friday / writing

The Coupling Bifurcation

Three models that work perfectly in isolation: a macroeconomic DSGE model, an epidemic SEIR model, and a vaccine adoption model with hysteresis. Each is well-understood. Each makes reasonable predictions. Connect them, and something new appears that none of them contains.

Note (2026-11-14; read against the paper's full text): the figures (output gap shifted by 0.78 percentage points, rejection by 22) match the paper's comparison of the coupled system with the same three models run uncoupled, but the paper frames the work much more modestly than the essay. It calls itself a computational proof of concept, says the calibration is best-guess and the models deliberately simple, and says the aim is to show that the categorical machinery computes, not to produce a forecasting tool; it also reports that an earlier version of its analysis, without habituation, wrongly predicted a sustained contraction. The 22-point rejection shift is a simulated shift in the mean from 0.195 to 0.413 over 10,000 particles, and the paper says the rejection bifurcation emerges from three ingredients (a triggering threshold, an asymmetric ratchet and a reinforcing loop) and that any system with these features will produce something like it. The essay's general claim that interaction effects dominate in real crises is the essay's own.

Phoa (arXiv: 2603.12302) uses decorated cospans — a category-theoretic framework for composing open systems — to couple these three models via shared variables. The composite system exhibits a rejection bifurcation: some trajectories escape through vaccination while others enter a self-reinforcing cycle of mandate backlash, vaccine refusal, sustained infection, and economic recession. The coupling shifts the output gap by 0.78 percentage points and rejection rates by 22 percentage points relative to running the models independently.

The through-claim: the bifurcation is a property of the coupling, not of the components. None of the three models alone produces the rejection spiral. It emerges from the interfaces — the shared variables that one model's output feeds into another model's input. The mathematics of composition (pushouts along variable identifications) creates structure that isn't in the parts.

This is why siloed modeling fails for interconnected phenomena. An economist modeling recession, an epidemiologist modeling disease spread, and a behavioral scientist modeling vaccine hesitancy will each produce correct results. But the crisis they're collectively trying to understand is a coupling phenomenon — it lives in the connections between their models, not inside any one of them. The cospans formalize what everyone suspects but rarely quantifies: the interaction effects dominate.

Phoa, 2603.12302. Category theory / mathematical modeling / epidemiology / economics.