PaperlayerView exact revision on arXiv

Source-linked AI summary

Massive MIMO Unsourced Random Access

Alexander Fengler, Giuseppe Caire, Peter Jung, Saeid Haghighatshoar

arXiv:1901.00828v1cs.ITmath.IT

TL;DR

Unsourced massive random access must decode sporadic transmissions from many users sharing a codebook, without relying on dedicated user pilots or known channel state. The paper combines concatenated coding with non-Bayesian non-coherent activity detection. It reports better performance than previous approaches and shows that required transmit Eb/N0 for reliable communication can become arbitrarily small as the number of receiver antennas grows.

  • Problem

    In unsourced massive access, dedicated user pilots and known channel information do not fit the shared-codebook setting, while practical coding schemes still face a large gap from random-coding achievability.

    Source: §I · pp. 1–2

  • Method

    The scheme combines an outer tree code with an inner non-Bayesian detector that treats activity and large-scale pathloss as a deterministic unknown sparse non-negative vector.

    Source: A. Our Contribution · p. 2, §III · p. 3

  • Results

    The proposed scheme performs substantially better than previous approaches, and reliable communication can require arbitrarily small transmit Eb/N0 as receiver antennas grow sufficiently large.

    Source: A. Our Contribution · p. 3

  • Takeaways & Limitations

    The authors present a practical coding solution for non-coherent massive-MIMO unsourced access whose required energy per bit can decrease without bound as antenna count grows.

    Source: A. Our Contribution · p. 3

    The inner-code reconstruction bound is conservative, and its universal constant κ is difficult to quantify sharply.

    Source: A. Inner Code · p. 6

Abstract

from arXiv · show

We consider an extension of the massive unsourced random access originally proposed by Polyanskiy to the case where the receiver has a very large number of antennas (a massive MIMO base station) and no channel state information is given to the receiver (fully non-coherent detection). Our coding approach borrows the concatenated coding idea from Amalladinne et. al., combined with a novel non-Bayesian `activity detection' algorithm for massive MIMO random access channels, that outperforms currently proposed Bayesian vector AMP (VAMP) schemes currently proposed for activity detection, and does not suffer from the numerical instabilities and requirement for accurate a priori statistics as VAMP. We show that the required transmit $E_b/N_0$ for reliable communication can be made arbitrarily small as the number of receiver antennas M grows sufficiently large.

I. INTRODUCTION

Unsourced massive random access targets sporadic IoT transmissions from many potential devices using a common codebook, with the receiver decoding active messages without identifying their senders. The section motivates grant-free simultaneous detection and decoding, and explains why existing coherent and pilot-based approaches do not directly fit this setting.

I. INTRODUCTION

  • The receiver must decode active messages up to permutations because users share one codebook and their identities are not known a priori.The setting is motivated by many objects sending sporadic data to a central collector.Source: §I · p. 1
  • Polyanskiy models the channel as a real adder with AWGN and uses per-user probability of error because conventional multiple-access error probability precludes reliable communication in the stated regime.Source: §I · p. 1
  • Grant-based access and dedicated pilots conflict with the common-codebook paradigm and add overhead, motivating simultaneous detection and decoding in one communication phase.Source: §I · pp. 1–2
  • Explicit coding schemes matter because the gap between random-coding achievability and known random-access schemes leaves substantial potential for improved energy efficiency.Source: §I · p. 1
  • The concatenated scheme in [2] combines an outer tree code with an inner compressed-sensing code that identifies transmitted columns before the tree decoder stitches submessages into valid paths.Source: §I · p. 2
  • Existing massive-MIMO activity detection commonly relies on individual pilots, while piggybacking approaches [13] retain user-specific codebooks and require 2^B pilot sequences for B-bit messages.Source: §I · p. 2

A. Our Contribution

The paper combines concatenated coding with non-Bayesian, non-coherent activity detection for massive-MIMO unsourced access. Its system model uses a shared codebook and block-fading channel, and the receiver outputs decoded messages rather than user identities.

