{"id":"cc9188c6-4681-40ab-9d74-bf66a8c70a5a","arxiv_id":"2411.14699","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-stage DNN compensates combined THz hardware imperfections, and combined slimming reduces parameters by 97.2% with near-identical symbol error rate.","lead":"This paper proposes a two-stage deep neural network method that compensates for multiple hardware imperfections in terahertz hybrid beamforming systems, and shows it can cut the network's parameter count by 97.2% while keeping the symbol error rate nearly unchanged. The work targets practical 6G THz transceivers where hardware non-idealities such as quantization noise, IQ imbalance, phase noise, and power amplifier nonlinearity otherwise degrade performance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stage-2 compensation network is diagonal per stream and cannot implement the general digital precoder FBB,c it claims, leaving the reported SER gains dependent on an undisclosed channel/beamforming realization.","rationale":"The reader's weakest assumption is sim-to-real transfer, which is a legitimate and important limitation. My concern is more internal: even within the paper's own simulation framework, the stage-2 compensation network may be unable to realize the digital precoder it claims to compute. The architecture in Sec. V.B processes each data stream independently, so it only implements a diagonal, memoryless nonlinear mapping from s to s1. A general digital precoder FBB,c is a complex matrix that mixes streams; unless the effective ideal channel C = W_BB^H W_RF^H H G_PA F_RF P_in F_BB is diagonal, a diagonal NNct cannot match the desired ideal received signal yi. The paper does not disclose H, FRF, WRF, FBB, or the channel model, and it uses one fixed hardware/channel realization for both training and testing (Sec. VII.A.2). The reported SER numbers therefore may be tied to a particular diagonalized setup rather than to a general compensation capability. This does not contradict the reader's CONDITIONAL verdict; it sharpens the condition that the authors must provide and justify the channel/beamforming setup, or the claims should be narrowed. If the off-diagonal test shows a non-negligible ratio and the compensation still works, the explanation must be that the hardware nonlinearities are themselves diagonalizable in that setup, which the paper should state explicitly. If the test instead shows the architecture cannot realize FBB,c in general, the verdict should move toward REJECT.","tokens_in":19630,"tokens_out":16972,"duration_ms":199678,"concrete_test":"Recreate the Sec. VII.A setup and compute the ideal noiseless end-to-end matrix C = W_BB^H W_RF^H H G_PA F_RF P_in F_BB for the exact H and beamformers used in the simulations; report ||C - diag(C)||_F / ||C||_F. If this off-diagonal ratio is non-negligible, then no diagonal NNct of the form in Eqs. (51)-(53) can make yct track yi for arbitrary s, and the claim that FBB,c = s1s^{-1} is a realized digital precoder fails. As a complementary check, retrain the Tx-compensation pipeline with an independent random H (e.g., a geometric THz channel with 3-5 paths) and a non-diagonal ideal FBB; if the SER at 10-15 dBm degrades substantially, the reported gains are setup-specific.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. V.B, the Tx-compensation network NNct is built from Lt independent sub-NNs, each with two input nodes for the real and imaginary parts of a single data stream and two output nodes; Eq. (51) defines c_{1,lt}=[Re(s_lt), Im(s_lt)]^T. The paper then states that 'FBB,c = s1s^{-1}' and treats this as a digital precoder. A digital precoder is a matrix in C^{Lt x Ns} that generally mixes all data streams; a per-stream, memoryless scalar network cannot represent off-diagonal entries of such a matrix. Consequently, NNct cannot compensate inter-stream coupling in the channel/beamforming product or implement general hybrid-beamforming precoding. The reported compensation gains in Figs. 7-10 therefore require that the effective ideal end-to-end mapping from s to yi be diagonal, or nearly so, for the specific H, FRF, WRF, and FBB used in Sec. VII.A. The paper never states the channel model, the beamformer design, or the diagonal dominance of that effective channel, and it trains and tests on a single fixed realization. Thus the central claim that the algorithm 'can effectively mitigate hardware imperfections' is not established for general THz hybrid beamforming even within simulation; the numerical SER values may be an artifact of a diagonalized setup rather than evidence that the proposed architecture can realize the claimed FBB,c.