Hermitian Positive Definite Neural Networks for Matrix-CFAR Geometric Signal Detection in Nonhomogeneous Clutter

Xiyao Chen, Takumi Sato, Linyu PengarXiv 2026 · arXiv

This paper, recently published by researchers at Keio University, tackles a major bottleneck in modern radar signal processing: the computational cost of "geometric" detection. By replacing slow, iterative math with a specialized neural network, they’ve found a way to make high-performance radar detection much faster without sacrificing accuracy.

TL;DR

The authors integrate Hermitian Positive Definite Neural Networks (HPDNet) into the Matrix-CFAR framework to accelerate signal detection in complex noise. By replacing computationally expensive iterative optimization with a learned, geometry-preserving mapping, they achieve significant speedups and robust performance on both simulated and real-world sea clutter.

What Problem Does This Paper Address?

In radar systems, the goal is to distinguish a target (like a ship or plane) from "clutter" (reflections from waves, rain, or ground). The standard approach is CFAR (Constant False Alarm Rate), which sets a detection threshold based on the surrounding noise.

Traditional CFAR looks at simple power levels, but Matrix-CFAR is more sophisticated. It treats the signal as a series of covariance matrices. These matrices don't live in a flat "Euclidean" space; they live on a curved mathematical surface called the Hermitian Positive Definite (HPD) manifold.

To detect a target, you calculate a "representative" matrix (a geometric mean or median) of the surrounding noise and compare it to the cell you're testing. The problem? Calculating these geometric averages usually requires iterative optimization algorithms. These are mathematically "heavy" and often too slow for real-time radar systems that need to process massive amounts of data instantly.

Methods

The authors propose using a Hermitian Positive Definite Neural Network (HPDNet) to do the heavy lifting.

The Architecture

Unlike standard neural networks that work on flat vectors, HPDNet is designed to stay on the "curved" manifold. It uses two specific types of layers:

  1. BiMap (Bilinear Mapping) Layer: Transforms the input matrices while preserving their structure.
  2. ReEig (Eigenvalue Rectification) Layer: Ensures the matrices remain "positive definite" (a strict requirement for covariance matrices) by nudging any tiny or negative eigenvalues back into the valid range.

The Training Strategy

Instead of training the network to classify "target vs. no target," they train it to be a geometric estimator. They feed it a set of noise matrices and use the actual Riemannian distance (the "distance" on the curved manifold) as the loss function. The network learns to output the geometric mean or median directly in a single forward pass, bypassing the need for iteration during testing.

The study evaluates three different geometric metrics:

  • AIRM (Affine-Invariant Riemannian Metric)
  • LE (Log-Euclidean)
  • BW (Bures–Wasserstein)

Datasets

The researchers used a two-pronged approach for evaluation:

  • Simulated Clutter: They generated "compound-Gaussian" clutter, which mimics the non-uniform, "spiky" noise often found in real environments.
  • Real-World Data: They used the IPIX sea clutter dataset, a famous benchmark in the radar community containing real radar returns from the ocean surface.
  • Generalization Test: Crucially, they trained the HPDNet only on simulated data and then tested it on the real IPIX data to see if the learned geometric mapping could generalize to the real world.

Key Insights and Ideas

  • Geometry as a Prior: The network isn't just "learning noise"; it's learning the underlying geometry of the HPD manifold. This is why a model trained on simulated math can work on real ocean waves.
  • Efficiency vs. Accuracy: For metrics like AIRM, which are notoriously slow, the HPDNet provided a massive speedup while actually improving detection probability in some scenarios.
  • Robustness to Outliers: Radar data is messy. Sometimes a target "leaks" into the cells used to estimate the noise. The authors showed that the HPDNet-based "median" estimator is just as robust against these outliers as traditional, slower robust statistics.

Key Results

  • Speedup: The HPDNet significantly reduced runtime for almost all metrics. For the AIRM and BW metrics—which usually require many iterations—the inference time was slashed by orders of magnitude.
  • Detection Performance: On simulated data, the HPDNet-based detector outperformed traditional Matrix-CFAR under the AIRM and LE metrics.
  • Real-World Success: When applied to the IPIX sea clutter, the HPDNet (trained only on simulations) achieved superior performance for the AIRM median and LE mean/median compared to classical methods.

Limitations and Future Work

  • Single Layer Focus: This study focused on a single-layer HPDNet to prove the concept of approximation. The authors suggest that deeper architectures (like Manifold-based ResNets or Transformers) could potentially capture even more complex clutter statistics.
  • Metric Expansion: While they tested three major metrics, the HPD manifold supports many others (like Bregman divergences) that could be integrated into this framework.
  • Hardware Deployment: While the algorithmic complexity is lower, the paper does not detail the power consumption or specific latency on embedded radar hardware (like FPGAs), which would be the next step for industrial application.

Summary updated

Abstract

Matrix information geometry (MIG) detectors exploit the Riemannian geometry of Hermitian positive definite (HPD) matrices to achieve robust object detection in nonhomogeneous clutter. The matrix constant false alarm rate (matrix-CFAR) method is a representative MIG detector that relies on the geometric structures with respect to specific metrics. In most cases, computing representative HPD covariance matrices requires numerical methods, typically involving computationally intensive iterative optimization algorithms. Although explicit formulas are available for certain cases, such as the log-Euclidean (LE) mean, they are generally unavailable for many other matrix means and medians. To address this challenge, we propose integrating the Hermitian positive definite neural network (HPDNet) into the matrix-CFAR framework, replacing conventional numerical optimization procedures with a learned, geometry-preserving mapping that directly estimates representative HPD covariance matrices from data. In the simulations, the framework is evaluated for MIG detectors using both means and medians associated with the affine-invariant Riemannian metric, the LE metric, and the Bures--Wasserstein metric, on both simulated compound-Gaussian clutter and real IPIX sea clutter. In both cases, the HPDNet is trained solely on simulated data to assess its applicability to the study of real data. The results show a significant reduction of computational cost in testing, while maintaining detection performance and robustness.
BibTeX
@misc{chen2026hermitian,
  author = {Xiyao Chen and Takumi Sato and Linyu Peng},
  title = {{Hermitian Positive Definite Neural Networks for Matrix-CFAR Geometric Signal Detection in Nonhomogeneous Clutter}},
  year = {2026},
  eprint = {2611.19820},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19820},
}
Figure