8.2 Uniform Boundedness, Open Mapping and Closed Graph Theorems
Three theorems that completeness buys: pointwise bounded families of operators are uniformly bounded, surjective bounded operators between Banach spaces are open, and closed graphs force boundedness — all from the Baire category theorem, with the counterexamples that show why every hypothesis is needed.
In finite dimensions every linear map is continuous, every injective linear map onto the target has a continuous inverse, and any two norms are equivalent. Section 8.1 showed that none of this survives in infinite dimensions. The remarkable discovery of Banach and Steinhaus in the 1920s is that a large part of it comes back as soon as the spaces are complete. If a family of bounded operators on a Banach space is bounded at each point, it is bounded uniformly; a bounded bijection between Banach spaces has a bounded inverse, so two comparable complete norms are equivalent; and a linear map between Banach spaces whose graph is closed is automatically bounded. All three theorems rest on one topological fact about complete metric spaces, the Baire category theorem.
These results are tools for proving that something is bounded without computing a bound, and — read backwards — for proving that a space is not complete or that a bad object exists. The uniform boundedness principle produces a continuous function whose Fourier series diverges at a point, without writing the function down. The bounded inverse theorem shows at once that $(C[0,1],\|\cdot\|_1)$ cannot be complete. GATE tests these theorems in MSQ form — which conclusions follow, and which hypotheses (completeness of the domain? of the range? surjectivity?) can be dropped — and through the standard examples: $c_{00}$, the diagonal operator $x\mapsto(x_n/n)$, and differentiation.
Baire's Theorem Revisited
We use the language of §4.1 (complete metric spaces) and of the Baire category theorem proved in CSIR Unit 1 · §7.1. For completeness we recall the statement and give a short proof.
📘 Definition 8.7 — Nowhere dense, meagre and residual sets
A subset $A$ of a metric space $X$ is nowhere dense if its closure has empty interior. $A$ is meagre (of the first category) if it is a countable union of nowhere dense sets, and of the second category otherwise. $A$ is residual if its complement is meagre. Examples in $\mathbb{R}$: $\mathbb{Z}$ and the Cantor set are nowhere dense; $\mathbb{Q}$ is meagre (a countable union of points) but dense; $\mathbb{R}\setminus\mathbb{Q}$ is residual. A proper closed subspace of a normed space is nowhere dense: if it contained a ball $B(x,r)$, it would contain $B(0,r)=B(x,r)-x$ and hence every vector.
📐 Theorem 8.12 — Baire category theorem
Let $(X,d)$ be a nonempty complete metric space. (a) If $U_1,U_2,\dots$ are open and dense in $X$, then $\bigcap_nU_n$ is dense in $X$. (b) Equivalently, $X$ is not meagre in itself: if $X=\bigcup_nF_n$ with every $F_n$ closed, then some $F_n$ contains an open ball.
Proof. (a) Let $W$ be a nonempty open set; we find a point of $W\cap\bigcap U_n$. Since $U_1$ is dense, $W\cap U_1$ is a nonempty open set, so it contains a closed ball $\overline B(x_1,r_1)$ with $r_1\lt1$. Inductively, $B(x_{n},r_{n})\cap U_{n+1}$ is nonempty and open and contains a closed ball $\overline B(x_{n+1},r_{n+1})$ with $r_{n+1}\lt r_n/2$. For $m\gt n$, $x_m\in\overline B(x_n,r_n)$, so $d(x_m,x_n)\le r_n\to0$: $(x_n)$ is Cauchy and converges to some $x$. Since $x_m\in\overline B(x_n,r_n)$ for all $m\ge n$ and the ball is closed, $x\in\overline B(x_n,r_n)\subseteq U_n$ for every $n$, and $x\in\overline B(x_1,r_1)\subseteq W$. (b) If no $F_n$ contained a ball, the sets $U_n=X\setminus F_n$ would be open and dense, and (a) would give a point in $\bigcap U_n=X\setminus\bigcup F_n=\varnothing$. ∎
✏️ Example 1 — No Banach space has a countably infinite Hamel basis
Prove that an infinite-dimensional Banach space cannot have a countable Hamel basis. Deduce that there is no norm on $c_{00}$, or on the space $\mathcal{P}$ of all polynomials, under which it is a Banach space.
💡 Show solution▶
Suppose $X$ is Banach with a Hamel basis $e_1,e_2,\dots$, so that every vector is a finite linear combination of the $e_k$. Let $F_n=\operatorname{span}\{e_1,\dots,e_n\}$. Then $X=\bigcup_nF_n$.
Each $F_n$ is small $F_n$ is finite-dimensional, hence closed (Theorem 8.6, §8.1), and it is a proper subspace ($e_{n+1}\notin F_n$), hence nowhere dense (Definition 8.7).
Contradiction A complete metric space is not a countable union of nowhere dense closed sets (Theorem 8.12(b)).
Consequence $c_{00}$ has the countable Hamel basis $\{e_n\}$ and $\mathcal{P}$ has $\{1,x,x^2,\dots\}$; so whatever norm we put on them, they are not complete. (A Hamel basis of an infinite-dimensional Banach space such as $\ell^2$ is therefore uncountable — and is never used in practice; "bases" in analysis are orthonormal bases (§8.3) or Schauder bases, which allow infinite sums.)
The Uniform Boundedness Principle
📐 Theorem 8.13 — Uniform boundedness principle
Let $X$ be a Banach space, $Y$ a normed space and $\{T_\alpha\}_{\alpha\in I}\subseteq B(X,Y)$ any family. If the family is pointwise bounded, i.e. $\sup_\alpha\|T_\alpha x\|\lt\infty$ for every $x\in X$, then it is uniformly bounded: $\sup_\alpha\|T_\alpha\|\lt\infty$.
Proof. For $n\in\mathbb{N}$ let $$F_n=\{x\in X:\|T_\alpha x\|\le n\text{ for all }\alpha\},$$ the intersection over $\alpha$ of the sets $\{x:\|T_\alpha x\|\le n\}$. Each of these is closed (preimage of $[0,n]$ under the continuous function $x\mapsto\|T_\alpha x\|$), so $F_n$ is closed, and pointwise boundedness says $X=\bigcup_nF_n$. By Baire (Theorem 8.12(b)) some $F_N$ contains a ball $\overline B(x_0,r)$. Now let $\|z\|\le r$. Both $x_0+z$ and $x_0$ lie in $F_N$, so for every $\alpha$
$$\|T_\alpha z\|=\|T_\alpha(x_0+z)-T_\alpha x_0\|\le N+N=2N.$$
For a unit vector $u$ take $z=ru$: $\|T_\alpha u\|\le2N/r$. Hence $\|T_\alpha\|\le2N/r$ for every $\alpha$. ∎
📐 Theorem 8.14 — Banach–Steinhaus theorem
Let $X$ be a Banach space, $Y$ a normed space and $T_n\in B(X,Y)$.
If $\lim_nT_nx$ exists for every $x\in X$, then $Tx=\lim_nT_nx$ defines $T\in B(X,Y)$, $\sup_n\|T_n\|\lt\infty$ and $\|T\|\le\liminf_n\|T_n\|$.
If $Y$ is a Banach space, $\sup_n\|T_n\|\lt\infty$, and $T_nx$ converges for every $x$ in a dense subset $D$ of $X$, then $T_nx$ converges for every $x\in X$.
Condensation of singularities. If $\sup_\alpha\|T_\alpha\|=\infty$ for a family $\{T_\alpha\}\subseteq B(X,Y)$, then the set $\{x:\sup_\alpha\|T_\alpha x\|=\infty\}$ is residual, in particular dense in $X$.
Proof. (1) A convergent sequence in $Y$ is bounded, so the family is pointwise bounded and Theorem 8.13 gives $M=\sup\|T_n\|\lt\infty$. $T$ is linear because limits are, and $\|Tx\|=\lim\|T_nx\|\le\liminf_n\|T_n\|\,\|x\|$.
(2) Let $M=\sup\|T_n\|$, $x\in X$ and $\varepsilon\gt0$. Choose $d\in D$ with $\|x-d\|\lt\varepsilon/(3M+1)$, and $N$ with $\|T_nd-T_md\|\lt\varepsilon/3$ for $n,m\ge N$. Then
$$\begin{aligned}\|T_nx-T_mx\|&\le\|T_n(x-d)\|+\|T_nd-T_md\|\\&\quad+\|T_m(d-x)\|\lt\varepsilon.\end{aligned}$$
So $(T_nx)$ is Cauchy, and converges because $Y$ is complete.
