Uniform Continuity

Uniform continuity is not a curiosity confined to introductory courses — it is a load-bearing hypothesis throughout mathematics and its applications.

• Riemann integration. The standard proof that every continuous function on $[a,b]$ is Riemann integrable uses uniform continuity to make upper and lower Riemann sums converge to the same value: uniform continuity lets you choose a single partition mesh size that controls the oscillation of $f$ on every subinterval simultaneously.
• Uniform approximation theory. The Weierstrass approximation theorem (polynomials are dense in continuous functions on $[a,b]$ under the sup norm) and Bernstein's explicit polynomial construction both rely on uniform continuity of the target function on the compact interval to get uniform (not just pointwise) convergence of the approximants.
• Numerical analysis. Discretization error bounds for numerical integration (Riemann sums, trapezoidal rule) and root-finding algorithms depend on a uniform modulus of continuity to guarantee that mesh refinement improves accuracy everywhere at a controlled, predictable rate.
• Metric space theory and completions. A uniformly continuous function on a dense subset of a metric space extends uniquely to a continuous function on the completion; this is how the exponential function, Fourier transforms on $L^2$, and many operators in functional analysis are rigorously constructed by extension from a dense, simpler subspace.
• Signal processing and control theory. Uniform continuity of a system's response function guarantees that small, uniformly bounded perturbations in an input signal (across all times, not just locally) produce uniformly bounded changes in output — a stability guarantee stronger than pointwise continuity provides.
• Probability and statistics. Uniform continuity (and its quantitative refinement, uniform equicontinuity) underlies tightness arguments and the Arzelà–Ascoli theorem, which are central to proving convergence of stochastic processes and to establishing the existence of limiting distributions in empirical process theory.
• Cryptography and complexity. In computable analysis, the notion of a function being "uniformly computable" on a compact domain (computable with a modulus of continuity independent of the specific input) parallels uniform continuity and is essential for complexity bounds on numerical algorithms operating over real-number representations.

Ask the tutor: Why does uniform continuity guarantee that continuous functions on $[a,b]$ are Riemann integrable?

Think of continuity at a point $x_0$ as a promise: "give me an error tolerance $\varepsilon$, and I will find a neighborhood of $x_0$ small enough that $f$ doesn't move by more than $\varepsilon$ inside it." Nothing forbids that neighborhood from shrinking to nothing as $x_0$ ranges over the domain. Uniform continuity is the stronger promise that one neighborhood-radius $\delta$ works simultaneously for every basepoint $x_0$ in the domain. A useful mental picture: slide a "$2\delta \times 2\varepsilon$" rectangle (or box) along the graph of $f$, centered at each point $(x_0, f(x_0))$. Continuity says that for each $x_0$ some box of height $2\varepsilon$ (with $\varepsilon$ given) has some width $2\delta$ narrow enough that the graph stays inside the box as it passes through. Uniform continuity says one width $2\delta$ works for the box centered at every point on the graph. For $f(x) = 1/x$ on $(0,1)$, the graph gets steeper and steeper as $x \to 0^+$, so the box must get narrower and narrower — no single width works for all boxes.

The picture changes entirely if the domain is compact — closed and bounded. On $[a,b]$, a continuous function cannot "run off to infinity in slope" the way $1/x$ does as it approaches a missing endpoint, because every point of $[a,b]$, including the endpoints, is actually in the domain and continuity applies there too. This is the seed of the Heine–Cantor theorem: compactness prevents the domain from having any "escape route" where local moduli of continuity could degenerate.

Remark on the proof technique. This is the standard "sequential compactness" argument: turn a failure of a uniform ($\forall\delta$) statement into a sequence, extract a convergent subsequence using compactness of the domain, and derive a contradiction from ordinary (pointwise) continuity at the limit point. The same schema proves the Extreme Value Theorem and many other compactness-based results in analysis; it never invokes uniform continuity itself, so there is no circularity.

