Neural-Path/Notes
45 min

Normed & Banach Spaces: Completeness, Compactness & the Hahn-Banach Theorem

A Banach space is a complete normed vector space — completeness ensures that Cauchy sequences converge, a property that makes iterative algorithms well-defined in infinite dimensions. Without it, a sequence of approximations can look convergent yet have no limit inside the space, breaking fixed-point arguments and the guarantees that underlie optimization and approximation theory.

Concepts

Lp Unit Ball — p = 2

x₁x₂-1-111v

L² (Euclidean)

‖v‖2 = 1.000v = (0.6, 0.8)

ML use

Ridge regression, weight decay, cosine similarity

Unit ball shape

Circle — perfectly symmetric; the only Lp ball invariant under rotation

The yellow vector v = (0.6, 0.8) lies on the L² unit sphere. Its Lp norm changes as p varies.

You have been working with vectors and measuring their size with the Euclidean norm all along. But the Euclidean norm is just one choice — 1\ell^1 (sum of absolute values), \ell^\infty (largest absolute value), and the whole p\ell^p family are equally valid norms, each inducing a different geometry and different unit balls. Banach spaces generalize this to infinite dimensions with one critical additional requirement: completeness. Without it, sequences that "should" converge have no limit in the space, and the iterative arguments that underlie optimization and approximation theory break down entirely.

Normed Vector Spaces

A normed vector space is a pair (X,)(X, \|\cdot\|) where XX is a vector space over R\mathbb{R} (or C\mathbb{C}) and :X[0,)\|\cdot\|: X \to [0,\infty) satisfies three axioms:

  1. Positive definiteness: x0\|x\| \geq 0, and x=0\|x\| = 0 if and only if x=0x = 0
  2. Homogeneity: αx=αx\|\alpha x\| = |\alpha|\,\|x\| for all scalars α\alpha
  3. Triangle inequality: x+yx+y\|x + y\| \leq \|x\| + \|y\|

Every norm induces a metric d(x,y)=xyd(x,y) = \|x - y\|, giving a topology on XX in which open balls are the unit of convergence.

Lp Norms and Classical Inequalities

For a sequence x=(x1,x2,)x = (x_1, x_2, \ldots) and 1p<1 \leq p < \infty, the p\ell^p norm is

xp=(ixip)1/p\|x\|_p = \left(\sum_i |x_i|^p\right)^{1/p}

The \ell^\infty norm is x=supixi\|x\|_\infty = \sup_i |x_i|. For functions, the Lp([a,b])L^p([a,b]) norm replaces the sum with an integral: fp=(abf(x)pdx)1/p\|f\|_p = \bigl(\int_a^b |f(x)|^p\,dx\bigr)^{1/p}.

Hölder's inequality. For conjugate exponents 1/p+1/q=11/p + 1/q = 1 with p,q1p, q \geq 1:

fg1fpgq\|fg\|_1 \leq \|f\|_p\,\|g\|_q

In finite dimensions this reads ixiyixpyq\sum_i |x_i y_i| \leq \|x\|_p\|y\|_q. The case p=q=2p = q = 2 is the Cauchy-Schwarz inequality.

Minkowski's inequality is the triangle inequality for LpL^p: f+gpfp+gp\|f+g\|_p \leq \|f\|_p + \|g\|_p. It follows from Hölder applied to (f+g)p1(f+g)^{p-1}.

The three axioms of a norm are not arbitrary choices — they encode exactly what is needed for distance to behave like distance: positivity ensures distinct points are distinguishable, homogeneity ensures scaling is consistent, and the triangle inequality ensures that detours don't help. Hölder's and Minkowski's inequalities are the proofs that these axioms survive when you raise vectors to powers, turning a metrically well-behaved space into an arithmetically useful one.

Banach Spaces: Definition and Examples

A Banach space is a normed vector space that is complete: every Cauchy sequence converges to an element of the space. Formally, if xnxm0\|x_n - x_m\| \to 0 as n,mn,m \to \infty, there exists xXx \in X with xnx0\|x_n - x\| \to 0.

