Ferkans — Interactive Telecom Tutor

Model Mismatches — When $\mathbf{A}_{\text{true}} \neq \mathbf{A}_{\text{assumed}}$

Every reconstruction algorithm assumes a specific forward model $\mathbf{y} = \mathbf{A}\mathbf{c} + \mathbf{w}$ . In practice, the true model always deviates:

$\mathbf{y} = (\mathbf{A} + \Delta\mathbf{A})\mathbf{c} + \mathbf{w}_{\text{eff}},$

where $\Delta\mathbf{A}$ captures model mismatch and $\mathbf{w}_{\text{eff}}$ may include clutter and other non-Gaussian components. This section catalogs the dominant sources of mismatch and strategies for robustness.

Definition:
Born Approximation Breakdown

The Born approximation assumes single scattering: each point in the scene interacts only with the incident field, not with fields scattered by other points. This fails when:

High contrast: $|\chi(\mathbf{p})| = |\varepsilon_r - 1| \gg 1$ (e.g., metal objects, human body at RF).
Electrically large objects: $\kappa a \gg 1$ , where $a$ is the object dimension.
Dense scenes: Multiple scatterers interact (multiple scattering).

The Born validity criterion (approximate rule of thumb):

$|\chi| \cdot \kappa a < C,$

where $C \approx 0.3$ - $1$ depending on geometry. When violated, the true received signal includes multiple-scattering terms that the linear model $\mathbf{y} = \mathbf{A}\mathbf{c} + \mathbf{w}$ cannot capture.

Alternative models for strong scatterers:

Rytov approximation: Replaces Born's additive error with a multiplicative one; better for smooth media.
Distorted Born iterative method (DBIM): Iteratively updates the background field to include multiple scattering.
Contrast source inversion (CSI): Jointly solves for the contrast and the total field.

,

Theorem: Reconstruction Error Under Model Mismatch

Let $\mathbf{c}^*$ be the true reflectivity and $\hat{\mathbf{c}}$ be the Tikhonov-regularized estimate using the mismatched sensing matrix $\mathbf{A}$ (instead of the true $\mathbf{A} + \Delta\mathbf{A}$ ):

$\hat{\mathbf{c}} = (\mathbf{A}^{H}\mathbf{A} + \lambda\mathbf{I})^{-1}\mathbf{A}^{H}\mathbf{y}.$

The reconstruction error satisfies:

$\|\hat{\mathbf{c}} - \mathbf{c}^*\|_2 \leq \underbrace{\frac{\|\Delta\mathbf{A}\|_2}{\lambda} \cdot \|\mathbf{c}^*\|_2}_{\text{mismatch error}} + \underbrace{\frac{\sigma^2}{\lambda} \cdot \|\mathbf{A}\|_2^{-1}}_{\text{noise amplification}} + \underbrace{\lambda\,\|\mathbf{c}^*\|_2}_{\text{regularization bias}}.$

The regularization parameter $\lambda$ trades off noise amplification against mismatch and bias: larger $\lambda$ reduces mismatch sensitivity but increases bias.

Over-regularization makes the reconstruction less sensitive to model errors (because the regularizer dominates over the data fidelity term), at the cost of a blurred, biased estimate. This is the fundamental tradeoff in model-mismatched imaging.

Proof

Substitute the true model

The measurements are $\mathbf{y} = (\mathbf{A} + \Delta\mathbf{A})\mathbf{c}^* + \mathbf{w}$ . Substituting into the Tikhonov formula:

$\hat{\mathbf{c}} = (\mathbf{A}^{H}\mathbf{A} + \lambda\mathbf{I})^{-1}\mathbf{A}^{H} [(\mathbf{A} + \Delta\mathbf{A})\mathbf{c}^* + \mathbf{w}].$

Separate the terms

$\hat{\mathbf{c}} = \underbrace{(\mathbf{A}^{H}\mathbf{A} + \lambda\mathbf{I})^{-1}\mathbf{A}^{H}\mathbf{A}}_{\mathbf{I} - \lambda(\mathbf{A}^{H}\mathbf{A} + \lambda\mathbf{I})^{-1}}\mathbf{c}^* + (\mathbf{A}^{H}\mathbf{A} + \lambda\mathbf{I})^{-1}\mathbf{A}^{H}\Delta\mathbf{A}\,\mathbf{c}^* + (\mathbf{A}^{H}\mathbf{A} + \lambda\mathbf{I})^{-1}\mathbf{A}^{H}\mathbf{w}.$ $

Bound each term

The first term gives the regularization bias: $\|(\mathbf{I} - \lambda(\mathbf{A}^{H}\mathbf{A} + \lambda\mathbf{I})^{-1})\mathbf{c}^* - \mathbf{c}^*\| \leq \lambda\|\mathbf{c}^*\|_2$ .

