⚡ electrical engineering

Hybrid Mamba-Attention Neural Architecture for Channel Estimation

This paper proposes a hybrid Mamba-attention neural architecture that utilizes a bidirectional selective scan to efficiently estimate OFDM channels with large numbers of subcarriers, achieving superior performance and generalization with lower complexity and fewer parameters than conventional transformer-based approaches.

Original authors: Dianxin Luan, Chengsi Liang, Jie Huang, Zheng Lin, Kaitao Meng, John Thompson, Cheng-Xiang Wang

Published 2026-07-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dianxin Luan, Chengsi Liang, Jie Huang, Zheng Lin, Kaitao Meng, John Thompson, Cheng-Xiang Wang

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 send a secret message across a crowded, noisy room using a series of flashlights. In the world of modern wireless communication, these flashlights are data signals, and the "room" is the air around us, filled with buildings, trees, and moving cars that bounce the signals around. This bouncing creates a messy puzzle called "channel distortion," where the message arrives scrambled. To fix this, the receiver needs to know exactly how the room messed up the signal—a process called "channel estimation."

For decades, engineers have tried to solve this puzzle with math formulas. Some formulas are fast but make mistakes (like guessing the weather based on a single cloud), while others are super accurate but require knowing the future (like needing to know exactly how every person in the room will move before the message is even sent). Recently, scientists started using "neural networks"—computer brains that learn from examples—to solve this. However, these computer brains have a problem: when the message gets huge (like in next-generation 6G networks with thousands of tiny signal pieces), the computer brain gets overwhelmed, needing too much memory and time to think. It's like trying to memorize a library of books by reading every single page one by one; it works for a few pages, but it breaks down for a whole library.

This is where a new paper steps in with a clever twist. The researchers, led by Dianxin Luan and his team, propose a new kind of computer brain architecture that mixes two powerful ideas: "Attention" (the ability to focus on important parts of a message) and "Mamba" (a new, super-efficient way to remember long sequences of information). They call their creation "MambaNet."

Here is the core problem they are tackling: In wireless signals, the different pieces of data (called subcarriers) are linked together in a way that doesn't follow a simple "first-in, first-out" order. It's like a conversation where the end of the sentence helps you understand the beginning. Traditional computer brains (called Transformers) try to look at every piece of data against every other piece to find these links. While this works well for short messages, it becomes a nightmare for long ones, requiring a massive amount of computing power that drains batteries and slows things down.

The paper argues that the old way of doing this is too heavy for the future of wireless tech. Instead of forcing the computer to look at everything at once, the authors suggest using a "Mamba" module. Think of Mamba as a highly efficient librarian who doesn't need to read every book on the shelf to find a connection. Instead, they walk down the aisle, remembering what they saw a moment ago and what they are about to see, creating a smooth flow of understanding.

However, there is a catch. In the world of wireless signals, the "past" and "future" of a signal are actually mixed together; the signal isn't strictly causal (it doesn't just flow one way). A standard Mamba usually only looks forward. To fix this, the team designed a "bidirectional" Mamba. Imagine a librarian who walks down the aisle twice: once from left to right, and once from right to left, combining both views to get the full picture. This allows the system to understand the complex, non-linear relationships between signal pieces without getting bogged down.

The researchers tested this new "MambaNet" architecture using simulations based on real-world standards (specifically the 3GPP TS 36.101 channel models). They compared their method against older, well-known techniques like Least Squares (LS), Minimum Mean Squared Error (MMSE), and other neural networks like InterpolateNet and Channelformer.

The results from their simulations show that MambaNet is a strong contender. When the signal-to-noise ratio (a measure of how clear the signal is) ranged from 5 dB to 25 dB, MambaNet consistently produced more accurate channel estimates than the other neural networks. For instance, at a high signal quality of 25 dB, MambaNet reduced the error (Mean Squared Error) by about 0.00072 compared to the HA02 model and 0.00033 compared to Channelformer. Even more impressively, it managed to outperform the theoretical "gold standard" (MMSE) in some high-signal scenarios, despite MMSE having the advantage of knowing the exact channel conditions in the simulation.

Beyond just being accurate, the paper highlights that MambaNet is much more efficient. While other advanced neural networks saw their complexity explode as the number of signal pieces increased, MambaNet kept its size manageable. It achieved these results with only 0.27 million tunable parameters, which is a 73% reduction compared to Channelformer and a 76% reduction compared to HA02. This means it could potentially run on smaller, battery-powered devices without needing a supercomputer.

The team also checked if their system could handle "unseen" situations. They trained the system on one type of environment and then tested it on completely different ones, like flat fading channels or two-path channels. The simulations showed that MambaNet remained robust and didn't crash, although it did perform slightly less perfectly than when it was tested on the exact environment it was trained for. This suggests that while the system is very good at generalizing, it still feels the effects of being trained on a specific "flavor" of noise.

In conclusion, the paper suggests that by combining the focusing power of attention mechanisms with the efficient, two-way memory of a bidirectional Mamba, we can build a smarter, lighter, and more accurate way to decode wireless signals. While these findings are currently based on computer simulations rather than real-world field tests, they offer a promising path forward for making future 6G networks faster and more reliable without burning out our devices.

