⚡ electrical engineering

Learning Successive Interference Cancellation for Low-Complexity Soft-Output MIMO Detection

The paper proposes recurSIC, a lightweight, learning-based MIMO detection framework inspired by successive interference cancellation that delivers reliable soft-output detection with minimal computational complexity and memory usage, making it ideal for edge-constrained 5G and IoT devices.

Original authors: Benedikt Fesl, Fatih Capar

Published 2026-01-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Benedikt Fesl, Fatih Capar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a conversation in a crowded room where four people are talking at once, and you only have two ears. This is the challenge of MIMO detection in modern wireless devices (like your phone or a smart sensor). The device receives a jumbled mix of signals and has to figure out exactly what each person said.

The paper introduces a new method called recurSIC to solve this problem. Here is how it works, explained through simple analogies:

The Problem: The "Crowded Room"

In wireless communication, devices send multiple streams of data at the same time to go faster. When they arrive at the receiver, they overlap and interfere with each other, like voices in a noisy party.

The Old Way (Linear Detectors): Imagine trying to guess what was said by just averaging the noise. It's fast, but you often get the words wrong.
The "Perfect" Way (Maximum Likelihood): Imagine trying to listen to every possible combination of voices simultaneously to find the perfect match. This is incredibly accurate, but it requires a supercomputer and takes too long for small, battery-powered devices.
The "Standard" Way (SIC): This is like listening to the loudest voice first, guessing what they said, and then "subtracting" their voice from the room to hear the next person. It's fast, but if you guess the first person wrong, you mess up the rest of the conversation (this is called "error propagation").

The Solution: recurSIC (The "Smart Detective")

The authors created recurSIC, which is a hybrid approach. It keeps the fast, step-by-step structure of the "Standard Way" but adds a Machine Learning brain to make better guesses.

Think of recurSIC as a detective solving a mystery layer by layer:

The "Lightweight" Brain:
Instead of using a massive, heavy computer brain (which uses too much memory and battery), recurSIC uses a tiny, specialized "micro-brain" (a small neural network).
- Analogy: Imagine a detective who carries a single, small notebook. Instead of rewriting the whole book for every new clue, they use the same notebook for every step of the investigation. This saves space and makes the process incredibly fast.
The "Learning" Step:
When the detective tries to guess what the first person said, they don't just guess randomly. They use their training (data) to look at the messy signal and say, "Based on what I've seen before, it's 80% likely they said 'Hello' and 20% 'Help'."
- This is better than the old way, which just picks the single most likely word without considering the "maybe."
The "Multi-Path" Safety Net:
Sometimes, the detective isn't sure. In the old "Standard Way," if you pick the wrong word, the whole case fails.
- The Innovation: recurSIC has a trick called Multi-Path. Instead of just following one guess, it follows a few of the most likely guesses at the same time.
- Analogy: Imagine the detective splits into three clones. Clone A follows the "Hello" path, Clone B follows the "Help" path, and Clone C follows a third option. They all walk down the hallway together. At the end, the detective checks which clone's path makes the most sense for the whole story.
- This allows the system to be very accurate (almost as good as the supercomputer) without needing the supercomputer's power.
The "Soft" Output:
Modern phones don't just need to know what was said; they need to know how confident the detective is.
- If the detective is 99% sure, the phone's error-correction software can relax. If the detective is only 51% sure, the software knows to work harder to fix potential mistakes.
- recurSIC is great at providing this "confidence score" (called soft information), which helps the phone decode messages perfectly even in bad signal conditions.

Why Does This Matter?

The paper shows that this method is perfect for IoT devices (like smart meters, sensors, or low-cost 5G gadgets) because:

It's Tiny: It uses very little memory (like a small app on your phone).
It's Fast: It doesn't need to run complex calculations over and over.
It's Reliable: It handles the "crowded room" noise much better than current cheap methods, and almost as well as the expensive, slow methods.

In short, recurSIC is a smart, lightweight detective that solves the puzzle of overlapping wireless signals by using a tiny, reusable brain and keeping a few backup plans ready, ensuring your smart devices stay connected without draining their batteries.

Technical Summary: Learning Successive Interference Cancellation for Low-Complexity Soft-Output MIMO Detection

Problem Statement
The design of Multiple-Input Multiple-Output (MIMO) detectors for emerging 5G Reduced Capability (RedCap) and Internet-of-Things (IoT) devices faces a critical trade-off between performance, computational complexity, and memory footprint. While Maximum Likelihood (ML) detection offers optimal performance, its exponential complexity renders it impractical for real-world systems. Conversely, linear detectors (e.g., Zero-Forcing, MMSE) offer low complexity but suffer significant performance degradation in non-overdetermined configurations and high-order modulation scenarios. Classical Successive Interference Cancellation (SIC) improves upon linear methods but is prone to error propagation and often relies on simplistic approximations for soft-output generation, which limits reliability when paired with modern soft-input channel decoders like Low-Density Parity-Check (LDPC). Furthermore, existing learning-based detectors often require large models, lack structural interpretability, or exhibit data-dependent, irregular complexity that hinders edge deployment.

