{
  "id": "differentiable-function",
  "title": "Differentiable Function",
  "lang": "en",
  "source_lang": "en",
  "summary": "A differentiable function admits a linear approximation at each point, providing the foundation for differential calculus and its applications.",
  "category": "mathematics",
  "subcategory": "analysis",
  "type": "concept",
  "aliases": [
    "Differentiable map"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/differentiable-function/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/differentiable-function/",
      "json": "https://aiwiki.page/zh/differentiable-function.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A differentiable function is a function whose change near a point can be approximated by a linear map, with an error that becomes negligible relative to the size of the displacement. For a real-valued function of one real variable, this is equivalent to the existence of a finite derivative. In several variables, differentiability requires a single linear approximation that works for displacements in every direction, rather than merely the existence of separate rates of change along coordinate axes.\n\nDifferentiability is a central concept of calculus and mathematical analysis. It makes precise the ideas of instantaneous velocity, tangent slope, local sensitivity, and first-order approximation. Its consequences include rules for differentiating compositions, criteria for extrema, local inversion theorems, and methods for solving equations. The concept extends from functions on Euclidean spaces to maps between infinite-dimensional spaces and smooth manifolds, while weaker notions of differentiation allow the treatment of functions that have corners, singularities, or only limited regularity."
    },
    {
      "heading": "Differentiability in one real variable",
      "text": "Let \\(f:U\\to\\mathbb R\\), where \\(U\\) is an open subset of the real numbers, and let \\(a\\in U\\). The function is differentiable at \\(a\\) if the limit\n\n\\[f'(a)=\\lim_{h\\to0}\\frac{f(a+h)-f(a)}{h}\\]\n\nexists as a finite real number. The quotient measures the average change in the function over a displacement \\(h\\); its limit gives the instantaneous rate of change. The requirement that the limit be finite excludes an infinite slope from ordinary real differentiability. For example, \\(f(x)=x^{1/3}\\) has a vertical tangent at zero, but it is not differentiable there in this sense.\n\nAn equivalent and more generalizable formulation is\n\n\\[f(a+h)=f(a)+f'(a)h+r(h), \\qquad \\lim_{h\\to0}\\frac{|r(h)|}{|h|}=0.\\]\n\nThe remainder is written \\(r(h)=o(|h|)\\), a form of asymptotic notation related to big-O notation. This expression identifies differentiability with the existence of a first-order linear approximation. The approximating function of \\(h\\), namely \\(f(a)+f'(a)h\\), is affine rather than linear unless \\(f(a)=0\\); its variable part \\(h\\mapsto f'(a)h\\) is linear.\n\nFor example, if \\(f(x)=x^2\\), then\n\n\\[f(a+h)=a^2+2ah+h^2.\\]\n\nThe remainder \\(h^2\\) satisfies \\(h^2/|h|\\to0\\), so \\(f'(a)=2a\\). More generally, every polynomial is differentiable everywhere. Rational functions are differentiable wherever their denominators are nonzero, and the exponential, logarithmic, and trigonometric functions are differentiable on their respective open domains.\n\nGeometrically, differentiability means that the graph has a tangent line whose equation is\n\n\\[y=f(a)+f'(a)(x-a).\\]\n\nA tangent is not required to intersect the graph at only one point, nor must it remain entirely above or below the graph. The defining property is local approximation. For \\(f(x)=x^3\\), the tangent at the origin is horizontal even though the graph crosses it.\n\nDifferentiability is initially a property at a point. A function is differentiable on an open set if it is differentiable at every point of that set. At an endpoint of an interval, one often uses a one-sided derivative, with \\(h\\) restricted to positive or negative values. Consequently, “differentiable on a closed interval” commonly means differentiable in the interior with appropriate one-sided derivatives at the endpoints. An extension to an open neighborhood is a separate condition and should not be silently assumed."