(3) The complement is $E=\{x:\sup_\alpha\|T_\alpha x\|\lt\infty\}=\bigcup_nF_n$ with $F_n$ as in the proof of Theorem 8.13. If some $F_n$ contained a ball, that proof would give $\sup_\alpha\|T_\alpha\|\lt\infty$. So every $F_n$ is closed with empty interior, i.e. nowhere dense, and $E$ is meagre. The set in question is $X\setminus E=\bigcap_n(X\setminus F_n)$, an intersection of dense open sets, hence dense by Theorem 8.12(a) (a dense $G_\delta$ set). ∎
📐 Theorem 8.15 — Weakly bounded sets are bounded
(a) Let $X$ be a normed space and $S\subseteq X$. If $\sup_{x\in S}|f(x)|\lt\infty$ for every $f\in X^*$, then $S$ is bounded in norm. (b) Let $X$ be a Banach space and $S\subseteq X^*$. If $\sup_{f\in S}|f(x)|\lt\infty$ for every $x\in X$, then $\sup_{f\in S}\|f\|\lt\infty$.
Proof. (b) is Theorem 8.13 with $Y=\mathbb{F}$. (a) Consider the functionals $Jx\in X^{**}=B(X^*,\mathbb{F})$, $x\in S$ (Theorem 8.10(4), §8.1). The domain $X^*$ is a Banach space even if $X$ is not (Theorem 8.5), and the hypothesis says $\sup_{x\in S}|(Jx)(f)|\lt\infty$ for each $f\in X^*$. By Theorem 8.13, $\sup_{x\in S}\|Jx\|\lt\infty$, and $\|Jx\|=\|x\|$. ∎
So a sequence $(x_n)$ for which $f(x_n)$ converges (or merely stays bounded) for every $f\in X^*$ — a weakly convergent sequence — is norm bounded. In $\ell^2$, $f(e_n)=y_n\to0$ for every $f=\varphi_y$, so $e_n\to0$ weakly while $\|e_n\|=1$: weak convergence does not give norm convergence, but it does give boundedness.
✏️ Example 2 — Completeness cannot be dropped
On $X=(c_{00},\|\cdot\|_\infty)$ let $f_n(x)=nx_n$ and $T_nx=(x_1,2x_2,\dots,nx_n,0,0,\dots)$. Show that $(f_n)$ is pointwise bounded but $\|f_n\|=n$, and that $T_nx$ converges for every $x$ to a limit $Tx$ which is not bounded. Why do Theorems 8.13 and 8.14 not apply?
💡 Show solution▶
Pointwise bounded Each $x\in c_{00}$ has $x_n=0$ for $n\gt N(x)$, so $|f_n(x)|\le\max_{k\le N}k|x_k|$ for all $n$ — finite.
Not uniformly bounded $|f_n(x)|\le n\|x\|_\infty$ with equality at $e_n$: $\|f_n\|=n\to\infty$.
Unbounded limit For $x\in c_{00}$, $T_nx=(x_1,2x_2,\dots,kx_k,\dots)=:Tx$ as soon as $n\ge N(x)$, so $T_nx\to Tx$ (eventually constant). But $\|Te_k\|_\infty=k$: $T$ is unbounded, and $\|T_n\|=n$.
Diagnosis $c_{00}$ is not complete (Theorem 8.3(2), §8.1), so the Baire argument is unavailable. On the completion $c_0$ the same $f_n$ are not pointwise bounded: $x=(n^{-1/2})\in c_0$ gives $f_n(x)=\sqrt n\to\infty$ — exactly as Theorem 8.14(3) predicts.
✏️ Example 3 — Multipliers: if $\sum x_ny_n$ converges for all $x$, then $y$ is small
(a) Suppose $\sum_nx_ny_n$ converges for every $x\in c_0$. Prove that $y\in\ell^1$. (b) Suppose $\sum_nx_ny_n$ converges for every $x\in\ell^p$, $1\le p\lt\infty$. Prove that $y\in\ell^q$ ($\tfrac1p+\tfrac1q=1$).
💡 Show solution▶
Set-up Let $\phi_N(x)=\sum_{n\le N}x_ny_n$. Each $\phi_N$ is a bounded functional — it is $\varphi_{y^{(N)}}$ for the truncation $y^{(N)}=(y_1,\dots,y_N,0,\dots)$ — and by Theorem 8.7 (§8.1) its norm is $\|y^{(N)}\|_1$ on $c_0$, $\|y^{(N)}\|_q$ on $\ell^p$.
Apply UBP The hypothesis says $\phi_N(x)$ converges, hence is bounded, for each $x$ in the Banach space $c_0$ (resp. $\ell^p$). By Theorem 8.13, $\sup_N\|\phi_N\|\lt\infty$, i.e. $\sup_N\sum_{n\le N}|y_n|\lt\infty$ in (a) and $\sup_N\|y^{(N)}\|_q\lt\infty$ in (b).
Conclude The partial sums of $\sum|y_n|$ (resp. $\sum|y_n|^q$, or $\sup|y_n|$ when $q=\infty$) are bounded, so $y\in\ell^1$ (resp. $y\in\ell^q$).
Remark A direct proof of (a) is possible (if $\sum|y_n|=\infty$, build $x_n\to0$ with $\sum x_ny_n=\infty$ by grouping blocks), but UBP turns a delicate construction into three lines. Part (b) is Landau's theorem.
✏️ Example 4 — A continuous function whose Fourier series diverges at $0$
Let $C(\mathbb{T})$ be the Banach space of continuous $2\pi$-periodic functions with $\|\cdot\|_\infty$, and $S_nf(0)=\sum_{|k|\le n}\hat f(k)$ the $n$-th partial sum of the Fourier series at $0$, where $\hat f(k)=\frac1{2\pi}\int_{-\pi}^{\pi}f(t)e^{-ikt}dt$. (a) Show $S_nf(0)=\frac1{2\pi}\int_{-\pi}^{\pi}f(t)D_n(t)\,dt$ with the Dirichlet kernel $D_n(t)=\sum_{|k|\le n}e^{ikt}=\dfrac{\sin((n+\frac12)t)}{\sin(t/2)}$. (b) Show that $\Lambda_n:f\mapsto S_nf(0)$ has norm $L_n=\frac1{2\pi}\int_{-\pi}^{\pi}|D_n|$ (the Lebesgue constant) and that $L_n\ge\frac4{\pi^2}\sum_{k=1}^n\frac1k$. (c) Conclude that there is $f\in C(\mathbb{T})$ whose Fourier series diverges at $0$ — indeed, such $f$ form a residual set.
💡 Show solution▶
(a) $\sum_{|k|\le n}\hat f(k)=\frac1{2\pi}\int f(t)\sum_{|k|\le n}e^{-ikt}dt$, and the sum is real and even, equal to $D_n(t)$. Summing the geometric series, $D_n(t)\sin\tfrac t2=\tfrac1{2i}\big(e^{i(n+\frac12)t}-e^{-i(n+\frac12)t}\big)$, which is $\sin\big((n+\tfrac12)t\big)$.
(b) Norm $\Lambda_n$ is the integral functional with continuous kernel $\frac1{2\pi}D_n$, so its norm is $\frac1{2\pi}\int|D_n|$ (Example 4(a) of §8.1, which works on any interval).
(b) Lower bound Since $|\sin\tfrac t2|\le\tfrac{|t|}2$ and $D_n$ is even, $$\begin{aligned}L_n&\ge\frac1\pi\int_0^{\pi}\frac{2|\sin((n+\frac12)t)|}{t}\,dt\\&=\frac2\pi\int_0^{(n+\frac12)\pi}\frac{|\sin s|}{s}\,ds.\end{aligned}$$ On $[(k-1)\pi,k\pi]$ we have $\frac1s\ge\frac1{k\pi}$ and $\int|\sin s|=2$, so the last integral is at least $\sum_{k=1}^n\frac2{k\pi}$, giving $L_n\ge\frac4{\pi^2}\sum_{k\le n}\frac1k\ge\frac4{\pi^2}\log(n+1)\to\infty$.
