Banach Spaces — Hahn-Banach, Open Mapping, Closed Graph

⚡ Key Takeaways

One-line: The three pillars of linear functional analysis — Hahn-Banach (existence of functionals), Open Mapping (surjective bounded operators are open), and Closed Graph (closed linear maps between Banach spaces are bounded) — form the structural backbone of the theory of Banach spaces. Key result: Continuous linear functionals separate points and extend from subspaces; bounded surjections are quotient maps; closedness of a graph implies continuity in the complete setting. Where it applies: Distribution theory, PDE solvability, spectral theory, approximation theory, optimization, and any analysis on infinite-dimensional normed spaces. Who/when: Hahn (1927), Banach (1929–1932), with the Open Mapping and Closed Graph theorems crystallizing in Banach's Théorie des opérations linéaires (1932); Baire category (1899) is the engine.

💡 The Core Insight

Functional analysis splits its foundational theorems into two families. The Hahn-Banach family (analytic/geometric, no completeness needed) produces functionals and separating hyperplanes by transfinite extension dominated by a convex gauge. The Baire family (Open Mapping, Closed Graph, Uniform Boundedness — completeness essential) extracts quantitative bounds from the impossibility of covering a complete metric space by countably many nowhere-dense sets. Together they say: Banach spaces have enough functionals to do geometry, and bounded operators between them are rigid enough that algebraic surjectivity/closedness force topological openness/continuity.

DefinitionNormed and Banach Spaces

A normed space $(X,\|\cdot\|)$ is a vector space over $\mathbb{K}\in\{\mathbb{R},\mathbb{C}\}$ with a map $\|\cdot\|:X\to[0,\infty)$ satisfying definiteness, homogeneity, and the triangle inequality. It is a Banach space if it is complete in the induced metric $d(x,y)=\|x-y\|$ . The dual space $X^* = B(X,\mathbb{K})$ is the set of bounded (equivalently continuous) linear functionals, normed by

1

\quad \|f\|_{X^*} = \sup_{\|x\|\le 1}|f(x)| = \sup_{x\ne 0}\frac{|f(x)|}{\|x\|}.

Derivation hint: Boundedness means $|f(x)|\le C\|x\|$ ; the smallest such $C$ is the operator norm, recovered as the supremum of $|f|$ on the unit ball. A map $p:X\to\mathbb{R}$ is sublinear if $p(x+y)\le p(x)+p(y)$ and $p(\lambda x)=\lambda p(x)$ for $\lambda\ge 0$ . A seminorm additionally satisfies $p(\lambda x)=|\lambda|p(x)$ .

TheoremHahn-Banach, Analytic Form (Real Case)

Let $X$ be a real vector space, $p:X\to\mathbb{R}$ sublinear, and $M\subseteq X$ a subspace with a linear $f:M\to\mathbb{R}$ satisfying $f(m)\le p(m)$ for all $m\in M$ . Then there exists a linear extension $F:X\to\mathbb{R}$ with $F|_M=f$ and

2

\quad F(x)\le p(x)\qquad\text{for all }x\in X.

Derivation hint: Extend one dimension at a time; the dominating inequality from above and below pins the new value into a nonempty interval, then apply Zorn's lemma to the poset of dominated extensions.

TheoremHahn-Banach, Normed Form

If $M\subseteq X$ is a subspace of a normed space and $f\in M^*$ , there exists $F\in X^*$ with $F|_M=f$ and $\|F\|_{X^*}=\|f\|_{M^*}$ . Consequently, for every $x_0\ne 0$ there is $F\in X^*$ with $\|F\|=1$ and

3

\quad F(x_0)=\|x_0\|.

Derivation hint: Take $p(x)=\|f\|\,\|x\|$ in the analytic form (complex case uses $f(x)=\operatorname{Re}g(x)-i\operatorname{Re}g(ix)$ ). For the second part, define $f(\lambda x_0)=\lambda\|x_0\|$ on $\mathbb{K}x_0$ and extend.

TheoremHahn-Banach, Geometric / Separation Form

Let $A,B$ be disjoint nonempty convex subsets of a normed space $X$ .

