Chapter 8 · Functional Analysis

8.1 Normed and Banach Spaces; Bounded Operators; Hahn–Banach

Spaces of sequences and functions with a norm: $\ell^p$, $c_0$, $C[a,b]$ and $L^p$ and why they are complete; bounded operators and the operator norm; the duals $(\ell^p)^*=\ell^q$ and $c_0^*=\ell^1$; and the Hahn–Banach theorem, which guarantees that a normed space has enough bounded linear functionals.

Linear algebra studies linear maps between finite-dimensional spaces; functional analysis studies linear maps between spaces of functions — the integral operator $f\mapsto\int_0^1K(\cdot,t)f(t)\,dt$, the differentiation operator $f\mapsto f'$, the map that sends a function to its sequence of Fourier coefficients. To speak of continuity, convergence and approximation for such maps we measure the size of a function by a norm. Two questions then dominate. Is the space complete, so that the limits we construct by iteration, by series or by approximation actually exist inside it? And is a given linear map bounded, so that small inputs produce small outputs? In finite dimensions both answers are always "yes"; in infinite dimensions they depend on the norm, and getting them right is most of the work.

The section ends with the first of the great theorems of the subject, the Hahn–Banach theorem. It says that a bounded linear functional defined on a subspace can always be extended to the whole space without increasing its norm, and, geometrically, that a convex set can be separated from a point outside it by a hyperplane. Its consequences — every normed space has many bounded functionals, $\|x\|=\max_{\|f\|\le1}|f(x)|$, closed subspaces are cut out by functionals — are used everywhere in analysis, optimisation and the theory of PDEs. GATE asks for the norms of concrete operators and functionals (often as NAT questions), for the completeness of concrete spaces, and for the precise statements of Hahn–Banach and its corollaries in MSQ form. Norms, $C[a,b]$ and equivalent norms are also developed, with further examples, in CSIR Unit 1 · §7.4; here we go further, to $L^p$, dual spaces and Hahn–Banach.

Normed Spaces and Banach Spaces

Throughout, $\mathbb{F}$ denotes $\mathbb{R}$ or $\mathbb{C}$, and all vector spaces are over $\mathbb{F}$. For $1\le p\le\infty$ the conjugate exponent $q$ is defined by $\tfrac1p+\tfrac1q=1$, with $q=\infty$ when $p=1$ and $q=1$ when $p=\infty$; for instance $p=2\leftrightarrow q=2$, $p=3\leftrightarrow q=\tfrac32$, $p=4\leftrightarrow q=\tfrac43$.

📘 Definition 8.1 — Norm, normed space, Banach space

A norm on a vector space $X$ is a function $\|\cdot\|:X\to[0,\infty)$ such that, for all $x,y\in X$ and $\lambda\in\mathbb{F}$,

Definiteness$\|x\|=0\iff x=0$.
Homogeneity$\|\lambda x\|=|\lambda|\,\|x\|$.
Triangle inequality$\|x+y\|\le\|x\|+\|y\|$.

A normed space is a vector space with a norm; it is a metric space with $d(x,y)=\|x-y\|$, and $x_n\to x$ means $\|x_n-x\|\to0$. A normed space that is complete (every Cauchy sequence converges) is a Banach space. From the triangle inequality, $\big|\|x\|-\|y\|\big|\le\|x-y\|$, so the norm is continuous, and so are addition and scalar multiplication.

📘 Definition 8.2 — The standard spaces
Sequence spacesFor $1\le p\lt\infty$, $\ell^p=\{x=(x_n)_{n\ge1}:\sum|x_n|^p\lt\infty\}$ with $\|x\|_p=\big(\sum_n|x_n|^p\big)^{1/p}$; $\ell^\infty$ = bounded sequences with $\|x\|_\infty=\sup_n|x_n|$.
Subspaces of $\ell^\infty$$c$ = convergent sequences, $c_0$ = sequences tending to $0$, $c_{00}$ = sequences with only finitely many nonzero terms, all with $\|\cdot\|_\infty$.
Continuous functions$C[a,b]$ with the sup norm $\|f\|_\infty=\max_{[a,b]}|f|$ (also written $\|f\|$); other norms on it, such as $\|f\|_1=\int_a^b|f|$, will be named explicitly.
Lebesgue spacesFor $1\le p\lt\infty$, $L^p[a,b]$ = Lebesgue measurable $f$ with $\int_a^b|f|^p\lt\infty$, $\|f\|_p=\big(\int_a^b|f|^p\big)^{1/p}$; $L^\infty[a,b]$ = essentially bounded $f$ with $\|f\|_\infty=\operatorname{ess\,sup}|f|$.

In $L^p$ two functions that are equal almost everywhere are identified (an element of $L^p$ is an equivalence class); otherwise $\|f\|_p=0$ would not force $f=0$. The unit vectors are $e_n=(0,\dots,0,1,0,\dots)$ with the $1$ in place $n$. The same definitions with $n\le N$ give the finite-dimensional spaces $\ell^p_N=(\mathbb{F}^N,\|\cdot\|_p)$.

📐 Theorem 8.1 — Hölder and Minkowski inequalities

Let $1\le p\le\infty$ with conjugate exponent $q$.

  1. Hölder: for sequences, $\sum_n|x_ny_n|\le\|x\|_p\|y\|_q$; for measurable functions on $[a,b]$, $\int_a^b|fg|\le\|f\|_p\|g\|_q$.
  2. Minkowski: $\|x+y\|_p\le\|x\|_p+\|y\|_p$, and $\|f+g\|_p\le\|f\|_p+\|g\|_p$.

Hence $\|\cdot\|_p$ is a norm on $\ell^p$ and on $L^p[a,b]$, and these are vector spaces.

Proof. The cases $p=1$ or $p=\infty$ follow from $|x_ny_n|\le|x_n|\,\|y\|_\infty$ and $|x_n+y_n|\le|x_n|+|y_n|$. Let $1\lt p\lt\infty$. Young's inequality: for $a,b\ge0$, $ab\le\frac{a^p}{p}+\frac{b^q}{q}$. This is clear if $ab=0$; otherwise, since $\log$ is concave and $\tfrac1p+\tfrac1q=1$, $$\log\Big(\tfrac1pa^p+\tfrac1qb^q\Big)\ge\tfrac1p\log a^p+\tfrac1q\log b^q,$$ and the right-hand side is $\log(ab)$. Hölder: if $\|x\|_p=0$ or $\|y\|_q=0$ there is nothing to prove. Otherwise apply Young to $a=|x_n|/\|x\|_p$, $b=|y_n|/\|y\|_q$ and sum over $n$: $$\frac{\sum|x_ny_n|}{\|x\|_p\|y\|_q}\le\frac1p\cdot\frac{\sum|x_n|^p}{\|x\|_p^p}+\frac1q\cdot\frac{\sum|y_n|^q}{\|y\|_q^q},$$ and the right-hand side equals $\tfrac1p+\tfrac1q=1$. Minkowski: first, $x+y\in\ell^p$, because $|x_n+y_n|^p\le\big(2\max(|x_n|,|y_n|)\big)^p$ and this is at most $2^p\big(|x_n|^p+|y_n|^p\big)$. Write $$\begin{aligned}|x_n+y_n|^p&\le|x_n|\,|x_n+y_n|^{p-1}\\&\quad+|y_n|\,|x_n+y_n|^{p-1}\end{aligned}$$ and apply Hölder to each sum; since $(p-1)q=p$, $$\begin{aligned}\|x+y\|_p^p&\le\big(\|x\|_p+\|y\|_p\big)\Big(\sum|x_n+y_n|^{p}\Big)^{1/q}\\&=\big(\|x\|_p+\|y\|_p\big)\|x+y\|_p^{p/q}.\end{aligned}$$ If $\|x+y\|_p\ne0$, divide by $\|x+y\|_p^{p/q}$ and use $p-\tfrac pq=1$. The function versions are proved word for word with sums replaced by integrals. ∎

📐 Theorem 8.2 — Completeness via absolutely convergent series
A normed space $X$ is a Banach space if and only if every absolutely convergent series converges: whenever $\sum_k\|x_k\|\lt\infty$, the partial sums $s_N=\sum_{k\le N}x_k$ converge in $X$.

Proof. ($\Rightarrow$) For $M\gt N$, $\|s_M-s_N\|\le\sum_{k=N+1}^M\|x_k\|$, which is small for large $N$; so $(s_N)$ is Cauchy and converges. ($\Leftarrow$) Let $(y_n)$ be Cauchy. Choose $n_1\lt n_2\lt\cdots$ with $\|y_m-y_n\|\lt2^{-k}$ for all $m,n\ge n_k$. The series $y_{n_1}+\sum_k(y_{n_{k+1}}-y_{n_k})$ is absolutely convergent ($\sum2^{-k}\lt\infty$), so it converges; its partial sums are $y_{n_{k+1}}$. Thus the subsequence $(y_{n_k})$ converges to some $y$, and a Cauchy sequence with a convergent subsequence converges to the same limit: $\|y_n-y\|\le\|y_n-y_{n_k}\|+\|y_{n_k}-y\|$. ∎

📐 Theorem 8.3 — The standard Banach spaces
  1. $\ell^p$ ($1\le p\le\infty$) is a Banach space. $c$ and $c_0$ are closed subspaces of $\ell^\infty$, hence Banach spaces.
  2. $c_{00}$ is not complete in any of the norms $\|\cdot\|_p$; its closure is $\ell^p$ in $\ell^p$ ($p\lt\infty$) and $c_0$ in $\ell^\infty$.
  3. $(C[a,b],\|\cdot\|_\infty)$ is a Banach space.
  4. Riesz–Fischer: $L^p[a,b]$ ($1\le p\lt\infty$) is a Banach space; so is $L^\infty[a,b]$.
  5. A subspace of a Banach space is a Banach space (with the restricted norm) if and only if it is closed.

Proof. (5) A complete subset of a metric space is closed (a limit of points of the subset is the limit of a Cauchy sequence in it), and a closed subset of a complete space is complete (its Cauchy sequences converge in the big space, and the limit stays in the closed subset).

(1) Let $p\lt\infty$ and let $(x^{(k)})_k$ be Cauchy in $\ell^p$. Since $|x^{(k)}_n-x^{(m)}_n|\le\|x^{(k)}-x^{(m)}\|_p$, each coordinate sequence is Cauchy in $\mathbb{F}$; let $x_n=\lim_kx^{(k)}_n$. Given $\varepsilon\gt0$ choose $K$ with $\|x^{(k)}-x^{(m)}\|_p\lt\varepsilon$ for $k,m\ge K$. For each fixed $N$, $$\sum_{n\le N}\big|x^{(k)}_n-x^{(m)}_n\big|^p\lt\varepsilon^p\qquad(k,m\ge K).$$ Let $m\to\infty$ (a finite sum), then $N\to\infty$: $\|x^{(k)}-x\|_p\le\varepsilon$ for $k\ge K$. So $x^{(K)}-x\in\ell^p$, hence $x=x^{(K)}-(x^{(K)}-x)\in\ell^p$, and $x^{(k)}\to x$. For $\ell^\infty$ the same argument works with "$\sup_{n\le N}$" in place of the sum. If $x^{(k)}\in c_0$ and $x^{(k)}\to x$ in $\ell^\infty$, pick $k$ with $\|x-x^{(k)}\|_\infty\lt\varepsilon$; then $|x_n|\le\varepsilon+|x^{(k)}_n|\lt2\varepsilon$ for large $n$, so $x\in c_0$. Similarly, if all $x^{(k)}\in c$, then for large $m,n$ we have $|x_m-x_n|\le2\varepsilon+|x^{(k)}_m-x^{(k)}_n|\lt3\varepsilon$, so $(x_n)$ is Cauchy and $x\in c$.

(2) The truncations $x^{(k)}=(1,\tfrac12,\dots,\tfrac1k,0,0,\dots)\in c_{00}$ converge in $\|\cdot\|_\infty$ to $(\tfrac1n)\notin c_{00}$, since $\|x^{(k)}-(\tfrac1n)\|_\infty=\tfrac1{k+1}$; in $\ell^p$ use $(n^{-2})$ instead. So $c_{00}$ is not closed in the Banach space, hence not complete by (5). For $x\in\ell^p$ ($p\lt\infty$), the truncations $P_kx=(x_1,\dots,x_k,0,\dots)$ satisfy $\|x-P_kx\|_p^p=\sum_{n\gt k}|x_n|^p\to0$ (tail of a convergent series), so $c_{00}$ is dense in $\ell^p$; for $x\in c_0$, $\|x-P_kx\|_\infty=\sup_{n\gt k}|x_n|\to0$. In $\ell^\infty$ the closure of $c_{00}$ is $c_0$, because $c_0$ is closed.

