CoupledSDEs

CoupledSDEs <: ContinuousTimeDynamicalSystem
CoupledSDEs(f, u0 [, p]; kwargs...)

(alg = SOSRA(), abstol = 1.0e-2, reltol = 1.0e-2)

Examples defining stochastic dynamics

Let's look at some examples of the different types of stochastic systems that can be defined.

For simplicity, we choose a slow exponential growth in 2 dimensions as the deterministic dynamics f:

using DynamicalSystemsBase, StochasticDiffEq, DiffEqNoiseProcess
using CairoMakie
f!(du, u, p, t) = du .= 1.01u # deterministic part

function plot_trajectory(Y, t)
    fig = Figure()
    ax = Axis(fig[1,1]; xlabel = "time", ylabel = "variable")
    for var in columns(Y)
        lines!(ax, t, var)
    end
    fig
end;

plot_trajectory (generic function with 1 method)

Additive noise

When $g(u, p, t)$ is independent of the state $u$, the noise is called additive; otherwise, it is multiplicative. We can define a simple additive noise system as follows:

sde = CoupledSDEs(f!, zeros(2));

2-dimensional CoupledSDEs
 deterministic: false
 discrete time: false
 in-place:      true
 dynamic rule:  f!
 SDE solver:    SOSRA
 SDE kwargs:    (abstol = 0.01, reltol = 0.01, dt = 0.1)
 Noise type:    (additive = true, autonomous = true, linear = true, invertible = true)
 parameters:    SciMLBase.NullParameters()
 time:          0.0
 state:         [0.0, 0.0]

which is equivalent to

t0 = 0.0; W0 = zeros(2);
W = WienerProcess(t0, W0, 0.0)
sde = CoupledSDEs(f!, zeros(2);
    noise_process=W, covariance=[1 0; 0 1], noise_strength=1.0
    );

2-dimensional CoupledSDEs
 deterministic: false
 discrete time: false
 in-place:      true
 dynamic rule:  f!
 SDE solver:    SOSRA
 SDE kwargs:    (abstol = 0.01, reltol = 0.01, dt = 0.1)
 Noise type:    (additive = true, autonomous = true, linear = true, invertible = true)
 parameters:    SciMLBase.NullParameters()
 time:          0.0
 state:         [0.0, 0.0]

We defined a Wiener process W, whose increments are vectors of normally distributed random numbers of length matching the output of g. The noise is applied element-wise, i.e., g.*dW. Since the noise processes are uncorrelated, meaning the covariance matrix is diagonal, this type of noise is referred to as diagonal.

We can sample a trajectory from this system using the trajectory function also used for the deterministic systems:

tr = trajectory(sde, 1.0)
plot_trajectory(tr...)

Correlated noise

In the case of correlated noise, the random numbers in a vector increment dW are correlated. This can be achieved by specifying the covariance matrix $\Sigma$ via the covariance keyword:

ρ = 0.3
Σ = [1 ρ; ρ 1]
diffeq = (alg = LambaEM(), dt=0.1)
sde = CoupledSDEs(f!, zeros(2); covariance=Σ, diffeq=diffeq)

2-dimensional CoupledSDEs
 deterministic: false
 discrete time: false
 in-place:      true
 dynamic rule:  f!
 SDE solver:    LambaEM
 SDE kwargs:    (dt = 0.1,)
 Noise type:    (additive = true, autonomous = true, linear = true, invertible = true)
 parameters:    SciMLBase.NullParameters()
 time:          0.0
 state:         [0.0, 0.0]

Alternatively, we can parametrise the covariance matrix by defining the diffusion function $g$ ourselves:

g!(du, u, p, t) = (du .= [1 p[1]; p[1] 1]; return nothing)
sde = CoupledSDEs(f!, zeros(2), (ρ); g=g!, noise_prototype=zeros(2, 2))

2-dimensional CoupledSDEs
 deterministic: false
 discrete time: false
 in-place:      true
 dynamic rule:  f!
 SDE solver:    SOSRA
 SDE kwargs:    (abstol = 0.01, reltol = 0.01, dt = 0.1)
 Noise type:    (additive = true, autonomous = true, linear = true, invertible = true)
 parameters:    0.3
 time:          0.0
 state:         [0.0, 0.0]

Here, we had to provide noise_prototype to indicate that the diffusion function g will output a 2x2 matrix.

Scalar noise

If all state variables are forced by the same single random variable, we have scalar noise. To define scalar noise, one has to give an one-dimensional noise process to the noise_process keyword of the CoupledSDEs constructor.

t0 = 0.0; W0 = 0.0;
noise = WienerProcess(t0, W0, 0.0)
sde = CoupledSDEs(f!, rand(2)/10; noise_process=noise)

tr = trajectory(sde, 1.0)
plot_trajectory(tr...)

We can see that noise applied to each variable is the same.

Multiplicative and time-dependent noise

In the SciML ecosystem, multiplicative noise is defined through the condition $g_i(t, u)=a_i u$. However, in the literature the name is more broadly used for any situation where the noise is non-additive and depends on the state $u$, possibly also in a non-linear way. When defining a CoupledSDEs, we can make the noise term time- and state-dependent by specifying an explicit time- or state-dependence in the noise function g, just like we would define f. For example, we can define a system with temporally decreasing multiplicative noise as follows:

function g!(du, u, p, t)
    du .= u ./ (1+t)
    return nothing
end
sde = CoupledSDEs(f!, rand(2)./10; g=g!)

2-dimensional CoupledSDEs
 deterministic: false
 discrete time: false
 in-place:      true
 dynamic rule:  f!
 SDE solver:    SOSRA
 SDE kwargs:    (abstol = 0.01, reltol = 0.01, dt = 0.1)
 Noise type:    (additive = false, autonomous = false, linear = true, invertible = false)
 parameters:    SciMLBase.NullParameters()
 time:          0.0
 state:         [0.07752822361739639, 0.0597275434811622]

Non-diagonal noise

Non-diagonal noise allows for the terms to be linearly mixed (correlated) via g being a matrix. Suppose we have two Wiener processes and two state variables such that the output of g is a 2x2 matrix. Therefore, we have

\[du_1 = f_1(u,p,t)dt + g_{11}(u,p,t)dW_1 + g_{12}(u,p,t)dW_2 \\ du_2 = f_2(u,p,t)dt + g_{21}(u,p,t)dW_1 + g_{22}(u,p,t)dW_2\]

To indicate the structure that g should have, we must use the noise_prototype keyword. Let us define a special type of non-diagonal noise called commutative noise. For this we can utilize the RKMilCommute algorithm which is designed to utilize the structure of commutative noise.

σ = 0.25 # noise strength
function g!(du, u, p, t)
  du[1,1] = σ*u[1]
  du[2,1] = σ*u[2]
  du[1,2] = σ*u[1]
  du[2,2] = σ*u[2]
    return nothing
end
diffeq = (alg = RKMilCommute(), reltol = 1e-3, abstol = 1e-3, dt=0.1)
sde = CoupledSDEs(f!, rand(2)./10; g=g!, noise_prototype = zeros(2, 2), diffeq = diffeq)

2-dimensional CoupledSDEs
 deterministic: false
 discrete time: false
 in-place:      true
 dynamic rule:  f!
 SDE solver:    RKMilCommute
 SDE kwargs:    (reltol = 0.001, abstol = 0.001, dt = 0.1)
 Noise type:    (additive = false, autonomous = true, linear = true, invertible = false)
 parameters:    SciMLBase.NullParameters()
 time:          0.0
 state:         [0.03806281791672672, 0.09681978040051963]