Differentiability is a local property of a function that makes it possible to approximate changes in its values by a linear rule. For a function of one real variable, it means that a finite derivative exists. For functions between higher-dimensional spaces, it normally means that a single linear transformation approximates the function in every direction simultaneously, with an error that becomes negligible compared with the size of the input change.
Differentiability supplies the mathematical basis for tangent lines and tangent planes, rates of change, sensitivity calculations, and many methods of optimization. It is stronger than continuity but weaker than continuous differentiability. Its precise meaning depends on the setting: real, complex, finite-dimensional, infinite-dimensional, or geometric. Extensions involving weak derivatives and almost-everywhere differentiability allow differential methods to remain useful even when classical derivatives fail to exist at some points.
Differentiability in one real variable
Let be defined on an open interval , and let . The function is differentiable at if the limit
exists as a finite real number. The quotient compares the change in output with the change in input. Geometrically, it is the slope of a secant line; its limit is the slope of the tangent line to the graph at . A divergent quotient is sometimes described informally as an infinite derivative, but it does not establish differentiability in the usual real-valued sense.
An equivalent and more generalizable definition is
The notation , associated with asymptotic notation, expresses that the error is of smaller order than the input increment. Thus differentiability is not merely the existence of an approximate slope: it requires the approximation to improve at a specified relative rate as the observation scale shrinks.
Every function differentiable at is continuous there, because both and tend to zero. The converse fails. The function is continuous everywhere, but its difference quotient at zero equals for positive increments and for negative increments. A corner therefore prevents differentiability. A cusp or a vertical tangent can also prevent a finite derivative: for , the quotient at zero is , which grows without bound.
Differentiability at a point does not require continuity of the derivative near that point. For example,
has , whereas
The derivative oscillates without approaching a limit at zero. Nevertheless, derivatives cannot have arbitrary discontinuities: Darboux’s theorem states that every derivative has the intermediate value property, even when it is not continuous.
At an endpoint of a closed interval, a one-sided derivative may be defined by restricting increments to the domain. Such derivatives are useful in boundary problems, but they must be distinguished from the ordinary two-sided derivative. The standard interior definition also presupposes that the function is defined throughout a neighborhood of the point; derivatives relative to more restricted domains require additional conventions.
Total differentiability in several variables
For a function , where is an open subset of , differentiability at means that there is a linear map such that
The map , written , is the total derivative or differential. It is unique: if two linear maps satisfied the condition, their difference, evaluated on increments and divided by , would have to vanish for every vector .
In coordinates, is represented by the Jacobian matrix,
For a scalar-valued function on Euclidean space, the differential is represented, using the usual inner product, by the gradient:
The differential is intrinsically a linear functional; the gradient is its vector representation after an inner product has been chosen.
Existence of all partial derivatives does not imply total differentiability. Partial derivatives test only coordinate directions. Even the existence of every directional derivative, with the values depending linearly on the direction, need not suffice. Define
The estimate
shows continuity at the origin. Every directional derivative there is zero: along a line , the quotient tends to zero. However, along , one has , so
Consequently, the zero linear map does not give the required approximation, and no total derivative exists.
A standard sufficient condition is the existence of all first partial derivatives in a neighborhood of , with those partial derivatives continuous at . The resulting differentiability follows by decomposing an increment into coordinate increments and controlling the errors. The distinction is one of uniformity: directional derivatives examine each fixed direction separately, whereas total differentiability controls all sufficiently small increments together. In finite-dimensional spaces, the definition is independent of the particular norm because all norms on a given finite-dimensional space are equivalent.
Higher differentiability and regularity classes
A differentiable function need not have a differentiable derivative. Repeated differentiability therefore produces a hierarchy. On an open set , the notation denotes continuous functions, functions whose first derivatives exist and are continuous, and functions with continuous derivatives through order . The class consists of smooth functions, meaning functions with continuous derivatives of every finite order. In particular, “differentiable” and “continuously differentiable” are not interchangeable.
Higher derivatives in several variables are naturally multilinear maps. For a twice continuously differentiable function ,
is bilinear in the increment vectors and . For a scalar-valued function, it is represented by the Hessian matrix. Continuity of the second partial derivatives guarantees equality of mixed derivatives and hence symmetry of the Hessian. Without suitable regularity, mixed partial derivatives can exist at a point without being equal.
Higher differentiability controls the accuracy of local polynomial approximations. If is near , then its Taylor expansion through order has the form
where the term is . This formulation generalizes the familiar one-variable Taylor expansion. Stronger assumptions permit more explicit remainder estimates. For example, if a suitable -st derivative is bounded on a neighborhood containing the relevant line segments, the remainder after order is bounded by a constant times .