    },
    {
      "heading": "Continuity, corners, and irregular derivatives",
      "text": "Differentiability at a point implies continuity there. Indeed, the expansion\n\n\\[f(a+h)-f(a)=f'(a)h+o(|h|)\\]\n\nshows that the change in the function tends to zero with \\(h\\). The converse is false. The function \\(f(x)=|x|\\) is continuous everywhere, but its difference quotient at zero equals \\(1\\) for positive \\(h\\) and \\(-1\\) for negative \\(h\\). Its corner therefore prevents a derivative from existing.\n\nNondifferentiability can also arise through rapid oscillation. Define \\(f(0)=0\\) and \\(f(x)=x\\sin(1/x)\\) for \\(x\\ne0\\). The bound \\(|f(x)|\\leq |x|\\) proves continuity at zero, but\n\n\\[\\frac{f(h)-f(0)}{h}=\\sin(1/h)\\]\n\nhas no limit. This behavior differs from the corner of the absolute-value function: there are not two stable one-sided slopes, but infinitely many oscillations at increasingly small scales.\n\nA differentiable function need not have a continuous derivative. For\n\n\\[g(x)= \\begin{cases} x^2\\sin(1/x),&x\\ne0,\\\\ 0,&x=0, \\end{cases}\\]\n\nthe derivative at zero is \\(g'(0)=0\\), whereas for \\(x\\ne0\\),\n\n\\[g'(x)=2x\\sin(1/x)-\\cos(1/x).\\]\n\nThe derivative oscillates without approaching zero. Replacing \\(x^2\\) by \\(x^2\\sin(1/x^2)\\) in the defining expression produces a differentiable function whose derivative is unbounded in every neighborhood of zero. Thus pointwise differentiability does not by itself provide a uniform bound on local rates of change.\n\nNevertheless, derivatives cannot behave arbitrarily. Darboux’s theorem states that every derivative has the intermediate value property: if \\(f'(a)<c<f'(b)\\), then \\(f'(x)=c\\) at some point between \\(a\\) and \\(b\\), assuming differentiability on an interval containing them. Consequently, a derivative cannot have a jump discontinuity, even though it may be discontinuous through oscillation. Derivatives are also pointwise limits of continuous difference-quotient functions locally, a further restriction on their possible irregularity.\n\nContinuity alone allows much more extreme behavior. A nowhere-differentiable function is continuous but has no finite derivative at any point of its domain. Classical constructions show that an apparently plausible connection between continuity and the existence of tangents is false. These examples are not merely exceptions produced by a few isolated corners: nondifferentiability can occur at every scale and every point. Their existence helped establish that continuity, differentiability, and continuity of the derivative are genuinely different levels of regularity."
    },
    {
      "heading": "Algebraic rules and theorems controlled by derivatives",
      "text": "Differentiable functions are stable under the basic algebraic operations. If \\(f\\) and \\(g\\) are differentiable at \\(a\\), then so are their sum and product, with\n\n\\[(f+g)'(a)=f'(a)+g'(a),\\]\n\n\\[(fg)'(a)=f'(a)g(a)+f(a)g'(a).\\]\n\nIf \\(g(a)\\ne0\\), continuity ensures that the quotient is defined near \\(a\\), and\n\n\\[\\left(\\frac fg\\right)'(a) = \\frac{f'(a)g(a)-f(a)g'(a)}{g(a)^2}.\\]\n\nThese rules follow by inserting first-order expansions and checking that products involving remainder terms remain negligible compared with \\(|h|\\).\n\nThe chain rule governs composition. If \\(g\\) is differentiable at \\(a\\) and \\(f\\) is differentiable at \\(g(a)\\), then\n\n\\[(f\\circ g)'(a)=f'(g(a))g'(a).\\]\n\nThis rule is more than a convenient symbolic identity: it expresses the composition of local linear approximations. It allows a change in an input to be propagated through successive dependencies, as in a physical measurement model or a computational graph.\n\nThe mean value theorem connects pointwise derivatives with changes over an entire interval. If \\(f\\) is continuous on \\([a,b]\\) and differentiable on \\((a,b)\\), there exists \\(c\\in(a,b)\\) such that\n\n\\[f(b)-f(a)=f'(c)(b-a).