Statistical Detection and Parameter Estimation

Introduction and Context

Radar detection is a statistical decision problem. Echoes are observed in the presence of receiver noise, clutter, interference, propagation fluctuations, and hardware imperfections. Consequently, a detector cannot usually determine with certainty whether a target is present. It must instead control two competing probabilities: the probability of detection, P_D, and the probability of false alarm, P_FA.

Parameter estimation is the related task of extracting quantities such as delay, Doppler frequency, amplitude, phase, angle, and radial velocity from the received data. Detection and estimation are coupled: poor estimation can reduce coherent integration gain, while an inappropriate detection threshold can bias the set of observations passed to an estimator.

This chapter develops the statistical foundations of radar detection, including likelihood-ratio testing, Neyman–Pearson design, coherent and noncoherent integration, target fluctuation models, CFAR processing, and estimation bounds.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Model radar observations under target-absent and target-present hypotheses.

2. Construct likelihood-ratio and matched-filter detectors.

3. Compute P_D and P_FA for common coherent and noncoherent detectors.

4. Explain Swerling target models and their effects on integration.

5. Distinguish coherent integration, noncoherent integration, and post-detection integration.

6. Design and analyze CFAR detectors.

7. Apply Fisher information and the Cramér–Rao lower bound to radar parameter estimation.

8. Evaluate practical limitations, implementation choices, and ethical constraints.

1. Statistical Observation Models

Let x denote the vector of complex samples collected over a fast-time range cell, pulse interval, or array snapshot. The basic binary hypotheses are

H0: x = n,

H1: x = s(theta) + n,

where n is noise or interference, s(theta) is the target-dependent signal, and theta contains unknown parameters such as delay, Doppler, amplitude, phase, or angle.

For thermal receiver noise, a widely used model is circular complex Gaussian noise,

n ~ CN(0, R),

where R = E[n n^H] is the covariance matrix. The notation CN(0, sigma^2) is used here specifically to mean E[|z|^2] = sigma^2 for a scalar complex variable z. Therefore, if z = a + jb, then a and b are independent real Gaussian variables with variance sigma^2/2.

For an N-dimensional vector,

p(x|H0) = 1/(pi^N det R) exp[-x^H R^(-1)x].

Under a known deterministic signal,

p(x|H1) = 1/(pi^N det R) exp[-(x-s)^H R^(-1)(x-s)].

These expressions assume known covariance, proper complex Gaussian noise, and nonsingular R. Non-Gaussian clutter and compound-Gaussian sea clutter require more advanced models.

Whitening and matched filtering

The inverse covariance R^(-1) appears in the likelihood, but R^(-1) alone does not whiten the data. Whitening requires a factor W satisfying W R W^H = I, commonly W = R^(-1/2), obtained from an eigendecomposition or Cholesky factorization. The whitened data are y = W x and the whitened signal is q = W s.

The optimum linear statistic for a known signal in colored Gaussian noise is

T = s^H R^(-1)x.

Equivalently, it is the matched filter applied after whitening. The normalized complex matched-filter output is

z = (s^H R^(-1)x) / sqrt(s^H R^(-1)s).

Under H0, z ~ CN(0,1). Thus Re{z} and Im{z} each have variance 1/2. A real, phase-known coherent statistic is

Y = sqrt(2) Re{exp(-j phi) z},

where phi is the known signal phase. Under H0, Y ~ N(0,1). This sqrt(2) factor is essential: without it, the real part has variance 1/2 rather than unity.

If the signal has unknown phase, the phase-invariant statistic is

U = |z|^2.

Under H0, U has an exponential distribution with unit mean:

P(U > gamma | H0) = exp(-gamma).

Under H1 with normalized signal-to-noise ratio rho = s^H R^(-1)s, U follows a noncentral chi-square law after the appropriate scaling. In particular, 2U is noncentral chi-square with two degrees of freedom and noncentrality parameter 2rho.

