IQ Imbalance

Occurs when the I-branch amplifier and Q-branch amplifier in the receiver have mismatched gain values. This causes the received constellation to stretch along one axis, transforming a perfect square QAM constellation into a rectangular shape. The gain mismatch parameter ε is defined as the ratio deviation from unity, typically expressed in decibels. Even a 1 dB gain imbalance can severely degrade the error vector magnitude (EVM) and increase the bit error rate for higher-order modulations like 64-QAM.

Arises when the local oscillator fails to maintain perfect 90-degree orthogonality between the I and Q mixing stages. Instead of a precise quadrature relationship, a phase error φ is introduced, causing the constellation to skew or shear into a parallelogram shape. This phase error rotates the Q component relative to the I component, creating correlation between branches that should be independent. The resulting image leakage folds energy from the positive frequency spectrum into the negative side, creating a mirror interference signal.

The primary metric for quantifying IQ imbalance severity. IRR measures the power ratio between the desired signal and the unwanted image signal created by the imbalance, expressed in decibels. A perfectly balanced receiver has infinite IRR. Practical direct-conversion receivers typically achieve 30-50 dB IRR without calibration. The relationship between IRR and the gain/phase errors is:

• IRR ≈ 10 log₁₀((ε² + φ²)/4) for small errors
• A 1° phase error and 0.1 dB gain error yields approximately 35 dB IRR
• Modulation classification algorithms must compensate when IRR falls below 25 dB

The combined effect of gain and phase imbalance transforms an ideal circular or square constellation into an elliptical shape in the IQ plane. The major and minor axes of the ellipse correspond to the eigenvectors of the imbalance matrix. This geometric distortion is mathematically modeled as:

• y(t) = α·x(t) + β·x*(t)
• Where α is the desired signal scaling and β is the image leakage coefficient
• The ratio |β/α| directly determines the IRR
• Blind estimation techniques can recover α and β from received samples to digitally compensate the distortion

IQ imbalance creates spurious constellation points that mislead feature-based classifiers. Key effects include:

• Higher-order cumulants become unreliable as the signal's statistical properties are altered by the image component
• Template matching fails because the distorted constellation no longer aligns with ideal reference templates
• Deep learning classifiers trained on ideal data suffer significant accuracy degradation when tested on imbalanced signals
• The image signal can mimic a different modulation format entirely, causing confusion between QPSK and 16-QAM in severe cases
• Robust classifiers must either pre-compensate the imbalance or train on augmented datasets that include various imbalance parameters

Modern receivers employ blind compensation algorithms that estimate and correct IQ imbalance without training sequences. Common approaches include:

• Circularity-based methods: Force the received signal to be proper (circularly symmetric) by adaptively filtering out the conjugate component
• Statistical decorrelation: Adjust gain and phase until the I and Q branches become statistically uncorrelated
• Adaptive filtering: Use a widely-linear filter structure with coefficients derived from the complementary autocorrelation function
• Neural network compensation: Train a small network to learn the inverse imbalance transformation directly from distorted constellation samples