IQ Data Type Compression

Unlike standard real-valued quantization, IQ data type compression must account for the circular symmetry of complex baseband signals. Direct scalar quantization of I and Q branches independently introduces phase distortion near the origin. Advanced techniques apply polar quantization, where magnitude and phase are quantized separately with non-uniform step sizes, or use vector quantization on the complex plane to preserve the relative angle between samples. This is critical for modulation schemes like QAM-256 where constellation point discrimination depends on precise phase relationships.

A shared-exponent compression scheme where a block of IQ samples shares a single exponent while maintaining individual mantissas. This exploits the temporal correlation in wireless signals where sample magnitudes within a coherence interval remain relatively stable. Key benefits:

• Reduces dynamic range overhead by 2-4x compared to per-sample floating-point
• Preserves relative precision within the block, critical for FFT operations
• Enables efficient SIMD processing on DSP architectures
• Commonly used in 5G NR baseband pipelines for fronthaul compression

Adapted from telephony codecs, mu-law companding applies a non-linear transfer function to IQ samples before uniform quantization. The logarithmic curve allocates more quantization levels to low-amplitude signals and fewer to high amplitudes, matching the crest factor characteristics of OFDM waveforms. This technique:

• Reduces quantization noise floor for weak signals by 6-10 dB
• Preserves EVM (Error Vector Magnitude) for constellation corners
• Can be implemented with a simple lookup table in FPGA fabric
• Pairs effectively with automatic gain control (AGC) front-ends

Exploits the fact that in many wideband spectrum sensing applications, only a fraction of frequency bins contain active signals. By applying structured sparsity directly to the time-domain IQ stream or its frequency-domain representation, irrelevant samples can be zeroed or skipped. Techniques include:

• Threshold-based gating: discard samples below a noise floor estimate
• Sub-Nyquist sampling: leverage compressed sensing to reconstruct sparse spectra from fewer IQ samples
• Attention-based masking: use a lightweight detector network to identify active signal regions before heavy processing This reduces memory bandwidth and MAC operations proportionally to spectrum occupancy.

Standard QAT simulates quantization during training using a straight-through estimator (STE). For IQ data, the QAT process must be modified to handle complex gradients. The real and imaginary components of the loss gradient are backpropagated independently through the quantization nodes, but the quantization grid itself is defined on the complex plane. Advanced IQ-QAT techniques:

• Use circular quantization grids that respect phase periodicity
• Apply magnitude-dependent step sizes to protect low-power samples
• Jointly optimize the dynamic range of I and Q branches
• Simulate hardware-specific rounding modes (round-to-nearest-even) during the forward pass

A noise-shaping technique borrowed from oversampling ADCs. By applying a sigma-delta modulator to the IQ stream at a higher sample rate, quantization noise is pushed out of the band of interest. The decimated output achieves higher effective number of bits (ENOB) within the signal bandwidth than the raw quantizer resolution. For RFML applications:

• Enables 1-2 bit IQ representation with high in-band SNR
• Naturally anti-aliases adjacent channel interference
• Pairs with polyphase filter banks for efficient multi-rate processing
• Particularly effective for narrowband IoT and LPWAN signal classification