PaperlayerView exact revision on arXiv

Source-linked AI summary

Uplink-aided High Mobility Downlink Channel Estimation over Massive MIMO-OTFS System

Yushan Liu, Shun Zhang, Feifei Gao, Jianpeng Ma, Xianbin Wang

arXiv:2003.07045v1eess.SP

TL;DR

High mobility creates rapid channel variation and substantial training overhead for massive MIMO-OTFS downlink estimation. The paper uses UL EM-VB parameter recovery, a fast low-complexity implementation, and UL–DL reciprocity to reconstruct downlink channels. Simulations report a valid and robust strategy with three-dimensional angle-delay-Doppler channel training schemes.

  • Problem

    Rapid channel variation makes existing approaches ineffective in high-mobility scenarios, while massive MIMO-OTFS downlink estimation can require substantial training overhead.

  • Method

    The paper formulates a UL time-domain massive MIMO-OTFS model, applies EM-VB and fast Bayesian inference, then exploits UL–DL reciprocity to reconstruct downlink parameters and channels.

  • Results

    The proposed strategy is reported as valid and strongly robust, with downlink channel estimation studied through three-dimensional angle-delay-Doppler channel training schemes.

  • Takeaways & Limitations

    UL-aided parameter transfer provides the paper’s route to high-mobility downlink channel estimation over the massive MIMO-OTFS delay-Doppler-angle domain.

Abstract

from arXiv · show

Although it is often used in the orthogonal frequency division multiplexing (OFDM) systems, application of massive multiple-input multiple-output (MIMO) over the orthogonal time frequency space (OTFS) modulation could suffer from enormous training overhead in high mobility scenarios. In this paper, we propose one uplink-aided high mobility downlink channel estimation scheme for the massive MIMO-OTFS networks. Specifically, we firstly formulate the time domain massive MIMO-OTFS signal model along the uplink and adopt the expectation maximization based variational Bayesian (EM-VB) framework to recover the uplink channel parameters including the angle, the delay, the Doppler frequency, and the channel gain for each physical scattering path. Correspondingly, with the help of the fast Bayesian inference, one low complex approach is constructed to overcome the bottleneck of the EM-VB. Then, we fully exploit the angle, delay and Doppler reciprocity between the uplink and the downlink and reconstruct the angles, the delays, and the Doppler frequencies for the downlink massive channels at the base station. Furthermore, we examine the downlink massive MIMO channel estimation over the delay-Doppler-angle domain. The channel dispersion of the OTFS over the delay-Doppler domain is carefully analyzed. Various numerical examples are presented to confirm the validity and robustness of the proposed scheme.

I. INTRODUCTION

High mobility makes conventional massive MIMO channel estimation difficult because block-fading assumptions and OFDM suffer under rapid variation and Doppler-induced interference. The paper proposes UL-aided massive MIMO-OTFS estimation using low-complexity Bayesian parameter recovery and UL–DL reciprocity.

  • Motivation: High-speed channel variation challenges methods designed for block-fading massive MIMO channels.Block-fading is considered reasonable at low velocity but may not apply in high-speed scenarios.
  • Motivation: OFDM can suffer severe inter-carrier interference from Doppler spread in time-variant frequency-selective channels.
  • OTFS rationale: OTFS converts time-variant channels into roughly constant channels in the delay-Doppler domain and can improve performance over OFDM.
  • Contributions: The paper develops a time-domain UL massive MIMO-OTFS model and uses EM-VB to recover angle, delay, Doppler, and gain parameters for scattering paths.
  • Contributions: A low-complexity EM-VB method uses fast Bayesian inference to avoid the large matrix inversion bottleneck.
  • Contributions: UL–DL angle, delay, and Doppler reciprocity enables downlink parameter reconstruction and delay-Doppler-angle channel estimation, including scheduling and reduced-dimension LS recovery.

II. SYSTEM MODEL

The system models high-mobility massive MIMO-OTFS channels through sparse propagation paths characterized by angle, delay, Doppler, and gain. OTFS processing yields delay-Doppler representations, while spatial transformation exposes a sparse delay-Doppler-angle channel that motivates UL-aided recovery.

