DDA-Net: Accurate TDD Channel Estimation via Deep Unfolding the Doppler-Delay-Angle Representation of Channel Signals

arXiv:2604.05389 · eess.SP · Submitted 2026-04-07 · Read on arXiv

Listen

Radio episode about this paper

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: "DDA-Net: Accurate TDD Channel Estimation via Deep Unfolding the Doppler-Delay-Angle Representation of Channel Signals".

Dev: Here's the gist of the paper. In TDD massive MIMO systems, channel estimation under sparse frequency-hopping pilots is challenging: each snapshot captures only one narrow pilot block that hops across frequency, with tens of milliseconds between adjacent snapshots The authors propose DDA-Net, a model-driven 3D deep unfolding network for joint multi-snapshot channel state reconstruction Problem Context Time-division duplex (TDD) massive multiple-input multiple-output (MIMO) systems rely on accurate uplink channel state information (CSI) for coherent combining, spatial multiplexing, and reciprocity-based precoding CSI acquisition becomes substantially harder when the channel evolves between pilot-bearing observations, since channel aging reduces the usefulness of stale CSI and calls for estimators that exploit temporal structure rather than treating each observation as static The setting involves a sounding reference signal (SRS) transmission occupying one OFDM symbol and covering one narrow contiguous pilot block, with adjacent observations separated by ∆t = 40 ms This deliberate sparse setting is motivated by capacity-oriented sounding, aiming to accommodate more UEs within a fixed sounding-resource budget at the cost of a harder full-band reconstruction problem DDA Representation and Unfolding The paper introduces the Doppler-delay-angle (DDA) representation as a structured transform-domain model to address limitations of classical sparse recovery methods The wideband massive MIMO channel at discrete snapshot q can be expressed as a superposition of L effective propagation paths, involving complex coefficient, delay, and pair of normalized receive spatial frequencies To obtain the DDA representation, a receiverside space-angle dictionary Fsa is introduced, which is square and unitary when ksa = 1 Along the frequency axis, an oversampled delay dictionary Ffd is used to represent fractional-delay paths The corresponding time-delay-angle representation is denoted by X in the DDA domain The temporal structure is further represented by applying a unitary Tw-point DFT along the discrete snapshot axis to obtain Xe DDA-Net Architecture DDA-Net truncates and unfolds the three ADMM updates into a finite-depth network with closed-form data consistency The architecture comprises one initialization stage, Niter intermediate stages, and one terminal unfolded stage followed by a DC-only output layer The central model-based component is the DC update, which admits a closed-form solution due to the row orthonormality of Aq The final time-delay-angle output Xˆ is mapped back to the TFS domain to obtain Hˆ Key Design Contributions The main contributions include: 1.

Rosa: DDA-Net is the first deep learning-based method for TDD massive MIMO channel estimation to use a full-window 3D DDA representation as its internal reconstruction state 2.

Dev: The authors combine a lightweight residual prior with a delayoversampled DDA representation to model local structure across angle, delay, and Doppler and to mitigate finitewindow and off-grid leakage 3.

Rosa: They show that coverage-aware hopping improves DDA-Net and the full-window classical solvers at the same pilot overhead Experimental Results DDA-Net consistently outperforms strong baselines across three channel settings from 3GPP TR 38.901: UMa-NLOS, UMi-NLOS, and CDL-B The Doppler-domain design proves consistently superior to its time-domain equivalent, with a wider margin after finetuning The MCD pilot substantially improves DDA-Net and the full-window classical solvers over the Standard hopping pilot at medium-to-high SNRs Ablation Insights Ablations confirm that window-level 3D processing and explicit Doppler modeling yield target-dependent improvements The gains from explicit Doppler-domain parameterization are 3.06, 1.12, and 0.19 dB before adaptation Delay oversampling improves all three domains, but the smaller and similar OOD gains suggest that it addresses a shared fractional-delay limitation The results show that source-domain pretraining substantially reduces the amount of target-domain data required for adaptation DDA-Net combines analytic data consistency with learned DDA-domain prior modeling The authors demonstrate that combining exact physical data consistency with a learned DDA-domain prior is an effective and sample-efficient approach to channel state acquisition under sparse frequency-hopping pilots The results support model-driven DDA reconstruction as an effective and sample-efficient framework for channel acquisition under sparse frequency-hopping pilots Reference List

1: T.