SpaceNormBanach?
Rn\mathbb{R}^nany p\ell^p normYes — all finite-dimensional normed spaces are Banach
p\ell^p, 1p1 \leq p \leq \inftyp\|\cdot\|_pYes
Lp([a,b])L^p([a,b]), 1p1 \leq p \leq \inftyp\|\cdot\|_pYes
C([a,b])C([a,b])$|f|_\infty = \supf
Polynomials on [0,1][0,1]\|\cdot\|_\inftyNo — LpL^p for 0<p<10 < p < 1 also fails

Four Foundational Theorems

Hahn-Banach Theorem. Let YXY \subset X be a subspace and f:YRf: Y \to \mathbb{R} a bounded linear functional with fYM\|f\|_Y \leq M. Then ff extends to f~:XR\tilde{f}: X \to \mathbb{R} with f~X=fY\|\tilde{f}\|_X = \|f\|_Y. This is the fundamental existence result for dual functionals.

Open Mapping Theorem. If T:XYT: X \to Y is a surjective bounded linear operator between Banach spaces, then TT is an open map. Corollary: a bijective bounded linear operator between Banach spaces has a bounded inverse.

Closed Graph Theorem. A linear operator T:XYT: X \to Y between Banach spaces is bounded if and only if its graph {(x,Tx):xX}\{(x, Tx) : x \in X\} is closed in X×YX \times Y.

Uniform Boundedness Principle (Banach-Steinhaus). Let {Tα}\{T_\alpha\} be a family of bounded linear operators from a Banach space XX to a normed space YY. If supαTαx<\sup_\alpha \|T_\alpha x\| < \infty for every xXx \in X, then supαTα<\sup_\alpha \|T_\alpha\| < \infty.

Dual Spaces

The dual space X=B(X,R)X^* = B(X, \mathbb{R}) consists of all bounded linear functionals f:XRf: X \to \mathbb{R}, with norm fX=supx1f(x)\|f\|_{X^*} = \sup_{\|x\| \leq 1} |f(x)|. It is always a Banach space regardless of whether XX is.

Riesz representation for LpL^p. For 1<p<1 < p < \infty and 1/p+1/q=11/p + 1/q = 1, every bounded functional on LpL^p has the form fgfdxf \mapsto \int g f\,dx for a unique gLqg \in L^q, giving an isometric isomorphism (Lp)Lq(L^p)^* \cong L^q. The dual of 1\ell^1 is \ell^\infty by the same argument applied termwise.

Worked Example

Proving Hölder's Inequality for Finite Sequences

For a,b0a, b \geq 0 and 1/p+1/q=11/p + 1/q = 1, Young's inequality states abap/p+bq/qab \leq a^p/p + b^q/q. Apply it to normalized terms a=xi/xpa = |x_i|/\|x\|_p and b=yi/yqb = |y_i|/\|y\|_q:

ixixpyiyqi(xippxpp+yiqqyqq)=1pxppxpp+1qyqqyqq=1p+1q=1\begin{aligned} \sum_i \frac{|x_i|}{\|x\|_p} \cdot \frac{|y_i|}{\|y\|_q} &\leq \sum_i \left(\frac{|x_i|^p}{p\,\|x\|_p^p} + \frac{|y_i|^q}{q\,\|y\|_q^q}\right) \\ &= \frac{1}{p} \cdot \frac{\|x\|_p^p}{\|x\|_p^p} + \frac{1}{q} \cdot \frac{\|y\|_q^q}{\|y\|_q^q} = \frac{1}{p} + \frac{1}{q} = 1 \end{aligned}

Multiplying through by xpyq\|x\|_p\|y\|_q gives ixiyixpyq\sum_i |x_i y_i| \leq \|x\|_p\|y\|_q.

