Direction-of-Arrival Accuracy
RMSE in degrees vs SNR and array size. Lower is better. 1,000 Monte Carlo trials per point. ULA, K=100 snapshots.
RMSE vs SNR
N=16 elements, K=100, 1 source, 1,000 trials per point. Cramer-Rao Lower Bound shown for reference.
Cramer-Rao Lower Bound (theoretical floor)
RMSE vs Array Size
SNR=10 dB, K=100, 1 source. Accuracy scales with aperture as expected.
Two-Source Resolution
Probability that two equal-power sources are reported as distinct, vs angular separation. N=16, SNR=10dB, 1,000 trials each.
Why this matters
Sub-degree RMSE down to 5dB SNR. 99.4% resolution rate at 2° separation. Accuracy scales smoothly with aperture from N=4 to N=64. The numbers are direct measurements; nothing is interpolated.
Emission Detection
Detection probability vs SNR for three array sizes. False alarm rate measured separately on pure-noise trials.
Probability of Detection vs SNR
ULA-8, ULA-16, ULA-32 — 1,000 trials per SNR point. Single source, K=100.
Track-Before-Detect Gain
8-frame TBD vs single-frame detection. TBD extends operational SNR floor by approximately 3 dB while preserving FAR=0.
FAR is structurally zero
10,000 pure-noise trials. Zero false alarms. 95% upper confidence bound: 0.038%. This is not a tuned threshold — the architecture cannot fire on noise.
Adaptive Beamforming
Null depth measurements vs MVDR (basic, military-grade, oracle). N=16 ULA, double-precision floating point.
TSD vs MVDR Across Scenarios
Null depth in dB. Lower (more negative) is better. MVDR Oracle uses TRUE covariance — theoretical ceiling.
Null Depth Distribution: 1J through 7J Jammers
Mean and percentiles for null depth across multiple simultaneous jammer scenarios. 500 trials per configuration.
Machine-precision nulling
Mean null depth around −325 dB across 1J–7J. The MVDR Oracle — using the true covariance matrix — achieves only −21 dB. The TSD engine exceeds the theoretical MVDR ceiling by more than 300 dB. This is design headroom for real-hardware impairments (calibration, coupling, quantization).
Overall System Performance
Throughput, stress test, and convergence proof.
Real-Time Throughput
Frames per second for the full Detect + DOA pipeline. 500 reps per configuration in a single process. Single compiled binary, commodity hardware.
Stress Test — 5,000 Scenarios
ULA + URA + UCA, randomized parameters. Detection rate by geometry.
Monte Carlo Convergence
Beamformer null depth mean and CI vs trial count. Demonstrates statistical convergence.
Reproducibility
Every number on this dashboard is a direct measurement from the validation Monte Carlo. Master seed 20260424. Total benchmark runtime: -- seconds. Run it again, get identical numbers.