2025-11-13T05:04:10.961087

Uncertainty Propagation in Finite Impulse Response Filters: Evaluating the Gaussian Assumption

Couchman, Stanley-Marbell
A common assumption in signal processing is that underlying data numerically conforms to a Gaussian distribution. It is commonly utilized in signal processing to describe unknown additive noise in a system and is often justified by citing the central limit theorem for sums of random variables, although the central limit theorem applies only to sums of independent identically distributed random variables. However, many linear operations in signal processing take the form of weighted sums, which transforms the random variables such that their distributions are no longer identical. One such operation is a finite impulse response (FIR) filter. FIR filters are commonly used in signal processing applications as a pre-processing step. FIR output noise is generally assumed to be Gaussian. This article examines the FIR output response in the presence of uniformly distributed quantization noise. We express the FIR output uncertainty in terms of the input quantization uncertainty and filter coefficients. We show that the output uncertainty cannot be assumed to be Gaussian, but depending on the application a Gaussian estimation may still be useful. Then, we show through detailed numerical simulations that the output uncertainty distribution of the filter can be estimated through its most dominant coefficients.
academic

Uncertainty Propagation in Finite Impulse Response Filters: Evaluating the Gaussian Assumption

Basic Information

  • Paper ID: 2510.11384
  • Title: Uncertainty Propagation in Finite Impulse Response Filters: Evaluating the Gaussian Assumption
  • Authors: Jennie Couchman, Phillip Stanley-Marbell
  • Classification: eess.SP (Signal Processing)
  • Publication Date: October 13, 2025 (arXiv preprint)
  • Paper Link: https://arxiv.org/abs/2510.11384v1

Abstract

This research challenges a common assumption in signal processing: that underlying data numerically follows a Gaussian distribution. While this assumption is typically justified by invoking the Central Limit Theorem (CLT), the CLT only applies to sums of independent and identically distributed random variables. However, many linear operations in signal processing take the form of weighted sums, which violates the identical distribution requirement. This paper specifically examines Finite Impulse Response (FIR) filtering, investigating FIR output responses in the presence of uniformly distributed quantization noise. The study demonstrates that output uncertainty cannot be assumed to follow a Gaussian distribution; however, depending on the application context, Gaussian approximation may still be useful.

Research Background and Motivation

Core Problem

This research addresses a fundamental theoretical assumption in signal processing: Does the uncertainty distribution of FIR filter outputs truly follow a Gaussian distribution?

Problem Significance

  1. Accuracy of Theoretical Foundations: The Gaussian assumption is ubiquitous in signal processing, affecting the design and performance evaluation of subsequent processing steps
  2. Reliability of Practical Applications: Incorrect distributional assumptions may lead to inaccurate uncertainty propagation analysis
  3. Guidance for Engineering Practice: Provides a more accurate theoretical foundation for practical system design

Limitations of Existing Methods

  1. Misapplication of Central Limit Theorem: Engineers frequently incorrectly apply CLT to weighted sums, whereas CLT only applies to independent and identically distributed variables
  2. Lack of Rigorous Verification: Existing research lacks rigorous mathematical analysis of FIR filter output distributions
  3. Disconnect Between Theory and Practice: Theoretical assumptions lack sufficient numerical verification

Core Contributions

  1. Mathematical Expression Derivation: First explicitly expresses FIR filter output uncertainty as a function of input uncertainty and filter coefficients
  2. Theoretical Proof: Rigorously proves that FIR filter outputs cannot be assumed to follow a Gaussian distribution
  3. Approximation Method: Proposes a practical method for estimating output uncertainty distribution using only the dominant filter coefficients
  4. Experimental Validation: Large-scale numerical simulations based on real EEG data validate the theoretical analysis results

Methodology Details

Task Definition

Investigate the propagation of quantization uncertainty in FIR filters, specifically analyzing the true characteristics of output distribution when input consists of uniformly distributed quantization noise.

Theoretical Framework

Quantization Uncertainty Model

The i-th sample of the input signal is modeled as a uniformly distributed random variable: Xi∼U(μi−δ2,μi+δ2)X_i \sim U\left(\mu_i - \frac{\delta}{2}, \mu_i + \frac{\delta}{2}\right)

Its probability density function is:

