A sequence of functions can converge at every single point and still misbehave everywhere at once. Uniform convergence — convergence in the sup norm — is the mode of convergence that analysis can trust.
Many of the most useful functions in mathematics are born as limits. The exponential is the limit of its Taylor polynomials; the solution of a differential equation is the limit of Picard iterates; a vibrating string is described by a Fourier series; a continuous nowhere-differentiable function is the sum of a series of smooth waves (§3.3). Each time we meet the same question: if every $f_n$ is continuous, or integrable, or differentiable, does the limit $f$ inherit the property — and may we compute $\int f$ or $f'$ from the $f_n$? With the most naive notion of limit, the answer is no: the continuous functions $x^n$ on $[0,1]$ converge to a function with a jump. The problem is that "converging at each point" allows the speed of convergence to deteriorate from point to point without any control.
This section introduces the two notions — pointwise and uniform convergence — and the tools for telling them apart: the sup-norm criterion, the Cauchy criterion, the Weierstrass M-test, the Dirichlet and Abel tests for series of functions, and Dini's theorem, which upgrades monotone pointwise convergence to uniform convergence on compact sets. What uniform convergence buys us (continuity, integrals, derivatives, power series) is the subject of §5.2. CSIR NET examines this material relentlessly: Part B asks for $\sup_x|f_n(x)-f(x)|$ or whether a named sequence converges uniformly on a named set; Part C offers four sequences or series and asks which converge uniformly — and the wrong options are always the standard examples collected below.
Pointwise Convergence and What It Loses
📘 Definition 5.1 — Pointwise convergence
Let $D\subseteq\mathbb{R}$ and let $f_n:D\to\mathbb{R}$ $(n\in\mathbb{N})$ and $f:D\to\mathbb{R}$. The sequence $(f_n)$ converges pointwise to $f$ on $D$ if $f_n(x)\to f(x)$ for every $x\in D$; that is,
The function $f$ is the pointwise limit; it is unique because limits of real sequences are unique. The integer $N=N(\varepsilon,x)$ is allowed to depend on both $\varepsilon$ and the point $x$.
📘 Definition 5.2 — Series of functions
Given $u_n:D\to\mathbb{R}$, the series of functions $\sum u_n$ is the sequence of its partial sums $S_N(x)=\sum_{n=1}^{N}u_n(x)$ (or $\sum_{n=0}^N$). It converges pointwise on $D$ to $S$ if $S_N\to S$ pointwise, and converges absolutely on $D$ if $\sum|u_n(x)|$ converges for every $x\in D$. For a convergent series, $R_N=S-S_N$, that is $R_N(x)=\sum_{n\gt N}u_n(x)$, is the $N$-th tail (remainder). Every statement about sequences of functions is a statement about series and conversely, exactly as for series of numbers (§2.3).
Pointwise convergence is simply convergence of real sequences, one for each $x$; all of Chapter 2 applies to each of these sequences. The trouble starts when we ask about properties that involve many points at once — continuity (nearby points), integrals (all points), derivatives (nearby points again). The next example shows each of them failing.
✏️ Example 1 — Four things a pointwise limit can destroy
(a) Find the pointwise limit of $f_n(x)=x^n$ on $[0,1]$. (b) Let $g_n(x)=n^2x(1-x^2)^n$ on $[0,1]$. Show $g_n\to0$ pointwise but $\int_0^1g_n\to\infty$. (c) Let $h_n(x)=\dfrac{x}{1+nx^2}$ on $\mathbb{R}$. Show that $h_n\to0$ pointwise while $h_n'(0)\not\to0$. (d) Show that $\displaystyle\lim_{x\to1^-}\lim_{n\to\infty}x^n\ne\lim_{n\to\infty}\lim_{x\to1^-}x^n$.
💡 Show solution▶
(a) Continuity lost For $0\le x\lt1$, $x^n\to0$ (§2.1); for $x=1$, $x^n=1$. So the limit is $f=0$ on $[0,1)$ and $f(1)=1$. Every $f_n$ is a polynomial, yet $f$ has a jump at $1$.
(b) Integrals lost $g_n(0)=g_n(1)=0$. For $0\lt x\lt1$ put $q=1-x^2\in(0,1)$; then $g_n(x)=x\cdot n^2q^n\to0$ because $n^kq^n\to0$ for $0\le q\lt1$ (a standard limit of §2.1). With $u=1-x^2$, $du=-2x\,dx$, $$\int_0^1g_n(x)\,dx=\frac{n^2}{2}\int_0^1u^n\,du=\frac{n^2}{2(n+1)}\to\infty,$$ while $\int_0^1\lim g_n=0$. The mass of $g_n$ concentrates in an ever taller, ever thinner hump near $x=0$.
(c) Derivatives lost $h_n(0)=0$ and, for $x\ne0$, $|h_n(x)|\le\dfrac{|x|}{nx^2}=\dfrac1{n|x|}\to0$. But $$h_n'(x)=\frac{1-nx^2}{(1+nx^2)^2},\qquad h_n'(0)=1\ \text{ for every }n,$$ so $\lim_n h_n'(0)=1\ne0=(\lim_n h_n)'(0)$. (We shall see that here $h_n\to0$ even uniformly — Example 2(c) — so uniform convergence alone does not protect derivatives; §5.2.)
(d) Limits do not commute For fixed $x\in(0,1)$, $\lim_nx^n=0$, so the left side is $\lim_{x\to1^-}0=0$. For fixed $n$, $\lim_{x\to1^-}x^n=1$, so the right side is $1$. Continuity of $f$ at $1$ (part (a)) would say exactly that these two iterated limits agree.
Uniform Convergence and the Sup-Norm Criterion
To keep the properties that pointwise limits destroy, we demand that one $N$ work for all $x$ simultaneously.
📘 Definition 5.3 — Uniform convergence
$(f_n)$ converges uniformly to $f$ on $D$, written $f_n\rightrightarrows f$ on $D$, if
Compared with Definition 5.1, the quantifier "$\forall x$" has moved from the front to the back: now $N=N(\varepsilon)$ depends on $\varepsilon$ only. Geometrically: for $n\ge N$ the whole graph of $f_n$ lies inside the $\varepsilon$-tube $\{(x,y):x\in D,\ |y-f(x)|\lt\varepsilon\}$ around the graph of $f$. A series $\sum u_n$ converges uniformly on $D$ if its partial sums do.
📘 Definition 5.4 — The sup norm
For $g:D\to\mathbb{R}$ put $\|g\|_D=\sup_{x\in D}|g(x)|\in[0,\infty]$ (written $\|g\|_\infty$ when $D$ is clear). It satisfies $\|g\|_D=0\iff g=0$, $\|cg\|_D=|c|\,\|g\|_D$ and $\|g+h\|_D\le\|g\|_D+\|h\|_D$; on the bounded functions it is a norm, the sup norm (or uniform norm), studied further in §7.4.
📐 Theorem 5.1 — The sup-norm criterion
Let $f_n,f:D\to\mathbb{R}$ and $M_n=\sup_{x\in D}|f_n(x)-f(x)|$ $(=\|f_n-f\|_D)$. Then $f_n\rightrightarrows f$ on $D$ if and only if $M_n\to0$ (in particular $M_n\lt\infty$ for all large $n$).
Proof. Suppose $f_n\rightrightarrows f$ and let $\varepsilon\gt0$. Choose $N$ as in Definition 5.3. For $n\ge N$ the number $\varepsilon$ is an upper bound of $\{|f_n(x)-f(x)|:x\in D\}$, so $M_n\le\varepsilon$. Hence $M_n\to0$. Conversely, if $M_n\to0$ and $\varepsilon\gt0$, choose $N$ with $M_n\lt\varepsilon$ for $n\ge N$; then $|f_n(x)-f(x)|\le M_n\lt\varepsilon$ for all $n\ge N$ and all $x\in D$. ∎
📐 Theorem 5.2 — Basic properties
Let $f_n,g_n,f,g:D\to\mathbb{R}$.
If $f_n\rightrightarrows f$ on $D$ then $f_n\to f$ pointwise; so the only candidate for a uniform limit is the pointwise limit.
If $f_n\rightrightarrows f$ on $D$ then $f_n\rightrightarrows f$ on every $E\subseteq D$. If $f_n\rightrightarrows f$ on each of finitely many sets $D_1,\dots,D_k$, then $f_n\rightrightarrows f$ on $D_1\cup\cdots\cup D_k$.
If $f_n\rightrightarrows f$ and $g_n\rightrightarrows g$ on $D$ and $a,b\in\mathbb{R}$, then $af_n+bg_n\rightrightarrows af+bg$ on $D$.
Proof. (1) For fixed $x$, $|f_n(x)-f(x)|\le\|f_n-f\|_D\to0$. (2) $\|f_n-f\|_E\le\|f_n-f\|_D$ because the supremum over a smaller set is smaller. For the union, $\|f_n-f\|_{D_1\cup\cdots\cup D_k}=\max_j\|f_n-f\|_{D_j}$, a maximum of finitely many sequences tending to $0$. (3) By the triangle inequality for the sup norm, $\|(af_n+bg_n)-(af+bg)\|_D$ is at most $|a|\,\|f_n-f\|_D+|b|\,\|g_n-g\|_D$, which tends to $0$. ∎
Part (2) fails for infinite unions: $x^n\rightrightarrows0$ on each $[0,1-\tfrac1k]$ but not on their union $[0,1)$ (Example 2(a)). To disprove uniform convergence one rarely computes a supremum exactly; it is enough to exhibit bad points.
📐 Theorem 5.3 — Detecting non-uniform convergence
Suppose $f_n\to f$ pointwise on $D$. Then $(f_n)$ fails to converge uniformly to $f$ on $D$ if and only if there are $\varepsilon_0\gt0$, indices $n_1\lt n_2\lt\cdots$ and points $x_k\in D$ with $|f_{n_k}(x_k)-f(x_k)|\ge\varepsilon_0$ for all $k$. In particular, if some sequence $(x_n)$ in $D$ has $f_n(x_n)-f(x_n)\not\to0$, the convergence is not uniform.
Proof. By Theorem 5.1, non-uniform convergence means $M_n\not\to0$: there is $\varepsilon_0\gt0$ with $M_n\gt\varepsilon_0$ for infinitely many $n$, say $n_1\lt n_2\lt\cdots$. Since $\varepsilon_0$ is smaller than the supremum $M_{n_k}$, it is not an upper bound, so some $x_k\in D$ has $|f_{n_k}(x_k)-f(x_k)|\gt\varepsilon_0$. Conversely, such points give $M_{n_k}\ge\varepsilon_0$ for all $k$, so $M_n\not\to0$. ∎
🧭 Problem-solving strategy — is $f_n\to f$ uniform on $D$?
Find the pointwise limit $f$ (fix $x$; split into cases such as $x=0$, $0\lt x\lt1$, $x=1$).
Compute or bound $M_n=\sup_D|f_n-f|$. Often calculus: differentiate $f_n-f$ in $x$, locate the maximum $x_n^*$, evaluate. Or an inequality such as AM–GM, $|\sin t|\le|t|$, $1+t\le e^t$.
Decide. $M_n\to0$ proves uniform convergence (Theorem 5.1). If $M_n\not\to0$, record the bad points $x_n^*$ (Theorem 5.3). A quick sufficient test for failure, proved in §5.2: the $f_n$ are continuous but $f$ is not.
Localise. Look at where the bad points go ($x_n^*\to1$, $x_n^*\to0$, $x_n^*\to\infty$). Removing a neighbourhood of that place usually restores uniform convergence: $[0,a]$ with $a\lt1$, $[a,\infty)$ with $a\gt0$, bounded sets.
✏️ Example 2 — The three standard examples
Decide where the convergence is uniform: (a) $f_n(x)=x^n$ on $[0,1]$ and on $[0,a]$, $0\lt a\lt1$; (b) $f_n(x)=nxe^{-nx}$ on $[0,\infty)$ and on $[a,\infty)$, $a\gt0$; (c) $f_n(x)=\dfrac{x}{1+nx^2}$ on $\mathbb{R}$.
💡 Show solution▶
(a) The limit is $f=0$ on $[0,1)$, $f(1)=1$ (Example 1). $|f_n-f|=x^n$ on $[0,1)$ and $0$ at $1$, so $M_n=\sup_{0\le x\lt1}x^n=1$ (approached as $x\to1^-$, not attained). Not uniform. Concretely, $x_n=2^{-1/n}\in(0,1)$ gives $f_n(x_n)-f(x_n)=\tfrac12$ for every $n$. On $[0,a]$: $M_n=a^n\to0$, so the convergence is uniform. The bad points $x_n=2^{-1/n}\to1$ escape every $[0,a]$.
(b) $f_n(0)=0$, and for $x\gt0$, $nxe^{-nx}=x\cdot ne^{-nx}\to0$. So $f=0$. Now $f_n'(x)=ne^{-nx}(1-nx)$ vanishes at $x=\tfrac1n$, where $f_n$ has its maximum $f_n(\tfrac1n)=e^{-1}$. So $M_n=e^{-1}$ for every $n$: not uniform on $[0,\infty)$ — nor on $[0,1]$. On $[a,\infty)$: once $n\ge1/a$ we have $\tfrac1n\le a$, $f_n$ is decreasing on $[a,\infty)$, and $M_n=f_n(a)=nae^{-na}\to0$. Uniform. The hump of fixed height $e^{-1}$ slides into $x=0$.
(c) $f_n\to0$ pointwise (Example 1(c)). By AM–GM, $1+nx^2\ge2\sqrt n\,|x|$, so $$|f_n(x)|\le\frac{|x|}{2\sqrt n\,|x|}=\frac1{2\sqrt n}\qquad(x\ne0),$$ with equality at $x=\pm\tfrac1{\sqrt n}$. Hence $M_n=\tfrac1{2\sqrt n}\to0$: uniform on all of $\mathbb{R}$.
Check $n=4$: $f_4(\tfrac12)=\dfrac{1/2}{1+1}=\tfrac14=\dfrac1{2\sqrt4}$ ✓. In (b) with $n=3$: $f_3(\tfrac13)=e^{-1}\approx0.3679$ ✓.
Interactive 1 — The $\varepsilon$-Tube Around the Limit
The thick dark curve is the pointwise limit $f$, the amber band is the $\varepsilon$-tube around it, and the red curve is $f_n$. The blue segment marks where $|f_n-f|$ is largest. Uniform convergence means: for every $\varepsilon$, from some $N$ on, the red curve stays entirely inside the tube. Press Play to increase $n$; shrink $\varepsilon$ and watch whether $N(\varepsilon)$ exists.
n—
ε—
Mₙ = sup |fₙ − f|—
attained at—
N(ε): Mₙ < ε for n ≥ N—
✏️ Example 3 — Prove or disprove
Let $f_n\rightrightarrows f$ and $g_n\rightrightarrows g$ on $D$. Prove or disprove: (a) if every $f_n$ is bounded, then $f$ is bounded and the $f_n$ are uniformly bounded ($|f_n(x)|\le K$ for all $n$ and $x$); (b) $f_ng_n\rightrightarrows fg$ on $D$; (c) if $\varphi:\mathbb{R}\to\mathbb{R}$ is uniformly continuous then $\varphi\circ f_n\rightrightarrows\varphi\circ f$, but "uniformly" cannot be weakened to "continuous"; (d) if $D=[0,1]$, each $f_n$ is continuous and $(f_n)$ converges uniformly on $\mathbb{Q}\cap[0,1]$, then $(f_n)$ converges uniformly on $[0,1]$.
💡 Show solution▶
(a) True Choose $N$ with $\|f_n-f\|_D\le1$ for $n\ge N$, and let $|f_N|\le K_N$. Then $|f|\le|f-f_N|+|f_N|\le1+K_N$: $f$ is bounded. For $n\ge N$, $|f_n|\le|f_n-f|+|f|\le2+K_N$. If $|f_j|\le K_j$ for $j\lt N$, then $K=\max(K_1,\dots,K_{N-1},\,2+K_N)$ bounds every $f_n$. (Pointwise convergence gives neither conclusion: Exercise IV.2.)
(b) False in general On $D=\mathbb{R}$ take $f_n(x)=x+\tfrac1n$, $g_n(x)=x$: both converge uniformly ($\|f_n-x\|=\tfrac1n$). But $f_ng_n-x^2=\tfrac xn$, whose supremum over $\mathbb{R}$ is $\infty$ for every $n$. True if $f$ and $g$ are bounded: then by (a) $|g_n|\le K$ for $n\ge N$, and $$|f_ng_n-fg|\le|f_n-f|\,|g_n|+|f|\,|g_n-g|\le K\|f_n-f\|+\|f\|\,\|g_n-g\|\to0 .$$
(c) True; and the weakening fails Given $\varepsilon\gt0$ choose $\delta$ from uniform continuity of $\varphi$ and $N$ with $\|f_n-f\|\lt\delta$ for $n\ge N$. Then $|\varphi(f_n(x))-\varphi(f(x))|\lt\varepsilon$ for all $x$ and $n\ge N$. With $\varphi(t)=t^2$ (continuous, not uniformly continuous) and $f_n(x)=x+\tfrac1n$ on $\mathbb{R}$: $\varphi(f_n)-\varphi(f)=\tfrac{2x}n+\tfrac1{n^2}$ is unbounded, so the convergence is not uniform.
(d) True Let $\varepsilon\gt0$ and choose $N$ with $|f_n(q)-f_m(q)|\lt\varepsilon$ for all rational $q\in[0,1]$ and $m,n\ge N$ (a uniformly convergent sequence is uniformly Cauchy — Theorem 5.4 below). For any $x\in[0,1]$ pick rationals $q_j\to x$; since $f_n-f_m$ is continuous, $|f_n(x)-f_m(x)|=\lim_j|f_n(q_j)-f_m(q_j)|\le\varepsilon$. So $(f_n)$ is uniformly Cauchy on $[0,1]$ and converges uniformly there by Theorem 5.4. A dense set and continuity are enough to control the whole interval.
