Hahn-Banach theorem (analytic and geometric forms)

Intuition Beginner

Suppose someone hands you a ruler that only measures lengths along a single line drawn through the origin. The ruler works perfectly on that line. The Hahn-Banach theorem promises that you can always extend that ruler to a measurement defined on the whole surrounding space, keeping every length on the original line unchanged and never reporting a larger length than the original ruler would have.

In plainer words: a length rule that is consistent on a small piece of a space can always be enlarged to a length rule on the whole space, without inflation.

This is the analytic version of the theorem. There is also a geometric version that sounds like a different statement but is really the same idea wearing different clothes: any two convex shapes that do not overlap can be separated by a flat wall.

Visual Beginner

Two convex blobs sitting in the plane without touching. A single straight line slides between them so that one blob sits entirely on one side and the other blob sits entirely on the other side.

The line is the geometric Hahn-Banach separator. In higher dimensions, replace "line" with "flat wall of one dimension less than the space." In infinite dimensions, the same picture holds, but the wall is built from a linear functional whose existence the theorem guarantees.

Worked example Beginner

Picture the plane with coordinates $x$ and $y$. On the $x$-axis alone, define a length rule that reports the absolute value of $x$. That rule is consistent on the axis.

To extend it to the whole plane, you need a measurement that agrees with absolute value on the axis and stays bounded by something reasonable everywhere else. One valid extension reports the absolute value of $x$ for every point in the plane, ignoring $y$. Another valid extension reports the larger of $|x|$ and $|y|$. Both extensions agree with the original rule on the axis. Hahn-Banach says at least one such extension always exists for the corresponding problem about linear measurements; the proof in the master tier explains why.

Why this matters Beginner

Three reasons the theorem appears in almost every functional analysis course:

First, it builds dual spaces. The dual of a space is the collection of all bounded linear measurements on that space. Without Hahn-Banach, the dual could be empty for unusual spaces. Hahn-Banach guarantees there are always plenty of measurements to work with.

Second, it separates objects. Optimisation, economics, and probability all reduce questions about feasibility or pricing to questions about whether two convex sets can be separated. The geometric Hahn-Banach gives that separation.

Third, it links algebra to geometry. A measurement is an algebraic object. A wall is a geometric object. The theorem says these two descriptions are equivalent.

Check your understanding Beginner

Formal definition Intermediate+

Let $X$ be a real vector space 01.01.03. A sublinear functional on $X$ is a map $p:X\to \mathbb{R}$ satisfying

$$p(\lambda x)=\lambda p(x)(\lambda \ge 0),p(x+y)\le p(x)+p(y).$$

A norm is sublinear, but sublinear functionals are more general: they need not be symmetric and need not vanish only at zero.

Theorem (analytic Hahn-Banach, real form). Let $X$ be a real vector space, $p$ a sublinear functional on $X$, $Y\subset X$ a linear subspace, and $f:Y\to \mathbb{R}$ a linear functional satisfying $f(y)\le p(y)$ for every $y\in Y$. Then there exists a linear functional $F:X\to \mathbb{R}$ with $F{|}_Y=f$ and $F(x)\le p(x)$ for every $x\in X$ ^{[Brezis §1]}.

The norm-preserving extension theorem for normed spaces is the corollary obtained by taking $p(x)=\Vert f{\Vert }_{Y^{\ast }}\Vert x\Vert$.

Theorem (norm-preserving extension). Let $X$ be a normed vector space 02.11.06, $Y\subset X$ a linear subspace, and $f\in Y^{\ast }$ a bounded linear functional. Then there is $F\in X^{\ast }$ with $F{|}_Y=f$ and $\Vert F{\Vert }_{X^{\ast }}=\Vert f{\Vert }_{Y^{\ast }}$.

The complex version replaces $\mathbb{R}$ with $\mathbb{C}$ and is due to Bohnenblust and Sobczyk: regard a complex-linear functional $f$ as $f(x)=u(x)-iu(ix)$ where $u=\mathrm{Re}f$, extend the real part using the real Hahn-Banach, and recover the complex extension by the same formula.

Theorem (geometric Hahn-Banach, first form). Let $X$ be a real normed space, $A\subset X$ open convex non-empty, $B\subset X$ convex non-empty, and $A\cap B=\varnothing$. Then there exist $f\in X^{\ast }\setminus \{0\}$ and $\alpha \in \mathbb{R}$ with

$$f(a)<\alpha \le f(b)(a\in A,b\in B).$$

Theorem (geometric Hahn-Banach, second form). Let $X$ be a real normed space, $A\subset X$ compact convex non-empty, $B\subset X$ closed convex non-empty, and $A\cap B=\varnothing$. Then there exist $f\in X^{\ast }\setminus \{0\}$ and ${\alpha }_1<{\alpha }_2$ in $\mathbb{R}$ with

$$f(a)\le {\alpha }_1<{\alpha }_2\le f(b)(a\in A,b\in B).$$

The first form gives a separating closed hyperplane, the second a strictly separating one.

Key theorem with proof Intermediate+

We prove the analytic Hahn-Banach for a real vector space using Zorn's lemma, then sketch the separable variant that avoids choice.

Proof of analytic Hahn-Banach (general case via Zorn). Let $\mathcal{E}$ be the collection of all pairs $(Z,g)$ where $Y\subset Z\subset X$ is a linear subspace and $g:Z\to \mathbb{R}$ is linear, extends $f$, and is dominated by $p$. Order $\mathcal{E}$ by extension: $(Z_1,g_1)⪯(Z_2,g_2)$ if $Z_1\subset Z_2$ and $g_2{|}_{Z_1}=g_1$.

The collection $\mathcal{E}$ is non-empty because $(Y,f)\in \mathcal{E}$. Any chain $\{(Z_{\alpha },g_{\alpha })\}$ has the upper bound $(\bigcup Z_{\alpha },g)$ where $g(x)=g_{\alpha }(x)$ for any $\alpha$ with $x\in Z_{\alpha }$; the dominance condition is inherited. Zorn's lemma supplies a maximal element $(M,F)$.

