{\\lambda\\downarrow0}x_\\lambda=A^+b,\\]\n\nand, at the matrix level,\n\n\\[A^+=\\lim_{\\lambda\\downarrow0}(A^*A+\\lambda I)^{-1}A^*.\\]\n\nThe regularized solution at positive \\(\\lambda\\) is generally not the Moore–Penrose solution of the original problem. Its purpose is precisely to replace unstable exact inversion with a controlled approximation."
    },
    {
      "heading": "Statistical and scientific applications",
      "text": "In linear regression, a real design matrix \\(X\\in\\mathbb R^{n\\times p}\\) relates observations to coefficients through\n\n\\[y=X\\beta+\\varepsilon.\\]\n\nThe ordinary least-squares coefficient vector can be chosen as\n\n\\[\\widehat\\beta=X^+y.\\]\n\nIf the columns of \\(X\\) are linearly independent, this equals \\((X^{\\mathsf T}X)^{-1}X^{\\mathsf T}y\\). If they are dependent, coefficients are not uniquely determined by least squares, but the fitted vector remains unique:\n\n\\[\\widehat y=Hy,\\qquad H=XX^+.\\]\n\nThe hat matrix \\(H\\) is the orthogonal projection onto the model’s column space, and\n\n\\[\\operatorname{tr}(H)=\\operatorname{rank}(X).\\]\n\nThe minimum-norm convention does not resolve statistical nonidentifiability. Under \\(E[\\varepsilon]=0\\),\n\n\\[E[\\widehat\\beta]=X^+X\\beta,\\]\n\nwhich equals \\(\\beta\\) only when \\(\\beta\\) lies in the row space of \\(X\\). A linear contrast \\(c^{\\mathsf T}\\beta\\) is estimable from the mean response when \\(c\\in\\mathcal R(X^{\\mathsf T})\\). If\n\n\\[\\operatorname{Cov}(\\varepsilon)=\\sigma^2I,\\]\n\nthen\n\n\\[\\operatorname{Cov}(\\widehat\\beta) = \\sigma^2X^+(X^+)^{\\mathsf T} = \\sigma^2(X^{\\mathsf T}X)^+.\\]\n\nZero variance in a null-space direction reflects the convention that the estimator sets that component to zero, not information supplied by observations.\n\nWeighted problems demonstrate the importance of specifying the geometry. For a positive-definite matrix \\(W\\), minimizing\n\n\\[(Ax-b)^*W(Ax-b)\\]\n\ncan be reduced to ordinary least squares by applying \\(W^{1/2}\\) to \\(A\\) and \\(b\\). If solution size is measured by \\(x^*Mx\\), with \\(M\\) positive definite, the minimum-\\(M\\)-norm weighted least-squares solution is\n\n\\[x=M^{-1/2} \\left(W^{1/2}AM^{-1/2}\\right)^+ W^{1/2}b.\\]\n\nSuch formulations arise when observation errors have unequal variances or when input coordinates have different physical scales.\n\nIn differential kinematics, a robot’s joint velocities \\(\\dot q\\) and task-space velocity \\(\\dot x\\) are related by a Jacobian matrix:\n\n\\[J\\dot q=\\dot x.\\]\n\nThe command \\(\\dot q=J^+\\dot x\\) has minimum Euclidean joint-velocity norm among exact solutions, or among least-squares solutions when the requested velocity is unattainable. The term \\((I-J^+J)z\\) describes motions that do not change task-space velocity to first order. Near kinematic singularities, small singular values can produce large commands, motivating damped least-squares variants.\n\nIn graph theory, a connected undirected graph has a singular Laplacian matrix \\(L\\), whose null space is spanned by the constant vector. Its pseudoinverse selects the mean-zero solution of \\(Lx=b\\) whenever \\(b\\) has zero sum. For a resistor network with positive edge conductances, the effective resistance between vertices \\(i\\) and \\(j\\) is\n\n\\[R_{ij}=(e_i-e_j)^{\\mathsf T}L^+(e_i-e_j).\\]\n\nThe missing inverse direction corresponds to the arbitrary additive constant in electric potential."
