Continuous Maps Between Metric Spaces

Continuous maps between metric spaces are functions

f : (X, d_{X}) \to (Y, d_{Y})

that preserve "closeness": small changes in input produce small changes in output. Key result:

f

is continuous iff the preimage of every open set is open — the topological characterization that frees continuity from the metric. Applies in: analysis, topology, geometry, numerical methods, dynamical systems. Origins: Cauchy and Weierstrass (rigorous

ε

–

δ

, 1820s–1870s), Fréchet (abstract metric spaces, 1906).

A map $f : (X, d_{X}) \to (Y, d_{Y})$ between metric spaces is continuous at a point $a \in X$ if inputs near $a$ map to outputs near $f (a)$ , quantified by the $ε$ – $δ$ condition. It is continuous if it is continuous at every point. This concept generalizes the familiar continuity of real functions to any setting equipped with a notion of distance — spaces of sequences, functions, matrices, or fractals. The deep insight is that continuity, originally defined metrically, has an equivalent formulation purely in terms of open sets, which makes it a topological property invariant under change of equivalent metric. A surprising fact: continuity does not preserve many properties one might expect — a continuous bijection need not have a continuous inverse, and continuous images of closed or bounded sets can fail to be closed or bounded.

The Core Insight

Continuity is about preservation of proximity. Three faces of the same idea coexist:

Metric ( $ε$ – $δ$ ): control the output error by controlling the input perturbation.
Sequential: $x_{n} \to a$ forces $f (x_{n}) \to f (a)$ .
Topological (open sets): $f^{- 1} (open)$ is open.

In metric spaces all three are equivalent. The topological version is the most powerful because it reveals continuity as the natural notion of a "structure-preserving map" for topology.

Formal Statement

Metric Space

A metric space is a pair $(X, d)$ where $X$ is a set and $d : X \times X \to [0, \infty)$ satisfies, for all $x, y, z \in X$ :

d (x, y) = 0 ⟺ x = y

⟳This is the identity of indiscernibles — distance zero means identical points.

d (x, y) = d (y, x)

⟳Symmetry — distance does not depend on the direction of measurement.

d (x, z) \leq d (x, y) + d (y, z)

⟳The triangle inequality — going via a third point cannot be shorter.

The open ball of radius $r > 0$ centered at $a$ is $B_{X} (a, r) = {x \in X : d_{X} (x, a) < r}$ .

Continuity at a Point — $\varepsilon$–$\delta$

Let $f : (X, d_{X}) \to (Y, d_{Y})$ . Then $f$ is continuous at $a \in X$ if

\forall ε > 0 \exists δ > 0 \forall x \in X : d_{X} (x, a) < δ ⟹ d_{Y} (f (x), f (a)) < ε

⟳Read it as "for every output tolerance

ε

there is an input tolerance

δ

that guarantees it." The

δ

may depend on both

ε

and

a

Equivalently, in ball notation:

\forall ε > 0 \exists δ > 0 : f (B_{X} (a, δ)) \subseteq B_{Y} (f (a), ε)

⟳Rewrite [4] by collecting the

x

with

d_{X} (x, a) < δ

into a ball and noting the conclusion places their images in a ball about

f (a)

f

is continuous if it is continuous at every

a \in X

Equivalent Characterizations of Continuity

For $f : (X, d_{X}) \to (Y, d_{Y})$ , the following are equivalent:

( $ε$ – $δ$ ) $f$ is continuous at every point per [4].
(Sequential) For every sequence $x_{n} \to a$ in $X$ , $f (x_{n}) \to f (a)$ in $Y$ .
(Open sets) $f^{- 1} (U)$ is open in $X$ for every open $U \subseteq Y$ .
(Closed sets) $f^{- 1} (C)$ is closed in $X$ for every closed $C \subseteq Y$ .
(Closure) $f (\overline{A}) \subseteq \overline{f (A)}$ for every $A \subseteq X$ .

Uniform Continuity (stronger notion)

f

is uniformly continuous if

\forall ε > 0 \exists δ > 0 \forall x, y \in X : d_{X} (x, y) < δ ⟹ d_{Y} (f (x), f (y)) < ε

⟳The crucial change from [4] is that

δ

depends only on

ε

, not on the point — the same

δ

works everywhere.

A Lipschitz map (constant $L$ ) satisfies $d_{Y} (f (x), f (y)) \leq L d_{X} (x, y)$ and is automatically uniformly continuous. Implications: Lipschitz $\Rightarrow$ uniformly continuous $\Rightarrow$ continuous. None of the reverse implications hold in general.