The Cauchy Criterion and the Weierstrass M-test
Very often the limit function is unknown — the sum of a series, say. Then we need a test for uniform convergence that does not mention $f$. As for sequences of numbers (§2.1), the answer is a Cauchy criterion.
📘 Definition 5.5 — Uniformly Cauchy sequence
$(f_n)$ is uniformly Cauchy on $D$ if for every $\varepsilon\gt0$ there is $N$ such that $|f_n(x)-f_m(x)|\lt\varepsilon$ for all $m,n\ge N$ and all $x\in D$; equivalently, $\sup_{m,n\ge N}\|f_n-f_m\|_D\to0$ as $N\to\infty$.
📐 Theorem 5.4 — Cauchy criterion for uniform convergence
A sequence $(f_n)$ of functions $D\to\mathbb{R}$ converges uniformly on $D$ (to some function $f$) if and only if it is uniformly Cauchy on $D$.
Proof. ($\Rightarrow$) If $f_n\rightrightarrows f$, choose $N$ with $\|f_n-f\|_D\lt\varepsilon/2$ for $n\ge N$. For $m,n\ge N$ and $x\in D$, $|f_n(x)-f_m(x)|$ is at most $|f_n(x)-f(x)|+|f(x)-f_m(x)|\lt\varepsilon$.
($\Leftarrow$) Suppose $(f_n)$ is uniformly Cauchy. For each fixed $x\in D$ the real sequence $(f_n(x))$ is Cauchy, hence convergent by the completeness of $\mathbb{R}$ (§2.1); call its limit $f(x)$. This defines $f:D\to\mathbb{R}$ with $f_n\to f$ pointwise. Now let $\varepsilon\gt0$ and choose $N$ with $|f_n(x)-f_m(x)|\lt\varepsilon$ for all $m,n\ge N$, $x\in D$. Fix $n\ge N$ and $x\in D$, and let $m\to\infty$ in this inequality: since $f_m(x)\to f(x)$ and $t\mapsto|f_n(x)-t|$ is continuous, $|f_n(x)-f(x)|\le\varepsilon$. As $N$ did not depend on $x$, this says $\|f_n-f\|_D\le\varepsilon$ for every $n\ge N$, i.e. $f_n\rightrightarrows f$. ∎
The key move — fix $x$, let $m\to\infty$, and notice that the bound did not depend on $x$ — is used again and again in this chapter. Applied to partial sums it gives the criterion for series.
📐 Theorem 5.5 — Cauchy criterion for series of functions
The series $\sum u_n$ converges uniformly on $D$ if and only if for every $\varepsilon\gt0$ there is $N$ with $$\Big|\sum_{k=m+1}^{n}u_k(x)\Big|\lt\varepsilon$$ for all $n\gt m\ge N$ and all $x\in D$. Equivalently, the series converges pointwise and its tails satisfy $\|R_N\|_D\to0$. Consequently, if $\sum u_n$ converges uniformly on $D$, then $u_n\rightrightarrows0$ on $D$, i.e. $\|u_n\|_D\to0$.
Proof. The displayed sum is $S_n(x)-S_m(x)$, so the first statement is Theorem 5.4 for the partial sums. The second is Theorem 5.1, because $S-S_N=R_N$. Finally, taking $n=m+1$ gives $|u_{m+1}(x)|\lt\varepsilon$ for all $m\ge N$ and $x\in D$. ∎
The last statement is a necessary condition only: in Example 4(e) the terms tend to $0$ uniformly but the series does not converge uniformly. The most useful sufficient condition compares the series with a convergent series of numbers.
📐 Theorem 5.6 — Weierstrass M-test
Let $u_n:D\to\mathbb{R}$ and suppose there are constants $M_n\ge0$ with $|u_n(x)|\le M_n$ for all $x\in D$ and all $n$, and $\sum_{n=1}^\infty M_n\lt\infty$. Then $\sum u_n$ and $\sum|u_n|$ both converge uniformly on $D$; in particular $\sum u_n$ converges absolutely and uniformly. Moreover, the tails satisfy $\|R_N\|_D\le\sum_{n\gt N}M_n$.
Proof. Let $\varepsilon\gt0$. By the Cauchy criterion for the convergent series $\sum M_n$ (§2.3) there is $N$ with $\sum_{k=m+1}^nM_k\lt\varepsilon$ for $n\gt m\ge N$. Then for every $x\in D$, $$\begin{aligned}\Big|\sum_{k=m+1}^nu_k(x)\Big|&\le\sum_{k=m+1}^n|u_k(x)|\\&\le\sum_{k=m+1}^nM_k\lt\varepsilon .\end{aligned}$$ By Theorem 5.5 both $\sum u_n$ and $\sum|u_n|$ converge uniformly. For the tail bound, let $n\to\infty$ in $|S_n(x)-S_N(x)|\le\sum_{k=N+1}^nM_k$. ∎
The best constants are $M_n=\|u_n\|_D$; so the M-test applies exactly when $\sum_n\|u_n\|_D\lt\infty$. When this sum is infinite the M-test is silent — it does not prove non-uniform convergence.
✏️ Example 4 — The M-test at work
Decide which series converge uniformly on the given set: (a) $\sum\dfrac{\cos nx}{n^2}$ on $\mathbb{R}$; how many terms make the error less than $10^{-3}$ everywhere? (b) $\sum\dfrac{x^n}{n^2}$ on $[-1,1]$; (c) $\sum\dfrac{x}{n(1+nx^2)}$ on $\mathbb{R}$; (d) $\sum_{n\ge1}x^2e^{-nx}$ on $[0,\infty)$; (e) $\sum_{n\ge0}x^n(1-x)$ on $[0,1]$.
💡 Show solution▶
(a) $\left|\frac{\cos nx}{n^2}\right|\le\frac1{n^2}=M_n$ and $\sum\frac1{n^2}\lt\infty$: uniform on $\mathbb{R}$. Tail: $\|R_N\|\le\sum_{n\gt N}\frac1{n^2}\lt\int_N^\infty\frac{dt}{t^2}=\frac1N$, so $N=1000$ terms guarantee error below $10^{-3}$ at every $x$ simultaneously.
(b) For $|x|\le1$, $\left|\frac{x^n}{n^2}\right|\le\frac1{n^2}$: uniform on $[-1,1]$ — including both end points.
(c) By Example 2(c), $\left|\frac{x}{1+nx^2}\right|\le\frac1{2\sqrt n}$, so $|u_n(x)|\le\frac1{2n^{3/2}}=M_n$ and $\sum M_n\lt\infty$: uniform on $\mathbb{R}$. Here a crude bound such as $|x|/n$ would have failed: the sup has to be computed.
(d) $\frac{d}{dx}\left(x^2e^{-nx}\right)=xe^{-nx}(2-nx)$, so the maximum on $[0,\infty)$ is at $x=\frac2n$: $\|u_n\|=\frac4{n^2e^2}$. $\sum\frac4{e^2n^2}\lt\infty$: uniform on $[0,\infty)$. (Its sum is $\frac{x^2}{e^x-1}$ for $x\gt0$ and $0$ at $x=0$ — a geometric series.)
(e) $\|u_n\|=\max_{[0,1]}x^n(1-x)=\frac{n^n}{(n+1)^{n+1}}=\frac1{n+1}\left(1+\frac1n\right)^{-n}\ge\frac1{e(n+1)}$, so $\sum\|u_n\|=\infty$ and the M-test says nothing. In fact the partial sums telescope: $S_N(x)=\sum_{n=0}^N(x^n-x^{n+1})=1-x^{N+1}$, so $S=1$ on $[0,1)$, $S(1)=0$, and $\|R_N\|_{[0,1)}=\sup_{x\lt1}x^{N+1}=1$: not uniform — although $\|u_n\|\to0$.
The M-test only sees absolute values. Series whose terms change sign can converge uniformly although $\sum\|u_n\|=\infty$; for them we need the function versions of the Dirichlet and Abel tests of §2.3.
📐 Theorem 5.7 — Dirichlet's and Abel's tests for uniform convergence
Let $a_n,b_n:D\to\mathbb{R}$, and suppose that for each $x\in D$ the sequence $(b_n(x))_n$ is monotone (the direction may depend on $x$).
Dirichlet. If the partial sums $A_n(x)=\sum_{k=1}^na_k(x)$ are uniformly bounded, $|A_n(x)|\le M$ for all $n$ and $x\in D$, and $b_n\rightrightarrows0$ on $D$, then $\sum a_nb_n$ converges uniformly on $D$.
Abel. If $\sum a_n$ converges uniformly on $D$ and the $b_n$ are uniformly bounded, $|b_n(x)|\le K$ for all $n$ and $x\in D$, then $\sum a_nb_n$ converges uniformly on $D$.
