Download as pdf or txt
Download as pdf or txt
You are on page 1of 34

Course Code: MA532; Title: Numerical Solution of Ordinary Differential Equations; Units: 4

Course Outline: Ordinary Differential Equations. Higher Order one –step methods; Taylor and Runge –Kutta

methods, Linear Multistep steps; Adams, Backward Differentiation Formula (BDF), Predictor –Corrector methods.

Convergence and Stability Analyses. Stiff systems and rational approximations. Continuous methods and

boundary value –problems (bvp), Fredholm and Voltera integral equations.

1. LINEAR MULTISTEP METHODS

The general Linear Multistep Method (LMM)

Consider the initial-value problem (ivp) for single first order differential equation

𝑦 ′ = 𝑓(𝑥, 𝑦), 𝑦(𝑎) = 𝑦0 (1.1)

We seek a solution in the range 𝑎 ≤ 𝑥 ≤ 𝑏, where 𝑎 and 𝑏 are finite and we assume that 𝑓 satisfies the conditions

stated in Theorem 1.1 below which guarantee that the problem has a unique continuously differentiable solution,

which we shall indicate by 𝑦(𝑥).

Theorem 1.1 (Existence and Uniqueness)

Let 𝑓(𝑥, 𝑦) be defined and continuous for all points (𝑥, 𝑦) in region 𝐷 defined by 𝑎 ≤ 𝑥 ≤ 𝑏, −∞ < 𝑦 < ∞, 𝑎 and

𝑏 finite, and let there exist a constant 𝐿 such that, for every 𝑥, 𝑦, 𝑦 ∗ such that (𝑥, 𝑦) and (𝑥, 𝑦 ∗ ) are both in 𝐷,

|𝑓(𝑥, 𝑦) − 𝑓(𝑥, 𝑦 ∗ )| ≤ 𝐿|𝑦 − 𝑦 ∗ |

Then if 𝑦0 is any given number, there exist a unique solution 𝑦(𝑥) of the ivp (1.1), where 𝑦(𝑥) is continuous and

differentiable for all (𝑥, 𝑦) in 𝐷.

Consider the sequence of points {𝑥𝑛 } defined by 𝑥𝑛 = 𝑎 + 𝑛ℎ, 𝑛 = 0,1,2, …, where the parameter is called the

step length or step size. An essential property of the majority of computational methods for the solution of (1.1) is

that of discretization; that is, we seek an approximate solution, not on the continuous interval 𝑎 ≤ 𝑥 ≤ 𝑏, but on

the discrete point set {𝑥𝑛 |𝑛 = 0,1, … , (𝑏 − 𝑎)⁄ℎ}. Let 𝑦𝑛 be an approximation to the theoretical solution at 𝑥𝑛 ,

that is, to 𝑦(𝑥𝑛 ), and let 𝑓𝑛 ≡ 𝑓(𝑥𝑛 , 𝑦𝑛 ). If a computational method for determining the sequence {𝑦𝑛 } takes the

form of a linear relationship between𝑦𝑛+𝑗 , 𝑓𝑛+𝑗 , 𝑗 = 0,1, … , 𝑘, we call it a linear multistep method of step number

𝑘 or a linear 𝑘 step method.


1
We wish to find the numerical solution of (1.1) on a given mesh/grid

𝑎 = 𝑥0 < 𝑥1 < 𝑥2 < ⋯ < 𝑥𝑛 < 𝑥𝑛+1 < ⋯ < 𝑥𝑁 = 𝑏

where ℎ = 𝑥𝑛+1 − 𝑥𝑛 is a constant step size and 𝑁 = (𝑏 − 𝑎)⁄ℎ.

