The Lotka-Volterra equations describe predator-prey dynamics deterministically: prey grow, predators eat prey and convert them into new predators, both populations oscillate. Stochastic versions add noise to capture the randomness of individual births, deaths, and encounters. The standard approach adds independent noise to each population — diagonal diffusion terms, one for prey fluctuations and one for predator fluctuations, uncoupled from each other. This is mathematically convenient and biologically intuitive: each species experiences its own demographic randomness.
Yu and Wang (arXiv 2602.22489, February 2026) show that this is wrong. Not approximately wrong — structurally wrong. The noise in predator and prey populations is inherently coupled, and the coupling has a specific sign that diagonal models cannot capture.
The argument starts from the microscopic events. A predation event simultaneously removes one prey individual and (after some conversion efficiency) adds a fraction of a predator individual. This single event decreases the prey count and increases the predator count. The fluctuations in the two populations from this event are anticorrelated: when one goes up, the other goes down. This anticorrelation is not a perturbation on top of independent fluctuations — it is built into the mechanism of predation itself. Every capture is a coupled event.
When the stochastic model is derived from first principles — starting from the Markov chain of individual events and taking the appropriate limit — the diffusion matrix that emerges has negative off-diagonal terms. The prey-predator covariance is structurally negative: a mathematical consequence of the fact that the same event (predation) decreases one population and increases the other. Diagonal noise models set these off-diagonal terms to zero, discarding a correlation that the biology requires.
The consequences are not merely quantitative. Diagonal noise models can produce qualitatively different behavior from the full covariance model: different stability properties, different extinction probabilities, different relationships between population size and fluctuation amplitude. The standard intuition that “adding noise” to a deterministic model captures demographic stochasticity holds only if the noise has the right structure. Independent noise for coupled populations is not the right structure when the coupling operates at the level of individual events.
The problem extends beyond predator-prey models. Any ecological interaction where a single event affects two populations simultaneously — parasitism, mutualism, competition for a shared resource — generates cross-covariance in the noise. Models that treat each population's fluctuations as independent are discarding information that the interaction itself produces. The standard modeling assumption of diagonal noise is not a simplification that preserves the essential dynamics — it is a simplification that removes the signature of the interaction from the fluctuations.
The fix is straightforward: derive the noise structure from the microscopic events rather than assuming it. The resulting non-diagonal diffusion matrix is no harder to work with computationally, and the derivation from the Markov chain is standard. The diagonal approximation was never necessary — it was a habit inherited from models of single-species dynamics, where demographic noise is genuinely independent because there is only one population. Extending that habit to multi-species systems assumed something false: that the noise in interacting populations is independent of the interaction.