Contents

# Overview¶

NodePy (Numerical ODEs in Python) is a Python package for designing, analyzing, and testing numerical methods for initial value ODEs. Its development was motivated by my own research in time integration methods for PDEs. I found that I was frequently repeating tasks that could be automated and integrated. Initially I developed a collection of MATLAB scripts, but this became unwieldy due to the large number of files that were necessary and the more limited capability for code reuse.

NodePy represents an object-oriented approach, in which the basic object is a numerical ODE solver. The idea is to design a laboratory for such methods in the same sense that MATLAB is a laboratory for matrices. Some distinctive design goals are:

Plug-and-play: any method can be applied to any problem using the same syntax. Also, properties of different kinds of methods are available through the same syntax. This makes it easy to compare different methods.Abstract representations: Generally, the most abstract (hence powerful) representaton of an object is used whenever possible. Thus, order conditions are generated using products on rooted trees (or other recursions) rather than being hard-coded.Numerical representation: The most precise representation possible is used for quantities such as coefficients: rational numbers (using SymPy’s Rational class) when available, floating-point numbers otherwise. Where necessary, method properties are determined by numerical calculations, using appropriate tolerances. Thus the “order” of a method with floating-point coefficients is determined by checking whether the order conditions are satisfied to within a small value (near machine-epsilon). For efficiency reasons, coefficients are always converted to floating-point for purposes of applying the method to a problem.

In general, user-friendliness of the interface and readability of the code are prioritized over performance.

NodePy includes capabilities for applying the methods to solve systems of ODEs. This is mainly intended for testing and comparison; for realistic problems of interest in most fields, time-stepping in Python will be too slow. One way around this is to wrap Fortran or C functions representing the right-hand-side of the ODE, and we are looking into this.

Note

The user guide is currently quite incomplete, and in general is not expected to keep pace with all NodePy development. However, NodePy includes substantial inline documentation within the various modules, and you are encouraged to look there.

## Dependencies¶

- Python 2.7
- Numpy, Matplotlib
- SymPy (note: NodePy is now compatible with SymPy 0.7.1)
- Optional: networkx (for some Runge-Kutta stage dependency graphing)

## Classes of Numerical ODE Solvers¶

- NodePy includes classes for the following types of methods:
- Runge-Kutta Methods
- Implicit
- Explicit
- Embedded pairs
- Low-storage methods
- Extrapolation methods
- Integral deferred correction methods
- Strong stability preserving methods
- Runge-Kutta-Chebyshev methods

Arbitrary methods in these classes can be instantiated by specifying their coefficients.

## Analysis of Methods¶

- NodePy includes functions for analyzing many properties, including:
- Stability:
- Absolute stability (e.g., plot the region of absolute stability)
- Strong stability preservation

- Accuracy
- Order of accuracy
- Error coefficients
- Relative accuracy efficiency
- Generation of Python and MATLAB code for order conditions

## Testing Methods¶

NodePy includes implementation of the actual time-stepping algorithms for the various classes of methods. A wide range of initial value ODEs can be loaded, including the DETEST suite of problems. Arbitrary initial value problems can be instantiated and solved simply by calling a method with the initial value problem as argument. For methods with error estimates, adaptive time-stepping can be used based on a specified error tolerance. NodePy also includes automated functions for convergence testing.

In the future, NodePy may also support solving semi-discretizations of initial boundary value PDEs.