Let T : X → Y {\displaystyle T:X\to Y} be a linear operator between Banach spaces (or more generally Fréchet spaces). Then the continuity of T {\displaystyle T} means that T x i → T x {\displaystyle Tx_{i}\to Tx} for each convergent sequence x i → x {\displaystyle x_{i}\to x} . On the other hand, the closedness of the graph of T {\displaystyle T} means that for each convergent sequence x i → x {\displaystyle x_{i}\to x} such that T x i → y {\displaystyle Tx_{i}\to y} , we have y = T x {\displaystyle y=Tx} . Hence, the closed graph theorem says that in order to check the continuity of T {\displaystyle T} , one can show T x i → T x {\displaystyle Tx_{i}\to Tx} under the additional assumption that T x i {\displaystyle Tx_{i}} is convergent.
In fact, for the graph of T to be closed, it is enough that if x i → 0 , T x i → y {\displaystyle x_{i}\to 0,\,Tx_{i}\to y} , then y = 0 {\displaystyle y=0} . Indeed, assuming that condition holds, if ( x i , T x i ) → ( x , y ) {\displaystyle (x_{i},Tx_{i})\to (x,y)} , then x i − x → 0 {\displaystyle x_{i}-x\to 0} and T ( x i − x ) → y − T x {\displaystyle T(x_{i}-x)\to y-Tx} . Thus, y = T x {\displaystyle y=Tx} ; i.e., ( x , y ) {\displaystyle (x,y)} is in the graph of T.
Note, to check the closedness of a graph, it’s not even necessary to use the norm topology: if the graph of T is closed in some topology coarser than the norm topology, then it is closed in the norm topology.1 In practice, this works like this: T is some operator on some function space. One shows T is continuous with respect to the distribution topology; thus, the graph is closed in that topology, which implies closedness in the norm topology and then T is a bounded by the closed graph theorem (when the theorem applies). See § Example for an explicit example.
Theorem—2 If T : X → Y {\displaystyle T:X\to Y} is a linear operator between Banach spaces (or more generally Fréchet spaces), then the following are equivalent:
The usual proof of the closed graph theorem employs the open mapping theorem. It simply uses a general recipe of obtaining the closed graph theorem from the open mapping theorem; see closed graph theorem § Relation to the open mapping theorem (this deduction is formal and does not use linearity; the linearity is needed to appeal to the open mapping theorem which relies on the linearity.)
In fact, the open mapping theorem can in turn be deduced from the closed graph theorem as follows. As noted in Open mapping theorem (functional analysis) § Statement and proof, it is enough to prove the open mapping theorem for a continuous linear operator that is bijective (not just surjective). Let T be such an operator. Then by continuity, the graph Γ T {\displaystyle \Gamma _{T}} of T is closed. Then Γ T ≃ Γ T − 1 {\displaystyle \Gamma _{T}\simeq \Gamma _{T^{-1}}} under ( x , y ) ↦ ( y , x ) {\displaystyle (x,y)\mapsto (y,x)} . Hence, by the closed graph theorem, T − 1 {\displaystyle T^{-1}} is continuous; i.e., T is an open mapping.
Since the closed graph theorem is equivalent to the open mapping theorem, one knows that the theorem fails without the completeness assumption. But more concretely, an operator with closed graph that is not bounded (see unbounded operator) exists and thus serves as a counterexample.
The Hausdorff–Young inequality says that the Fourier transformation ⋅ ^ : L p ( R n ) → L p ′ ( R n ) {\displaystyle {\widehat {\cdot }}:L^{p}(\mathbb {R} ^{n})\to L^{p'}(\mathbb {R} ^{n})} is a well-defined bounded operator with operator norm one when 1 / p + 1 / p ′ = 1 {\displaystyle 1/p+1/p'=1} . This result is usually proved using the Riesz–Thorin interpolation theorem and is highly nontrivial. The closed graph theorem can be used to prove a soft version of this result; i.e., the Fourier transformation is a bounded operator with the unknown operator norm.3
Here is how the argument would go. Let T denote the Fourier transformation. First we show T : L p → Z {\displaystyle T:L^{p}\to Z} is a continuous linear operator for Z = the space of tempered distributions on R n {\displaystyle \mathbb {R} ^{n}} . Second, we note that T maps the space of Schwarz functions to itself (in short, because smoothness and rapid decay transform to rapid decay and smoothness, respectively). This implies that the graph of T is contained in L p × L p ′ {\displaystyle L^{p}\times L^{p'}} and T : L p → L p ′ {\displaystyle T:L^{p}\to L^{p'}} is defined but with unknown bounds. Since T : L p → Z {\displaystyle T:L^{p}\to Z} is continuous, the graph of T : L p → L p ′ {\displaystyle T:L^{p}\to L^{p'}} is closed in the distribution topology; thus in the norm topology. Finally, by the closed graph theorem, T : L p → L p ′ {\displaystyle T:L^{p}\to L^{p'}} is a bounded operator.
