deeepeshthakur/sdelab2

A Julia package for numerically approximating stochastic differential equations (SDEs)

SDELAB2

This is a reimplementation of SDELAB, a MATLAB package for generating sample paths of initial-value problems for SDEs numerically, as described in

H. Gilsing and T. Shardlow. SDELab: a package for solving stochastic differential equations in MATLAB. In: J. Comput. Appl. Math. 205.2 (2007), pp. 1002–1018. issn: 0377-0427. doi: 10.1016/j.cam.2006.05.037 .

SDELAB2 is implemented in Julia and provides the functionality described in this paper, though there are changes in syntax. The key differences are

• drift and diffusion are autonomous (non-automous functions needs to disguised as an autonomous SDE)

• you specify the drift function (y->f(y)), diffusion-matrix function ((y,dw)->g(y) dw), and diffusion-vector function (dif_vec[i]=(y->gi(y)) (i.e., a vector of functions y->gi(y).

• automatic differentiation creates Jacobian fns for f (for nonlinear solve) and for the diffusion-vector function gi (for Milstein methods)

• the main driver routine is

t,y,W=sde_strong_solution(fcn, tspan, y0, opt)

where the return values are a vector of times t and corresponding matrix of solutions y (with rows corresponding to times), and the position W of the driving Brownian motion at the final time,

and the inputs are:

• fcn is a special data type for describing the SDE coefficient functions, created by

fcn=set_fcn(d,m,y->drift_f(y,p),(y,dw)->diff_g_mat(y,dw,sig),diff_g_vecs(sig), noise_type)

(here p and sig are parameters)

for d=phase-space dimension, m=number of Brownian motions,

• drift_f, diff_g_mat, diff_g_vecs are the coefficient functions for the SDE,

• noise_type=1 (unstructured), =2 (diagonal), =3 (commutative) (see the above paper for the meaning of this).

• tspan=well-ordered vector of times where the solution is required

• y0=d-vector of initial data

• opt=set_opt(dt,method), where dt=maximum time step,

• method="SIE0" or "SIE1" or "SIE" (Euler-Maruyama for explicit, implicit, or alpha-implicit),

• "SSEH0" or "SSEH1" or "SSEH" (explicit Stratonovich Heun for explicit, implicit, or alpha-implicit),

• "SIM0" or "SIM1" or "SIM" (for Ito Milstein),

• "SSM0" or "SSM1" or "SSM" (Stratonovich Milstein), or

• "SIBDF" (second-order BDF).

See above-referenced paper for details. All use auto-computed derivatives.

opt can be further adjusted to set

opt["Alpha"]=parameter alpha for alpha-implicit methods (default alpha=0.5; trapezium rule)

opt["nlsolve_param"].maxfeval and .ftol and .xtol to control nonlinear sovler (default 10,1e-6,1e-8)

opt["calc_final_W"] gives final position W of driving Wiener process (useful for calculating exact soln of geometric BM)

opt["do_plotting"]=true or false, for automatic plotting via PyPlot

opt["always_clear"]=true or false, to clear current plot each time

opt["output_sel"]= vector of indicies to plot (default to [1,2])

opt["output_plot_type"]="path_plot" or "phase_plot" ("time_phase" and "phase3_plot" not yet implemented).

SDELab2 depends on the Julia packages NLsolve and PyPlot.

The package is currently being refined. The four examples presented (geometric Brownian motion, Van der Pol, FitzHugh-Nagumo PDE, and Hodkin-Huxley) illustrate its use.