We claim $M=X$. Suppose otherwise; pick $x_0\in X\setminus M$ and set $M^{\prime }=M+\mathbb{R}{x}_0$. Any linear extension $F^{\prime }$ of $F$ to $M^{\prime }$ is determined by the scalar $c=F^{\prime }(x_0)$. The dominance condition

$$F(m)+\lambda c\le p(m+\lambda x_0)(m\in M,\lambda \in \mathbb{R})$$

splits by sign of $\lambda$. Writing $\lambda >0$ as $\lambda$ and $\lambda <0$ as $-\mu$ with $\mu >0$, the condition becomes

$$\frac{F(m)-p(m-\mu x_0)}{\mu }\le c\le \frac{p(m+\lambda x_0)-F(m)}{\lambda }(m\in M,\lambda ,\mu >0).$$

So we need

$$\underset{m\in M,\mu >0}{\sup }\frac{F(m)-p(m-\mu x_0)}{\mu }\le \underset{m\in M,\lambda >0}{\inf }\frac{p(m+\lambda x_0)-F(m)}{\lambda }.$$

For any $m_1,m_2\in M$ and $\mu ,\lambda >0$, the dominance of $F$ on $M$ and subadditivity of $p$ yield

$$\lambda F(m_1)+\mu F(m_2)=F(\lambda m_1+\mu m_2)\le p(\lambda m_1+\mu m_2),$$

and adding and subtracting the term $\mu \lambda x_0$ inside $p$ gives

$$F(\lambda m_1+\mu m_2)\le \lambda p(m_1+\mu x_0)+\mu p(m_2/\mu \cdot 1-x_0)\dots$$

More cleanly: pick $\lambda ,\mu >0$ and $m_1,m_2\in M$ and observe

$$\mu F(m_1)+\lambda F(m_2)=F(\mu m_1+\lambda m_2)\le p(\mu m_1+\lambda m_2).$$

Since $\mu m_1+\lambda m_2=\mu (m_1-\mu x_0\cdot 0)+\lambda (m_2+\lambda x_0\cdot 0)$ we use the identity

$$\mu m_1+\lambda m_2=\mu (m_1-\lambda x_0/\mu \cdot \mu )+\lambda (m_2+\mu x_0/\lambda \cdot \lambda )$$

after introducing $z=\mu \lambda x_0$ into both summands. The cleanest way to do this is the following: for any $m_1,m_2\in M$ and any $\lambda ,\mu >0$,

$$\mu F(m_1)+\lambda F(m_2)=F(\mu m_1+\lambda m_2)\le p(\mu m_1+\lambda m_2).$$

We split $\mu m_1+\lambda m_2$ as $\mu (m_1-\nu x_0)+\lambda (m_2+\sigma x_0)$ where $\mu \nu =\lambda \sigma$; the cross terms cancel. Choosing $\nu =\lambda$ and $\sigma =\mu$ yields

$$\mu F(m_1)+\lambda F(m_2)\le \mu p(m_1-\lambda x_0)+\lambda p(m_2+\mu x_0),$$

so

$$\frac{F(m_1)-p(m_1-\lambda x_0)}{\lambda }\le \frac{p(m_2+\mu x_0)-F(m_2)}{\mu }.$$

The supremum on the left is below the infimum on the right, and any $c$ in between works. Choosing one such $c$ produces a strict extension $(M^{\prime },F^{\prime })\in \mathcal{E}$ with $(M,F)\prec (M^{\prime },F^{\prime })$, contradicting maximality. So $M=X$ and $F$ is the desired extension. $\square$

Proof of separable case (no Zorn needed). Suppose $X$ is separable: pick a countable dense set $\{x_n\}$ and let $X_0=Y$, $X_n=X_{n-1}+\mathbb{R}{x}_n$. Extend $f$ from $X_{n-1}$ to $X_n$ by the one-step construction above at each stage. Set $f^{\ast }={\bigcup }_n{f}_n$ on ${\bigcup }_n{X}_n$, which is dense in $X$. The estimate $|f^{\ast }(x)|\le p(x)$ guarantees uniform continuity, so $f^{\ast }$ extends by uniform continuity to all of $X$ when $p$ is continuous (in particular, when $p$ is a norm). No choice principle stronger than dependent choice is used.

Bridge. The norm-preserving extension corollary forces $X^{\ast }\ne \{0\}$ for every normed space with at least one non-zero vector, which is the foundation for the open mapping theorem 02.11.09 and Banach-Steinhaus 02.11.10 pending. The geometric form supplies the separating hyperplane underpinning the weak-topology theory 02.11.11 pending and the duality machinery in convex optimisation ^{[Brezis §1]}.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: partial — Mathlib supplies the analytic Hahn-Banach as ContinuousLinearMap.exists_extension_norm_eq and the geometric forms via Mathlib.Analysis.NormedSpace.HahnBanach.Separation. The companion module records curriculum-facing aliases for the analytic and the two geometric forms so that downstream units cite a stable name even while the underlying proofs ride on Mathlib's API. A full proof of the sublinear-dominated version requires importing the order-theoretic Zorn machinery and re-deriving the one-step extension lemma, which the companion leaves as an exercise to plug in later.

import Mathlib.Analysis.NormedSpace.HahnBanach.Extension
import Mathlib.Analysis.NormedSpace.HahnBanach.Separation

namespace Codex.Analysis.FunctionalAnalysis

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

abbrev HahnBanachAnalytic : Prop :=
  ∀ (p : Subspace ℝ E) (f : p →L[ℝ] ℝ),
    ∃ g : E →L[ℝ] ℝ, (∀ x : p, g x = f x) ∧ ‖g‖ = ‖f‖

end Codex.Analysis.FunctionalAnalysis

Complex Hahn-Banach via Bohnenblust-Sobczyk Master

The original Hahn-Banach proofs by Hahn (1927) and Banach (1929) handled real scalar fields. The complex case required a separate argument because the dominance condition $f(x)\le p(x)$ does not make sense for complex-valued functionals. Bohnenblust and Sobczyk (1938) gave the now-standard reduction.

Theorem (complex Hahn-Banach). Let $X$ be a complex vector space, $p:X\to [0,\infty )$ a seminorm, $Y\subset X$ a complex-linear subspace, and $f:Y\to \mathbb{C}$ a complex-linear functional with $|f(y)|\le p(y)$ for every $y\in Y$. Then $f$ extends to a complex-linear functional $F:X\to \mathbb{C}$ with $|F(x)|\le p(x)$ for every $x\in X$.

Proof. Write $f(x)=u(x)-iu(ix)$ where $u(x)=\mathrm{Re}f(x)$. The map $u:Y\to \mathbb{R}$ is $\mathbb{R}$-linear and satisfies $u(y)\le |f(y)|\le p(y)$. The real Hahn-Banach applied to the underlying real vector space $X_{\mathbb{R}}$, with sublinear functional $p$, extends $u$ to an $\mathbb{R}$-linear $U:X\to \mathbb{R}$ dominated by $p$.

Define $F(x)=U(x)-iU(ix)$. Linearity in real scalars is clear. For multiplication by $i$:

$$F(ix)=U(ix)-iU(-x)=U(ix)+iU(x)=i(U(x)-iU(ix))=iF(x),$$

so $F$ is complex-linear. On $Y$, $F(y)=U(y)-iU(iy)=u(y)-iu(iy)=f(y)$.

For the bound on $|F|$, pick $x\in X$ with $F(x)\ne 0$ and let $\theta \in \mathbb{R}$ with $e^{-i\theta }F(x)=|F(x)|$. Then

$$|F(x)|=e^{-i\theta }F(x)=F(e^{-i\theta }x)=U(e^{-i\theta }x)\le p(e^{-i\theta }x)=p(x),$$

where the second equality uses complex linearity, the third uses that $F(e^{-i\theta }x)$ is real positive (hence equal to its real part $U$), and the inequality is the dominance of $U$ by $p$. $\square$

The same trick works for the norm-preserving variant: every bounded complex-linear functional on a subspace of a complex normed space extends to the whole space with the same operator norm.