Why It Holds / Derivation

($\varepsilon$–$\delta$) $\iff$ (Open sets)

Step 1 — Continuity $\Rightarrow$ preimages of open sets are open. Assume

f

continuous (per [4]) and let

U \subseteq Y

be open. Take any

a \in f^{- 1} (U)

, so

f (a) \in U

. Since

U

is open, there is

ε > 0

with

B_{Y} (f (a), ε) \subseteq U

. Step 2. By continuity at

a

, choose

δ > 0

so that

f (B_{X} (a, δ)) \subseteq B_{Y} (f (a), ε) \subseteq U

. Hence

B_{X} (a, δ) \subseteq f^{- 1} (U)

. Step 3. Every point of

f^{- 1} (U)

is interior, so

f^{- 1} (U)

is open. Step 4 — Preimages of open sets open $\Rightarrow$ continuity. Fix

a \in X

and

ε > 0

. The ball

U = B_{Y} (f (a), ε)

is open, so by hypothesis

f^{- 1} (U)

is open and contains

a

. Step 5. Openness gives

δ > 0

with

B_{X} (a, δ) \subseteq f^{- 1} (U)

, i.e.

d_{X} (x, a) < δ \Rightarrow f (x) \in B_{Y} (f (a), ε) \Rightarrow d_{Y} (f (x), f (a)) < ε

. This is exactly [4].

■

($\varepsilon$–$\delta$) $\iff$ (Sequential)

Step 1 — $(\Rightarrow)$ . Assume

f

continuous at

a

and

x_{n} \to a

. Given

ε > 0

, pick

δ

from [4]. Since

x_{n} \to a

, there is

N

with

d_{X} (x_{n}, a) < δ

for

n \geq N

, whence

d_{Y} (f (x_{n}), f (a)) < ε

. Thus

f (x_{n}) \to f (a)

. Step 2 — $(\Leftarrow)$ , contrapositive. Suppose

f

is not continuous at

a

. Then there is

ε_{0} > 0

such that for every

δ > 0

some

x

with

d_{X} (x, a) < δ

has

d_{Y} (f (x), f (a)) \geq ε_{0}

. Step 3. Take

δ = 1/ n

and choose

x_{n}

with

d_{X} (x_{n}, a) < 1/ n

but

d_{Y} (f (x_{n}), f (a)) \geq ε_{0}

. Step 4. Then

x_{n} \to a

yet

f (x_{n}) \neq \to f (a)

, contradicting the sequential hypothesis. Hence sequential continuity implies

ε

–

δ

continuity.

■

(Note: this equivalence uses that metric spaces are first-countable; in general topological spaces sequential continuity is weaker.)

Picture an $ε$ -target circle around $f (a)$ in $Y$ . Continuity is a game: an adversary picks any $ε$ (how close you must land), and you must respond with a $δ$ (a circle around $a$ ) so small that everything inside it lands inside the target. If you can always win, $f$ is continuous at $a$ .

Uniform continuity raises the stakes: you must produce a single

δ

that works at every point at once. Where the graph steepens without bound (like

1/ x

near

0

), you lose — no uniform

δ

survives.

The $\varepsilon$–$\delta$ game: a $\delta$-ball maps into an $\varepsilon$-ball

Maurice Fréchet | 1878–1973

Fréchet introduced abstract metric spaces in his 1906 doctoral thesis Sur quelques points du calcul fonctionnel, coining the term écart (deviation) for the distance and recognizing that continuity, completeness, and compactness could be studied axiomatically without reference to the real line. His framework unified the $ε$ – $δ$ analysis of Cauchy and Weierstrass with the emerging functional analysis of spaces of functions, paving the way for Hausdorff's general topology (1914). Fréchet also pioneered the notion of compactness via sequential limits and the study of function spaces as metric spaces — directly enabling the modern theory of continuous maps.

