{
  "id": "linear-approximation",
  "title": "Linear Approximation",
  "lang": "en",
  "source_lang": "en",
  "summary": "Linear approximation replaces a differentiable function near a point by its value plus a linear function of the displacement, providing a local model for calculation and analysis.",
  "category": "mathematics",
  "subcategory": "analysis",
  "type": "concept",
  "aliases": [
    "First-order approximation",
    "Tangent-line approximation",
    "Local linear approximation"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/linear-approximation/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/linear-approximation/",
      "json": "https://aiwiki.page/zh/linear-approximation.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Linear approximation is a method of approximating a function near a chosen point by retaining its value and its first-order variation. For a real-valued function \\(f\\) differentiable at \\(a\\), the approximation is\n\n\\[f(x)\\approx f(a)+f'(a)(x-a).\\]\n\nGeometrically, it replaces the graph near \\(a\\) by its tangent line. In several variables, the derivative supplies a linear transformation that describes small changes in the output. Linear approximation is fundamental to calculus, numerical computation, uncertainty propagation, and the local analysis of nonlinear systems."
    },
    {
      "heading": "Definition and mathematical meaning",
      "text": "For a function \\(f:\\mathbb{R}\\to\\mathbb{R}\\) with a derivative at \\(a\\), its linear approximation at \\(a\\) is\n\n\\[L_a(x)=f(a)+f'(a)(x-a).\\]\n\nWriting \\(h=x-a\\) gives the equivalent form\n\n\\[f(a+h)\\approx f(a)+f'(a)h.\\]\n\nThe approximation is justified by the defining property of differentiability:\n\n\\[f(a+h)=f(a)+f'(a)h+r(h), \\qquad \\lim_{h\\to0}\\frac{|r(h)|}{|h|}=0.\\]\n\nThus, the approximation error becomes negligible relative to the size of the displacement \\(h\\). This is stronger than merely requiring the absolute error to tend to zero.\n\nThe function \\(L_a\\) is generally an affine map, rather than a linear map in the strict algebraic sense: it need not send zero to zero. The expression “linear approximation” refers to its degree-one dependence on the variable, or to the fact that the output change\n\n\\[L_a(a+h)-L_a(a)=f'(a)h\\]\n\nis a linear map of the input change.\n\nThe derivative is the unique coefficient with this first-order approximation property. If a number \\(m\\) satisfies\n\n\\[f(a+h)=f(a)+mh+o(|h|),\\]\n\nthen \\(m=f'(a)\\)."
    },
    {
      "heading": "Geometric interpretation and examples",
      "text": "The graph of \\(L_a\\) passes through \\((a,f(a))\\) and has slope \\(f'(a)\\). It is therefore the tangent line to the graph of \\(f\\) at that point. Differentiability means that, under sufficiently strong magnification near the point, the difference between the curve and this line is small compared with horizontal displacement.\n\nFor example, let \\(f(x)=\\sqrt{x}\\) and choose \\(a=4\\). Since\n\n\\[f(4)=2,\\qquad f'(4)=\\frac14,\\]\n\nthe approximation is\n\n\\[\\sqrt{x}\\approx 2+\\frac{x-4}{4}.\\]\n\nConsequently,\n\n\\[\\sqrt{4.1}\\approx 2.025,\\]\n\nwhereas the actual value is approximately \\(2.024846\\).\n\nSeveral familiar small-argument formulas are linear approximations at zero:\n\n\\[\\sin h\\approx h,\\qquad e^h\\approx 1+h,\\qquad \\ln(1+h)\\approx h.\\]\n\nThe trigonometric formula assumes that angles are measured in radians. Each approximation is local; its usefulness depends on the size of \\(h\\) and the accuracy required.\n\nWhen \\(f'(a)=0\\), the linear approximation is constant. For \\(f(x)=x^2\\) at \\(a=0\\), it gives \\(L_0(x)=0\\). This remains a valid first-order approximation because \\(x^2/|x|\\to0\\), although it does not describe the leading nonzero variation of the function."