    },
    {
      "heading": "Hilbert-space operators and infinite-dimensional limits",
      "text": "The geometric construction extends from matrices to bounded linear operators between Hilbert spaces, but finite dimensionality conceals an important condition: every finite-dimensional range is closed. Let \\(T:H\\to K\\) be bounded. If \\(\\mathcal R(T)\\) is closed, then\n\n\\[H=\\mathcal N(T)^\\perp\\oplus\\mathcal N(T), \\qquad K=\\mathcal R(T)\\oplus\\mathcal R(T)^\\perp.\\]\n\nThe restriction of \\(T\\) to \\(\\mathcal N(T)^\\perp\\) is a bijection onto \\(\\mathcal R(T)\\). Its inverse, extended by zero on \\(\\mathcal R(T)^\\perp\\), is a bounded operator \\(T^+\\). It has the same minimum-norm and projection interpretations as a matrix pseudoinverse.\n\nIf the range is not closed, a Moore–Penrose inverse can still be defined as a densely defined operator with domain\n\n\\[\\mathcal D(T^+)=\\mathcal R(T)\\oplus\\mathcal R(T)^\\perp,\\]\n\nbut it is unbounded. This changes the existence of least-squares solutions. An element \\(y\\) may lie in the closure of \\(\\mathcal R(T)\\) without lying in the range itself. Then the infimum of \\(\\|Tx-y\\|\\) is zero, yet no vector attains it. Finite-dimensional least squares always has a minimizer because a matrix range is closed.\n\nA concrete example acts on the space \\(\\ell^2\\) of square-summable sequences:\n\n\\[T(x_1,x_2,\\ldots) = \\left(x_1,\\frac{x_2}{2},\\frac{x_3}{3},\\ldots\\right).\\]\n\nThis operator is bounded and injective, and its range is dense but not closed. Its generalized inverse acts as\n\n\\[T^+(y_1,y_2,\\ldots) = (y_1,2y_2,3y_3,\\ldots)\\]\n\non the domain for which the resulting sequence is square summable. For the unit coordinate vector \\(e_k\\), \\(\\|T^+e_k\\|=k\\), proving that \\(T^+\\) is unbounded. The sequence \\(y=(1,1/2,1/3,\\ldots)\\) belongs to \\(\\ell^2\\) but not to \\(\\mathcal R(T)\\), since its formal preimage is the nonsummable constant sequence.\n\nFor an operator admitting a singular system,\n\n\\[Tv_i=\\sigma_i u_i,\\]\n\nthe finite-dimensional formula becomes\n\n\\[T^+y=\\sum_i\\frac{\\langle y,u_i\\rangle}{\\sigma_i}v_i,\\]\n\nwhere the series converges only for appropriate \\(y\\). In particular, its range component must satisfy\n\n\\[\\sum_i\\frac{|\\langle y,u_i\\rangle|^2}{\\sigma_i^2}<\\infty.\\]\n\nCompact operators of infinite rank have singular values tending to zero, so their inverses on the range are unbounded. This phenomenon appears in integral equations and other problems in functional analysis. Finite discretizations replace unbounded inversion with matrices having increasingly small singular values, linking the operator-theoretic closed-range condition directly to numerical ill-conditioning and the need for regularization."
    }
  ],
  "links": [
    "matrix",
    "inverse-matrix",
    "system-of-linear-equations",
    "least-squares",
    "conjugate-transpose",
    "matrix-transpose",
    "linear-map",
    "inner-product",
    "image-linear-map",
    "null-space",
    "orthogonal-projection",
    "idempotence",
    "singular-value-decomposition",
    "unitary-matrix",
    "matrix-rank",
    "gram-matrix",
    "spectral-theorem",
    "floating-point-arithmetic",
    "qr-decomposition",
    "condition-number",
    "inverse-problem",
    "regularization",
    "linear-regression",
    "ordinary-least-squares",
    "statistical-identification",
    "positive-definite-matrix",
    "jacobian-matrix",
    "graph-theory",
    "graph-laplacian",
    "hilbert-space",
    "functional-analysis"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}