{"id":"1481a0b0-dd16-4860-a039-eaf00768a206","arxiv_id":"2607.15441","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An offline LASSO+MRMR feature-selection pipeline lets a neural-network digital predistorter match phase-normalized-network accuracy with 25-32% fewer FLOPs on measured FR3 power amplifiers.","lead":"Digital predistortion corrects power-amplifier distortion, but neural-network versions are too costly for real-time radio. This paper proposes an offline feature-selection pipeline (LASSO plus MRMR) that feeds a compact set of Volterra-style features into the network, cutting runtime FLOPs by up to 30% at similar linearization quality on measured FR3 amplifiers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FLOP comparison omits cost of generating selected Volterra features; Table I lacks an input-feature-generation row despite §III claiming it is counted, so the 25–30% savings may be an artifact.","rationale":"The reader's weakest assumption precisely identifies the omission of feature-generation cost in the FLOP model. My review confirms this is the load-bearing issue: the manuscript explicitly claims this cost is counted (§III) but never provides a formula, and the proposed method's input features are much more expensive to construct than the PNN's raw I/Q delay lines. The magnitude of the omitted cost is plausible to be on the order of hundreds of FLOPs, which could eliminate the claimed 25–32% savings. This is an internal accounting inconsistency, not a disagreement with the field's consensus. The verdict should remain CONDITIONAL: the performance measurements (EVM/ACLR) may be sound, but the headline complexity claim is unverifiable until the missing feature-generation cost is supplied and the comparison is recomputed. No other concern is as load-bearing; the linearization results and public dataset are positives. I recommend no change to the CONDITIONAL verdict until the test is run.","tokens_in":8748,"tokens_out":4540,"duration_ms":44780,"concrete_test":"Derive and compute C_FEAT for the actual top-k feature sets used for Figs. 5 and 7 (e.g., k=100–200). For each selected feature type in Eqs. (7)–(10), count the real FLOPs to generate it from the raw I/Q stream (magnitude, powers, cross-products, delay-line access), sum these to the Table I totals, and re-plot the NMSE-vs-FLOP curves. If the Feature Selection NN no longer shows a positive FLOP advantage over PNN at equal NMSE (e.g., −37 dB or −34 dB), the paper's central claim is invalidated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a 25–32% FLOP reduction at fixed NMSE. The complexity model in Section II-C/Table I lists only phase normalization (CPN), fully connected (CFC), and tanh (CTANH) for NN models. However, Section III explicitly states that 'only the top-k feature generation and NN model inference are counted' in the runtime complexity. Table I has no row for input feature generation, and no formula is derived anywhere for the cost of computing the selected features from Eqs. (7)–(10). This is not a minor omission: for the proposed Feature Selection NN, the input is a vector of k≈100–200 features including envelope terms |x(n−m)|^p, I/Q cross-products, and products with envelope terms. Computing each such feature from raw I/Q samples requires real multiplications, additions, possibly magnitude/square-root operations, and exponentiations. Even at 2–3 FLOPs per selected feature on average, this adds 200–600 FLOPs per sample, which is comparable to the entire reported 595-FLOP budget of the proposed model at −37 dB NMSE. The PNN baseline consumes raw delayed I/Q samples, whose generation is essentially free (or a memory read). Thus the comparison as reported may be systematically under-counting the proposed method. Because the headline contribution is specifically a complexity reduction, an unaccounted cost of this magnitude could erase or invert the claimed advantage. The text gives no way to recover the missing cost, so the main quantitative claim is unverifiable from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Feature Selection Neural Network (NN) architecture for digital predistortion (DPD) of wideband FR3 power amplifiers. An offline pipeline uses LASSO to prune a large Volterra-inspired feature library and MRMR to rank and select a compact set of k real-valued features; these features are then fed into a residual phase-normalized NN. The authors report measured NMSE, EVM, and ACLR results on two 15 GHz devices with 100 MHz signals, and claim up to 30% FLOP reduction relative to a PNN baseline at comparable linearization performance. The measured FR3 PA datasets are publicly released.","tokens_in":9096,"tokens_out":6163,"duration_ms":68693,"significance":"If the complexity accounting is correct, the paper makes a practically