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a narrowband THz UM-MIMO hybrid-beamforming system with hardware imperfections at both the transmitter and receiver, including DAC/ADC quantization, IQ imbalance, phase noise, phase-shifter amplitude/phase errors, and PA nonlinearity. It proposes a two-stage deep-learning compensation scheme: stage 1 trains a modular DNN, built from per-RF-chain, per-antenna, and per-RF-chain sub-networks, to represent the combined impairments from pilot data; stage 2 trains either a transmitter-side NN (used in place of the digital precoder) or a receiver-side NN (used in place of the digital combiner) to minimize the MSE with respect to the ideal received signal. Three network-slimming methods are proposed and combined: reducing hidden-layer neurons, sharing parameters among the PA sub-networks, and removing the PA sub-network at low transmit power. Simulation results in uncoded and coded systems are reported to show that transmitter compensation outperforms receiver compensation and that the slimmed networks retain most of the compensation gain, with parameter reductions up to 97.2% and runtime reductions up to 39.2%.","tokens_in":19920,"tokens_out":11881,"duration_ms":121582,"significance":"If the central claim holds, the paper offers a practical data-driven alternative to per-imperfection calibration in THz systems: it treats the combined hardware impairments as a black box, uses a modular architecture whose parameter count is easy to analyze, and demonstrates substantial complexity reductions with explicit parameter and runtime tables. Strengths include the explicit consideration of a broad set of impairments, the coded-system evaluation, and a comparison against a direct-DNN baseline. However, the significance is conditional: the per-stream architecture of the transmitter compensation network, the absence of a specified channel and beamforming model, and the simulation-only training data all limit the generality of the demonstrated gains. The paper is a potentially useful contribution if these points are resolved or if the claims are appropriately restricted.","major_comments":[{"comment":"The transmitter-compensation network N Nct is built from Lt independent sub-NNs, each taking the real and imaginary parts of a single data stream as input and producing two outputs (Eq. (51)). Such a network implements a diagonal, memoryless, per-stream transformation and cannot represent a general digital precoder FBB,c in C^{Lt x Ns} that mixes the streams; the formula FBB,c = s1 s^{-1} is not a well-defined linear precoder when s1 is the output of a nonlinear network. Since the full-connected analog precoder FRF and the per-antenna PA nonlinearity in Eq. (24) create inter-stream coupling, this architecture cannot in general compensate the combined hardware imperfections of a hybrid-beamforming system. The reported SER gains in Figs. 7(c), 9, and 10 therefore rest on an unstated assumption that the effective ideal channel W_BB^H W_RF^H H F_RF P_in F_BB is diagonal, or nearly so, for the particular realization used. The authors should either redesign stage 2 to process all streams jointly or restrict the claim to a clearly specified diagonalized setup.","section":"Sec. V.B, Eq. (51)"},{"comment":"The simulation setup does not specify how H, FRF, WRF, FBB, and WBB are generated: no THz channel model, no beamformer design rule, and no number of channel realizations are given. The entire evaluation is performed on data produced by the authors' own simulation model (Sec. VII.A.2), with training and testing on a single fixed realization. Without the channel model and the effective-channel structure, the reader cannot determine whether the setup is representative of THz UM-MIMO or whether the per-stream compensation in Sec. V.B is adequate. The authors should report the channel model, the beamforming algorithm, the realized effective channel's off-diagonal strength, and results averaged over multiple channel realizations.","section":"Sec. VII.A"},{"comment":"The training data are generated from the same simulation model that is used for testing, and the text explicitly notes that real THz hardware data are unavailable. The numerical results therefore demonstrate that the DNN can