Proof. Both rest on Abel's summation by parts (§2.3): if $T_k=\sum_{j=m}^ka_j$ for $k\ge m$ (and $T_{m-1}=0$), then $$\sum_{k=m}^na_kb_k=T_nb_n+\sum_{k=m}^{n-1}T_k\,(b_k-b_{k+1}),$$ obtained by writing $a_k=T_k-T_{k-1}$ and regrouping. Since $(b_k(x))$ is monotone, the differences $b_k-b_{k+1}$ all have the same sign, so $\sum_{k=m}^{n-1}|b_k-b_{k+1}|=|b_m-b_n|$.
Dirichlet. Here $T_k=A_k-A_{m-1}$, so $|T_k(x)|\le2M$. A monotone sequence with limit $0$ has $|b_n(x)|\le|b_m(x)|$ for $n\ge m$. Hence $$\begin{aligned}\Big|\sum_{k=m}^na_k(x)b_k(x)\Big|&\le2M|b_n|+2M|b_m-b_n|\\&\le6M|b_m(x)|\le6M\|b_m\|_D ,\end{aligned}$$ which tends to $0$ as $m\to\infty$, uniformly in $x$ and $n$. Theorem 5.5 applies.
Abel. Given $\varepsilon\gt0$, the uniform Cauchy criterion for $\sum a_n$ gives $N$ with $|T_k(x)|\lt\varepsilon$ for all $k\ge m\ge N$ and $x\in D$. Then $$\begin{aligned}\Big|\sum_{k=m}^na_kb_k\Big|&\le\varepsilon|b_n|+\varepsilon|b_m-b_n|\\&\le\varepsilon K+2\varepsilon K=3K\varepsilon\end{aligned}$$ for $n\ge m\ge N$, and Theorem 5.5 applies again. ∎
The two trigonometric estimates of §2.3 become uniform estimates on sets that stay away from the zeros of $\sin\frac x2$: for $\delta\le x\le2\pi-\delta$ (with $0\lt\delta\lt\pi$), $$\Big|\sum_{k=1}^n\sin kx\Big|\le\frac1{\sin(\delta/2)},$$ $$\Big|\sum_{k=1}^n\cos kx\Big|\le\frac1{\sin(\delta/2)} .$$ Abel's test is the engine behind Abel's theorem on power series (§5.2).
✏️ Example 5 — Beyond the M-test
(a) Show that $\sum_{n\ge1}\dfrac{(-1)^{n+1}}{n+x^2}$ converges uniformly on $\mathbb{R}$ but absolutely at no point. (b) Show that $\sum\dfrac{\sin nx}{n}$ converges uniformly on $[\delta,2\pi-\delta]$ for each $0\lt\delta\lt\pi$, but not on $(0,2\pi)$. (c) Find the sum of $\sum_{n\ge0}\dfrac{x^2}{(1+x^2)^n}$ on $\mathbb{R}$ and decide where the convergence is uniform.
💡 Show solution▶
(a) For each fixed $x$, $b_n=\frac1{n+x^2}$ decreases to $0$, so the Leibniz test applies, with the remainder bound $|R_N(x)|\le b_{N+1}=\frac1{N+1+x^2}\le\frac1{N+1}$ (§2.3). The bound does not depend on $x$: $\|R_N\|_{\mathbb{R}}\le\frac1{N+1}\to0$, uniform convergence. But $\sum\frac1{n+x^2}$ diverges for every $x$ (limit comparison with $\sum\frac1n$), and $\|u_n\|=\frac1n$, so the M-test could never have worked.
(b) Uniform on $[\delta,2\pi-\delta]$ Dirichlet's test with $a_n(x)=\sin nx$ (partial sums bounded by $1/\sin(\delta/2)$ there) and $b_n=\frac1n$ (constant in $x$, decreasing, $\|b_n\|=\frac1n\to0$).
(b) Not uniform on $(0,2\pi)$ We violate the Cauchy criterion. Given $m$, take $x_m=\frac{\pi}{4m}$. For $m+1\le k\le2m$, $kx_m\in\left(\frac\pi4,\frac\pi2\right]$, so $\sin kx_m\ge\frac1{\sqrt2}$, and $$\sum_{k=m+1}^{2m}\frac{\sin kx_m}{k}\ge\frac1{\sqrt2}\sum_{k=m+1}^{2m}\frac1k\ge\frac1{\sqrt2}\cdot m\cdot\frac1{2m}=\frac1{2\sqrt2}.$$ This block does not tend to $0$ uniformly, so by Theorem 5.5 the series does not converge uniformly on $(0,2\pi)$. The bad points $x_m$ crowd into $0$, where the sum $\frac{\pi-x}2$ jumps.
(c) At $x=0$ every term is $0$, so $S(0)=0$. For $x\ne0$ it is a geometric series with ratio $r=\frac1{1+x^2}\in(0,1)$: $S(x)=\frac{x^2}{1-r}=1+x^2$. The tail after the term $n=N$ is $\sum_{n\gt N}x^2r^n=(1+x^2)r^{N+1}=(1+x^2)^{-N}$. Its supremum over $x\ne0$ is $1$ (let $x\to0$), so the convergence is not uniform on $\mathbb{R}$ (nor on any interval containing $0$). On $\{|x|\ge a\}$, $a\gt0$, the tail is at most $(1+a^2)^{-N}\to0$: uniform. Note that the M-test fails here too: $\|u_n\|\ge u_n\big(\tfrac1{\sqrt{n-1}}\big)\sim\frac1{en}$.
Interactive 2 — Partial Sums, Tails and the M-test
The thick dark curve is the sum $S$ of the series, the red curve is the partial sum $S_N$, and the green background marks a smaller set. The readouts give the worst error $\sup|S-S_N|$ on the whole set and on the green part, and the M-test bound $\sum_{n\gt N}M_n$ when the M-test applies. Press Play to add terms. Which series converge uniformly without the M-test? Where does uniformity break, and does it come back on the green part?
Pointwise convergence on its own never implies uniform convergence. But on a compact set, if both the $f_n$ and the limit are continuous and the convergence is monotone, there is no room for a hump to escape: this is Dini's theorem, a favourite of CSIR NET because each of its four hypotheses has a famous counterexample.
📐 Theorem 5.8 — Dini's theorem
Let $K\subseteq\mathbb{R}$ be compact (closed and bounded), and let $f_n,f:K\to\mathbb{R}$ be continuous. Suppose that for every $x\in K$ the sequence $(f_n(x))$ is monotone and converges to $f(x)$. Then $f_n\rightrightarrows f$ on $K$. Series form. If $u_n\ge0$ are continuous on $K$ and $\sum u_n$ converges pointwise on $K$ to a continuous function, then the convergence is uniform. (The theorem holds verbatim on any compact metric space, §7.2.)
Proof. Put $g_n=|f_n-f|$. Each $g_n$ is continuous, and for each $x$ the sequence $g_n(x)$ is decreasing (a monotone sequence approaches its limit from one side, so its distance to the limit decreases) with $g_n(x)\to0$. Fix $\varepsilon\gt0$ and let $U_n=\{x\in K:g_n(x)\lt\varepsilon\}$.
Each $U_n$ is open in $K$: if $x\in U_n$, continuity of $g_n$ gives $\delta_x\gt0$ with $g_n(y)\lt\varepsilon$ for $y\in K\cap(x-\delta_x,x+\delta_x)$. So $U_n=K\cap V_n$, where $V_n$ is the union of these open intervals. The $U_n$ increase: $g_{n+1}\le g_n$ gives $U_n\subseteq U_{n+1}$. They cover $K$: for each $x$, $g_n(x)\to0$, so $x\in U_n$ for some $n$.
Thus $\{V_n\}$ is an open cover of the compact set $K$. By the Heine–Borel theorem (§3.1) finitely many of them, say $V_{n_1},\dots,V_{n_r}$, cover $K$; with $N=\max n_j$ and the nesting, $K=U_N$. Hence for every $n\ge N$ and $x\in K$, $g_n(x)\le g_N(x)\lt\varepsilon$, i.e. $\|f_n-f\|_K\le\varepsilon$. The series form is the sequence form applied to the partial sums, which increase because $u_n\ge0$. ∎
A second proof, by sequences. If the convergence were not uniform, Theorem 5.3 would give $\varepsilon_0\gt0$ and points $x_k\in K$ with $g_{n_k}(x_k)\ge\varepsilon_0$. By Bolzano–Weierstrass (§2.2) a subsequence of $(x_k)$ converges to some $x^*\in K$ ($K$ is closed). For a fixed $m$ and all $k$ with $n_k\ge m$, monotonicity gives $g_m(x_k)\ge g_{n_k}(x_k)\ge\varepsilon_0$; letting $k\to\infty$ along the subsequence, continuity gives $g_m(x^*)\ge\varepsilon_0$ — for every $m$, contradicting $g_m(x^*)\to0$.
Every hypothesis is needed.Compactness: $x^n\downarrow0$ on $[0,1)$, continuous limit, not uniform. Continuity of $f$: $x^n$ on $[0,1]$. Monotonicity: $nxe^{-nx}\to0$ on $[0,1]$ with continuous terms and limit, but for fixed $x$ the values rise until $n\approx1/x$ and then fall; $M_n=e^{-1}$. Continuity of $f_n$: $f_n=\mathbf 1_{(0,1/n)}$ decreases to $0$ on $[0,1]$, but $M_n=1$.
