L.S.

Omalanga Study Guide Material (APM-4814)

UNIT 4: DISTINGUISH BETWEEN THE NOTIONS OF

CONVERGENCE AND STABILITY OF PARABOLIC PDEs SOLUTION.

Learning objectives:
After studying this chapter, you will able to distinguish between the notions of
convergence and stability of parabolic PDEs solution.

1. Introduction
It has been noticed that two difficulties usually arise in the numerical solution of
partial differential equations of parabolic type by means of difference equations.
The first is the question of whether the solution of the difference equation
converges to that of the differential equation as the mesh size decreases; Some
researchers have found that certain ratios often must hold between the increments
of the independent variables for convergence to take place [1]. Later, as high-
speed computers were developed, the second problem of stability against error
growth (usually thought of as round-off) occurred. Von Neumann suggested a
criterion for answering this question in some cases, and his criterion explained
that stability occurred when precisely the same ratios as called for by Courant,
Friedrichs, and Lewy held [2].

Leutert’s example [3] had investigated if the two notions of convergence and
stability are equivalent. It was found that there is an unstable difference analogue
of the heat flow equation. This leaves the implication of convergence by stability
open, and it is this question that will be treated in this chapter. It has been found
for wide classes of difference analogues of linear parabolic differential equations
with time-independent coefficients that stability implies convergence in the
mean.

In this chapter, we will start with a brief definition of the notion of stability in the
sense of von Neumann. Then, some examples in the parabolic case will be treated
in order to clarify the assumptions made in the abstraction to follow, the

1|P ag e
convergence proof for the parabolic case then follows, and then the results are
applied to the examples.

2. Stability:
Let us consider for the moment, the numerical solution of a linear partial
differential equation in the region 0 ≤ x ≤ 1, t ≥ 0 for which the solution is
specified on the boundaries x = 0 and x = 1 and sufficient initial conditions (i.e.,
value of function only for parabolic case).

Let (N + 1) ∆x = 1 and ∆t be given, let 𝑥𝑖 = i∆x and 𝑡𝑛 = 𝑛∆𝑡, and denote f (𝑥𝑖 , 𝑡𝑛 )
by 𝑓𝑖𝑛 . Let 𝑤𝑖𝑛 be the solution of the difference analogue of the differential
equation. Further, let 𝑤𝑛 represent the column vector (𝑤1𝑛 , 𝑤2𝑛 , … , 𝑤𝑁𝑛 ). Then,
any linear partial difference equation may be written in the following matrix
form:

(1) (0) (−1) (−𝑞)

𝐴𝑛 𝑤𝑛+1 = 𝐴𝑛 𝑤𝑛 + 𝐴𝑛 𝑤𝑛−1 + ⋯ + 𝐴𝑛 𝑤𝑛−𝑞 + 𝑏𝑛 (n ≥ q), (1)

(1)
where 𝐴𝑛 is a non-singular matrix and 𝑏𝑛 contains the boundary conditions,
(1)
among other things. (𝐴𝑛 is the identity for explicit equations.) For (1) to be
useful, it is necessary that 𝑤0 , 𝑤1, ..., 𝑤𝑞 be known initially. We should assume
that these starting values can be obtained by some procedure, and, moreover, in
the convergence analysis to follow that we know how close they are to the values
of the solution of the differential equation.

We stated in the introduction that we were going to study only the case of time-
independent coefficients for the differential equation; consequently, with fixed
∆𝑥 and ∆𝑡 it is reasonable to limit ourselves to time-independent matrices in (1).
Thus (1) reduces to:

𝐴(1) 𝑤𝑛+1 = 𝐴(0) 𝑤𝑛 + 𝐴(1)𝑤𝑛−1 + ⋯ + 𝐴(−𝑞) 𝑤𝑛−𝑞 + 𝑏𝑛 (𝑛 ≥ 𝑞), (2)

