Chapter Summary

Chapter 1 Summary — Functional Analysis for Imaging

Key Points

• 1.

  Normed spaces generalise $\mathbb{C}^n$ to infinite dimensions. The $L^p$ family provides the standard function spaces for imaging: $L^2$ for energy and least-squares, $L^1$ for sparsity and edge preservation. In finite dimensions all norms are equivalent; in infinite dimensions they are not — the choice of norm determines the character of every reconstruction.

• 2.

  Banach spaces (complete normed spaces) are the right setting for iterative algorithms. Completeness ensures that Cauchy sequences converge to a limit within the space, guaranteeing that iterative algorithms (ISTA, ADMM, Landweber) have well-defined fixed points.

• 3.

  Hilbert spaces add the geometry of angles, projections, and orthogonality to Banach spaces. The orthogonal projection theorem guarantees a unique best approximation in any closed subspace, underpinning matched filtering, MMSE estimation, and Tikhonov regularisation. Parseval's identity equates spatial energy and spectral energy — the foundation of diffraction tomography.

• 4.

  Bounded linear operators generalise matrices to function spaces. The adjoint $\mathcal{A}^*$ (generalising $\mathbf{A}^{H}$) is the back-projection operator and appears in every iterative reconstruction algorithm. The operator norm $\|\mathcal{A}\|_{\mathrm{op}} = \sigma_1(\mathcal{A})$ controls the step size for gradient descent.

• 5.

  Compact operators — which include all integral operators with square-integrable kernels, hence all Born-approximation RF imaging operators — are not boundedly invertible on infinite-dimensional spaces. This is the mathematical origin of ill-posedness. The forward operator $\mathcal{A}$ is compact; its pseudoinverse is unbounded.

• 6.

  The singular system $\{(\sigma_n, v_n, u_n)\}$ of a compact operator provides the complete spectral picture: singular values $\sigma_n \to 0$ (the decay rate classifies ill-posedness as mild, moderate, or severe), right singular functions $v_n$ are the "scene modes" visible to the sensor, and left singular functions $u_n$ are the data modes.

• 7.

  The Picard condition $\sum_n |\langle \mathbf{y}, u_n \rangle|^2 / \sigma_n^2 < \infty$ is necessary and sufficient for the inverse problem to have a solution. Noise always violates it at high frequencies (where $\sigma_n$ is small), which is why regularisation is mandatory. The index $n^*$ where the noise level crosses the singular value curve determines the achievable reconstruction bandwidth.

• 8.

  Distributions and Sobolev spaces extend classical analysis to handle point scatterers ($\delta$ functions), sharp edges (Heaviside), and the singularities of the Helmholtz Green's function $G = e^{j\kappa r}/(4\pi r)$ — the kernel of the forward operator in the Born approximation. Sobolev regularisation $H^s$ encodes prior smoothness assumptions about the scene.

• 9.

  The forward model $\mathbf{y} = \mathbf{A}\,\mathbf{c} + \mathbf{w}$ is the golden thread of the book. Every subsequent chapter either builds the sensing matrix $\mathbf{A}$ (Ch~5–9), analyses its properties (Ch~10–11), develops methods to invert it (Ch~12–23), or applies the framework to a specific RF system (Ch~24–32).