The closed graph theorem can be generalized from Banach spaces to more abstract topological vector spaces in the following ways.
Theorem—A linear operator from a barrelled space X {\displaystyle X} to a Fréchet space Y {\displaystyle Y} is continuous if and only if its graph is closed.
There are versions that does not require Y {\displaystyle Y} to be locally convex.
Theorem—A linear map between two F-spaces is continuous if and only if its graph is closed.45
This theorem is restated and extend it with some conditions that can be used to determine if a graph is closed:
Theorem—If T : X → Y {\displaystyle T:X\to Y} is a linear map between two F-spaces, then the following are equivalent:
Every metrizable topological space is pseudometrizable. A pseudometrizable space is metrizable if and only if it is Hausdorff.
Closed Graph Theorem7—Also, a closed linear map from a locally convex ultrabarrelled space into a complete pseudometrizable TVS is continuous.
Closed Graph Theorem—A closed and bounded linear map from a locally convex infrabarreled space into a complete pseudometrizable locally convex space is continuous.8
Theorem9—Suppose that T : X → Y {\displaystyle T:X\to Y} is a linear map whose graph is closed. If X {\displaystyle X} is an inductive limit of Baire TVSs and Y {\displaystyle Y} is a webbed space then T {\displaystyle T} is continuous.
Closed Graph Theorem10—A closed surjective linear map from a complete pseudometrizable TVS onto a locally convex ultrabarrelled space is continuous.
An even more general version of the closed graph theorem is
Theorem11—Suppose that X {\displaystyle X} and Y {\displaystyle Y} are two topological vector spaces (they need not be Hausdorff or locally convex) with the following property:
Under this condition, if T : X → Y {\displaystyle T:X\to Y} is a linear map whose graph is closed then T {\displaystyle T} is continuous.
Main article: Borel Graph Theorem
The Borel graph theorem, proved by L. Schwartz, shows that the closed graph theorem is valid for linear maps defined on and valued in most spaces encountered in analysis.12 Recall that a topological space is called a Polish space if it is a separable complete metrizable space and that a Souslin space is the continuous image of a Polish space. The weak dual of a separable Fréchet space and the strong dual of a separable Fréchet-Montel space are Souslin spaces. Also, the space of distributions and all Lp-spaces over open subsets of Euclidean space as well as many other spaces that occur in analysis are Souslin spaces. The Borel graph theorem states:
Borel Graph Theorem—Let u : X → Y {\displaystyle u:X\to Y} be linear map between two locally convex Hausdorff spaces X {\displaystyle X} and Y . {\displaystyle Y.} If X {\displaystyle X} is the inductive limit of an arbitrary family of Banach spaces, if Y {\displaystyle Y} is a Souslin space, and if the graph of u {\displaystyle u} is a Borel set in X × Y , {\displaystyle X\times Y,} then u {\displaystyle u} is continuous.13
An improvement upon this theorem, proved by A. Martineau, uses K-analytic spaces.
A topological space X {\displaystyle X} is called a K σ δ {\displaystyle K_{\sigma \delta }} if it is the countable intersection of countable unions of compact sets.
A Hausdorff topological space Y {\displaystyle Y} is called K-analytic if it is the continuous image of a K σ δ {\displaystyle K_{\sigma \delta }} space (that is, if there is a K σ δ {\displaystyle K_{\sigma \delta }} space X {\displaystyle X} and a continuous map of X {\displaystyle X} onto Y {\displaystyle Y} ).
