In mathematics, a function is a rule that assigns to each element of one set, called the domain, exactly one element of another set, called the codomain. Formally, it can be defined as a binary relation in which no input has two outputs, which places it within set theory and the Cartesian product. This precise notion is the basis of most of modern mathematics.
Functions matter because they are the common language of calculus, analysis, linear algebra, and probability, where a random variable is itself a function. They also describe physical states such as the wave function, economic relationships such as demand and supply, and the layered computations of an artificial neural network.
This entry traces the historical development of the concept, from early tables and curves to the modern set-theoretic definition. It then covers notation and the basic classes of functions, including injective, surjective, and bijective maps. It proceeds through elementary families, limits, differentiation and integration, and functions of several variables. The later sections address series, complex-valued functions, function spaces, special functions, foundations, and applications in probability, the sciences, computation, and machine learning, and it closes with generalized notions of function.
Formal definition and basic terminology
In modern mathematics a function is a special kind of binary relation. Given two sets X and Y, a function f from X to Y assigns to each element of X exactly one element of Y. Formally, the function is identified with its graph, a subset G of the Cartesian product X × Y such that for every x in X there is one and only one y in Y with . The existence part makes the relation total, since every input receives some output. The uniqueness part makes it single-valued, since no input receives two outputs. A relation that violates the existence condition is a partial function, and one that violates uniqueness is a multivalued relation. Because the graph alone does not record the target set, the formal definition is often given as a triple (X, Y, G), so that the codomain is part of the data.
Functions are written to indicate that f maps X into Y. For an input x, the element of Y assigned to x is written , read "f of x". The distinction between the two is essential: f is the function, a single object, whereas is a value, an element of Y. Many applied texts blur this by writing "the function f(x)", but the careful convention is to reserve the symbol f for the whole map. The maps-to arrow specifies the rule for a generic input without naming the function. For example, describes the squaring rule, read "x maps to x squared". The words "map" and "mapping" are commonly used as synonyms for "function".
For a function , X is the domain and Y is the codomain. The image of f, sometimes called its range, is the set , a subset of Y. The term "range" is ambiguous, since some authors use it to mean the codomain, so the image is the safer term. The preimage of a subset B of Y is . This set is defined for every function, whether or not f has an inverse. The preimage of a single point is called a fiber of f.
The codomain need not equal the image. The function given by has codomain ℝ, but its image is the set of nonnegative reals, because a square is never negative. Similarly, has image [−1, 1]. The codomain is a chosen target set, while the image is determined entirely by the rule. The condition that the image fills the whole codomain is surjectivity, which is treated in a separate section.
Two functions and are equal exactly when X = X′, Y = Y′, and for every x in X. Both the domain and the codomain must match, not only the rule. The functions from ℝ to ℝ and from ℝ to [0, ∞) agree on every input, yet they are distinct functions because their codomains differ. Under a bare set-of-pairs definition they would have identical graphs, which is why the codomain is included in the definition. Likewise, on ℝ with 0 removed and on ℝ agree wherever both are defined, but they have different domains and so are not equal. Equality of functions therefore depends on the domain, the codomain, and the rule together, and changing any one of these produces a different function.
Historical development
The practice of pairing inputs with outputs is much older than the word "function." Old Babylonian scribes compiled tables of squares, cubes and reciprocals of natural numbers, and these tables can be read as maps from natural numbers to natural numbers. Historians today see that the Babylonians were dealing with functions, but they would not have framed the work in those terms. The scribes were recording results, not studying a rule that links an input to an output, so the concept itself is usually dated later.
Medieval scholars also described quantities that vary along a line or a curve in graphical terms, which prepared the ground for later work. The decisive step came in the seventeenth century, when Descartes published La Géométrie in 1637. He showed that any point in a plane could be identified by two numbers: its horizontal distance from a fixed reference point and its vertical distance from the same point. Once points had numerical coordinates, any algebraic equation in two variables could now be seen as a curve on the plane. Descartes also sorted curves by the degree of their equations. His curves were algebraic, and he rejected the transcendental curves that later analysis took for granted, so his framework covered only part of what would later count as functions.
The word itself was introduced by Leibniz. In a 1673 manuscript he used "function" for a quantity that varies from point to point along a curve, such as the length of a tangent or a normal segment. He also introduced the terms "constant," "variable" and "parameter." Johann Bernoulli, in his correspondence with Leibniz between 1694 and 1698, adopted "function" for expressions built from a single variable, and later Johann Bernoulli applied the term to any quantity formed in an algebraic or transcendental manner. In the eighteenth century, the emphasis shifted to analytic expressions. Euler worked within this view in his 1748 Introductio in analysin infinitorum, where algebraic operations could be repeated infinitely to produce series and products. Euler introduced the notation f(x) in 1734 and gave a broader definition in 1755, in which one quantity is a function of others when it changes whenever they change. Euler's own practice blurred the line between a function and the formula representing it. Trigonometric quantities such as sine were regarded as line segments tied to a circle until Euler treated them as functions. Euler's definition still included only analytic functions, a restricted subset of the continuous functions.
The decisive break came with Dirichlet. His study of Fourier series in 1829 required him to reason about functions without assuming an analytic formula, and in 1837 he defined a function as an arbitrary correspondence between two sets of numbers, such that each value of the independent variable determines exactly one value of the dependent variable. This separated the concept of function from any particular representation. Later critics objected that Dirichlet's wording allowed the correspondence to be given in any way at all, which pointed toward a more abstract notion. In the late nineteenth and early twentieth centuries, as set theory was developed, the function came to be identified with a set of ordered pairs drawn from the Cartesian product of its domain and codomain. This formulation, which removes any reference to rules or formulas, became the standard definition in twentieth-century mathematics.
Ways of specifying functions
The most common way to specify a function is an explicit formula that computes the output from the input, such as or . A formula does not fix a function until a domain is stated: defines a real-valued function only on . When one rule cannot cover all inputs, a function may be defined piecewise, with a different formula on each part of the domain. The absolute value is the standard example:
The floor function , which returns the greatest integer not exceeding , is piecewise with infinitely many pieces, so that but . For a piecewise definition to be well formed, the pieces must not overlap in conflicting ways.
Some functions are known only through finitely many values, as in a table of measurements, a printed table of logarithms, or experimental data. A table specifies a function on its finite set of inputs and says nothing about other points. To obtain values in between, one uses interpolation, which extends the data to a function on an interval. Linear interpolation joins consecutive points by line segments, so that for ,
Polynomial interpolation seeks a polynomial of degree at most passing through points with distinct abscissae; such a polynomial is unique and can be written explicitly by Lagrange's formula. Spline interpolation instead uses low-degree polynomial pieces joined smoothly. Different methods applied to the same table generally produce different functions, which is why the table alone does not determine a unique rule. The study of such approximations belongs to approximation theory.
Functions can also be specified indirectly. An equation may relate input and output without solving for either: does not define as a single-valued function of , because each in corresponds to two values of . Restricting to yields the function . Inverse functions are another indirect method. A function has an inverse only when it is injective on its domain, so the sine function is first restricted to , and its inverse is the arcsine. Likewise, the natural logarithm is defined as the inverse of the exponential function: is the unique with . The Lambert W function, defined by the relation , shows the same issue in a sharper form. For some arguments this relation has several solutions, so a function is obtained only after choosing a branch.