\frac{1}{\delta} & \text{if } |x_i - \mu_i| < \frac{\delta}{2} \\ 0 & \text{otherwise} \end{cases}$$ #### FIR Filter Output Modeling The FIR filter output is expressed as a weighted sum: $$Y_n = \sum_{i=0}^{N} b_i X_{n-i}$$ The output mean and variance are respectively: $$\mu = \sum_{i=0}^{N} b_i \mu_{i-n}$$ $$\sigma^2 = \sum_{i=0}^{N} b_i^2 \frac{\delta^2}{12}$$ #### Exact Distribution Derivation Based on the closed-form solution from Kamgar-Parsi et al., the output probability density function is: $$f_{Y_n}(y_n) = \frac{(-1)^{N+1}}{N! \tilde{b}} \sum_{s_0=\pm1} \cdots \sum_{s_N=\pm1} \tilde{s} \times (y_n - \mu - S)^N \times \Theta(y_n - \mu - S)$$ Where: - $\tilde{b} = \prod_{i=0}^{N} b_i \delta$ - $\tilde{s} = \prod_{i=0}^{N} s_i$ - $S = \sum_{i=0}^{N} s_i b_i \frac{\delta}{2}$ - $\Theta(t)$ is the Heaviside step function ### Technical Innovations 1. **Dominant Coefficient Approximation**: Since complete computation is infeasible for high-order filters, this approach uses only dominant coefficients exceeding 5% of the maximum coefficient magnitude 2. **Distribution Comparison Framework**: Employs Jensen-Shannon distance to quantitatively compare differences between actual and theoretical distributions 3. **Statistical Testing Method**: Applies D'Agostino & Pearson test to verify non-Gaussianity ## Experimental Setup ### Dataset - **Data Source**: Physionet EEGBCI dataset - **Configuration**: 64-electrode EEG system, sampling frequency 160 Hz, amplitude resolution 1 μV - **Task**: Motor imagery task (left/right hand clenching imagination) - **Simulation Scale**: 5,000 repeated simulations generating 75,735 distinct output distributions ### Filter Configuration - **Type**: Bandpass FIR filter - **Frequency Band**: 7-35 Hz (corresponding to mu and beta bands for motor imagery) - **Parameters**: - Low cutoff frequency: 7 Hz, transition bandwidth 2 Hz - High cutoff frequency: 35 Hz, transition bandwidth 8.75 Hz - Window function: Hamming window - Filter length: 265 samples (1.656 seconds) ### Evaluation Metrics 1. **Jensen-Shannon Distance**: Quantifies distribution differences 2. **D'Agostino & Pearson Test**: Tests for normality 3. **Kurtosis Analysis**: Evaluates distribution shape characteristics 4. **Data Proportion Within Standard Deviation**: Compares distribution concentration ### Quantization Noise Simulation Uniform quantization noise is added: $U(-0.5\mu V, 0.5\mu V)$ ## Experimental Results ### Main Results #### Non-Gaussianity Verification - **Statistical Significance**: 99.7% of output distributions reject the Gaussian null hypothesis at p<0.05 level - **Kurtosis Characteristics**: All output distributions are platykurtic, with average excess kurtosis of -0.2924 #### Distribution Fitting Comparison - **Weighted Sum Distribution vs. Gaussian Distribution**: - Average Jensen-Shannon distance to weighted sum theoretical distribution: 0.03184 - Average Jensen-Shannon distance to Gaussian distribution: 0.03661 - Weighted sum distribution is significantly closer to actual output distribution #### Distribution Characteristics Analysis - **Gaussian Distribution**: 38.29% of values fall within mean ± 0.5 standard deviations - **Actual Output**: 36.81% of values fall within mean ± 0.5 standard deviations - **Weighted Sum Distribution**: 37.00% of values fall within mean ± 0.5 standard deviations ### Effectiveness of Dominant Coefficient Approximation By using only 19 dominant coefficients (selected from 265), the method successfully approximates the complete theoretical distribution, demonstrating the feasibility of the computational simplification approach. ### Case Analysis Figure 2 presents 9 representative cases covering different electrodes (C3, C4, Cz), different time points, and different trials, consistently showing: 1. Actual distributions clearly deviate from the Gaussian assumption 2. Weighted sum theoretical distribution fits actual data well 3. Jensen-Shannon distance quantitatively confirms this observation ## Related Work ### Central Limit Theorem Extensions for Weighted Sums The paper cites research by Cuzick (1995), Weber (2006), Avena & da Costa (2024) on conditions for weighted sum central limit theorems, but notes these analyses have not been applied to FIR filters. ### Filtering Applications in EEG Signal Processing - **Motor Imagery Research**: Work by McFarland et al. (2000), ter Horst et al. (2013), Yu et al. (2022) established associations between mu/beta bands and motor imagery - **BCI Systems**: Schalk et al. (2004) BCI2000 system provides standard platform for EEG data acquisition ### Uncertainty Propagation Theory Based on Kamgar-Parsi et al. (1995) distribution theory for weighted sums of uniform random variables, this work applies it for the first time to digital signal processing. ## Conclusions and Discussion ### Main Conclusions 1. **Theoretical Breakthrough**: FIR filter output uncertainty distribution is indeed not Gaussian but rather a platykurtic distribution with specific shape characteristics 2. **Practical Value**: Although non-Gaussian, Gaussian approximation may still be acceptable in certain applications, requiring evaluation for specific use cases 3. **Computational Optimization**: The dominant coefficient method provides a practical computational