If $A$ is open, there exist $f\in X^*$ , $\alpha\in\mathbb{R}$ with $\operatorname{Re}f(a)<\alpha\le\operatorname{Re}f(b)$ for all $a\in A,b\in B$ .
If $A$ is compact, $B$ closed, they are strictly separated: $\operatorname{Re}f(a)\le\alpha-\varepsilon<\alpha+\varepsilon\le\operatorname{Re}f(b)$ .

4

\quad \exists\, f\in X^*,\ \alpha\in\mathbb{R}:\quad \sup_{a\in A}\operatorname{Re}f(a)\ \le\ \alpha\ \le\ \inf_{b\in B}\operatorname{Re}f(b).

Derivation hint: Use the Minkowski gauge $p_C(x)=\inf\{t>0:x/t\in C\}$ of an open convex neighborhood of $0$ ; it is sublinear, and a functional dominated by it gives the separating hyperplane.

TheoremOpen Mapping Theorem

Let $X,Y$ be Banach spaces and $T\in B(X,Y)$ surjective. Then $T$ is an open map: $T(U)$ is open in $Y$ whenever $U$ is open in $X$ . Equivalently there exists $c>0$ with

5

\quad B_Y(0,c)\subseteq T\big(B_X(0,1)\big).

Derivation hint: Surjectivity gives $Y=\bigcup_n \overline{T(B_X(0,n))}$ ; Baire forces one closure to have interior, yielding the inclusion up to closure, then a successive-approximation series (using completeness of $X$ ) removes the closure.

TheoremBounded Inverse Theorem

If $T\in B(X,Y)$ is a bijection between Banach spaces, then $T^{-1}\in B(Y,X)$ . In particular two complete norms on the same space, one dominating the other, are equivalent.

6

\quad T\text{ bijective bounded}\ \Longrightarrow\ \exists\, m>0:\ \|Tx\|\ge m\|x\|.

Derivation hint: Immediate corollary of Open Mapping: openness of $T$ means $T^{-1}$ is continuous.

TheoremClosed Graph Theorem

Let $X,Y$ be Banach spaces and $T:X\to Y$ linear with closed graph

7

\quad \Gamma(T)=\{(x,Tx):x\in X\}\ \text{closed in}\ X\times Y.

Derivation hint: $\Gamma(T)$ is a subspace; equip $X\times Y$ with $\|(x,y)\|=\|x\|+\|y\|$ . Closedness + completeness make $\Gamma(T)$ a Banach space; the projection $\pi_X:\Gamma(T)\to X$ is a bounded bijection, so by Bounded Inverse it has bounded inverse, hence $T=\pi_Y\circ\pi_X^{-1}$ is bounded. Then $T$ is bounded. Concretely: if $x_n\to x$ and $Tx_n\to y$ together imply $y=Tx$ , then $T$ is continuous.

ProofHahn-Banach Analytic Form, Real Case

Step 1 (one-step extension). Suppose $f$ is defined on $M\subsetneq X$ and $z\notin M$ . We seek $\alpha=F(z)$ so that for all $m\in M$ and scalars, $F(m+tz)\le p(m+tz)$ . For $t>0$ and $t<0$ this reduces to needing, for all $u,v\in M$ ,

8

\quad f(u)-p(u-z)\ \le\ \alpha\ \le\ p(v+z)-f(v).

Derivation hint: Split the dominance inequality $f(m\pm z)+\text{terms}\le p$ by sign of $t$ and rescale by $|t|$ using sublinearity and linearity.

Step 2 (interval is nonempty). For any $u,v\in M$ ,

9

\quad f(u)+f(v)=f(u+v)\le p(u+v)\le p(u-z)+p(v+z),

so $\sup_u[f(u)-p(u-z)]\le\inf_v[p(v+z)-f(v)]$ , guaranteeing a valid $\alpha$ . Derivation hint: Add and subtract $z$ inside $p$ and use subadditivity; rearrange.

