Complex Baseband

The defining property of a complex baseband signal is the strict orthogonality between its In-Phase (I) and Quadrature (Q) components. These two real-valued streams are modulated onto carrier waves that are exactly 90 degrees out of phase. This mathematical independence allows a single signal to carry two independent data streams simultaneously—one on the cosine carrier (I) and one on the sine carrier (Q)—effectively doubling spectral efficiency without requiring additional bandwidth. Any deviation from this 90-degree relationship constitutes I/Q imbalance, a critical hardware impairment.

Complex baseband is a frequency-translated version of a real bandpass signal. The process of quadrature downconversion shifts the carrier frequency to exactly 0 Hz (DC). This eliminates the carrier from the mathematical model, leaving only the modulating information. The resulting spectrum is no longer symmetric; it is a single-sided representation where negative and positive frequencies carry distinct information. This zero-centering is essential for simplifying digital processing algorithms and is the native format for software-defined radio (SDR) hardware.

Every complex sample encodes two physical quantities simultaneously:

• Instantaneous Amplitude (Envelope): Calculated as the magnitude (\sqrt{I^2 + Q^2}). This represents the signal's power at that exact moment.
• Instantaneous Phase: Calculated as the arctangent (\arctan(Q/I)). This represents the angular position of the signal vector. This polar representation is the geometric basis for all digital modulation constellations (e.g., QPSK, 16-QAM), where each symbol is a distinct point in the complex plane.

Any real bandpass signal (s(t)) can be mathematically reconstructed from its complex baseband equivalent (s_b(t)) using the relation: (s(t) = \Re{s_b(t) e^{j2\pi f_c t}}) The term (s_b(t)) is the complex envelope, a low-pass signal containing all amplitude and phase modulation information. The exponential term (e^{j2\pi f_c t}) represents the carrier. This analytic representation is lossless; no information is discarded during downconversion, making it the preferred format for digital signal processing and machine learning feature extraction.

A residual Carrier Frequency Offset (CFO) between the transmitter and receiver local oscillators manifests as a continuous rotation of the entire complex baseband constellation. Mathematically, this is a multiplication by (e^{j2\pi \Delta f t}), where (\Delta f) is the frequency error. In the complex plane, a static QPSK constellation will appear to spin at a rate proportional to the offset. This rotation is a primary impairment that must be estimated and corrected via I/Q centering or blind synchronization algorithms before a neural network classifier can reliably identify the modulation scheme.

For neural network input, the complex baseband stream can be handled in two distinct ways:

• Dual-Channel Real Input: The I and Q streams are treated as two independent real-valued channels, analogous to the red and blue channels of an image. Standard real-valued convolutions are applied.
• Complex-Valued Input: The network uses complex-valued weights and activation functions (e.g., modReLU) that natively preserve the phase relationships. This approach is mathematically more faithful to the signal's structure but requires specialized deep learning frameworks. The choice impacts the model's ability to learn phase-sensitive features.

Complex baseband IQ samples serve as the raw input tensor for deep learning classifiers. The network learns hierarchical features directly from the time-domain complex stream, eliminating the need for hand-crafted expert features like cyclic cumulants or spectral correlation functions. A dual-channel CNN treats the In-Phase (I) and Quadrature (Q) components as separate real-valued input channels, analogous to the RGB channels of an image, allowing the model to learn spatial relationships in the constellation geometry.

When plotted on the complex plane, the baseband IQ stream forms a constellation diagram—a scatter plot of Q versus I. This geometric representation directly reveals the modulation scheme:

• QPSK: Four distinct clusters at 45°, 135°, 225°, and 315°
• 16-QAM: A 4×4 grid of points with varying amplitude and phase
• BPSK: Two antipodal points along the I-axis CNNs trained on constellation images can achieve high classification accuracy by learning these spatial patterns.

From the complex baseband stream, three critical instantaneous parameters are derived for feature-based classifiers:

• Instantaneous Amplitude: |x[n]| = √(I[n]² + Q[n]²), revealing amplitude-shift keying patterns
• Instantaneous Phase: ∠x[n] = arctan(Q[n]/I[n]), capturing phase transitions
• Instantaneous Frequency: f[n] = d(∠x[n])/dt, discriminating FSK variants Statistical moments of these parameters—variance, kurtosis, skewness—form compact feature vectors for lightweight classifiers.

Applying the Short-Time Fourier Transform (STFT) to the complex baseband stream produces a time-frequency representation called a spectrogram. This 2D image reveals:

• Frequency-hopping patterns in spread-spectrum signals
• Subcarrier activity in OFDM waveforms
• Spectral shape differences between PSK, FSK, and QAM families Spectrograms are fed into image-classification CNNs like ResNet or EfficientNet, treating modulation recognition as a visual pattern recognition problem.

Complex baseband representations enable adversarial perturbation analysis for AMC systems. Researchers inject carefully crafted perturbations directly into the IQ stream—imperceptible phase rotations or amplitude variations—to test classifier robustness. A perturbation of just 0.01 radians in instantaneous phase can cause a high-confidence misclassification from QPSK to 8-PSK, revealing vulnerabilities that must be hardened through adversarial training on the baseband domain.

Before classification, the complex baseband stream undergoes digital preprocessing to correct hardware and channel distortions:

• Carrier Frequency Offset (CFO) correction: Derotates the constellation to stop continuous spinning
• I/Q imbalance compensation: Restores orthogonality between I and Q branches
• DC offset removal: Centers the constellation at the origin These corrections ensure the classifier sees the true modulation structure rather than receiver artifacts, critical for low-SNR scenarios below 0 dB.