✏️ Example 6 — Dini at work: a polynomial staircase to $\sqrt x$
(a) Define polynomials by $p_0(x)=0$ and $p_{n+1}(x)=p_n(x)+\tfrac12\big(x-p_n(x)^2\big)$. Prove that $p_n\rightrightarrows\sqrt x$ on $[0,1]$, and deduce that $|t|$ is a uniform limit of polynomials on $[-1,1]$. (b) Prove that $\left(1+\frac xn\right)^n\rightrightarrows e^x$ on $[0,a]$ for every $a\gt0$, but not on $[0,\infty)$.
💡 Show solution▶
(a) Bounds Fix $x\in[0,1]$ and write $s=\sqrt x\in[0,1]$. We prove $0\le p_n(x)\le s$ by induction; it holds for $n=0$. If it holds for $n$, then $$s-p_{n+1}=(s-p_n)-\tfrac12(s-p_n)(s+p_n)=(s-p_n)\Big(1-\frac{s+p_n}2\Big).$$ Both factors are $\ge0$ because $s+p_n\le2s\le2$; so $p_{n+1}\le s$. Also $p_{n+1}-p_n=\tfrac12(x-p_n^2)\ge0$ since $p_n^2\le s^2=x$.
(a) Limit So $(p_n(x))$ is increasing and bounded by $s$; by the monotone convergence theorem it has a limit $L\in[0,s]$. Letting $n\to\infty$ in the recursion, $L=L+\tfrac12(x-L^2)$, so $L^2=x$ and $L=\sqrt x$.
(a) Dini $[0,1]$ is compact, the $p_n$ are polynomials (continuous), the limit $\sqrt x$ is continuous, and the convergence is monotone. By Theorem 5.8, $p_n\rightrightarrows\sqrt x$. The first few are $p_1=\tfrac x2$, $p_2=x-\tfrac{x^2}8$, $p_3=\tfrac{3x}2-\tfrac{5x^2}8+\tfrac{x^3}8-\tfrac{x^4}{128}$. Numerically $\|\sqrt x-p_{10}\|\approx0.0547$ (attained near $x=0$, where $\sqrt x$ is steep). Finally, $q_n(t)=p_n(t^2)$ are polynomials with $\sup_{|t|\le1}\big||t|-q_n(t)\big|=\sup_{0\le x\le1}|\sqrt x-p_n(x)|\to0$. This is the seed of the Weierstrass approximation theorem (§5.2).
(b) For $x\ge0$, AM–GM applied to the $n+1$ numbers $1+\frac xn$ ($n$ times) and $1$ gives $\left(1+\frac xn\right)^n\le\left(\frac{n(1+x/n)+1}{n+1}\right)^{n+1}=\left(1+\frac x{n+1}\right)^{n+1}$: the sequence increases, and its limit $e^x$ is continuous. On the compact set $[0,a]$ Dini's theorem gives uniform convergence. On $[0,\infty)$ take $x=n$: $e^n-2^n\to\infty$, so $\sup_{x\ge0}\big|e^x-(1+\frac xn)^n\big|=\infty$ for every $n$. Here compactness is exactly what is lost.
⚠️ Counterexamples and pitfalls
Pointwise limits of continuous functions can be discontinuous: $x^n$ on $[0,1]$; $\frac{nx}{1+nx}$ on $[0,1]$ (limit $1$ for $x\gt0$, $0$ at $0$).
The sliding hump: $nxe^{-nx}\to0$ pointwise on $[0,\infty)$ with $M_n=e^{-1}$; also $\frac{nx}{1+n^2x^2}$ with $M_n=\frac12$ at $x=\frac1n$. Every fixed $x$ is eventually passed, but the hump never shrinks.
The moving bump: $\frac1{1+(x-n)^2}\to0$ pointwise on $\mathbb{R}$, $M_n=1$; it converges uniformly on every $(-\infty,a]$.
The growing spike: $n^2x(1-x^2)^n\to0$ pointwise on $[0,1]$ but $\int_0^1=\frac{n^2}{2(n+1)}\to\infty$.
Uniform convergence does not carry derivatives: $\frac{x}{1+nx^2}\rightrightarrows0$ on $\mathbb{R}$ but $f_n'(0)=1$; $\frac{\sin nx}{\sqrt n}\rightrightarrows0$ but $f_n'(0)=\sqrt n\to\infty$.
Uniform on each piece of an infinite family is not uniform on the union: $x^n$ on each $[0,1-\tfrac1k]$ but not on $[0,1)$; finite unions are fine.
Terms tending to $0$ uniformly do not make a series uniformly convergent: $\sum x^n(1-x)$ on $[0,1]$ has $\|u_n\|\to0$ and tails of sup $1$.
The M-test is sufficient, not necessary: $\sum\frac{(-1)^{n+1}}{n+x^2}$ converges uniformly on $\mathbb{R}$ and absolutely nowhere.
"Absolutely and uniformly" is not "absolutely uniformly": $\sum(-1)^nx^n(1-x)$ converges uniformly on $[0,1]$ (alternating tails $\le\frac1{N+2}$) and absolutely at each point, yet $\sum x^n(1-x)$ does not converge uniformly.
Uniform convergence needs finite sups: $\sum\frac{x^2}{n^2}$ converges at every $x$ but its tails $x^2\sum_{n\gt N}\frac1{n^2}$ are unbounded on $\mathbb{R}$.
Products and compositions: $x+\frac1n\rightrightarrows x$ on $\mathbb{R}$ but $(x+\frac1n)^2\not\rightrightarrows x^2$; it works for bounded limits or uniformly continuous outer functions.
Pointwise limits of bounded functions can be unbounded: $\min(n,\tfrac1x)$ on $(0,1]$ $\to\tfrac1x$.
Dini needs all four hypotheses: compact domain ($x^n$ on $[0,1)$), continuous limit ($x^n$ on $[0,1]$), monotonicity ($nxe^{-nx}$ on $[0,1]$), continuous terms ($\mathbf 1_{(0,1/n)}$).
Exercises
I
Pointwise limits and sup norms. In each part find the pointwise limit, decide whether the convergence is uniform on the given set, and when it is not, find a natural subset on which it is.
1 $f_n(x)=\dfrac{x^n}{1+x^n}$ on $[0,\infty)$.
2 $f_n(x)=\dfrac{nx}{1+n^2x^2}$ on $[0,1]$.
3 $f_n(x)=x^n(1-x)$ on $[0,1]$.
4 $f_n(x)=nx^n(1-x)$ on $[0,1]$.
5 $f_n(x)=\dfrac{\sin nx}{\sqrt n}$ on $\mathbb{R}$.
6 $f_n(x)=\dfrac xn$ on $\mathbb{R}$.
7 $f_n(x)=\sqrt{x^2+\tfrac1{n^2}}$ on $\mathbb{R}$.
8 $f_n(x)=n\big(x^{1/n}-1\big)$ on $[1,2]$ and on $[1,\infty)$.
9 $f_n(x)=\dfrac1{1+(x-n)^2}$ on $\mathbb{R}$.
10 $f_n(x)=\dfrac{x^2}{x^2+(1-nx)^2}$ on $[0,1]$.
💡 Show solutions▶
1 Limit $0$ on $[0,1)$, $\tfrac12$ at $1$, $1$ on $(1,\infty)$. Not uniform: $x_n=2^{-1/n}$ gives $f_n(x_n)=\frac{1/2}{3/2}=\frac13$ while $f(x_n)=0$. Uniform on $[0,a]$, $a\lt1$ ($|f_n|\le a^n$) and on $[b,\infty)$, $b\gt1$ ($|f_n-1|=\frac1{1+x^n}\le\frac1{1+b^n}$).
2 Limit $0$ ($f_n(0)=0$; $\le\frac1{nx}$ for $x\gt0$). $f_n(\frac1n)=\frac12$, so $M_n=\frac12$ (AM–GM: $1+n^2x^2\ge2nx$): not uniform. On $[a,1]$: $f_n\le\frac{1}{na}\to0$, uniform.
3 Limit $0$; the maximum is at $x=\frac n{n+1}$, $M_n=\frac{n^n}{(n+1)^{n+1}}\lt\frac1{n+1}\to0$: uniform on $[0,1]$.
4 Limit $0$ ($nx^n\to0$ for $0\le x\lt1$; value $0$ at $1$). $M_n=\frac{n^{n+1}}{(n+1)^{n+1}}=\left(1+\frac1n\right)^{-(n+1)}\to e^{-1}$: not uniform. Uniform on $[0,a]$, $a\lt1$ ($\le na^n\to0$).
5 $|f_n|\le\frac1{\sqrt n}$: uniform to $0$ on $\mathbb{R}$.
6 Limit $0$; $\sup_{\mathbb{R}}|x/n|=\infty$: not uniform. On $[-A,A]$, $M_n=\frac An\to0$: uniform on every bounded set.