2. Likelihood-Ratio and Neyman–Pearson Detection

The likelihood ratio is

Lambda(x) = p(x|H1) / p(x|H0).

The likelihood-ratio test decides H1 when Lambda(x) > eta. The threshold eta depends on prior probabilities and decision costs in a Bayesian detector. In a Neyman–Pearson detector, the threshold is selected to satisfy a specified false-alarm probability while maximizing detection probability.

For equal covariance under H0 and H1 and a known deterministic signal, taking logarithms gives

ln Lambda(x) = 2 Re{s^H R^(-1)x} - s^H R^(-1)s.

Consequently, the decision reduces to thresholding the matched-filter output. For known phase, the real matched-filter output is optimal. For unknown phase, the magnitude-squared output is used, causing a noncoherent decision rule.

For a real normalized coherent statistic Y ~ N(0,1) under H0,

P_FA = P(Y > gamma | H0) = Q(gamma),

where Q(.) is the upper tail probability of a standard normal variable. Hence gamma = Q^(-1)(P_FA). If the signal produces normalized mean mu = sqrt(2rho), then

P_D = Q(gamma - sqrt(2rho)).

This formula assumes known phase, known signal shape, Gaussian noise, and a single threshold test. It should not be applied directly to unknown-phase or fluctuating-target cases.

Worked Example 1: coherent detection

Suppose the desired single-test false-alarm probability is 10^(-6), and the normalized coherent signal mean is sqrt(2rho) = 6. The threshold is approximately gamma = 4.753 because Q(4.753) is about 10^(-6). Therefore,

P_D = Q(4.753 - 6) = Q(-1.247) ≈ 0.894.

The result is approximately 89.4 percent detection probability under the stated ideal assumptions.

3. Energy Detection and Integration

3.1 Noncoherent single-sample detection

When phase is unknown or varies randomly, energy detection is appropriate. For one normalized complex sample,

U = |z|^2.

Under H0, U is exponential, so the threshold required for false-alarm probability alpha is

gamma = -ln(alpha).

For alpha = 10^(-6), gamma = 13.816.

Under H1 with a deterministic signal and unknown phase, the detection probability is the complementary cumulative distribution of a noncentral chi-square variable:

P_D = Q_1(sqrt(2rho), sqrt(2gamma)),

where Q_1 is the first-order Marcum Q-function.

3.2 Noncoherent integration

Suppose M independent energy measurements are summed:

V = sum from k=1 to M of |z_k|^2.

Under H0, V has a gamma distribution with shape M and unit scale. Its false-alarm probability is

P_FA = exp(-gamma) sum from k=0 to M-1 of gamma^k/k!.

Under a deterministic target with equal per-pulse normalized energy rho, 2V is noncentral chi-square with 2M degrees of freedom and noncentrality 2M rho. Noncoherent integration improves detection because independent energy evidence is accumulated, but it does not preserve phase information and is generally less efficient than ideal coherent integration.

The independence assumption is important. Correlated pulses, fluctuating clutter, or receiver memory alter the distribution and reduce the nominal gain.

3.3 Coherent integration

For M pulses with known phase progression, the samples are phase-aligned and summed before magnitude detection:

z_c = (1/sqrt(M)) sum from k=1 to M of z_k exp(-j phi_k).

Under H0, z_c ~ CN(0,1). If each pulse has normalized signal energy rho, the integrated signal energy is M rho. Coherent integration therefore produces an SNR gain of M relative to one pulse, provided phase, Doppler, timing, and amplitude are accurately compensated.

Coherent integration is highly sensitive to Doppler mismatch. If the phase error from pulse to pulse is not corrected, the gain is reduced by the squared magnitude of the normalized coherent sum. For a frequency mismatch f_e over M uniformly spaced pulses with pulse repetition interval T_r, the loss factor is

L_c = |sin(pi M f_e T_r)/(M sin(pi f_e T_r))|^2,

provided the denominator is nonzero; the limiting value at zero mismatch is one.