The general form of Linear Multistep Method (LMM)

A 𝑘 −step LMM for (1.1) is defined by a difference equation

𝛼𝑘 𝑦𝑛+𝑘 + 𝛼𝑘−1 𝑦𝑛+𝑘−1 + ⋯ + 𝛼1 𝑦𝑛+1 + 𝛼0 𝑦𝑛 = ℎ[𝛽𝑘 𝑓𝑛+𝑘 + 𝛽𝑘−1 𝑓𝑛+𝑘−1 + ⋯ + 𝛽1 𝑓𝑛+1 + 𝛽0 𝑓𝑛 ]

or

𝑘 𝑘

∑ 𝛼𝑗 𝑦𝑛+𝑗 = ℎ ∑ 𝛽𝑗 𝑓𝑛+𝑗 (1.2)


𝑗=0 𝑗=0

Where 𝛼𝑗 and 𝛽𝑗 are constants, called the coefficients of the method. We assume that 𝛼𝑘 ≠ 0 and that not both

𝛼0 and 𝛽0 are zero, that is, 𝛼02 + 𝛽02 > 0. Since (1.2) can be multiplied on both sides by the same constant without

altering the relationship, the coefficients 𝛼𝑗 and 𝛽𝑗 are arbitrary to the extent of a constant multiplier. We remove

this arbitrariness by assuming throughout that 𝛼𝑘 = 1. Some examples of LMM are the mid –point method,

Simpson’s method, Adams methods and Backward Differentiation Formulae (BDF) methods.

The problem of determining the solution 𝑦(𝑥) of (1.1) is replaced by that of finding a sequence {𝑦𝑛 } which

satisfies the difference equation (1.2). The 𝑦𝑛 values which are intended to be estimates of 𝑦(𝑎 + 𝑛ℎ) can be

computed recursively from (1.2) if in addition to the initial value 𝑦0 = 𝑦(𝑎), 𝑘 − 1 more values, that is,

𝑦1 , 𝑦2 , … , 𝑦𝑘−1 are given.

2
We say that the method (1.2) is explicit if 𝛽𝑘 = 0, and implicit if 𝛽𝑘 ≠ 0. For an explicit method, (1.2) yields the

current value 𝑦𝑛+𝑘 directly in terms of 𝑦𝑛+𝑗 , 𝑓𝑛+𝑗 , 𝑗 = 0,1, … , 𝑘, which at this stage of computation have been

calculated. An implicit method, however, will call for the solution at each stage of computation, of the equation

𝑦𝑛+𝑘 = ℎ𝛽𝑘 𝑓(𝑥𝑛+𝑘 , 𝑦𝑛+𝑘 ) + 𝑔 where 𝑔 is a known function of the previously calculated values 𝑦𝑛+𝑗 , 𝑓𝑛+𝑗 , 𝑗 =

0,1, … , 𝑘 − 1. Thus implicit methods in general entail a substantially greater computational effort than do explicit

methods. On the other hand, for a given step number 𝑘, implicit methods can be made more accurate than

explicit ones and moreover, enjoy more favourable stability properties.

Example 1.1: Consider the method

𝛼2 𝑦𝑛+2 + 𝛼1 𝑦𝑛+1 + 𝛼0 𝑦𝑛 = ℎ[𝛽2 𝑓𝑛+2 + 𝛽1 𝑓𝑛+1 + 𝛽0 𝑓𝑛 ].

If 𝛽2 ≠ 0, the above method is implicit, however, if 𝛽2 = 0, then the method is explicit

𝛼2 𝑦𝑛+2 + 𝛼1 𝑦𝑛+1 + 𝛼0 𝑦𝑛 = ℎ[𝛽1 𝑓𝑛+1 + 𝛽0 𝑓𝑛 ]

We now consider the problem of determining the coefficients 𝛼𝑗 , 𝛽𝑗 appearing in (1.2). Any specific LMM may be

derived in a number of different ways and we shall consider a selection of different approaches.