Title and authors: Rosa: So this paper is talking about this DDA-Net thing for channel estimation in TDD massive MIMO when you're using sparse frequency hopping pilots, right? It suggests they can build a whole three-dimensional deep unfolding network to reconstruct the channel across multiple snapshots at once.

Dev: Right. Basically, the challenge with these sparse pilots is that each snapshot only gets one narrow block of frequency data and then there’s this gap of tens of milliseconds before the next one arrives, which makes it really hard to get good channel information because things change between snapshots.

Taro: So what does this DDA representation actually mean? Is it a better way to look at that sparse channel data?

Rosa: It introduces the Doppler-delay-angle, or DDA, representation as a structured model. Think of it as breaking down the wideband massive MIMO channel into these effective propagation paths, which involve complex coefficients and spatial frequencies. They even introduce dictionaries for the delay and frequency parts to represent those fractional delays.

Dev: And then they apply a unitary two point discrete Fourier transform along the snapshot axis to get that time-delay-angle representation, which is what they call X in the DDA domain. That sounds like a pretty structured way to handle the temporal evolution of the channel.

Taro: But how does this structure help when we only get one narrow pilot block at a time? Does it solve that problem directly?

Rosa: It helps by allowing them to use this structured model within their deep unfolding network. They use a model-driven approach where they truncate and unfold three alternating direction method updates into a finite depth network, which gives them that closed-form data consistency update that avoids tensor inversion.

Dev: That sounds like they are enforcing the physics of the problem directly into the learning process instead of just letting the network guess everything from scratch. And they mention using a lightweight Doppler-domain learned prior to model local structure across angle, delay, and Doppler.

Taro: So this prior is helping them deal with those issues like finite window leakage and off-grid effects that they mentioned?

Rosa: Exactly. They use this combination of a residual prior and the delay oversampled DDA representation to model the local structure across angle, delay, and Doppler. They also show that coverage-aware hopping designs, specifically using minimum coverage distance pilots, improves both their network and the classical full-window solvers at the same pilot overhead.

Dev: That's interesting because it means you can get better performance without increasing how much you talk over the air with pilots. And they showed that explicitly modeling Doppler domain parameters gives gains of three point zero six, one point one two, and zero point one nine decibels before adaptation when compared to standard hopping pilots across three channel settings from 3GPP TR thirty-eight point nine zero one: UMa-NLOS, UMi-NLOS, and CDL-B <ref:2604.05389#pg1,across three channel settings from 3GPP TR 38.901: UMa-NLOS, UMi>.

Taro: So the Doppler domain modeling is actually giving tangible improvements in terms of signal quality? That’s a solid number to have when you’re dealing with real constraints.

Rosa: It is, but they also noted that delay oversampling helps all three domains, though the gains from explicitly parameterizing Doppler domain are bigger. They found that source-domain pretraining substantially cuts down on the amount of target domain data needed for adaptation, which is a big deal for sample efficiency.

Dev: So we’re talking about a model where you can combine exact physical consistency with a learned DDA-domain prior to get this sample efficient channel acquisition under sparse frequency hopping pilots. What's the actual implication here for deploying these systems?

Taro: It means that instead of needing huge amounts of data to train an estimator, we can pretrain on one system and adapt it quickly to another system with just a little bit of new data. That sounds way more practical for real-world deployment scenarios.

Rosa: And because they use this DDA-Net approach, which combines analytic data consistency with learned prior modeling, the overall framework seems to be very sample efficient for acquiring the channel under these tricky sparse pilots.

Dev: The paper concludes that model-driven DDA reconstruction is an effective and sample efficient way to get channel state information when you are dealing with sparse frequency hopping pilots. It’s a framework built on combining exact physical data consistency with a learned DDA-domain prior, which they show works well across the UMa-NLOS, UMi-NLOS, and CDL-B settings.