Worked Examples

Verifying continuity of $f(x)=x^2$ at $a=3$

Given:

f : R \to R

f (x) = x^{2}

, with

d (x, y) = ∣ x - y ∣

. Target tolerance

ε = 0.01

. Find a valid

δ

a = 3

. Steps:

$∣ f (x) - f (3) ∣ = ∣ x^{2} - 9∣ = ∣ x - 3∣ ∣ x + 3∣$ .
Restrict $∣ x - 3∣ < 1$ so $2 < x < 4$ , giving $∣ x + 3∣ < 7$ .
Then $∣ x^{2} - 9∣ < 7∣ x - 3∣$ . To force this below $0.01$ , require $∣ x - 3∣ < 0.01/7 \approx 0.001428$ .
Combine both: $δ = min {1, 0.01/7} = 0.001428 \dots$

∣ x - 3∣ < δ ⟹ ∣ x^{2} - 9∣ < 7∣ x - 3∣ < 7 δ = ε

⟳Bound the variable factor

∣ x + 3∣

by a constant via a preliminary restriction, then choose

δ

as the minimum of the restriction radius and

ε / (bound)

Answer:

δ = 0.001428

(dimensionless) guarantees

∣ x^{2} - 9∣ < 0.01

:::example[Showing $f (x) = 1/ x$ is continuous but NOT uniformly continuous on $(0, 1]$ ] Given: $f : (0, 1] \to R$ , $f (x) = 1/ x$ .

Pointwise continuity at $a \in (0, 1]$ :

$\frac{1}{x} - \frac{1}{a} = \frac{∣ x - a ∣}{∣ x ∣ a}$ .
Restrict $∣ x - a ∣ < a /2$ so $x > a /2$ , giving $\frac{1}{∣ x ∣ a} < \frac{2}{a ^{2}}$ .
Choose $δ = min {a /2, ε a^{2} /2}$ ; then $∣ f (x) - f (a) ∣ < \frac{2}{a ^{2}} ∣ x - a ∣ < ε$ . ✓ Continuous.

Failure of uniformity — concrete numbers: Take

ε = 1

. Try the points

x_{n} = \frac{1}{n}

y_{n} = \frac{1}{n + 1}

∣ x_{n} - y_{n} ∣ = \frac{1}{n ( n + 1 )} \to 0, ∣ f (x_{n}) - f (y_{n}) ∣ = ∣ n - (n + 1) ∣ = 1

⟳As points crowd toward

0

their separation shrinks but the difference of reciprocals stays fixed — no single

δ

can control all pairs.

With $n = 100$ : $∣ x_{n} - y_{n} ∣ = 1/10100 \approx 9.9 \times 1 0^{- 5}$ yet $∣ f (x_{n}) - f (y_{n}) ∣ = 1$ .

Answer:

f is continuous on (0, 1] but not uniformly continuous

; the obstruction is the unbounded steepness near

0

Common Misconceptions

"Continuous = uniformly continuous." False on non-compact domains.

1/ x

(0, 1]

and

x^{2}

R

are continuous but not uniformly so. However, by the Heine–Cantor theorem, a continuous map on a compact metric space is automatically uniformly continuous. " $δ$ can always be chosen independent of the point." No — in [4] the

δ

legitimately depends on both

ε

and

a

. Independence from

a

is precisely the extra demand of uniform continuity [6]. "Continuity preserves Cauchy sequences." Only uniform continuity guarantees that.

f (x) = 1/ x

maps the Cauchy sequence

1/ n

n

, which is not Cauchy. "A continuous bijection is a homeomorphism." No — the inverse may fail to be continuous (see below). It is a homeomorphism only if both

f

and

f^{- 1}

are continuous.

Continuous bijection whose inverse is discontinuous

Let $X = [0, 2 π)$ with the usual metric and $Y = S^{1} \subset R^{2}$ the unit circle. Define

f (t) = (cos t, sin t)

⟳This wraps the half-open interval once around the circle; it is a continuous bijection because each angle appears exactly once.

f

is a continuous bijection. But

f^{- 1}

is discontinuous at

(1, 0) = f (0)

: points just below

2 π

map near

(1, 0)

from below, yet their preimages (

\approx 2 π

) are far from

0

. Formally, the sequence

p_{n} = f (2 π - \frac{1}{n}) \to (1, 0) = f (0)

, but

f^{- 1} (p_{n}) = 2 π - \frac{1}{n} \to 2 π \neq = 0

. So continuity of

f

does not imply continuity of

f^{- 1}

Continuous image of a closed/bounded set need not be closed/bounded

Closedness fails: $f (x) = arctan x$ maps the closed set $R$ onto $(- π /2, π /2)$ , which is not closed.
Boundedness fails: $f (x) = 1/ x$ maps the bounded set $(0, 1]$ onto the unbounded $[1, \infty)$ .

The remedy is compactness: continuous images of compact sets are compact (hence closed and bounded in $R^{n}$ ) — this is the Extreme Value Theorem's engine.

Connections

Continuity sits at the crossroads of analysis and topology.

