This article may be too long to read and navigate comfortably. Consider splitting content into sub-articles, condensing it, or adding subheadings. Please discuss this issue on the article's talk page. (December 2020)

Distributions, also known as Schwartz distributions or generalized functions, are objects that generalize the classical notion of functions in mathematical analysis. Distributions make it possible to differentiate functions whose derivatives do not exist in the classical sense. In particular, any locally integrable function has a distributional derivative.

Distributions are widely used in the theory of partial differential equations, where it may be easier to establish the existence of distributional solutions (weak solutions) than classical solutions, or where appropriate classical solutions may not exist. Distributions are also important in physics and engineering where many problems naturally lead to differential equations whose solutions or initial conditions are singular, such as the Dirac delta function.

A function is normally thought of as acting on the points in the function domain by "sending" a point in the domain to the point Instead of acting on points, distribution theory reinterprets functions such as as acting on test functions in a certain way. In applications to physics and engineering, test functions are usually infinitely differentiable complex-valued (or real-valued) functions with compact support that are defined on some given non-empty open subset . (Bump functions are examples of test functions.) The set of all such test functions forms a vector space that is denoted by or

Most commonly encountered functions, including all continuous maps if using can be canonically reinterpreted as acting via "integration against a test function." Explicitly, this means that such a function "acts on" a test function by "sending" it to the number which is often denoted by This new action of defines a scalar-valued map whose domain is the space of test functions This functional turns out to have the two defining properties of what is known as a distribution on : it is linear, and it is also continuous when is given a certain topology called the canonical LF topology. The action (the integration ) of this distribution on a test function can be interpreted as a weighted average of the distribution on the support of the test function, even if the values of the distribution at a single point are not well-defined. Distributions like that arise from functions in this way are prototypical examples of distributions, but there exist many distributions that cannot be defined by integration against any function. Examples of the latter include the Dirac delta function and distributions defined to act by integration of test functions against certain measures on Nonetheless, it is still always possible to reduce any arbitrary distribution down to a simpler family of related distributions that do arise via such actions of integration.

More generally, a distribution on is by definition a linear functional on that is continuous when is given a topology called the canonical LF topology. This leads to the space of (all) distributions on , usually denoted by (note the prime), which by definition is the space of all distributions on (that is, it is the continuous dual space of ); it is these distributions that are the main focus of this article.

Definitions of the appropriate topologies on spaces of test functions and distributions are given in the article on spaces of test functions and distributions. This article is primarily concerned with the definition of distributions, together with their properties and some important examples.


The practical use of distributions can be traced back to the use of Green's functions in the 1830s to solve ordinary differential equations, but was not formalized until much later. According to Kolmogorov & Fomin (1957), generalized functions originated in the work of Sergei Sobolev (1936) on second-order hyperbolic partial differential equations, and the ideas were developed in somewhat extended form by Laurent Schwartz in the late 1940s. According to his autobiography, Schwartz introduced the term "distribution" by analogy with a distribution of electrical charge, possibly including not only point charges but also dipoles and so on. Gårding (1997) comments that although the ideas in the transformative book by Schwartz (1951) were not entirely new, it was Schwartz's broad attack and conviction that distributions would be useful almost everywhere in analysis that made the difference.


The following notation will be used throughout this article:

Definitions of test functions and distributions

In this section, some basic notions and definitions needed to define real-valued distributions on U are introduced. Further discussion of the topologies on the spaces of test functions and distributions is given in the article on spaces of test functions and distributions.

  1. Let
  2. Let denote the vector space of all k-times continuously differentiable real or complex-valued functions on U.
  3. For any compact subset let and both denote the vector space of all those functions such that
    • If then the domain of is U and not K. So although depends on both K and U, only K is typically indicated. The justification for this common practice is detailed below. The notation will only be used when the notation risks being ambiguous.
    • Every contains the constant 0 map, even if
  4. Let denote the set of all such that for some compact subset K of U.
    • Equivalently, is the set of all such that has compact support.
    • is equal to the union of all as ranges over all compact subsets of
    • If is a real-valued function on , then is an element of if and only if is a bump function. Every real-valued test function on is also a complex-valued test function on
The graph of the bump function where and This function is a test function on and is an element of The support of this function is the closed unit disk in It is non-zero on the open unit disk and it is equal to 0 everywhere outside of it.

