Ferkans — Interactive Telecom Tutor

Why Well-Posedness Matters for Imaging

Every imaging system can be abstractly described by a forward model

$y = \mathcal{A} x,$

where $x$ is the unknown scene (reflectivity, permittivity, etc.) and $y$ is the measured data (scattered field, projections, etc.). The inverse problem is to recover $x$ from $y$ .

In practice, we never observe $y$ exactly — we have noisy data $y^\delta = y + \eta$ with $\|\eta\| \leq \delta$ . The central question of this chapter is: does a small perturbation in the data lead to a small perturbation in the reconstruction? If not, the problem is ill-posed, and naive inversion will catastrophically amplify noise.

Hadamard formalised these ideas in 1902, establishing three conditions that a well-posed problem must satisfy. Most imaging inverse problems fail at least one of these conditions, motivating the entire theory of regularization developed in this chapter.

Historical Note: Jacques Hadamard and the Origins of Well-Posedness

1902–1963

Jacques Hadamard (1865–1963) introduced the concept of well-posedness in his 1902 paper Sur les problèmes aux dérivées partielles et leur signification physique, in the context of partial differential equations. He observed that physically reasonable problems should have solutions that exist, are unique, and depend continuously on the data — otherwise the model fails to capture reality.

For nearly half a century, ill-posed problems were considered pathological and unworthy of study. It was only with Tikhonov's work in the 1940s–1960s that a systematic theory for solving ill-posed problems was developed, transforming them from curiosities into the central challenge of inverse problems — and of computational imaging.

Definition:
Hadamard Well-Posedness

Let $\mathcal{A} \colon \mathcal{X} \to \mathcal{Y}$ be an operator between normed spaces. The equation $\mathcal{A} x = y$ is well-posed in the sense of Hadamard if all three of the following conditions hold:

Existence: For every $y \in \mathcal{Y}$ , there exists at least one $x \in \mathcal{X}$ such that $\mathcal{A} x = y$ (equivalently, $\mathcal{A}$ is surjective).
Uniqueness: For every $y \in \mathcal{Y}$ , there is at most one $x \in \mathcal{X}$ such that $\mathcal{A} x = y$ (equivalently, $\mathcal{A}$ is injective).
Stability (continuous dependence on data): The inverse mapping $\mathcal{A}^{-1}$ is continuous: if $y_n \to y$ in $\mathcal{Y}$ , then $\mathcal{A}^{-1} y_n \to \mathcal{A}^{-1} y$ in $\mathcal{X}$ .

A problem that fails any of these conditions is called ill-posed.

In finite dimensions, if $\mathcal{A}$ is a square invertible matrix, all three conditions hold automatically — the inverse is a matrix, hence continuous. Ill-posedness is fundamentally an infinite-dimensional phenomenon, though discrete approximations inherit it through large condition numbers.

,

Well-Posed Problem

A problem $\mathcal{A}x = y$ is well-posed (in the sense of Hadamard) if a solution exists for every datum $y$ , the solution is unique, and it depends continuously on the data. A problem failing any condition is ill-posed.

Degree of Ill-Posedness

The degree of ill-posedness of a linear inverse problem is characterised by the decay rate of the singular values $\sigma_k$ of the forward operator $\mathcal{A}$ : polynomial decay $\sigma_k \sim k^{-p}$ gives mildly ill-posed problems; exponential decay $\sigma_k \sim e^{-ck^\beta}$ gives severely ill-posed problems.

Related: Well-Posed Problem

Example: Failure of Existence: Limited-Angle Tomography

In X-ray computed tomography, the forward operator is the Radon transform $\mathcal{R}$ , which maps a 2D density function $f(x,y)$ to its line integrals. In limited-angle tomography, measurements are available only for projection angles $\theta \in [\theta_0, \theta_1]$ with $\theta_1 - \theta_0 < \pi$ .

Explain why the limited-angle Radon transform is not surjective onto $L^2$ , so the existence condition fails for generic data.

Solution

Characterise the range via the Fourier Slice Theorem

The Fourier Slice Theorem states that the 1D Fourier transform of $\mathcal{R}f$ at angle $\theta$ equals a slice of the 2D Fourier transform $\hat{f}$ through the origin at the same angle:

$\widehat{\mathcal{R}f}(\theta, \omega) = \hat{f}(\omega\cos\theta,\, \omega\sin\theta).$

Limited-angle data determines $\hat{f}$ only on a wedge of the frequency plane, corresponding to angles in $[\theta_0, \theta_1]$ .

Identify the gap

Any function $g \in L^2$ whose sinogram has nonzero energy at angles outside $[\theta_0, \theta_1]$ lies outside the range of the limited-angle operator. Therefore, for generic $y \in L^2([\theta_0,\theta_1]\times\mathbb{R})$ , there need not exist an $f$ whose limited-angle projections equal $y$ — existence fails.

Imaging interpretation

The frequency components of the scene in the missing angular wedge are simply not measured. This explains the characteristic streak artefacts of limited-angle CT: the reconstruction algorithm tries to invent data that was never acquired.