7 Limit $|x|$; $0\le f_n(x)-|x|=\dfrac{1/n^2}{\sqrt{x^2+1/n^2}+|x|}\le\dfrac{1/n^2}{1/n}=\dfrac1n$, with equality at $x=0$: uniform on $\mathbb{R}$, $M_n=\frac1n$. (Smooth functions converging uniformly to a function with a corner.)
8 With $t=\log x$: $n(e^{t/n}-1)-t=\sum_{k\ge2}\frac{t^k}{k!\,n^{k-1}}$, which lies between $0$ and $\frac{t^2}{n}e^{t}$. On $[1,2]$, $0\le t\le\log2$, so $M_n\le\frac{2(\log2)^2}{n}\to0$: uniform to $\log x$. On $[1,\infty)$ take $x=e^n$: $n(e-1)-n=n(e-2)\to\infty$: not uniform.
9 Limit $0$; $f_n(n)=1$, so $M_n=1$: not uniform on $\mathbb{R}$. On $(-\infty,a]$, for $n\gt a$: $M_n=\frac1{1+(n-a)^2}\to0$, uniform.
10 $f_n(0)=0$; for $x\gt0$, $(1-nx)^2\sim n^2x^2\to\infty$, so $f_n(x)\to0$. But $f_n(\tfrac1n)=1$: $M_n=1$, not uniform. On $[a,1]$ with $n\ge2/a$: $nx-1\ge\frac{nx}2$, so $f_n\le\frac{x^2}{n^2x^2/4}=\frac4{n^2}$, uniform.
II
Series of functions: the Cauchy criterion, the M-test, Dirichlet and Abel.
1 Prove that $\sum\dfrac{\cos nx}{n^{3/2}}$ converges uniformly on $\mathbb{R}$, and bound its tail after $N$ terms.
2 Show that $\sum\dfrac{x}{1+n^2x^2}$ converges uniformly on $[a,\infty)$ for every $a\gt0$ but not on $(0,\infty)$.
3 Show that $\sum\dfrac{x^n}{n^2}$ converges uniformly on $[-1,1]$ and that $\sum\dfrac{x^n}{n}$ converges uniformly on $[-1,0]$.
4 Show that $\sum n^{-x}$ converges uniformly on $[a,\infty)$ for $a\gt1$ but not on $(1,\infty)$.
5 Show that $\sum\dfrac{(-1)^n}{n}e^{-nx}$ converges uniformly on $[0,\infty)$, although it does not converge absolutely at $x=0$.
6 Prove that $\sum_{n\ge0}x^n(1-x)$ does not converge uniformly on $[0,1]$, whereas its terms tend to $0$ uniformly. Is the convergence uniform on $[0,a]$, $a\lt1$?
7 Let $0\lt\delta\lt\pi$. Prove that $\sum\dfrac{\cos nx}{n}$ converges uniformly on $[\delta,2\pi-\delta]$.
8 Show that $\sum\dfrac{x}{n(1+nx^2)}$ converges uniformly on $\mathbb{R}$ and that $\sum\dfrac{x}{1+nx^2}$ does not converge at any $x\ne0$.
9 Use Abel's test to show that $\sum\dfrac{(-1)^n}{n}\cdot\dfrac{x^n}{1+x^n}$ converges uniformly on $[0,\infty)$.
💡 Show solutions▶
1 M-test with $M_n=n^{-3/2}$. Tail: $\|R_N\|\le\sum_{n\gt N}n^{-3/2}\lt\int_N^\infty t^{-3/2}dt=\frac2{\sqrt N}$.
2 For $x\ge a$: $\frac{x}{1+n^2x^2}\le\frac{x}{n^2x^2}=\frac1{n^2x}\le\frac1{n^2a}$; M-test. On $(0,\infty)$, at $x=\frac1N$: $\sum_{n=N+1}^{2N}\frac{1/N}{1+n^2/N^2}\ge N\cdot\frac{1/N}{1+4}=\frac15$, so the Cauchy blocks do not tend to $0$ uniformly (Theorem 5.5).
3 $|x^n/n^2|\le1/n^2$ on $[-1,1]$: M-test. On $[-1,0]$ write $x=-y$, $0\le y\le1$: $\sum(-1)^n\frac{y^n}n$ alternates with $\frac{y^n}n$ decreasing to $0$ in $n$, so $|R_N|\le\frac{y^{N+1}}{N+1}\le\frac1{N+1}$ (Leibniz), uniformly.
4 On $[a,\infty)$, $n^{-x}\le n^{-a}$ and $\sum n^{-a}\lt\infty$. On $(1,\infty)$: $\sum_{n=N+1}^{2N}n^{-x}\ge N(2N)^{-x}\to\frac12$ as $x\to1^+$, so for each $N$ there is $x\gt1$ with block $\gt\frac14$: the Cauchy criterion fails.
5 For fixed $x\ge0$, $\frac{e^{-nx}}n$ decreases to $0$ in $n$; Leibniz gives $|R_N(x)|\le\frac{e^{-(N+1)x}}{N+1}\le\frac1{N+1}$. At $x=0$ the absolute series is $\sum\frac1n=\infty$.
6 As in Example 4(e): $S_N=1-x^{N+1}$, the tail on $[0,1)$ is $x^{N+1}$ with supremum $1$. $\|u_n\|=\frac{n^n}{(n+1)^{n+1}}\to0$. On $[0,a]$ the tail is $\le a^{N+1}\to0$: uniform.
7 Dirichlet (Theorem 5.7): $|\sum_{k\le n}\cos kx|\le\frac1{\sin(\delta/2)}$ on $[\delta,2\pi-\delta]$, and $b_n=\frac1n\downarrow0$ uniformly.
8 $\big|\frac{x}{1+nx^2}\big|\le\frac1{2\sqrt n}$ (Example 2(c)), so the terms of the first series are $\le\frac1{2n^{3/2}}$: M-test. For the second, with $x\ne0$ fixed, $\frac{x}{1+nx^2}\sim\frac1{nx}$, and limit comparison with $\sum\frac1n$ gives divergence.
9 $\sum\frac{(-1)^n}n$ converges, and since its terms do not depend on $x$ it converges uniformly. $b_n(x)=\frac{x^n}{1+x^n}=1-\frac1{1+x^n}$ is monotone in $n$ for each $x\ge0$ (increasing if $x\gt1$, decreasing if $x\lt1$, constant if $x=1$) and $0\le b_n\le1$. Theorem 5.7(2) applies.
III
Dini's theorem and the algebra of uniform convergence.
1 Let $f_n(x)=x^{1/n}$ on $[0,1]$. Find the pointwise limit, show the convergence is monotone but not uniform, and name the hypothesis of Dini's theorem that fails. Is it uniform on $[a,1]$, $0\lt a\lt1$?
2 Show that $p_n(x)$ of Example 6 satisfies $0\le\sqrt x-p_n(x)\le\dfrac{2\sqrt x}{2+n\sqrt x}\le\dfrac2n$ on $[0,1]$.
3 Suppose $f_n\rightrightarrows f$ on $D$ and $|f(x)|\ge c\gt0$ for all $x\in D$. Prove that $\frac1{f_n}\rightrightarrows\frac1f$ on $D$ (for $n$ large enough that $f_n\ne0$).
4 Prove that if $f_n\rightrightarrows f$ and $g_n\rightrightarrows g$ on $D$, then $\max(f_n,g_n)\rightrightarrows\max(f,g)$ and $|f_n|\rightrightarrows|f|$.
5(Pólya.) Let each $f_n:[a,b]\to\mathbb{R}$ be an increasing function of $x$ (not necessarily continuous), and suppose $f_n\to f$ pointwise with $f$ continuous. Prove that $f_n\rightrightarrows f$ on $[a,b]$.
6 Show that $f_n(x)=\dfrac{nx}{1+nx}$ does not converge uniformly on $[0,1]$, using Theorem 5.3, but converges uniformly on $[a,1]$ for $a\gt0$.
7 Prove that on a finite set $D$ pointwise convergence implies uniform convergence. Give an example on $D=\mathbb{N}$ where it does not.
8 Let $u_n\ge0$ be continuous on $[0,1]$ with $\sum u_n$ pointwise convergent to a function $S$. Show by example that $S$ need not be continuous, and explain why Dini's theorem forces uniform convergence when $S$ is continuous.
💡 Show solutions▶
1 $f_n(0)=0$, and for $x\gt0$, $x^{1/n}=e^{(\log x)/n}\to1$. So $f=\mathbf 1_{(0,1]}$, discontinuous. For $0\lt x\lt1$ the values $x^{1/n}$ increase in $n$. $\sup_{(0,1]}|1-x^{1/n}|=1$ (let $x\to0^+$): not uniform. Dini fails because the limit is discontinuous. On $[a,1]$: $M_n=1-a^{1/n}\to0$, uniform (or Dini with $K=[a,1]$).
