A radar decides whether a target is present in a test cell by comparing the cell against an estimate of the background interference covariance. That estimate is built from a few neighbouring "secondary" cells, which are assumed to contain clutter only. Three problems make this hard in practice. There are often fewer secondary cells than array channels. The clutter is spiky rather than Gaussian. And the operator needs the false alarm rate to stay fixed at a known value, which is the constant false alarm rate (CFAR) requirement. Each problem has a known remedy. Regularized Tyler estimators handle spikes and few samples. Persymmetry is the mirror symmetry that clutter covariance has for symmetric arrays and evenly spaced pulses, and it effectively doubles the number of samples. Random-matrix and central-limit-theorem results let the detection threshold be predicted in closed form. Nobody has combined all three, and the combination fails in a specific way: the obvious combined estimator is easy to build, but every existing tool for choosing its shrinkage and setting its threshold assumes independent samples. The persymmetric estimator violates that assumption by construction.
This note proposes to build that estimator and then derive the shrinkage rule and false alarm threshold that correctly account for its paired-sample structure. The goal is a detector that meets its nominal false alarm rate in heavy-tailed clutter, including in the regime where secondary cells number between half and all of the channels, without the heuristic subspace padding currently used there. The testable claim is that naive plug-in calibration miscalibrates in a measurable way and the corrected theory does not.
Core Claim
Pailloux et al. (Persymmetric Adaptive Radar Detectors) map a complex persymmetric covariance to a real symmetric one with a unitary transform. They derive a persymmetric fixed-point Tyler estimator and a texture-CFAR GLRT, but only for $K \ge N$ secondary samples, where $N$ is the number of channels. Chen, Wiesel and Hero (Robust Shrinkage Estimation of High-Dimensional Covariance Matrices) add shrinkage to Tyler's iteration so that it converges uniquely for any sample size, but they ignore structure.
The first step is to observe that the two combine trivially. Augment each secondary sample $\mathbf{x}_i$ with its "backward" copy $\mathbf{J}\mathbf{x}_i^*$, where $\mathbf{J}$ is the anti-identity matrix and $^*$ denotes complex conjugation. Then run the ordinary regularized Tyler iteration on the $2K$ augmented vectors:
$$\hat{\mathbf{\Sigma}} = (1-\rho)\frac{N}{2K}\sum_{i=1}^{K}\frac{\mathbf{x}_i\mathbf{x}_i^H + \mathbf{J}\mathbf{x}_i^*\mathbf{x}_i^T\mathbf{J}}{\mathbf{x}_i^H\hat{\mathbf{\Sigma}}^{-1}\mathbf{x}_i} + \rho\mathbf{I}.$$
Here $\rho\in(0,1]$ is the shrinkage weight. The update maps persymmetric matrices to persymmetric matrices, and the Chen–Wiesel–Hero fixed point is unique. Together these imply that the fixed point is persymmetric, that the two copies of each sample receive the same Tyler weight, and that the result equals a regularized version of the Pailloux estimator, in which the real and imaginary parts share one texture.
We therefore do not claim the estimator as the contribution. The contribution is what follows from this construction. The $2K$ augmented vectors come in dependent pairs. The shrinkage oracles in current use (Chen's CWH and Ollila's OTy) and the Couillet–Kammoun–Pascal central limit theorem for bilinear forms (Second order statistics of robust estimators of scatter) all assume independent samples. If one plugs in "$2K$ samples", the threshold is set for the wrong null distribution. We claim three things. The size of this miscalibration depends on the unknown covariance through how strongly $\mathbf{x}$ and $\mathbf{J}\mathbf{x}^*$ are correlated. It can be corrected in closed form. And once corrected, the detector holds its nominal false alarm probability for $N/2 < K < N$ without the $\hat{\alpha}_0\mathbf{I}_{N-K}$ padding that Hoarau et al. (A subspace approach for shrinkage parameter selection) describe as heuristic.