2|P ag e
We are now in a position to perform a stability analysis on (2). Let an error be
∗
introduced by round-off or some other source into the vector 𝑤𝑚 and call it 𝑤𝑚 .
Then, step ahead using (2) and call the perturbed solution 𝑤𝑛∗ . The resulting
equations become:

𝐴(1) 𝑤𝑚+1
∗
= 𝐴(0) 𝑤𝑚
∗
+ 𝐴(−1)𝑤𝑚−1 + ⋯ + 𝐴(−𝑞) 𝑤𝑚−𝑞 + 𝑏𝑚

𝐴(1) 𝑤𝑚+2
∗
= 𝐴(0) 𝑤𝑚+1
∗
+ 𝐴(−1)𝑤𝑚
∗
+ ⋯ + 𝐴(−𝑞)𝑤𝑚−𝑞+1 + 𝑏𝑚

……………………………………………………………………(3)

𝐴(−1)𝑤𝑛+1
∗
= 𝐴(0)𝑤𝑛∗ + 𝐴(−1)𝑤𝑛−1
∗
+ ⋯ + 𝐴(−𝑞)𝑤𝑛−𝑞
∗
+ 𝑏𝑛

Let (n ≥ m+q)

𝑒𝑛 = 𝑤𝑛∗ − 𝑤𝑛 (4)

Then 𝐴(1) 𝑒𝑛+1 = 𝐴(0)𝑒𝑛 + 𝐴(−1)𝑒𝑛−1 + ⋯ + 𝐴(−𝑞) 𝑒𝑛−𝑞 (5)

(n ≥ m + q),

𝑒𝑚 , 𝑒𝑚+1 , … , 𝑒𝑚+𝑞 determined by (3)

Note that (5) is nothing more than the homogeneous equation corresponding to
(2). As 𝐴(𝑖) is time-independent, m may be taken to be zero.

Von Neumann's technique is at this point to apply the method of separation of

variables to (5). For separation of variables to be applicable, it is necessary that
the matrices 𝐴(𝑖) , i = 1, 0, … , -q, possess a common set {𝜌𝑝 }, p = 1 ,…, N, of
eigenfunctions. (We allow the identity matrix to have any complete orthonormal
set as eigenfunctions.) Consequently, we must restrict ourselves to difference
equations whose coefficient matrices satisfy this condition.

3|P ag e
Definition: Equation (2) is stable if and only if max |𝜌𝑝𝑗 | ≤ 1, j =1, …, q+1, p=1,
…, N.

In the sections to follow, we shall consider in detail only the case where the roots
are distinct, as a minute change in the time step is usually sufficient to obtain this
behavior. The results for the remaining cases will be stated.

Note that equations (1) and (2) are unaltered in form when more than one space
variable occur if 𝑤𝑛 has the same number of components as there are lattice points
in the multi-dimensional region. Consequently, the stability definition and
convergence proofs to follow are independent of the number of space variables
involved.

3. Example: Parabolic case: Let us consider the boundary value problem for
the heat flow equation,
𝑢𝑥𝑥 = 𝑢𝑡 , (0 < x < 1, 0 < t ≤ T),
u (x, 0) = f (x)
u (0, t) = 𝑔0 (t),
u (1, t) = 𝑔1 (t), (6)
Numerous difference analogues of (1) have been studied; among these are the
following:
∆𝑡
𝑤𝑖,𝑛+1 = 𝑤𝑖𝑛 + [𝑤𝑖−1,𝑛 − 2 𝑤𝑖𝑛 + 𝑤𝑖+1,𝑛 ] (7) (Forward)
(∆𝑥)2

