Banach fixed-point theorem: Definition & Examples

The Banach fixed-point theorem guarantees that a contraction of a nonempty complete metric space into itself has exactly one fixed point, and repeated application converges to it from every starting point in that space. For a quant using an iterative valuation or feedback model, it can certify both a unique consistent answer and a bound on numerical error.

The guarantee has three ingredients

Let $X$ be a nonempty set with distance $d$ . Require:

Completeness: every Cauchy sequence in $X$ has its limit in $X$ . Informally, a sequence whose late terms become arbitrarily close cannot converge to a missing point.
A self-map: $T:X\to X$ , so every update stays admissible.
A uniform contraction: one constant $0\le q<1$ works for every pair:

d(Tx,Ty)\le q\,d(x,y),\qquad x,y\in X.

Then

T(x^*)=x^*

for exactly one

x^*\in X

, and fixed-point iteration,

x_{n+1}=T(x_n)

, converges to it for any

x_0\in X

. These are the hypotheses of the contraction principle, not conclusions inferred from a few successful iterations.

A hypothetical valuation update

Suppose a deliberately simplified, dimensionless valuation state obeys

T(x)=0.5x+0.1,\qquad X=[0,1],\qquad d(x,y)=\lvert x-y\rvert.

This is a synthetic feedback equation, not a pricing formula fitted to markets. The closed interval is complete, $T(X)=[0.1,0.6]\subset X$ , and differences shrink by exactly $q=0.5$ . All three ingredients hold. Solving $x=0.5x+0.1$ gives $x^*=0.2$ .

Starting at zero produces:

Updates $n$	State $x_n$	Distance to $0.2$
0	0	0.2
1	0.1	0.1
2	0.15	0.05
3	0.175	0.025

The theorem covers every starting state in the interval. Applying $T$ to the whole interval makes this visible:

Synthetic map T(x)=0.5x+0.1 sends the interval [0,1] to [0.1,0.6], then [0.15,0.4], [0.175,0.3] and [0.1875,0.25]. All contain the fixed point 0.2; their widths halve at each update.

Original synthetic illustration. Each bar contains the iterates from every initial state in [0,1] after the stated number of updates. The left endpoints are the iterates starting at zero.

From convergence to a stopping certificate

A useful solver needs a computable error bound. The prior and posterior estimates give, for

n\ge1

\begin{aligned} d(x_n,x^*)&\le \frac{q^n}{1-q}d(x_1,x_0),\\ d(x_n,x^*)&\le \frac{q}{1-q}d(x_n,x_{n-1}). \end{aligned}

The first plans the iteration count from the initial step; the second uses the latest completed step. Their underlying mechanism is a geometric tail: each successive step length is at most $q$ times the previous one, so the remaining travel is bounded by their infinite sum. Completeness keeps the limit inside $X$ ; contraction makes that limit fixed and rules out a second fixed point.

In the example, after three updates the prior bound is

\frac{0.5^3}{1-0.5}(0.1)=0.025.

The posterior bound is also $0.025$ , equal to the actual error here. This equality is special to the example; bounds can be conservative. To certify an error at most $0.001$ from zero, eight updates suffice: the bound is $0.00078125$ , whereas after seven it is $0.0015625$ .

There is also a residual certificate for any candidate $x\in X$ :

d(x,x^*)\le\frac{d(x,T(x))}{1-q}.

Indeed, the triangle inequality gives $d(x,x^*)\le d(x,T(x))+q\,d(x,x^*)$ . A small residual alone is less informative when $q$ is close to one. These bounds assume exact updates; rounding or approximate evaluations need their own error allowance.

Where the guarantee stops

The domain matters. On the open interval $(0,1)$ , $T(x)=x/2$ is a self-map with $q=0.5$ , but its only real fixed point is zero, outside the interval. Completeness is missing. At the other boundary, allowing $q=1$ is insufficient: $T(x)=1-x$ on $[0,1]$ has a unique fixed point $0.5$ , yet iteration from zero cycles between zero and one.

For a differentiable scalar map on a closed interval, a bound

\lvert T'(x)\rvert\le q<1

throughout the interval establishes contraction by the mean value theorem. You must still verify that

T

maps the interval into itself. A derivative check only at the proposed solution supports a local analysis, not the global guarantee above. Driscoll and Braun develop the connection between derivatives and iteration behavior.

Contraction is sufficient, not necessary for convergence. If it cannot be established, choose another analysis or solver rather than treating the theorem as evidence of failure. Root-Finding and Numerical Integration develops bisection and Newton methods for scalar equations, including bond yield. The full lesson requires membership; its preview is public.

Sources

Stephen New, University of Waterloo, Lecture Notes for PMATH 351, Real Analysis, §6.1–6.3, printed page 37: contraction, fixed point, existence, uniqueness and convergence.
George Voutsadakis, Introduction to Functional Analysis, Chapter 5, May 2023, slides 6–11: contraction theorem, geometric-tail proof and prior/posterior error estimates.
T. A. Driscoll and R. J. Braun, Fundamentals of Numerical Computation, §4.2, especially §4.2.1 and §4.2.3: local derivative analysis and contraction conditions.

The guarantee has three ingredients

A hypothetical valuation update

From convergence to a stopping certificate

Where the guarantee stops

Sources

Related terms

Fixed point

Fixed-point iteration

Metric space and completeness