Open mapping and closed graph theorems

Intuition Beginner

Imagine you have two complete spaces, $X$ and $Y$, and a continuous linear rule $T$ that turns vectors in $X$ into vectors in $Y$. Suppose $T$ uses every vector of $Y$: each $y$ in $Y$ is the image of at least one $x$ in $X$. The open mapping theorem says that this rule cannot crush open neighbourhoods. A small wiggle in $X$ produces a small wiggle in $Y$ in the strong sense that the image of an open ball around the origin contains an open ball around the origin.

A consequence of that promise is what feels at first like magic: if $T$ is a one-to-one continuous linear correspondence between two complete spaces, then the inverse map is also continuous. Continuity for free, in one direction, given completeness in both.

A second consequence dresses the same idea in a different costume. A linear map between complete spaces is automatically continuous as soon as its graph is a closed set in the product space. You never have to check the continuity directly; closure of the graph suffices.

Visual Beginner

Picture two large round disks side by side. The left disk is $X$ and the right disk is $Y$. A continuous linear map $T$ aims arrows from the left to the right, hitting every point in $Y$. Centre a small open ball in $X$ around the origin. Its image is a blob in $Y$ around the origin. The open mapping theorem says this blob is fat enough to fully contain its own open ball around the origin.

In the closed graph picture, draw the product space $X\times Y$. Plot the graph of $T$, the set of pairs $(x,Tx)$. If that plotted curve is a closed set in the product, then $T$ itself is continuous as a map from $X$ to $Y$.

Worked example Beginner

Picture the space of continuous functions on the closed interval from $0$ to $1$ with the supremum norm. This is a Banach space; call it $X$. Consider the operator that doubles a function: $Tf$ is the function whose value at each point is twice the value of $f$ there.

The operator $T$ is a bijection from $X$ to $X$: every continuous function $g$ is the double of $g$ divided by two. Continuity of $T$ is immediate, since doubling the supremum doubles the norm. The open mapping theorem promises something stronger: the inverse $T$ to the minus one is also continuous. And indeed, the inverse is just the halving map, whose continuity is also direct.

The interest of the theorem appears when the operator is more complicated. Consider an operator built from an integral, like sending a function $f$ to the function whose value at $x$ is the integral of $f$ from $0$ to $x$. If you restrict attention to functions whose integral from $0$ to $1$ vanishes, the integral operator becomes a bijection with the space of continuously differentiable functions vanishing at $0$. The inverse map is differentiation, which is bounded with respect to a stronger norm. Completeness of both sides plus the abstract theorem produce the continuity, which a direct calculation would not.

Why this matters Beginner

Three reasons the open mapping and closed graph results appear in every functional analysis course.

First, they give continuity for free. Once you know an operator is a linear bijection between Banach spaces and one direction is continuous, the other direction is too. This converts existence statements about solutions into continuity statements about solution operators, exactly the form needed for stability results in differential equations.

Second, they reduce continuity checks to graph checks. To prove a linear operator is bounded, you would normally need to estimate norms. The closed graph theorem lets you check instead that the operator's graph in the product space is a closed set. For many constructed operators, closure of the graph follows directly from how the operator was built, while a direct norm bound would be painful.

Third, together with Hahn-Banach, they form the three pillars of Banach-space theory: extension of functionals, openness of surjections, automatic continuity from graph closure. Almost every classical result in functional analysis decomposes into uses of these three pillars.

Check your understanding Beginner

Formal definition Intermediate+

Throughout this section, $X$ and $Y$ denote Banach spaces 02.11.04 over the same scalar field $\mathbb{K}$ (real or complex), and $T:X\to Y$ a linear map.

Definition (open map). A map $T:X\to Y$ between topological spaces is open if for every open $U\subset X$, the image $T(U)$ is open in $Y$. Equivalently, for every $x\in X$ and every neighbourhood $U$ of $x$, the image $T(U)$ is a neighbourhood of $T(x)$.

Definition (graph of an operator). The graph of a map $T:X\to Y$ is the subset $$ \Gamma(T) = {(x, T x) : x \in X} \subset X \times Y. $$ The product space $X\times Y$ carries the product norm $\Vert (x,y)\Vert =\Vert x\Vert +\Vert y\Vert$ (or any equivalent norm such as $\max (\Vert x\Vert ,\Vert y\Vert )$); when $X$ and $Y$ are Banach spaces, so is $X\times Y$.

Theorem (open mapping; Banach-Schauder). Let $X,Y$ be Banach spaces and $T:X\to Y$ a bounded linear surjection. Then $T$ is an open map.

Theorem (bounded inverse). Let $X,Y$ be Banach spaces and $T:X\to Y$ a bounded linear bijection. Then $T$ to the minus one is a bounded linear map from $Y$ to $X$.

Theorem (closed graph). Let $X,Y$ be Banach spaces and $T:X\to Y$ a linear map. Then $T$ is bounded if and only if its graph $\Gamma (T)$ is a closed subset of $X\times Y$.

Theorem (Banach-Steinhaus; uniform boundedness principle). Let $X$ be a Banach space, $Y$ a normed space, and $\{T_{\alpha }{\}}_{\alpha \in A}$ a family of bounded linear maps from $X$ to $Y$ such that ${\sup }_{\alpha }\Vert T_{\alpha }x{\Vert }_Y<\infty$ for every $x\in X$. Then ${\sup }_{\alpha }\Vert T_{\alpha }\Vert <\infty$.

These four statements, together with the Hahn-Banach theorem 02.11.02, are the four classical pillars of Banach-space theory. The first three are direct corollaries of one another given completeness; the fourth flows from the same Baire-category mechanism.

Key theorem with proof Intermediate+

We prove the open mapping theorem via the Baire category theorem, then derive the bounded inverse and closed graph as corollaries.