(3) A Cauchy sequence in $\|\cdot\|_\infty$ is uniformly Cauchy, so it converges uniformly; a uniform limit of continuous functions is continuous (§4.2).

(4) We use Theorem 8.2. Let $f_k\in L^p$ with $\sum_k\|f_k\|_p=S\lt\infty$, and put $g_N=\sum_{k\le N}|f_k|$. By Minkowski $\|g_N\|_p\le S$. The functions $g_N$ increase to $g=\sum_k|f_k|$ (possibly $+\infty$), and the monotone convergence theorem (§4.3) gives $\int g^p=\lim\int g_N^p\le S^p$. Hence $g\lt\infty$ almost everywhere, so the series $\sum f_k(x)$ converges absolutely for almost every $x$; call its sum $f(x)$ (and $f=0$ elsewhere). Then $|f|\le g$, so $f\in L^p$, and with $s_N=\sum_{k\le N}f_k$ we have $|f-s_N|^p\le(2g)^p$, an integrable bound, while $|f-s_N|^p\to0$ a.e. By dominated convergence $\|f-s_N\|_p\to0$. For $L^\infty$: if $(f_k)$ is Cauchy, discard the null set where some $|f_k-f_m|\gt\|f_k-f_m\|_\infty$; off it the convergence is uniform. ∎

✏️ Example 1 — Complete or not? Prove or disprove
Decide which of the following normed spaces are complete: (a) $C[0,1]$ with $\|f\|_1=\int_0^1|f|$; (b) the space $\mathcal{P}$ of polynomials on $[0,1]$ with $\|\cdot\|_\infty$; (c) $\ell^1$ with the norm $\|\cdot\|_\infty$; (d) $C^1[0,1]$ with $\|f\|_{C^1}=\|f\|_\infty+\|f'\|_\infty$.
(a) Not complete Let $f_n=0$ on $[0,\tfrac12]$, $f_n(x)=n(x-\tfrac12)$ on $[\tfrac12,\tfrac12+\tfrac1n]$ and $f_n=1$ afterwards (ramps, $n\ge2$). For $m\gt n$, $f_m$ and $f_n$ agree outside $[\tfrac12,\tfrac12+\tfrac1n]$ and differ by at most $1$ there, so $\|f_m-f_n\|_1\le\tfrac1n$: the sequence is Cauchy. Suppose $\|f_n-f\|_1\to0$ with $f$ continuous. Then $\int_0^{1/2}|f|\le\|f_n-f\|_1\to0$, so $f=0$ on $[0,\tfrac12]$; and for each $\delta\gt0$, $f_n=1$ on $[\tfrac12+\delta,1]$ for large $n$, so $f=1$ there. A continuous $f$ cannot jump from $0$ to $1$ at $\tfrac12$. (In $L^1[0,1]$ the ramps do converge, to the step function $\mathbf 1_{(1/2,1]}$: $L^1$ is the completion.)
(b) Not complete The Taylor polynomials $p_n(x)=\sum_{k\le n}x^k/k!$ converge uniformly on $[0,1]$ to $e^x$, so they are Cauchy in $\|\cdot\|_\infty$; their only possible limit in $C[0,1]$ is $e^x$, which is not a polynomial (a polynomial of degree $d$ has vanishing $(d+1)$-st derivative). So $\mathcal{P}$ is not closed in $C[0,1]$; by Weierstrass its closure is all of $C[0,1]$. By Theorem 8.3(5) it is not complete.
(c) Not complete $\ell^1\subseteq c_0$, and $c_{00}\subseteq\ell^1$ is already dense in $(c_0,\|\cdot\|_\infty)$, so the closure of $\ell^1$ in $\ell^\infty$ is $c_0\ne\ell^1$ (e.g. $(\tfrac1n)\in c_0\setminus\ell^1$). A non-closed subspace of the Banach space $\ell^\infty$ is not complete.
(d) Complete Let $(f_n)$ be Cauchy in $\|\cdot\|_{C^1}$. Then $(f_n)$ and $(f_n')$ are uniformly Cauchy, so $f_n\to f$ and $f_n'\to g$ uniformly, with $f,g$ continuous. Pass to the limit in $f_n(x)=f_n(0)+\int_0^xf_n'(t)\,dt$ (uniform convergence lets us exchange limit and integral): $f(x)=f(0)+\int_0^xg(t)\,dt$. By the fundamental theorem of calculus $f\in C^1$ and $f'=g$, so $\|f_n-f\|_{C^1}=\|f_n-f\|_\infty+\|f_n'-f'\|_\infty\to0$.
Moral Completeness is a property of the norm, not of the set: $C[0,1]$ is complete under $\|\cdot\|_\infty$ but not under $\|\cdot\|_1$, and $C^1[0,1]$ is complete under $\|\cdot\|_{C^1}$ but not under $\|\cdot\|_\infty$ (its closure in $C[0,1]$ is $C[0,1]$).
✏️ Example 2 — How the $\ell^p$ and $L^p$ spaces sit inside each other
(a) For which $p$ is $x=(n^{-\alpha})_{n\ge1}$ in $\ell^p$ ($\alpha\gt0$)? (b) Prove that $\ell^p\subseteq\ell^r$ with $\|x\|_r\le\|x\|_p$ when $1\le p\lt r\le\infty$. (c) Prove that $L^r[0,1]\subseteq L^p[0,1]$ with $\|f\|_p\le\|f\|_r$ when $1\le p\lt r\lt\infty$, and decide for which $p$ the function $f(x)=x^{-1/2}$ lies in $L^p[0,1]$. (d) Show that on $(0,\infty)$ neither of $L^1$, $L^2$ contains the other.
(a) $\sum n^{-\alpha p}\lt\infty\iff\alpha p\gt1$ ($p$-series). So $x\in\ell^p\iff p\gt1/\alpha$; and $x\in\ell^\infty$ (indeed $c_0$) for every $\alpha\gt0$. For $\alpha=\tfrac12$: $x\in\ell^p$ exactly for $p\gt2$.
(b) The claim is homogeneous, so assume $\|x\|_p=1$. Then $|x_n|\le1$ for all $n$, so $|x_n|^r\le|x_n|^p$ and $\sum|x_n|^r\le1$, i.e. $\|x\|_r\le1$. For $r=\infty$, $|x_n|\le\|x\|_p$ is immediate. Sequences get more permissive as $p$ grows: $\ell^1\subsetneq\ell^2\subsetneq\cdots\subsetneq c_0\subsetneq\ell^\infty$.
(c) Apply Hölder with the exponents $s=r/p\gt1$ and its conjugate $s'$ to $|f|^p\cdot1$: $$\int_0^1|f|^p\le\Big(\int_0^1|f|^{r}\Big)^{p/r}\Big(\int_0^11\Big)^{1/s'}=\|f\|_r^p.$$ So on an interval of length $1$, functions get less permissive as $p$ grows. For $f=x^{-1/2}$: $\int_0^1x^{-p/2}dx\lt\infty\iff p/2\lt1$, so $f\in L^p[0,1]$ exactly for $p\lt2$. (On an interval of length $\ell$ the constant is $\ell^{1/p-1/r}$.)
(d) $f=x^{-1/2}\mathbf 1_{(0,1)}$ has $\int f=2$ but $\int f^2=\int_0^1\tfrac{dx}x=\infty$: in $L^1\setminus L^2$. $g=x^{-1}\mathbf 1_{(1,\infty)}$ has $\int g^2=1$ but $\int g=\infty$: in $L^2\setminus L^1$. Singularities are punished by large $p$, slow decay at infinity by small $p$; sequences have no singularities, so only (b) survives.

Bounded Linear Operators and the Operator Norm

📘 Definition 8.3 — Bounded operator, operator norm, dual space

Let $X,Y$ be normed spaces. A linear map $T:X\to Y$ is bounded if there is a constant $C$ with $\|Tx\|\le C\|x\|$ for all $x\in X$. Its operator norm is

$$\|T\|=\sup_{\|x\|\le1}\|Tx\|.$$

$B(X,Y)$ is the vector space of bounded linear maps $X\to Y$, and $B(X)=B(X,X)$. A linear map $X\to\mathbb{F}$ is a linear functional; the dual space of $X$ is $X^*=B(X,\mathbb{F})$, with the norm $\|f\|=\sup_{\|x\|\le1}|f(x)|$.

📐 Theorem 8.4 — Bounded = continuous
For a linear map $T:X\to Y$ between normed spaces the following are equivalent: (i) $T$ is continuous at $0$; (ii) $T$ is continuous; (iii) $T$ is Lipschitz; (iv) $T$ is bounded. When they hold, $\|Tx\|\le\|T\|\,\|x\|$ for all $x$, and for $X\ne\{0\}$ $$\|T\|=\sup_{\|x\|=1}\|Tx\|=\sup_{x\ne0}\frac{\|Tx\|}{\|x\|},$$ and $\|T\|$ is the least constant $C\ge0$ with $\|Tx\|\le C\|x\|$ for all $x$.

Proof. (iv)$\Rightarrow$(iii): $\|Tx-Ty\|=\|T(x-y)\|\le C\|x-y\|$. (iii)$\Rightarrow$(ii)$\Rightarrow$(i) is clear. (i)$\Rightarrow$(iv): continuity at $0$ with $\varepsilon=1$ gives $\delta\gt0$ with $\|Tz\|\le1$ whenever $\|z\|\le\delta$. For $x\ne0$ put $z=\delta x/\|x\|$; by linearity $\|Tx\|=\tfrac{\|x\|}{\delta}\|Tz\|\le\tfrac1\delta\|x\|$. For the formulas: if $x\ne0$ then $x/\|x\|$ is a unit vector, so $\|Tx\|/\|x\|=\|T(x/\|x\|)\|$; this shows $\sup_{x\ne0}\|Tx\|/\|x\|=\sup_{\|x\|=1}\|Tx\|=:s$, and $\|Tx\|\le s\|x\|$, so $s$ is the least admissible $C$ (any admissible $C$ satisfies $C\ge s$). Finally $\sup_{\|x\|\le1}\|Tx\|=s$: the unit sphere is inside the ball, and for $0\lt\|x\|\le1$, $\|Tx\|=\|x\|\,\|T(x/\|x\|)\|\le s$. ∎

📐 Theorem 8.5 — $B(X,Y)$ is a Banach space when $Y$ is
$(B(X,Y),\|\cdot\|)$ is a normed space and $\|ST\|\le\|S\|\,\|T\|$ whenever the composition makes sense. If $Y$ is a Banach space then so is $B(X,Y)$ — whether or not $X$ is complete. In particular the dual $X^*$ of every normed space is a Banach space.

Proof. The norm axioms: $\|T\|=0$ means $Tx=0$ for all $\|x\|\le1$, hence for all $x$; $\|(\lambda T)x\|=|\lambda|\|Tx\|$; and $\|(S+T)x\|\le\|Sx\|+\|Tx\|\le\|S\|+\|T\|$ for $\|x\|\le1$. Also $\|STx\|\le\|S\|\,\|Tx\|\le\|S\|\,\|T\|\,\|x\|$.

Completeness: let $(T_n)$ be Cauchy in $B(X,Y)$. For each $x$, $\|T_nx-T_mx\|\le\|T_n-T_m\|\,\|x\|$, so $(T_nx)$ is Cauchy in $Y$; let $Tx=\lim T_nx$. $T$ is linear (limits respect linear combinations). Given $\varepsilon\gt0$ choose $N$ with $\|T_n-T_m\|\lt\varepsilon$ for $n,m\ge N$; then $\|T_nx-T_mx\|\le\varepsilon\|x\|$, and letting $m\to\infty$ (the norm is continuous) gives $\|T_nx-Tx\|\le\varepsilon\|x\|$ for $n\ge N$. Thus $T_n-T$ is bounded, so $T=T_n-(T_n-T)\in B(X,Y)$, and $\|T_n-T\|\le\varepsilon$ for $n\ge N$. The last statement is the case $Y=\mathbb{F}$, which is complete. ∎

📐 Theorem 8.6 — Finite versus infinite dimension
(a) If $\dim X\lt\infty$, all norms on $X$ are equivalent, $X$ is complete, and every linear map from $X$ to any normed space is bounded; consequently every finite-dimensional subspace of a normed space is closed. (b) If $\dim X=\infty$, there is a linear functional on $X$ that is not bounded.

