Uniform Convergence

Uniform convergence is the load-bearing hypothesis behind nearly every "interchange of limits" theorem in analysis: term-by-term differentiation and integration of series, the continuity of sums of power series on compact subsets of the disk of convergence, and the construction of continuous nowhere-differentiable functions (Weierstrass's function) as a uniformly convergent series of continuous pieces. In numerical analysis, uniform error bounds for approximating a function by polynomials (Weierstrass approximation theorem) or by truncated Fourier/Taylor series are precisely statements about uniform, not just pointwise, convergence — this is what guarantees an approximation is accurate everywhere on an interval simultaneously, not just at scattered sample points. In probability and statistics, the Glivenko–Cantelli theorem strengthens pointwise convergence of empirical distribution functions to a uniform statement (the basis of the Kolmogorov–Smirnov test). In numerical PDE and functional analysis, uniform convergence on compact sets underlies the topology of spaces like $C(K)$ and the Arzelà–Ascoli compactness criterion, and in complex analysis, uniform convergence on compact subsets of the domain is exactly the hypothesis needed for term-by-term differentiation of convergent series of holomorphic functions.

:::quiz[If $f_n \to f$ pointwise on $[0,1]$ and each $f_n$ is continuous, which conclusion is guaranteed?]

• $f$ is continuous on $[0,1]$
• $f$ is bounded on $[0,1]$
• $\int_0^1 f_n \to \int_0^1 f$
• None of these are guaranteed without an additional hypothesis such as uniform convergence ✓

:::quiz[For $f_n(x) = x^n$ on $[0,1]$, what is $\|f_n - f\|_\infty$ where $f$ is the pointwise limit?]

• It equals $0$ for all $n$
• It equals $1$ for every $n$, so convergence is not uniform ✓
• It tends to $0$ as $n \to \infty$
• It is undefined because $f$ is discontinuous

Ask the tutor: Can you show me step-by-step how to check whether $f_n(x) = \frac{x}{1+nx^2}$ converges uniformly on $\mathbb{R}$?

Picture the graph of $f$ as a fixed curve, and imagine an "$\varepsilon$-tube" of vertical half-width $\varepsilon$ drawn around it: the region $\{(x,y) : |y - f(x)| < \varepsilon\}$. Pointwise convergence says that for each individual $x$-coordinate, the graphs $y = f_n(x)$ eventually land inside the tube above that $x$ — but different $x$'s may take arbitrarily long to get there. Uniform convergence says something much stronger: from some index $N$ onward, the entire graph of $f_n$ lies inside the tube, all at once, over the whole domain. This is exactly why uniform convergence is the right tool for preserving continuity: if every $f_n$ is continuous and the whole graph of $f_n$ eventually squeezes into a thin tube around $f$, then $f$ itself cannot have a jump, because a thin enough tube around a jump cannot contain a continuous graph. Pointwise convergence gives no such global squeeze — the tube can be violated at points that keep sliding around as $n$ grows, which is precisely the mechanism behind the standard counterexamples.

:::svg[The $\varepsilon$-tube picture: $f_n(x)=x^n$ pointwise vs. uniform failure on $[0,1]$]

Example 1 (uniform convergence on a restricted domain, with rate computed explicitly). Let $f_n(x) = \dfrac{nx}{1+n^2x^2}$ for $x \in [1, \infty)$. Pointwise limit. For fixed $x > 0$, as $n \to \infty$,

So $f_n \to f \equiv 0$ pointwise on $[1,\infty)$ (and in fact on all of $\mathbb{R}$). Claim: the convergence is uniform on $[1,\infty)$. We compute $\|f_n - f\|_\infty = \sup_{x \ge 1} |f_n(x)|$ exactly. Treat $f_n$ as a function of the real variable $x>0$:

Differentiate in $x$:

This vanishes when $x = 1/n$, which is a maximum (check sign: $f_n'>0$ for $x<1/n$, $f_n'<0$ for $x>1/n$), with maximum value $f_n(1/n) = \frac{n \cdot (1/n)}{1 + n^2(1/n^2)} = \frac{1}{2}$. For $n \ge 1$, the critical point $x = 1/n \le 1$ lies outside the domain $[1,\infty)$ (except $n=1$, where $1/n=1$ is the left endpoint). Since $f_n$ is decreasing on $[1/n, \infty) \supseteq [1,\infty)$ for every $n\ge 1$, the supremum of $f_n$ on $[1,\infty)$ is attained at the left endpoint $x=1$:

Now compute the limit of this exact quantity:

since $\frac{n}{1+n^2} \le \frac{n}{n^2} = \frac1n \to 0$. By the sup-norm criterion, $f_n \to 0$ uniformly on $[1,\infty)$. Given $\varepsilon > 0$, an explicit valid $N$: since $\frac{n}{1+n^2} < \frac 1n$, it suffices to take $N > 1/\varepsilon$, e.g. $N = \lceil 1/\varepsilon\rceil + 1$; then for $n \ge N$, $\sup_{x\ge1}|f_n(x)| < \varepsilon$, and this $N$ manifestly does not depend on $x$.

Example 2 (uniform convergence fails on a larger domain — same formula, different domain). Now consider the same sequence $f_n(x) = \dfrac{nx}{1+n^2x^2}$, but on the domain $(0,\infty)$. Pointwise, we still get $f_n \to 0$ on $(0,\infty)$ by the identical computation as above (for each fixed $x>0$). But now the critical point $x = 1/n$ lies inside the domain $(0,\infty)$ for every $n$, and we found $f_n(1/n) = 1/2$. Hence

which does not tend to $0$. So $f_n \to 0$ fails to be uniform on $(0,\infty)$, even though it converges pointwise there. Concretely, taking $\varepsilon = 1/4$: no matter how large $N$ is, choose $x = 1/N$; then $f_N(x) = 1/2 > 1/4 = \varepsilon$. So no single $N$ works for all $x$ simultaneously — the definition of uniform convergence genuinely fails, and the failure is located exactly at the point $x = 1/n$, which "escapes to $0$" as $n \to \infty$ and is different for every $n$. This illustrates precisely the mechanism: pointwise convergence can hold everywhere while uniform convergence fails, because the "bad point" where $f_n$ stays far from $f$ is allowed to migrate with $n$.