Yonsei University
Chae-Yeon Xon
Yonsei University, School of Economics
Home/Resources/Geometry, Dynamics, and Operators for Economic Models/III-11
Function Spaces, Inverse Problems, and Risk Certificates · Chapter III-11

Integral, Multiplication, Adjoint, and Smoothing Operators

Conditional expectation and measurement systems naturally produce integral and multiplication operators. Their adjoints identify which queries can be represented, while smoothing attenuates high-frequency components and makes inversion unstable. The legal norm depends on the probability space and cannot be transferred casually between L1 and L2 arguments.

Conceptual map

  1. III-11.01Probability integral operators
  2. III-11.02Multiplication operators
  3. III-11.03Adjoints
  4. III-11.04Fourier smoothing
  5. III-11.05Inverse domains

1. Integral kernels define maps between measured spaces

Conditional expectation and measurement error produce integral operators. Their norm and adjoint depend on the source and target measures carried by the two function spaces.

Let (\(X,\mu\)) and (\(Y,\nu\)) be \(\sigma\)-finite measure spaces. For a measurable kernel \(k:Y\times X\to\mathbb{C}\), define

\[(Kf)(y)=\int _{X}k(y,x)f(x)d\mu (x).\](1)
Proposition 1 · Hilbert–Schmidt bound and adjoint

If \(k\in L^{2}(\nu\otimes\mu)\), then \(K:L^{2}(\mu)\to L^{2}(\nu)\) is bounded and compact, with \(\lVert K\rVert \le \lVert k\rVert_{L^{2}}\). Its adjoint is

\[(K^{*}g)(x)=\int _{Y} \operatorname{conj}(k(y,x))g(y)d\nu (y).\](2)

Proof. Cauchy–Schwarz in \(x\) bounds \(|Kf(y)|^{2}\) by \(\int|k(y,x)|^{2}\,d\mu\cdot\lVert f\rVert^{2}\). Integrating in \(y\) gives the norm bound. Fubini's theorem applied to the integrable product \(\overline{g(y)}k(y,x)f(x)\) yields (2). Approximation of \(k\) in \(L^{2}\) by finite sums of product functions gives compactness. ∎

2. Multiplication and conditional weighting

For a measurable function \(m\), the multiplication operator \(M_{m}f=mf\) is bounded from \(L^{2}(\mu)\) to itself exactly when \(m\in L^{\infty}(\mu)\), and its norm is the essential supremum of \(\lvert m\rvert\). If \(m\) approaches zero, the inverse multiplication operator has domain {\(g\): \(g/m\in L^{2}\)} and can be unbounded.

Changing from \(L^{2}\) to \(L^{1}\) changes the available adjoint and compactness arguments. A probability density or likelihood ratio can act as a bounded multiplier in one space and fail the required bound in another.

3. Gaussian smoothing in a Fourier basis

On the unit circle equipped with normalized measure \(dx/(2\pi)\), let \(e_{j}(x)=\operatorname{exp}(ijx)\) be the orthonormal Fourier basis. A periodic heat-kernel convolution with scale \(\tau>0\) satisfies

\[Ke_{j}=\operatorname{exp}(-\tau j^{2})e_{j}.\](3)

At \(\tau=0.02\), frequency 10 is multiplied by \(\operatorname{exp}(-2)=0.135335\), while frequency 20 is multiplied by \(\operatorname{exp}(-8)=0.00033546\). Exact inversion amplifies the corresponding coefficients by 7.389 and 2,980.96. Noise of size \(10^{-4}\) in frequency 20 becomes an inverted error of about 0.298.

Failure case · Injectivity read as stable identification

Every Fourier multiplier in (3) is positive, so the operator is injective. The multipliers converge to zero, which makes the inverse unbounded on the range and exposes high-frequency queries to arbitrarily large noise amplification.

4. Truncation and Tikhonov filters

A spectral cutoff at \(\lvert j\rvert \le J\) uses inverse multiplier \(\operatorname{exp}(\tau j^{2})\) on retained frequencies and zero elsewhere. Tikhonov regularization with parameter \(\alpha>0\) replaces 1/\(\sigma _{j}\) by \(\sigma _{j}/(\sigma _{j}^{2}+\alpha)\). Both procedures trade approximation bias against noise amplification.

The estimand matters. A linear query \(\langle q,f\rangle\) is stably representable from \(Kf\) when \(q\) lies in \(\operatorname{range}(K^{*})\) with a controlled representing norm. Full recovery of \(f\) imposes a stronger requirement than recovery of one smooth query.

5. Implementation, exercises, and sources

Record the measures and basis normalization, compute the forward multipliers before inversion, and plot retained signal and amplified noise separately. Numerical underflow in \(\operatorname{exp}(-\tau j^{2})\) should be treated as a resolution limit and reported with the cutoff rule.

Download the volume verification script →

Exercises

  1. Verify the adjoint identity in (2) for a finite-rank kernel \(k(y,x)=\sum _{r}a_{r}(y)\overline{b_{r}(x)}\).
  2. Reproduce the attenuation and amplification values following (3).
  3. For coefficients \(\lvert f_{j}\rvert \le C|j|^{-s}\), derive a cutoff error bound containing a tail-bias term and an amplified-noise term.
Partial solutions

1. Substitute the finite kernel sum in (1), move the finite summation through the integral, and use the defining inner products to identify \(K^{*}g=\sum b_{r}\langle a_{r},g\rangle\). 2. Insert \(\tau=0.02\) and \(j=10\),20 into \(\operatorname{exp}(-\tau j^{2})\); invert the two values and multiply the second by \(10^{-4}\).

  1. John B. Conway, A Course in Functional Analysis, Chapter II.Hilbert-space adjoints and compact operators.
  2. Rainer Kress, Linear Integral Equations, Chapters 2–4.Integral operators and ill-posed equations.
  3. Heinz W. Engl, Martin Hanke, and Andreas Neubauer, Regularization of Inverse Problems, Chapters 2–4.Spectral regularization filters.

6. Audit checkpoint

State source and target measures, function spaces, kernel integrability, adjoint convention, singular multipliers, inverse domain, regularization rule, and the estimand-specific representing norm.

7. Scope boundary

The chapter covers Hilbert–Schmidt integral operators, multiplication operators, adjoints, and Fourier smoothing witnesses. General harmonic analysis and minimax inverse-problem rates require separate treatment.

Prerequisites