\\]\n\nIt follows that a function with zero derivative throughout an interval is constant there. A nonnegative derivative implies that the function is nondecreasing, while a strictly positive derivative implies that it is strictly increasing. These conclusions depend on working on an interval; a function with zero derivative on a disconnected open domain may take different constant values on different components.\n\nDerivative bounds also control continuity quantitatively. If \\(|f'(x)|\\le M\\) on an interval, then\n\n\\[|f(x)-f(y)|\\le M|x-y|.\\]\n\nThus \\(f\\) satisfies Lipschitz continuity with constant \\(M\\). A continuously differentiable function is locally Lipschitz, because its derivative is locally bounded. Mere differentiability is insufficient, as functions with locally unbounded derivatives demonstrate.\n\nThe fundamental theorem of calculus links differentiation to integration. If \\(g\\) is continuous and\n\n\\[F(x)=\\int_a^x g(t)\\,dt,\\]\n\nthen \\(F'(x)=g(x)\\). Conversely, if \\(f\\) is differentiable and \\(f'\\) is Riemann integrable on \\([a,b]\\), then\n\n\\[f(b)-f(a)=\\int_a^b f'(x)\\,dx.\\]\n\nThe integrability qualification matters: differentiability alone does not guarantee that the derivative is bounded or Riemann integrable. More general versions use absolute continuity and Lebesgue integration, with hypotheses designed to ensure that integration recovers all of the function’s change."
    },
    {
      "heading": "Total differentiability in several variables",
      "text": "For a function \\(f:U\\to\\mathbb R^m\\), where \\(U\\) is an open subset of Euclidean space \\(\\mathbb R^n\\), differentiability at \\(a\\) means that there is a linear map \\(A:\\mathbb R^n\\to\\mathbb R^m\\) satisfying\n\n\\[\\lim_{h\\to0} \\frac{\\|f(a+h)-f(a)-Ah\\|}{\\|h\\|}=0.\\]\n\nThe map \\(A\\), uniquely determined when it exists, is denoted \\(Df(a)\\). This definition is often called total differentiability or Fréchet differentiability. It requires the approximation to work as \\(h\\) approaches zero along arbitrary paths, not just along individual lines.\n\nIn standard coordinates, \\(Df(a)\\) is represented by the Jacobian matrix\n\n\\[J_f(a)= \\left( \\frac{\\partial f_i}{\\partial x_j}(a) \\right)_{\\substack{1\\le i\\le m\\\\1\\le j\\le n}}.\\]\n\nFor a scalar-valued function, this is a row matrix. Its coefficients can equivalently be collected in the gradient vector, giving\n\n\\[Df(a)h=\\nabla f(a)\\cdot h.\\]\n\nThe differential itself is a linear functional; its identification with a gradient vector uses the Euclidean inner product. This distinction becomes important in non-Euclidean geometry and infinite-dimensional spaces.\n\nExistence of all partial derivatives at a point does not imply differentiability or even continuity there. For example, set\n\n\\[p(x,y)= \\begin{cases} \\dfrac{xy}{x^2+y^2},&(x,y)\\ne(0,0),\\\\ 0,&(x,y)=(0,0). \\end{cases}\\]\n\nBoth coordinate partial derivatives at the origin are zero, since the function vanishes on the coordinate axes. Along \\(y=x\\), however, its value is \\(1/2\\) away from the origin, so it is not continuous there.\n\nEven the existence of every directional derivative is insufficient. Define\n\n\\[q(x,y)= \\begin{cases} \\dfrac{x^3}{x^2+y^2},&(x,y)\\ne(0,0),\\\\ 0,&(x,y)=(0,0). \\end{cases}\\]\n\nFor a nonzero direction \\(v=(v_1,v_2)\\), the directional derivative at zero is\n\n\\[D_vq(0)=\\frac{v_1^3}{v_1^2+v_2^2}.\\]\n\nThis dependence on \\(v\\) is not linear, so no total derivative can represent it. Moreover, even when directional derivatives depend linearly on direction, their existence alone does not generally supply the uniform error estimate required by total differentiability.\n\nA standard sufficient condition is that all first partial derivatives exist in a neighborhood of \\(a\\) and are continuous at \\(a\\). Then \\(f\\) is differentiable at \\(a\\). Continuity of the partial derivatives is sufficient, not necessary. Once total differentiability is established, the multivariable chain rule takes the compact form\n\n\\[D(f\\circ g)(a)=Df(g(a))\\,Dg(a),\\]\n\nrepresented in coordinates by multiplication of the corresponding Jacobian matrices."