(∆𝑥)2
−𝑤𝑖−1,𝑛+1 + (2 + )𝑤𝑖,𝑛+1 − 𝑤𝑖+1,𝑛+1
∆𝑡
(∆𝑥)2
= 𝑤𝑖𝑛 (Backward) (8)
∆𝑇
(∆𝑥)2
−𝑤𝑖−1,𝑛+1 + 2(1 + ) 𝑤𝑖,𝑛+1− 𝑤𝑖+1,𝑛+1
∆𝑡
(∆𝑥)2
𝑤𝑖−1,𝑛 − 2(1 − ) 𝑤𝑖,𝑛 − 𝑤𝑖+1,𝑛 (9) (Crank – Nicolson)
∆𝑡
(∆𝑥)2 5(∆𝑥)2 (∆𝑥)2
-(1 - ) 𝑤𝑖−1, 𝑛−1 + (2 + ) 𝑤𝑖,𝑛+1 - (1 - )𝑤𝑖+1, 𝑛+1
6∆𝑡 3∆𝑡 ∆𝑡

4|P ag e
 (∆𝑥)2 5(∆𝑥)2 (∆𝑥)2
= (1 + ) 𝑤𝑖−1, 𝑛 – (2 - ) 𝑤𝑖𝑛 + (1 + ) 𝑤𝑖+1, 𝑛 (10)
6∆𝑡 3∆𝑡 ∆𝑡

(High Order Correct Implicit)

It is to be understood that the initial values and boundary values of 𝑤𝑖𝑛 are to
agree with u (x, t).
In each of the above examples the eigenfunctions in the stability analysis may
be taken to be:
21/2
ρp = sin πpx (p = 1,…, N, (N+1) ∆x=1). (11)
(N+1)1/2

It may be seen readily that the eigenvalues of the A(1) matrix are, respectively,
𝜆𝑃 = 1 (12) (Forward)
(∆𝑥)2 𝜋𝑃
𝜆𝑃 = + 4 𝑠𝑖𝑛2 (13) (Backward)
∆𝑡 2 (𝑁+1)

(∆𝑥)2 𝜋𝑃
𝜆𝑃 = 2 + 4 𝑠𝑖𝑛2 (14) (Crank-Nicolson)
∆𝑡 2 (𝑁+1)

(∆𝑥)2 (∆𝑥)2 𝜋𝑃
𝜆𝑃 = 2 + (4 − 2/3 ) 𝑠𝑖𝑛2 (15) (H.O.C.I.),
∆𝑡 ∆𝑡 2 (𝑁+1)

Moreover, the stability ratios are:

𝜋𝑃
ρ = 1 - r 𝑠𝑖𝑛2 (16) (Forward)
2 (𝑁+1)
1
𝜌= (17) (Backward)
1+𝑟 𝑠𝑖𝑛2 𝜋𝑃/2(𝑁+1)
2−𝑟𝑠𝑖𝑛2 𝜋𝑝
2 (𝑁+1)
ρ= 2+𝑟𝑠𝑖𝑛2 𝜋𝑝
(18) (Crank – Nicolson)
2 (𝑁+1)

2
2−(1+3𝑟)𝑟𝑠𝑖𝑛2 𝜋𝑝/2(𝑁+1)
ρ= 2 (19) ( H.O.C.I)
2+(1− )𝑟𝑠𝑖𝑛2 𝜋𝑝/2(𝑁+1)
3𝑟

Where, r = 4∆t / (∆𝑥)2 (20)

5|P ag e
Thus, (7) is stable for 0 ≤ r ≤ 2, and (7), (8), and (9) for any r > 0. Let us note
one point concerning the eigenvalues 𝜆 . In each case, for any fixed r > 0 there
exists a positive constant K independent of N such that:
|𝜆𝑃 | > K > 0, (21)

i.e., the norm of the inverse of 𝐴(1) is bounded independently of the finesse of
subdivision. It should be assumed this condition to hold for all difference
equations to be treated below. Equation (21) holds for any of the standard
difference approximations to the heat flow equation in two or three space
dimensions as well as those for certain linear equations with variable coefficients.
While the solution u of (6) does not satisfy exactly any of the difference equations
listed for arbitrary boundary conditions, it is easy to see that u satisfies at interior
lattice points a difference equation closely associated with each one. Writing any
of the examples in the form:

𝐴𝑤𝑛+1 = 𝐵𝑤𝑛 + 𝑏𝑛 , (22)

