To address the problem of the great variability of the mechanical state of subduction zones, we investigate the mechanics of back arc spreading and seismic decoupling. Back arc spreading is assumed to be due to rifting of the upper plate and hence occurs when trench-normal tension reaches a critical value. Seismic decoupling is assumed to occur when the normal stress at the frictional interface is decreased by an amount sufficient to cross the friction stability transition. Two forces are important in this problem. The first is a small component of the slab pull force which remains unbalanced by the subduction resistance and exerts a vertical suction force at the trench. The second is a sea anchor force exerted on the slab that resists its lateral motion, assumed to occur at the upper plate velocity. Both forces contribute to the coupling problem: only the sea anchor force is responsible for back arc spreading. The unbalanced slab pull force is determined from a force balance for subduction, the sea anchor force is computed as the hydrodynamic resistance to the facewise translation of an elliptical disc through a viscous fluid. The model predicts three regimes: seismically coupled compressional arcs with advancing upper plates; seismically decoupled extensional arcs with retreating upper plates, and strongly extensional arcs which also have back arc spreading. This model is applied in detail to the Izu-Bonin-Marianas system. It shows that back arc spreading occurs when the integrated tension in the upper plate exceeds a value of about 1×1013 N m−1 and requires a residual tension of about a third that to drive the back arc spreading once rifting is completed. It shows why the plate interface near Guam is seismically coupled, while the plate boundary everywhere farther north is decoupled. When applied globally, it successfully predicts the state of seismic coupling and back arc spreading in more than 80% of the world's subduction zones. Of the remaining, half can be seen to contain additional complications not included in the model. About 10% of cases remain unexplained, but some of these may have incorrectly determined seismic coupling coefficients.