The Uniform Limit of Continuous Functions

This theorem is the load-bearing wall underneath enormous portions of analysis. Power series are uniformly convergent on compact subsets of their disk of convergence, which is exactly why functions like $e^x$, $\sin x$, and $\log(1+x)$, defined by their Taylor series, are guaranteed continuous — an argument that generalizes to the fully rigorous theory of analytic functions in complex analysis. <a class="wiki-link" href="/wiki/math-analysis-uniform-convergence-integration" data-track="wiki_article.body.wiki_link">Uniform Convergence and Integration</a> builds on the same $\varepsilon/3$ template to show that uniform limits can be integrated term by term, which is foundational for constructing solutions to differential equations via Picard iteration. Fourier series convergence theory constantly wrestles with exactly this dichotomy: pointwise Fourier convergence can badly fail continuity (Gibbs phenomenon at jump discontinuities), while Fejér's theorem restores good behavior by using Cesàro-averaged partial sums that do converge uniformly for continuous periodic functions. In approximation theory, the <span class="wiki-link-missing">Stone-Weierstrass Theorem</span> guarantees that continuous functions on a compact set can be uniformly approximated by polynomials (or more general algebras of functions) — and it is precisely because uniform limits preserve continuity that such approximations are meaningful replacements, not just numerically close but structurally faithful, for the original function. In numerical analysis, understanding whether an approximation scheme (splines, Bernstein polynomials, neural network approximants) converges uniformly rather than merely pointwise determines whether error bounds hold uniformly across the input domain — a concern central to reliable computation. In probability theory, the analogous fact underlies why uniform convergence of distribution functions (as opposed to merely pointwise) is the right notion for many convergence theorems, and in functional analysis, the theorem is the reason $C(K)$, continuous functions on a compact space with the sup norm, forms a Banach space at all — completeness of that space is inseparable from this continuity-preservation fact.

:::quiz[If $f_n \to f$ pointwise (not necessarily uniformly) on $[0,1]$ and each $f_n$ is continuous, which statement is correct?]

• $f$ must be continuous on $[0,1]$
• $f$ must be discontinuous somewhere on $[0,1]$
• $f$ may or may not be continuous; pointwise convergence alone does not decide it ✓
• $f$ must be bounded on $[0,1]$

:::quiz[Consider $f_n(x) = x^n$ on $[0, 1-\delta]$ for a fixed constant $0 < \delta < 1$. What can we correctly conclude?]

• The convergence is uniform there and the limit function is continuous (in fact identically $0$) ✓
• The convergence is not uniform there because it fails to be uniform on all of $[0,1]$
• The limit function has a jump discontinuity at $x = 1-\delta$
• Nothing can be concluded without checking $x=1$ explicitly

Ask the tutor: Can you show me why uniform convergence of the derivatives is needed for term-by-term differentiation, using the same $\varepsilon/3$ style argument?

The difference between pointwise and uniform convergence is the difference between "each point eventually settles down" and "all points settle down at the same speed." In pointwise convergence, the required index $N$ is allowed to depend on $x$: points near a boundary or near a kink can take arbitrarily long to converge, and this is exactly the loophole that lets discontinuities leak into the limit. Uniform convergence closes the loophole by demanding a single $N$ that works everywhere on $E$ simultaneously. Geometrically, uniform convergence $f_n \to f$ means: for every $\varepsilon > 0$, eventually the entire graph of $f_n$ lies inside an $\varepsilon$-tube around the graph of $f$ — not just that each vertical slice individually gets close. This tube picture is exactly what lets continuity pass through the limit: if $f_N$ is trapped in an $\varepsilon$-tube around $f$, and $f_N$ doesn't jump, then $f$ can't have "used" the tube's slack to hide a jump, because the tube has uniform width across all $x$.

