Completeness, Cauchy Sequences, and Banach Spaces

Completeness is not a museum piece; it is the load-bearing assumption under working analysis and computation. Every proof that a differential equation has a solution — from Picard–Lindelöf for ODEs to fixed-point formulations of nonlinear PDEs — cashes out as a contraction on a complete function space, and the same a-priori bound $q^n/(1-q)$ we checked above tells a numerical analyst how many iterations a solver needs to hit a tolerance. In quantum mechanics the state space is a complete inner-product space (a Hilbert space) precisely so limits of physically realizable states are themselves states; otherwise the spectral theorem for self-adjoint operators — the content of "observables have real eigenvalues" — would have no ground to stand on. The Neumann series is the rigorous backbone of time-independent perturbation theory and scattering resolvent expansions, and its operator-norm bound decides when those expansions converge. Wherever an answer is built as a limit, completeness is the silent clause that promises it exists.

Picture the number line and imagine you may only stand on rational points. Walk out the decimal truncations of $\sqrt2$ — $1,\,1.4,\,1.41,\,1.414,\,1.4142,\dots$ Each step lands on a legal rational, and the steps huddle ever more tightly; the gap between the $m$-th and $n$-th terms shrinks below any tolerance you fix. Every instinct says this sequence is going somewhere — and it is, but the destination $\sqrt2$ is not rational. There is a hole in the floor exactly where the limit should be, and that hole is what completeness forbids. A complete space has the floor filled in: everywhere a Cauchy sequence tries to land, there is genuinely a point to land on.

The Cauchy test earns its keep because it is intrinsic. In a function space the object you are chasing — the solution of a differential equation, the sum of a Fourier series — is unknown by construction; you can only build a sequence of approximations and hope. Completeness lets you cash that hope: assemble a Cauchy sequence of approximants and the limit is guaranteed to exist as a genuine element of the space. This is why every serious setting for analysis is taken complete, and why incomplete ones are quietly replaced by their completions. The most operational face of completeness is the fixed-point machine. A map $T$ that shrinks distances by a factor $q<1$ — a contraction — turns any starting guess into a sequence $x_{n+1}=Tx_n$ whose steps shrink geometrically, so the iterates are Cauchy. In a complete space that sequence must converge, and its limit is the unique fixed point $Tx^\ast=x^\ast$. The staircase below climbs to the fixed point of $T(x)=\tfrac12 x+1$, its error halved at every step and pinned exactly by the a-priori bound.

The whole subject arranges into a short hierarchy of increasing analytic power: a bare normed space lets you measure lengths; a Banach space (complete normed) guarantees limits exist; a Hilbert space (complete inner-product) adds orthogonal geometry. Each inclusion buys a tool — first guaranteed limits, then the Pythagorean structure that quantum mechanics runs on.

Everything begins with the intrinsic convergence test. A sequence is Cauchy when, past a threshold index, all its terms lie within any prescribed tolerance of each other — a statement about the sequence alone, mentioning no limit.

where: $x_n$ — the $n$-th term of a sequence in a normed space $V$; $\lVert\cdot\rVert$ — the norm supplying distances; $\varepsilon$ — the tolerance; $N$ — the threshold index witnessing the bunching. Reading it: no matter how tight a target $\varepsilon$ you name, eventually every pair of terms is closer than that. Ordinary convergence, by contrast, refers to an external limit point $x$: the distance from the terms to $x$ must vanish.

where: $x$ — the limit, a fixed element of $V$. Reading it: convergence names a destination; the Cauchy condition (1) does not. This is the entire difference, and completeness is what reconciles them. One direction is free and holds in every normed space: a sequence that converges is automatically Cauchy, because two terms both near $x$ are near each other.

where: the implication follows from the triangle inequality alone. Reading it: convergent $\Rightarrow$ Cauchy always; completeness is exactly the converse, and it can fail. Completeness names the spaces where the converse holds — where the intrinsic test (1) is as good as convergence (2).