Topological Spaces: The open-set characterization (item 3 of the theorem) defines continuity in general topology, with metrics dispensed entirely.
Compactness in Metric Spaces: Continuous images of compact sets are compact; continuity on a compact domain upgrades to uniform continuity (Heine–Cantor).
Connectedness: Continuous images of connected sets are connected — the abstract Intermediate Value Theorem.
Complete Metric Spaces: Lipschitz/contraction maps (special continuous maps) drive the Banach Fixed-Point Theorem.
Homeomorphisms: A bicontinuous bijection; the equivalence relation of topology.
Uniform Convergence: A uniform limit of continuous functions is continuous, but a pointwise limit may not be.

Real-World Applications

Numerical analysis & error control: Lipschitz continuity bounds how input rounding errors propagate; condition numbers quantify the effective $L$ for solving $A x = b$ .
Differential equations: The Picard–Lindelöf existence theorem requires the vector field to be Lipschitz (a strong continuity) to guarantee unique solutions.
Machine learning: Lipschitz-constrained neural networks (e.g., spectral normalization in GANs, Wasserstein GAN critics) enforce stability and bound generalization error.
Computer graphics & morphing: Continuous deformations between shapes underlie animation, mesh interpolation, and texture mapping.
Signal processing: Continuity of filters ensures small input perturbations don't cause large output artifacts; stability analysis relies on it.
Robotics & control: Continuous feedback maps from sensor states to actuator commands guarantee smooth, non-chattering behavior.
Economics: Brouwer/Kakutani fixed-point theorems (requiring continuity) prove the existence of market equilibria and Nash equilibria.

Historical Development

1817 — Bolzano gives an early correct definition of continuity and proves the Intermediate Value Theorem, ahead of his time and largely unnoticed.
1821 — Cauchy publishes Cours d'analyse, defining continuity via infinitesimal increments — a major step toward rigor, though still verbally framed.
1854 — Riemann uses continuity carefully in the foundations of integration and function theory.
1861–1872 — Weierstrass formalizes the modern $ε$ – $δ$ definition in his Berlin lectures and exhibits a continuous but nowhere-differentiable function, shattering naive intuition.
1872 — Heine introduces uniform continuity and proves (with Cantor's compactness ideas) that continuous functions on closed intervals are uniformly continuous (Heine–Cantor theorem).
1906 — Fréchet defines abstract metric spaces, recasting continuity, completeness, and compactness axiomatically.
1914 — Hausdorff develops general topology in Grundzüge der Mengenlehre, giving the open-set definition of continuity its definitive home.
1922 — Banach establishes the fixed-point theorem for contraction maps in complete metric spaces, a cornerstone of applied continuity.

Subtleties & Limits

Beyond metric spaces. In general topological spaces, sequential continuity is strictly weaker than continuity (the equivalence relies on first-countability, which metric spaces always have). Characterizing which spaces are sequential (where sequences suffice) leads to the theory of Fréchet–Urysohn and sequential spaces. Extension problems. Given

f

uniformly continuous on a dense subset

D \subseteq X

into a complete space

Y

, it extends uniquely to all of

X

— but mere continuity on

D

does not. The precise interplay of density, completeness, and continuity is delicate. Regularity of inverses. Determining when a continuous bijection is a homeomorphism is nontrivial in general; Invariance of Domain (Brouwer, 1912) shows a continuous injection

R^{n} \to R^{n}

is open, but analogous statements fail across dimensions and in infinite dimensions, where compactness is scarce. Modulus of continuity. Quantifying how continuous a map is (Hölder, Lipschitz, log-Lipschitz exponents) connects to active research in PDE regularity and geometric measure theory.

Continuous Maps Between Metric Spaces

The Core Insight

Formal Statement

Metric Space

Continuity at a Point — $\varepsilon$–$\delta$

Equivalent Characterizations of Continuity

Uniform Continuity (stronger notion)

Why It Holds / Derivation

($\varepsilon$–$\delta$) $\iff$ (Open sets)

($\varepsilon$–$\delta$) $\iff$ (Sequential)

The $\varepsilon$–$\delta$ game: a $\delta$-ball maps into an $\varepsilon$-ball

Maurice Fréchet | 1878–1973

Worked Examples

Verifying continuity of $f(x)=x^2$ at $a=3$

Common Misconceptions

Continuous bijection whose inverse is discontinuous

Continuous image of a closed/bounded set need not be closed/bounded

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also

Before this

Where this leads