5.2 Uniform Convergence, Power Series and Approximation
What uniform convergence buys: continuous limits, integrals and derivatives term by term, power series that may be differentiated freely inside their radius, Abel's theorem at the edge — and the Weierstrass theorem that every continuous function on [a, b] is a uniform limit of polynomials.
Section 5.1 taught us to recognise uniform convergence. This section collects the dividends. Uniform limits of continuous functions are continuous; integrals of uniformly convergent sequences converge; derivatives can be taken term by term once the derivatives converge uniformly. Applied to power series these three facts produce the whole calculus of Taylor series — radius of convergence, term-by-term differentiation and integration, uniqueness of coefficients — and, at the edge of the interval of convergence, Abel's theorem, which explains why $1-\tfrac12+\tfrac13-\cdots=\log2$ and $1-\tfrac13+\tfrac15-\cdots=\tfrac\pi4$.
The section closes with one of the great theorems of analysis: Weierstrass's theorem that every continuous function on a closed bounded interval is a uniform limit of polynomials, proved here with Bernstein's explicit polynomials. In CSIR NET this material appears as "find the radius of convergence", "evaluate $\lim_n\int f_n$", "is the limit function continuous / differentiable", "which of these interchanges of limit and integral are valid" and, in Part C, as statements about polynomial approximation and moments — every one of which is decided by checking the hypotheses of the theorems below.
Continuity, Limits and Integrals under Uniform Convergence
Proof (the "$\varepsilon/3$ argument"). Let $\varepsilon\gt0$. Choose $n$ with $\|f_n-f\|_D\lt\varepsilon/3$. Since $f_n$ is continuous at $x_0$, there is $\delta\gt0$ with $|f_n(x)-f_n(x_0)|\lt\varepsilon/3$ for $x\in D$, $|x-x_0|\lt\delta$. For such $x$,
$$\begin{aligned}|f(x)-f(x_0)|&\le|f(x)-f_n(x)|\\&\quad+|f_n(x)-f_n(x_0)|\\&\quad+|f_n(x_0)-f(x_0)|\lt\varepsilon .\end{aligned}$$The first and last terms are small because the convergence is uniform: one $n$ serves both $x$ and $x_0$. ∎
The theorem is mostly used backwards: if continuous $f_n$ converge pointwise to a discontinuous $f$, the convergence is not uniform (nor on any set $E$ with the point of discontinuity $x_0$ in its closure, provided $f$ restricted to $E\cup\{x_0\}$ is still discontinuous at $x_0$ — e.g. $x^n$ on $[0,1)$; see Exercise I.6). This settles $x^n$ on $[0,1]$, $x^{1/n}$ on $[0,1]$, $\sum x^n(1-x)$ and $\sum\frac{x^2}{(1+x^2)^n}$ at once. The proof uses only the values of $f$ near $x_0$, so it also gives the more general interchange of limits.
Proof. For all $x\in D$, $|f_n(x)-f_m(x)|\le\|f_n-f_m\|_D$; letting $x\to c$ gives $|L_n-L_m|\le\|f_n-f_m\|_D$. The right side tends to $0$ (Theorem 5.4), so $(L_n)$ is Cauchy, with a limit $L$. Given $\varepsilon\gt0$, fix $n$ with $\|f_n-f\|_D\lt\varepsilon/3$ and $|L_n-L|\lt\varepsilon/3$, and then choose a neighbourhood of $c$ on which $|f_n(x)-L_n|\lt\varepsilon/3$. There $$\begin{aligned}|f(x)-L|&\le|f(x)-f_n(x)|\\&\quad+|f_n(x)-L_n|+|L_n-L|\lt\varepsilon .\end{aligned}$$ So $\lim_{x\to c}f(x)=L$. ∎
Proof. Put $\varepsilon_n=\|f_n-f\|$, so $f_n-\varepsilon_n\le f\le f_n+\varepsilon_n$; in particular $f$ is bounded. For every partition $P$ of $[a,b]$ the upper and lower Darboux sums (§4.1) therefore satisfy $U(f,P)\le U(f_n,P)+\varepsilon_n(b-a)$ and $L(f,P)\ge L(f_n,P)-\varepsilon_n(b-a)$, so $$\begin{aligned}U(f,P)-L(f,P)&\le U(f_n,P)-L(f_n,P)\\&\quad+2\varepsilon_n(b-a).\end{aligned}$$ Given $\varepsilon\gt0$, choose $n$ with $2\varepsilon_n(b-a)\lt\varepsilon/2$ and then (Riemann's criterion for $f_n$) a partition with $U(f_n,P)-L(f_n,P)\lt\varepsilon/2$. Riemann's criterion now shows that $f$ is integrable. Finally, for every $x\in[a,b]$, $$\begin{aligned}\Big|\int_a^xf_n-\int_a^xf\Big|&\le\int_a^x|f_n-f|\\&\le\varepsilon_n(b-a),\end{aligned}$$ a bound independent of $x$ that tends to $0$. For series apply this to the partial sums. ∎
Uniform convergence is sufficient but far from necessary. The sharp results belong to Lebesgue theory: the dominated convergence theorem (§4.5). Its Riemann-integral special case is worth knowing by name: Arzelà's bounded convergence theorem — if $f_n$ and $f$ are Riemann integrable on $[a,b]$, $|f_n(x)|\le M$ for all $n,x$, and $f_n\to f$ pointwise, then $\int_a^bf_n\to\int_a^bf$. (It follows from the dominated convergence theorem because Riemann and Lebesgue integrals agree, §4.5.)