Baire category theorem (statement used). In a complete metric space, the countable union of closed sets with empty interior has empty interior. Equivalently, a countable intersection of dense open sets is dense.

Proof of the open mapping theorem. Let $T:X\to Y$ be a bounded linear surjection between Banach spaces. Write $B_X(r)=\{x\in X:\Vert x\Vert <r\}$ and similarly for $B_Y$. The conclusion is: there exists $\delta >0$ with $B_Y(\delta )\subset T(B_X(1))$.

Step 1 (Baire input). Since $T$ is surjective, $Y={\bigcup }_{n\ge 1}T(B_X(n))={\bigcup }_{n\ge 1}n\cdot T(B_X(1))$. By Baire applied to $Y$, at least one $\overline{T(B_X(n))}$ has non-empty interior. Scaling, $\overline{T(B_X(1))}$ has non-empty interior: there exist $y_0\in Y$ and $r>0$ with $B_Y(y_0,r)\subset \overline{T(B_X(1))}$.

Step 2 (centre at the origin). By symmetry of $T(B_X(1))$ under negation (the ball $B_X(1)$ is symmetric, $T$ is linear), $\overline{T(B_X(1))}$ is also symmetric. So $B_Y(-y_0,r)\subset \overline{T(B_X(1))}$ too. Convexity of $\overline{T(B_X(1))}$ (closure of a convex set is convex) plus the midpoint of $y_0$ and $-y_0$ gives $B_Y(0,r)\subset \overline{T(B_X(2))}$. Scaling, $B_Y(0,r/2)\subset \overline{T(B_X(1))}$.

Step 3 (remove the closure). This is the heart of the argument. Set $\eta =r/2$, so that $$ B_Y(0, \eta) \subset \overline{T(B_X(1))}. $$ We upgrade closure to actual containment: $B_Y(0,\eta /2)\subset T(B_X(1))$.

Fix $y\in B_Y(0,\eta /2)$. By Step 2 with the scaled inclusion $B_Y(0,\eta /2)\subset \overline{T(B_X(1/2))}$, there exists $x_1\in B_X(1/2)$ with $\Vert y-T{x}_1\Vert <\eta /4$. Iterating: $y-T{x}_1\in B_Y(0,\eta /4)\subset \overline{T(B_X(1/4))}$, so there exists $x_2\in B_X(1/4)$ with $\Vert (y-T{x}_1)-T{x}_2\Vert <\eta /8$. Continuing produces a sequence $(x_n)$ with $\Vert x_n\Vert <2^{-n}$ and $$ \left| y - T \left( \sum_{k=1}^n x_k \right) \right| < \eta \cdot 2^{-(n+1)}. $$ The partial sums $s_n={\sum }_{k=1}^n{x}_k$ form a Cauchy sequence in $X$ because ${\sum }_n\Vert x_n\Vert <{\sum }_n{2}^{-n}=1$ converges. Completeness of $X$ gives $s_n\to x\in X$ with $\Vert x\Vert \le 1$. Continuity of $T$ gives $T{s}_n\to Tx$, and the estimate $\Vert y-T{s}_n\Vert <\eta \cdot 2^{-(n+1)}$ forces $Tx=y$. So $y\in T(B_X(1))$, with the slight inequality $\Vert x\Vert \le 1$ replaced by strict inequality after a rescaling argument: pick a slightly smaller ball $B_X(1+\epsilon )$ in the iteration and adjust.

Step 4 (open at the origin implies open everywhere). We have $B_Y(0,\delta )\subset T(B_X(1))$ for $\delta =\eta /2=r/4$. For any open $U\subset X$ and any $x_0\in U$, pick $\rho >0$ with $B_X(x_0,\rho )\subset U$. Then $$ T(U) \supset T(B_X(x_0, \rho)) = T x_0 + \rho \cdot T(B_X(0, 1)) \supset T x_0 + B_Y(0, \rho \delta). $$ So $T(U)$ contains an open ball around every one of its points, hence is open. This completes the proof. $\square$

Bounded inverse from open mapping. If $T:X\to Y$ is a bounded linear bijection, then $T$ is open, so $T$ to the minus one (which exists set-theoretically) is continuous: the preimage under $T$ to the minus one of an open set $U\subset X$ is $T(U)$, which is open by the open mapping theorem. Continuity of a linear map equals boundedness, so $T$ to the minus one is bounded.

Closed graph from bounded inverse. Suppose $T:X\to Y$ is linear with closed graph $\Gamma (T)\subset X\times Y$. Then $\Gamma (T)$ is a closed subspace of the Banach space $X\times Y$, hence itself a Banach space under the inherited norm. Consider the two projection maps $$ \pi_X : \Gamma(T) \to X, \quad (x, T x) \mapsto x, \qquad \pi_Y : \Gamma(T) \to Y, \quad (x, T x) \mapsto T x. $$ Both are bounded linear maps (the projections have norm at most $1$). The map ${\pi }_X$ is a bijection. By the bounded inverse theorem, ${\pi }_X$ to the minus one is bounded. Then $T={\pi }_Y\circ {\pi }_X$ to the minus one is a composition of bounded maps, hence bounded.

For the reverse direction, a bounded linear map has closed graph automatically: if $(x_n,T{x}_n)\to (x,y)$ in $X\times Y$, then $x_n\to x$ and continuity of $T$ gives $T{x}_n\to Tx$; uniqueness of limits forces $y=Tx$, so $(x,y)\in \Gamma (T)$.

