Chapter 11
MATH 6371
Annalisa Quaini
[email protected]
Office : PGH 662
Lecture : TuTh 2:30 PM - 4:00 PM in S 101
Office hours : Th 1:00 PM - 2:30 PM or BY APPOINTMENT
A. Quaini, UH MATH 6371 1/6
Analysis of One-Step Methods
Section 11.3
c 2021, A. Quaini, UH MATH 6371 2/1
Analysis of one-step methods
A generic explicit method for (CP) can be written as
Definition
Φ is called increment function and it fully characterizes the
explicit method.
Ex: for (FE), Φ(tn , un , fn ; h) = fn .
c 2021, A. Quaini, UH MATH 6371 3/1
Analysis of one-step methods
Now, force the exact solution in the numerical scheme:
Let’s write the residual
n+1 = hτn+1 (h)
and set
τ (h) = max |τn+1 (h)|.
0≤n≤Nh −1
Definition
τn+1 (h) is called local truncation error. τ (h) is called global
truncation error.
c 2021, A. Quaini, UH MATH 6371 4/1
Analysis of one-step methods
Definition
A method is said to be consistent if the global truncation errors is
s.t.:
lim τ (h) = 0.
h→0
Moreover, a scheme has order p if ∀ ∈ I , the solution y (t) of (CP)
is s.t.:
τ (h) = O(hp ) as h → 0.
If
lim Φ(tn , yn , f (tn , yn ); h) = f (tn , yn ) ∀tn ≥ t0 ,
h→0
it is easy to show that the method is consistent.
c 2021, A. Quaini, UH MATH 6371 5/1
The Zero-Stability
c 2021, A. Quaini, UH MATH 6371 6/1
The Zero-Stability
Both C and h0 may depend on the problem’s data (t0 , T , y0 , f ).
The zero-stability deals with the behavior of (EM) in the limit case
as h → 0 (thus the name). It is a property of the numerical
method, not of the (CP). For the (CP) we had Liapunov stability.
c 2021, A. Quaini, UH MATH 6371 7/1
Convergence Analysis
Definition
A method is said to be convergent if
|un − yn | ≤ C (h) ∀n = 0, . . . , Nh
where C (h) is an infinitesimal wrt h. It is said to be convergent
with order p if ∃const > 0 s.t. C (h) = (const)hp .
c 2021, A. Quaini, UH MATH 6371 8/1
Convergence Analysis
This theorem is called Lax-Richtmeyer.
c 2021, A. Quaini, UH MATH 6371 9/1
Convergence Analysis
The Lax-Richtmeyer theorem states a fundamental result:
CONSISTENCY + ZERO STABILITY ⇒ CONVERGENCE
The converse CONVERGENCE ⇒ ZERO STABILITY is obviously
true.
Now, let’s see the convergence analysis for the Forward Euler
method.
c 2021, A. Quaini, UH MATH 6371 10 / 1
Convergence analysis for the Forward Euler method
We have
yn+1 − yn
= f (tn , yn ) + τn+1 n = 0, . . . , Nh − 1.
h
Set en = un − yn . By subtracting the above equation from the
Forward Euler method, we get:
en+1 − en
= f (tn , un ) − f (tn , yn ) − τn+1
h
which means
en+1 = en + h[f (tn , un ) − f (tn , yn )] − hτn+1 .
c 2021, A. Quaini, UH MATH 6371 11 / 1
Convergence analysis for the Forward Euler method
Starting from this equation and for the Lipschitz continuity of f ,
we have:
c 2021, A. Quaini, UH MATH 6371 12 / 1
Convergence analysis for the Forward Euler method
Since e0 = u0 − y0 = 0, we can write
c 2021, A. Quaini, UH MATH 6371 13 / 1
Convergence analysis for the Forward Euler method
c 2021, A. Quaini, UH MATH 6371 14 / 1
Convergence analysis for the Forward Euler method
c 2021, A. Quaini, UH MATH 6371 15 / 1
Convergence analysis for the Forward Euler method
It is possible to show that for the (FE) method:
h 00
τn (h) = y (ξ)
2
with ξ ∈ (tn−1 , tn ). From this, we see that the (FE) method is of
first order. Moreover,
M
τ (h) < h with M = max |y 00 (ξ)|
2 ξ∈I
and
e LT − 1 M
|en | ≤ h.
L 2
c 2021, A. Quaini, UH MATH 6371 16 / 1
The (absolute) stability properties
(Chapt. 7.6 in the book)
For given λ < 0, we consider the model problem:
(
y ′ (t) = λy(t) for t ∈ R+
(21)
y(0) = 1
The solution is
y(t) = eλt . In particular, lim y(t) = 0.
t→∞
Let 0 = t0 < t1 < . . . < tn < tn+1 < . . . such that tn = nh and where the time
step h > 0 is fixed.
We say that a numerical scheme associated to the model problem is
absolutely stable if limn→∞ un = 0.
Ordinary differential equations – p. 26/85
Chapter 11
Absolute stability
Generalization of the test problem:
c 2021, A. Quaini, UH MATH 6371 17 / 31
Chapter 11
Absolute stability
The numerical solution un of (TP) depends on h (discretization
step). Thus, a method will be absolutely stable for certain values
of h and λ and (maybe) not for others.