  • Channel model: The considered single-cell system has a massive-antenna BS, single-antenna users, frequency-selective fading, and time-selective channels caused by mobility.
  • Channel model: Each scattering path is associated with a direction of departure, Doppler frequency shift, time delay, and channel gain.
  • OTFS signal processing: The OTFS transmitter maps delay-Doppler data through Fourier processing, OFDM modulation, cyclic-prefix insertion, and vectorization into a time-domain signal.
  • OTFS signal processing: At reception, cyclic-prefix removal, DFT processing, and SFFT post-processing recover a two-dimensional delay-Doppler data block.
  • Sparse channel representation: Spatial DFT reveals that the channel is sparse over the delay-Doppler-angle domain, with dominant elements determined by the physical path parameters.
  • Estimation motivation: Direct downlink recovery requires high-complexity sparse recovery and large training overhead, motivating UL-aided estimation through angle, delay, and Doppler reciprocity even in FDD.

III. SBL-BASED CHANNEL PARAMETER CAPTURING ALONG UL

The uplink procedure uses short, orthogonal user training sequences to capture channel observations at the base station, then represents sparse paths over angle and delay grids for parameter extraction.

  • UL training: Users transmit short training sequences separately, with orthogonal time-domain sequences assigned to different users.Each sequence occupies (Lcp + Nt)Ts, and user k starts at (n1 + (Lcp + Nt)(k −1))Ts.
  • UL observations: The base station discards cyclic-prefix samples and collects M-dimensional antenna outputs over Nt valid time samples.The collected vectors y^ul_k,n are formed after removing the first Lcp samples.
  • UL observations: The uplink observations include additive white Gaussian noise across the M receiving antennas, with independent noise vectors across distinct time samples.The noise has zero mean and covariance σ^2I_M.
  • Sparse representation: Uniform sampling grids are constructed over angle and delay, with N angle grids and L delay grids, where L exceeds the maximum delay index.The sparse matrix G^ul_k maps P nonzero channel paths onto these grids.
  • Sparse representation: The sparse matrix identifies path locations by associating each scattering path with nearby delay and angle grid points.The closest delay and angle grids are selected for each path when determining nonzero entries.

1. Then, it can be concluded

The uplink signal model converts time-domain observations into a sparse angle-delay representation and uses sparse Bayesian learning to recover channel parameters under practical Doppler and grid-mismatch conditions.

  • Doppler approximation: For a 6 GHz carrier, Nt = 40, sampling rate 1/Ts = 20 MHz, and speed 300 km/h, the maximum phase accumulation over NtTs is 0.021.Because this is much smaller than 1, the Doppler-dependent exponential is approximated using a Taylor expansion.
  • Off-grid modeling: The dictionary A^ul is built on uniform angle grids, but true directions of arrival may fall between predefined grid points.This off-grid mismatch is addressed by approximating the practical steering vector around its nearest angle grid.
  • SBL formulation: The channel parameter extraction task is formulated as sparse recovery and implemented with a sparse Bayesian learning framework.The model uses hierarchical priors, including Gaussian channel coefficients with a diagonal precision matrix and Gamma-distributed hyperparameters.
  • SBL formulation: Integrating out the Gamma hyperparameters yields a Student-t marginal distribution that is strongly peaked around zero, promoting sparse channel coefficients.The hierarchical prior is used to simplify posterior inference while favoring sparse solutions.

C. Solving SBL Using EM-VB

The EM-VB procedure approximates the posterior distribution of hidden channel variables by iteratively updating variational factors and their posterior statistics.

  • Variational inference: Variational Bayesian inference seeks a tractable distribution q(H_k) that closely approximates the posterior over all hidden variables.The hidden-variable set includes the equivalent sparse channel representation and related latent parameters.
  • EM-VB updates: The maximization step derives posterior statistics for each hidden variable in H_k.These statistics are then used in the expectation phase to update the approximate posterior.
  • EM-VB updates: The variational distribution for the sparse channel coefficients is complex Gaussian with an updated mean and covariance.The coefficient mean is also used when deriving the variational distribution of the precision hyperparameters.
  • Hyperparameter updates: The precision variables follow Gamma distributions whose parameters are updated from the coefficient posterior quantities.The resulting updates use eak = ak + 1 together with an updated scale parameter ebk.

