In functional analysis, a recurring task is to construct bounded linear functionals with prescribed behaviour. You may need a functional that attains a certain value on a specific vector. You may need one that vanishes on a given subspace but not on a particular element outside it. You may need to certify that two distinct points in a [Banach space](/page/Banach%20Space) are genuinely different by exhibiting a functional that separates them. In finite dimensions, such constructions are routine: extend a basis, define the functional by its values on each basis vector, and check linearity. In infinite dimensions, this strategy collapses. There is no countable algebraic basis to work with in general, and even when one exists (a Hamel basis), the resulting functional may fail to be bounded, rendering it useless for analysis. The Hahn-Banach theorem resolves this difficulty. It guarantees that a bounded linear functional defined on a subspace of a [normed vector space](/page/Normed%20Vector%20Space) can always be extended to the entire space without increasing its norm. The theorem comes in several forms --- an algebraic version for arbitrary real vector spaces, an analytic version for normed spaces, and a geometric version that separates convex sets by hyperplanes --- and each form serves as a foundational tool for different parts of the theory. The consequences are profound. The theorem ensures that the [dual space](/page/Dual%20Space) $X^*$ of any normed space $X$ is rich enough to separate points, to recover the norm of every element, and to support a robust duality theory. Without it, functionals on subspaces would be trapped there, dual spaces might be trivial, and the entire program of weak topologies, reflexivity, and variational methods would be untenable. [example: The Extension Problem in Finite Dimensions] Consider $\mathbb{R}^3$ with the Euclidean norm and the subspace $W = \{(x_1, x_2, 0) : x_1, x_2 \in \mathbb{R}\}$. Define the functional $f \in W^*$ by $f(x_1, x_2, 0) = 3x_1 - x_2$. Its operator norm on $W$ is: \begin{align*} \|f\|_{W^*} = \sup_{\substack{(x_1, x_2, 0) \neq 0}} \frac{|3x_1 - x_2|}{(x_1^2 + x_2^2)^{1/2}} = (3^2 + (-1)^2)^{1/2} = \sqrt{10}. \end{align*} Any extension $\tilde{f}: \mathbb{R}^3 \to \mathbb{R}$ must take the form $\tilde{f}(x_1, x_2, x_3) = 3x_1 - x_2 + cx_3$ for some $c \in \mathbb{R}$. Its norm on $\mathbb{R}^3$ is $\|\tilde{f}\|_{(\mathbb{R}^3)^*} = (9 + 1 + c^2)^{1/2} = (10 + c^2)^{1/2}$, which is minimised when $c = 0$. The unique norm-preserving extension is $\tilde{f}(x_1, x_2, x_3) = 3x_1 - x_2$. In finite dimensions, the extension is straightforward and the norm-preserving choice is often unique (here, by orthogonality). In infinite dimensions, neither the construction nor the uniqueness is guaranteed without the Hahn-Banach theorem. Extensions always exist algebraically (via a Hamel basis), but they may increase the norm or fail to be continuous altogether. [/example] ## Definition The central objects in the Hahn-Banach theorem are sublinear functionals, which serve as the dominating functions that control the extension. A sublinear functional is not a norm --- it need not be symmetric --- but it captures just enough structure (positive homogeneity and the triangle inequality) to constrain the values of a linear functional from above. [definition: Sublinear Functional] Let $V$ be a real vector space. A function $p: V \to \mathbb{R}$ is called a **sublinear functional** if it satisfies: 1. **Positive homogeneity:** $p(\lambda v) = \lambda \, p(v)$ for all $\lambda > 0$ and $v \in V$. 