Chapter Summary

Chapter 17 Summary: AMP and OAMP for RF Imaging

Key Points

• 1.

  AMP fails for RF imaging because the sensing matrix $\mathbf{A} = \mathbf{A}_{1} \otimes \mathbf{A}_{2}$ has structured (non-i.i.d.) entries. The scalar Onsager correction cannot decorrelate the residual when the singular values deviate from the Marchenko-Pastur distribution, causing AMP to diverge.

• 2.

  OAMP replaces the scalar Onsager correction with an LMMSE orthogonalization step that uses the full SVD of $\mathbf{A}$. This produces estimation errors that are orthogonal and Gaussian, enabling state evolution for right-rotationally invariant matrices — a much larger class than i.i.d. Gaussian.

• 3.

  Kronecker structure in the RF sensing operator ($\mathbf{A} = \mathbf{A}_{1} \otimes \mathbf{A}_{2}$) reduces the LMMSE step from $O(N^3)$ to $O(N_1^3 + N_2^3)$, making OAMP practical for large imaging problems. The SVDs of the small factor matrices are computed once, and all subsequent iterations use efficient matrix-matrix products.

• 4.

  The denoiser is the modeling knob: soft thresholding gives LASSO, BG-MMSE achieves the Bayes-optimal MSE for sparse scenes, and learned denoisers (DnCNN, U-Net) handle non-sparse, realistic RF scenes with 3–5 dB improvement over BG-MMSE.

• 5.

  Mismatch analysis shows that OAMP is moderately robust to prior misspecification. State evolution remains valid under mismatch and enables offline robustness analysis. Underestimating sparsity is more harmful than overestimating. EM-based learning provides an automatic tuning mechanism.