Step 3 (Zorn). Order the set of dominated extensions $(N,g)$ by inclusion-and-restriction. Every chain has an upper bound (union of domains). A maximal element $(N_0,F)$ must have $N_0=X$ , else Step 1 extends it, contradicting maximality. $\square$

∎

ProofOpen Mapping Theorem

Step 1 (Baire gives interior). $X=\bigcup_{n\ge1}B_X(0,n)$ , so $Y=T(X)=\bigcup_n \overline{T(B_X(0,n))}$ . Since $Y$ is complete (nonmeager in itself), some $\overline{T(B_X(0,n_0))}$ has nonempty interior; by scaling and symmetry $\overline{T(B_X(0,1))}\supseteq B_Y(0,2c)$ for some $c>0$ .

Step 2 (remove the closure). Fix $y$ with $\|y\|<c$ . Find $x_1$ , $\|x_1\|<\tfrac12$ , with $\|y-Tx_1\|<c/2$ . Iterate: at step $k$ choose $x_k$ with $\|x_k\|<2^{-k}$ and

10

\quad \Big\|y-T\!\sum_{j=1}^k x_j\Big\|<\frac{c}{2^k}.

Derivation hint: At each stage the remaining residual lies in a scaled copy of $\overline{T(B(0,1/2))}$ , allowing the next correction term to be chosen with geometrically small norm.

Step 3 (sum and conclude). By completeness of $X$ , $x=\sum_j x_j$ converges with $\|x\|<\sum 2^{-j}=1$ , and continuity gives $Tx=y$ . Hence $B_Y(0,c)\subseteq T(B_X(0,1))$ , proving openness. $\square$

∎

ProofClosed Graph from Open Mapping

Step 1. Endow $\Gamma(T)\subseteq X\times Y$ with the product norm. Since $\Gamma(T)$ is a closed subspace of the Banach space $X\times Y$ , it is itself Banach.

Step 2. The projection $\pi_X:\Gamma(T)\to X$ , $(x,Tx)\mapsto x$ , is linear, bounded ( $\|x\|\le\|(x,Tx)\|$ ), and bijective (linearity + single-valuedness of $T$ ).

Step 3. By the Bounded Inverse Theorem $\pi_X^{-1}:x\mapsto(x,Tx)$ is bounded, so $\|Tx\|\le\|(x,Tx)\|=\|\pi_X^{-1}x\|\le C\|x\|$ . Thus $T$ is bounded. $\square$

∎

SB

Stefan Banach

1892–1945

A largely self-taught mathematician from Kraków, Banach was "discovered" by Steinhaus on a park bench discussing Lebesgue integration. He founded the Lwów School of Mathematics, whose collaborative culture was immortalized in the Scottish Café and its problem book, the Scottish Book. His 1932 monograph Théorie des opérations linéaires organized the field around complete normed spaces — now bearing his name — and contains the Open Mapping, Closed Graph, and Uniform Boundedness theorems in essentially modern form. Hans Hahn had independently proved the extension theorem in 1927; Eduard Helly's earlier sequence-space work (1912, 1921) anticipated key cases. Banach survived WWII feeding lice for a typhus-vaccine lab and died of lung cancer in 1945.

🧠 Intuition

Think of Hahn-Banach as a "reach" guarantee: in any normed space, no matter how exotic, you can always build a continuous measuring device (a functional) that reads off the size of a chosen vector and that respects any prescribed convex constraint $p$ . It is the analytic incarnation of "convex sets can be separated by hyperplanes." The Baire-family theorems are about rigidity through completeness. Imagine completeness as a "no-gaps" guarantee that forbids pathological thinning. Open Mapping says: if a bounded operator hits everything in $Y$ , it cannot do so "infinitely inefficiently" — there is a uniform $c$ so that solving $Tx=y$ costs at most $\|y\|/c$ in norm. Closed Graph says: to check continuity you only need to verify that the map cannot jump — if inputs converge and outputs converge, the output limit is forced to be correct. In incomplete spaces this rigidity evaporates, which is exactly why each theorem fails without completeness.

Figure

The three pillars and their dependence on Baire vs. Zorn

ExampleExtending a functional with controlled norm

Given: On $X=\mathbb{R}^2$ with the $\ell^1$ norm $\|(a,b)\|_1=|a|+|b|$ , let $M=\{(t,0):t\in\mathbb{R}\}$ and $f(t,0)=3t$ .

Step 1 — norm on $M$ . $\|f\|_{M^*}=\sup_{|t|\le 1}|3t|=3$ .

Step 2 — find an extension $F(a,b)=3a+\beta b$ . The dual of $\ell^1$ is $\ell^\infty$ , so $\|F\|=\max(|3|,|\beta|)$ . To preserve $\|F\|=3$ we need $|\beta|\le 3$ .

Step 3 — verify Hahn-Banach is non-unique. Any $\beta\in[-3,3]$ works, e.g. $\beta=0$ gives $F(a,b)=3a$ , and $\beta=3$ gives $F(a,b)=3a+3b$ , both with $\|F\|_{X^*}=3$ .

Boxed answer:

11

\quad \boxed{F(a,b)=3a+\beta b,\quad \beta\in[-3,3],\quad \|F\|_{X^*}=3.}

Derivation hint: $(\ell^1)^*=\ell^\infty$ with the max/sup norm controlling the operator norm.

ExampleOpen Mapping constant for a concrete operator

Given: $T:\ell^2\to\ell^2$ , $T(x_1,x_2,x_3,\dots)=(x_1,\tfrac12 x_2,\tfrac13 x_3,\dots)$ is bounded with $\|T\|=1$ . Is $T$ open? Is $T^{-1}$ bounded?

Step 1 — range. $T$ is injective. Its range is $\{y:\sum n^2|y_n|^2<\infty\}$ , a dense proper subspace, hence not surjective onto $\ell^2$ .

Step 2 — apply the theorem. Open Mapping requires surjectivity. Since $T$ is not surjective, the theorem does not apply, and indeed the formal inverse $T^{-1}y=(y_1,2y_2,3y_3,\dots)$ is unbounded: take $y=e_n/n$ , then $\|y\|=1/n\to 0$ but $\|T^{-1}y\|=1$ .

Step 3 — contrast. Replace the diagonal by bounded-below entries, e.g. $S(x)=(x_1,\tfrac12 x_2+\tfrac12 x_1,\dots)$ designed surjective with $\inf$ singular value $m=1/2$ ; then $S$ is open with constant $c=m=1/2$ , i.e. $B(0,1/2)\subseteq S(B(0,1))$ .

Boxed answer:

12

\quad \boxed{T\text{ not surjective}\Rightarrow\text{Open Mapping fails};\ T^{-1}\text{ unbounded},\ \|T^{-1}e_n/n\|=1\text{ while }\|e_n/n\|\to 0.}

Derivation hint: A diagonal operator with diagonal $\to 0$ has range strictly smaller than the whole space and an unbounded inverse on its range — completeness of the range fails.

⚠️ Common mistake

Hahn-Banach does not need completeness. It holds in any normed (indeed vector) space. Only the Baire-family theorems require completeness of both spaces.
The Hahn-Banach extension is generally not unique. Uniqueness holds iff the space is smooth (the norm is Gâteaux differentiable); in Hilbert space and reflexive smooth spaces it is unique, but not in $\ell^1$ (Example 11).
Open Mapping needs surjectivity, Closed Graph needs both spaces complete. Dropping either hypothesis breaks the theorem (see counterexamples).
"Closed operator" ≠ "bounded operator" in general. They coincide only when the domain is the whole Banach space. Differential operators are the classic closed-but-unbounded examples — and they live on proper dense domains, so the Closed Graph theorem does not contradict their unboundedness.
Open map ≠ continuous map ≠ closed map. These are independent notions; a homeomorphism is all three, but isolated implications fail in general.

CounterexampleClosed Graph fails without completeness of the domain

Let $X=C^1[0,1]$ equipped with the sup norm $\|\cdot\|_\infty$ (incomplete!) and $Y=C[0,1]$ with $\|\cdot\|_\infty$ (complete). Define $T=\dfrac{d}{dx}:X\to Y$ . $T$ has closed graph: if $f_n\to f$ uniformly and $f_n'\to g$ uniformly, then $f$ is $C^1$ with $f'=g$ (a standard uniform-convergence theorem). Yet $T$ is unbounded: $f_n(x)=\sin(nx)$ has $\|f_n\|_\infty=1$ but $\|Tf_n\|_\infty=\|n\cos(nx)\|_\infty=n\to\infty$ . The Closed Graph conclusion (" $T$ bounded") fails because $X$ is not complete under $\|\cdot\|_\infty$ . Re-norming $X$ with $\|f\|_{C^1}=\|f\|_\infty+\|f'\|_\infty$ makes $X$ complete and then $T$ is bounded — illustrating that the completeness hypothesis cannot be dropped.

