Hahn-Banach Theorem

Why Hahn-Banach matters

Without Hahn-Banach the dual space X* of a normed space could be the zero space — there would be no general reason to believe non-trivial continuous linear functionals exist. Hahn-Banach turns the dual into a rich object that mirrors X faithfully. Almost every theorem in functional analysis that uses the dual either invokes Hahn-Banach directly or relies on a downstream consequence (Banach-Alaoglu, James's theorem, the bipolar theorem).

• Existence of non-trivial functionals. For any non-zero x ∈ X, Hahn-Banach gives a functional Φ ∈ X* with Φ(x) = ‖x‖ and ‖Φ‖ = 1. The dual space therefore separates points: distinct vectors are distinguished by some functional, so the canonical embedding X → X** is injective.
• Weak topologies and Banach-Alaoglu. The weak topology on X is the coarsest topology making every Φ ∈ X* continuous. Hahn-Banach ensures this topology has enough functionals to be Hausdorff and to genuinely refine the weak-* topology on X*. The unit ball of X* is weak-*-compact (Banach-Alaoglu), a theorem central to existence proofs in PDE and calculus of variations.
• Convex separation. The geometric Hahn-Banach (separation theorem) says any two disjoint convex sets — one of them open — sit on opposite sides of a closed hyperplane. This is the foundation of Lagrangian duality, the Karush-Kuhn-Tucker conditions, support functions in convex analysis, and the bipolar theorem A° ° = closed convex hull of A.
• Distributions and generalized functions. Schwartz's space of tempered distributions S′(ℝⁿ) is the topological dual of the rapidly-decreasing functions S(ℝⁿ). Distributions like δ, principal-value 1/x, and the Fourier transform of x^k are made precise as continuous linear functionals; their existence and extension properties rest on Hahn-Banach machinery.
• Optimization and economics. The first and second welfare theorems in microeconomics, Lagrangian duality in convex optimization, the separating-hyperplane theorem in game theory, and Farkas's lemma in linear programming are all instances of geometric Hahn-Banach applied to convex cones and polyhedra.
• Control theory. The Pontryagin maximum principle proves optimality of a control trajectory by separating the reachable set from the target with a hyperplane; the costate (adjoint variable) is the normal to that hyperplane. Without Hahn-Banach this central argument breaks.
• Generalized limits and Banach limits. A Banach limit is a translation-invariant linear functional on ℓ^∞(ℕ) extending the ordinary limit on convergent sequences. It is constructed via Hahn-Banach: the sublinear functional p(x) = lim sup (1/n) Σ x_k bounds the partial-Cesàro averages, and the extension gives Lim(x) for every bounded sequence including non-convergent ones.

Precise statements

Three commonly used forms — all equivalent, each useful in different settings.

Analytic form (sublinear dominated). Let X be a real vector space, p: X → ℝ a sublinear functional (p(x + y) ≤ p(x) + p(y) and p(λx) = λ p(x) for λ ≥ 0), Y ⊂ X a linear subspace, and φ: Y → ℝ linear with φ(y) ≤ p(y) for all y ∈ Y. Then there exists a linear extension Φ: X → ℝ with Φ|_Y = φ and Φ(x) ≤ p(x) for all x ∈ X.

Normed-space form. Let X be a normed vector space (real or complex), Y ⊂ X a linear subspace, and φ: Y → 𝕂 a continuous linear functional. Then there exists a continuous linear extension Φ: X → 𝕂 with Φ|_Y = φ and ‖Φ‖_{X*} = ‖φ‖_{Y*}.

Geometric form (Hahn-Banach separation). Let X be a normed space and A, B ⊂ X be disjoint non-empty convex sets with A open. Then there exists Φ ∈ X* and α ∈ ℝ with Re Φ(a) < α ≤ Re Φ(b) for all a ∈ A, b ∈ B. If both A and B are closed and one is compact, the inequality can be made strict on both sides.

The complex form is reduced to the real one by treating ℂ as ℝ² (real and imaginary parts of φ) and using the fact that a complex-linear functional is determined by its real part.

Proof sketch via Zorn's lemma

The classical proof goes by transfinite induction or, equivalently, Zorn's lemma. The argument has two parts.

