ConNet: Complex Orthogonal Network With Low-Rank Regularization for CSI Localization

Channel State Information (CSI) fingerprinting can achieve high in-domain accuracy for WiFi localization, yet its performance often degrades severely after deployment in a new environment, hindering rapid transfer. For cross-domain passive localization, we first provide a closed-form prediction error bound (PEB) as an analytical design reference, showing that localization is primarily governed by relative phase structures across antennas and subcarriers, whereas packet-wise global complex scaling acts as a nuisance transformation. Guided by this insight, we propose ConNet, a CSI-specific complex-valued model consisting of a complex-scale-aware backbone and a hard-coupled transfer module that applies Gram-Schmidt orthogonalization before nuclear norm regularization. The resulting design preserves informative complex phase structure, reduces multipath-induced feature redundancy, and promotes a compact domain-robust subspace. Experiments under five-domain source rotation demonstrate improved cross-domain transfer performance and effective location-wise few-shot adaptation.