Coherent Direct D-MIMO Localization
Distributed multiple-input multiple-output (D-MIMO) is envisioned as a key deployment architecture for future wireless systems, offering improved coverage and robustness through spatial separation, and favorable geometry for localization and sensing. Its greatest potential for localization lies in joint coherent processing across distributed antenna panels. However, stringent frequency-synchronization and phase-calibration requirements, together with multimodal likelihood functions, hinder the estimation process. Consequently, most existing algorithms process the panels noncoherently, potentially sacrificing localization accuracy. We present a unified family of Bayesian state-space filters that are based on concentrated Type-I and marginal Type-II likelihoods for wideband near-field D-MIMO systems and operate directly on noisy channel observations. The Type-I filters explicitly realize (i) noncoherent, (ii) coherent, and (iii) carrier-phase-based processing. For Type-II filtering, we show that a zero-mean model is inherently noncoherent under distributed processing, whereas observation stacking restores coherence. A nonzero-mean model can automatically adapt to the coherence available in the data, a property that we term ``soft coherence''. We derive posterior Cramér-Rao lower bounds (PCRLBs) for all three coherence levels and show that each level is fundamentally tied to the number of phase parameters used for positioning or treated as nuisance parameters. Numerical results show that the coherence-specific filters closely approach their respective PCRLBs and that coherent processing can substantially outperform noncoherent processing. We derive particle-based belief propagation methods, which parallelize over particles and distributed panels, scale linearly with the observed data, and achieve runtimes of tens of milliseconds per time step in a GPU-accelerated implementation.