PaperlayerView exact revision on arXiv

Source-linked AI summary

Low-Rank Prior-Guided Rank-One Sensing for Efficient CSI Feedback

Shunpu Tang, Qianqian Yang, Seung-Woo Ko, Jihong Park

arXiv:2609.28156v1eess.SP

TL;DR

Wideband FDD massive MIMO requires CSI feedback, but overhead grows with antenna and subcarrier counts, and existing approaches do not jointly exploit low-rank multipath structure for compression and direct reconstruction. The paper introduces LRP rank-one sensing and synthesis, then extends it with DCRNetV2 residual paths. LRP outperforms iterative recovery at lower complexity, while DCRNetV2 achieves a better accuracy-complexity tradeoff than learning-based methods.

  • Problem

    Wideband FDD massive MIMO has growing CSI feedback overhead, and prior methods rarely use path-wise rank-one structure as a shared prior for compression and direct BS reconstruction.

    Source: §1 · p. 1

  • Method

    LRP learns rank-one sensing at the UE and directly synthesizes CSI from rank-one factors at the BS, while DCRNetV2 adds dilated-convolutional residual paths to the LRP backbone.

    Source: §1 · p. 1, §3.2 · p. 2, §3.3 · p. 2

  • Results

    LRP outperforms iterative recovery with lower complexity, and DCRNetV2 achieves a better accuracy-complexity tradeoff than existing learning-based methods.

    Source: §1 · p. 1, §5.2 · p. 4

  • Takeaways & Limitations

    DCRNetV2’s residual paths capture residual structures and reduce low-rank approximation errors in the reported ablation.

    Source: §5.3 · p. 4

    The low-rank multipath prior is presented under a standard far-field geometric channel model with frequency-invariant array responses.

    Source: §3.1 · p. 2

Abstract

from arXiv · show

Downlink channel state information (CSI) feedback is essential for frequency-division duplex massive MIMO, yet the feedback overhead grows rapidly with the numbers of antennas and subcarriers. To reduce this overhead, most deep learning approaches compress CSI by treating it as a generic image, leaving its low-rank multipath structure unexploited. In contrast, model-driven alternatives explicitly embed this structure in iterative recovery, but at the cost of high computational complexity and latency. To overcome these limitations, we propose a low-rank prior-guided (LRP) framework that performs learnable rank-one sensing at the user equipment to compress CSI into low-dimensional codewords, and reconstructs CSI by direct rank-one synthesis at the base station. Both compression and reconstruction are jointly optimized, and fully exploit the low-rank structure of CSI. We further develop DCRNetV2, which preserves LRP as its backbone and uses gated dilated-convolutional residual paths to compensate for finite-rank errors. Experimental results show that LRP outperforms iterative model-based methods with lower complexity, and DCRNetV2 achieves a better accuracy-complexity tradeoff than existing learning-based methods.

1. INTRODUCTION

Wideband FDD massive MIMO makes CSI feedback costly, while existing compression approaches either underuse CSI’s low-rank multipath structure or retain expensive iterative recovery. The paper proposes shared rank-one sensing and synthesis, reporting lower-complexity gains over iterative methods and a better accuracy-complexity tradeoff for DCRNetV2.

1. INTRODUCTION

  • CSI feedback overhead grows with antenna and subcarrier counts in wideband FDD massive MIMO, despite its importance for beamforming and interference mitigation.The UE must feed CSI back because channel reciprocity does not hold.Source: §1 · p. 1
  • Compressed sensing can degrade when angular-delay sparsity mismatches practical channels, while iterative recovery adds BS computation; learning-based autoencoders treat CSI as a generic image.Model-driven designs embed physical priors but often retain iterative optimization through algorithm unfolding.Source: §1 · p. 1
  • Prior work rarely uses path-wise rank-one structure as a shared prior to guide both CSI compression and direct BS reconstruction.Each wideband propagation path is described as a rank-one outer product, and a limited number of dominant paths makes the channel matrix naturally low-rank.Source: §1 · p. 1
  • LRP learns UE-side delay- and angular-domain sensing vectors and directly synthesizes CSI from rank-one factors at the BS; DCRNetV2 adds gated dilated-convolutional residual paths.On COST 2100 channels, LRP outperforms iterative recovery with lower complexity, while DCRNetV2 achieves a better accuracy-complexity tradeoff than existing learning-based methods.Source: §1 · p. 1

2. SYSTEM MODEL

The system is an FDD massive MIMO-OFDM downlink whose spatial-frequency CSI is transformed to an angular-delay representation. An encoder maps that representation to a feedback codeword, and a decoder reconstructs it under end-to-end MSE training.

2. SYSTEM MODEL

  • The system has an Nt-antenna BS, a single-antenna UE, and Nc subcarriers, with spatial-frequency CSI represented by H.DFTs and a truncation matrix retain the first Na delay bins to form the angular-delay CSI Ha.Source: §2 · p. 1
  • The encoder maps Ha to an M-dimensional codeword z, which the BS decoder uses to reconstruct Ha; the compression ratio is η = M/(2NaNt).Encoder and decoder parameters are trained end-to-end to minimize mean squared error.Source: §2 · p. 2

3. PROPOSED LRP

LRP exploits the low-rank multipath structure of angular-delay CSI through learnable rank-one sensing at the UE and rank-one synthesis at the BS. The sensing and synthesis components are jointly optimized, with DCR modules refining residual structure.

