PaperlayerView exact revision on arXiv

Source-linked AI summary

Subspace Estimation and Decomposition for Large Millimeter-Wave MIMO systems

Hadi Ghauch, Taejoon Kim, Mats Bengtsson, Mikael Skoglund

arXiv:1507.00287v2cs.ITmath.IT

TL;DR

Hybrid analog-digital mmWave MIMO still needs channel estimation and precoder/combiner design before its anticipated gains can be harnessed. The procedure uses BS-initiated echoing to estimate the dominant eigenvectors of H†H, and MS-initiated echoing to estimate those of HH†. Simulations show performance similar to the ideal fully digital perfect-CSI case using only a few RF chains.

  • Problem

    Hybrid analog-digital mmWave MIMO still needs channel estimation and precoder/combiner design before its anticipated gains can be harnessed.

    Source: §I · p. 1

  • Method

    The procedure uses BS-initiated echoing to estimate the dominant eigenvectors of H†H, and MS-initiated echoing to estimate those of HH†.

    Source: B. Arnoldi Iteration for Subspace… · p. 4

  • Results

    Simulations show performance similar to the ideal fully digital perfect-CSI case using only a few RF chains.

    Source: §VI · p. 13

  • Takeaways & Limitations

    The simulations indicate that this performance can be achieved with few RF chains, saving energy and cost over conventional MIMO systems.

    Source: §VI · p. 13

    The objective metric R is generally not an achievable rate for the system, so it provides an upper bound on achievable rates.

    Source: F. Discussion · p. 10

Abstract

from arXiv · show

Channel estimation and precoding in hybrid analog-digital millimeter-wave (mmWave) MIMO systems is a fundamental problem that has yet to be addressed, before any of the promised gains can be harnessed. For that matter, we propose a method (based on the well-known Arnoldi iteration) exploiting channel reciprocity in TDD systems and the sparsity of the channel's eigenmodes, to estimate the right (resp. left) singular subspaces of the channel, at the BS (resp. MS). We first describe the algorithm in the context of conventional MIMO systems, and derive bounds on the estimation error in the presence of distortions at both BS and MS. We later identify obstacles that hinder the application of such an algorithm to the hybrid analog-digital architecture, and address them individually. In view of fulfilling the constraints imposed by the hybrid analog-digital architecture, we further propose an iterative algorithm for subspace decomposition, whereby the above estimated subspaces, are approximated by a cascade of analog and digital precoder / combiner. Finally, we evaluate the performance of our scheme against the perfect CSI, fully digital case (i.e., an equivalent conventional MIMO system), and conclude that similar performance can be achieved, especially at medium-to-high SNR (where the performance gap is less than 5%), however, with a drastically lower number of RF chains (4 to 8 times less).

I. INTRODUCTION

The paper addresses channel estimation and hybrid precoder/combiner design for mmWave MIMO, where many antennas make RF-chain power and hardware cost a concern. It focuses on estimating dominant channel subspaces and approximating them with hybrid analog-digital processing, rather than directly maximizing the rate metric.

I. INTRODUCTION

  • Large antenna arrays can exploit abundant mmWave spectrum, but the RF chains needed to drive them increase power consumption and hardware cost.Hybrid processing moves part of the signal processing to analog phase-shifter networks.Source: §I · p. 1
  • Hybrid mmWave systems need channel estimation and corresponding precoder/combiner design, while classical MIMO training schemes are not applicable to this setting.The paper notes that its technique encompasses both beamforming and precoding, independent of the number of streams.Source: §I · p. 1

A. Signal Model

  • The signal model assumes TDD reciprocity, with constant-modulus analog precoders and combiners alongside digital processing.It considers a single-user system with d data streams and r RF chains, where d ≤ r.Source: A. Signal Model · p. 2

B. Motivation

  • With no channel information at either terminal, the paper estimates dominant subspaces and approximates them with hybrid precoder/combiner cascades, rather than directly optimizing the rate metric.The stated focus is subspace estimation and decomposition under hybrid-architecture constraints.Source: B. Motivation · p. 3

III. EIGENVALUE ALGORITHMS AND SUBSPACE ESTIMATION

The section contrasts low-dimensional subspace estimation with estimating the full channel, then describes a reciprocal, distributed Arnoldi method for estimating dominant channel modes. The method uses echoing and Ritz eigenpairs to recover the right and left singular subspaces.