E. Maximization Step of EM-VB

The EM-VB maximization step updates Doppler-, gain-, and sparsity-related parameters using current posterior quantities, and the alternating procedure is summarized in Algorithm 1.

  • Parameter updates: Posterior means are arranged into blocks to derive the updates for β^ul_k and related channel parameters.The block structure collects coefficient means across angle and delay indices.
  • Parameter updates: The maximization step separately estimates the channel gain-related parameter β^ul_k and the Doppler parameter υ^ul_k.The Doppler update uses the matrix D_k to organize variables across iterations.
  • Algorithm: The expectation and maximization steps are implemented iteratively, updating the sparse coefficients and hyperparameters until the EM procedure terminates.The iterative scheme increments the EM iteration counter and is summarized in Algorithm 1.
  • Algorithm: Algorithm 1 initializes the training vector and hyperparameters, then repeatedly updates the precision variables within a bounded EM iteration loop.The listed inputs include t, ak, and bk, while αk is updated by equation (25).

F. Low Complex EM-VB

The low-complex EM-VB procedure addresses the computational bottleneck of standard EM-VB by updating sparse model components efficiently rather than repeatedly inverting the full matrix.

  • Standard EM-VB requires an NL×NL matrix inversion at each iteration, motivating a fast Bayesian inference procedure.The computational bottleneck arises in the expectation step.
  • The fast procedure updates one entry of αk at a time while retaining the expectation and maximization structure of EM-VB.Only the recovery of αk in the expectation step is modified.
  • Matrix identities allow αk,i to be updated using quantities independent of αk,i, supporting efficient coordinate-wise Bayesian inference.The derivation isolates the changing component while holding αk,−i fixed.
  • Sparse basis pruning removes inactive columns from the observation space, reducing the dimension of subsequent matrix inversions.The effective space collects only basis vectors associated with nonzero sparse channel components.
  • The resulting low-complex EM-VB preserves the EM-VB framework while using substantially simpler updates based on sparse active components.The procedure is summarized as an iterative algorithm with basis-vector addition, retention, or pruning.

G. Complexity Analysis

The complexity analysis attributes the main cost of variational Bayesian inference to large matrix inversion and shows how fast Bayesian updates and sparsity reduce this burden.

  • Variational Bayesian inference is dominated by computing the matrix inverse Σk, requiring complexity on the order of O(N^3L^3).
  • Updating a single αk,i per iteration instead of the entire vector produces more efficient parameter updates.
  • Because the channel representation is highly sparse, Σk can be constructed in fewer dimensions than NL×NL, further reducing computational complexity.
  • The extracted uplink parameters are subsequently used to construct downlink delay-Doppler-angle channel parameters.

1) Deriving the Parameters in TDD Mode:

In TDD mode, uplink and downlink channel parameters are reciprocal, enabling downlink reconstruction from uplink estimates and structured delay-Doppler-angle training.

  • 1) Deriving the Parameters in TDD Mode:: TDD reciprocity makes the downlink channel model parameters equal to the uplink parameters.
  • 1) Deriving the Parameters in TDD Mode:: The delay-Doppler-angle channel is represented as a sparse 3D structure whose dominant coordinates correspond to physical path signatures.
  • 1) Deriving the Parameters in TDD Mode:: OTFS symbols disperse over the delay-Doppler domain, but known delays and Doppler frequencies determine the exact dispersion locations.This structure can reduce the required observation dimension.
  • 1) Deriving the Parameters in TDD Mode:: Training schemes optimize pilots over a 3D cubic resource space with delay, Doppler, and angle dimensions.
  • 1) Deriving the Parameters in TDD Mode:: For orthogonal delay-Doppler paths, one embedded pilot can estimate the channel using observations associated with the P scattering paths.
  • 1) Deriving the Parameters in TDD Mode:: In the multi-user case, users are separated first by angle and then by delay-Doppler signatures, with guard gaps preventing overlap.The scheduling assigns distinct 3D resources so users can decode without inter-user interference.

