PaperlayerView exact revision on arXiv

Source-linked AI summary

A Tutorial on the Optimization of Amplify-and-Forward MIMO Relay Systems

Luca Sanguinetti, Antonio A. D'Amico, Yue Rong

arXiv:1303.2817v1cs.IT

TL;DR

MIMO relay optimization extends single-hop transceiver design to amplify-and-forward relay architectures with varied operating conditions. This tutorial reviews the principal optimization structures, practical algorithms, and unresolved challenges across these systems.

  • Problem

    Optimization of MIMO relay networks has received less comprehensive treatment than point-to-point MIMO systems, despite their promise for wireless reliability and coverage.

  • Method

    The tutorial synthesizes optimization results and implementation issues for AF MIMO relay systems across one-way, multi-hop, parallel-relay, direct-link, nonlinear, and robust settings.

  • Results

    The tutorial reports that one-way two-hop MIMO optimization is well understood under perfect channel knowledge, while robust solutions and several other architectures remain unresolved.

  • Takeaways & Limitations

    The reviewed systems motivate decentralized algorithms that could trade performance against scalability, alongside further work on channel uncertainty and unsolved architectures.

Abstract

from arXiv · show

The remarkable promise of multiple-input multiple-output (MIMO) wireless channels has motivated an intense research activity to characterize the theoretical and practical issues associated with the design of transmit (source) and receive (destination) processing matrices under different operating conditions. This activity was primarily focused on point-to-point (single-hop) communications but more recently there has been an extensive work on two-hop or multi-hop settings in which single or multiple relays are used to deliver the information from the source to the destination. The aim of this tutorial is to provide an up-to-date overview of the fundamental results and practical implementation issues of designing amplify-and-forward MIMO relay systems.

I. INTRODUCTION

The tutorial surveys optimization results and implementation issues for amplify-and-forward MIMO relay systems, emphasizing one-way two-hop architectures and extensions across operating conditions.

  • MIMO relay communications are presented as a promising approach for improving wireless-system reliability and coverage.
  • Amplify-and-forward uses nonregenerative relays that linearly process and retransmit received signals, trading noise propagation against implementation practicality.
  • The tutorial focuses on half-duplex systems because power consumption, implementation costs, and spatial efficiency make them appealing for wireless applications.
  • Its one-way two-hop analysis begins with linear architectures, frequency-flat channels, negligible direct links, and source and relay average-power constraints.
  • The tutorial also covers QoS-constrained total-power minimization, nonlinear architectures, frequency-selective channels, channel-state-information acquisition, and robust optimization.
  • The one-way two-hop model uses source processing U, relay processing F, and destination processing G across source-relay and relay-destination links.

B. Problem formulation

The section formulates two AF MIMO relay optimization problems: optimizing an objective under source and relay power budgets, or minimizing total power under stream-level MSE requirements.

  • B. Problem formulation: The general objective f is increasing in each MSE argument and is optimized over destination, source, and relay processing matrices.
  • B. Problem formulation: Problem P1 minimizes or maximizes f subject to average power constraints at the source and relay nodes.
  • B. Problem formulation: Problem P2 minimizes total power consumption while requiring each stream's MSE to remain below its specified QoS threshold η_k.
  • B. Problem formulation: The formulation normally assumes perfect knowledge of the source-relay and relay-destination channel matrices at all nodes.
  • C. Design of (G, U, F) for problem P1: For fixed source and relay processing, the optimal destination matrix is the Wiener filter because f increases with every diagonal MSE.
  • C. Design of (G, U, F) for problem P1: With the Wiener filter, each stream's SINR is related to its MSE, so P1 also encompasses criteria expressed through SINRs.
  • C. Design of (G, U, F) for problem P1: Closed-form U and F designs are available for additively Schur-concave or Schur-convex objectives, which cover diverse wireless-system criteria.
  • C. Design of (G, U, F) for problem P1: Examples include mutual-information maximization and product-of-MSE minimization for additively Schur-concave objectives.

1) Additively Schur-concave functions:

For additively Schur-concave objectives, the design aligns processing with channel singular vectors, diagonalizes the relay system into parallel SISO channels, and allocates power across them.