2. **Subadditivity:** $p(v_1 + v_2) \le p(v_1) + p(v_2)$ for all $v_1, v_2 \in V$. [/definition] Every norm is a sublinear functional (with the additional property that $p(-v) = p(v)$), but the class is strictly larger. For instance, if $f: V \to \mathbb{R}$ is any linear functional, then $p(v) = f(v)$ is sublinear (it satisfies both conditions by linearity), even though $p(-v) = -p(v) \neq p(v)$ in general. The Minkowski functional (gauge) of a convex set containing the origin is another important example, and it plays a key role in the geometric form of the theorem. [example: The Minkowski Functional] Let $X$ be a real [normed vector space](/page/Normed%20Vector%20Space), and let $C \subset X$ be a convex, open set with $0 \in C$. The **Minkowski functional** (or gauge) of $C$ is \begin{align*} \mu_C: X &\to \mathbb{R} \\ x &\mapsto \inf\{t > 0 : x \in tC\}. \end{align*} We verify that $\mu_C$ is sublinear. For positive homogeneity: given $\lambda > 0$, \begin{align*} \mu_C(\lambda x) = \inf\{t > 0 : \lambda x \in tC\} = \inf\{t > 0 : x \in (t/\lambda) C\} = \lambda \inf\{s > 0 : x \in sC\} = \lambda\, \mu_C(x), \end{align*} where we substituted $s = t/\lambda$. For subadditivity: given $x_1, x_2 \in X$ and $\varepsilon > 0$, choose $t_1, t_2 > 0$ with $x_1 \in t_1 C$, $x_2 \in t_2 C$, $t_1 < \mu_C(x_1) + \varepsilon$, and $t_2 < \mu_C(x_2) + \varepsilon$. Write $x_1 = t_1 c_1$ and $x_2 = t_2 c_2$ with $c_1, c_2 \in C$. Then \begin{align*} x_1 + x_2 = (t_1 + t_2)\left(\frac{t_1}{t_1 + t_2} c_1 + \frac{t_2}{t_1 + t_2} c_2\right). \end{align*} The term in parentheses is a convex combination of $c_1$ and $c_2$, so it lies in $C$ by convexity. Therefore $x_1 + x_2 \in (t_1 + t_2)C$, giving $\mu_C(x_1 + x_2) \le t_1 + t_2 < \mu_C(x_1) + \mu_C(x_2) + 2\varepsilon$. Since $\varepsilon > 0$ was arbitrary, $\mu_C(x_1 + x_2) \le \mu_C(x_1) + \mu_C(x_2)$. The Minkowski functional also characterises membership in $C$: since $C$ is open, $x \in C$ if and only if $\mu_C(x) < 1$. This equivalence is what makes $\mu_C$ the bridge between the analytic and geometric forms of the Hahn-Banach theorem. [/example] ## The Algebraic Extension Theorem The core of the Hahn-Banach theory is a purely algebraic result that requires no topology. It states that a linear functional dominated by a sublinear functional on a subspace can be extended to the full vector space while preserving the domination. The proof relies on Zorn's lemma (equivalently, the Axiom of Choice), and this non-constructive character is an essential feature: in general, the extension cannot be made explicit, and it is not unique. [quotetheorem:879] Several aspects of this statement deserve emphasis. First, the domination condition $f(w) \le p(w)$ is a *one-sided* inequality --- it does not assert $|f(w)| \le p(w)$. This asymmetry is deliberate: a sublinear functional $p$ satisfies $p(-v) \neq p(v)$ in general, so the two-sided bound is not available from $p$ alone. However, when $p$ happens to be a seminorm (so that $p(-v) = p(v)$), the one-sided bound $\tilde{f}(v) \le p(v)$ applied to both $v$ and $-v$ gives the absolute bound $|\tilde{f}(v)| \le p(v)$. Second, the theorem asserts existence but not uniqueness. The extension depends on the choice of $\tilde{f}(v_0)$ at each step of the inductive process (or, in the Zorn's lemma formulation, on the choice of maximal element). Different choices of $\tilde{f}(v_0)$ at the codimension-one step yield different extensions. Uniqueness holds only when $W$ is dense in $V$ and $p$ is continuous (forcing the extension to be determined by continuity). Third, the theorem is purely algebraic: no topology is assumed on $V$, no continuity is required of $f$ or $p$, and the extension $\tilde{f}$ is not claimed to be continuous. Continuity enters only in the normed-space version, where the choice of a specific $p$ (namely, $p(v) = \|f\|_{W^*} \cdot \|v\|_V$) ensures it. ## The Norm-Preserving Extension The algebraic theorem becomes a powerful analytic tool once we specialise the sublinear functional $p$ to the norm. The resulting statement answers the motivating question directly: bounded linear functionals on subspaces extend to the whole space with no increase in norm. To pass from the algebraic to the analytic version, the key observation is that if $f: W \to \mathbb{R}$ is a bounded linear functional on a subspace $W$ of a normed space $V$, then the function \begin{align*} p: V &\to \mathbb{R} \\ v &\mapsto \|f\|_{W^*} \cdot \|v\|_V \end{align*} is a sublinear functional (positive homogeneity comes from the homogeneity of the norm, and subadditivity from the triangle inequality). Moreover, $f(w) \le |f(w)| \le \|f\|_{W^*} \cdot \|w\|_V = p(w)$ for all $w \in W$, so the domination condition of the algebraic theorem is satisfied. The algebraic extension $\tilde{f}$ then satisfies $\tilde{f}(v) \le \|f\|_{W^*} \cdot \|v\|_V$ for all $v \in V$. Applying this to $-v$ as well gives $|\tilde{f}(v)| \le \|f\|_{W^*} \cdot \|v\|_V$, which is precisely the statement that $\tilde{f}$ is bounded with $\|\tilde{f}\|_{V^*} \le \|f\|_{W^*}$. The reverse inequality $\|\tilde{f}\|_{V^*} \ge \|f\|_{W^*}$ is immediate since $W \subset V$. [quotetheorem:880] The norm-preserving property $\|\tilde{f}\|_{V^*} = \|f\|_{W^*}$ is optimal: any extension must have norm at least $\|f\|_{W^*}$ (since $V$ contains $W$), and the Hahn-Banach extension achieves this lower bound. The extension is generally not unique when $W$ is a proper closed subspace of an infinite-dimensional space. In a [Hilbert space](/page/Hilbert%20Space), however, the Riesz representation theorem provides a canonical extension (via orthogonal projection), and in strictly convex spaces, the norm-preserving extension is unique. [remark: The Complex Case] The theorem extends to complex normed spaces, but the argument requires an additional step. If $V$ is a complex normed space and $f: W \to \mathbb{C}$ is a bounded complex-linear functional, one first restricts to the real part $g := \operatorname{Re}(f): W \to \mathbb{R}$, which is a bounded real-linear functional with $\|g\|_{W^*} \le \|f\|_{W^*}$. One extends $g$ to $\tilde{g}: V \to \mathbb{R}$ by the real Hahn-Banach theorem, and then recovers the complex extension via \begin{align*} \tilde{f}(v) := \tilde{g}(v) - i\,\tilde{g}(iv) \quad \text{for all } v \in V. \end{align*} This formula exploits the identity $f(v) = \operatorname{Re}(f)(v) - i\,\operatorname{Re}(f)(iv)$, which holds for any complex-linear functional $f$ (since $f(iv) = if(v)$ implies $\operatorname{Re}(f(iv)) = -\operatorname{Im}(f(v))$). The resulting $\tilde{f}$ is complex-linear with $\|\tilde{f}\|_{V^*} = \|f\|_{W^*}$. [/remark] ## Consequences for the Dual Space The norm-preserving extension theorem has three immediate consequences that establish the dual space as a faithful and powerful representation of the original space. These results answer the question posed at the outset: is $X^*$ rich enough to detect the structure of $X$? ### Existence of Support Functionals The first question is quantitative: given $x \in X$ with $x \neq 0$, can we find $f \in X^*$ that not only distinguishes $x$ from zero but actually *achieves* the norm $\|x\|_X$ as its value? The answer is yes, and the construction is a direct application of the norm-preserving extension. Start with the one-dimensional subspace $W = \operatorname{span}\{x\}$ and define the functional $f: W \to \mathbb{R}$ by $f(\lambda x) = \lambda \|x\|_X$. Then $f$ is linear with $\|f\|_{W^*} = 1$ (since $|f(\lambda x)| = |\lambda| \cdot \|x\|_X = \|\lambda x\|_X$ for all $\lambda \in \mathbb{R}$). By the norm-preserving extension theorem, $f$ extends to $f_x \in X^*$ with $\|f_x\|_{X^*} = 1$ and $f_x(x) = \|x\|_X$. [quotetheorem:881] The support functional has a geometric interpretation: the hyperplane $H = \{y \in V : f_v(y) = \|v\|_V\}$ is tangent to the closed ball $\overline{B}(0, \|v\|_V)$ at the point $v$. In a [Hilbert space](/page/Hilbert%20Space), the support functional is given by the Riesz representative $f_v = (\cdot, v/\|v\|_V)_H$, but in a general Banach space, the support functional may not be unique (non-uniqueness is equivalent to the norm not being Gateaux differentiable at $v$). The theorem immediately yields the **norm recovery formula**: \begin{align*} \|x\|_X = \max_{\substack{f \in X^* \\ \|f\|_{X^*} \le 1}} f(x) \quad \text{for all } x \in X. \end{align*} The supremum is attained (hence the maximum) because the support functional $f_x$ achieves it. This formula shows that the norm on $X$ is entirely determined by the dual space: it is the supremum of a family of continuous affine functions, which immediately implies that the norm is convex and [weakly](/page/Weak%20Topology) lower semicontinuous. ### Separation of Points The existence of support functionals answers a qualitative question as well: the dual space separates points. [quotetheorem:882] Part (ii) follows from part (i) applied to $v - w \neq 0$: the support functional $f_{v-w}$ satisfies $f_{v-w}(v) - f_{v-w}(w) = f_{v-w}(v-w) = \|v - w\|_V > 0$. This result is the foundation for the entire duality program. It guarantees: - The canonical embedding $J: X \to X^{**}$ defined by $J(x)(f) = f(x)$ is injective (and isometric, by the norm recovery formula). This makes $X$ identifiable with a subspace of its bidual, and the question of whether $J$ is surjective defines [reflexivity](/page/Reflexive%20Space). - The [weak topology](/page/Weak%20Topology) $\sigma(X, X^*)$ is Hausdorff. If weak limits are unique, it is because functionals in $X^*$ separate points. - A bounded linear operator $T: X \to Y$ between normed spaces is determined by its action via duality: $T = 0$ if and only if $f(Tx) = 0$ for all $f \in Y^*$ and $x \in X$. [example: Separation Fails Without a Norm] The Hahn-Banach theorem requires a sublinear dominating function, which in the normed-space setting comes from the norm itself. Without a norm, the dual space can collapse entirely. Consider the space $X = L^p([0,1], \mathcal{L}^1)$ for $0 < p < 1$. This is a complete metric space under the metric $d(f,g) = \int_0^1 |f(x) - g(x)|^p \, d\mathcal{L}^1(x)$ (the integral converges because $|f-g|^p \le |f-g|$ for small values). However, $L^p$ for $p < 1$ is not a normed space: the function $\|f\|_p = \left(\int_0^1 |f|^p \, d\mathcal{L}^1\right)^{1/p}$ does not satisfy the triangle inequality. A theorem of Day (1940) shows that the only continuous linear functional on $L^p([0,1])$ for $0 < p < 1$ is the zero functional: \begin{align*} (L^p([0,1]))^* = \{0\} \quad \text{for } 0 < p < 1. \end{align*} The dual space is trivial and cannot separate any two distinct points. This does not contradict the Hahn-Banach theorem because $L^p$ for $p < 1$ is not a normed space (nor even a locally convex topological vector space), so the dominating sublinear functional required by the theorem does not exist. The underlying geometric reason is that $L^p$ for $p < 1$ has no nontrivial convex open sets: the unit ball $\{f : \|f\|_p < 1\}$ is not convex when $p < 1$, and there is nothing for a separating hyperplane to "lean on." [/example] ### The Norm Recovery Formula and the Bidual The support functional theorem provides even more: it gives a quantitative duality between $X$ and $X^*$ that is captured by two complementary norm formulas. In addition to the formula $\|x\|_X = \max_{\|f\| \le 1} f(x)$ for elements of $X$, there is a parallel formula for elements of the dual: \begin{align*} \|f\|_{X^*} = \sup_{\substack{x \in X \\ \|x\|_X \le 1}} |f(x)| \quad \text{for all } f \in X^*. \end{align*} This is the definition of the operator norm. The significance is that while the second formula holds by definition, the first formula is a *theorem* (it requires Hahn-Banach). Together, the two formulas establish a perfect symmetry: the norm on $X$ is recovered from $X^*$, and the norm on $X^*$ is recovered from $X$. This symmetry is formalised by the canonical embedding $J: X \to X^{**}$, defined by \begin{align*} J: X &\to X^{**} \\ x &\mapsto \hat{x}, \quad \text{where } \hat{x}(f) := f(x) \text{ for all } f \in X^*. \end{align*} The norm recovery formula shows that $\|\hat{x}\|_{X^{**}} = \|x\|_X$, so $J$ is an isometric embedding. The space $X$ is called [reflexive](/page/Reflexive%20Space) when $J$ is surjective --- that is, when every element