Detection and estimation of spikes in presence of noise and interference

Damien Passemier, Abla Kammoun, Mérouane DebbaharXiv 2014 · DOI · arXiv

Abstract

In many practical situations, the useful signal is contained in a low-dimensional subspace, drowned in noise and interference. Many questions related to the estimation and detection of the useful signal arise. Because of their particular structure, these issues are in connection to the problem that the mathematics community refers to as "spike detection and estimation". Previous works in this direction have been restricted to either determining the number of spikes or estimating their values while knowing their multiplicities. This motivates our work which considers the joint estimation of the number of spikes and their corresponding orders, a problem which has not been yet investigated to the best of our knowledge.

What Problem Does This Paper Address?

In signal processing, a common challenge is identifying a "useful" signal buried under significant noise and interference. Mathematically, this is often modeled using a spiked covariance model. Imagine a high-dimensional space where most of the data is random noise (the "bulk"), but specific directions contain the actual signal (the "spikes").

Empirical histogram of eigenvalues showing the noise bulk and signal spikes relative to the Marchenko-Pastur Law.

Existing research typically focuses on either counting the number of spikes or calculating their power assuming the count and grouping are already known. This paper addresses the joint estimation problem: in real-world scenarios like cognitive radio or sensor arrays, one often knows neither the number of distinct power levels ($K$) nor the multiplicities ($m_k$) of those levels. This research provides a framework to solve for both simultaneously.

Methods

The approach is based on Random Matrix Theory (RMT), specifically the behavior of eigenvalues as the number of observations ($n$) and the dimension of the signal ($p$) both grow to infinity.

  • The "Bulk" vs. "Spikes": According to the Marchenko-Pastur law, noise eigenvalues cluster into a predictable "bulk." If a signal is strong enough (above a specific phase-transition threshold), its corresponding eigenvalue will "pop out" of this bulk, as seen in the empirical histogram.
  • Joint Estimation Strategy: The authors propose a Bayesian/Maximum Likelihood approach. They prove that the sum of eigenvalues within a cluster follows a Gaussian distribution. By testing different values for $K$ (up to a maximum bound) and calculating which one maximizes the likelihood function based on these Gaussian distributions, they can infer the most probable number of sources and their multiplicities.

Key Insights and Results

The paper relies on synthetic simulations with dimensions up to $p=500$ and $n=1000$ samples to validate the theory. They tested three configurations (Models A, B, and C) with varying numbers of spikes and multiplicities.

Frequency of correct estimation for different signal models (A, B, and C) as a function of dimension/sample size.
  • High Accuracy and Consistency: The proposed estimator is "strongly consistent," meaning that as more data is collected, the probability of correct identification approaches 100%. In low-noise environments (SNR > 1.5 dB), the empirical probability of correctly identifying $K$ was nearly 99%.
  • Model Complexity: The method successfully distinguished between different models even as the total number of spikes increased, though convergence was slightly slower for more complex configurations like Model C.
  • Phase Transition Sensitivity: A spike is only detectable if its power $\alpha_k$ satisfies $\alpha_k > \sigma^2\sqrt{\gamma}$. Performance drops sharply once the noise variance approaches this RMT threshold, illustrating the "all-or-nothing" nature of phase transitions in high-dimensional data.

Limitations and Future Work

The method cannot detect signals that fall below the RMT phase-transition threshold; if the signal is too weak, the spikes remain merged with the noise bulk. Additionally, the Bayesian approach currently requires an a priori distribution for power levels, and the algorithm requires a predefined upper bound ($K_{max}$) for the number of clusters. Future work may focus on scenarios with completely unknown power distributions or higher computational efficiency for very large numbers of sources.

Summary updated

Abstract

In many practical situations, the useful signal is contained in a low-dimensional subspace, drown in noise and interference. Many questions related to the estimation and detection of the useful signal arise. Because of their particular structure, these issues are in connection to the problem that the mathematics community refers to as "spike detection and estimation". Previous works in this direction have been restricted to either determining the number of spikes or estimating their values while knowing their multiplicities. This motivates our work which considers the joint estimation of the number of spikes and their corresponding orders, a problem which has not been yet investigated to the best of our knowledge.
BibTeX
@misc{passemier2014detection,
  author = {Damien Passemier and Abla Kammoun and Mérouane Debbah},
  title = {{Detection and estimation of spikes in presence of noise and interference}},
  year = {2014},
  eprint = {1405.4846},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1405.4846},
}
Figure