relevant contribution: it addresses the gap between NN-DPD accuracy and real-time implementation cost, and the use of measured FR3 data with public datasets is a genuine strength. The offline feature-selection idea is sensible, and the measured EVM/ACLR results (Table II) support the linearization capability of the proposed model. However, the central quantitative claim—a 25–32% complexity reduction—depends on a FLOP model that is currently incomplete. Because the paper's headline contribution is specifically a complexity reduction, this accounting issue is load-bearing and must be resolved before the claim can be accepted.","major_comments":[{"comment":"The text states that \"only the top-k feature generation and NN model inference are counted\" in the runtime complexity, but Table I has no term for input feature generation. The proposed NN input consists of features of the forms (8)–(10), which require computing envelope magnitudes |x(n−m)|, odd powers, and real multiplications between I/Q and envelope terms. For k≈100–200, even a low estimate of 2–3 FLOPs per selected feature adds 200–600 FLOPs per sample—comparable to the 595-FLOP point used for the 25% saving in §IV-A and to the 407-FLOP point in Table II. The PNN baseline consumes delayed raw I/Q samples, whose generation cost is essentially a memory read. Without deriving and adding C_feature_gen, the claimed 25–32% reduction may shrink or vanish. This is load-bearing because the headline claim is a complexity reduction.","section":"II-C/Table I; §III"},{"comment":"It is unclear how phase normalization is applied to the proposed feature set. Table I says phase normalization is applied to \"all input features\" and to the NN output, but §III-D says it is kept \"for I and Q inputs.\" For envelope terms such as |x(n−m)|^p and products like x_I(n−l)|x(n−m)|^p, phase normalization is not a well-defined complex multiplication. The formula C_PN(I)=6(I+1) therefore may over- or under-count the operations for the proposed model. This ambiguity affects the FLOP comparison and must be clarified, with a precise definition of the phase-normalization operation for each feature type.","section":"III-D; II-C"},{"comment":"The headline Pareto points (e.g., 595 vs. 797 FLOPs for the first DUT, and 234 FLOPs for the second DUT) are reported without the corresponding architecture details: the value of k, the hidden-layer sizes, the number of layers, and the composition of the selected feature set. Since the central claim is a quantitative complexity comparison, these details are necessary for reproducibility. Please include a table of the exact architectures and feature indices for the reported Pareto points, or otherwise make the selected features publicly available.","section":"IV-A; Figs. 5 and 7"}],"minor_comments":[{"comment":"The total number of candidate features is stated as F=321200 for M=200 and P=7, but no formula for F is provided. Give the general expression to make the feature-basis construction reproducible.","section":"III-A"},{"comment":"The MRMR step uses mutual information between continuous features and the target, but no estimator or discretization scheme is specified. This is important for reproducibility of the selected feature ranking.","section":"III-C"},{"comment":"The \"No DPD\" row reports EVM and ACLR but not NMSE; consider aligning the columns or explaining why NMSE is not applicable.","section":"Table II"},{"comment":"The abstract states \"up to 30% reduction\", while §IV-A reports 25% for DUT1 and 32% for DUT2. Clarify the distinction between the two datasets in the abstract or conclusion.","section":"Abstract / §IV-A"},{"comment":"The 10-FLOP tanh cost is attributed to a single technical report [17]. A brief justification of why this is representative for low-complexity hardware implementations would be helpful.","section":"II-C"}],"recommendation":"major_revision","confidential_remarks":"The core concern is whether the reported FLOP savings survive once the cost of generating the selected Volterra-like features is included. This is not a question of experimental quality—the measured results are credible—but of the accounting premise of the paper's main claim. If the feature-generation cost is added and the advantage persists, the paper would be a solid empirical contribution; if not, the central claim must be revised. I also recommend checking the phase-normalization definition for envelope and cross-product features."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful engineering paper with a plausible complexity-saving pipeline, but the headline 25–30% FLOP reduction is not checkable from the manuscript. The missing piece is the feature-generation cost that separates the proposed method from the baseline PNN.