  • 1) Additively Schur-concave functions:: Mutual-information maximization and product-of-MSE minimization are examples of additively Schur-concave design criteria.
  • 1) Additively Schur-concave functions:: The optimal source and relay processing matrices use the singular vectors associated with the largest singular values of the channel matrices.
  • 1) Additively Schur-concave functions:: The coefficients A_k and B_k represent source and relay power required by stream k and are obtained by solving the associated allocation problem.
  • 1) Additively Schur-concave functions:: Matching corresponding channel singular vectors pairs the strongest spatial channels of the source-relay and relay-destination links.
  • 1) Additively Schur-concave functions:: The resulting overall channel and MSE matrices are diagonal, with their entries determined by the per-stream processing and channel quantities.
  • 1) Additively Schur-concave functions:: The optimized AF relay system is equivalent to parallel SISO channels when the unitary matrix S is selected as the identity.
  • 1) Additively Schur-concave functions:: Power allocation over the parallel channels is non-convex; alternating updates can monotonically converge to a local optimum when properly initialized.
  • 1) Additively Schur-concave functions:: A closed-form low-complexity approximation works properly only when source and relay power budgets sufficiently exceed the noise variance.

2) Additively Schur-convex functions:

Additively Schur-convex objectives yield diagonal relay structures up to a unitary rotation, with power allocation independent of the specific objective function. In uncoded 4-QAM simulations, minimizing maximum MSE outperforms the compared Schur-concave criteria and naive AF.

  • The considered additively Schur-convex objectives include minimizing maximum MSE, harmonic-mean SINR, and negative minimum SINR.
  • The power allocation problem is independent of the particular choice of objective function f.
  • The optimal relay structure is diagonal up to a unitary matrix S, chosen so the MSE diagonal entries equal the arithmetic mean of its eigenvalues.
  • In Fig. 3, minimizing maximum MSE outperforms mutual-information, product-SINR, and sum-MSE designs, while all outperform naive AF.The comparison uses 4-QAM with NS = NR = 3, K = 2, and relay-destination SNR fixed at 20 dB.

D. Design of (G, U, F) for problem P2

For QoS-constrained power minimization, the optimized structure remains diagonal up to a unitary matrix satisfying the target MSEs. Geometric-programming and reduced-complexity procedures are compared, with the latter requiring nearly minimum power.

  • The optimal matrices use a diagonal-up-to-unitary structure, with S selected so each MSE diagonal entry equals its QoS target ηk.
  • The power-allocation problem is non-convex, motivating geometric-programming bounds and a reduced-complexity algorithm.The geometric-programming and dual-decomposition solutions have relatively high computational complexity for practical implementation.
  • In Fig. 4, GP and RC are evaluated against the suboptimal SA design under equal QoS constraints and 0 dB noise variances.The setup uses NS = NR = K = 3 and η1 = η2 = η3 = η.
  • RC requires power very close to the minimum and substantially the same as GP, while both outperform the corresponding linear-receiver designs.
  • Practical systems may be unable to meet all QoS requirements because source and relay power resources or regulations limit transmit power.

E. Extension to non-linear architectures

For DFE receivers, multiplicatively Schur-concave objectives reduce the optimized nonlinear architecture to the linear one, whereas multiplicatively Schur-convex objectives produce a distinct design. The latter uses Wiener filtering and a structured backward matrix.

  • A DFE receiver introduces a strictly upper-triangular backward matrix B into the decision-device input model.Under correct previous decisions, y = (G^H U − B)s + Gn.
  • For multiplicatively Schur-concave objectives, the optimized nonlinear architecture reduces to the previously discussed linear architecture.
  • For multiplicatively Schur-convex objectives, the optimal receive filter is the Wiener filter and B has the form B = DL^H − I_K.Here L is lower triangular and D scales selected diagonal entries to unity.
  • The resulting multiplicatively Schur-convex design has equal MSE diagonal entries set to the geometric mean of the MSE eigenvalues.
  • The global optimization is non-convex, so the cited procedures obtain locally optimal solutions rather than guaranteed global optima.
  • In the reported BER experiment, DFE provides better performance than a linear receiver for the maximum-MSE design, but not when a multiplicatively Schur-concave function is used.

2) Design of (G, U, F, B) for problem P2:

QoS-constrained design for nonlinear relay architectures extends the optimization framework through non-convex power allocation, geometric programming, reduced-complexity methods, and frequency-selective multicarrier formulations. Cooperative subcarriers perform especially well in highly selective channels.