Proof. (a) Fix a basis $e_1,\dots,e_n$ and write $x=\sum a_ie_i$, $\|x\|_1:=\sum|a_i|$. The equivalence of all norms with $\|\cdot\|_1$ (proved by the compactness of the $\|\cdot\|_1$-unit sphere; see CSIR Unit 1 · §7.4) gives completeness, since $(\mathbb{F}^n,\|\cdot\|_1)$ is complete. For linear $T$, $\|Tx\|\le\sum|a_i|\,\|Te_i\|$, which is at most $\big(\max_i\|Te_i\|\big)\|x\|_1\le C\|x\|$. A finite-dimensional subspace is complete in the restricted norm, hence closed (Theorem 8.3(5)). (b) Choose linearly independent unit vectors $u_1,u_2,\dots$ and extend them to a Hamel basis $\{u_i\}\cup\{v_j\}$ of $X$ (Zorn's lemma). Define $f$ on the basis by $f(u_n)=n$, $f(v_j)=0$, and extend linearly. Then $|f(u_n)|=n$ with $\|u_n\|=1$, so $f$ is unbounded. ∎

Part (b) is an existence statement — nobody can write such an $f$ down on, say, $\ell^2$. In practice, the unbounded operators one meets are defined on non-complete spaces or on dense subspaces: differentiation on $(C^1[0,1],\|\cdot\|_\infty)$, or $x\mapsto\sum_nnx_n$ on $(c_{00},\|\cdot\|_\infty)$. Section 8.2 explains why a "natural" unbounded operator can never be defined everywhere on a Banach space with a closed graph.

✏️ Example 3 — Norms of a matrix
Let $A=\begin{pmatrix}1&2\\0&2\end{pmatrix}$, acting on $\mathbb{R}^2$. Find $\|A\|$ when $\mathbb{R}^2$ carries (a) $\|\cdot\|_1$, (b) $\|\cdot\|_\infty$, (c) $\|\cdot\|_2$, proving the general formula in each case, and find a unit vector at which the norm is attained.
(a) Column sums For an $m\times n$ matrix, $\|Ax\|_1=\sum_i\big|\sum_ja_{ij}x_j\big|\le\sum_j|x_j|\sum_i|a_{ij}|$, which is at most $\big(\max_j\sum_i|a_{ij}|\big)\|x\|_1$, with equality at $x=e_{j_0}$ for a column $j_0$ of largest absolute sum. Here the column sums are $1$ and $4$: $\|A\|_{1}=4$, attained at $e_2$ ($Ae_2=(2,2)$, $\|Ae_2\|_1=4$).
(b) Row sums $|(Ax)_i|\le\sum_j|a_{ij}|\,|x_j|\le\big(\sum_j|a_{ij}|\big)\|x\|_\infty$, so $\|A\|_\infty\le\max_i\sum_j|a_{ij}|$; equality at $x_j=\overline{\operatorname{sgn}a_{i_0j}}$ (a vector of $\|\cdot\|_\infty$-norm $1$) for the largest row $i_0$. Row sums $3$ and $2$: $\|A\|_\infty=3$, attained at $x=(1,1)$, $Ax=(3,2)$.
(c) Spectral norm $\|Ax\|_2^2=\langle A^TAx,x\rangle$, and for the symmetric matrix $A^TA$ the maximum of $\langle A^TAx,x\rangle$ over unit vectors is its largest eigenvalue (Rayleigh quotient, §3.3). Here $A^TA=\begin{pmatrix}1&2\\2&8\end{pmatrix}$ with trace $9$ and determinant $4$, eigenvalues $\tfrac{9\pm\sqrt{65}}2$, so $$\|A\|_2=\sqrt{\tfrac{9+\sqrt{65}}{2}}\approx2.9208,$$ attained at the corresponding unit eigenvector of $A^TA$, approximately $(0.2567,\,0.9665)$.
Check The three norms differ, as they must — they are norms on different spaces of operators. All three dominate the spectral radius: the eigenvalues of $A$ are $1$ and $2$, and $2\le2.92\le3\le4$. (Values checked numerically.)
✏️ Example 4 — Integral functionals and integral operators on $C[a,b]$
(a) Let $g\in C[a,b]$ and $\varphi_g(f)=\int_a^bg(t)f(t)\,dt$ on $(C[a,b],\|\cdot\|_\infty)$. Prove $\|\varphi_g\|=\int_a^b|g|$. (b) For $g(t)=2t-1$ on $[0,1]$, find $\|\varphi_g\|$ and show that it is not attained. (c) Let $K$ be continuous on $[a,b]^2$ and $(Tf)(x)=\int_a^bK(x,t)f(t)\,dt$. Prove $\|T\|=\max_x\int_a^b|K(x,t)|\,dt$, and compute it for $K(x,t)=x+t$ on $[0,1]^2$. (d) Is differentiation $D f=f'$ bounded from $(C^1[0,1],\|\cdot\|_\infty)$ to $(C[0,1],\|\cdot\|_\infty)$?
(a) Upper bound $|\varphi_g(f)|\le\int|g|\,|f|\le\|f\|_\infty\int|g|$, so $\|\varphi_g\|\le\int_a^b|g|$.
(a) Lower bound The ideal test function $\overline{\operatorname{sgn}g}$ is discontinuous, so smooth it: $f_n=\dfrac{\overline{g}}{|g|+1/n}$ is continuous with $|f_n|\lt1$, and $$\varphi_g(f_n)=\int_a^b\frac{|g|^2}{|g|+1/n}\longrightarrow\int_a^b|g|$$ by monotone convergence (the integrands increase to $|g|$). Hence $\|\varphi_g\|=\int_a^b|g|$ — the $L^1$ norm of $g$.
(b) Not attained $\|\varphi_g\|=\int_0^1|2t-1|\,dt=\tfrac12$. Suppose $\|f\|_\infty\le1$ and $|\varphi_g(f)|=\tfrac12$; replacing $f$ by $cf$ with $|c|=1$ we may assume $\varphi_g(f)=\tfrac12$. Then $\int_0^1\big(|2t-1|-(2t-1)\operatorname{Re}f(t)\big)dt=0$ with a continuous non-negative integrand, so $(2t-1)\operatorname{Re}f(t)=|2t-1|$ for all $t$: $\operatorname{Re}f=1$ on $(\tfrac12,1]$ and $\operatorname{Re}f=-1$ on $[0,\tfrac12)$ — impossible for a continuous $f$. The supremum is approached, e.g. by $f_n=\max(-1,\min(1,n(2t-1)))$, but never reached. (On $L^\infty[0,1]$ the same functional attains its norm at the step function $\operatorname{sgn}(2t-1)$.)
(c) Operators $|Tf(x)|\le\|f\|_\infty\int|K(x,t)|\,dt$, so $\|T\|\le M:=\max_x\int|K(x,t)|\,dt$ (the maximum exists because $x\mapsto\int|K(x,t)|\,dt$ is continuous on $[a,b]$). If $x_0$ attains $M$, the functional $f\mapsto Tf(x_0)$ is $\varphi_g$ with $g=K(x_0,\cdot)$, of norm $M$ by (a); since $|Tf(x_0)|\le\|Tf\|_\infty$, $\|T\|\ge M$. For $K=x+t\ge0$: $\int_0^1(x+t)\,dt=x+\tfrac12$, maximal at $x=1$, so $\|T\|=\tfrac32$, attained at $f\equiv1$ ($Tf=x+\tfrac12$).
(d) Unbounded $f_n(x)=x^n$ has $\|f_n\|_\infty=1$ and $\|f_n'\|_\infty=\max nx^{n-1}=n$. No constant $C$ works, so $D$ is unbounded (discontinuous: $\tfrac1{\sqrt n}x^n\to0$ but $\|D(\tfrac1{\sqrt n}x^n)\|_\infty=\sqrt n\to\infty$). With the norm $\|\cdot\|_{C^1}$ of Example 1(d) on the domain, $\|Df\|_\infty\le\|f\|_{C^1}$, and $D$ is bounded with norm $1$.
Interactive 1 — The Operator Norm as the Largest Stretch of the Unit Sphere

Left: a unit vector $x$ travels round the unit sphere of $\|\cdot\|_p$ on $\mathbb{R}^2$. Right: its image $Ax$ traces the curve $A(S)$, and the dashed curve is the smallest $p$-ball that contains it — its radius is $\|A\|_p=\sup_{\|x\|_p=1}\|Ax\|_p$. Watch where the maximum is reached: at a vertex $\pm e_j$ for $p=1$, at a corner $(\pm1,\pm1)$ for $p=\infty$, along a singular vector for $p=2$.

x (unit vector)—
‖Ax‖_p—
largest seen so far—
‖A‖_p (formula)—

Dual Spaces of the Classical Spaces

For a concrete space we want to describe $X^*$: every $y$ in a suitable sequence space defines a functional on $\ell^p$ by "pairing", and the question is whether all functionals arise this way.

📐 Theorem 8.7 — The duals of $\ell^p$, $c_0$ and $L^p$

Let $1\le p\lt\infty$ with conjugate exponent $q$. For a sequence $y$ put $\varphi_y(x)=\sum_nx_ny_n$.

  1. $y\mapsto\varphi_y$ is an isometric isomorphism of $\ell^q$ onto $(\ell^p)^*$: every bounded functional on $\ell^p$ is $\varphi_y$ for exactly one $y\in\ell^q$, and $\|\varphi_y\|=\|y\|_q$. In short, $(\ell^p)^*=\ell^q$; in particular $(\ell^1)^*=\ell^\infty$ and $(\ell^2)^*=\ell^2$.
  2. The same map is an isometric isomorphism of $\ell^1$ onto $c_0^*$: $\ c_0^*=\ell^1$.
  3. (Statement.) $(L^p[a,b])^*=L^q[a,b]$ for $1\le p\lt\infty$, every functional being $f\mapsto\int_a^bfg$ for a unique $g\in L^q$, with norm $\|g\|_q$.
  4. $y\mapsto\varphi_y$ maps $\ell^1$ isometrically into $(\ell^\infty)^*$, but not onto (Example 7); likewise $L^1\subsetneq(L^\infty)^*$.

Proof. (1), case $1\lt p\lt\infty$. By Hölder, $|\varphi_y(x)|\le\|x\|_p\|y\|_q$, so the series converges and $\|\varphi_y\|\le\|y\|_q$. For the reverse inequality fix $N$ and let $x^{(N)}_n=|y_n|^{q-1}\,\overline{\operatorname{sgn}y_n}$ for $n\le N$ and $0$ for $n\gt N$, where $\operatorname{sgn}z=z/|z|$ ($\operatorname{sgn}0=0$). Since $(q-1)p=q$, $$\varphi_y(x^{(N)})=\sum_{n\le N}|y_n|^q=:S_N,\qquad\|x^{(N)}\|_p=S_N^{1/p}.$$ So $\|\varphi_y\|\ge S_N^{1-1/p}=S_N^{1/q}$ whenever $S_N\gt0$; letting $N\to\infty$ gives $\|\varphi_y\|\ge\|y\|_q$. Onto: let $\varphi\in(\ell^p)^*$ and put $y_n=\varphi(e_n)$. For $x\in\ell^p$ the truncations $\sum_{n\le N}x_ne_n$ converge to $x$ in $\ell^p$ (Theorem 8.3(2) — here $p\lt\infty$ is essential), so by continuity and linearity $\varphi(x)=\lim_N\sum_{n\le N}x_ny_n=\sum_nx_ny_n$. It remains to see $y\in\ell^q$: with $x^{(N)}$ as above, $S_N=\varphi(x^{(N)})\le\|\varphi\|\,S_N^{1/p}$, so $S_N^{1/q}\le\|\varphi\|$ for every $N$. Uniqueness: $y_n=\varphi_y(e_n)$ is determined by $\varphi_y$.

Case $p=1$, $q=\infty$: $|\varphi_y(x)|\le\|y\|_\infty\|x\|_1$, and $|\varphi_y(e_n)|=|y_n|$ with $\|e_n\|_1=1$ gives $\|\varphi_y\|=\|y\|_\infty$. For $\varphi\in(\ell^1)^*$, $y_n=\varphi(e_n)$ satisfies $|y_n|\le\|\varphi\|$, and $\varphi=\varphi_y$ by the same truncation argument.

