Sparse Solution Trade-offs in GMP DPD: A Least Squares Thresholding Approach
summary
The gist
Sparse Solution Trade-offs in GMP DPD: A Least Squares Thresholding Approach is a practical case study on sparse kernel selection for radiofrequency DPD for power amplifiers which achieves a 20×…
In short
Researchers studied how to simplify Digital Pre-Distortion (DPD) for power amplifiers using Generalized Memory Polynomials (GMP). They compared Least Squares Thresholding (LST) against other greedy methods like Orthogonal Matching Pursuit (OMP), finding LST achieved a 20x complexity reduction while maintaining near-identical performance. This suggests LST is a practical, efficient method for selecting sparse kernel coefficients in DPD.
Key concepts
- Generalized Memory Polynomial (GMP)
- A specialized mathematical model used to describe the nonlinear distortion in power amplifiers. It extends standard Volterra series by including both aligned memory terms and cross-terms involving signal and envelope delays, which are crucial for accurately modeling complex nonlinear behavior.
- Least Squares Thresholding (LST)
- A method proposed to find a sparse solution (a small set of important coefficients) for the GMP model. It works by sorting the kernel columns once and then greedily selecting them based on a fixed order, offering a computationally cheaper alternative to methods that re-correlate at every step.
- Orthogonal Matching Pursuit (OMP)
- A greedy algorithm used to select the most relevant terms (kernels) from the GMP model. OMP repeatedly finds the term that best correlates with the remaining signal error, which is effective but computationally more intensive than LST.
- Performance Metrics
- Quantitative measures used to judge DPD quality, including Adjacent Channel Power Ratio (ACPR) for spectral regrowth, Error Vector Magnitude (EVM) for constellation fidelity, and Normalized Mutual Information (NMI) to assess overall signal statistical dependence.
Terminology used across episodes
This episode discusses
The paper
Sparse Solution Trade-offs in GMP DPD: A Least Squares Thresholding Approach · Read on arXiv
Ellison Murray, Morriel Kasher, Predrag Spasojevic
Rutgers University
Power amplifiers (PAs) in satellite communication systems introduce nonlinear distortion, degrading spectral fidelity. Digital pre-distortion linearizes the PA response, but full-complexity solutions are prohibitive under strict size, weight, and power (SWaP) constraints. We propose the use of Least Squares Thresholding (LST) and compare it against Orthogonal Matching Pursuit (OMP) and Matching Pursuit. LST achieves a 2.77x complexity reduction while maintaining near-identical linearization performance to OMP.
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Sparse Solution Trade-offs in GMP DPD".
Dev: Here's the gist of the paper. Sparse Solution Trade-offs in GMP DPD: A Least Squares Thresholding Approach is a practical case study on sparse kernel selection for radiofrequency DPD for power amplifiers which achieves a 20× complexity reduction while maintaining near-identical linearization performance to OMP and confirms MP’s poor performance for GMP-DPD.
Rosa: Problem Context Power amplifiers in satellite communication systems introduce nonlinear distortion, degrading spectral fidelity.
Dev: Digital pre-distortion (DPD) is utilized to linearize the PA response, but non-sparse solutions are prohibitive under strict size, weight, and power (SWaP) constraints.
Rosa: The power amplifier introduces severe nonlinear distortion to the transfer characteristic of the input and output magnitude.
Title and authors: Dev: The paper's title sets up the whole trade-off: Sparse Solution Trade-offs in GMP DPD: A Least Squares Thresholding Approach. It tells you immediately we are dealing with finding a sparse solution for Generalized Memory Polynomials using a specific method, LST.
Rosa: And the implication is that this LST approach actually maintains near-identical linearization performance to OMP, which is what makes it so attractive when you need to keep the signal quality high.
Dev: But they also confirm that Matching Pursuit has poor performance for GMP DPD because of its structure, and LST manages to bridge that gap while saving computation time.
Taro: It sounds like they're showing how a simpler selection method can beat more complex greedy searches in this specific context, which is interesting when you think about the real world failure modes we test on.
The paper's summary: Rosa: They summarize by saying that the power amplifier introduces severe distortion to the transfer characteristic of the input and output magnitude, and that spectral regrowth means emissions at out-of-band frequencies are seen.
