{
  "id": "matrix-norm",
  "title": "Matrix Norm",
  "lang": "en",
  "source_lang": "en",
  "summary": "A matrix norm measures matrix size, supporting quantitative analysis of linear transformations, approximation errors, numerical conditioning, and convergence.",
  "category": "mathematics",
  "subcategory": "algebra",
  "type": "concept",
  "aliases": [
    "Norm of a matrix"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/matrix-norm/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/matrix-norm/",
      "json": "https://aiwiki.page/zh/matrix-norm.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A matrix norm is a norm on a vector space of matrices, assigning each matrix a nonnegative number that measures its size. Depending on the norm, this size may describe the magnitude of its entries, its maximum amplification of vectors, or the aggregate size of its singular values. Matrix norms are fundamental in linear algebra and numerical analysis, where they quantify errors, sensitivity, approximation quality, and convergence."
    },
    {
      "heading": "Definition and conventions",
      "text": "Let \\(\\mathbb F\\) be either the real or complex numbers. A norm on \\(\\mathbb F^{m\\times n}\\) is a function \\(A\\mapsto\\|A\\|\\) satisfying:\n\n- Positive definiteness: \\(\\|A\\|\\geq 0\\), with equality if and only if \\(A=0\\).\n- Absolute homogeneity: \\(\\|\\alpha A\\|=|\\alpha|\\|A\\|\\) for every scalar \\(\\alpha\\).\n- Triangle inequality: \\(\\|A+B\\|\\leq\\|A\\|+\\|B\\|\\).\n\nThe third property is the triangle inequality for matrix addition.\n\nTerminology varies. Some authors call any norm on a matrix space a matrix norm. Others additionally require submultiplicativity:\n\n\\[\\|AB\\|\\leq\\|A\\|\\,\\|B\\|.\\]\n\nFor square matrices, this requirement concerns products within one matrix space. For rectangular matrices, it concerns a compatible family of norms across the dimensions involved.\n\nThe distinction is important: a norm on matrix entries need not be submultiplicative. Unless otherwise specified, properties involving products must be checked rather than inferred from the three norm axioms alone."
    },
    {
      "heading": "Induced norms",
      "text": "A matrix \\(A\\in\\mathbb F^{m\\times n}\\) represents a linear map from \\(\\mathbb F^n\\) to \\(\\mathbb F^m\\). Given norms on these two vector spaces, its induced norm, also called its operator norm or subordinate norm, is\n\n\\[\\|A\\|_{\\alpha\\to\\beta} = \\sup_{x\\neq 0} \\frac{\\|Ax\\|_\\beta}{\\|x\\|_\\alpha} = \\max_{\\|x\\|_\\alpha=1}\\|Ax\\|_\\beta.\\]\n\nThe maximum exists because the unit sphere is compact in finite dimensions.\n\nThis norm measures the greatest amplification of an input vector:\n\n\\[\\|Ax\\|_\\beta\\leq \\|A\\|_{\\alpha\\to\\beta}\\|x\\|_\\alpha.\\]\n\nIt is the smallest constant for which this inequality holds for every \\(x\\).\n\nWhen the domain and codomain use the same type of vector norm, the notation is commonly shortened to \\(\\|A\\|_p\\). Induced norms obey\n\n\\[\\|AB\\|_{\\alpha\\to\\gamma} \\leq \\|A\\|_{\\beta\\to\\gamma}\\, \\|B\\|_{\\alpha\\to\\beta}.\\]\n\nIn particular, each induced \\(p\\)-norm is submultiplicative. The identity matrix satisfies \\(\\|I\\|_p=1\\).\n\nInduced \\(1\\)-norm\n\nFor the vector norm \\(\\|x\\|_1=\\sum_j|x_j|\\), the induced matrix norm is the largest absolute column sum:\n\n\\[\\|A\\|_1=\\max_{1\\leq j\\leq n}\\sum_{i=1}^m|a_{ij}|.\\]\n\nInduced infinity norm\n\nFor the vector norm \\(\\|x\\|_\\infty=\\max_j|x_j|\\), the induced matrix norm is the largest absolute row sum:\n\n\\[\\|A\\|_\\infty=\\max_{1\\leq i\\leq m}\\sum_{j=1}^n|a_{ij}|.\\]\n\nSpectral norm\n\nThe induced norm associated with the Euclidean vector norm is the spectral norm:\n\n\\[\\|A\\|_2 = \\sigma_{\\max}(A) = \\sqrt{\\lambda_{\\max}(A^*A)}.\\]\n\nHere \\(A^*\\) is the conjugate transpose, \\(\\sigma_{\\max}(A)\\) is the largest singular value, and \\(\\lambda_{\\max}(A^*A)\\) is the largest eigenvalue of the positive semidefinite matrix \\(A^*A\\).\n\nGeometrically, the spectral norm is the largest factor by which \\(A\\) stretches a vector in Euclidean space. It can be obtained from the singular value decomposition. Despite its name, it is not generally the largest absolute eigenvalue of \\(A\\)."
