# Unbounded operator

In mathematics, more specifically functional analysis and operator theory, the notion of **unbounded operator** provides an abstract framework for dealing with differential operators, unbounded observables in quantum mechanics, and other cases.

The term "unbounded operator" can be misleading, since

- "unbounded" should sometimes be understood as "not necessarily bounded";
- "operator" should be understood as "linear operator" (as in the case of "bounded operator");
- the domain of the operator is a linear subspace, not necessarily the whole space;
- this linear subspace is not necessarily closed; often (but not always) it is assumed to be dense;
- in the special case of a bounded operator, still, the domain is usually assumed to be the whole space.

In contrast to bounded operators, unbounded operators on a given space do not form an algebra, nor even a linear space, because each one is defined on its own domain.

The term "operator" often means "bounded linear operator", but in the context of this article it means "unbounded operator", with the reservations made above. The given space is assumed to be a Hilbert space. Some generalizations to Banach spaces and more general topological vector spaces are possible.

## Contents

## Short history

The theory of unbounded operators developed in the late 1920s and early 1930s as part of developing a rigorous mathematical framework for quantum mechanics.^{[1]} The theory's development is due to John von Neumann^{[2]} and Marshall Stone.^{[3]} Von Neumann introduced using graphs to analyze unbounded operators in 1936.^{[4]}

## Definitions and basic properties

Let *X*, *Y* be Banach spaces. An **unbounded operator** (or simply *operator*) *T* : *X* → *Y* is a linear map Template:Mvar from a linear subspace *D*(*T*) ⊆ *X* — the domain of Template:Mvar — to the space *Y*.^{[5]} Contrary to the usual convention, Template:Mvar may not be defined on the whole space Template:Mvar. Two operators are equal if they have the common domain and they coincide on that common domain.^{[5]}

An operator Template:Mvar is said to be **closed** if its graph Γ(*T*) is a closed set.^{[6]} (Here, the graph Γ(*T*) is a linear subspace of the direct sum *X* ⊕ *Y*, defined as the set of all pairs (*x*, *Tx*), where Template:Mvar runs over the domain of Template:Mvar ). Explicitly, this means that for every sequence {*x _{n}*} of points from the domain of Template:Mvar such that

*x*→

_{n}*x*and

*Tx*→

_{n}*y*, it holds that Template:Mvar belongs to the domain of Template:Mvar and

*Tx*=

*y*.

^{[6]}The closedness can also be formulated in terms of the

*graph norm*: an operator Template:Mvar is closed if and only if its domain

*D*(

*T*) is a complete space with respect to the norm:

^{[7]}

An operator Template:Mvar is said to be **densely defined** if its domain is dense in Template:Mvar.^{[5]} This also includes operators defined on the entire space Template:Mvar, since the whole space is dense in itself. The denseness of the domain is necessary and sufficient for the existence of the adjoint and the transpose; see the next section.

If *T* : *X* → *Y* is closed, densely defined and continuous on its domain, then it is defined on Template:Mvar.^{[8]}

A densely defined operator Template:Mvar on a Hilbert space Template:Mvar is called **bounded from below** if *T* + *a* is a positive operator for some real number Template:Mvar. That is, ⟨*Tx*|*x*⟩ ≥ −*a* Template:!!*x*Template:!!^{2} for all Template:Mvar in the domain of Template:Mvar.^{[9]} If both Template:Mvar and −*T* are bounded from below then Template:Mvar is bounded.^{[9]}

## Example

Let *C*([0, 1]) denote the space of continuous functions on the interval, and let *C*^{1}([0, 1]) denote the space of continuously differentiable functions. Define the classical differentiation operator {{ safesubst:#invoke:Unsubst||$B=*d*/*dx*}} : *C*^{1}([0, 1]) → *C*([0, 1]) by the usual formula:

Every differentiable function is continuous, so *C*^{1}([0, 1]) ⊆ *C*([0, 1]). Consequently, {{ safesubst:#invoke:Unsubst||$B=*d*/*dx*}} : *C*([0, 1]) → *C*([0, 1]) is a well-defined unbounded operator, with domain *C*^{1}([0, 1]).

