Hierarchical Cumulant Classifier

The architecture begins by separating broad modulation families before refining to specific orders. At the root node, a Gaussianity test using fourth-order cumulants distinguishes single-carrier modulations from OFDM or noise. Subsequent nodes apply cumulant ratios like |C40|/|C42| to split the remaining candidates into PSK, QAM, and ASK subsets. This hierarchical strategy reduces the number of pairwise comparisons from O(n²) to O(log n), dramatically lowering computational complexity while maintaining classification accuracy.

Each node in the tree applies a binary hypothesis test comparing an estimated cumulant value against a theoretically derived threshold. For example:

• |C42| > 0.6: Signal is likely QAM (super-Gaussian)
• |C42| < 0.3: Signal is likely PSK (sub-Gaussian)
• 0.3 ≤ |C42| ≤ 0.6: Ambiguous region requiring additional samples or alternative features Thresholds are computed analytically from the theoretical cumulant values of candidate modulations under ideal conditions, with guard bands added to account for finite-sample estimation variance.

The hierarchical tree leverages cumulant invariants—mathematical transformations that remain constant under nuisance parameters like phase rotation, frequency offset, and amplitude scaling. Key invariants include:

• |C40|/|C42|: Phase-invariant ratio for QAM order identification
• Normalized sixth-order cumulant: Robust to both phase and scale variations
• Cyclic cumulant magnitudes: Invariant to timing offsets when extracted at the correct cycle frequency These invariants eliminate the need for precise synchronization before classification, making the tree robust to real-world non-cooperative reception conditions.

The tree dynamically adjusts its depth based on confidence metrics computed at each node. When a cumulant estimate falls within the ambiguous guard band, the classifier can:

• Extend observation time to reduce estimation variance
• Branch to alternative features such as sixth-order cumulants or cyclic statistics
• Return a partial classification (e.g., 'QAM family, order ≥ 16') rather than forcing an incorrect decision This adaptive mechanism prevents error propagation—a critical failure mode in fixed-depth decision trees where an early misclassification cascades through all subsequent nodes.

Hierarchical cumulant classifiers achieve sub-millisecond classification latency on FPGA and embedded ARM platforms by:

• Incremental cumulant estimation: Updating running moment accumulators with each new IQ sample rather than batch recomputation
• Early termination: Exiting the tree as soon as a leaf node is reached, avoiding unnecessary feature calculations
• Fixed-point arithmetic: Implementing cumulant estimators using integer math on FPGA DSP slices, eliminating floating-point overhead A typical 8-modulation tree requires only 3-4 cumulant computations total, compared to computing a full 20-element feature vector for a flat classifier.

The hierarchical structure inherently resists common channel impairments without explicit compensation:

• Phase offset: Higher-order cumulants of PSK and QAM signals are naturally phase-invariant when using magnitude operations
• Frequency offset: Cyclic cumulants extracted at the symbol-rate cycle frequency are immune to carrier offset
• Gaussian noise: All cumulants of order > 2 are theoretically zero for Gaussian processes, providing inherent noise rejection
• Multipath fading: The normalized cumulant ratios remain stable under flat fading when sufficient samples are averaged This robustness reduces the preprocessing pipeline complexity compared to likelihood-based classifiers that require precise channel estimation.

Deployed in ESM receivers to rapidly identify threat radar and communication emitters in dense electromagnetic environments. The hierarchical tree structure enables real-time triage by first separating continuous-wave from pulsed signals using kurtosis, then refining to specific intrapulse modulation types.

• Threat library matching: Compares extracted cumulant fingerprints against known hostile emitter databases
• Low-SNR operation: Functions below 0 dB SNR where constellation-based methods fail
• Platforms: Airborne RWR, naval ESM, ground-based SIGINT

Enables opportunistic spectrum access by identifying the modulation format of primary users before secondary transmission. The classifier's blind operation—requiring no pilot tones or synchronization—makes it ideal for non-cooperative spectrum monitoring.

• IEEE 802.22 WRAN: Identifies wireless microphone (FM) vs. DTV (OFDM) signals
• Dynamic spectrum sharing: Triggers spectrum evacuation when high-order QAM indicates active primary use
• Energy-efficient: Cumulant computation avoids expensive FFT operations

Used in ground station telemetry analysis to autonomously classify downlink modulation formats without prior coordination. The hierarchical structure handles the wide variety of APSK and QAM schemes common in DVB-S2/S2X standards.

• Blind feed detection: Identifies modulation changes due to adaptive coding and modulation (ACM)
• Interference geolocation: Classifies adjacent satellite interferers by their modulation fingerprint
• ITU compliance monitoring: Verifies that satellite operators adhere to filed modulation parameters

Integrated into vector signal analyzers and production-line test systems to automate physical layer validation. The classifier provides a reference-free modulation check that operates without known bit sequences.

• Keysight/Anritsu integration: Embedded in instrument firmware for one-button modulation ID
• Manufacturing QA: Detects incorrect modulation due to DAC nonlinearity or IQ modulator imbalance
• Field troubleshooting: Allows technicians to identify unknown signals without protocol decoders

National regulatory authorities deploy hierarchical cumulant classifiers in remote monitoring stations to detect unauthorized transmissions and verify license compliance. The blind nature of cumulant analysis avoids the legal complexity of demodulating private communications.

• Pirate radio detection: Identifies unauthorized FM/AM broadcasts by modulation signature
• License violation alerts: Flags operators exceeding authorized bandwidth or modulation complexity
• Automated logging: Generates timestamped modulation records for evidentiary use

Modern cognitive jammers use hierarchical classification to identify the target waveform and select the optimal jamming strategy. The tree structure enables rapid decisions: PSK signals trigger phase-noise jamming, while QAM signals trigger amplitude-distortion techniques.

• Modulation-specific jamming: Matches jamming waveform to target's statistical vulnerabilities
• Low probability of intercept: Classifier operates on short signal bursts before frequency hopping
• Resource allocation: Prioritizes high-order modulations carrying more critical data