Application: two complete norms are equivalent or incomparable. Suppose $\Vert \cdot {\Vert }_1$ and $\Vert \cdot {\Vert }_2$ are two norms on the same vector space $X$ that both make $X$ complete, and suppose $\Vert x{\Vert }_2\le C\Vert x{\Vert }_1$ for some constant $C$ and all $x$. Then the identity map from $(X,\Vert \cdot {\Vert }_1)$ to $(X,\Vert \cdot {\Vert }_2)$ is a bounded linear bijection between Banach spaces. By the bounded inverse theorem, the identity in the other direction is also bounded: $\Vert x{\Vert }_1\le C^{\prime }\Vert x{\Vert }_2$ for some constant $C^{\prime }$. So the two norms are equivalent.

The contrapositive is striking: any two complete norms on the same space are either fully equivalent (mutually bounded with constants) or fully incomparable (neither dominates the other). There is no in-between, no "one direction only" possibility. This is the rigidity that the open mapping theorem buys.

Bridge. With Hahn-Banach 02.11.02 supplying duals, the open mapping family supplies inverses. The Banach-Steinhaus principle below adds equicontinuity from pointwise bounds. Together, these are the three pillars on which the spectral theory of compact operators 02.11.12 pending, the weak-topology machinery 02.11.11 pending, and the existence theory for linear PDE all rest.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: partial — Mathlib supplies the open mapping theorem as ContinuousLinearMap.isOpenMap_of_surjective and the closed graph theorem via LinearMap.continuous_of_isClosed_graph. The companion module records curriculum-facing aliases so downstream units can cite stable handles. The Banach-Steinhaus principle lives in Mathlib as banach_steinhaus in the file Mathlib.Analysis.NormedSpace.Banach.

import Mathlib.Analysis.NormedSpace.Banach
import Mathlib.Analysis.NormedSpace.BanachSteinhaus

namespace Codex.Analysis.FunctionalAnalysis

variable (E F : Type*)
variable [NormedAddCommGroup E] [NormedSpace ℝ E] [CompleteSpace E]
variable [NormedAddCommGroup F] [NormedSpace ℝ F] [CompleteSpace F]

/-- Open mapping theorem (Banach-Schauder): a surjective continuous linear map
between Banach spaces is an open map. -/
abbrev OpenMappingTheorem : Prop :=
  ∀ (T : E →L[ℝ] F), Function.Surjective T → IsOpenMap T

/-- Bounded inverse theorem: a bijective continuous linear map between Banach
spaces has continuous inverse. -/
abbrev BoundedInverseTheorem : Prop :=
  ∀ (T : E →L[ℝ] F) (hT : Function.Bijective T),
    Continuous (Function.invFun T)

/-- Closed graph theorem: a linear map between Banach spaces is continuous
iff its graph is closed. -/
abbrev ClosedGraphTheorem : Prop :=
  ∀ (T : E →ₗ[ℝ] F), IsClosed { p : E × F | p.2 = T p.1 } → Continuous T

end Codex.Analysis.FunctionalAnalysis

Baire category theorem Master

Every proof in this unit channels through the Baire category theorem, the topological fact that drives the whole apparatus.

Theorem (Baire category, complete metric form). Let $(M,d)$ be a complete metric space. Then $M$ is not a countable union of closed sets with empty interior. Equivalently, a countable intersection of dense open sets in $M$ is dense.

Proof. Suppose $M={\bigcup }_n{F}_n$ with each $F_n$ closed and having empty interior, aiming for contradiction. Equivalently, the open sets $U_n=M\setminus F_n$ are dense, and we show that ${\bigcap }_n{U}_n$ is non-empty (in fact dense, but non-emptiness contradicts ${\bigcup }_n{F}_n=M$).

Pick any $x_0\in M$ and $r_0>0$. Density of $U_1$ in $M$ supplies $x_1\in U_1\cap B(x_0,r_0)$. Since $U_1$ is open, there is $r_1\in (0,r_0/2)$ with $\overline{B(x_1,r_1)}\subset U_1\cap B(x_0,r_0)$. Inductively, density of $U_{n+1}$ in $\overline{B(x_n,r_n)}$ supplies $x_{n+1}$ and $r_{n+1}\in (0,r_n/2)$ with $\overline{B(x_{n+1},r_{n+1})}\subset U_{n+1}\cap B(x_n,r_n)$.

The radii satisfy $r_n<r_0/2^n$, so the sequence $(x_n)$ is Cauchy: $d(x_n,x_m)<r_{\min (n,m)}<r_0/2^{\min (n,m)}$. Completeness yields $x_{\infty }=\lim x_n\in M$. The closed-ball construction forces $x_{\infty }\in \overline{B(x_n,r_n)}\subset U_n$ for every $n$. So $x_{\infty }\in {\bigcap }_n{U}_n$, contradicting the supposed coverage by the $F_n$. $\square$

Locally compact Hausdorff form. A locally compact Hausdorff space is also a Baire space: the same nested-non-empty-compact construction works without a metric, using a base of relatively compact open neighbourhoods. This is the form that drives applications in $C^{\ast }$-algebra and harmonic analysis on locally compact groups.

Failure of Baire on incomplete spaces. The rationals $\mathbb{Q}$ with the subspace metric are a counterexample: $\mathbb{Q}={\bigcup }_{q\in \mathbb{Q}}\{q\}$ writes a complete-looking metric space as a countable union of closed singletons, each with empty interior. Completeness is essential. So is the use of countable unions: any space is the union of uncountably many singletons.

Generic-point philosophy. Baire's theorem is the foundation of the "generic point" style of proof. A property holds at a generic point of a complete metric space if it holds on a dense $G_{\delta }$ set, the countable intersection of dense open sets. Examples: a generic continuous function on $[0,1]$ is nowhere differentiable (Banach 1931, Mazurkiewicz 1931); a generic bounded sequence has no Cesaro mean; a generic compact subset of $[0,1]$ has Lebesgue measure zero. Each generic-point statement decomposes into showing the relevant set is a countable intersection of dense open sets, then invoking Baire.