where: $V$ — a normed vector space over $\mathbb{R}$ or $\mathbb{C}$; a complete normed space is a Banach space. Reading it: no Cauchy sequence is allowed to escape; there are no holes. The workhorse restatement of completeness is a series test: absolute convergence forces ordinary convergence. This is the form you actually reach for in proofs.

where: $\sum x_n$ — the limit of the partial sums $s_N=\sum_{n\le N}x_n$. Reading it: if the lengths add up, the vectors add up. Remarkably this implication is equivalent to completeness (Derivation 1), turning an abstract property into a checkable one. Two structures ride on completeness. A contraction uniformly shrinks distances, and in a complete space it has a unique fixed point.

where: $T$ — a self-map of the complete space $V$; $q$ — the contraction (Lipschitz) constant, strictly below $1$. Reading it: each application of $T$ pulls any two points a fixed fraction closer. Finally, the operator incarnation: a bounded operator smaller than the identity can be inverted by a geometric series, and the operators themselves form a Banach space.

where: $A$ — a bounded operator with operator norm $\lVert A\rVert=\sup_{\lVert x\rVert=1}\lVert Ax\rVert$; $I$ — the identity; $\mathcal{B}(V,W)$ — bounded linear maps $V\to W$ under the operator norm. Reading it: completeness of the target $W$ is inherited by the whole space of operators, which is why operator series make sense at all.

Whether a space is complete is decided by the norm, not the set. The same collection of vectors can sit on either side of the line depending on how you measure distance — the single most important thing to internalize about completeness.

1 · A contraction and its guaranteed fixed point (fully worked)

Take $T(x)=\tfrac12 x+1$ on the complete space $\mathbb{R}$ and iterate from $x_0=0$. Show $T$ is a contraction, find its fixed point, and check the iterates against the a-priori error bound of the fixed-point theorem.

Step 1. Contraction constant: $|T(x)-T(y)|=\big|\tfrac12(x-y)\big|=\tfrac12|x-y|$, so $T$ contracts with $q=\tfrac12<1$. Because $\mathbb{R}$ is complete, the theorem promises a unique fixed point reached from any start.

Step 2. Iterate from $x_0=0$: $x_1=1,\ x_2=1.5,\ x_3=1.75,\ x_4=1.875,\ x_5=1.9375,\dots$ These climb steadily; solving $x^\ast=\tfrac12 x^\ast+1$ directly gives the fixed point $\mathbf{x^\ast=2}$, exactly the value the staircase figure converges to.

Step 3. The a-priori bound is $\lVert x_n-x^\ast\rVert\le\dfrac{q^n}{1-q}\lVert x_1-x_0\rVert$. Here $\lVert x_1-x_0\rVert=1$ and $q=\tfrac12$, so the bound is $\dfrac{(1/2)^n}{1/2}=(1/2)^{n-1}$. The true errors are $|x_1-2|=1,\ |x_2-2|=0.5,\ |x_3-2|=0.25,\ |x_4-2|=0.125$ — i.e. $(1/2)^{n-1}$ exactly. The bound is not merely satisfied but tight: the error halves every step, $|x_n-2|=(1/2)^{n-1}$, and completeness of $\mathbb{R}$ is what guarantees the iterates land on $x^\ast=2$ rather than at a hole.

2 · The Neumann series inverts a $2\times2$ operator (fully worked)

Let $A=\begin{pmatrix}0&\tfrac12\\[2pt]\tfrac12&0\end{pmatrix}$ act on $\mathbb{R}^2$ with the Euclidean norm. Confirm $\lVert A\rVert<1$, sum the Neumann series for $(I-A)^{-1}$, and compare with the direct inverse and the norm bound.

Step 1. Operator norm: $A$ is symmetric with eigenvalues $\pm\tfrac12$, so $\lVert A\rVert=\tfrac12<1$ and the Neumann series applies. Powers are clean: $A^2=\tfrac14 I$, hence $A^{2k}=(\tfrac14)^k I$ and $A^{2k+1}=(\tfrac14)^k A$.

