Uniform Convergence and Differentiation

Term-by-term differentiation under the correct hypotheses is the rigorous backbone of solving ordinary and partial differential equations by series methods, including power series solutions and Frobenius-type expansions, where one must justify that a formally differentiated series really does represent the derivative of the sum. In Fourier analysis, this theorem (or its refinements) governs when a Fourier series can be differentiated term by term to recover the derivative of the represented function — a delicate question central to the study of Convergence of Fourier Series, since Fourier coefficients typically decay only like $1/k$ or $1/k^2$, right at the edge of what uniform convergence of the derivative series demands. In numerical analysis and physics, the same logic justifies differentiating asymptotic expansions and perturbation series term by term, provided the derivative series can be shown to converge uniformly (often via comparison tests analogous to the Weierstrass $M$-test used above). In functional analysis, the phenomenon that uniform convergence of functions does not control derivatives motivates the study of stronger topologies (e.g. $C^1$-norm convergence) and compactness criteria such as the Arzelà–Ascoli Theorem, which characterizes when a family of functions has uniformly convergent subsequences with controlled derivatives. In probability and statistics, uniform convergence of derivative sequences underlies rigorous justifications for differentiating moment generating functions or characteristic functions term by term to obtain moments.

:::quiz[If $f_n \to f$ uniformly on $[a,b]$ and each $f_n$ is differentiable, which additional hypothesis guarantees $f' = \lim f_n'$?]

• Nothing further is needed; uniform convergence of $f_n$ alone suffices
• $f_n'$ converges uniformly on $[a,b]$, and $f_n(x_0)$ converges at some point ✓
• $f_n'$ converges pointwise on $[a,b]$
• Each $f_n$ is a polynomial

:::quiz[In the counterexample $f_n(x) = \sin(nx)/\sqrt n$ on $[0,2\pi]$, which statement is correct?]

• $f_n \to 0$ pointwise but not uniformly
• $f_n' \to 0$ uniformly while $f_n$ diverges
• $f_n \to 0$ uniformly, but $f_n'(0) = \sqrt n$ diverges to infinity ✓
• Both $f_n$ and $f_n'$ converge uniformly to $0$

:::faq[Does the theorem require the interval $[a,b]$ to be bounded?] Yes, the standard statement requires a closed bounded interval, because the proof multiplies the uniform derivative bound by the factor $(b-a)$, which must be finite for the argument to close. On an unbounded domain like $\mathbb{R}$, one typically applies the theorem separately on each bounded subinterval $[-M,M]$ and observes that differentiability is a local property, so the conclusions patch together consistently across all reals, provided the hypotheses hold on every bounded subinterval.

Ask the tutor: Can you show me a sequence where $f_n'$ converges pointwise but not uniformly, and the limit's derivative genuinely disagrees with $\lim f_n'$?

Differentiation is a local slope operation, so it is exquisitely sensitive to small, fast oscillations that uniform convergence of the function values does nothing to suppress. A sequence of functions can be squeezed uniformly close to a limit — within a band of height $\varepsilon$ — while still wiggling with unbounded steepness inside that band. The graphs converge, but the slopes of the graphs do not. Uniform convergence of the derivatives is a fundamentally stronger condition: it forces the tangent lines of $f_n$ to converge uniformly to the tangent lines of a limiting slope field $g$. Once the slope field is under uniform control, and we anchor the family at a single point $x_0$ where we know $f_n(x_0)$ converges, the Mean Value Theorem lets us "integrate" that uniform control on slopes back into uniform control on the function values themselves — like reconstructing a road's elevation profile from a GPS-accurate uniformly convergent record of its grade, plus one known altitude checkpoint.

The picture shows the mechanism of every counterexample in this territory: squeeze the amplitude to zero while raising the frequency, and uniform convergence of function values is trivially preserved while the derivative diverges everywhere. This motivates why the theorem below places its central hypothesis on $f_n'$, not on $f_n$.

