Fast Timescales in Stochastic Population Dynamics

  • George Constable

Student thesis: Phd


In this thesis, I present two methods of fast variable elimination in stochastic systems. Their application to models of population dynamics from ecology, epidemiology and population genetics, is explored. In each application, care is taken to develop the models at the microscale, in terms of interactions between individuals. Such an approach leads to well-defined stochastic systems for finite population sizes. These systems are then approximated at the mesoscale, and expressed as stochastic differential equations. It is in this setting the elimination techniques are developed. In each model a deterministically stable state is assumed to exist, about which the system is linearised. The eigenvalues of the system's Jacobian are used to identify the existence of a separation of timescales. The fast and slow directions are then given locally by the associated eigenvectors. These are used as approximations for the fast and slow directions in the full non-linear system. The general aim is then to remove these fast degrees of freedom and thus arrive at an approximate, reduced-variable description of the dynamics on a slow subspace of the full system. In the first of the methods introduced, the conditioning method, the noise of the system is constrained so that it cannot leave the slow subspace. The technique is applied to an ecological model and a susceptible-exposed-infectious-recovered epidemiological model, in both instances providing a reduced system which preserves the behaviour of the full model to high precision. The second method is referred to as the projection matrix method. It isolates the components of the noise on the slow subspace to provide its reduced description. The method is applied to a generalised Moran model of population genetics on islands, between which there is migration. The model is successfully reduced from a system in as many variables as there are islands, to an effective description in a single variable. The same methodology is later applied to the Lotka-Volterra competition model, which is found under certain conditions to behave as a Moran model. In both cases the agreement between the reduced system and stochastic simulations of the full model is excellent.It is stressed that the ideas behind both the conditioning and projection matrix methods are simple, their application systematic, and the results in very good agreement with simulations for a range of parameter values. When the methods are compared however, the projection matrix method is found in general to provide better results.
Date of Award1 Aug 2015
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorAlan Mckane (Supervisor)


  • population genetics
  • timescale separation
  • stochastic dynamics
  • population dynamics

Cite this