In this paper, we advocate the Decoupled Error and State (DES) methodology for state estimation, which uses distinct representations for error and state estimates and updates the state through tailored functions based on the selected representations. Focusing on Visual-Inertial Navigation Systems (VINS), for the first time, we analytically discover the connections between the prominent VINS estimators, offering a unified view and insightful understanding of the SOTA algorithms. Building upon this discovery along with the proposed DES idea, we further develop the DES-VINS. The proposed estimator adopts a global-centric state to naturally represent the physical quantities concerned by the underlying navigation system, while designing a new error representation by lifting orientation to mitigate the issues caused by linearization and ensure proper observability properties. Interestingly, despite not being constrained by the Lie-group affine properties (which are often challenging to ensure in practice), the DES-VINS estimator is shown to share the identical properties of the linearized error-state system as the invariant EKF. However, the DES-VINS algorithm allows efficient and consistent integration of long-tracked SLAM features (which are almost always needed in practice), being 3 \(\times \) faster than the invariant VINS. Extensive numerical studies and real-world experiments are presented to compare with SOTA VINS estimators, providing valuable insights into their performance and applicability.