This is a linear operator, since a linear combination *a f * + *bg* of two continuously differentiable functions *f* , *g* is also continuously differentiable, and

The operator is not bounded. For example,

satisfy

but

The operator is densely defined, and closed.

The same operator can be treated as an operator *Z* → *Z* for many Banach spaces Template:Mvar and is still not bounded. However, it is bounded as an operator *X* → *Y* for some pairs of Banach spaces *X*, *Y*, and also as operator *Z* → *Z* for some topological vector spaces Template:Mvar. As an example let *I* ⊂ **R** be an open interval and consider

where:

## Adjoint

The adjoint of an unbounded operator can be defined in two equivalent ways. First, it can be defined in a way analogous to how we define the adjoint of a bounded operator. Namely, the adjoint *T* ^{∗} : *H*_{2} → *H*_{1} of Template:Mvar is defined as an operator with the property:

More precisely, *T* ^{∗} is defined in the following way. If Template:Mvar is such that is a continuous linear functional on the domain of Template:Mvar, then, after extending it to the whole space via the Hahn–Banach theorem, we can find a *z* such that

since the dual of a Hilbert space can be identified with the set of linear functionals given by the inner product. For each *y*, *z* is uniquely determined if and only if the linear functional is densely defined; i.e., Template:Mvar is densely defined. Finally, we let *T* ^{∗}*y* = *z*, completing the construction of *T* ^{∗}.^{[10]} Note that *T* ^{∗} exists if and only if Template:Mvar is densely defined.

By definition, the domain of *T* ^{∗} consists of elements Template:Mvar in *H*_{2} such that is continuous on the domain of Template:Mvar. Consequently, the domain of *T* ^{∗} could be anything; it could be trivial (i.e., contains only zero).^{[11]} It may happen that the domain of *T*^{∗} is a closed hyperplane and *T* ^{∗} vanishes everywhere on the domain.^{[12]}^{[13]} Thus, boundedness of *T* ^{∗} on its domain does not imply boundedness of Template:Mvar. On the other hand, if *T* ^{∗} is defined on the whole space then Template:Mvar is bounded on its domain and therefore can be extended by continuity to a bounded operator on the whole space.^{[14]} If the domain of *T* ^{∗} is dense, then it has its adjoint *T* ^{∗∗}.^{[15]} A closed densely defined operator Template:Mvar is bounded if and only if *T* ^{∗} is bounded.^{[16]}

The other equivalent definition of the adjoint can be obtained by noticing a general fact. Define a linear operator Template:Mvar as follows:^{[15]}

Since Template:Mvar is an isometric surjection, it is unitary. We then have: *J*(Γ(*T*))^{⊥} is the graph of some operator Template:Mvar if and only if Template:Mvar is densely defined.^{[17]} A simple calculation shows that this "some" Template:Mvar satisfies:

for every Template:Mvar in the domain of Template:Mvar. Thus, Template:Mvar is the adjoint of Template:Mvar.

It follows immediately from the above definition that the adjoint *T* ^{∗} is closed.^{[15]} In particular, a self-adjoint operator (i.e., *T* = *T* ^{∗}) is closed. An operator Template:Mvar is closed and densely defined if and only if *T* ^{∗∗} = *T*.^{[18]}

Some well-known properties for bounded operators generalize to closed densely defined operators. The kernel of a closed operator is closed. Moreover, the kernel of a closed densely defined operator *T* : *H*_{1} → *H*_{2} coincides with the orthogonal complement of the range of the adjoint. That is,^{[19]}

von Neumann's theorem states that *T* ^{∗}*T* and *TT* ^{∗} are self-adjoint, and that *I* + *T* ^{∗}*T* and *I* + *TT* ^{∗} both have bounded inverses.^{[20]} If *T* ^{∗} has trivial kernel, Template:Mvar has dense range (by the above identity.) Moreover:

- Template:Mvar is surjective if and only if there is a
*K*> 0 such that Template:!!*f*Template:!!_{2}≤*K*Template:!!*T*^{∗}*f*Template:!!_{1}for all*f*in*D*(*T*^{∗}).^{[21]}(This is essentially a variant of the so-called closed range theorem.) In particular, Template:Mvar has closed range if and only if*T*^{∗}has closed range.

