Stiff Equations - What to plot as a qualitative or quantitative measure of stiffness

Question

On a recommendation from Mathematica.SE, I am posting this on Computational Science.SE:

I am trying to quantify stiffness of an ODE by relating it to the fine-ness with which NDSolve treats it's time step.

BACKGROUND

For the uninitiated:

Essentially, when an ordinary differential equation is stiff and we attempt to solve it with NDSolve in Mathematica, it would employ a stiff solver to detect and reconcile stiffness.

I use the stiff equation shown on wikipedia to demonstrate this. This equation is solved using BDF (backward difference formulation/formula) which is a recognized method to solve stiff equations.

Equation and solution using NDSolve with BDF

tMax=100;
rSol=
r/.NDSolve[
r'[t]==-15 r[t]&&r[0]==1,r,
{t,0,tMax},Method->{"BDF","MaxDifferenceOrder"->5}][[1]]

Plots of solution r vs t

{Plot[
rSol[t],
{t,0,100},
PlotRange->{{0,tMax},Automatic}
],
LogPlot[
rSol[t],
{t,0,100},
PlotRange->{{0,tMax},Automatic}
]}

Fig 1

Fig 2

Fig 1 shows that up to t=29.9 or so, there is considerable stiffness (unstable oscillations) that is reconciled by ensuring that the time step is sufficiently small.

THE REAL QUESTION:

What would be a good plot to quantify this stiffness and internal reconciliation/changing of time step? How can I best quantify stiffness graphically in Mathematica?

Answer 1

Stiffness has no precise quantitative definition. (See my answer here.)

Common numerical proxies would be things like:

• stiffness ratio: find the (unsigned) ratio of largest magnitude eigenvalue to smallest magnitude eigenvalue of the Jacobian matrix of your right-hand side; estimates are probably fine
• ratio of time step taken to maximum stable time step: a ratio closer to 1 indicates a stability-limited time step, so your problem is probably stiff; a ratio closer to 0 indicates an accuracy-limited time step, and probably a non-stiff problem

There may be other metrics out there; these are the first two that come to mind.