(c) $C(\mathbb{T})$ is a Banach space and $\sup_n\|\Lambda_n\|=\infty$. By Theorem 8.13 the family cannot be pointwise bounded: some $f$ has $\sup_n|S_nf(0)|=\infty$, so its Fourier series diverges at $0$; by Theorem 8.14(3) the set of such $f$ is residual. Numerically $L_1=\tfrac13+\tfrac{2\sqrt3}\pi\approx1.436$, $L_{10}\approx2.223$, $L_{100}\approx3.139$: the growth is logarithmic, $L_n\approx\frac4{\pi^2}\log n+1.27$, slow but unbounded.
Contrast The Cesàro means $\sigma_nf=\frac1{n+1}\sum_{m\le n}S_mf$ have the Fejér kernel $F_n=\frac1{n+1}\Big(\frac{\sin((n+1)t/2)}{\sin(t/2)}\Big)^2\ge0$ with $\frac1{2\pi}\int F_n=1$, so $\|\sigma_n\|=1$ for all $n$. They converge to $P$ uniformly for every trigonometric polynomial $P$ (each coefficient gets the factor $1-\frac{|k|}{n+1}\to1$), a dense set (§4.2); by Theorem 8.14(2) $\sigma_nf$ converges for every $f\in C(\mathbb{T})$, and the limit is $f$ (Fejér's theorem; Exercise II.4). Uniform bounds are the whole difference.
Interactive 1 — Dirichlet versus Fejér: Why One Family Is Unbounded
Top: the kernel $K_n$ on $[-\pi,\pi]$, with the area under $|K_n|$ shaded. The norm of the functional $f\mapsto\frac1{2\pi}\int fK_n$ on $C(\mathbb{T})$ is that area divided by $2\pi$. Bottom: these norms for $0\le k\le n$. For the Dirichlet kernel they grow like $\frac4{\pi^2}\log n$ (dashed curve), so by the uniform boundedness principle some continuous function has a divergent Fourier series. For the Fejér kernel they all equal $1$ — and every continuous function is recovered.
n—
norm = (1/2π)∫|K_n|—
(4/π²) log n + 1.27—
max K_n = K_n(0)—
The Open Mapping and Bounded Inverse Theorems
📘 Definition 8.8 — Open map
A map $T:X\to Y$ between metric spaces is open if $T(U)$ is open in $Y$ for every open $U\subseteq X$. For a linear map between normed spaces this holds if and only if $T(B_X)$ contains a ball $\delta B_Y$ around $0$, where $B_X=\{x:\|x\|\lt1\}$: then $T(x+rB_X)=Tx+rT(B_X)\supseteq B(Tx,r\delta)$. An open linear map is surjective ($T(X)$ contains a ball around $0$, hence all of $Y$ by scaling).
📐 Theorem 8.16 — Open mapping theorem (Banach–Schauder)
Let $X$ and $Y$ be Banach spaces and $T\in B(X,Y)$ surjective. Then $T$ is an open map: there is $\delta\gt0$ with $T(B_X)\supseteq\delta B_Y$.
Proof.Step 1 (Baire): $\overline{T(B_X)}$ contains a ball around $0$. Since $T$ is onto, $Y=\bigcup_nT(nB_X)=\bigcup_n\overline{T(nB_X)}$. $Y$ is complete, so by Theorem 8.12(b) some $\overline{T(nB_X)}=n\overline{T(B_X)}$ contains a ball, and hence (scaling) $\overline{T(B_X)}$ contains a ball $B(y_0,\varepsilon)$. The set $\overline{T(B_X)}$ is convex and symmetric ($-T(B_X)=T(B_X)$), so for $\|y\|\lt\varepsilon$,
$$y=\tfrac12(y_0+y)+\tfrac12(-y_0+y)\in\overline{T(B_X)},$$
because $y_0+y\in B(y_0,\varepsilon)$ and $-(y_0-y)\in-\overline{T(B_X)}=\overline{T(B_X)}$. Thus $\overline{T(B_X)}\supseteq2\delta B_Y$ with $2\delta=\varepsilon$, and by scaling $\overline{T(rB_X)}\supseteq2\delta rB_Y$ for every $r\gt0$.
Step 2 (completeness of $X$): remove the closure. Let $\|y\|\lt\delta$. By Step 1 with $r=\tfrac12$ there is $x_1\in\tfrac12B_X$ with $\|y-Tx_1\|\lt\delta/2$. Applying Step 1 with $r=\tfrac14$ to $y-Tx_1$ gives $x_2\in\tfrac14B_X$ with $\|y-Tx_1-Tx_2\|\lt\delta/4$, and inductively $x_k\in2^{-k}B_X$ with $\|y-T(x_1+\dots+x_k)\|\lt\delta2^{-k}$. Since $\sum\|x_k\|\lt\sum2^{-k}=1$ and $X$ is complete, $x=\sum_kx_k$ exists (Theorem 8.2) with $\|x\|\lt1$, and by continuity $Tx=\lim T(x_1+\dots+x_k)=y$. Hence $\delta B_Y\subseteq T(B_X)$, and $T$ is open by Definition 8.8. ∎
📐 Theorem 8.17 — Bounded inverse theorem and its corollaries
Bounded inverse theorem. If $X,Y$ are Banach spaces and $T\in B(X,Y)$ is a bijection, then $T^{-1}$ is bounded; equivalently there is $c\gt0$ with $\|Tx\|\ge c\|x\|$ for all $x$.
Comparable complete norms are equivalent. If $\|\cdot\|_a$ and $\|\cdot\|_b$ are norms on $X$, both complete, and $\|x\|_a\le C\|x\|_b$ for all $x$, then also $\|x\|_b\le C'\|x\|_a$.
Closed range. If $X,Y$ are Banach spaces and $T\in B(X,Y)$ is injective, then $T^{-1}:T(X)\to X$ is bounded if and only if the range $T(X)$ is closed in $Y$.
Proof. (1) $T$ is open by Theorem 8.16, i.e. $(T^{-1})^{-1}(U)=T(U)$ is open for every open $U$: $T^{-1}$ is continuous, hence bounded (Theorem 8.4), and $\|x\|=\|T^{-1}Tx\|\le\|T^{-1}\|\,\|Tx\|$ gives $c=1/\|T^{-1}\|$. (2) The identity $(X,\|\cdot\|_b)\to(X,\|\cdot\|_a)$ is a bounded bijection between Banach spaces; by (1) its inverse is bounded. (3) If $T(X)$ is closed it is a Banach space (Theorem 8.3(5)) and (1) applies to $T:X\to T(X)$. Conversely, if $\|Tx\|\ge c\|x\|$ and $Tx_n\to y$, then $\|x_n-x_m\|\le\tfrac1c\|Tx_n-Tx_m\|$, so $(x_n)$ is Cauchy, $x_n\to x$, and $y=\lim Tx_n=Tx\in T(X)$. ∎
✏️ Example 5 — Using the bounded inverse theorem backwards
(a) Show that $\|f\|_1\le\|f\|_\infty$ on $C[0,1]$ but that the two norms are not equivalent, and deduce (again) that $(C[0,1],\|\cdot\|_1)$ is not complete. (b) Show that $\|x\|_2\le\|x\|_1$ on $\ell^1$ but that the norms are not equivalent there; what follows about $(\ell^1,\|\cdot\|_2)$? (c) Does the bounded inverse theorem say that every bounded linear bijection between normed spaces has a bounded inverse?
💡 Show solution▶
(a) $\int_0^1|f|\le\max|f|$. Not equivalent: the tent $f_k$ of height $1$ on $[0,\tfrac2k]$ (peak at $\tfrac1k$) has $\|f_k\|_\infty=1$ and $\|f_k\|_1=\tfrac1k$, so no $C$ gives $\|f\|_\infty\le C\|f\|_1$. If $(C[0,1],\|\cdot\|_1)$ were complete, Theorem 8.17(2) — with $\|\cdot\|_\infty$ complete and $\|\cdot\|_1\le\|\cdot\|_\infty$ — would force equivalence. So it is not complete, without constructing a single non-convergent Cauchy sequence.
(b) $\|x\|_2\le\|x\|_1$ (Example 2, §8.1). For $x=e_1+\dots+e_n$: $\|x\|_1=n$, $\|x\|_2=\sqrt n$, so the ratio $\sqrt n$ is unbounded. Since $(\ell^1,\|\cdot\|_1)$ is complete, Theorem 8.17(2) shows that $(\ell^1,\|\cdot\|_2)$ is not complete — consistent with $\ell^1$ being a dense, proper subspace of $\ell^2$.
