⚡ electrical engineering

Low Complexity Neural Network Digital Predistortion of Wideband Power Amplifiers through Feature Selection

This paper proposes a low-complexity neural network-based digital predistortion architecture that utilizes an offline LASSO and MRMR feature selection pipeline to achieve up to a 30% reduction in computational cost while maintaining linearization performance comparable to polynomial methods for wideband power amplifiers.

Original authors: Cel Thys, Rodney Martinez Alonso, Ali H Alsarraf, Dominique Schreurs, Sofie Pollin

Published 2026-07-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Cel Thys, Rodney Martinez Alonso, Ali H Alsarraf, Dominique Schreurs, Sofie Pollin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Digital Tuner: Why Your Phone Needs Smarter Math

Imagine you are trying to shout a secret message across a noisy, crowded stadium. If you just yell at full volume, your voice might distort, crack, or get lost in the echo, making it impossible for your friend to understand. In the world of wireless communication, this "yelling" is done by a device called a Power Amplifier (PA). Its job is to boost weak signals from your phone or a cell tower so they can travel far. But here's the catch: when these amplifiers work hard to be efficient, they get "grumpy" and start twisting the signal, adding unwanted noise and distortion. This is like your voice cracking when you try to hit a high note.

To fix this, engineers use a clever trick called Digital Predistortion (DPD). Think of DPD as a "pre-correction" step. Before the signal hits the amplifier, a computer intentionally messes it up in the exact opposite way the amplifier will. When the grumpy amplifier then tries to distort the signal, it accidentally fixes the mess the computer made, leaving a clean, perfect message on the other side. As our world demands faster internet and wider channels (like the new "FR3" bands used for future 6G networks), the signals get more complex, and the amplifiers get even more grumpy. This forces the computers doing the pre-correction to become incredibly complex, requiring massive amounts of computing power—like hiring a thousand mathematicians just to fix one sentence. The big question becomes: Can we get the same perfect signal without needing a supercomputer?

The Paper's Big Idea: Picking the Best Ingredients

This paper, titled "Low Complexity Neural Network Digital Predistortion of Wideband Power Amplifiers through Feature Selection," tackles that exact problem. The authors, a team from Belgium, noticed that while Neural Networks (NNs) are amazing at fixing these distorted signals—often doing a better job than older methods—they are also incredibly heavy and slow. They require thousands of mathematical steps (called FLOPs) for every single piece of data, which is too much for the tiny chips inside our devices.

The team's solution is a bit like cooking a gourmet meal but only using the absolute best ingredients. Instead of throwing every possible spice and vegetable into the pot (which is what current complex neural networks do), they created a smart "shopping list" generator. They call this a Feature Selection pipeline.

Here is how their "smart shopping" works:

The Big Pile: First, they generate a massive list of potential clues (features) that could help the computer understand how the amplifier is misbehaving. This list is based on old-school math models and is huge, containing over 300,000 potential clues.
The First Cut (LASSO): They use a tool called LASSO to quickly chop off the useless clues. It's like a strict chef who says, "We don't need 50 types of salt; let's keep only the ones that actually add flavor." This step cuts the list down to a more manageable size.
The Final Pick (MRMR): Next, they use a method called MRMR (Minimum Redundancy Maximum Relevance). This is the ultimate filter. It looks at the remaining clues and asks, "Which of these tell us something new and important that the others don't?" It removes any clues that are just repeating the same information.

By the end of this process, they have a tiny, super-efficient list of the top 100 to 200 most important clues. They feed this short list into a Neural Network that is much smaller and faster than the usual ones.

What They Found: Faster, Smarter, and Just as Good

The researchers tested this new "Feature Selection" method on real hardware, specifically two different power amplifiers operating in the 15 GHz range (part of the FR3 band). They compared their new method against the standard, heavy-duty neural networks and the older polynomial models.

The results were quite promising. In their tests, the new method managed to reduce the computational cost by up to 30% compared to the standard neural network, while still achieving the same level of signal clarity.

For one of the amplifiers, they achieved a signal quality (measured as NMSE) of -37 dB using only 595 FLOPs (mathematical steps). The standard neural network needed 797 FLOPs to get a similar result.
On the second, more difficult amplifier, they reached -34 dB quality with just 234 FLOPs, a 32% reduction in effort compared to the standard approach.

When they actually ran the signals through the physical hardware and measured the real-world performance, the new method shined again. It produced lower error rates (EVM) and cleaner signals (ACLR) than the standard neural network, even when using fewer resources. For instance, one of their models achieved an error rate of just 1.38% with only 407 FLOPs, whereas the standard model needed more power to reach a slightly higher error rate of 1.88%.

The Takeaway

The authors suggest that by doing the hard work of picking the right "ingredients" before the neural network starts cooking (an offline process), we can build a much lighter, faster system that runs in real-time. They didn't just simulate this on a computer; they measured it on actual hardware and released their data for others to check.

While they don't claim to have solved every problem in the world, they have shown a clear path forward: we don't need to make our neural networks bigger and heavier to get better results. Sometimes, the secret is just knowing which clues to ignore. This approach could help make future 6G networks more efficient, allowing our devices to talk to each other faster without burning out their batteries or needing massive, expensive chips.