Differentiation Term by Term
Uniform convergence of $f_n$ says nothing about $f_n'$: $\frac{\sin nx}{\sqrt n}\rightrightarrows0$ while $f_n'(0)=\sqrt n$. The right hypothesis is on the derivatives.
- $(f_n(x_0))$ converges for at least one $x_0\in[a,b]$, and
- $f_n'\rightrightarrows g$ on $[a,b]$.
Proof. Uniform convergence of $f_n$. Let $\varepsilon\gt0$ and choose $N$ such that for $m,n\ge N$: $|f_n(x_0)-f_m(x_0)|\lt\frac\varepsilon2$ and $\|f_n'-f_m'\|\lt\frac{\varepsilon}{2(b-a)}$. Apply the mean value theorem (§3.3) to $h=f_n-f_m$: for $x,t\in[a,b]$, $$\begin{aligned}|h(x)-h(t)|&\le|x-t|\cdot\frac{\varepsilon}{2(b-a)}\\&\le\frac\varepsilon2 .\qquad(\ast)\end{aligned}$$ Hence $|h(x)|\le|h(x)-h(x_0)|+|h(x_0)|\lt\varepsilon$ for all $x$: $(f_n)$ is uniformly Cauchy and converges uniformly to some $f$ (Theorem 5.4).
The derivative. Fix $x\in[a,b]$ and define, for $t\in[a,b]$, $t\ne x$, $$\varphi_n(t)=\frac{f_n(t)-f_n(x)}{t-x},$$ $$\varphi(t)=\frac{f(t)-f(x)}{t-x}.$$ Then $\varphi_n\to\varphi$ pointwise, and $\lim_{t\to x}\varphi_n(t)=f_n'(x)$. Dividing $(\ast)$ by $|t-x|$ gives $|\varphi_n(t)-\varphi_m(t)|\le\frac{\varepsilon}{2(b-a)}$ for $m,n\ge N$, so $\varphi_n\rightrightarrows\varphi$ on $[a,b]\setminus\{x\}$. By the Moore–Osgood theorem (Theorem 5.10), $$\lim_{t\to x}\varphi(t)=\lim_{n\to\infty}f_n'(x)=g(x).$$ Thus $f$ is differentiable at $x$ with $f'(x)=g(x)$. ∎
If the $f_n'$ are continuous, there is a shorter proof: $f_n(x)=f_n(x_0)+\int_{x_0}^xf_n'$, and Theorem 5.11 lets us pass to the limit, giving $f(x)=f(x_0)+\int_{x_0}^xg$; then the fundamental theorem of calculus (§4.1) gives $f'=g$. The proof above avoids continuity of $f_n'$. Hypothesis 1 cannot be dropped ($f_n=n$ has $f_n'=0\rightrightarrows0$ but diverges), and hypothesis 2 cannot be weakened to pointwise convergence (Example 2(a), (b)).
Power Series and the Radius of Convergence
A power series about $c\in\mathbb{R}$ is a series of functions $\sum_{n\ge0}a_n(x-c)^n$ with real coefficients $a_n$. Its radius of convergence is the number $R\in[0,\infty]$ of Theorem 5.13; the set of $x$ where it converges is its interval of convergence.
A sequence or series of functions converges uniformly on compact subsets (locally uniformly) of an open set $U$ if it converges uniformly on every compact $K\subseteq U$. By Theorem 5.9 the limit is then continuous on $U$, because each point of $U$ has a compact neighbourhood in $U$.
