Chapter Summary

Chapter 8 Summary: Scene Representation and Forward Model Variations

Key Points

• 1.

  Point-scatterer models represent the scene as isolated reflectors with continuous positions. The forward model is non-linear in position but inherently sparse and free of basis mismatch.

• 2.

  Grid-based reflectivity maps discretize $c(\mathbf{p})$ on a regular grid, yielding the linear model $\mathbf{y} = \mathbf{A}\mathbf{c} + \mathbf{w}$. The complex phase $\arg(c_q)$ encodes sub-grid position and must not be discarded.

• 3.

  Basis mismatch (spectral leakage from off-grid targets) is the dominant error in grid-based compressed sensing. Mitigation includes oversampling, off-grid methods, and dictionary refinement.

• 4.

  Deterministic channel models (Born, ray tracing, PO, full-wave) construct $\mathbf{A}$ from physics. Multipath extends $\mathbf{A}$ with ghost target columns. Through-wall propagation adds frequency-dependent attenuation and range shift.

• 5.

  Stochastic models (Swerling I-IV, Rayleigh/Rician speckle) describe target fluctuation and clutter statistics but cannot replace deterministic $\mathbf{A}$ for imaging. In telecom the channel is a nuisance; in imaging it is the signal.

• 6.

  Near-field effects become significant when $R < 2D^2/\lambda$ (Fraunhofer distance). The quadratic phase correction destroys the Kronecker structure, increasing computational cost from $O(M+Q)$ to $O(MQ)$.

• 7.

  Model mismatch ($\Delta\mathbf{A}$) from Born breakdown, gridding errors, and calibration imperfections degrades all reconstruction algorithms. Over-regularization trades resolution for robustness. The inverse crime hides mismatch effects in simulations.