In contrast to the bounded case, it is not necessary that we have: (*TS*)^{∗} = *S* ^{∗}*T* ^{∗}, since, for example, it is even possible that (*TS*)^{∗} doesn't exist.{{ safesubst:#invoke:Unsubst||date=__DATE__ |$B=
{{#invoke:Category handler|main}}{{#invoke:Category handler|main}}^{[citation needed]}
}} This is, however, the case if, for example, Template:Mvar is bounded.^{[22]}

A densely defined, closed operator Template:Mvar is called *normal* if it satisfies the following equivalent conditions:^{[23]}

*T*^{∗}*T*=*TT*^{∗};

- the domain of Template:Mvar is equal to the domain of
*T*^{∗}, and Template:!!*Tx*Template:!! = Template:!!*T*^{∗}*x*Template:!! for every Template:Mvar in this domain;

- there exist self-adjoint operators
*A*,*B*such that*T*=*A*+*iB*,*T*^{∗}=*A*–*iB*, and Template:!!*Tx*Template:!!^{2}= Template:!!*Ax*Template:!!^{2}+ Template:!!*Bx*Template:!!^{2}for every Template:Mvar in the domain of Template:Mvar. Every self-adjoint operator is normal.

## Transpose

Let *T* : *B*_{1} → *B*_{2} be an operator between Banach spaces. Then the *transpose* (or *dual*) of *T* is an operator satisfying:

for all *x* in *B*_{1} and *y* in B_{2}^{*}. Here, we used the notation: .^{[24]}

The necessary and sufficient condition for the transpose of *T* to exist is that *T* is densely defined (for essentially the same reason as to adjoints, as discussed above.)

For any Hilbert space *H*, there is the anti-linear isomorphism:

given by *Jf* = *y* where .
Through this isomorphism, the transpose *T*^{'} relates to the adjoint *T*^{∗} in the following way:

where . (For the finite-dimensional case, this corresponds to the fact that the adjoint of a matrix is its conjugate transpose.) Note that this gives the definition of adjoint in terms of a transpose.

## Closed linear operators

Closed linear operators are a class of linear operators on Banach spaces. They are more general than bounded operators, and therefore not necessarily continuous, but they still retain nice enough properties that one can define the spectrum and (with certain assumptions) functional calculus for such operators. Many important linear operators which fail to be bounded turn out to be closed, such as the derivative and a large class of differential operators.

Let *X*, *Y* be two Banach spaces. A linear operator *A* : *D*(*A*) ⊂ *X* → *Y* is **closed** if for every sequence {*x _{n}*} in

*D*(

*A*) converging to Template:Mvar in Template:Mvar such that

*Ax*→

_{n}*y*∈

*Y*as

*n*→ ∞ one has

*x*∈

*D*(

*A*) and

*Ax*=

*y*. Equivalently, Template:Mvar is closed if its graph is closed in the direct sum

*X*⊕

*Y*.

Given a linear operator Template:Mvar, not necessarily closed, if the closure of its graph in *X* ⊕ *Y* happens to be the graph of some operator, that operator is called the **closure** of Template:Mvar, and we say that Template:Mvar is **closable**. Denote the closure of Template:Mvar by Template:Overline. It follows easily that Template:Mvar is the restriction of Template:Overline to *D*(*A*).

A **core** of a closable operator is a subset Template:Mvar of *D*(*A*) such that the closure of the restriction of Template:Mvar to Template:Mvar is Template:Overline.