Taro: I just want to say that their finding that full channel reconstruction can happen at sampled sounding times rather than continuous path Dopplers is physically meaningful because alias equivalent Dopplers generate the same sampled channel and therefore correspond to the same reconstruction target. That’s a nice physical grounding for this approach.

Rosa: Yeah, it gives it real meaning in terms of how we model what we actually observe, not just theoretical math. So that's what they did with DDA-Net, combining the structure of DDA with deep learning to handle the sparsity and temporal gaps in these massive MIMO systems.

The paper's summary: Rosa: So we've seen how they built this DDA-Net thing to reconstruct the channel across multiple time snapshots using this three dimensional representation, and now we're going to talk about what that whole approach actually means for us in terms of getting reliable channel information in massive MIMO.

Dev: Right. The core idea here is that they take those sparse measurements—just one narrow pilot block at a time, with gaps between them—and instead of treating each measurement as completely independent static data, they use this structured DDA domain to model how the channel evolves across angles, delays, and Doppler shifts simultaneously.

Rosa: Exactly. They essentially turn a messy problem of tracking a time-varying signal into a more organized mathematical structure where they can apply deep learning techniques to solve it, which is what the deep unfolding network does. It's like they're not just guessing the channel at each moment, but they’re building a model that understands the physical rules governing how that channel changes over time and space.

Dev: And what's really interesting from their methodology is how they use this representation to enforce physical constraints during the learning process. They use something called closed form data consistency updates in the time delay angle domain, which means instead of letting the network learn everything implicitly through just trial and error, they’re using an update that already incorporates exactly how a real pilot is sampled, solving a quadratic penalty problem.

Rosa: That’s smart because it keeps the model grounded in reality. It stops the network from learning nonsense and forces it to respect the known physics of how those pilots are actually being measured, even when the measurements are sparse and infrequent.

Dev: And they combine that with a lightweight prior that learns local structure across angle, delay, and Doppler—this helps them deal with those tricky issues like getting stuck in a finite window or having errors from off-grid effects where the signal doesn't perfectly match their model.

Rosa: So the result is this model-driven reconstruction framework. It’s not just another black box deep learning system; it has this analytic consistency baked in, and then it learns the specific nuances of how different massive MIMO systems behave through that learned prior.

Dev: And they showed that you can use this approach to get better results without needing a ton of new target domain data. Source domain pretraining allows them to adapt quickly when moving from one massive MIMO setup to another, which is pretty big for real-world deployment where you can't always afford massive retraining datasets.

Rosa: It suggests that the future of channel estimation in these complex TDD systems isn't just about throwing bigger neural networks at it, but using structured transforms like DDA to give those networks a better starting point and a clearer physical understanding of what they are reconstructing.

Dev: It means for operators, we might see systems that can maintain high quality channel state information even when pilots are very sparse and frequency hopping is dense because the model has already captured the underlying physics.

Rosa: So if you think about this, it means that next time you’re designing a massive MIMO system for a cell with very specific pilot constraints, this DDA-Net idea suggests we can get much more accurate channel estimation with far less overhead and much faster adaptation.

The paper's improvements: Taro: So we've seen how they built this DDA-Net thing to reconstruct the channel across multiple time snapshots using this three dimensional representation, and now we're going to talk about what that whole approach actually means for us in terms of getting reliable channel information in massive MIMO.

Rosa: Right. The core idea here is that they take those sparse measurements—just one narrow pilot block at a time, with gaps between them—and instead of treating each measurement as completely independent static data, they use this structured DDA domain to model how the channel evolves across angles, delays, and Doppler shifts simultaneously.

Dev: Exactly. They essentially turn a messy problem of tracking a time-varying signal into a more organized mathematical structure where they can apply deep learning techniques to solve it, which is what the deep unfolding network does. It's like they're not just guessing the channel at each moment, but they’re building a model that understands the physical rules governing how that channel changes over time and space.

