Functional Analysis: Harmonic Analysis¶

Job: develop harmonic analysis on groups and interpolation at frontier depth — locally compact abelian (LCA) groups and Pontryagin duality, the Hausdorff–Young theorem, the Calderón–Zygmund singular-integral theory (decomposition + $L^p$ boundedness), the T1 theorem of David–Journé, and the Hardy–Littlewood maximal function with the Lebesgue differentiation theorem.

Core¶

LCA groups and Pontryagin duality¶

A locally compact abelian (LCA) group $G$ is a locally compact abelian topological group (with Haar measure $dx$). The dual group $\hat G=\mathrm{Hom}(G,\mathbb{T})$ is the group of continuous characters $\chi:G\to\mathbb{T}$ (the unit circle), with pointwise multiplication and the compact-open topology. The Fourier transform of $f\in L^1(G)$ is $\hat f(\chi)=\int_G f(x)\overline{\chi(x)}\,dx\in\hat G$.

Theorem (Pontryagin duality). The natural map $G\to\widehat{\hat G}$, $x\mapsto(\chi\mapsto\chi(x))$, is an isomorphism of topological groups: $G\cong\widehat{\hat G}$. The Fourier transform is an $L^2$ isometry (the Plancherel theorem): for $f\in L^1\cap L^2(G)$, $\|f\|_{L^2(G)}=\|\hat f\|_{L^2(\hat G)}$ (with Haar measures normalized), extending to a unitary $L^2(G)\to L^2(\hat G)$. The Fourier inversion $f(x)=\int_{\hat G}\hat f(\chi)\chi(x)\,d\chi$ holds for $f\in L^1(G)$ with $\hat f\in L^1(\hat G)$.

LCA group $G$	Dual $\hat G$	Haar / Fourier
$\mathbb{R}^n$	$\mathbb{R}^n$	Lebesgue; $\hat f(\xi)=\int f(x)e^{-ix\xi}\,dx$
$\mathbb{T}$	$\mathbb{Z}$	normalized Haar on $\mathbb{T}$; $\hat f(n)=\frac1{2\pi}\int_0^{2\pi}f(e^{i\theta})e^{-in\theta}\,d\theta$ (Fourier series)
$\mathbb{Z}$	$\mathbb{T}$	counting measure; $\hat f(z)=\sum_n f(n)z^n$ (generating function / Laurent series)
$\mathbb{Z}/N\mathbb{Z}$	$\mathbb{Z}/N\mathbb{Z}$	counting; the finite DFT
$\mathbb{Q}_p$ (p-adics)	$\mathbb{Q}_p$	self-dual (the p-adic Fourier transform)
finite abelian $G$	the dual finite abelian group	character table; Pontryagin duality $G\cong\hat{\hat G}$

The Hausdorff–Young theorem¶

Theorem (Hausdorff–Young). For $1\le p\le 2$ and $1/p+1/q=1$ (so $q\ge 2$), the Fourier transform on an LCA group satisfies $$ |\hat f|{L^q(\hat G)}\le|f|. $$ The bound is sharp at the endpoints (Beckner–Babenko sharp Hausdorff–Young: $\|\hat f\|_{L^{q}}\le (p^{1/p}/q^{1/q})^{n/2}\|f\|_{L^p}$ on $\mathbb{R}^n$). The proof interpolates between the $L^1\to L^\infty$ bound $\|\hat f\|_\infty\le\|f\|_1$ and the $L^2\to L^2$ Plancherel isometry via the Riesz–Thorin theorem.