2 With $s=\sqrt x$ and $e_n=s-p_n$, the proof of Example 6 gives $e_{n+1}=e_n\left(1-s+\tfrac{e_n}2\right)$; the map $e\mapsto e(1-s+\frac e2)$ is increasing for $e\ge0$. Induction on $e_n\le\frac{2s}{2+ns}=:c_n$: it holds at $n=0$, and $e_{n+1}\le c_n(1-s+\frac{c_n}2)\le c_{n+1}$; writing $a=2+ns\ge2$, the last inequality reduces to $s\big((a-1)^2-1+s(a-1)\big)\ge0$, which holds. Finally $\frac{2s}{2+ns}\le\frac{2s}{ns}=\frac2n$.
3 Choose $N$ with $\|f_n-f\|\lt\frac c2$ for $n\ge N$; then $|f_n|\ge\frac c2$, and $\left|\frac1{f_n}-\frac1f\right|=\frac{|f-f_n|}{|f_n||f|}\le\frac{2}{c^2}\|f_n-f\|\to0$.
4 $|\max(a,b)-\max(c,d)|\le\max(|a-c|,|b-d|)$ (check the four cases, or use $\max(a,b)=\frac{a+b+|a-b|}2$). So $\|\max(f_n,g_n)-\max(f,g)\|\le\max(\|f_n-f\|,\|g_n-g\|)\to0$. Also $\big||f_n|-|f|\big|\le|f_n-f|$.
5 $f$ is increasing (a pointwise limit of increasing functions) and continuous on $[a,b]$, hence uniformly continuous: choose $a=x_0\lt\cdots\lt x_k=b$ with $f(x_i)-f(x_{i-1})\lt\varepsilon$. Choose $N$ with $|f_n(x_i)-f(x_i)|\lt\varepsilon$ for all $i$ and $n\ge N$ (finitely many points). For $x\in[x_{i-1},x_i]$: $f_n(x)\le f_n(x_i)\lt f(x_i)+\varepsilon\le f(x)+2\varepsilon$ and $f_n(x)\ge f_n(x_{i-1})\gt f(x_{i-1})-\varepsilon\ge f(x)-2\varepsilon$. So $\|f_n-f\|\le2\varepsilon$.
6 Limit $1$ for $x\gt0$ and $0$ at $0$. With $x_n=\frac1n$: $f_n(x_n)-f(x_n)=\frac12-1\not\to0$, so not uniform. On $[a,1]$: $|1-f_n|=\frac1{1+nx}\le\frac1{1+na}\to0$.
7 If $D=\{x_1,\dots,x_k\}$, choose $N_i$ for each point and $N=\max N_i$. On $D=\mathbb{N}$, $f_n=\mathbf 1_{\{n\}}$ (value $1$ at $n$, $0$ elsewhere) $\to0$ pointwise, yet $\|f_n\|=1$.
8 $u_n(x)=x^{n}(1-x)\ge0$ on $[0,1]$ ($n\ge0$) has sum $\mathbf 1_{[0,1)}$, discontinuous. If $S$ is continuous, the partial sums are continuous and increase (as $u_n\ge0$) to the continuous $S$ on the compact $[0,1]$: Theorem 5.8 gives uniform convergence.
IV
Prove or disprove, and two applications.
1 A uniform limit of uniformly continuous functions on $D$ is uniformly continuous.
2 If every $f_n$ is bounded on $(0,1]$ and $f_n\to f$ pointwise, then $f$ is bounded.
3 If $\sum|u_n|$ converges uniformly on $D$, then $\sum u_n$ converges uniformly on $D$.
4 If $\sum u_n$ converges uniformly on $D$ and absolutely at every point of $D$, then $\sum|u_n|$ converges uniformly on $D$.
5 If polynomials $p_n$ converge uniformly on $\mathbb{R}$ to $f$, then $f$ is a polynomial.
6Applied (numerical approximation). An engineer replaces $e^{-x}$ on $[0,5]$ by its Taylor polynomial $T_N(x)=\sum_{k=0}^N\frac{(-x)^k}{k!}$. Use the Lagrange remainder to find the smallest $N$ that the remainder bound certifies for $\|e^{-x}-T_N\|_{[0,5]}\lt10^{-6}$.
7Applied (heat flow). The temperature in a rod is modelled by $u(x,t)=\sum_{n\ge1}b_ne^{-n^2t}\sin nx$ with $|b_n|\le1$. Show that the series converges uniformly on $\mathbb{R}\times[t_0,\infty)$ for every $t_0\gt0$, and find how many terms guarantee a truncation error below $10^{-4}$ for all $x$ and all $t\ge0.1$.
💡 Show solutions▶
1True Given $\varepsilon$, choose $n$ with $\|f_n-f\|\lt\frac\varepsilon3$, then $\delta$ with $|f_n(x)-f_n(y)|\lt\frac\varepsilon3$ whenever $|x-y|\lt\delta$. Then $|f(x)-f(y)|\le|f(x)-f_n(x)|+|f_n(x)-f_n(y)|+|f_n(y)-f(y)|\lt\varepsilon$.
2False $f_n(x)=\min(n,\frac1x)$ is bounded by $n$, and $f_n(x)=\frac1x$ once $n\ge\frac1x$; the limit $\frac1x$ is unbounded on $(0,1]$.
3True $\left|\sum_{k=m+1}^nu_k(x)\right|\le\sum_{k=m+1}^n|u_k(x)|$; the uniform Cauchy criterion (Theorem 5.5) for $\sum|u_n|$ gives it for $\sum u_n$.
4False $u_n(x)=(-1)^nx^n(1-x)$ on $[0,1]$: uniformly convergent (alternating, tail $\le\max x^{N+1}(1-x)\lt\frac1{N+2}$) and absolutely convergent at each $x$, but $\sum|u_n|=\sum x^n(1-x)$ is not uniformly convergent (Example 4(e)).
5True By the Cauchy criterion there is $N$ with $\|p_n-p_N\|_{\mathbb{R}}\le1$ for $n\ge N$. A polynomial bounded on $\mathbb{R}$ is constant, so $p_n=p_N+c_n$ with constants $c_n$. Then $c_n=p_n(0)-p_N(0)$ converges to some $c$, and $f=p_N+c$, a polynomial. (So the Weierstrass theorem of §5.2 genuinely needs a compact interval.)
6 $|e^{-x}-T_N(x)|=\frac{e^{-\xi}x^{N+1}}{(N+1)!}\le\frac{5^{N+1}}{(N+1)!}$ on $[0,5]$. This is $2.12\times10^{-6}$ for $N=21$ and $4.61\times10^{-7}$ for $N=22$: $N=22$. (The true error at $N=21$ is $1.74\times10^{-6}$, so here the bound is sharp enough to give the true minimum.)
7 For $t\ge t_0$: $|b_ne^{-n^2t}\sin nx|\le e^{-n^2t_0}$, and $\sum e^{-n^2t_0}\lt\infty$ (compare with $\sum e^{-nt_0}$): M-test on $\mathbb{R}\times[t_0,\infty)$. With $t_0=0.1$: $\sum_{n\gt8}e^{-0.1n^2}\approx3.6\times10^{-4}$ but $\sum_{n\gt9}e^{-0.1n^2}\approx5.2\times10^{-5}$, so $N=9$ terms suffice. (At $t_0=0$ the M-test fails: near $t=0$ the series is a general Fourier series.)
🎯 Exam corner CSIR Part BCSIR Part CGATE MA
E1.CSIR Part BGATE MA Let $f_n(x)=x^n(1-x^n)$ for $x\in[0,1]$. Then (A) $f_n\to0$ uniformly on $[0,1]$; (B) $f_n\to0$ pointwise on $[0,1]$, but not uniformly; (C) $(f_n(1))$ does not converge; (D) $f_n$ converges uniformly to a discontinuous function.
💡 Show solution▶
Idea Substitute $t=x^n\in[0,1]$: $f_n=t(1-t)$, whose maximum is $\frac14$ at $t=\frac12$.
Work Pointwise: $x^n\to0$ for $x\lt1$ and $f_n(1)=0$, so the limit is $0$ (this rules out (C), (D)). At $x_n=2^{-1/n}$, $f_n(x_n)=\frac14$ for every $n$: $M_n=\frac14\not\to0$, ruling out (A). (On $[0,a]$, $a\lt1$, it is uniform.)
Answer (B).
E2.CSIR Part B Let $M_n=\sup_{x\in\mathbb{R}}\big|xe^{-nx^2}\big|$. Then $\lim_{n\to\infty}\sqrt n\,M_n$ equals (A) $0$; (B) $\dfrac1{\sqrt{2e}}$; (C) $\dfrac1{\sqrt e}$; (D) $\infty$.
💡 Show solution▶
Idea Maximise by calculus.
Work $\frac{d}{dx}xe^{-nx^2}=e^{-nx^2}(1-2nx^2)=0$ at $x=\frac1{\sqrt{2n}}$, giving $M_n=\frac{e^{-1/2}}{\sqrt{2n}}$. So $\sqrt nM_n=\frac1{\sqrt{2e}}\approx0.4289$ for every $n$. In particular $M_n\to0$: $xe^{-nx^2}\rightrightarrows0$ on $\mathbb{R}$, at rate $n^{-1/2}$.
Answer (B).