Geometric Hahn-Banach proofs Master

The two geometric forms follow from the analytic form via the Minkowski functional of a convex set, the operation that converts geometry into a sublinear functional.

Definition. Let $C$ be a convex subset of a real vector space $X$ with $0\in C$ and $C$ absorbing (every $x\in X$ has $tx\in C$ for some $t>0$). The Minkowski functional (or gauge) of $C$ is

$$p_C(x)=\inf \{t>0:x\in tC\}.$$

It satisfies $p_C(\lambda x)=\lambda p_C(x)$ for $\lambda \ge 0$ and $p_C(x+y)\le p_C(x)+p_C(y)$, hence is sublinear. If $C$ is open, $p_C(x)<1$ if and only if $x\in C$. If $C$ is open and contains a ball around $0$, $p_C$ is continuous and bounded by $\Vert x\Vert /r$ where $r$ is the radius of the ball.

Proof of geometric Hahn-Banach, first form. Translate so that $A$ is an open convex neighbourhood of the origin; pick $x_0\in B$ and replace $A$ by $A-B+x_0$ if needed. Let $C=A-B+x_0$, which is convex (sum of convex sets), open (sum with an open set), and contains $0$ (because $A\cap B=\varnothing$ forces $0\notin A-B$, but $C=A-B+x_0$ shifts that; the careful setup is in Brezis §1.2).

Concretely: take $A$ open convex non-empty and $B$ convex non-empty disjoint from $A$. Pick $a_0\in A$, $b_0\in B$, and let $z_0=b_0-a_0$ and $C=A-B+z_0$. Then $C$ is open convex, $0\in C$ (since $a_0-b_0+z_0=0$), and $z_0\notin C$ (because $z_0\in C$ would yield $z_0=a-b+z_0$ for some $a\in A,b\in B$, so $a=b$, contradicting $A\cap B=\varnothing$).

Define $g:\mathbb{R}{z}_0\to \mathbb{R}$ by $g(\lambda z_0)=\lambda$. For $\lambda \ge 0$, $g(\lambda z_0)=\lambda \le p_C(\lambda z_0)$ since $z_0\notin C$ forces $p_C(z_0)\ge 1$. For $\lambda <0$, $g(\lambda z_0)=\lambda <0\le p_C(\lambda z_0)$.

The analytic Hahn-Banach extends $g$ to a linear $G:X\to \mathbb{R}$ dominated by $p_C$. Continuity of $p_C$ (it is bounded by a multiple of $\Vert \cdot \Vert$ because $C$ contains a ball around $0$) gives continuity of $G$, hence $G\in X^{\ast }$.

For $a\in A,b\in B$: $a-b+z_0\in C$, so $G(a)-G(b)+G(z_0)\le p_C(a-b+z_0)<1=G(z_0)$, hence $G(a)<G(b)$. Setting $f=G$ and $\alpha ={\sup }_{a\in A}G(a)\le {\inf }_{b\in B}G(b)$ gives the separator. $\square$

Proof of geometric Hahn-Banach, second form. Let $A$ be compact convex, $B$ closed convex, $A\cap B=\varnothing$. The distance $d={\inf }_{a\in A,b\in B}\Vert a-b\Vert$ is positive: continuity of $(a,b)\mapsto \Vert a-b\Vert$ on the compact-times-closed set, plus the disjointness, give a positive infimum (a compact-set / closed-set Bolzano-Weierstrass argument).

Let $A^{\prime }=A+\overline{B(0,d/2)}$, the closed Minkowski sum of $A$ with the closed ball of radius $d/2$, and $B^{\prime }=B+\overline{B(0,d/2)}$. Both are convex; $A^{\prime }$ is compact and $B^{\prime }$ is closed. The interiors of $A^{\prime }$ and $B^{\prime }$ are disjoint: any common interior point $z$ would satisfy $z=a+v=b+w$ with $\Vert v\Vert ,\Vert w\Vert <d/2$, hence $\Vert a-b\Vert =\Vert w-v\Vert <d$, contradiction.

Apply the first form to $\mathrm{int}{A}^{\prime }$ (open convex non-empty) and $B^{\prime }$ (convex non-empty disjoint): there exist $f\in X^{\ast }\setminus \{0\}$ and $\alpha \in \mathbb{R}$ with $f(x)<\alpha \le f(y)$ for $x\in \mathrm{int}{A}^{\prime },y\in B^{\prime }$. By a translation argument and the fact that $A\subset A^{\prime }$ with a margin of $d/2$ to the boundary, $f(a)\le {\alpha }_1<{\alpha }_2\le f(b)$ for $a\in A,b\in B$ where ${\alpha }_2-{\alpha }_1=\Vert f\Vert \cdot d/2>0$. $\square$

Corollaries Master

The corollaries below all flow directly from norm-preserving extension or from geometric separation.

(C1) Richness of the dual. If $X$ is a normed space and $X\ne \{0\}$, then $X^{\ast }\ne \{0\}$. Concretely, for each $x_0\ne 0$ there is $f\in X^{\ast }$ with $f(x_0)=\Vert x_0\Vert$ and $\Vert f\Vert =1$. The dual norm is realised: $\Vert x\Vert ={\max }_{f\in B_{X^{\ast }}}|f(x)|$.

(C2) Density via annihilators. A linear subspace $Y\subset X$ is dense in $X$ if and only if every $f\in X^{\ast }$ that vanishes on $Y$ is zero. Equivalently, the annihilator $Y^{\perp }=\{f\in X^{\ast }:f{|}_Y=0\}$ vanishes if and only if $\overline{Y}=X$.

(C3) Existence of supporting hyperplanes. If $C\subset X$ is a closed convex set and $x_0\in \partial C$ (boundary point) with $\mathrm{int}C\ne \varnothing$, then there exists $f\in X^{\ast }\setminus \{0\}$ with $f(x_0)\ge f(x)$ for every $x\in C$. The hyperplane $\{x:f(x)=f(x_0)\}$ supports $C$ at $x_0$.

(C4) Biduality and reflexivity. The canonical embedding $J:X\to X^{\ast \ast }$, $J(x)(f)=f(x)$, is an isometry by (C1). $X$ is reflexive if $J$ is surjective. Hilbert spaces, finite-dimensional spaces, $L^p$ for $1<p<\infty$, and ${\ell }^p$ for $1<p<\infty$ are reflexive. The spaces $L^1$, $L^{\infty }$, ${\ell }^1$, ${\ell }^{\infty }$, and $c_0$ are not. Reflexivity controls weak compactness via the Banach-Alaoglu and Kakutani theorems.

(C5) Quotient duality. For a closed subspace $Y\subset X$, $(X/Y{)}^{\ast }\cong Y^{\perp }$ and $Y^{\ast }\cong X^{\ast }/Y^{\perp }$ as isometric isomorphisms. The first follows from extending continuous functionals on $X/Y$ to $X$ (vanishing on $Y$); the second extends functionals on $Y$ to $X$ and quotients out the annihilator.

