⚡ electrical engineering

Statistical Characterization and Block-EM Estimation of Frequency-Domain NSI for OFDM Systems in Bursty Impulsive Noise

This paper proposes a frequency-domain, block-based framework for mitigating impulsive noise in OFDM systems that derives a transformed Gaussian mixture model for noise statistics and introduces an unsupervised block-EM estimation algorithm to achieve optimal performance while preserving subcarrier orthogonality and avoiding the complexity of time-domain processing.

Original authors: Chin-Hung Chen, Wim van Houtum, Yan Wu, Alex Alvarado

Published 2026-08-19

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Chin-Hung Chen, Wim van Houtum, Yan Wu, Alex Alvarado

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

Modern communication systems, from the power lines that bring electricity to our homes to the wireless networks that connect our devices, face a persistent and invisible enemy: sudden, high-energy bursts of static. Unlike the steady, predictable hiss of background noise that engineers have long learned to manage, this interference arrives in sharp, unpredictable spikes. It is generated by the very electronics we rely on—switching power supplies, electric motors, and digital converters—that create chaotic electrical surges. When these surges hit a signal, they can overwhelm the delicate information being carried, causing data to vanish or become corrupted. For decades, engineers have tried to clean up these signals by looking at the raw electrical wave before it is processed, attempting to cut out the spikes or reconstruct the original wave. However, these methods often struggle because they try to fix the problem in the time domain, a realm where the signal's structure is complex and easily damaged by imperfect repairs.

A team of researchers at Eindhoven University of Technology and NXP Semiconductors has proposed a different approach, one that works with the natural architecture of modern digital transmission rather than against it. They focused on a system called Orthogonal Frequency-Division Multiplexing, or OFDM, which is the backbone of technologies like digital television and power-line internet. In this system, data is sent not as a single stream, but as a collection of many parallel channels, each carrying a small piece of information. The researchers realized that by waiting until the signal has been converted into these parallel frequency channels, they could describe the chaotic noise using a statistical model that is far more accurate than previous assumptions. Instead of trying to guess and subtract the noise, they developed a method to identify exactly which parts of the signal are being hit by a spike and which are safe, allowing the receiver to adjust its confidence in the data accordingly.

The core of their work involves understanding how these noise spikes behave when they pass through the mathematical transformation that turns the raw signal into frequency channels. The researchers found that while the noise looks chaotic in the time domain, it organizes itself into a predictable pattern once it enters the frequency domain. They modeled this pattern as a mixture of different states, where each state represents a different level of noise intensity. In a perfect world, a receiver would know exactly which state the noise is in at any given moment, allowing it to decode the message with near-perfect accuracy. In reality, this information is hidden. The researchers' breakthrough was creating a set of algorithms that can learn this hidden information on the fly, without needing a pre-recorded map of the noise.

They developed three distinct methods to solve this learning problem, each designed for different levels of complexity and resource availability. The first method treats each block of data independently, updating its understanding of the noise based on the current signal. The second method adds a layer of memory, recognizing that noise spikes often last for several moments in a row, and uses this continuity to make better predictions. The third and most sophisticated method is designed to be self-correcting; it starts with a large number of possible noise states and automatically prunes away the ones that are not needed, ensuring the system remains efficient even as the complexity of the signal grows. These methods rely on a specific feature of the transmission system: a small number of "silent" channels that carry no data. By observing only these silent channels, the algorithms can deduce the nature of the noise affecting the entire signal without ever needing to know the original message.

Through extensive computer simulations, the team demonstrated that their approach offers a significant advantage over traditional methods, particularly when the noise is highly bursty and the system is constrained. In scenarios where the noise is intense and unpredictable, their new algorithms could reduce the rate of errors by a factor of ten or more compared to standard receivers that assume the noise is uniform. The simulations showed that the more advanced methods, which account for the memory of the noise, were especially effective when the system had very few silent channels to observe, a common constraint in real-world applications. However, the researchers also found that as the size of the data blocks increased, the benefits of these complex methods began to fade, as the noise naturally smoothed out into a more predictable pattern. This suggests that the best approach depends heavily on the specific conditions of the network.