Rosa: And what's really interesting from their methodology is how they use this representation to enforce physical constraints during the learning process. They use something called closed form data consistency updates in the time delay angle domain, which means instead of letting the network learn everything implicitly through just trial and error, they’re using an update that already incorporates exactly how a real pilot is sampled, solving a quadratic penalty problem.

Dev: That’s smart because it keeps the model grounded in reality. It stops the network from learning nonsense and forces it to respect the known physics of how those pilots are actually being measured, even when the measurements are sparse and infrequent.

Rosa: And they combine that with a lightweight prior that learns local structure across angle, delay, and Doppler—this helps them deal with those tricky issues like getting stuck in a finite window or having errors from off-grid effects where the signal doesn't perfectly match their model.

Dev: So the result is this model-driven reconstruction framework. It’s not just another black box deep learning system; it has this analytic consistency baked in, and then it learns the specific nuances of how different massive MIMO systems behave through that learned prior.

Rosa: And they showed that you can use this approach to get better results without needing a ton of new target domain data. Source domain pretraining allows them to adapt quickly when moving from one massive MIMO setup to another, which is pretty big for real-world deployment where you can't always afford massive retraining datasets.

Dev: It means for operators, we might see systems that can maintain high quality channel state information even when pilots are very sparse and frequency hopping is dense because the model has already captured the underlying physics.

Rosa: So if you think about this, it means that next time you’re designing a massive MIMO system for a cell with very specific pilot constraints, this DDA-Net idea suggests we can get much more accurate channel estimation with far less overhead and much faster adaptation.

Conclusion: Rosa: So to wrap things up, this paper on DDA-Net shows how combining exact physical consistency with learned structural knowledge in the Doppler-Delay-Angle representation can make channel estimation under sparse frequency hopping pilots much more sample efficient and accurate.

Dev: It boils down to using that three dimensional model not just as a fancy way to look at the data, but as a guide for the AI to follow during training so it respects physical constraints while still learning what it needs from the measurements.

Taro: I'm curious about how robust this whole thing is when things get messy outside of the perfect simulation environment. Does this DDA-Net setup handle real-world noise or unexpected signal behavior gracefully?

Rosa: That’s a fair question, Taro. The authors show it works well across three different channel environments from 3GPP TR thirty-eight point nine zero one, like UMa NLOS and CDL B, which means the structure they learned isn't just specific to one test setup <ref:2604.05389#pg1>.

Dev: And we saw the experimental results consistently outperform stronger baselines across those settings, even when comparing it to classical full-window solvers at the same pilot overhead. That suggests a real performance gain in terms of how much channel information you can get for your communication budget.

Taro: So it’s not just theoretical accuracy; it’s showing better recovery on actual complex channels while keeping the pilot usage low, which is what we need for practical massive MIMO deployment.

Rosa: Exactly. They demonstrated that by using coverage-aware hopping pilots and incorporating explicit Doppler modeling, the DDA-Net framework actually improves things even more than just the raw deep learning approach alone.

Dev: The safety case here seems solid because they are using an iterative unfolding process with known constraints, which gives us a good idea of where the latency might be in a real time loop. We're talking about a closed form data consistency update that runs predictably.

Taro: It sounds like it’s handling those complex temporal dependencies very well, allowing for that full channel reconstruction at sampled times instead of having to track continuous path Dopplers constantly.

Rosa: Right, so the big idea is that this DDA-Net framework provides a powerful way to get high-quality channel state information efficiently by anchoring the deep learning in known physical models.

Dev: It’s a solid piece of work on structured compressive sensing for TDD massive MIMO systems with sparse pilots.

Taro: I think we need to keep watching how they tackle the latency in that iterative unfolding process, because if it stays low, this could be a serious tool for real-time channel tracking in 5G networks <ref:2604.05389#pg2>.