### Basic Properties

Any closed linear operator defined on the whole space Template:Mvar is bounded. This is the closed graph theorem. Additionally, the following properties are easily checked:

- If Template:Mvar is closed then
*A*−*λI*is closed where Template:Mvar is a scalar and Template:Mvar is the identity function; - If Template:Mvar is closed, then its kernel (or nullspace) is a closed subspace of Template:Mvar;
- If Template:Mvar is closed and injective, then its inverse
*A*^{−1}is also closed; - An operator Template:Mvar admits a closure if and only if for every pair of sequences {
*x*} and {_{n}*y*} in_{n}*D*(*A*) both converging to Template:Mvar, such that both {*Ax*} and {_{n}*Ay*} converge, one has lim_{n}_{n}*Ax*= lim_{n}_{n}*Ay*._{n}

### Example

Consider the derivative operator *A* = {{ safesubst:#invoke:Unsubst||$B=*d*/*dx*}} where *X* = *Y* = *C*([*a*, *b*]) is the Banach space of all continuous functions on an interval [*a*, *b*]. If one takes its domain *D*(*A*) to be *C*^{1}([*a*, *b*]), then Template:Mvar is a closed operator, which is not bounded. On the other hand if *D*(*A*) = *C*^{∞}([*a*, *b*]), then Template:Mvar will no longer be closed, but it will be closable, with the closure being its extension defined on *C*^{1}([*a*, *b*]).

## Symmetric operators and self-adjoint operators

{{#invoke:main|main}}

An operator *T* on a Hilbert space is *symmetric* if and only if for each *x* and *y* in the domain of Template:Mvar we have . A densely defined operator Template:Mvar is symmetric if and only if it agrees with its adjoint *T*^{∗} restricted to the domain of *T*, in other words when *T*^{∗} is an extension of Template:Mvar.^{[26]}

In general, the domain of the adjoint *T*^{∗} need not equal the domain of *T*. If the domain of *T* and the domain of the adjoint coincide, then we say that *T* is *self-adjoint*.^{[27]} Note that, when *T* is self-adjoint, the existence of the adjoint implies that *T* is dense and since *T*^{∗} is necessarily closed, *T* is closed.

A densely defined operator *T* is *symmetric*, if the subspace Γ(*T*) is orthogonal to its image *J*(Γ(*T*)) under *J* (where *J*(*x*,*y*):=(*y*,-*x*)).^{[28]}

Equivalently, an operator *T* is *self-adjoint* if it is densely defined, closed, symmetric, and satisfies the fourth condition: both operators *T* – *i*, *T* + *i* are surjective, that is, map the domain of *T* onto the whole space *H*. In other words: for every *x* in *H* there exist *y* and *z* in the domain of *T* such that *Ty* – *iy* = *x* and *Tz* + *iz* = *x*.^{[29]}

An operator *T* is *self-adjoint*, if the two subspaces Γ(*T*), *J*(Γ(*T*)) are orthogonal and their sum is the whole space ^{[15]}

This approach does not cover non-densely defined closed operators. Non-densely defined symmetric operators can be defined directly or via graphs, but not via adjoint operators.

A symmetric operator is often studied via its Cayley transform.

An operator *T* on a complex Hilbert space is symmetric if and only if its quadratic form is real, that is, the number is real for all *x* in the domain of *T*.^{[26]}

A densely defined closed symmetric operator *T* is self-adjoint if and only if *T*^{∗} is symmetric.^{[30]} It may happen that it is not.^{[31]}^{[32]}

A densely defined operator *T* is called *positive*^{[9]} (or *nonnegative*^{[33]}) if its quadratic form is nonnegative, that is, for all *x* in the domain of *T*. Such operator is necessarily symmetric.

The operator *T*^{∗}*T* is self-adjoint^{[34]} and positive^{[9]} for every densely defined, closed *T*.

The spectral theorem applies to self-adjoint operators ^{[35]} and moreover, to normal operators,^{[36]}^{[37]} but not to densely defined, closed operators in general, since in this case the spectrum can be empty.^{[38]}^{[39]}

A symmetric operator defined everywhere is closed, therefore bounded,^{[6]} which is the Hellinger–Toeplitz theorem.^{[40]}

{{#invoke:see also|seealso}}

By definition, an operator *T* is an *extension* of an operator *S* if Γ(*S*) ⊆ Γ(*T*).^{[41]} An equivalent direct definition: for every *x* in the domain of *S*, *x* belongs to the domain of *T* and *Sx* = *Tx*.^{[5]}^{[41]}

Note that an everywhere defined extension exists for every operator, which is a purely algebraic fact explained at Discontinuous linear map#General existence theorem and based on the axiom of choice. If the given operator is not bounded then the extension is a discontinuous linear map. It is of little use since it cannot preserve important properties of the given operator (see below), and usually is highly non-unique.

