Carrier Frequency Offset (CFO)

CFO introduces a linear phase rotation that accumulates over time across OFDM symbols. The received baseband signal is multiplied by a complex exponential term e^(j2πΔft), where Δf represents the frequency mismatch between transmitter and receiver local oscillators. This causes the entire IQ constellation to spin at a constant angular velocity proportional to the offset magnitude. For a CFO of 1 kHz, the constellation completes one full rotation every millisecond, rendering higher-order QAM constellations like 64-QAM or 256-QAM completely unintelligible without correction.

In multi-carrier systems like OFDM, CFO destroys orthogonality between subcarriers. The frequency offset causes energy from one subcarrier to leak into adjacent subcarriers, creating inter-carrier interference. This ICI power increases with both the magnitude of the offset and the number of subcarriers. Even a small fractional offset of 1-2% of the subcarrier spacing can cause a severe signal-to-interference-plus-noise ratio (SINR) floor, beyond which increasing transmit power provides no BER improvement.

CFO arises from two primary physical mechanisms: oscillator mismatch and Doppler shift. Manufacturing tolerances in crystal oscillators typically produce offsets of 1-20 parts per million (ppm), translating to several kHz at gigahertz carrier frequencies. Additionally, relative motion between transmitter and receiver induces a Doppler frequency shift proportional to the radial velocity and carrier wavelength. In high-mobility scenarios like vehicle-to-everything (V2X) or high-speed rail communications, Doppler-induced CFO can dominate and vary rapidly over time.

CFO estimation is typically divided into acquisition and tracking stages. Coarse estimation often uses a data-aided approach with known training sequences or preambles, exploiting the phase difference between identical repeated symbols (e.g., Schmidl-Cox algorithm). Fine tracking employs decision-directed loops that measure the phase rotation of decoded symbols. Correction is applied by multiplying the received samples by a complex sinusoid of the estimated negative frequency, effectively derotating the constellation back to its intended reference positions.

In learned communication systems, CFO presents a significant challenge for complex-valued neural networks (CVNNs) and autoencoder-based transceivers. A neural network trained on CFO-free data suffers catastrophic performance degradation when exposed to even small offsets. Robust solutions include data augmentation during training with a distribution of synthetic CFO values, dedicated CFO correction pre-processing layers, or end-to-end learning where the receiver jointly learns to estimate and compensate for the offset implicitly within its weight parameters.

CFO is commonly expressed in normalized form as ε = Δf / f_sub, where f_sub is the subcarrier spacing. This dimensionless quantity is decomposed into an integer part (ε_i) and a fractional part (ε_f). The fractional CFO causes the ICI and constellation rotation described above. The integer CFO shifts the subcarrier indices by a whole number of positions, causing a complete misalignment of the demapping process without destroying orthogonality. Estimation algorithms must resolve both components for successful demodulation.