approximate and invert the particular simulated impairment model (AQNM quantization, Gaussian phase noise, Rapp PA, and the assumed phase-shifter error statistics), not that the method transfers to physical THz hardware if any of these models is inaccurate. The abstract and conclusion should qualify the claim to the simulated model, or the authors should add a mismatched-model robustness study, for example training under one PA or phase-noise parameter set and testing under another.","section":"Sec. VII.A.2"},{"comment":"Phase noise and quantization noise are described as varying across symbols in the dataset generation (Sec. VII.A.2), but the DNNs in both stages are memoryless and have no input representing the instantaneous phase-noise or quantization realization. Under the MSE losses, such a network can only learn the conditional-mean deterministic distortion, not invert the symbol-varying random phase rotations. The paper should clarify that the compensation targets only the deterministic part of the impairments and should quantify the residual phase-noise and quantization floor; otherwise the claim that the algorithm 'can effectively mitigate the hardware imperfections' (abstract) is broader than what the architecture can achieve.","section":"Sec. V.A, Eqs. (45), (54), (58)"}],"minor_comments":[{"comment":"The term Γt2 x*2 should presumably be Γt2 x*1 or Γt2 (Pin x1)*; as written, x2 appears on the right-hand side of its own definition.","section":"Eq. (23b)"},{"comment":"The subscripts are inconsistent in the notation: for example, the network N N1 is described with 'lth r sub-NN' before Eq. (36), although the index should be the lth t sub-NN, and similar inconsistencies appear in the N N3 description.","section":"Sec. V.A"},{"comment":"What is called 'pruning' is actually a reduction of the number of hidden neurons before training, not pruning a trained network; renaming this to 'hidden-size reduction' or describing the pruning criterion would avoid confusion.","section":"Sec. VI.A"},{"comment":"The training details are incomplete: learning rate, optimizer, batch size, number of epochs, initialization, and random seeds are not reported, which prevents reproduction of the convergence curves in Fig. 7(a) and of the SER results.","section":"Sec. VII.A.2"},{"comment":"The signal ys is not defined in the text; the reader can infer that it is the noise-free received signal, but it should be stated explicitly.","section":"Sec. VII.C.2"},{"comment":"The abstract's 'combined slimming methods' reduce parameters by 97.2%, but parameter sharing alone already reduces them by 96.6%; the incremental gain of the combination is small and should be described more carefully.","section":"Table II and Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and addresses a timely problem. My main concern is the mismatch between the per-stream NN architecture and the claimed general digital precoder, together with the missing channel-model specification; these make the reported SER values hard to interpret. I would like the revision to address these issues or narrow the claims. The simulation-only limitation is acknowledged in the text, but it should be made more prominent in the abstract and conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my take. The genuinely new thing is the joint treatment of five THz hardware impairments in one hybrid-beamforming model and a DNN whose structure follows the signal chain, plus the three slimming tricks. That is a reasonable engineering contribution, and the reported 97.2% parameter reduction at small SER cost is internally consistent, not hand-wavy. The paper also does something right that many ML-for-comms papers skip: it reports convergence, uncoded and coded SER, and compares against a direct-DNN baseline. The impairment models are cited to relevant sources, and the slimming ablation is shown rather than just claimed.\n\nThe soft spot is load-bearing. In Sec. V.B, NNct is built from Lt independent sub-networks, each with two inputs and two outputs on a single stream. Such a network produces a diagonal mapping from s to s1, so FBB,c = s1 s^{-1} is diagonal, not a general Lt x Ns digital precoder. The compensation problem in (32) allows a full precoder; the implemented network cannot realize that. The reported SER numbers therefore depend on the effective channel/beamforming product being diagonal or nearly so for the single simulated realization. The paper never specifies the channel model, the beamformer design, or the diagonal dominance of that effective channel. That is a real gap, and it makes the conclusion that the algorithm \"can effectively mitigate hardware imperfections\" too broad for general THz hybrid beamforming. The stress-test note is right about this.