(C6) Mazur's theorem on weakly closed convex sets. A convex subset of a normed space is closed if and only if it is weakly closed. Proof: closure in the weak topology contains closure in the norm topology; for the reverse, a norm-closed convex set $C$ and an external point $x_0$ are separated by some $f\in X^{\ast }$ via (C3), which witnesses $x_0$ outside the weak closure too.

Applications to optimisation Master

Convex optimisation lives on the geometric Hahn-Banach. Three flagship results:

Lagrangian duality. Consider the primal problem $\inf \{f(x):x\in X,g_i(x)\le 0\}$ where $f,g_i$ are convex. Form the perturbation function $v(u)=\inf \{f(x):g_i(x)\le u_i\}$. Geometric Hahn-Banach applied to the epigraph $\mathrm{epi}v$ (convex if all data are convex) and the point $(0,v(0)-\epsilon )$ produces a supporting hyperplane parameterised by Lagrange multipliers ${\lambda }_i\ge 0$. The hyperplane's slope gives the dual functional $L(\lambda )={\inf }_x(f(x)+{\sum }_i{\lambda }_i{g}_i(x))$. Strong duality $\inf f=\sup L$ holds under a Slater interior-point condition; this is the constraint qualification that ensures the supporting hyperplane is not vertical ^{[Brezis §1.3]}.

Minimax theorems. Von Neumann's minimax theorem and its convex-analytic generalisations (Sion's theorem, Ky Fan's inequality) reduce to existence of saddle points for convex-concave functions on convex sets. The proof factors through the geometric Hahn-Banach: separate the lower-level set of the gain from the upper-level set of the loss, and read off the saddle multiplier.

Pricing in arbitrage-free markets. A finite or infinite-asset market is arbitrage-free if and only if the set of attainable gains is disjoint from the strictly positive cone. Geometric Hahn-Banach separates the two by a linear functional, normalised to a probability measure equivalent to the original; this is the fundamental theorem of asset pricing. The same logic recurs in mathematical economics under the names "second welfare theorem" and "Arrow-Debreu equilibrium pricing."

Best-approximation problems. Given a closed convex set $C$ in a reflexive Banach space and a point $x_0\notin C$, the projection $P_C(x_0)$ exists and is characterised by the variational inequality $⟨x_0-P_C(x_0),y-P_C(x_0)⟩\le 0$ for all $y\in C$. The supporting hyperplane at $P_C(x_0)$ is the geometric witness.

Failure without choice Master

Hahn-Banach is strictly weaker than the axiom of choice but strictly stronger than Zermelo-Fraenkel set theory alone. The status of HB in choice-restricted set theories is a delicate piece of reverse mathematics.

Theorem (Pincus 1972). Hahn-Banach is independent of Zermelo-Fraenkel set theory: there is a model of ZF in which Hahn-Banach fails. In particular, there is a normed vector space $X$ and a bounded linear functional $f$ on a subspace $Y\subset X$ with no norm-preserving extension to $X$.

Pincus showed Hahn-Banach is equivalent (over ZF + dependent choice) to the Boolean prime ideal theorem (BPI), which is strictly weaker than full choice. So the cleaner statement is:

(Pincus, Łoś-Ryll-Nardzewski). In ZF + dependent choice, the analytic Hahn-Banach theorem is equivalent to the Boolean prime ideal theorem (every Boolean algebra with more than one element has a prime ideal). Both are strictly weaker than the axiom of choice, and both are strictly stronger than ZF + dependent choice alone.

Theorem (Foreman and Wehrung 1991). Hahn-Banach implies the existence of a non-measurable subset of the real line. So if every subset of $\mathbb{R}$ is Lebesgue-measurable (a consequence of certain large-cardinal axioms via Solovay's model), then Hahn-Banach fails. The geometric form fails accordingly: there exist convex sets in a normed space that cannot be separated by a continuous linear functional.

The practical implication: working mathematicians use Hahn-Banach freely in ZFC, but choice-sensitive constructions in descriptive set theory and analysis must track which form of Hahn-Banach is invoked. The separable case avoids all of this because the inductive proof uses only countable dependent choice ^{[Conway §III]}.

Ordered vector spaces and Kantorovich extension Master

The analytic Hahn-Banach generalises beyond real or complex scalars to positive linear functionals on ordered vector spaces, a setting initiated by Kantorovich in the 1930s as part of his work on linear programming and Soviet planning theory.

Definition. A Riesz space (or vector lattice) is a real vector space $X$ with a partial order $\le$ compatible with the vector operations and such that every pair $x,y\in X$ has a supremum $x\vee y$ and infimum $x\wedge y$.

Theorem (Kantorovich extension). Let $X$ be a Riesz space, $Y\subset X$ a vector sublattice, and $f:Y\to \mathbb{R}$ a positive linear functional (so $f(y)\ge 0$ when $y\ge 0$). If $X$ has the Riesz decomposition property and $f$ is dominated by a positive sublinear functional $p$, then $f$ extends to a positive linear functional $F:X\to \mathbb{R}$ with $F\le p$.

The proof parallels the analytic Hahn-Banach but tracks order-positivity at each step of the Zorn induction. The Kantorovich extension drives spectral theory of positive operators (Krein-Rutman theorem), duality in linear programming, and the modern theory of operator algebras (positive functionals on $C^{\ast }$-algebras extending to states on the enveloping von Neumann algebra). The link with $C^{\ast }$-algebras 02.11.15 pending is the Gelfand-Naimark-Segal construction, whose first step is precisely a Hahn-Banach extension of a state.

The same machinery gives the Hahn-Banach-Lagrange theorem (Simons), unifying analytic Hahn-Banach, Lagrange duality, and the Fenchel duality of convex analysis into a single statement parameterised by sublinear majorants.

Banach limits and finitely additive measures Master

A direct application of analytic Hahn-Banach produces objects that the axioms of countably additive measure theory forbid: shift-invariant means on bounded sequences, finitely additive translation-invariant probability measures on the integers, and ultrafilters on infinite sets that are not principal. These objects share the same set-theoretic strength as Hahn-Banach itself (Boolean prime ideal theorem) and are constructively unavailable.

Theorem (Banach limit). There exists a linear functional $L:{\ell }^{\infty }(\mathbb{N})\to \mathbb{R}$ with the following four properties.

1. Positivity. $L(x)\ge 0$ whenever $x_n\ge 0$ for every $n$.
2. Normalisation. $L(1,1,1,\dots )=1$.
3. Shift invariance. $L(x_2,x_3,x_4,\dots )=L(x_1,x_2,x_3,\dots )$.
4. Extension of $\lim$. $L(x)={\lim }_{n\to \infty }{x}_n$ whenever the limit exists.

