PaperlayerView exact revision on arXiv

Source-linked AI summary

Successive Bayesian Reconstructor for Channel Estimation in Fluid Antenna Systems

Zijian Zhang, Jieao Zhu, Linglong Dai, Robert W. Heath Jr

arXiv:2312.06551v5cs.ITcs.SYeess.SPeess.SYmath.IT

TL;DR

FAS channel estimation is difficult because dense port spaces and limited RF connections make CSI acquisition challenging, while existing estimators rely on channel assumptions. S-BAR uses Bayesian regression with an experiential kernel to select measurements and reconstruct channels. Simulations report higher estimation accuracy than existing assumption-based schemes in both model-mismatched and model-matched cases.

  • Problem

    Dense FAS port channels and limited RF connections make CSI acquisition difficult, and existing estimators depend on assumptions such as slow variation or angular-domain sparsity.

    Source: A. Prior Work · p. 2

  • Method

    S-BAR models FAS channels as a stochastic process and successively reconstructs them through kernel-based sampling and regression using an experiential kernel.

    Source: B. Our Contributions · p. 2, §VI · p. 12

  • Results

    S-BAR achieves higher estimation accuracy than existing schemes relying on channel assumptions in both model-mismatched and model-matched cases.

    Source: §VI · p. 12

  • Takeaways & Limitations

    The proposed experiential-kernel approach provides a non-parametric FAS channel estimator that selects a few informative channels for measurement and reconstructs high-dimensional channels.

    Source: §VI · p. 12

Abstract

from arXiv · show

Fluid antenna systems (FASs) can reconfigure their antenna locations freely within a spatially continuous space. To keep favorable antenna positions, the channel state information (CSI) acquisition for FASs is essential. While some techniques have been proposed, most existing FAS channel estimators require several channel assumptions, such as slow variation and angular-domain sparsity. When these assumptions are not reasonable, the model mismatch may lead to unpredictable performance losses. In this paper, we propose the successive Bayesian reconstructor (S-BAR) as a general solution to estimate FAS channels. Unlike model-based estimators, the proposed S-BAR is prior-aided, which builds the experiential kernel for CSI acquisition. Inspired by Bayesian regression, the key idea of S-BAR is to model the FAS channels as a stochastic process, whose uncertainty can be successively eliminated by kernel-based sampling and regression. In this way, the predictive mean of the regressed stochastic process can be viewed as a Bayesian channel estimator. Simulation results verify that, in both model-mismatched and model-matched cases, the proposed S-BAR can achieve higher estimation accuracy than the existing schemes.

I. INTRODUCTION

Fluid antenna systems (FASs) move a small number of antennas among densely spaced ports, making CSI acquisition important but difficult. The paper introduces S-BAR, a Bayesian-regression-based estimator, and considers its setting alongside prior FAS work and channel-estimation methods.

I. INTRODUCTION

  • FASs use a few movable antennas to explore a space of closely spaced ports, and their transmission performance depends on antenna positions and channel state information.The system model considers an uplink with an N-port base station, M fluid antennas (M ≪ N), and a single-antenna user.Source: §I · p. 1, A. System Model · p. 3

A. Prior Work

  • Dense port channels and limited RF connections make real-time CSI acquisition difficult, while existing estimators rely on assumptions such as slow variation, angular sparsity, or known angles of arrival.The cited estimators include SeLMMSE, compressed sensing, and least squares with known AoAs.Source: A. Prior Work · p. 2

B. Our Contributions

  • The paper introduces Bayesian regression to FAS channel estimation, using port switching as sampling and channel correlation to motivate a kernel-based, non-parametric reconstruction.The experiential kernel and selected ports are used to characterize channel correlations and reconstruct FAS channels.Source: B. Our Contributions · p. 2
  • S-BAR models FAS channels as a stochastic process and uses kernel-based sampling and regression to reconstruct channels without assuming a channel model.Its two stages design a switch matrix by maximizing mutual-information increment between adjacent sampling points, then regress channel observations using the designed matrix and experiential kernel.Source: B. Our Contributions · p. 2
  • Simulations report higher S-BAR estimation accuracy than benchmark schemes in both model-mismatched and model-matched cases, with accuracy increasing by an order of magnitude.The paper also derives an MSE that depends only on channel-covariance eigenvalues and noise power, and reports online computational complexity linear in the number of ports.Source: B. Our Contributions · p. 2

A. System Model

  • The system model represents each timeslot’s antenna locations with a switch matrix and reconstructs the N-dimensional channel from PM-dimensional noisy pilots, where PM ≪ N.Each of the M antennas switches among N ports to receive pilots during the coherence-time frame.Source: A. System Model · p. 3

B. Existing Methods for FAS Channel Estimation

The compressed-sensing estimator reconstructs a high-dimensional FAS channel from low-dimensional pilots by assuming spatial sparsity and using the switch matrix as a virtual analog precoder. One described implementation, FAS-OMP, estimates channel parameters before recovering the channel explicitly.