of $X^{**}$ is the evaluation functional at some point of $X$. The Hahn-Banach theorem does not by itself determine whether $X$ is reflexive, but it guarantees that the canonical embedding is always well-defined and isometric. ## Geometric Separation of Convex Sets The algebraic and analytic forms of the Hahn-Banach theorem concern the extension of functionals. The *geometric* form takes a different perspective: instead of extending a functional from a subspace, it *constructs* a functional that separates two disjoint convex sets. The connection between the two viewpoints is the Minkowski functional: to separate a convex set $C$ from a point outside it, one uses the gauge of $C$ as the dominating sublinear functional in the algebraic Hahn-Banach theorem. The fundamental difficulty that the geometric form addresses is this: in finite dimensions, the separating hyperplane theorem for convex sets is a consequence of elementary convex geometry (and can be proved by projection arguments or by the supporting hyperplane theorem). In infinite dimensions, projections onto closed convex sets may not exist (they do in Hilbert spaces, but not in general Banach spaces), and supporting hyperplanes cannot be constructed by geometric intuition alone. The Hahn-Banach theorem provides the analytic substitute. [quotetheorem:974] The hyperplane $H = \{x \in X : f(x) = \alpha\}$ separates $A$ from $B$: the set $A$ lies strictly in the open half-space $\{f < \alpha\}$, while $B$ lies in the closed half-space $\{f \ge \alpha\}$. The separation is *strict* on the $A$ side (because $A$ is open) but only *weak* on the $B$ side. Several features of the hypotheses and conclusion deserve attention. **Why $A$ must be open.** The openness of $A$ is essential for the strict inequality $f(a) < \alpha$. Without openness, one can only guarantee $f(a) \le \alpha \le f(b)$, which is the *weak* separation theorem. To see why openness matters, consider $X = \mathbb{R}^2$, $A = \{(x_1, 0) : x_1 \le 0\}$ (a closed ray), and $B = \{(x_1, x_2) : x_1 > 0, \; x_2 \ge 1/x_1\}$ (the region above a hyperbola). These are disjoint convex sets, but no hyperplane strictly separates them: any separating line must be the $x_2$-axis ($x_1 = 0$), and the infimum of $f(b) = b_1$ over $B$ is $0 = \sup f(a)$, so strict separation fails on the $A$ side. **Why convexity is needed.** The conclusion is false for non-convex sets, even in $\mathbb{R}^2$. Take $A = \{(x_1, x_2) : x_1^2 + x_2^2 < 1\} \setminus \{(x_1, 0) : x_1 > 0\}$ (the open unit disk with a radius removed) and $B = \{(1/2, 0)\}$. Then $B$ is "inside" $A$ (surrounded by it), and no affine hyperplane can separate them, despite $A \cap B = \varnothing$. When both sets are closed and one is compact, the separation can be upgraded to *strict* on both sides. The key geometric fact is that $\operatorname{dist}(A, B) > 0$ when $A$ is compact and $B$ is closed, so one can place an open ball around $A$ that is still disjoint from $B$ and apply the basic separation theorem to the enlarged set. This strict separation version is the one most commonly used in optimisation, convex analysis, and Lagrange multiplier theory. [quotetheorem:1044] The strict gap $\alpha_2 - \alpha_1 > 0$ is what distinguishes this from the basic separation theorem. Neither compactness nor closedness can be dropped: the counterexample above (a closed ray and a hyperbola region in $\mathbb{R}^2$) shows that two closed convex sets may only admit weak separation. [example: Separating a Point from a Closed Convex Set] Let $X$ be a real normed space, let $C \subset X$ be a nonempty closed convex set, and let $x_0 \in X \setminus C$. We show that $x_0$ can be strictly separated from $C$. Since $C$ is closed and $x_0 \notin C$, there exists $r > 0$ with $B(x_0, r) \cap C = \varnothing$. Set $A = B(x_0, r)$ (open, convex) and $B = C$ (convex). Since $A \cap B = \varnothing$, the Hahn-Banach Separation Theorem provides $f \in X^*$ and $\alpha \in \mathbb{R}$ with \begin{align*} f(a) < \alpha \le f(c) \quad \text{for all } a \in B(x_0, r), \; c \in C. \end{align*} In particular, $f(x_0) < \alpha \le f(c)$ for all $c \in C$. Moreover, the strict inequality $f(a) < \alpha$ for all $a \in B(x_0, r)$ implies $\sup_{a \in B(x_0, r)} f(a) \le \alpha$. Computing this supremum: \begin{align*} \sup_{\|a - x_0\| < r} f(a) = f(x_0) + r \|f\|_{X^*}. \end{align*} So $f(x_0) + r\|f\|_{X^*} \le \alpha \le f(c)$ for all $c \in C$, giving $f(c) - f(x_0) \ge r\|f\|_{X^*} > 0$ (the functional $f$ is not identically zero, since it separates nonempty sets). This is strict separation: the continuous linear functional $f$ separates $x_0$ from $C$ with a gap of at least $r\|f\|_{X^*}$. [/example] ## Applications to Specific Spaces The abstract Hahn-Banach theorem acquires concrete force when applied to specific function spaces. In each case, the theorem produces functionals whose existence would be difficult or impossible to establish by direct construction. ### Functionals on $L^p$ Spaces For $1 \le p < \infty$, the [dual space](/page/Dual%20Space) of $L^p(U, \mathcal{L}^n)$ is isometrically isomorphic to $L^q(U, \mathcal{L}^n)$ where $1/p + 1/q = 1$ (with $q = \infty$ when $p = 1$). This identification is the Riesz representation theorem for $L^p$ spaces. The Hahn-Banach theorem enters the proof of this identification in an essential way: to show that every bounded linear functional on $L^p$ is represented by integration against some $g \in L^q$, one first constructs $g$ on simple functions (a dense subspace), uses the norm-preserving extension to lift the functional to all of $L^p$, and then identifies $g$ via the Radon-Nikodym theorem. The Hahn-Banach theorem also constructs functionals on $L^p$ that could not be found by inspection. Consider the following application to non-trivial annihilators. [example: Annihilating a Closed Subspace of $L^p$] Let $U = (0, 2\pi)$ and consider $X = L^2(U, \mathcal{L}^1)$. Let $M \subset X$ be the closed subspace of functions whose first Fourier coefficient vanishes: \begin{align*} M = \left\{f \in L^2(0, 2\pi) : \int_0^{2\pi} f(x) e^{-ix} \, d\mathcal{L}^1(x) = 0 \right\}. \end{align*} Since $L^2$ is a Hilbert space, we can identify $X^* \cong X$ via the Riesz representation. The function $g(x) = \frac{1}{2\pi} e^{ix}$ lies in $X$ and satisfies $\int_0^{2\pi} f(x) g(x) \, d\mathcal{L}^1(x) = 0$ for all $f \in M$ (by definition of $M$). So the functional $T_g(f) = \int_0^{2\pi} fg \, d\mathcal{L}^1$ annihilates $M$. Now take any $h \notin M$, say $h(x) = e^{ix}$. Then $T_g(h) = \int_0^{2\pi} e^{ix} \cdot \frac{1}{2\pi} e^{ix} \, d\mathcal{L}^1(x) = \frac{1}{2\pi} \int_0^{2\pi} e^{2ix} \, d\mathcal{L}^1(x) = 0$. This particular $g$ does not separate $h$ from $M$. Instead, define $g_1(x) = \frac{1}{2\pi} e^{-ix}$. Then for $h(x) = e^{ix}$: \begin{align*} T_{g_1}(h) = \frac{1}{2\pi}\int_0^{2\pi} e^{ix} e^{-ix} \, d\mathcal{L}^1(x) = \frac{1}{2\pi}\int_0^{2\pi} 1 \, d\mathcal{L}^1(x) = 1 \neq 0. \end{align*} For any $f \in M$: $T_{g_1}(f) = \frac{1}{2\pi}\int_0^{2\pi} f(x) e^{-ix} \, d\mathcal{L}^1(x) = 0$ by definition of $M$. So $T_{g_1}$ is a bounded linear functional that annihilates $M$ but not $h$. In a Hilbert space, the construction of annihilating functionals is straightforward (via orthogonal complements). In a general Banach space such as $L^p$ with $p \neq 2$, no inner product is available, and the Hahn-Banach theorem is the only tool that guarantees the existence of functionals separating a closed subspace from a point outside it. [/example] ### Functionals on $C(K)$ For a compact Hausdorff space $K$, the Riesz-Markov-Kakutani representation theorem identifies $C(K)^*$ with the space $\mathcal{M}(K)$ of signed Radon measures on $K$. The Hahn-Banach theorem enters the proof indirectly (through the construction of extensions and the separation of convex sets) and directly in applications: it guarantees that for any $g \in C(K)$ with $g \not\equiv 0$, there exists