The second (mismatch) term: $\|(\mathbf{A}^{H}\mathbf{A} + \lambda\mathbf{I})^{-1}\mathbf{A}^{H}\Delta\mathbf{A}\mathbf{c}^*\| \leq \frac{\|\Delta\mathbf{A}\|}{\lambda}\|\mathbf{c}^*\|$ .

The third (noise) term is bounded similarly. Combining gives the stated bound. $\blacksquare$

Definition:
Gridding Error (Basis Mismatch)

When targets do not lie on the assumed grid, the sensing matrix $\mathbf{A}$ does not perfectly model the measurements. The gridding error has two components:

Spectral leakage: Energy from an off-grid target leaks to neighboring grid points, creating a sinc-like artifact pattern. The leakage energy is:

$E_{\text{leak}} \approx 1 - \text{sinc}^2(\delta/\Delta x),$

where $\delta$ is the offset from the nearest grid point and $\Delta x$ is the grid spacing.
Phase mismatch: The off-grid target has a different round-trip phase than any grid point, creating a structured model error:

$\Delta\mathbf{A}_{(i,j,k),q} = [\mathbf{A}]_{(i,j,k),q}\, (e^{-j\boldsymbol{\kappa}_{i,j,k}^T \delta\mathbf{p}} - 1).$

For compressed sensing, the gridding error destroys the sparsity assumption. A truly sparse scene (few point targets) becomes non-sparse in the grid basis due to leakage.

Definition:
Calibration Errors

Calibration errors arise from imperfect knowledge of the system parameters used to construct $\mathbf{A}$ :

Error source	Model effect	Typical magnitude
Antenna position	Phase error in steering vector	$< \lambda/20$ with GPS/INS
Mutual coupling	Effective pattern distortion	$-15$ to $-25$ dB (element-dependent)
Phase noise	Random phase on each measurement	$-80$ to $-100$ dBc/Hz at 1 MHz
I/Q imbalance	Image at conjugate frequency	1-5% gain, 1-5 degrees phase
Frequency offset	Range shift	PPM-level with modern oscillators

These errors are modeled as a perturbation to the sensing matrix:

$\mathbf{A}_{\text{actual}} = \mathbf{A}_{\text{ideal}} + \Delta\mathbf{A}_{\text{cal}},$

and reconstruction quality depends on $\|\Delta\mathbf{A}_{\text{cal}}\| / \|\mathbf{A}\|$ .

Example: Born Approximation Validity for a Dielectric Cylinder

A dielectric cylinder of radius $a = \lambda/4$ and contrast $\chi = \varepsilon_r - 1$ is illuminated at frequency $f_0$ .

(a) Compute the Born validity parameter $|\chi| \cdot \kappa a$ for $\varepsilon_r \in \{1.5, 3, 6, 10\}$ .

(b) Using the rule of thumb $|\chi|\kappa a < 0.5$ , determine which materials are within the Born regime.

(c) For concrete ( $\varepsilon_r = 6$ ), at what maximum object size $a$ does the Born approximation remain valid at $f_0 = 5$ GHz?

Solution

Born parameter computation

With $a = \lambda/4$ , $\kappa a = 2\pi/\lambda \cdot \lambda/4 = \pi/2 \approx 1.57$ .

$\varepsilon_r$	$\|\chi\| = \|\varepsilon_r - 1\|$	$\|\chi\|\kappa a$
1.5	0.5	0.79
3	2	3.14
6	5	7.85
10	9	14.1

Born validity check

Only $\varepsilon_r = 1.5$ is close to the threshold $|\chi|\kappa a < 0.5$ (it gives $0.79$ , slightly above). Materials with $\varepsilon_r \geq 3$ are well beyond the Born regime for objects of size $\lambda/4$ .

Maximum size for concrete

For $\varepsilon_r = 6$ at $f_0 = 5$ GHz: $\lambda = 6$ cm, $\kappa = 104.7$ rad/m, $|\chi| = 5$ .

Solve $5 \times 104.7 \times a < 0.5$ : $a < 0.5/(5 \times 104.7) = 9.5 \times 10^{-4}$ m $\approx 1$ mm.

The Born approximation holds for concrete objects smaller than about 1 mm at 5 GHz — essentially, it fails for any macroscopic concrete structure. This is why through-wall imaging requires more sophisticated models (Section s02, DBIM, or ray tracing).

,

Born Approximation Validity

Explore the Born approximation error as a function of contrast strength $|\chi|$ and electrical size $\kappa a$ . The color map shows the relative error between the Born approximation and a reference multi-scattering solution.

Parameters

Maximum contrast

|\chi|

5

Maximum

\kappa a

3

Reconstruction Robustness vs. Model Mismatch

Compare the reconstruction error of back-projection, Tikhonov, and $\ell_1$ minimization as the model mismatch $\|\Delta\mathbf{A}\|_F / \|\mathbf{A}\|_F$ increases. Over-regularized Tikhonov trades resolution for robustness.