We can observe that:
𝐴𝑢𝑛+1 = 𝐵𝑢𝑛 + 𝑏𝑛 + ℎ𝑛 (23)
Where the components of ℎ𝑛 are remainders in Taylor series developments.
Denote the truncation error by:
𝑣𝑛 = 𝑢𝑛 − 𝑤𝑛 (24)
Then,
𝐴𝑣𝑛+1 = 𝐵𝑣𝑛 + ℎ𝑛 (25)

If the solution of (6) is sufficiently regular in the closed region 0 ≤ x ≤ 1, 0 ≤ t

≤ T, then ℎ𝑛 . will be small. The components for fixed r are O ((∆𝑡)2) for our first
two examples if 𝑢𝑥𝑥𝑥𝑥 exists and is bounded in the region, and they are O((∆t)2)
and O ((∆t)3) respectively for (9) and (10) if sixth derivatives exist

6|P ag e
boundedly:
1 𝜕2 4 1 𝜕2 𝑢
ℎ𝑖𝑛 = − 4
(∆𝑥 )2 ∆𝑡 + (∆𝑡)2 ( 𝐹𝑜𝑟𝑤𝑎𝑟𝑑 ),
12 𝜕𝑥 2 𝜕𝑡 2
1 𝜕4 𝑢 1 𝜕2 𝑢
ℎ𝑖𝑛 = − (∆𝑥 )4 − (∆𝑥 )2 ∆𝑡 (𝐵𝑎𝑘𝑤𝑎𝑟𝑑 ),
12 𝜕𝑥 4 2 𝜕𝑡 2
1 𝜕4 𝑢 1 𝜕4 𝑢 1 𝜕3 𝑢
ℎ𝑖𝑛 = − (∆𝑥 )4 − (∆𝑥 )2(∆𝑡)2 + (∆𝑥 )2(∆𝑡)2 (𝐶𝑟𝑎𝑛𝑘 − 𝑁𝑖𝑐𝑜𝑙𝑠𝑜𝑛)
12 𝜕𝑥 4 8 𝜕𝑥 2 𝜕𝑡 2 12 𝜕𝑡 3
{ ℎ𝑖𝑛 = 0 ((∆𝑡)3 ) ( 𝐻. 𝑂. 𝐶. 𝐼 ) (26)

The restriction that the derivatives appearing in (26) be bounded in the closed
region 0 ≤ x ≤ 1, 0 ≤ t ≤ T, appears somewhat strong; however, in a certain
sense it is not. It is an immediate application of the maximum principle for
parabolic differential equations that the solution of (6) is stable with respect to
small changes in either initial or boundary conditions. It is also easily seen, from
the Green's function representation of the solution of (6), that, if the initial
condition and boundary conditions are approximated by bounded functions in the
L2 sense, then the solution for the approximating data also approximates L2-wise
the solution of the original problem. As the data for any reasonable physical
problem can be approximated in this manner by data leading to a solution of (6)
infinitely' differentiable in the closed region, it is possible in the problems of
greatest interest to alter the original data slightly so as to lead to a problem of the
type to be considered below without seriously altering the result. In fact, the'
altered problem may well represent the physical problem better than the original
mathematical model, as it is rarely possible physically to obtain the
discontinuities or inconsistencies in the data which lead to singularities ill the
solution of the mathematical abstraction. In multi-dimensional problems another
source of singularities arises; this is the shape of the region. Again, the solution
is stable against small alterations in the region.