Functions on the natural numbers are often defined recursively, by giving initial values and a rule that expresses each value in terms of earlier ones. The factorial is given by and , and the Fibonacci numbers by , , and . Such definitions are sequences specified by a recurrence relation. The recursion theorem guarantees that the initial values and the rule determine exactly one function. Linear recurrences with constant coefficients often have closed forms; for the Fibonacci numbers, Binet's formula gives with . A function can also be specified by a procedure. The Euclidean algorithm computes the greatest common divisor through the rule , with . A procedure defines a total function only if it halts on every input in the domain. Which procedures can be carried out is formalized in computability theory, using the Turing machine as a model, and the halting problem shows that some well-defined functions have no such procedure.
Injective, surjective, and bijective functions
Three properties describe how a function distributes its values over the codomain. They determine when a function can be reversed and when two sets can be said to have the same size.
Definitions and examples
A function is injective (one-to-one) when distinct inputs always give distinct outputs: forces . It is surjective (onto) when every element of the codomain is a value of , so that the image equals the codomain. A function that is both is bijective, and it pairs the elements of and so that each element on either side has exactly one partner.
The examples show how the properties depend on the rule and on the chosen domain and codomain together. The map from to is bijective. The squaring map on is neither: breaks injectivity, and no real number maps to , which breaks surjectivity. Restricting the domain to makes it injective, and restricting the codomain to makes it surjective. The exponential function from to is injective but not surjective, since it takes only positive values; from to it is bijective. The floor function from to is surjective but not injective, because each interval is sent to the single integer . The successor map on the natural numbers is injective but misses .
For finite sets of equal size, the three properties coincide: a map between them is injective exactly when it is surjective, a consequence of the pigeonhole principle. The successor map shows that this fails for infinite sets.
Inverse functions and one-sided inverses
A bijection has an inverse , where is the unique with . The two composites satisfy and . Conversely, a two-sided inverse can exist only for a bijection. If a map satisfies both identities, then is injective, because gives , and surjective, because . Thus a two-sided inverse exists exactly when is bijective.
Weaker inverses separate the two properties. A left inverse of is a map with . Such a map exists exactly when is injective, provided is nonempty: one sets when and assigns an arbitrary fixed value otherwise. A right inverse is a map with , and it exists exactly when is surjective, since one may choose a preimage for each . Injectivity is thus equivalent to the existence of a left inverse, and surjectivity to the existence of a right inverse.
Comparing sizes of sets
Injections and surjections allow sets to be compared without counting their elements. One writes when there is an injection , and when there is a bijection. The Cantor–Schröder–Bernstein theorem states that injections in both directions yield a bijection, so the comparison is antisymmetric; this proof does not require the axiom of choice. A set is countable when it admits an injection into , which is equivalent to being finite or in bijection with . The integers are countable by a bijection with , and the rationals are countable because a surjection from onto them exists. The real numbers are not countable: Cantor's diagonal argument shows that no map from onto is surjective, so is strictly larger than . The notion of cardinality formalizes these comparisons.
A surjection also suggests , since one would like to choose an element of each fiber to form an injection . For well-orderable this choice can be made without extra assumptions, but for arbitrary sets it cannot.
The role of the axiom of choice
Choosing a right inverse for a surjection means selecting one element from each fiber , and the family of all these fibers may be infinite and unstructured. In ZF set theory, the axiom of choice is exactly what guarantees such simultaneous selections. The statement that every surjection has a right inverse is equivalent to the axiom of choice, and so is the statement that a surjection always yields an injection . Without the axiom, some surjections may admit no right inverse, and the comparison of sizes via surjections becomes unreliable. Consequently, many standard results in set theory and analysis distinguish between statements that use only the existence of injections and those that require selecting preimages.
Composition, restriction, and algebra of functions
If and , their composition is defined by . The output of becomes the input of , so the rule applies to every input only when the image of lies inside the domain of . Otherwise the composite is defined only on those for which falls in that domain. For example, and give , which is defined exactly where .
Composition is associative: , because both sides send to . Parentheses can therefore be omitted in iterated composites such as . Iterates are defined by and , a notation that is central to the study of dynamical systems. The identity function acts as a neutral element: and .
Composition is not commutative. Take and . Then , while . Even when both composites are defined, they generally differ, and they can coincide only in special cases. Composition also fails to distribute over addition from the left. With , , and , one gets , which is not equal to . Composition on the right does distribute over pointwise operations, so . Composition also respects the properties studied below: a composite of injective maps is injective, and a composite of surjective maps is surjective. Among bijections , composition is associative, has an identity, and gives every element an inverse, which makes these bijections a group, the symmetric group on .
Functions with a common domain can be combined pointwise. If and map a set into , one defines , , , and for a constant . The quotient is defined wherever . When the domains differ, the natural domain of the combined function is the intersection of the two domains. Because these operations are computed value by value, they inherit the algebraic laws of the codomain. The set of all functions from a fixed set into therefore forms a commutative ring under pointwise operations, and a vector space when only addition and scalar multiplication are used. The zero function and the constant function serve as the additive and multiplicative identities. This vector-space structure is the starting point for the function spaces treated later in the article.
A function is a restriction of to a subset of the domain when for all . It is written . Restriction discards information. An injective function remains injective after restriction, but surjectivity can be lost, because values attained only outside disappear from the image. The converse operation is extension: extends when . Extensions are generally not unique. The function on extends to as , as , or as any function that agrees with on .
Restriction leads naturally to partial functions, which are defined on a subset of their nominal domain. Examples include , defined on , and restricted to when only real values are wanted. Composition of partial functions is defined on the set of points where every step is defined, namely . Partial functions allow an expression to be meaningful on part of its natural domain without forcing a value everywhere. The formal treatment of such generalizations is taken up in a later section.
Basic properties of real-valued functions
Many properties of a function , where is a subset of the real line, can be read from its values and its graph without computing anything in detail. The function is increasing on when forces , and strictly increasing when the inequality is strict. Decreasing and strictly decreasing are defined in the same way with the inequality reversed. A function that is increasing or decreasing in one of these senses is called monotonic. On an interval, a derivative that is nonnegative everywhere gives a nondecreasing function, a consequence usually proved with the mean-value-theorem. Monotonicity also has structural consequences: a monotone function on an interval has one-sided limits at every interior point and can have discontinuities only of the jump type, and there are at most countably many of them.
Boundedness concerns the range of values rather than their order. The function is bounded above if some constant satisfies for all in the domain, bounded below if a similar lower constant exists, and bounded if both hold. Equivalently, the function is bounded when its values lie in a finite interval. The sine and cosine functions are bounded by 1 in absolute value, while is bounded below but not above on the real line. The supremum and infimum of the set of values serve as the least upper and greatest lower bounds whenever they exist, and they give the sharpest constants in these estimates.
Even and odd symmetry describe how a function behaves under reflection in the origin. A function is even when for every in its domain, so its graph is symmetric about the y-axis; it is odd when , so its graph is symmetric under a half-turn about the origin.Even and odd functions are defined based on their symmetry properties; a function f is odd if f(−x) = −f(x) for all x in its domain, while a function f is even if f(−x) = f(x) for all x in its domain. Examples include and (even), and , and (odd). Many functions are neither: has no such symmetry, and it can be split into an even part and an odd part . This decomposition is and , which sum to . The parity rules are multiplicative: the product of two even functions, or of two odd functions, is even, and the product of an even and an odd function is odd. Parity also constrains expansions. The Maclaurin series of an even function contains only even powers, and the fourier-series of a periodic odd function contains only sine terms.