\n\nThe second soft spot is the standard one: training and testing data both come from the paper's own simulator. The hardware models are reasonable and cited, but there is no external or measured validation. The authors are upfront about this in Sec. VII.A.2. That does not kill the paper, but it means the numbers are self-consistent rather than predictive of real hardware behavior.\n\nMinor: training hyperparameters, the hidden-layer count, and the 5 dBm threshold for removing NN2 look post hoc. I would not call that fatal; the ablation is at least visible.\n\nWho is this for? People building THz link-level simulators or prototyping baseband compensation schemes. It deserves a serious referee: the diagonal-vs-full precoding question needs to be settled, and the channel model needs disclosure, but the work is coherent and reproducible in principle. I would send it out and ask for those revisions. If the authors explicitly show that their setup is diagonalized, the results are useful. As it stands, I would not cite the general claim in my own work unless the scope is narrowed.","headline":"A competent, fully simulated DNN compensation study for THz hybrid beamforming whose headline claim outruns its architecture: the stage-2 precoder is per-stream, so the general-digital-precoder claim is not established.","tokens_in":20439,"tokens_out":2620,"would_cite":false,"duration_ms":27884,"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":"A two-stage neural algorithm learns all hardware imperfections of a THz hybrid beamforming link as a black box and compensates them at baseband, cutting symbol error rate to 0.0005 at 15 dBm.","keywords":["Terahertz communications","ultra-massive MIMO","hybrid beamforming","hardware imperfection","deep neural network","symbol error rate","network slimming","digital precoder compensation"],"falsifier":"Run the two-stage algorithm on a THz hybrid beamforming testbed with real measured impairments (or on a simulator using a different PA model, such as a memory polynomial, and different phase-noise statistics) and compare the SER at 15 dBm with the reported 0.0005 for transmitter compensation; if the SER rises to a clearly higher level, the simulation-to-reality transfer claim fails.","tokens_in":19424,"feed_emoji":"📡","tokens_out":12240,"duration_ms":99239,"temperature":0.7,"pith_summary":"The paper claims that the combined hardware impairments of a THz ultra-massive MIMO hybrid-beamforming link—DAC/ADC quantization, IQ imbalance, phase noise, imperfect phase shifters, and PA nonlinearity—can be treated as a single learnable black box for compensation, so a data-driven scheme can undo their joint effect without needing separate models for each imperfection. A two-stage algorithm first trains a structured DNN to reproduce the distorted received signal from pilot transmissions, then trains a second network to design the digital precoder or combiner that undoes the learned distortion. In simulation with 256 antennas and 4-bit converters, transmitter-side compensation lowers the symbol error rate to 0.0018 at 10 dBm and 0.0005 at 15 dBm, outperforming receiver-side compensation and a generic fully connected DNN baseline. The paper also shows that pruning, parameter sharing, and conditional removal of a sub-network cut parameters by 97.2% and runtime by 39.2% with nearly unchanged SER. A reader should care because this points to a practical, low-complexity way to keep THz links usable despite hardware that is far from ideal.","feed_headline":"Simulation: two-stage DNN cuts THz beamforming symbol errors to 0.0005","feed_subtitle":"In simulation, neural compensation restores 16-QAM links at 15 dBm while slimming the DNN by 97.2%.","key_machinery":"The load-bearing object is the stage-1 DNN, written as $\\hat{y} = \\mathbf{W}_{\\mathrm{BB}}^H \\mathcal{NN}_3(\\mathbf{W}_{\\mathrm{RF}}^H \\mathbf{H}\\, \\mathcal{NN}_2(\\mathbf{F}_{\\mathrm{RF}}\\mathbf{P}_{\\mathrm{in}} \\mathcal{NN}_1(\\mathbf{F}_{\\mathrm{BB}} \\mathbf{s})))$, a chain of small sub-networks (two real inputs from the I/Q components, one hidden layer with $\\mathrm{Tanh}$ activation, two outputs) arranged