4. Convergence theorem: Parabolic case. Consider the parabolic boundary

value problem:

7|P ag e
𝜕𝑢 𝜕2 𝑢 𝜕𝑢
= ∑𝑚
𝑖,𝑗=1 𝑎𝑖,𝑗 (𝑥𝑖 , … , 𝑥𝑚 ) + ∑𝑚
𝑖=1 𝑏𝑖 (𝑥𝑖,…, 𝑥𝑚 ) + 𝐶 (𝑥1 , … , 𝑥𝑚 )𝑢 +
𝜕𝑡 𝜕𝑥𝑖𝜕𝑥𝑗 𝜕𝑥𝑖

𝑑 (𝑥𝑖 , … , 𝑥𝑚 ) (( 𝑥1 , … , 𝑥𝑚 ) 𝜖 ℝ, 0 < 𝑡 ≤ 𝑇), 𝑢 ( 𝑥1 , … , 𝑥𝑚,0) =

𝑓 ( 𝑥1, … , 𝑥𝑚 ) (( 𝑥1 , … , 𝑥𝑚 ) 𝜖 ℝ), 𝑢 (𝑥1, … , 𝑥𝑚,𝑡 ) =
𝑔 ( 𝑥1, … , 𝑥𝑚,𝑡 ) ((𝑥1 , … , 𝑥𝑚 ) ϵ S, 0 < t ≤ T), (27)

where R is connected open set in Euclidean m-space and S is its boundary. Let
{R U S} X {0 ≤ t ≤ T} = D. Moreover, let the space increments be functions of
∆t as follows:
∆𝑥𝑖 = ∆𝑥𝑖 (∆t) ( i = 1, …, m). (28)
We shall assume that for each ∆t of some sequence {∆𝑡𝛽 }, ∆𝑡𝛽 → 0, there exists
a translation of the lattice of points (𝛼1∆𝑥1 , … , 𝛼𝑚 ∆𝑥𝑚 ), 𝛼𝑖 = 0, ±1,…, such that
the neighboring lattice points of any lattice point in R are in R U S.
In order that the storage requirements for machine calculation be minimized it
advantageous that the difference equation analogue of (4.1) involve the values of
its solution for t = 𝑡𝑛 and t = 𝑡𝑛+1 only. That this does not necessarily restrict the
order of convergence will result from example (10). This section will be limited
to equations of the form (22). The truncation error equation has been shown (25)
to be:
𝐴𝑢𝑛+1 = 𝐵𝑢𝑛 + ℎ𝑛 (29)
𝑤0 = 𝑢0 is required, 𝑣0 = 0.

Two problems now arise, first, to solve (29), and, second, to demonstrate the
smallness of 𝑣𝑛 in some topology. The solution of (29) will be obtained by
means of a principle of superposition, a metric will be introduced, and two
convergence theorems will be proved.

Theorem 1: If there exists a constant A, independent of ∆t such that ‖ℎ𝑛 ‖ ≤

𝐴1 (∆𝑡) 𝑠+1, m = 0,…, [ T/∆t] the solution 𝑤𝑛 of difference equation (22)

8|P ag e
converges uniformly with respect to n in the topology of the 𝑙2 – space to the
solution of the differential equation (27). Moreover, there exists a constant 𝐵1 ,
independent of ∆t such that ‖𝑢𝑛 − 𝑤𝑛 ‖ ≤ 𝐵1 (∆𝑡)s.
The convergence obtained above is essentially convergence in the mean along
successive horizontal lines, the density of which increases as ∆t decreases.

Theorem 2: If all second partial derivatives of u are bounded in D and ℎ𝑛 =

O ((∆𝑡)𝑠+1), then the multilinearly interpolated solution of (22) converges in
the mean to that of (27) in D. Moreover, the 𝐿2 norm of u (x, t) – w (x, t) is:
O ((∆𝑥1 )2 + … + (∆𝑥𝑚 )2 + (∆𝑡)2+ (∆𝑡)𝑠 ).