Proof sketch. Define the sublinear functional $p(x)={\mathrm{limsup}}_{n\to \infty }\frac{1}{n}{\sum }_{k=1}^n{x}_k$. The functional $p$ is sublinear on ${\ell }^{\infty }(\mathbb{N})$, agrees with $\lim$ on convergent sequences, and is shift-invariant. The analytic Hahn-Banach extends $\lim$ from the closed subspace $c$ of convergent sequences to a linear $L$ on all of ${\ell }^{\infty }(\mathbb{N})$ dominated by $p$. Shift-invariance survives because $p$ is shift-invariant. $\square$

The Banach limit gives a meaning to "the average value of a bounded sequence" even when no average in the usual sense exists. It is used to define states on ${\ell }^{\infty }(\mathbb{N})$ in operator algebra, to construct finitely additive translation-invariant measures on $\mathbb{Z}$ in amenable-group theory, and to handle Cesaro divergence in number theory and ergodic theory.

Corollary (finitely additive measure on $\mathbb{N}$). Define $\mu (A)=L({\chi }_A)$ for $A\subset \mathbb{N}$, where ${\chi }_A$ is the indicator sequence. Then $\mu$ is a finitely additive probability measure on $\mathbb{N}$ that assigns zero mass to finite sets and is invariant under translation by any fixed integer. No countably additive measure on $\mathbb{N}$ with all four properties exists: the union of singletons ${\bigcup }_n\{n\}=\mathbb{N}$ would force $1={\sum }_n\mu (\{n\})=0$.

Connection to amenability. A discrete group $G$ is amenable if ${\ell }^{\infty }(G)$ admits a left-invariant mean, namely a positive normalised linear functional invariant under left translation. The Banach-limit construction shows that $\mathbb{Z}$ and all abelian groups are amenable; the construction fails for free groups on two or more generators (a result of von Neumann). The Hahn-Banach extension of a translation-invariant sublinear majorant is the recurring ingredient.

Subdifferentials and Fenchel duality Master

The geometric Hahn-Banach is the engine of subdifferential calculus, the convex-analytic substitute for the gradient when functions fail to be differentiable.

Definition. Let $f:X\to (-\infty ,+\infty ]$ be a convex function on a normed space $X$. The subdifferential of $f$ at $x_0\in \mathrm{dom}f$ is the set

$$\partial f(x_0)=\{\phi \in X^{\ast }:f(x)\ge f(x_0)+\phi (x-x_0)\text{ for every }x\in X\}.$$

Each $\phi \in \partial f(x_0)$ is a slope of a supporting affine functional to the epigraph of $f$ at the point $(x_0,f(x_0))$.

Theorem (existence of subgradients). If $f$ is convex, lower-semicontinuous, proper (not identically $+\infty$), and continuous at $x_0$, then $\partial f(x_0)\ne \varnothing$.

Proof. Continuity at $x_0$ gives an open neighbourhood $U$ of $(x_0,f(x_0))$ in $X\times \mathbb{R}$ disjoint from the strict epigraph ${\mathrm{epi}}^{>}f=\{(x,t):t>f(x)\}$. The strict epigraph is open and convex; its complement contains the point $(x_0,f(x_0))$. Geometric Hahn-Banach separates them by a closed hyperplane $\{(\phi ,c):\phi (x)-ct=\alpha \}$ for some $\phi \in X^{\ast }$, $c\in \mathbb{R}$. Vertical separation is ruled out by openness; rescaling gives $c=1$ and $\phi (x)-t\le \alpha$ on the epigraph with equality at $(x_0,f(x_0))$. The functional $\phi$ is the subgradient. $\square$

Fenchel conjugate. For $f:X\to (-\infty ,+\infty ]$, the Fenchel conjugate $f^{\ast }:X^{\ast }\to (-\infty ,+\infty ]$ is

$$f^{\ast }(\phi )=\underset{x\in X}{\sup }(\phi (x)-f(x)).$$

The Young inequality $f(x)+f^{\ast }(\phi )\ge \phi (x)$ holds by definition. Equality holds at $(x,\phi )$ if and only if $\phi \in \partial f(x)$.

Theorem (Fenchel-Moreau biduality). If $f$ is proper, convex, and lower-semicontinuous, then $f^{\ast \ast }=f$ under the canonical embedding $X\hookrightarrow X^{\ast \ast }$.

The proof uses the second geometric Hahn-Banach to separate the closed convex set $\mathrm{epi}f$ from a point $(x_0,t_0)\notin \mathrm{epi}f$, recovering $f(x_0)$ as the supremum over all such separating affine functionals; this supremum equals $f^{\ast \ast }(x_0)$.

Fenchel duality theorem. Let $f,g$ be proper convex functions on $X$ with $\mathrm{dom}f\cap \mathrm{int}\mathrm{dom}g\ne \varnothing$. Then

$$\underset{x\in X}{\inf }(f(x)+g(x))=\underset{\phi \in X^{\ast }}{\sup }(-f^{\ast }(-\phi )-g^{\ast }(\phi )),$$

with attainment of the supremum and a primal-dual optimality condition $-\phi \in \partial f(x)$ and $\phi \in \partial g(x)$ at the optimum.

This is the convex-analytic master statement: linear programming duality, quadratic programming duality, semidefinite programming duality, and the entire optimisation literature on saddle-point conditions and KKT multipliers are special cases or refinements. The Slater interior-point assumption $\mathrm{dom}f\cap \mathrm{int}\mathrm{dom}g\ne \varnothing$ is exactly what is needed to invoke the geometric Hahn-Banach with an open convex set on one side.

Weak topologies and Banach-Alaoglu Master

The dual space $X^{\ast }$ built by Hahn-Banach carries a natural topology beyond the norm topology, the weak- topology* of pointwise convergence on $X$.

Definition. A net $({\phi }_{\alpha })$ in $X^{\ast }$ converges to $\phi$ in the weak-* topology if ${\phi }_{\alpha }(x)\to \phi (x)$ for every $x\in X$. The weak-* topology is the coarsest topology on $X^{\ast }$ for which all evaluation maps ${\mathrm{ev}}_x:\phi \mapsto \phi (x)$ are continuous.

Theorem (Banach-Alaoglu). The closed unit ball $B_{X^{\ast }}=\{\phi \in X^{\ast }:\Vert \phi \Vert \le 1\}$ is compact in the weak-* topology.

Proof. Embed $B_{X^{\ast }}$ into the product ${\prod }_{x\in X}[-\Vert x\Vert ,\Vert x\Vert ]$ via $\phi \mapsto (\phi (x){)}_{x\in X}$. The product is compact by Tychonoff's theorem (equivalent to the axiom of choice). The image of $B_{X^{\ast }}$ is the subset cut out by linearity and dominance conditions, each of which is a closed condition in the product topology. So $B_{X^{\ast }}$ embeds homeomorphically as a closed subset of a compact space, hence is compact. $\square$

The Banach-Alaoglu theorem is the workhorse compactness result of infinite-dimensional analysis. It is the reason existence proofs in the calculus of variations and PDE can be carried out by extracting weak-* convergent subsequences from norm-bounded sequences in $L^{\infty }$, $L^p$ duals, and similar spaces.

Theorem (Kakutani characterisation of reflexivity). A Banach space $X$ is reflexive if and only if its closed unit ball $B_X$ is weakly compact (compact in the weak topology induced by $X^{\ast }$ on $X$).