Periodicity means that a function repeats its values at a fixed shift. A function is periodic with period when for all . The smallest positive period, when it exists, is called the fundamental period. The sine and cosine functions have fundamental period , while has period . A sum of two periodic functions is periodic only when their periods are commensurable, meaning their ratio is rational; is not periodic, for instance. The indicator function of the rationals is periodic with every rational number as a period, so it has no fundamental period at all, which shows that the notion can behave unexpectedly.
Graph transformations show how formulas change the picture. For a constant , the graph of is the graph of moved up by , and moves it right by , so the sign inside the argument runs against intuition. Multiplying the output by stretches the graph vertically by , and a negative also reflects it across the x-axis. Replacing by compresses or stretches the graph horizontally by a factor of . The map reflects across the y-axis, which is why an even function is unchanged by it. For example, has period , and is the parabola shifted two units right and three units up.
Convexity is a shape condition. A function on an interval is convex when, for any two points in the interval and any , ; the chord between two points on the graph never falls below the graph. Concavity reverses the inequality. For twice differentiable functions, convexity is equivalent to a nonnegative second derivative, which gives the standard test for , and as convex functions. Convex functions are closed under sums, multiplication by nonnegative constants and pointwise maxima. Convexity is the hypothesis behind Jensen's inequality, which has its origins in work by Otto Hölder in 1889 and was proved in its general form by Johan Jensen in 1906.
The finite form of Jensen's inequality states that for a convex , positive weights summing to 1 and points in the domain, . In probability, for a random variable with a convex function , the inequality takes the form , and the difference between the two sides is called the Jensen gap. Equality in the strictly convex case holds only when all the coincide, or when is almost surely constant. For concave functions the inequality is reversed. A standard application uses the concavity of : applying the inequality with equal weights to and yields for positive , which is the arithmetic–geometric mean inequality. Jensen's inequality is widely used in statistics, information theory and economics, where it expresses the effect of curvature on averages.
Elementary functions
Elementary functions are the functions that appear most often in algebra, geometry, and calculus. They are built from a small set of basic types by arithmetic operations, composition, and taking inverses. Their domains and behavior are well understood, and they serve as the reference cases for the general theory of functions.
Polynomial functions have the form with . The degree governs end behavior, because for large the leading term dominates. A nonzero polynomial of degree has at most real roots, counted with multiplicity. Over the complex numbers it has exactly roots, which is the content of the fundamental theorem of algebra. Polynomials are continuous everywhere and differentiable everywhere.
A rational function is a quotient of two polynomials, defined wherever . Its roots are the zeros of that are not cancelled by a zero of . A pole occurs at a zero of where does not vanish, and there grows without bound. Where a common factor cancels, the discontinuity is removable: equals except at . Asymptotic behavior follows from comparing degrees. If , the line is a horizontal asymptote. If the degrees are equal, the horizontal asymptote is , where and are the leading coefficients of and . If , the graph has a slant asymptote. The function illustrates the first case, with a pole at 0 and horizontal asymptote .
For with , the exponential function is strictly monotone with range . Its inverse is the logarithm , defined on positive reals. The base is the value of , and it is distinguished by the fact that is its own derivative. The natural logarithm can be written as . The change-of-base formula and the identity reduce every exponential and logarithm to the base . This is why calculus texts almost always work with and .
The circular functions , , and are defined on the unit circle and are periodic with period . Their basic identities include and the addition formula . Euler's formula connects them to the exponential, a link developed further in the complex-valued section. The hyperbolic functions are defined by , , and . They satisfy , which parallels the circular identity with a hyperbola in place of a circle. The graph of is the curve formed by a freely hanging cable, called a catenary. The inverse shows how these functions reduce to logarithms.
Elementary functions are also classified as algebraic or transcendental. A function is algebraic if there is a nonzero polynomial with identically. Polynomials, rational functions, and roots such as or are algebraic, and so is any finite combination of them under the four arithmetic operations and root extraction. A function that is not algebraic is transcendental. The exponential , the logarithm , and the trigonometric functions and are transcendental. The same distinction applies to numbers: Lindemann proved in 1882 that is transcendental, and is also transcendental. Transcendental functions cannot be defined by finitely many algebraic operations, which is one reason they are typically introduced through limits, integrals, or power series.
Limits and continuity
The limit of a function describes the value that approaches as approaches a point , and it does not require to be defined at . Formally, means that for every there is a such that implies . The condition excludes itself, so the expression has the limit 1 at even though it is undefined there. Restricting to one side of gives one-sided limits, written and . A two-sided limit exists exactly when both one-sided limits exist and agree. The Heaviside step function, which is 0 for and 1 for , has left limit 0 and right limit 1 at 0, so no two-sided limit exists there.
A function is continuous at a point of its domain when , and continuous on a set when it is continuous at every point of that set. Sums, products, and quotients (with nonzero denominator) of continuous functions are continuous. Continuity is also preserved under composition: if is continuous at and is continuous at , then is continuous at . This is why polynomials, rational functions, and the exponential and trigonometric functions are continuous wherever they are defined.
Discontinuities are classified by the behaviour of one-sided limits. At a removable discontinuity the two-sided limit exists but either is undefined or differs from it, so redefining the value at removes the defect. At a jump discontinuity both one-sided limits exist and are finite but differ, as with the floor function at integers. At a discontinuity of the second kind at least one one-sided limit fails to exist or is infinite. Examples include the pole of at 0 and the oscillation of at 0. The Dirichlet function, equal to 1 on the rationals and 0 elsewhere, is discontinuous at every point.
For a function continuous on a closed bounded interval , two global theorems hold. The intermediate value theorem states that if lies strictly between and , then for some in . Its proof depends on the completeness of the real numbers, and it has a standard consequence: every real polynomial of odd degree has a real root. The extreme value theorem states that such a function attains a maximum and a minimum on . Both hypotheses are needed. The function on is unbounded, and on the open interval attains neither a maximum nor a minimum. Both theorems are instances of more general results about connected and compact sets.
Continuity is a pointwise property, since the in the definition may depend on the point. Uniform continuity requires that for every there be one such that implies for all in the set. The function is continuous on but not uniformly continuous there, because equal steps produce ever larger changes as grows. The Heine–Cantor theorem states that a function continuous on a compact set is uniformly continuous. The function on shows that the compactness hypothesis cannot simply be dropped.
A stronger condition is Lipschitz continuity: there is a constant with for all in the domain. Geometrically, every secant slope of the graph is bounded by in absolute value. A Lipschitz function is uniformly continuous, since works. The converse fails. The function is uniformly continuous on , but its secant slopes near 0 are unbounded, so it is not Lipschitz. The function is Lipschitz on every bounded interval but not on . If a function is differentiable with , the mean value theorem gives a Lipschitz constant . When , a Lipschitz map is a contraction, which is the setting of the Banach fixed-point theorem.