1.1: Derivation through Taylor Expansion

Consider the Taylor expansion for 𝑦(𝑥𝑛 + ℎ) about 𝑥𝑛 ;

2 𝑞
𝑦(𝑥𝑛 + ℎ) = 𝑦(𝑥𝑛 ) + ℎ𝑦 ′ (𝑥𝑛 ) + ℎ2! 𝑦 ′′ (𝑥𝑛 ) + ⋯ + ℎ𝑞! 𝑦 (𝑞) (𝑥𝑛 ) + ⋯

where

𝑑 𝑞 𝑦(𝑥𝑛 )
𝑦 (𝑞) (𝑥𝑛 ) = 𝑑𝑥 𝑞
,𝑞 = 1,2, …

If we truncate this expansion after two terms and substitute for 𝑦 ′ (𝑥) from the differential equation (1.1), we

have

𝑦(𝑥𝑛 + ℎ) ≅ 𝑦(𝑥𝑛 ) + ℎ𝑓(𝑥𝑛 , 𝑦(𝑥𝑛 )) (1.3)

a relation that is in error by

ℎ2 3
2!
𝑦 ′′ (𝑥𝑛 ) + ℎ3! 𝑦 ′′′ (𝑥𝑛 ) + ⋯ (1.4)

Equation (1.3) expresses an approximate relation between exact values of the solution of (1.1). We can also

interpret it as an exact relation between approximate values of the solution of (1.1) if we replace 𝑦(𝑥𝑛 ), 𝑦(𝑥𝑛 +

ℎ) by 𝑦𝑛 , 𝑦𝑛+1 respectively, yielding

3
𝑦𝑛+1 = 𝑦𝑛 + ℎ𝑓𝑛

an explicit linear one –step method. It is, in fact, Euler’s explicit rule, the simplest of all LMM.

The error associated with it is the expression (1.4) (multiplied by ±1 according to the definition of error) and is

called the local truncation error (LTE) or local discretization error. For this method, the local truncation error is

𝑜(ℎ2 ) and that it is exactly zero if the solution of (1.1) is a polynomial of degree not exceeding one. The first term
2
of the LTE ℎ2 𝑦 ′′ (𝑥𝑛 ), is called the principal part of the LTE.

Consider now Taylor expansions for 𝑦(𝑥𝑛 + ℎ) and 𝑦(𝑥𝑛 − ℎ) about 𝑥𝑛 :


2 3
𝑦(𝑥𝑛 + ℎ) = 𝑦(𝑥𝑛 ) + ℎ𝑦 ′ (𝑥𝑛 ) + ℎ2! 𝑦 ′′ (𝑥𝑛 ) + ℎ3! 𝑦 ′′′ (𝑥𝑛 ) …

2 3
𝑦(𝑥𝑛 − ℎ) = 𝑦(𝑥𝑛 ) − ℎ𝑦 ′ (𝑥𝑛 ) + ℎ2! 𝑦 ′′ (𝑥𝑛 ) − ℎ3! 𝑦 ′′′ (𝑥𝑛 ) …

Subtracting, we get
3
𝑦(𝑥𝑛 + ℎ) − 𝑦(𝑥𝑛 − ℎ) = 2ℎ𝑦 ′ (𝑥𝑛 ) + ℎ3 𝑦 ′′′ (𝑥𝑛 ) …

Using the same argument as previously yields the associated LMM

𝑦𝑛+1 − 𝑦𝑛−1 = 2ℎ𝑓𝑛

This can be brought into the standard form (1.2) by replacing 𝑛 by 𝑛 + 1 to give

𝑦𝑛+2 − 𝑦𝑛 = 2ℎ𝑓𝑛+1
3
which is the Mid –point rule and its local truncation error is ±ℎ3! 𝑦 ′′′ (𝑥𝑛 ) + ⋯ .

