The paper studies a mathematical object called Dyson Brownian motion, which describes how a collection of particles evolve randomly over time while repelling each other. This kind of process appears naturally in the study of large random matrices, where the particles represent eigenvalues. The central question is: where do these particles actually live at any given moment? There is a well-known theoretical prediction, given by something called the semicircular free-convolution flow, which tracks how the overall density of particles should evolve. The paper proves rigorously that the actual particles stay extremely close to the predicted support region (the set where the density is nonzero) with overwhelming probability, at all times up to any fixed horizon.
What makes this result notable is its generality. Previous results often required the particle configuration to start in a nice, well-behaved state, especially near the edges of the support. This paper imposes essentially no such conditions. The starting configuration just needs to have bounded spread. The result also handles situations where the support breaks into several disconnected pieces with genuine gaps between them, which is technically much harder than the single connected case. The guarantee holds from time zero onward, not just after some warm-up period.
The key technical innovation is what the authors call a "local resolvent exclusion principle." The resolvent is a complex-analytic tool commonly used to study eigenvalues indirectly, by analyzing certain transforms of the spectral measure rather than the eigenvalues themselves. The authors show that if this transform looks small above a particular point on the real line, and that point is separated from where eigenvalues are supposed to be, then no eigenvalues can actually be near that point. This turns abstract analytic estimates into concrete statements about particle locations, and because it is formulated in a model-independent way, it should be useful well beyond this specific setting.