Differentiation and integration
The derivative as best linear approximation
For a real-valued function defined near a point , the difference quotient gives the average rate of change over an interval of length . The derivative is the limit of this quotient as tends to zero, when that limit exists. The same fact can be stated as an approximation: is differentiable at exactly when some number satisfies
The tangent line is therefore the unique straight line that matches to first order at . This view extends to several variables, where the derivative becomes a linear map represented by the Jacobian matrix.
Notation reflects different traditions. Leibniz wrote and , which treats differentiation as an operator and makes the chain rule look like cancellation of differentials. Lagrange introduced the prime , Newton wrote a dot over a variable for time derivatives in mechanics, and Euler used the operator .
Differentiation rules and the mean value theorem
The basic rules follow from the limit definition. The sum rule gives , and the product rule gives . Where , the quotient rule gives . The chain rule states that . Its proof reflects the approximation viewpoint: composing two linear approximations yields a linear approximation whose slope is the product of the slopes.
The mean value theorem states that if is continuous on and differentiable on , then there is a point in with
Its special case in which is Rolle's theorem. Several consequences follow: a function whose derivative vanishes on an interval is constant there, a nonnegative derivative forces the function to be nondecreasing, and a derivative bounded by a constant makes the function Lipschitz continuous with constant .
Definite integral and Riemann sums
The definite integral is intended to measure the signed area between the graph of and the horizontal axis. Area above the axis counts positively and area below it counts negatively. The rigorous construction, formulated by Bernhard Riemann in the 1850s, proceeds as follows. One partitions into points , chooses a tag in each subinterval , and forms the Riemann sum
If these sums approach one value as the largest subinterval length tends to zero, regardless of the choice of tags, that value is the Riemann integral. Every continuous function on a closed bounded interval is integrable in this sense. Lebesgue's criterion gives the exact characterization: a bounded function is Riemann integrable if and only if its set of discontinuities has Lebesgue measure zero.
Fundamental theorem and the Lebesgue extension
The fundamental theorem of calculus has two parts. First, if is continuous and , then is differentiable with . Second, if is any antiderivative of on , then . Together these show that integration and differentiation are inverse operations up to an additive constant, and the second part turns the computation of integrals into the search for antiderivatives. Newton and Leibniz developed this link independently in the late seventeenth century.
Riemann integration has limits. The Dirichlet function, equal to 1 on rational numbers and 0 elsewhere, has upper Riemann sums equal to 1 and lower sums equal to 0 on every interval, so it is not Riemann integrable. The Lebesgue integral instead partitions the range of the function and weights each level set by its measure, which is defined in measure theory and uses a σ-algebra of measurable sets. Under this definition the Dirichlet function is integrable with integral 0, because the rationals have measure zero. The Lebesgue framework also makes limits and integrals interchangeable under the monotone and dominated convergence theorems.
The fundamental theorem survives in a weaker form for Lebesgue integrals. If is Lebesgue integrable, then is differentiable almost everywhere with almost everywhere. Almost-everywhere differentiability is not enough to recover from its derivative, however. The Cantor function is nondecreasing and has derivative zero almost everywhere, yet it rises from 0 to 1. The correct condition is absolute continuity, under which equals the integral of .
Functions of several variables and vector-valued functions
For a function of several inputs, , the derivative is taken one coordinate at a time. The partial derivative of with respect to at a point is where is the -th standard basis vector and all other variables are held fixed. For , this gives and . The curly-d symbol is usually credited to Legendre. Collecting the partials into a vector gives the gradient, . The gradient points in the direction of steepest increase of , and its length is the rate of that increase. For a unit vector , the directional derivative is . If is differentiable at , this equals .
Existence of partial derivatives does not imply differentiability. The function with has both partials equal to zero at the origin. Yet along the diagonal it takes the constant value , so it is discontinuous there and not differentiable. Its directional derivatives in directions other than the coordinate axes do not exist.
Jacobian matrix and differentiable maps
A map is differentiable at if there is a linear map with . This is unique and is represented by the Jacobian matrix , whose entries are . A convenient sufficient condition is that all partial derivatives exist and are continuous near . The chain rule then becomes a matrix product, . The determinant of the Jacobian measures local volume change. For polar coordinates , the determinant equals , which accounts for the factor in . The matrix of second partial derivatives of a scalar function is the Hessian matrix, which is used to classify critical points.
Level sets, the implicit function theorem, and Lagrange multipliers
A level set is a set of the form . Where , the gradient is perpendicular to the level set. The implicit function theorem states that if , is continuously differentiable, and is invertible at a solution point, then near that point the solution set is the graph of a continuously differentiable function . For , at we have , so the unit circle is locally . At the derivative vanishes and no such graph exists, because the circle has a vertical tangent there.
In constrained optimization, suppose is a local extremum of on the set , and . Then for some scalar , the Lagrange multiplier. Equivalently, is a stationary point of the Lagrangian . To maximize subject to , solve together with . This gives and a maximum value of 25. Methods of this kind are central to optimization.
Vector-valued functions, curves, and integrals
A vector-valued function , written , traces a parametrized curve. Its velocity is , and its arc length over is . For the helix , the speed is constantly , so the length over is .
A scalar line integral along is . For a vector field , the line integral computes work done by a force along the path. If , the value depends only on the endpoints and equals . A surface integral over a surface parametrized by is . For the unit sphere, , and integrating over and gives the area . The flux of a vector field through a surface is . Green's, Stokes', and the divergence theorem relate these integrals to integrals over the boundary of the region.
Series and sequences of functions
Power series
A power series centred at has the form . For each such series there is a radius of convergence , given by the Cauchy–Hadamard formula . The series converges absolutely when and diverges when ; the behaviour on the boundary must be checked separately. The geometric series has and sums to on , while the exponential series has . Inside its disc of convergence, a power series defines a function that is infinitely differentiable, and it may be differentiated and integrated term by term. This is why power series are a principal means of representing analytic functions. Over the complex numbers the radius is determined by the nearest point where the sum fails to extend, which links the theory to complex-analysis.
Taylor and Fourier expansions
If is infinitely differentiable at , its Taylor series is . Taylor's theorem with the Lagrange remainder shows that the th Taylor polynomial approximates with an error controlled by the th derivative, but this does not imply that the series converges to . The standard counterexample is for with : every derivative vanishes at 0, so the Taylor series at 0 is identically zero, although is not. Taylor series therefore describe a function locally only when a remainder estimate holds.
Fourier series expand periodic functions in trigonometric terms. For a function of period , one writes , with coefficients and . Pointwise convergence holds under mild conditions such as piecewise smoothness; at a jump discontinuity the series converges to the average of the one-sided limits, and partial sums overshoot near the jump, a phenomenon known as the Gibbs effect. Fejér showed that the Cesàro means of the partial sums converge uniformly for every continuous periodic function, which removes these difficulties for averaged sums. For square-integrable functions, the coefficients satisfy Parseval's identity, and the series converges in the sense. Convergence almost everywhere for functions was proved by Carleson in 1966, whereas Kolmogorov constructed in 1923 an function whose Fourier series diverges everywhere. The expansions connect to the fourier-transform as a continuous analogue.