The study explicitly argues against the prevailing reliance on time-domain reconstruction techniques, which attempt to rebuild the original signal by mathematically subtracting the estimated noise. The authors found that these methods are prone to error; if the subtraction is even slightly imperfect, it destroys the delicate mathematical relationship between the different channels, creating new interference that is harder to fix than the original problem. Their frequency-domain approach avoids this pitfall entirely by never trying to reconstruct the raw wave, instead focusing on adjusting the statistical confidence of the received data. They also ruled out the idea that a simple, static model of noise is sufficient for modern systems, showing that the noise's behavior changes dynamically and must be tracked in real time.

The results, derived from simulations of a coded communication system, indicate that this framework can bridge the gap between theoretical perfection and practical implementation. By using the silent channels to learn the noise profile, the system can adapt to changing environments without requiring extra bandwidth or complex hardware. The researchers noted that while their most advanced algorithm offers the best performance, it comes with a higher computational cost, and in many cases, a simpler version of the algorithm is sufficient. This trade-off between complexity and performance is a key finding, providing engineers with a toolkit to choose the right level of sophistication for their specific needs. The work confirms that by shifting the perspective from time to frequency and embracing the statistical nature of the interference, it is possible to build receivers that are far more robust against the chaotic electrical noise of the modern world.

Technical Summary: Statistical Characterization and Block-EM Estimation of Frequency-Domain NSI for OFDM Systems in Bursty Impulsive Noise

Problem Statement
Impulsive noise (IN), characterized by high power and non-Gaussian distribution, poses a significant challenge to Orthogonal Frequency-Division Multiplexing (OFDM) systems, particularly in environments with electronic interference (e.g., power lines, switching gears). While existing mitigation techniques often rely on time-domain processing (e.g., adaptive clipping, blanking, or sparse reconstruction), these methods introduce several limitations:

Complexity and Alignment: They operate before the Discrete Fourier Transform (DFT), failing to align with OFDM's inherent frequency-domain processing flow.
Orthogonality Risks: Imperfect subtraction of time-domain impulses can destroy subcarrier orthogonality, generating self-induced Inter-Carrier Interference (ICI).
Computational Burden: Advanced reconstruction methods (e.g., Compressed Sensing, Sparse Bayesian Learning) require iterative processing and significant null-tone overhead, which is often impractical.
Statistical Mismatch: Conventional frequency-domain receivers typically assume Additive White Gaussian Noise (AWGN), which is suboptimal when the noise exhibits bursty, non-Gaussian characteristics.

The core challenge addressed is the lack of a robust, frequency-domain framework that can accurately estimate Noise State Information (NSI) without the complexity of time-domain reconstruction, thereby enabling optimal detection in bursty IN environments.

Methodology
The paper proposes a frequency-domain, block-based framework that shifts the focus from reconstructing time-domain impulses to statistically characterizing the post-DFT noise distribution.

Statistical Characterization:
- The authors model time-domain IN using a Markov-Middleton model (a Hidden Markov Model combined with a Gaussian Mixture Model).
- They derive the statistical representation of the noise in the frequency domain. By applying the DFT to a block of time-domain samples, the frequency-domain noise is shown to follow a transformed Gaussian Mixture Model (GMM).
- The number of frequency-domain states ( $L_f$ ) is derived based on the block size ( $M$ ) and the number of time-domain states ( $L$ ). As $M$ increases, the Central Limit Theorem (CLT) suggests the noise converges toward a single Gaussian distribution, but for finite block sizes, the non-Gaussian mixture structure remains significant.
Optimal Receiver Benchmark:
- An optimal receiver is derived assuming perfect NSI (knowing the exact noise state and variance for each symbol). This serves as an upper bound for achievable performance, demonstrating that precise NSI is critical for non-Gaussian environments, especially when block sizes are small or temporal correlation is high.
Block-EM Estimation Framework:
To bridge the gap between theory and practice, the paper proposes three unsupervised Expectation-Maximization (EM) algorithms to estimate NSI using null tones (guard bands) within the OFDM frame:
- SBS-EM (Symbol-by-Symbol): A simple EM algorithm that estimates a single state for the entire OFDM symbol block, updating variances and mixing coefficients iteratively.
- HMM-EM (Hidden Markov Model): An enhanced algorithm that models the temporal correlation between consecutive OFDM symbols. It utilizes the BCJR algorithm to compute joint state probabilities, explicitly learning transition probabilities to track bursty noise patterns.
- MAP-EM (Maximum A Posteriori): A variant designed to address the "state-space explosion" problem in large DFT systems. It incorporates a sparsity-promoting Dirichlet prior on mixing coefficients. This allows the algorithm to dynamically prune redundant states and merge similar variances, adapting the model complexity to the observation window without a fixed, high-dimensional state space.