- converges absolutely for $|x-c|\lt R$;
- diverges for $|x-c|\gt R$ (its terms do not tend to $0$);
- converges absolutely and uniformly on $[c-r,c+r]$ for every $0\lt r\lt R$, hence uniformly on compact subsets of $(c-R,c+R)$, and its sum is continuous on $(c-R,c+R)$.
Proof. Apply the root test (§2.3) to the series of absolute values, $\sum|a_n||x-c|^n$. We have $\limsup\big(|a_n||x-c|^n\big)^{1/n}=\rho|x-c|$. If $\rho|x-c|\lt1$ the series converges; if $\rho|x-c|\gt1$ then $|a_n(x-c)^n|\gt1$ for infinitely many $n$, so the terms do not tend to $0$. This proves 1 and 2 (with the conventions when $\rho=0$ or $\infty$). For 3, let $|x-c|\le r\lt R$: then $|a_n(x-c)^n|\le|a_n|r^n=:M_n$, and $\sum M_n\lt\infty$ by part 1 at the point $c+r$. The M-test gives absolute uniform convergence on $[c-r,c+r]$. A compact $K\subseteq(c-R,c+R)$ attains $r=\max_K|x-c|\lt R$, so $K\subseteq[c-r,c+r]$. Continuity follows from Theorem 5.9. ∎
Proof. The chain of inequalities is the theorem that ratios control roots (§2.2). When the outer two terms agree and equal $L$, so does $\limsup|a_n|^{1/n}=\rho$, and $R=1/L$ by Theorem 5.13. ∎
- Radius. Use $R=\lim|a_n/a_{n+1}|$ when the ratio has a limit (factorials, binomials, $n^k$); use $R=1/\limsup|a_n|^{1/n}$ for $n$-th powers ($(1+\frac1n)^{n^2}$) and for coefficients that switch between two patterns. For series in $x^2$, $x^3$ or $(2x+1)$, substitute $t=x^2$ etc., find the radius in $t$, and translate back.
- End points. Test $x=c\pm R$ separately with the tools of §2.3 (the $n$-th term test, $p$-series, Leibniz). All four combinations occur: $\sum x^n$ (neither), $\sum\frac{x^n}n$ (only $-1$), $\sum\frac{(-x)^n}n$ (only $1$), $\sum\frac{x^n}{n^2}$ (both).
- Uniformity. Uniform on every $[c-r,c+r]$ with $r\lt R$; on the whole interval only with an extra argument — the M-test when $\sum|a_n|R^n\lt\infty$, Abel's theorem (Theorem 5.16) at a convergent end point.
- the series $\sum na_n(x-c)^{n-1}$ and $\sum\frac{a_n}{n+1}(x-c)^{n+1}$ also have radius $R$;
- $f$ is differentiable on $(c-R,c+R)$ with $f'(x)=\sum_{n\ge1}na_n(x-c)^{n-1}$, and $\int_c^xf=\sum_{n\ge0}\frac{a_n}{n+1}(x-c)^{n+1}$ there;
- $f$ is infinitely differentiable on $(c-R,c+R)$ and $a_n=\dfrac{f^{(n)}(c)}{n!}$: a power series is the Taylor series of its sum;
- (uniqueness) if $\sum a_n(x-c)^n=\sum b_n(x-c)^n$ for all $|x-c|\lt\delta$, $\delta\gt0$, then $a_n=b_n$ for every $n$.
Proof. (1) $(n|a_n|)^{1/n}=n^{1/n}|a_n|^{1/n}$ and $n^{1/n}\to1$ (§2.1). If $t_n\to1$ and $s_n\ge0$, then $\limsup t_ns_n=\limsup s_n$ (for large $n$, $(1-\eta)s_n\le t_ns_n\le(1+\eta)s_n$). So $\sum na_nx^n$ has the same radius as $\sum a_nx^n$, and dividing by $x$ does not change convergence for $x\ne0$. The same argument with $(n+1)^{-1/n}\to1$ handles the integrated series. (2) Fix $0\lt r\lt R$. On $[c-r,c+r]$ the differentiated series converges uniformly by (1) and Theorem 5.13, and the series for $f$ converges at $x=c$; Theorem 5.12 gives $f'$ termwise on $[c-r,c+r]$, hence on $(c-R,c+R)$. The integral statement is (2) applied to the integrated series, whose derivative is $f$. (3) Apply (2) repeatedly: $$f^{(k)}(x)=\sum_{n\ge k}\frac{n!}{(n-k)!}\,a_n(x-c)^{n-k};$$ at $x=c$ only the term $n=k$ survives, giving $f^{(k)}(c)=k!\,a_k$. (4) Both sums are the same function near $c$, so by (3) $a_n=\frac{f^{(n)}(c)}{n!}=b_n$. ∎
Abel's Theorem: Behaviour at the End Points
Inside $(c-R,c+R)$ everything is smooth; the end points are where the subtle questions live. If the series happens to converge at an end point, is its value there the limit of $f(x)$ from inside? Abel's answer is yes.
Proof. Use Abel's test for uniform convergence (Theorem 5.7(2), §5.1) on $D=[0,1]$ with the constant functions $a_n$ and with $b_n(x)=x^n$. The series $\sum a_n$ converges, and since its terms do not depend on $x$ it converges uniformly on $D$. For each $x\in[0,1]$ the sequence $x^n$ is decreasing in $n$, and $|x^n|\le1$. So $\sum a_nx^n$ converges uniformly on $[0,1]$. Its terms are continuous, so the sum is continuous on $[0,1]$ (Theorem 5.9), in particular at $1$. The general statement follows by the substitution $x=c+Rt$, $t\in[0,1]$, and $x=c-Rt$ at the left end point. ∎
Proof. Since $\sum a_n$ converges, $(a_n)$ is bounded, so $\limsup|a_n|^{1/n}\le1$ and $\sum a_nx^n$ has radius $\ge1$; likewise for $b_n$ and for $c_n$ (as $\sum c_n$ converges). For $0\le x\lt1$ both $\sum a_nx^n$ and $\sum b_nx^n$ converge absolutely, so by Mertens' theorem (§2.3) $$\sum c_nx^n=\Big(\sum a_nx^n\Big)\Big(\sum b_nx^n\Big)$$ for $0\le x\lt1$, because the Cauchy product of $\sum a_nx^n$ and $\sum b_nx^n$ has terms $c_nx^n$. Let $x\to1^-$: by Theorem 5.16 applied three times, the left side tends to $\sum c_n$ and the right side to $AB$. ∎
The converse of Abel's theorem is false: $\sum(-1)^nx^n=\frac1{1+x}\to\frac12$ as $x\to1^-$, but $\sum(-1)^n$ diverges. A limit $\lim_{x\to1^-}\sum a_nx^n$ is called the Abel sum of $\sum a_n$. Partial converses are called Tauberian theorems; the first, Tauber's theorem, says that if $na_n\to0$ and the Abel sum is $s$, then $\sum a_n$ converges to $s$ (statement only).
The Weierstrass Approximation Theorem
Example 6 of §5.1 built polynomials converging uniformly to $|t|$. Weierstrass proved in 1885 that every continuous function on a closed bounded interval is a uniform limit of polynomials. Among many proofs, Bernstein's (1912) is explicit and short, and it has a probabilistic meaning: $B_nf(x)$ is the expected value of $f(S/n)$ when $S$ is the number of successes in $n$ independent trials with success probability $x$.
Proof. The first identity is the binomial theorem for $(x+(1-x))^n$. Using $k\binom nk=n\binom{n-1}{k-1}$, $$\begin{aligned}\sum_kk\,p_{n,k}(x)&=nx\sum_{k\ge1}\tbinom{n-1}{k-1}x^{k-1}(1-x)^{n-k}\\&=nx .\end{aligned}$$ Similarly $k(k-1)\binom nk=n(n-1)\binom{n-2}{k-2}$ gives $\sum_kk(k-1)p_{n,k}=n(n-1)x^2$. Hence $$\begin{aligned}\sum_k(k-nx)^2p_{n,k}&=\big[n(n-1)x^2+nx\big]\\&\quad-2nx\cdot nx+n^2x^2\\&=nx(1-x).\end{aligned}$$ Divide by $n^2$; finally $x(1-x)\le\frac14$. ∎
Proof. Let $M=\|f\|$ and $\varepsilon\gt0$. Since $f$ is uniformly continuous on $[0,1]$ (§3.2), there is $\delta\gt0$ with $|f(s)-f(x)|\lt\varepsilon$ whenever $|s-x|\lt\delta$. Fix $x\in[0,1]$. By the first identity of Theorem 5.18, $$\begin{aligned}B_nf(x)&-f(x)\\&=\sum_{k=0}^n\Big(f\big(\tfrac kn\big)-f(x)\Big)p_{n,k}(x).\end{aligned}$$ Split the sum. For the $k$ with $|\frac kn-x|\lt\delta$, each bracket is $\lt\varepsilon$, and these terms contribute at most $\varepsilon\sum p_{n,k}=\varepsilon$. For the other $k$, $|f(\frac kn)-f(x)|\le2M\le2M\frac{(k/n-x)^2}{\delta^2}$, so by Theorem 5.18 they contribute at most $$\begin{aligned}\frac{2M}{\delta^2}\sum_k\Big(\frac kn-x\Big)^2p_{n,k}(x)&\le\frac{2M}{\delta^2}\cdot\frac1{4n}\\&=\frac{M}{2n\delta^2}.\end{aligned}$$ Therefore $\|B_nf-f\|\le\varepsilon+\frac{M}{2n\delta^2}\lt2\varepsilon$ for all $n\gt\frac{M}{2\delta^2\varepsilon}$. For a general interval $[a,b]$, apply this to $g(t)=f(a+(b-a)t)$ on $[0,1]$ and substitute back $t=\frac{x-a}{b-a}$; a polynomial in $t$ is a polynomial in $x$. ∎
- There are polynomials with rational coefficients converging uniformly to $f$ on $[a,b]$. (So $C[a,b]$ with the sup norm has a countable dense subset — §7.4.)