Interpolation theorem	Method	Hypotheses	Conclusion
Riesz–Thorin (complex)	complex interpolation	$T$ strong-type $(p_0,q_0)$ and $(p_1,q_1)$	$T$ strong-type $(p_\theta,q_\theta)$, $\\|T\\|\le\\|T\\|_0^{1-\theta}\\|T\\|_1^\theta$
Marcinkiewicz (real)	real interpolation	$T$ weak-type $(p_0,q_0)$ and $(p_1,q_1)$	$T$ strong-type $(p_\theta,q_\theta)$ (with $p_i<q_i$)
Stein (analytic family)	complex, analytic family $T_z$	$T_z$ bounded at $\mathrm{Re}\,z=0,1$	$T_\theta$ bounded at $\mathrm{Re}\,z=\theta$
Hausdorff–Young	Riesz–Thorin	$\hat{\cdot}:L^1\to L^\infty$ and $L^2\to L^2$	$\hat{\cdot}:L^p\to L^{p'}$ for $1\le p\le 2$

Calderón–Zygmund singular integrals¶

A Calderón–Zygmund kernel $K:\mathbb{R}^n\setminus\{0\}\to\mathbb{C}$ is a function homogeneous of degree $-n$ ($K(\lambda x)=\lambda^{-n}K(x)$), smooth off $0$, with zero average $\int_{S^{n-1}}K\,d\sigma=0$. The associated singular integral operator is $Tf(x)=\mathrm{p.v.}\int K(x-y)f(y)\,dy$ (the principal value).

Theorem (Calderón–Zygmund decomposition). For $f\in L^1(\mathbb{R}^n)$ and $\lambda>0$, there is a decomposition $f=g+b$ with $g\in L^1\cap L^\infty$, $|g|\le 2^n\lambda$ a.e., and $b=\sum_j b_j$ supported on disjoint dyadic cubes $Q_j$ with $\int_{Q_j}b_j=0$ (the cancellation), $\|b_j\|_{L^1}\le 2^n\lambda|Q_j|$, and $\sum_j|Q_j|\le\|f\|_{L^1}/\lambda$. The decomposition splits $f$ into a "good" bounded part $g$ and a "bad" part $b$ supported on small cubes with zero mean — the input to the $L^p$ theory.

Theorem (Calderón–Zygmund $L^p$). For a CZ singular integral $T$ with kernel $K$ as above, $T$ is bounded on $L^p(\mathbb{R}^n)$ for $1<p<\infty$, with $\|Tf\|_{L^p}\le C_p\|f\|_{L^p}$. The bound fails at $p=1$ and $p=\infty$: at $p=1$, the weak-type bound $\|Tf\|_{L^{1,\infty}}\le C\|f\|_{L^1}$ holds (the CZ decomposition gives the weak-$L^1$ estimate); at $p=\infty$, the bound is replaced by BMO (the second derivatives of $-\Delta^{-1}$ are in BMO when $f\in L^\infty$ — the John–Nirenberg / BMO estimate, see Function Spaces).

Theorem (T1, David–Journé 1984). A CZ operator $T$ (with a kernel satisfying the standard size and smoothness estimates) is bounded on $L^2(\mathbb{R}^n)$ iff the T1 condition holds: $T(1),T^*(1)\in\mathrm{BMO}(\mathbb{R}^n)$ and the weak boundedness property ($|T(\varphi_{x,r})(\psi_{x,r})|\le C$ for normalized bumps $\varphi_{x,r},\psi_{x,r}$ at all scales). The T1 theorem is the criterion for $L^2$ boundedness of singular integrals without explicit kernel bounds — it characterizes the bounded CZ operators by the BMO condition on $T(1)$ and $T^*(1)$.

Maximal functions and differentiation¶

Theorem (Hardy–Littlewood maximal). The maximal function $Mf(x)=\sup_{r>0}\frac1{|B(x,r)|}\int_{B(x,r)}|f(y)|\,dy$ satisfies $\|Mf\|_{L^p}\le C_p\|f\|_{L^p}$ for $1<p\le\infty$ (the strong-$L^p$ bound) and the weak-$L^1$ bound $\|Mf\|_{L^{1,\infty}}=\sup_\lambda\lambda|\{Mf>\lambda\}|\le C\|f\|_{L^1}$. The constant $C_p\sim 1/(p-1)$ as $p\to 1^+$.

