Universality classes work because they assume a single characteristic length dominates. Marquis, Gallotti, and Barthelemy show what happens when that assumption fails.
Their model deposits blobs onto a growing surface with power-law size distributions. When the exponent is steep enough (τ ≥ 3), rare large events are suppressed and the system falls into Kardar-Parisi-Zhang universality as expected. But when τ drops below 3, the largest deposited cluster introduces its own dynamical length scale — ζ, the linear size of the biggest blob — alongside the standard correlation length ξ. Two rulers measuring the same surface at different rates.
The Family-Vicsek scaling framework assumes you can collapse all the physics onto one rescaled variable. That collapse requires one length to dominate. When ζ and ξ coexist, critical exponents stop being fixed and vary continuously with τ. Universality doesn't fail because the system becomes random or chaotic. It fails because the system develops a second opinion about what “large” means.
The through-claim: universality is not a property of the dynamics. It is a property of scale separation — specifically, the condition that one scale wins. When rare events are rare enough to ignore, one scale wins and the standard classes apply. When they are not, the mathematical machinery that produces universality (the renormalization group, the scaling collapse) loses its footing. The breakdown is not in the physics but in the assumption that there is only one relevant way to measure the system.
This is worth remembering whenever universality feels like inevitability. It works when it works because something is negligible. The rare event is the thing you decided not to measure.