Theorem (Eberlein-Smulian). In a Banach space, a subset is weakly compact if and only if it is weakly sequentially compact. Equivalently, weak compactness is determined by sequences rather than nets.

Theorem (Goldstine). The canonical embedding $J:X\to X^{\ast \ast }$ has $J(B_X)$ weak-* dense in $B_{X^{\ast \ast }}$. Reflexivity is equivalent to $J(B_X)=B_{X^{\ast \ast }}$, which by Banach-Alaoglu plus Goldstine is equivalent to $J(B_X)$ being weak-* closed in $X^{\ast \ast }$.

The chain Hahn-Banach $\Rightarrow$ rich dual $\Rightarrow$ weak / weak-* topologies $\Rightarrow$ Banach-Alaoglu $\Rightarrow$ Kakutani $\Rightarrow$ existence of weak limits in reflexive spaces is the standard route to existence theorems in modern analysis. Every link in the chain rests on Hahn-Banach.

Connections Master

• Banach spaces 02.11.04 — Hahn-Banach assumes nothing more than a normed structure for its analytic form, but its corollaries (reflexivity, biduality, weak compactness) make essential use of completeness.

• Normed vector space 02.11.06 — the natural setting for both the analytic and geometric statements; the norm provides the sublinear functional in the most common applications.

• Set theory and Zorn's lemma — the general analytic Hahn-Banach uses Zorn's lemma; the separable case uses only dependent choice. The Pincus-Foreman-Wehrung results pin down the exact choice strength.

• Open mapping theorem 02.11.09 — together with Hahn-Banach, the open mapping theorem and closed graph theorem form the three pillars of Banach-space theory. Hahn-Banach builds duals; the open mapping theorem builds inverses.

• Banach-Steinhaus / uniform boundedness 02.11.10 pending — uses Hahn-Banach via the duality $\Vert x\Vert ={\sup }_{f\in B_{X^{\ast }}}|f(x)|$ to convert pointwise boundedness into uniform boundedness over the unit ball.

• Weak topologies and reflexivity 02.11.11 pending — the weak topology is defined by the family $X^{\ast }$, which is non-empty thanks to Hahn-Banach. Banach-Alaoglu, Eberlein-Smulian, and the Kakutani characterisation of reflexivity all rest on this dual.

• Convex analysis and optimisation — Lagrange duality, Fenchel duality, and minimax theorems all reduce to geometric Hahn-Banach plus the Legendre transform.

• C algebras 02.11.15 pending* — states on $C^{\ast }$-algebras are extended via Hahn-Banach in the GNS construction; the order-theoretic refinement (Kantorovich extension) is what guarantees positivity is preserved.

Representation theorems via Hahn-Banach Master

A family of celebrated representation theorems in measure theory, probability, and operator theory rest on Hahn-Banach as their existence engine.

Theorem (Riesz-Markov-Kakutani, compact form). Let $K$ be a compact Hausdorff space and let $C(K)$ denote the Banach space of real continuous functions on $K$ under the supremum norm. The dual $C(K{)}^{\ast }$ is isometrically isomorphic to the space $M(K)$ of finite signed regular Borel measures on $K$, the isomorphism sending a measure $\mu$ to the functional $f\mapsto {\int }_{K}fd\mu$.

The existence direction takes real work: given a bounded linear functional $\Lambda$ on $C(K)$, one must construct a signed measure $\mu$. The Hahn decomposition splits $\Lambda$ into a positive and a negative part; the positive part is extended to a positive linear functional on the upper-semicontinuous envelope $C(K{)}^{\uparrow }$ via Hahn-Banach, and the Daniell-Stone construction promotes that functional to a positive Borel measure. The geometric Hahn-Banach reappears in the regularity proof: open sets are approximated from below by compact sets via separation arguments.

Theorem (Krein-Milman). A non-empty compact convex subset $K$ of a locally convex Hausdorff topological vector space is the closed convex hull of its extreme points. Symbolically, $K=\overline{\mathrm{co}}(\mathrm{ext}K)$.

Proof outline. Apply Zorn's lemma to the collection of non-empty closed convex extremal subsets of $K$, ordered by reverse inclusion, to extract a minimal extremal subset, which must be a single point and therefore an extreme point of $K$. For the equality, suppose $\overline{\mathrm{co}}(\mathrm{ext}K)⊊K$: a point $x_0\in K$ outside the closed convex hull is separated from it by the geometric Hahn-Banach via a continuous linear functional $\phi$, and the face $\{y\in K:\phi (y)={\max }_K\phi \}$ is a closed extremal subset disjoint from $\overline{\mathrm{co}}(\mathrm{ext}K)$, contradicting that every extremal subset contains an extreme point.

Theorem (Stone-Weierstrass). Let $K$ be a compact Hausdorff space and let $A\subset C(K)$ be a subalgebra containing the constants and separating points. Then $A$ is dense in $C(K)$.

The standard proof goes through Dini's theorem and lattice operations, but a Hahn-Banach proof is cleaner conceptually: if $A$ is not dense, the density criterion (C2) produces a non-zero $\mu \in M(K)$ vanishing on $A$. Restricting attention to the support of $|\mu |$ and using point-separation plus the algebra property of $A$ forces $\mu$ to vanish on a dense subalgebra, hence on all of $C(K)$, contradiction.

Theorem (Choquet's representation theorem). Let $K$ be a metrizable compact convex subset of a locally convex space. For each $x\in K$, there is a Borel probability measure ${\mu }_x$ supported on $\mathrm{ext}K$ with $x={\int }_{\mathrm{ext}K}yd{\mu }_x(y)$.

Choquet sharpens Krein-Milman: the closed-convex-hull statement is upgraded to a unique-in-many-cases integral representation by extreme points. The proof glues a Krein-Milman-style Hahn-Banach separation argument with the Riesz-Markov-Kakutani correspondence between functionals and measures.

Hahn-Banach in functional-analytic optimisation Master

Two further consequences of the theorem deserve specific mention because they recur throughout applied analysis.

Theorem (closed range theorem of Banach). Let $T:X\to Y$ be a bounded linear operator between Banach spaces. The following are equivalent.

1. The range $T(X)$ is closed in $Y$.
2. The range of the adjoint $T^{\ast }(Y^{\ast })$ is closed in $X^{\ast }$.
3. $T(X)=(\ker T^{\ast }{)}^{\perp }$.
4. $T^{\ast }(Y^{\ast })=(\ker T{)}^{\perp }$.

The annihilator identities (3) and (4) are direct Hahn-Banach corollaries: $f\in T(X{)}^{\perp }$ iff $f\circ T=0$ iff $T^{\ast }f=0$, and the second annihilator equals the closure for closed convex sets via the geometric form. The closed range theorem underpins solvability theorems for linear PDE, where "the equation $Tu=v$ has a solution" reduces to "$v$ annihilates every solution of the adjoint equation $T^{\ast }\phi =0$." Fredholm alternatives are the most familiar instance.

