De Filippis and Mingione solved a class of elliptic partial differential equations that had resisted direct analysis for decades. The equations describe how gradients behave in materials with multiple phases — situations where the mathematical structure changes character depending on the local conditions. The difficulty was that the equations couldn't be manipulated by standard techniques; the nonlinearity was too severe, the growth conditions too irregular.
Their method was to stop trying to solve the original equation. Instead, they derived a shadow equation — a “ghost” that preserved the essential structural features of the original while being tractable. They solved the ghost, used its solution to constrain the original, then iteratively refined the ghost to make the constraints tighter, eventually recovering the gradient regularity they needed from the original equation.
The strategy is: when the function is intractable, study a derivative object. Not the derivative in the calculus sense, but an object that is derived from the original — that inherits some of its structure while shedding the properties that make it impossible to work with. The ghost equation has the same scaling behavior, the same critical exponents, the same phase transition structure as the original, but it's amenable to linearization. The information lost in the derivation is exactly the information that was causing the obstruction.
This is a general strategy disguised as a specific proof. When a system is too complex to characterize directly, construct a simplified object that preserves the structural features you care about and discard the rest. Measure the simplified object. Transfer the measurements back to the original. The key insight is that the simplification isn't an approximation — it's a projection onto a subspace where the relevant quantities are computable. The ghost equation doesn't approximate the solution; it exactly captures the gradient behavior while discarding the pointwise solution values.
The same logic operates outside mathematics. Perturbation analysis studies how systems change rather than what they are — the derivative of behavior with respect to input, which is smoother and more tractable than the behavior itself. Reaction-time priming measures unconscious representations by their effect on subsequent processing, not by introspection. Radiocarbon dating measures the absence of an isotope rather than the presence of a fossil. In each case, the derivative object — the thing constructed from the original — is more informative than the original because the construction strips away the noise that makes direct measurement impossible.