Technical Summary: Hybrid Mamba-Attention Neural Architecture for Channel Estimation

Problem Statement
Accurate Channel State Information (CSI) acquisition is critical for sixth-generation (6G) orthogonal frequency-division multiplexing (OFDM) systems to ensure effective compensation for channel impairments and support high data rates. Conventional estimation methods, such as Least Squares (LS) and Minimum Mean Squared Error (MMSE), face significant limitations: LS performs poorly, while MMSE relies on unavailable actual channel statistics. Furthermore, precise interpolation for doubly-selective channels is challenging due to inaccurate statistical models. While deep learning solutions (e.g., ChannelNet, ReEsNet) offer locally optimal solutions, they often suffer from performance degradation on unseen channels. Although Transformer-based architectures have shown promise, their quadratic computational and spatial complexity with respect to the number of subcarriers ( $N_f$ ) makes them computationally prohibitive for large-scale subcarrier configurations, posing challenges for battery-powered devices.

Methodology
The paper proposes MambaNet, a hybrid neural architecture that integrates a multi-head attention mechanism with a customized Mamba module to address channel estimation in OFDM systems with large $N_f$ .

Input Processing: The network takes the LS estimates at pilot positions ( $\hat{H}^{pilot}_{LS}$ ) as input. These are concatenated into a column vector, with real and imaginary parts split into two channels to form "channel tokenization."
Attention-Assisted Mamba Module:
- Self-Attention: A multi-head attention module first captures correlations among the LS inputs. It projects the input to generate Query ( $Q$ ), Key ( $K$ ), and Value ( $V$ ) matrices, computing scaled dot-product attention to refine feature representations.
- Customized Mamba: The output of the attention module is processed by a Mamba module based on the Selective State Space Model (SSM). Unlike standard Mamba implementations, this paper introduces a bidirectional selective state scan. This is necessary because channel gains across subcarriers are inherently non-causal. The module performs both forward and backward scans to propagate information in both directions, capturing long-range dependencies efficiently.
- Gate Mechanism: The Mamba module utilizes gate coefficients generated via depthwise convolution and activation functions (SiLU, Sigmoid) to dynamically control state updates and information flow.
Residual Architecture and Upsampling: The features from the Mamba module are reshaped and passed through a residual architecture consisting of convolutional layers and residual blocks. Finally, a bilinear interpolation upsampling module and a final coupled convolutional layer generate the complete channel matrix estimate for the entire slot.

Key Contributions

Hybrid Architecture: The integration of a multi-head attention mechanism with a state-space model (Mamba) to balance correlation capture and computational efficiency.
Bidirectional Selective Scan: A novel adaptation of the Mamba structure that employs bidirectional scanning to handle the non-causal nature of subcarrier channel correlations, a departure from the causal assumptions in standard Mamba applications.
Complexity Reduction: The proposed method replaces the quadratic complexity ( $O(N^2)$ ) of self-attention with the linear complexity ( $O(N)$ ) of the selective scan, significantly reducing space and computational complexity for large $N_f$ .
Robust Generalization: The framework is designed to perform robustly on previously unseen channel realizations, leveraging synthetic training datasets generated via specific design criteria.

Experimental Results
Simulations were conducted using the 3GPP TS 36.101 channel models (specifically the Extended Typical Urban - ETU channel) with QPSK modulation.

Performance: MambaNet outperforms baseline methods, including LS, MMSE (with linear interpolation), InterpolateNet, HA02, and Channelformer, across a Signal-to-Noise Ratio (SNR) range of 5 dB to 25 dB.
- At 25 dB SNR, MambaNet achieves a Mean Squared Error (MSE) lower than HA02 and Channelformer.
- In terms of Bit Error Ratio (BER), MambaNet achieves the lowest BER among neural network baselines, though the MMSE method (with perfect channel knowledge) still holds a slight edge in BER at high SNR.
Generalization: When tested on unseen channels (Flat Fading, Two-path, EPA, EVA, and ETU), the offline-trained MambaNet demonstrates robust performance, though a performance gap remains compared to the design channel, indicating moderate degradation due to distribution mismatch.
Complexity: MambaNet achieves a significant reduction in tunable parameters compared to Transformer-based baselines. It has 73% fewer parameters than Channelformer and 76% fewer than HA02. However, it requires approximately twice the running time of Channelformer, though it remains more efficient in terms of parameter scaling as $N_f$ increases.

Significance and Claims
The paper claims that the proposed hybrid Mamba-attention architecture offers a practical solution for channel estimation in large-scale OFDM systems. By leveraging the state-space model, the framework efficiently captures long-distance dependencies among subcarriers without the prohibitive computational cost of full Transformer architectures. The bidirectional scan specifically addresses the non-causal characteristics of channel data, enabling superior estimation performance with fewer tunable parameters. The authors assert that this approach provides a viable path for accurate channel estimation in future 6G systems where subcarrier counts are high and computational resources are constrained.

More like this