(c) No The identity $(C[0,1],\|\cdot\|_\infty)\to(C[0,1],\|\cdot\|_1)$ is a bounded linear bijection whose inverse is unbounded (part (a)). Completeness of both spaces is essential; here the target fails it.
✏️ Example 6 — An injective operator with dense, non-closed range
Let $T:\ell^2\to\ell^2$, $Tx=\big(x_1,\tfrac{x_2}2,\tfrac{x_3}3,\dots\big)$. Prove: $T$ is bounded with $\|T\|=1$, injective, with dense range; the range is not closed; $T$ is not surjective; and $T^{-1}$ (defined on the range) is unbounded. Which hypothesis of the bounded inverse theorem fails?
💡 Show solution▶
Bounded, injective $T=D_\lambda$ with $\lambda_n=\tfrac1n$, so $\|T\|=\sup\tfrac1n=1$ (Exercise II.3 of §8.1). $Tx=0$ forces every $x_n=0$.
Dense range $T(ne_n)=e_n$, so the range contains $c_{00}$, which is dense in $\ell^2$.
Not surjective $y=(\tfrac1n)\in\ell^2$ would need $x_n=ny_n=1$, and $(1,1,\dots)\notin\ell^2$. A dense subspace that is not everything is not closed.
Unbounded inverse $\|Te_n\|=\tfrac1n$ while $\|e_n\|=1$: no $c\gt0$ has $\|Tx\|\ge c\|x\|$. Both spaces are Banach and $T$ is bounded and injective; what fails is surjectivity — and, by Theorem 8.17(3), that is the same as the range not being closed. (Operators like this, with "small" singular values tending to $0$, are compact operators — §8.4 — and they are exactly why inverse problems such as deblurring are unstable.)
The Closed Graph Theorem
📘 Definition 8.9 — Graph and closed operator
The graph of a linear map $T:D\to Y$, defined on a subspace $D\subseteq X$, is $G(T)=\{(x,Tx):x\in D\}\subseteq X\times Y$, where $X\times Y$ carries the norm $\|(x,y)\|=\|x\|+\|y\|$ (complete if $X$ and $Y$ are). $T$ is closed if $G(T)$ is a closed subset, i.e.
$$x_n\in D,\quad x_n\to x,\quad Tx_n\to y$$ $$\Longrightarrow\quad x\in D\ \text{ and }\ Tx=y.$$
Every bounded $T$ defined on all of $X$ is closed (then $Tx_n\to Tx$, so $y=Tx$). The point of the definition is that one need not know in advance that $Tx_n$ converges.
📐 Theorem 8.18 — Closed graph theorem
Let $X$ and $Y$ be Banach spaces and $T:X\to Y$ a linear map defined on all of $X$. If $T$ is closed, then $T$ is bounded.
Proof. $G(T)$ is a closed subspace of the Banach space $X\times Y$, hence a Banach space (Theorem 8.3(5)). The projection $P:G(T)\to X$, $P(x,Tx)=x$, is linear, bounded ($\|x\|\le\|x\|+\|Tx\|$) and bijective (its inverse is $x\mapsto(x,Tx)$). By the bounded inverse theorem (Theorem 8.17(1)), $P^{-1}$ is bounded: $\|x\|+\|Tx\|\le C\|x\|$ for all $x$. In particular $\|Tx\|\le C\|x\|$. ∎
The norm $\|x\|+\|Tx\|$ is the graph norm. The proof shows: $T$ is closed iff $X$ is complete under the graph norm, and $T$ is bounded iff the graph norm is equivalent to the original norm.
📐 Theorem 8.19 — Two applications of the closed graph theorem
(a) Hellinger–Toeplitz. Let $\langle x,y\rangle=\sum_nx_n\overline{y_n}$ on $\ell^2$. If $T:\ell^2\to\ell^2$ is linear (no continuity assumed) and $\langle Tx,y\rangle=\langle x,Ty\rangle$ for all $x,y$, then $T$ is bounded. (b) Projections. Let $X$ be a Banach space and $X=M\oplus N$ (every $x$ is uniquely $m+n$, $m\in M$, $n\in N$) with $M,N$ closed subspaces. Then the projection $P(m+n)=m$ is bounded.
Proof. (a) Let $x_k\to x$ and $Tx_k\to z$. For every $y$, using the continuity of $\langle\cdot,y\rangle$ (Cauchy–Schwarz, the case $p=q=2$ of Hölder),
$$\begin{aligned}\langle z,y\rangle&=\lim_k\langle Tx_k,y\rangle=\lim_k\langle x_k,Ty\rangle\\&=\langle x,Ty\rangle=\langle Tx,y\rangle.\end{aligned}$$
With $y=z-Tx$ this gives $\|z-Tx\|^2=0$, so $z=Tx$: $T$ is closed, hence bounded (Theorem 8.18). (b) Let $x_k\to x$ and $Px_k\to m$. Since $Px_k\in M$ and $M$ is closed, $m\in M$; since $x_k-Px_k\in N$, $x_k-Px_k\to x-m$ and $N$ is closed, $x-m\in N$. So $x=m+(x-m)$ is the decomposition of $x$, and $Px=m$. $P$ is closed, hence bounded. ∎
✏️ Example 7 — Differentiation is closed but unbounded; when evaluations decide the norm
(a) Let $D=C^1[0,1]\subseteq C[0,1]$ and $Tf=f'$, with $\|\cdot\|_\infty$ on both sides. Show that $T$ is closed and unbounded, and explain why this does not contradict Theorem 8.18. (b) Let $\|\cdot\|$ be a norm on $C[0,1]$ under which $C[0,1]$ is a Banach space and every point evaluation $f\mapsto f(t)$ is continuous. Prove that $\|\cdot\|$ is equivalent to $\|\cdot\|_\infty$.
💡 Show solution▶
(a) Closed Let $f_n\in C^1$, $f_n\to f$ and $f_n'\to g$ uniformly. Passing to the limit in $f_n(x)=f_n(0)+\int_0^xf_n'$ gives $f(x)=f(0)+\int_0^xg$, so $f\in C^1$ and $f'=g$ (the theorem on uniform convergence of derivatives, §4.2).
(a) Unbounded $f_n(x)=\sin(n\pi x)/\sqrt n$: $\|f_n\|_\infty\le\tfrac1{\sqrt n}\to0$ while $\|f_n'\|_\infty=\pi\sqrt n\to\infty$.
(a) No contradiction The closed graph theorem needs $T$ defined on a Banach space. The domain $(C^1[0,1],\|\cdot\|_\infty)$ is not complete (its closure in $C[0,1]$ is $C[0,1]$). With the graph norm $\|f\|_\infty+\|f'\|_\infty$ the domain is complete (Example 1(d), §8.1) and $T$ is bounded — as the theorem predicts.
(b) Closed graph Consider the identity $I:(C[0,1],\|\cdot\|)\to(C[0,1],\|\cdot\|_\infty)$, a linear map between Banach spaces. If $\|f_n-f\|\to0$ and $\|f_n-g\|_\infty\to0$, then for each $t$, $f_n(t)\to f(t)$ (evaluation is $\|\cdot\|$-continuous) and $f_n(t)\to g(t)$; so $f=g$ and $I$ is closed. By Theorem 8.18, $\|f\|_\infty\le C\|f\|$; by Theorem 8.17(2) the norms are equivalent.
Moral On $C[0,1]$ there is essentially only one complete norm compatible with pointwise evaluation. The closed graph theorem often turns "the obvious map is well defined" into "the obvious map is bounded".
Interactive 2 — When Completeness Fails: Three Pictures
Each mode shows a sequence in $C[0,1]$ for which one of the big theorems would give the wrong answer if completeness were dropped. Spikes: $\|f_k\|_1\to0$ while $\|f_k\|_\infty=1$ — the identity $(C,\|\cdot\|_\infty)\to(C,\|\cdot\|_1)$ is a bounded bijection with unbounded inverse. Ramps: a $\|\cdot\|_1$-Cauchy sequence (shaded area $=\|f_k-f_{2k}\|_1$) heading for a step function, so $(C,\|\cdot\|_1)$ is incomplete. Waves: $f_k=\sin(k\pi x)/\sqrt k$ shrinks while its derivative (dashed, scaled down) grows — $d/dx$ is closed but unbounded on the incomplete space $(C^1,\|\cdot\|_\infty)$.
k—
——
——
ratio—
🧭 Problem-solving strategy — which big theorem?