- Moments determine $f$: if $\int_a^bf(x)x^n\,dx=0$ for every $n=0,1,2,\dots$, then $f\equiv0$.
Proof. (1) Take a polynomial $p=\sum_{j\le d}c_jx^j$ with $\|f-p\|\lt\varepsilon/2$, and rationals $q_j$ with $|c_j-q_j|\lt\frac{\varepsilon}{2(d+1)K^j}$, where $K$ is the largest of the numbers $1$, $|a|$ and $|b|$. Then $\big\|p-\sum q_jx^j\big\|\le\sum|c_j-q_j|K^j\lt\varepsilon/2$. (2) By linearity $\int_a^bfp=0$ for every polynomial $p$. Choose polynomials $p_n\rightrightarrows f$. Then $$\begin{aligned}\int_a^bf^2&=\int_a^bf(f-p_n)\\&\le\|f\|\,\|f-p_n\|\,(b-a)\to0,\end{aligned}$$ so $\int_a^bf^2=0$. A continuous non-negative function with zero integral vanishes identically (§4.1), so $f^2\equiv0$. ∎
- Pointwise limits do not commute with integrals: $n^2x(1-x)^n\to0$ but $\int_0^1\to1$.
- Uniform convergence is sufficient, not necessary: $nx(1-x)^n\to0$ non-uniformly with $\int_0^1\to0$ (bounded convergence).
- Unbounded intervals: $\frac1ne^{-x/n}\rightrightarrows0$ on $[0,\infty)$ but $\int_0^\infty=1$. Theorem 5.11 needs $[a,b]$.
- Uniform convergence of $f_n$ does not give $f_n'\to f'$: $\frac{\sin nx}{\sqrt n}$; $\frac{x}{1+nx^2}$ at $0$.
- Theorem 5.12 needs convergence at one point: $f_n=n$ has $f_n'\rightrightarrows0$.
- Convergence of the differentiated series is not enough: $\sum\frac{\sin nx}{n^3}$ has no second derivative at $0$, though $-\sum\frac{\sin nx}{n}$ converges there.
- Power series converge uniformly on compact subsets, not on the open interval: $\sum x^n$ on $(-1,1)$; $\sum(-1)^nx^{2n}$ has tails with sup $\frac12$ on $(-1,1)$.
- End points can do anything: $\sum x^n$, $\sum\frac{x^n}n$, $\sum\frac{(-x)^n}n$, $\sum\frac{x^n}{n^2}$ realise all four cases; and differentiation can lose an end point ($\sum\frac{x^n}{n^2}$ vs $\sum\frac{x^{n-1}}n$ at $1$).
- The ratio test can fail to give $R$: $\sum(2+(-1)^n)^nx^n$ — use $\limsup|a_n|^{1/n}$.
- Smooth on $\mathbb{R}$ but radius $1$: $\frac1{1+x^2}=\sum(-1)^nx^{2n}$ only for $|x|\lt1$ (the obstruction is the complex zeros $\pm i$ of $1+z^2$ — Unit 2).
- $C^\infty$ but not analytic: $e^{-1/x^2}$ (with value $0$ at $0$) has all derivatives $0$ at $0$, so its Taylor series is $0$ and represents it only at $x=0$.
- The converse of Abel's theorem fails: $\sum(-1)^nx^n\to\frac12$ as $x\to1^-$, but $\sum(-1)^n$ diverges.
- Weierstrass needs a closed bounded interval: no polynomials converge uniformly to $\frac1x$ on $(0,1)$ (uniform limits of bounded functions are bounded) or to $e^x$ on $\mathbb{R}$ (uniform limits of polynomials on $\mathbb{R}$ are polynomials, §5.1 Exercise IV.5).