exactly as the signal flows through transmit RF chains, antennas, and receive RF chains. This network is trained with an MSE loss to reproduce the distorted received signal of the nonideal model, so it becomes a differentiable black-box model of the combined hardware imperfections. Stage 2 keeps that network fixed and trains a separate small NN to produce the compensating digital precoder (Tx) or combiner (Rx), reading off $\\mathbf{F}_{\\mathrm{BB},c}$ or $\\mathbf{W}_{\\mathrm{BB},c}$ from the network's input–output relation under the per-symbol power constraint. The slimming methods act on this machinery: pruning cuts hidden neurons, parameter sharing reuses one sub-network across antennas because all PAs share the same nonlinearity, and removing NN2 drops PA modeling when transmit power is low.","core_discovery":"The paper argues that in a THz UM-MIMO hybrid beamforming system the aggregate distortion produced by nonideal DACs and ADCs, IQ imbalance, phase noise, imperfect phase shifters, and nonlinear power amplifiers can be represented accurately by a DNN whose sub-networks follow the physical signal path: one sub-network per transmit RF chain, one per transmit antenna, and one per receive RF chain. Trained on pairs of transmitted pilots and the corresponding received signals, this network acts as a differentiable surrogate for the nonideal link. A second neural network then uses that surrogate to optimize either the digital precoder at the transmitter or the digital combiner at the receiver, minimizing mean-square error to the ideal received signal. Reported results for a 256-antenna, 4-stream, 4-bit-converter system show transmitter compensation reaching SER 0.0018 at 10 dBm and 0.0005 at 15 dBm, clearly better than receiver compensation and than a direct fully connected DNN baseline, which suggests the structured architecture and the clean training inputs at the transmitter are what make the compensation effective.","pith_inferences":["Editorial inference: the modular, signal-flow-matched architecture is why the method beats a generic fully connected DNN at high power; if so, the same modularity should transfer to other impairment sets, such as wideband beam squint or mutual coupling, by adding corresponding sub-networks.","Editorial inference: the strongest test of the paper's practical value is a hardware-transfer experiment; because the training data are simulated, a real THz testbed with measured impairments would reveal how much of the reported SER gain survives model mismatch.","Editorial inference: the learned surrogate could be reused beyond compensation, for example to monitor hardware aging or to supply gradients for model-based precoding design, since the DNN makes the distorted link differentiable with respect to baseband inputs."],"forward_implications":["If the claimed result holds, THz hardware-imperfection compensation becomes a data-driven baseband problem: collect pilot–received pairs, learn the distortion as a black box, and optimize a low-rank digital precoder or combiner against it.","Transmitter-side compensation is the preferred design point, because its network inputs are clean 16-QAM symbols, whereas receiver-side compensation must contend with noise entering the compensation network input.","Combining pruning, parameter sharing, and conditional removal of the PA sub-network yields a 97.2% parameter reduction and a 39.2% runtime reduction with nearly unchanged SER, making the scheme plausible for resource-constrained THz transceivers.","Below 5 dBm transmit power the PA nonlinearity is weak enough that the PA sub-network can be removed entirely; above that threshold it must be kept, giving a simple transmit-power-driven complexity control.","In a coded 16-QAM system with a rate-2/3 convolutional code, the slimmed networks still reach very low SER at 15 dBm and outperform the direct fully connected DNN baseline."],"supporting_citations":[{"why":"Supplies the additive quantization noise model (AQNM) used for both DAC and ADC distortion.","marker":"[9]"},{"why":"Provides the transmitter IQ imbalance model with gain and phase error parameters used in the combined impairment chain.","marker":"[12]"},{"why":"Supplies the Gaussian phase-noise model for THz oscillators used to generate training data.","marker":"[14]"},{"why":"Provides the Rapp-model parameters for the 300 GHz nonlinear PA used in simulation.","marker":"[18]"},{"why":"Supplies amplitude and phase error