Theorem (Lebesgue differentiation). For $f\in L^1_{\mathrm{loc}}(\mathbb{R}^n)$, $\frac1{|B(x,r)|}\int_{B(x,r)}f(y)\,dy\to f(x)$ as $r\to 0$ for a.e. $x$, equivalently $Mf$ (the maximal function) controls the a.e. convergence of the averages. The maximal function is the operator that encodes the differentiation theorem — the $L^p$ bounds on $M$ are the content of the differentiation theorem.

Theorem (Coifman–McIntosh–Meyer, Cauchy integral on Lipschitz curves). The Cauchy integral $C_\varphi f(x)=\mathrm{p.v.}\frac1{\pi i}\int\frac{f(y)}{y-x+i(\varphi(x)-\varphi(y))}\,dy$ on a Lipschitz curve $\Gamma=\{x+i\varphi(x):\varphi\in \mathrm{Lip}\}$ is bounded on $L^p(\Gamma)$ for $1<p<\infty$, with the bound $\|C_\varphi f\|_{L^p}\le C_p\|\varphi\|_{\mathrm{Lip}}\|f\|_{L^p}$. This resolved the Calderón conjecture (the Cauchy integral on Lipschitz curves is a CZ operator), the foundational result for the $L^p$ theory of singular integrals on non-smooth curves.

See it¶

The Calderón–Zygmund decomposition and the weak-$L^1$ bound for a CZ singular integral. Let $f\in L^1(\mathbb{R}^n)$ and $\lambda>0$. The CZ decomposition:

Apply the dyadic stopping-time / the one-third trick: select a maximal disjoint family of dyadic cubes $\{Q_j\}$ with $\frac1{|Q_j|}\int_{Q_j}|f|>\lambda$ (the maximal cubes where the average exceeds $\lambda$). By the Hardy–Littlewood weak-$L^1$ bound (or directly), $\sum_j|Q_j|\le\|f\|_{L^1}/\lambda$.
Define $b_j=(f-\frac1{|Q_j|}\int_{Q_j}f)\,\mathbf{1}_{Q_j}$ (the "bad" part on $Q_j$); by construction $\int_{Q_j}b_j=0$ (the cancellation) and $\|b_j\|_{L^1}\le 2\int_{Q_j}|f|\le 2\lambda|Q_j|$, so $\|b\|_{L^1}\le 2\lambda\sum_j|Q_j|\le 2\|f\|_{L^1}$.
Define $g=f-b$ (the "good" part); $g$ satisfies $|g|\le 2^n\lambda$ a.e. (the average over the parent cube is $\le 2^n\lambda$ by maximality) and $\|g\|_{L^1}\le\|f\|_{L^1}$.

The decomposition $f=g+b$ splits $f$ into a bounded piece $g$ (handled by the $L^2$ theory: $\|Tg\|_{L^2}\le C\|g\|_{L^2}\le C\lambda^{1/2}\|g\|_{L^1}^{1/2}\le C\lambda^{1/2}\|f\|_{L^1}^{1/2}$, by interpolation) and a piece $b$ supported on small cubes with zero mean (handled by the cancellation: $\int_{Q_j}b_j=0$ makes the principal-value integral over $Q_j$ bounded by the smoothness of $K$, giving $\|Tb\|_{L^1}\le C\sum_j|Q_j|\lambda\le C\|f\|_{L^1}$ via the geometric / dyadic estimates). Combining, the weak-$L^1$ bound $\|Tf\|_{L^{1,\infty}}=\sup_\lambda\lambda|\{|Tf|>\lambda\}|\le C\|f\|_{L^1}$ follows from $\|Tg\|_{L^2}$ (Chebyshev) and $\|Tb\|_{L^1}$ (the dyadic / geometric estimate), giving the weak-$L^1$ estimate for the CZ operator. This single decomposition — the splitting of $f$ into a bounded part and a cancellation-supported part, with the $L^2$ and cancellation arguments applied respectively — is the operational content of the Calderón–Zygmund theory: the $L^p$ boundedness ($1<p<\infty$) follows by interpolating the weak-$L^1$ and $L^2$ bounds (Marcinkiewicz), and the failure at $L^1,L^\infty$ is the obstruction that the weak-$L^1$ and BMO estimates replace.

