Chapter 7

T. H. Pulliam
NASA Ames

1
 Stability of Linear Systems

• Stability will be defined in terms of ODE’s and O∆E ’s

– ODE: Couples System

d~u
= A ~u − ~f (t) (1)
dt
– O∆E : Matrix form from applying Eq. 1
~un+1 = C ~un − ~g n (2)

– For Example: Euler Explicit C = [I − hA]

2
 Inherent Stability of ODE’s

• Stability of Eq. 1 depends entirely on the eigensystem of A.

• λm -spectrum of A: function of finite-difference scheme,BC

For a stationary matrix A, Eq. 1 is inherently stable if,

(3)
when ~f is constant, ~u remains bounded as t → ∞.

• Note that inherent stability depends only on the transient

solution of the ODE’s.
λ1 h n ~ λm h n ~



~u(t) = c1 e x1 + · · · + cm e xm + · · ·
λM h n ~


+ cM e xM + P.S. (4)

3
• ODE’s are inherently stable if and only if

<(λm ) ≤ 0 for all m (5)

• For inherent stability, all of the λ eigenvalues must lie on, or to

the left of, the imaginary axis in the complex λ plane.
• This criterion is satisfied for the model ODE’s representing both
diffusion and biconvection.

4
 Numerical Stability of O∆E ’s

• Stability of Eq. 2 related to the eigensystem of its matrix, C.

• σm -spectrum of C: determined by the O∆E and are a function
of λm
n n
~un = c1 (σ1 ) ~x1 + · · · + cm (σm ) ~xm + · · ·
n
+ cM (σM ) ~xM + P.S. (6)

• Spurious roots play a similar role in stability.

• The O∆E companion to Statement 3 is

For a stationary matrix C, Eq. 2 is numerically stable

(7)
if, when ~g is constant, ~un remains bounded as n → ∞.

5
• Definition of stability: referred to as asymptotic or time stability.
• Time-marching method is numerically stable if and only if

|(σm )k | ≤ 1 for all m and k (8)

• This condition states that, for numerical stability, all of the σ

eigenvalues (both principal and spurious, if there are any) must
lie on or inside the unit circle in the complex σ-plane.
• This definition of stability for O∆E’s is consistent with the
stability definition for ODE’s.

6
 Review

• Our Approach leads to

– The PDE’s are converted to ODE’s by approximating the
space derivatives on a finite mesh.
– Inherent stability of the ODE’s is established by guaranteeing
that <(λ) ≤ 0.
– Time-march methods are developed which guarantee that
|σ(λh)| ≤ 1 and this is taken to be the condition for numerical
stability.

7
Time-Space Stability and Convergence of O∆E’s

• A more classical view (but consistent) in the time-space sense.

– The homogeneous part of Eq. 2, ~un+1 = C~un
– Applying simple recursion ~un = C n ~u0
– Using vector and matrix p-norms
n n
||~un || = ||C ~u0 || ≤ ||C || · ||~u0 || ≤ ||C||n · ||~u0 || (9)

– Assume that the initial data vector is bounded, the solution

vector is bounded if

||C|| ≤ 1 (10)

where ||C|| represents any p-norm of C.

8
 – This is often used as a sufficient condition for stability.
– Well known relation between spectral radii and matrix norms
∗ The spectral radius of a matrix is its L2 norm when the
matrix is normal, i.e., it commutes with its transpose.
∗ The spectral radius is the lower bound of all norms.
• The matrix norm approach and the σ − λ analysis are consistent
when both A and C have a complete eigensystem.

9
Numerical Stability Concepts: Complex σ-Plane

• σ-Root Traces Relative to the Unit Circle

• The O∆E solution to the homogeneous part

~un = c1 σ1n ~x1 + · · · + cm σm

n n
~xm + · · · + cM σM ~xM

• Semi-discrete approach leads to a relation between the σ and the

λ eigenvalues.
• Numerical stability of the O∆E requires that σ-roots lie within
unit circle in the complex σ-plane.
• Trace the locus of the σ-roots as a function of the parameter λh

10
 Stability in the Complex-σ Plane

• Define σexact = eλh and Separate:

– Dissipation (λh = β < 0) —— Convection (λh = iω)
• Plot Real(σ) and Imag(σ) for varying λh of both types.

Ι(σ) Ι(σ)

λ h= 0 ωh= 0

R (σ) R (σ)
λ h= οο

i ωh
σ = e λ h , λh - oo σ=e , ωh oo

a) Dissipation b) Convection

Figure 1: Exact traces of σ-roots for model equations.

11
 σ − λ Relations for Various Schemes

1. σ − 1 − λh = 0 Explicit Euler
2. σ 2 − 2λ h σ − 1 = 0 Leapfrog
3. σ 2 − (1 + 3 λh)σ + 1 λh = 0 AB2
2 2
4. 3
σ − (1 + 23 2
λh)σ + 16 λhσ − 5 λ h = 0 AB3
12 12 12
5. σ ( 1 − λ h) − 1 = 0 Implicit Euler
6. σ (1 − 1 λ h ) − (1 + 1 λh ) = 0 Trapezoidal
2 2
7. σ 2 (1 − 2 λ h ) − 4 σ + 1 = 0 2nd-Order
3 3 3
Backward
8. σ 2 (1 − 5 λ h ) − (1 + 8 λh)σ + 1 λ h = 0 AM3
12 12 12
9. σ 2 − (1 + 13 λh + 15 λ2 h2 )σ + 1 λh(1 + 5 λh) = 0 ABM3
12 24 12 2
10. σ 3 − (1 + 2λ h )σ 2 + 3 λhσ − 1 λ h = 0 Gazdag
2 2
11. σ − 1 − λ h − 1 λ2 h2 = 0 RK2
2
12. σ − 1 − λ h − λ2 h2 − 1 λ3 h3 − 1 λ4 h4 = 0
1 RK4
2 6 24
13. σ 2 (1 − 1 λ h ) − 4 λ hσ − (1 + 1 λ h ) = 0 Milne 4th
3 3 3

Table 7.1. Some λ − σ Relations

12
Traces of σ-roots for various methods.

σ1

σ1

λ h=-2
a) Euler Explicit

ωh = 1

σ2 σ1
σ2 σ1

b) Leapfrog

σ1
σ2 σ1 σ2

λ h=-1
c) AB2
Diffusion Convection

13
Traces of σ-roots for various methods.

σ1 σ1

λ h=- οο d) Trapezoidal ω h = 2/3

σ2

σ3 σ1
σ1
σ2 σ3
e) Gazdag

σ1 σ1

λ h=-2 f) RK2

σ1 σ1

g) RK4
λ h=-2.8 ωh= 2 2

Diffusion Convection

14
 Types of Stability

• Conditional Stability: Explicit Methods

– O∆E ’s where λh ≤ Constant
ah
– λ spectrum, e.g. λb h = − ∆x (1 − cos(k∆x) + isin(k∆x))
– Given ∆x, wave speed a, and difference scheme: λ fixed
– Adjust h = ∆t to satisfy stability bound
– Time accuracy: use an appropriate h
– Mildly-unstable: Prof. Milton VanDyke
Lock bike fork and peddle as fast as you can, you may cross
the street before you fall over and a truck hits you.

15
• Un-Conditional Stability: Implicit Methods

A numerical method is unconditionally stable if it is

stable for all ODE’s that are inherently stable.

– O∆E ’s where λh → ∞ is stable

– Time accuracy: use an appropriate h
– Steady-State: any h which converges fast.
– Computationally expensive compared with Explicit Methods

16
 Stability Contours in the Complex λh Plane.

• Another view of stability properties of a time-marching method

is to plot the locus of the complex λh for which |σ| = 1
• |σ| refers to the maximum absolute value of any σ, principal or
spurious, that is a root to the characteristic polynomial for a
given λh.
• Inherently stable ODE’s lies in the left half complex-sigma plane

17
 Example for Euler Explicit

• Euler explicit: σee = 1 + hλ

– Wave equation: central differencing, λc = −ai sin(k∆x)
∆x

ah
σee =1− isin(k∆x)
∆x
– |σee | > 1.0 for all h, unconditionally unstable
• Wave equation: 1st order backward differencing,
ah
λb h = − ∆x (1 − cos(k∆x) + isin(k∆x))
– |σee | ≤ 1.0 for all some h, conditionally stable
ah
– Note: CF L = ∆x , CFL Number

18
• Complex λ-plane, Euler explcit, σee = 1 + λh
– Let λh = x + iy, then σee = 1 + x + iy
p
|σee | = (1 + x)2 + y 2