For all and any compact subsets and of , we have:

Definition: Elements of are called test functions on U and is called the space of test functions on U. We will use both and to denote this space.

Distributions on U are continuous linear functionals on when this vector space is endowed with a particular topology called the canonical LF-topology. The following proposition states two necessary and sufficient conditions for the continuity of a linear function on that are often straightforward to verify.

Proposition: A linear functional T on is continuous, and therefore a distribution, if and only if any of the following equivalent conditions is satisfied:

  1. For every compact subset there exist constants and (dependent on ) such that for all with support contained in ,[1][2]
  2. For every compact subset and every sequence in whose supports are contained in , if converges uniformly to zero on for every multi-index , then

Topology on Ck(U)

We now introduce the seminorms that will define the topology on Different authors sometimes use different families of seminorms so we list the most common families below. However, the resulting topology is the same no matter which family is used.

Suppose and is an arbitrary compact subset of Suppose an integer such that [note 1] and is a multi-index with length For define:

while for define all the functions above to be the constant 0 map.

All of the functions above are non-negative -valued[note 2] seminorms on As explained in this article, every set of seminorms on a vector space induces a locally convex vector topology.

Each of the following sets of seminorms

generate the same locally convex vector topology on (so for example, the topology generated by the seminorms in is equal to the topology generated by those in ).

The vector space is endowed with the locally convex topology induced by any one of the four families of seminorms described above. This topology is also equal to the vector topology induced by all of the seminorms in

With this topology, becomes a locally convex Fréchet space that is not normable. Every element of is a continuous seminorm on Under this topology, a net in converges to if and only if for every multi-index with and every compact the net of partial derivatives converges uniformly to on [3] For any any (von Neumann) bounded subset of is a relatively compact subset of [4] In particular, a subset of is bounded if and only if it is bounded in for all [4] The space is a Montel space if and only if [5]

A subset of is open in this topology if and only if there exists such that is open when is endowed with the subspace topology induced on it by

Topology on Ck(K)

As before, fix Recall that if is any compact subset of then

Assumption: For any compact subset we will henceforth assume that is endowed with the subspace topology it inherits from the Fréchet space

If is finite then is a Banach space[6] with a topology that can be defined by the norm

And when then is even a Hilbert space.[6]

Trivial extensions and independence of Ck(K)'s topology from U

Suppose is an open subset of and is a compact subset. By definition, elements of are functions with domain (in symbols, ), so the space and its topology depend on to make this dependence on the open set clear, temporarily denote by Importantly, changing the set to a different open subset (with ) will change the set from to [note 3] so that elements of will be functions with domain instead of Despite depending on the open set (), the standard notation for makes no mention of it. This is justified because, as this subsection will now explain, the space is canonically identified as a subspace of (both algebraically and topologically).

It is enough to explain how to canonically identify with when one of and is a subset of the other. The reason is that if and are arbitrary open subsets of containing then the open set also contains so that each of and is canonically identified with and now by transitivity, is thus identified with So assume are open subsets of containing

Given its trivial extension to is the function defined by:

This trivial extension belongs to (because has compact support) and it will be denoted by (that is, ). The assignment thus induces a map that sends a function in to its trivial extension on This map is a linear injection and for every compact subset (where is also a compact subset of since ),
If is restricted to then the following induced linear map is a homeomorphism (linear homeomorphisms are called TVS-isomorphisms):
and thus the next map is a topological embedding:
Using the injection
the vector space is canonically identified with its image in Because through this identification, can also be considered as a subset of Thus the topology on is independent of the open subset of that contains [7] which justifies the practice of writing instead of

Canonical LF topology

Main article: Spaces of test functions and distributions

See also: LF-space and Topology of uniform convergence

Recall that denotes all functions in that have compact support in where note that is the union of all as ranges over all compact subsets of Moreover, for each is a dense subset of The special case when gives us the space of test functions.

is called the space of test functions on and it may also be denoted by Unless indicated otherwise, it is endowed with a topology called the canonical LF topology, whose definition is given in the article: Spaces of test functions and distributions.

The canonical LF-topology is not metrizable and importantly, it is strictly finer than the subspace topology that induces on However, the canonical LF-topology does make into a complete reflexive nuclear[8] Montel[9] bornological barrelled Mackey space; the same is true of its strong dual space (that is, the space of all distributions with its usual topology). The canonical LF-topology can be defined in various ways.


See also: Continuous linear functional

As discussed earlier, continuous linear functionals on a are known as distributions on Other equivalent definitions are described below.

By definition, a distribution on is a continuous linear functional on Said differently, a distribution on is an element of the continuous dual space of when is endowed with its canonical LF topology.

There is a canonical duality pairing between a distribution on and a test function which is denoted using angle brackets by

One interprets this notation as the distribution acting on the test function to give a scalar, or symmetrically as the test function acting on the distribution

Characterizations of distributions

Proposition. If is a linear functional on then the following are equivalent:

  1. T is a distribution;
  2. T is continuous;
  3. T is continuous at the origin;
  4. T is uniformly continuous;
  5. T is a bounded operator;
  6. T is sequentially continuous;
    • explicitly, for every sequence in that converges in to some [note 4]
  7. T is sequentially continuous at the origin; in other words, T maps null sequences[note 5] to null sequences;
    • explicitly, for every sequence in that converges in to the origin (such a sequence is called a null sequence),
    • a null sequence is by definition any sequence that converges to the origin;
  8. T maps null sequences to bounded subsets;
    • explicitly, for every sequence in that converges in to the origin, the sequence is bounded;
  9. T maps Mackey convergent null sequences to bounded subsets;
    • explicitly, for every Mackey convergent null sequence in the sequence is bounded;
    • a sequence is said to be Mackey convergent to the origin if there exists a divergent sequence of positive real numbers such that the sequence is bounded; every sequence that is Mackey convergent to the origin necessarily converges to the origin (in the usual sense);
  10. The kernel of T is a closed subspace of
  11. The graph of T is closed;
  12. There exists a continuous seminorm on such that
  13. There exists a constant and a finite subset (where is any collection of continuous seminorms that defines the canonical LF topology on ) such that [note 6]
  14. For every compact subset there exist constants and such that for all [1]
  15. For every compact subset there exist constants and such that for all with support contained in [10]
  16. For any compact subset and any sequence in if converges uniformly to zero for all multi-indices then

Topology on the space of distributions and its relation to the weak-* topology

The set of all distributions on is the continuous dual space of which when endowed with the strong dual topology is denoted by Importantly, unless indicated otherwise, the topology on is the strong dual topology; if the topology is instead the weak-* topology then this will be indicated. Neither topology is metrizable although unlike the weak-* topology, the strong dual topology makes into a complete nuclear space, to name just a few of its desirable properties.

Neither nor its strong dual is a sequential space and so neither of their topologies can be fully described by sequences (in other words, defining only what sequences converge in these spaces is not enough to fully/correctly define their topologies). However, a sequence in converges in the strong dual topology if and only if it converges in the weak-* topology (this leads many authors to use pointwise convergence to define the convergence of a sequence of distributions; this is fine for sequences but this is not guaranteed to extend to the convergence of nets of distributions because a net may converge pointwise but fail to converge in the strong dual topology). More information about the topology that is endowed with can be found in the article on spaces of test functions and distributions and the articles on polar topologies and dual systems.

A linear map from into another locally convex topological vector space (such as any normed space) is continuous if and only if it is sequentially continuous at the origin. However, this is no longer guaranteed if the map is not linear or for maps valued in more general topological spaces (for example, that are not also locally convex topological vector spaces). The same is true of maps from (more generally, this is true of maps from any locally convex bornological space).

Localization of distributions

There is no way to define the value of a distribution in at a particular point of U. However, as is the case with functions, distributions on U restrict to give distributions on open subsets of U. Furthermore, distributions are locally determined in the sense that a distribution on all of U can be assembled from a distribution on an open cover of U satisfying some compatibility conditions on the overlaps. Such a structure is known as a sheaf.

Extensions and restrictions to an open subset

Let be open subsets of Every function can be extended by zero from its domain V to a function on U by setting it equal to on the complement This extension is a smooth compactly supported function called the trivial extension of to and it will be denoted by This assignment defines the trivial extension operator which is a continuous injective linear map. It is used to canonically identify as a vector subspace of (although not as a topological subspace). Its transpose (explained here)

is called the restriction to of distributions in [11] and as the name suggests, the image of a distribution under this map is a distribution on called the restriction of to The defining condition of the restriction is:
If then the (continuous injective linear) trivial extension map is not a topological embedding (in other words, if this linear injection was used to identify as a subset of then 's topology would strictly finer than the subspace topology that induces on it; importantly, it would not be a topological subspace since that requires equality of topologies) and its range is also not dense in its codomain [11] Consequently if then the restriction mapping is neither injective nor surjective.[11] A distribution is said to be extendible to U if it belongs to the range of the transpose of and it is called extendible if it is extendable to [11]

Unless the restriction to V is neither injective nor surjective. Lack of surjectivity follows since distributions can blow up towards the boundary of V. For instance, if and then the distribution

is in but admits no extension to

Gluing and distributions that vanish in a set

Theorem[12] — Let be a collection of open subsets of For each let and suppose that for all the restriction of to is equal to the restriction of to (note that both restrictions are elements of ). Then there exists a unique such that for all the restriction of T to is equal to

Let V be an open subset of U. is said to vanish in V if for all such that we have T vanishes in V if and only if the restriction of T to V is equal to 0, or equivalently, if and only if T lies in the kernel of the restriction map

Corollary[12] — Let be a collection of open subsets of and let if and only if for each the restriction of T to is equal to 0.

Corollary[12] — The union of all open subsets of U in which a distribution T vanishes is an open subset of U in which T vanishes.

Support of a distribution

This last corollary implies that for every distribution T on U, there exists a unique largest subset V of U such that T vanishes in V (and does not vanish in any open subset of U that is not contained in V); the complement in U of this unique largest open subset is called the support of T.[12] Thus

If is a locally integrable function on U and if is its associated distribution, then the support of is the smallest closed subset of U in the complement of which is almost everywhere equal to 0.[12] If is continuous, then the support of is equal to the closure of the set of points in U at which does not vanish.[12] The support of the distribution associated with the Dirac measure at a point is the set [12] If the support of a test function does not intersect the support of a distribution T then A distribution T is 0 if and only if its support is empty. If is identically 1 on some open set containing the support of a distribution T then If the support of a distribution T is compact then it has finite order and there is a constant and a non-negative integer such that:[7]

If T has compact support, then it has a unique extension to a continuous linear functional on ; this function can be defined by where is any function that is identically 1 on an open set containing the support of T.[7]

If and then and Thus, distributions with support in a given subset form a vector subspace of [13] Furthermore, if is a differential operator in U, then for all distributions T on U and all we have and [13]

Distributions with compact support

Support in a point set and Dirac measures

For any let denote the distribution induced by the Dirac measure at For any and distribution the support of T is contained in if and only if T is a finite linear combination of derivatives of the Dirac measure at [14] If in addition the order of T is then there exist constants such that:[15]

Said differently, if T has support at a single point then T is in fact a finite linear combination of distributional derivatives of the function at P. That is, there exists an integer m and complex constants such that

where is the translation operator.

Distribution with compact support

Theorem[7] — Suppose T is a distribution on U with compact support K. There exists a continuous function defined on U and a multi-index p such that

where the derivatives are understood in the sense of distributions. That is, for all test functions on U,

Distributions of finite order with support in an open subset

Theorem[7] — Suppose T is a distribution on U with compact support K and let V be an open subset of U containing K. Since every distribution with compact support has finite order, take N to be the order of T and define There exists a family of continuous functions defined on U with support in V such that

where the derivatives are understood in the sense of distributions. That is, for all test functions on U,

Global structure of distributions

The formal definition of distributions exhibits them as a subspace of a very large space, namely the topological dual of (or the Schwartz space for tempered distributions). It is not immediately clear from the definition how exotic a distribution might be. To answer this question, it is instructive to see distributions built up from a smaller space, namely the space of continuous functions. Roughly, any distribution is locally a (multiple) derivative of a continuous function. A precise version of this result, given below, holds for distributions of compact support, tempered distributions, and general distributions. Generally speaking, no proper subset of the space of distributions contains all continuous functions and is closed under differentiation. This says that distributions are not particularly exotic objects; they are only as complicated as necessary.

Distributions as sheaves

Theorem[16] — Let T be a distribution on U. There exists a sequence in such that each Ti has compact support and every compact subset intersects the support of only finitely many and the sequence of partial sums defined by converges in to T; in other words we have:

Recall that a sequence converges in (with its strong dual topology) if and only if it converges pointwise.