Peking University

eess.SP

Submitted: 2026-04-07

Updated: 2026-11-05

Comments: 16 pages, 9 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 99/100

The gist: In TDD massive MIMO systems, channel estimation under sparse frequency-hopping pilots is challenging: each snapshot captures only one narrow pilot block that hops across frequency…

Key concepts

Doppler-Delay-Angle (DDA) Representation
This is a structured mathematical model used to represent the wideband massive MIMO channel. It decomposes the channel into components related to complex coefficients, time delays, and normalized spatial frequencies (angles). This representation allows the network to capture the physical structure of how signals propagate across frequency and time.
Deep Unfolding Network
This is a deep learning architecture where a complex mathematical optimization problem is 'unfolded' into a sequence of trainable layers. DDA-Net uses this to solve the channel estimation problem by iteratively applying updates derived from the underlying physical model, ensuring data consistency and leveraging analytic solutions.
Data Consistency Update
This refers to an update step within the network that is mathematically guaranteed to maintain consistency with the physical model constraints. In DDA-Net, this update has a closed-form solution because of row orthonormality, meaning it exactly incorporates the pilot sampling operator without relying solely on implicit learning.
Coverage-aware Hopping
This technique involves designing the frequency hopping pattern (pilot placement) to ensure better coverage. By using minimum-coverage-distance (MCD) pilots, the system achieves a more balanced distribution of frequency coverage within each reconstruction window, leading to improved channel estimation accuracy.

Terminology

Summary

In TDD massive MIMO systems, channel estimation under sparse frequency-hopping pilots is challenging: each snapshot captures only one narrow pilot block that hops across frequency, with tens of milliseconds between adjacent snapshots The authors propose DDA-Net, a model-driven 3D deep unfolding network for joint multi-snapshot channel state reconstruction

Problem Context

Time-division duplex (TDD) massive multiple-input multiple-output (MIMO) systems rely on accurate uplink channel state information (CSI) for coherent combining, spatial multiplexing, and reciprocity-based precoding CSI acquisition becomes substantially harder when the channel evolves between pilot-bearing observations, since channel aging reduces the usefulness of stale CSI and calls for estimators that exploit temporal structure rather than treating each observation as static The setting involves a sounding reference signal (SRS) transmission occupying one OFDM symbol and covering one narrow contiguous pilot block, with adjacent observations separated by ∆t = 40 ms This deliberate sparse setting is motivated by capacity-oriented sounding, aiming to accommodate more UEs within a fixed sounding-resource budget at the cost of a harder full-band reconstruction problem

DDA Representation and Unfolding

The paper introduces the Doppler-delay-angle (DDA) representation as a structured transform-domain model to address limitations of classical sparse recovery methods The wideband massive MIMO channel at discrete snapshot q can be expressed as a superposition of L effective propagation paths, involving complex coefficient, delay, and pair of normalized receive spatial frequencies To obtain the DDA representation, a receiverside space-angle dictionary Fsa is introduced, which is square and unitary when ksa = 1 Along the frequency axis, an oversampled delay dictionary Ffd is used to represent fractional-delay paths The corresponding time-delay-angle representation is denoted by X in the DDA domain The temporal structure is further represented by applying a unitary Tw-point DFT along the discrete snapshot axis to obtain Xe

DDA-Net Architecture

DDA-Net truncates and unfolds the three ADMM updates into a finite-depth network with closed-form data consistency The architecture comprises one initialization stage, Niter intermediate stages, and one terminal unfolded stage followed by a DC-only output layer The central model-based component is the DC update, which admits a closed-form solution due to the row orthonormality of Aq The final time-delay-angle output Xˆ is mapped back to the TFS domain to obtain Hˆ

Key Design Contributions

The main contributions include:

  1. DDA-Net is the first deep learning-based method for TDD massive MIMO channel estimation to use a full-window 3D DDA representation as its internal reconstruction state

  2. The authors combine a lightweight residual prior with a delayoversampled DDA representation to model local structure across angle, delay, and Doppler and to mitigate finitewindow and off-grid leakage

  3. They show that coverage-aware hopping improves DDA-Net and the full-window classical solvers at the same pilot overhead

Experimental Results

DDA-Net consistently outperforms strong baselines across three channel settings from 3GPP TR 38.901: UMa-NLOS, UMi-NLOS, and CDL-B The Doppler-domain design proves consistently superior to its time-domain equivalent, with a wider margin after finetuning The MCD pilot substantially improves DDA-Net and the full-window classical solvers over the Standard hopping pilot at medium-to-high SNRs