3) General Case:

The general case addresses users whose paths are not fully separated by angle or delay-Doppler by using continuous observation regions and a feasible channel-recovery method.

  • 3) General Case:: The general case arises when all paths of one user are not distinguished over the angle or delay-Doppler domains.
  • 3) General Case:: For this case, the paper states that fully exploiting super-resolution to schedule delay-Doppler-angle resources is beyond its scope.
  • 3) General Case:: The paper nevertheless presents one feasible channel-recovering method for the general case.
  • 3) General Case:: The method fixes effective pilots within a bounded region and models each user’s pilot observations as a continuous region.
  • 3) General Case:: Received observations are arranged into a vector-matrix model with additive white Gaussian noise, after which least squares estimates the channel.

C. Pilot Overhead Analysis

The scheme estimates uplink channel parameters, exploits uplink–downlink reciprocity, and designs delay-Doppler-angle training schemes whose overhead depends on path separability and sparsity. Simulations assess estimation accuracy across mobility, SNR, observations, sparsity, and channel models.

  • Training architecture: The method extracts uplink channel parameters in the frequency-time-antenna domain using a time-domain training sequence of length K(Lcp + Nt).It then reconstructs downlink channels in the delay-Doppler-angle domain.
  • Training architecture: Three downlink training schemes are constructed over the delay-Doppler-angle domain according to scattering-path characteristics.The schemes account for whether paths can be separated over the delay-Doppler domain.
  • Pilot overhead: When each user’s paths are separable over delay-Doppler, at most min{KP, M} delay-Doppler-angle grids are required.A fully distinguished case uses KP^2 grids in the three-dimensional resource cube.
  • Pilot overhead: The general-case pilot overhead is at most min{KPHdHD, MHdHD}, with Hd and HD selected according to the training design.The paper contrasts this with prior work focused only on downlink channel estimation.
  • Multi-user operation: For multiple users, feasible schemes and path scheduling are provided to avoid inter-user interference.The paper frames these schemes within massive MIMO-OTFS data transmission.
  • Numerical evaluation: Fast EM-VB performance is close to EM-VB, with a 2 ∼3 dB loss, while requiring only five iterations to reach steady states.Across simulations, MSE decreases with SNR and increases only slightly with user speed.
  • Numerical evaluation: Increasing observations reduces MSE and accelerates convergence at higher SNR, while increasing path count helps only when P remains below Nt.When P exceeds Nt, insufficient independent observation equations limit estimation even if supports are correctly identified.
  • Numerical evaluation: The proposed downlink estimator outperforms the block-fading method and is not closely coupled with users’ mobility speeds.This comparison uses the delay-Doppler-angle channel representation.

VI. CONCLUSION

The paper develops an uplink-aided high-mobility downlink estimator for massive MIMO-OTFS. It combines EM-VB-based uplink parameter recovery, fast Bayesian inference, uplink–downlink reciprocity, and three-dimensional downlink training, with simulations finding the strategy valid and robust.

  • Scope: The paper studies uplink-aided high-mobility downlink channel estimation for massive MIMO-OTFS in the three-dimensional angle-delay-Doppler domain.The conclusion presents this as the paper’s central estimation problem.
  • Method: EM-VB recovers uplink angles, delays, Doppler frequencies, and gains, while fast Bayesian inference supplies a lower-complexity EM-VB approach.The recovered parameters describe physical scattering paths.
  • Method: The method exploits angle, delay, and Doppler reciprocity to reconstruct downlink channel parameters at the base station.Downlink estimation is examined over the delay-Doppler-angle domain.
  • Method: The paper analyzes OTFS channel dispersion over delay-Doppler and designs three downlink three-dimensional training schemes based on scattering characteristics.These analyses support the downlink training design.
  • Conclusion: Simulation results indicate that the proposed strategy is valid and robust.The conclusion states this outcome without narrowing it to a single simulation condition.
Loading 2003.07045v1…