E3.CSIR Part B Let $f_n:[0,1]\to\mathbb{R}$ be continuous with $f_n(x)\ge f_{n+1}(x)\ge0$ for all $n$ and $x$, and $f_n(x)\to0$ for every $x$. Which statement is necessarily true? (A) $\sup_{[0,1]}f_n\to0$; (B) $f_n\to0$ uniformly on $[a,1]$ for each $a\gt0$, but some such sequence does not converge uniformly on $[0,1]$; (C) $\int_0^1f_n$ need not tend to $0$; (D) the conclusion $\sup f_n\to0$ requires each $f_n$ to be differentiable.
Work Theorem 5.8 gives $\|f_n\|_{[0,1]}\to0$ for every such sequence, so (A) holds and (B) is wrong; (D) is wrong because Dini needs only continuity. (C) is wrong because $0\le\int_0^1f_n\le\|f_n\|\to0$.
Answer (A).
E4.CSIR Part C Which of the following sequences converge uniformly on $[0,\infty)$? (One or more options may be correct.) (A) $\dfrac{x}{n+x}$; (B) $xe^{-nx}$; (C) $\dfrac{nx^2}{1+nx}$; (D) $\dfrac1{1+(x-n)^2}$.
💡 Show solution▶
(A) No Pointwise limit $0$, but $\frac{x}{n+x}\to1$ as $x\to\infty$ for each $n$: $M_n=1$.
(B) Yes Maximum at $x=\frac1n$: $M_n=\frac1{ne}\to0$.
(C) Yes The limit is $x$ (for $x\gt0$, $\frac{nx^2}{1+nx}\to x$; both are $0$ at $0$), and $\left|\frac{nx^2}{1+nx}-x\right|=\frac{x}{1+nx}\lt\frac1n$. A limit that is unbounded is no obstacle.
(D) No The moving bump: limit $0$, value $1$ at $x=n$.
Answer (B) and (C).
E5.CSIR Part C Which of the following series converge uniformly on $\mathbb{R}$? (A) $\sum\dfrac{\cos nx}{n^2+x^2}$; (B) $\sum\dfrac{(-1)^n}{n+x^2}$; (C) $\sum\dfrac1{1+n^2x^2}$; (D) $\sum\dfrac{x^2}{n^2}$.
(C) No At $x=0$ every term is $1$: the series diverges there, so it does not even converge pointwise on $\mathbb{R}$.
(D) No It converges to $\frac{\pi^2x^2}6$ at each $x$, but the tail $x^2\sum_{n\gt N}\frac1{n^2}$ is unbounded on $\mathbb{R}$. (It is uniform on bounded sets.)
Answer (A) and (B).
E6.CSIR Part C Let $f_n,g_n,f,g:(0,1]\to\mathbb{R}$. Which of the following are true? (A) If $f_n\rightrightarrows f$ and every $f_n$ is bounded, then $f$ is bounded. (B) If $f_n\rightrightarrows f$ and $g_n\rightrightarrows g$, then $f_ng_n\rightrightarrows fg$. (C) If $f_n\rightrightarrows f$ on $(0,\frac12]$ and on $[\frac12,1]$, then $f_n\rightrightarrows f$ on $(0,1]$. (D) If $f_n\rightrightarrows f$ on $[\frac1k,1]$ for every $k\in\mathbb{N}$, then $f_n\rightrightarrows f$ on $(0,1]$.
💡 Show solution▶
(A) True Example 3(a): $|f|\le|f-f_N|+|f_N|\le1+\sup|f_N|$.
(B) False $f_n=g_n=\frac1x+\frac1n$, $f=g=\frac1x$: $f_ng_n-fg=\frac2{nx}+\frac1{n^2}$ is unbounded on $(0,1]$ for every $n$.
(C) True Finite unions (Theorem 5.2(2)): the sup over the union is the larger of the two sups.
(D) False $f_n(x)=(1-x)^n$, $f=0$: on $[\frac1k,1]$, $M_n=(1-\frac1k)^n\to0$, but on $(0,1]$, $M_n=\sup_{x\gt0}(1-x)^n=1$.
Answer (A) and (C).
🧑🏫 For the teacher
Suggested plan — 2 lectures × 60 min.
Pointwise versus uniform. Open with Example 1 on the board — ask the class to predict each answer before computing. Definitions 5.1 and 5.3 side by side; underline the moved quantifier. Theorem 5.1 (sup norm), Theorem 5.3 (bad points) and the strategy box. Example 2 in full, then project Interactive 1: $x^n$ on $[0,1]$ versus $[0,0.9]$, then the sliding hump $nxe^{-nx}$ on $[0,4]$ versus $[1,4]$, then $\sqrt{x^2+1/n^2}$ (uniform, corner in the limit). Finish with Example 3(b) as a quick poll.
Series and Dini. Theorem 5.4 with its proof (the "let $m\to\infty$ with $x$ fixed" step deserves time), Theorem 5.5 and the M-test; Example 4 — ask which $M_n$ to take in (c) and (d). Dirichlet/Abel (Theorem 5.7) and Example 5; Interactive 2 with series 1, 3 and 5 to show uniform convergence without the M-test, series 2 and 4 to show the green part rescuing uniformity. Dini's theorem with the open-cover proof and its four counterexamples; Example 6(a) as the bridge to Weierstrass. Close with E4 or E6.
Discussion questions. (1) Uniform convergence on $D$ means $\|f_n-f\|_D\to0$. What is the analogous "norm" statement for pointwise convergence — is there one? (2) Why does the M-test prove absolute and uniform convergence, while Dirichlet's test gives only uniform convergence? (3) In Dini's theorem, where exactly does the proof use that the limit $f$ is continuous?
Misconceptions to watch for. Computing $\lim_n f_n(x)$ and declaring the convergence uniform because the limit "looks nice"; bounding $|f_n(x)-f(x)|$ by something that still depends on $x$; believing that $\|u_n\|\to0$ suffices for a series; thinking that failure of the M-test proves non-uniform convergence; forgetting the isolated point $x=1$ in $x^n$; applying Dini on a non-compact interval.
🔬 Research–Teaching NexusResearch-tutored · optional, about 45 minutes
Research question. A Fourier series of a function with a jump converges at every point but not uniformly. Exactly how badly does it fail — and does the failure disappear as more terms are added?
Why it matters. Magnetic resonance imaging reconstructs a picture from finitely many Fourier coefficients. Near a sharp edge — the boundary between the skull and the brain, or between cartilage and fluid in a knee scan — truncating the series produces bright and dark "ringing" bands parallel to the edge, the Gibbs artefact, which radiologists must not mistake for anatomy. Audio engineers meet the same overshoot when a square wave is band-limited.
What to do.
For $S_N(x)=\frac4\pi\sum_{k=1}^N\frac{\sin(2k-1)x}{2k-1}$ (the square wave equal to $1$ on $(0,\pi)$ and $-1$ on $(-\pi,0)$), compute $\max_{0\lt x\lt\pi}S_N(x)$ for $N=10,100,1000$. Does the overshoot above $1$ tend to $0$?
Express the overshoot as a percentage of the jump (which is $2$). Compare with $\frac{2}{\pi}\int_0^\pi\frac{\sin t}{t}\,dt-1$.
Fix $\delta=0.2$ and compute $\sup_{\delta\le x\le\pi-\delta}|S_N(x)-1|$ for the same $N$. How does it scale with $N$? Relate your answer to Dirichlet's test (Theorem 5.7) and Example 5(b).
Smooth the jump: average the partial sums (Fejér means $\sigma_N=\frac1N\sum_{n\lt N}S_n$). Repeat step 1. What happened to the overshoot?
Tools: Python with NumPy, or a spreadsheet with a fine grid of $x$ values.
Hand in: A table of overshoots and interior errors, a one-paragraph explanation in terms of uniform versus pointwise convergence, and one sentence on how MRI software suppresses ringing (windowing the coefficients).
Going further. Read about Fejér's theorem: the Cesàro means of the Fourier series of a continuous periodic function converge uniformly — a second route to the Weierstrass approximation theorem of §5.2.
Instructor note: the maxima are about $1.17981$ ($N=10$), $1.17899$ ($N=100$), $1.17898$ ($N=1000$) — the overshoot tends to $\frac2\pi\operatorname{Si}(\pi)-1\approx0.1790$, i.e. about $8.95\%$ of the jump, and never to $0$; it only moves closer to the jump (its location is about $\frac{\pi}{2N}$). The interior error on $[\delta,\pi-\delta]$ decays like $1/N$ (Dirichlet's test). Fejér means are positive averages and do not overshoot at all.
🧪 Self-check quiz
0/5 correct
Q1. On $[0,1)$ the sequence $f_n(x)=x^n$
Q2. $\sup_{x\ge0}\dfrac{nx}{1+n^2x^2}$ equals
Q3. If $|u_n(x)|\le M_n$ on $D$ and $\sum M_n\lt\infty$, then $\sum u_n$
Q4. $f_n(x)=x^n$ decreases to the continuous function $0$ on $[0,1)$, yet not uniformly. Which hypothesis of Dini's theorem fails?
Q5. If $\sum u_n$ converges uniformly on $D$, which of the following must hold?