Counterexamples and the role of completeness Master

The three theorems of this unit fail outside the Banach setting. The counterexamples sharpen the working analyst's instinct for when completeness is genuinely load-bearing.

(N1) Open mapping fails without completeness of the codomain. Let $X=c_{00}$, the space of finitely supported real sequences with the supremum norm, and let $Y=c_{00}$ with the ${\ell }^1$ norm. The identity map $X\to Y$ is a continuous linear bijection with respect to the inequality $\Vert x{\Vert }_1\ge \Vert x{\Vert }_{\infty }$, but its inverse is not continuous: the sequence $e_1+e_2+\cdots +e_n$ has ${\ell }^{\infty }$ norm $1$ and ${\ell }^1$ norm $n$. Neither $X$ nor $Y$ is complete; both completions repair to $c_0$ and ${\ell }^1$ respectively, which are no longer in bijection by the identity.

(N2) Closed graph fails without completeness of the domain. Let $X$ be the space of polynomials on $[0,1]$ with the supremum norm, $Y=X$, and $T:X\to Y$ be differentiation, $Tp=p^{\prime }$. The map $T$ is linear with everywhere-defined domain. Its graph is closed: if $p_n\to p$ uniformly and $p_n^{\prime }\to q$ uniformly, then $p$ is continuously differentiable with $p^{\prime }=q$, so $(p,q)\in \Gamma (T)$. But $T$ is not bounded: $p(x)=x^n$ has supremum norm $1$ but $p^{\prime }=n{x}^{n-1}$ has supremum norm $n$. The domain $X$ is not complete (its completion is $C([0,1])$, where $T$ is no longer everywhere defined), so closed graph fails.

(N3) Bounded inverse fails without completeness. Take $X$ to be the space of continuously differentiable functions on $[0,1]$ with the supremum norm, and define $T:X\to X$ by $Tf(x)={\int }_0^{x}f(t)dt$. The image of $T$ consists of continuously differentiable functions vanishing at $0$, a proper subspace. Restricting codomain, $T$ becomes a continuous linear bijection between $X$ and a subspace of $X$. The inverse is differentiation, which is unbounded with respect to the supremum norm. Both spaces are incomplete; the closed-graph or open-mapping conclusion fails accordingly.

(N4) Pathological complete norms. On a Hamel basis $\{e_i{\}}_{i\in I}$ of an infinite-dimensional Banach space $X$, choose any rescaling $\{c_i{e}_i\}$ with $c_i>0$ unbounded. The rescaled norm $\Vert x{\Vert }^{\prime }={\sup }_i|c_i⟨x,e_i^{\ast }⟩|$ may fail to be complete and may be incomparable to the original norm. Two complete norms on $X$ in general bijection by the identity are equivalent by the bounded inverse theorem, but two arbitrary norms with no completeness assumption need not be.

(N5) Baire's role in the closed graph proof is essential. A linear map between two incomplete normed spaces can have closed graph without being bounded (see (N2)), but the closure of the graph is not a Banach space in the inherited norm. The closed-graph proof of boundedness invokes the bounded inverse theorem on the projection ${\pi }_X:\Gamma (T)\to X$, which is a continuous bijection between Banach spaces only when both $X$ and $\Gamma (T)$ are complete. Closure of $\Gamma (T)$ in $X\times Y$ inherits completeness when $X\times Y$ is complete, namely when $X$ and $Y$ both are.

The pattern is uniform: Baire requires completeness, completeness propagates through closed subspaces and finite products, and the three theorems propagate through that completeness. Drop completeness on either side, and you can construct counterexamples by hand from the function spaces of classical analysis.

Banach-Steinhaus / uniform boundedness Master

The fourth pillar deserves its own statement and proof. Where the open mapping theorem turns a single surjective map into an open map, Banach-Steinhaus turns a family of pointwise bounded maps into a uniformly bounded family.

Theorem (Banach-Steinhaus; uniform boundedness principle). Let $X$ be a Banach space, $Y$ a normed space, and $\{T_{\alpha }{\}}_{\alpha \in A}$ a (possibly uncountable) family of bounded linear maps $X\to Y$. Suppose ${\sup }_{\alpha }\Vert T_{\alpha }x\Vert <\infty$ for every $x\in X$. Then ${\sup }_{\alpha }\Vert T_{\alpha }\Vert <\infty$.

Proof (via Baire). Define $E_n=\{x\in X:{\sup }_{\alpha }\Vert T_{\alpha }x\Vert \le n\}$. Each $E_n$ is closed because $E_n={\bigcap }_{\alpha }\{x:\Vert T_{\alpha }x\Vert \le n\}$ and each $\{x:\Vert T_{\alpha }x\Vert \le n\}$ is the preimage of $[0,n]$ under the continuous map $x\mapsto \Vert T_{\alpha }x\Vert$. By hypothesis $X={\bigcup }_n{E}_n$. Baire applied to the Banach space $X$ forces some $E_N$ to have non-empty interior: there are $x_0\in X$ and $r>0$ with $B(x_0,r)\subset E_N$.

For any $z\in B(0,r)$ and any $\alpha$, $z=(x_0+z)-x_0$ gives $$ |T_\alpha z| \leq |T_\alpha(x_0 + z)| + |T_\alpha x_0| \leq N + N = 2N. $$ Scaling: for arbitrary $x\ne 0$, $rx/(2\Vert x\Vert )\in B(0,r)$, so $$ |T_\alpha x| = \frac{2 |x|}{r} \cdot |T_\alpha(r x / (2 |x|))| \leq \frac{2 |x|}{r} \cdot 2N = \frac{4N}{r} |x|. $$ So $\Vert T_{\alpha }\Vert \le 4N/r$ for every $\alpha$. $\square$

Contrapositive (condensation of singularities). If ${\sup }_{\alpha }\Vert T_{\alpha }\Vert =\infty$, then there exists $x\in X$ with ${\sup }_{\alpha }\Vert T_{\alpha }x\Vert =\infty$. In fact, the set of such $x$ is a dense $G_{\delta }$ subset of $X$, hence generic in the Baire sense.