  • 2) Design of (G, U, F, B) for problem P2:: The QoS-constrained power-allocation variables {Ak} and {Bk} solve a problem subject to ordered eigenvalue conditions 0 < λk ≤ λk+1 ≤ 1.
  • 2) Design of (G, U, F, B) for problem P2:: Upper and lower bounds can be computed for the non-convex problem, while a reduced-complexity procedure offers an alternative.
  • 2) Design of (G, U, F, B) for problem P2:: GP-DFE and RC-DFE require substantially the same power and outperform the corresponding linear-receiver solutions across the investigated QoS values.
  • F. Extension to frequency selective fading channels: With cooperative subcarriers, each subproblem can be computed in closed form as in the linear flat-fading case.
  • F. Extension to frequency selective fading channels: Independent per-subcarrier processing can be decomposed into N subproblems controlled by a master problem when the objective functions increase in each argument.
  • F. Extension to frequency selective fading channels: Carrier cooperation performs better than noncooperative approaches, especially for highly frequency-selective channels.Frequency diversity provides additional degrees of freedom that cooperating subcarriers can exploit to improve system performance.

G. Acquisition of channel state information and robust optimization

The tutorial examines channel-state acquisition and robust transceiver optimization for AF MIMO relay systems, including distributed information sharing and statistical treatment of channel-estimation errors.

  • Channel-state acquisition: Closed-loop designs estimate HSR at the relay and HRD at the destination, then share channel information for processing-matrix computation.Distributed algorithms use only locally available channel state information, whereas centralized processing requires all propagation-channel information at one computation node.
  • Channel-state acquisition: Perfect or partial propagation-channel knowledge is assumed for the tutorial’s transceiver-design problems.The paper notes that space-time coding is outside its detailed scope when channel state information is unavailable.
  • Robust optimization: Using estimated channels directly in place of true channels causes substantial performance degradation relative to perfect channel knowledge.Robust optimization instead incorporates channel-estimation errors into the design.
  • Robust optimization: The robust error model uses Gaussian-Kronecker channel uncertainty with covariance expressions depending on the channel-estimation algorithm.The model includes Bayesian MMSE estimators under the stated covariance assumptions.
  • Robust optimization: For scaled-identity error covariances, the Wiener filter is optimal for G, while explicit U and F structures are available for additively Schur-concave objectives.The diagonal power-loading matrices can be obtained using an iterative method; Schur-convex objectives require a unitary choice equalizing the MSE diagonal.
  • Robust optimization: Robust designs outperform schemes that simply substitute estimated channel matrices for the true matrices without modeling estimation errors.The robust formulation computes statistical expectations with respect to both channel distributions and estimation-error distributions.

III. OPTIMIZATION OF A ONE-WAY MULTI-HOP MIMO

For one-way multi-hop MIMO systems, the tutorial reviews signal modeling and optimization across multiple relay hops. Despite coupled power constraints, optimal processing retains a channel-diagonalizing form, while practical implementation can require centralized computation or simplified local algorithms.

  • System model: A linear L-hop system uses one source, L−1 relays, channel matrices Hi, and processing matrices F0 and Fi across the hops.The hop recursion is ri = Hixi−1 + ni, with x0 = F0s and relay-transmitted vectors generated by the relay processing matrices.
  • Existing results: Prior work derives asymptotic capacity, capacity scaling, and diversity results for multi-hop systems under restricted relay-processing structures.These results include scaled-identity and diagonal relay matrices, as well as optimization of relay matrices under simplified assumptions.
  • Optimization: The first joint design minimizes additively Schur-concave or Schur-convex objectives under average source and relay power constraints.Unlike the two-hop case, each relay constraint depends on processing matrices at all preceding nodes.
  • Optimization: The optimal multi-hop solution has the same channel-diagonalizing structure as the two-hop solution for any arbitrary number L of hops.G is the Wiener filter, while source and relay processing matrices use SVD-derived factors and diagonal power loading.
  • Practical issues: Implementing the optimized system is challenging because centralized processing is required to compute the diagonal power-loading matrices.Simplified algorithms optimize relay matrices locally while maintaining comparable performance by decomposing the overall MSE into relay-level contributions.
  • Practical issues: For long source-destination distances requiring many hops, noise propagation can make practical data recovery nearly impossible at typical SNRs.A combination of regenerative and non-regenerative relays is proposed as a trade-off between end-to-end delay and error-rate performance.

IV. OPTIMIZATION OF A ONE-WAY TWO-HOP MIMO SYSTEM WITH MULTIPLE (PARALLEL) RELAYS

The tutorial reviews one-way two-hop MIMO systems with multiple parallel relays, modeling their stacked channels and block-diagonal relay processing. Joint optimization can diagonalize the overall channel, but unitary design and non-convex power allocation remain difficult.