Dev: To fight that, digital pre-distortion is used, modeling the amplifier with Volterra series or Memory Polynomials. But Volterra is too complex due to exponential complexity, so they focus on Generalized Memory Polynomials or GMP.
Taro: And the GMP model itself incorporates not just aligned memory terms but also cross-terms between the signal and its envelope at both leading and lagging delays. That’s a lot of terms to manage.
Rosa: They show that these coefficients are put into one big vector, which is then used to build a matrix B where each column represents one term in the polynomial model. The size of this GMP, which is the number of kernels, is determined by three parameters: H, L and J.
Dev: And they define this size as M equals L plus one times H plus one over two plus J times H minus one. That’s how many terms you have to potentially select from when you are trying to find the sparse solution x.
Taro: The paper explains that kernel selection itself is an NP-Hard problem, which is why they look at search-based methods like OMP and MP instead of just solving it directly.
The paper's improvements: Rosa: What they propose as the improvement is this Least Squares Thresholding approach, LST, which sorts columns once at initialization and then adds greedily based on that fixed order.
Dev: They claim this LST method achieves a twenty times complexity reduction compared to OMP when assuming a near-orthogonal kernel dictionary, which is a big win for the size weight and power constraints.
Taro: From my side, I see it as an improvement in terms of graceful recovery because if you can find the solution much faster in real time, that means less latency and potentially better handling of unexpected signal behavior.
Rosa: They also note that they use four different performance measures to evaluate this DPD: Adjacent Channel Power Ratio or ACPR, Normalized Mean Squared Error or NMSE, Error Vector Magnitude or EVM, and Normalized Mutual Information or NMI.
Dev: And what’s important here is the results show that LST matches the full solution's performance when they use twenty to twenty-five kernels.
Taro: The paper also points out that training is done offline, and a simpler training procedure doesn't necessarily help runtime efficiency in deployed systems, which is a practical limitation we have to keep in mind.
Conclusion: Rosa: To wrap up the paper "Sparse Solution Trade-offs in GMP DPD: A Least Squares Thresholding Approach," they show that LST achieves an EVM twenty-five point seven percent lower than the three gigahertz sixty four qam limit of negative twenty-one point nine four decibels, which is eight.
Dev: And they also show that LST gets an ACPR of negative thirty-one point nine seven decibels, which is lower than the no pre-distortion baseline's negative thirty-one point nine seven decibels.
Taro: The paper confirms that for the NMI metric, LST gives zero point seven four zero nine, which is comparable to OMP's zero point seven four zero nine, while Matching Pursuit gets a lower score of zero point six three four three.
Rosa: So the main thing is that LST offers computational savings reflected in Table two, and it proves that this sparse kernel selection method works for radio frequency DPD for power amplifiers.
Dev: The paper confirms the performance metrics, like ACPR and NMI, are good indicators of selecting the optimal sparse solution x hat for DPD kernel coefficients. This helps engineers quantify trade-offs between spectral regrowth control, symbol fidelity, and overall information retention.
Taro: I just think the future work mentioned about integrating this into an iterative learning loop or on FPGA hardware is where we need to look next to see how this actually lives in a deployed system.
Rosa: Right. It’s a practical case study that shows how we can get significant computational savings without losing the critical linearization performance of GMP DPD.
More episodes
- Daily Summary for 2026-11-19
- 2111.07910-Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- 2203.02149-HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- 2205.10102-Degradation-Aware Unfolding Half-Shuffle Transformer for Spectral Compressive Imaging
- 2311.10959-Structure-Aware Sparse-View X-ray 3D Reconstruction
- 2403.04116-Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis
- 2410.02572-Combining Pre- and Post-Demosaicking Noise Removal for RAW Video
- 2604.17453-Learned Nonlocal Feature Matching and Filtering for RAW Image Denoising
- 2603.20387-End-to-End Multi-Task Learning for Adjustable Joint Noise Reduction and Hearing Loss Compensation
- 2507.20825-Chirp-Permuted AFDM: A Versatile and Multifunctional Waveform for 6G