Corollary (continuity of pointwise limits of bounded operators). Let $X,Y$ be Banach spaces and $(T_n)$ a sequence of bounded linear maps $X\to Y$ with $T_{n}x\to Tx$ for every $x\in X$. Then $T$ is a bounded linear map. Linearity is preserved by pointwise convergence. For boundedness, Banach-Steinhaus applied to $\{T_n\}$ gives ${\sup }_n\Vert T_n\Vert =M<\infty$, and $\Vert Tx\Vert ={\lim }_n\Vert T_{n}x\Vert \le M\Vert x\Vert$.

The corollary is the basis for the standard construction of operators by approximation: build a sequence of finite-rank or otherwise simple operators that converges pointwise, and Banach-Steinhaus guarantees the limit is bounded with norm at most the supremum of the approximating norms.

Application: convergence of Fourier series Master

The historically dramatic application of Banach-Steinhaus is the proof that there exist continuous periodic functions whose Fourier series diverges at a given point. This was a stunning negative result when Du Bois-Reymond first constructed an explicit example in 1873; Banach-Steinhaus gives an abstract proof that not only is there one such function, but the set of "bad" continuous functions is generic.

Setup. Let $C(\mathbb{T})$ denote the Banach space of continuous functions on the circle $\mathbb{T}=\mathbb{R}/2\pi \mathbb{Z}$ with the supremum norm. For $f\in C(\mathbb{T})$ and $n\ge 0$, the $n$-th partial sum of the Fourier series at the point $0$ is $$ (S_n f)(0) = \sum_{k = -n}^{n} \hat f(k) = \frac{1}{2 \pi} \int_{-\pi}^{\pi} f(t) D_n(t) dt $$ where $D_n$ is the Dirichlet kernel $D_n(t)={\sum }_{k=-n}^n{e}^{ikt}=\sin ((n+1/2)t)/\sin (t/2)$. The evaluation $f\mapsto (S_{n}f)(0)$ is a bounded linear functional ${\Lambda }_n:C(\mathbb{T})\to \mathbb{C}$ with norm equal to the $L^1$-norm of $D_n/(2\pi )$, namely the Lebesgue constant $L_n=\Vert D_n{\Vert }_{L^1}/(2\pi )$.

Lemma (growth of Lebesgue constants). $L_n\sim (4/{\pi }^2)\log n$ as $n\to \infty$. In particular, ${\sup }_n{L}_n=\infty$.

Du Bois-Reymond theorem. There exists $f\in C(\mathbb{T})$ such that the partial sums $(S_{n}f)(0)$ do not converge.

Proof via Banach-Steinhaus. Suppose for contradiction that $(S_{n}f)(0)$ converges for every $f\in C(\mathbb{T})$. Then ${\sup }_n|{\Lambda }_{n}f|<\infty$ for every $f$. By Banach-Steinhaus, ${\sup }_n\Vert {\Lambda }_n\Vert ={\sup }_n{L}_n<\infty$, contradicting the lemma. So there must exist $f$ for which ${\sup }_n|{\Lambda }_{n}f|=\infty$, hence the sequence does not converge.

Genericity. The contrapositive of Banach-Steinhaus is the condensation of singularities principle: the set of $f\in C(\mathbb{T})$ whose Fourier series fails to converge at $0$ is the complement of a meagre set, hence a generic point in the Baire sense. Iterating over a countable dense subset of $\mathbb{T}$, the set of $f$ whose Fourier series fails to converge on a dense set of points is also generic. Pointwise convergence of Fourier series for continuous functions is a structurally rare phenomenon, not a typical one.

Carleson's positive result. Carleson 1966 proved the deep theorem that the Fourier series of an $L^2$ function converges almost everywhere. The Banach-Steinhaus failure for continuous functions at a fixed point and the Carleson convergence almost everywhere are not contradictory: "everywhere convergence at a fixed point for every continuous function" and "almost-everywhere convergence for $L^2$ functions" are distinct statements with different quantifier structures.

Closability and unbounded operators Master

The closed graph theorem says a linear operator with closed graph between Banach spaces is bounded. Many natural operators in analysis are unbounded (differentiation on $C([0,1])$, multiplication by $x$ on $L^2$, the Laplacian on $L^2$), so their graphs cannot be closed. The fix is the notion of a closable operator: an operator whose graph closure is still the graph of a function.

Definition (closability). A linear operator $T:D(T)\subset X\to Y$ with domain $D(T)\subset X$ a linear subspace is closable if the closure $\overline{\Gamma (T)}$ in $X\times Y$ is the graph of a function. Equivalently, whenever $(x_n)\subset D(T)$ with $x_n\to 0$ and $T{x}_n\to y$, then $y=0$.

If $T$ is closable, the closure $\overline{T}$ is the operator whose graph is $\overline{\Gamma (T)}$. It is the smallest closed extension of $T$.

Example: differentiation on $L^2$. Let $X=Y=L^2([0,1])$ and $Tf=f^{\prime }$ defined on $D(T)=C^{\infty }([0,1])$. $T$ is closable: if $f_n\to 0$ in $L^2$ and $f_n^{\prime }\to g$ in $L^2$, then for any $\phi \in C_c^{\infty }((0,1))$, $\int g\phi =\lim \int f_n^{\prime }\phi =-\lim \int f_n{\phi }^{\prime }=0$. So $g=0$ almost everywhere, hence $g=0$ in $L^2$. The closure $\overline{T}$ has domain the Sobolev space $H^1([0,1])$ 02.13.09 pending, with $\overline{T}f=f^{\prime }$ (weak derivative).

