A power amplifier, the last analog stage of a radio transmitter, is run close to saturation to save energy. Near saturation it bends the signal: large amplitudes get compressed and rotated in phase, and the output depends on recent past inputs. Transmitters undo this with digital predistortion (DPD), a model that pre-warps the signal so that the amplifier's distortion cancels it. For two decades the standard model has been the generalized memory polynomial (GMP), a fixed polynomial expansion over current and delayed samples that is fitted by ordinary least squares. A growing literature reports that neural networks do clearly better. Residual Neural Networks for Digital Predistortion reports −38.0 dB normalized mean square error (NMSE) against GMP's −33.41 dB on a real GaN amplifier. In each of these papers, however, the polynomial baseline is evaluated at one hand-picked memory depth and nonlinear order. The neural model, by contrast, receives the authors' full design attention.
Language modeling and recommender systems have already shown that margins of this kind can disappear once baselines get the same tuning effort (On the State of the Art of Evaluation in Neural Language Models). We propose to measure whether that happens in DPD. Neural and polynomial predistorters would be compared under explicitly matched tuning budgets and matched inference cost, and we report how each family's best achievable result grows with tuning effort. The cost of tuning, which is what normally makes such audits impractical, becomes part of what we measure rather than a reason to skip the comparison.
Core Claim
The claim to be tested is specific. On public measured-amplifier data, the reported advantage of neural DPD over polynomial DPD at equal inference cost (measured in floating-point operations, FLOPs, per sample) is partly an artifact of unequal hyperparameter search. A GMP whose memory depths, cross-term lags, nonlinear orders, and ridge penalty are searched as thoroughly as a neural model's widths, learning rates, and memory taps will close a measurable fraction of the published gap.
We commit in advance to a quantitative test. Take the gap between the best neural and best polynomial predistorter at matched FLOPs and matched tuning compute. If that gap falls below 1 dB NMSE, or below about 1 dB in ACPR (adjacent-channel power ratio, the power leaked into neighboring frequency channels), the "neural beats polynomial" narrative needs heavy qualification. If the gap stays near the published 4.5 dB, the neural claim is validated for the first time under fair conditions. Both outcomes are useful, and the study is designed so that neither is presupposed.
What is new here is not a new predistorter. It is the first set of budget-versus-performance curves for DPD model families, together with hyperparameter-sensitivity measurements and a check that cheap, short-signal tuning trials rank configurations correctly.
Relation to Prior Work
Three neural DPD papers motivate the audit, because each compares against a fixed-configuration baseline:
- The residual time-delay network R2TDNN (Residual Neural Networks for Digital Predistortion) routes the input through an identity shortcut so the hidden layers learn only the nonlinear residual.
- The cascaded amplifier and predistortion networks of Learning to Compensate use frequency-domain losses whose weights were set by trial and error. They are compared against a fixed memory polynomial.
- TRes-DeltaGRU in OpenDPDv2 jointly optimizes quantization and temporal sparsity, but only for inference cost. It leaves training hyperparameters and baseline tuning unexamined.
We reuse their architectures and data. We differ in giving every model family the same search.
Two papers supply the evaluation method. Melis et al. show that a well-tuned LSTM reaches 58.3 test perplexity on Penn Treebank, below the 62.4 of a cell found by neural architecture search. Quantifying the Hyperparameter Sensitivity of Neural Networks for Character-level Sequence-to-Sequence Tasks supplies concrete sensitivity metrics: relative performance at rank k, and the rank correlation between distance in configuration space and accuracy. We port both into the radio-frequency setting.
The tuning machinery comes from BOHB. BOHB combines early stopping of weak trials with model-guided sampling of new ones, and it is our main tuner. Melis et al. also constrain the design: they warn that repeated evaluation on one validation set lets the tuner fit noise. We therefore hold out test signal frames that no tuner ever sees.
Research Program
We build one shared tuning harness that wraps four model families behind identical data splits and an identical tuner interface:
- GMP fitted by ridge least squares;
- sparse GMP, in which terms are pruned by an L1 penalty;
- R2TDNN;
- a recurrent DPD (a GRU and TRes-DeltaGRU from OpenDPDv2).
Each family gets a deliberately generous search space. For GMP this covers aligned and cross-term memory depths, cross-lag counts, nonlinear order, and ridge strength. For the neural families it covers widths, depth, memory taps, learning rate, weight decay, and, where relevant, loss weights.
Tuning budget is defined in two ways, and both are reported. One is equal wall-clock compute on the same hardware. The other is an equal number of full-fidelity trials. GMP trials are cheap, so under equal compute the polynomials effectively get an exhaustive search. That asymmetry is genuine and favors them, so we present it alongside the trial-matched view rather than hide it.
Each family is summarized by a curve: the expected best held-out NMSE and ACPR after k trials, estimated by bootstrap resampling over a large pool of random-search trials. These curves are plotted against both k and compute. The full comparison is repeated within FLOPs bins, so that no family wins simply by being bigger. The result is an accuracy-versus-FLOPs Pareto front for each family under each budget.
The overhead question enters through the fidelity schedule. Neural trials on full captures are what make fair tuning expensive. We therefore test whether short-capture training (a fraction of the samples) combined with few epochs ranks configurations the way full training does, using Spearman correlation between low-fidelity and full-fidelity scores. As an exploratory arm, we also test training on a narrower-bandwidth version of the signal. This may change the amplifier's memory behavior and so break the ranking, which is exactly what we want to find out. The cheaper fidelity enters BOHB only if its rank correlation is high. Otherwise BOHB uses epochs alone.
