Leibenson's equation on graphs

What the math gives to ML

The paper studies a nonlinear diffusion flow on graphs in which node states are first transformed pointwise by a power and then diffused through a discrete p-Laplacian. This gives a conservative message-passing primitive whose behavior differs from ordinary linear graph convolution: p controls sensitivity to edge gradients, while q controls state-dependent diffusivity. The direct neural transfer is a graph layer implemented as several stable explicit diffusion steps, with nonnegativity and antisymmetric edge fluxes providing architectural stability; suitable parameter regimes may also promote sparsity or reduce harmful oversmoothing.