a measure $\mu \in \mathcal{M}(K)$ with $\int_K g \, d\mu \neq 0$. A striking consequence is the characterisation of when a continuous function belongs to a closed subspace of $C(K)$. If $M \subset C(K)$ is a closed subspace and $g \notin M$, the Hahn-Banach theorem provides $\mu \in \mathcal{M}(K)$ with $\int_K f \, d\mu = 0$ for all $f \in M$ but $\int_K g \, d\mu \neq 0$. Equivalently: $g \in M$ if and only if $\int_K g \, d\mu = 0$ for every $\mu \in M^\perp$, where $M^\perp = \{\mu \in C(K)^* : \mu(f) = 0 \text{ for all } f \in M\}$ is the annihilator of $M$. ## The Role in Weak Topologies and Reflexivity The Hahn-Banach theorem is not merely a source of individual functionals; it underpins the entire theory of [weak topologies](/page/Weak%20Topology) and [reflexivity](/page/Reflexive%20Space). The separation-of-points property guarantees that the weak topology on $X$ and the weak-$*$ topology on $X^*$ are Hausdorff, which is the minimum requirement for these topologies to be useful. But the theorem's influence runs deeper. ### Mazur's Theorem and the Weak Closure of Convex Sets A central question in the theory of weak topologies is: how do weakly closed sets relate to norm-closed sets? In general, the weak topology is coarser than the norm topology, so every weakly closed set is norm-closed, but not conversely. A remarkable theorem of Mazur asserts that for *convex* sets, the two notions coincide. [quotetheorem:985] The "if" direction is immediate (the weak topology is coarser, so weakly closed implies norm-closed). The "only if" direction is the content: a norm-closed convex set is weakly closed. The key step in the proof is the Hahn-Banach Separation Theorem. If $C$ is norm-closed and convex, and $x_0 \notin C$, then (as we showed above) there exists $f \in X^*$ and $\alpha \in \mathbb{R}$ with $f(x_0) < \alpha \le f(c)$ for all $c \in C$. The set $\{x : f(x) < \alpha\}$ is a weakly open neighbourhood of $x_0$ that is disjoint from $C$, proving $x_0 \notin \overline{C}^{\,\sigma(X, X^*)}$. Mazur's theorem has far-reaching consequences for the calculus of variations and PDE theory. When a minimising sequence $u_n \rightharpoonup u$ converges weakly and the constraint set is closed and convex, Mazur's theorem guarantees that $u$ remains in the constraint set. Without it, weak limits might escape the feasible region. ### The Weak-$*$ Topology and the Banach-Alaoglu Theorem The Hahn-Banach theorem also plays a fundamental role in the compactness theory of dual spaces. The Banach-Alaoglu theorem states that the closed unit ball $\overline{B}_{X^*} = \{f \in X^* : \|f\|_{X^*} \le 1\}$ is compact in the weak-$*$ topology $\sigma(X^*, X)$. The proof embeds $\overline{B}_{X^*}$ into a product of compact intervals via the evaluation maps $f \mapsto (f(x))_{x \in X}$ and uses Tychonoff's theorem. While the proof itself does not directly invoke Hahn-Banach, the *usefulness* of the Banach-Alaoglu theorem depends critically on it. Specifically, the Banach-Alaoglu theorem guarantees weak-$*$ cluster points, but one needs the Hahn-Banach theorem to know that these cluster points are meaningful --- that the weak-$*$ topology separates points of $X^*$, and that weak-$*$ limits are unique. Moreover, the Hahn-Banach theorem is essential for relating weak-$*$ compactness in $X^*$ to weak compactness in $X$. Goldstine's lemma (which states that $J(\overline{B}_X)$ is weak-$*$ dense in $\overline{B}_{X^{**}}$, where $J$ is the canonical embedding) is proved using the Hahn-Banach theorem, and it connects the two compactness theories: $X$ is reflexive if and only if the weak-$*$ topology on $X^*$ coincides with the weak topology $\sigma(X^*, X^{**})$ on the unit ball. ## Standard Techniques Using Hahn-Banach The Hahn-Banach theorem appears in analysis through a small number of recurring argument patterns. This section collects the most important techniques. ### Constructing Functionals by Extension The most direct application: define a linear functional on a subspace where its values are forced by the problem, then extend. **Pattern:** 1. Identify the subspace $W$ where the desired functional's behaviour is determined (often one-dimensional: $W = \operatorname{span}\{x_0\}$). 