  • System model: With Q parallel relays, the source-relay and relay-destination channels are stacked, while the relay processing matrix is block diagonal.Each relay contributes channel matrices HSRq and HRqD, and F = diag{F1, F2, …, FQ}.
  • Existing optimization results: Optimizing only the relay matrix has closed-form solutions for single-antenna relays, while the multiple-antenna case requires different treatment.Related designs also consider total power consumption and MSE objectives under relay power constraints.
  • Joint optimization: The overall channel is diagonalized up to an arbitrary unitary matrix P, but its optimal structure is unknown.The diagonal power-loading matrices require solving a non-convex power-allocation problem.
  • Numerical results: Increasing the number of relays improves BER in the illustrated 4-QAM system as source-relay-link SNR varies.The figure compares Q = 2, 3, and 5 with NS = NR = K = 3 and relay-destination-link SNR fixed at 20 dB.
  • Extensions: The channel-diagonalizing structure remains optimal when a decision-feedback equalizer is used at the destination, provided source and destination unitary rotations are included.The source rotates transmitted symbols and the destination counter-rotates received symbols.
  • Extensions: MSE-based QoS constraints can be used to optimize power consumption with either optimal or non-optimal source precoding.This extends the optimization objectives beyond the relay-processing design alone.

V. OPTIMIZATION OF A ONE-WAY TWO-HOP MIMO

With a direct link, one-way two-hop MIMO optimization must jointly account for source, relay, and destination processing, but closed-form solutions remain limited. Existing approaches use Wiener filtering, structured relay beamforming, and numerical optimization, with direct-link processing providing some BER improvement.

  • Optimization with the direct link: Jointly designing (G, U, F) with a direct link remains extremely challenging, so prior work often fixes G and U or optimizes only subsets of the matrices.Some formulations use mutual-information bounds or omit relay power constraints, while others set U equal to the identity matrix.
  • Optimization with the direct link: The optimal receiver G is the Wiener filter in the considered sum-MSE formulations.This result appears in both the design using a generic increasing objective and the related joint-design formulation.
  • Relay processing structure: The optimal relay matrix F is matched to the singular vectors of the source-relay and relay-destination channels, but its arbitrary matrix A is generally non-diagonal.Consequently, the optimized structure does not diagonalize the overall communication system; closed-form A is available only for a single source antenna, while multiple-antenna cases use numerical methods.
  • Numerical solutions and performance: Gradient-based numerical methods are used when the joint design lacks a closed-form solution.The cited algorithm optimizes the remaining design variables numerically and evaluates BER versus source-relay-link SNR under fixed relay-destination and source-destination SNRs.
  • Numerical solutions and performance: Accounting for the direct link achieves some BER improvement relative to the no-operation-and-forwarding baseline and the solution that neglects the direct link.The comparison is made for a one-way two-hop MIMO system with a 4-QAM constellation.

VI. OPTIMIZATION OF A TWO-WAY TWO-HOP MIMO

Two-way two-hop relaying reduces the channel-resource penalty of conventional four-phase one-way relaying by allowing both users to transmit simultaneously. The resulting optimization exploits self-interference cancellation and structured relay processing, but the joint design remains nonconvex and generally requires iterative methods.

  • Protocol and signal model: Two-way relaying lets the two users transmit simultaneously, reducing the resource penalty of four-phase one-way relaying and achieving the same spectral efficiency as direct communications.The relay forwards the jointly received signal in a second phase, while users remove their self-interference when the required channel knowledge is available.
  • Protocol and signal model: Self-interference cancellation requires each user to know its relay-to-user channel multiplied by the corresponding user-to-relay channel and relay matrix.After cancellation, each user detects the other user's signal through an effective channel.
  • Relay optimization: When NR ≥ 2NS, the optimal relay matrix for achievable-sum-rate maximization has a structured form based on semi-unitary matrices from two QR decompositions.The remaining factor is an arbitrary matrix A ∈ C2NS×2NS.

R1 HH

The tutorial identifies robust design under channel uncertainty, decentralized scalable algorithms, and broader relay architectures as important unresolved directions. Two-way systems additionally depend critically on channel knowledge for both matrix design and self-cancellation.

  • Open optimization issues: The two-way optimization problem remains difficult because its processing-matrix optimization is nonconvex and lacks a closed-form optimal structure.Iterative alternating optimization is therefore used for A and (G, U).
  • Channel knowledge: Channel knowledge is required not only to design the processing matrices but also to perform self-cancellation in two-way relay systems.This makes channel estimation and the delivery of channel state information central implementation concerns.
  • Future directions: The tutorial concludes that robust solutions for channel uncertainties and decentralized algorithms balancing performance with scalability remain important open problems.It also identifies unresolved optimization aspects in architectures beyond the well-understood perfect-channel one-way two-hop case.
Loading 1303.2817v1…