4. Target Fluctuation and Swerling Models

Radar cross section and complex scattering amplitude fluctuate because a target contains multiple scattering centers whose relative phases change with aspect, frequency, and motion. Swerling models approximate these fluctuations and are central to realistic detection analysis.

Let the instantaneous power or radar cross section be sigma. A common model assumes exponential power fluctuations, equivalent to a zero-mean complex Gaussian scattering amplitude. The mean SNR is denoted bar-rho.

Swerling I

The target power is constant during one scan or dwell but changes independently from scan to scan. Within a dwell, repeated pulses experience the same random amplitude. This model is often associated with slowly fluctuating targets having many comparable scatterers.

For a single noncoherent pulse, averaging the conditional exponential-noise detection law over exponential target power yields

P_D = exp[-gamma/(1 + bar-rho)].

For multiple pulses within the same dwell, the observations are conditionally independent given the common target amplitude, but are statistically dependent after averaging. Therefore, formulas derived for independently fluctuating pulses must not be used without modification.

Swerling II

The target power changes independently from pulse to pulse. This model represents faster fluctuation than Swerling I. For a single pulse, the same expression applies:

P_D = exp[-gamma/(1 + bar-rho)].

For M-pulse noncoherent integration, the averaged detection probability is obtained from the distribution of a sum of exponential noise energies with independent exponential target powers. Its closed form can be expressed through incomplete gamma functions; numerical evaluation is often preferred. The important engineering consequence is that integration gains differ from the constant-amplitude case because each pulse may be weak even when the average target is detectable.

Swerling III and IV

Swerling III and IV use a four-degree-of-freedom chi-square power distribution, which has less severe fluctuations than the exponential model. Swerling III remains constant over a dwell and changes between dwells; Swerling IV changes independently from pulse to pulse. For the same mean SNR, these models generally provide higher detection probability than Swerling I and II because deep fades are less frequent.

Swerling V and model selection

Swerling V is sometimes used for nonfluctuating target amplitude, particularly as an idealized deterministic reference. Real targets may not fit one model over all aspect angles or waveforms. Model uncertainty should be included in performance studies by averaging over plausible models or by using measured high-resolution radar-cross-section data.

The practical ranking is not simply “more pulses always solve fluctuation.” Coherent integration is excellent when phase and amplitude are stable; noncoherent integration is more robust to phase uncertainty; and fluctuation can produce diminishing returns or scan-to-scan variability. A detector designed for a fixed target can substantially overestimate performance against Swerling I or II targets.

5. CFAR Detection

A fixed threshold is unsuitable when clutter or interference varies with range, azimuth, time, or frequency. Constant-false-alarm-rate processing estimates the local background level and adapts the threshold.

In cell-averaging CFAR, the cell under test is surrounded by reference cells. Guard cells reduce contamination from the target sidelobes. If the reference-cell powers are X_1,...,X_N and the estimated noise level is

Z_hat = (1/N) sum X_i,

the threshold is T = K Z_hat, where K is selected to meet the desired P_FA under the assumed background distribution.

For exponentially distributed reference-cell power and no target contamination, the CA-CFAR false-alarm probability is

P_FA = (1 + K/N)^(-N).

Therefore,

K = N[P_FA^(-1/N) - 1].

Worked Example 2: CA-CFAR multiplier

Let N = 24 reference cells and desired P_FA = 10^(-4). Then

K = 24[(10^4)^(1/24) - 1].

Since (10^4)^(1/24) = exp(9.21034/24) ≈ 1.467,

K ≈ 24(0.467) = 11.21.

This multiplier applies only to the stated exponential, independent, homogeneous-background model.

CA-CFAR loses performance when a reference window contains interfering targets or clutter edges. Ordered-statistic CFAR is more resistant to multiple-target contamination; greatest-of and smallest-of CFAR are useful near clutter transitions, with different trade-offs. Modern systems may use trimmed means, censored estimates, adaptive mixture models, or machine-learning classifiers, but any learned detector must be calibrated against operational false-alarm requirements and distribution shift.

6. Parameter Estimation

Detection asks whether a signal is present; estimation asks what signal parameters best explain the data. The maximum-likelihood estimator is

hat(theta)_ML = arg max_theta p(x; theta).

For Gaussian noise with known covariance, maximizing likelihood is equivalent to minimizing the whitened squared error

(x - s(theta))^H R^(-1)(x - s(theta)).

This emphasizes why covariance modeling matters: errors in R change both the detector and the estimator.

Delay and Doppler estimation

For a known waveform, delay can be estimated by maximizing the matched-filter magnitude over candidate delays. Doppler can be estimated by maximizing coherent response over a Doppler-frequency grid. The resulting range-Doppler map is a bank of approximate generalized likelihood-ratio tests.

For monostatic radar, range and radial velocity are related to delay tau and Doppler frequency f_D by

R = c tau/2,

v_r = lambda f_D/2,

where c is propagation speed and lambda is wavelength. These equations assume a monostatic geometry and radial motion; bistatic geometry requires different factors and projections.

Fisher information and the Cramér–Rao lower bound

For parameter vector theta, the Fisher information matrix is

J_ij(theta) = E[(partial ln p/partial theta_i)(partial ln p/partial theta_j)].

Under regularity conditions, unbiased estimators satisfy

Cov(hat(theta)) >= J(theta)^(-1).

The inequality means that the difference Cov(hat(theta)) - J^(-1) is positive semidefinite. The bound presumes a correctly specified model, sufficient differentiability, and parameter-independent support. It is a lower bound, not a guarantee that an estimator achieves it.

For a single complex delayed waveform in white Gaussian noise with known amplitude and phase, delay information increases with effective signal bandwidth and SNR. A commonly used form is

var(tau_hat) >= 1/[8 pi^2 beta^2 rho],

where beta is the rms bandwidth defined consistently from the waveform spectrum and rho is the matched-filter SNR. The exact numerical factor depends on the definition of beta and on whether amplitude, phase, and frequency are nuisance parameters. Thus this expression should not be transferred uncritically between conventions.

At high SNR and for well-resolved parameters, maximum-likelihood estimators often approach the CRLB. At low SNR, threshold effects, ambiguity, sidelobes, and biased estimates can make the bound optimistic. Bayesian or Ziv–Zakai bounds are valuable when prior information, ambiguity, or threshold behavior is important.

7. Applied Case: Maritime Surveillance Radar

Background

Consider an X-band coastal radar monitoring a shipping channel. Sea clutter varies with range and wind state, while small vessels may have fluctuating radar cross sections. The system must maintain a specified false-alarm rate without missing low-observable vessels.

Problem

A fixed energy threshold produces excessive alarms during high sea state and inadequate sensitivity during quiet periods. Coherent integration is also unreliable because vessel micro-motion and uncompensated Doppler cause phase changes across pulses.

Analysis

The engineering team models receiver noise as complex Gaussian and treats clutter locally rather than globally. Swerling II is used as a conservative first approximation for pulse-to-pulse target fluctuation, while measured data are retained for validation. A range-Doppler processor estimates Doppler before integration. Because phase coherence is only partial, noncoherent integration is selected for the final detection channel.

Approach and solution

The detector uses guard cells and a CA-CFAR processor in homogeneous regions. At suspected clutter edges, a greatest-of CFAR branch is selected. Tracks are initiated only after detections persist over multiple scans, reducing isolated false alarms. The estimator refines range and Doppler by local interpolation around the strongest range-Doppler cell. Calibration data continuously check the empirical P_FA across operating conditions.

Outcome

The adaptive detector maintains a more stable false-alarm rate than a fixed threshold and reduces false track initiation. However, performance is reported as a probability distribution over sea states and target classes rather than as one universal P_D value. This is essential because target fluctuation, clutter nonstationarity, and model mismatch dominate the uncertainty.