Pointwise versus uniform convergence
A sequence converges pointwise to on a set if for each , with the rate allowed to depend on . It converges uniformly if , so that one index serves for all points at once. The difference matters. The functions on converge pointwise to a function that equals 0 on and 1 at ; each is continuous, but the limit is not, so the convergence cannot be uniform. Uniform convergence preserves continuity: a uniform limit of continuous functions is continuous. It also justifies exchanging limits with operations. On a compact interval, uniform convergence of to gives , and if the are continuously differentiable and converges uniformly, then is differentiable with . Pointwise convergence alone permits neither conclusion. For integrals over unbounded domains or for more general functions, the lebesgue-integral offers the dominated convergence theorem, which replaces uniform control with an integrable majorant. Power series converge uniformly on every closed disc strictly inside their radius of convergence, and the Weierstrass M-test provides a convenient criterion for uniform convergence of series. The concepts are developed in mathematical-analysis and in the theory of uniform-convergence and pointwise-convergence.
The Weierstrass approximation theorem
Weierstrass proved in 1885 that every continuous real function on a closed interval is the uniform limit of polynomials: for every there is a polynomial with . Bernstein later gave a constructive proof using the polynomials , which converge uniformly to on although they converge slowly and need not approximate derivatives. Stone's 1937 generalization extends the result to subalgebras of continuous functions on compact Hausdorff spaces that contain constants and separate points. A periodic version holds for trigonometric polynomials, a consequence of Fejér's theorem. As a result, polynomials are dense in the space of continuous functions with the supremum norm, and also in each space for finite , which makes polynomials and trigonometric polynomials basic tools in approximation-theory and in the theory of Banach spaces of functions.
Complex-valued functions
Holomorphic functions and the Cauchy–Riemann equations
A function of a complex variable can be written as , where and and are real-valued. Complex differentiability at a point requires the difference quotient to approach one limit as the complex increment tends to zero, regardless of direction. Functions with this property on an open set are called holomorphic, and the term is used to distinguish them from real-analytic functions. In complex analysis, the term holomorphic function is often used to distinguish it from a real analytic function.
Differentiability forces a pair of first-order partial differential equations on and :
These are the Cauchy–Riemann equations. Under the assumption that all four first-order partial derivatives are continuous, they are equivalent to the complex function being holomorphic. Their history runs through several figures: the system first appeared in the work of Jean le Rond d'Alembert in 1752. Cauchy in 1814 then used these equations to construct his theory of functions. Riemann's dissertation on the theory followed in 1851.
The equations have strong consequences. The real and imaginary parts of a holomorphic function are harmonic functions, so they satisfy the Laplace equation. Holomorphy is also far more rigid than it looks: holomorphic functions are , and as a consequence of the Cauchy–Riemann equations, a real-valued holomorphic function must be constant.
The complex exponential and Euler's formula
The exponential extends to complex arguments by . Setting gives Euler's formula, , which links the exponential function to trigonometric functions. Because and , trigonometric identities can be derived from the algebra of exponentials. The function is holomorphic on the entire plane and is periodic with period , which explains the periodicity of the real trigonometric functions when they are restricted to real arguments. The same identities also carry over to the hyperbolic functions, since and .
Analytic continuation and the Riemann zeta function
The Riemann zeta function is first defined by the series for . That series diverges elsewhere, so the function must be extended to be defined on the rest of the plane. Riemann's 1859 theorem shows that the function extends to a meromorphic function on the whole complex plane, regular except for simple poles at and , and satisfies . Here , where the factor is the gamma function. The factor removes the poles, but for most uses it turns out to be more convenient not to use this factor and tolerate the poles at s = 0, 1.
From this, also extends to a meromorphic function on C, which is regular except for a simple pole at s = 1, and this analytic continuation of ζ has simple zeros at the negative even integers. These zeros are called the trivial zeros. The continuation gives values that the series cannot: ζ(−1) = −1/12. Because a holomorphic continuation of this kind is unique wherever it exists, the value is forced once the functional equation is accepted.
Contour integration and real integrals
Contour integration evaluates integrals along paths in the complex plane, and it is also a standard way to compute real integrals. The trick is to convert the definite integral into a contour integral, and then solve the contour integral using the residue theorem. The residue theorem states that if a function f(z) is analytic inside and on a simple closed contour C, except for a finite number of isolated singularities inside C, then the integral of f(z) around C is 2πi times the sum of the residues of f at those singularities.
For a rational integrand on the real line, one usually chooses a closed contour made of the real segment and a large semicircle in the upper half-plane. The arc contribution vanishes in the limit, and since the real part or imaginary part of the original function f(x) will be left, its integration can be obtained. A rational integrand must satisfy conditions: the degree of the denominator must be at least 2 higher than the degree of the numerator, and the rational function must have no poles on the real axis. When poles lie on the real axis, the contour is indented around them and the result is a Cauchy principal value.
As a simple case, the integral has a single enclosed pole at , with residue . The residue theorem gives . Oscillatory integrals such as can be handled the same way by a semicircular contour.
Function spaces and functional analysis
Vector spaces of functions
Functions with a common domain and values in the real or complex numbers can be added and scaled pointwise, and the collection of all such functions is a vector space. Its subsets that are closed under these operations are the working objects of analysis. The continuous functions on a closed interval , written , form such a subspace, and the polynomials of degree at most form a finite-dimensional subspace of it. All polynomials together form an infinite-dimensional subspace, but their span is not all of : the algebraic dimension of is uncountable, so the density result of the Weierstrass theorem approximates continuous functions by polynomials without describing them exactly. Spaces of differentiable functions, functions with prescribed boundary values, and integrable functions are built in the same way. Once a space is infinite-dimensional, finite-dimensional linear-algebra arguments no longer apply directly, and the choice of topology becomes essential. This is the reason for the norms introduced below.
Norms, metrics, and completeness
A norm assigns each function a nonnegative size and must satisfy positivity, homogeneity, and the triangle inequality. On the supremum norm,
measures the largest deviation of a function from zero, and the distance makes the space a metric space. Convergence in this norm is uniform convergence, so it is strictly stronger than pointwise convergence: a sequence can converge in norm while its values at a particular point behave differently from the limit at that point in the sense of the pointwise definition. A normed space in which every Cauchy sequence converges is a Banach space. Stefan Banach introduced this concept and studied it systematically in 1920–1922, together with Hans Hahn and Eduard Helly. with the supremum norm is a Banach space, since a uniform limit of continuous functions is continuous.
For a measure space and , the norm is
and the space consists of the measurable functions for which this integral is finite. Its elements are equivalence classes of functions that agree almost everywhere, so changing a function on a set of measure zero does not change it as an element of . Each space is a Banach space. For it is not a Hilbert space, but is one, with inner product (for complex-valued functions, with a complex conjugate on one factor). A Hilbert space is a complete inner-product space; every Hilbert space is a Banach space, but the converse fails. The inner product gives the geometric tools of orthogonality and projection onto closed subspaces, and a normed space is a Hilbert space exactly when the parallelogram law holds:
Sobolev spaces combine the norms of a function and of its derivatives up to a given order, and they are Banach spaces under this combined norm. Their derivatives are understood in the weak sense, which allows functions that are not differentiable in the classical sense to be included.