    },
    {
      "heading": "Higher derivatives and degrees of smoothness",
      "text": "Differentiability can be iterated. A function is twice differentiable if its first derivative is itself differentiable, and analogous definitions apply to higher orders. The notation \\(C^k(U)\\) means that all derivatives through order \\(k\\) exist and are continuous on \\(U\\). The class \\(C^0\\) consists of continuous functions, while \\(C^\\infty\\) consists of smooth functions with continuous derivatives of every finite order.\n\nThese classes form a hierarchy, with each higher regularity condition imposing additional restrictions. A differentiable function need not belong to \\(C^1\\), and a \\(C^1\\) function need not be twice differentiable. For example,\n\n\\[f(x)=x|x|\\]\n\nhas derivative \\(f'(x)=2|x|\\), including \\(f'(0)=0\\). It is therefore \\(C^1\\), but its first derivative has a corner at zero, so \\(f''(0)\\) does not exist. Powers and absolute values provide many explicit examples separating successive regularity classes.\n\nIn several variables, the second derivative is a bilinear map. For a scalar-valued function, it is represented by the Hessian matrix\n\n\\[H_f(a)= \\left( \\frac{\\partial^2 f}{\\partial x_i\\partial x_j}(a) \\right).\\]\n\nUnder the usual hypothesis that second partial derivatives are continuous near the point, the mixed partial derivatives agree, so the Hessian is symmetric. Without suitable regularity, the two orders of mixed differentiation can produce different values. Higher derivatives are naturally represented by multilinear maps rather than simply by vectors or matrices.\n\nTaylor’s theorem expresses local behavior through derivatives. If \\(f\\) is \\(C^k\\) near \\(a\\), then\n\n\\[f(a+h)= \\sum_{j=0}^{k}\\frac{f^{(j)}(a)}{j!}h^j +o(|h|^k).\\]\n\nIf one more derivative exists with appropriate bounds, a more explicit remainder estimate is available. For a \\(C^2\\) scalar-valued function in several variables,\n\n\\[f(a+h)=f(a)+\\nabla f(a)\\cdot h +\\frac12 h^{\\mathsf T}H_f(a)h +o(\\|h\\|^2).\\]\n\nThe first-order term gives a tangent approximation; the second-order term describes curvature and often determines local stability or the type of a critical point.\n\nSmoothness does not imply representation by a convergent Taylor series. The function\n\n\\[\\phi(x)= \\begin{cases} e^{-1/x^2},&x\\ne0,\\\\ 0,&x=0 \\end{cases}\\]\n\nis \\(C^\\infty\\), and every derivative at zero vanishes. Its Taylor series there is identically zero, although the function is positive at every nonzero real argument. An analytic function, by contrast, agrees locally with a convergent power series. This distinction between smooth and analytic regularity is fundamental in differential equations, approximation, and geometry."
