Skip to content
TheChatArchiveThe Chat Archive
Grok 4.1 Fast· xaiHugging Face Frontier Reasoning Traces

Prove the Hahn-Banach theorem.

  1. User

    Prove the Hahn-Banach theorem.

  2. Assistant

    Hahn-Banach Theorem (Extension Theorem for Sublinear Functionals)

    Theorem. Let XX be a real vector space and p:X→Rp: X \to \mathbb{R} a sublinear functional on XX (i.e., p(αx)=αp(x)p(\alpha x) = \alpha p(x) for α≥0\alpha \geq 0, and p(x+y)≤p(x)+p(y)p(x+y) \leq p(x) + p(y) for all x,y∈Xx,y \in X). Let M⊆XM \subseteq X be a subspace and f:M→Rf: M \to \mathbb{R} a linear functional such that f(x)≤p(x)f(x) \leq p(x) for all x∈Mx \in M. Then there exists a linear functional F:X→RF: X \to \mathbb{R} such that F∣M=fF|_M = f and F(x)≤p(x)F(x) \leq p(x) for all x∈Xx \in X.

    (The complex case follows similarly, with minor adjustments using ∣⋅∣≤1| \cdot | \leq 1 multipliers; the normed-space corollary is a special case with p(x)=∥x∥p(x) = \|x\|.)

    Proof

    We use Zorn's lemma. Consider the collection [ \mathcal{E} = { (N, g) \mid M \subseteq N \subseteq X,, N\text{ subspace},, g: N \to \mathbb{R}\text{ linear},, g|_M = f,, g(x) \leq p(x)\ \forall x \in N }. ] Partially order E\mathcal{E} by (N,g)≤(N′,g′)(N,g) \leq (N',g') if N⊆N′N \subseteq N' and g′∣N=gg'|_N = g.

    Step 1: E\mathcal{E} is nonempty. Indeed, (M,f)∈E(M,f) \in \mathcal{E}.

    Step 2: E\mathcal{E} is inductive. Let {(Nα,gα)}α∈A\{ (N_\alpha, g_\alpha) \}_{\alpha \in A} be a chain. Let N=⋃αNαN = \bigcup_{\alpha} N_\alpha (a subspace) and define g:N→Rg: N \to \mathbb{R} by g(x)=gα(x)g(x) = g_\alpha(x) for x∈Nαx \in N_\alpha. Linearity and g∣M=fg|_M = f follow from the chain property. For x∈Nx \in N, say x∈Nαx \in N_\alpha, then g(x)=gα(x)≤p(x)g(x) = g_\alpha(x) \leq p(x). Thus (N,g)∈E(N,g) \in \mathcal{E} is an upper bound.

    Step 3: Zorn's lemma yields a maximal element (Y,h)∈E(Y,h) \in \mathcal{E}.

    Step 4: Y=XY = X. Suppose not; pick z∈X∖Yz \in X \setminus Y. We extend hh to Y+Rz:={y+λz∣y∈Y, λ∈R}Y + \mathbb{R} z := \{ y + \lambda z \mid y \in Y,\, \lambda \in \mathbb{R} \}, a subspace properly containing YY.

    Define [ \alpha = \sup_{y \in Y} \frac{p(y + z) - h(y)}{1}, \qquad \beta = \inf_{y \in Y} \frac{p(y - z) + h(y)}{1}. ] We claim α≤β\alpha \leq \beta. Indeed, for any y1,y2∈Yy_1, y_2 \in Y, [ p(y_1 + z) - h(y_1) \leq p( (y_1 + y_2) + z ) - h(y_1 + y_2) + p(y_2 - z) - h(y_2) + h(y_2) \leq p( (y_1 + y_2) + z ) - h(y_1 + y_2) + p(y_2 - z) + h(y_2), ] by sublinearity: p(y1+z)≤p((y1+y2)+z)+p(−y2)≤p((y1+y2)+z)+p(y2−z)p(y_1 + z) \leq p( (y_1 + y_2) + z ) + p(-y_2) \leq p( (y_1 + y_2) + z ) + p(y_2 - z) (since p(−y2)≤p(y2−z)+p(z)≤p(y2−z)+p(z)p(-y_2) \leq p(y_2 - z) + p(z) \leq p(y_2 - z) + p(z), wait—actually, the standard inequality uses p(y1+z)≤p(y1+y2+z)+p(−y2)≤p(y1+y2+z)+p(y2−z)p(y_1 + z) \leq p(y_1 + y_2 + z) + p(-y_2) \leq p(y_1 + y_2 + z) + p(y_2 - z), since p(−y2)=p(y2−2y2)≤p(y2−z)+p(−z)≤p(y2−z)+p(z)p(-y_2) = p(y_2 - 2y_2) \leq p(y_2 - z) + p(-z) \leq p(y_2 - z) + p(z), but this is messy. Correctly:

    For y1,y2∈Yy_1, y_2 \in Y, [ p(y_1 + z) \leq p(y_1 + y_2 + z) + p(- y_2) \leq p(y_1 + y_2 + z) + p(y_2 - z), ] since p(−y2)≤p(y2−z)+p(−z)≤p(y2−z)+p(z)p(-y_2) \leq p(y_2 - z) + p(-z) \leq p(y_2 - z) + p(z), but to bound tightly, note the direct way: [ [p(y_1 + z) - h(y_1)] + [p(y_2 - z) + h(y_2)] \geq p( (y_1 + z) + (y_2 - z) ) - h(y_1) + h(y_2) \geq h(y_1 + y_2) - h(y_1) + h(y_2) = h(y_1 + y_2), ] no—sublinearity upper bounds p, but here:

    Actually, the precise argument is: [ p(y_1 + z) - h(y_1) \leq p( (y_1 + z) + (y_2 - z) ) + p( - (y_2 - z) ) - h(y_1) \leq \cdots ] Standard proof shows α≤β\alpha \leq \beta by: Take y1,y2∈Yy_1, y_2 \in Y. Then [ p(y_1 + z) \leq p(y_1 - y_2 + z) + p(2 y_2) \leq p(y_1 - y_2 + z) + 2 p(y_2), ] but the clean way is: Consider for any $y_1, y_

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.