2. Define $f$ on $W$ to achieve the prescribed behaviour (e.g., $f(x_0) = \|x_0\|$, or $f(x_0) = 1$, $f|_M = 0$). 3. Verify that $f$ is bounded on $W$ (compute $\|f\|_{W^*}$). 4. Extend to $\tilde{f} \in X^*$ by Hahn-Banach. [example: Separating an Element from a Closed Subspace] Let $X$ be a [Banach space](/page/Banach%20Space), $M \subset X$ a closed subspace, and $x_0 \in X \setminus M$. We construct $f \in X^*$ with $f|_M = 0$ and $f(x_0) = 1$. **Step 1.** Define $W = M \oplus \operatorname{span}\{x_0\}$. Every element of $W$ can be written uniquely as $w = m + \lambda x_0$ with $m \in M$ and $\lambda \in \mathbb{R}$ (the sum is direct because $x_0 \notin M$ and $M$ is a subspace). **Step 2.** Define $f: W \to \mathbb{R}$ by $f(m + \lambda x_0) = \lambda$. **Step 3.** We compute $\|f\|_{W^*}$. Let $d = \operatorname{dist}(x_0, M) = \inf_{m \in M} \|x_0 - m\|_X > 0$ (positive because $M$ is closed and $x_0 \notin M$). For $w = m + \lambda x_0$ with $\lambda \neq 0$: \begin{align*} \frac{|f(w)|}{\|w\|_X} = \frac{|\lambda|}{\|m + \lambda x_0\|_X} = \frac{1}{\|m/\lambda + x_0\|_X} \le \frac{1}{d}, \end{align*} since $-m/\lambda \in M$ and $\|x_0 - (-m/\lambda)\|_X \ge d$. Moreover, this bound is sharp: taking $m_k \in M$ with $\|x_0 - m_k\|_X \to d$, we get $|f(m_k + x_0)| / \|m_k + x_0\|_X = 1/\|x_0 - (-m_k)\|_X \to 1/d$. So $\|f\|_{W^*} = 1/d$. **Step 4.** Extend $f$ to $\tilde{f} \in X^*$ by Hahn-Banach with $\|\tilde{f}\|_{X^*} = 1/d$. Then $\tilde{f}(x_0) = 1$, $\tilde{f}|_M = 0$, and $\|\tilde{f}\|_{X^*} = 1/\operatorname{dist}(x_0, M)$. [/example] ### The Duality Pairing Test Many results in functional analysis reduce to the principle: *to prove something about $x \in X$, test it against all $f \in X^*$*. The Hahn-Banach theorem ensures that $X^*$ is rich enough for such tests to be conclusive. **Pattern:** 1. Express the desired conclusion as a statement about $f(x)$ for all $f \in X^*$ (e.g., $x = 0$ iff $f(x) = 0$ for all $f$). 2. Use specific properties of $x$ (it solves a PDE, satisfies an integral identity, etc.) to estimate $|f(x)|$. 3. Invoke the norm recovery formula $\|x\| = \sup_{\|f\| \le 1} |f(x)|$ to extract the conclusion. The typical instance is proving that a weakly convergent sequence $x_n \rightharpoonup x$ in a reflexive Banach space satisfies $\|x\| \le \liminf_n \|x_n\|$ (weak lower semicontinuity of the norm). For any $f \in X^*$ with $\|f\| \le 1$, we have $f(x) = \lim_n f(x_n) \le \liminf_n \|x_n\|$. Taking the supremum over $f$ and using the norm recovery formula gives $\|x\| \le \liminf_n \|x_n\|$. ### Density Arguments via Hahn-Banach A powerful indirect technique: to show that a subspace $M$ is dense in $X$, show that no nonzero functional in $X^*$ can annihilate $M$. **Pattern:** 1. Suppose $M$ is not dense in $X$. Then $\overline{M} \subsetneq X$. 2. Pick $x_0 \in X \setminus \overline{M}$. 3. By the Hahn-Banach theorem, construct $f \in X^*$ with $f|_{\overline{M}} = 0$ and $f(x_0) \neq 0$. 4. Derive a contradiction from the properties of $f$ and $M$. Equivalently: $\overline{M} = X$ if and only if $M^\perp := \{f \in X^* : f(m) = 0 \text{ for all } m \in M\} = \{0\}$. This technique is used throughout approximation theory. For instance, the Weierstrass approximation theorem can be proved by showing that any Borel measure on $[a, b]$ that annihilates all polynomials must be zero (via the moment problem and analytic continuation). ## References 1. Brezis, H., *Functional Analysis, Sobolev Spaces and Partial Differential Equations* (2011). 2. Conway, J. B., *A Course in Functional Analysis* (1990). 3. Rudin, W., *Functional Analysis* (1991). 4. Lax, P. D., *Functional Analysis* (2002). 5. Megginson, R. E., *An Introduction to Banach Space Theory* (1998).