A. Subspace Estimation vs. Channel Estimation

  • Subspace estimation targets a low-dimensional channel subspace rather than the entire channel, making it more communication-efficient when that subspace is small.The paper highlights large, low-rank systems such as mmWave channels as a particularly relevant setting.Source: A. Subspace Estimation vs. Channel… · p. 3

B. Arnoldi Iteration for Subspace Estimation

  • Echoing noise can violate the tridiagonal requirement of Lanczos, so the paper uses Arnoldi for its subspace-estimation process.The paper states that noise inherent to echoing makes Lanczos inapplicable in practice.Source: B. Arnoldi Iteration for Subspace… · p. 4
  • The distributed Arnoldi process exploits TDD reciprocity to estimate the d largest eigenvectors of H†H at the BS while accounting for distortion.The BS sends beamforming vectors and the MS echoes received signals in an amplify-and-forward manner to estimate H†H products.Source: B. Arnoldi Iteration for Subspace… · p. 4
  • Arnoldi builds a Krylov basis and reduced Hessenberg matrix; their Ritz eigenvectors yield estimates of the dominant right and left channel subspaces.The BS obtains the right subspace from H†H, and MS-initiated echoing can estimate the left subspace from HH†.Source: B. Arnoldi Iteration for Subspace… · pp. 3–4

C. Perturbation Analysis

The perturbation analysis relates the eigenpairs of the Arnoldi output to those of the channel-related matrix despite distortion. It bounds eigenvalue error, while noting that eigenvector sensitivity is more difficult to assess.

C. Perturbation Analysis

  • As m approaches M, the eigenpairs of C_m approximate those of A, but the BS cannot compute C_m because the distortion term is unknown.The method therefore uses the Arnoldi output T_m to approximate C_m and, in turn, A.Source: C. Perturbation Analysis · p. 5
  • The eigenvalue error between T_m and C_m is bounded by √m times the Frobenius norm of the distortion matrix.The bound is |λ~ − λ| ≤ √m ||W~_m||_F.Source: C. Perturbation Analysis · p. 5
  • Simulations indicate that good approximations can occur even when m ≪ M, although the analysis’s vanishing-error result is asymptotic in m.The stated eigenpair approximation becomes exact as m approaches M, up to round-off errors.Source: C. Perturbation Analysis · p. 5
  • The relevant estimation-error metric is distance between the true and estimated subspaces, while analogous eigenvector sensitivity analysis is harder because it depends on eigenvalue clustering.The estimate is formed from the Arnoldi basis and eigenvectors of T_m.Source: C. Perturbation Analysis · p. 5

IV. HYBRID ANALOG-DIGITAL PRECODING FOR MMWAVEMIMO SYSTEMS

The hybrid architecture requires estimated singular subspaces to be represented by cascaded analog and digital precoders or combiners. The section formulates this subspace decomposition problem and motivates an iterative method that respects the analog hardware constraint.

IV. HYBRID ANALOG-DIGITAL PRECODING FOR MMWAVE MIMO SYSTEMS

A. Preliminaries: Subspace Decomposition

  • For r ≥ 2d, a decomposition exists with zero cost, and the paper calls this an optimal decomposition.Earlier results cited in the passage establish existence of such a factorization.Source: A. Preliminaries: Subspace Decomposition · p. 5
  • The paper reports the same sum-rate for its decomposition using a subset of RF chains as for the cited schemes using all available RF chains.Source: A. Preliminaries: Subspace Decomposition · p. 6
  • The decomposition resembles dictionary-learning and related formulations, but the analog hardware constraint on F prevents using those existing tools directly.The paper therefore develops its own iterative approximate solution to the non-convex, non-separable problem.Source: A. Preliminaries: Subspace Decomposition · p. 6

1) Block Coordinate Descent for Subspace Decomposition:

BCD-SD alternates optimization of the analog and digital factors while enforcing the analog hardware constraint through projection. The approach uses fewer than all available RF chains and implicitly meets the power constraint.