Smoothness does not imply analyticity over the real numbers. The function
is smooth, and every derivative at zero is zero, but the function is positive away from zero. Its Taylor series at zero therefore represents the zero function rather than on any neighborhood. A real analytic function satisfies the stronger condition of agreeing locally with a convergent power series.
Intermediate regularity is often expressed by Hölder classes. A function belongs locally to , for , if its derivatives of order satisfy estimates of the form
on suitable neighborhoods. The case is Lipschitz regularity. Such distinctions matter in differential equations and approximation theory: two functions can both be , yet have substantially different derivative variation, numerical approximation errors, and stability properties.
Differentiation rules and local structure
The algebraic rules of differential calculus follow from the linear-approximation definition. If and are differentiable scalar functions, then
For composable differentiable maps, the chain rule states
It expresses a basic structural fact: the best linear approximation to a composition is the composition of the best linear approximations. In coordinates this becomes matrix multiplication of Jacobians, with the order determined by the order of the maps.
The mean value theorem connects pointwise differentiability with behavior across intervals. If a real-valued function is continuous on and differentiable on , then some satisfies
Consequently, a derivative bounded in absolute value by yields the estimate . A derivative identically zero on an interval forces the function to be constant there. For vector-valued functions, a single equality with one intermediate point generally fails, although norm inequalities remain available under appropriate hypotheses.
The inverse function theorem gives a particularly strong consequence of continuous differentiability. If is near and is invertible, then has a local inverse near , with
Invertibility of the derivative is a sufficient condition for local smooth invertibility, not a necessary condition for local one-to-one behavior. The function is one-to-one near zero despite having derivative zero there, but its inverse is not differentiable at zero.
The related implicit function theorem explains when an equation determines some variables as differentiable functions of others. If , with continuously differentiable and the derivative with respect to invertible at a solution, then locally , and
This provides a rigorous foundation for differentiating constrained relationships without first solving them explicitly.
Differentiability also constrains extrema. If a differentiable scalar function has an interior local minimum or maximum, its differential is zero. The condition is necessary but not sufficient: has derivative zero at zero without an extremum. Second derivatives distinguish many cases, but a degenerate Hessian can require examination of higher-order terms or the function itself.
Complex differentiability
For a function , where is open, complex differentiability at means that
exists as approaches zero through arbitrary nonzero complex values. Although this resembles the real definition, the denominator may approach zero in every direction in the complex plane. The linear approximation must be multiplication by a single complex number, rather than an arbitrary real-linear transformation of the plane.
Writing and makes the restriction explicit. If is complex differentiable at a point, its real partial derivatives there satisfy the Cauchy–Riemann equations:
When the partial derivatives are continuous in a neighborhood, these equations are also sufficient for complex differentiability there. Mere existence of the partial derivatives and validity of the equations at one point are not sufficient unless additional control, such as real differentiability at that point, is available.
For example, complex conjugation is differentiable as a map , with real derivative represented by
It is nowhere complex differentiable: its difference quotient is along real increments and along imaginary increments. Conversely, is complex differentiable at zero, with derivative zero, but not at any nonzero point. Complex differentiability at an isolated point therefore does not imply differentiability throughout a neighborhood.
A function complex differentiable at every point of an open set is called holomorphic. A central result of complex analysis is that holomorphic functions are automatically analytic: they possess convergent local power-series expansions and derivatives of every order. This is much stronger than real differentiability, where even infinitely many continuous derivatives do not guarantee analyticity.
At a point where the derivative of a holomorphic function is nonzero, its real derivative is a rotation followed by a positive scaling. It therefore preserves oriented angles locally. This links differentiability to conformal geometry. Holomorphic functions also satisfy strong uniqueness principles and integral identities; these arise from differentiability on open sets, rather than from the existence of one derivative at a single point. The contrast between real and complex differentiability is thus not simply notational but reflects a major difference in mathematical rigidity.
Differentiability in normed spaces
The linear-approximation definition extends to normed vector spaces. Let and be such spaces, open, and . The function is Fréchet differentiable at if a bounded linear operator satisfies
Boundedness of the derivative is essential in infinite dimensions, where linearity alone does not guarantee continuity. In finite-dimensional spaces, this definition reduces to ordinary total differentiability.