CounterexampleOpen Mapping fails without completeness

Let $X=(c_{00},\|\cdot\|_\infty)$ , finitely supported sequences (incomplete), and $T:X\to X$ , $T(x_n)=(x_n/n)$ . $T$ is a bounded bijection of $c_{00}$ onto $c_{00}$ , but $T^{-1}(x_n)=(nx_n)$ is unbounded, so $T$ is not open. Completeness of $X$ fails, so the Bounded Inverse / Open Mapping theorem does not apply.

connection

These theorems weave through all of functional analysis. Hahn-Banach underwrites the Dual Space and Weak Topologies: it makes the canonical embedding $X\hookrightarrow X^{**}$ isometric and defines Reflexive Banach Spaces. The separation form is the analytic source of Convex Optimization Duality and the supporting-hyperplane theorem. The Baire trio (Open Mapping, Closed Graph, Uniform Boundedness Principle) jointly control operator regularity and undergird the Spectral Theory of Bounded Operators (closedness of resolvent sets) and Unbounded Operators and Self-Adjointness (closed operators as the right category). In Distribution Theory, the Closed Graph theorem (Schwartz-kernel context) and Banach-space duality justify defining operations by transposition. The Open Mapping theorem is the abstract reason behind a priori estimates and solvability in linear Elliptic PDE Theory.

🔬 Where this appears

⏳ History

- 1899 — René Baire establishes the Baire Category Theorem in his thesis, the future engine of the Open Mapping / Closed Graph theorems.

- 1912 — Eduard Helly proves an extension result for continuous linear functionals on sequence spaces, a special case of Hahn-Banach.

- 1921 — Helly's "Über lineare Funktionaloperationen" sharpens the duality of sequence spaces and norm-preserving extension.

- 1922 — Banach's doctoral thesis introduces axiomatic complete normed spaces ("espaces de type B").

- 1927 — Hans Hahn proves the general analytic extension theorem for real normed spaces (the "Hahn-Banach theorem" name later honors both).

- 1929 — Banach independently generalizes the extension theorem and develops the duality and category-based theorems.

- 1932 — Banach's Théorie des opérations linéaires systematizes Open Mapping, Closed Graph, Bounded Inverse, and Uniform Boundedness in mature form.

- 1938–40 — Mazur, Bohnenblust, and Sobczyk extend Hahn-Banach to the complex case via the real-part trick.

- 1950s onward — Schwartz, Grothendieck generalize the closed-graph / open-mapping circle to Fréchet, (LF), and locally convex spaces (Closed Graph Theorem of Pták, De Wilde webbed spaces).

open-problem

Beyond Banach spaces and the role of choice.

Logical strength. Full Hahn-Banach is strictly weaker than the Axiom of Choice but not provable in ZF; it follows from the Boolean Prime Ideal theorem (Luxemburg). Banach measures (finitely additive, translation-invariant) on $\mathbb{R}$ exist precisely because of Hahn-Banach, exhibiting non-constructive consequences.
Generalizing the closed-graph circle. For locally convex spaces the Open Mapping and Closed Graph theorems require structural hypotheses (e.g. De Wilde's webbed spaces as domain, ultrabornological ranges). Identifying the optimal pairing of source/target classes for which closed graph ⇒ continuity remains a refined research theme.
Vector-valued extensions. The Hahn-Banach property for $X^*$ -valued (operator-valued) extensions characterizes injective Banach spaces; classifying spaces with norm-preserving operator extension (the $\mathcal{P}_\lambda$ and $1$ -injective spaces; $L^\infty$ -type) connects to deep Banach-space geometry that is still incompletely mapped.
Quantitative separation. Effective (computable) bounds in separation theorems and their role in proof mining for nonlinear analysis is an active program (Kohlenbach).

Banach Spaces — Hahn-Banach, Open Mapping, Closed Graph

Formal Statement

Why It Holds

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also