Check¶

Explain why the Calderón–Zygmund $L^p$ theorem fails at $p=1$ and $p=\infty$, and identify the substitutes at each endpoint.

The $L^1$ failure: the CZ kernel $K(z)=z/|z|^n$ (a Hessian Riesz transform, $n=1$ analog: the Hilbert transform $Hf=\mathrm{p.v.}\frac1\pi\int\frac{f(y)}{x-y}dy$) is unbounded on $L^1$ — concretely $H\chi_{[0,1]}(x)=\frac1\pi\log|\frac{x}{x-1}|$ is unbounded at $0,1$, so $\|H\chi_{[0,1]}\|_\infty=\infty$ while $\|\chi_{[0,1]}\|_1=1$; the same phenomenon occurs in $n$ dimensions for the Riesz transforms. The substitute at $L^1$ is the weak-$L^1$ bound $\|Tf\|_{L^{1,\infty}}\le C\|f\|_{L^1}$ (the CZ decomposition gives this — the singular integral maps $L^1$ to weak-$L^1$ but not to $L^1$). The $L^\infty$ failure: the Riesz transform of an $L^\infty$ function is unbounded (the same example, $H\chi_{[0,1]}$ is unbounded), but the second derivatives of $-\Delta^{-1}f$ for $f\in L^\infty$ are in BMO (the John–Nirenberg inequality / the CZ theory gives the BMO estimate), not $L^\infty$. Hence the substitutes: weak-$L^1$ at $p=1$, BMO at $p=\infty$, with the $L^p$ boundedness for $1<p<\infty$ by interpolation (Riesz–Thorin / Marcinkiewicz). The $L^1$ and $L^\infty$ failures are the same phenomenon: the CZ singular integral amplifies the endpoint $L^p$ spaces to unbounded functions, and the weak-$L^1$ / BMO are the correct endpoint substitutes.

State the T1 theorem and explain why the BMO condition (not just kernel estimates) is necessary for $L^2$ boundedness.

The T1 theorem (David–Journé): a CZ operator $T$ (with a kernel $K(x,y)$ satisfying the standard size $|K(x,y)|\le C|x-y|^{-n}$ and smoothness $|K(x,y)-K(x',y)|\le C|x-x'||x-y|^{-n-1}$ estimates) is bounded on $L^2$ iff (i) $T(1),T^*(1)\in\mathrm{BMO}$ and (ii) the weak boundedness property ($|T(\varphi_{x,r})(\psi_{x,r})|\le C$ for normalized bumps at all scales). The BMO condition is necessary because the kernel estimates alone do not control the action of $T$ on the constant function $1$: a CZ operator with a good kernel may map $1$ to an unbounded function (the kernel smoothness controls $Tf$ for $f$ with compact support, but the constant $1$ has non-compact support and the kernel estimates do not bound $T(1)$). The BMO condition $T(1)\in\mathrm{BMO}$ is the precise regularity needed — BMO is the largest space to which $T$ can map $L^\infty$ (the endpoint of the $L^p$ scale), and the T1 theorem reduces the $L^2$ boundedness to checking $T(1)$ and $T^*(1)$ are in BMO. The Cauchy integral on a Lipschitz curve (Coifman–McIntosh–Meyer) is the prototypical application: the T1 theorem gives the $L^2$ boundedness (hence $L^p$ by interpolation) once $T(1)\in\mathrm{BMO}$ is verified, which reduces to the Lipschitz regularity of the curve.