Example 1: A uniformly convergent sequence, verified directly, with the $\varepsilon/3$ machinery made explicit. Let $E = [0,1]$ and define

Each $f_n$ is a ratio of polynomials with denominator $1 + x^n \geq 1 > 0$ on $[0,1]$, hence continuous on $[0,1]$. Claim: $f_n \to 0$ uniformly on $[0,1]$, where $f \equiv 0$. To see this, bound $f_n$ uniformly. For $x \in [0,1]$ we have $x^n \in [0,1]$ and $1 - x \in [0,1]$, so

since $\frac{1}{1+x^n}\le 1$ and $1-x\le1$. Now maximize $x^n$ on $[0,1]$: it attains its maximum value $1$ at $x=1$, but at $x=1$ the numerator $x^n(1-x)=0$, so we should be more careful and directly bound $\sup_x f_n(x)$. Differentiate: for $x\in(0,1)$, write $g_n(x) = x^n(1-x) = x^n - x^{n+1}$, so $g_n'(x) = n x^{n-1} - (n+1)x^n = x^{n-1}\big(n - (n+1)x\big)$. Setting $g_n'(x)=0$ gives the critical point $x^* = \frac{n}{n+1} \in (0,1)$. There,

since $\left(\frac{n}{n+1}\right)^n < 1$. Since $f_n(x) \le g_n(x)$ (as $\tfrac{1}{1+x^n}\le 1$), and $g_n$ attains its maximum on $[0,1]$ at $x^*$ (checking endpoints: $g_n(0)=g_n(1)=0$), we get

Given $\varepsilon > 0$, choose $N > 1/\varepsilon$; then for $n \geq N$, $\sup_x |f_n(x)| \le \frac{1}{n+1} < \varepsilon$. This $N$ does not depend on $x$, confirming uniform convergence to $f \equiv 0$. By the theorem, since each $f_n$ is continuous on $[0,1]$ and $f_n \to 0$ uniformly, the limit $f \equiv 0$ is continuous on $[0,1]$ — which we can also verify directly, since it's the zero function. This example is deliberately simple in its conclusion but shows the full uniform bound must be computed by calculus (a maximization), not guessed, which is typical of genuine uniform-convergence estimates.

Example 2: Applying the theorem to detect non-uniform convergence via its contrapositive. Let $E = [0,1]$ and $f_n(x) = x^n$. Each $f_n$ is a polynomial, hence continuous on $[0,1]$. Pointwise,

since for fixed $x \in [0,1)$, $x^n \to 0$, while $1^n = 1$ for all $n$. The pointwise limit $f$ has a jump discontinuity at $x_0 = 1$: $\lim_{x \to 1^-} f(x) = 0 \neq 1 = f(1)$. Now apply the contrapositive of the theorem: each $f_n$ is continuous on $[0,1]$, but the pointwise limit $f$ is not continuous at $x=1$. If the convergence $f_n \to f$ were uniform on $[0,1]$, the theorem would force $f$ to be continuous on $[0,1]$ — a contradiction. Therefore, the convergence $f_n \to f$ is not uniform on $[0,1]$, and we have deduced this purely from the theorem, without directly manipulating the definition of uniform convergence. (One can double-check this by exhibiting the failure explicitly: for $\varepsilon = 1/2$, no single $N$ works, since for any $n$, choosing $x = (1/2)^{1/n}$ close to $1$ gives $x^n = 1/2$, so $|f_n(x) - f(x)| = |1/2 - 0| = 1/2$ arbitrarily close to $x=1$, no matter how large $n$ is — this is the standard direct verification, but the point of this example is that the theorem itself, used contrapositively, is a fast diagnostic tool.) However, note that the convergence is uniform on $[0, 1-\delta']$ for any fixed $0 < \delta' < 1$: there $\sup_{x \in [0,1-\delta']} x^n = (1-\delta')^n \to 0$. This illustrates that uniformity is a property of convergence on a specific set, and restricting the domain to avoid the point where continuity would need to "jump" restores uniformity.