simplification pathway ### Limitations 1. **Specific Noise Model**: Study only considers uniformly distributed quantization noise 2. **Linear System Restriction**: Analysis applies only to linear FIR filters 3. **Application Scenario Evaluation**: Requires case-by-case assessment of Gaussian approximation acceptability ### Future Directions 1. **Extension to Other Noise Models**: Investigate input uncertainty with other distribution types 2. **Nonlinear System Analysis**: Extend to IIR filters and nonlinear systems 3. **Real-Time Computational Optimization**: Develop more efficient online uncertainty propagation algorithms ## In-Depth Evaluation ### Strengths #### Theoretical Contributions 1. **Rigorous Mathematical Analysis**: First provides rigorous mathematical framework for FIR filter uncertainty propagation 2. **Practical Computational Method**: Dominant coefficient approximation method addresses computational complexity 3. **Comprehensive Experimental Validation**: Large-scale simulations provide strong empirical support #### Methodological Advantages 1. **Statistical Test Rigor**: Employs multiple statistical methods to verify conclusion reliability 2. **Real Data Validation**: Verification using standard EEG datasets enhances result credibility 3. **Quantitative Analysis**: Jensen-Shannon distance provides objective distribution comparison standard #### Practical Value 1. **Engineering Guidance**: Provides important theoretical correction for signal processing engineers 2. **Computational Feasibility**: Proposed approximation method has practical application value ### Weaknesses #### Theoretical Limitations 1. **Single Noise Model**: Only considers uniform distribution; actual system noise may be more complex 2. **Linear Assumption**: Does not address effects of nonlinear processing stages 3. **Independence Assumption**: Assumes input samples are mutually independent, which may not hold in some practical situations #### Experimental Design 1. **Limited Application Domain**: Validation only on EEG data, lacking verification on other signal types 2. **Single Filter Type**: Tests only one specific FIR filter configuration 3. **Lack of Comparative Baselines**: No comparison with other uncertainty propagation methods #### Practical Considerations 1. **Computational Complexity**: Although approximation method is proposed, challenges remain for ultra-large-scale systems 2. **Parameter Selection**: The 5% threshold for dominant coefficients lacks theoretical justification 3. **Real-Time Applicability**: Does not discuss feasibility for real-time system applications ### Impact #### Academic Contribution 1. **Theoretical Correction Value**: Corrects long-standing erroneous assumptions with significant academic value 2. **Interdisciplinary Impact**: Provides insights relevant to signal processing, statistics, and metrology 3. **Methodological Innovation**: Dominant coefficient approximation may inspire analysis of other complex systems #### Practical Value 1. **Engineering Application Guidance**: Provides more accurate theoretical foundation for filter design and uncertainty analysis 2. **Standards Development Reference**: May influence development and revision of relevant technical standards 3. **Educational Value**: Provides important theoretical supplement to signal processing education #### Reproducibility 1. **Open-Source Tools**: Based on open-source tools like MNE, facilitating reproduction 2. **Public Datasets**: Uses standard public datasets enhancing reproducibility 3. **Detailed Method Description**: Provides sufficient implementation details ### Applicable Scenarios #### Direct Application Domains 1. **Biomedical Signal Processing**: Filtering of EEG, ECG, and other biological signals 2. **Communication Systems**: Channel filtering and equalization in digital communications 3. **Audio Processing**: Audio signal denoising and frequency selection #### Extended Application Potential 1. **Control Systems**: Uncertainty analysis of filtering stages in feedback control 2. **Image Processing**: Uncertainty propagation in digital image filtering 3. **Sensor Networks**: Uncertainty management in distributed sensor data fusion #### Theoretical Research Value 1. **Stochastic Process Theory**: Provides new insights into random process behavior in linear systems 2. **Measurement Uncertainty**: Supplements uncertainty propagation theory in measurement science 3. **Numerical Analysis**: Provides reference for error propagation analysis in numerical computation ## References The paper cites 12 important references covering weighted sum distribution theory, EEG signal processing, and statistical testing methods, providing solid theoretical foundation. Particularly noteworthy is the pioneering work by Kamgar-Parsi et al. (1995), which establishes the theoretical framework for this research. --- **Overall Assessment**: This is a high-quality research paper with significant theoretical value and practical significance. It not only corrects a common misconception in signal processing but also provides practical solutions. While there are certain limitations in research scope and experimental design, the core contributions are sufficiently important to positively advance development in related fields.