1) Block Coordinate Descent for Subspace Decomposition:

  • The BCD approach implicitly enforces the power constraint, which can therefore be dropped without changing the problem.The resulting solution can be scaled to meet the power constraint with equality without loss of optimality for the stated performance metric.Source: 1) Block Coordinate Descent for… · p. 6
  • BCD-SD alternates updates of F and G while projecting each F solution onto the hardware-feasible set S.The method relaxes the hardware constraint during optimization and then applies the projection.Source: 1) Block Coordinate Descent for… · p. 6
  • Each analog-factor update projects onto S by setting every element’s magnitude to 1/√M.This gives the closest point in S to an arbitrary F under the stated Euclidean projection.Source: 1) Block Coordinate Descent for… · p. 6

2) One-dimensional case:

The paper gives a globally optimal one-dimensional subspace decomposition, compares decomposition methods, and adapts reciprocal echoing to hybrid hardware. The hybrid echo process remains distorted by analog processing and decomposition error.

2) One-dimensional case:

  • For d = 1, the decomposition has a globally optimal solution whose analog phases match those of the target vector.The stated approximation error at each element is the difference between its magnitude and ||γ~_1||_1/M, with the target phase retained.Source: 2) One-dimensional case · p. 7

3) Numerical Results:

  • The column-wise decomposition performs surprisingly well against OMP in the average subspace-distance comparison.The comparison uses M = 64 and r = 10 and averages over 500 random realizations; OMP is stopped after r iterations.Source: Figure 2 · p. 7, 3) Numerical Results · p. 7

B. Echoing in the Hybrid Analog-Digital Architecture

  • The BS approximates each echo vector as f~_l g~_l, and the MS reuses its analog combiner as the uplink precoder before the BS forms p_l.This emulates the amplify-and-forward step using operations possible in the hybrid architecture.Source: B. Echoing in the Hybrid… · pp. 7–8
  • Hybrid echoing does not yield a good estimate of H†Hq_l because analog processing distorts the effective channel and produces a low-dimensional observation, while decomposition error adds further distortion.The BS observation has dimension r, with r < M.Source: B. Echoing in the Hybrid… · p. 8

1) Cancellation of Analog-Processing Impairments:

RAID echoing uses repeated measurements and linear combining to suppress analog-processing impairments in the hybrid architecture. It yields a desired channel estimate with residual decomposition-induced distortions and incurs channel-use overhead.

1) Cancellation of Analog-Processing Impairments:

  • Each downlink echo uses N/r repetitions, and the uplink uses M/r repetitions, with measurements linearly combined to form the estimate.The analog combiners are drawn from DFT-matrix columns and reused for linear combining.Source: 1) Cancellation of Analog-Processing… · p. 8
  • RAID echoing suppresses impairments from analog precoding and combining, leaving the desired estimate corrupted by BS- and MS-side decomposition-induced distortions.The downlink repeats the approximated vector across measurements processed with DFT-derived analog combiners; the uplink uses a similar process.Source: 1) Cancellation of Analog-Processing… · p. 8
  • The procedure reduces the hybrid architecture to a conventional MIMO-like channel at a cost of Kr channel uses in the downlink, with the same applying to the uplink.The cited passage states the same cost for the uplink using its corresponding echoing relation.Source: 1) Cancellation of Analog-Processing… · p. 8
  • The overhead for each echoing is (M + N)/r channel uses and decreases as more RF chains are used.The passage says that d RF chains are active at the transmitting side while all r RF chains are used at the receiving side to minimize overhead.Source: 1) Cancellation of Analog-Processing… · pp. 8–9

2) Imperfect Compensation of Analog-Processing Impairments:

The section considers reducing RAID overhead with truncated DFT matrices, then presents hybrid-architecture algorithms that combine subspace estimation with analog-digital decomposition. It also describes a RAID-adapted two-way QR alternative.