    },
    {
      "heading": "Local inversion, constraints, and optimization",
      "text": "An invertible derivative can make a nonlinear map locally behave like an invertible linear map. The inverse function theorem states that if \\(f:U\\subseteq\\mathbb R^n\\to\\mathbb R^n\\) is \\(C^1\\) and \\(Df(a)\\) is invertible, then \\(f\\) has a continuously differentiable local inverse near \\(a\\). Its derivative satisfies\n\n\\[D(f^{-1})(f(a))=[Df(a)]^{-1}.\\]\n\nThe theorem is local: it does not assert that \\(f\\) is one-to-one on all of \\(U\\). For instance, \\(x\\mapsto x^2\\) has an invertible derivative at every nonzero point, but it is not globally one-to-one on the real line.\n\nThe implicit function theorem treats equations in which some variables are determined indirectly. Suppose \\(F(x,y)=0\\), where \\(F\\) has \\(m\\) components and \\(y\\in\\mathbb R^m\\). If \\(F\\) is \\(C^1\\), \\(F(a,b)=0\\), and the partial derivative \\(D_yF(a,b)\\) is invertible, then the equation locally defines \\(y=g(x)\\), with\n\n\\[Dg(a)=-[D_yF(a,b)]^{-1}D_xF(a,b).\\]\n\nFor the circle \\(x^2+y^2=1\\), a local expression of \\(y\\) in terms of \\(x\\) exists wherever \\(y\\ne0\\). At \\((1,0)\\), this particular choice of dependent variable fails, although \\(x\\) can instead be expressed smoothly in terms of \\(y\\).\n\nDifferentiability also provides necessary conditions for optimization. If a differentiable scalar-valued function has a local maximum or minimum at an interior point \\(a\\), then\n\n\\[Df(a)=0, \\qquad\\text{or equivalently}\\qquad \\nabla f(a)=0.\\]\n\nThis follows by restricting the function to every line through \\(a\\). The condition is not sufficient: \\(f(x)=x^3\\) has zero derivative at zero without having a local extremum there. Boundary extrema require separate treatment because not every direction is available.\n\nSecond derivatives sharpen the classification. At a stationary point of a \\(C^2\\) function, a positive-definite Hessian implies a strict local minimum, a negative-definite Hessian implies a strict local maximum, and an indefinite Hessian gives a saddle point. A singular or merely semidefinite Hessian may be inconclusive. The functions \\(x^4\\), \\(-x^4\\), and \\(x^3\\) all have vanishing first and second derivatives at zero, yet exhibit different local behavior.\n\nFor a differentiable convex function on an open convex domain, the first-order inequality\n\n\\[f(y)\\ge f(x)+\\nabla f(x)\\cdot(y-x)\\]\n\nholds throughout the domain. Consequently, a stationary point is a global minimum. Differentiability also supports numerical algorithms such as Newton’s method, which repeatedly replaces a nonlinear equation by its local linear approximation. The accuracy and convergence of such methods depend on more than derivative existence: continuity, nonsingularity, and bounds on derivative variation control whether the approximations remain reliable."