Similar techniques can be used to derive any LMM of given specification. Thus if we wish to find the most

accurate one –step implicit method

𝑦𝑛+1 + 𝛼0 𝑦𝑛 = ℎ[𝛽1 𝑓𝑛+1 + 𝛽0 𝑓𝑛 ]

We write down the associated approximate relationship

𝑦(𝑥𝑛 + ℎ) + 𝛼0 𝑦(𝑥𝑛 ) = ℎ[𝛽1 𝑦 ′ (𝑥𝑛 + ℎ) + 𝛽0 𝑦 ′ (𝑥𝑛 )] (1.5)

and we choose 𝛼0 , 𝛽1 , 𝛽0 so as to make the approximation as accurate as possible. The following expansions are

used:
2 3
𝑦(𝑥𝑛 + ℎ) = 𝑦(𝑥𝑛 ) + ℎ𝑦 ′ (𝑥𝑛 ) + ℎ2! 𝑦 ′′ (𝑥𝑛 ) + ℎ3! 𝑦 ′′′ (𝑥𝑛 ) + ⋯

2
𝑦 (1) (𝑥𝑛 + ℎ) = 𝑦 ′ (𝑥𝑛 ) + ℎ𝑦 ′′ (𝑥𝑛 ) + ℎ2! 𝑦 ′′′ (𝑥𝑛 ) + ⋯

Substituting in (1.5) and collecting the terms on LHS gives

4
𝑐0 𝑦(𝑥𝑛 ) + 𝑐1 ℎ𝑦 ′ (𝑥𝑛 ) + 𝑐2 ℎ2 𝑦 ′′ (𝑥𝑛 ) + 𝑐3 ℎ3 𝑦 ′′′ (𝑥𝑛 ) + ⋯ ≅ 0

where

𝑐0 = 1 + 𝛼0 , 𝑐1 = 1 − 𝛽1 − 𝛽0

𝑐2 = 12−𝛽1, 𝑐3 = 16 − 12𝛽1

Thus in order to make the approximation in (1.5) as accurate as possible, we choose 𝛼0 = −1, 𝛽1 = 𝛽0 = 12. 𝑐3

1
then takes the value −12 . The LMM is now

𝑦𝑛+1 − 𝑦𝑛 = ℎ2[𝑓𝑛+1 + 𝑓𝑛 ]

1 3 ′′′
the Trapezoidal rule, its local truncation error is −12ℎ 𝑦 (𝑥𝑛 ) + ⋯.

Exercise 1.1: Find the most accurate implicit linear two –step method. Find also the first term in the local

truncation error.

1.2: Derivation through Numerical Integration

Consider the identity


𝑥
𝑦(𝑥𝑛+2 ) − 𝑦(𝑥𝑛 ) ≡ ∫𝑥 𝑛+2 𝑦 ′ (𝑥)𝑑𝑥 (1.6)
𝑛

Using (1.1), we can replace 𝑦 (1) (𝑥) by 𝑓(𝑥, 𝑦) in the integrand. If our aim is to derive, say a linear two –step

method, then the only available data for the approximate evaluation of the integral will be the values

(𝑥𝑛 , 𝑓𝑛 ), (𝑥𝑛+1 , 𝑓𝑛+1 ), (𝑥𝑛+2 , 𝑓𝑛+2 ). By the Newton –Gregory Forward interpolation formula,

𝑝(𝑥) = 𝑝(𝑥𝑛 + 𝑟ℎ) = 𝑓𝑛 + 𝑟Δ𝑓𝑛 + 𝑟(𝑟−1)


2!
Δ2 𝑓𝑛

We now make the approximation

𝑥 𝑥 2
∫𝑥 𝑛+2 𝑦 ′ (𝑥)𝑑𝑥 ≅ ∫𝑥 𝑛+2 𝑝(𝑥)𝑑𝑥 = ∫0 [𝑓𝑛 + 𝑟Δ𝑓𝑛 + 𝑟(𝑟−1)
2!
1 2
Δ2 𝑓𝑛 ]ℎ𝑑𝑟 = ℎ(2𝑓𝑛 + 2Δfn + Δ 𝑓𝑛 )
3
𝑛 𝑛

Expanding Δ𝑓𝑛 and Δ2 𝑓𝑛 in terms of 𝑓𝑛 , 𝑓𝑛+1 , 𝑓𝑛+2 and substituting in (1.6) gives

𝑦𝑛+2 − 𝑦𝑛 = ℎ3[𝑓𝑛+2 + 4𝑓𝑛+1 + 𝑓𝑛 ]

which is Simpson’s rule, the most accurate linear two –step method.