– Contour of |σee | = 0.8 leads to (1 + x)2 + y 2 = (0.8)2 : circle in

x, y centered at x = −1 with radius 0.8 , Stable
– Contour of |σee | = 1.2 leads to (1 + x)2 + y 2 = (1.2)2 ’: circle
in x, y centered at x = −1 with radius 1.2, Un-Stable

19
 Example for Euler Implicit

1
• Euler implicit: σei = 1−λh

– Wave equation: central differencing, λc = −ai sin(k∆x)

∆x

1
σei = ah
1 + ∆x isin(k∆x)

– |σei | < 1.0 for all h, unconditionally stable

– Even for Compex λh: unconditional stability

20
 1
• Complex λ-plane, Euler Implcit, σei = 1−λh
1
– Let λh = x + iy, then σei = 1−x−iy

1
|σei | = p
(1 − x)2 + y 2
1 2
– Contour of |σei | = 0.8 leads to (1 − x)2 + y 2 = ( 0.8 ) : circle in
1
x, y centered at x = 1 with radius 0.8 , Stable
1 2
– Contour of |σei | = 1.2 leads to (1 − x)2 + y 2 = ( 1.2 ) : circle in
1
x, y centered at x = 1 with radius 1.2 < 1.0, Un-Stable
– The unstable contours are in the right half of the inherent
stable of the ODE’s

21
 I( λ h) I(λ h) I(λ h)

Stable

Unstable

1 1
Stable Unstable

Stable
R(λ h) R(λ h) Unstable R( λ h)

a) Euler Explicit θ=0 b) Trapezoid Implicit θ = 1/2 c) Euler Implicit θ=1

Figure 2: Stability contours for the θ-method.

22
Stability contours for some explicit methods.

23
Stability contours for Runge-Kutta methods.

I( λh)

3.0
Stable RK4
Regions
RK3
2.0

RK2

1.0

RK1
R(λ h)

-3.0 -2.0 -1.0

24
 Fourier Stability Analysis

• Classical stability analysis for numerical schemes

• Fourier or von Neumann approach.
– Periodic in space derivative, similar to modified wave number
– Usually carried out on point operators
– Does not depend on an intermediate stage of ODE’s.
• Strictly speaking it applies only to difference approximations of
PDE’s that produce O∆E’s
• Serves as a fairly reliable necessary stability condition, but it is
by no means a sufficient one.

25
 The Basic Procedure

• Impose a spatial harmonic as an initial value on the mesh

• Will its amplitude grow or decay in time?
• Determined by finding the conditions under which

u(x, t) = eαt · eiκx (11)

• Is a solution to the difference equation, where κ is real and κ∆x

lies in the range 0 ≤ κ∆x ≤ π.
• For the general term,
(n+`) (n)
uj+m = eα(t+`∆t) · eiκ(x+m∆x) = eα`∆t · eiκm∆x · uj

(n)
• uj is common to every term and can be factored out.

26
• Find the term eα∆t , which we represent by σ, thus:

σ ≡ eα∆t

α∆t n
αt


• Since e = e = σn

For numerical stability |σ| ≤ 1 (12)

• Solve for the σ’s produced by any given method

• A necessary condition for stability, make sure that, in the worst
possible combination of parameters, condition 12 is satisfied.

27
 Example 1

• Finite-difference approximation to the model diffusion equation

• Richardson’s method of overlapping steps.
(n+1) (n−1) 2∆t  (n) (n) (n)

uj = uj +ν u − 2uj + uj−1 (13)
∆x2 j+1
– Substitution of Eq. 11 into Eq. 13
−1 2∆t iκ∆x −iκ∆x


σ =σ +ν 2
e −2+e
∆x
or
 
2 4ν∆t
σ + (1 − cos κ∆x) σ − 1 = 0 (14)
∆x2
| {z }
2b
– Eq. 11 is a solution of Eq. 13 if σ is a root of Eq. 14.

28
– The two roots of Eq. 14 are
p
σ1,2 = −b ± b2 + 1

– One |σ| is always > 1.

– Therefore, that by the Fourier stability test, Richardson’s
method of overlapping steps is unstable for all ν, κ and ∆t.

29
 Example 2

• Finite-difference approximation for the model biconvection

equation
(n+1) (n) a∆t  (n) (n)

uj = uj − u − uj−1 (15)
2∆x j+1
a∆t
σ =1− · i · sin κ∆x
∆x
• |σ| > 1 for all nonzero a and κ.
• Thus we have another finite-difference approximation that, by
the Fourier stability test, is unstable for any choice of the free
parameters.

30