MathLabs

Analysis

Singular integrals and Calderón–Zygmund theory

The study of integral operators with singular kernels, central to modern harmonic analysis and PDE.

IntuitionA kernel too singular to be an ordinary integral

A photo-editing 'sharpen' filter, an audio phase-shifter, and a crack in an elastic material all hide the same mathematical object: a convolution against a kernel that blows up at the origin, like 1/x1/x in one dimension. Such a kernel is not absolutely integrable near 0, so ∫K(x−y)f(y) dy\int K(x-y)f(y)\,dy does not converge in the ordinary sense — yet the positive and negative parts of the kernel on either side of the singularity cancel almost perfectly, and a principal-value limit exists for reasonable ff. Singular integral operators are exactly the operators built this way, and Calderón–Zygmund theory is the machinery that makes sense of them and proves they behave as well as one could hope.

Interactive complex-domain widget visualizing a holomorphic map on the upper half-plane, related to the Hilbert transform.
The Hilbert transform arises as the boundary values of a holomorphic function on the upper half-plane; explore how this complex map behaves as you zoom near the boundary.

UndergraduateThe prototype: the Hilbert transform

Definition: Hilbert transform

For a nice function ff on R\mathbb{R}, its Hilbert transform is the principal-value convolution Hf(x)=p.v. 1π∫−∞∞f(y)x−y dyHf(x) = \text{p.v.}\ \dfrac{1}{\pi}\int_{-\infty}^{\infty} \dfrac{f(y)}{x-y}\,dy: the singularity at y=xy=x is excised symmetrically before taking the limit, which is exactly what lets the cancellation between y<xy<x and y>xy>x produce a finite value.

Hf(x)=p.v. 1π∫−∞∞f(y)x−y dyHf(x) = \text{p.v.}\ \dfrac{1}{\pi}\int_{-\infty}^{\infty} \dfrac{f(y)}{x-y}\,dy

On the Fourier side the Hilbert transform is remarkably simple: it multiplies each frequency by a pure phase, Hf^(ξ)=−i sgn⁡(ξ) f^(ξ)\widehat{Hf}(\xi) = -i\,\operatorname{sgn}(\xi)\, \hat f(\xi), flipping sign at ξ=0\xi=0. Every frequency is rotated by 90∘90^\circ (with the sign of the rotation depending on whether the frequency is positive or negative) but not rescaled — which already hints that HH should be an isometry on L2L^2.

Hf^(ξ)=−i sgn⁡(ξ) f^(ξ)\widehat{Hf}(\xi) = -i\,\operatorname{sgn}(\xi)\, \hat f(\xi)

Definition: Calderón–Zygmund operator

More generally, a Calderón–Zygmund operator is a convolution Tf(x)=p.v.∫RnK(x−y)f(y) dyTf(x) = \text{p.v.}\int_{\mathbb{R}^n} K(x-y) f(y)\,dy on Rn\mathbb{R}^n whose kernel KK satisfies a size bound and a smoothness (Hörmander) bound away from the origin, ∣K(x)∣≤C∣x∣n,∣∇K(x)∣≤C∣x∣n+1|K(x)| \le \dfrac{C}{|x|^n}, \qquad |\nabla K(x)| \le \dfrac{C}{|x|^{n+1}}, together with the assumption that TT is already known to be bounded on L2L^2. The Hilbert transform is the case n=1n=1, K(x)=1/(πx)K(x)=1/(\pi x); the Riesz transforms Rjf=p.v. cn∫xj−yj∣x−y∣n+1f(y) dyR_j f = \text{p.v.}\, c_n \int \frac{x_j-y_j}{|x-y|^{n+1}} f(y)\,dy are the higher-dimensional analogues used below in elliptic PDE.

From one dimension to many
OperatorDimensionKernel K(x)K(x)Typical use
Hilbert transformn=1n=11/(πx)1/(\pi x)Analytic signal, phase shifting
Riesz transformsn≥2n\ge 2(xj−yj)/∣x−y∣n+1(x_j-y_j)/|x-y|^{n+1}Second-derivative estimates for elliptic PDE
General Calderón–Zygmund TTn≥1n\ge 1Size + Hörmander smoothnessCommon framework for all of the above

ResearchKey theorems

For every f∈L2(R)f \in L^2(\mathbb{R}), ∥Hf∥L2=∥f∥L2\|Hf\|_{L^2} = \|f\|_{L^2}: the Hilbert transform is not merely bounded but an isometry on L2L^2.

