Widely Linear Filtering

Unlike strictly linear filters that operate solely on the complex signal z(t), a widely linear filter forms its output as y = w₁ᵀz + w₂ᵀz*, where z* is the complex conjugate. This dual-path architecture explicitly models the correlation between the signal and its conjugate, capturing the full second-order statistical information. The filter employs two distinct weight vectors—w₁ for the standard signal path and w₂ for the conjugate path—enabling it to exploit both the covariance matrix R = E[zzᴴ] and the pseudo-covariance matrix P = E[zzᵀ]. This structure collapses to a strictly linear filter when P = 0, meaning the signal is circular.

A complex random signal is termed improper or non-circular when its probability distribution is not rotationally invariant, meaning E[z²] ≠ 0. Improper signals arise in many practical scenarios:

• BPSK and ASK modulations: Real-valued constellations exhibit maximum impropriety
• IQ imbalance: Hardware mismatches induce conjugate correlation
• Offset QPSK: Non-circular due to staggered I/Q transitions For improper signals, the strictly linear minimum mean-square error (MSE) estimator is suboptimal. The widely linear Wiener filter achieves the theoretically minimum MSE by incorporating the pseudo-covariance, delivering superior interference suppression and equalization performance.

Widely linear filtering operates within the framework of augmented complex statistics, where the original complex vector z is mapped to an augmented vector z̲ = [zᵀ, zᴴ]ᵀ. The augmented covariance matrix R̲ is a 2N × 2N block matrix:

• Top-left block: Standard covariance R
• Top-right block: Pseudo-covariance P
• Bottom-left block: Conjugate pseudo-covariance P*
• Bottom-right block: Conjugate covariance R* This formulation transforms the widely linear estimation problem into a standard linear estimation problem in the augmented space, enabling the direct application of classical Wiener-Hopf equations and adaptive algorithms like the augmented complex LMS (ACLMS).

The circularity coefficient ρ = |E[z²]| / E[|z|²] quantifies the degree of impropriety, ranging from 0 (perfectly circular) to 1 (maximally improper). This metric serves as a critical design parameter:

• ρ ≈ 0: Strictly linear filtering suffices; widely linear offers negligible gain
• ρ > 0.5: Widely linear processing provides substantial MSE improvement
• ρ = 1: Real-valued signals like BPSK; widely linear filtering is essential for optimality In adaptive implementations, tracking the circularity coefficient enables algorithm switching between strictly and widely linear modes, optimizing computational resources while maintaining performance.

Widely linear filters are deployed across multiple physical layer functions where signal impropriety is inherent or induced:

• Widely Linear MMSE Equalizers: Outperform conventional equalizers for BPSK and offset QAM in frequency-selective channels
• Interference Rejection: Exploit the non-circularity of co-channel interferers for enhanced spatial filtering in multi-antenna systems
• IQ Imbalance Compensation: Jointly estimate and correct gain/phase mismatches in direct-conversion receivers using the conjugate signal path
• Channel Estimation: Improve pilot-based estimation accuracy for improper training sequences in massive MIMO systems
• Spectrum Sensing: Detect primary users with non-circular modulation (e.g., BPSK) at low SNR by testing for pseudo-covariance structure

Several adaptive filtering algorithms extend classical approaches to the widely linear model:

• Augmented CLMS (ACLMS): Minimizes the instantaneous squared error using gradient descent in the augmented space, requiring 4N complex multiplications per iteration
• Augmented RLS: Provides faster convergence at the cost of O(N²) complexity, suitable for rapidly time-varying channels
• Normalized ACLMS: Improves convergence stability by normalizing the step size by the augmented input power
• Kalman-based WL Filters: Model the optimal weight vector as a state-space process for tracking non-stationary improper signals These algorithms converge to the widely linear Wiener solution, with the ACLMS being the most widely adopted due to its balance of simplicity and performance.

Widely linear filtering provides the optimal minimum mean square error (MMSE) solution for compensating frequency-selective IQ imbalance in direct-conversion receivers. By modeling the received improper signal as a widely linear transformation of the ideal transmitted signal, the filter jointly processes the received IQ data and its conjugate to suppress the mirror-frequency interference caused by gain and phase mismatches. This approach outperforms strictly linear equalizers, which cannot exploit the statistical impropriety introduced by the imbalance.

Binary Phase Shift Keying (BPSK) and other real-valued modulation schemes generate maximally improper signals whose real and imaginary parts are perfectly correlated. Widely linear filters exploit this impropriety to achieve superior co-channel interference suppression compared to conventional linear receivers. By incorporating the conjugate signal path, the filter can cancel interference that occupies the same spectral band but exhibits different impropriety characteristics, a technique known as improper signaling for interference alignment.

In wireless communications, transmitted signals such as offset QPSK (OQPSK) and Gaussian minimum shift keying (GMSK) are inherently improper. A widely linear equalizer at the receiver jointly filters the received signal and its conjugate to exploit this non-circularity, achieving a lower symbol error rate than a strictly linear equalizer of equivalent complexity. This is particularly effective in frequency-selective fading channels where the improper nature of the source provides an additional degree of freedom for signal reconstruction.

Meteorological data, such as wind velocity represented as a complex-valued vector (speed and direction), often exhibits improper or non-circular statistics due to prevailing directional patterns. Widely linear autoregressive (AR) and Kalman filtering models outperform their strictly linear counterparts by modeling the correlation between the wind's complex components and their conjugates. This enables more accurate short-term forecasting for applications in wind farm power optimization and aviation safety systems.

In array signal processing, widely linear minimum variance distortionless response (MVDR) beamformers exploit the impropriety of communication signals like BPSK to achieve enhanced interference nulling. By augmenting the array observation vector with its complex conjugate, the beamformer effectively doubles the degrees of freedom available for suppressing interferers. This technique, known as widely linear Capon beamforming, provides superior spatial filtering when the signal-of-interest is non-circular.

In convolutive blind source separation for audio and biomedical signals, the frequency-domain mixing model is inherently widely linear due to the conjugate symmetry of real-valued time-domain signals. Applying a widely linear demixing matrix in each frequency bin correctly models the complete second-order statistics, preventing the permutation and scaling ambiguities that plague strictly linear complex-valued independent component analysis (ICA) algorithms when separating real-world acoustic mixtures.