- Even moments do not determine $f$ on $[-1,1]$: $f(x)=x$.
Exercises
- 1 Use Theorem 5.9 to show that $\sum_{n\ge0}\frac{x^2}{(1+x^2)^n}$ does not converge uniformly on $[-1,1]$.
- 2 Show that $f_n(x)=\frac{nx}{1+n^2x^2}$ does not converge uniformly on $[0,1]$, but $\int_0^1f_n\to\int_0^1\lim f_n$.
- 3 Justify integrating term by term to show $\int_0^1\sum_{n\ge1}\frac{x^n}{n^2}\,dx=\frac{\pi^2}6-1$.
- 4 Find $\lim_n\int_0^1nxe^{-nx^2}\,dx$ and compare it with $\int_0^1\lim_nnxe^{-nx^2}\,dx$. Is the convergence uniform?
- 5 Let $f(x)=\sum_{n\ge1}\frac1{n^2}\cdot\frac{x}{x+n}$ for $x\ge0$. Prove that $\lim_{x\to\infty}f(x)=\frac{\pi^2}6$.
- 6 Let $f_n$ be continuous on $[a,b]$ and $f_n\rightrightarrows f$ on $(a,b)$. Prove that $(f_n)$ converges uniformly on $[a,b]$.
- 7 Let $f_n$ be Riemann integrable on $[a,b]$ with $f_n\rightrightarrows f$. Prove that $\int_a^b|f_n-f|\to0$ and $\int_a^bf_n^2\to\int_a^bf^2$.
- 8 Use Arzelà's bounded convergence theorem to find $\lim_n\int_0^1\frac{dx}{1+nx^3}$.
- 1 Show that $f(x)=\sum_{n\ge1}\frac{\cos nx}{n^4}$ is twice continuously differentiable on $\mathbb{R}$ and find $f''$ as a series.
- 2 Show that $\zeta(x)=\sum_{n\ge1}n^{-x}$ is differentiable on $(1,\infty)$ with $\zeta'(x)=-\sum_{n\ge2}\frac{\log n}{n^x}$.
- 3 Let $f_n(x)=\frac1n\log(1+n^2x^2)$ on $[-1,1]$. Show $f_n\rightrightarrows0$ and $f_n'\to0$ pointwise but not uniformly. Does $\lim f_n'=(\lim f_n)'$ hold anyway?
- 4 Let $f(x)=\sum_{n\ge1}\frac{e^{-nx}}{1+n^2}$. Show that $f$ is continuous on $[0,\infty)$ and differentiable on $(0,\infty)$. Show that $\frac{f(h)-f(0)}h\to-\infty$ as $h\to0^+$.
- 5 Prove that $y(x)=\sum_{n\ge0}\frac{x^{2n}}{4^n(n!)^2}$ converges for all $x$ and satisfies $xy''+y'-xy=0$.
- 6 Give an example of differentiable $f_n$ on $[0,1]$ with $f_n'\rightrightarrows0$ for which $(f_n)$ does not converge at any point.
- 1 Interval of convergence of $\sum\frac{n^2}{3^n}x^n$.
- 2 Interval of convergence of $\sum\frac{x^n}{\sqrt n\,2^n}$.
- 3 Interval of convergence of $\sum\frac{(x-1)^n}{n^2}$.
- 4 Interval of convergence of $\sum\binom{2n}{n}x^n$. (Use $\binom{2n}n4^{-n}\sim\frac1{\sqrt{\pi n}}$ and that $\binom{2n}{n}4^{-n}$ decreases.)
- 5 Radius of $\sum a_nx^n$ where $a_n=2^n$ for even $n$ and $a_n=3^{-n}$ for odd $n$. Does the ratio formula apply?
- 6 Radius of $\sum(\sin n)\,x^n$.
- 7 Find $\sum_{n\ge1}n^2x^n$ for $|x|\lt1$ and evaluate $\sum\frac{n^2}{3^n}$.
- 8 Show $\sum_{n\ge0}\frac{x^{2n+1}}{2n+1}=\frac12\log\frac{1+x}{1-x}$ for $|x|\lt1$ and evaluate $\sum_{n\ge0}\frac1{(2n+1)3^{2n+1}}$.
- 9 If $\sum a_nx^n$ and $\sum b_nx^n$ have radii $R\ne S$, show that $\sum(a_n+b_n)x^n$ has radius $\min(R,S)$. What can happen when $R=S$?
