PaperlayerView exact revision on arXiv

Source-linked AI summary

Compressive channel estimation and tracking for large arrays in mm wave picocells

Zhinus Marzi, Dinesh Ramasamy, Upamanyu Madhow

arXiv:1506.05367v1cs.IT

TL;DR

The paper tackles sparse-channel estimation and tracking for large mm-wave RF-beamformed arrays, where standard least-squares methods cannot access individual antenna elements. It uses pseudorandom-phase compressive beacons and a sequential spatial-frequency estimator, achieving low overhead while addressing beacon-SNR and interference design.

  • Problem

    Large RF-beamformed mm-wave arrays make standard least-squares channel estimation inapplicable because individual antenna-element signals are inaccessible.

  • Method

    The paper uses pseudorandom-phase compressive beacons and mobile-array measurements, with sequential 2D spatial-frequency estimation that exploits channel continuity.

  • Results

    Less than 1% overhead is reported for compressive beaconing, while estimated channels yield less than 0.3 dB SNR loss with ideal beamforming and less than 1 dB with four-phase control.

  • Takeaways & Limitations

    Compressive measurements can super-resolve sparse mm-wave spatial channels with coarse phase-only RF beamforming and scale to very large arrays.

Abstract

from arXiv · show

We propose and investigate a compressive architecture for estimation and tracking of sparse spatial channels in millimeter (mm) wave picocellular networks. The base stations are equipped with antenna arrays with a large number of elements (which can fit within compact form factors because of the small carrier wavelength) and employ radio frequency (RF) beamforming, so that standard least squares adaptation techniques (which require access to individual antenna elements) are not applicable. We focus on the downlink, and show that "compressive beacons," transmitted using pseudorandom phase settings at the base station array, and compressively processed using pseudorandom phase settings at the mobile array, provide information sufficient for accurate estimation of the two-dimensional (2D) spatial frequencies associated with the directions of departure of the dominant rays from the base station, and the associated complex gains. This compressive approach is compatible with coarse phase-only control, and is based on a near-optimal sequential algorithm for frequency estimation which can exploit the geometric continuity of the channel across successive beaconing intervals to reduce the overhead to less than 1% even for very large (32 x 32) arrays. Compressive beaconing is essentially omnidirectional, and hence does not enjoy the SNR and spatial reuse benefits of beamforming obtained during data transmission. We therefore discuss system level design considerations for ensuring that the beacon SNR is sufficient for accurate channel estimation, and that inter-cell beacon interference is controlled by an appropriate reuse scheme.

I. INTRODUCTION

The paper addresses channel estimation and tracking for large mm-wave arrays using RF beamforming, where individual antenna-element access is unavailable. It proposes compressive measurements that exploit sparse spatial channels and reports low overhead with system-level considerations for beaconing.

  • Motivation: Large mm-wave arrays provide compact, highly directive form factors and support spatial reuse, but create new channel-estimation challenges.The small carrier wavelength enables many antenna elements in compact arrays, while the paper focuses on associated signal-processing and system-design issues.
  • Motivation: RF beamforming routes a common baseband signal through the array, so standard least squares techniques requiring per-element baseband access do not apply.The setting assumes fixed per-element amplitude and coarse four-phase control.
  • Contributions: The proposed architecture uses pseudorandom-phase compressive beacons, compressive mobile measurements, and feedback for estimating and tracking dominant paths.The base station uses the feedback to estimate and track paths to mobiles, with feedback potentially carried over LTE.
  • Contributions: A sequential 2D spatial-frequency estimation algorithm exploits geometric channel continuity across beaconing intervals to reduce measurement overhead.Directions of arrival and departure map to 2D spatial frequencies, and the algorithm is described as near-optimal in related work.
  • Background and related work: The channel is modeled as a small number of discrete rays with continuous delays and angles, enabling super-resolution beyond spatial bandwidth-based limits.The paper identifies measurement-based validation of this sparse ray model as an important future need.
  • System model: The parametric ray model is more efficient than directly estimating individual channel-matrix entries and can drastically reduce the required number of measurements.The simulations use dominant paths including line-of-sight and single-bounce reflections, while the algorithm discovers and tracks paths without assuming their number.

IV. COMPRESSIVE CHANNEL ESTIMATION

The paper replaces inaccessible per-antenna least-squares measurements with compressive sounding using randomized transmit and receive phase settings. These measurements form a virtual MIMO channel from which the spatial channel can be estimated despite coarse phase-only RF control.

  • Channel sounding: Compressive beacons use known transmit signals with independently randomized coarse phase weights.The transmit weights are selected from a small phase set such as {±1, ±j}.
  • Channel sounding: Each beacon is repeated across L measurements while the mobile uses randomized receive weights to create virtual receive antennas.The resulting responses sample each transmit beacon with multiple compressive receive settings.
  • Channel sounding: The measurements form an M × L virtual MIMO channel matrix V between virtual transmit and receive antennas.The virtual channel is constructed from the measured responses for all beacon and receive-weight pairs.
  • Channel sounding: The virtual channel matrix satisfies V = AHBT, linking compressive measurements to transmit and receive weight matrices.The matrix representation follows from the virtual-pair channel relation vi,j = ai^Hbj.
  • Channel sounding: The measurements include independent identically distributed measurement noise.Noise perturbs the observed compressive channel responses used for estimation.

B. Feedback strategies

Feedback conveys the information needed to estimate the base-station-side spatial channel through either the full measured virtual channel or an energy-preserving low-dimensional representation. The same estimation algorithm applies to both feedback forms.

  • Feedback strategies: The target parameters are the dominant path gains and spatial frequencies describing the channel seen from the base station.Each mobile feeds back information supporting tracking of these parameters.
  • Feedback strategies: Feedback can consist of the entire measured virtual channel matrix Y.This is the direct feedback strategy.
  • Feedback strategies: Alternatively, the receiver feeds back Q strongest left singular vectors of Y scaled by their singular values.This representation identifies the Q-dimensional column-space subspace with maximum energy.
  • Feedback strategies: The estimation algorithm operates with either the full virtual MIMO matrix Y or the dominant weighted left singular vectors D.Thus, the algorithm is shared across both feedback strategies.
  • Feedback strategies: Unknown receive weights remain usable because randomized receive settings make the per-path coefficients independent realizations with known expected squared magnitude.This permits estimation of path-gain magnitudes from the resulting coefficients.

A. Single path

For a single path, the estimator first searches a discrete oversampled 2D frequency grid and then refines the detected frequency and gains toward the maximum-likelihood solution. The procedure removes the detected path and retains the residual for subsequent multipath detection.

  • A. Single path: The single-path model writes each measurement as yk = hkx(ω) + zk.The measurements combine a frequency-dependent spatial response, path-specific gain, and noise.
  • A. Single path: For any candidate frequency, the gain estimates are obtained by least squares, yielding the maximum-likelihood frequency objective.The frequency estimate is then selected by maximizing the resulting criterion.
  • A. Single path: Detection evaluates an oversampled 2D FFT grid and selects the frequency maximizing the single-path objective.The grid is Φ = {(2πk/T, 2πl/T)} with T = RN1D,t.
  • A. Single path: The corresponding gains are estimated at the detected frequency, and the detected sinusoid is removed from the measured response.The resulting residual supports sequential detection when multiple paths are present.
  • A. Single path: A Newton-based refinement alternates frequency and gain updates after grid detection.The refinement uses local derivatives and residual updates for several iterations.

B. Multiple paths

The multipath algorithm sequentially detects new paths from residuals, refines all current paths in round-robin fashion, and stops when adding another path yields insufficient residual reduction. Prior frequency estimates initialize tracking when channel geometry changes little between rounds.

