A fox catches a rabbit. That's one event. The rabbit population drops by one, and the fox gains a meal that will fund reproduction. Two variables change — prey count and predator resource — but the cause is a single, indivisible act. The statistical signature of that act is a negative correlation: when prey decreases, predator food increases.
Standard ecological models don't represent it that way. The conventional approach adds noise to predator and prey equations independently — diagonal diffusion, in the technical language. Each species gets its own random perturbation, uncorrelated with the other. Yu and Wang (arXiv 2602.22489) show that this is not an approximation. It's a structural error. They derive the correct noise from the underlying Markov chain of individual predation and conversion events, and the result has off-diagonal terms — negative cross-covariance between predator and prey fluctuations — that diagonal models discard entirely.
What's lost isn't a small correction. The cross-covariance carries the causal coupling. It determines extinction dynamics, stability boundaries, and the shape of population trajectories under uncertainty. Removing it doesn't simplify the truth; it replaces it with a different truth — one where predator and prey fluctuate for independent reasons that happen to coexist. The model says: these two populations live in the same space. The reality says: these two populations are the same event viewed from opposite sides.
The error is seductive because it's convenient. Diagonal noise matrices are easier to handle analytically. Stochastic differential equations with off-diagonal diffusion require more careful boundary treatment, different numerical schemes, more parameters to estimate. The shortcut saves work at every stage of the analysis. It also happens to remove exactly the information that distinguishes coupled dynamics from coincidental proximity.
This is not unique to ecology. Any system with inherently coupled events — an infection that simultaneously removes a susceptible and creates an infected, a trade that simultaneously debits a buyer and credits a seller, a chemical reaction that consumes reactants and produces products at the same instant — has off-diagonal noise structure. The diagonal approximation is the same error repeated across fields, each time justified by convenience, each time discarding the same kind of information: the fact that some fluctuations have a common cause.
The general lesson: when a model treats a shared event as two independent perturbations, the lost information is the coupling itself. And the coupling is usually what you most want to understand.