Example: multiplication by $x$ on $L^2(\mathbb{R})$. Let $Tf(x)=xf(x)$ defined on $D(T)=\{f\in L^2:xf\in L^2\}$. $T$ is already closed: if $f_n\to f$ and $x{f}_n\to g$ in $L^2$, then along a subsequence $f_n\to f$ and $x{f}_n\to g$ almost everywhere, so $g(x)=xf(x)$ almost everywhere, hence $f\in D(T)$ and $Tf=g$. $T$ is unbounded because $\Vert Tf\Vert$ can be arbitrarily large for $\Vert f\Vert =1$ (concentrate $f$ near large $|x|$).

The closed graph theorem flips for unbounded operators. For an unbounded operator $T:D(T)\to Y$ with $D(T)⊊X$ a proper subspace, closedness of $\Gamma (T)$ in $X\times Y$ is the substitute for boundedness: it gives enough structural rigidity to develop spectral theory, but the operator itself cannot be extended to all of $X$ while remaining a function. This is the Hellinger-Toeplitz phenomenon below.

Hellinger-Toeplitz theorem Master

A celebrated corollary of the closed graph theorem rules out a class of "good" operators in quantum mechanics: everywhere-defined symmetric operators on a Hilbert space are automatically bounded.

Theorem (Hellinger-Toeplitz, 1910). Let $\mathcal{H}$ be a Hilbert space 02.11.08 and $T:\mathcal{H}\to \mathcal{H}$ a linear map (everywhere defined) satisfying $⟨Tx,y⟩=⟨x,Ty⟩$ for all $x,y\in \mathcal{H}$. Then $T$ is bounded.

Proof via closed graph. We show $\Gamma (T)$ is closed. Suppose $x_n\to x$ and $T{x}_n\to y$ in $\mathcal{H}$. For any $z\in \mathcal{H}$, $$ \langle y, z \rangle = \lim_n \langle T x_n, z \rangle = \lim_n \langle x_n, T z \rangle = \langle x, T z \rangle = \langle T x, z \rangle. $$ This holds for every $z$, so $y=Tx$. The graph is closed; the closed graph theorem (Hilbert spaces are Banach) gives boundedness. $\square$

Quantum-mechanical implication. Position $\hat{X}$, momentum $\hat{P}$, and the Hamiltonian $\hat{H}$ of a quantum particle are all symmetric (more precisely, self-adjoint) and unbounded on $L^2(\mathbb{R})$. Hellinger-Toeplitz forces these operators to be undefined on some vectors in $L^2(\mathbb{R})$: if they were everywhere-defined, they would be bounded, which they are not. So the rigorous formulation of quantum mechanics distinguishes domains: $\hat{X}$ is defined on $\{\psi :x\psi \in L^2\}$, $\hat{P}$ on the Sobolev space $H^1(\mathbb{R})$, and $\hat{H}={\hat{P}}^2/2m+V(\hat{X})$ on a more restrictive domain depending on $V$. Domain issues are not pedantic; they are forced by Hellinger-Toeplitz.

Generalisation to self-adjoint operators. The same closed-graph argument shows that any everywhere-defined operator $T$ on $\mathcal{H}$ whose adjoint $T^{\ast }$ is also everywhere defined must be bounded. The hypothesis of Hellinger-Toeplitz, symmetry, gives $T=T^{\ast }$ as operators with the same domain, so $T^{\ast }$ is everywhere defined.

Sobolev embeddings and closed graph Master

The closed graph theorem provides a clean route to many embedding theorems in functional analysis where a direct estimate would require substantial calculation.

Setup. Let $\Omega \subset {\mathbb{R}}^n$ be a bounded domain with $C^1$ boundary, and let $H^s(\Omega )$ denote the $L^2$-Sobolev space of functions with $s$ weak derivatives in $L^2$ 02.13.09 pending. The Sobolev embedding theorem says that for $s>n/2$, $H^s(\Omega )\subset C^0(\overline{\Omega })$ continuously: there is a constant $C$ with ${\sup }_{\Omega }|f|\le C\Vert f{\Vert }_{H^s}$ for $f\in H^s(\Omega )\cap C^{\infty }$.

Embedding via closed graph. Once you know set-theoretically that every $f\in H^s$ has a continuous representative, the embedding $H^s\hookrightarrow C^0$ becomes a linear map between Banach spaces. To check it is continuous (which is what "embedding" really means), the closed graph theorem suffices: if $f_n\to f$ in $H^s$ and $f_n\to g$ in $C^0$, then both convergences are in $L^2$ (since $C^0$ embeds in $L^2$ on a bounded domain), so $f=g$ in $L^2$, hence pointwise almost everywhere. Both are continuous, so $f=g$ pointwise. The graph is closed; continuity follows.

This style of argument bypasses the technical Sobolev inequality $\Vert f{\Vert }_{\infty }\le C\Vert f{\Vert }_{H^s}$ for the abstract embedding statement, although the explicit constant $C$ still requires direct estimation. Most embedding theorems in PDE follow this two-step pattern: prove set-theoretic inclusion by hand, then deduce continuity from the closed graph theorem.

Rellich-Kondrachov via closed graph. The Rellich-Kondrachov theorem says the embedding $H^1(\Omega )\hookrightarrow L^2(\Omega )$ is compact for bounded $\Omega$. Closed graph plus the abstract observation that a continuous linear map between Banach spaces is compact iff it sends bounded sets to relatively compact sets reduces the proof to the equicontinuity calculation in the unit ball of $H^1$, which is the Frechet-Kolmogorov criterion applied to Sobolev derivatives.