  • B. Multiple paths: For multiple paths, the algorithm detects a new sinusoid from the residual left by the currently estimated paths.The new path is added to the current parameter set before joint refinement.
  • B. Multiple paths: After each detection, all estimated paths are refined one by one in a repeated round-robin sequence.Each path is updated using residual measurements formed by excluding that path.
  • B. Multiple paths: The algorithm terminates when the reduction in total residual from a newly added path falls below τ.The simulations use τ = 30σ2 log (20Nt,1D).
  • B. Multiple paths: Sounding is frequent enough that spatial frequencies remain nearly unchanged between successive estimation cycles, even when path gains vary.This preserves the usefulness of prior angle-of-departure estimates during beamforming.
  • B. Multiple paths: Prior frequency estimates initialize the next round through a constructed matrix X and initial gain estimates before refinement and new-path search.This replaces empty initialization with continuity-based initialization.
  • B. Multiple paths: Weak paths from previous rounds are deleted using the stopping criterion when they are no longer viable.The paper gives blockage as an example of why previously estimated paths may become stale.

VI. PROTOCOL PARAMETER CHOICES

The protocol chooses transmitter beacons and receive measurements to approximately preserve the geometry of sparse spatial-frequency estimation. These choices determine how many compressive measurements are needed relative to full-array access.

  • The protocol selects M, L, and the minimum effective SNR to configure compressive channel estimation and sounding bandwidth.These parameters also determine how frequently the channel must be sounded.
  • Compressive measurements preserve the ML estimation cost structure when the measurement matrix preserves relevant pairwise geometry.The resulting estimation is similar to using all Nt,1D^2 measurements, apart from an SNR gain given by M.
  • A 2K-isometry criterion preserves the geometry of the spatial-frequency estimation problem for K-sparse channels.Random measurement matrices satisfy the relevant isometry with high probability when their dimensions scale with sparsity and logarithmic problem size.
  • 30 random transmitter beacons keep the measurement-energy ratio within [−5, 3] dB for estimating K = 4 paths with a 32 × 32 array.This approximates measuring all 1024 antenna elements individually.

B. Number of compressive receive measurements

The receive array uses a small number of carefully selected random projections to preserve path energy across candidate spatial frequencies. For a 4 × 4 receive array, about five projections limit the worst-case SNR degradation to 3 dB.

  • The receive measurements must avoid realizations in which all projections contribute negligibly to a path.Otherwise, the corresponding path provides insufficient information for estimating transmitter spatial frequencies.
  • For L = O(log R) = O(log Nr,1D), random receive projections preserve the norm of candidate array responses with high probability.This applies the JL lemma to the oversampled DFT-grid responses and the zero vector.
  • Around 5 carefully chosen projections limit SNR degradation to no greater than 3dB for a 4 × 4 receive array.The selected matrix is the best of 10^4 random instances.

C. SNR for successful estimation

Successful frequency estimation requires the effective path SNR to exceed a ZZB threshold. The threshold is nearly unchanged between 8 × 8 and 32 × 32 arrays, while sounding bandwidth controls the time overhead needed to reach it.

  • SNRth = 16.04dB for an 8 × 8 array and SNRth = 16.13dB for a 32 × 32 array.The threshold is defined as the SNR where the ZZB is within 0.1dB of the CRB.
  • The per-measurement noise variance scales with the number of isolated receive antennas because independent thermal-noise contributions are summed after phase weighting.With σ2e = N0Ws, the aggregate variance is σ2 = Nr,1D^2 × (N0Ws).
  • The effective SNR of the i-th sinusoid must exceed SNRth for successful estimation.The effective SNR is determined from the path energy collected across compressive measurements and the per-measurement noise variance.
  • The sounding bandwidth Ws determines the time overhead ML/Ws required to collect the measurements for a given path gain.The cell size supplies a lower bound on path gain used to choose Ws.

D. Sounding rate

The sounding rate is chosen so that spatial-frequency drift between sounding rounds stays within the beamforming-loss tolerance. The resulting rate depends on array size, user speed, distance, and wavelength.

  • The worst-case spatial-frequency change between rounds is 2πdvmax/fBRλ.It occurs when the user is at distance R on array boresight and moves along an array axis.
  • The sounding rate must be high enough that 2πdvmax/fBRλ ≤ π/Nt,1D.This condition ensures beamforming losses between sounding phases remain smaller than 3dB.

