PaperlayerView exact revision on arXiv

Source-linked AI summary

Efficient Coordinated Recovery of Sparse Channels in Massive MIMO

Mudassir Masood, Laila H. Afify, Tareq Y. Al-Naffouri

arXiv:1409.4671v2stat.APcs.IT

TL;DR

The paper addresses sparse massive-MIMO channel estimation with few pilots despite unknown or non-Gaussian channel-tap distributions. It proposes distributed Bayesian algorithms that exploit approximately common support across neighboring antennas, with simulations showing good recovery performance and further gains from data-aided processing.

  • Problem

    Sparse channels and approximately common support across neighboring antennas motivate reducing the pilots needed for massive-MIMO channel estimation, while channel distributions may be unknown or difficult to characterize.

  • Method

    The proposed distributed Bayesian algorithms combine channel sparsity, approximately shared support, neighbor information exchange, and a data-aided extension without assuming a distribution for active channel taps.

  • Results

    The algorithms show good performance across scenarios using relatively few pilots; MB-R achieves nearly two orders of magnitude lower BER than BR at SNR = 35 dB and Q = 4.

  • Takeaways & Limitations

    Neighbor collaboration supports sparse-channel recovery with limited pilots, while data-aided processing further reduces pilot requirements and enhances channel estimates.

  • Takeaways & Limitations

    MB algorithms can improve performance when pilots are scarce, but their advantage comes with relatively high computational and communication costs.

Abstract

from arXiv · show

This paper addresses the problem of estimating sparse channels in massive MIMO-OFDM systems. Most wireless channels are sparse in nature with large delay spread. In addition, these channels as observed by multiple antennas in a neighborhood have approximately common support. The sparsity and common support properties are attractive when it comes to the efficient estimation of large number of channels in massive MIMO systems. Moreover, to avoid pilot contamination and to achieve better spectral efficiency, it is important to use a small number of pilots. We present a novel channel estimation approach which utilizes the sparsity and common support properties to estimate sparse channels and require a small number of pilots. Two algorithms based on this approach have been developed which perform Bayesian estimates of sparse channels even when the prior is non-Gaussian or unknown. Neighboring antennas share among each other their beliefs about the locations of active channel taps to perform estimation. The coordinated approach improves channel estimates and also reduces the required number of pilots. Further improvement is achieved by the data-aided version of the algorithm. Extensive simulation results are provided to demonstrate the performance of the proposed algorithms.

I. INTRODUCTION

Massive MIMO requires estimating many channels while limiting pilot overhead, centralized processing, and communication costs. The paper proposes distributed Bayesian algorithms that exploit sparse channel impulse responses and approximately common support among neighboring antennas.

  • Motivation: Massive MIMO channel estimation must reduce pilots to limit pilot contamination while processing many channels efficiently.The paper identifies distributed computation and low communication overhead as additional requirements.
  • Channel structure: Wireless channel impulse responses are sparse because only a few multipath paths are significant.The paper treats sparsity as a central property for reducing pilot overhead.
  • Channel structure: Neighboring antennas observe channels with approximately common support but different active-tap amplitudes and phases.The shared support arises because closely spaced antennas experience similar path arrival times.
  • Proposed approach: The proposed algorithms perform distributed Bayesian channel estimation while remaining agnostic to the active-tap distribution.Neighboring antennas share support beliefs with limited communication overhead.
  • Proposed approach: A data-aided extension identifies reliable carriers to further reduce pilots and improve channel impulse-response estimates.The algorithms are based on support-agnostic Bayesian matching pursuit.

C. Spatial Channel Model

The spatial channel model distinguishes space-invariant and space-variant antenna arrays according to how channel support changes across the array. In both settings, nearby antennas provide approximately shared support that coordinated estimation can exploit.