Ablation Insights

Ablations confirm that window-level 3D processing and explicit Doppler modeling yield target-dependent improvements The gains from explicit Doppler-domain parameterization are 3.06, 1.12, and 0.19 dB before adaptation Delay oversampling improves all three domains, but the smaller and similar OOD gains suggest that it addresses a shared fractional-delay limitation The results show that source-domain pretraining substantially reduces the amount of target-domain data required for adaptation

Conclusion

DDA-Net combines analytic data consistency with learned DDA-domain prior modeling The authors demonstrate that combining exact physical data consistency with a learned DDA-domain prior is an effective and sample-efficient approach to channel state acquisition under sparse frequency-hopping pilots The results support model-driven DDA reconstruction as an effective and sample-efficient framework for channel acquisition under sparse frequency-hopping pilots

Reference List

[1] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,” IEEE Transactions on Wireless Communications, vol. p. 0 9, no. 11, pp. 3590–3600, 2010

[2] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive MIMO for next generation wireless systems,” IEEE Communications Magazine, vol. p. 0 52, no. 2, pp. 186–195, 2014

[3] K. T. Truong and J. Robert W. Heath, “Effects of channel aging in massive MIMO systems,” Journal of Communications and Networks, vol. p. 0 15, no. 4, pp. 338–351, 2013

[4] 3rd Generation Partnership Project (3GPP), “NR; radio resource control (rrc); protocol specification,” ETSI, Technical Specification TS 38.331, Version 18. p. 06.0, Release 18, 2025

[5] ——, “NR; physical channels and modulation,” ETSI, Technical Specification TS 38.211, Version 18. p. 0 p. 05.0, Release 18, 2025

[6] Z. Gao, L. Dai, W. Dai, B. Shim, and Z. Wang, “Structured compressive sensing-based spatio-temporal joint channel estimation for FDD massive MIMO,” IEEE Transactions on Communications, vol. p. 0 64, no. 2, pp. 601–617, 2016

[7] A. Liu, V. K. N. Lau, and W. Dai, “Exploiting burst-sparsity in massive MIMO with partial channel support information,” IEEE Transactions on Wireless Communications, vol. p. 0 15, no. 11, pp. 7820–7830, 2016

[8] W. Shen, L. Dai, J. An, P. Fan, and J. Robert W. Heath, “Channel estimation for orthogonal time frequency space (OTFS) massive MIMO,” IEEE Transactions on Signal Processing, vol. p. 0 67, no. 16, pp. 4204–4217, 2019

[9] H. Yin, H. Wang, Y. Liu, and D. p. 1 Gesbert, “Addressing the curse of mobility in massive MIMO with Prony-based angular-delay domain channel predictions,” IEEE Journal on Selected Areas in Communications, vol. p. 0 38, no. 12, pp. 2903–2917

[10] Y. Chi, L. L. Scharf, A. Pezeshki, and A. R. Calderbank, “Sensitivity to basis mismatch in compressed sensing,” IEEE Transactions on Signal Processing, vol. p. 3 59, no. 5, pp. 2182–2195

[11] Z. Wei, W. Yuan, S. Li, J. Yuan, and D. W. K. Ng, “Off-grid channel estimation with sparse Bayesian learning for OTFS systems,” IEEE Transactions on Wireless Communications, vol. p. 0 21, no. 9, pp. 7407–7426

[12] Y. Zhu, J. Zhuang, G. Sun, H. Hou, L. You, and W. Wang, “Joint channel estimation and prediction for massive MIMO with frequency hopping sounding,” IEEE Transactions on Communications, vol. p. 0 73, no. 7, pp. 5139–5154

[13] G. Tang, B. N. Bhaskar, P. Shah, and B. Recht, “Compressed sensing off the grid,” IEEE Transactions on Information Theory, vol. p. 0 59, no. 11, pp. 7465–7490