Technical Summary: Low Complexity Neural Network Digital Predistortion of Wideband Power Amplifiers through Feature Selection

Problem Statement
The evolution toward 6G wireless systems, particularly in the Frequency Range 3 (FR3, 7–24 GHz), necessitates the use of highly efficient power amplifiers (PAs) operating near saturation. This operation introduces significant nonlinear distortion, requiring Digital Predistortion (DPD) for linearization. While Neural Network (NN) based DPD approaches, such as Phase-Normalized Neural Networks (PNN), have demonstrated superior linearization performance compared to traditional Volterra-based models (e.g., Memory Polynomial (MP) and Generalized Memory Polynomial (GMP)), they face a critical barrier to practical deployment: computational complexity. NN models typically require thousands of Floating-point Operations (FLOPs) per IQ sample, significantly exceeding the few hundred FLOPs typical of polynomial models. This high computational cost hinders real-time implementation in resource-constrained digital frontend hardware.

Methodology
The authors propose a "Feature Selection NN-DPD" architecture designed to bridge the gap between the high performance of NNs and the low complexity of polynomial models. The methodology decouples the DPD problem into two distinct stages:

Offline Feature Engineering Pipeline:
- Feature Generation: The system generates a comprehensive, Volterra-inspired feature space from the input signal. This includes real and imaginary parts of delayed samples, envelope terms ( $|x(n)|^p$ ), and cross-terms combining I/Q components at different memory depths. This creates a large candidate feature matrix ( $F \approx 321,200$ features).
- LASSO Pruning: To address the high mutual correlation among high-order Volterra basis functions, the Least Absolute Shrinkage and Selection Operator (LASSO) is applied. This $\ell_1$ -regularized regression promotes sparsity, reducing the candidate set to approximately 2,000 non-zero features.
- MRMR Ranking: The Minimum Redundancy Maximum Relevance (MRMR) algorithm is then used to rank the remaining features. It selects a subset of features that maximize mutual information with the target output while minimizing redundancy among the features themselves. A hyperparameter $k$ (typically 100–200 features) determines the final input size.
Online Low-Complexity Inference:
- The selected top- $k$ features serve as the input to a compact, real-valued Time-Delay Neural Network (TDNN).
- The architecture utilizes a residual PNN structure with phase-normalization operations applied to the inputs and outputs.
- Crucially, the computationally intensive feature selection occurs offline. During real-time operation, only the generation of the top- $k$ features and the NN inference are performed, drastically reducing the runtime FLOP count.

Key Contributions

Novel Architecture: Introduction of a Feature Selection NN-DPD that replaces the standard, computationally heavy data-driven feature extraction of NNs with a pre-computed, optimized feature set.
Hybrid Pipeline: Integration of LASSO and MRMR algorithms to construct a compact, informative, and non-redundant input representation derived from the same Volterra features used in GMP models.
Open Benchmarking: The authors publicly released measured wideband FR3 PA datasets (two 15 GHz devices with 100 MHz signals) to support reproducible research and open benchmarking.

Results
The proposed method was validated using measured data from two distinct PAs (a low-gain GaAs amplifier and a high-gain amplifier) driven by 100 MHz 64-QAM signals.

Complexity Reduction: The Feature Selection NN achieved a reduction in computational complexity of up to 30% compared to the baseline PNN while maintaining comparable linearization performance. Specifically, for a target Normalized Mean Square Error (NMSE) of -37 dB, the proposed method required 595 FLOPs versus 797 FLOPs for the baseline PNN (a ~25% improvement). In a second device test, it achieved -34 dB NMSE with 234 FLOPs, a 32% reduction over the PNN baseline.
Linearization Performance: In terms of Error Vector Magnitude (EVM) and Adjacent Channel Leakage Ratio (ACLR), the proposed model consistently outperformed or matched the PNN baseline at equivalent complexity levels. For instance, with 407 FLOPs, the proposed model achieved 1.38% EVM compared to 1.88% for the PNN. It also demonstrated the lowest worst-case ACLR, improving by up to 10 dB over the PA without DPD and by 3.8 dB over a comparable PNN model.
Trade-off Optimization: The study utilized Bayesian optimization to tune hyperparameters (including the number of input features $k$ ), demonstrating that the proposed approach shifts the Pareto front, offering better performance for a given complexity or lower complexity for a given performance.

Significance and Claims
The paper claims that the proposed approach successfully improves the complexity-performance trade-off of NN-based DPD. By targeting the input features rather than the network architecture alone, the method allows NN models to match the linearization accuracy of state-of-the-art PNNs while significantly reducing the computational burden. This reduction makes NN-based DPD more viable for real-time deployment in resource-constrained hardware, addressing the core obstacle that has limited the adoption of NN DPD in practical systems. The authors note that the work is validated on measured FR3 data and that the public availability of these datasets supports further research in open benchmarking. They modestly suggest that future work could involve inspecting the selected features to develop more interpretable NN DPD algorithms.

The Digital Tuner: Why Your Phone Needs Smarter Math

The Paper's Big Idea: Picking the Best Ingredients

What They Found: Faster, Smarter, and Just as Good

The Takeaway

More like this