  • Array configurations: A uniform rectangular array is used for convenience, although the algorithm is intended to support linear, planar, and cylindrical configurations.The paper states that the approach applies to one-, two-, or three-dimensional antenna configurations.
  • Space-invariant arrays: Space-invariant arrays have the same channel support across antennas, while tap strengths may vary.This corresponds to closely spaced elements whose path arrival times remain similar.
  • Space-variant arrays: Space-variant arrays have support that changes slowly across the array, so each central antenna and its four neighbors have approximately common support.The model links support commonality to antenna separation, bandwidth, and operating frequency.
  • Model scope: The proposed approach addresses both space-invariant and space-variant antenna-array cases despite limited prior work on space-variant sparse-channel estimation.The paper distinguishes the cases using the maximum resolvable distance relative to array dimensions.

D. Pilots

The pilot model reserves K subcarriers for pilots and uses the remaining N − K carriers for data. Exploiting sparse channels and approximately shared support across adjacent antennas can reduce the pilots needed for channel estimation.

  • Pilot allocation: Pilots occupy K subcarriers, while the remaining N − K carriers transmit data.The received pilot vectors and sensing matrix are formed by selecting entries and rows indexed by the pilot set P.
  • Pilot requirement: Directly solving for a channel with delay spread L requires at least K ≥ L pilots, reducing spectral efficiency.This requirement applies without exploiting channel sparsity.
  • Sparsity-aware estimation: Sparse channels and approximately common support across adjacent antennas enable substantially fewer pilots under compressed-sensing methods.The approach uses both sparsity and neighboring-antenna support similarity.
  • Estimation procedure: The channel-estimation procedure is organized into Bayesian estimation at each antenna, distributed estimation, and data-aided estimation.These are presented as three successive stages of the channel-estimation approach.
  • Observation model: For the linear model, Y and W have dimension K × 1, h has dimension L × 1, and A has dimension K × L.The Bayesian estimator targets the wireless channel impulse response h.

A. Simple Channel Estimation using SABMP

SABMP estimates sparse channel impulse responses through a greedy Bayesian search over dominant support sets. Its modified form accommodates non-identically distributed tap activity and supplies quantities needed for distributed and data-aided estimation.

  • Bayesian channel estimation: SABMP seeks an MMSE estimate of the sparse channel h from K observations in the linear model Y = Ah + W.The algorithm estimates h given the observation vector Y and the sensing matrix A.
  • Bayesian channel estimation: Because enumerating all 2^L support sets is impractical for large L, SABMP approximates the estimate using supports with significant posterior probability.The dominant support set S_d is selected to retain computational feasibility.
  • Greedy support search: SABMP greedily selects the best support of size one, then expands it incrementally by adding the tap that maximizes the support-selection metric.This reduces the search compared with exhaustive support enumeration.
  • Greedy support search: The dominant support set is an incremental sequence S_d = {S_1, S_2, …, S_Tmax} of increasingly large supports.The sequence starts with {α_1} and adds one detected location at each step.
  • Algorithm modifications: The modified algorithm assigns independent, non-identically distributed Bernoulli probabilities to taps instead of a common activation probability.Tap-specific probabilities λ_i can reflect differing prior likelihoods of activation.
  • Algorithm outcome: The resulting greedy procedure estimates each antenna’s sparse CIR in a distribution-agnostic manner.The algorithm can remain useful when the active-tap distribution is unknown.
  • Algorithm modifications: The modification also outputs estimation-error covariance and marginal activation probabilities for distributed and data-aided channel estimation.These outputs support the later coordinated recovery algorithms.

B. Error Covariance and Estimation Error

The estimation error is represented by an error vector and covariance matrix, whose trace gives the MMSE estimation error. These quantities are computed over dominant support sets and support the data-aided extension.

  • Error metrics: The trace of the error covariance matrix R_eh gives the MMSE estimation error.The conditional error vector and covariance are defined for a given support S.
  • Covariance construction: The overall error covariance R_eh is obtained by combining conditional covariance matrices R_eh|S over the dominant support set S_d.The conditional estimate is defined as the posterior mean for each support.
  • Covariance construction: Replacing the conditional posterior mean with a BLUE estimate yields a covariance expression involving the sensing matrix and noise covariance.The resulting expression is combined with the dominant-support formulation.
  • Computational cost: Although covariance calculation includes matrix inversion, the required inverses are already available from intermediate SABMP computations.The authors therefore state that covariance evaluation adds no additional computational burden.
  • Data-aided extension: The error covariance and estimation error are used to develop the data-aided approach.They provide the uncertainty information needed by that extension.

