Instantaneous Frequency

Instantaneous frequency is formally defined as the derivative of the unwrapped instantaneous phase with respect to time:

f(t) = (1 / 2π) * dφ(t) / dt

• φ(t) is the instantaneous phase obtained from the arctangent of Q/I
• Phase unwrapping is mandatory to remove 2π discontinuities before differentiation
• The result is a time series where each sample represents the frequency deviation at that instant
• For a pure tone, IF is constant; for an FM signal, it varies proportionally to the modulating waveform

Raw instantaneous phase from arctan2(Q, I) is bounded to [-π, π], creating artificial jumps. Phase unwrapping restores continuity:

• Detects phase jumps exceeding a threshold (typically π radians)
• Adds or subtracts 2π multiples to create a continuous phase trajectory
• Failure to unwrap causes impulse-like spikes in the IF estimate
• Critical for high-SNR signals where phase wraps occur frequently over long observation windows

IF analysis directly reveals the modulation structure of angle-modulated signals:

• Analog FM: IF varies continuously, tracking the audio or baseband waveform
• Binary FSK (2-FSK): IF toggles between two discrete frequency values, creating a square-wave pattern
• 4-FSK / M-FSK: IF steps between M distinct levels corresponding to symbol states
• CPFSK: IF shows continuous phase transitions between frequency states, producing smoother trajectories
• Linear FM chirp: IF ramps linearly, useful for radar pulse classification

IF estimation is highly sensitive to additive noise because differentiation amplifies high-frequency fluctuations:

• At low SNR, phase errors from noise produce wild IF excursions that obscure modulation patterns
• Smoothing filters (moving average, Savitzky-Golay) applied post-differentiation reduce variance
• Carrier Frequency Offset (CFO) introduces a constant bias in the IF estimate—centering is essential
• Alternative approach: compute IF from the time-derivative of the complex signal directly, avoiding explicit phase extraction: f(t) = (1 / 2π) * Im[(dx/dt) / x(t)]

IF-derived features serve as inputs to both statistical and deep learning modulation classifiers:

• Statistical moments: mean, variance, skewness, and kurtosis of the IF sequence distinguish constant-frequency from varying-frequency modulations
• Histogram binning: the distribution of IF values reveals the number of discrete frequency states in FSK
• Zero-crossing rate of the centered IF indicates symbol rate for FSK signals
• IF spectrogram: a time-frequency representation of the IF sequence itself, capturing modulation patterns as 2D images for CNN-based classifiers
• Wavelet decomposition of IF extracts transient features for identifying modulation changes

IF and instantaneous amplitude are complementary features that together characterize a modulated signal:

• Constant-envelope modulations (FM, FSK, GMSK): amplitude is flat while IF carries all information
• Amplitude-modulated signals (AM, QAM): IF may be constant while amplitude varies
• Joint AM-FM signals: both amplitude and frequency vary, requiring combined analysis
• The ratio of IF variance to amplitude variance is a powerful feature for discriminating between modulation families
• In I/Q preprocessing pipelines, IF and amplitude are often concatenated as a dual-stream feature vector

The angular component of a complex IQ sample at a given instant, calculated as the arctangent of Q/I. It represents the signal's momentary phase angle and is the direct input for computing instantaneous frequency. Phase unwrapping is essential to remove discontinuities when the angle crosses the ±π boundary, producing a continuous phase trajectory for differentiation.

The absolute magnitude of the complex IQ sample, calculated as √(I² + Q²). This represents the signal envelope and is critical for distinguishing constant-envelope modulations like FSK and PSK from amplitude-varying schemes like QAM. Key derived statistics include:

• γ_max: Maximum normalized instantaneous amplitude
• σ_aa: Standard deviation of absolute instantaneous amplitude

A feature measuring the ratio of power in the upper and lower sidebands relative to the carrier. Calculated as P = (P_L - P_U) / (P_L + P_U), this metric discriminates between double-sideband (DSB) and single-sideband (SSB) modulated signals. Values near zero indicate symmetric spectra typical of AM and FM, while strongly positive or negative values suggest SSB or vestigial sideband modulation.

A computationally efficient proxy for instantaneous frequency that counts how many times the I or Q component crosses zero within a signal segment. While less precise than phase differentiation, it provides a robust estimate of dominant frequency content and is highly effective for distinguishing FSK modulation orders (2-FSK vs. 4-FSK) based on the number of distinct zero-crossing rates present.

A higher-order statistical measure of the peakedness of the signal's power spectral density. It quantifies how much of the signal energy is concentrated in narrow frequency components versus spread across the band. High kurtosis indicates tonal, narrowband signals like unmodulated carriers, while low kurtosis suggests broadband noise-like modulations such as spread spectrum or high-order OFDM.

A time-frequency representation generated by applying the Short-Time Fourier Transform (STFT) to an IQ stream. This converts raw time-domain samples into a 2D image where instantaneous frequency variations appear as visible contours. Convolutional neural networks (CNNs) process these spectrograms as images, learning to recognize modulation-specific patterns such as linear chirps in LFM or frequency hops in FHSS signals.