IQ Sample

An IQ sample is a complex number where the real part is the In-Phase (I) component and the imaginary part is the Quadrature (Q) component. This mathematical structure captures both amplitude and phase simultaneously.

• Amplitude: sqrt(I² + Q²)
• Phase: arctan(Q/I)
• Eliminates the carrier frequency, representing only the modulating information
• Enables efficient digital processing at lower sample rates than RF sampling would require

IQ samples are generated by a quadrature demodulator that mixes the incoming RF signal with two local oscillator signals offset by exactly 90 degrees.

• I channel: Mixed with cos(ωt), producing the in-phase component
• Q channel: Mixed with sin(ωt), producing the quadrature component
• The 90-degree phase offset ensures the two channels are orthogonal
• Orthogonality prevents information loss during downconversion
• Hardware imperfections in this 90-degree relationship cause I/Q imbalance

Each IQ sample represents the signal state at a specific instant in time, with the sample rate determining temporal resolution. The Nyquist criterion requires sampling at least twice the signal bandwidth.

• Typical sample rates range from kS/s to GS/s depending on application
• Sample synchronization recovers the optimal sampling instant at symbol centers
• I/Q resampling adjusts the rate through decimation or interpolation
• I/Q segmentation divides continuous streams into fixed-length inference windows
• Overlapping segments can increase temporal coverage for real-time classification

Modern deep learning classifiers accept IQ samples directly as raw I/Q input, eliminating manual feature extraction. The network learns optimal representations from the time-domain complex data.

• Dual-channel input: Treats I and Q as separate real-valued channels (like image RGB)
• Complex-valued input: Processes IQ natively with complex weights and activations
• Preserves phase relationships that would be lost in magnitude-only representations
• Enables end-to-end learning from waveform to modulation classification
• Requires I/Q normalization to prevent numerical instability during training

Raw IQ samples carry the imprint of all channel effects encountered during transmission. These impairments must be addressed through preprocessing or learned compensation.

• Carrier Frequency Offset (CFO) causes continuous constellation rotation
• DC Offset manifests as a non-zero mean in the IQ stream
• Additive White Gaussian Noise (AWGN) degrades the signal-to-noise ratio
• I/Q imbalance distorts the constellation geometry
• I/Q correction applies inverse filtering to restore orthogonality
• I/Q augmentation deliberately adds impairments during training for robustness

Labeled collections of IQ samples form the I/Q dataset used to train and benchmark modulation classifiers. Both real-world captures and synthetic generation play critical roles.

• Synthetic I/Q provides perfectly labeled data for rare or classified signal types
• Channel simulation applies fading, multipath, and noise models to synthetic signals
• Real-world datasets like RadioML contain millions of labeled IQ segments
• I/Q augmentation expands training diversity through phase rotation and noise addition
• Dataset quality directly determines classifier generalization to field conditions