WorksAgricultural and Biological Sciences
Fractional Order or False Memory: Identifiability of Caputo SIR Models from Incidence Data
Abstract
When a fractional order improves the fit of an epidemic model, it is usually taken as evidence of memory. We ask whether the order of the Caputo susceptible-infectious-removed (SIR) model can actually be recovered from incidence data. If the transmission rate is allowed to vary freely in time, as in hybrid forecasting models, every admissible order, removal rate and initial count reproduces the data exactly. With constant transmission and exact data the order is identifiable in principle, but the dominant term of early linearized growth depends on only two parameter combinations: the growth rate r = (beta - gamma)^(1/alpha) and the amplitude beta*I0/alpha. What is left is transient, and it is limited. We prove that parameter sets sharing these two combinations give Poisson count distributions whose Kullback-Leibler divergence stays bounded no matter how long the linear phase is observed, and that the nonlinear model inherits this bound as the population grows. In simulations, wrong orders were rejected in at most 5% of replicates when the data ended two weeks or more before the epidemic peak, and in 95-100% once the peak was included; separation began only a few days before the peak. With mild overdispersion the order was identified only once about 60% of the population had been infected, and with stronger overdispersion not even then. Fits that the early data could not tell apart predicted peaks that differed by 75%, and plug-in forecasts that ignored parameter uncertainty were badly overconfident. Data generated without any memory produced fitted orders below one when transmission changed and when the fitted model was discretized in daily steps. In early COVID-19 data from Bangladesh and the United States, orders from 0.6 to 1 gave almost the same growth rate, whatever value we assumed for the removal rate.