Decomposition of distributions as sums of derivatives of continuous functions

By combining the above results, one may express any distribution on U as the sum of a series of distributions with compact support, where each of these distributions can in turn be written as a finite sum of distributional derivatives of continuous functions on U. In other words, for arbitrary we can write:

where are finite sets of multi-indices and the functions are continuous.

Theorem[17] — Let T be a distribution on U. For every multi-index p there exists a continuous function on U such that

  1. any compact subset K of U intersects the support of only finitely many and

Moreover, if T has finite order, then one can choose in such a way that only finitely many of them are non-zero.

Note that the infinite sum above is well-defined as a distribution. The value of T for a given can be computed using the finitely many that intersect the support of

Operations on distributions

Many operations which are defined on smooth functions with compact support can also be defined for distributions. In general, if is a linear map that is continuous with respect to the weak topology, then it is not always possible to extend to a map by classic extension theorems of topology or linear functional analysis.[note 7] The “distributional” extension of the above linear continuous operator A is possible if and only if A admits a Schwartz adjoint, that is another linear continuous operator B of the same type such that , for every pair of test functions. In that condition, B is unique and the extension A’ is the transpose of the Schwartz adjoint B. [citation needed][18][clarification needed]

Preliminaries: Transpose of a linear operator

Main article: Transpose of a linear map

Operations on distributions and spaces of distributions are often defined using the transpose of a linear operator. This is because the transpose allows for a unified presentation of the many definitions in the theory of distributions and also because its properties are well-known in functional analysis.[19] For instance, the well-known Hermitian adjoint of a linear operator between Hilbert spaces is just the operator's transpose (but with the Riesz representation theorem used to identify each Hilbert space with its continuous dual space). In general, the transpose of a continuous linear map is the linear map

or equivalently, it is the unique map satisfying for all and all (the prime symbol in does not denote a derivative of any kind; it merely indicates that is an element of the continuous dual space ). Since is continuous, the transpose is also continuous when both duals are endowed with their respective strong dual topologies; it is also continuous when both duals are endowed with their respective weak* topologies (see the articles polar topology and dual system for more details).

In the context of distributions, the characterization of the transpose can be refined slightly. Let be a continuous linear map. Then by definition, the transpose of is the unique linear operator that satisfies:

Since is dense in (here, actually refers to the set of distributions ) it is sufficient that the defining equality hold for all distributions of the form where Explicitly, this means that a continuous linear map is equal to if and only if the condition below holds:

where the right-hand side equals

Differential operators

Differentiation of distributions

Let be the partial derivative operator To extend we compute its transpose:

Therefore Thus, the partial derivative of with respect to the coordinate is defined by the formula

With this definition, every distribution is infinitely differentiable, and the derivative in the direction is a linear operator on

More generally, if is an arbitrary multi-index, then the partial derivative of the distribution is defined by

Differentiation of distributions is a continuous operator on this is an important and desirable property that is not shared by most other notions of differentiation.

If is a distribution in then

where is the derivative of and is a translation by thus the derivative of may be viewed as a limit of quotients.[20]

Differential operators acting on smooth functions

A linear differential operator in with smooth coefficients acts on the space of smooth functions on Given such an operator we would like to define a continuous linear map, that extends the action of on to distributions on In other words, we would like to define such that the following diagram commutes:

where the vertical maps are given by assigning its canonical distribution which is defined by:
With this notation, the diagram commuting is equivalent to:

To find the transpose of the continuous induced map defined by is considered in the lemma below. This leads to the following definition of the differential operator on called the formal transpose of which will be denoted by to avoid confusion with the transpose map, that is defined by

Lemma — Let be a linear differential operator with smooth coefficients in Then for all we have

which is equivalent to:


As discussed above, for any the transpose may be calculated by:

For the last line we used integration by parts combined with the fact that and therefore all the functions have compact support.[note 8] Continuing the calculation above, for all

The Lemma combined with the fact that the formal transpose of the formal transpose is the original differential operator, that is, [21] enables us to arrive at the correct definition: the formal transpose induces the (continuous) canonical linear operator defined by We claim that the transpose of this map, can be taken as To see this, for every compute its action on a distribution of the form with :

We call the continuous linear operator the differential operator on distributions extending .[21] Its action on an arbitrary distribution is defined via:

If converges to then for every multi-index converges to

Multiplication of distributions by smooth functions

A differential operator of order 0 is just multiplication by a smooth function. And conversely, if is a smooth function then is a differential operator of order 0, whose formal transpose is itself (that is, ). The induced differential operator maps a distribution to a distribution denoted by We have thus defined the multiplication of a distribution by a smooth function.

