PaperlayerView exact revision on arXiv

Source-linked AI summary

Sparse Activity Detection for Massive Connectivity

Zhilin Chen, Foad Sohrabi, Wei Yu

arXiv:1801.05873v1cs.IT

TL;DR

Massive connectivity requires detecting sporadic user activity and estimating channels despite non-orthogonal signatures. The paper formulates this joint problem as compressed sensing and solves it with AMP designed around wireless-channel statistics, with state evolution predicting false-alarm and missed-detection probabilities; simulations show gains from statistical channel information and multiple antennas, while exact large-scale fading knowledge offers negligible improvement.

  • Problem

    Massive connectivity requires accurate activity detection and channel estimation for many potential devices with sporadic traffic and non-orthogonal signatures.

  • Method

    The paper formulates joint detection and estimation as SMV or MMV compressed sensing and develops AMP with channel-statistics-based MMSE denoisers and state-evolution analysis.

  • Results

    Simulations validate the analysis, show significant gains from exploiting channel statistics and multiple antennas, and find negligible improvement from exact large-scale fading knowledge.

  • Takeaways & Limitations

    Compressed sensing with AMP is a viable strategy for sporadic activity detection in massive connectivity using random non-orthogonal signatures.

Abstract

from arXiv · show

This paper considers the massive connectivity application in which a large number of potential devices communicate with a base-station (BS) in a sporadic fashion. The detection of device activity pattern together with the estimation of the channel are central problems in such a scenario. Due to the large number of potential devices in the network, the devices need to be assigned non-orthogonal signature sequences. The main objective of this paper is to show that by using random signature sequences and by exploiting sparsity in the user activity pattern, the joint user detection and channel estimation problem can be formulated as a compressed sensing single measurement vector (SMV) problem or multiple measurement vector (MMV) problem, depending on whether the BS has a single antenna or multiple antennas, and be efficiently solved using an approximate message passing (AMP) algorithm. This paper proposes an AMP algorithm design that exploits the statistics of the wireless channel and provides an analytical characterization of the probabilities of false alarm and missed detection by using the state evolution. We consider two cases depending on whether the large-scale component of the channel fading is known at the BS and design the minimum mean squared error (MMSE) denoiser for AMP according to the channel statistics. Simulation results demonstrate the substantial advantage of exploiting the statistical channel information in AMP design; however, knowing the large-scale fading component does not offer tangible benefits. For the multiple-antenna case, we employ two different AMP algorithms, namely the AMP with vector denoiser and the parallel AMP-MMV, and quantify the benefit of deploying multiple antennas at the BS.

I. INTRODUCTION

Massive connectivity requires detecting sporadically active users and estimating their channels despite non-orthogonal signatures. The paper formulates this as sparse compressed sensing and develops AMP methods with channel-statistics-aware analysis for single- and multiple-antenna base stations.

  • Networks may contain 10^4 to 10^6 devices, while only a small fraction are active at any given time.
  • Because potential users outnumber available pilot dimensions, signatures cannot be mutually orthogonal and create multi-user interference.
  • The paper models activity detection and channel estimation as compressed-sensing SMV or MMV problems and solves them with computationally efficient AMP.
  • Its AMP design exploits wireless-channel statistics and uses state evolution to characterize false-alarm and missed-detection probabilities.
  • For multiple antennas, the study evaluates vector-denoiser AMP and parallel AMP-MMV, while single-antenna results show exact large-scale fading offers little improvement over its statistics.

III. AMP ALGORITHM

AMP iteratively recovers sparse signals by transforming matched-filtered observations into a signal-plus-Gaussian-noise model and applying designed denoisers. State evolution tracks its asymptotic per-coordinate performance for SMV and MMV settings.

  • AMP is an iterative compressed-sensing algorithm that recovers sparse signals for both SMV and MMV problems.
  • At each iteration, AMP uses a denoiser and an Onsager correction term to update the estimate and residual.
  • The matched-filtered output is modeled as the signal plus Gaussian noise, including multi-user interference, enabling denoiser-based error reduction.
  • State evolution analyzes AMP as L, N →∞ with fixed N/L and predicts per-coordinate performance at each iteration.
  • For MMV recovery, a vector denoiser operates on each row of the matched-filtered output, with Bayesian design available to minimize estimation error.

2) Parallel AMP-MMV:

Parallel AMP-MMV decomposes the multiple-antenna problem into per-antenna AMP-SMV stages and exchanges soft activity information across antennas. The paper combines this distributed structure with channel-statistics-aware denoising and analyzes its modeling assumptions.

  • Parallel AMP-MMV solves the MMV problem iteratively using multiple parallel AMP-SMV stages that exchange soft information.
  • Its per-antenna operation can support distributed implementation, which is computationally advantageous when the number of antennas is large.
  • Each stage receives beliefs about user activity, applies conventional AMP, and returns channel-gain information to refine those beliefs.
  • The denoiser depends on activity beliefs shared across all AMP-SMV stages and is indexed by outer iteration and antenna stage.
  • The design considers channel distributions with either unknown large-scale fading statistics or known user-specific large-scale fading.
  • For large-scale fading, the model derives distributions from path loss, shadowing, device locations, and Rayleigh fading before designing MMSE denoisers.

1) With Statistical Knowledge of Large-Scale Fading Only:

When only the distribution of large-scale fading is known, AMP uses a common MMSE denoiser derived from the channel prior. The denoiser shrinks weak inputs toward zero and substantially outperforms soft thresholding in the reported simulations.

  • With only statistical large-scale-fading knowledge, the denoiser is identical for every matched-filtered entry because user-specific fading values are unavailable.
  • The MMSE denoiser is defined as the conditional expectation under AMP’s signal-plus-noise model.
  • The iteration-dependent noise state τ_t can be estimated empirically, and the denoiser can be precomputed as a lookup table without adding runtime complexity.
  • The MMSE denoiser shrinks small inputs toward the origin, thereby promoting sparsity similarly to soft thresholding.
  • The known-fading formulation adds g_n to emphasize dependence on user-specific prior information.

B. User Activity Detection

After AMP converges, user activity is detected with a likelihood-ratio test. The paper derives detection rules and false-alarm and missed-detection probabilities for statistical or exact large-scale-fading knowledge.

  • Detection rule: The likelihood-ratio test detects activity after AMP convergence using a threshold determined by the decision cost.The log-likelihood ratio is monotonic in the magnitude of the AMP estimate, so detection can use that magnitude alone.
  • Performance metrics: The false-alarm probability is the probability of declaring an inactive device active, while missed detection is declaring an active device inactive.These are the two performance metrics used to characterize the detector.
  • Statistical fading knowledge: With only statistical large-scale-fading knowledge, the derived false-alarm and missed-detection probabilities are averaged and do not depend on the individual fading coefficient g_n.The corresponding likelihood model uses the distribution of the large-scale fading rather than its exact realization.
  • Exact fading knowledge: With exact g_n known, the channel coefficient distribution becomes Bernoulli-Gaussian and the likelihood ratio explicitly depends on that prior information.Detection still uses the magnitude of the AMP estimate and a threshold l_n.
  • Threshold design: A common target false-alarm probability can be enforced by designing the threshold l_n as a function of each user’s known large-scale fading.The paper also notes that users could instead receive individually optimized thresholds based on their own cost functions.

C. State Evolution Analysis

State evolution tracks the AMP residual variance and thereby predicts denoiser MSE and detection-error probabilities across iterations. The analysis compares statistical versus exact large-scale-fading knowledge.