Pitfalls¶

The CZ theorem fails at $L^1$/$L^\infty$ (replaced by weak-$L^1$ / BMO): the $L^p$ boundedness holds for $1<p<\infty$; the endpoint substitutes are weak-$L^1$ (at $p=1$) and BMO (at $p=\infty$). Citing $L^\infty$ boundedness of a CZ operator is wrong.
The T1 theorem needs the BMO condition (not just kernel estimates): the David–Journé criterion requires $T(1),T^*(1)\in\mathrm{BMO}$ and the weak boundedness property; kernel estimates alone (the standard size + smoothness) do not give $L^2$ boundedness without the BMO / WBP conditions. The BMO is the endpoint regularity.
Pontryagin duality is for LCA groups, not for non-abelian groups: for non-abelian locally compact groups, the dual is not a group but a representation category (the unitary dual $\hat G$ with the Fell topology); the Fourier transform is operator-valued (the Plancherel theorem for non-abelian groups is more subtle — e.g. the Plancherel theorem for reductive $p$-adic groups, the Harish-Chandra / Waldspurger theorem).
The Hausdorff–Young theorem holds only for $1\le p\le 2$ (not for $p>2$): the Fourier transform $L^p\to L^{p'}$ with $p'\ge 2$ is bounded only for $p\le 2$; for $p>2$ the Fourier transform of an $L^p$ function is a distribution (not a function), and the bound fails. The interpolation (Riesz–Thorin) goes from $L^1\to L^\infty$ and $L^2\to L^2$ to the intermediate $L^p\to L^{p'}$ ($1<p<2$), and the direction $p\to p'$ is fixed.
The Hardy–Littlewood maximal function is unbounded on $L^1$ (only weak-$L^1$): $\|Mf\|_{L^1}\le C\|f\|_{L^1}$ fails (e.g. $f=\chi_{[0,1]}$ gives $Mf(x)\sim 1/|x|$ for large $x$, not in $L^1$); the weak-$L^1$ bound is the correct one, and it is the substitute that gives the Lebesgue differentiation theorem.
The Coifman–McIntosh–Meyer theorem is for Lipschitz curves (not for general / rough curves): the Cauchy integral is bounded on $L^p$ for Lipschitz curves; for rougher curves (e.g. chord-arc / general rectifiable curves), the $L^p$ boundedness is more subtle (the David semmes / the T(b) theorem).

Refs¶

Folland, A Course in Abstract Harmonic Analysis (LCA groups, Pontryagin duality); Rudin, Fourier Analysis on Groups
Stein, Singular Integrals and Differentiability Properties of Functions; Harmonic Analysis (CZ, maximal functions, T1)
Grafakos, Classical / Modern Fourier Analysis (CZ decomposition, T1, Hausdorff–Young, interpolation)
David, Wavelets and Singular Integrals on Curves and Surfaces; Journé, Calderón–Zygmund Operators; Coifman–McIntosh–Meyer (Cauchy integral)

Next¶

Operator Algebras · Spectral Theory · Function Spaces · FA index · Foundations · Fourier analysis · PDE / elliptic theory · Measure theory · MT / harmonic & singular — the maximal / the CZ / the BMO / the $A_p$ · FA / Hardy & BMO

LCA group \(G\)	Dual \(\hat G\)	Haar / Fourier
\(\mathbb{R}^n\)	\(\mathbb{R}^n\)	Lebesgue; \(\hat f(\xi)=\int f(x)e^{-ix\xi}\,dx\)
\(\mathbb{T}\)	\(\mathbb{Z}\)	normalized Haar on \(\mathbb{T}\); \(\hat f(n)=\frac1{2\pi}\int_0^{2\pi}f(e^{i\theta})e^{-in\theta}\,d\theta\) (Fourier series)
\(\mathbb{Z}\)	\(\mathbb{T}\)	counting measure; \(\hat f(z)=\sum_n f(n)z^n\) (generating function / Laurent series)
\(\mathbb{Z}/N\mathbb{Z}\)	\(\mathbb{Z}/N\mathbb{Z}\)	counting; the finite DFT
\(\mathbb{Q}_p\) (p-adics)	\(\mathbb{Q}_p\)	self-dual (the p-adic Fourier transform)
finite abelian \(G\)	the dual finite abelian group	character table; Pontryagin duality \(G\cong\hat{\hat G}\)