Example 1: A polynomial on a closed bounded interval, verified directly (not just via the theorem). Let $f(x) = x^3 - 2x$ on $[-2, 3]$. We verify uniform continuity directly with an explicit $\delta(\varepsilon)$, illustrating the mechanism the Heine–Cantor theorem guarantees abstractly. For $x, y \in [-2,3]$,

Since $x, y \in [-2,3]$, we have $|x| \le 3$ and $|y| \le 3$, so

Hence

Given $\varepsilon > 0$, choose

Then for all $x,y \in [-2,3]$ with $|x-y| < \delta$,

This $\delta$ depends only on $\varepsilon$ (not on $x$ or $y$), confirming uniform continuity directly — exactly as the theorem predicts, since $f$ is continuous (a polynomial) on the closed bounded interval $[-2,3]$. Notice the bound $29$ came from the boundedness of the domain: on an unbounded domain like all of $\mathbb{R}$, the analogous bound $|x^2+xy+y^2-2|$ would be unbounded, and indeed $f$ is not uniformly continuous on $\mathbb{R}$ (cubics grow too fast — see the counterexample below for the general phenomenon with $x^2$).

Example 2: Uniform continuity of $\sqrt{x}$ on $[0, 100]$, where the naive Lipschitz bound fails. Let $g(x) = \sqrt{x}$ on $[0, 100]$. This example is instructive because $g$ is not Lipschitz near $x=0$ (its derivative $\tfrac{1}{2\sqrt x}$ is unbounded as $x \to 0^+$), so the slick "bound the difference quotient" trick from Example 1 does not directly apply — yet the Heine–Cantor theorem still guarantees uniform continuity, and we can exhibit $\delta(\varepsilon)$ by a case split. $g$ is continuous on $[0,100]$ (standard fact about the square root function, continuous on its whole domain $[0,\infty)$, hence continuous on $[0,100]$ including the endpoint $x=0$ via a one-sided limit). By the Heine–Cantor theorem, $g$ is uniformly continuous on $[0,100]$. Let us verify with an explicit $\delta$. Fix $\varepsilon > 0$. We split into two regimes. Case A (both points near 0). If $x, y \in [0, \varepsilon^2]$, then $\sqrt{x}, \sqrt{y} \in [0, \varepsilon]$, so trivially

(We will absorb the factor of $2$ by using $\varepsilon/2$ in place of $\varepsilon$ throughout; see below.) Case B (at least one point bounded away from 0). Suppose $x, y \in [0,100]$ with $\max(x,y) \geq \varepsilon^2$ (so we are not both in the tiny interval above). Using the identity

if additionally $y \geq \varepsilon^2/4$ say, then $\sqrt x + \sqrt y \geq \varepsilon/2$, giving $|g(x)-g(y)| \le \tfrac{2}{\varepsilon}|x-y|$, so a Lipschitz-type bound does hold once we are bounded away from $0$. To assemble a single clean $\delta$, take (working with tolerance $\varepsilon$ from the start, redoing constants carefully): Given $\varepsilon>0$, let $\delta = \delta(\varepsilon) := \varepsilon^2$. We claim this works for all $x,y \in [0,100]$ with $|x-y|<\delta$. Assume without loss of generality $x \le y$ (the roles of $x,y$ are symmetric in $|g(x)-g(y)|$), so $0 \le y - x < \varepsilon^2$.

• If $x = 0$: then $y < \varepsilon^2$, so $|g(x)-g(y)| = \sqrt y < \sqrt{\varepsilon^2} = \varepsilon$.
• If $x > 0$: then since $(\sqrt y - \sqrt x)(\sqrt y + \sqrt x) = y - x$,

This bound alone isn't uniform as $x\to0$, so instead compare $\sqrt y - \sqrt x$ to $\sqrt{y-x}$ directly: since $t \mapsto \sqrt t$ is subadditive on $[0,\infty)$ (i.e. $\sqrt{a+b} \le \sqrt a + \sqrt b$ for $a,b\ge 0$, proven by squaring both sides: $a + b \le a + 2\sqrt{ab}+b$), we get

so

In both cases $|g(x)-g(y)| < \varepsilon$ whenever $|x-y|<\delta=\varepsilon^2$, and $\delta$ depends only on $\varepsilon$. This confirms uniform continuity of $\sqrt{x}$ on $[0,100]$ directly, matching the guarantee of the Heine–Cantor theorem, and shows the theorem's power: it spares us this case analysis whenever we only need existence of some uniform $\delta$, not an explicit formula.