    },
    {
      "heading": "Error estimates and Taylor expansion",
      "text": "Linear approximation is the first-degree truncation of a Taylor expansion, but its validity requires only differentiability, not the existence or convergence of an infinite Taylor series.\n\nDifferentiability alone gives\n\n\\[f(a+h)-L_a(a+h)=o(|h|).\\]\n\nIt does not, by itself, imply that the error is proportional to \\(h^2\\).\n\nIf \\(f\\) has a continuous second derivative on an interval containing \\(a\\) and \\(x\\), Taylor’s theorem gives\n\n\\[f(x)-L_a(x)=\\frac12 f''(\\xi)(x-a)^2\\]\n\nfor some \\(\\xi\\) between \\(a\\) and \\(x\\). Hence, if \\(|f''(t)|\\le M\\) throughout that interval,\n\n\\[|f(x)-L_a(x)|\\le \\frac{M}{2}|x-a|^2.\\]\n\nIn asymptotic notation, the error is then \\(O(|x-a|^2)\\).\n\nThe second derivative measures how rapidly the slope changes. A large displacement or substantial variation in slope can make the tangent-line model inaccurate. For a differentiable convex function on an interval, every tangent line lies at or below the graph throughout that interval; for a concave function, it lies at or above the graph.\n\nAbsolute and relative accuracy must also be distinguished. If \\(f(a)=0\\), an approximation can have a small absolute error while having a large relative error. For example, the constant approximation \\(L_0(x)=0\\) to \\(x^2\\) has relative error \\(1\\) at every nonzero \\(x\\)."
    },
    {
      "heading": "Several variables",
      "text": "For a differentiable function \\(f:\\mathbb{R}^n\\to\\mathbb{R}\\), the approximation at a point \\(a\\) is\n\n\\[f(a+h)\\approx f(a)+\\nabla f(a)\\cdot h,\\]\n\nwhere \\(\\nabla f(a)\\) is the gradient and the dot denotes the dot product. In coordinates,\n\n\\[f(a+h)\\approx f(a)+ \\sum_{i=1}^{n}\\frac{\\partial f}{\\partial x_i}(a)h_i.\\]\n\nFor example, if \\(f(x,y)=xy\\), then near \\((a,b)\\),\n\n\\[f(a+h,b+k)=ab+bh+ak+hk.\\]\n\nThe linear approximation retains \\(ab+bh+ak\\) and discards the second-order product \\(hk\\).\n\nGeometrically, for a scalar-valued function of \\(n\\) variables, this affine approximation describes a tangent hyperplane to its graph.\n\nMore generally, let \\(F:\\mathbb{R}^n\\to\\mathbb{R}^m\\). Its derivative at \\(a\\) is represented by the Jacobian matrix \\(J_F(a)\\), giving\n\n\\[F(a+h)=F(a)+J_F(a)h+r(h), \\qquad \\frac{\\|r(h)\\|}{\\|h\\|}\\to0.\\]\n\nThe derivative therefore describes the local action of a possibly nonlinear function through a matrix.\n\nThe existence of all partial derivatives at a point is not sufficient to establish this approximation. Differentiability requires a single linear map whose error is negligible for displacements approaching zero in every direction. Continuous partial derivatives in a neighborhood provide a standard sufficient condition.\n\nIf a scalar-valued function has continuous second derivatives near \\(a\\), its Hessian matrix describes the leading error:\n\n\\[f(a+h)=f(a)+\\nabla f(a)\\cdot h +\\frac12 h^{\\mathsf T}H_f(a)h+o(\\|h\\|^2).\\]"
    },
    {
      "heading": "Applications",