Why is it true?

This is the base case every general Calderón–Zygmund theorem builds on: before controlling a singular integral operator on L1L^1 or LpL^p, one first needs it to be well-behaved on L2L^2, where Plancherel gives direct access via the Fourier transform.

Proof

Step 1 (find the multiplier via a regularization). The kernel 1/(πx)1/(\pi x) is not integrable, so approximate it: for ϵ>0\epsilon>0 let kϵ(x)=1πxx2+ϵ2k_\epsilon(x) = \frac{1}{\pi}\frac{x}{x^2+\epsilon^2}, an odd, integrable approximation that converges to the principal-value kernel as ϵ→0\epsilon\to 0. A direct (contour or tables) computation gives k^ϵ(ξ)=−i sgn⁡(ξ) e−2πϵ∣ξ∣\hat k_\epsilon(\xi) = -i\,\operatorname{sgn}(\xi)\, e^{-2\pi\epsilon|\xi|}.

Step 2 (pass to the limit). As ϵ→0+\epsilon \to 0^+, e−2πϵ∣ξ∣→1e^{-2\pi\epsilon|\xi|} \to 1 for every fixed ξ≠0\xi \ne 0, and kϵ∗f→Hfk_\epsilon * f \to Hf for nice ff. Passing to the limit inside the convolution theorem gives the multiplier identity Hf^(ξ)=−i sgn⁡(ξ) f^(ξ)\widehat{Hf}(\xi) = -i\,\operatorname{sgn}(\xi)\, \hat f(\xi): the Hilbert transform acts on the frequency side by multiplication by −i sgn⁡(ξ)-i\,\operatorname{sgn}(\xi).

Step 3 (apply Plancherel). Since ∣−i sgn⁡(ξ)∣=1|-i\,\operatorname{sgn}(\xi)| = 1 for every ξ≠0\xi \ne 0 (a single point has measure zero and does not affect the integral), ∣Hf^(ξ)∣2=∣f^(ξ)∣2|\widehat{Hf}(\xi)|^2 = |\hat f(\xi)|^2 for a.e. ξ\xi. Integrating and invoking Plancherel's theorem on both sides gives ∥Hf∥L2=∥f∥L2\|Hf\|_{L^2} = \|f\|_{L^2}: the Hilbert transform is an isometry on L2(R)L^2(\mathbb{R}), in particular bounded.

If a convolution operator TT is bounded on L2(Rn)L^2(\mathbb{R}^n) and its kernel satisfies the Calderón–Zygmund conditions ∣K(x)∣≤C∣x∣n,∣∇K(x)∣≤C∣x∣n+1|K(x)| \le \dfrac{C}{|x|^n}, \qquad |\nabla K(x)| \le \dfrac{C}{|x|^{n+1}}, then TT is of weak type (1,1), ∣{x:∣Tf(x)∣>λ}∣≤Cλ∥f∥L1\big|\{x : |Tf(x)| > \lambda\}\big| \le \dfrac{C}{\lambda}\|f\|_{L^1}, and consequently (by the Marcinkiewicz interpolation theorem, plus duality) TT is bounded on Lp(Rn)L^p(\mathbb{R}^n) for every 1<p<∞1<p<\infty.

Why is it true?

This single theorem explains why singular integrals — despite their kernels not being absolutely integrable — behave like ordinary bounded operators on almost every LpL^p space: the only casualty is the endpoint p=1p=1, where boundedness weakens from strong to merely weak type.

Proof

Step 1 (Calderón–Zygmund decomposition). Fix f∈L1f \in L^1 and λ>0\lambda>0. Using a stopping-time argument on dyadic cubes (equivalently, the Hardy–Littlewood maximal function MfMf), split f=g+bf = g + b where the 'good' part satisfies ∥g∥∞≤Cλ\|g\|_\infty \le C\lambda and ∥g∥1≤∥f∥1\|g\|_1 \le \|f\|_1, and the 'bad' part b=∑jbjb = \sum_j b_j is a sum of pieces supported on pairwise disjoint dyadic cubes QjQ_j with mean zero, ∫Qjbj=0\int_{Q_j} b_j = 0, and total measure ∑j∣Qj∣≤Cλ∥f∥1\sum_j |Q_j| \le \frac{C}{\lambda}\|f\|_1 (this is exactly the weak-(1,1) bound for MM applied to the stopping cubes).

