Find your next learning step
Two radar cells can become asymptotically independent and still exceed together far more often than an independence calculation suggests. This guide makes that residual tail risk measurable before it distorts a CFAR threshold.
Open the chapter contents · Radar root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The teaching chapters remain open; these grouped routes form the index.
Route 1 · Read the tail question
Route 2 · Build the hidden model
Route 3 · Connect it to CFAR
Route 4 · Estimate and challenge
Route 5 · Check and transfer
CHAPTER 1 OF 12 · READ THE TAIL QUESTION
1. When asymptotic independence is not ordinary independence
Radar false-alarm models often ask whether two high clutter returns occur together. “Asymptotically independent” means that, as the threshold moves without bound, the conditional chance that the second variable also exceeds the threshold tends to zero. It does not mean that finite high thresholds behave like independent draws.
That distinction is operational. A million-cell range–Doppler map can contain joint exceedances at levels far beyond the data used to fit a threshold. If a model sets the joint probability equal to the product of marginal probabilities too early, it can understate clustered or paired alarms even when the limiting tail-dependence coefficient is zero.
CHAPTER 2 OF 12 · READ THE TAIL QUESTION
2. Put the margins on a common tail scale
Dependence diagnostics should not be confused with different marginal scales. Transform each margin to a common rank, uniform or standard-Pareto scale before examining the joint tail. On a uniform scale, let u be a high quantile and define chi of u as the probability that V exceeds u given that U exceeds u.
If chi of u approaches a positive number, the pair is asymptotically dependent. If it approaches zero, the pair is asymptotically independent. A declining estimate is evidence to investigate, not proof by itself: few exceedances, threshold drift and serial clustering can all create a fall.
CHAPTER 3 OF 12 · READ THE TAIL QUESTION
3. Worked threshold sequence: chi falls but risk remains
Suppose 100,000 paired cell observations are converted to uniform ranks. At u equals 0.99, there are about 1,000 U exceedances and 80 joint exceedances, so chi hat is 0.080. At 0.995, the counts are 500 and 26, giving 0.052. At 0.999, they are 100 and 2, giving 0.020.
| Uniform threshold u | U exceedances | Joint exceedances | Estimated chi of u | Joint exceedances expected under independence |
|---|---|---|---|---|
| 0.990 | 1,000 | 80 | 0.080 | 10.0 |
| 0.995 | 500 | 26 | 0.052 | 2.5 |
| 0.999 | 100 | 2 | 0.020 | 0.1 |
The last column uses n times one minus u squared. The observed joint counts fall relative to the anchor exceedances but remain much larger than this product rule. That is the residual tail-risk signal hidden by the binary label “asymptotically independent.”
CHAPTER 4 OF 12 · BUILD THE HIDDEN MODEL
4. The residual coefficient and hidden regular variation
Ledford and Tawn model the joint survivor probability on a standard tail scale as a slowly varying factor times a power governed by eta. Under their common convention, eta equals one for asymptotic dependence, one half for independent standard-Pareto components, and lies between one half and one for positively associated asymptotic independence. Some fields use reciprocal or transformed tail-order parameters, so the equation and convention must accompany the number.
Hidden regular variation refines the model on a smaller cone where both components are large. Ordinary multivariate regular variation may place its limiting mass on the axes and therefore say little about simultaneous extremes. The hidden limit describes the thinner interior tail. It is “hidden” because the first-order limit discards that scale, not because the risk is unobservable.
The foundational sources are Ledford and Tawn’s near-independence model and Maulik and Resnick’s characterisations of hidden regular variation.
CHAPTER 5 OF 12 · BUILD THE HIDDEN MODEL
5. What this changes in a CFAR model
Constant false alarm rate (CFAR) processing calibrates a threshold from reference cells under a clutter model. Marginal calibration controls the probability that one test statistic crosses its threshold. It does not automatically control the chance that neighbouring cells, channels or polarimetric features cross together.
A hidden-tail component can therefore matter to alarm multiplicity, track initiation and post-detection grouping even when the single-cell false-alarm rate is correct. It should not be inserted as an arbitrary threshold multiplier. First define the joint event that matters: two adjacent range cells, range and Doppler channels, two looks, or two derived statistics. Then estimate its finite-threshold probability and propagate it through the downstream decision rule.
The extreme-value CFAR guide covers marginal peaks-over-threshold modelling. This article supplies the missing joint-tail layer.
CHAPTER 6 OF 12 · CONNECT IT TO CFAR
6. Direction, lag and conditioning are part of the event
Joint tails can be directional. Conditioning on a large leading cell and asking about its neighbour need not match the reverse direction when the margins, geometry or processing differ. A common rank scale removes marginal units, but it does not remove physical asymmetry.