Parameters

Relative mismatch

\|\Delta A\|_F / \|A\|_F

0.1

Regularization factor (relative to optimal)1

Strategies for Robustness to Model Mismatch

Over-regularization: Increase $\lambda$ beyond the noise-optimal value. Reduces sensitivity to $\Delta\mathbf{A}$ at the cost of resolution (bias).
Robust optimization: Minimax formulations that optimize for the worst-case $\Delta\mathbf{A}$ within a bounded set: $\min_{\mathbf{c}} \max_{\|\Delta\mathbf{A}\| \leq \epsilon} \|\mathbf{y} - (\mathbf{A} + \Delta\mathbf{A})\mathbf{c}\|^2$ .
Joint estimation: Simultaneously estimate $\mathbf{c}$ and $\Delta\mathbf{A}$ (auto-calibration, autofocus in SAR).
Learned robustness: Train neural networks on data with realistic model mismatches (Part IV). The network implicitly learns to be robust.

Understanding the magnitude and structure of model mismatch is essential for choosing reconstruction algorithms — a theme that pervades Part III.

🎓CommIT Contribution(2026)

Caire's Unified Illumination and Sensing Model

G. Caire — CommIT Research Note

Caire's unified framework derives the forward model $\mathbf{y} = \mathbf{A}\mathbf{c} + \mathbf{w}$ from first principles (diffraction tomography), connecting the illumination pattern to the sensing operator through wavenumber-domain analysis. The derivation in this chapter (Sections s01-s02) follows Caire's framework, which provides the Kronecker-structured operator, the far-field Taylor approximation, and the link budget normalization that unifies diverse sensing geometries under a single formalism.

rf-imagingforward-modelsensing-operator

⚠️Engineering Note

Calibration — Closing the Model-Reality Gap

Calibration measures and corrects hardware impairments so that $\mathbf{A}_{\text{actual}} \approx \mathbf{A}_{\text{ideal}}$ :

Calibration type	What it corrects	Method
Antenna calibration	Pattern errors, coupling	Anechoic chamber measurement
Channel calibration	Gain/phase across channels	Known target (corner reflector)
Motion calibration	Platform trajectory errors	INS + GPS + autofocus
System calibration	End-to-end transfer function	Loop-back or reference target

For MIMO radar arrays, online calibration techniques estimate the coupling matrix and channel gain/phase errors from the radar data itself, analogous to self-calibration in radio astronomy.

Practical Constraints

•
Anechoic chamber measurements are expensive and one-time
•
Online calibration requires known reference targets in the scene
•
Autofocus adds computational cost to reconstruction

Common Mistake: The Inverse Crime

Mistake:

Testing a reconstruction algorithm using synthetic data generated from the same forward model used for reconstruction (same grid, same $\mathbf{A}$ , same discretization).

Correction:

Always generate test data from a finer grid or a different forward model than the one used for reconstruction. Otherwise, the results are overly optimistic because the model mismatch that dominates real-world performance is absent. This is called the "inverse crime" in the inverse problems literature.

Quick Check

Increasing the regularization parameter $\lambda$ in Tikhonov regularization makes the reconstruction:

More sensitive to model mismatch but sharper

Less sensitive to model mismatch but more blurred

Both sharper and more robust

Correction:

Less sensitive to model mismatch but more blurred

Larger $\lambda$ increases the regularization bias (blur) but reduces the amplification of model errors $\Delta\mathbf{A}$ . This is the resolution-robustness tradeoff.

Key Takeaway

Model mismatch is the dominant practical challenge in RF imaging. Born approximation breakdown, gridding errors, and calibration imperfections all introduce structured perturbations $\Delta\mathbf{A}$ . Over-regularization trades resolution for robustness. The inverse crime — testing with the same model used for reconstruction — hides the true impact of mismatch.

Model Mismatches and Robustness

Model Mismatches — When Atrue≠Aassumed\mathbf{A}_{\text{true}} \neq \mathbf{A}_{\text{assumed}}Atrue​=Aassumed​

Definition: Born Approximation Breakdown

Theorem: Reconstruction Error Under Model Mismatch

Substitute the true model

Separate the terms

Bound each term

Definition: Gridding Error (Basis Mismatch)

Definition: Calibration Errors

Example: Born Approximation Validity for a Dielectric Cylinder

Born parameter computation

Born validity check

Maximum size for concrete

Born Approximation Validity

Parameters

Reconstruction Robustness vs. Model Mismatch

Parameters

Strategies for Robustness to Model Mismatch

Caire's Unified Illumination and Sensing Model

Calibration — Closing the Model-Reality Gap

Common Mistake: The Inverse Crime

Quick Check

Key Takeaway

Model Mismatches — When $\mathbf{A}_{\text{true}} \neq \mathbf{A}_{\text{assumed}}$

Definition:
Born Approximation Breakdown

Definition:
Gridding Error (Basis Mismatch)

Definition:
Calibration Errors