Theorem (Bishop-Phelps). In every Banach space $X$, the set of functionals in $X^{\ast }$ that attain their norm on the closed unit ball $B_X$ is norm-dense in $X^{\ast }$.

The proof constructs a "petal" of $X^{\ast }$ near a given functional, applies geometric Hahn-Banach to support the petal at an extreme point, and iterates a contraction-like argument. The result tells us that even when the dual norm fails to be uniformly Frechet-differentiable, generic functionals behave as if there were a maximiser; this is the starting point for the Bishop-Phelps-Bollobas theorem, James's reflexivity criterion (a Banach space is reflexive iff every continuous linear functional attains its norm), and the variational principle of Ekeland.

Theorem (Mazur-Orlicz inequality). Let $X$ be a real vector space, $p:X\to \mathbb{R}$ sublinear, and $T\subset X$ a non-empty convex set. Then there exists a linear functional $\ell :X\to \mathbb{R}$ with $\ell \le p$ on $X$ and ${\inf }_{t\in T}\ell (t)={\inf }_{t\in T}p(t)$.

This is a sandwich theorem with adjustable lower bound; analytic Hahn-Banach is the special case $T=\{0\}$ and Mazur-Orlicz is the corresponding constraint-relaxation result that drives modern variational analysis (Borwein, Lewis). Many appearances of Hahn-Banach in convex optimisation are most naturally stated as Mazur-Orlicz invocations rather than as bare Hahn-Banach extensions.

Counterexamples and limits of the theorem Master

Hahn-Banach is sharp: weakening its hypotheses or asking for too much from its conclusion both lead to counterexamples that reveal the precise content of the theorem.

(L1) Sublinear majorant cannot be dropped. A linear functional $f$ on a subspace $Y$ of a real vector space $X$ with no sublinear majorant on $X$ need not extend linearly to $X$. For example, an unbounded linear functional on a dense subspace of an incomplete normed space typically has no norm-bounded extension; the norm is the natural sublinear majorant, and "unbounded" means no such majorant dominates the original.

(L2) Extension is not unique. Hahn-Banach guarantees existence but not uniqueness of the extension. In general, the collection of norm-preserving extensions of a given functional is a non-empty closed convex set in the dual ball, and it reduces to a single point exactly when the original functional has a unique Hahn-Banach extension. Spaces in which every functional on every subspace has a unique norm-preserving extension are called smooth or Phelps spaces; the dual characterisation is that $X$ is smooth iff $X^{\ast }$ is strictly convex. Smoothness fails for ${\ell }^{\infty }$, $L^{\infty }$, and $C(K)$ for most $K$, so Hahn-Banach extensions in these spaces involve a genuine choice.

(L3) The geometric form needs an interior condition. Two disjoint closed convex sets in an infinite-dimensional Banach space need not be separable by a continuous linear functional. Classic counterexample: in ${\ell }^2$, let $A$ be the closed convex hull of $\{e_n+n\cdot e_1:n\ge 2\}$ and let $B=\{0\}$. The sets are disjoint and closed, but no continuous linear functional separates them, because the closure of $A-B$ in the weak topology contains $0$. The compactness hypothesis in the second geometric form is genuinely needed.

(L4) The Hahn-Banach extension property characterises injective Banach spaces. A Banach space $X$ is injective if for every Banach space $Z$, every closed subspace $Y\subset Z$, and every bounded linear $T:Y\to X$, there is a bounded linear extension $\tilde{T}:Z\to X$ with $\Vert \tilde{T}\Vert \le \lambda \Vert T\Vert$ for some fixed $\lambda \ge 1$. Hahn-Banach is the statement that $\mathbb{R}$ and $\mathbb{C}$ are injective with $\lambda =1$; the analogous statement for vector-valued functionals does not hold for most Banach spaces. The injective spaces with $\lambda =1$ are exactly the spaces of continuous functions on extremally disconnected compact Hausdorff spaces (Goodner-Nachbin-Kelley theorem); $L^{\infty }(\mu )$ and ${\ell }^{\infty }$ are the prototype examples.

These limits sharpen the working analyst's intuition about when to expect Hahn-Banach to deliver and when to look for stronger structural input (compactness, smoothness, injectivity, reflexivity).

Hahn-Banach over locally convex spaces and ordered duals Master

The cleanest formulation of Hahn-Banach is not over a Banach space but over a locally convex topological vector space, the natural setting in which both analytic and geometric statements take their most general form. A locally convex space is a topological vector space whose topology is generated by a separating family of seminorms.

Theorem (Hahn-Banach for locally convex spaces). Let $X$ be a real locally convex space with a continuous seminorm $p$, let $Y\subset X$ be a linear subspace, and let $f:Y\to \mathbb{R}$ be a linear functional with $|f(y)|\le p(y)$ for $y\in Y$. Then $f$ extends to a continuous linear functional $F:X\to \mathbb{R}$ with $|F(x)|\le p(x)$ for every $x\in X$.

This statement underlies the duality theory of distribution spaces, tempered distributions ${\mathcal{S}}^{\prime }({\mathbb{R}}^n)$, test-function spaces $\mathcal{D}({\mathbb{R}}^n)$ and ${\mathcal{D}}^{\prime }({\mathbb{R}}^n)$, currents, hyperfunctions, and the entire framework of microlocal analysis.

Geometric form for locally convex spaces. Two disjoint convex sets in a real locally convex Hausdorff space, one open and non-empty, can be separated by a continuous linear functional. Two disjoint convex sets, one compact and one closed, can be strictly separated. The proofs proceed exactly as in the normed-space case, with the Minkowski functional of a convex neighbourhood of zero replacing the role of the norm.

Order duals and the Hahn-Banach-Kantorovich theorem. Let $X$ be an ordered topological vector space with a positive cone $X_{+}$ that is closed, convex, and pointed (meaning $X_{+}\cap (-X_{+})=\{0\}$). Define the order dual $X^{\sim }$ as the space of continuous linear functionals $\phi$ that are bounded on order intervals $[a,b]=\{x\in X:a\le x\le b\}$.

The Kantorovich-Riesz decomposition theorem says: if $X$ is a Riesz space with the order-continuous decomposition property, then every $\phi \in X^{\sim }$ decomposes uniquely as $\phi ={\phi }^{+}-{\phi }^{-}$ with ${\phi }^{+},{\phi }^{-}\in X_{+}^{\sim }$. The order-Hahn-Banach extension theorem says: a positive linear functional on an order ideal of $X$ that is dominated by a positive sublinear-monotone majorant on $X$ extends to a positive linear functional on $X$.

This machinery is what allows the GNS construction on $C^{\ast }$-algebras to deliver a faithful representation as bounded operators on a Hilbert space, what makes the Markov-Kakutani fixed-point theorem work for positive operators, and what powers the modern theory of operator spaces and operator systems.

Application to Lebesgue integration. The Daniell-Stone construction of the Lebesgue integral is a Hahn-Banach extension at heart: define the elementary integral on simple functions, extend it via positivity and dominated convergence to the upward-closure of simple functions, then to all integrable functions. The positivity condition is exactly Kantorovich extension. This route bypasses Caratheodory's outer-measure construction and gives the integral as the primitive object with measure recovered as $\mu (A)=\int {\chi }_A$.

