The diffusion equation is the most well-behaved partial differential equation in physics — until you try to make it relativistic. Apply a Lorentz boost and the equation develops exponential instabilities. Solutions that decay in the rest frame grow without bound in the moving frame. The initial-value problem becomes ill-posed. This isn't a technicality. It means that Fick's law of diffusion, which works perfectly well in a lab, is fundamentally incompatible with special relativity as a standalone equation.
The problem has been known for decades. The usual resolution is to replace the diffusion equation with something more complicated — a hyperbolic equation, a telegrapher's equation, something that propagates signals at finite speed. These replacements work mathematically but feel like admissions of defeat. The diffusion equation is simple and correct in its domain. The question is whether it can be saved.
Gavassino (2602.21254) saves it by embedding it. Diffusion is not a fundamental equation — it's the hydrodynamic sector of a deeper kinetic theory. Specifically, Fick-type diffusion arises exactly from the relativistic Fokker-Planck equation in the long-wavelength limit. The kinetic theory is relativistically well-posed. The diffusion equation is its low-energy projection. The instabilities appear when you boost the projection without boosting the full theory.
The fix is to formulate the initial-value problem within the space of density profiles that admit a kinetic-theory realization. Not every density profile does — only those that can be the hydrodynamic projection of a well-behaved distribution function. These turn out to be band-limited functions. The ultraviolet modes that cause the instabilities are precisely the modes that have no kinetic realization. They are mathematical ghosts — solutions of the diffusion equation that don't correspond to any physical state.
Within this restricted space, the evolution is well-posed both forward and backward in time. Gavassino obtains a closed-form Green function — a Shannon-Whittaker-type kernel defined on the full Minkowski plane. The diffusion equation, properly embedded, doesn't just survive the Lorentz boost. It becomes time-reversible within its physical domain.
The Shannon-Whittaker connection is deep. Band-limited functions are exactly the functions that can be reconstructed from discrete samples — the sampling theorem of signal processing. So the physical initial conditions for boosted diffusion are precisely those that can be specified by their values at a discrete lattice of points. The resolution of the lattice is set by the kinetic theory, not by the diffusion equation. The macroscopic equation inherits a microscopic length scale through its embedding.
This is a case study in the relationship between effective theories and their parent theories. The diffusion equation is an effective theory — valid in a limited regime, derived by discarding microscopic details. In its own regime, it's exact. But it lacks the information needed to define its own boundaries. The instabilities under boosting are what happens when you push an effective theory past its domain without knowing where the domain ends. The parent theory (kinetic theory) knows the boundary because it contains the microscopic information. The embedding doesn't add physics to the diffusion equation — it adds the information about what the diffusion equation is allowed to say.
The backward-in-time well-posedness is the most surprising consequence. Ordinary diffusion is irreversible — entropy increases, information is lost, you can't unmix the dye. But this irreversibility is a property of the full space of initial conditions. Restricted to band-limited functions, the evolution is invertible. The irreversibility comes not from the dynamics but from the mismatch between the mathematical space (all functions) and the physical space (band-limited functions). Physics is reversible. The irreversibility is in the description, not the world.