statistics for imperfect phase shifters at THz frequencies.","marker":"[35]"},{"why":"The direct fully connected DNN (D-DNN) baseline that the proposed two-stage method is compared against.","marker":"[30]"},{"why":"A prior DNN method for combined PA nonlinearity and IQ imbalance, motivating the DNN-based treatment of combined impairments.","marker":"[32]"},{"why":"A receiver IQ imbalance pre-compensation scheme that the paper extends by considering transmitter-side distortions and THz frequencies.","marker":"[11]"}],"fun_headline_variants":["Two-stage DNN fixes THz beamforming hardware errors, SER to 0.0005","Neural compensation cuts THz hybrid beamforming SER to 0.0005","Slimmed DNN tames THz hardware imperfections in beamforming","DNN-based two-stage compensation beats THz hardware distortion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulation models used for training faithfully represent real THz hardware imperfections, because the paper generates all training data by simulation rather than from a real hybrid beamforming testbed, and if the actual phase noise, PA behavior, or quantization differs from the Gaussian, Rapp, and AQNM models, the reported SER gains may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage DNN fixes THz beamforming hardware errors, SER to 0.0005","Neural compensation cuts THz hybrid beamforming SER to 0.0005","Slimmed DNN tames THz hardware imperfections in beamforming","DNN-based two-stage compensation beats THz hardware distortion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000338,"raw_usage":{"total_tokens":1941,"prompt_tokens":1090,"completion_tokens":851,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":706,"completion_tokens_details":{"reasoning_tokens":768}},"tokens_in":706,"tokens_out":851,"duration_ms":7590,"temperature":1.0,"reasoning_tokens":768,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:01:16.127447+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two-stage algorithm on a THz hybrid beamforming testbed with real measured impairments (or on a simulator using a different PA model, such as a memory polynomial, and different phase-noise statistics) and compare the SER at 15 dBm with the reported 0.0005 for transmitter compensation; if the SER rises to a clearly higher level, the simulation-to-reality transfer claim fails.","supporting_citations":[{"cited_title":"Low Power Analog-to-digital Conversion in Millimeter Wave Systems: Impact of Resolution and Bandwidth on Performance,","cited_arxiv_id":null,"evidence_quote":"Supplies the additive quantization noise model (AQNM) used for both DAC and ADC distortion."},{"cited_title":"Transmitter IQ Imbalance Pre- Compensation for mm-wave Hybrid Beamforming Systems,","cited_arxiv_id":null,"evidence_quote":"Provides the transmitter IQ imbalance model with gain and phase error parameters used in the combined impairment chain."},{"cited_title":"Phase Noise Robust Terahertz Communications,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian phase-noise model for THz oscillators used to generate training data."},{"cited_title":"Parametrization of simplified memoryless amplifier models at 300 ghz,","cited_arxiv_id":null,"evidence_quote":"Provides the Rapp-model parameters for the 300 GHz nonlinear PA used in simulation."},{"cited_title":"A 220–320-GHz Vector-Sum Phase Shifter Using Single Gilbert-Cell Structure With Lossy Output Matching,","cited_arxiv_id":null,"evidence_quote":"Supplies amplitude and phase error statistics for imperfect phase shifters at THz frequencies."},{"cited_title":"Composite Neural Network Digital Predistortion Model for Joint Mitigation of Crosstalk, I/Q Imbalance, Nonlinearity in MIMO Transmitters,","cited_arxiv_id":null,"evidence_quote":"The direct fully connected DNN (D-DNN) baseline that the proposed two-stage method is compared against."},{"cited_title":"Signal Detection in MIMO Systems With Hardware Imperfections: Message Passing on Neural Networks,","cited_arxiv_id":null,"evidence_quote":"A prior DNN method for combined PA nonlinearity and IQ imbalance, motivating the DNN-based treatment of combined impairments."},{"cited_title":"Downlink Multi- User Hybrid Beamforming Systems With Receiver IQ Imbalance: Pre- Compensation and Performance Analysis,","cited_arxiv_id":null,"evidence_quote":"A receiver IQ imbalance pre-compensation scheme that the paper extends by considering transmitter-side distortions and THz frequencies."}],"review_version":1}