Technical Summary: Learning-Based Signal Recovery in Nonlinear Systems with Spectrally Separated Interference

1. Problem Statement

The paper addresses the challenge of recovering a desired signal in Upper Mid-Band (FR3, 7–24 GHz) 6G receivers operating in spectrally dense environments. In these scenarios, strong adjacent-band (out-of-band, OOB) interferers are inevitable. While conventional linear receivers rely on frequency-selective filtering to suppress disjoint-band interference, practical hardware limitations—specifically front-end nonlinearities (e.g., amplifier saturation) and finite-resolution quantization—cause spectral leakage.

When a strong OOB interferer drives the receiver into a nonlinear regime, it generates in-band distortion that renders classical linear filtering ineffective. The core problem is to estimate the desired signal $x_0$ from receiver observations $y$ (or quantized $y_q$) that are corrupted by:

1. Strong OOB Interference: Occupying a disjoint frequency band.
2. Nonlinear Saturation: Modeled as a smooth, memoryless nonlinearity (soft-thresholding).
3. Quantization: Optional coarse quantization by the Analog-to-Digital Converter (ADC).

2. Methodology

System Model

The authors adopt a discrete-time signal model where the received vector $r$ is the inverse DFT of frequency-domain sources $x_\ell$ occupying disjoint bands $B_\ell$. The receiver front-end is modeled as a memoryless nonlinear function $\Phi(\cdot)$ followed by additive noise and optional quantization:
$$y = \Phi(r, w)$$
Specifically, a soft saturation model is used, where the amplitude is compressed via a $\tanh$ function while preserving phase, followed by additive noise and a uniform quantizer.

Proposed Algorithm: Learned ML-VAMP

The paper proposes a Learned Multi-Layer Vector Approximate Message Passing (LMLVAMP) algorithm. This approach integrates model-based Bayesian inference with data-driven adaptability.

• Structure: The algorithm treats the system as a two-layer network:
  1. Layer 1 (Frequency Domain): Maps the frequency-domain signal $x$ to the time-domain signal $r$ via the IDFT matrix.
  2. Layer 2 (Time Domain): Maps $r$ to the nonlinear observation $y$.
• Iterative Inference: The algorithm iteratively updates estimates $\hat{x}$ and $\hat{r}$ using two denoising functions:
  1. Spectral Denoiser ($G_0$): Operates in the frequency domain. For the desired signal, this is a linear MMSE estimator based on known spectral priors (mean $\mu$ and variance $S$).
  2. Nonlinear Denoiser ($G_1$): Operates in the time domain to invert the nonlinear distortion.
• Learning Component:
  • In classical ML-VAMP, the nonlinear denoiser $G_1$ requires complex multidimensional numerical integration, which is computationally prohibitive.
  • The proposed Learned approach replaces the intractable analytical denoisers with compact, trainable neural networks ($f_0$ and $f_1$).
  • These networks are trained end-to-end on simulated data to approximate the optimal MMSE denoisers. They take inputs such as the current estimate, noise variance, and system parameters (e.g., saturation power $P_{sat}$) and output refined estimates and precision parameters.
• Training Objective: The model is trained using a composite loss function that penalizes the frequency-domain error at intermediate iterations (to guide convergence) and the final iteration (to ensure accuracy).

Complexity Advantages

Unlike general deep learning approaches that might require $O(N^2)$ parameters for $N$-dimensional inputs, the LMLVAMP framework utilizes component-wise neural networks. The complexity does not scale with the signal dimension $N$, making it suitable for wideband systems.

3. Key Contributions

1. Novel Framework: Introduction of a Learned ML-VAMP framework specifically designed for signal recovery in the presence of strong OOB interference, receiver saturation, and quantization.
2. Hybrid Architecture: Integration of spectral priors (for the desired signal band) with neural network-based denoising to handle nonlinear distortions without requiring explicit analytical models of the hardware nonlinearity.
3. Interpretability: Unlike "black-box" deep learning networks, the proposed method retains the iterative structure of message passing, offering interpretability through its Bayesian inference roots.
4. Efficiency: Demonstration that replacing analytical denoisers with lightweight neural networks avoids the computational burden of high-dimensional integration while maintaining scalability.

4. Performance Results

Simulations were conducted under FR3-relevant conditions (SNR 10–20 dB, INR 30–80 dB, 10-bit quantization). The proposed LMLVAMP was compared against:

• LINEAR-K/U: Linear frequency-domain Wiener estimators (with/without interferer knowledge).
• ORACLE: An idealized estimator with perfect system knowledge (upper bound).

Key Findings:

• Robustness to Interference: Linear estimators degrade sharply when the Interference-to-Noise Ratio (INR) exceeds 50 dB due to nonlinear spectral leakage. In contrast, LMLVAMP maintains strong performance, achieving achievable rates significantly higher than linear baselines (gaps exceeding 20 dB in some regimes).
• Convergence: The proposed method approaches the performance of the ORACLE benchmark within two iterations.
• Generalization: The "Unknown" variant (LMLVAMP-U), which does not assume knowledge of the interferer's signal, significantly outperforms linear baselines, demonstrating the ability to generalize to unknown interference conditions.
• Quantization Resilience: The algorithm retains a substantial advantage over linear alternatives even under 10-bit quantization.

5. Significance and Claims

The paper claims that the proposed LMLVAMP framework offers a viable solution for designing robust and cost-efficient FR3 receivers for future 6G systems. By jointly modeling nonlinearity, noise, and quantization, the method overcomes the limitations of classical linear filtering in dense spectral environments.

The authors emphasize that this approach bridges the gap between rigorous model-based inference and the adaptability of data-driven methods. It provides a practical alternative to complex polynomial modeling (which is hard to estimate) and black-box neural networks (which lack interpretability). The results highlight that iterative inference combined with learned denoisers is essential for handling the severe nonlinear distortions characteristic of upper mid-band coexistence scenarios.

Future work, as noted by the authors, includes extending the framework to time-varying interference, memory effects, and real-time hardware implementations.