Linear operators and functionals
A linear map between normed spaces is bounded when there is a constant with for all . For linear maps, boundedness is equivalent to continuity, and the smallest such is the operator norm. A linear functional is a linear map into the scalar field, and the set of bounded linear functionals on is the dual space , which is itself a Banach space. The Riesz representation theorem identifies these dual spaces concretely. For , the dual of is , where . In a Hilbert space, every bounded functional is given by taking the inner product with a fixed vector, so a Hilbert space is isomorphic to its own dual.
Differential and integral operators
Operators acting on function spaces are the main objects in the theory of differential equations. Differentiation is a standard example of an operator that is not bounded: on , the functions satisfy , while . Differential operators are therefore usually defined on dense subspaces, and their domains are part of their definition. By contrast, an integral operator with a continuous kernel ,
is a bounded operator on , with . A linear differential equation can be studied as an equation between elements of function spaces, where the solutions form an affine subspace whose homogeneous part is the kernel of . For elliptic problems, such as the Poisson equation with zero boundary values, one often seeks weak solutions in a Sobolev space. There the existence of a solution can follow from the Riesz representation theorem applied to a suitable bilinear form, which is why the Hilbert-space structure is central to the modern theory of partial differential equations.
Special functions and integral transforms
The gamma function extends the factorial from the natural numbers to the complex plane. For positive real , it is defined by , and it satisfies , so for positive integers . The functional equation alone does not pin the function down. Bohr and Mollerup showed in 1922 that the unique log-convex function with that satisfies this equation is the gamma function. It has simple poles at the non-positive integers, takes the value , and obeys the reflection formula . Its normalizing constants appear throughout probability and statistics, for example in the densities of the gamma, beta and Student t distributions.
Bessel functions were first studied systematically by Friedrich Wilhelm Bessel in 1824, in the context of planetary perturbations. They are the solutions of , a differential equation that arises whenever a problem with circular or cylindrical symmetry is separated into variables. For integer order , the function has the integral representation . They describe the vibration modes of circular membranes, heat flow in cylinders, and electromagnetic waves in cylindrical waveguides. Legendre polynomials play a parallel role in spherical geometry. They are generated by , and they appear when Laplace's equation is separated in spherical coordinates. This is why they occur in the multipole expansion of gravitational and electrostatic potentials, a line of work that Legendre pursued from 1782 onward.
A generating function packs an entire sequence into a formal power series , so that combinatorial questions become questions about coefficients. The Fibonacci numbers with and have generating function , which yields the closed form by coefficient extraction. The partition function is generated by , and the Catalan generating function satisfies . Products of series correspond to convolutions of sequences, which is why multiplying generating functions counts ways of combining structures. Exponential generating functions are used instead when the objects are labeled, as in set partitions.
Integral transforms convert operations such as differentiation and convolution into simpler ones. The Laplace transform is , and it satisfies . Because of this rule, a linear differential equation with constant coefficients and given initial conditions becomes an algebraic equation in . The solution is then inverted, usually by partial fractions or a table of transforms. This method is standard in circuit analysis and control engineering. The Fourier transform uses as a common convention. It turns differentiation into multiplication by and convolution into pointwise multiplication. The Plancherel theorem makes it an isometry on , and it is the basic tool for solving the heat and wave equations on the whole line.
Orthogonal polynomial families are sequences , with , that satisfy for with respect to a positive weight on an interval. The classical examples are the Legendre polynomials (weight on ), the Chebyshev polynomials (weight on ), the Hermite polynomials (weight on ) and the Laguerre polynomials (weight on ). Each such family obeys a three-term recurrence of the form . Favard's theorem (1935) gives the converse: any sequence satisfying a suitable three-term recurrence is orthogonal with respect to some positive measure. Bochner's classification (1929) identifies the families that are eigenfunctions of a second-order differential operator, namely the Hermite, Laguerre and Jacobi polynomials. Hermite polynomials give the eigenfunctions of the quantum harmonic oscillator, and the zeros of orthogonal polynomials serve as nodes in Gaussian quadrature.
Functions in set theory, logic, and category theory
Within set theory, a function is not a primitive object but a particular kind of set. Using the Kuratowski definition of an ordered pair, , every element of a Cartesian product is itself a set. A function can therefore be identified with its graph, a subset in which every occurs in exactly one pair . In Zermelo–Fraenkel set theory, the axiom of replacement supplies what is needed for functions defined by formulas. If a formula assigns a unique output to each element of a set , the collection of those outputs is again a set. Without this guarantee, the image of a set under a definable operation would not be known to exist as an object, and the standard constructions of images and domains of such maps would lack a foundation.
Every function also determines an equivalence relation on its domain: holds exactly when . Reflexivity, symmetry, and transitivity follow directly from the properties of equality in the codomain. The equivalence classes partition , and the resulting quotient set is in bijection with the image through the map that sends each class to its common value. This observation is the set-theoretic core of the first isomorphism theorem in algebra: for a homomorphism , the quotient of by the kernel is isomorphic to the image of . The algebraic proof consists of this bijection together with a check that it preserves the structure. The correspondence runs in both directions, since any partition of yields a surjection from onto its quotient set. Choosing one representative from each class, which is needed to build a right inverse of a surjection, requires the axiom of choice in general.
Category theory, introduced by Samuel Eilenberg and Saunders Mac Lane in 1945, reverses the emphasis. Instead of building functions out of their elements, it characterizes morphisms by how they compose. A category consists of objects and morphisms, together with an associative composition and an identity morphism for each object. The category Set has sets as objects and functions as morphisms, with ordinary composition and identity maps supplying the structure. Inside Set, the category-theoretic notions match the familiar ones: monomorphisms are exactly the injective maps, epimorphisms are exactly the surjective maps, and isomorphisms are exactly the bijective maps. In general categories these notions can come apart, so they are defined by cancellation and invertibility properties, and the set-theoretic versions are special cases. A functor is a morphism between categories. It assigns objects to objects and morphisms to morphisms while preserving identities and composition. The power-set operation is a covariant functor on Set, sending to the direct image map on subsets. The preimage operation is contravariant: it sends subsets of to subsets of and reverses the direction of the arrow, so it is a contravariant functor on Set. Forgetful functors, such as the one that sends a group to its underlying set and discards the operation, show how functors can also simply strip away structure.
In first-order logic, function symbols belong to the signature of a language. Terms are built from variables and constants by applying function symbols, and in a structure with domain , each -ary function symbol is interpreted as a genuine map . The syntactic symbol and the semantic function are therefore distinct, and the same symbol may be interpreted differently in different structures. A function is called definable in a structure when there is a formula , possibly with parameters, such that for every there is exactly one satisfying . Every function symbol can be replaced by a relation symbol denoting its graph, at the cost of longer formulas and extra axioms stating that the relation is total and single-valued. The notion of definability is central to model theory, where the functions a structure can express determine much of its behavior.
Random variables and probability functions
A random variable is formally a measurable function defined on a probability space . Measurability requires that for every Borel set , the preimage belongs to the σ-algebra , so that is defined. The assignment is the distribution of , a probability measure on . Because the random variable is itself a function, its probabilistic behaviour is described by functions of a real variable. The three standard descriptions, the mass function, the density, and the distribution function, are different encodings of the same distribution.
For a discrete random variable taking values in a countable set, the probability mass function is , which satisfies and gives . For an absolutely continuous distribution, the probability density function satisfies , where the integral is a Lebesgue integral in general. A density value is not a probability and may exceed 1; a uniform distribution on , for instance, has density 2 there. The density is also determined only up to changes on sets of measure zero.
The cumulative distribution function is defined for every random variable. It is nondecreasing and right-continuous, and it tends to 0 at and to 1 at . The three functions are linked by summation and integration. In the discrete case , and in the absolutely continuous case . Conversely, is recovered as the derivative of wherever that derivative exists, which is almost everywhere. Interval probabilities follow directly, since .
Given a parametric family with mass or density and an observed value , the likelihood function is , viewed as a function of with held fixed. The same expression plays two roles. As a function of it is a density, but as a function of it need not integrate to one, and it is defined only up to a positive constant, so only ratios of its values are meaningful. For an independent sample , the likelihood is , and the log-likelihood converts the product into a sum. In maximum likelihood estimation, one chooses the parameter value that maximizes , often by setting its derivative to zero. For a Bernoulli sample with successes in trials, the estimate is . For a normal sample with known variance, the estimate of the mean is the sample mean.
If and is a differentiable, strictly monotone map between intervals with differentiable inverse , then
The factor records how the map stretches or compresses probability mass, and the formula is the probabilistic form of substitution in an integral. For example, if is uniform on and , then and for , which is the exponential distribution with rate 1. When is not monotone, the density is summed over all preimages, . If is standard normal and , this yields the chi-square density with one degree of freedom, for . For vector-valued variables, the factor is replaced by the absolute value of the determinant of the Jacobian matrix of the inverse map, as in the change-of-variables formula.
Functions in physics, chemistry, and economics
In quantum physics, the central function is the wave function, a complex-valued function of position and time. The output, Ψ, is the wave function: a complex-valued function of position and time. Its evolution follows the Schrödinger equation:
Here is the Hamiltonian operator, which contains the potential energy of the system, itself a function of position. The choice of determines the physical system: a free particle, a harmonic oscillator, a hydrogen atom, and a spin-1⁄2 particle in a magnetic field each have a distinct Ĥ. When the potential does not depend on time, the time-independent form gives stationary states, and for bound states (like a particle trapped in a box), the energy spectrum is discrete, meaning only certain energies are allowed. The wave function itself is not observed directly. Its squared modulus gives a probability density, a rule that was formalized by Max Born in 1926 and is now called the Born rule.
In chemistry, the analogous functional dependence is the rate law, which expresses a reaction rate as a function of reactant concentrations. Enzyme-catalysed reactions do not fit simple power laws. Instead the reaction reaches a limiting or saturation rate. The standard model for this behaviour is the Michaelis–Menten equation, which gives the initial rate as a function of substrate concentration :
Here is the limiting rate when the enzyme is saturated, and is the substrate concentration at which the reaction rate is 50% of the Vmax. The function is a rectangular hyperbola: the reaction rate approaches as substrate concentration increases. Its two regimes correspond to different effective orders, since at low substrate concentrations, the rate is almost first order, while at very high substrate concentrations the rate is almost zero^th order. The model rests on a steady-state assumption, as it assumes that the reaction is in the steady state, where the ES concentration remains constant. Its validity is also bounded: recent experiments have measured fluctuations in the catalytic rate to reveal that the MM equation, though valid for bulk amounts, is not obeyed at the molecular level.
In economics, several basic models are functions of a single economic variable or of a bundle of goods. A demand function gives the quantity buyers want at price , typically decreasing in , while a supply function is typically increasing. Market equilibrium is the price at which . A utility function assigns a number to each consumption bundle so that preferences can be ranked. The utility function expresses utility as a function of consumption of real goods (pounds, gallons, kilograms, liters) as opposed to nominal goods (dollars, euros). The cost side uses a total cost function . In economics, the cost curve expresses production costs in terms of the amount produced. Dividing by output gives the average cost , and the marginal cost is the derivative of total cost with respect to output, which links these models to differentiation.
Empirical work fits model functions to data. Given observations , one chooses a family and estimates the parameters by minimizing a loss, most commonly the sum of squared residuals:
When is linear in its parameters, this is linear regression, which has a closed-form solution. Under a probabilistic error model, the same estimation problem can be stated in terms of the likelihood of the parameters, which connects the least-squares criterion to maximum likelihood estimation.
Computation, programming, and machine learning
In computer science, "function" names a reusable block of code that accepts arguments and returns a result. Real programs do not always follow the mathematical ideal. A pure function depends only on its arguments and returns the same output for the same inputs, without side effects such as writing to a file, modifying a global variable, or printing to a screen. Functional programming languages favour pure functions because they are predictable and easy to test, and their behaviour can be reasoned about in a way that resembles equational reasoning in mathematics. Imperative code, by contrast, updates shared state freely, so one call can change the result of a later call.
A higher-order function takes one or more functions as arguments, returns a function as its result, or does both. Mathematicians often call such objects operators or functionals, and the differential operator, which sends a function to its derivative, is the standard example. In programming, operations such as map, filter, and reduce apply a supplied function to the elements of a collection, and most modern languages provide them. In the untyped lambda calculus, every function is higher-order. Typed variants of the calculus underlie much of the design of functional languages, and John McCarthy's Lisp, introduced in 1958, was built on the lambda calculus.
The theoretical foundations of computing rest on the work of 1936 and 1937. Alonzo Church proposed the lambda calculus, Kurt Gödel and Stephen Kleene worked with recursive functions, and Emil Post and Alan Turing devised abstract machines. The lambda calculus has only three ingredients: variables, abstraction (forming a function such as λx.M), and application (applying one term to another). It contains no built-in numbers or arithmetic, yet arithmetic and recursion can be encoded within it. These formalisms were shown to define the same class of functions. This convergence led to the Church–Turing thesis, which identifies the intuitively computable functions with those that a Turing machine can compute. The thesis has no formal proof, but every realistic model of computation proposed so far has proved equivalent to it. Systems able to express all such functions are called Turing complete. Church and Turing used these models to show independently that the Entscheidungsproblem of deciding the validity of first-order statements has no general algorithmic solution.
Modern machine learning uses functions in a different way. An artificial neural network is a composition of parametrized functions. Each layer computes a weighted sum of its inputs and adds a bias, then applies a nonlinear activation function such as the sigmoid, the hyperbolic tangent, or the rectified linear unit (ReLU). For a three-layer network with layer functions f₁, f₂, and f₃, the output is f₃(f₂(f₁(x))), and each layer's weights are the parameters of that layer's function. The nonlinearity is essential. Without it, a composition of affine maps is itself affine, so additional layers would add no representational power beyond a single linear layer.
Training a network means adjusting its parameters to reduce a loss function, which maps the network's predictions and the target values to a single real number measuring their mismatch. Common choices include mean squared error for regression and cross-entropy for classification. The loss is the outermost function in a very long composition, so its dependence on each weight is found by the chain rule. backpropagation organizes this computation: it evaluates the network forward, stores intermediate values, and then propagates derivatives backward from the output layer to the first layer. The resulting gradients feed into gradient-based optimization methods such as gradient descent, which update the weights in small steps. Because the method only needs the activation and loss functions and their derivatives to be efficiently computable, the same procedure applies to many different architectures.