You know…
You want…
Use (check the hypotheses!)
$\sup_\alpha\|T_\alpha x\|\lt\infty$ for each $x$
$\sup_\alpha\|T_\alpha\|\lt\infty$
UBP — domain Banach
$T_nx$ converges for each $x$
the limit is bounded
Banach–Steinhaus — domain Banach
$\sup\|T_n\|=\infty$
a "bad" $x$ exists
condensation (8.14(3)) — domain Banach
$T$ bounded, bijective
$T^{-1}$ bounded
bounded inverse — both Banach
$\|\cdot\|_a\le C\|\cdot\|_b$
equivalence, or incompleteness
8.17(2) — both complete
$T$ linear, everywhere defined, closed
$T$ bounded
closed graph — both Banach
To test closedness, assume both $x_n\to x$ and $Tx_n\to y$ and show $y=Tx$ — usually by applying a continuous functional (evaluation at a point, a coordinate, an inner product) to both limits.
⚠️ Counterexamples and pitfalls
UBP needs a complete domain. On $(c_{00},\|\cdot\|_\infty)$, $f_n(x)=nx_n$ is pointwise bounded with $\|f_n\|=n$; and $T_nx=(x_1,\dots,nx_n,0,\dots)$ converges pointwise to an unbounded $T$.
Pointwise is not norm convergence. Banach–Steinhaus gives a bounded limit, not $\|T_n-T\|\to0$: the truncations $P_n$ on $\ell^2$ satisfy $P_nx\to x$ but $\|P_n-I\|=1$.
Open mapping needs surjectivity and completeness of both spaces. $x\mapsto(x_n/n)$ on $\ell^2$ is bounded and injective with dense range but not open; the identity $(C,\|\cdot\|_\infty)\to(C,\|\cdot\|_1)$ is a bounded bijection that is not open.
A bounded injective operator can have an unbounded inverse — exactly when its range is not closed (Theorem 8.17(3)).
Closed graph needs an everywhere-defined map on a Banach space. $d/dx$ on $C^1[0,1]\subseteq C[0,1]$ is closed and unbounded; so is $x\mapsto(nx_n)$ on $\{x\in\ell^2:(nx_n)\in\ell^2\}$.
"Closed operator" is not "maps closed sets to closed sets" and not "has closed range": the bounded operator $x\mapsto(x_n/n)$ has closed graph but non-closed range.
Comparability matters. Two complete norms on the same space need not be equivalent if neither dominates the other (such pairs exist, via Hamel bases); Theorem 8.17(2) needs $\|\cdot\|_a\le C\|\cdot\|_b$.
Baire fails without completeness: $\mathbb{Q}=\bigcup_q\{q\}$ is a countable union of nowhere dense sets. And meagre is not the same as "measure zero": $\mathbb{R}$ is the union of a meagre set and a set of measure zero.
Divergence is generic. Fourier series of continuous functions can diverge at a point (indeed for a residual set of $f$), while their Cesàro means always converge uniformly — the difference is $\sup_n\|S_n\|=\infty$ versus $\|\sigma_n\|=1$.
In finite dimensions all three theorems are empty: every linear map is bounded and every norm complete. Exam options that cite them for matrices are distractors.
Exercises
I
Baire category and uniform boundedness.
1 Show that $\mathbb{Q}$ is meagre in $\mathbb{R}$ but not nowhere dense, and use the Baire category theorem to show that $\mathbb{R}\setminus\mathbb{Q}$ is not meagre.
2 Show that $c_{00}$ is a meagre subset of $\ell^2$.
3 Let $(x_n)$ be a sequence in a normed space $X$ with $f(x_n)\to0$ for every $f\in X^*$. Prove that $\sup_n\|x_n\|\lt\infty$. Show that $(e_n)$ has this property in $\ell^2$ and in $c_0$, but not in $\ell^1$.
4 Suppose $\sum_na_nx_n$ converges for every $x\in\ell^2$. Prove that $(a_n)\in\ell^2$.
5 Suppose $\sum_na_nx_n$ converges for every $x\in\ell^1$. Prove that $(a_n)$ is bounded, once with the UBP and once directly.
6 On $(c_{00},\|\cdot\|_\infty)$ find bounded functionals $f_n$ with $f_n(x)\to0$ for every $x$ but $\|f_n\|\to\infty$.
7 Let $X$ be a Banach space, $Y$ a normed space and $B:X\times Y\to\mathbb{F}$ bilinear and continuous in each variable separately. Prove that $|B(x,y)|\le C\|x\|\,\|y\|$ for some $C$, so $B$ is jointly continuous.
8 Let $X$ be a Banach space, $T_n,T\in B(X,Y)$ with $T_nx\to Tx$ for every $x$, and $x_n\to x$. Prove that $T_nx_n\to Tx$.
💡 Show solutions▶
1 $\mathbb{Q}=\bigcup_q\{q\}$ is a countable union of closed sets with empty interior. It is not nowhere dense: its closure is $\mathbb{R}$. If $\mathbb{R}\setminus\mathbb{Q}$ were also meagre, $\mathbb{R}$ would be a countable union of nowhere dense sets, contradicting Theorem 8.12 ($\mathbb{R}$ is complete).
2 $c_{00}=\bigcup_nF_n$ with $F_n=\operatorname{span}\{e_1,\dots,e_n\}$, a finite-dimensional (so closed) proper subspace of $\ell^2$, hence nowhere dense.
3 For each $f$, $(f(x_n))$ converges, so it is bounded; Theorem 8.15(a) gives $\sup\|x_n\|\lt\infty$. In $\ell^2$, $f=\varphi_y$ with $y\in\ell^2$ and $f(e_n)=y_n\to0$; in $c_0$, $y\in\ell^1$ and again $y_n\to0$. In $\ell^1$, $f=\varphi_{\mathbf 1}$ ($\mathbf 1\in\ell^\infty=(\ell^1)^*$) gives $f(e_n)=1$ for all $n$.
4 Example 3(b) with $p=q=2$: the functionals $\phi_N(x)=\sum_{n\le N}a_nx_n$ have $\|\phi_N\|=\big(\sum_{n\le N}|a_n|^2\big)^{1/2}$ and are pointwise bounded on the Banach space $\ell^2$, so these partial sums are bounded.
5 UBP: $\|\phi_N\|_{(\ell^1)^*}=\max_{n\le N}|a_n|$, and pointwise boundedness gives $\sup_N\max_{n\le N}|a_n|\lt\infty$. Directly: if $(a_n)$ were unbounded, choose $n_1\lt n_2\lt\cdots$ with $|a_{n_k}|\ge k^2$ and put $x_{n_k}=k^{-2}\overline{\operatorname{sgn}a_{n_k}}$, $x_n=0$ otherwise; then $x\in\ell^1$ but infinitely many terms $a_nx_n$ have modulus $\ge1$, so the series diverges.
6 $f_n(x)=nx_n$: for fixed $x\in c_{00}$, $f_n(x)=0$ once $n$ exceeds the support of $x$, yet $\|f_n\|=n$. (Impossible on a Banach space by the UBP.)
7 For $\|y\|\le1$ let $T_y=B(\cdot,y)\in X^*$. For fixed $x$, $y\mapsto B(x,y)$ is a bounded functional on $Y$, so $\sup_{\|y\|\le1}|T_y(x)|\lt\infty$. By the UBP on the Banach space $X$, $C:=\sup_{\|y\|\le1}\|T_y\|\lt\infty$, so $|B(x,y)|\le C\|x\|$ when $\|y\|\le1$; scale $y$. Joint continuity: $B(x,y)-B(x_0,y_0)$ $=B(x-x_0,y)+B(x_0,y-y_0)$.
8 $M=\sup\|T_n\|\lt\infty$ by Theorem 8.14(1). Then $\|T_nx_n-Tx\|\le M\|x_n-x\|$ $+\,\|T_nx-Tx\|\to0$.
II
Fourier series and summability.
