PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning-to-Transition for Large-scale and High-Order MIMO Detection

Yubo Zhang, Yiyao Liu, Xiaodong Wang

arXiv:2608.14511v1cs.ITcs.AI

TL;DR

High-order MIMO detection must search an exponentially large symbol space efficiently while supporting reliable soft information for decoding. This paper learns complete-vector transitions with a channel-coupled Transformer and transfers the hard detector into an iterative soft receiver, but the supplied passages do not report empirical outcomes.

  • Problem

    High-order MIMO maximum-likelihood detection searches an exponentially large feasible symbol space, motivating learned amortized optimization across channel instances.

  • Method

    The framework recursively learns stochastic complete-vector transitions with a channel-coupled Transformer, then transfers the hard policy into untied soft-input soft-output IDD layers.

  • Results

    The supplied passages describe BER, residual, BLER, GMI, and Brier-score evaluations but do not report their empirical outcomes.

  • Takeaways & Limitations

    The paper presents a unified learned search process spanning hard detection and decoder-informed soft iterative reception.

Abstract

from arXiv · show

High-order multiple-input multiple-output (MIMO) detection requires efficient search over a large discrete symbol space while producing reliable soft information for channel decoding. This paper develops a learning-to-transition (L2T) framework that formulates MIMO detection as a stochastic sequence of complete-vector transitions. At each transition, a channel-coupled Transformer updates both the instance embedding and the sampling policy, while a blockwise autoregressive factorization captures inter-stream dependence with moderate sequential complexity. For hard-output detection, a transition network is applied recursively and trained through a residual-to-BER curriculum, which first learns the MIMO search geometry from the exact residual metric and then aligns the policy with transmitted-bit accuracy. For soft-output reception, the well-trained hard policy is cloned at the parameter level into every layer of an untied soft-input soft-output iterative detection and decoding (IDD) receiver. This tied-to-untied transfer preserves the learned zero-prior search dynamics while enabling layer- and round-specific specialization under decoder feedback. Within each IDD round, decoder priors tilt candidate generation according to Bayes' rule, and likelihood-weighted terminal hypotheses produce posterior and extrinsic log-likelihood ratios for LDPC decoding. A multi-stage training strategy further stabilizes the hard-to-soft transfer by progressively exposing the receiver to synthetic and in-loop decoder-generated priors.

I. INTRODUCTION · II. PROBLEM STATEMENT · A. MIMO Detection

The paper formulates high-order MIMO detection as learned stochastic transitions between complete transmit vectors, addressing exhaustive-search complexity through a channel-coupled Transformer and amortized optimization. It further aligns hard detection with BER and transfers the learned policy to prior-conditioned soft IDD reception.

  • I. INTRODUCTION: High-order MIMO detection becomes difficult as system loading and modulation order grow because ML search spans Q^Nt transmit vectors.This exponential search is prohibitive for large or square high-order systems.
  • I. INTRODUCTION: Existing approximate, model-driven, and learned detectors improve performance-complexity tradeoffs but generally do not learn how to traverse the complete-vector discrete search space.Related combinatorial and Ising formulations expose the discrete structure but typically use specialized optimization rather than an amortized transition policy.
  • I. INTRODUCTION: The proposed L2T framework recursively applies a channel-coupled Transformer to map each complete MIMO vector to a distribution over the next complete vector.The Transformer jointly updates the instance embedding and sampling policy, while blockwise autoregression captures inter-stream dependence.
  • I. INTRODUCTION: Hard-output training uses trajectory-level reinforcement learning with a residual-to-BER curriculum that shifts from exact residual geometry to transmitted-bit BCE alignment.The residual-driven policy-gradient objective teaches coupled search geometry before the transmitted-bit objective targets BER performance.
  • I. INTRODUCTION: Soft reception transfers the tied hard policy into independently trainable soft-detector layers, preserving zero-prior initialization while enabling specialization across depth and IDD rounds.Bayesian prior-tilted sampling and likelihood-weighted hypotheses provide posterior and extrinsic LLRs for decoding.
  • A. MIMO Detection: The MIMO problem uses Nt transmit antennas, Nr receive antennas, a channel matrix H, and a unit-energy Gray-coded QAM constellation of order Q.Each QAM symbol contains B = log2 Q coded bits, with B/2 bits on each PAM axis.
  • A. MIMO Detection: ML detection exhaustively searches 2^NtB feasible bit assignments, becoming computationally prohibitive as Nt or B increases.Repeatedly solving similar problems is also inefficient because the channel H frequently varies.
  • A. MIMO Detection: The paper therefore trains one neural network across channel instances H ∼ pH, shifting computation from online optimization to offline amortized learning.This formulation remains challenging because symbol detection is an NP-hard combinatorial optimization problem lacking a high-quality general-purpose solver.

B. Learning to Transition

The L2T method learns a transition policy over the full MIMO solution space using channel features derived from matched-filter and Gram-matrix statistics. It recursively updates both candidate solutions and channel embeddings across T transitions to amortize the search while reflecting explored regions.

  • Transition policy: L2T trains a neural network to learn an effective transition policy over the entire solution space B.The approach is motivated by the importance of expressive channel embedding for detection performance.
  • Channel embedding: Matched-filter entries z and diagonal Gram-matrix entries form node inputs, while off-diagonal entries encode complex inter-stream couplings as edge inputs.A GNN encoder Gφ maps these sufficient channel statistics to the initial channel embedding.
  • Recursive transitions: The L2T scheme performs T updates, amortizing the search across transition steps and reducing the difficulty of learning the policy.This recursive formulation distributes the solution-space search across all steps.
  • Recursive transitions: At each transition, bit-wise sampling produces the next solution b(t) ∼p(t), while a second output updates the feature embedding at transition step t.Updating the embedding prevents a static representation from ignoring which search-space regions the evolving candidate has already explored.

III. HARD-DECISION RECEIVER · A. Transformer-based Policy Network

The hard-decision receiver trains a learning-based solver for direct bit decisions in the NP-hard MIMO detection problem. Its Transformer policy recursively refines sampled solutions and channel embeddings across transitions.

  • III. HARD-DECISION RECEIVER: The non-iterative receiver directly obtains transmitted bits through hard decisions while learning to solve the NP-hard combinatorial optimization problem at the population level.The design first specifies the policy network and then introduces policy-gradient training.
  • A. Transformer-based Policy Network: The transition policy Πθ0 is parameterized by a Transformer that processes the previous sampled solution together with its channel embedding.The previous solution is reshaped and concatenated with the corresponding embedding to form the next layer’s input token sequence.
  • A. Transformer-based Policy Network: Layer normalization and multi-head attention transform the channel-coupled token sequence into attention outputs for subsequent policy updates.The architecture defines query, key, and value projections for each attention head, then combines the head outputs through concatenation and projection.
  • A. Transformer-based Policy Network: The projected attention representation includes channel embedding residuals, and the next discrete solution is sampled from the resulting policy.This branch couples continuous channel information with stochastic solution transitions.
  • A. Transformer-based Policy Network: A parallel branch updates the channel embedding through residual connections and an MLP.The embedding update is produced alongside the solution-sampling branch within the same Transformer layer.
  • A. Transformer-based Policy Network: Unrolling the Transformer layer over T transitions recursively refines both the discrete solution and the continuous embedding.The two output branches therefore evolve jointly across the transition sequence.

B. Blockwise Autoregressive Sampling

The framework uses blockwise autoregressive sampling to balance mean-field speed against bit-by-bit modeling of conditional correlations. Blocks correspond to QAM symbols, preserving inter-antenna dependence while sampling within-symbol bits in parallel.

  • B. Blockwise Autoregressive Sampling: Blockwise autoregressive sampling generates complete solutions while trading off mean-field parallelism against the sequential cost of bit-by-bit autoregression.Mean-field sampling ignores conditional correlations among N bits, whereas bit-by-bit sampling captures them but requires many sequential forward processes.
  • B. Blockwise Autoregressive Sampling: Each block is a B-bit QAM symbol from one transmit antenna, retaining dependence across Nt symbols while generating each symbol’s bits in parallel.Dense off-diagonal entries in the Gram matrix R couple detections across antennas; the factorization avoids a bit-by-bit chain of length NtB.
  • B. Blockwise Autoregressive Sampling: Randomizing the block ordering n = 1, 2, . . . , Nt during training prevents persistent ordering bias.Before sampling each data symbol, undetermined symbols are represented by an all-zero bmask vector of length B.
  • B. Blockwise Autoregressive Sampling: After Nt blockwise samplings, the policy produces the complete solution b(t), with M = 1 yielding mean-field sampling and M = N yielding bit-by-bit autoregression.At each step, the next block’s variables are sampled in parallel from the partial solution, prior solution, and current channel embedding.

C. Policy-Gradient Training Scheme

The policy-gradient scheme trains a stochastic L2T trajectory policy using an entropy-regularized objective over channel instances and sampled solution trajectories. Its step-wise loss transitions from a residual surrogate to BER-oriented BCE, while inference recursively generates candidate solutions and selects the one with the smallest residual loss.

  • Trajectory objective: The L2T process forms a complete stochastic solution trajectory τ=(b(0), b(1), …, b(T)) with a computable sampling probability.The trajectory consists of the initial solution and T transition outputs.
  • Trajectory objective: The trajectory-wise objective combines the step-wise loss g(τ; y, H) with entropy regularization λ log pθ0(τ | y, H), where λ>0 encourages early exploration.The entropy term is the negative expected log-policy term, so positive λ promotes exploration during initial training.
  • Population training: Population training averages over channel instances and stochastic trajectories, approximating the inner trajectory expectation by sampling K trajectories initialized from randomly generated valid solutions.The outer expectation targets a solver that generalizes across the channel distribution, while the inner expectation reflects policy-induced trajectory stochasticity.
  • Step-wise loss: The hybrid step-wise loss combines residual loss f(b(t); y, H) and BER-oriented BCE, with ρs decreasing from 1 to 0 to shift training toward BCE.Residual loss provides a smooth early-stage surrogate, while BCE uses known transmitted bits to target detection BER.
  • Inference: During inference, the trained policy is recursively applied for T transitions from random initializations, and the final hard decision is the candidate with the smallest residual loss.The resulting final solution set contains K transitioned candidates, from which the minimum-residual candidate is selected.

IV. SOFT-INPUT SOFT-OUTPUT RECEIVER · A. Iterative Detection and Decoding

The iterative detection and decoding receiver exchanges soft information between a MIMO detector and an LDPC decoder across IDD rounds. Interleaving, LLR conventions, decoder mapping, initialization, and extrinsic-information subtraction define the exchange protocol.

  • A. Iterative Detection and Decoding: IDD exchanges soft information between the MIMO detector and a soft-output LDPC decoder.This establishes the receiver’s iterative detector-decoder architecture.
  • A. Iterative Detection and Decoding: An interleaver distributes one LDPC codeword across a batch of MIMO observations.The interleaver is denoted by Γ.
  • A. Iterative Detection and Decoding: After detection, LLRs from the observations are reassembled in codeword order before LDPC decoding.The reassembly reverses the interleaved observation arrangement for decoding.
  • A. Iterative Detection and Decoding: The receiver uses the convention L(b) = log, with P(b = 1) = sigmoid(−L(b)).The supplied passage states this LLR convention and probability relation.
  • A. Iterative Detection and Decoding: At the mth IDD round, the complete turbo exchange uses the LDPC belief-propagation mapping DBP.The mapping relates the detector-decoder exchange at each round.
  • A. Iterative Detection and Decoding: The LLR subscripts A, P, and E denote a-priori, posterior, and extrinsic information, respectively.These labels identify the information types exchanged by the modules.
  • A. Iterative Detection and Decoding: The exchange is initialized with LA,(0) = 0.The initial a-priori LLR is zero.
  • A. Iterative Detection and Decoding: Two subtractions ensure that each module passes only newly created information instead of returning its input prior as new evidence.The detector and decoder contribute information from their own observations or parity constraints.

B. LLR Computations

The detector incorporates decoder priors through Bayes-tilted sampling probabilities and computes posterior and extrinsic LLRs from likelihood-weighted terminal trajectories. The resulting extrinsic subtraction is only approximate after mixture aggregation, with residual prior dependence diagnosed explicitly.

  • LLR Computations: Bayes-tilted sampling probabilities incorporate the a-priori LLR, while LA = 0 recovers the previous hard-decision detector.The sampling bias follows the posterior-to-policy relation in (29).
  • LLR Computations: Because p(y | b, H) ∝ exp[−Nr f(b; ξ)], τw = 1/Nr is physically matched, whereas other values explicitly temper the likelihood.The posterior and extrinsic LLRs are computed from this likelihood-weighted mixture.
  • LLR Computations: The detector combines K sampled trajectories using exact normalized residuals and terminal Bernoulli uncertainty to estimate posterior bit probabilities.This forms a Rao–Blackwellized terminal-mixture estimator whose weights are evaluated at sampled terminal vectors.
  • LLR Computations: After residual-weighted mixing, extrinsic subtraction removes the bit’s own prior only approximately because both mixture marginals and trajectory weights depend on LA.The residual dependence shrinks as trajectories agree and is measured by a prior-sensitivity diagnostic.

C. IDD Training Scheme · 1) Hard-decision initialization:

The soft IDD training scheme addresses drifting decoder priors by transferring the recursively shared hard-decision policy into an untied detector unfolded across transition steps and IDD rounds. Subsequent soft training then specializes layers to search position and round-specific decoder feedback.

  • C. IDD Training Scheme: Soft IDD training must handle nonstationary decoder priors whose distributions drift across IDD rounds as detector parameters evolve.The hard-decision training scheme assumes equal priors and cannot directly adapt to practical IDD settings.
  • 1) Hard-decision initialization:: The hard-decision receiver recursively reuses the same policy Πθ∗0 at all T transitions.Parameter sharing supports learning a stable transition rule but constrains expressive power.
  • 1) Hard-decision initialization:: The soft-IDD receiver is unfolded across both the T transition steps and the M IDD rounds.This creates a more expressive construction than the recursively shared hard-decision receiver.
  • 1) Hard-decision initialization:: The resulting soft parameter family ΘS = {θ(m,t)}m,t is untied, allowing each layer to adapt to search position and decoder-prior reliability.Untying provides layer-specific flexibility across the unfolded detector.
  • 1) Hard-decision initialization:: The trained hard policy supplies the common initialization for every untied soft layer.This transfers the stable hard-decision solver into the higher-capacity soft receiver.
  • 1) Hard-decision initialization:: Parameter sharing is released only after initialization, allowing later soft training to specialize rounds as decoder feedback becomes increasingly informative.The procedure continues from a stable, parameter-efficient hard-decision solver to an iterative soft receiver.

2) Synthetic-prior training: · 3) In-loop training:

Synthetic-prior training exposes the detector to Gaussian-model priors across the required reliability range without LDPC-decoder drift. In-loop training then inserts the fixed 5G NR LDPC decoder, uses decoder-generated extrinsic priors, and trains from final-transition posterior probabilities with straight-through backpropagation.