[14] Z. Yang, L. Xie, and P. Stoica, “Vandermonde decomposition of multilevel toeplitz matrices with application to multidimensional superresolution,” IEEE Transactions on Information Theory, vol. p. 0 62, no. 6, pp. 3685–3701

[15] Y. Wan and A. Liu, “A two-stage 2d channel extrapolation scheme for TDD 5g NR systems,” IEEE Transactions on Wireless Communications, vol. p. 0 p. 0 23, no. 8, pp. 8497–8511

[16] Y. Wan, A. Liu, and T. p. 1 p. 1 Q. S. Quek, “Multi-user pilot pattern optimization for channel extrapolation in 5g NR systems,” IEEE Transactions on Wireless Communications, vol. p. 0 p. 0 24, no. 7, pp. 6166–6179

[17] H. Ye, G. Y. Li, and B.-H. F. Juang, “Power of deep learning for channel estimation and signal detection in OFDM systems,” IEEE Wireless Communications Letters, vol. p. 0 7, no. 1, pp. 114–117

[18] M. Soltani, V. Pourahmadi, A. Mirzaei, and H. Sheikhzadeh, “Deep learning-based channel estimation,” IEEE Communications Letters, vol. p. 0 23, no. 4, pp. 652–655

[19] D. Luan and J. Thompson, “Channelformer: Attention based neural solution for wireless channel estimation and effective online training,” IEEE Transactions on Wireless Communications, vol. p. 0 22, no. 10, pp. 6562–6577

[20] C. Liu, W. Jiang, and X. Yuan, “Learning-based block-wise planar channel estimation for time-varying MIMO OFDM,” IEEE Wireless Communications Letters, vol. p. 0 13, no. 8, pp. 2125–2129

[21] C. J. Chun, J. M. Kang, and I. M. Kim, “Deep learning-based channel estimation for massive MIMO systems,” IEEE Wireless Communications Letters, vol. p. 0 8, no. 4, pp. 1228–1231

[22] H. He, S. Jin, C.-K. Wen, F. Gao, G. Y. Li, and Z. Xu, “Modeldriven deep learning for physical layer communications,” IEEE Wireless Communications, vol. p. 0 26, no. 5, pp. 77–83

[23] A. Balatsoukas-Stimming and C. Studer, “Deep unfolding for communications systems: A survey and some new directions,” in 2019 IEEE International Workshop on Signal Processing Systems (SiPS), 2019, pp. p. 0 266–271

[24] M. Borgerding, P. Schniter, and S. Rangan, “AMP-inspired deep networks for sparse linear inverse problems,” IEEE Transactions on Signal Processing, vol. p. 0 65, no. 16, pp. 4293–4308

[25] C. A. Metzler, A. Mousavi, and R. G. Baraniuk, “Learned D-AMP: Principled neural network based compressive image recovery,” in Advances in Neural Information Processing Systems 30, 2017

[26] Y. Yang, J. Sun, H. Li, and Z. Xu, “ADMM-CSNet: A deep learning approach for image compressive sensing,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 42, no. 3, pp. 521–538

[27] H. He, C.-K. Wen, S. Jin, and G. Y. Li, “Deep learning-based channel estimation for beamspace mmwave massive MIMO systems,” IEEE Wireless Communications Letters, vol. p. 0 7, no. 5, pp. 852–855

[28] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed optimization and statistical learning via the alternating direction method of multipliers,” Foundations and Trends in Machine Learning, vol. p. 0 3, no. 1, pp. 1–122

[29] X. Zhu and T. Li, “ST-DDA: Dynamic channel estimation in the doppler-delay-angle domain via sparse subspace tracking for TDD systems,” 2026

[30] S. Jaeckel, L. Raschkowski, K. Borner, and L. Thiele, “QuaDRiGa: A ¨ 3-D multi-cell channel model with time evolution for enabling virtual field trials,” IEEE Transactions on Antennas and Propagation, vol. p. 0 p. 0 p. 0 62, no. 6, pp. 3242–3256