Generalized and alternative notions of function
Partial and multivalued functions
A partial function assigns at most one output to each input but need not be defined on the whole of its input set. The set where it is defined is its domain of definition. Such maps arise naturally: division by a variable is undefined at zero, and the real square root is defined only on the nonnegative reals. Analysts usually treat such an object as a function on its natural domain, so that statements about continuity or differentiability refer to that domain.
A multivalued function assigns a set of outputs to each input, so it fails the single-output requirement of the strict definition. The relation is the standard example, since every positive has two square roots. Analysts handle such relations in two main ways. The first is to choose a branch, a single-valued function on a region that selects one value at each point. The principal logarithm, with , is an example. The second is to keep all values and follow them continuously. Analytic continuation along paths in the complex plane shows that the logarithm and the square root become single-valued only on suitable Riemann surfaces, which are multi-sheeted covers of the plane.
Distributions
A distribution is a continuous linear functional on a space of smooth test functions, usually the smooth functions with compact support. Distributions do not take values at individual points. The Dirac delta is the best-known example. Dirac used it in quantum mechanics in the 1920s and 1930s, and it is characterized by for every test function . The expression "" is therefore a heuristic and not a value of a function. Laurent Schwartz gave the systematic theory in the early 1950s. Distributions can be differentiated without restriction; the derivative of the Heaviside step function, for instance, is in this sense. The theory supplies the Fourier transform of non-integrable functions, Green's functions for differential equations, and the weak formulations used in Sobolev spaces.
Functions on discrete and nonnumerical sets
Many functions are defined on finite or symbolic domains. A Boolean function maps to . There are such functions, which gives 16 for two variables and 256 for three. Boolean functions underlie digital circuit design and Boolean algebra, and each non-constant one can be written as a propositional formula in disjunctive normal form.
Functions over a finite field form another important class. For a prime , Fermat's little theorem gives for every in . Consequently, distinct polynomials such as and induce the same function on . A function over a finite field therefore has no unique polynomial expression. Standard practice selects reduced polynomials, which have degree less than in each variable, as canonical representatives. Such functions appear in coding theory and cryptography.
Debates over the definition
Euler's 1748 Introductio in analysin infinitorum described a function of a variable quantity as an analytic expression built from that quantity and constants. On this view, functions are in principle given by formulas. Dirichlet's formulation, which descends into the set-theoretic definition used in most modern texts, requires only that each input be assigned exactly one output, whatever the rule. His example, for rational and for irrational , is nowhere continuous and has no analytic expression of the kind Euler had in mind. The two views therefore lead to different classes of objects.
The difference affects teaching. Courses that begin with formulas tend to present a function as a rule for computation. Courses that begin with sets and correspondences tend to emphasize ordered pairs, graphs, and the general mapping concept. Educators continue to disagree about which approach better prepares students for calculus and for the set-based language of higher mathematics, and that disagreement concerns pedagogical emphasis rather than mathematical content.
References
- function concept - MacTutor History of Mathematicsmathshistory.st-andrews.ac.uk
- René Descartes Mathematics: How Algebra Met Geometry - Kronecker Walliskroneckerwallis.com
- History of the function concepten.wikipedia.org
- (PDF) The History of the Concept of Function and Some Educational Implicationsresearchgate.net
- Chapter 5 Functions: How they have changed through Historyms.uky.edu
- Mathematics and languagearxiv.org
- (PDF) Evolution of the Function Concept: A Brief Surveyresearchgate.net
- Functions, what are they?cut-the-knot.org
- History of function conceptsummaryplanet.com
- Geometry in the Age of Enlightenmentarxiv.org
- Rene Descartesslideshare.net
- Even and Odd Function - an overviewsciencedirect.com
- Jensen's inequalityen.wikipedia.org
- Part 2.6: Even and Odd Functionscfm.brown.edu
- Jensen's inequalitystatlect.com
- Symmetry - Ximera - The Ohio State Universityximera.osu.edu
- Cauchy-Riemann Equation - an overviewsciencedirect.com
- On the nonlinear Cauchy-Riemann equations of structural transformation and nonlinear Laplace equationarxiv.org
- A rapid review of complex function theory 1 Holomorphic ...web.ma.utexas.edu
- Holomorphic function - Citizendiumen.citizendium.org
- Introduction to Analytic Number Theory The Riemann zeta ...people.math.harvard.edu
- On Riemann's Paper, "On the Number of Primes Less Than a Given Magnitude"arxiv.org
- 9: Contour Integration - Physics LibreTextsphys.libretexts.org
- Residue Theorem - GeeksforGeeksgeeksforgeeks.org
- Applying Residue Theorem to Compute Real Definite Integraliopscience.iop.org
- Evaluating real integrals using residuesfiveable.me
- Applications and Techniques of Contour Integration in ...ijltemas.in
- MEROMORPHIC CONTINUATION AND FUNCTIONAL EQUATION OF RIEMANN ZETApeople.reed.edu
- Function Spaces 1 Hilbert Spacesstat.cmu.edu
- Chapter 2 Banach and Hilbert space analysis 2.1 Hilbert Spaces 2.1.1dpmms.cam.ac.uk
- Banach Space - an overviewsciencedirect.com
- graphsearch.epfl.chgraphsearch.epfl.ch
- Sobolev spaceen.wikipedia.org
- L^p-Space -- from Wolfram MathWorldmathworld.wolfram.com
- Lp Space - an overviewsciencedirect.com
- 8.1. Riesz Representation Theorem 1 Chapter 8. The Lp Spaces:faculty.etsu.edu
- 245B, notes 3: L^p spacesterrytao.wordpress.com
- The Schrödinger Equation: Wave Functions and Quantum Statesquantumphysicsauthority.com
- Schrödinger Equation and Wave Functionsfiveable.me
- Enzyme Kinetics & The Michaelis-Menten Equationecampusontario.pressbooks.pub
- Enzyme Kinetics - Structure - Function - Michaelis-Menten Kineticsteachmephysiology.com
- Michaelis-Menten Equation Explained: Definition, Examples, Practice & Video Lessonspearson.com
- Enzyme kinetics at the molecular levelarxiv.org
- Definition of Utility Functioneconmodel.com
- Cost functionen.wikipedia.org
- The Original Michaelis Constant: Translation of the 1913 Michaelis-Menten Paper - PMCpmc.ncbi.nlm.nih.gov
- The Michaelis–Menten Modelmicrobenotes.com
- Michaelis-Menten Kinetics: Vmax, Km, Equation Explainedgasgasgas.uk
- Higher-order functionen.wikipedia.org
- Computability and Complexity (Stanford Encyclopedia of Philosophy)plato.stanford.edu
- Church-Turing Thesis -- from Wolfram MathWorldmathworld.wolfram.com
- Physics, Topology, Logic and Computation: A Rosetta Stonearxiv.org
- Mass problems and intuitionistic higher-order logicarxiv.org
- Very Simple Chaitin Machines for Concrete AITarxiv.org
- Alonzo Church and Lambda Calculus: Foundations of Computing - Kronecker Walliskroneckerwallis.com
- Pure Functions and Higher-Order Functions - Cocoacastscocoacasts.com
- Backpropagationen.wikipedia.org
- What is Backpropagation?ibm.com
- Neural networks: training with backpropagation. - Jeremy Jordanjeremyjordan.me
- Backpropagation in Neural Networkscodesignal.com