Estimate ordered conditional exceedance curves and several lags. Compare range, Doppler and time directions rather than pooling them by default. The tail-copula CFAR guide shows how directional joint surfaces reveal these differences, while the extremogram guide tracks lagged tail dependence.
CHAPTER 7 OF 12 · CONNECT IT TO CFAR
7. Changed context: extrapolation beyond the fitted threshold
Imagine a system evaluating one million cell pairs. Two candidate tail models agree at u equals 0.99 but differ higher up. Model A uses the independence product. Model B is asymptotically independent with residual eta above one half, calibrated to the observed joint tail. The table uses teaching probabilities to expose the consequence rather than claim a fitted radar law.
| Threshold u | Independence joint probability | Residual-tail joint probability | Expected joints in one million pairs, A | Expected joints in one million pairs, B |
|---|---|---|---|---|
| 0.990 | 0.000100 | 0.000800 | 100 | 800 |
| 0.999 | 0.000001 | 0.000020 | 1 | 20 |
| 0.9999 | 0.00000001 | 0.00000050 | 0.01 | 0.5 |
The transfer question is not “does chi equal zero?” It is “at the deployed map size and threshold, how many joint events does each plausible residual-tail model predict?”
CHAPTER 8 OF 12 · CONNECT IT TO CFAR
8. A practical estimation workflow
Begin with quality-controlled clutter-only data representative of the operating regime. Remove deterministic trends that would make ranks incomparable, while retaining the dependence the decision must face. Transform margins, choose a grid of high thresholds, and plot chi of u with block or cluster-aware uncertainty.
Estimate eta or an equivalent tail-order parameter over a range where the curve is reasonably stable. Repeat by direction, lag and regime. Use held-out windows to compare predicted and observed joint counts. Threshold choice is a bias–variance decision: too low mixes ordinary dependence with tail behaviour; too high leaves too few joint events.
Do not treat every reference-cell pair as independent evidence. The extremal-index CFAR guide explains how clusters reduce the effective number of tail events.
CHAPTER 9 OF 12 · ESTIMATE AND CHALLENGE
9. From a tail estimate to a threshold check
For each candidate threshold, calculate the marginal false-alarm rate, the joint exceedance rate for the event of interest, and the distribution of alarm-cluster sizes. Compare these with explicit system tolerances. A threshold can pass the single-cell requirement and fail the paired-alarm or track-initiation requirement.
Calibration should be conditional on relevant regimes such as sea state, weather, scan angle or processing mode. The conditional-coverage CFAR guide provides the monitoring logic. Hidden-tail diagnostics add a joint coverage panel, not a replacement for marginal calibration.
CHAPTER 10 OF 12 · ESTIMATE AND CHALLENGE
10. Limits, evidence and common misreadings
Hidden regular variation is an asymptotic model. Finite samples may not reveal a stable hidden scale, and extrapolation can be more uncertain than a point estimate suggests. Marginal transformation error, nonstationarity, target contamination and threshold selection all affect the result. Confidence intervals should reflect clustering and repeated model choices where possible.
Asymptotic independence does not mean no simultaneous extremes. Eta above one half is not the same as positive limiting chi. A good-looking straight line on a log plot is not sufficient validation. Radar-specific copula and CFAR work, including dependence modelling for CFAR and copula-based radar detection research, supports modelling dependence explicitly; the deployed joint event and clutter regime still need their own evidence.
CHAPTER 11 OF 12 · CHECK AND TRANSFER
11. Questions readers usually ask
Why not fit one copula and read its tail coefficient?
A copula family imposes a tail shape. Threshold curves and residual-order diagnostics show whether that shape is credible before far-tail extrapolation.
Can chi equal zero while eta is informative?
Yes. Chi records the limiting conditional probability, while eta describes the rate at which the joint survivor probability decays when that limit is zero.
Is eta below one half possible?
Some models permit negative tail association, but interpretation depends on the convention and support. State the formula and do not compare bare parameter values across papers.
Does this replace CFAR loss analysis?
No. It adds a joint false-alarm and clustering check. Detection loss, target fluctuation and threshold estimation remain separate parts of the radar budget.
CHAPTER 12 OF 12 · CHECK AND TRANSFER
12. A transfer routine and connected reading
Define the simultaneous event, standardise the margins, plot chi across thresholds, estimate residual tail order with cluster-aware uncertainty, and convert the fitted joint probability into expected operational events. Repeat by direction and regime. Reject any conclusion that changes when one or two joint exceedances are removed.
Continue with compound-Gaussian clutter and false-alarm inflation, posterior-predictive CFAR calibration and order-statistic CFAR. The reusable lesson is precise: a vanishing limiting coefficient can still leave a consequential finite joint tail.
Top · Contents · Previous · How X Works