Abstract
We consider the problem of recovering an unknown signal ${\mathbf x}\in {\mathbb R}^n$ from general nonlinear measurements obtained through a generalized linear model (GLM), i.e., ${\mathbf y}= f\left({\mathbf A}{\mathbf x}+{\mathbf w}\right)$, where $f(\cdot)$ is a componentwise nonlinear function. Based on the unitary transform approximate message passing (UAMP) and expectation propagation, a unitary transform based generalized approximate message passing (GUAMP) algorithm is proposed for general measurement matrices $\bf{A}$, in particular highly correlated matrices. Experimental results on quantized compressed sensing demonstrate that the proposed GUAMP significantly outperforms state-of-the-art GAMP and GVAMP under correlated matrices $\bf{A}$.
Introduction
Signal recovery from nonlinear observations is a fundamental challenge in fields such as wireless communications, medical imaging, and machine learning. This process is typically modeled as a Generalized Linear Model (GLM): $\mathbf{y} = f(\mathbf{Ax} + \mathbf{w})$, where $\mathbf{A}$ is a measurement matrix and $f(\cdot)$ is a nonlinear function like a quantizer. While Generalized Approximate Message Passing (GAMP) is highly effective for independent, identically distributed (i.i.d.) Gaussian matrices, it often diverges when the matrix $\mathbf{A}$ is highly correlated. Even advanced variants like GVAMP struggle in these challenging, real-world scenarios.
Methodology: The GUAMP Algorithm
The proposed GUAMP algorithm bridges the gap between Unitary Approximate Message Passing (UAMP)—known for handling correlated matrices in linear models—and GAMP's ability to handle nonlinearities.
The core of GUAMP lies in a modular inference framework. By performing a Singular Value Decomposition (SVD) on the measurement matrix $\mathbf{A} = \mathbf{U\Sigma V}^T$ and applying a unitary transform, the problem is converted into a more stable form. The inference is then split into two interacting modules:
- Module A (AMP Module): Handles the linear transform and signal prior.
- Module B (GAMP Module): Manages the nonlinear measurement function.
These modules iteratively exchange "extrinsic information" (means and variances) using the principles of Expectation Propagation (EP) until convergence. This decomposition allows the algorithm to maintain stability even when the original matrix $\mathbf{A}$ is ill-conditioned or highly correlated.
Experimental Results
The researchers evaluated GUAMP using quantized compressed sensing simulations with sparse Bernoulli-Gaussian signals. They compared GUAMP against GAMP and GVAMP across varying correlation coefficients ($\rho$) and quantization levels (1-bit and 2-bit).
As shown in the results, all algorithms perform similarly under low correlation ($\rho \approx 0$). However, as correlation increases ($\rho > 0.15$), GAMP and GVAMP performance degrades rapidly, often failing to recover the signal entirely (with NMSE spiking toward 0 dB). In contrast, GUAMP remains remarkably robust, providing accurate signal recovery even at high correlation levels ($\rho = 0.35$ or higher). Furthermore, GUAMP exhibits efficient convergence, typically reaching a stable solution within 10–20 iterations.
Conclusion
GUAMP offers a significant advancement in robust signal recovery for nonlinear systems. By leveraging unitary transforms and a modular EP-based framework, it overcomes the convergence issues that plague traditional message-passing algorithms in the presence of correlated measurement matrices. Future work will explore the integration of damping techniques and adaptive step-sizes to further enhance stability and speed.
Summary updated