Step 1 — extension by one dimension. Given the functional φ on Y, pick x₀ ∉ Y. The subspace Y₁ = Y ⊕ ℝ·x₀ has dimension one more than Y. Any extension Φ₁ of φ to Y₁ is determined by the single number c = Φ₁(x₀); the requirement Φ₁ ≤ p forces c into a non-empty closed interval [sup_{y ∈ Y} (φ(y) − p(y − x₀)), inf_{y ∈ Y} (p(y + x₀) − φ(y))]. Sublinearity of p makes this interval non-empty; any c in it gives a one-step extension.

Step 2 — Zorn's lemma. Consider the partially ordered set P of all dominated extensions (Z, ψ) where Y ⊂ Z ⊂ X and ψ extends φ with ψ ≤ p on Z, ordered by extension. Every chain has an upper bound (the union). By Zorn, P has a maximal element (Z*, Φ). If Z* ≠ X, pick x₀ ∉ Z* and use Step 1 to extend — contradicting maximality. So Z* = X, and Φ is the desired extension.

For separable normed spaces, Step 1 alone suffices (no Zorn needed): enumerate a countable dense subset and extend one direction at a time. Separable Hahn-Banach is provable in ZF + countable choice. The non-separable case genuinely needs Zorn (or the Boolean prime ideal theorem BPI, which is strictly weaker than AC).

Worked example — extending from a line in ℝ²

Consider X = ℝ² with the ℓ^∞ norm ‖(a, b)‖_∞ = max(|a|, |b|). Let Y = {(t, 0) : t ∈ ℝ} be the x-axis, and define φ(t, 0) = t. Then ‖φ‖_{Y*} = 1.

An extension is determined by Φ(0, 1) = c for some c ∈ ℝ; linearity gives Φ(a, b) = a + cb. The dual norm of Φ in the ℓ^∞ → ℝ pairing is ‖Φ‖ = sup_{‖(a, b)‖_∞ ≤ 1} |a + cb| = 1 + |c|. The constraint ‖Φ‖ = 1 forces c = 0 — but only if we insist on equality. If we allow any c ∈ [−1, 1] the operator norm is 1 + |c| ≤ 2, which is fine if our sublinear bound was p(a, b) = 2 max(|a|, |b|), but is too large for norm preservation.

Replacing ℓ^∞ with ℓ^1: ‖(a, b)‖_1 = |a| + |b|, dual norm ‖Φ‖ = max(1, |c|). Norm preservation ‖Φ‖ = 1 now allows any c ∈ [−1, 1] — a whole interval of valid extensions. This illustrates non-uniqueness exactly when the dual unit ball (here, the ℓ^∞-disc {|a| ≤ 1, |c| ≤ 1}) has a flat face at the relevant point.

Analytic vs geometric forms

The two faces of the theorem look different but encode the same content. A geometric separation gives an analytic extension by taking the linear functional whose level set is the separating hyperplane; an analytic extension gives a geometric separation by taking the level set of the extended functional. The table summarizes the parallel structure.

Form	Input	Output	Key inequality	Where used	Equivalent to
Analytic (sublinear)	p sublinear on X, φ linear on Y ⊂ X with φ ≤ p	Linear extension Φ on X with Φ ≤ p	Φ(x) ≤ p(x)	Construction of Banach limits, invariant means	Mazur's theorem on barreled spaces
Normed (norm-preserving)	Continuous linear φ on Y ⊂ normed X	Continuous linear Φ on X with ‖Φ‖_{X*} = ‖φ‖_{Y*}	|Φ(x)| ≤ ‖φ‖·‖x‖	Duality theory, X separates points	Existence of supporting functionals on convex bodies
Geometric (separation)	Disjoint convex A, B with A open	Closed hyperplane Φ = α between them	Φ(a) < α ≤ Φ(b)	Convex optimization, welfare theorems	Bipolar theorem A°° = co̅(A ∪ {0})
Strict separation	Disjoint closed convex A, B with one compact	Strict closed hyperplane	Φ(a) < α₁ < α₂ < Φ(b)	Constructing dual variables, support hyperplanes	Closure of A − B does not contain 0
Mazur's theorem	x not in closed convex hull of S	Functional strictly separating x from co̅(S)	Φ(x) < inf_S Φ	Weak closure equals strong closure for convex sets	Strong-closed convex = weak-closed convex
Complex form	ℂ-linear φ on Y ⊂ complex X	ℂ-linear Φ on X with ‖Φ‖ preserved	|Φ(x)| ≤ ‖φ‖·‖x‖	Complex Banach space duality, holomorphic extension	Real Hahn-Banach applied to Re φ

Non-uniqueness of the extension

Hahn-Banach guarantees existence; uniqueness is a separate question. The set of norm-preserving extensions of φ is always a non-empty convex set, often a single point but sometimes a continuum.