Example: Failure of Uniqueness: Null Space of the Forward Operator

Consider 1D deconvolution where the forward operator is convolution with a band-limited kernel $h$ satisfying $\hat{h}(\omega) = 0$ for $|\omega| > \omega_c$ . Show that uniqueness fails.

Solution

Identify the null space

By the convolution theorem, $\widehat{\mathcal{A}x}(\omega) = \hat{h}(\omega)\cdot\hat{x}(\omega)$ .

Any function $x_0$ whose Fourier transform is supported entirely in $\{|\omega| > \omega_c\}$ satisfies $\hat{h}(\omega)\cdot\hat{x}_0(\omega) = 0$ for all $\omega$ , so $\mathcal{A} x_0 = 0$ . Thus $x_0 \in \mathcal{N}(\mathcal{A})$ and the null space is infinite-dimensional.

Imaging interpretation

In RF imaging, the system's point spread function (PSF) always has finite bandwidth. High-frequency details of the scene beyond the system's resolution limit are in the null space — they produce zero response and are fundamentally invisible to the measurement system. This is why super-resolution is an inherently ill-posed problem.

Example: Failure of Stability: Compact Operators

Let $\mathcal{A} \colon \mathcal{X} \to \mathcal{Y}$ be an injective compact operator between infinite-dimensional Hilbert spaces. Show that $\mathcal{A}^{-1} \colon \mathcal{R}(\mathcal{A}) \to \mathcal{X}$ is unbounded, so stability fails.

Solution

Use the singular value decomposition

By the spectral theory of compact operators (TSpectral Theorem for Compact Self-Adjoint Operators), $\mathcal{A}$ has singular values $\sigma_1 \geq \sigma_2 \geq \cdots > 0$ with $\sigma_k \to 0$ as $k \to \infty$ .

For the singular vector $v_k$ , we have $\mathcal{A} v_k = \sigma_k u_k$ and therefore

$\frac{\|\mathcal{A}^{-1}(\sigma_k u_k)\|}{\|\sigma_k u_k\|} = \frac{\|v_k\|}{\sigma_k \|u_k\|} = \frac{1}{\sigma_k} \to \infty.$

Conclude instability

Since $\sigma_k \to 0$ , the ratio $\|\mathcal{A}^{-1} y_k\|/\|y_k\|$ is unbounded. Thus $\mathcal{A}^{-1}$ is an unbounded operator and Hadamard's stability condition fails.

Physical interpretation: Data components $\sigma_k u_k$ become arbitrarily small for large $k$ , but the corresponding solution components $v_k$ do not. Noise at frequency $k$ gets amplified by $1/\sigma_k$ , which grows without bound.

Theorem: Compact Operators Yield Ill-Posed Problems

Let $\mathcal{A} \colon \mathcal{X} \to \mathcal{Y}$ be a compact linear operator between infinite-dimensional Hilbert spaces with $\mathcal{N}(\mathcal{A}) = \{0\}$ . Then the equation $\mathcal{A} x = y$ is ill-posed: the inverse $\mathcal{A}^{-1} \colon \mathcal{R}(\mathcal{A}) \to \mathcal{X}$ is discontinuous.

Compact operators compress infinite-dimensional information into a rapidly decaying sequence of singular values. Inverting this compression requires dividing by these vanishing singular values, which amplifies any perturbation without bound.

Proof

Singular value decay

Since $\mathcal{A}$ is compact and the spaces are infinite-dimensional, it has infinitely many positive singular values $\sigma_k \to 0$ (TSpectral Theorem for Compact Self-Adjoint Operators).

Unboundedness of the inverse

Consider the sequence $y_k = \sigma_k u_k \in \mathcal{R}(\mathcal{A})$ . Then $\|y_k\| = \sigma_k \to 0$ , but $\|\mathcal{A}^{-1} y_k\| = \|v_k\| = 1$ for all $k$ .

If $\mathcal{A}^{-1}$ were continuous at $0$ , then $y_k \to 0$ would imply $\mathcal{A}^{-1} y_k \to 0$ , contradicting $\|\mathcal{A}^{-1} y_k\| = 1$ . Therefore $\mathcal{A}^{-1}$ is discontinuous. $\blacksquare$

,

Ill-Posedness in Standard Imaging Operators

Modality	Forward Operator	Failing Condition	Singular Value Decay	Degree
SAR / Stripmap Radar	Band-limited Fourier restriction	Uniqueness (null space of high frequencies)	$\sigma_k \sim k^{-1/2}$ (mild)	Mildly ill-posed
X-ray CT (full angle)	Radon transform	Stability	$\sigma_k \sim k^{-1/2}$	Mildly ill-posed
Limited-angle CT	Truncated Radon transform	Existence + Stability	Exponential in angular gap	Severely ill-posed
Gaussian Deblurring	Convolution with $e^{-\sigma^2\omega^2/2}$	Stability	$\sigma_k \sim e^{-ck^2}$	Severely ill-posed
Microwave Tomography	Volume integral equation (Born)	Stability	$\sigma_k \sim k^{-1}$	Mildly ill-posed

Why This Matters: Ill-Posedness in RF Imaging Systems

Virtually all RF imaging modalities involve compact forward operators, making them ill-posed:

Synthetic Aperture Radar (SAR): The scattering operator maps the scene reflectivity to the received signal. The finite aperture and bandwidth limit the observable spatial frequencies, creating a band-limited (hence compact) forward operator.
Ground-Penetrating Radar (GPR): Subsurface permittivity profiles are mapped to surface measurements through a Green's function integral operator, which is compact due to the smoothing effect of wave propagation.
Microwave Tomography: The scattered field is related to the contrast function through a volume integral equation. The integral operator is compact, and the nonlinear dependence of the scattered field on the contrast makes the problem doubly challenging.
Near-Field to Far-Field Transformation: The near-to-far-field operator is compact, and exponential decay of evanescent modes means that near-field details are severely attenuated in far-field data.

In every case, the singular values of the discretized forward operator exhibit rapid decay, and the condition number grows with the problem size — a hallmark of ill-posedness.

See full treatment in Kronecker Product Structure of the Sensing Matrix

Common Mistake: Discrete Systems Can Be Invertible Yet Severely Ill-Conditioned

Mistake:

Believing that because a discretized imaging system $\mathbf{A}\mathbf{x} = \mathbf{y}$ has a unique solution (when $\mathbf{A}$ is square and invertible), the problem is well-posed and can be solved by direct inversion.

Correction:

A finite-dimensional system is always well-posed in Hadamard's sense if $\mathbf{A}$ is invertible — the inverse is automatically continuous in finite dimensions. However, the condition number $\kappa(\mathbf{A}) = \sigma_1/\sigma_n$ can be astronomically large, making the problem effectively ill-posed for any practical noise level.

As the discretization is refined ( $n \to \infty$ ), $\kappa(\mathbf{A}_n) \to \infty$ , reflecting the unboundedness of $\mathcal{A}^{-1}$ . Regularization is necessary for the discrete problem just as for the continuous one.

Quick Check

For the integral operator $(\mathcal{A}x)(t) = \int_0^1 e^{-(t-s)^2} x(s)\,ds$ on $L^2([0,1])$ , which Hadamard condition fails first?

Existence

Uniqueness

Stability

All three conditions fail

Correction:

Stability

The Gaussian kernel yields exponentially decaying singular values $\sigma_k \sim e^{-ck^2}$ , making this a severely ill-posed problem. The inverse $\mathcal{A}^{-1}$ is unbounded — stability fails catastrophically.

⚠️Engineering Note

Condition Number as a Practical Ill-Posedness Metric

In software-defined radar and microwave imaging systems, the condition number $\kappa(\mathbf{A}) = \sigma_1/\sigma_n$ of the discretized sensing matrix is routinely computed before choosing a reconstruction strategy:

$\kappa < 100$ : Standard least-squares or CGLS without regularization.
$\kappa \in [100, 10^6]$ : Tikhonov regularization with the L-curve or discrepancy principle for parameter selection.
$\kappa > 10^6$ : Truncated SVD or sparsity-promoting regularization required; standard Tikhonov may over-smooth.

For SAR systems, the condition number grows with the scene size $N$ as $\kappa \sim N^{0.5}$ (mildly ill-posed), while microwave tomography in strongly scattering media can reach $\kappa \sim N^{1.5}$ due to the multiple-scattering contribution.

Key Takeaway

Hadamard's three conditions — existence, uniqueness, stability — define well-posedness. Most RF imaging inverse problems fail the stability condition because their forward operators are compact with decaying singular values. The degree of ill-posedness is quantified by the decay rate of $\sigma_k$ : polynomial decay gives mild ill-posedness; exponential decay gives severe ill-posedness. Discrete ill-posedness manifests as large condition numbers — regularization is needed regardless of whether one works in the continuous or discrete setting.

Hadamard Well-Posedness

Why Well-Posedness Matters for Imaging

Historical Note: Jacques Hadamard and the Origins of Well-Posedness

Definition: Hadamard Well-Posedness

Well-Posed Problem

Degree of Ill-Posedness

Example: Failure of Existence: Limited-Angle Tomography

Characterise the range via the Fourier Slice Theorem

Identify the gap

Imaging interpretation

Example: Failure of Uniqueness: Null Space of the Forward Operator

Identify the null space

Imaging interpretation

Example: Failure of Stability: Compact Operators

Use the singular value decomposition

Conclude instability

Theorem: Compact Operators Yield Ill-Posed Problems

Singular value decay

Unboundedness of the inverse

Ill-Posedness in Standard Imaging Operators

Why This Matters: Ill-Posedness in RF Imaging Systems

Common Mistake: Discrete Systems Can Be Invertible Yet Severely Ill-Conditioned

Quick Check

Condition Number as a Practical Ill-Posedness Metric

Key Takeaway

Definition:
Hadamard Well-Posedness