The Dual of 1\ell^1 Is \ell^\infty

Every bounded linear functional f:1Rf: \ell^1 \to \mathbb{R} satisfies f(x)=iaixif(x) = \sum_i a_i x_i where ai=f(ei)a_i = f(e_i) and eie_i is the ii-th standard basis vector. Boundedness gives ak=f(ek)fek1=f|a_k| = |f(e_k)| \leq \|f\|\,\|e_k\|_1 = \|f\| for each kk, so (ak)(a_k) \in \ell^\infty with (ak)f\|(a_k)\|_\infty \leq \|f\|. Conversely any (ak)(a_k) \in \ell^\infty defines a bounded functional via this formula with f=(ak)\|f\| = \|(a_k)\|_\infty. The Hahn-Banach theorem guarantees functionals defined on subspaces extend to all of 1\ell^1, completing the isometric isomorphism (1)(\ell^1)^* \cong \ell^\infty.

Closed Graph Theorem: Why Completeness Is Essential

Consider the identity T=id:(C([0,1]),L1)(C([0,1]),)T = \mathrm{id}: (C([0,1]), \|\cdot\|_{L^1}) \to (C([0,1]), \|\cdot\|_\infty). Its graph is closed: if fnff_n \to f in L1L^1 and fngf_n \to g uniformly, then f=gf = g a.e. so the graph closes up. Yet TT is not bounded: take fn(x)=min(nx,1)f_n(x) = \min(nx, 1) truncated to a triangle of height 11 and base 1/n1/n — then fnL1=1/2\|f_n\|_{L^1} = 1/2 while fn=1\|f_n\|_\infty = 1 stays constant, but scaling to a spike gives fn/fnL1\|f_n\|_\infty / \|f_n\|_{L^1} \to \infty. The closed graph theorem does not apply because the domain (C([0,1]),L1)(C([0,1]), \|\cdot\|_{L^1}) is not complete — it is not a Banach space.

Connections

Where Your Intuition Breaks

The Uniform Boundedness Principle says: if a family of bounded linear operators is pointwise bounded (each operator is bounded on each individual input), then the operators are uniformly bounded (there is a single constant that bounds all of them). The surprising direction is the contrapositive: if no such uniform bound exists, there must be a single input xx where the operators are unbounded. The existence of such an xx is proved non-constructively via the Baire category theorem — the set of "bad" inputs is a residual set (countable intersection of dense open sets), which in a complete metric space is non-empty. This is why completeness of the domain is essential: the Baire argument fails in incomplete spaces, and there exist examples of pointwise-bounded families that are not uniformly bounded when the domain is not Banach.

💡Intuition

Completeness is what separates "every Cauchy sequence has a limit" from "every Cauchy sequence looks like it should converge." The rationals Q\mathbb{Q} with |\cdot| form a normed space but not a Banach space — the sequence 3,3.1,3.14,3, 3.1, 3.14, \ldots is Cauchy in Q\mathbb{Q} but its limit πQ\pi \notin \mathbb{Q}. In functional analysis, this is the exact failure mode that breaks Picard iteration for ODEs: the iterates live in a function space, and if that space has "holes," the limit of approximations may not exist inside it.

💡Intuition

The Banach-Steinhaus theorem is the key tool for proving that pointwise-convergent sequences of linear operators are uniformly bounded. In approximation theory, if partial Fourier sums SnfS_n f converge pointwise for every ff in some Banach space, Banach-Steinhaus immediately forces supnSn<\sup_n \|S_n\| < \infty. Contrapositive: if supnSn=\sup_n \|S_n\| = \infty, there must exist some ff for which SnfS_n f diverges — and du Bois-Reymond's theorem that continuous functions exist with divergent Fourier series is proved exactly this way.

⚠️Warning

Not all Banach spaces are reflexive (XXX^{**} \cong X). The space 1\ell^1 is not reflexive: its double dual is ()(\ell^\infty)^*, which is strictly larger than 1\ell^1 (it contains finitely additive measures on N\mathbb{N}). Reflexivity matters for optimization: in a reflexive Banach space, every bounded sequence has a weakly convergent subsequence (Kakutani's theorem), which is used to prove existence of minimizers via the direct method of the calculus of variations. In non-reflexive spaces this fails and existence proofs require separate arguments.

Enjoying these notes?

Get new lessons delivered to your inbox. No spam.