- 1 Evaluate $\sum_{n\ge0}\frac{(-1)^n}{(2n+1)(2n+2)}$.
- 2 Show that $\sum_{n\ge1}\frac{x^n}n$ converges uniformly on $[-1,r]$ for every $r\lt1$, but not on $[-1,1)$.
- 3 Find the Abel sum of $1-2+3-4+\cdots$, i.e. $\lim_{x\to1^-}\sum_{n\ge1}(-1)^{n+1}nx^{n-1}$. Does the series converge?
- 4 Compute $B_3f$ for $f(x)=|x-\frac12|$ and compare it with $B_2f$ (Example 7(b)).
- 5 Show that $B_n(x^3)=x^3+\frac{x(1-x)\big((3n-2)x+1\big)}{n^2}$, and deduce $B_n(x^3)\rightrightarrows x^3$ directly.
- 6 Let $f\in C[0,1]$ and $k\in\mathbb{N}$. If $\int_0^1f(x)x^n\,dx=0$ for every $n\ge k$, prove $f\equiv0$.
- 7 Prove that no sequence of polynomials converges uniformly to $\frac1x$ on $(0,1)$, nor to $\sin x$ on $\mathbb{R}$.
- 8 Let $f\in C[a,b]$. Show that there are polynomials $p_n\rightrightarrows f$ on $[a,b]$ with $p_n(a)=f(a)$ and $p_n(b)=f(b)$ for every $n$.
- 9 Applied (design curves). A designer approximates the profile $f(x)=\sin\pi x$ on $[0,1]$ by $B_nf$. Using Theorem 5.18 and $|f''|\le\pi^2$, show that $\|B_nf-f\|\le\frac{\pi^2}{8n}$, and find $n$ guaranteeing an error below $0.01$.
E1. CSIR Part B GATE MA The radius of convergence of $\sum_{n\ge1}\left(1+\frac1n\right)^{n^2}x^n$ is (A) $1$; (B) $e$; (C) $\frac1e$; (D) $0$.
E2. CSIR Part B $\displaystyle\lim_{n\to\infty}\int_0^1\frac{nx^{n-1}}{1+x}\,dx$ equals (A) $0$; (B) $\frac12$; (C) $\log2$; (D) $1$.
E3. CSIR Part B Let $f(x)=\sum_{n\ge1}\frac{x^n}{n^2}$. Which statement is true? (A) The series converges uniformly on $[-1,1]$ and $f'(x)\to\infty$ as $x\to1^-$; (B) the series converges uniformly on $[-1,1]$ and $f'$ is bounded on $(-1,1)$; (C) the series converges only on $(-1,1)$; (D) the series converges uniformly on $[-1,1]$, but $f$ is not continuous at $1$.
E4. CSIR Part C Let $\sum a_nx^n$ have radius of convergence $1$. Which of the following are true? (One or more options may be correct.) (A) $\sum na_nx^n$ has radius of convergence $1$. (B) If $\sum a_n$ converges, then $\sum a_nx^n$ converges uniformly on $[0,1]$. (C) $\sum a_nx^n$ converges uniformly on $(-1,1)$. (D) If $\sum a_nx^n$ converges at $x=1$, then it converges at $x=-1$.
E5. CSIR Part C Let $f:[-1,1]\to\mathbb{R}$ be continuous. Which of the following conditions force $f\equiv0$? (A) $\int_{-1}^1f(x)x^n\,dx=0$ for all $n\ge0$. (B) $\int_{-1}^1f(x)x^{2n}\,dx=0$ for all $n\ge0$. (C) $\int_{-1}^1f(x)x^n\,dx=0$ for all $n\ge1$. (D) $\int_{-1}^1f(x)x^n\,dx\to0$ as $n\to\infty$.
E6. CSIR Part C Let $f_n(x)=\frac{\sin(n^2x)}{n}$ for $x\in\mathbb{R}$. Which of the following are true? (A) $f_n\to0$ uniformly on $\mathbb{R}$. (B) $f_n'(x)\to0$ for every $x\in\mathbb{R}$. (C) $\sum_{n\ge1}\frac{f_n(x)}{n}$ converges uniformly on $\mathbb{R}$ to a continuous function. (D) $\sum_{n\ge1}\frac{f_n(x)}{n}$ may be differentiated term by term at every $x\in\mathbb{R}$.
Suggested plan — 3 lectures × 60 min.
- The three interchange theorems. Theorem 5.9 with the $\varepsilon/3$ proof (draw the three gaps), Moore–Osgood, Theorem 5.11 (prove integrability of the limit carefully). Example 1 as a "predict, then compute" sequence; state Arzelà's theorem. Theorem 5.12 with the MVT proof; Example 2(a), (b) as warnings.