    },
    {
      "heading": "Complex differentiability and analytic rigidity",
      "text": "For a function \\(f:U\\to\\mathbb C\\), where \\(U\\) is open in the complex numbers, complex differentiability at \\(z_0\\) means that\n\n\\[f'(z_0)= \\lim_{\\substack{h\\to0\\\\h\\in\\mathbb C}} \\frac{f(z_0+h)-f(z_0)}{h}\\]\n\nexists. The increment \\(h\\) can approach zero through any direction in the plane. A function complex differentiable at every point of an open set is called holomorphic there. This apparently modest change from real to complex scalars produces a much stronger theory.\n\nWriting \\(f=u+iv\\) and \\(z=x+iy\\), complex differentiability implies the Cauchy–Riemann equations\n\n\\[u_x=v_y, \\qquad u_y=-v_x.\\]\n\nIf \\(u\\) and \\(v\\) are real differentiable at a point, these equations are also sufficient for complex differentiability there. They require the real derivative matrix to have the form\n\n\\[\\begin{pmatrix} a&-b\\\\ b&a \\end{pmatrix},\\]\n\nwhich is exactly the matrix representing multiplication by the complex number \\(a+ib\\). Thus the real linear approximation must respect complex multiplication, not merely real addition and scaling.\n\nComplex conjugation illustrates the distinction. The function \\(f(z)=\\overline z\\) is a smooth map from \\(\\mathbb R^2\\) to itself, but\n\n\\[\\frac{\\overline h}{h}\\]\n\nequals \\(1\\) for real \\(h\\) and \\(-1\\) for purely imaginary \\(h\\). It is therefore nowhere complex differentiable. Conversely, differentiability at one isolated point is weak: \\(f(z)=|z|^2\\) has complex derivative zero at the origin but is not holomorphic on any neighborhood of it.\n\nThe decisive rigidity arises from differentiability on an open set. Fundamental results of complex analysis imply that every holomorphic function has derivatives of all orders and admits a local convergent power series. In particular,\n\n\\[f(z)=\\sum_{n=0}^{\\infty} \\frac{f^{(n)}(z_0)}{n!}(z-z_0)^n\\]\n\nthroughout a sufficiently small disk about each point \\(z_0\\). Unlike the real case, there is no distinction between holomorphicity and local complex analyticity.\n\nThis rigidity has substantial consequences. On a connected domain, holomorphic functions that agree on a set with an accumulation point inside the domain agree everywhere. Their real and imaginary parts are harmonic, and nonconstant holomorphic functions cannot attain a local maximum of their modulus at an interior point. These properties explain why complex differentiability is not merely real differentiability written with complex notation: it imposes strong constraints on both local and global behavior."
    },
    {
      "heading": "Differentiation on normed spaces and manifolds",
      "text": "The linear-approximation definition extends to maps between normed vector spaces. Let \\(X\\) and \\(Y\\) be such spaces, \\(U\\subseteq X\\) open, and \\(F:U\\to Y\\). The Fréchet derivative at \\(a\\) is a bounded linear operator \\(A:X\\to Y\\) such that\n\n\\[\\lim_{\\|h\\|_X\\to0} \\frac{\\|F(a+h)-F(a)-Ah\\|_Y}{\\|h\\|_X}=0.\\]\n\nBoundedness is automatic for linear maps between finite-dimensional normed spaces, but not in infinite dimensions. It ensures that the proposed first-order approximation is itself continuous.\n\nA weaker approach examines one direction at a time. The directional limit\n\n\\[\\lim_{t\\to0}\\frac{F(a+th)-F(a)}{t}\\]\n\nmay exist for each fixed \\(h\\). When these limits define a linear operator, one obtains a notion commonly called the Gâteaux derivative, although conventions differ over whether continuity of that operator is included. Fréchet differentiability implies the corresponding directional derivatives. The converse generally fails because directional limits need not control the approximation uniformly as the direction changes.\n\nSuch distinctions are central in functional analysis, where the variables may themselves be functions. On a real Hilbert space, for example,\n\n\\[J(u)=\\frac12\\|u\\|^2\\]\n\nhas derivative\n\n\\[DJ(u)h=\\langle u,h\\rangle.\\]\n\nThe remainder is \\(\\frac12\\|h\\|^2\\), which is negligible relative to \\(\\|h\\|\\). The derivative is a bounded linear functional, and the inner product represents it by the vector \\(u\\), the gradient of \\(J\\). This framework underlies variational formulations in which an entire curve, field, or state function is the unknown.\n\nOn a smooth manifold, subtraction of nearby points is generally not intrinsically defined. Differentiability is therefore described using coordinate charts. A map is \\(C^k\\) if its coordinate expressions are \\(C^k\\) maps between Euclidean open sets. Smooth compatibility between charts ensures that this property does not depend on the coordinate choice. For the standard smooth category, \\(k=\\infty\\).\n\nThe derivative of a smooth map \\(F:M\\to N\\) at \\(p\\) is an intrinsic linear map\n\n\\[dF_p:T_pM\\to T_{F(p)}N\\]\n\nbetween tangent spaces. Coordinate changes alter its matrix representation but not the underlying map. The chain rule remains valid, making differentiation compatible with successive maps between manifolds. For a real-valued function, \\(df_p\\) is naturally a covector; converting it into a gradient vector requires a chosen metric. This distinction is essential in differential geometry, where different metrics can assign different gradients to the same differential."