1 Compute the Lebesgue constants $L_0$ and $L_1$ exactly.
2 Show that $\max_tD_n(t)=D_n(0)=2n+1$ and $\frac1{2\pi}\int_{-\pi}^{\pi}D_n=1$.
3 Show that $F_n=\frac1{n+1}\sum_{m=0}^nD_m$ equals $\frac1{n+1}\Big(\frac{\sin((n+1)t/2)}{\sin(t/2)}\Big)^2$, so $F_n\ge0$ and $\frac1{2\pi}\int F_n=1$.
4 Complete the proof of Fejér's theorem: $\sigma_nf\to f$ uniformly for every $f\in C(\mathbb{T})$.
5 Use $L_n\ge\frac4{\pi^2}\sum_{k\le n}\frac1k$ to bound $L_{10}$ from below, and compare with $L_{10}\approx2.223$.
6 On the space $\mathcal{T}$ of trigonometric polynomials with the sup norm, is the family $\Lambda_n(P)=S_nP(0)$ pointwise bounded? Why does this not contradict $\|\Lambda_n\|=L_n\to\infty$?
💡 Show solutions▶
1 $D_0=1$, so $L_0=1$. $D_1=1+2\cos t\ge0$ exactly for $|t|\le\tfrac{2\pi}3$. With $\int_{-\pi}^{\pi}D_1=2\pi$ and $\int_{2\pi/3}^{\pi}(1+2\cos t)\,dt=\tfrac\pi3-\sqrt3$, $$\int_{-\pi}^{\pi}|D_1|=2\pi-4\Big(\tfrac\pi3-\sqrt3\Big)=\tfrac{2\pi}3+4\sqrt3,$$ so $L_1=\tfrac13+\tfrac{2\sqrt3}\pi\approx1.4360$ (checked numerically).
2 $|D_n(t)|=\big|\sum_{|k|\le n}e^{ikt}\big|\le2n+1$ with equality at $t=0$. Integrating term by term, only $k=0$ survives: $\frac1{2\pi}\int D_n=1$.
3 $\sin\tfrac t2\,D_m(t)=\sin\big((m+\tfrac12)t\big)$, and $2\sin\tfrac t2\sin\big((m+\tfrac12)t\big)$ $=\cos(mt)-\cos((m+1)t)$. Summing over $m\le n$ telescopes to $1-\cos((n+1)t)=2\sin^2\tfrac{(n+1)t}2$. Divide by $2\sin^2\tfrac t2$ and by $n+1$. Positivity is visible; the integral is the average of the integrals of $D_m$, all equal to $2\pi$.
4 $\sigma_nf=\frac1{2\pi}\int f(t)F_n(\cdot-t)\,dt$, so $\|\sigma_nf\|_\infty\le\|f\|_\infty$: $\|\sigma_n\|\le1$. For a trigonometric polynomial $P=\sum_{|k|\le d}c_ke^{ikx}$, $\sigma_nP=\sum_{|k|\le d}\big(1-\tfrac{|k|}{n+1}\big)c_ke^{ikx}\to P$ uniformly. Trigonometric polynomials are dense in $C(\mathbb{T})$ (Stone–Weierstrass, §4.2). By Theorem 8.14(2) (with $Y=C(\mathbb{T})$ complete), $\sigma_nf$ converges for every $f$; the limit operator $S$ is bounded and equals $I$ on a dense set, so $S=I$.
5 $H_{10}=1+\tfrac12+\dots+\tfrac1{10}\approx2.929$, so $L_{10}\ge\tfrac4{\pi^2}\cdot2.929\approx1.187$. The true value $2.223$ is about $1.04$ larger — the bound has the right growth ($\frac4{\pi^2}\log n$) but not the right constant ($\approx1.27$).
6 Yes: for $P$ of degree $d$, $S_nP=P$ for all $n\ge d$, so $\sup_n|S_nP(0)|\lt\infty$. There is no contradiction because $(\mathcal{T},\|\cdot\|_\infty)$ is not complete (it is dense in $C(\mathbb{T})$, with a countable Hamel basis); the UBP only applies on the completion, where divergence appears.
III
Open mapping and bounded inverse.
1 For $A=\begin{pmatrix}1&1\\0&1\end{pmatrix}$ on $(\mathbb{R}^2,\|\cdot\|_2)$, find the largest $r$ such that $A$ maps the closed unit disc onto a set containing the closed disc of radius $r$.
2 Show that $\|\cdot\|_1$ and $\|\cdot\|_\infty$ are not equivalent on $c_{00}$. Is $c_{00}$ complete under either?
3 Let $X,Y$ be Banach spaces and $T\in B(X,Y)$ surjective. Prove there is $C$ such that every $y\in Y$ equals $Tx$ for some $x$ with $\|x\|\le C\|y\|$.
4 Let $T:\ell^1\to\ell^1$, $Tx=(x_n/n)$. Show that $T$ is bounded and injective with dense range, but not surjective.
5 Let $(Vf)(x)=\int_0^xf$ on $C[0,1]$. Find the range of $V$, show that it is not closed, and conclude that $V^{-1}$ is unbounded on it.
6 Let $X$ be a Banach space and $T\in B(X)$ with $\|Tx\|\ge c\|x\|$ ($c\gt0$) and dense range. Prove that $T$ is invertible in $B(X)$.
7 Let $\|\cdot\|$ be a norm on $\ell^1$ under which $\ell^1$ is complete and $\|x\|_1\le\|x\|$ for all $x$. Prove that $\|\cdot\|$ is equivalent to $\|\cdot\|_1$.
8 Let $X$ be a Banach space and $P\in B(X)$ with $P^2=P$. Prove that $P(X)$ and $\ker P$ are closed and $X=P(X)\oplus\ker P$.
💡 Show solutions▶
1 $A$ maps the unit disc onto an ellipse with half-axes the singular values $\sigma_1\ge\sigma_2$; the largest inscribed disc has radius $\sigma_2=1/\|A^{-1}\|_2$. $A^TA=\begin{pmatrix}1&1\\1&2\end{pmatrix}$ has eigenvalues $\tfrac{3\pm\sqrt5}2$, so $\sigma_2=\sqrt{\tfrac{3-\sqrt5}2}=\tfrac{\sqrt5-1}2\approx0.618$.
2 $x=e_1+\dots+e_n$: $\|x\|_1=n$, $\|x\|_\infty=1$. Neither norm is complete on $c_{00}$ (Example 1 — or Theorem 8.3(2), §8.1); so Theorem 8.17(2) gives no information here, and none is needed.
3 By Theorem 8.16, $T(B_X)\supseteq\delta B_Y$. For $y\ne0$, $y'=\tfrac{\delta}{2\|y\|}y\in\delta B_Y$, so $y'=Tx'$ with $\|x'\|\lt1$; then $x=\tfrac{2\|y\|}{\delta}x'$ has $Tx=y$ and $\|x\|\le\tfrac2\delta\|y\|$. ($C=2/\delta$.)
4 $\|Tx\|_1\le\|x\|_1$; injective; $T(ne_n)=e_n$, so the range contains $c_{00}$, dense in $\ell^1$. $y=(1/n^2)\in\ell^1$ is not in the range: it would need $x=(1/n)\notin\ell^1$. By Theorem 8.17(3) the range is not closed and $T^{-1}$ is unbounded ($\|Te_n\|_1=\tfrac1n$).
5 $V(C[0,1])=\{g\in C^1[0,1]:g(0)=0\}$ (fundamental theorem of calculus). Its closure in $\|\cdot\|_\infty$ is $\{g\in C[0,1]:g(0)=0\}$ — e.g. $|x-\tfrac12|-\tfrac12$ is a uniform limit of such $g$ but is not differentiable — so the range is not closed, and $V^{-1}=d/dx$ is unbounded on it (Example 7(a)). $V$ is injective because $Vf=0$ gives $f=(Vf)'=0$.
6 The range is closed (proof of Theorem 8.17(3)) and dense, hence all of $X$. So $T$ is a bounded bijection and Theorem 8.17(1) gives $T^{-1}\in B(X)$ with $\|T^{-1}\|\le1/c$.
7 Both $\|\cdot\|$ and $\|\cdot\|_1$ are complete and $\|x\|_1\le\|x\|$; Theorem 8.17(2) gives $\|x\|\le C'\|x\|_1$.
8 $P(X)=\ker(I-P)$ (if $y=Px$ then $Py=P^2x=y$) and $\ker P$ are kernels of bounded operators, hence closed. $x=Px+(x-Px)$ with $P(x-Px)=0$, and $P(X)\cap\ker P=\{0\}$ since $y=Py=0$. (Theorem 8.19(b) is the converse.)