1) Compressed Sensing Based Estimator:

  • Under spatial sparsity, the FAS channel is represented by a small number of significant angular-domain components, enabling compressed-sensing reconstruction from low-dimensional pilots.The received signal uses the sensing matrix Ψ = S^H F, with S randomly selected subject to the switch-matrix constraints.Source: 1) Compressed Sensing Based Estimator · pp. 3–4
  • FAS-OMP iteratively selects sensing-matrix columns to estimate path gains and incident angles, then substitutes those parameters to recover the channel explicitly.The algorithm uses orthogonal projection and residual updates during its iterative selection.Source: 1) Compressed Sensing Based Estimator · p. 4

2) Alternating Optimization Based Estimator:

The FAS-ML reconstructor addresses off-grid error by estimating path gains and incident angles through maximum likelihood. Because those variables are coupled, it updates them alternately rather than solving for them simultaneously.

2) Alternating Optimization Based Estimator:

  • When true channel parameters fall between grid points, gridded compressed sensing can incur estimation error, motivating maximum-likelihood optimization of path gains and incident angles.The FAS-ML reconstructor is introduced as an alternating-optimization method for this objective.Source: 2) Alternating Optimization Based… · p. 4
  • Because path gains and incident angles are coupled, FAS-ML updates one while fixing the other, monotonically increasing the likelihood toward a suboptimal solution.The method initializes these variables using a compressed-sensing algorithm such as OMP.Source: 2) Alternating Optimization Based… · p. 4

C. Challenge and Opportunity

Existing FAS estimators can lose accuracy under model mismatch and may need many pilots, while successive measurement of correlated ports offers a non-parametric reconstruction opportunity. The paper introduces S-BAR through Bayesian regression, which samples uncertain points and uses posterior regression to estimate the channel.

C. Challenge and Opportunity

  • Most existing estimators assume spatial sparsity, so model mismatch can cause unpredictable performance loss; sparse switch matrices also mean high pilot overhead may be needed.The received pilots contain information about only the channels at selected ports.Source: C. Challenge and Opportunity · p. 4
  • Because FAS ports are closely spaced and their channels strongly correlated, sampling one port can partially reduce uncertainty at nearby ports.Carefully chosen successive measurements and regression can therefore reconstruct channels non-parametrically.Source: C. Challenge and Opportunity · p. 4

III. PROPOSED SUCCESSIVE BAYESIAN RECONSTRUCTOR

  • The paper proposes S-BAR as a high-accuracy solution for FAS channel estimation, using Bayesian regression to match the antennas' successive sampling and correlated channels.S-BAR builds experiential kernels for the objective and uses kernel-based sampling and regression.Source: §III · p. 5

A. Bayesian Linear Regression

  • Bayesian linear regression models the unknown function as a Gaussian process whose kernel encodes smoothness, then updates its posterior from noisy measurements.For one-by-one sampling, the next point is chosen at maximum posterior variance; sampling continues until uncertainty falls below the tolerance, and the posterior mean estimates the function.Source: A. Bayesian Linear Regression · p. 5

B. Working Principle of the Proposed S-BAR

S-BAR treats the channel as a Gaussian process with a kernel and designs switch-matrix columns greedily to reduce channel uncertainty. It uses posterior covariance to guide port selection and the posterior mean as the channel reconstructor.

B. Working Principle of the Proposed S-BAR

  • With a Gaussian-process channel model, the posterior mean is the optimal reconstructor for a given switch matrix, and expected squared error is the trace of posterior covariance.Thus, switch-matrix design can be framed as minimizing that covariance trace.Source: B. Working Principle of the Proposed… · pp. 5–6
  • S-BAR designs switch-matrix columns greedily by maximizing each column's mutual-information increment between the received pilots and channels.The method adopts the greedy-sampling idea from Bayesian linear regression.Source: B. Working Principle of the Proposed… · p. 6
  • The greedy choice is equivalent to minimizing posterior covariance, but port selection must exclude ports already assigned to another antenna in the same timeslot.The sampling sequence tracks selected ports so the zero-one switch constraint is satisfied.Source: B. Working Principle of the Proposed… · p. 6
  • Algorithm 4 designs the switch matrix and reconstructor weights from the kernel, then applies them to received pilots to reconstruct the channel online.The algorithm separates offline design from online regression.Source: B. Working Principle of the Proposed… · p. 6

C. Hybrid Offline and Online Implementation of S-BAR

S-BAR separates kernel-dependent design from channel measurement: the switch matrix and reconstruction weights are prepared offline, then deployed for online regression. This split is intended to reduce online computational complexity.