2) Imperfect Compensation of Analog-Processing Impairments:

  • Truncating the DFT matrices can reduce the RAID overhead, but perfect analog-processing-impairment cancellation relies on measurements spanning the full square DFT matrices.The proposed truncated-matrix design is left to stochastic optimization, such as simulated annealing, which can be performed offline; further investigation is outside the work’s scope.Source: 2) Imperfect Compensation of… · p. 9

C. Proposed Algorithms

  • SED estimates the right and left singular subspaces with SE-ARN and then approximates them using cascades of analog and digital precoders or combiners.The subspace-estimation mechanism is decoupled from decomposition, so the latter can be substituted.Source: C. Proposed Algorithms · p. 9
  • Earlier PM and TQR algorithms are not directly applicable because they require the mobile station to amplify and forward received signals at its antennas.The paper describes adapting their echoing structure with RAID for hybrid analog-digital systems.Source: C. Proposed Algorithms · p. 9
  • MTQR adapts two-way QR to the hybrid architecture by using RAID, since conventional TQR assumes fully digital transmission.MTQR iteratively sends signals in downlink and uplink and applies QR to the received signals.Source: C. Proposed Algorithms · p. 9

D. Bounds on Eigenvalue Perturbation

The analysis relates eigenvalues produced by SE-ARN to dominant channel eigenvalues, while noting a pessimistic bound and practical stability concerns. The discussion frames the user-rate metric and specifies the sparse-channel simulation setting.

D. Bounds on Eigenvalue Perturbation

  • The eigenvalue analysis relates the eigenvalues of T_m to dominant eigenvalues of C_m and, asymptotically, of A, with good approximations reported even for m much smaller than M.The bound ignores the effect of DID compensation in RAID echoing and is therefore described as pessimistic.Source: D. Bounds on Eigenvalue Perturbation · p. 10

E. Practical Implementation Aspects

  • The work focuses on numerical investigation and simulations, rather than a stable, low-complexity implementation ready for practice.Floating-point errors can reduce Arnoldi-basis orthogonality, and echoing noise further amplifies the effect; restarted Arnoldi enhances stability but greatly increases overhead, making real-world implementation quite impractical.Source: E. Practical Implementation Aspects · p. 10

F. Discussion

  • The reported user-rate metric uses the actual channel with precoders and combiners designed from the estimated channel.The paper says its objective is generally not an achievable rate, but serves as an upper bound on achievable rates under partial CSI.Source: F. Discussion · p. 10

A. Simulation Setup

  • The simulations use a sparse mmWave channel model in which L scatterers contribute to the received signal, with uniform linear arrays and half-wavelength element spacing.The stated setup assumes M/r = 8 and N/r = 4.Source: A. Simulation Setup · pp. 10–11

1) Benchmarks/Upper bounds:

The evaluation compares the proposed schemes with benchmark and optimal performance, then examines robustness, scalability, and operating trade-offs. The discussion highlights low-SNR performance and differences in how the two proposed algorithms are analytically characterized.