An operator *T* is called *closable* if it satisfies the following equivalent conditions:^{[6]}^{[41]}^{[42]}

*T*has a closed extension;- the closure of the graph of
*T*is the graph of some operator; - for every sequence (
*x*) of points from the domain of_{n}*T*such that*x*→ 0 and also_{n}*Tx*→_{n}*y*it holds that*y*= 0.

Not all operators are closable.^{[43]}

A closable operator *T* has the least closed extension called the *closure* of *T*. The closure of the graph of *T* is equal to the graph of ^{[6]}^{[41]}

Other, non-minimal closed extensions may exist.^{[31]}^{[32]}

A densely defined operator *T* is closable if and only if *T*^{∗} is densely defined. In this case and ^{[15]}^{[44]}

If *S* is densely defined and *T* is an extension of *S* then *S*^{∗} is an extension of *T*^{∗}.^{[45]}

Every symmetric operator is closable.^{[46]}

A symmetric operator is called *maximal symmetric* if it has no symmetric extensions, except for itself.^{[26]}

Every self-adjoint operator is maximal symmetric.^{[26]} The converse is wrong.^{[47]}

An operator is called *essentially self-adjoint* if its closure is self-adjoint.^{[46]}

An operator is essentially self-adjoint if and only if it has one and only one self-adjoint extension.^{[30]}

An operator may have more than one self-adjoint extension, and even a continuum of them.^{[32]}

A densely defined, symmetric operator *T* is essentially self-adjoint if and only if both operators *T* – *i*, *T* + *i* have dense range.^{[48]}

Let *T* be a densely defined operator. Denoting the relation "*T* is an extension of *S*" by *S* ⊂ *T* (a conventional abbreviation for Γ(*S*) ⊆ Γ(*T*)) one has the following.^{[49]}

- If
*T*is symmetric then*T*⊂*T*^{∗∗}⊂*T*^{∗}. - If
*T*is closed and symmetric then*T*=*T*^{∗∗}⊂*T*^{∗}. - If
*T*is self-adjoint then*T*=*T*^{∗∗}=*T*^{∗}. - If
*T*is essentially self-adjoint then*T*⊂*T*^{∗∗}=*T*^{∗}.

## Importance of self-adjoint operators

The class of **self-adjoint operators** is especially important in mathematical physics. Every self-adjoint operator is densely defined, closed and symmetric. The converse holds for bounded operators but fails in general. Self-adjointness is substantially more restricting than these three properties. The famous spectral theorem holds for self-adjoint operators. In combination with Stone's theorem on one-parameter unitary groups it shows that self-adjoint operators are precisely the infinitesimal generators of strongly continuous one-parameter unitary groups, see Self-adjoint operator#Self adjoint extensions in quantum mechanics. Such unitary groups are especially important for describing time evolution in classical and quantum mechanics.