1) Stage 1 (Offline Design):

C. Hybrid Offline and Online Implementation of S-BAR

A. Performance Analysis of S-BAR

The analysis characterizes S-BAR's estimation error and achievable minimum MSE, then describes its offline and online computational costs. The minimum depends on channel-covariance eigenvalues, and the reported implementation can reduce practical online complexity through offline design.

2) Computational Complexity of S-BAR:

1) Estimation Accuracy of S-BAR:

  • The error characterization permits evaluation under kernel–covariance mismatch and is independent of the real channel mean; with a matched kernel, full observation recovers the LMMSE posterior-covariance result.The paper identifies the perfect kernel-based Bayesian reconstructor as equivalent to an LMMSE estimator.Source: 1) Estimation Accuracy of S-BAR · p. 7
  • For rank-K channel covariance, Emin = Σ_k λ_kσ²/(λ_k + σ²); equality requires the kernel to match the real covariance and complete channel observation, PM = N.The minimum depends on the positive eigenvalues of the channel covariance and noise power.Source: 1) Estimation Accuracy of S-BAR · p. 7

2) Computational Complexity of S-BAR:

  • Stage 1 complexity is high but can be performed offline; practical S-BAR use is then linear in port count N, and its kernel-based designs are user-independent.The paper also states that online weighted-sum reconstruction can be parallelized, making time complexity dimension-independent.Source: 2) Computational Complexity of S-BAR · pp. 7–8

B. Kernel Selection of S-BAR

Kernel selection determines the flexibility of S-BAR's regression model. The section contrasts a covariance kernel trained from channel data with experiential kernels for scenarios where explicit CSI is difficult to obtain.

1) Covariance Kernel:

  • An ideal covariance kernel is the real channel covariance, which can be approximated from CSI in existing channel datasets before S-BAR is used.The kernel is treated as prior covariance.Source: 1) Covariance Kernel · p. 8

2) Experiential Kernel:

  • When explicit CSI is difficult to acquire, experiential kernels trained from received pilots offer a more practical option.The section recommends balancing complexity and practicality when choosing them.Source: 2) Experiential Kernel · p. 8
  • The exponential kernel is suited to channels without obvious regularity because it is not sensitive to outliers.It is also called the Laplacian kernel.Source: 2) Experiential Kernel · p. 8
  • The Bessel kernel is suited to complex-valued channels with oscillatory or periodic patterns.Its flexibility allows it to adapt to repeating fluctuations.Source: 2) Experiential Kernel · p. 8
  • The hyperparameter η is selected by maximum likelihood using R channel realizations, received pilots, and a one-dimensional search over the kernel candidates.The experiential kernel can be exponential or Bessel.Source: 2) Experiential Kernel · p. 8

V. SIMULATION RESULTS

  • The simulation section evaluates S-BAR behavior, receiver-noise effects, pilot-overhead effects, and electromagnetic-coupling effects.It first specifies the setup and benchmarks, then examines these four aspects.Source: §V · p. 8

A. Simulation Setup and Benchmarks

Across QuaDRiGa and SSC channels, S-BAR achieves the highest estimation accuracy among the evaluated schemes, while its experiential kernels perform similarly to a covariance kernel. Simulations also illustrate how additional samples reduce channel uncertainty and where sampling is especially useful.

1) Simulation Setup:

  • The simulations use 256 ports, four fluid antennas, a 10λ array, ten pilots, and 100 training timeslots, with QuaDRiGa and SSC channel models.Source: 1) Simulation Setup · p. 9

2) Simulation Schemes:

  • The benchmarks are FAS-OMP, FAS-ML, SeLMMSE, and S-BAR with a specified kernel.FAS-OMP and FAS-ML use sparsity-based reconstruction; SeLMMSE measures PM equally spaced ports and applies zero-order interpolation.Source: 2) Simulation Schemes · p. 9

C. NMSE versus the Receiver SNR

B. Estimation Behavior of the Proposed S-BAR

C. NMSE versus the Receiver SNR

  • Experiential exponential and Bessel kernels achieve performance similar to the covariance kernel, with nearly coincident curves for SSC channels.The experiential kernels can provide useful prior knowledge even when the real channel covariance is unknown.Source: C. NMSE versus the Receiver SNR · p. 10

D. NMSE versus the Number of Pilots

The pilot study compares S-BAR with benchmark estimators under QuaDRiGa and SSC channels, then examines estimation under electromagnetic coupling. The conclusion presents S-BAR as a general Bayesian approach and identifies extensions for future work.