We now give an alternative presentation of the multiplication of a distribution on by a smooth function The product is defined by

This definition coincides with the transpose definition since if is the operator of multiplication by the function (that is, ), then

so that

Under multiplication by smooth functions, is a module over the ring With this definition of multiplication by a smooth function, the ordinary product rule of calculus remains valid. However, some unusual identities also arise. For example, if is the Dirac delta distribution on then and if is the derivative of the delta distribution, then

The bilinear multiplication map given by is not continuous; it is however, hypocontinuous.[22]

Example. The product of any distribution with the function that is identically 1 on is equal to

Example. Suppose is a sequence of test functions on that converges to the constant function For any distribution on the sequence converges to [23]

If converges to and converges to then converges to

Problem of multiplying distributions

It is easy to define the product of a distribution with a smooth function, or more generally the product of two distributions whose singular supports are disjoint.[24] With more effort, it is possible to define a well-behaved product of several distributions provided their wave front sets at each point are compatible. A limitation of the theory of distributions (and hyperfunctions) is that there is no associative product of two distributions extending the product of a distribution by a smooth function, as has been proved by Laurent Schwartz in the 1950s. For example, if is the distribution obtained by the Cauchy principal value

If is the Dirac delta distribution then

so the product of a distribution by a smooth function (which is always well-defined) cannot be extended to an associative product on the space of distributions.

Thus, nonlinear problems cannot be posed in general and thus are not solved within distribution theory alone. In the context of quantum field theory, however, solutions can be found. In more than two spacetime dimensions the problem is related to the regularization of divergences. Here Henri Epstein and Vladimir Glaser developed the mathematically rigorous (but extremely technical) causal perturbation theory. This does not solve the problem in other situations. Many other interesting theories are non-linear, like for example the Navier–Stokes equations of fluid dynamics.

Several not entirely satisfactory[citation needed] theories of algebras of generalized functions have been developed, among which Colombeau's (simplified) algebra is maybe the most popular in use today.

Inspired by Lyons' rough path theory,[25] Martin Hairer proposed a consistent way of multiplying distributions with certain structures (regularity structures[26]), available in many examples from stochastic analysis, notably stochastic partial differential equations. See also Gubinelli–Imkeller–Perkowski (2015) for a related development based on Bony's paraproduct from Fourier analysis.

Composition with a smooth function

Let be a distribution on Let be an open set in and If is a submersion then it is possible to define

This is the composition of the distribution with , and is also called the pullback of along , sometimes written

The pullback is often denoted although this notation should not be confused with the use of '*' to denote the adjoint of a linear mapping.

The condition that be a submersion is equivalent to the requirement that the Jacobian derivative of is a surjective linear map for every A necessary (but not sufficient) condition for extending to distributions is that be an open mapping.[27] The Inverse function theorem ensures that a submersion satisfies this condition.

If is a submersion, then is defined on distributions by finding the transpose map. The uniqueness of this extension is guaranteed since is a continuous linear operator on Existence, however, requires using the change of variables formula, the inverse function theorem (locally), and a partition of unity argument.[28]

In the special case when is a diffeomorphism from an open subset of onto an open subset of change of variables under the integral gives:

In this particular case, then, is defined by the transpose formula:


Under some circumstances, it is possible to define the convolution of a function with a distribution, or even the convolution of two distributions. Recall that if and are functions on then we denote by the convolution of and defined at to be the integral

provided that the integral exists. If are such that then for any functions and we have and [29] If and are continuous functions on at least one of which has compact support, then and if then the value of on do not depend on the values of outside of the Minkowski sum [29]

Importantly, if has compact support then for any the convolution map is continuous when considered as the map or as the map [29]

Translation and symmetry

Given the translation operator sends to defined by This can be extended by the transpose to distributions in the following way: given a distribution the translation of by is the distribution defined by [30][31]

Given define the function by Given a distribution let be the distribution defined by