Baselines
The naive reference is each family exactly as published: GMP at the configuration used in Residual Neural Networks for Digital Predistortion, R2TDNN at its published architecture, and TRes-DeltaGRU at its OpenDPDv2 defaults. Each is retrained on our splits. This reproduces the "illusion" condition and anchors the size of the published gap.
- The reference numbers on RF WebLab are NMSE −38.0 dB for R2TDNN versus −33.41 dB for GMP, with R2TDNN at an ACPR of −42.5 dBc and about 500 FLOPs (Residual Neural Networks for Digital Predistortion).
- On APA_200MHz the reference is the FP32 TRes-DeltaGRU at ACPR −59.9 dBc and EVM (error vector magnitude, the in-band error) −42.1 dB (OpenDPDv2).
The strongest existing competitor is the BOHB-tuned neural model, specifically TRes-DeltaGRU on APA_200MHz and R2TDNN on WebLab. BOHB is the tuner with the strongest reported efficiency in the file: 20× to 55× speedups over standard baselines (BOHB). It runs on the same splits, the same hardware, and the same compute cap as the polynomial search. If the tuned neural model no longer beats −59.9 dBc after our tuning, we report that rather than an inflated number. A tuned polynomial "beats" or "matches" this competitor if, within the same FLOPs bin, its held-out NMSE and ACPR lie within 1 dB of the competitor's across three random splits.
The key ablation separates the effect of tuning from the effect of model family. We cross tuned and untuned conditions with both families. The difference between untuned GMP and tuned GMP measures how much of the published gap comes from the baseline being under-tuned. The difference between untuned and tuned neural models measures how much the authors' own design effort contributed.
Sparse GMP isolates whether any polynomial recovery comes from selecting the right cross terms or simply from adding more of them.
Why not just…?
Why not just fit a large GMP with L1 pruning once and skip hyperparameter search? L1 pruning is part of the search, not a substitute for it. The penalty strength and the candidate term set (maximum lags and order) are themselves hyperparameters, and pruning cannot add terms that the candidate set excludes. The sparse-GMP arm runs at the same budget, so if a single large pruned GMP matches the full search, the curves will show it, and that would itself be a reportable simplification.
Why not just run plain random search on everything and skip multi-fidelity tuning? This is a real threat to the overhead component, and it is why the cheap-fidelity arm is conditional rather than central. Random search over a large trial pool is the source of our budget curves anyway. BOHB earns its place only if it reaches the random-search curve's final value with clearly less compute. If short-capture fidelities do not preserve rankings, we drop them, and the audit conclusions still stand.
Why not tune online with hypergradients during a single training run? Online Learning Rate Adaptation with Hypergradient Descent adapts the learning rate by taking dot products of successive gradients. Self-Tuning Networks tune regularizers inside one run. Both handle continuous hyperparameters, but the variables that matter most here are discrete: memory depth, polynomial order, and widths. Hypergradients also add their own step size, which must itself be chosen. We include hypergradient learning-rate adaptation as an optional inner optimizer for the neural arms. It cannot replace the search over structure.
Why not just trust the reported results, since GMP's floor under deep saturation is physically plausible? Learning to Compensate does report up to 34% gains in severe-distortion regimes. A physical floor may well exist. The point of this study is to locate that floor with a tuned baseline instead of assuming where it is.
Proposed Experiments
The primary testbeds are APA_200MHz and the Chalmers RF WebLab GaN amplifier.
- APA_200MHz: 98,304 samples from a 3.5 GHz GaN Doherty amplifier driven by a 200 MHz 256-QAM OFDM signal, as used by OpenDPDv2.
- RF WebLab: measurements as used in Residual Neural Networks for Digital Predistortion.
Metrics are held-out NMSE, ACPR, EVM, and FLOPs per sample, each reported across three temporal train/validation/test splits with confidence intervals.
There are four decisive checks:
- Gap shrinkage. Within matched FLOPs bins, does the tuned-GMP versus tuned-neural gap fall below 1 dB from the published roughly 4.5 dB NMSE?
- Sensitivity. Do neural families show lower relative performance at rank k and rougher configuration landscapes than GMP, as Transformers did relative to LSTMs in Wiemerslage et al.? Rougher landscapes would explain why they need more tuning.
- Fidelity ranking. Does a short-capture fidelity preserve configuration rankings at a Spearman correlation of 0.8 or higher, which would justify cheaper fair tuning?
- Tuner overfitting. Is the gap between validation and test performance comparable across families?
If the data from Learning to Compensate is available, we use it as a third, supply-voltage-varying test of whether conclusions transfer across operating conditions. If it is not available, we restrict claims to the two testbeds above.
References
- Residual Neural Networks for Digital Predistortion
- On the State of the Art of Evaluation in Neural Language Models
- Learning to Compensate: A Deep Neural Network Framework for 5G Power Amplifier Compensation
- OpenDPDv2: A Unified Learning and Optimization Framework for Neural Network Digital Predistortion
- Quantifying the Hyperparameter Sensitivity of Neural Networks for Character-level Sequence-to-Sequence Tasks
- BOHB: Robust and Efficient Hyperparameter Optimization at Scale
- Online Learning Rate Adaptation with Hypergradient Descent
- Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions