Neural Augmentation of MIMO-OFDM Receivers for Universal LLR Reconstruction

Abstract: The growing demands for higher throughput and cost-efficient wireless communications drive the need for receivers that are both simple to deploy and robust to hardware impairments and nonlinear environments. While classical model-based receivers and recently proposed deep neural network ( DNN) architectures provide complementary benefits, they either rely on simplified linear Gaussian assumptions, require considerable computational resources, or are tailored for a given setting and modulation. In this work, we propose a compact and modular DNN augmentation that universally refines the soft outputs of existing receivers (model-based or data-driven), addressing two distinct operating regimes: structurally incomplete soft information arising from reduced-complexity detectors, and degraded soft outputs caused by hardware impairments and synchronization errors. A key property of the proposed framework is its task-agnostic nature: operating without any knowledge of the specific source of unreliability, it produces well-calibrated log-likelihood ratios (LLRs) suitable for channel decoding. Our design leverages an element-wise scaled convolutional neural network tailored to perform learned interference cancellation across users and neighboring subcarriers, combined with a training algorithm that encourages accurate LLR s for soft channel decoding. Numerical results demonstrate that the proposed augmentation consistently improves diverse receiver algorithms in challenging channel conditions while incurring minimal overhead.

• The paper introduces a compact ESCNN post-processor that refines corrupted LLRs and reconstructs missing bit metrics across detectors, modulation orders, and channel conditions.
• The method restores near-unity mutual information under CFO and IQ mismatch, enables reduced-bit sphere decoders to achieve reliable LDPC decoding, and can deliver BLER below 10⁻³ in combined impairment scenarios.
• The modular design uses 80,400 parameters and adds only 5–6% latency to LMMSE and DeepSIC, while remaining limited by severe clipping, simulated validation, and the need for labeled retraining data.

The paper "Neural Augmentation of MIMO-OFDM Receivers for Universal LLR Reconstruction" (2606.29345) proposes a lightweight, modular deep learning framework that refines the soft outputs of an arbitrary primary detector in a multi-user MIMO-OFDM uplink, rather than replacing the receiver altogether. The work, authored by Ory Eger and Nir Shlezinger of Ben-Gurion University and partially presented at IEEE ICC 2026, addresses two distinct failure modes of existing receivers: structurally incomplete log-likelihood ratios (LLRs) produced by reduced-complexity detectors, and corrupted LLRs caused by hardware impairments and synchronization errors. A single trained augmentation module handles both regimes without knowledge of the impairment source, producing calibrated LLRs suitable for LDPC channel decoding.

Motivation and problem formulation

The authors position the work against the two dominant receiver design paradigms. Classical model-based detectors—LMMSE/ZF equalizers, sphere decoders, and soft interference cancellation—are analytically grounded but rely on a linear, (pseudo-)time-invariant channel with additive Gaussian noise. Practical systems violate these assumptions through carrier frequency offset (CFO), which induces inter-carrier interference (ICI), IQ mismatch (IQMM), amplifier nonlinearities, and clipping. Compensating within the model-based paradigm requires impairment-specific designs or computationally expensive generalizations such as Volterra series. Conversely, fully data-driven receivers (DeepRx, DeepSIC, GNN- and attention-based architectures) can learn in mismatched environments but carry high parameter counts, substantial training requirements, and limited flexibility across modulations and site conditions.

The paper formalizes the system as a multi-user uplink with $K$ single-antenna users, $n_r$ receive antennas, and $B$ subcarriers, characterized by a general conditional distribution $p_{Y|S}$ that subsumes both the canonical linear Gaussian model and CFO-affected channels with ICI leakage coefficients $c_{l,b}(\varepsilon)$. Any primary detector is treated abstractly as a mapping $\psi$ producing bit-wise LLRs for $N' \leq N$ bits per symbol. The augmentation module must satisfy four requirements: channel-model-agnostic operation, architecture-agnostic compatibility, low latency, and modulation scaling/LLR completion (handling $N' \neq N_s$). The latter covers two scenarios: neural receivers trained for a fixed modulation being reused at higher orders, and sphere decoders whose complexity grows as $|\mathcal{S}|^K$, motivating a reduced-bit sphere decoder (RBSD) that computes only a subset of LLRs.

ESCNN architecture and training

The core building block is an element-wise scaled CNN (ESCNN), one module per user, operating in parallel. Each module receives a structured input combining the sigmoid-transformed primary-detector LLRs (zero-filled where missing, so a zero LLR represents an uninformative prior) with the real and imaginary parts of the received signal, forming a $B \times (2n_r + K\cdot N)$ matrix. A per-user element-wise scale matrix $n_r$0 adapts to subcarrier-specific variation, followed by a compact three-layer one-dimensional CNN with kernels spanning adjacent subcarriers. This design exploits local spectral correlations to perform learned interference cancellation across users and neighboring subcarriers—extending the DeepSIC principle to the multi-carrier setting—while shared kernels and short convolutions keep the parameter count low. The CNN outputs $n_r$1 channels; only the $n_r$2 channels matching the active modulation are used as LLRs.