IV
Closed graphs.
1 Prove that every $T\in B(X,Y)$ has a closed graph, and give a bounded operator on $\ell^2$ whose graph is closed but whose range is not closed.
2 Let $X,Y$ be Banach spaces and $T:X\to Y$ linear with $g\circ T\in X^*$ for every $g\in Y^*$. Prove that $T$ is bounded.
3 Let $(a_{ij})_{i,j\ge1}$ be an infinite matrix such that for every $x\in\ell^2$ each series $(Ax)_i=\sum_ja_{ij}x_j$ converges and $Ax\in\ell^2$. Prove that $A$ is a bounded operator on $\ell^2$.
4 Suppose $(a_nx_n)\in\ell^1$ for every $x\in\ell^1$. Use the closed graph theorem to show that $M_a:x\mapsto(a_nx_n)$ is bounded on $\ell^1$, and find $\|M_a\|$.
5 On $D(T)=\{x\in\ell^2:(nx_n)\in\ell^2\}$ let $Tx=(nx_n)$. Show that $T$ is closed, $D(T)$ is dense and $T$ is unbounded. Why is there no contradiction with Theorem 8.18?
6 In $X=C[-1,1]$ let $M$ and $N$ be the even and the odd functions. Show $X=M\oplus N$ with $M,N$ closed, write down the projection $P$ onto $M$ along $N$, and find $\|P\|$.
7 Let $\|\cdot\|_a$ and $\|\cdot\|_b$ be complete norms on $X$ such that $\|x_n\|_a\to0$ and $\|x_n-y\|_b\to0$ imply $y=0$. Prove that the norms are equivalent.
💡 Show solutions▶
1 If $x_n\to x$ and $Tx_n\to y$, continuity gives $Tx_n\to Tx$, so $y=Tx$. Example: $x\mapsto(x_n/n)$ (Example 6).
2 Closed graph: if $x_n\to x$ and $Tx_n\to y$, then for every $g\in Y^*$, $g(y)=\lim g(Tx_n)=(g\circ T)(x)=g(Tx)$, because $g$ and $g\circ T$ are continuous. Functionals separate points (Theorem 8.10(1)), so $y=Tx$. Theorem 8.18 applies.
3 Each row $(a_{ij})_j$ lies in $\ell^2$ (Exercise I.4), so the coordinate maps $x\mapsto(Ax)_i$ are bounded functionals. If $x_n\to x$ and $Ax_n\to y$ in $\ell^2$, then for each $i$, $y_i=\lim(Ax_n)_i=(Ax)_i$ (coordinates are continuous on $\ell^2$). So $y=Ax$: $A$ is closed, hence bounded.
4 If $x^{(k)}\to x$ and $M_ax^{(k)}\to y$ in $\ell^1$, then coordinatewise $y_n=\lim a_nx^{(k)}_n=a_nx_n$, so $y=M_ax$. By Theorem 8.18 $M_a$ is bounded; $\|M_ae_n\|_1=|a_n|$ shows $|a_n|\le\|M_a\|$, and conversely $\|M_ax\|_1\le\|a\|_\infty\|x\|_1$. So $\|M_a\|=\|a\|_\infty$.
5 If $x^{(k)}\to x$ and $Tx^{(k)}\to y$ in $\ell^2$, then coordinatewise $y_n=\lim nx^{(k)}_n=nx_n$, so $(nx_n)=y\in\ell^2$: $x\in D(T)$ and $Tx=y$. $D(T)\supseteq c_{00}$ is dense. $\|Te_n\|=n$. No contradiction: $T$ is not defined on all of $\ell^2$ (e.g. $(1/n)\notin D(T)$), and $D(T)$ with the $\ell^2$ norm is not complete.
6 $f(x)=\tfrac{f(x)+f(-x)}2+\tfrac{f(x)-f(-x)}2$, even plus odd, uniquely (a function both even and odd is $0$). $M=\ker(f\mapsto f-f(-\cdot))$ and $N$ similarly are closed. $Pf(x)=\tfrac{f(x)+f(-x)}2$, $\|Pf\|_\infty\le\|f\|_\infty$, equality for $f\equiv1$: $\|P\|=1$.
7 The identity $I:(X,\|\cdot\|_a)\to(X,\|\cdot\|_b)$ is closed: if $x_n\to x$ in $\|\cdot\|_a$ and $x_n\to y$ in $\|\cdot\|_b$, apply the hypothesis to $x_n-x$ (which tends to $0$ in $\|\cdot\|_a$ and to $y-x$ in $\|\cdot\|_b$): $y-x=0$. By Theorem 8.18, $\|x\|_b\le C\|x\|_a$, and Theorem 8.17(2) gives equivalence.
🎯 GATE corner GATE MA · MCQGATE MA · MSQGATE MA · NAT
E1.GATE MA · MCQ Let $X$ be the space of real sequences with only finitely many nonzero terms, with $\|x\|_\infty=\max_n|x_n|$, and let $f_n(x)=nx_n$. Which one of the following is true? (A) $(f_n)$ is not pointwise bounded on $X$. (B) $(f_n)$ is pointwise bounded and $\sup_n\|f_n\|\lt\infty$. (C) $(f_n)$ is pointwise bounded, $\|f_n\|=n$, and this is consistent with the uniform boundedness principle because $X$ is not complete. (D) $\|f_n\|=1$ for every $n$.
💡 Show solution▶
Idea For fixed $x$ the values $f_n(x)$ are eventually $0$; on unit vectors $|f_n(e_n)|=n$ (Example 2).
Options (A) false; (B) false, $\|f_n\|=n$; (D) false; (C) true — the UBP needs a Banach domain.
Answer (C).
E2.GATE MA · MCQ Let $\|\cdot\|_a$ and $\|\cdot\|_b$ be two norms on a real vector space $X$ such that $X$ is a Banach space under each of them and $\|x\|_a\le3\|x\|_b$ for all $x\in X$. Then necessarily (A) $\|x\|_b\le3\|x\|_a$ for all $x$; (B) there is $C\gt0$ with $\|x\|_b\le C\|x\|_a$ for all $x$; (C) $\|x\|_a=\|x\|_b$ for all $x$; (D) $X$ is finite-dimensional.
💡 Show solution▶
Idea Theorem 8.17(2): comparable complete norms are equivalent, but nothing controls the constant.
Distractors (A) fails for $\|x\|_b=10\|x\|_a$ on $\mathbb{R}$ (then $\|x\|_a\le3\|x\|_b$ holds, but $\|x\|_b\le3\|x\|_a$ does not); this also kills (C). (D): $X=\ell^2$ with $\|\cdot\|_a=\|\cdot\|_b=\|\cdot\|_2$.
Answer (B).
E3.GATE MA · MSQCSIR Part C Which of the following linear maps are necessarily bounded? (One or more options may be correct.) (A) A linear $T:\ell^2\to\ell^2$ with $\langle Tx,y\rangle=\langle x,Ty\rangle$ for all $x,y\in\ell^2$. (B) $f\mapsto f'$ from $(C^1[0,1],\|\cdot\|_\infty)$ to $(C[0,1],\|\cdot\|_\infty)$. (C) A linear $T:\ell^1\to\ell^1$, defined on all of $\ell^1$, whose graph is closed. (D) A linear functional on $\ell^2$ whose kernel is closed.
💡 Show solution▶
(A) Bounded Hellinger–Toeplitz (Theorem 8.19(a)).
(B) Unbounded $x^n\mapsto nx^{n-1}$ (Example 7(a)); the domain is not complete.
(C) Bounded Closed graph theorem: $\ell^1$ is a Banach space and $T$ is everywhere defined.
(D) Bounded A functional is bounded iff its kernel is closed (Exercise II.7, §8.1).
Answer (A), (C) and (D).
E4.GATE MA · MSQ Let $X,Y$ be Banach spaces and $T\in B(X,Y)$. Which of the following statements are true? (One or more options may be correct.) (A) If $T$ is surjective, then $T$ maps open subsets of $X$ to open subsets of $Y$. (B) If $T$ is injective, then $T^{-1}:T(X)\to X$ is bounded. (C) If $T$ is bijective, then there is $c\gt0$ with $\|Tx\|\ge c\|x\|$ for all $x$. (D) If $T$ has dense range and $\|Tx\|\ge c\|x\|$ for some $c\gt0$ and all $x$, then $T$ is surjective.
💡 Show solution▶
(A) True Open mapping theorem (Theorem 8.16).
(B) False $x\mapsto(x_n/n)$ on $\ell^2$ is injective with unbounded inverse on its (non-closed) range (Example 6).
(D) True The lower bound makes the range closed (Theorem 8.17(3)); closed and dense means everything.
Answer (A), (C) and (D).
E5.GATE MA · NAT On $C[-\pi,\pi]$ with the sup norm, let $\Lambda(f)=\dfrac1{2\pi}\displaystyle\int_{-\pi}^{\pi}f(t)\,(1+2\cos t+2\cos2t)\,dt$. The value of $\|\Lambda\|$, correct to two decimal places, is ______.
💡 Show solution▶
Idea $\|\Lambda\|=\frac1{2\pi}\int_{-\pi}^{\pi}|D_2|$ with $D_2=1+2\cos t+2\cos2t=\dfrac{\sin(5t/2)}{\sin(t/2)}$ — the Lebesgue constant $L_2$.
Sign changes On $(0,\pi)$, $D_2$ vanishes at $t=\tfrac{2\pi}5,\tfrac{4\pi}5$: positive, negative, positive. An antiderivative is $G(t)=t+2\sin t+\sin2t$.
Compute With $a=\tfrac{2\pi}5$, $b=\tfrac{4\pi}5$: $G(a)\approx3.7465$, $G(b)\approx2.7378$, $G(\pi)=\pi$. Then $$\int_0^\pi|D_2|=2G(a)-2G(b)+G(\pi)\approx5.1590,$$ and by evenness $\|\Lambda\|=\frac1\pi\int_0^\pi|D_2|\approx1.6422$ (checked numerically).
Answer $1.64$ (accept $1.63$ to $1.65$). Using $f\equiv1$ would give only $1$, and $\max|D_2|=5$ is the norm of $D_2$ in $C$, not of the functional.
E6.GATE MA · NAT Let $A=\begin{pmatrix}3&1\\1&1\end{pmatrix}$ act on $\mathbb{R}^2$ with the Euclidean norm. The largest $r$ such that the image of the closed unit disc under $A$ contains the closed disc of radius $r$ centred at $0$, correct to two decimal places, is ______.
💡 Show solution▶
Idea The quantitative open mapping constant: $A(\overline B)\supseteq r\overline B$ iff $\|A^{-1}y\|\le1$ whenever $\|y\|\le r$, i.e. $r\le1/\|A^{-1}\|$ — the smallest singular value.
Compute $A$ is symmetric with trace $4$ and determinant $2$, eigenvalues $2\pm\sqrt2$; its singular values are $|2\pm\sqrt2|$, the smallest $2-\sqrt2\approx0.5858$.
Answer $0.59$ (accept $0.58$ to $0.59$).
🧑🏫 For the teacher
Suggested plan — 3 lectures × 60 min.
Baire and the UBP. Definition 8.7 and Theorem 8.12 (a ten-minute recap with the nested-balls picture); Example 1 as the first "Baire proves something" moment. Theorem 8.13 in full, Theorem 8.14 (prove (1) and (2); state (3)), Theorem 8.15. Example 2 — ask the class which hypothesis fails before revealing it.
UBP at work. Examples 3 and 4; project Interactive 1, first with the Dirichlet kernel (let the class guess whether $L_n$ stays bounded), then switch to Fejér. Prove $L_n\ge\frac4{\pi^2}H_n$ on the board. Close with GATE corner E5 as a calculator exercise.
Open mapping and closed graph. Theorem 8.16 (Step 1 with a picture of the convex symmetric set, Step 2 as "successive approximation"), Theorem 8.17 and Examples 5–6; Interactive 2 in "Spikes" mode. Theorem 8.18 via the graph norm, Theorem 8.19, Example 7 with Interactive 2 in "Waves" mode. Finish with E3 and E4 live.
Discussion questions. (1) The UBP proves that a continuous function with divergent Fourier series exists, but does not exhibit one. Is that a satisfactory proof? What would a constructive one need? (2) Why must the open mapping theorem fail for a bounded bijection onto a non-complete space — which step of the proof breaks? (3) The closed graph theorem says "closed + everywhere defined on a Banach space ⇒ bounded". Physicists' operators (momentum, energy) are unbounded; what does the theorem say about their domains?
Misconceptions to watch for. "Pointwise bounded implies uniformly bounded" without completeness; "Banach–Steinhaus gives norm convergence"; "an injective bounded operator has a bounded inverse"; "closed operator means closed range"; applying the closed graph theorem to $d/dx$ on $(C^1,\|\cdot\|_\infty)$; forgetting that Theorem 8.17(2) needs the two norms to be comparable.
🔬 Research–Teaching NexusResearch-tutored · optional, about 60 minutes
Research question. Polynomial interpolation through $n+1$ nodes defines a bounded projection $I_n$ on $C[-1,1]$. How fast does $\|I_n\|$ (the interpolation Lebesgue constant) grow for equally spaced nodes and for Chebyshev nodes — and what does the uniform boundedness principle then say about interpolating every continuous function?
Why it matters. Interpolation turns samples into curves: reconstructing temperature profiles from weather stations, trajectories from GPS fixes, or shapes in computer-aided design. Since $\|f-I_nf\|_\infty\le(1+\|I_n\|)\operatorname{dist}(f,\mathcal{P}_n)$, a small Lebesgue constant means interpolation is nearly as good as the best polynomial approximation, while a large one means tiny measurement errors are amplified enormously (Runge's phenomenon).
What to do.
Show that $\|I_n\|=\max_{x\in[-1,1]}\sum_j|\ell_j(x)|$, where $\ell_j$ are the Lagrange basis polynomials (the argument is the "smoothed sign" idea of §8.1, now with point evaluations).
Compute this number on a fine grid for $n=5,10,20$ with equally spaced nodes and with Chebyshev nodes $\cos\frac{(2j+1)\pi}{2n+2}$. Tabulate and plot $\log\|I_n\|$ against $n$ and against $\log n$.
Compare the Chebyshev values with $\frac2\pi\log(n+1)+1$. Interpolate $f(x)=\frac1{1+25x^2}$ at both node families and plot the errors.
Use the UBP (as in Example 4) to prove Faber's theorem in the form: for any fixed triangular array of nodes, $\sup_n\|I_n\|=\infty$ (take this growth as known), so some continuous $f$ has $\|I_nf\|_\infty$ unbounded.
Tools: Python with NumPy (or MATLAB/Octave); a plotting tool.
Hand in: A one-page report: the table, the two plots, and a paragraph explaining to an engineer why "more data points" can make an interpolating polynomial worse, and why Chebyshev nodes help but cannot help for every continuous function.
Going further. Read about Bernstein's and Faber's results on the impossibility of a universally convergent interpolation scheme, and about why splines (piecewise polynomials) avoid the problem.
Instructor note: numerically, equispaced nodes give $\|I_5\|\approx3.106$, $\|I_{10}\|\approx29.90$, $\|I_{20}\|\approx1.099\times10^4$ (exponential growth, roughly $2^{n+1}/(en\log n)$), while Chebyshev nodes give $2.104$, $2.489$, $2.901$, just below $\frac2\pi\log(n+1)+1\approx2.141,\ 2.527,\ 2.938$. For every node system $\|I_n\|\ge\frac2\pi\log(n+1)-C$ for an absolute constant $C$ (results of Faber, Bernstein and Erdős), so logarithmic growth is the best possible and Chebyshev nodes are nearly optimal.
🧪 Self-check quiz
0/5 correct
Q1. The uniform boundedness principle for a family $\{T_\alpha\}\subseteq B(X,Y)$ requires
Q2. Which of these operators is closed but unbounded?
Q3. Let $X,Y$ be Banach spaces and $T\in B(X,Y)$ bijective. Then
Q4. The Lebesgue constants $L_n=\frac1{2\pi}\int|D_n|$ tend to $\infty$. By the uniform boundedness principle this shows that
Q5. On $C[0,1]$, the norms $\|f\|_\infty$ and $\|f\|_1=\int_0^1|f|$ satisfy $\|f\|_1\le\|f\|_\infty$. What follows?