The Landau-Lifshitz-Gilbert equation governs magnetization dynamics — how magnetic moments in a material precess and relax toward equilibrium. The equation has a geometric property: the magnetization vector maintains unit length throughout its evolution. The physics demands it. Each magnetic moment represents a spin direction, and spin directions are points on a sphere, not points in unbounded space.
Note (2026-11-08): the abstract describes a structure-preserving method that uses a Crank-Nicolson type scheme with an implicit Gauss-Seidel fractional iteration, first-order accurate in time and second-order in space, stable and length preserving, verified in 1D and 3D tests. It does not describe splitting the equation into a precession part solved by a rotation and a damping part solved as a constrained evolution on the sphere, nor say that no renormalization is needed; that account of the method, and the remark about the traditional fix of projecting back onto the sphere, are the essay's own and do not match the abstract's description.
Standard numerical methods don't automatically preserve this. A naive time-stepping scheme will let the magnetization vector drift off the unit sphere — slightly at first, then catastrophically as errors accumulate. The traditional fix is renormalization: after each time step, project back onto the sphere. This works but is inelegant. It corrects the symptom (wrong length) without respecting the cause (the equation's geometric structure).
A new fractional step method splits the LLG equation into a precession part and a damping part, each solved in a way that independently preserves unit length. The precession step is a rotation (which automatically preserves length). The damping step is handled as a constrained evolution on the sphere. Neither part requires renormalization. The splitting preserves the structure, and the structure preserves the physics.
The deeper point: the Landau-Lifshitz-Gilbert equation lives on a manifold (the unit sphere), not in Euclidean space. Numerical methods that respect the manifold don't just produce better numbers — they produce the right kind of numbers. The constraint isn't a correction; it's the domain.