      "text": "Numerical root finding\n\nNewton’s method approximates a function locally and finds where that approximation vanishes. Setting\n\n\\[0=f(x_k)+f'(x_k)(x_{k+1}-x_k)\\]\n\ngives\n\n\\[x_{k+1}=x_k-\\frac{f(x_k)}{f'(x_k)}.\\]\n\nFor systems of equations, the corresponding step is obtained by solving a linear system involving the Jacobian. This is a local procedure: convergence depends on the function, the starting point, and the behavior of its derivative.\n\nSensitivity and uncertainty propagation\n\nIn sensitivity analysis, a small input perturbation \\(\\delta x\\) produces the approximate output change\n\n\\[\\delta y\\approx J_F(a)\\delta x.\\]\n\nThe derivative identifies which input changes most strongly affect the output near the operating point.\n\nIf a random input \\(X\\) is concentrated near its mean \\(\\mu\\), with covariance matrix \\(\\Sigma\\), first-order uncertainty propagation gives\n\n\\[\\operatorname{Cov}(F(X)) \\approx J_F(\\mu)\\Sigma J_F(\\mu)^{\\mathsf T}.\\]\n\nFor scalar output, this becomes\n\n\\[\\operatorname{Var}(f(X)) \\approx \\nabla f(\\mu)^{\\mathsf T}\\Sigma\\nabla f(\\mu).\\]\n\nThese formulas neglect nonlinear contributions, which may be important when uncertainty is large or the derivative vanishes.\n\nDifferential equations and local dynamics\n\nFor a dynamical system\n\n\\[\\dot{x}=F(x),\\]\n\nlet \\(a\\) be an equilibrium, so that \\(F(a)=0\\). Writing \\(x=a+h\\) yields the linearized system\n\n\\[\\dot{h}=J_F(a)h.\\]\n\nThe eigenvalues of the Jacobian help determine local stability. For a continuously differentiable system, eigenvalues with strictly negative real parts imply local asymptotic stability; an eigenvalue with positive real part implies instability. When eigenvalues have zero real part and none have positive real part, linearization alone generally does not decide stability.\n\nOptimization\n\nIn optimization, the local model\n\n\\[f(x+h)\\approx f(x)+\\nabla f(x)\\cdot h\\]\n\nidentifies first-order directions of increase and decrease. It underlies gradient-based algorithms and linearized constraints. At a point where the gradient vanishes, second-order or higher-order information is generally needed to distinguish a minimum, maximum, or saddle point."
    },
    {
      "heading": "Distinctions and limitations",
      "text": "Linear approximation is not the same as linear regression. A tangent approximation is determined by a function’s value and derivative at one point. A regression line is fitted to observations according to a statistical criterion and need not be tangent to an underlying function.\n\nIt also differs from linear interpolation. An interpolating line passes through two specified points; a tangent approximation matches the value and slope at one point. Piecewise linear approximations use multiple affine pieces to cover a wider domain.\n\nA local model need not remain accurate far from its expansion point. It can miss curvature, turning points, constraints, and changes in qualitative behavior. At a nondifferentiable point, such as \\(f(x)=|x|\\) at zero, no single linear map provides a first-order approximation valid from both sides. Even where an approximation exists, the identity function\n\n\\[\\frac{1}{1-x}\\approx 1+x\\]\n\nnear zero illustrates why local agreement cannot justify extrapolation toward a singularity at \\(x=1\\)."
    }
  ],
  "links": [
    "mathematical-function",
    "calculus",
    "derivative",
    "affine-map",
    "linear-map",
    "graph-of-a-function",
    "radian",
    "taylor-series",
    "big-o-notation",
    "convex-function",
    "gradient",
    "dot-product",
    "hyperplane",
    "jacobian-matrix",
    "partial-derivative",
    "hessian-matrix",
    "newtons-method",
    "sensitivity-analysis",
    "covariance-matrix",
    "measurement-uncertainty",
    "dynamical-system",
    "eigenvalues-and-eigenvectors",
    "mathematical-optimization",
    "linear-regression",
    "interpolation"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}