Trace theorems. The trace map $\gamma :H^1(\Omega )\to H^{1/2}(\partial \Omega )$, sending a Sobolev function to its boundary values, is continuous. Set-theoretically, the trace of a Sobolev function exists by a density argument; continuity follows from the closed graph theorem. Direct estimates are available (Adams-Fournier) but the abstract approach delivers the qualitative result without them.

F-spaces and Frechet generalisation Master

The classical open mapping, closed graph, and Banach-Steinhaus theorems extend from Banach spaces to F-spaces (complete metrisable topological vector spaces) and to Frechet spaces (locally convex F-spaces), the natural setting where the Baire-category proof still goes through.

Definition (F-space). A topological vector space $X$ is an F-space if its topology is induced by a complete translation-invariant metric. F-spaces are exactly the complete metrisable topological vector spaces.

Definition (Frechet space). A Frechet space is a locally convex F-space, equivalently a complete metrisable topological vector space whose topology is generated by a countable family of seminorms $\{p_n\}$ separating points.

Open mapping for F-spaces. A continuous linear surjection between two F-spaces is open. The proof is identical to the Banach case, replacing "open ball" with "neighbourhood of zero" and using completeness of the codomain for the iteration step.

Closed graph for F-spaces. A linear map between two F-spaces is continuous iff its graph is closed. Same proof: the graph closure is itself an F-space, and the projection-bijection-inverse argument goes through.

Banach-Steinhaus for F-spaces. A pointwise bounded family of continuous linear maps from an F-space to a topological vector space is equicontinuous. Same Baire-based proof, with "equicontinuous" replacing "uniformly bounded in norm" because the latter does not make sense without a norm.

Examples of Frechet spaces.

• $C^{\infty }(\Omega )$ for an open $\Omega \subset {\mathbb{R}}^n$: smooth functions with the topology of uniform convergence of all derivatives on compact subsets. Seminorms $p_{n,K}(f)={\sup }_K|{\partial }^{\alpha }f|$ for multi-index $\alpha$ with $|\alpha |\le n$ and compact $K\subset \Omega$.
• The Schwartz space $\mathcal{S}({\mathbb{R}}^n)$: smooth functions all of whose derivatives decay faster than any polynomial. Seminorms $p_{n,m}(f)={\sup }_x(1+|x{|}^2{)}^{n/2}|{\partial }^{\alpha }f|$ for $|\alpha |\le m$.
• $H(\Omega )$ for an open $\Omega \subset \mathbb{C}$: holomorphic functions on $\Omega$ with uniform convergence on compact subsets. Cauchy estimates show this is the same as uniform convergence of all derivatives.

Distributions as duals. The dual space ${\mathcal{S}}^{\prime }({\mathbb{R}}^n)$ of tempered distributions and the dual ${\mathcal{D}}^{\prime }(\Omega )$ of general distributions are not themselves Frechet but are the continuous duals of Frechet (respectively LF) spaces. The Banach-Steinhaus theorem in the Frechet setting is what allows pointwise limits of distributions to remain distributions, the foundation of the theory of distributional convergence used in PDE and microlocal analysis.

LF-spaces. The test-function space $\mathcal{D}(\Omega )=C_c^{\infty }(\Omega )$ is not Frechet (it cannot be metrised by a single sequence of seminorms in a useful way) but is an LF-space, the strict inductive limit of an increasing sequence of Frechet spaces ${\mathcal{D}}_K(\Omega )=\{f\in C_c^{\infty }:\mathrm{supp}f\subset K\}$. The open mapping and closed graph theorems extend to LF-spaces with appropriate hypotheses (Komatsu 1967, Schwartz's original distribution theory).

The pattern is consistent: the Baire-based theorems generalise as far as Baire generalises, which is to complete metric spaces and locally compact Hausdorff spaces. Beyond these, more delicate analogues exist (Pták's open mapping theorem in non-metrisable settings) but require additional structural hypotheses.

Connections Master

• Hahn-Banach theorem 02.11.02 — Hahn-Banach builds dual spaces; the open mapping theorem builds inverses. Together with Banach-Steinhaus, they form the three pillars on which Banach-space theory rests.

• Banach spaces 02.11.04 — completeness is the essential hypothesis. Every theorem of this unit fails without it, and the counterexamples N1-N5 illustrate the failure in concrete function spaces.

• Normed vector space 02.11.06 — the natural setting for the statements; completeness elevates a normed space to the Banach setting where the theorems apply.

• Bounded linear operators 02.11.01 — the open mapping theorem is a statement about a surjective bounded operator; the closed graph theorem characterises bounded operators by graph closure; Banach-Steinhaus bounds families of bounded operators.

• Baire category theorem — every theorem of this unit derives from Baire applied to a Banach space or to a closed subspace thereof. The Baire theorem is the topological engine; the Banach-space results are the analytic payoff.

• Banach-Steinhaus / uniform boundedness 02.11.10 pending — the fourth pillar, proved here via the same Baire mechanism and applied in the Fourier-series example.

• Weak topologies and reflexivity 02.11.11 pending — Banach-Alaoglu and Goldstine work alongside the Baire-based theorems; weak compactness arguments often use Banach-Steinhaus to convert pointwise convergence into uniform bounds.

• Spectral theory of compact operators 02.11.12 pending — the Riesz-Schauder theory of compact operators uses the closed graph theorem to identify the spectrum and the bounded inverse theorem to construct resolvents.

• Sobolev spaces and PDE 02.13.09 pending — Sobolev embeddings $H^s\hookrightarrow C^k$ and trace theorems $\gamma :H^1\to H^{1/2}(\partial \Omega )$ rely on closed graph for the continuity step; Rellich-Kondrachov compactness uses closed graph plus equicontinuity.