Learning points

1. A statistical model must match the operating environment.

2. CFAR protects false-alarm control but cannot eliminate target masking.

3. Integration strategy must reflect phase stability and target fluctuation.

4. Empirical validation is necessary even when analytical formulas are available.

8. Limitations, Governance, and Best Practices

Analytical detection formulas may fail under receiver saturation, quantization, correlated clutter, interference, non-Gaussian sea returns, multipath, or inaccurate calibration. Monte Carlo simulation should therefore accompany closed-form analysis. Simulations must report confidence intervals, especially when estimating very small P_FA values; observing zero false alarms does not prove that the true probability is zero.

Operational radar systems should document threshold logic, calibration procedures, data provenance, and performance across environmental conditions. Adaptive and AI-assisted detectors introduce governance concerns: training-data bias can cause unequal detection performance across target categories, black-box decisions complicate certification, and adversarial or deceptive emissions can exploit learned weaknesses. Human oversight, audit logs, controlled updates, and fail-safe behavior are appropriate for safety-critical or regulated applications. Privacy and spectrum regulations must also be respected when radar observations involve civilian activity or protected frequency bands.

Best practice is to specify P_FA per resolution cell, per dwell, and per scan separately; state the noise and clutter assumptions; distinguish conditional from averaged P_D; include Swerling and mismatch studies; validate CFAR empirically; and report estimator bias and variance rather than only a nominal CRLB.

Chapter Summary

Radar detection is a statistical decision process governed by target, noise, clutter, and interference distributions. The likelihood-ratio test provides the general decision framework, while the Neyman–Pearson criterion maximizes detection probability subject to a prescribed false-alarm probability. In colored Gaussian noise, whitening uses R^(-1/2), after which matched filtering produces the appropriate sufficient statistic.

Normalization is important: if CN(0,sigma^2) means E|z|^2 = sigma^2, a complex unit-variance output has real and imaginary variances of 1/2. A phase-known real statistic must therefore be scaled by sqrt(2) to become standard normal. Coherent integration provides an SNR gain proportional to the number of phase-aligned pulses; noncoherent integration is more tolerant of phase uncertainty but generally less efficient. Swerling models describe target fluctuation and significantly affect integration performance.

CFAR processing adapts thresholds to local background conditions, but its guarantees depend on the assumed distribution and uncontaminated reference cells. Maximum-likelihood estimation, matched-filter searches, Fisher information, and the Cramér–Rao bound provide a rigorous foundation for estimating delay, Doppler, and other parameters. In practice, model mismatch, nonstationarity, low-SNR threshold effects, and governance requirements must be addressed through simulation, measurement, calibration, and responsible system design.

Key Terms

Likelihood ratio: Ratio of target-present to target-absent probability densities.

Neyman–Pearson test: Detector that maximizes P_D for a specified P_FA.

False alarm: Declaration of a target when no target is present.

Detection probability: Probability of declaring a target when it is present.

Whitening: Transformation that converts colored noise covariance into the identity matrix.

Matched filter: Filter maximizing output SNR for a known signal in Gaussian noise.

Coherent integration: Phase-aligned accumulation of complex samples.

Noncoherent integration: Accumulation of detected energies or magnitudes.

Swerling model: Statistical model for radar-cross-section fluctuation.

CFAR: Adaptive thresholding method designed to maintain local false-alarm rate.

Cell under test: Resolution cell whose target status is being decided.

Guard cell: Cell excluded from background estimation near the test cell.

Reference cell: Neighboring cell used to estimate background level.

Maximum-likelihood estimator: Parameter value maximizing the likelihood function.

Fisher information: Expected curvature or score-product measure of parameter information.

Cramér–Rao lower bound: Lower bound on covariance of unbiased estimators.

SNR: Ratio of signal energy or power to noise energy or power under a stated convention.

Marcum Q-function: Tail probability function used in noncentral chi-square detection analysis.

Range-Doppler map: Two-dimensional representation of matched-filter response versus delay and Doppler.

Review Questions

Short Answer Questions

1. Why is R^(-1/2), rather than R^(-1), used for whitening?

2. State the difference between coherent and noncoherent integration.

3. What is the purpose of guard cells in a CFAR processor?

4. Define a Swerling II target model.

5. What does the Cramér–Rao lower bound quantify?

Long Answer Questions

1. Derive the matched-filter statistic for a known signal in colored complex Gaussian noise.

2. Explain the normalization of a complex Gaussian matched-filter output and derive the standard-normal coherent statistic.

3. Compare Swerling I, II, III, and IV models and discuss their implications for pulse integration.

4. Derive the CA-CFAR threshold multiplier for exponentially distributed reference-cell powers.

5. Discuss the assumptions and limitations of using the CRLB for radar delay estimation.

Analytical / Application Questions

1. A detector uses a normalized phase-known coherent statistic with P_FA = 10^(-5) and mean under H1 equal to 5. Determine the threshold and estimate P_D using the Gaussian Q-function.

2. Design a CA-CFAR processor with 32 reference cells and P_FA = 10^(-6). Calculate its threshold multiplier and list two conditions under which the result is unreliable.

3. A radar experiences rapid pulse-to-pulse target fluctuation and significant Doppler phase uncertainty. Recommend an integration and CFAR strategy, justify it statistically, and identify measurements needed to validate the design.

Multiple Choice Questions with Answers

Question 1

For z ~ CN(0,1) with E|z|² = 1, what is the variance of Re{z}?

A. 1

B. 1/2

C. 2

D. 0

Correct Answer: B. 1/2

Question 2

Which operation actually whitens colored data with covariance R?

A. Multiplication by R

B. Multiplication by R⁻¹ only

C. Multiplication by a factor W satisfying WRWᴴ = I

D. Addition of the mean signal

Correct Answer: C. Multiplication by a factor W satisfying WRWᴴ = I

Question 3

For a phase-known normalized coherent statistic Y ~ N(0,1) under H0, which expression gives P_FA?

A. Q(γ)

B. exp(−γ)

C. 1 − Q(γ)

D. γ²

Correct Answer: A. Q(γ)

Question 4

For one normalized complex energy sample U = |z|² under H0, what is P(U > γ)?

A. Q(γ)

B. exp(−γ)

C. 1 − exp(−γ)

D. exp(−γ²)

Correct Answer: B. exp(−γ)

Question 5

What is the ideal coherent-integration SNR gain for M phase-aligned pulses?

A. M

B. sqrt(M)

C. 1/M

D. M²

Correct Answer: A. M

Question 6

Which Swerling model has independently fluctuating target power from pulse to pulse?

A. Swerling I

B. Swerling II

C. Swerling III

D. Swerling V

Correct Answer: B. Swerling II

Question 7

For CA-CFAR with N reference cells and exponential background, which formula gives the threshold multiplier?

A. K = N[P_FA^(1/N) − 1]

B. K = N[P_FA^(−1/N) − 1]

C. K = P_FA/N

D. K = N/P_FA

Correct Answer: B. K = N[P_FA^(−1/N) − 1]

Question 8

What is the primary purpose of guard cells in CFAR?

A. Increase transmitter power

B. Estimate Doppler frequency

C. Prevent target energy from biasing the background estimate

D. Whiten receiver noise

Correct Answer: C. Prevent target energy from biasing the background estimate

Question 9

The Cramér–Rao lower bound is primarily a bound on the covariance of which estimators?

A. Only biased estimators

B. Unbiased estimators under regularity conditions

C. Only Bayesian estimators

D. Only threshold detectors

Correct Answer: B. Unbiased estimators under regularity conditions

Question 10

For a monostatic radar, which relation connects radial velocity v_r and Doppler frequency f_D?

A. v_r = c f_D

B. v_r = 2 f_D/lambda

C. v_r = lambda f_D/2

D. v_r = lambda/(2 f_D)

Correct Answer: C. v_r = lambda f_D/2