Research Program
The first piece is existence and equivalence. We will prove that the augmented regularized iteration and the transform-domain regularized persymmetric fixed point coincide. We will then determine when the unregularized limit ($\rho\to 0$) exists. In real space there are $2K$ real vectors of dimension $N$, but each pair is constrained to share a texture, so it is not obvious that the existence condition is simply $2K > N$. The answer determines exactly where shrinkage becomes necessary rather than merely helpful.
The second piece is fluctuation theory. Couillet, Kammoun and Pascal show that bilinear forms $\mathbf{a}^H\hat{\mathbf{C}}(\rho)^{-1}\mathbf{b}$ of the regularized Tyler estimator fluctuate like those of a simpler, tractable matrix, at rate $N^{-1/2}$. We will redo this argument for the paired model. The pairing enters through the cross-moment $\mathbb{E}[\mathbf{x}\mathbf{x}^T]\mathbf{J}$. That moment vanishes only in special cases, so it should add a correction term to the limiting variance of the GLRT statistic. That variance is what sets the threshold through the predicted false alarm probability. We will also check whether the effect weakens or strengthens when the steering vector is itself persymmetric, as it is for a ULA steered off broadside.
Kammoun et al. (Convergence and Fluctuations of Regularized Tyler Estimators) argue that the fixed-$N$, large-$K$ approximation is more accurate at small dimensions. Robinson et al. (Space-Time Adaptive Detection at Low Sample Support) and Vallet et al. (An Improved Low Rank Detector in the High Dimensional Regime) insist that the $N/K\to c$ regime is necessary in space-time adaptive processing. We will derive both versions of the persymmetric correction and let the experiments decide which one predicts false alarms at realistic radar sizes. That turns the methodological tension in the literature into a measurement.
The third piece is shrinkage selection. We will derive a CWH-style oracle for $\rho$ that uses the paired second-order statistics. The aim is to show that it stays inside $[0,1]$ down to about $K\approx N/2$. Hoarau et al. found that the unstructured oracle leaves this interval when $K\ll N$, which is why they resort to subspace projection.
Baselines
The naive reference is the adaptive matched filter built on the sample covariance with diagonal loading. Its load is set by grid search on simulated clutter, using the same tuning budget we give our own $\rho$. The detector is then run on the identical secondary cells.
The strongest structured Gaussian competitor is the persymmetric adaptive matched filter from Pailloux et al. with forward-backward averaging, which is defined only for $K\ge N/2$ or loaded below that. The strongest robust competitor at full support is their GLRT with the persymmetric fixed-point estimator, which is valid for $K\ge N$.
In the undersampled regime, the strongest existing method is unstructured regularized Tyler with Hoarau et al.'s subspace-projected CWH oracle. On the real ENGIE ground-penetrating-radar data, they report $P_{FA}=1.4\times10^{-3}$ for subspace CWH, $2.1\times10^{-3}$ for subspace OTy, and $1.15\times10^{-2}$ for standard OTy. The file does not state the nominal false alarm target for those figures, so we will read them as calibration error relative to a target we recover from their paper. We will beat this baseline only if our empirical-to-nominal false alarm ratio is closer to one at the same $K/N$ and our detection probability at matched false alarm rate is no lower.
We also include Robinson et al.'s Ledoit–Wolf Direct nonlinear shrinkage, as the strongest Gaussian-model low-sample method. The file has no reported scalar result for it on KASSPER I, so we will rerun it. Ginolhac et al.'s persymmetric low-rank filter is included with the rank given to it as an oracle (Exploiting persymmetry for low-rank Space Time Adaptive Processing).
The decisive ablation keeps our estimator fixed and swaps only the calibration. One variant applies the existing independent-sample oracle and threshold with $2K$ as the sample size. The other applies our paired correction. As an upper reference, a Monte Carlo threshold computed with the true covariance shows how much calibration error any closed form leaves.
Why not just…?