Computational and constructive perspectives Master

Despite its non-constructive character, Hahn-Banach admits constructive refinements under specific structural assumptions on the space and the functional, and these refinements drive computational analysis.

Constructive Hahn-Banach in separable spaces. If $X$ is a separable Banach space with a Schauder basis $(e_n)$ and $f$ is a bounded linear functional defined on the closed linear span of finitely many basis vectors, the inductive construction of the proof gives an explicit recursive formula for the extension: at each step, choose the midpoint of the admissible interval $[{\sup }_m(F(m)-p(m-\mu x_0))/\mu ,{\inf }_m(p(m+\lambda x_0)-F(m))/\lambda ]$. This is a computable Hahn-Banach extension and is the basis of the Weihrauch hierarchy classification of the analytic Hahn-Banach principle.

Computable analysis perspective (Weihrauch). The analytic Hahn-Banach for separable Banach spaces is Weihrauch-reducible to the closed-choice principle on the natural numbers, $C_{\mathbb{N}}$, but not Weihrauch-reducible to weaker principles. The general (non-separable) version is Weihrauch-equivalent to the choice principle on $2^{\mathbb{N}}$, namely WKL (Weak Konig's lemma) over second-order arithmetic. The geometric form has the same complexity. This places Hahn-Banach precisely in the reverse-mathematics zone between ${\mathbf{RCA}}_0$ and ${\mathbf{ACA}}_0$.

Constructive convex analysis (Bishop). Bishop's constructive analysis develops a version of Hahn-Banach for "located" convex sets, where a set is located if there is a uniformly computable distance function. The constructive separation theorem requires both convex sets to be located and the closure of their difference to be located away from zero, conditions that are vacuous in classical mathematics but cutting in constructive settings. Specker sequences and recursive counterexamples show these hypotheses cannot be dropped without losing the theorem.

Algorithmic implications. In linear programming, the Hahn-Banach proof is replaced by Farkas' lemma, which gives a constructive certificate of infeasibility. The simplex algorithm and interior-point methods are both algorithmic realisations of duality statements that, in their continuous infinite-dimensional analogue, would invoke Hahn-Banach. The duality gap in mixed-integer programming and the polynomial-time hierarchy of convex programming both have measure-theoretic interpretations that lean on the classical theorem at the limit.

Reverse mathematics summary. Working in the standard hierarchy of subsystems of second-order arithmetic (Friedman, Simpson), the analytic Hahn-Banach for separable spaces is provable in ${\mathbf{WKL}}_0$ (Weak Konig's lemma) and not provable in ${\mathbf{RCA}}_0$ alone. The general form (over non-separable spaces) is not directly comparable to the standard hierarchy because non-separable Banach spaces are not natively coded; in set-theoretic terms, the general form is exactly the strength of the Boolean prime ideal theorem.

Historical and philosophical context Master

Hans Hahn's 1927 paper "Uber lineare Gleichungssysteme in linearen Raumen" introduced the extension theorem in the real case, motivated by the moment problem and the systematic treatment of bounded linear functionals on sequence spaces. Stefan Banach independently rediscovered and strengthened the theorem in his 1929 papers in Studia Mathematica, packaging it as a foundational tool of the new theory of linear operators. Banach's 1932 monograph "Theorie des operations lineaires" placed the result at the centre of functional analysis ^{[Brezis §1]}.

The complex case waited until 1938, when Bohnenblust and Sobczyk supplied the now-standard reduction by separating real and imaginary parts of a complex-linear functional. The geometric reformulation in terms of separating hyperplanes is implicit in Minkowski's late-nineteenth-century work on convex bodies and was developed systematically by Mazur, Eidelheit, and Krein in the 1930s, who recognised the duality between extension and separation.

Philosophically, Hahn-Banach is the first place a student of analysis encounters non-constructive existence: the theorem says an extension exists without telling you how to find it. The proof via Zorn's lemma is genuinely non-constructive, and Pincus's 1972 work and the later Foreman-Wehrung result clarified that this is not an accident of the proof but an essential feature of the theorem itself. Working in ZF alone, one cannot prove Hahn-Banach; working in ZF + dependent choice, one can prove it for separable spaces but not in general; the full theorem requires the Boolean prime ideal theorem, which sits strictly between dependent choice and full choice.

This sensitivity to set-theoretic foundations is more than a curiosity. It explains why constructive and computable analysis (Bishop, Weihrauch) struggles with infinite-dimensional dualities, why the descriptive set theory of Banach spaces (Bossard, Argyros) is delicate, and why certain seemingly natural functionals (Banach limits, finitely additive measures on $\mathbb{N}$, ultrafilters) all share the same "Hahn-Banach scent": their existence is provable from BPI but not from ZF alone.

The unifying viewpoint that emerged in the second half of the twentieth century, due to Simons, Konig, Borwein, and others, presents analytic Hahn-Banach, geometric Hahn-Banach, Lagrange duality, Fenchel duality, and the minimax theorem as five faces of one underlying sandwich theorem about sublinear and superlinear functions. This unification clarifies the logical content of each form and the price each pays in side conditions.

Bibliography Master

• Hahn, H. Uber lineare Gleichungssysteme in linearen Raumen. Journal fur die reine und angewandte Mathematik 157 (1927), 214-229.
• Banach, S. Sur les fonctionnelles lineaires. Studia Mathematica 1 (1929), 211-216 and 223-239.
• Banach, S. Theorie des operations lineaires. Warsaw, 1932.
• Bohnenblust, H. F. and Sobczyk, A. Extensions of functionals on complex linear spaces. Bulletin of the American Mathematical Society 44 (1938), 91-93.
• Brezis, H. Functional Analysis, Sobolev Spaces and Partial Differential Equations. Springer, 2011. §1.
• Conway, J. B. A Course in Functional Analysis. 2nd ed., Springer, 1990. §III.
• Rudin, W. Functional Analysis. 2nd ed., McGraw-Hill, 1991. §3.
• Reed, M. and Simon, B. Methods of Modern Mathematical Physics, Vol. I. Academic Press, 1980. §III.
• Pincus, D. The strength of the Hahn-Banach theorem. Lecture Notes in Mathematics 369, Springer, 1974, 203-248.
• Foreman, M. and Wehrung, F. The Hahn-Banach theorem implies the existence of a non-Lebesgue measurable set. Fundamenta Mathematicae 138 (1991), 13-19.
• Kantorovich, L. V. Functional Analysis and Applied Mathematics. Uspekhi Mat. Nauk 3 (1948), 89-185 (English trans. Nat. Bur. Standards, 1952).
• Simons, S. From Hahn-Banach to Monotonicity. Lecture Notes in Mathematics 1693, 2nd ed., Springer, 2008.

---

Wave A unit, missing-units backfill. Produced as the Hahn-Banach foundation for the open mapping, Banach-Steinhaus, weak-topology, and convex-optimisation strands.