## See also

## Notes

- ↑ Template:Harvnb
- ↑ {{#invoke:citation/CS1|citation |CitationClass=citation }}
- ↑ {{#invoke:citation/CS1|citation |CitationClass=book }}
- ↑ {{#invoke:citation/CS1|citation |CitationClass=citation }}
- ↑
^{5.0}^{5.1}^{5.2}^{5.3}Template:Harvnb - ↑
^{6.0}^{6.1}^{6.2}^{6.3}^{6.4}Template:Harvnb - ↑ Template:Harvnb
- ↑ Suppose
*f*is a sequence in the domain of Template:Mvar that converges to_{j}*g*∈*X*. Since Template:Mvar is uniformly continuous on its domain,*Tf*is Cauchy in Template:Mvar. Thus, (_{j}*f*,_{j}*T f*) is Cauchy and so converges to some (_{j}*f*,*T f*) since the graph of Template:Mvar is closed. Hence,*f*=*g*, and the domain of Template:Mvar is closed. - ↑
^{9.0}^{9.1}^{9.2}^{9.3}Template:Harvnb - ↑ Verifying that
*T*^{∗}is linear trivial. - ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Proof: being closed, the everywhere defined
*T*^{∗}is bounded, which implies boundedness of*T*^{∗∗}, the latter being the closure of Template:Mvar. See also Template:Harv for the case of everywhere defined Template:Mvar. - ↑
^{15.0}^{15.1}^{15.2}^{15.3}^{15.4}Template:Harvnb - ↑ Proof: We have:
*T*^{∗∗}=*T*. So, if*T*^{∗}is bounded, then its adjoint Template:Mvar is bounded. - ↑ Template:Harvnb
- ↑ Proof: If Template:Mvar is closed densely defined, then
*T*^{∗}exists and is densely defined. Thus,*T*^{∗∗}exists. The graph of Template:Mvar is dense in the graph of*T*^{∗∗}; hence,*T*=*T*^{∗∗}. Conversely, since the existence of*T*^{∗∗}implies that that of*T*^{∗}, which in turn implies Template:Mvar is densely defined. Since*T*^{∗∗}is closed, Template:Mvar is densely defined and closed. - ↑ Brezis, pp. 28.
- ↑ Yoshida, pp. 200.
- ↑ If Template:Mvar is surjective, then
*T*: (ker*T*)^{⊥}→*H*_{2}has bounded inverse, which we denote by Template:Mvar. The estimate then follows since Conversely, suppose the estimate holds. Since*T*^{∗}has closed range then, we have: ran(*T*) = ran(*TT*^{*}). Since ran(*T*) is dense, it suffices to show that*TT*^{∗}has closed range. If*TT*^{∗}*f*is convergent, then_{j}*f*is convergent by the estimate since Say,_{j}*f*→_{j}*g*. Since*TT*^{∗}is self-adjoint; thus, closed, (von Neumann's theorem),*TT*^{∗}*f*→_{j}*TT*^{∗}*g*. QED - ↑ Yoshida, pp. 195.
- ↑ Template:Harvnb
- ↑ Yoshida, pp. 193.
- ↑ Yoshida, pp. 196.
- ↑
^{26.0}^{26.1}^{26.2}^{26.3}Template:Harvnb - ↑ Template:Harvnb
- ↑ Follows from Template:Harv and the definition via adjoint operators.
- ↑ Template:Harvnb
- ↑
^{30.0}^{30.1}Template:Harvnb - ↑
^{31.0}^{31.1}Template:Harvnb - ↑
^{32.0}^{32.1}^{32.2}Template:Harvnb - ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑
^{41.0}^{41.1}^{41.2}^{41.3}Template:Harvnb - ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑
^{46.0}^{46.1}Template:Harvnb - ↑ Template:Harvnb
- ↑ Template:Harvnb
- ↑ Template:Harvnb

## References

- {{#invoke:citation/CS1|citation

|CitationClass=citation }} (see Chapter 12 "General theory of unbounded operators in Hilbert spaces").

- {{#invoke:citation/CS1|citation

|CitationClass=citation }}

- {{#invoke:citation/CS1|citation

|CitationClass=citation }}

- {{#invoke:citation/CS1|citation

|CitationClass=citation }}

- {{#invoke:citation/CS1|citation

|CitationClass=citation }}

- {{#invoke:citation/CS1|citation

|CitationClass=citation }} (see Chapter 5 "Unbounded operators").

- {{#invoke:citation/CS1|citation

|CitationClass=citation }} (see Chapter 8 "Unbounded operators").

- {{#invoke:citation/CS1|citation

|CitationClass=citation }}

*This article incorporates material from Closed operator on PlanetMath, which is licensed under the Creative Commons Attribution/Share-Alike License.*