- Power series. Example 3 (the Basel sum from term-by-term calculus) is a memorable start. Cauchy–Hadamard and the ratio formula, the strategy box, Example 4. Theorem 5.15 and Example 5. Project Interactive 1: $\sum x^n$ with $r=0.8$, then slide $r$ to $1$; switch to $\frac1{1+x^2}$ (smooth function, finite radius) and to $\arctan$ and $\sum\frac{x^n}{n^2}$ (uniform on the closed interval).
- Abel and Weierstrass. Abel's theorem via Abel's test, Example 6, Abel's product theorem and the remark on Tauber. Bernstein polynomials: prove Theorem 5.18 and 5.19 in full (the probabilistic picture helps). Interactive 2: $x^2$ (rate $\frac1{4n}$), then $|x-\frac12|$ (rate $n^{-1/2}$). Theorem 5.20 and Example 7(c), (d). Close with E5.
Discussion questions. (1) Why does Theorem 5.12 put the uniform convergence hypothesis on $f_n'$ rather than on $f_n$? (2) A power series converges uniformly on every $[-r,r]$ with $r\lt R$. Why does this already give continuity on the open interval, even when the convergence is not uniform there? (3) Bernstein polynomials converge slowly even for $x^2$. Why might they still be preferred in computer graphics?
Misconceptions to watch for. Interchanging limit and integral on an unbounded interval because the convergence is uniform; differentiating a uniformly convergent series term by term; using the ratio formula when the ratios oscillate; forgetting to test the end points (or testing them with the ratio test, which is always inconclusive there); believing that a function smooth on $\mathbb{R}$ has an everywhere convergent Taylor series; reading Abel's theorem backwards.
Suggested homework. I(2, 4, 5, 8), II(1, 3, 4), III(2, 4, 5, 8), IV(1, 2, 6, 7, 9); Exam corner E2, E4, E5.
Research question. Bernstein polynomials converge slowly, yet they are the backbone of computer-aided design. What do they do well that faster approximations do not?
Why it matters. In the early 1960s two engineers in the French car industry, Paul de Casteljau at Citroën and Pierre Bézier at Renault, independently designed car bodies with curves of the form $\sum_kP_k\,p_{n,k}(t)$, where the $P_k$ are "control points" — exactly Bernstein polynomials with vector coefficients. Today the same Bézier curves describe fonts on every screen, vector graphics, and the motion paths of robots and animations.
What to do.
- Implement de Casteljau's algorithm: repeated linear interpolation $b_k^{(r)}=(1-t)b_k^{(r-1)}+tb_{k+1}^{(r-1)}$ starting from $b_k^{(0)}=f(\frac kn)$. Check that it reproduces $B_nf(t)$.
- For $f(x)=x^2$, $\sin\pi x$ and $|x-\frac12|$, tabulate $\|B_nf-f\|$ for $n=10,20,40,80,160$ and estimate the rate (the slope on a log–log plot).
- Verify Voronovskaya's theorem numerically for $f=\sin\pi x$ at $x=\frac12$: $n\big(B_nf(x)-f(x)\big)$ tends to $\tfrac12x(1-x)f''(x)$ (test it at $x=\tfrac12$).
- Shape preservation: take an increasing, convex $f$ and check that every $B_nf$ is increasing and convex. Compare with the polynomial that interpolates $f$ at $n+1$ equally spaced points for Runge's function $\frac1{1+25(2x-1)^2}$.
Tools: Python with NumPy (or a spreadsheet), and a plotting tool.
Hand in: Your table of errors with estimated rates, one plot comparing $B_{20}f$ and the degree-20 interpolant of Runge's function, and a short paragraph explaining why designers accept slow convergence in exchange for stability and shape preservation.
Going further. Read about the variation-diminishing property of Bernstein polynomials, and about Chebyshev interpolation, which converges much faster for smooth functions.
🧪 Self-check quiz
Q1. Which sequence of continuous functions cannot converge uniformly on $[0,1]$, by Theorem 5.9 alone?
Q2. The radius of convergence of $\sum\frac{n^3}{3^n}x^n$ is
Q3. To conclude $\big(\sum u_n\big)'=\sum u_n'$ on $[a,b]$ it suffices that
Q4. $\displaystyle\lim_{x\to1^-}\sum_{n\ge0}\frac{(-1)^nx^n}{n+1}$ equals
Q5. For $f(x)=x^2$ on $[0,1]$, $\|B_nf-f\|_\infty$ equals