Training minimizes binary cross-entropy over the active bit outputs only, using a dataset $n_r$3 with varying modulation orders, via mini-batch SGD. The authors emphasize that this loss encourages well-calibrated probabilistic outputs whose logits serve as LLRs, and that the bit-wise formulation decouples the module from the constellation of the primary detector. The complexity analysis is concrete: for a three-layer $n_r$4-kernel CNN, the per-module parameter count is a few thousand—under 1% of a DeepRx parameterization for the same scenario—and total ESCNN parameters are 80,400 versus 909,696 for DeepSIC and 12,377,600 for DeepRx.

Numerical results

The evaluation uses 3GPP-compliant channels (TDL-C via Sionna; QuaDRiGa RMa and UMi), $n_r$5 subcarriers, $n_r$6 users, $n_r$7 antennas, LDPC coding, and two metrics: histogram-based bit-wise mutual information (MI) and CRC-verified BLER.

LLR refinement under impairments. On a clean TDL-C channel, augmentation offers little gain, as expected. With a 4.5 kHz CFO, the unaugmented LMMSE, sphere decoder, and DeepSIC fail to reach unity MI and cannot decode reliably, while augmented versions are driven close to unity MI with very low BLER; DeepRx, which already handles ICI, is neither improved nor degraded. The CFO results replicate consistently across RMa and UMi channels. Under IQMM (1 dB gain, 5° phase mismatch), augmentation substantially improves the LMMSE and sphere decoders and further improves DeepRx BLER.

Transferability. A notable cross-detector result: an ESCNN trained on sphere-decoder or DeepSIC LLRs but tested with LMMSE LLRs still achieves near-optimal performance (11.8 and 12.1 dB SNR at 10% BLER, versus 11.7 dB for the matched LMMSE-trained module, on a channel where the unaugmented receiver never reaches 0.1 BLER). DeepRx-trained augmentation transfers less well (15.3 dB). For modulation scaling, a single modulation-agnostic module trained jointly on QPSK/16QAM/64QAM slightly outperforms dedicated per-modulation modules (8.8 vs. 9.5 dB for 16QAM; 17.4 vs. 18.1 dB for 64QAM at 10% BLER), which the authors attribute to the larger effective training set.

LLR completion. In a demanding 64QAM, $n_r$8 setting with medium spatial correlation, an RBSD computing only 4 of 6 LLRs runs roughly an order of magnitude faster than a full sphere decoder but its MI saturates at 2/3 and BLER saturates entirely when unaugmented. With ESCNN augmentation, the two missing LLRs are reconstructed, MI reaches unity, and the augmented RBSD outperforms both augmented and unaugmented LMMSE by 2–4 dB. In combined scenarios—RBSD plus 3 dB clipping (UMi channel) or plus 2.25 kHz CFO (TDL-C)—augmentation restores waterfall behavior across all four detectors, achieving BLER below $n_r$9. The authors candidly note that under 3 dB clipping, MI plateaus above 0.8 but below unity, likely due to distortion severity, and that in the clipping case a small MI gap between unaugmented and augmented LMMSE coexists with a stark BLER difference (error floor versus waterfall), showing that MI alone does not fully predict BLER because decoding depends on the LDPC graph structure.

Latency. On an NVIDIA RTX 4060, augmentation adds 5–6% latency over LMMSE and DeepSIC (70→74 ms and 75→79 ms per 5000 OFDM symbols) and under 1% over the sphere decoder (~100 s) and DeepRx (800→804 ms).

Limitations and open questions

The paper concedes several boundaries of its claims. The universality of the augmentation is demonstrated empirically rather than established theoretically; the task-agnostic property is argued to follow from the data-driven formulation, but no calibration or convergence guarantee is provided. Performance under severe clipping does not fully recover unity MI, indicating a residual gap for strong nonlinearities. The evaluation is confined to simulated 3GPP channel models with synthetic impairments—no over-the-air or recorded-channel validation is presented—and to $B$0 users with $B$1 bits per symbol. Training assumes access to labeled data (transmitted bits) and to the primary detector during training; the cost of retraining when the primary detector or deployment environment changes is not quantified. The authors themselves flag as open directions the cascading of multiple augmentation modules in an unfolding-style architecture, and integration with online learning, modular adaptation, and drift-detection mechanisms for non-stationary channels.

Conclusion

This paper contributes a compact, receiver-agnostic neural post-processor for multi-user MIMO-OFDM soft detection, unifying LLR refinement under hardware impairments with LLR completion for reduced-complexity detectors in a single modulation-scalable module. The empirical evidence—near-unity MI restoration under CFO and IQMM, successful reconstruction of missing LLRs from a reduced-bit sphere decoder, cross-detector and cross-modulation transferability, and 5–6% or lower latency overhead—supports the central claim that neural augmentation is a practical mechanism for extending the operating range of existing receivers. The main open questions concern theoretical guarantees on LLR calibration, robustness under more severe nonlinearities, and validation on measured channels and non-stationary deployments.