Analysis on the real line rests on one quiet miracle: there are no holes. Chase a sequence whose terms crowd ever closer together and the number line always has a point waiting at the end for it to land on. That single guarantee — completeness — is what lets you take limits without fear, sum infinite series, solve differential equations, and know that the answer you are converging toward actually exists.

Modern applied mathematics does not live on the number line. It lives in spaces whose "points" are entire functions: a signal, an image, a quantum wavefunction, the state of a vibrating string. To do calculus on such a space you need a way to measure the size of a function (a norm) — and you need the same no-holes guarantee, so that when a sequence of functions crowds together, a limiting function is really there. A normed vector space that has that guarantee is a Banach space, and it is the arena where infinite-dimensional analysis is done.

The whole subject turns on that second bullet. The vectors and the norm are the easy part; whether the space is complete is where the mathematics — and all the drama — lives.

Cauchy and complete, in a normed setting

Recall the two definitions, now written with a norm instead of an absolute value. Let (x_n) be a sequence of vectors in a normed space (X, \lVert \cdot \rVert).

• Convergent. x_n \to x means \lVert x_n - x \rVert \to 0: for every \varepsilon > 0 there is an N with \lVert x_n - x \rVert < \varepsilon for all n \ge N. Note the limit x is named — you must already possess it.
• Cauchy. (x_n) is Cauchy when its own terms bunch together: for every \varepsilon > 0 there is an N with \lVert x_m - x_n \rVert < \varepsilon for all m, n \ge N. No limit is mentioned — this is a condition the sequence checks about itself.

In any normed space, convergent \Rightarrow Cauchy: if the terms all approach one point they must approach each other, by the triangle inequality \lVert x_m - x_n \rVert \le \lVert x_m - x \rVert + \lVert x - x_n \rVert. The converse is exactly what can fail. A space where the converse always holds — where Cauchy \Rightarrow convergent — is complete:

The rationals \mathbb{Q} are the cautionary tale one dimension down: 1, 1.4, 1.41, 1.414, \dots is Cauchy, its terms crowd together perfectly, yet the point they crowd toward — \sqrt{2} — is not rational. The sequence is trying to converge, but its destination is missing from the space. Completeness is precisely the promise that this never happens: the destination is always in the building.

A gallery of Banach spaces

Almost every space you compute in is a Banach space. A few workhorses:

• \mathbb{R}^n (any norm). With the Euclidean norm \lVert x \rVert_2 = \big(\sum_i x_i^2\big)^{1/2} — or the 1-norm, or the max-norm — \mathbb{R}^n is complete. A Cauchy sequence of vectors is Cauchy in each coordinate, each coordinate converges in the complete \mathbb{R}, and assembling the limits gives a vector in \mathbb{R}^n. In fact every finite-dimensional normed space is a Banach space — incompleteness is a purely infinite-dimensional disease.
• The sequence spaces \ell^p, 1 \le p \le \infty. Points are infinite sequences x = (x_1, x_2, \dots) with \lVert x \rVert_p = \big(\sum_k |x_k|^p\big)^{1/p} finite (and \lVert x \rVert_\infty = \sup_k |x_k| for p = \infty). Every one of these is complete: a Banach space.
• C[a, b] with the supremum norm. The continuous functions on [a, b] with \lVert f \rVert_\infty = \sup_{x \in [a,b]} |f(x)|. This is the first genuinely infinite-dimensional Banach space most people meet, and its completeness is the uniform-limit theorem in disguise (next card).
• The Lebesgue spaces L^p(a,b), 1 \le p \le \infty. Measurable functions with \lVert f \rVert_p = \big(\int_a^b |f|^p\big)^{1/p} < \infty. That these are complete — the Riesz–Fischer theorem — is one of the crowning payoffs of Lebesgue integration, and the reason the integral norm below gets rescued.

(The Hilbert spaces \ell^2 and L^2 are the p = 2 members — Banach spaces whose norm also comes from an inner product, so they carry angles and orthogonality on top of completeness.)

Worked example: C[a,b] with the sup norm is complete

This is the archetype, so it is worth seeing the argument. Let (f_n) be a Cauchy sequence in \big(C[a,b], \lVert \cdot \rVert_\infty\big). We must produce a continuous limit.