Definition
The region of stability of a numerical method is
A = {z = hλ ∈ C : |un | → 0 as tn → ∞}.
Definition
A method is called A-stable if A ∩ C− = C− .
Let’s check the absolute stability of one-step methods.
c 2021, A. Quaini, UH MATH 6371 18 / 31
• For the forward Euler:
un+1 = (1 + λh)un , where un = (1 + λh)n , ∀n ≥ 0. (22)
If 1 + λh < −1, then |un | → ∞ when n → ∞, therefore forward Euler is
unstable.
To ensure stability, we need to limit the time step h, by imposing the stability
condition :
|1 + λh| < 1 hence h < 2/|λ|.
• For the backward Euler:
n
1 1
un+1 = un and therefore un = , ∀n ≥ 0.
1 − λh 1 − λh
Because limn→∞ un = 0, it is unconditionally stable (it is stable for any h > 0).
Ordinary differential equations – p. 27/85
Chapter 11
Absolute stability for (FE)
c 2021, A. Quaini, UH MATH 6371 19 / 31
Chapter 11
Absolute stability for (FE)
c 2021, A. Quaini, UH MATH 6371 20 / 31
Chapter 11
Absolute stability for (FE)
c 2021, A. Quaini, UH MATH 6371 21 / 31
Chapter 11
Absolute stability for (FE)
c 2021, A. Quaini, UH MATH 6371 22 / 31
• For the forward Euler:
un+1 = (1 + λh)un , where un = (1 + λh)n , ∀n ≥ 0. (22)
If 1 + λh < −1, then |un | → ∞ when n → ∞, therefore forward Euler is
unstable.
To ensure stability, we need to limit the time step h, by imposing the stability
condition :
|1 + λh| < 1 hence h < 2/|λ|.
• For the backward Euler:
n
1 1
un+1 = un and therefore un = , ∀n ≥ 0.
1 − λh 1 − λh
Because limn→∞ un = 0, it is unconditionally stable (it is stable for any h > 0).
Ordinary differential equations – p. 27/85
Chapter 11
Absolute stability for (BE)
c 2021, A. Quaini, UH MATH 6371 23 / 31
Chapter 11
Absolute stability for (BE)
c 2021, A. Quaini, UH MATH 6371 24 / 31
Chapter 11
Absolute stability for (BE)
c 2021, A. Quaini, UH MATH 6371 25 / 31
Example 4. Let solve the problem (21) for λ = −2 and y0 = 1 at interval
[0, 10] using forward and backward Euler methods with h = 0.9 and h = 1.1.
Here are the Matlab/Octave commands for the case h = 0.9. Note that, even
if f (t, y) does not depend on t, it must be defined in Matlab/Octave as a
function of (t, y).
>> f = @(t,x) -2*x; h=0.9; tspan=[0 10]; Nh = 10/h; y0=1;
>> [t_ep, y_ep] = feuler(f, tspan, y0, Nh);
>> [t_er, y_er] = beuler(f, tspan, y0, Nh);
>> t = linspace(0, 10, 11); sol_ex = @(t) exp(-2*t);
>> plot(t, sol_ex(t), ’b’, t_ep, y_ep, ’ro-’, t_er, y_er’, ’go-’)
The following figure shows obtained solutions for h = 0.9 (on the left) and
h = 1.1 (on the right) and the exact solution.
Ordinary differential equations – p. 28/85
y ’ (t) = −2 y(t)
y ’ (t) = −2 y(t) 6
1
sol−ex sol−ex
Euler prog. Euler prog.
0.8 4 Euler retr
Euler retr
0.6
2
0.4
0.2
y(t)
0
y(t)
−2
−0.2
−0.4
−4
−0.6
−0.8 −6
0 2 4 6 8 10 0 2 4 6 8 10
t t
Comparison of solutions that we obtain by the forward and backward Euler
methods for h = 0.9 (on the left, stable) and h = 1.1 (on the right, unstable)
(stability condition for forward Euler: |λ| = 2 ⇒ h < 2/|λ| = 1).
Ordinary differential equations – p. 29/85
Chapter 11
Absolute stability for (CN)
c 2021, A. Quaini, UH MATH 6371 26 / 31
Chapter 11
Absolute stability for (CN)
c 2021, A. Quaini, UH MATH 6371 27 / 31
Chapter 11
Absolute stability for Heun
c 2021, A. Quaini, UH MATH 6371 28 / 31
Chapter 11
Absolute stability for Heun
c 2021, A. Quaini, UH MATH 6371 29 / 31
Chapter 11
Absolute stability for Heun
c 2021, A. Quaini, UH MATH 6371 30 / 31
Chapter 11
Summarizing
(FE) is conditionally stable in C− ← explicit
(BE) is absolutely stable in C− ← implicit
(CN) is absolutely stable in C− ← implicit
Heun is conditionally stable in C− ← explicit
Note: This is not a general rule, that is in general it is not said
that implicit methods are unconditionally absolutely stable. There
exists implicit unstable or conditionally stable schemes. On the
other hand, there are no explicit unconditionally absolutely stable
methods.
c 2021, A. Quaini, UH MATH 6371 31 / 31