  • 2) Synthetic-prior training:: Synthetic priors are sampled from the classical Gaussian LLR model without invoking the LDPC decoder.This provides the detector’s training priors before decoder feedback is introduced.
  • 2) Synthetic-prior training:: The sampled reliability σA is obtained from target a priori mutual information through σA = J−1(IA).The Gaussian prior’s mean has the transmitted-bit sign, with magnitude increasing with reliability.
  • 2) Synthetic-prior training:: Covering the required reliability range exposes the detector to uninformative, moderate, and confident priors without decoder-induced drift.This range is achieved by sampling the target a priori mutual information.
  • 3) In-loop training:: In-loop training inserts the fixed LDPC decoder into the forward pass and unrolls prior exchange from LA,(0) = 0.The resulting priors match those encountered by the deployed receiver, closing the remaining synthetic-prior gap.
  • 3) In-loop training:: Only the posterior probability from the final transition of each IDD-round network enters the training loss.The subsequent passages define the round-m posterior BCE and the ultimate soft-IDD training loss.
  • 3) In-loop training:: The forward pass uses the actual 5G NR LDPC decoder, and the next detector receives its extrinsic mapping EBP(l) = DBP(l) −l.The decoder is configured according to the Sionna link-level implementation [22].
  • 3) In-loop training:: Straight-through backpropagation copies gradients through the composite extrinsic module while retaining the true decoder posterior and subtraction in the forward pass.Thus, forward computation uses the decoder’s actual extrinsic operation, while backward computation passes gradients through the whole module.

V. PERFORMANCE EVALUATION … C. Overall Performance

The evaluation studies L2T-based hard and soft MIMO detection under standardized simulation conditions, with ablations targeting curriculum learning and hard-to-soft receiver transfer. Overall experiments assess uncoded BER, post-LDPC BLER, and LLR quality across 64-QAM and 256-QAM, including round-by-round IDD behavior.

  • A. Simulation Settings: Simulation uses perfect-CSI Rayleigh downlink MIMO with normalized Gray-coded 64-QAM and 256-QAM symbols.Channel coefficients are i.i.d. CN(0, 1/N_t), and symbols have unit average energy.
  • A. Simulation Settings: Unless otherwise stated, L2O uses T = 8 transitions, embedding dimension d = 256, eight attention heads, d_ff = 256, and K = 16 trajectories.One trajectory starts from the LMMSE estimate and the others from randomized perturbations; residual-only post-processing is disabled.
  • A. Simulation Settings: The hard detector is trained at 20 dB with AdamW, residual-to-BER curriculum thresholds s_0 = 10^3 and s_1 = 5 × 10^3, and modulation-specific batch sizes.Soft detector layers are initialized from the same hard checkpoint and trained with synthetic-prior and decoder-in-the-loop stages; SNR points use at least five seeds.
  • A. Simulation Settings: Hard detection compares LMMSE, QR-domain K-best, OAMP-Net2, RE-MIMO, and SGT, while soft IDD compares soft LMMSE, MMSE-PIC, K-best list-MAP, SGT, and DUIDD.All methods share channel realizations, modulation, code, decoder iterations, and SNR definition.
  • B. Ablation Studies: The curriculum ablation compares smooth residual-to-BER training with residual-only, BCE-only, and abrupt residual-to-BCE switching using BER, normalized residual, and seed variation.The full model also tracks normalized Hamming distance between consecutive vectors across transition indices.
  • B. Ablation Studies: The soft-receiver ablation removes or constrains hard-checkpoint initialization, parameter untiedness, prior tilting, synthetic-prior training, decoder-in-the-loop fine-tuning, and gradient flow.These variants are evaluated near the coded waterfall from the complete three-stage training scheme.
  • C. Overall Performance: Uncoded BER-versus-SNR experiments test whether complete-vector transitions improve over linear and neural detectors and approach explicit tree-search K-best performance.The comparison covers both 64-QAM and 256-QAM.
  • C. Overall Performance: Post-LDPC BLER-versus-SNR experiments show the proposed receiver after one, two, and three IDD rounds, with baselines matched in decoder iterations and applicable detector–decoder exchanges.Final-round information-bit BER and detector GMI diagnose whether Bayesian prior-tilted sampling turns decoder feedback into increasingly informative candidate sets.
Loading 2608.14511v1…