Example 1 — a convergent series of functions, differentiated term by term. Let

We show that $f(x) = \sum_{k=1}^\infty \frac{\sin(kx)}{k^3}$ is differentiable with derivative $f'(x) = \sum_{k=1}^\infty \frac{\cos(kx)}{k^2}$, on any interval $[a,b]$, e.g. $[a,b] = [-10, 7]$. Setup. Each $f_n$ is a finite sum of differentiable functions, hence differentiable everywhere, with

Convergence at a point. Take $x_0 = 0$: then $f_n(0) = \sum_{k=1}^n \frac{\sin(0)}{k^3} = 0$ for every $n$, so trivially $f_n(0) \to 0$. Uniform convergence of derivatives. For any $x \in \mathbb{R}$ and any $n < m$,

using $|\cos(kx)| \le 1$. The bound on the right is independent of $x$ and is the tail of the convergent series $\sum 1/k^2$ (a $p$-series with $p=2>1$), so it $\to 0$ as $n \to \infty$ uniformly in $m > n$. Hence $(f_n')$ is uniformly Cauchy on all of $\mathbb{R}$, and by completeness converges uniformly to some function

Apply the theorem. On $[-10,7]$ (or any bounded interval, or indeed all of $\mathbb{R}$ by applying the theorem on each $[-M,M]$), the hypotheses hold: convergence at $x_0=0$, and uniform convergence of derivatives. The theorem gives: $f_n \to f$ uniformly, $f$ is differentiable, and

Notice this argument used the Weierstrass M-test ($|\cos(kx)/k^2| \le 1/k^2$, $\sum 1/k^2 < \infty$) only on the derivative series — exactly the point of the theorem. We never needed to prove uniform convergence of $f_n$ itself directly; it fell out of the argument for free.

Example 2 — a case where the hypothesis must be checked carefully, with a genuinely asymmetric sequence. Let

This is a standard example, but examined here on the asymmetric interval $[1,3]$ (not centered anywhere near $0$), which forces us to actually use the endpoint behavior rather than symmetry. Step A: compute $f_n'$. By the quotient rule,

Step B: pointwise limits. For fixed $x \in [1,3]$, as $n \to \infty$,

(the second limit: numerator is $O(n)$, denominator is $O(n^2)$, so the ratio is $O(1/n) \to 0$). So pointwise $f_n \to f \equiv 0$ and $f_n' \to g \equiv 0$ on $[1,3]$, consistent with $f' = 0 = g$ — but we must check the convergence of $f_n'$ is genuinely uniform on $[1,3]$ before invoking the theorem, since that is its essential hypothesis. Step C: uniform bound on $|f_n'(x)|$ for $x \in [1,3]$. We estimate

For $n \geq 1$ and $x \ge 1$, we have $nx^2 \ge n \ge 1$, so for $n\ge 1$, $nx^2 - 1 \ge 0$ eventually (for $n\ge 1$ and $x\ge1$, $nx^2\ge 1$), giving $|1-nx^2| = nx^2-1 \le nx^2$. Hence

using $x^2 \ge 1$ on $[1,3]$. This bound $1/n$ is independent of $x$ and $\to 0$, so

confirming $f_n' \to 0$ uniformly on $[1,3]$. Step D: apply the theorem. Take $x_0 = 1 \in [1,3]$: $f_n(1) = \frac{1}{1+n} \to 0$, convergent. Both hypotheses hold on $[1,3]$, so the theorem gives $f_n \to f$ uniformly on $[1,3]$ (here $f \equiv 0$, consistent with the direct pointwise computation, and one can also check directly that $\sup_{x\in[1,3]}|f_n(x)| = \sup_x \frac{x}{1+nx^2}\to 0$), $f$ is differentiable, and

This matches direct computation, but the value of the theorem is that it guarantees this identification without needing an independent proof that $f=0$ is differentiable with derivative $g$ — the argument is completely mechanical once the two hypotheses are checked.