A weaker notion is the Gâteaux derivative, based on limits along fixed directions:
Terminology varies: some treatments require these directional limits to define a linear map, while others discuss the limits first and impose linearity separately. In either convention, fixed-direction convergence is weaker than the uniform control required for Fréchet differentiability. Fréchet differentiability implies Gâteaux differentiability with the same derivative, but the converse fails without further assumptions.
A basic example in a real Hilbert space is
Expanding the square gives
so . The quadratic remainder establishes Fréchet differentiability directly. At nonzero , the norm itself is differentiable, with
At zero the norm is not differentiable on a nontrivial real normed space: the two-sided directional quotient along a nonzero vector has incompatible signs. In a general normed space, differentiability of the norm away from zero depends on the geometry of its unit ball.
These concepts are central in functional analysis and the calculus of variations, where an input may be an entire function. For example, on the real space , the energy
has derivative
Here the derivative is a continuous linear functional acting on perturbations . Setting the derivative of an energy with forcing equal to zero yields variational equations that may be meaningful even when classical second derivatives do not exist.
Many finite-dimensional results extend to Banach spaces, including suitable inverse and implicit function theorems. Their hypotheses involve bounded linear isomorphisms and completeness. Infinite-dimensional distinctions between operator topologies, boundedness, and uniform approximation make the precise choice of derivative substantially more important than it is in elementary calculus.
Differentiability on manifolds
A manifold is locally modeled on Euclidean space but need not possess a single global coordinate system. Differentiability is defined through coordinate charts. For a map , charts near and near produce the coordinate representation
The map is differentiable at when this representation is differentiable at . Compatibility of the charts and the chain rule ensure that the definition does not depend on which admissible charts are chosen.
The derivative becomes a map between tangent spaces:
Tangent vectors can be represented by velocities of curves. If , then
This identifies the differential as the operation that carries an infinitesimal velocity in the domain to the corresponding velocity in the target. Coordinate Jacobians are matrix representations of this intrinsic map, rather than separate geometric objects.
For a scalar-valued function , the differential belongs to the dual of , the cotangent space. It is therefore naturally a covector. A gradient requires an additional choice of metric: a Riemannian metric determines the unique vector satisfying
for every tangent vector . This distinction explains why a differential can be defined on a smooth manifold without measuring lengths or angles, whereas a gradient cannot.
Differentiability also identifies geometric subsets. The sphere in is the level set of
At a point on the unit sphere,
so its tangent space is
More generally, when a smooth map has surjective derivative at points of a level set, that level set is locally a manifold of dimension .
A bijection that is smooth and has a smooth inverse is a diffeomorphism. Such maps preserve differentiable structure, a stronger requirement than the continuity preserved by a homeomorphism. A topological manifold does not automatically come with a unique smooth structure, and in some settings compatible smooth structures may fail to exist or may be inequivalent. Consequently, differential geometry studies not only shapes and coordinate formulas but also the structures that make differentiation meaningful in the first place.
Almost-everywhere and weak differentiability
Classical differentiability can fail at exceptional points while remaining available on most of a domain. In this setting, “most” is commonly interpreted using Lebesgue measure. A property holds almost everywhere if the set where it fails has measure zero. Rademacher’s theorem states that a locally Lipschitz map between finite-dimensional Euclidean spaces is differentiable almost everywhere. The conclusion concerns total differentiability, not merely the existence of coordinate derivatives.
The hypothesis of Lipschitz continuity gives a uniform local bound
It allows corners, as in , but prevents sufficiently violent local oscillation. In one variable, monotone real-valued functions are also differentiable with a finite derivative almost everywhere. Neither result says that the derivative determines the original function without additional assumptions. The Cantor function, for example, is continuous, nonconstant, and has derivative zero almost everywhere.
The relevant stronger property in one dimension is absolute continuity. An absolutely continuous function on is differentiable almost everywhere, has an integrable derivative, and satisfies
Conversely, indefinite integrals of Lebesgue-integrable functions are absolutely continuous. This is a precise form of the fundamental theorem of calculus. It explains why almost-everywhere differentiability alone is insufficient for recovering a function by integration: singular variation can be invisible to the ordinary derivative almost everywhere.
Weak differentiation uses integration identities instead of pointwise limits. A locally integrable function is the weak partial derivative if
for every smooth test function compactly supported in . The identity is modeled on integration by parts. For , the weak derivative is the sign function, with its value at zero irrelevant. Thus a classical derivative may fail at a corner while a useful weak derivative exists.