(2) For $y\in\ell^1$, $|\varphi_y(x)|\le\|x\|_\infty\|y\|_1$. The vector $x^{(N)}=(\overline{\operatorname{sgn}y_1},\dots,\overline{\operatorname{sgn}y_N},0,\dots)\in c_0$ has norm at most $1$ and $\varphi_y(x^{(N)})=\sum_{n\le N}|y_n|$, so $\|\varphi_y\|=\|y\|_1$. Given $\varphi\in c_0^*$, set $y_n=\varphi(e_n)$; then $\sum_{n\le N}|y_n|=\varphi(x^{(N)})\le\|\varphi\|$, so $y\in\ell^1$, and since truncations converge in $c_0$, $\varphi=\varphi_y$. (3) needs the Radon–Nikodym theorem and is beyond our scope; (4) is proved in Example 7. ∎

Notice where $p\lt\infty$ entered: the vectors $e_n$ span a dense subspace of $\ell^p$ and of $c_0$, so a bounded functional is determined by its values on them. In $\ell^\infty$ they do not: the constant sequence $(1,1,1,\dots)$ is at distance $1$ from every finitely supported sequence.

✏️ Example 5 — Norms of functionals on sequence spaces
Let $\varphi(x)=\sum_{n\ge1}x_n/2^n$. Find $\|\varphi\|$ as a functional on (a) $\ell^1$, (b) $\ell^2$, (c) $\ell^{3/2}$, (d) $c_0$, (e) $\ell^\infty$, and say in each case whether the norm is attained. (f) Find the norm of $\psi(x)=x_1+x_2$ on $\ell^p$.
Idea $\varphi=\varphi_y$ with $y=(2^{-n})$, and by Theorem 8.7 its norm on $\ell^p$ is $\|y\|_q$, on $c_0$ it is $\|y\|_1$.
(a) $\ell^1$ $\|y\|_\infty=\tfrac12$, attained at $x=e_1$.
(b) $\ell^2$ $\|y\|_2=\big(\sum4^{-n}\big)^{1/2}=\big(\tfrac{1/4}{1-1/4}\big)^{1/2}=\tfrac1{\sqrt3}\approx0.5774$, attained at $x=y/\|y\|_2$ (equality in Cauchy–Schwarz).
(c) $\ell^{3/2}$ $q=3$: $\|y\|_3=\big(\sum8^{-n}\big)^{1/3}=\big(\tfrac17\big)^{1/3}\approx0.5228$, attained at $x_n=|y_n|^{q-1}=4^{-n}$ (normalised), the extremal vector of the proof.
(d) $c_0$ $\|y\|_1=\sum2^{-n}=1$, not attained: if $\|x\|_\infty\le1$ and $\sum x_n2^{-n}=1=\sum2^{-n}$, then $\sum(1-\operatorname{Re}x_n)2^{-n}=0$ with non-negative terms, so $x_n=1$ for all $n$ — not in $c_0$.
(e) $\ell^\infty$ Also $1$ (the same two estimates), now attained at $x=(1,1,1,\dots)\in\ell^\infty$.
(f) $\psi=\varphi_y$ with $y=(1,1,0,0,\dots)$: $\|\psi\|=\|y\|_q=2^{1/q}=2^{1-1/p}$. So it is $1$ on $\ell^1$, $\sqrt2$ on $\ell^2$, $2$ on $\ell^\infty$ (where $y\in\ell^1$ still gives $\|\varphi_y\|=\|y\|_1$).

The Hahn–Banach Theorem

Does a normed space have any nonzero bounded functionals at all? For $\ell^p$ and $C[a,b]$ we can write them down, but for an abstract space we need a theorem. The Hahn–Banach theorem builds functionals by extending them one dimension at a time, controlled by a "sublinear" bound.

📘 Definition 8.4 — Sublinear functionals and seminorms
A function $p:X\to\mathbb{R}$ on a real vector space is sublinear if $p(x+y)\le p(x)+p(y)$ and $p(tx)=t\,p(x)$ for all $x,y\in X$ and $t\ge0$. A seminorm on a real or complex space is a function $p\ge0$ with $p(x+y)\le p(x)+p(y)$ and $p(\lambda x)=|\lambda|\,p(x)$ for all scalars $\lambda$. Examples: every norm, $c\|x\|$ for $c\ge0$, $p(x)=\limsup_nx_n$ on real $\ell^\infty$ (sublinear, not a seminorm), $p(f)=|f(0)|$ on $C[0,1]$ (a seminorm, not a norm).
📐 Theorem 8.8 — Hahn–Banach theorem (real vector spaces)
Let $X$ be a real vector space, $p$ a sublinear functional on $X$, $M$ a subspace and $g:M\to\mathbb{R}$ linear with $g(m)\le p(m)$ for all $m\in M$. Then there is a linear $F:X\to\mathbb{R}$ with $F|_M=g$ and $F(x)\le p(x)$ for all $x\in X$.

Proof. Step 1: one more dimension. Let $x_0\notin M$ and $M_1=M+\mathbb{R}x_0$. Every element of $M_1$ is uniquely $m+tx_0$, so an extension is determined by the number $c=F(x_0)$, and we need $g(m)+tc\le p(m+tx_0)$ for all $m\in M$, $t\in\mathbb{R}$. Dividing by $|t|$ and using positive homogeneity, this means $$c\le p(m+x_0)-g(m)\quad(m\in M),$$ $$c\ge g(m')-p(m'-x_0)\quad(m'\in M).$$ Such a $c$ exists if and only if $g(m')-p(m'-x_0)\le p(m+x_0)-g(m)$ for all $m,m'$. And this holds: $$\begin{aligned}g(m)+g(m')&=g(m+m')\le p(m+m')\\&\le p(m+x_0)+p(m'-x_0).\end{aligned}$$ So $a=\sup_{m'}\big[g(m')-p(m'-x_0)\big]$ is at most $b=\inf_m\big[p(m+x_0)-g(m)\big]$, and any $c\in[a,b]$ works.

Step 2: Zorn's lemma. Let $\mathcal{E}$ be the set of pairs $(N,h)$ with $M\subseteq N$ a subspace, $h:N\to\mathbb{R}$ linear, $h|_M=g$ and $h\le p$ on $N$, ordered by extension: $(N,h)\preceq(N',h')$ if $N\subseteq N'$ and $h'|_N=h$. $\mathcal{E}\ni(M,g)$ is nonempty, and every chain $\{(N_i,h_i)\}$ has an upper bound: $N=\bigcup N_i$ is a subspace (any two of its elements lie in a common $N_i$) and $h=h_i$ on $N_i$ is well defined, linear and $\le p$. By Zorn's lemma there is a maximal $(N,h)$. If $N\ne X$, Step 1 extends $h$ to $N+\mathbb{R}x_0$, contradicting maximality. So $N=X$ and $F=h$. ∎

For a separable normed space Zorn is not needed: extend step by step along a dense sequence $x_1,x_2,\dots$ to the subspace they span with $M$, and then by continuity to its closure, which is $X$.

📐 Theorem 8.9 — Hahn–Banach: complex spaces and normed spaces
(a) Let $X$ be a real or complex vector space, $p$ a seminorm on $X$, $M$ a subspace and $g:M\to\mathbb{F}$ linear with $|g(m)|\le p(m)$ on $M$. Then $g$ has a linear extension $F:X\to\mathbb{F}$ with $|F(x)|\le p(x)$ for all $x$. (b) Normed version. If $X$ is a normed space, $M$ a subspace and $g\in M^*$, there is $F\in X^*$ with $F|_M=g$ and $\|F\|=\|g\|$.

Proof. (a) Real case: Theorem 8.8 gives $F\le p$, and then $-F(x)=F(-x)\le p(-x)=p(x)$, so $|F|\le p$. Complex case: $u=\operatorname{Re}g$ is real-linear on $M$ with $u\le|g|\le p$, and $g$ is recovered from $u$, because $\operatorname{Im}g(m)=-\operatorname{Re}\big(ig(m)\big)=-u(im)$: $$g(m)=u(m)-i\,u(im).$$ Regard $X$ as a real space, extend $u$ to a real-linear $U\le p$ (Theorem 8.8), and define $F(x)=U(x)-iU(ix)$. Then $F$ is real-linear, and $F(ix)=U(ix)-iU(-x)$ $=U(ix)+iU(x)=iF(x)$, so $F$ is complex-linear; on $M$ it equals $g$. Finally, given $x$, write $F(x)=|F(x)|e^{i\theta}$; then $|F(x)|=F(e^{-i\theta}x)$ is real, so it equals $U(e^{-i\theta}x)\le p(e^{-i\theta}x)=p(x)$. (b) Apply (a) with the seminorm $p(x)=\|g\|\,\|x\|$: $|F(x)|\le\|g\|\|x\|$ gives $\|F\|\le\|g\|$, and $\|F\|\ge\|g\|$ because $F$ extends $g$. ∎

✏️ Example 6 — Hahn–Banach extensions need not be unique
On $\mathbb{R}^2$ every linear functional is $F(x,y)=ax+by$. (a) Show that its norm is $|a|+|b|$ for $\|\cdot\|_\infty$, $\max(|a|,|b|)$ for $\|\cdot\|_1$, and $\sqrt{a^2+b^2}$ for $\|\cdot\|_2$. (b) In $(\mathbb{R}^2,\|\cdot\|_\infty)$ let $M=\{(t,t)\}$ and $g(t,t)=t$. Find $\|g\|$ and all norm-preserving extensions. (c) In $(\mathbb{R}^2,\|\cdot\|_1)$ let $M=\{(t,0)\}$ and $g(t,0)=t$. Same question. (d) Same as (c) for $\|\cdot\|_2$.
(a) These are the cases $p=\infty,1,2$ of Theorem 8.7 in two dimensions: $|ax+by|\le(|a|+|b|)\max(|x|,|y|)$ with equality at $(\operatorname{sgn}a,\operatorname{sgn}b)$; $|ax+by|\le\max(|a|,|b|)(|x|+|y|)$ with equality at $e_1$ or $e_2$; and Cauchy–Schwarz.
(b) Infinitely many $\|(t,t)\|_\infty=|t|$, so $\|g\|=1$. An extension $F=ax+by$ needs $a+b=1$ (to agree with $g$ at $(1,1)$) and $|a|+|b|\le1$. Since $1=a+b\le|a|+|b|\le1$, both $a,b\ge0$: the extensions are $F_s(x,y)=sx+(1-s)y$, $0\le s\le1$ — a whole segment. Geometrically: the level line $F_s=1$ passes through the corner $(1,1)$ of the unit square and every line through a corner that misses the interior is a supporting line.
(c) Infinitely many $\|g\|=1$; extensions $F=x+by$ with $\max(1,|b|)\le1$, i.e. $|b|\le1$. Again a segment: the point $(1,0)$ is a corner of the $\ell^1$ diamond.
(d) Unique $F=x+by$ with $\sqrt{1+b^2}\le1$ forces $b=0$. The disc has exactly one tangent line at $(1,0)$. (In every Hilbert space the norm-preserving extension is unique — §8.3.)
Moral Hahn–Banach is an existence theorem. Uniqueness depends on the shape of the unit ball: extensions of a norm-one $g$ correspond to the supporting hyperplanes of the ball at a point where $g$ is "normed", and a corner has many. Interactive 2 shows this for every direction of $M$.
Interactive 2 — Hahn–Banach: Extending Functionals and Separating Convex Sets

Extension mode. $M$ is the line through $0$ in the direction shown, and $g$ is the norm-one functional on $M$ with $g(x_0)=1$ at the point $x_0\in M$ of the unit sphere. Left: the unit ball of $\|\cdot\|_p$ and the level lines $\{F=1\}$ of all norm-preserving extensions $F$. Right: the dual unit ball; the extensions form its intersection with the line $\{F:F(x_0)=1\}$ — one point, or a whole segment at a corner. Separation mode. A point $x_0$ circles a closed convex set $C$; the functional $f=\langle n,\cdot\rangle$ ($n$ the unit vector from the nearest point $c$ towards $x_0$) strictly separates them.

direction of M—
x₀—
extensions—
unique?—
📐 Theorem 8.10 — Consequences of Hahn–Banach
Let $X$ be a normed space.
  1. For every $x_0\ne0$ there is $f\in X^*$ with $\|f\|=1$ and $f(x_0)=\|x_0\|$. In particular $X^*\ne\{0\}$ when $X\ne\{0\}$, and $X^*$ separates points: if $x\ne y$, some $f\in X^*$ has $f(x)\ne f(y)$.
  2. $\|x\|=\max\{|f(x)|:f\in X^*,\ \|f\|\le1\}$ — the maximum is attained.
  3. Let $M$ be a subspace and $x_0\in X$ with $d=\operatorname{dist}(x_0,M)\gt0$. Then there is $f\in X^*$ with $f|_M=0$, $\|f\|=1$ and $f(x_0)=d$. Consequently a subspace $M$ is dense in $X$ if and only if the only $f\in X^*$ vanishing on $M$ is $f=0$.
  4. The canonical map $J:X\to X^{**}$, $(Jx)(f)=f(x)$, is a linear isometry.
  5. If $X^*$ is separable, then $X$ is separable.