\n\nWhat is actually new: using LASSO plus MRMR offline to select a compact set of Volterra-like features as the input to a residual phase-normalized NN, instead of feeding raw delayed I/Q samples. That combination has been used for polynomial DPD but not, as far as I can tell, for NN-DPD input design. The authors also release measured FR3 PA datasets, which is real and welcome. The measured EVM/ACLR results on two 15 GHz devices support the linearization claim, and the reported Pareto curves show the proposed model reaching similar NMSE at lower listed FLOPs than PNN.\n\nThe soft spot is the central one. Section III explicitly says runtime complexity counts top-k feature generation plus NN inference, but Table I only includes phase normalization, fully connected layers, and tanh. There is no expression for computing envelope terms, I/Q cross-products, or products with envelope terms—the actual selected features. With k in the 100–200 range, even a few FLOPs per feature adds hundreds of FLOPs per sample, comparable to the claimed 595-FLOP budget. The PNN baseline takes raw delayed I/Q, which is nearly free, so the comparison may be systematically under-counting the proposed method. I don't think the saving necessarily disappears, but the paper as written gives no way to check. That is not a minor omission; it is the load-bearing claim.\n\nOther concerns are minor: single-point measurements without error bars, no code or exact feature/network configuration beyond the dataset, and only two DUTs. The citation pattern looks fine. The paper is honest about what is offline and what is real-time, and the dataset is a genuine contribution.\n\nWho this is for: people working on low-complexity NN DPD for FR3/6G will want the datasets and the measured linearization results. A serious referee should push for a complete FLOP derivation—either an honest accounting of feature generation or a hardware/fixed-point implementation cost—before the complexity claim can be accepted.\n\nRecommendation: send it to peer review, but make the FLOP accounting fix a condition. If the 25–30% saving survives that, this is a solid, practical result. If not, the linearization results still stand, but the paper's main claim needs to be rewritten.","headline":"Sensible feature-selection DPD with good measured linearization, but the central FLOP-savings claim is unverifiable because the cost of generating the selected features—the very thing that distinguishes it from the baseline—is missing from the complexity model.","tokens_in":9616,"tokens_out":3100,"would_cite":true,"duration_ms":35614,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Offline feature selection lets compact neural networks linearize FR3 power amplifiers at 25–32% lower computational cost.","keywords":["digital predistortion","neural network DPD","feature selection","LASSO","MRMR","FR3 power amplifier","Volterra features","NMSE-complexity tradeoff"],"falsifier":"Implement the selected top-k feature set (terms of the forms |x(n−m)|^p, xI(n−l)xQ(n−m), xI(n−l)|x(n−m)|^p, and their Q counterparts) in a cycle-accurate DSP/RTL model that counts every real multiply and add per sample, then recompute the NMSE-vs-FLOP Pareto curves. If the added feature-generation FLOPs exceed the roughly 200-FLOP gap between the proposed model and PNN at matched NMSE, the central complexity claim collapses.","tokens_in":8591,"feed_emoji":"📡","tokens_out":6300,"duration_ms":57880,"temperature":0.7,"pith_summary":"The paper claims that neural-network digital predistortion (DPD) can be made much cheaper without losing accuracy, if the expensive part — finding which inputs matter — is moved offline. It proposes a two-stage pipeline: build a huge library of Volterra-inspired real-valued input features, prune and rank them with LASSO and MRMR, and feed only the top ~100–200 features into a small residual phase-normalized network. On measured 15 GHz FR3 power amplifiers with 100 MHz signals, this Feature Selection NN reaches −37 dB NMSE at 595 FLOPs and −34 dB NMSE at 234 FLOPs, corresponding to roughly 25% and 32% reductions versus a phase-normalized NN baseline at comparable accuracy. Measured EVM and ACLR on one device confirm the trend. If correct, the approach narrows the long-standing complexity gap between neural-network DPD and polynomial DPD, removing a barrier to real-time deployment.","feed_headline":"Feature selection cuts NN predistortion cost by up to 30%","feed_subtitle":"An offline LASSO+MRMR step picks compact features so a small network rivals PNN linearization on FR3 power amps.","key_machinery":"The machinery is an offline feature-engineering pipeline followed by a small network. LASSO (ℓ1-regularized least squares) prunes a 321,200-column library of real-valued Volterra-inspired features to roughly 2,000 survivors, and MRMR greedily