IQ Correction

Corrects amplitude mismatches between the I and Q branches caused by component tolerances in mixers, filters, and ADCs. Gain imbalance creates an elliptical distortion of the constellation, where one axis is stretched relative to the other. Correction involves estimating the gain ratio and applying a multiplicative scaling factor to equalize the branch amplitudes. Even a 0.5 dB imbalance can degrade EVM by several percent in high-order QAM schemes.

Compensates for deviations from the ideal 90-degree phase offset between the I and Q local oscillator paths. Quadrature error causes a skewing of the constellation, where symbols rotate out of their ideal positions. Correction applies an adaptive phase rotation matrix to restore orthogonality. In direct-conversion receivers, phase errors as small as 1 degree can significantly raise the bit error rate for 256-QAM and higher modulations.

Estimates imbalance parameters without requiring known training sequences, using only the statistical properties of the received signal. Blind IQ correction leverages the fact that ideal complex baseband signals exhibit circularity—they are uncorrelated with their own complex conjugate. The circularity coefficient measures deviation from this property, and widely linear filtering can restore properness without pilot symbols. This is critical for passive signal intelligence and non-cooperative receiver scenarios.

Addresses mismatches that vary across the signal bandwidth, common in wideband systems where analog filter responses differ between I and Q paths. Frequency-selective imbalance cannot be corrected by a single scalar or phase rotation; it requires adaptive FIR filtering on both the direct signal and its complex conjugate. Modern correction architectures use widely linear equalizers that jointly process the signal and its conjugate to flatten the frequency-dependent image response.

The primary metric for quantifying IQ correction performance, measured as the power ratio between the desired signal and the unwanted image frequency. IRR is expressed in dB, with higher values indicating better suppression. Uncorrected consumer-grade receivers may achieve only 25-35 dB IRR, while professional correction algorithms can push this beyond 60 dB. Each 3 dB improvement in IRR roughly corresponds to halving the image interference power.

Applies complex-valued neural networks (CVNNs) to learn non-linear IQ imbalance patterns that linear estimators miss. CVNN-based correction processes raw IQ samples directly in the complex domain using Wirtinger calculus for backpropagation, preserving phase relationships. These models can jointly compensate for IQ imbalance, PA non-linearity, and LO leakage in a single learned transformation, outperforming cascaded classical correction stages in highly impaired channels.