Proof. (1) On $M=\mathbb{F}x_0$ define $g(tx_0)=t\|x_0\|$; then $|g(tx_0)|=\|tx_0\|$, so $\|g\|=1$. Extend by Theorem 8.9(b). For separation apply this to $x_0=x-y$. (2) $|f(x)|\le\|f\|\|x\|\le\|x\|$ when $\|f\|\le1$, and (1) gives equality (for $x=0$ take $f=0$).

(3) Every element of $N=M+\mathbb{F}x_0$ is uniquely $m+tx_0$ (because $x_0\notin M$, as $d\gt0$). Define $g(m+tx_0)=td$. For $t\ne0$, since $-m/t\in M$, $$|g(m+tx_0)|=|t|\,d\le|t|\,\big\|x_0+\tfrac mt\big\|=\|m+tx_0\|,$$ so $\|g\|\le1$. Choosing $m_k\in M$ with $\|x_0-m_k\|\to d$ gives $g(x_0-m_k)=d$ and $\|g\|\ge d/\|x_0-m_k\|\to1$. Extend $g$ to $f$ with $\|f\|=1$. For the density criterion: if $\overline{M}\ne X$, pick $x_0\notin\overline M$; then $d\gt0$ and $f\ne0$ vanishes on $M$. Conversely, if $M$ is dense and $f$ is continuous and zero on $M$, then $f=0$ on $\overline M=X$.

(4) $|(Jx)(f)|=|f(x)|\le\|x\|\,\|f\|$, so $Jx\in X^{**}$ with $\|Jx\|\le\|x\|$, and (2) gives $\|Jx\|=\|x\|$. Linearity is clear.

(5) Let $\{f_n\}$ be dense in $X^*$ and choose unit vectors $x_n$ with $|f_n(x_n)|\ge\tfrac12\|f_n\|$. Let $M$ be the closed linear span of $\{x_n\}$; it is separable (finite rational linear combinations are dense in it). If $M\ne X$, (3) gives $f\in X^*$ with $\|f\|=1$ and $f|_M=0$. Pick $f_n$ with $\|f-f_n\|\lt\tfrac13$. Then $\|f_n\|\gt\tfrac23$, while $|f_n(x_n)|=|(f_n-f)(x_n)|\lt\tfrac13$, contradicting $|f_n(x_n)|\ge\tfrac12\|f_n\|\gt\tfrac13$. So $M=X$. ∎

📘 Definition 8.5 — Bidual and reflexive spaces
$X^{**}=(X^*)^*$ is the bidual of $X$. $X$ is reflexive if the canonical isometry $J:X\to X^{**}$ of Theorem 8.10(4) is onto. Facts (statements): finite-dimensional spaces, $\ell^p$ and $L^p[a,b]$ for $1\lt p\lt\infty$, and all Hilbert spaces are reflexive; $c_0$, $\ell^1$, $\ell^\infty$, $L^1[a,b]$ and $C[a,b]$ are not. A reflexive space is complete (it is isometric to $X^{**}$, a dual space, which is Banach by Theorem 8.5).

For example, $\ell^1$ is not reflexive: it is separable, but its dual $\ell^\infty$ is not (the uncountably many $0$–$1$ sequences are at mutual distance $1$). If $\ell^1$ were reflexive, $(\ell^\infty)^*\cong(\ell^1)^{**}\cong\ell^1$ would be separable, and Theorem 8.10(5) applied to $X=\ell^\infty$ would make $\ell^\infty$ separable — a contradiction. Likewise $c_0^{**}=(\ell^1)^*=\ell^\infty\ne c_0$.

Separation of convex sets

The geometric form of Hahn–Banach replaces the norm by a convex set. A real hyperplane in $X$ is a set $\{x:\operatorname{Re}f(x)=t\}$ with $f\in X^*\setminus\{0\}$ and $t\in\mathbb{R}$; it separates $A$ and $B$ if they lie in the two closed half-spaces it bounds.

📘 Definition 8.6 — Minkowski functional (gauge)
Let $C$ be a convex subset of a normed space with $0$ an interior point of $C$. Its Minkowski functional is $p_C(x)=\inf\{t\gt0:x\in tC\}$. For the open unit ball, $p_C=\|\cdot\|$; for $C=\{(x,y):x^2/4+y^2\lt1\}$, $p_C(x,y)=\sqrt{x^2/4+y^2}$.
📐 Theorem 8.11 — Separation theorems (geometric Hahn–Banach)
Let $A,B$ be nonempty disjoint convex subsets of a normed space $X$ (real or complex).
  1. If $A$ is open, there are $f\in X^*$ and $t\in\mathbb{R}$ with $\operatorname{Re}f(a)\lt t\le\operatorname{Re}f(b)$ for all $a\in A$, $b\in B$.
  2. If $A$ is closed and $B$ is compact, there are $f\in X^*$ and $t_1\lt t_2$ with $\operatorname{Re}f(a)\le t_1\lt t_2\le\operatorname{Re}f(b)$: strict separation. In particular, a point outside a closed convex set can be strictly separated from it, and every closed convex set is the intersection of the closed half-spaces containing it.

Proof. Work first with real scalars. Gauge lemma: if $C$ is open and convex with $B(0,r)\subseteq C$, then $p_C$ is sublinear, $0\le p_C(x)\le\|x\|/r$, and $C=\{x:p_C(x)\lt1\}$. Indeed, $\tfrac{r}{\|x\|+\varepsilon}x\in B(0,r)\subseteq C$ gives $p_C(x)\le(\|x\|+\varepsilon)/r$. If $x\in sC$ and $s\lt t$, then $x/t=\tfrac st(x/s)+(1-\tfrac st)0\in C$, so $x\in tC$. Given $\varepsilon\gt0$ put $s=p_C(x)+\varepsilon$, $s'=p_C(y)+\varepsilon$; then $x/s,\,y/s'\in C$ and, by convexity, $$\frac{x+y}{s+s'}=\frac{s}{s+s'}\cdot\frac xs+\frac{s'}{s+s'}\cdot\frac y{s'}\in C,$$ so $p_C(x+y)\le p_C(x)+p_C(y)+2\varepsilon$. Positive homogeneity is clear. If $x\in C$, openness gives $(1+\delta)x\in C$, so $p_C(x)\le\tfrac1{1+\delta}\lt1$; if $p_C(x)\lt1$, then $x\in tC$ for some $t\lt1$, hence $x\in C$.

A point and an open convex set. Let $C$ be open convex and $x_0\notin C$; translating, assume $0\in C$. On $\mathbb{R}x_0$ let $g(sx_0)=s$. Then $g\le p_C$ there: for $s\gt0$, $p_C(sx_0)=s\,p_C(x_0)\ge s$ because $x_0\notin C$; for $s\le0$, $g(sx_0)\le0\le p_C$. Theorem 8.8 gives a linear $F\le p_C$ extending $g$; $F$ is bounded since $|F(x)|\le\max(p_C(x),p_C(-x))\le\|x\|/r$, and $F(x)\le p_C(x)\lt1=F(x_0)$ for $x\in C$.

(1) $C=A-B=\bigcup_{b\in B}(A-b)$ is open and convex, and $0\notin C$. The previous step (with $x_0=0$, after translating $C$) gives $F\in X^*$ with $F(a-b)\lt F(0)=0$, i.e. $F(a)\lt F(b)$. Let $t=\sup_AF\le\inf_BF$. The set $F(A)$ is open in $\mathbb{R}$ (a nonzero functional maps open sets to open sets: if $F(z)=1$ then $F(a+sz)=F(a)+s$), so the supremum is not attained: $F(a)\lt t\le F(b)$. (2) The function $b\mapsto\operatorname{dist}(b,A)$ is continuous and positive on the compact set $B$ ($A$ is closed), so it has a minimum $\delta\gt0$. The set $A_\delta=A+B(0,\delta/2)$ is open, convex and disjoint from $B$; by (1) there is $F\ne0$ with $F(a)+F(z)\lt t\le F(b)$ whenever $\|z\|\lt\delta/2$. Taking the supremum over $z$, $F(a)+\tfrac\delta2\|F\|\le t$, so $t_1=t-\tfrac\delta2\|F\|\lt t_2=t$ works.

Complex scalars: apply the real result to $X$ regarded as a real space to get a real functional $u$, and put $f(x)=u(x)-iu(ix)$; as in Theorem 8.9, $f\in X^*$ and $\operatorname{Re}f=u$. ∎

✏️ Example 7 — $(\ell^\infty)^*$ is bigger than $\ell^1$
Work in real $\ell^\infty$. Prove that there is $f\in(\ell^\infty)^*$ with $\|f\|=1$, $f=0$ on $c_0$ and $f(\mathbf 1)=1$, where $\mathbf 1=(1,1,1,\dots)$, and deduce that $f$ is not of the form $\varphi_y$ with $y\in\ell^1$. Show moreover that $\liminf x_n\le f(x)\le\limsup x_n$ for every $x\in\ell^\infty$.
Distance For $y\in c_0$, $\|\mathbf 1-y\|_\infty\ge\limsup_n|1-y_n|=1$, and $y=0$ gives $1$; so $\operatorname{dist}(\mathbf 1,c_0)=1$.
Hahn–Banach Theorem 8.10(3) with $M=c_0$, $x_0=\mathbf 1$ gives $f\in(\ell^\infty)^*$, $\|f\|=1$, $f|_{c_0}=0$, $f(\mathbf 1)=1$.
Not in $\ell^1$ If $f=\varphi_y$ with $y\in\ell^1$, then $y_n=f(e_n)=0$ for all $n$ ($e_n\in c_0$), so $f=0$ — contradicting $f(\mathbf 1)=1$. Hence $\ell^1\to(\ell^\infty)^*$ is not onto, proving Theorem 8.7(4).
Between liminf and limsup Let $L=\limsup x_n$. Given $\varepsilon\gt0$, $x_n\le L+\varepsilon$ for $n\ge N$. Change the first $N-1$ terms (an element of $c_{00}\subseteq c_0$, so $f$ does not notice) to get $x'$ with $f(x')=f(x)$ and $x'_n\le L+\varepsilon$ for all $n$. Then $(L+\varepsilon)\mathbf 1-x'$ has non-negative real entries, say $z$, and $f(z)\ge0$: indeed $\|\|z\|_\infty\mathbf 1-z\|_\infty\le\|z\|_\infty$, so $\|z\|_\infty-f(z)=f(\|z\|_\infty\mathbf 1-z)\le\|z\|_\infty$. Hence $f(x)=f(x')\le L+\varepsilon$; letting $\varepsilon\to0$, $f(x)\le\limsup x_n$, and applying this to $-x$ gives the other inequality. So $f$ extends "$\lim$" from $c$ to all of $\ell^\infty$. (A shift-invariant version of such an $f$ is called a Banach limit.)
🧭 Problem-solving strategy — computing $\|T\|$ and $\|\varphi\|$ in exam problems
  1. Upper bound first: estimate $\|Tx\|\le C\|x\|$ with Hölder (sequences, $L^p$), the kernel bound $\max_x\int|K(x,t)|\,dt$ ($C[a,b]$), or row/column sums (matrices).
  2. Then an extremal vector: $e_n$ for $\ell^1$; the sign pattern $\overline{\operatorname{sgn}y_n}$ for $c_0$/$\ell^\infty$; $|y_n|^{q-1}\overline{\operatorname{sgn}y_n}$ for $\ell^p$; $f\equiv1$ when the kernel is $\ge0$; a smoothed $\overline{\operatorname{sgn}g}$ on $C[a,b]$.
  3. Read off known formulas: $\|\varphi_y\|_{(\ell^p)^*}=\|y\|_q$, $\|\varphi_y\|_{c_0^*}=\|y\|_1$, $\|\varphi_g\|_{C[a,b]^*}=\int|g|$, point evaluations $\delta_t(f)=f(t)$ have norm $1$, so $\|\sum c_i\delta_{t_i}\|=\sum|c_i|$ for distinct $t_i$.
  4. Attained? Check whether the extremal vector lies in the space: it fails for sign changes on $C[a,b]$ and for infinitely many nonzero $y_n$ on $c_0$; it always succeeds on $\ell^p$, $1\lt p\lt\infty$.
  5. Unbounded? Exhibit unit vectors with $\|Tx_n\|\to\infty$ ($x^n$ for $d/dx$, $e_n$ for $x\mapsto\sum nx_n$).
⚠️ Counterexamples and pitfalls
  1. Incomplete normed spaces are everywhere. $(c_{00},\|\cdot\|_p)$, $(C[0,1],\|\cdot\|_1)$, polynomials with $\|\cdot\|_\infty$, $(C^1[0,1],\|\cdot\|_\infty)$ and $(\ell^1,\|\cdot\|_\infty)$ are not complete. A subspace of a Banach space is Banach iff it is closed.
  2. Inclusions go opposite ways. $\ell^p\subseteq\ell^r$ for $p\lt r$ (sequences), but $L^r[0,1]\subseteq L^p[0,1]$ for $p\lt r$ (finite measure), and on $\mathbb{R}$ no inclusion holds: $x^{-1/2}\mathbf 1_{(0,1)}\in L^1\setminus L^2$, $x^{-1}\mathbf 1_{(1,\infty)}\in L^2\setminus L^1$.
  3. Norms need not be attained. $f\mapsto\int_0^1(2t-1)f$ on $C[0,1]$ (norm $\tfrac12$); $x\mapsto\sum x_n2^{-n}$ on $c_0$ (norm $1$); $f\mapsto\int_0^1f-f(0)$ (norm $2$).
  4. Linear does not imply bounded in infinite dimensions: $d/dx$ on $(C^1[0,1],\|\cdot\|_\infty)$, $x\mapsto\sum nx_n$ on $c_{00}$, and (by Zorn) functionals on every infinite-dimensional normed space. A functional is bounded iff its kernel is closed; an unbounded one has a dense kernel.
  5. Pointwise vs norm convergence of operators. The left shift $L$ on $\ell^2$ has $L^nx\to0$ for every $x$, yet $\|L^n\|=1$ for all $n$. The truncations $P_k$ satisfy $P_kx\to x$ in $\ell^2$ but $\|I-P_k\|=1$.
  6. The duality is not symmetric. $(\ell^1)^*=\ell^\infty$ but $(\ell^\infty)^*\ne\ell^1$; $c_0^*=\ell^1$ although $c_0\ne\ell^\infty$; $(L^1)^*=L^\infty$ but $(L^\infty)^*\ne L^1$. $\ell^1$, $c_0$, $\ell^\infty$, $L^1$, $C[a,b]$ are not reflexive.
  7. Hahn–Banach extensions are not unique in general: in $(\mathbb{R}^2,\|\cdot\|_\infty)$, $g(t,t)=t$ has the extensions $sx+(1-s)y$, $s\in[0,1]$ (unique in Hilbert spaces).
  8. Functionals extend, operators need not. Hahn–Banach extends scalar-valued functionals. The identity $c_0\to c_0$ has no bounded extension $\ell^\infty\to c_0$ (Phillips' theorem, quoted): $c_0$ is not complemented in $\ell^\infty$.
  9. Separation needs openness or compactness. $A=\{(x,y):x\gt0,\ y\ge1/x\}$ and $B=\{(x,0):x\in\mathbb{R}\}$ are disjoint closed convex sets in $\mathbb{R}^2$ at distance $0$; the line $y=0$ separates them, but no line separates them strictly.
  10. $X^*$ is always complete, even when $X$ is not: $(c_{00},\|\cdot\|_\infty)^*=\ell^1$ (via the completion $c_0$).

Exercises

I
Norms and completeness.
  • 1 Which of the following are norms on $\mathbb{R}^2$? (a) $|x|+2|y|$; (b) $\max(|x+y|,|x-y|)$; (c) $|x|^2+|y|$; (d) $\big(\sqrt{|x|}+\sqrt{|y|}\big)^2$; (e) $|x-y|$.
  • 2 For $x=(1,-2,2)\in\mathbb{R}^3$ compute $\|x\|_1$, $\|x\|_2$ and $\|x\|_\infty$.
  • 3 For which $\alpha\gt0$ is $(n^{-\alpha})$ in $\ell^3$? For which $p\in[1,\infty]$ is $(n^{-1/2})$ in $\ell^p$?
  • 4 For which $p\in[1,\infty)$ is $f(x)=x^{-1/3}$ in $L^p[0,1]$? Compute $\|f\|_2$.
  • 5 Show that $M=\{x\in\ell^\infty:x_1=x_2\}$ is a closed subspace of $\ell^\infty$ and $N=\{x\in\ell^2:\sum_nx_n/n=0\}$ is a closed subspace of $\ell^2$.
  • 6 Prove that $N(f)=|f(0)|+\|f'\|_\infty$ is a norm on $C^1[0,1]$ equivalent to $\|f\|_{C^1}=\|f\|_\infty+\|f'\|_\infty$, and deduce that $(C^1[0,1],N)$ is a Banach space.
  • 7 Prove that $\ell^1\subsetneq\ell^2$ and that the inclusion map $(\ell^1,\|\cdot\|_1)\to(\ell^2,\|\cdot\|_2)$ is bounded with norm $1$.
  • 8 Use Theorem 8.2 to give a second proof that $\ell^1$ is complete.
  • 9 Show that $C[0,1]$ with $\|f\|_2=\big(\int_0^1|f|^2\big)^{1/2}$ is not complete.
1 (a) Yes — a weighted $\ell^1$ norm. (b) Yes: $\max(|x+y|,|x-y|)^2=x^2+y^2+2|xy|$, which is $(|x|+|y|)^2$, so it is $\|\cdot\|_1$. (c) No: $N(2,0)=4\ne2N(1,0)$. (d) No: $N(1,0)+N(0,1)=2\lt N(1,1)=4$. (e) No — only a seminorm, $N(1,1)=0$.
2 $\|x\|_1=5$, $\|x\|_2=\sqrt{1+4+4}=3$, $\|x\|_\infty=2$; note $\|x\|_\infty\le\|x\|_2\le\|x\|_1$.
3 $\sum n^{-3\alpha}\lt\infty\iff\alpha\gt\tfrac13$. $(n^{-1/2})\in\ell^p\iff p\gt2$ (including $p=\infty$).
4 $\int_0^1x^{-p/3}dx\lt\infty\iff p\lt3$. $\|f\|_2^2=\int_0^1x^{-2/3}dx=3$, so $\|f\|_2=\sqrt3\approx1.732$.
5 $M=\ker\phi$ with $\phi(x)=x_1-x_2$, $|\phi(x)|\le2\|x\|_\infty$; $N=\ker\varphi_y$ with $y=(1/n)\in\ell^2$, bounded on $\ell^2$ by Hölder. The kernel of a bounded functional is the preimage of the closed set $\{0\}$ under a continuous map, hence a closed subspace.
6 If $N(f)=0$ then $f'\equiv0$, so $f$ is constant, equal to $f(0)=0$; the other axioms are clear. From $f(x)=f(0)+\int_0^xf'$, $\|f\|_\infty\le N(f)$, so $\|f\|_{C^1}\le2N(f)$; and $N(f)\le\|f\|_{C^1}$. Equivalent norms have the same Cauchy sequences and the same limits, and $\|\cdot\|_{C^1}$ is complete (Example 1(d)).
7 If $\|x\|_1=1$ then $|x_n|\le1$, so $\sum|x_n|^2\le\sum|x_n|=1$: $\|x\|_2\le\|x\|_1$ (Example 2(b)), with equality at $e_1$; so the norm is $1$. $(1/n)\in\ell^2\setminus\ell^1$.
8 Let $x^{(k)}\in\ell^1$ with $\sum_k\|x^{(k)}\|_1\lt\infty$. For each $n$, $\sum_k|x^{(k)}_n|\le\sum_k\|x^{(k)}\|_1$, so $s_n=\sum_kx^{(k)}_n$ converges absolutely. Then $\sum_n|s_n-\sum_{k\le K}x^{(k)}_n|\le\sum_{k\gt K}\|x^{(k)}\|_1\to0$ (exchange the order of summation of non-negative terms). So the series converges in $\ell^1$ to $s$, and $\ell^1$ is complete.
9 The ramps $f_n$ of Example 1(a) satisfy $\|f_m-f_n\|_2^2\le\tfrac1n$ for $m\gt n$, so they are $\|\cdot\|_2$-Cauchy; exactly as in Example 1(a), a continuous limit would have to be $0$ on $[0,\tfrac12]$ and $1$ on $(\tfrac12,1]$. ($L^2[0,1]$ is the completion.)
II
Bounded operators and their norms.
  • 1 For $A=\begin{pmatrix}3&-1\\2&2\end{pmatrix}$ find $\|A\|_1$, $\|A\|_\infty$ and $\|A\|_2$.
  • 2 On $\ell^2$ let $R(x_1,x_2,\dots)=(0,x_1,x_2,\dots)$ and $L(x_1,x_2,\dots)=(x_2,x_3,\dots)$. Find $\|R\|$, $\|L\|$, $LR$ and $RL$, and show that $L^nx\to0$ for every $x$ while $\|L^n\|=1$.
  • 3 For a sequence $\lambda$ let $D_\lambda x=(\lambda_nx_n)$. Prove that $D_\lambda$ maps $\ell^p$ into itself boundedly if and only if $\lambda\in\ell^\infty$, and then $\|D_\lambda\|=\|\lambda\|_\infty$.
  • 4 Find the norm of $(Tf)(x)=\int_0^1e^{x-t}f(t)\,dt$ on $C[0,1]$.
  • 5 Find the norm of $\phi(f)=f(\tfrac12)-2f(1)$ on $C[0,1]$. Is it attained?
  • 6 Find the norm of $\phi(f)=\int_0^1f(t)\,dt-f(0)$ on $C[0,1]$. Is it attained?
  • 7 Prove that a linear functional $f$ on a normed space is bounded if and only if $\ker f$ is closed.
  • 8 For $g\in C[0,1]$ let $M_gf=gf$ on $C[0,1]$. Prove $\|M_g\|=\|g\|_\infty$ and evaluate it for $g(x)=x-x^2$.
  • 9 Let $X$ be a Banach space and $T\in B(X)$ with $\|T\|\lt1$. Prove that $I-T$ is invertible in $B(X)$, with $(I-T)^{-1}=\sum_{n\ge0}T^n$. Use it to solve $u(x)-\tfrac12\int_0^1xt\,u(t)\,dt=x$ in $C[0,1]$.
1 Column sums $5,3$: $\|A\|_1=5$. Row sums $4,4$: $\|A\|_\infty=4$. $A^TA=\begin{pmatrix}13&1\\1&5\end{pmatrix}$, eigenvalues $9\pm\sqrt{17}$, so $\|A\|_2=\sqrt{9+\sqrt{17}}\approx3.6226$.
2 $\|Rx\|=\|x\|$, so $\|R\|=1$; $\|Lx\|\le\|x\|$ with equality at $e_2$: $\|L\|=1$. $LR=I$ but $RL(x)=(0,x_2,x_3,\dots)\ne x$. $\|L^nx\|^2=\sum_{k\gt n}|x_k|^2\to0$, yet $\|L^ne_{n+1}\|=\|e_1\|=1$, so $\|L^n\|=1$: $L^n\to0$ pointwise but not in norm.
3 If $\lambda\in\ell^\infty$, $\|D_\lambda x\|_p\le\|\lambda\|_\infty\|x\|_p$, and $\|D_\lambda e_n\|_p=|\lambda_n|$ gives $\|D_\lambda\|=\sup|\lambda_n|$. If $\lambda$ is unbounded, $\|D_\lambda e_n\|_p=|\lambda_n|$ is unbounded on unit vectors; in fact $D_\lambda$ does not even map $\ell^p$ into $\ell^p$: choose $n_k$ with $|\lambda_{n_k}|\ge k^{2}$, take $x=\sum_kk^{-2}e_{n_k}\in\ell^p$; then $|(D_\lambda x)_{n_k}|\ge1$ for all $k$.
4 Kernel $e^{x-t}\gt0$: $\|T\|=\max_x\int_0^1e^{x-t}dt$ $=\max_xe^x(1-e^{-1})=e-1\approx1.718$, attained at $f\equiv1$ (Example 4(c)).
5 $|\phi(f)|\le|f(\tfrac12)|+2|f(1)|\le3\|f\|_\infty$; a continuous $f$ with $\|f\|_\infty=1$, $f(\tfrac12)=1$, $f(1)=-1$ (piecewise linear) gives $\phi(f)=3$. Norm $3$, attained.
6 $|\phi(f)|\le2\|f\|_\infty$. With $f_n=-1+2nt$ on $[0,\tfrac1n]$ and $f_n=1$ after, $\int f_n=1-\tfrac1n$ and $\phi(f_n)=2-\tfrac1n\to2$. Norm $2$, not attained: $\phi(f)=2$ with $|f|\le1$ forces $\int_0^1f=1$ (so $f\equiv1$, by continuity) and $f(0)=-1$.
7 Bounded $\Rightarrow$ $\ker f=f^{-1}(\{0\})$ is closed. Conversely let $\ker f$ be closed and $f\ne0$; pick $x_0$ with $f(x_0)=1$, so $d=\operatorname{dist}(x_0,\ker f)\gt0$. If $f(x)\ne0$ then $x_0-x/f(x)\in\ker f$, so $\|x/f(x)\|=\|x_0-(x_0-x/f(x))\|\ge d$, i.e. $|f(x)|\le\|x\|/d$. So $\|f\|\le1/d$.
8 $\|gf\|_\infty\le\|g\|_\infty\|f\|_\infty$, and $f\equiv1$ gives equality. For $g=x-x^2$: $\max_{[0,1]}(x-x^2)=\tfrac14$ at $x=\tfrac12$.
9 $\sum\|T^n\|\le\sum\|T\|^n\lt\infty$, so by Theorems 8.2 and 8.5 the series $S=\sum T^n$ converges in $B(X)$; $(I-T)\sum_{n\le N}T^n=I-T^{N+1}\to I$, and the same on the other side, so $S=(I-T)^{-1}$ with $\|S\|\le\frac1{1-\|T\|}$. Here $(Tu)(x)=\tfrac12x\int_0^1tu(t)\,dt$ has $\|T\|=\max_x\int_0^1\tfrac12xt\,dt=\tfrac14\lt1$. The solution has the form $u=x+cx$; substituting, $c=\tfrac12\int_0^1t^2(1+c)\,dt=\tfrac{1+c}6$, so $c=\tfrac15$ and $u(x)=\tfrac65x$. Check: $\tfrac65x-\tfrac12x\cdot\tfrac65\cdot\tfrac13=x$. ✓
III
Dual spaces.
  • 1 Find the norm of $\phi(x)=x_1-x_2+x_3$ on $\ell^1$, on $\ell^2$ and on $\ell^\infty$.
  • 2 Let $\phi(x)=\sum_nx_n/n$. On which of the spaces $\ell^1$, $\ell^2$, $\ell^p$ ($1\lt p\lt\infty$), $c_0$ is $\phi$ a bounded functional? Find $\|\phi\|$ on $\ell^1$ and on $\ell^2$.
  • 3 Find the norm of $\phi(x)=\sum_nx_n3^{-n}$ on $\ell^2$ and on $c_0$, and decide whether each is attained.
  • 4 Find the norm of $x\mapsto x_1+x_2+x_3+x_4$ on $\ell^4$.
  • 5 Show that $\lim:c\to\mathbb{F}$, $x\mapsto\lim_nx_n$, is a bounded functional of norm $1$, and that it is not of the form $x\mapsto\sum x_ny_n$ for any $y\in\ell^1$.
  • 6 Which of the following are reflexive: $\ell^3$, $c_0$, $L^2[0,1]$, $\ell^1$, $C[0,1]$, $\mathbb{R}^5$ with $\|\cdot\|_\infty$?
  • 7 Show that every $\phi\in(\ell^p)^*$ attains its norm when $1\lt p\lt\infty$, while $\varphi_y\in c_0^*$ ($y\in\ell^1$) attains its norm if and only if $y$ has only finitely many nonzero terms.
1 $\phi=\varphi_y$, $y=(1,-1,1,0,\dots)$: on $\ell^1$, $\|y\|_\infty=1$; on $\ell^2$, $\|y\|_2=\sqrt3$; on $\ell^\infty$, $\|y\|_1=3$.
2 $y=(1/n)$ lies in $\ell^q$ for every $q\gt1$ and in $\ell^\infty$, but not in $\ell^1$. So $\phi$ is bounded on $\ell^1$ (norm $\|y\|_\infty=1$) and on $\ell^p$ for $1\lt p\lt\infty$ (norm $\|y\|_q$); on $\ell^2$, $\|\phi\|=\big(\sum n^{-2}\big)^{1/2}=\pi/\sqrt6\approx1.2825$. On $c_0$ it is not even defined everywhere: $x_n=1/\log(n+1)$ gives $\sum1/(n\log(n+1))=\infty$.
3 On $\ell^2$: $\big(\sum9^{-n}\big)^{1/2}=\big(\tfrac18\big)^{1/2}=\tfrac1{2\sqrt2}\approx0.3536$, attained at $x=y/\|y\|_2$. On $c_0$: $\sum3^{-n}=\tfrac12$, not attained (attaining would need $x_n=1$ for all $n$).
4 $q=\tfrac43$: $\|(1,1,1,1)\|_{4/3}=4^{3/4}=2\sqrt2\approx2.828$.
5 $|\lim x_n|\le\sup|x_n|$, with equality at $\mathbf 1$: norm $1$. If $\lim x_n=\sum x_ny_n$ on $c$, then $y_n=\lim(e_n)=0$ for every $n$, so the functional would vanish — but $\lim\mathbf 1=1$. (In fact $c^*\cong\ell^1\oplus\mathbb{F}$: every functional is $x\mapsto a\lim x_n+\sum x_ny_n$.)
6 Reflexive: $\ell^3$, $L^2[0,1]$ and $\mathbb{R}^5$ (every finite-dimensional space). Not reflexive: $c_0$ ($c_0^{**}=\ell^\infty$), $\ell^1$ (its dual $\ell^\infty$ is not separable — see the remark after Definition 8.5), $C[0,1]$.
7 On $\ell^p$, the vector $x_n=|y_n|^{q-1}\overline{\operatorname{sgn}y_n}$ (all $n$) lies in $\ell^p$ because $\sum|x_n|^p=\sum|y_n|^q\lt\infty$, and it gives equality in Hölder. On $c_0$: if $\|x\|_\infty\le1$ and $\sum x_ny_n=\sum|y_n|$, then $\sum(|y_n|-\operatorname{Re}(x_ny_n))=0$ with non-negative terms, forcing $x_n=\overline{\operatorname{sgn}y_n}$, of modulus $1$, whenever $y_n\ne0$. That is compatible with $x_n\to0$ only if $y_n\ne0$ for finitely many $n$; conversely, if $y\in c_{00}$, $x=(\overline{\operatorname{sgn}y_n})$ is in $c_{00}\subseteq c_0$.
IV
Hahn–Banach and its consequences.
  • 1 Let $M=\{(t,2t)\}\subseteq\mathbb{R}^2$. (a) With $\|\cdot\|_1$ on $\mathbb{R}^2$ and $g(t,2t)=3t$, find $\|g\|$ and all norm-preserving extensions. (b) With $\|\cdot\|_\infty$ and $g(t,2t)=2t$, do the same. Explain the answers with the shape of the unit balls.
  • 2 Prove that if $x\in X$ and $f(x)=0$ for every $f\in X^*$, then $x=0$.
  • 3 In $C[0,1]$ let $M=\{h:h(0)=0\}$. Find $\operatorname{dist}(\mathbf 1,M)$ and an explicit $f\in C[0,1]^*$ with $\|f\|=1$, $f|_M=0$ and $f(\mathbf 1)=\operatorname{dist}(\mathbf 1,M)$.
  • 4 Prove that $\operatorname{dist}(x_0,\ker f)=|f(x_0)|/\|f\|$ for $f\in X^*\setminus\{0\}$. Compute $\operatorname{dist}(e_1,H)$ for $H=\{x:x_1+x_2+x_3=0\}$ in $\ell^1$, in $\ell^2$ and in $\ell^\infty$.
  • 5 Show that the Minkowski functional of the open unit ball of a normed space is the norm, and compute the Minkowski functional of $C=\{(x,y):|x|\lt1,\ |y|\lt3\}$.
  • 6 Prove that every finite-dimensional subspace $M$ of a normed space $X$ is complemented: there is a closed subspace $N$ with $X=M\oplus N$.
  • 7 For $T\in B(X,Y)$ prove that $\|T\|$ equals the supremum of $|g(Tx)|$ over $\|x\|\le1$ and $g\in Y^*$ with $\|g\|\le1$, and deduce that the adjoint $T^*:Y^*\to X^*$, $T^*g=g\circ T$, has $\|T^*\|=\|T\|$.
  • 8 In $(\mathbb{R}^2,\|\cdot\|_2)$ let $C$ be the closed unit disc and $x_0=(3,4)$. Find a unit vector $n$ such that $f(x)=n\cdot x$ strictly separates $x_0$ from $C$, and compute $\sup_Cf$ and $f(x_0)$.
1 (a) $\|(1,2)\|_1=3$, so $\|g\|=1$. Extensions $ax+by$ need $a+2b=3$ and $\max(|a|,|b|)\le1$; then $a=3-2b\le1$ forces $b\ge1$, so $b=1$, $a=1$: the unique extension is $x+y$. (b) $\|(1,2)\|_\infty=2$, $\|g\|=1$; need $a+2b=2$ and $|a|+|b|\le1$. Since $2=a+2b\le|a|+2|b|\le1+|b|$, $|b|\ge1$, so $b=\pm1$; $b=-1$ gives $a=4$, too big; $b=1$, $a=0$: unique extension $y$. In both cases the normalised point $\tfrac13(1,2)$ (resp. $\tfrac12(1,2)$) lies in the middle of an edge of the unit ball, not at a corner, so the supporting line is unique (compare Example 6).
2 If $x\ne0$, Theorem 8.10(1) gives $f$ with $f(x)=\|x\|\ne0$.
3 For $h\in M$, $\|\mathbf 1-h\|_\infty\ge|1-h(0)|=1$, and $h=0$ gives $1$: $\operatorname{dist}=1$. The point evaluation $\delta_0(h)=h(0)$ has norm $1$, vanishes on $M$ and $\delta_0(\mathbf 1)=1$.
4 For $m\in\ker f$, $|f(x_0)|=|f(x_0-m)|\le\|f\|\,\|x_0-m\|$, so $\operatorname{dist}\ge|f(x_0)|/\|f\|$. For any $x$ with $f(x)\ne0$, $m=x_0-\tfrac{f(x_0)}{f(x)}x\in\ker f$ and $\|x_0-m\|=|f(x_0)|\,\tfrac{\|x\|}{|f(x)|}$; the infimum of $\|x\|/|f(x)|$ is $1/\|f\|$. For $f=\varphi_y$, $y=(1,1,1,0,\dots)$, $f(e_1)=1$: in $\ell^1$, $\|f\|=\|y\|_\infty=1$, distance $1$; in $\ell^2$, $\|f\|=\sqrt3$, distance $1/\sqrt3\approx0.577$; in $\ell^\infty$, $\|f\|=\|y\|_1=3$, distance $\tfrac13$.
5 $x\in tB(0,1)\iff\|x\|\lt t$, so $\inf\{t\gt0:\|x\|\lt t\}=\|x\|$. For the rectangle, $(x,y)\in tC\iff|x|\lt t$ and $|y|\lt3t$, so $p_C(x,y)=\max(|x|,|y|/3)$.
6 Let $e_1,\dots,e_n$ be a basis of $M$ and $g_i$ the coordinate functionals on $M$ ($g_i(\sum a_je_j)=a_i$), bounded by Theorem 8.6. Extend them to $f_i\in X^*$ (Hahn–Banach). Then $Px=\sum_if_i(x)e_i$ is a bounded projection of $X$ onto $M$ ($P^2=P$, $P|_M=I$), and $N=\ker P=\bigcap_i\ker f_i$ is closed with $X=M\oplus N$: $x=Px+(x-Px)$.
7 By Theorem 8.10(2), $\|Tx\|=\max_{\|g\|\le1}|g(Tx)|$; take the supremum over $\|x\|\le1$. Then $\|T^*\|=\sup_{\|g\|\le1}\|g\circ T\|$, which is $\sup_{\|g\|\le1}\sup_{\|x\|\le1}|g(Tx)|=\|T\|$.
8 The nearest point of $C$ to $x_0$ is $c=x_0/\|x_0\|=(\tfrac35,\tfrac45)$, and $n=(x_0-c)/\|x_0-c\|=(\tfrac35,\tfrac45)$. So $f(x)=\tfrac{3x+4y}5$: $\sup_Cf=\|n\|_2=1$ (Cauchy–Schwarz) and $f(x_0)=5$. Any $t\in(1,5)$ gives a separating line $3x+4y=5t$; the gap $5-1=4$ is $\operatorname{dist}(x_0,C)$.
🎯 GATE corner GATE MA · MCQGATE MA · MSQGATE MA · NAT

E1. GATE MA · MCQ Let $\phi:C[0,1]\to\mathbb{R}$ be $\phi(f)=\int_0^1t\,f(t)\,dt-f(1)$, where $C[0,1]$ carries the sup norm. Then $\|\phi\|$ equals (A) $\tfrac12$ (B) $1$ (C) $\tfrac32$ (D) $2$.

