Circularity

A circular signal's probability density function remains unchanged when multiplied by a unit-magnitude complex exponential (e^jθ). This means the I and Q components have equal variance and are uncorrelated. Visually, the scatter plot of a circular signal appears as a cloud with no preferred orientation. This property is essential for modulation schemes like QPSK and QAM, where phase rotations from channel effects must not alter the signal's statistical structure.

The defining mathematical test for circularity is that the pseudo-covariance (also called the complementary covariance or relation matrix) must equal zero: E[zz^T] = 0, where z is a zero-mean complex random vector. This condition means the signal is uncorrelated with its own complex conjugate. In contrast, a non-circular (improper) signal has a non-zero pseudo-covariance, indicating correlation between the I and Q branches that carries additional statistical information exploitable by widely linear filtering.

A proper (circular) signal has a covariance matrix that fully describes its second-order statistics. An improper (non-circular) signal requires both the standard covariance matrix E[zz^H] and the pseudo-covariance matrix E[zz^T] for a complete statistical description. Examples of improper signals include:

• BPSK and ASK modulated signals (real-valued in baseband)
• IQ-imbalanced signals where gain and phase mismatches create correlation
• Rectilinear signals where I and Q have unequal power

For circular signals, the classical Wiener filter and minimum mean square error (MMSE) estimator are statistically optimal. For non-circular signals, however, a widely linear (WL) filter that processes both the signal and its complex conjugate achieves superior performance. The WL-MMSE exploits the additional information in the pseudo-covariance, providing a gain of up to 3 dB in signal-to-interference-plus-noise ratio (SINR) for highly improper signals like BPSK in interference-limited scenarios.

The degree of non-circularity (impropriety) is quantified by the circularity coefficient κ, ranging from 0 (perfectly circular) to 1 (maximally improper, e.g., a real-valued signal). It is derived from the canonical correlation between the signal and its complex conjugate. This coefficient is critical in cognitive radio and automatic modulation classification, where identifying the impropriety of an unknown signal reveals its modulation format—BPSK (κ=1) versus QPSK (κ=0)—even under low signal-to-noise ratio conditions.

Circularity assumptions directly influence complex-valued neural network (CVNN) architecture. Standard complex backpropagation using Wirtinger calculus implicitly assumes circularity when using split activation functions (applying real-valued activations separately to magnitude and phase). For non-circular data, augmented complex backpropagation that computes gradients with respect to both the variable and its conjugate is required. Preserving or exploiting impropriety through the network layers can improve classification accuracy for signals like BPSK and GMSK.

The foundational mathematical framework that models a bandpass signal as a complex-valued signal centered at zero hertz. This representation captures both amplitude and phase in a single complex number, enabling the analysis of circularity. A signal is circular if its baseband representation is uncorrelated with its own complex conjugate, meaning the real and imaginary parts have equal variance and are independent.

A proper (circular) complex random signal has a vanishing pseudo-covariance, meaning the real and imaginary parts are uncorrelated with equal variance. An improper (non-circular) signal violates this condition, exhibiting correlation between its components. Key distinctions:

• Proper: Rotationally invariant probability density function
• Improper: Orientation-dependent statistics, common in BPSK and ASK modulations
• Pseudo-covariance: The statistical measure that quantifies the degree of impropriety

An augmented filtering paradigm that processes both the complex signal and its complex conjugate to achieve optimal performance for improper data. Standard linear filtering is sufficient only for circular signals. Widely linear models, such as the augmented complex least mean square (ACLMS) algorithm, exploit the full second-order statistical information, providing superior interference cancellation when circularity is violated.

Hardware impairments in direct-conversion transceivers, such as gain and phase mismatches between the I and Q branches, inherently generate improper signals. The resulting mirror-frequency interference creates a conjugate correlation that destroys circularity. Correcting IQ imbalance is mathematically equivalent to restoring the properness of the received complex baseband signal through widely linear compensation.

Neural architectures that process data directly in the complex domain using complex-valued weights and activation functions. For non-circular signals, standard real-valued networks applied to concatenated I/Q data are suboptimal. CVNNs trained with Wirtinger calculus inherently preserve phase information and can learn widely linear transformations, making them ideal for tasks like channel equalization and beamforming where circularity assumptions fail.

Many communication signals exhibit periodically time-varying autocorrelation functions, a property called cyclostationarity. The conjugate cyclic autocorrelation function directly measures the degree of impropriety at specific cycle frequencies. This feature is exploited for:

• Blind modulation classification: Distinguishing BPSK (improper) from QPSK (proper)
• Signal detection: Identifying signals buried below the noise floor
• Parameter estimation: Extracting symbol rate and carrier frequency without prior knowledge