A. Our Contribution

  • The scheme combines the concatenated approach of [2] with non-coherent activity detection that treats activity and large-scale pathloss coefficients as a deterministic unknown sparse non-negative vector.Source: A. Our Contribution · p. 2
  • The proposed non-Bayesian detector outperforms Bayesian VAMP even with fully known pathloss, avoids its instability at large antenna counts, and improves as antennas grow for fixed pilot dimension.The approach is slightly more complex than VAMP and requires less a priori knowledge.Source: A. Our Contribution · p. 3
  • The authors report substantially better performance than previous approaches and show that reliable communication can require arbitrarily small transmit Eb/N0 as receiver antennas grow sufficiently large.Source: A. Our Contribution · p. 3

II. SYSTEM MODEL

  • The model has M antennas and Ka active users sharing a codebook; it assumes spatially white zero-mean Rayleigh fading, with channel coefficients fixed across coding dimensions.The receiver outputs a list of decoded messages and evaluates misdetection and false alarms rather than assigning messages to users.Source: §II · p. 3

III. PROPOSED CODING SCHEME

The proposed concatenated code splits each message into blocks, appends parity bits to later blocks, and uses the resulting sequence as a path in a shared outer tree code.

III. PROPOSED CODING SCHEME

  • Each message is divided into L blocks, with the first block of size J and each later block smaller than J before parity augmentation.Source: §III · p. 3
  • Each later block is padded to J bits using pseudo-random linear combinations of information bits from preceding blocks, producing a depth-L tree.Source: §III · pp. 3–4
  • The pseudo-random parity-check equations are identical for all users, so every user employs the same outer code.Source: §III · p. 4

A. Inner Code

The inner code uses a common matrix whose columns are codewords sent across consecutive subslots. The resulting received matrices combine user activity, channel gains, fading, and noise.

A. Inner Code

  • Each active user sends a sequence of columns from the common coding matrix, one per consecutive subslot.Each column is an inner codeword representing a J-bit outer-encoded message.Source: A. Inner Code · p. 4
  • The received signal in each subslot is the coding matrix multiplied by the activity, gain, and channel matrices, plus noise.The activity matrix selects each active user's transmitted column; inactive users have zero activity vectors.Source: A. Inner Code · p. 4
  • The model permits channel matrices to change across subslots, provided their marginal entries remain i.i.d. Gaussian.Their correlation across subslots does not change the stated results.Source: A. Inner Code · p. 4
  • When the number of candidate columns substantially exceeds the number of active users, the effective activity matrix is row-sparse.The probability that a row is zero is (1 − 1/2^J)^K_a.Source: A. Inner Code · p. 4
  • The received-signal matrix in the stated model has no dependence on the total user population K_tot.The passage states that none of the matrices in the model depends on K_tot.Source: A. Inner Code · p. 4

B. Inner decoding as a massive MIMO activity detection problem

The inner decoder detects active coding-matrix columns by estimating their associated activity levels, then passes subslot message lists to an outer decoder. The outer decoder uses parity checks to select valid message sequences.

B. Inner decoding as a massive MIMO activity detection problem

  • The inner decoder identifies active columns of the coding matrix and passes the resulting subslot message lists to the outer decoder.The detection task is equivalent to identifying nonzero diagonal elements of Γ.Source: B. Inner decoding as a massive MIMO… · p. 4
  • Because each diagonal activity value is a random sum of user gains with unknown values and statistics, the method treats the activity vector as deterministically unknown.The passage says Bayesian VAMP approaches do not apply straightforwardly; it also reports numerical instability when the subslot dimension is small relative to the antenna count.Source: B. Inner decoding as a massive MIMO… · p. 4

C. Outer Code

  • The outer decoder extends only message paths that pass parity checks as it processes the subslot lists in order.The subblock messages contain parity bits, so not every combination of listed messages is possible.Source: C. Outer Code · p. 5

IV. MAXIMUM LIKELIHOOD ACTIVITY DETECTION

The activity detector estimates a deterministic activity vector using a likelihood based on the received-signal covariance. Its coordinate-descent minimization is independent of antenna count, although forming the covariance scales with that count.