B. Performance Evaluation

  • The proposed schemes perform close to the optimal bound, especially above −10 dB, while has a larger gap; optimal decomposition adds no gains.The comparison uses ACE and the fully digital perfect-CSI bound, with overhead matched to ACE; the tested overheads are 72 and 144 channel uses. Classical pilot estimation would require 2MN/r^2 total pilot slots in the described setting.Source: Figure 4 · p. 11, B. Performance Evaluation · p. 11, 1) Benchmarks/Upper bounds · p. 11
  • As the number of paths L increases, low-SNR performance degrades only modestly; simulations report no such degradation at medium-to-high SNR.The authors describe the approach as best suited to channels with few paths, such as mmWave channels.Source: Figure 5 · p. 11, B. Performance Evaluation · p. 12
  • Both SED and MTQR show better subspace-estimation accuracy as SNR increases, with similar behavior.Accuracy is evaluated using the average subspace angle between the channel subspace and its estimate.Source: Figure 6 · p. 11, B. Performance Evaluation · p. 12
  • With M/r and N/r fixed, Algorithm 1 closely follows the optimal bound as system dimensions grow, while retaining small overhead.For the 16 × 8 case the performance may not offset overhead, whereas it does for 256 × 128; the benchmark's complexity prevented results above 128 × 64.Source: B. Performance Evaluation · p. 12, Figure 7 · p. 12
  • In the SCM simulation, Algorithm 1 provides relatively high throughput at low overhead, especially for d = 3, although its gap from the benchmark increases with d.The setup uses 60 GHz, four paths, ULAs, and an urban-micro scenario.Source: B. Performance Evaluation · p. 12, Figure 8 · p. 12
  • Increasing Arnoldi iterations m improves estimation accuracy but raises communication overhead; simulations indicate 2d ≤ m ≤ 3d is a good trade-off.The marginal improvement from increasing m decreases.Source: B. Performance Evaluation · pp. 12–13

C. Discussions

  • At normalized overhead, SED and MTQR have similar performance profiles, but only SED's behavior is analytically characterized in this work.The authors describe MTQR as a heuristic variation on TQR whose behavior they have not modeled analytically.Source: B. Performance Evaluation · p. 12, C. Discussions · p. 13
  • The proposed schemes have poor low-SNR performance, as does ACE; the conclusion reports similar-to-ideal performance with few RF chains and lower energy and cost.The authors report that using more than r RF chains during RAID echoing can greatly boost low-SNR performance.Source: C. Discussions · p. 13, §VI · p. 13

APPENDIX

The appendix develops proof steps for the Arnoldi-based subspace-estimation bounds, including a matrix relation and an eigenvalue perturbation argument.

APPENDIX

  • Using the definitions of T̃_m and W̃_m in (9), the proof rewrites the relation in matrix form and states AQ_m − Q_m C_m = [T̃_m]_{m+1,m} q_{m+1} b†.Source: APPENDIX · p. 13
  • The stated limiting behavior as m approaches M makes the residual terms tend to zero in the proof's convergence step.The passage refers to both the Frobenius-norm residual and the subdiagonal entry [T̃_m]_{m+1,m}.Source: APPENDIX · p. 13

B. Proof of Lemma 2

  • Lemma 2's proof applies Bauer-Fike to bound perturbed eigenvalues, then uses the Frobenius norm to express the bound.The proof treats T_m as C_m plus a perturbation and notes that the spectral norm is no greater than the Frobenius norm.Source: B. Proof of Lemma 2 · p. 13

C. Proof of Lemma 3

The appendix gives an optimization argument for Lemma 3 and derives distortion-norm bounds for Corollary 1.

C. Proof of Lemma 3

  • The Lemma 3 proof restricts g to be real and nonnegative, then uses the structure of h_o to optimize f and g in turn for a globally optimal solution.After substituting the resulting f into the original problem, optimizing g is an unconstrained quadratic problem.Source: C. Proof of Lemma 3 · pp. 13–14

D. Proof of Corrollary 1

  • The Corollary 1 proof expresses W̃_m in terms of the error e_l(t), using the relation W̃w_l = −H†He_l(t).The derivation then combines norm properties with the bounds on the error terms.Source: D. Proof of Corrollary 1 · p. 14
  • The proof bounds the distortion terms by ∥E(t)∥_F ≤ 2√m and ∥E(r)∥_F ≤ √m(1 + d∥H∥_F).Source: D. Proof of Corrollary 1 · p. 14
Loading 1507.00287v2…