, 94 min read
Stability Mountains for Fredebeul's Formulas
Beware, this three-dimensional graphic needs some time to load. You can rotate the graphic around any axis.
Here we analyze the three methods from Christoph Fredebeul, which were mentioned in Albrecht (1989). These are three cyclic linear multistep methods of order $p=3$ and $p=4$. They should not to be confused with the A-BDF from Fredebeul.
Bibliography.
- Peter Albrecht: “Elements of a General Theory of Composite Integration Methods”, Applied Mathematics and Computation, Vol. 31, May 1989, pp. 1–17
- Christoph Fredebeul: “A-BDF: A Generalization of the Backward Differentiation Formulae”, SIAM Journal on Numerical Analysis, Vol. 35, Issue 5, pp.1917–1938
Order 3. The right one is insofar special as the three implicit formulas can be computed in parallel. This is due to the diagonal structure of the $A_1$ and $B_1$ matrix, see below. That makes this cycle particular attractive for GPUs. However, the predictor will likely give worse estimates for the first iterate in Newton iteration.
Order 4.
The method, written as
has the stability polynomial
For $\mathop{\rm Re} H\to -\infty$ the matrix $B_0$ will dominate.
The methods are all $A[\alpha]$-stable but not $A_\infty^0[\alpha]$-stable. See Tendler-like Formulas for Stiff ODEs for the definitions of all these stability properties.
Widlund-wedge angle and distance are as below. For comparison the Widlund-wedge and distance from the corresponding Tendler formulas.
| Method | α | δ | r(0) | r(∞) | Method | α | δ | r(0) | r(∞) |
|---|---|---|---|---|---|---|---|---|---|
| Fred3 | 88.165° | 0.0282 | 0 | 0.85539679 | Tendler3 | 89.427° | 0.0048 | 0.55371901 | 0 |
| Fred3p | 61.526° | 1.0474 | 0 | 0.94061062 | eTendler3 | 89.724° | 0.0016 | 0.70756795 | 0 |
| Fred4 | 77.345° | 0.3766 | 0 | 0.88440508 | Tendler4 | 80.882° | 0.2441 | 0.35406989 | 0 |
The very high absolute radius at infinity makes the Fredebeul formulas less suitable for stiff system. However, the Fredebeul formulas have two highly sought properties:
- Zero parasitic roots
- After $k=p$ steps the formulas are stable in step-changing!
The comparison with the Tischer formulas:
- All Tischer formulas up to $p\le4$ are A-stable with $r_\infty=0$.
- All parasitic roots are zero.
The error constant is
The order 3 method.
Fredebeul3, p=3, k=3, l=3
12.0000 12.0000 12.0000
0.0000 0.0000 0.0000
-72.0000 -72.0000 -72.0000
60.0000 0.0000 0.0000
0.0000 60.0000 0.0000
0.0000 0.0000 60.0000
-9.0000 -14.0000 -19.0000
-6.0000 4.0000 14.0000
21.0000 16.0000 11.0000
30.0000 60.0000 60.0000
0.0000 30.0000 60.0000
0.0000 0.0000 30.0000
rho_0 0.000000000 0.000000000 0.000000000
rho_1 0.000000000 0.000000000 0.000000000
rho_2 0.000000000 0.000000000 0.000000000
rho_3 0.000000000 0.000000000 0.000000000
rho_4 -0.125000000 -0.333333333 -0.625000000 <-----
The parallel order 3 method.
Fredebeul3p, p=3, k=3, l=3
30.0000 30.0000 30.0000
0.0000 0.0000 0.0000
-90.0000 -90.0000 -90.0000
60.0000 0.0000 0.0000
0.0000 60.0000 0.0000
0.0000 0.0000 60.0000
-3.0000 40.0000 71.0000
-66.0000 -200.0000 -310.0000
51.0000 190.0000 305.0000
18.0000 0.0000 0.0000
0.0000 30.0000 0.0000
0.0000 0.0000 54.0000
rho_0 0.000000000 0.000000000 0.000000000
rho_1 0.000000000 0.000000000 0.000000000
rho_2 0.000000000 0.000000000 0.000000000
rho_3 0.000000000 0.000000000 0.000000000
rho_4 0.075000000 0.666666667 0.375000000 <-----
The order 4 method.
Fredebeul4, p=4, k=3, l=3
300.0000 300.0000 300.0000
0.0000 0.0000 0.0000
-900.0000 -900.0000 -900.0000
600.0000 0.0000 0.0000
0.0000 600.0000 0.0000
0.0000 0.0000 600.0000
-75.0000 0.0000 113.0000
-525.0000 -800.0000 -1132.0000
375.0000 700.0000 923.0000
225.0000 400.0000 543.0000
0.0000 300.0000 408.0000
0.0000 0.0000 345.0000
rho_0 0.000000000 0.000000000 0.000000000
rho_1 0.000000000 0.000000000 0.000000000
rho_2 0.000000000 0.000000000 0.000000000
rho_3 0.000000000 0.000000000 0.000000000
rho_4 0.000000000 0.000000000 0.000000000
rho_5 -0.020833333 -0.172222222 -0.586944444 <-----
1. Stability regions
Using stabregion3 to graph the stability regions.
stabregion3 -f Fredebeul3 -oj -r1800
stabregion3 -f Fredebeul3p -oj -r1800
stabregion3 -f Fredebeul4 -oj -r1800
2. Stability mountain
Below is the output of:
stabregion3 -f Fredebeul3 -o3
Fredebeul3 stability mountain. So visually it is obvious that the method is not $A_\infty^0[\alpha]$-stable.
Output for the "parallel" order 3 method.
stabregion3 -f Fredebeul3p -o3 -r100 -L4:-15:10:25
Again, this method is not $A_\infty^0[\alpha]$-stable.
Output for the order 4 method.
stabregion3 -f Fredebeul4 -o3 -r100 -L5:-20:20:20
This method is not $A_\infty^0[\alpha]$-stable.