Step 2 (the good part). Since TT is bounded on L2L^2 and ∥g∥22≤∥g∥∞∥g∥1≤Cλ∥f∥1\|g\|_2^2 \le \|g\|_\infty \|g\|_1 \le C\lambda\|f\|_1, Chebyshev's inequality gives ∣{∣Tg∣>λ/2}∣≤4λ2∥Tg∥22≤Cλ∥f∥1|\{|Tg| > \lambda/2\}| \le \frac{4}{\lambda^2}\|Tg\|_2^2 \le \frac{C}{\lambda}\|f\|_1: the good part alone already satisfies the weak-(1,1) bound.

Step 3 (the bad part, off the doubled cubes). Let Qj∗Q_j^* be the cube with the same center as QjQ_j and twice the side length; the union ⋃jQj∗\bigcup_j Q_j^* has measure at most Cλ∥f∥1\frac{C}{\lambda}\|f\|_1 as well. Away from Qj∗Q_j^*, the mean-zero condition on bjb_j lets one subtract a constant from the kernel: Tbj(x)=∫Qj(K(x−y)−K(x−cj))bj(y) dyTb_j(x) = \int_{Q_j} \big(K(x-y)-K(x-c_j)\big) b_j(y)\,dy for the center cjc_j of QjQ_j, and the Hörmander smoothness bound ∣K(x)∣≤C∣x∣n,∣∇K(x)∣≤C∣x∣n+1|K(x)| \le \dfrac{C}{|x|^n}, \qquad |\nabla K(x)| \le \dfrac{C}{|x|^{n+1}} shows this difference is small enough that ∫(Qj∗)c∣Tbj(x)∣ dx≤C∥bj∥1\int_{(Q_j^*)^c} |Tb_j(x)|\,dx \le C\|b_j\|_1.

Step 4 (combine). Summing Step 3 over jj and applying Chebyshev off the doubled cubes gives ∣{x∉⋃jQj∗:∣Tb(x)∣>λ/2}∣≤Cλ∑j∥bj∥1≤Cλ∥f∥1\big|\{x \notin \bigcup_j Q_j^* : |Tb(x)|>\lambda/2\}\big| \le \frac{C}{\lambda}\sum_j \|b_j\|_1 \le \frac{C}{\lambda}\|f\|_1. Adding the measure of ⋃jQj∗\bigcup_j Q_j^* from Step 1, the bad part contributes O(∥f∥1/λ)O(\|f\|_1/\lambda) to the level set as well, and combining with Step 2 for the good part gives the full estimate ∣{x:∣Tf(x)∣>λ}∣≤Cλ∥f∥L1\big|\{x : |Tf(x)| > \lambda\}\big| \le \dfrac{C}{\lambda}\|f\|_{L^1}.

UndergraduateReal-World Applications and Worked Examples

Calderón–Zygmund theory earns its central place in analysis by controlling two very different-looking objects with the same tool: the second derivatives of solutions to elliptic PDE, and the 'quadrature' signal engineers need to extract instantaneous amplitude and phase from a real-world waveform.

Example: Second derivatives of the Newtonian potential

In the theory of elliptic PDE, one wants to solve Δu=f\Delta u = f on Rn\mathbb{R}^n and control the second derivatives of uu in terms of ff. If f∈Lp(Rn)f \in L^p(\mathbb{R}^n) for some 1<p<∞1<p<\infty, why does ∥D2u∥Lp≤Cp∥f∥Lp\|D^2 u\|_{L^p} \le C_p \|f\|_{L^p} hold?

Solution

The solution is given by convolution with the Newtonian potential, u=f∗Nu = f * N where N(x)∼cn∣x∣2−nN(x) \sim c_n|x|^{2-n} for n≥3n\ge 3. Differentiating twice under the integral sign formally gives ∂i∂ju=f∗(∂i∂jN)\partial_i\partial_j u = f * (\partial_i\partial_j N), but ∂i∂jN(x)\partial_i\partial_j N(x) is homogeneous of degree −n-n with mean zero on spheres around the origin — exactly a Calderón–Zygmund kernel, up to a Dirac mass term that only matters for i=ji=j.