D. NMSE versus the Number of Pilots

  • At NMSE −15 dB in SSC channels, S-BAR with the preferred kernel needs 5 pilots, versus 13 for FAS-ML and 20 each for FAS-OMP and SeLMMSE; overall pilot overhead falls by about 50%.The comparison is plotted for QuaDRiGa and SSC channels in Figs. 6 and 7.Source: D. NMSE versus the Number of Pilots · p. 11, Figure 6 · p. 11, Figure 7 · p. 11
  • With non-preferred kernels, S-BAR reaches −21 dB at P = 10 in QuaDRiGa, 6 dB lower than the best SeLMMSE, and −20 dB in SSC, about 10 dB lower than the best FAS-ML.The respective non-preferred kernels are Σbes for QuaDRiGa and Σexp for SSC.Source: D. NMSE versus the Number of Pilots · p. 11
  • In SSC, S-BAR’s error falls quickly for P < 6 as sampling captures spatial features, then more slowly as measurement noise dominates; this pattern is not observed for QuaDRiGa.The paper attributes the difference to SSC’s periodicity, which is less evident in QuaDRiGa channels.Source: D. NMSE versus the Number of Pilots · p. 11

E. Impact of EM Coupling on Estimation Accuracy

  • Despite electromagnetic coupling, S-BAR achieves NMSE below −15 dB; without suppression, coupling limits accuracy below 30λ, while weaker spatial correlation limits it above 30λ.The study varies normalized array length W/λ and plots NMSE in Fig. 8.Source: Figure 8 · p. 11, E. Impact of EM Coupling on Estimation… · p. 12
  • For the no-suppression and coupling-suppression estimators, increasing port spacing weakens spatial correlation and raises NMSE; the suppression scheme models the equivalent channel with a modified kernel.The ideal without-coupling case assumes hardware isolation eliminates inter-port coupling.Source: E. Impact of EM Coupling on Estimation… · p. 12

VI. CONCLUSIONS

  • S-BAR uses an experiential kernel and selected informative channel measurements to reconstruct high-dimensional FAS channels; simulations report higher accuracy than assumption-based schemes in matched and mismatched cases.The paper identifies wideband FAS estimation, phased-array estimation, and RIS–FAS cooperation as follow-up directions.Source: §VI · p. 12

APPENDIX A PROOF OF Lemma 1

The appendices establish the mutual-information sampling criterion, derive a posterior-covariance recursion, and rewrite the S-BAR square error for its proof.

APPENDIX A PROOF OF Lemma 1

  • Maximizing the information gain from the next observation is reduced to choosing s that maximizes sᴴΣₜs.The proof uses matrix partition operations and triangularization.Source: Appendix A · p. 12

APPENDIX B PROOF OF Lemma 2

  • Block-matrix inversion is used to derive the recursion for Σₜ, which is then substituted into the posterior covariance expression.The proof invokes the definitions of Sₜ and the selected sample index nₜ.Source: Appendix B · pp. 12–13

APPENDIX C PROOF OF Lemma 3

  • At t = PM, the S-BAR square error is rewritten using the stated channel and estimator definitions.The proof then uses the distribution of z to complete the calculation.Source: Appendix C · p. 13

APPENDIX D PROOF OF Lemma 4

The proof characterizes the error as a quadratic form, identifies the covariance kernel that minimizes it, and relates the minimum to observing all port channels.

APPENDIX D PROOF OF Lemma 4

  • The MSE is a quadratic form in Π, so its minimum is sought by setting its partial derivative with respect to Π to zero.The proof derives the derivative and reduces the claim to a further identity.Source: Appendix D · p. 13
  • Because S-BAR does not depend on noise power, the proof considers arbitrary σ² and uses the limiting case σ² → ∞.Source: Appendix D · p. 13
  • The proof identifies Σ = Σcov as necessary for the identity and shows that the MSE is minimized when all port channels are observed, PM = N, though this may be impractical.The argument relates the MSE to the trace of the posterior covariance and then uses matrix inversion and eigenvalue decomposition.Source: Appendix D · p. 13
Loading 2312.06551v5…