Key Contributions

Frequency-Domain Statistical Derivation: The paper rigorously derives the frequency-domain GMM representation of bursty IN, establishing the conditions under which the CLT allows for simplified AWGN assumptions and where precise NSI is necessary.
Block-EM Algorithms: The development of three distinct EM variants (SBS, HMM, MAP) that leverage the block-correlated nature of frequency-domain IN. These operate after the DFT, preserving subcarrier orthogonality.
Adaptive Complexity: The introduction of the MAP-EM variant, which uses sparsity-promoting priors to automatically reduce the number of estimated states, balancing performance and computational complexity.
System-Level Trade-off Analysis: A comprehensive evaluation identifying the optimal use-cases for each algorithm variant based on DFT size, null-tone overhead, and noise correlation parameters.

Simulation Results
The proposed framework was evaluated using a coded OFDM system with QPSK modulation under a Markov-Middleton noise channel.

Optimal vs. Conventional: The optimal NSI-aware receiver significantly outperforms the conventional AWGN receiver, particularly at small DFT sizes ( $M$ ) and high impulsive-to-background power ratios ( $\Lambda$ ). For instance, at $M=1$ and $\Lambda=5$ , the optimal receiver achieves a ~10 dB SNR gain at $10^{-4}$ BER compared to the AWGN receiver.
Impact of Block Size: As $M$ increases, the performance gap between the optimal and AWGN receivers narrows due to the CLT effect, unless high temporal correlation ( $r$ ) is present. High correlation preserves state stability, allowing the optimal receiver to maintain advantages even at larger block sizes.
EM Algorithm Performance:
- HMM-EM excels in scenarios with high temporal correlation ( $r \approx 0.99$ ) and low null-tone overhead, outperforming SBS-EM by accurately tracking state transitions.
- MAP-EM demonstrates robustness by converging to a parsimonious number of active states (typically 4–6 for the tested scenarios) from an initial 10, matching the performance of fixed 10-state models while reducing complexity.
- Overhead Sensitivity: In low-overhead scenarios (e.g., 12.5% null tones), HMM-based tracking is essential. In high-overhead scenarios, simpler SBS estimators suffice.
- Comparison to Benchmarks: The block-EM approaches significantly outperform the Block-Average (BAVG) variance estimator, which fails under low-overhead conditions due to high estimation uncertainty.

Significance and Claims
The paper claims that its frequency-domain design offers a robust alternative to time-domain mitigation by seamlessly integrating with the OFDM processing chain. By avoiding time-domain reconstruction, it eliminates the risk of destroying subcarrier orthogonality and reduces the need for complex iterative reconstruction algorithms.

The authors emphasize that while simplified models (like SBS-EM) are sufficient under high-overhead conditions, sophisticated variants (HMM-EM and MAP-EM) are critical for resource-constrained systems with limited null tones or highly correlated bursty noise. The work provides a practical pathway to achieving near-optimal performance in non-Gaussian environments without the immense complexity of time-domain impulse reconstruction. Future work is suggested to focus on optimizing initialization strategies and extending the framework to semi-blind, pilot-aided systems.

Technical Summary: Statistical Characterization and Block-EM Estimation of Frequency-Domain NSI for OFDM Systems in Bursty Impulsive Noise

More like this