So away from that constant multiple of ff itself, ∂i∂ju=Rijf\partial_i \partial_j u = R_{ij}f where RijR_{ij} is (a Riesz-transform-type) Calderón–Zygmund operator with a kernel satisfying the size and Hörmander smoothness conditions.

By the Calderón–Zygmund theorem, RijR_{ij} is bounded on LpL^p for every 1<p<∞1<p<\infty, so ∥∂i∂ju∥Lp≤C∥f∥Lp\|\partial_i\partial_j u\|_{L^p} \le C\|f\|_{L^p} for each pair i,ji,j, which is exactly ∥D2u∥Lp≤Cp∥f∥Lp\|D^2 u\|_{L^p} \le C_p \|f\|_{L^p}. This is the classical Calderón–Zygmund estimate that underlies elliptic regularity theory: it is precisely what fails at p=1p=1 and p=∞p=\infty, forcing PDE theorists to work in LpL^p, 1<p<∞1<p<\infty, or in Hölder spaces instead.

Example: Building the analytic signal for a radio engineer

A communications engineer has a real carrier signal f(t)=cos⁡(2πνt)f(t) = \cos(2\pi\nu t) with ν>0\nu>0, and wants its 90°-shifted 'quadrature' companion to build the analytic signal f(t)+iHf(t)f(t) + iHf(t) used to extract instantaneous amplitude and phase (as in single-sideband radio and envelope detectors). Compute HfHf.

Solution

Write cos⁡(2πνt)=12e2πiνt+12e−2πiνt\cos(2\pi\nu t) = \tfrac12 e^{2\pi i \nu t} + \tfrac12 e^{-2\pi i \nu t}, whose Fourier transform is a pair of spikes 12δν+12δ−ν\tfrac12\delta_\nu + \tfrac12\delta_{-\nu} at ξ=±ν\xi = \pm\nu.

Apply the Hilbert transform multiplier Hf^(ξ)=−i sgn⁡(ξ) f^(ξ)\widehat{Hf}(\xi) = -i\,\operatorname{sgn}(\xi)\, \hat f(\xi): at ξ=ν>0\xi=\nu>0 it multiplies by −i-i, and at ξ=−ν<0\xi=-\nu<0 it multiplies by +i+i. So on the frequency side the transformed signal is −i2δν+i2δ−ν-\tfrac{i}{2}\delta_\nu + \tfrac{i}{2}\delta_{-\nu}.

Inverting, this is −i2e2πiνt+i2e−2πiνt=i2(e−2πiνt−e2πiνt)=sin⁡(2πνt)-\tfrac{i}{2}e^{2\pi i\nu t} + \tfrac{i}{2}e^{-2\pi i\nu t} = \tfrac{i}{2}\big(e^{-2\pi i\nu t}-e^{2\pi i\nu t}\big) = \sin(2\pi\nu t) (using e−iθ−eiθ=−2isin⁡θe^{-i\theta}-e^{i\theta}=-2i\sin\theta), giving H(cos⁡(2πνt))=sin⁡(2πνt)H(\cos(2\pi\nu t)) = \sin(2\pi\nu t).

So the analytic signal is cos⁡(2πνt)+isin⁡(2πνt)=e2πiνt\cos(2\pi\nu t) + i\sin(2\pi\nu t) = e^{2\pi i \nu t} — a single positive-frequency spinning exponential, whose modulus (here constantly 1) and phase directly give the instantaneous amplitude and phase the engineer wants, with the redundant negative-frequency component discarded entirely.

The Fourier multiplier of the Hilbert transform, defined by Hf^(ξ)=−i sgn⁡(ξ) f^(ξ)\widehat{Hf}(\xi) = -i\,\operatorname{sgn}(\xi)\, \hat f(\xi), is:

The Calderón–Zygmund theorem says a singular integral operator bounded on L2L^2 with a suitable kernel is, at the endpoint p=1p=1:

In single-sideband (SSB) radio and envelope detection, engineers build the 'analytic signal' f(t)+iHf(t)f(t) + iHf(t) because:

For the Poisson equation Δu=f\Delta u = f with f∈Lp(Rn)f\in L^p(\mathbb{R}^n), 1<p<∞1<p<\infty, the Calderón–Zygmund estimate controls:

References

  1. Tuomas P. Hytönen (2012). The sharp weighted bound for general Calderón–Zygmund operators · arXiv:1007.4330
  2. Elias M. Stein (1970). Singular Integrals and Differentiability Properties of Functions