Find your next learning step
A Bayesian CFAR detector can integrate uncertainty inside one clutter family, yet still be overconfident if the family itself is uncertain. Bayesian model averaging carries several plausible clutter models into one predictive tail and derives the threshold from that mixture.
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 · Build the model mixture
Route 2 · Work the threshold
Route 3 · Stress the weights
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · BUILD THE MODEL MIXTURE
1. Model uncertainty sits above parameter uncertainty
The Bayesian CFAR guide integrated uncertainty in an unknown clutter rate. It still conditioned on one chosen likelihood. In practice, exponential, compound-Gaussian, Pareto-like or other clutter families may all look plausible over a limited reference window.
Bayesian model averaging, or BMA, assigns posterior weights to candidate models and averages their predictive distributions. A CFAR threshold is then the quantile of the predictive mixture, not a vote among separately selected detectors.
This article extends the Radar root, Bayesian posterior-predictive CFAR and compound-Gaussian clutter diagnostics.
CHAPTER 2 OF 12 · BUILD THE MODEL MIXTURE
2. Average predictive tails before finding the threshold
Let m index candidate clutter models and D denote reference data. The averaged predictive density is p(x|D) = Σₘ p(m|D)p(x|D,m). For a desired clutter tail α, select t so that Σₘ wₘP(X>t|D,m) = α, where wₘ = p(m|D).
This order matters. Quantiles are nonlinear, so averaging model-specific thresholds generally does not produce the mixture quantile. First average the tail probabilities at a proposed t; then solve for t.
Posterior model weights combine prior model probabilities with marginal likelihoods. They are not permanent performance scores. A changed sea state, resolution cell or preprocessing chain can change both candidate adequacy and weights.
CHAPTER 3 OF 12 · WORK THE THRESHOLD
3. Worked two-model threshold
Use two invented predictive tails. Model A is the eight-cell exponential result A(t) = [40/(40+t)]⁸. Model B is a heavier Lomax-style tail B(t) = (1+t/20)⁻³. Give them posterior weights 0.70 and 0.30.
| Quantity | Model A | Model B | Mixture |
|---|---|---|---|
| Posterior weight | 0.70 | 0.30 | 1.00 |
| Individual α = 0.01 threshold | 31.13 | 72.83 | Not averaged directly |
| Tail at mixture threshold 45.77 | 0.00224 | 0.02811 | 0.01000 |
The mixture threshold is approximately 45.77. Model A contributes little tail probability there, while the smaller weight on the heavy-tailed model still matters strongly.
CHAPTER 4 OF 12 · WORK THE THRESHOLD
4. Averaging thresholds misses the target
The weighted average of individual thresholds is 0.70×31.13 + 0.30×72.83 = 43.64. At that value, the predictive mixture tail is about 0.01123, not 0.01000.
| Rule | Threshold | Mixture tail |
|---|---|---|
| Correct predictive-mixture inversion | 45.77 | 0.01000 |
| Weighted average of thresholds | 43.64 | 0.01123 |
| Model A alone | 31.13 | Mixture tail above target |
| Model B alone | 72.83 | Mixture tail below target |
The numerical gap depends on model separation and weights. Near-identical models may hide the mistake; that does not make the operation valid.
CHAPTER 5 OF 12 · WORK THE THRESHOLD
5. Model weights need an honest evidence ledger
Marginal likelihood rewards predictive fit while integrating parameter uncertainty, but it can be sensitive to priors. Record the model list, parameter priors, prior model weights and how marginal likelihoods were computed.
Do not include ten slight variants of one family and one representative of another without considering model duplication. Splitting one idea into many near-copies can give it excessive aggregate weight.
Use held-out log predictive scores as a companion check. Pure posterior model probabilities and stacking-style predictive weights answer related but different questions; name the method actually used.
CHAPTER 6 OF 12 · STRESS THE WEIGHTS
6. Changed context: the heavy-tail weight rises
Suppose new reference windows increase Model B’s weight from 0.30 to 0.60. The mixture threshold rises because more posterior mass supports a heavy tail. Recompute the mixture rather than interpolating old thresholds.
Track threshold against w_B from zero to one. A smooth curve is useful; a sharp rise shows that a small weight change has operational consequences in the target tail.
The weight change may indicate a real clutter transition or contaminated reference cells. BMA represents uncertainty among supplied models; it does not diagnose why evidence changed.
CHAPTER 7 OF 12 · STRESS THE WEIGHTS
7. Reference-window contamination can fool every model
An interfering target in the clutter reference process can increase support for a heavy-tailed family. The averaged detector may respond by raising its threshold, protecting false alarms while worsening masking.
Stress tests should insert controlled interferers and clutter edges. Compare threshold, achieved false-alarm rate and target detection. Link the result to order-statistic CFAR and variability-index switching.
If all candidate models assume homogeneous uncontaminated cells, averaging them cannot create a contamination model. Expand the model set or change the reference-data treatment.
CHAPTER 8 OF 12 · STRESS THE WEIGHTS
8. Threshold stability needs more than one posterior
Repeat the analysis across range, azimuth, time and labelled regimes. Plot model weights and thresholds together. A stable threshold with unstable weights can indicate compensating models; stable weights with a drifting threshold can indicate parameter change inside a family.
Report effective sample size or Monte Carlo error when model weights or predictive tails are simulated. Tail quantiles can be sensitive to small numerical errors, especially when α is very small.
Use prequential or rolling validation only when training and evaluation order matches deployment. Randomly mixing adjacent radar cells can leak dependence and make every model appear more stable.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and common misconceptions
BMA does not guarantee CFAR under arbitrary clutter. The nominal tail is calibrated under the candidate set, priors and reference assumptions. If the true regime lies outside that set, the mixture can still fail.
Posterior model probability is not the probability that a clutter family is literally true. It is conditional support within the defined set. A tiny weight on a heavy tail can materially affect an extreme quantile, so pruning rules need tail checks.
This educational example does not specify an operational radar detector. Hardware, coherent processing, correlation, guard geometry and safety requirements demand system-specific engineering.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Evidence and reproducibility route
Hoeting, Madigan, Raftery and Volinsky explain model uncertainty and coherent predictive averaging in Bayesian model averaging: a tutorial. Weinberg, Howard and Tran develop Bayesian predictive CFAR reasoning for exponential and Pareto clutter in Bayesian framework for detector development in Pareto distributed clutter.
Archive candidate families, statistic definitions, priors, model weights, evidence calculations, reference and guard geometry, numerical tail solver, α and validation regimes. Preserve tail probability contributions from each model at the final threshold.
Reproduce the 45.77 solution by evaluating the mixture tail on a bracket and using bisection. A solver result without a tail check is incomplete.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Why not choose the highest-weight model?
Selection discards residual model uncertainty and can understate tail risk.
Can model-specific thresholds be averaged?
Not generally. Average predictive tails, then invert the mixture.
Does adding models always improve detection?
No. Poor or duplicated candidates can distort weights and thresholds.
Is a heavy-tail model always safer?
It can raise thresholds and reduce false alarms, but may also mask targets. Safety depends on the complete decision objective.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Final transfer check and connected reading
Verify the mixture tail at 45.77, then evaluate it at the weighted-average threshold 43.64. Increase the heavy-tail weight and explain why the threshold should move upward. Finally contaminate one reference cell and ask whether the weight change represents clutter or interference.
Continue through the Radar root, Bayesian CFAR, compound-Gaussian clutter, order-statistic CFAR and variability-index CFAR.
Top · Contents · Previous · How X Works