1: \begin{abstract}

2: 

3: We explore the use of probability models for uncertainty arising from discretization of system states defined implicitly by ordinary or partial differential equations.  Accounting for this uncertainty is vital for characterising potential inferential bias incurred when likelihoods are constructed based on a numerical approximation over a finite discretization grid.  A formalism for inferring fixed but a priori unknown model trajectories is proposed within the forward problem through Bayesian updating of sequentially refined prior models conditional on model interrogations.  A one-step-ahead sampling scheme for interrogating the model is studied in detail, its consistency and first order convergence properties are proved.  The proposed approach is demonstrated to capture the functional structure and magnitude of the discretization error, while attaining computational scaling of the same order as traditional first order numerical methods, and providing a formal statistical trade-off between accuracy and discretization grid size.  Examples illustrate the flexibility of this framework to deal with a wide variety of models that include initial value, delay, and boundary value ordinary differential equations, as well as partial differential equations.  Scalability is illustrated for a model of fluid dynamics consisting of over 16,000 coupled, stiff ordinary differential equations.  Within the inverse problem, discretization uncertainty is defined as an additional layer in the Bayesian hierarchy, and a Markov chain Monte Carlo sampling scheme that targets the resulting posterior distribution is developed.  This formalism is used to infer states and parameters for the JAK-STAT delay differential equation network of protein dynamics from indirectly observed measurements. Finally, the discussion outlines a wide range of emerging research themes in the new field of probabilistic numerics that naturally follow from the work presented.

4: 

5: \end{abstract}

6: