Cumulant-Based Modulation Set Partitioning

Partitioning organizes modulation candidates into a decision tree where each node applies a specific cumulant threshold test. The root node separates QAM from PSK constellations using the fourth-order cumulant ratio |C40|/|C42|, while subsequent nodes refine classification to specific orders (e.g., QPSK vs. 8-PSK). This structure reduces an N-class problem to a series of binary decisions, dramatically lowering computational complexity.

The fundamental partitioning principle exploits deviation from Gaussianity as measured by higher-order cumulants. Linear digital modulations exhibit non-zero cumulants, while Gaussian noise and OFDM signals have near-zero higher-order cumulants. This enables a primary partition separating structured communication signals from noise-like waveforms using a Gaussianity test threshold on estimated C40 or C42 values.

Partition boundaries rely on normalized cumulant ratios that are inherently immune to amplitude scaling. Key discriminants include:

• |C40|/|C42|: Separates PSK (≈1.0) from QAM (≈0.68 for 16-QAM)
• |C63|²/|C42|³: Distinguishes 16-QAM from 64-QAM
• |C80|/|C40|²: Identifies 8-PSK vs. QPSK These ratios remain constant regardless of received signal power, eliminating the need for precise automatic gain control.

Without partitioning, a brute-force multi-class classifier must evaluate all M candidate modulations for each decision. Hierarchical cumulant partitioning reduces this to O(log M) comparisons by pruning entire branches at each node. For a 12-modulation recognition task, partitioning typically requires only 3-4 cumulant computations versus 12 full likelihood evaluations, enabling real-time operation on FPGA and embedded platforms.

Cumulant-based partitions are designed to be invariant to common channel impairments:

• Phase rotation: Cumulant magnitudes are phase-blind
• Frequency offset: Higher-order cumulants are insensitive to slow frequency drift
• Timing offset: Sample cumulants converge correctly with sufficient samples This robustness means partition boundaries remain stable without requiring perfect synchronization before classification.

Each partition level has distinct sample requirements based on the cumulant order used. Lower-order cumulants (C40, C42) stabilize with fewer samples, enabling coarse QAM/PSK separation early. Higher-order cumulants (C63, C80) require more samples for reliable estimation, but are only computed on the surviving subset. This progressive refinement optimizes the observation length versus classification accuracy trade-off.

A decision tree architecture that uses specific cumulant thresholds at each node to partition the modulation candidate set. The process begins with coarse separation—such as distinguishing PSK from QAM constellations—and refines to specific orders.

• Root node: Uses |C42| to separate constant-modulus (PSK, FSK) from multi-amplitude (QAM, APSK) signals
• Intermediate nodes: Apply C40 thresholds to determine PSK order (BPSK vs. QPSK vs. 8PSK)
• Leaf nodes: Final classification using combined cumulant ratio tests

This approach reduces computational complexity by avoiding exhaustive comparison against all candidate modulations simultaneously.

A discriminative feature formed by dividing two different cumulant orders, such as |C40|/|C42|, to create a modulation fingerprint inherently robust to phase and frequency offsets.

• Amplitude invariance: Ratios cancel out signal power scaling, eliminating the need for precise gain calibration
• Phase robustness: The magnitude operation removes carrier phase dependency
• Key ratios for partitioning:
  • |C40|/|C42| ≈ 0 for QPSK, ≈ 1 for 16QAM, ≈ 0.68 for 64QAM
  • |C61|/|C42| separates inner/outer constellation points in cross-QAM formats

These ratios form the decision boundaries that define set partitions in hierarchical classification trees.

A specific higher-order statistic measuring the normalized fourth-order moment minus the squared second-order moment. Used as a robust feature to classify QAM, PSK, and ASK modulations by their deviation from Gaussianity.

• C40 (kurtosis): Measures the peakedness of the distribution; negative for sub-Gaussian PSK signals, positive for super-Gaussian multi-amplitude signals
• C42 (variance of squared magnitude): Captures the dispersion of instantaneous power; near-zero for constant-envelope modulations, large for high-order QAM
• Gaussian noise insensitivity: Theoretical cumulants of Gaussian noise above second order are zero, making these features naturally noise-resistant

These two cumulants alone can partition most standard linear digital modulation sets.

A scale-invariant cumulant value obtained by dividing a higher-order cumulant by a power of the signal variance. This ensures the classification feature is independent of received signal amplitude.

• Normalization formula: C̃₄₀ = C₄₀ / (C₂₁)² where C₂₁ is the signal power estimate
• Eliminates dependency on automatic gain control (AGC) settings
• Enables consistent decision thresholds across varying link budgets
• Critical for blind modulation identification where no prior signal level calibration exists

Without normalization, cumulant values would scale with signal power, making fixed-threshold partitioning impossible across different receivers and ranges.

A statistical hypothesis test using sample cumulants to determine if a signal's distribution deviates from Gaussian. This enables the hierarchical separation of linear modulations from Gaussian noise or OFDM signals.

• Test statistic: Estimated C40 compared against theoretical zero for Gaussian processes
• OFDM detection: OFDM signals with sufficient subcarriers approximate Gaussian distributions due to the Central Limit Theorem, yielding C40 ≈ 0
• Threshold setting: Derived from the variance of the sample cumulant estimator under the null hypothesis
• Partitioning role: First-stage filter that separates structured digital modulations from noise-like waveforms before detailed classification

This test prevents wasted computation on signals that cannot be meaningfully classified by cumulant methods.

A structured set of estimated cumulants and their ratios concatenated into a single input vector for a machine learning classifier to perform automatic modulation recognition.

• Typical composition: [|C̃₂₀|, |C̃₂₁|, |C̃₄₀|, |C̃₄₁|, |C̃₄₂|, |C̃₆₀|, |C̃₆₁|, |C̃₆₂|, |C̃₆₃|]
• Ratio features: [|C40|/|C42|, |C61|/|C42|, |C63|/|C42|³]
• Dimensionality: Typically 8-15 features, dramatically smaller than raw IQ sample inputs
• Compatible with lightweight classifiers: SVM, k-NN, or small feedforward neural networks

This compact representation bridges classical statistical signal processing with modern machine learning, enabling efficient training and inference on resource-constrained hardware.