Chapter Summary

Chapter 13 Summary: Matched Filter and Backpropagation Imaging

Key Points

• 1.
  The Matched Filter Estimator

  The matched filter $\hat{\mathbf{c}}^{\text{BP}} = \mathbf{A}^{H} \mathbf{y}$ is the adjoint (not inverse) of the forward model. It maximizes per-pixel SNR but produces an image convolved with the Gram matrix $\mathbf{G} = \mathbf{A}^{H} \mathbf{A}$ (the PSF). It is the negative gradient of the data-fidelity cost at the origin and the starting point for all iterative reconstruction algorithms.

• 2.
  Three Names, One Operation

  The matched filter (estimation theory), delay-and-sum beamforming (radar), and backpropagation (diffraction tomography) are mathematically identical: $\mathbf{A}^{H} \mathbf{y}$. In the k-space view, backpropagation is an inverse Fourier transform of the sampled spatial spectrum.

• 3.
  PSF, Resolution, and Sidelobes

  Resolution is fundamentally limited by bandwidth (range: $c/2B$) and aperture (cross-range: $\lambda/2D_{\text{eff}}$). Sidelobes at $-13$ dB limit the dynamic range. Windowing improves sidelobes at the cost of resolution -- the resolution-dynamic range tradeoff.

• 4.
  Filtered Backpropagation

  The Ram-Lak ramp filter corrects the non-uniform k-space sampling density, sharpening the PSF compared to plain backpropagation. This is the RF analogue of filtered back-projection in CT.

• 5.
  Adaptive Beamforming

  Capon (MVDR) achieves data-dependent super-resolution via covariance inversion. MUSIC exploits the signal/noise subspace decomposition. Both require covariance estimation and are limited by SNR. Sparse recovery offers superior single-snapshot performance.

• 6.
  Limitations and the MF-to-U-Net Challenge

  The matched filter cannot deconvolve the PSF, recover true reflectivities, or resolve sub-diffraction targets. A critical CommIT finding: the sidelobe structure of physical sensing matrices corrupts MF-to-U-Net pipelines, motivating model-based architectures for learned reconstruction.