• Uniqueness holds. When the dual space X* is strictly convex (no flat segments on the boundary of the unit ball). Hilbert spaces, L^p for 1 < p < ∞, and uniformly convex Banach spaces have this property — every functional on a subspace has exactly one norm-preserving extension. The geometric reason: the support hyperplane to the dual ball at a boundary point is unique iff the boundary is strictly convex.
• Uniqueness fails. In ℓ^1, ℓ^∞, c₀, L^1, L^∞, and C(K) for general compact K. The dual unit ball has flat faces; multiple distinct functionals achieve the same maximum on the unit ball of Y, and any of them lifts to a valid extension. The set of extensions is an exposed face of the dual unit ball, parametrized by the kernel direction of the original functional's failure to determine the extension uniquely.
• Bishop-Phelps theorem. Even when extensions are non-unique, the set of functionals attaining their norm on the unit ball is norm-dense in X* (Bishop-Phelps 1961). So every functional can be norm-approximated by one with a clean support hyperplane.
• Banach limits as extreme non-uniqueness. The Banach limit on ℓ^∞(ℕ) extends the ordinary limit on c (convergent sequences) — but there are uncountably many distinct Banach limits, differing on non-Cesàro-convergent sequences. The set of Banach limits is a weak-* compact convex face of the unit ball of ℓ^∞(ℕ)*, parametrized in some sense by translation-invariant probability measures on βℕ \ ℕ.

Hahn-Banach and the axiom of choice

Hahn-Banach occupies an interesting place on the choice spectrum. It is strictly stronger than countable choice (which suffices for separable spaces) but strictly weaker than the full axiom of choice.

• Reduces to Zorn's lemma. The transfinite-induction proof uses Zorn to extend one direction at a time. Zorn is equivalent to full AC, so this is the strongest proof needed.
• Boolean prime ideal theorem (BPI). Łoś & Ryll-Nardzewski (1951) showed Hahn-Banach is equivalent over ZF to BPI: every non-trivial Boolean algebra has a prime ideal. BPI is strictly weaker than AC (Halpern-Lévy 1971) but not provable in ZF alone.
• Non-constructive consequences. Hahn-Banach plus measurability arguments construct non-Lebesgue-measurable sets in ℝ (Pincus-Solovay): there is a model of ZF + Hahn-Banach in which every set of reals is measurable, but the construction of a non-measurable set using Hahn-Banach proceeds via choice-style arguments.
• Separable case is constructive. For X separable, the proof by enumerating a countable dense subspace and extending one direction at a time uses only countable dependent choice (DC). Most applications in concrete spaces (ℓ^p, L^p separable) need only this weaker fragment.
• Costed claim. Hahn-Banach in full generality needs the axiom of choice. Specifically, full Hahn-Banach is equivalent to the Boolean prime ideal theorem BPI; BPI is independent of ZF, strictly weaker than AC but unprovable in ZF alone.

Related results and corollaries

• Banach-Alaoglu theorem. The closed unit ball of X* is compact in the weak-* topology. The proof embeds the unit ball into a product of intervals (one for each x ∈ X) and uses Tychonoff. Hahn-Banach is needed to show this embedding is into the right ambient space — the dual has enough functionals to make the embedding non-trivial.
• Goldstine's theorem. The canonical image of the closed unit ball of X in X** is weak-*-dense in the closed unit ball of X**. Combined with Banach-Alaoglu, this provides the workhorse weak compactness arguments in reflexivity proofs.
• James's theorem. X is reflexive iff every continuous linear functional on X attains its norm on the unit ball. A deep theorem (James 1957) characterizing reflexivity in terms of norm-attaining functionals; Hahn-Banach provides the dual setting in which the statement makes sense.
• Bipolar theorem. For A ⊂ X, A°° (the bipolar) equals the weak-* closed convex hull of A ∪ {0}. The proof uses Hahn-Banach separation between a hypothetical point outside the bipolar and the polar set's annihilator.
• Mazur's theorem. A convex set is strongly closed iff it is weakly closed. Hahn-Banach separation between a point not in the strong closure and the closed convex set gives the separating functional that witnesses non-membership in the weak closure.
• Farkas's lemma. A linear inequality Ax ≤ b has a solution iff every non-negative combination y ≥ 0 with y^T A = 0 satisfies y^T b ≥ 0. This is finite-dimensional Hahn-Banach applied to polyhedral cones; the proof is the same separation argument.