C. Finding Marginals

The marginal-finding procedure modifies SABMP to evaluate posterior probabilities for relevant support sets and infer tap-activation beliefs. These beliefs are then shared among neighboring antennas for coordinated channel recovery.

  • Marginal probabilities: Marginal activation probabilities are computed from the support posteriors p(S|Y) over relevant support sets.The marginal probabilities are not directly available from standard SABMP output.
  • Marginal probabilities: Only the 2^Tmax − 1 support sets formed from the T_max detected locations are treated as having significant posterior mass.The method assumes other support sets have very small posterior values.
  • Algorithm modification: SABMP is modified to output posteriors for all support sets in S_r^Tmax rather than only its original T_max supports.This provides the terms needed to compute marginal probabilities for detected nonzero locations.
  • Computational cost: The additional computation for marginalized posteriors is described as not significantly increasing complexity because SABMP intermediate information is reused.The authors illustrate this with the extra posterior required in the T_max = 3 example.
  • Coordinated recovery: In coordinated recovery, antennas exchange information so neighboring observations jointly improve estimates of most-significant-tap locations.The sharing strategy is implemented through processors associated with the receive antennas.
  • Information sharing: Repeated sharing diffuses information across antenna tiers, reaching four neighbors after one iteration and twelve antennas after two iterations.The number of incorporated antennas grows with repeated application of the sharing step.
  • Information sharing: The stepwise collaboration controls the number of collaborators and reduces communication overhead by avoiding simultaneous all-to-all collaboration.These properties support the proposed distributed estimation algorithms.

A. Algorithm 1: Marginal-based Channel Estimation using Pilots

Algorithm 1 estimates sparse channels locally, shares tap marginals among neighboring antennas, and uses the averaged beliefs as priors for final recovery. The sharing depth D is selected according to antenna support similarity and pilot-based recovery requirements.

  • Algorithm 1 procedure: Each antenna acts as a central node, collects marginals from its 4-neighbors, and computes an updated marginal for each tap.Undetected taps receive a small probability, reflecting their presumed near-zero gains.
  • Selecting D: For space-invariant arrays, D may equal max(M, G), allowing each antenna to receive information from every other antenna.This choice is motivated by support locations that do not vary across the array.
  • Selecting D: A noise-free unique-recovery condition relates the observation size K, channel sparsity n, participating antennas q, and sharing depth D.The number of antennas participating after D steps is 2D(D+1)+1, and the pilot count must also satisfy the stated recovery condition.
  • Selecting D: For space-variant arrays, increasing D can degrade estimates when support changes rapidly, whereas slowly changing support makes local collaboration beneficial.The selected neighborhood should contain antennas with approximately similar support.
  • Algorithm 1 procedure: Algorithm 1 initializes equal tap-active probabilities, runs RS1 independently, shares marginals with neighboring antennas, averages them, and repeats this process D times.The resulting marginals become new priors for final SABMP-based sparse CIR estimation.
  • Algorithm 1 procedure: After D sharing and averaging iterations, antennas use the new marginals as priors with SABMP to obtain a more accurate sparse CIR estimate.Information sharing strengthens beliefs about the locations of active taps.

B. Algorithm 2: Reduced Communication and Computational Cost – Integer-based Channel Estimation

Algorithm 2 replaces floating-point marginal exchange with integer scores and belief vectors, reducing communication and computation while retaining coordinated sparse-channel estimation. The data-aided extension further selects reliable carriers and combines their observations with pilots for refined recovery.