  • State evolution: State evolution determines the residual-noise standard deviation τ_t, which converges to a fixed point τ_∞ as AMP converges.The resulting τ_t values allow false-alarm and missed-detection probabilities to be evaluated at each iteration.
  • MSE analysis: The MSE of the MMSE denoiser is expressed through conditional variance, with separate characterizations for statistical and exact large-scale-fading knowledge.The paper gives propositions for both denoiser cases.
  • MSE decomposition: The state-evolution MSE separates estimation error under known activity from the additional error caused by unknown activity.This decomposition identifies the cost of jointly detecting activity and estimating the channel.
  • Impact of fading knowledge: Knowing the large-scale fading can improve estimation in theory, but simulations show only minor improvement for the fading model considered.The paper concludes that exact g_n is not crucial for activity detection; statistical fading information is sufficient.

V. USER ACTIVITY DETECTION: MULTIPLE-ANTENNA CASE

For multiple antennas, the paper develops channel-statistics-aware AMP denoisers and likelihood-based activity detection for vector observations. It derives error probabilities for unknown and known large-scale fading.

  • MMSE denoisers: The multiple-antenna formulation models each user’s channel as a row vector and designs MMSE denoisers for statistical or exact large-scale-fading knowledge.The vector denoiser uses conditional expectations under the corresponding row-vector distributions.
  • Covariance tracking: The covariance matrix Σ_t is tracked by state evolution and remains diagonal with identical diagonal entries when initialized in that form.The initial covariance after matched filtering has this structure, enabling simplified denoisers.
  • Activity detection: After AMP convergence, likelihood-ratio tests use mixed-Gaussian likelihoods to detect activity under unknown or known large-scale fading.The detector’s false-alarm and missed-detection probabilities are derived from these likelihoods.
  • Threshold design: The threshold l_n controls the trade-off between false-alarm and missed-detection probabilities, and the single-antenna thresholding strategy extends to multiple antennas.This applies to both large-scale-fading knowledge cases.
  • Performance characterization: The multiple-antenna error probabilities reduce to the single-antenna formulas when M = 1.The derivations use the chi-square distribution with 2M degrees of freedom for the vector-observation magnitude.

B. User Activity Detection by Parallel AMP-MMV

Parallel AMP-MMV divides the multiple-measurement-vector problem into antenna-wise AMP-SMV problems and exchanges soft activity information. The approach supports parallel computation but complicates analytic performance prediction.

  • Algorithm: Parallel AMP-MMV solves M parallel AMP-SMV problems and exchanges soft estimates of user activity probabilities between antennas.Its scalar denoiser uses improved activity-probability estimates instead of the fixed prior λ.
  • Activity detection: The algorithm performs likelihood-ratio activity detection after the parallel AMP iterations terminate.Its LLR has a form similar to the vector-denoiser case, yielding analogous false-alarm and missed-detection expressions.
  • Performance analysis: Deriving an analytic state evolution for τ^2_T,I is difficult because the antennas exchange soft information.Complete analytic performance prediction therefore requires determining a parameter whose evolution is not straightforward.
  • Empirical comparison: Numerical experiments show that parallel AMP-MMV performs very similarly to AMP with a vector denoiser.This observation suggests that its final residual-variance parameter should be close to the vector-denoiser result.
  • Complexity: Both AMP implementations have per-iteration complexity O(NLM), while parallel AMP-MMV additionally permits parallel computation across the antenna-wise subproblems.The computational advantage comes from dividing the MMV problem into several SMV problems.

VI. SIMULATION RESULTS

Simulations show that AMP’s predicted activity-detection performance closely matches experiments, while channel-statistics-aware MMSE denoising improves detection and multiple antennas reduce the pilot length needed for near-zero errors.

  • Simulation setup: The evaluation uses a cell with R = 1000m, N = 4000 potential users, 200 active users, and λ = 0.05.The channel parameters are α = 15.3, β = 37.6, and σSF = 8, with background noise of −169dBm/Hz over 10MHz.
  • Single-antenna detection: Predicted false-alarm and missed-detection probabilities closely match simulations for AMP with an MMSE denoiser using only statistical large-scale-fading knowledge.With L = 800 and transmit powers of 5dBm, 15dBm, and 25dBm, the lower bounds are close to actual performance; residual error is dominated by background noise.
  • Single-antenna detection: Knowing exact large-scale fading coefficients provides negligible improvement over knowing only their distribution.The predicted curves with exact coefficients closely match simulations and are nearly the same as those obtained with statistical knowledge alone.
  • Denoiser comparison: AMP with an MMSE denoiser outperforms both CoSaMP and AMP with a soft-thresholding denoiser.The comparison attributes this advantage partly to MMSE denoising’s use of statistical knowledge that the other methods do not exploit; soft-thresholding implicitly solves LASSO.
  • Denoiser comparison: The minimum pilot length for driving PF and PM to zero is between 300 and 400 for MMSE denoising, versus 600 and 800 for soft thresholding.This comparison demonstrates the advantage of accounting for channel statistics in detector design.
  • Multiple-antenna detection: AMP with a vector denoiser and parallel AMP-MMV achieve approximately the same performance across antenna counts.For the vector denoiser, simulated results closely match the predicted results.
  • Multiple-antenna detection: Increasing pilot length L or antenna count M can drive PF and PM to zero as transmit power increases.For L = 300 and M = 1, both probabilities remain unchanged with increasing power; increasing either L or M improves performance, reducing the minimum required L.

VII. CONCLUSION

The paper concludes that AMP with channel-statistics-aware denoisers supports sparse activity detection with random non-orthogonal signatures, while multiple antennas improve performance.

  • Compressed sensing is viable for sporadic device activity detection in massive connectivity with random non-orthogonal signature sequences.
  • The proposed AMP framework covers single- and multiple-antenna base stations and uses state evolution to predict false-alarm and missed-detection probabilities.
  • Channel-statistics-aware AMP denoisers significantly improve the detection threshold.
  • Knowing the exact large-scale fading coefficient provides no substantial improvement over using only its statistical information.
  • Deploying multiple antennas at the base station brings significant detection-performance improvement.

APPENDIX

The appendix derives channel distributions and the conditional moments required for AMP’s MMSE denoisers, covering both known and unknown large-scale fading.

  • The appendix defines distance, shadowing, path-loss, and large-scale fading variables, then derives their probability density functions.Large-scale fading is defined as g ≜ yz.
  • For a uniformly distributed user location in a circular cell, the distance density is derived from the cell radius and location assumption.
  • The channel coefficient combining large-scale and Rayleigh fading is modeled as h = gx, and its density is obtained using a change of variables and Jacobian determinant.When g is known, the conditional channel distribution is complex Gaussian with zero mean and variance g^2.
  • The conditional expectation and variance of the channel coefficient are derived from these densities for use in the MMSE denoiser and MSE calculation.The derivations substitute the observation density into the conditional-moment expressions and integrate the resulting variance.
  • For unknown large-scale fading, the appendix derives the observation density by separating inactive-user noise from active-user channel-plus-noise distributions.The active observation is represented as Y ≜ H + W, while the inactive observation is W ≜ τV.
  • For known large-scale fading, analogous conditional moments are derived by conditioning on g and averaging over its distribution when required.

G. Proof of Proposition 8

The proof derives the conditional covariance structure for the multiple-antenna MMSE denoiser and simplifies state evolution using symmetry across antenna dimensions.

  • For unknown large-scale fading, the proof extends the scalar derivation to random vectors and obtains the required conditional distributions and moments.The known-fading case can instead use a conditional-expectation result from prior work.
  • The proof assumes an induction form for the state covariance and derives the distribution of the effective observation conditioned on large-scale fading.
  • The conditional covariance matrix is evaluated by integrating over the effective observation and conditioning on the large-scale fading coefficient.
  • The covariance matrix is diagonal because off-diagonal integrals vanish as odd functions over symmetric domains.
  • Its diagonal elements are identical, so the expected covariance has the form of a scalar multiple of the identity matrix.
  • Substituting this covariance structure into the AMP state-evolution recursion yields the simplified recursion and completes the induction.
Loading 1801.05873v1…