    },
    {
      "heading": "Almost-everywhere and weak differentiability",
      "text": "Classical differentiability requires a pointwise linear approximation, but many useful functions fail this condition on exceptional sets. Rademacher’s theorem states that a locally Lipschitz map from an open subset of \\(\\mathbb R^n\\) to \\(\\mathbb R^m\\) is differentiable almost everywhere, meaning outside a set of Lebesgue measure zero. Thus a Lipschitz function may have corners, but its first-order behavior is well defined at almost every point.\n\nThis result does not mean that arbitrary continuous functions are almost everywhere differentiable. Nor does it make all singular sets harmless for every purpose. A single nondifferentiable point can prevent a classical optimization condition from applying at a minimizer. Nevertheless, almost-everywhere differentiability is sufficient for many integral identities and variational arguments. It also explains why piecewise linear functions can support a substantial differential theory despite their corners.\n\nA different extension is the weak derivative, defined through integration rather than pointwise difference quotients. If \\(u\\) and \\(v\\) are locally integrable on an open set \\(U\\subseteq\\mathbb R^n\\), then \\(v\\) is the weak derivative \\(\\partial_i u\\) if\n\n\\[\\int_U u\\,\\partial_i\\varphi\\,dx = -\\int_U v\\,\\varphi\\,dx\\]\n\nfor every smooth test function \\(\\varphi\\) with compact support in \\(U\\). This identity transfers differentiation from the possibly irregular function \\(u\\) to the smooth test function. It reproduces integration by parts without boundary terms.\n\nFor \\(u(x)=|x|\\), the weak first derivative is the sign function, with its value at zero irrelevant to the integral identity. The function is not classically differentiable at zero, but it has a locally bounded weak derivative. By contrast, the unit step function has a distributional derivative equal to a Dirac point mass; that derivative is not represented by an ordinary locally integrable function. Distributional differentiation therefore extends beyond weak derivatives represented by functions.\n\nThe spaces \\(W^{k,p}\\), known as Sobolev spaces, consist of functions whose weak derivatives through order \\(k\\) belong to suitable \\(L^p\\) spaces. They are fundamental in partial differential equations, where solutions may exist naturally with only weak regularity. Weak differentiability alone does not imply classical differentiability everywhere, and the extent to which it gives continuous or smoother representatives depends on the dimension, derivative order, and integrability exponent.\n\nIn one dimension, absolute continuity supplies a particularly precise bridge. An absolutely continuous function on \\([a,b]\\) is differentiable almost everywhere, its derivative is Lebesgue integrable, and\n\n\\[f(x)=f(a)+\\int_a^x f'(t)\\,dt.\\]\n\nContinuity and almost-everywhere differentiability alone do not guarantee this representation. The Cantor function, for example, is continuous and has derivative zero almost everywhere, yet increases from zero to one. Absolute continuity excludes this kind of change concentrated on a measure-zero set."