    },
    {
      "heading": "Entrywise norms and the Frobenius norm",
      "text": "Matrices can also be treated as vectors of entries. For \\(1\\leq p<\\infty\\), the entrywise \\(p\\)-norm is\n\n\\[\\|A\\|_{\\mathrm{entry},p} = \\left(\\sum_{i,j}|a_{ij}|^p\\right)^{1/p},\\]\n\nand the entrywise maximum norm is\n\n\\[\\|A\\|_{\\max}=\\max_{i,j}|a_{ij}|.\\]\n\nThese are norms, but they generally differ from induced \\(p\\)-norms. In particular, “matrix \\(1\\)-norm” ordinarily means the maximum column sum, not the sum of all absolute entries. Likewise, “matrix infinity norm” ordinarily means the maximum row sum, not the largest absolute entry.\n\nThe maximum-entry norm is not generally submultiplicative with constant \\(1\\). For example, if \\(J\\) is the \\(n\\times n\\) matrix whose entries are all \\(1\\), then\n\n\\[\\|J\\|_{\\max}=1, \\qquad \\|J^2\\|_{\\max}=n.\\]\n\nFor compatible matrices with inner dimension \\(n\\), the general estimate is\n\n\\[\\|AB\\|_{\\max} \\leq n\\|A\\|_{\\max}\\|B\\|_{\\max}.\\]\n\nFrobenius norm\n\nThe entrywise \\(2\\)-norm is the Frobenius norm, also called the Hilbert–Schmidt norm:\n\n\\[\\|A\\|_F = \\left(\\sum_{i,j}|a_{ij}|^2\\right)^{1/2}.\\]\n\nIt is induced by the matrix inner product\n\n\\[\\langle A,B\\rangle_F=\\operatorname{tr}(A^*B),\\]\n\nwhere \\(\\operatorname{tr}\\) denotes the trace. Consequently,\n\n\\[\\|A\\|_F^2 = \\operatorname{tr}(A^*A) = \\sum_j\\sigma_j(A)^2.\\]\n\nThe Frobenius norm is submultiplicative:\n\n\\[\\|AB\\|_F\\leq\\|A\\|_F\\|B\\|_F.\\]\n\nIt is not generally induced by a vector norm on the underlying input and output spaces: for an \\(n\\times n\\) identity matrix,\n\n\\[\\|I\\|_F=\\sqrt n,\\]\n\nwhereas every same-space induced norm gives \\(\\|I\\|=1\\)."
    },
    {
      "heading": "Unitarily invariant norms and Schatten norms",
      "text": "A norm is unitarily invariant if\n\n\\[\\|UAV\\|=\\|A\\|\\]\n\nfor every pair of appropriately sized unitary matrices \\(U,V\\). Over the real numbers, the corresponding transformations use orthogonal matrices.\n\nSuch norms depend only on the singular values. Important examples belong to the family of Schatten norms:\n\n\\[\\|A\\|_{S_p} = \\left(\\sum_j\\sigma_j(A)^p\\right)^{1/p}, \\qquad 1\\leq p<\\infty,\\]\n\nwith\n\n\\[\\|A\\|_{S_\\infty}=\\max_j\\sigma_j(A).\\]\n\nThe principal cases are:\n\n- \\(p=1\\): the nuclear norm or trace norm,\n\n\\[\\|A\\|_*=\\sum_j\\sigma_j(A);\\]\n- \\(p=2\\): the Frobenius norm;\n- \\(p=\\infty\\): the spectral norm.\n\nThus the notation must distinguish the induced \\(1\\)-norm from the Schatten \\(1\\)-norm, and the induced \\(2\\)-norm from the Schatten \\(2\\)-norm.\n\nThe nuclear norm is important in convex optimization because it is a convex surrogate for matrix rank. Penalizing it encourages solutions with relatively few significant singular values, although it does not always produce the same solution as direct rank minimization."
    },
    {
      "heading": "Comparison and equivalence",
      "text": "For \\(A\\in\\mathbb F^{m\\times n}\\) of rank \\(r\\),\n\n\\[\\|A\\|_2\\leq\\|A\\|_F\\leq\\sqrt r\\,\\|A\\|_2,\\]\n\nand\n\n\\[\\|A\\|_F\\leq\\|A\\|_* \\leq\\sqrt r\\,\\|A\\|_F.\\]\n\nThese inequalities follow by comparing the maximum, Euclidean norm, and sum of the singular values.\n\nOther useful bounds are\n\n\\[\\|A\\|_2 \\leq\\sqrt{\\|A\\|_1\\|A\\|_\\infty}\\]\n\nand\n\n\\[\\|A\\|_{\\max} \\leq\\|A\\|_F \\leq\\sqrt{mn}\\,\\|A\\|_{\\max}.\\]\n\nAll norms on a fixed finite-dimensional matrix space are equivalent: for any two norms \\(\\|\\cdot\\|_a,\\|\\cdot\\|_b\\), constants \\(c,C>0\\) exist such that\n\n\\[c\\|A\\|_a\\leq\\|A\\|_b\\leq C\\|A\\|_a\\]\n\nfor every matrix in that space. They therefore define the same topology and the same convergent sequences.\n\nThis does not make norms quantitatively interchangeable. Equivalence constants may depend on dimension and grow without bound when matrix sizes increase."
    },
    {
      "heading": "Eigenvalues, powers, and convergence",
      "text": "The spectral radius of a square matrix is\n\n\\[\\rho(A)=\\max_j|\\lambda_j(A)|.\\]\n\nFor every submultiplicative matrix norm,\n\n\\[\\rho(A)\\leq\\|A\\|.\\]\n\nFor a normal matrix, meaning \\(A^*A=AA^*\\), the spectral norm equals the spectral radius. This equality can fail substantially for nonnormal matrices. For example,\n\n\\[A= \\begin{pmatrix} 0&M\\\\ 0&0 \\end{pmatrix}\\]\n\nhas \\(\\rho(A)=0\\), but \\(\\|A\\|_2=|M|\\).\n\nSubmultiplicativity gives\n\n\\[\\|A^k\\|\\leq\\|A\\|^k.\\]\n\nThus \\(\\|A\\|<1\\) is a sufficient condition for \\(A^k\\to0\\). The exact condition is \\(\\rho(A)<1\\); a particular chosen norm can exceed \\(1\\) even when the powers eventually decay.\n\nMore generally, Gelfand’s formula states that, for any norm on a fixed finite-dimensional square matrix space,\n\n\\[\\rho(A)=\\lim_{k\\to\\infty}\\|A^k\\|^{1/k}.\\]\n\nWhen \\(\\|A\\|<1\\) in a submultiplicative norm, the matrix geometric series satisfies\n\n\\[(I-A)^{-1}=\\sum_{k=0}^{\\infty}A^k.\\]\n\nFor an induced norm, it also yields\n\n\\[\\|(I-A)^{-1}\\| \\leq\\frac{1}{1-\\|A\\|}.\\]"
    },
    {
      "heading": "Numerical applications",
      "text": "Conditioning and error bounds\n\nFor an invertible square matrix, the normwise condition number is\n\n\\[\\kappa(A)=\\|A\\|\\,\\|A^{-1}\\|,\\]\n\nusually defined using an induced norm. In the spectral norm,\n\n\\[\\kappa_2(A)=\\frac{\\sigma_{\\max}(A)}{\\sigma_{\\min}(A)}.\\]\n\nConsider a linear system \\(Ax=b\\), and an approximate solution \\(\\widehat x\\) with residual \\(r=b-A\\widehat x\\). For nonzero \\(b\\), the corresponding induced norm gives\n\n\\[\\frac{\\|x-\\widehat x\\|}{\\|x\\|} \\leq \\kappa(A)\\frac{\\|r\\|}{\\|b\\|}.\\]\n\nA small residual therefore need not imply a small relative solution error when the matrix is ill-conditioned. Conditioning describes sensitivity of the mathematical problem; it is distinct from the numerical stability of an algorithm.\n\nLow-rank approximation\n\nIf\n\n\\[A=U\\Sigma V^*\\]\n\nis a singular value decomposition, retaining only the largest \\(k\\) singular values produces a matrix \\(A_k\\) of rank at most \\(k\\). The Eckart–Young–Mirsky theorem gives\n\n\\[\\min_{\\operatorname{rank}(B)\\leq k}\\|A-B\\|_2 = \\sigma_{k+1}(A)\\]\n\nand\n\n\\[\\min_{\\operatorname{rank}(B)\\leq k}\\|A-B\\|_F = \\left(\\sum_{j>k}\\sigma_j(A)^2\\right)^{1/2}.\\]\n\nThese results underpin matrix compression and methods related to principal component analysis.\n\nRegularization and computational cost\n\nDifferent norms impose different notions of error or model complexity. A Frobenius penalty distributes a squared penalty across all entries or singular values; a spectral-norm constraint controls worst-case Euclidean amplification; a nuclear-norm penalty encourages low rank. Their roles in regularization are therefore not interchangeable.\n\nFor a dense \\(m\\times n\\) matrix, the induced \\(1\\)-norm, infinity norm, maximum-entry norm, and Frobenius norm can be evaluated directly in \\(O(mn)\\) arithmetic operations. Spectral and nuclear norms generally require singular-value computations or iterative approximations. For large sparse matrices, spectral-norm estimates can often be obtained using products with \\(A\\) and \\(A^*\\), without explicitly forming \\(A^*A\\)."
    }
  ],
  "links": [
    "norm-mathematics",
    "vector-space",
    "matrix",
    "linear-algebra",
    "numerical-analysis",
    "triangle-inequality",
    "linear-map",
    "identity-matrix",
    "conjugate-transpose",
    "eigenvalues-and-eigenvectors",
    "euclidean-space",
    "singular-value-decomposition",
    "inner-product",
    "matrix-trace",
    "unitary-matrix",
    "convex-optimization",
    "matrix-rank",
    "topology",
    "geometric-series",
    "condition-number",
    "system-of-linear-equations",
    "numerical-stability",
    "principal-component-analysis",
    "regularization",
    "sparse-matrix"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}