3.1. Low-rank multipath prior

  • Under a far-field geometric channel model with frequency-invariant array responses, each path preserves a rank-one outer-product structure, so Ha has rank at most L and can be approximated by R rank-one components.R controls approximation capacity and need not equal the physical path count L.Source: §3.1 · p. 2

3.2. Rank-one sensing encoder

  • The encoder uses learnable rank-one sensing branches whose delay and angular vectors probe the channel, then projects the stacked measurements’ real and imaginary parts to the feedback dimension.A branch measurement acts like an attention-like similarity response, becoming strong when a path matches both queries.Source: §3.2 · p. 2

3.3. Low-rank multipath reconstruction

  • The decoder maps the codeword to delay- and angular-domain factors and synthesizes CSI from rank-one branches, capturing dominant low-rank structure without one-to-one path correspondence.The complex gains are absorbed into the factors.Source: §3.3 · p. 2

3.4. End-to-end training

  • Sensing and synthesis are jointly trained by minimizing expected reconstruction risk, while DCRNetV2 keeps LRP as its main low-rank path and uses DCR modules to encode and refine residual structure.Mini-batch stochastic gradient descent optimizes the objective coupling sensing and synthesis.Source: §3.4 · p. 2, Figure 1 · p. 3

4. PROPOSED DCRNETV2

DCRNetV2 retains LRP as its main path and adds dilated-convolutional residual paths. Its encoder fuses residual features with LRP measurements, while its decoder synthesizes additional rank-one components and applies a final residual correction.

4. PROPOSED DCRNETV2

  • DCRNetV2 keeps LRP as its main path and adds dilated-convolution blocks to capture residual channel structure.The residual paths are designed to address structure beyond the dominant low-rank component.Source: §4 · p. 3

4.1. Residual encoder and fusion

  • The UE encoder processes angular-delay CSI in parallel: LRP produces global rank-one measurements, while DCR extracts local residual features using asymmetric, dilated convolutions.A learned fusion MLP combines the branches while additive paths preserve their direct information; initialization keeps the initial codeword close to LRP.Source: §4.1 · p. 3

4.2. Hybrid decoder and refinement

  • At the BS, LRP first forms a direct estimate; a residual decoder maps the codeword into additional rank-one components, whose contributions are gated and merged with the direct path.Increasing Rdec improves performance by capturing weak multipath components and residual energy, but directly scaling Rdec increases projection parameter overhead.Source: §4.2 · p. 3
  • A final DCR module predicts a residual correction to refine the hybrid estimate, controlled by a learnable scalar gate.The passage notes that this refinement step is also used in existing CSI feedback networks to capture structure beyond the low-rank approximation.Source: §4.2 · p. 3

4.3. Configurations

  • Mini and Unified use factored residual decoders, while Small and Base omit them; the four variants trade architecture size through sensing branches, direct synthesis branches, residual components, and bottleneck width.Mini uses (128, 10, 8, 8), Unified (512, 32, 16, 16), Small (256, 16, 0, 0), and Base (512, 16, 0, 0).Source: §4.3 · p. 3

5. EXPERIMENTS

Experiments on COST 2100 compare LRP and DCRNetV2 with classical, model-based, and learning-based methods. LRP outperforms iterative methods at lower complexity, while DCRNetV2 offers a favorable accuracy-complexity trade-off and benefits from both residual paths.

5.1. Experimental settings

  • Evaluations use COST 2100 indoor (5.3 GHz) and outdoor (0.3 GHz) scenarios, with NMSE and FLOPs reported for networks trained under common settings.The channel dimensions are (Nt, Nc, Na) = (32, 1024, 32); training uses 1500 epochs, Adam, MSE loss, batch size 200, and initial learning rate 2 × 10^-3.Source: §5.1 · pp. 3–4

5.2. Performance comparison

  • LRP outperforms iterative methods at lower complexity; DCRNetV2-mini has the lowest FLOPs across compression ratios, while unified achieves the best overall performance at most ratios and a superior accuracy-complexity trade-off.The comparisons include FLOPs versus indoor NMSE at η = 1/4 and η = 1/16.Source: §5.2 · p. 4, Figure 2 · p. 4, Table 2 · p. 4

5.3. Ablation studies

  • At indoor η = 1/4, NMSE improves from −21.62 dB to −29.34 dB with only the decoder path, −32.00 dB with only the encoder path, and −32.84 dB with both.The ablation evaluates architectural components of DCRNetV2-unified.Source: §5.3 · p. 4, Table 3 · p. 4

6. CONCLUSION

The conclusion frames CSI feedback as low-rank matrix sensing and reconstruction. It presents LRP as a low-complexity approach, extended by DCRNetV2 with dilated-convolutional residual paths, and reports experiments on COST 2100 verifying the methods' effectiveness.

  • The paper reframes CSI feedback as a low-rank matrix sensing and reconstruction problem.Source: §6 · p. 4
  • LRP achieves strong performance with low complexity.Source: §6 · p. 4
  • DCRNetV2 extends LRP by combining it with dilated-convolutional residual paths, and COST 2100 experiments verify the effectiveness of the proposed methods.Source: §6 · p. 4
Loading 2609.28156v1…