Every compact set is K-analytic so that there are non-separable K-analytic spaces. Also, every Polish, Souslin, and reflexive Fréchet space is K-analytic as is the weak dual of a Frechet space. The generalized Borel graph theorem states:
Generalized Borel Graph Theorem14—Let u : X → Y {\displaystyle u:X\to Y} be a linear map between two locally convex Hausdorff spaces X {\displaystyle X} and Y . {\displaystyle Y.} If X {\displaystyle X} is the inductive limit of an arbitrary family of Banach spaces, if Y {\displaystyle Y} is a K-analytic space, and if the graph of u {\displaystyle u} is closed in X × Y , {\displaystyle X\times Y,} then u {\displaystyle u} is continuous.
If F : X → Y {\displaystyle F:X\to Y} is closed linear operator from a Hausdorff locally convex TVS X {\displaystyle X} into a Hausdorff finite-dimensional TVS Y {\displaystyle Y} then F {\displaystyle F} is continuous.15
Notes
Theorem 4 of Tao. NB: The Hausdorffness there is put to ensure the graph of a continuous map is closed. ↩
Vogt 2000, Theorem 1.8. - Vogt, Dietmar (2000), Lectures on Fréchet spaces (PDF) https://www2.math.uni-wuppertal.de/~vogt/vorlesungen/fs.pdf ↩
Tao, Example 3 - Tao, Terence, 245B, Notes 9: The Baire category theorem and its Banach space consequences https://terrytao.wordpress.com/2009/02/01/245b-notes-9-the-baire-category-theorem-and-its-banach-space-consequences/ ↩
Schaefer & Wolff 1999, p. 78. - Schaefer, Helmut H.; Wolff, Manfred P. (1999). Topological Vector Spaces. GTM. Vol. 8 (Second ed.). New York, NY: Springer New York Imprint Springer. ISBN 978-1-4612-7155-0. OCLC 840278135. https://search.worldcat.org/oclc/840278135 ↩
Trèves (2006), p. 173 - Trèves, François (2006) [1967]. Topological Vector Spaces, Distributions and Kernels. Mineola, N.Y.: Dover Publications. ISBN 978-0-486-45352-1. OCLC 853623322. https://search.worldcat.org/oclc/853623322 ↩
Rudin 1991, pp. 50–52. - Rudin, Walter (1991). Functional Analysis. International Series in Pure and Applied Mathematics. Vol. 8 (Second ed.). New York, NY: McGraw-Hill Science/Engineering/Math. ISBN 978-0-07-054236-5. OCLC 21163277. https://archive.org/details/functionalanalys00rudi ↩
Narici & Beckenstein 2011, pp. 474–476. - Narici, Lawrence; Beckenstein, Edward (2011). Topological Vector Spaces. Pure and applied mathematics (Second ed.). Boca Raton, FL: CRC Press. ISBN 978-1584888666. OCLC 144216834. https://search.worldcat.org/oclc/144216834 ↩
Narici & Beckenstein 2011, p. 479-483. - Narici, Lawrence; Beckenstein, Edward (2011). Topological Vector Spaces. Pure and applied mathematics (Second ed.). Boca Raton, FL: CRC Press. ISBN 978-1584888666. OCLC 144216834. https://search.worldcat.org/oclc/144216834 ↩
Trèves 2006, p. 169. - Trèves, François (2006) [1967]. Topological Vector Spaces, Distributions and Kernels. Mineola, N.Y.: Dover Publications. ISBN 978-0-486-45352-1. OCLC 853623322. https://search.worldcat.org/oclc/853623322 ↩
Trèves 2006, p. 549. - Trèves, François (2006) [1967]. Topological Vector Spaces, Distributions and Kernels. Mineola, N.Y.: Dover Publications. ISBN 978-0-486-45352-1. OCLC 853623322. https://search.worldcat.org/oclc/853623322 ↩
Trèves 2006, pp. 557–558. - Trèves, François (2006) [1967]. Topological Vector Spaces, Distributions and Kernels. Mineola, N.Y.: Dover Publications. ISBN 978-0-486-45352-1. OCLC 853623322. https://search.worldcat.org/oclc/853623322 ↩
Narici & Beckenstein 2011, p. 476. - Narici, Lawrence; Beckenstein, Edward (2011). Topological Vector Spaces. Pure and applied mathematics (Second ed.). Boca Raton, FL: CRC Press. ISBN 978-1584888666. OCLC 144216834. https://search.worldcat.org/oclc/144216834 ↩