Hahn-Banach versus the other pillars

Hahn-Banach is one of the four classical pillars of functional analysis; the others are open mapping, closed graph, and Banach-Steinhaus. The table contrasts what each needs and produces.

Theorem	Setting needed	Foundational tool	What it produces	Choice level
Hahn-Banach	Normed space + subspace + functional	Zorn's lemma / sublinear domination	Norm-preserving extension Φ ∈ X*	BPI (weaker than AC)
Open mapping	Surjective bounded T: X → Y, both complete	Baire category theorem	T is an open map; bounded inverse if bijective	Dependent choice (DC)
Closed graph	Linear T: X → Y, both complete	Open mapping applied to graph	T bounded ⇔ graph closed in X × Y	Dependent choice (DC)
Banach-Steinhaus	Family of bounded operators, X complete	Baire category theorem	Pointwise-bounded ⇒ uniformly bounded	Dependent choice (DC)
Banach-Alaoglu	Normed space X (no completeness needed)	Tychonoff + Hahn-Banach	Unit ball of X* is weak-*-compact	Equivalent to BPI
Krein-Milman	Compact convex set in locally convex space	Hahn-Banach separation	Compact convex set = closed convex hull of its extreme points	BPI

Common pitfalls

• "Extension is unique." Only in strictly convex duals. In general, multiple norm-preserving extensions exist — ℓ^∞ on ℝ² already shows a full interval of valid extensions. The set of extensions is always a convex face of the dual unit ball.
• "Hahn-Banach gives a constructive recipe." The proof is non-constructive: Zorn's lemma asserts existence of a maximal extension without specifying which one. For separable spaces a recursive choice along a dense subset gives a specific extension, but the choices made at each step have to be supplied externally.
• "Sublinear means convex and homogeneous." Sublinear means subadditive (p(x + y) ≤ p(x) + p(y)) and positively homogeneous (p(λx) = λp(x) for λ ≥ 0). Norms are sublinear but so are p(x) = lim sup x_n on ℓ^∞ (which is not a norm). Sublinear is weaker than norm — it can be 0 or negative on non-zero vectors.
• "Separation works for any disjoint convex sets." You need at least one to be open, or one to be compact and both closed. Two disjoint closed convex sets in infinite dimensions can fail to be separable by a closed hyperplane — examples exist in ℓ^∞.
• "Hahn-Banach works in any vector space." The analytic form (sublinear dominated) works in any real vector space without topology. The normed-space form requires a norm; the geometric separation form requires at least a topological vector space structure. The proofs differ slightly: the analytic case is "no topology" Zorn, while the geometric case relies on the Minkowski functional construction.
• "Banach limit is a unique extension of lim." No — there are uncountably many distinct Banach limits. They all agree on convergent sequences (giving the ordinary limit) but disagree on bounded oscillating ones like (1, 0, 1, 0, …). The set of Banach limits is a non-trivial weak-* compact convex face of the unit ball of ℓ^∞(ℕ)*.

History

Hahn proved the real case for normed spaces in 1927 ("Über lineare Gleichungssysteme in linearen Räumen"), motivated by problems on moment sequences and integral equations. Banach independently rediscovered and extended the theorem in his 1929 paper and his 1932 monograph "Théorie des opérations linéaires," giving the modern formulation. The sublinear form (general analytic version) is due to Banach.

The geometric separation version was developed by Banach in the same period and refined by Mazur, Krein, Šmulian, and others in the Lwów school. The complex case was added by Bohnenblust and Sobczyk (1938). The connection with the axiom of choice was worked out by Łoś and Ryll-Nardzewski (1951), placing Hahn-Banach at the BPI level — strictly between countable choice and full AC. Pincus and Solovay (1972) gave models of ZF where Hahn-Banach holds but full AC fails.

Modern uses span convex optimization (Rockafellar's 1970 monograph systematized convex duality on these foundations), control theory (Pontryagin maximum principle, 1956), and mathematical economics (Arrow-Debreu equilibrium, 1954, used separation to prove existence of equilibrium prices). The theorem remains an almost daily tool of working analysts.