Functions with weak derivatives satisfying specified integrability conditions form Sobolev spaces. These are fundamental in the study of partial differential equations, because physically or geometrically natural solutions need not have continuous classical derivatives. Weak derivatives are also distinguished from general distributional derivatives: a jump discontinuity produces a Dirac delta term, which is not represented by an ordinary locally integrable function. Weak differentiability therefore enlarges classical calculus without making every singularity disappear.
Historical development and pathological examples
The origins of differentiability lie in problems of tangents, instantaneous motion, and extrema. Seventeenth-century techniques developed by Pierre de Fermat, Isaac Newton, and Gottfried Wilhelm Leibniz gave increasingly systematic procedures for calculating derivatives. Newton’s fluxions emphasized changing quantities and velocities; Leibniz’s differential notation made relations between infinitesimal changes especially explicit. These methods preceded the modern separation between a computational rule for derivatives and a precise local property of a function.
During the nineteenth century, limits became central to the foundations of analysis. Augustin-Louis Cauchy presented differentiation through limiting difference quotients in his early nineteenth-century work, including his 1823 treatment of infinitesimal calculus. Later developments associated with Karl Weierstrass strengthened the use of explicit approximation and limit arguments. Differentiability increasingly became a condition to be proved, rather than a geometric regularity assumed from a curve’s appearance.
Continuous nowhere-differentiable functions were especially important in this change. Bernard Bolzano constructed an early example in work that remained unpublished during his lifetime. Weierstrass presented his celebrated example in 1872. A classical Weierstrass function has the form
with , an odd positive integer, and suitable additional restrictions. The classical sufficient condition
ensures nowhere differentiability. The series converges uniformly because its terms are bounded by the summable geometric sequence , so its sum is continuous.
The example separates two kinds of control. Small amplitudes can guarantee continuity of a convergent series, while rapidly increasing frequencies destroy the existence of limiting slopes. It also shows why uniform convergence of differentiable functions does not, by itself, preserve differentiability. To pass derivatives through a limit, further hypotheses are needed, typically controlling the convergence of the derivatives as well as the values at at least one point.
Twentieth-century work connected these phenomena to measure theory, functional analysis, and probability. Henri Lebesgue’s integration theory clarified the distinction between pointwise behavior and almost-everywhere behavior. Maurice Fréchet and René Gâteaux developed derivative concepts suited to spaces of functions. Hans Rademacher proved his almost-everywhere differentiability theorem in 1919. In probability, paths of standard Brownian motion are almost surely continuous and nowhere differentiable. Lack of ordinary derivatives thus appears not only in specially constructed counterexamples but also in central models of random motion.
Computation, optimization, and differential models
Differentiability makes local calculations possible, but it does not guarantee that a derivative is easy to compute or numerically stable. A finite-difference approximation,
contains both approximation error and floating-point error. If has a bounded second derivative near , the forward-difference truncation error is . A central difference has error under suitable third-derivative assumptions. Making extremely small can worsen the result because subtraction of nearly equal values amplifies rounding effects.
Automatic differentiation uses a different principle: it applies the chain rule to the elementary operations of a program. Forward mode propagates directional derivatives, while reverse mode propagates sensitivities from outputs toward inputs. For a scalar output depending on many inputs, reverse mode can compute a full gradient at a cost typically comparable to a small multiple of evaluating the program. It computes derivatives of the operations actually executed; it does not by itself prove that an underlying mathematical function is differentiable at a branch boundary or other singular point.
In optimization, a differentiable objective supplies first-order information for descent methods. Newton’s method uses second-order information: for solving , its basic iteration is
when the derivative is invertible. Local quadratic convergence follows under suitable regularity, such as a locally Lipschitz derivative near a solution with invertible Jacobian. Mere differentiability is not enough to guarantee that behavior. Similarly, an invertible but nearly singular derivative can produce large sensitivity, connecting differentiation to conditioning rather than guaranteeing numerical reliability.
Nondifferentiable models often retain useful structure. For a convex function, a subgradient at satisfies
throughout the domain. For , every is a subgradient at zero. Such objects replace a unique tangent slope with a family of supporting slopes. This is relevant to absolute-value penalties, piecewise-linear models, and optimization problems whose solutions may occur precisely at corners.
Differential equations and statistical models also depend on the distinction between formal derivatives and justified differentiability. A classical solution of a differential equation must possess the derivatives appearing in the equation, whereas a weak solution satisfies an integrated formulation. In statistical estimation, differentiating an integral or expectation requires hypotheses permitting interchange of differentiation and integration. Sensitivity formulas, gradient calculations, and parameter-dependent solutions therefore rest not only on symbolic rules but on regularity conditions that ensure those rules describe genuine local behavior.