A. Transmit power

The protocol sets transmit and sounding parameters to provide adequate estimation SNR while minimizing sounding overhead for 8 × 8 and 32 × 32 arrays.

  • 40 dBm EIRP yields total transmit powers of 22 dBm for Nt,1D = 8 and 10 dBm for Nt,1D = 32.
  • Fixed per-element powers are 4 dBm and −20 dBm for Nt,1D = 8 and Nt,1D = 32, respectively, including beaconing.
  • The sounding protocol uses bandwidth Ws, beacon count M, measurements L, and sounding rate fB, with M, L, and Ws determining effective sounding SNR.
  • The effective-sounding requirement is that ML/Ws compensate for the absence of beamforming during sounding, with picocell range canceling from the overhead condition.
  • M = 24 for Nt,1D = 8 and M = 30 for Nt,1D = 32 are selected using a geometry-preservation criterion for spatial channel estimation.
  • With R = 20 m and vmax = 20 m/s, the selected sounding rates produce overheads of 0.0131% and 0.8542% for 8 × 8 and 32 × 32 arrays.The corresponding rates are fB ≥ 8 Hz and fB ≥ 32 Hz, with Ws = 8.8124 MHz and 674.34 KHz, respectively.

D. Reuse analysis for channel sounding

The reuse analysis models inter-cell interference during aligned channel sounding and selects reuse factors to maintain effective SIR while keeping aggregate sounding bandwidth small.

  • Aligned sounding rounds across basestations require spatial reuse of the 2 GHz spectrum to limit neighboring-picocell interference during channel estimation.
  • Interfering basestations using the same sounding bandwidth are modeled at distances {kRfS, k ∈ Z \ {0}} in a narrow urban canyon.
  • Four equal-power paths per interfering basestation are assumed, a pessimistic choice because NLOS paths have greater attenuation and reflection losses.
  • Effective per-element noise combines thermal noise N0Ws with interference I, yielding SINReff for channel estimation.
  • For SIReff > 26 dB, Rf = 4 is needed at S = 50 m, whereas Rf = 3 suffices at S = 200 m because oxygen absorption attenuates interference.
  • At S = 50 m, system-level sounding bandwidth is 35.2 MHz for 8 × 8 arrays and 2.7 MHz for 32 × 32 arrays, within 2 GHz.

VIII. SIMULATION RESULTS

Simulations evaluate channel tracking for six moving users using 8 × 8 and 32 × 32 arrays, comparing full-matrix and reduced singular-vector feedback. Dominant singular-vector feedback matches full feedback with one-third overhead, while estimated-path beamforming incurs limited SNR loss and larger arrays better suppress undesired taps.

  • Simulation setup: Six mobile users move through the urban canyon at speeds from 1.5 to 20 meters per second during the 7 second simulation interval.The scenario includes both vehicular and pedestrian settings, with users starting at positions marked in Figure 10.
  • Simulation setup: The simulations target K = 4 paths for six users and compare the proposed algorithm with full-matrix and dominant-singular-vector feedback.The full strategy feeds back Y, while the svd strategy feeds back two dominant left singular vectors scaled by their singular values.
  • Feedback comparison: One-third feedback overhead: dominant singular-vector feedback performs just as well as feeding back the entire matrix Y.Figure 11 reports CCDFs of estimation errors and PDFs of the estimated path count for 8 × 8 and 32 × 32 systems.
  • Beamforming performance: Less than 0.3 dB SNR loss occurs with ideal beamforming from estimated paths, compared with less than 1 dB under four-phase control for the 8 × 8 array.The 32 × 32 results are reported as entirely similar and are not plotted.
  • Channel impulse response: The 32 × 32 array attenuates an undesired channel tap to one-ninth of the desired path, whereas the 8 × 8 array fails to resolve the two nearby paths.The larger array has a 4° half-power beamwidth in the simulated setting.
  • Conclusions: The compressive approach is compatible with coarse phase-only RF beamforming and can super-resolve mm wave spatial channels using relatively few measurements.The conclusion also identifies comprehensive experimental validation and broader network coordination as future work.
Loading 1506.05367v1…