Upper bound $|\phi(f)|\le\|f\|_\infty\int_0^1t\,dt+|f(1)|\le\tfrac32\|f\|_\infty$.
Approach it Let $f_n=1$ on $[0,1-\tfrac1n]$, linear down to $f_n(1)=-1$. Then $\int_0^1tf_n\to\tfrac12$ and $-f_n(1)=1$, so $\phi(f_n)\to\tfrac32$. (Not attained: $\int tf=\tfrac12$ with $|f|\le1$ forces $f=1$ on $(0,1]$, contradicting $f(1)=-1$.)
Answer (C) $\tfrac32$. (A) forgets the point mass, (B) forgets the integral, (D) adds the wrong weights.

E2. GATE MA · MCQ Which one of the following normed spaces is not a Banach space? (A) $\{x\in\ell^\infty:x_1=x_2\}$ with $\|\cdot\|_\infty$ (B) $c_0$ with $\|\cdot\|_\infty$ (C) the space of sequences with only finitely many nonzero terms, with $\|\cdot\|_1$ (D) $C[0,1]$ with $\|f\|=\max_{[0,1]}|f|+|f(0)|$.

(A), (B) Closed subspaces of $\ell^\infty$ (a kernel of a bounded functional; Theorem 8.3(1)) — Banach.
(C) $c_{00}$ is dense in $\ell^1$ but not equal to it, so it is not closed and not complete: $(1,\tfrac14,\dots,\tfrac1{k^2},0,\dots)$ is Cauchy with limit $(n^{-2})\notin c_{00}$.
(D) $\|f\|_\infty\le\|f\|\le2\|f\|_\infty$: an equivalent norm, so complete.
Answer (C).

E3. GATE MA · MSQ CSIR Part C Let $X$ be a normed space and $M$ a proper closed subspace of $X$. Which of the following statements are true? (One or more options may be correct.) (A) There is $f\in X^*$, $f\ne0$, with $f=0$ on $M$. (B) For every $x_0\notin M$ there is $f\in X^*$ with $\|f\|=1$, $f|_M=0$ and $f(x_0)=\operatorname{dist}(x_0,M)$. (C) Every $g\in M^*$ has exactly one extension $F\in X^*$ with $\|F\|=\|g\|$. (D) Every $f\in X^*$ attains its norm on the closed unit ball of $X$.

(A) True Pick $x_0\notin M$; since $M$ is closed, $\operatorname{dist}(x_0,M)\gt0$, and (B) produces such an $f$.
(B) True Theorem 8.10(3).
(C) False In $(\mathbb{R}^2,\|\cdot\|_\infty)$, $g(t,t)=t$ on $M=\{(t,t)\}$ has the extensions $sx+(1-s)y$, $0\le s\le1$ (Example 6).
(D) False $f\mapsto\int_0^1(2t-1)f(t)\,dt$ on $C[0,1]$ has norm $\tfrac12$, never attained (Example 4(b)).
Answer (A) and (B).

E4. GATE MA · MSQ Let $T:\ell^2\to\ell^2$ be $T(x_1,x_2,x_3,\dots)=\big(0,\,x_1,\,\tfrac{x_2}2,\,\tfrac{x_3}3,\dots\big)$. Which of the following are true? (One or more options may be correct.) (A) $\|T\|=1$. (B) $T$ is injective. (C) $T$ is surjective. (D) $\|T^2\|=\tfrac12$.

(A) True $(Tx)_{n+1}=x_n/n$, so $\|Tx\|^2=\sum|x_n|^2/n^2\le\|x\|^2$, with equality at $e_1$.
(B) True $Tx=0$ forces $x_n/n=0$ for all $n$.
(C) False Every $Tx$ has first coordinate $0$, so $e_1$ is not in the range.
(D) True $(T^2x)_{n+2}=(Tx)_{n+1}/(n+1)=\dfrac{x_n}{n(n+1)}$, a weighted shift with weights $\tfrac1{n(n+1)}$, so $\|T^2\|=\max_n\tfrac1{n(n+1)}=\tfrac12$ (at $n=1$). Note $\|T^2\|\lt\|T\|^2$.
Answer (A), (B) and (D).

E5. GATE MA · NAT Let $\phi:\ell^{3/2}\to\mathbb{R}$, $\phi(x)=\sum_{n=1}^{\infty}x_n\,2^{-n}$. The value of $\|\phi\|$, correct to two decimal places, is ______.

Idea $(\ell^{3/2})^*=\ell^3$ isometrically (Theorem 8.7): $\|\phi\|=\|(2^{-n})\|_3$.
Compute $\sum_n2^{-3n}=\sum_n8^{-n}=\tfrac{1/8}{1-1/8}=\tfrac17$, so $\|\phi\|=7^{-1/3}\approx0.5228$.
Answer $0.52$ (accept $0.51$ to $0.53$).

E6. GATE MA · NAT On $C[0,1]$ with the sup norm let $(Tf)(x)=\int_0^1\sin\big(\pi(x+t)\big)f(t)\,dt$. The operator norm $\|T\|$, correct to three decimal places, is ______.

Idea For a continuous kernel, $\|T\|=\max_x\int_0^1|K(x,t)|\,dt$ (Example 4(c)).
Compute For fixed $x$, substitute $s=x+t$: $\int_0^1|\sin\pi(x+t)|\,dt=\int_x^{x+1}|\sin\pi s|\,ds$. The integrand has period $1$, so this is $\int_0^1\sin\pi s\,ds=\tfrac2\pi$ for every $x$.
Answer $\|T\|=\tfrac2\pi\approx0.6366$: $0.637$ (accept $0.636$ to $0.637$). (A trap: $f\equiv1$ gives only $\max_x|\int_0^1\sin\pi(x+t)\,dt|$ $=\max_x\tfrac2\pi|\cos\pi x|=\tfrac2\pi$ here, by luck; for kernels that change sign the constant function usually underestimates.)
🧑‍🏫 For the teacher

Suggested plan — 3 lectures × 60 min.

  1. Banach spaces. Definitions 8.1–8.2 with the "zoo" of spaces on the board; Hölder and Minkowski (Theorem 8.1) in full; Theorem 8.2 and the $\ell^p$ completeness proof (the template for every completeness proof); Riesz–Fischer as an application of MCT/DCT. Example 1 as a class poll ("complete or not?") and Example 2 for the inclusions. Project the chapter teaser (unit balls) for two minutes at the start.
  2. Operators and duals. Theorems 8.4–8.6; Example 3 with Interactive 1 (ask the class to predict where the maximum is reached before switching $p$); Example 4 — insist on the "smoothed sign" trick and on checking attainment; Theorem 8.7 with the extremal vector $|y_n|^{q-1}\overline{\operatorname{sgn}y_n}$; Example 5 done as rapid NAT drill.
  3. Hahn–Banach. Theorem 8.8 (Step 1 slowly, Zorn quickly), Theorem 8.9, Example 6 with Interactive 2 in extension mode at $p=\infty$, angle $45^\circ$; Theorem 8.10 (prove (1), (3), (4)); separation (Theorem 8.11) with the widget in separation mode; Example 7. End with GATE corner E3 and E6 live.

Discussion questions. (1) Why does the proof that $(\ell^p)^*=\ell^q$ break down for $p=\infty$, and which step uses $p\lt\infty$? (2) Hahn–Banach produces a functional on $\ell^\infty$ that vanishes on $c_0$; can you "write it down"? What does that say about proofs that use Zorn's lemma? (3) In $\mathbb{R}^2$, for which norms are all Hahn–Banach extensions unique? Relate the answer to corners of the unit ball.

Misconceptions to watch for. "Every linear map is continuous"; "$\ell^p$ gets smaller as $p$ grows" (it gets larger; $L^p[0,1]$ gets smaller); "the norm of a functional is always attained"; using $f\equiv1$ as the extremal function when the kernel changes sign; "the dual of $\ell^\infty$ is $\ell^1$"; "Hahn–Banach gives a unique extension"; forgetting that $X^*$ is complete even when $X$ is not.

Suggested homework. I(1, 3, 4, 6, 9), II(1, 2, 5, 6, 7, 9), III(1, 2, 3, 7), IV(1, 3, 4, 6); GATE corner E1, E4, E5.

🔬 Research–Teaching NexusResearch-tutored · optional, about 60 minutes

Research question. How does the norm $\|A\|_{p}$ of a fixed matrix, acting on $(\mathbb{R}^n,\|\cdot\|_p)$, change as $p$ runs from $1$ to $\infty$ — and is there a hidden regularity in the curve?

Why it matters. Engineers who design digital filters, image-deblurring kernels or numerical schemes for PDEs need to know how much an operator can amplify a signal measured in different norms: $\ell^\infty$ (peak value, relevant for clipping in audio and for maximum stress in structures), $\ell^2$ (energy) or $\ell^1$ (total variation, used in compressed sensing for medical MRI). The Riesz–Thorin interpolation theorem says that $\log\|A\|_p$ is a convex function of $1/p$ — so the cheap end-point norms $\|A\|_1$ (column sums) and $\|A\|_\infty$ (row sums) control every $\|A\|_p$ in between.

What to do.

  1. Write a short program that estimates $\|A\|_p=\max_{\|x\|_p=1}\|Ax\|_p$ for a $2\times2$ matrix by sampling the unit sphere finely (parametrise as in Interactive 1), and for a $3\times3$ matrix by random search plus local refinement.
  2. For $A=\begin{pmatrix}1&2\\0&2\end{pmatrix}$ tabulate $\|A\|_p$ for $1/p=0,\tfrac14,\tfrac12,\tfrac34,1$ and plot $\log\|A\|_p$ against $1/p$. Check convexity (second differences) and the Riesz–Thorin bound $\|A\|_p\le\|A\|_1^{1/p}\|A\|_\infty^{1-1/p}$.
  3. For a symmetric matrix, show that $\|A\|_p=\|A\|_q$ for conjugate exponents (use $\|A^T\|_q=\|A\|_p$, a consequence of Exercise IV.7) and verify it numerically; where is the minimum of the curve?
  4. Repeat for a $1$-D blurring kernel (a circulant matrix with rows $\tfrac14,\tfrac12,\tfrac14$) and explain why its norm is $1$ for every $p$ (Young's inequality $\|k*x\|_p\le\|k\|_1\|x\|_p$).

Tools: Python with NumPy (or any language), a plotting tool.

Hand in: A one-page note with your table and plot, the convexity check, and one sentence explaining what an engineer gains from knowing $\|A\|_1$ and $\|A\|_\infty$ only.

Going further. Read about the Riesz–Thorin theorem and its proof by the three-lines lemma from complex analysis; computing $\|A\|_p$ exactly is NP-hard in general for $p\ne1,2,\infty$, which is why the interpolation bound matters in practice.

Instructor note: for the matrix of step 2 the values are $\|A\|_\infty=3$, $\|A\|_4\approx2.7327$, $\|A\|_2\approx2.9208$, $\|A\|_{4/3}\approx3.3669$, $\|A\|_1=4$ (checked numerically). The logarithms $1.0986,\ 1.0053,\ 1.0719,\ 1.2140,\ 1.3863$ have positive second differences, and the curve has its minimum strictly inside — at about $p\approx4$ — not at $p=2$, because the matrix is not symmetric. For a symmetric matrix the curve is symmetric about $1/p=\tfrac12$.

🧪 Self-check quiz

0/5 correct

Q1. Which of these normed spaces is not complete?

Q2. The dual space of $\ell^3$ is (isometrically isomorphic to)

Q3. The norm of $f\mapsto f(0)+f(1)$ on $C[0,1]$ (sup norm) is

Q4. The Hahn–Banach theorem guarantees that

Q5. For a linear map $T$ between normed spaces, boundedness is equivalent to