• Unbounded operators in quantum mechanics — Hellinger-Toeplitz forces self-adjoint operators (position, momentum, Hamiltonian) to be defined on dense subspaces rather than everywhere. Closed graph plus closability is the technical foundation of the spectral theorem 02.11.03.

Historical and philosophical context Master

The open mapping theorem and its corollaries were established in the early years of functional analysis as the discipline crystallised in interwar Poland under Stefan Banach and Juliusz Schauder. Banach's 1922 thesis "Sur les operations dans les ensembles abstraits et leur application aux equations integrales" introduced the abstract framework of normed complete vector spaces (later named after him) and proved early forms of fixed-point theorems in the setting. The open mapping theorem itself first appeared in Banach's 1929 paper and was strengthened and given its modern statement in his collaboration with Schauder, published in Comptes Rendus in 1932 and consolidated in Banach's 1932 monograph "Theorie des operations lineaires" ^{[Brezis §2]}.

Schauder's 1930 paper "Uber die Umkehrung linearer, stetiger Funktionaloperationen" gave the bounded inverse theorem the form now standard. The closed graph theorem followed quickly as a corollary. Hugo Steinhaus had already proved in 1919 a version of the uniform boundedness principle for sequences of linear functionals, and Banach and Steinhaus published the general principle in 1927 in Fundamenta Mathematicae; this became the fourth pillar joining Hahn-Banach and the open mapping family.

The unifying engine throughout is the Baire category theorem, proved by Rene Baire in 1899 in his thesis on real functions. Baire's original motivation was the classification of functions on the real line into transfinite categories (now the Baire hierarchy in descriptive set theory). The application to functional analysis came two decades later, when Banach and Steinhaus recognised that pointwise boundedness combined with Baire's theorem gives uniform boundedness essentially for free. The structural insight, that completeness plus countable union forces somewhere-fatness of one piece, was the conceptual breakthrough.

Hellinger and Toeplitz proved their theorem in 1910 in the context of bilinear forms on infinite sequences, before the abstract Hilbert-space framework existed. Their original proof was direct; the closed-graph proof presented here is a much later abstraction (probably von Neumann or Stone in the early 1930s) that simplifies the argument enormously and clarifies why the result is forced. The quantum-mechanical interpretation — that observables like position and momentum cannot be everywhere-defined operators — was articulated by John von Neumann in his 1932 monograph "Mathematische Grundlagen der Quantenmechanik," which formalised quantum mechanics on a Hilbert space and made domain questions for unbounded self-adjoint operators a central theme.

Philosophically, the three theorems of this unit (plus Banach-Steinhaus) illustrate a recurring pattern in twentieth-century analysis: completeness as a structural hypothesis converts qualitative statements (existence) into quantitative ones (estimates). The Banach setting is the minimum framework in which this conversion is automatic. Outside the Banach setting, the same conversions are possible but require explicit construction. The theorems are also a paradigm of "soft analysis": no explicit constants appear, no estimates are written down, and yet quantitative information (boundedness of inverses, uniform boundedness of families, continuity of limits) is extracted from purely topological and algebraic hypotheses.

This soft-analysis style influenced the development of category theory, operator algebras, and the abstract approach to PDE. Bourbaki's encyclopedic project in the 1940s and 1950s adopted the Banach-Steinhaus / open mapping / closed graph triple as the canonical illustration of the Baire-based axiomatic method. The four theorems together are sometimes called the Banach trilogy plus Hahn-Banach, and they appear in every textbook on functional analysis as the gateway from elementary normed-space theory to the deeper theorems on spectra, weak topologies, and applications to PDE.

Bibliography Master

• Baire, R. Sur les fonctions de variables reelles. Annali di Matematica Pura ed Applicata 3 (1899), 1-123.
• Hellinger, E. and Toeplitz, O. Grundlagen fur eine Theorie der unendlichen Matrizen. Mathematische Annalen 69 (1910), 289-330.
• Steinhaus, H. Additive und stetige Funktionaloperationen. Mathematische Zeitschrift 5 (1919), 186-221.
• Banach, S. Sur les operations dans les ensembles abstraits et leur application aux equations integrales. Fundamenta Mathematicae 3 (1922), 133-181.
• Banach, S. and Steinhaus, H. Sur le principe de la condensation de singularites. Fundamenta Mathematicae 9 (1927), 50-61.
• Schauder, J. Uber die Umkehrung linearer, stetiger Funktionaloperationen. Studia Mathematica 2 (1930), 1-6.
• Banach, S. and Schauder, J. Sur les ensembles convexes mesurables. Comptes Rendus de l'Academie des Sciences Paris 194 (1932), 633-635.
• Banach, S. Theorie des operations lineaires. Warsaw, 1932.
• von Neumann, J. Mathematische Grundlagen der Quantenmechanik. Springer, 1932.
• Brezis, H. Functional Analysis, Sobolev Spaces and Partial Differential Equations. Springer, 2011. §2.
• Conway, J. B. A Course in Functional Analysis. 2nd ed., Springer, 1990. §III.
• Rudin, W. Functional Analysis. 2nd ed., McGraw-Hill, 1991. §2.
• Reed, M. and Simon, B. Methods of Modern Mathematical Physics, Vol. I. Academic Press, 1980. §III.
• Schwartz, L. Theorie des distributions. Hermann, 1950 / 1966 revised ed.

---

Wave A unit, missing-units backfill. Produced as the Banach-Schauder pillar joining Hahn-Banach 02.11.02, Banach-Steinhaus 02.11.10 pending, and the spectral / weak-topology / Sobolev applications downstream.