5. Application of convergence theorems to parabolic case: Consider first the

forward and backward difference equations given by (22) and (23).
As 𝑢𝑥𝑥 , 𝑢𝑥𝑡 , 𝑢𝑡𝑡 are required to be bounded in order that ‖ℎ𝑛 ‖ be O ((∆t)2),
‖𝑢 − 𝑤‖ 𝐿2 = O (∆t). (30)

if r < 2 for the forward equation and r an arbitrary constant for the backward
equation. The same result holds for the Crank-Nicolson equation when ∆t/(∆x)2
is held constant. However, a better result can be obtained. The smaller the expo-
nent on ∆x in the ratio of ∆t to (∆x)α, the fewer time steps are required to complete
the numerical solution in D; moreover, under the restrictions we have made,
convergence is obtained for any 𝛼 > 0, (see the second line of (33) below). As
the elementary truncation error ℎ𝑛 involves ∆x and ∆t in like manner, it seems
natural to attempt to set:
∆𝑡
= R (31) (a constant).
∆𝑥

Note, though, that 𝜆𝑃 is no longer bounded away from zero as ∆t → 0. Thus,

direct application of theorems 1 and 2 is not possible; however, the method of
proof carries over.
(∆𝑥)2
𝜆𝑃 > 2 ,
(∆𝑡)

9|P ag e
‖ℎ𝑛 ‖ < 𝐴1 (∆𝑥 )2 ((∆𝑥 )2 + (∆𝑡)2 )), (32)
It is readily shown that:
𝑇 ∆𝑡
‖𝑣𝑛 ‖ ≤ 𝐴, (∆𝑥 )2 ((∆𝑥 )2 + (∆𝑡)2)
∆𝑡 2(∆𝑥)2

= 𝐵1 ((∆𝑥)2 + (∆𝑡)2)
= 𝐶1 (∆𝑡)2, (33)

if (31) holds. The argument for ‖𝑣‖ 𝑙2 follows immediately.

MORE DEFINITIONS:
1. Consistency: For a method to be consistent, the truncation error must become
zero when the mesh spacing ∆t → 0 and/or ∆x →0.
2. Stability: Stability means the errors at any stage of the computation are
attenuated, not amplified as the computation progresses.
Furthermore, a numerical solution method is said to be stable if it does not magni-
fy the errors that appear in the course of numerical solution process. The most
widely used approach to studying stability of numerical schemes is the Von
Neumann’s method.
The von Neumann condition is a sufficient condition for the l2 stability of any
scheme to which the Fourier analysis can be applied, if the scheme is a one-step
scheme for a scalar equation so that G is scalar. It is also sufficient if G is a normal
matrix so that its subordinate norms can be bounded by its spectral radius.

3. Convergence: Convergence means that as ∆x and ∆t approach zero the results

of the numerical technique approach the true result.
Furthermore, a numerical method is said to be convergent if the solution of
the discretized equations tend to the exact solution of the differential equation
as the grid spacing tends to zero. If the method is stable and if all approximations
used in the discretization process are consistent, we usually find that the solution
does converge to a grid-independent solution.
10 | P a g e
Supplementary problem:
1. Consider the classical solution of the heat equation:
𝜕𝑇 𝜕2 𝑇
=𝐾
𝜕𝑡 𝜕𝑥 2

Analyse the stability of the finite difference method schemes to heat equation.

Recommended readings
1. Courant, R., Friedrichs, K., Lewy, H. (1928). Uber die partiellen
differenzengleichungen der mathematischen Physik, Math. Ann., 100 (1928), 32-
74.
2. O’Brien, G.G., Hyman, M.A., Kaplan, S. (1951). A study of numerical solution
of partial differential equations, J. Math. And Physics, 29, 223-251.
3. Leutert, W. (1951). On the convergence of approximate solutions of heat
equation to the exact solution. Proc. Amer. Math. Soc., 2, 433-439.
4. Morton, K.W., Mayers, D. (2005), Numerical solution of partial differential
equations. Cambridge University Press.

5. Buchanan, M.L. (1963), A necessary and sufficient condition for stability of

difference schemes for second order initial value problems, J. Soc. Indust. Appl.
Math. 11, 474–501.

6. Buchanan, M.L. (1963), A necessary and sufficient condition for stability of

difference schemes for initial value problems, J. Soc. Indust. Appl. Math. 11, 919–
35.

11 | P a g e