[31] 3rd Generation Partnership Project (3GPP), “Study on channel model for frequencies from 0.5 to 100 GHz,” ETSI, Technical Report TR 38. p. 6901, Version 18. p. 00.0, Release 18, 2024

[32] X. Zhu, Y. Zeng, and T. Li, “Coverage-distance and collinearityminimizing pilots for channel estimation in TDD systems,” IEEE Wireless Communications Letters, vol. p. 0 15, pp. 3741–3745

[33] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm for linear inverse problems,” SIAM Journal on Imaging Sciences, vol. p. 7 2, no. 1, pp. 183–202, 2009

[34] Yufei Ma received the BS degree in mathematics from Peking University in 2022 and is currently pursuing the Ph.D. degree in computational mathematics at Peking University, with research interests in machine learning and wireless communications

[35] Xu Zhu received the BS degree in mathematics in 2021 from Peking University, where he is currently pursuing the Ph.D. degree in mathematics with an emphasis on optimization, signal processing and wireless communication

[36] Tiejun Li received the BS and MS degrees in mathematics from Tsinghua University, Beijing, in 1995 and 1998, respectively, and the PhD degree in mathematics from Peking University, in 2001

[37] S. Jaeckel, L. Raschkowski, K. Borner, and L. Thiele, “QuaDRiGa: A ¨ 3-D multi-cell channel model with time evolution for enabling virtual field trials,” IEEE Transactions on Antennas and Propagation, vol. 62, no. 6, pp. 3242–3256

[38] 3rd Generation Partnership Project (3GPP), “NR; radio resource control (rrc); protocol specification,” ETSI, Technical Specification TS 38.311, Version 18. p. 06.0, Release 18, 2025

[39] ——, “NR; physical channels and modulation,” ETSI, Technical Specification TS 38.211, Version 18.5.0, Release 18, 2025

[40] S. Jaeckel, L. Raschkowski, K. Borner, and L. Thiele, “QuaDRiGa: A ¨ 3-D multi-cell channel model with time evolution for enabling virtual field trials,” IEEE Transactions on Antennas and Propagation, vol. 62, no. 6, pp. 3242–3256

[41] Y. Wan and A. Liu, “A two-stage 2d channel extrapolation scheme for TDD 5g NR systems,” IEEE Transactions on Wireless Communications, vol. 23, no. 8, pp. 8497–8511

[42] Y. Wan, A. Liu, and T. Q. S. Quek, “Multi-user pilot pattern optimization for channel extrapolation in 5g NR systems,” IEEE Transactions on Wireless Communications, vol. 24, no.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Window-level 3D processing enables joint multi-snapshot channel state reconstruction by exploiting common sparse supports across successive OFDM symbols, which is more effective than purely data-driven approaches lacking an explicit structured transform-domain model.

  2. The DDA representation, which integrates a lightweight Doppler-domain learned prior, allows the network to model local angledelay-Doppler structure and mitigate finite window leakage and off-grid effects.

  3. The DDA-Net architecture utilizes a closed-form data consistency update in the time delay angle domain that exactly incorporates the pilot sampling operator of each snapshot by solving a quadratic penalty problem, enforcing physical model constraints without implicit learning of the measurement process.

  4. Delay oversampling refines the representation by increasing dictionary size to N tau = kfd Nf, which is motivated because continuous path delays determine the phase evolution across subcarriers, thereby improving recovery of unobserved frequency responses and preserving the closed-form structure of the unfolded DC step.

  5. Coverage-aware pilot design, specifically the minimum-coverage-distance (MCD) pilot, substantially improves DDA-Net and full-window classical solvers by providing more balanced frequency coverage within each reconstruction window.

  6. Few-shot target domain adaptation allows pretraining on a source domain model (like UMa) to adapt to new target domains (UMi or CDL) with minimal data, demonstrating that source-domain pretraining substantially reduces the amount of target-domain data required for adaptation.

  7. The system can perform full channel reconstruction at sampled sounding times rather than continuous path Dopplers, which is physically meaningful because alias equivalent Dopplers generate the same sampled channel and therefore correspond to the same reconstruction target.

Sources

Related papers