Why not just run ordinary regularized Tyler on forward-backward augmented data and use the existing threshold formula? This is the most serious alternative, because, as argued above, it produces exactly our estimator. The only question is whether the independent-sample theory still sets the threshold correctly. If our ablation shows the plug-in false alarm rate stays within Monte Carlo error of nominal across textures and covariances, the paired correction is unnecessary. We would report that as the result and stop. Our expectation is that it fails when $\mathbf{x}$ and $\mathbf{J}\mathbf{x}^*$ are strongly correlated, for example with narrow clutter ridges near broadside. The experiment will settle this.
Why not just set the threshold empirically from clutter-only cells? At $P_{FA}=10^{-4}$ this needs tens of thousands of clutter-only snapshots drawn from the same covariance. Such snapshots are not available in nonhomogeneous scenes. That data scarcity is the very reason for estimating covariance from a handful of cells. We will nonetheless include the empirical threshold wherever the real datasets have enough range-Doppler cells, to show its variance.
Why not just use the Hoarau subspace method? It works below $N/2$, where persymmetry cannot help, and we keep it as the method of choice there. Between $N/2$ and $N$, however, it discards the mirror structure and relies on identity padding of the unobserved subspace, which its authors call heuristic. The head-to-head comparison over $K/N$ will locate where each method is better calibrated.
Why not just use a learned detector such as the FAR-controllable deep forest? That model reports average accuracy above 99.19% on IPIX (Multi-Feature Fusion for Weak Target Detection). It needs labeled data and offers no invariance guarantee under shifts in clutter texture or power. We will include it on IPIX as a detection-probability reference, but not as a calibration competitor.
Proposed Experiments
The synthetic experiments use Toeplitz covariances, which are Hermitian Toeplitz and therefore persymmetric. Textures are Gamma, Weibull and Student-$t$, with $N\in\{16,32,64\}$ and $K/N$ from $0.4$ to $3$. Each configuration estimates false alarm probability down to $10^{-4}$ over at least $10^{6}$ trials.
The primary check is calibration. Across all textures and covariances, the empirical-to-nominal false alarm ratio of the corrected detector should stay near one and flat. The plug-in variant should drift systematically with the covariance's forward-backward correlation. A second check compares fixed-$N$ and large-dimensional predictions at each $N$. A third tests robustness by perturbing persymmetry with simulated array imperfections, to find the mismatch level at which the structural gain disappears.
On real data, we use clutter that the file lists with persymmetric structure: the Thales ground clutter, the ONERA Nostradamus sea clutter, and the DGA/MI space-time datacube with injected targets. For all three, access must be negotiated with the owners. We also use the public IPIX database, where temporal persymmetry under constant pulse spacing can be tested directly before the method is applied.
Metrics on real data are empirical false alarm rate against nominal, detection probability versus signal-to-clutter ratio at matched false alarm rate, and SINR loss. On the DGA/MI datacube, Ginolhac et al.'s reported clutter reduction of up to 6 dB serves as the bar for the structured baseline.
The proposal is confirmed if the corrected detector matches nominal false alarm probability within Monte Carlo error for $N/2<K<N$ in non-Gaussian clutter, while the plug-in calibration and the subspace-padding method do not. The proposal is falsified if plug-in calibration performs just as well.
References
- Persymmetric Adaptive Radar Detectors
- Robust Shrinkage Estimation of High-Dimensional Covariance Matrices
- Second order statistics of robust estimators of scatter. Application to GLRT detection for elliptical signals
- A subspace approach for shrinkage parameter selection in undersampled configuration for Regularised Tyler Estimators
- Convergence and Fluctuations of Regularized Tyler Estimators
- Space-Time Adaptive Detection at Low Sample Support
- An Improved Low Rank Detector in the High Dimensional Regime
- Exploiting persymmetry for low-rank Space Time Adaptive Processing
- Multi-Feature Fusion for Weak Target Detection on Sea-Surface Based on FAR Controllable Deep Forest Model