Methodology: recurSIC
The authors propose recurSIC, a lightweight, learning-based MIMO detection framework designed to balance high performance with minimal resource consumption. The methodology is grounded in the following principles:

Physics-Informed Architecture: Unlike purely data-driven detectors, recurSIC preserves the recursive, layer-wise structure of classical SIC. It operates on a signal model preprocessed via QR Decomposition (QRD) of an extended channel matrix.
Lightweight Learning Blocks: The framework replaces the standard decision logic in each SIC stage with a shared, lightweight Neural Network (NN) block. This block takes a scalar input (the interference-cancelled observation for a specific layer) and outputs a probability vector over the modulation alphabet.
Shared Weights and SNR Embedding: A single NN block is trained and reused across all detection layers and Signal-to-Noise Ratios (SNRs). To accommodate varying SNR conditions without retraining, the network utilizes sinusoidal position embeddings of the SNR value, which are transformed into scaling and bias vectors. This ensures a constant, minimal memory footprint independent of the number of transmit layers or operating SNR.
Multi-Path Hypothesis Tracking: To generate reliable soft information and mitigate error propagation, the framework extends to a multi-path formulation. Instead of selecting a single symbol per layer, the detector tracks a tunable number ( $K$ ) of the most probable symbol hypotheses. These hypotheses propagate through independent SIC recursions, forming a tree of candidate symbol combinations.
Soft-Output Generation: Soft information is derived by approximating bit-wise Log-Likelihood Ratios (LLRs) using the max-log approximation over the tracked candidate paths. To ensure stability, the authors employ LLR clipping and a fallback mechanism for bits where counter-hypotheses are missing from the tracked set.

Key Contributions

RecurSIC Framework: Introduction of a physics-informed, learning-based detector that retains the low-complexity, recursive nature of SIC while enhancing each stage via learned, lightweight modules.
Tunable Multi-Path Extension: A novel extension that enables reliable soft-output detection by tracking a fixed, small number of symbol hypotheses per layer. This allows for a tunable performance-complexity trade-off without increasing the model size or requiring multiple forward passes.
Minimalist Architecture: Design of a shallow NN with shared weights across all stages and SNRs, resulting in a constant and minimal parameter count (approximately $10^3$ to $2 \times 10^3$ parameters) suitable for resource-constrained devices.
Empirical Validation: Demonstration of strong hard- and soft-detection performance in realistic 5G NR scenarios (TDL-A and TDL-B channels) with 16QAM and 64QAM modulation, outperforming conventional linear and classical SIC baselines.

Results
Numerical results in realistic 5G NR scenarios (2x2 MIMO) indicate the following:

Hard-Decision Performance: Even with a single path ( $K=1$ ), recurSIC significantly outperforms linear MMSE equalizers and performs close to MMSE-SIC. Increasing the path count ( $K$ ) allows the detector to approach ML performance; for instance, $K=2$ yields near-ML Bit Error Rate (BER) in TDL-A channels.
Soft-Output Performance: In terms of Block Error Rate (BLER) with LDPC decoding, recurSIC consistently outperforms classical SIC variants (MMSE-SIC, ZF-SIC), which suffer from error propagation and overconfident LLR approximations. Multi-path tracking ( $K \in \{4, 6, 8\}$ ) consistently enhances soft-detection performance, approaching the ML benchmark.
Complexity: The approach achieves these results with ultra-low complexity. The computational cost scales primarily linearly with the constellation size ( $M$ ), requiring only approximately $10^3$ to $2 \times 10^3$ multiply-accumulates (MACs) per block evaluation. This is orders of magnitude lower than iterative learning-based detectors or Sphere Decoding.
Robustness: The framework demonstrates robustness under more challenging channel conditions (TDL-B with higher delay spread and Doppler), where it maintains strong performance with moderate increases in the number of tracked paths.

Significance
The paper posits that recurSIC is a significant step toward deploying machine learning on edge devices for wireless communications. By explicitly addressing the constraints of IoT and RedCap devices—specifically stringent limits on memory, power, and computational complexity—the work offers a viable alternative to both traditional signal processing and heavy deep learning models. The ability to generate high-quality soft information for LDPC decoding while maintaining a single forward pass and a minimal parameter count makes the framework particularly well-suited for practical, edge-constrained MIMO receivers. The authors suggest that future work could further optimize the performance-complexity trade-off through adaptive path management and joint learning of detection ordering.

The Problem: The "Crowded Room"

The Solution: recurSIC (The "Smart Detective")

Why Does This Matter?

Technical Summary: Learning Successive Interference Cancellation for Low-Complexity Soft-Output MIMO Detection

More like this