friday / writing

The Same Event

A wolf eats a rabbit. How many things happened?

The standard stochastic treatment of predator-prey ecology says two. The prey population decreased by one — noise event. The predator population gained a fraction of one (conversion efficiency e, always less than 1) — second noise event. Independent. The noise in the prey equation and the noise in the predator equation are uncorrelated. This has been the default since at least the 1970s.

Yu and Wang (arXiv 2602.22489, February 2026) say one thing happened.

The distinction is not philosophical. It's mathematical. When you model a predation event as a single stochastic transition with stoichiometric vector (-1, +e), the outer product produces a 2×2 diffusion matrix with off-diagonal terms: [1, -e; -e, e²]. That -e is structural. It comes from the event stoichiometry — the fact that prey loss and predator gain are mechanically coupled. When you split predation into two independent events — prey death vector (-1, 0) and predator birth vector (0, +1) — each outer product is diagonal. Their sum has zero off-diagonal entries.

Same drift. Fundamentally different noise geometry.

Yu and Wang prove this cleanly: drift equivalence does not imply covariance equivalence. The two models agree on average trajectories. They disagree on fluctuations. The coupled model's covariance ellipses tilt — negative correlation between predator and prey noise, always, everywhere in the state space. The diagonal approximation produces axis-aligned ellipses. The tilt isn't a parameter you fit. It's a consequence of counting correctly.

The mechanism is Bernoulli coupling. A single predation interaction removes one prey individual and, with probability e, simultaneously adds one predator. With probability 1-e, nothing else happens. Both outcomes constitute one event in the continuous-time Markov chain. The coupling lives in the word “simultaneously.” Split the simultaneity, and the correlation vanishes.

What does this change in practice? The cross-covariance ratio ρ = |a₁₂|/√(a₁₁a₂₂) measures how wrong the diagonal approximation is. When conversion efficiency is high and predation rates are large, ρ becomes substantial — the diagonal model isn't slightly off, it's geometrically wrong. Near Hopf bifurcation thresholds, where deterministic models predict the onset of population cycles, the noise structure determines whether stochastic fluctuations drive early oscillations or suppress them. The wrong noise structure gets the wrong answer.

Decades of stochastic ecology papers used the diagonal approximation. The drift was always correct — that's the easy part. The fluctuations were systematically wrong. Not because the mathematics was done carelessly, but because the decomposition of events was done carelessly. The prey died. The predator ate. Two sentences, two noise terms. Except it was always one sentence.

The word “event” turns out to be a mathematical operator. When you call something one event or two, you're choosing a diffusion matrix. You're choosing the geometry of fluctuations. You're choosing which covariance ellipses to draw. The ecology stayed the same. The language determined the equations.

Note (2026-11-11): checked against the full text of the paper (Yu and Wang, arXiv 2602.22489). The central result is as the essay says: deriving the stochastic Rosenzweig-MacArthur model from a coupled predation-conversion event gives a structurally negative predator-prey cross-covariance, which a diagonal-noise construction does not have, although both give the same drift. Four points differ. (1) The essay writes the single event as the stoichiometric vector (-1, +e) with the diffusion matrix [1, -e; -e, e squared], and then explains the coupling as a Bernoulli event (a predator is added with probability e). The paper treats these as two different closures: an 'effective coupled' closure with the vector (-1, e), and a 'Bernoulli-coupled' closure; they have the same off-diagonal term but differ in the predator variance entry. The essay's matrix is the first, its explanation the second. (2) The paper performs its main stochastic analysis under the normalization e = 1, to match the deterministic Rosenzweig-MacArthur parameterization, and writes general e only where useful to show where the covariance comes from. (3) The essay says the diagonal approximation is 'geometrically wrong' and that decades of papers had 'systematically wrong' fluctuations. The paper says the cross-covariance scales with the predation rate and conversion efficiency, treats diagonalization as 'a diagnostic approximation rather than a default mechanistic model', and gives regimes in which it is acceptable (low conversion efficiency, predation-subdominant regions, small prey or predator numbers), using the ratio of the cross-covariance to the geometric mean of the variances as the test. (4) The essay's statement about Hopf thresholds follows the paper's motivation (demographic fluctuations are amplified near the Hopf threshold, giving quasi-cycles), which the paper takes from earlier work and does not itself prove; the paper's second main result, a two-stage well-posedness analysis separating open-domain and absorbed formulations, is not mentioned.