  • Integer-based channel estimation: Algorithm 2 ranks detected taps by estimated amplitude, assigns integer scores, averages scores across neighboring antennas, and repeats sharing for D iterations.Undetected tap locations receive score zero.
  • Integer-based channel estimation: Each antenna converts the shared scores into a belief vector and uses those beliefs as Bernoulli priors for final RS1 channel re-estimation.This preserves the coordinated-support mechanism while replacing marginal probabilities with integer-derived beliefs.
  • Integer-based channel estimation: Integer-based estimation reduces communication cost by avoiding floating-point exchanges and lowers computational complexity by avoiding marginal-probability calculations.These are the stated advantages over the marginal-based algorithm.
  • Data-aided channel estimation: Reliable carriers are selected using a metric that compares the probability of correct decoding with probabilities of incorrect constellation decisions.The metric accounts for combined channel-estimation error and noise through the distortion variable Z.
  • Data-aided channel estimation: The data-aided procedure intersects reliable-carrier sets across neighboring antennas, retains carriers with agreeing data, and combines them with pilots for SABMP recovery.Algorithm 3 first obtains CIR estimates, exchanges carrier indices and data, and then re-estimates the channel using the pruned set R⋆.

A. System Setup

The simulations evaluate sparse-channel recovery using unknown-statistics channels, multiple proposed estimators, and NMSE and BER metrics across pilot, algorithm, sparsity, and collaboration experiments. Results show successful recovery with few pilots, graceful degradation as channels become denser, and scenario-dependent benefits from information sharing.

  • Simulation setup: The simulations use sparse Rayleigh channels with unknown channel statistics and compare oracle-LS, BR, MB, and IB estimators, including pilot- and data-aided variants.The evaluation averages NMSE and BER over 100 trials, with zero-forcing equalization for recovered data.
  • Experiment 1 - How many pilots?: 12 pilots achieve a 100% recovery success rate for both MB-R and IB-R in SIA channels of length L = 64.Six pilots exceed 50% success; the 12-pilot requirement is 18.75% of the channel length under the stated experiment conditions.
  • Experiment 2 - Comparison between MB and IB: MB and IB perform equally well across the tested SIA and SVA configurations, while MB offers little advantage in the stated setting.MB can become more advantageous with fewer pilots, but at relatively high computational and communication cost.
  • Experiment 3 - Comparison with BR and oracle-LS: Nearly two orders of magnitude separate MB-R and BR in BER at SNR = 35 dB and Q = 4, with MB-R performing best among the compared algorithms.This comparison uses QAM orders 4 and 16, channel length L = 32, sparsity n = 3, K = 8 pilots, and the SIA case.
  • Experiment 4 - Effect of sparsity rate: The proposed IB-R and MB-R algorithms degrade gracefully as sparsity decreases, with reconstruction accuracy worsening as the number of non-zero taps increases.The experiment uses L = 64, n = 3, 5, and 7, corresponding to sparsity rates of 4.7%–11%, with K = 16 pilots.
  • Experiment 5 - Effect of D: Sharing across neighbor levels improves SIA BER, but gains become insignificant beyond D = 3, whereas higher D does not improve SVA estimation accuracy.For reduced pilots in SIA, additional neighbors may provide significant improvement beyond D = 3; in SVA, D = 1 or 2 may suffice.

VII. CONCLUSION AND FUTURE WORK

The paper presents collaborative estimation in which neighboring antennas help estimate many massive-MIMO channels. Its three algorithms perform well across different scenarios while using relatively few pilots.

  • Conclusion: Massive MIMO channel estimation can be performed collaboratively by sharing information among neighboring antennas.The paper presents three algorithms based on this collaborative method.
  • Conclusion: The proposed algorithms show good performance across different scenarios while using a relatively small number of pilots.

APPENDIX A CHANNEL MODELS

The paper discusses channel-modeling considerations for large antenna arrays and contrasts space-invariant and space-variant support behavior. Its experiments show that neighbor sharing helps in the former case but not the latter.

  • Array configurations: Uniform linear arrays are often impractical for large-scale antenna modeling, motivating consideration of two- and three-dimensional array configurations.Existing modeling efforts include tools and extensions for 2D and 3D antenna arrays.
  • Space-invariant arrays: Information sharing across neighbor levels improves channel-estimation accuracy in the space-invariant-array case.
  • Space-variant arrays: Information sharing does not improve channel-estimation accuracy in the space-variant-array case.
Loading 1409.4671v2…