IV. MAXIMUM LIKELIHOOD ACTIVITY DETECTION

  • The columns of the received matrix are modeled as i.i.d. Gaussian vectors with a conditional covariance determined by the activity estimate.The likelihood is based on this conditional distribution and the empirical covariance of the received columns.Source: §IV · p. 5
  • The likelihood function is reported to have only global minima when K_a ≳ O(n0^2).The work derives a low-complexity coordinate-descent algorithm and reformulates it to avoid explicit matrix inversion.Source: §IV · p. 5
  • The likelihood minimization is independent of M, while forming the empirical covariance scales with M; VAMP complexity scales at least linearly with M.The likelihood depends on the received data only through the n0 × n0 empirical covariance matrix.Source: §IV · p. 5
  • The coordinate-descent algorithm takes the sample covariance as input and initializes the activity estimate to zero.It initializes the inverse covariance estimate to (1/N0)I_n0.Source: §IV · p. 5

V. DISCUSSION AND ANALYSIS

The supplied passage identifies Figure 1 as the proposed coding scheme.

V. DISCUSSION AND ANALYSIS

A. Inner Code

The inner decoder's support-reconstruction error vanishes as the antenna count grows for any positive SNR under the stated active-user scaling. The bound is conservative, and its extension to maximum-likelihood decoding remains unproved in the passage.

A. Inner Code

  • The support-estimation error vanishes as M grows for any positive SNR when the active-user count follows the stated scaling.The result is based on the reconstruction-error bound for NNLS; the passage notes that applying the bound to maximum-likelihood decoding remains work in progress.Source: A. Inner Code · pp. 5–6
  • The NNLS reconstruction-error bound is conservative, and its universal constant κ is difficult or impossible to quantify sharply.The passage says numerical results suggest maximum-likelihood reconstruction is at least as good as NNLS, typically much better, but does not establish the bound for ML.Source: A. Inner Code · p. 6

B. Outer Code

The outer code treats each subslot's detected support as a vector OR-MAC output and combines those observations to decode messages. The analysis identifies different active-user limits for long and short blocks, while sufficiently many antennas can achieve a desired PUPE at any positive Eb/N0 under the stated conditions.

B. Outer Code

  • With error-free support recovery, each subslot's estimated support is the output of a vector OR-MAC whose inputs are the activity-matrix columns.The OR operation is component-wise, and the vector output entropy is bounded using the entropies of scalar OR-MACs.Source: B. Outer Code · p. 6
  • The outer code sends Ka JRout information bits per subslot, so reliable decoding requires that the OR-MAC output carry enough information for those bits.The passage states the necessary condition but the supplied excerpt does not include its full expression.Source: B. Outer Code · p. 6
  • For large block lengths, the outer code limits the active-user count; for small block lengths, inner-decoder detection limits it.The antenna analysis considers fixed Ka and block lengths large enough for the outer code to be limiting.Source: B. Outer Code · p. 6
  • For any Eb/N0 > 0, J, and Rout, if (25) holds and Ka ≤ 2J(1−Rout)+1, a sufficiently large M achieves a desired PUPE reliability level.Source: B. Outer Code · p. 6

A. Thresholding and Power Allocation

Thresholds convert detected supports into binary decisions without requiring knowledge of the active-user count, while power allocation can trade early false alarms against later ones. In the reported simulation, 400 receive antennas bring error probability below 10^-2 for 300 active users at Eb/N0 = 0 dB.

A. Thresholding and Power Allocation

  • Thresholding each section's support avoids requiring Ka, unlike selecting the Ka + ∆ largest entries, and balances missed detections against false alarms.A missed detection necessarily removes a message from the final list, whereas the outer decoder can correct a relatively large number of false alarms.Source: A. Thresholding and Power Allocation · p. 6
  • Decaying power across subslots can reduce early false alarms and support an increasing parity profile while keeping decoder complexity manageable.The total power allocation remains LP; the passage describes this as a compromise between parity placement and complexity.Source: A. Thresholding and Power Allocation · p. 7

B. Results

  • The simulation uses 96-bit payloads over 3200 complex channel uses, with L = 32 and J = 12.Its inner coding matrix has dimensions 100×4096; the passage contrasts this with 296 columns per user for the cited piggybacking scheme [13].Source: B. Results · p. 7
  • With M = 400, error probability falls below 10^-2 for 300 active users at Eb/N0 = 0 dB.The reported spectral efficiency is 9 bit per channel use; Figure 2 plots error probability against active-user count for different receive-antenna counts.Source: B. Results · p. 7, Figure 2 · p. 7
Loading 1901.00828v1…