ranks those survivors by mutual information with the target minus redundancy with already-chosen features, yielding a top-k input set with k around 100–200. The online stage is a residual phase-normalized neural network whose per-sample cost is counted as feature generation plus phase normalization plus fully connected multiply-adds plus tanh activations; the offline LASSO/MRMR search is excluded because it does not run under real-time constraints.","core_discovery":"The central discovery is that the input representation, not the network size, is what makes NN DPD expensive. A residual phase-normalized network with only 100–200 pre-selected features — I/Q delay taps, odd-order envelope magnitudes, and I/Q cross-products of the kind used in GMP models — matches or beats the linearization accuracy of a phase-normalized NN while using roughly 25% fewer FLOPs on a 15 GHz GaAs PA (595 vs 797 FLOPs at −37 dB NMSE) and 32% fewer on a high-gain PA (234 FLOPs at −34 dB NMSE). The feature selection itself runs once offline on 15,000 samples from 321,200 candidates; only generation of the selected top-k features and NN inference count toward runtime complexity. Mea","pith_inferences":["A natural next step, left implicit, is to inspect which features MRMR selects across different PAs; a recurring small set could be hard-coded, removing the need for per-device offline selection.","The same offline-selection idea could be applied adaptively: periodically re-running LASSO/MRMR on new captures would keep the feature set compact as the PA ages or its operating point drifts, without raising per-sample cost.","A hardware-level test — synthesis or cycle-accurate RTL for the selected features and network — would show whether the FLOP savings survive real implementation overheads such as memory access and activation-table lookups.","Because the basis includes envelope and I/Q cross-terms, the approach is likely to extend to wider bandwidths (e.g., 200–400 MHz) where memory effects are stronger, though the optimal k may need to grow."],"forward_implications":["Neural DPD can operate at a few hundred FLOPs per sample while matching phase-normalized NN accuracy, bringing it into the same complexity regime as MP/GMP polynomial models.","The winning features are standard Volterra-like terms, so no new nonlinear basis is needed; the gain comes from selecting a compact subset instead of using all candidate terms.","The NMSE-vs-FLOPs Pareto front places the Feature Selection NN below and to the left of the PNN baseline in the low-complexity region on both measured devices.","Measured EVM and ACLR on the first DUT confirm the simulated advantage at roughly 400 FLOPs, with the proposed model reaching 1.38% EVM versus 1.88% for the PNN.","The released FR3 PA datasets let other groups reproduce the comparison and benchmark future DPD algorithms against the same devices."],"fun_headline_variants":["Feature selection cuts NN predistortion cost by 30%","Slimmer DPD via LASSO+MRMR: same linearity, 30% fewer FLOPs","Input pruning, not network size, drives DPD efficiency","Offline feature pick trims wideband PA predistortion costs","NN predistortion: 30% cheaper with smart feature selection"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The FLOP reduction hinges on a complexity model that counts the selected features' generation and the NN inference but excludes the offline selection cost; if generating the top-k Volterra-like features (envelope powers, I/Q cross-products) in real time costs more than Table I assigns, the claimed 25–32% saving shrinks.","fun_headline_variants_meta":{"raw":{"variants":["Feature selection cuts NN predistortion cost by 30%","Slimmer DPD via LASSO+MRMR: same linearity, 30% fewer FLOPs","Input pruning, not network size, drives DPD efficiency","Offline feature pick trims wideband PA predistortion costs","NN predistortion: 30% cheaper with smart feature selection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000539,"raw_usage":{"total_tokens":2413,"prompt_tokens":726,"completion_tokens":1687,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":1603}},"tokens_in":470,"tokens_out":1687,"duration_ms":11551,"temperature":1.0,"reasoning_tokens":1603,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T23:21:02.948318+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement the selected top-k feature set (terms of the forms |x(n−m)|^p, xI(n−l)xQ(n−m), xI(n−l)|x(n−m)|^p, and their Q counterparts) in a cycle-accurate DSP/RTL model that counts every real multiply and add per sample, then recompute the NMSE-vs-FLOP Pareto curves. If the added feature-generation FLOPs exceed the roughly 200-FLOP gap between the proposed model and PNN at matched NMSE, the central complexity claim collapses.","supporting_citations":[],"review_version":1}