Doppler Shift Estimation

The maximum Doppler shift is defined as f_d = (v * f_c) / c, where v is the relative velocity, f_c is the carrier frequency, and c is the speed of light. This parameter defines the rate of channel variation.

• High-Speed Rail: At 500 km/h and 3.5 GHz, f_d ≈ 1.6 kHz.
• V2X Communication: At 120 km/h and 5.9 GHz, f_d ≈ 656 Hz.
• Coherence Time: The channel's stationary interval is inversely proportional to f_d (T_c ≈ 0.423 / f_d).

The Level Crossing Rate quantifies how often the signal envelope crosses a specified amplitude threshold in the positive direction. It is a second-order statistic crucial for designing adaptive modulation and interleaving depth.

• Rayleigh Fading: LCR is directly proportional to f_d and the target threshold normalized to the RMS signal level.
• Burst Error Design: High LCR values dictate shorter packet lengths to avoid deep fades.
• Estimator Input: Used as a feature in classical maximum-likelihood velocity estimators.

Classical velocity estimation relies on the autocorrelation function of the channel impulse response. The zeroth-order Bessel function of the first kind, J_0(2π f_d τ), models the temporal correlation.

• Jakes' Model: Assumes isotropic scattering, allowing f_d to be solved by inverting the Bessel function at a known lag τ.
• Covariance Fitting: Modern estimators fit the empirical channel covariance to theoretical models to jointly estimate Doppler spread and K-factor.
• Limitations: Performance degrades in non-isotropic scattering (e.g., highway scenarios with dominant line-of-sight).

Neural networks are increasingly used to predict Doppler shift directly from raw channel estimates, bypassing explicit statistical modeling.

• Recurrent Neural Networks (RNNs): LSTMs process sequential Channel Impulse Response (CIR) samples to forecast phase evolution.
• Complex-Valued Networks: Preserve phase information crucial for Doppler estimation, outperforming real-valued equivalents.
• Transformer Architectures: Self-attention mechanisms capture long-range dependencies in time-varying channels, enabling robust prediction even during deep fades.

It is critical to distinguish between Doppler shift (a deterministic frequency translation) and Doppler spread (a stochastic spectral broadening).

• Doppler Shift: Caused by dominant line-of-sight motion; corrected via frequency offset estimation (e.g., CFO compensation).
• Doppler Spread: Caused by multipath reflections arriving at different angles; defines the coherence bandwidth of the time-varying channel.
• Estimation Impact: Shift requires phase-locked loop tracking; spread requires pilot density adaptation to sample the channel above the Nyquist rate in the time domain.