Step 2. Sum the series by even/odd parts: $\sum_{k\ge0}A^k=\big(\sum_k 4^{-k}\big)I+\big(\sum_k 4^{-k}\big)A=\dfrac{1}{1-\tfrac14}\,(I+A)=\dfrac{4}{3}\begin{pmatrix}1&\tfrac12\\[2pt]\tfrac12&1\end{pmatrix}=\begin{pmatrix}\tfrac43&\tfrac23\\[2pt]\tfrac23&\tfrac43\end{pmatrix}$.

Step 3. Check directly: $I-A=\begin{pmatrix}1&-\tfrac12\\[2pt]-\tfrac12&1\end{pmatrix}$ has determinant $1-\tfrac14=\tfrac34$, so $(I-A)^{-1}=\tfrac{1}{3/4}\begin{pmatrix}1&\tfrac12\\[2pt]\tfrac12&1\end{pmatrix}=\begin{pmatrix}\tfrac43&\tfrac23\\[2pt]\tfrac23&\tfrac43\end{pmatrix}$ — identical to the series sum. The Neumann series reproduces the exact inverse, and its norm $\lVert(I-A)^{-1}\rVert=2$ saturates the bound $(1-\lVert A\rVert)^{-1}=(1-\tfrac12)^{-1}=2$. Completeness of $\mathbb{R}^{2\times2}$ (a finite-dimensional, hence Banach, operator space) is what lets the infinite operator series converge in the first place.

3 · Why $C[0,1]$ is not complete in the $L^2$ norm

Exhibit a Cauchy sequence of continuous functions, under $\lVert f\rVert_2=(\int_0^1|f|^2)^{1/2}$, whose limit is not continuous.

Step 1. Let $f_n$ ramp linearly from $0$ to $1$ across $[\tfrac12,\tfrac12+\tfrac1n]$: it is $0$ on $[0,\tfrac12]$, rises with slope $n$, then equals $1$ on $[\tfrac12+\tfrac1n,1]$. Every $f_n$ is continuous.

Step 2. For $m>n$ the two ramps differ only on a window of width $\le\tfrac1n$, where $|f_m-f_n|\le1$, so $\lVert f_m-f_n\rVert_2^2\le\tfrac1n\to0$. The sequence is Cauchy in $\lVert\cdot\rVert_2$.

Step 3. Pointwise the ramps approach the unit step $H(x-\tfrac12)$, discontinuous at $x=\tfrac12$. No continuous function can be the $L^2$-limit, since any limit would have to agree with the step almost everywhere. So $(C[0,1],\lVert\cdot\rVert_2)$ has a hole where the step sits; completing the space fills it with $L^2[0,1]$, where quantum wavefunctions live.

Completeness went from an implicit assumption to a named axiom in the hands of one school: the Lwów mathematicians of interwar Poland, who did much of their work in the marble-topped tables of the Scottish Café. Its central figure gave his name to the objects.

Texts. Kreyszig, Introductory Functional Analysis with Applications, Ch. 1–2 and §5.1 (completeness, contraction mapping) · Reed & Simon, Methods of Modern Mathematical Physics I, §I.3 (Banach and Hilbert spaces) · Stone & Goldbart, Mathematics for Physics, Ch. 1 (function spaces) · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 5 · Hassani, Mathematical Physics, Ch. 5–6 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed. Key checkpoints in the material above: the $2^{-j}$ subsequence construction makes the norm-series sum exactly $1$; the telescoping identity $(I-A)\sum_{k\le N}A^k=I-A^{N+1}$ underlies the Neumann series argument; the numeric contraction $q=\tfrac12$ yields error $(1/2)^{n-1}$ with fixed point $2$; and the $2\times2$ Neumann sum equals $\begin{pmatrix}4/3&2/3\\2/3&4/3\end{pmatrix}$ with norm $2=1/(1-\tfrac12)$.

Prerequisites. Inner-Product Spaces and Norms · Vector Spaces and Subspaces · Linear Independence, Basis, and Dimension. Leads to. Adjoint and Hermitian Operators · Linear Operators and Matrix Representations · Diagonalization and the Spectral Decomposition.

Ask the tutor: Want me to walk through why $C[0,1]$ splits into complete and incomplete under different norms, or to run the contraction and Neumann numbers for a case of your choosing?