    },
    {
      "heading": "Historical development and computational realization",
      "text": "The concept of differentiability emerged from problems about tangents, motion, and extrema. Seventeenth-century methods associated with Pierre de Fermat, Isaac Newton, and Gottfried Wilhelm Leibniz developed procedures for calculating rates of change before the modern limit definition had been fully established. Newton’s fluxions treated quantities as changing with time, while Leibniz’s differential notation emphasized relations between infinitesimal changes. Their approaches helped unify problems that had previously been treated separately.\n\nDuring the nineteenth century, limit-based analysis sharpened the distinction between intuitive smoothness and mathematically defined differentiability. Augustin-Louis Cauchy contributed influential formulations of limits, continuity, and derivatives. Later work associated with Karl Weierstrass emphasized precise limiting arguments and exposed the limitations of geometric intuition. Weierstrass presented his continuous nowhere-differentiable example in 1872. Darboux’s 1875 theorem showed, in a different direction, that derivatives retain an intermediate-value property even without continuity.\n\nThe twentieth-century development of functional analysis and the work associated with Maurice Fréchet and René Gâteaux extended differentiation to spaces of functions. This made it possible to treat energies, trajectories, and solutions of differential equations as variables in their own right. Measure-theoretic and distributional approaches further separated pointwise differentiability from the forms of differentiability needed in integration and differential equations. The resulting framework includes several related notions rather than a single universal definition.\n\nComputational practice distinguishes symbolic differentiation, numerical differentiation, and automatic differentiation. Symbolic differentiation transforms mathematical expressions into expressions for their derivatives. Numerical differentiation estimates derivatives from function values, for example through\n\n\\[f'(x)\\approx\\frac{f(x+h)-f(x)}{h}.\\]\n\nFor sufficiently smooth functions, the truncation error of this forward difference is typically \\(O(h)\\). The centered difference\n\n\\[\\frac{f(x+h)-f(x-h)}{2h}\\]\n\ntypically has error \\(O(h^2)\\) under appropriate higher-derivative assumptions. Smaller \\(h\\) does not always improve accuracy because subtraction cancellation and floating-point arithmetic introduce competing errors.\n\nAutomatic differentiation instead applies the chain rule to the elementary operations in a program. Forward mode propagates directional derivatives; reverse mode propagates sensitivities from outputs back to inputs. For a scalar output depending on many inputs, reverse mode can compute the full gradient with a cost comparable, up to implementation-dependent factors, to a small number of evaluations of the original computation. This makes differentiability operational in large-scale scientific computing and machine learning.\n\nA program can nevertheless contain nondifferentiable operations or branches. The rectified linear function \\(r(x)=\\max(0,x)\\), for example, is not differentiable at zero, although software commonly assigns a convention there for computational purposes. Such a convention does not create a classical derivative. Differentiating the operations along an executed branch must also be distinguished from proving differentiability of the full input-output map, especially when branch changes or discontinuities occur at the point under consideration."
    }
  ],
  "links": [
    "mathematical-function",
    "derivative",
    "calculus",
    "mathematical-analysis",
    "real-number",
    "limit",
    "big-o-notation",
    "polynomial",
    "graph-of-a-function",
    "continuous-function",
    "darboux-theorem-analysis",
    "chain-rule",
    "function-composition",
    "mean-value-theorem",
    "lipschitz-continuity",
    "fundamental-theorem-of-calculus",
    "integral",
    "open-set",
    "euclidean-space",
    "linear-map",
    "jacobian-matrix",
    "gradient",
    "partial-derivative",
    "directional-derivative",
    "hessian-matrix",
    "taylor-series",
    "analytic-function",
    "inverse-function-theorem",
    "inverse-function",
    "mathematical-optimization",
    "saddle-point",
    "convex-function",
    "newtons-method",
    "complex-number",
    "cauchy-riemann-equations",
    "complex-analysis",
    "power-series",
    "normed-vector-space",
    "functional-analysis",
    "hilbert-space",
    "manifold",
    "tangent-space",
    "differential-geometry",
    "almost-everywhere",
    "lebesgue-measure",
    "sobolev-space",
    "partial-differential-equation",
    "absolute-continuity",
    "gottfried-wilhelm-leibniz",
    "augustin-louis-cauchy",
    "karl-weierstrass",
    "automatic-differentiation",
    "floating-point-arithmetic",
    "machine-learning"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}