Step 1 — a pointwise limit. Fix a point x \in [a, b]. Since |f_m(x) - f_n(x)| \le \lVert f_m - f_n \rVert_\infty, the numbers \big(f_n(x)\big) form a Cauchy sequence in \mathbb{R}, which is complete. So the limit f(x) := \lim_n f_n(x) exists for every x, defining a function f.

Step 2 — the convergence is uniform. Given \varepsilon > 0, choose N so that \lVert f_m - f_n \rVert_\infty < \varepsilon for all m, n \ge N. Then |f_m(x) - f_n(x)| < \varepsilon for every x; hold n \ge N fixed and let m \to \infty, giving |f(x) - f_n(x)| \le \varepsilon for all x at once. That is \lVert f - f_n \rVert_\infty \le \varepsilon: the convergence f_n \to f is uniform.

Step 3 — the limit is continuous. A uniform limit of continuous functions is continuous (the classic \varepsilon/3 argument). Hence f \in C[a, b], and f_n \to f in the sup norm. Every Cauchy sequence converged, inside the space — C[a,b] is a Banach space. The linchpin was Step 3: the sup norm's convergence is uniform, and uniformity is exactly what preserves continuity. Change the norm and that protection can vanish — which is the whole point of the next card.

The key non-example: continuous functions under the integral norm

Completeness is not automatic. Take the very same vector space, C[-1, 1], but measure size with the integral (L^1 or L^2) norm instead of the sup norm:

These are perfectly good norms — but now the space is not complete. Here is the witness. Let

a family of smooth, steepening logistic curves, each one continuous. As n grows they flatten to 0 on the left, to 1 on the right, and turn ever more sharply through the origin. Two things happen at once:

• They are Cauchy in the integral norm. For large m, n the graphs of f_m and f_n differ only in a shrinking sliver around x = 0, so \lVert f_m - f_n \rVert_2 \to 0. The area between them collapses.
• Their only candidate limit is discontinuous. Pointwise, f_n(x) \to g(x) where g is the step function g(x) = 0 for x < 0, g(x) = 1 for x > 0, and one can check \lVert f_n - g \rVert_2 \to 0. But g has a jump at 0 — it is not continuous, so g \notin C[-1, 1].

So (f_n) is a Cauchy sequence with no limit inside the space. It is straining toward a point that has been torn out — a hole, exactly like \sqrt{2} was a hole in \mathbb{Q}. And no continuous function can serve as the limit either: if some h \in C[-1,1] had \lVert f_n - h \rVert_2 \to 0, then \lVert h - g \rVert_2 = 0, forcing h = g wherever both are continuous — impossible for a continuous h. The space \big(C[-1,1], \lVert \cdot \rVert_2\big) is incomplete.

This is not a curiosity — it is the historical reason for Lebesgue's integral. To make the integral norms complete you must enlarge the space, throwing in the missing discontinuous limits (like g) as bona-fide points. The completion of \big(C[a,b], \lVert \cdot \rVert_p\big) is exactly the Lebesgue space L^p(a,b). Banach's completeness is the property that L^p has and continuous functions, under the integral norm, lack.

What completeness buys you: absolutely convergent series

Why care so much about completeness? Because it hands you a clean, powerful test for summing infinite series of vectors — the workhorse of the whole theory. Say a series \sum_{n=1}^{\infty} x_n is absolutely convergent if the numerical series of norms converges:

Why complete \Rightarrow the series test. Suppose \sum \lVert x_n \rVert < \infty. For N > M the triangle inequality gives

and the right-hand side is a tail of a convergent real series, so it \to 0. Hence (s_N) is Cauchy; completeness supplies a limit s \in X. The convergent series of numbers dominates and drags the vector partial sums into a Cauchy pattern, and completeness finishes the job. The reverse implication builds a limit for an arbitrary Cauchy sequence by summing a rapidly-shrinking telescoping subsequence — so this innocuous-looking series test is logically equivalent to completeness itself.

Concretely: in any Banach space, a geometric-style series such as \sum_n \tfrac{1}{2^n}\, u_n with all \lVert u_n \rVert \le 1 automatically converges, because \sum_n \tfrac{1}{2^n} = 1 < \infty. This is the mechanism behind Neumann series (I - T)^{-1} = \sum_n T^n for operators, power series of matrices like e^A = \sum_n A^n/n!, and countless existence proofs — all of which quietly require the underlying space to be Banach.