If we replace (1.6) by the identity


𝑥
𝑦(𝑥𝑛+2 ) − 𝑦(𝑥𝑛+1 ) ≡ ∫𝑥 𝑛+2 𝑦 ′ (𝑥)𝑑𝑥
𝑛+1

and replace 𝑦 ′ (𝑥) by 𝑝(𝑥), defined as above, we derive the method

5

𝑦𝑛+2 − 𝑦𝑛+1 = 12[5𝑓𝑛+2 + 8𝑓𝑛+1 − 𝑓𝑛 ]

the two –step Adams –Moulton Method.

Clearly this technique can be used to derive only a subclass of LMM consisting of hose methods for which

𝛼𝑘 = 1, 𝛼𝑗 = −1, 𝛼𝑖 = 0, 𝑖 = 0,1, … , 𝑗 − 1, 𝑗 + 1, … , 𝑘, 𝑗 ≠ 𝑘

The importance of this technique is that it establishes a link between the concepts of polynomial interpolation

and LMM.

Convergence

A basic property which we shall demand of an acceptable LMM is that the solution {𝑦𝑛 } generated by the method

converges, in some sense, to the theoretical solution 𝑦(𝑥) as the step length tends to zero.

Order and Error Constant

With the LMM (1.2), we associate the linear difference operator defined by

ℒ[𝑦(𝑥); ℎ] = ∑[ 𝛼𝑗 𝑦(𝑥 + 𝑗ℎ) − ℎ𝛽𝑗 𝑦 ′ (𝑥 + 𝑗ℎ)] (1.7)


𝑗=0

where 𝑦(𝑥) is an arbitrary function, continuously differentiable on [𝑎, 𝑏]. Expanding 𝑦(𝑥 + 𝑗ℎ) and its derivative

𝑦 ′ (𝑥 + 𝑗ℎ) as Taylor series about 𝑥, and collecting terms in (1.7) give

ℒ[𝑦(𝑥); ℎ] = 𝑐0 𝑦(𝑥) + 𝑐1 ℎ𝑦 ′ (𝑥) + … + 𝑐𝑞 ℎ𝑞 𝑦 (𝑞) (𝑥) + ⋯

where

𝑐0 = 𝛼0 + 𝛼1 + ⋯ + 𝛼𝑘

𝑐𝑞 = 𝑞!1 (𝛼1 + 2𝑞 𝛼2 + ⋯ + 𝑘 𝑞 𝛼𝑘 ) − (𝑞−1)!


1
(𝛽1 + 2𝑞−1 𝛽2 + ⋯ + 𝑘 𝑞−1 𝛽𝑘 )

𝑞 = 2,3, …

Definition 1.1 (Order of LMM)

The LMM (1.2) is said to be of order 𝑝 if 𝑐0 = 0, 𝑐1 = 0, … , 𝑐𝑝 = 0 but 𝑐𝑝+1 ≠ 0 and 𝑐𝑝+1 is called the error

constant.

Example 1.2: Find the order and error constant of the trapezoidal method 𝑦𝑛+1 − 𝑦𝑛 = ℎ2[𝑓𝑛+1 + 𝑓𝑛 ].

Solution: 𝛼1 = 1, 𝛼0 = −1, 𝛽0 = 𝛽1 = 12

6
𝑐0 = 𝛼0 + 𝛼1 = 0, 𝑐1 = 𝛼1 − (𝛽0 + 𝛽1 ) = 0, 𝑐2 = 12𝛼1 − 𝛽1 = 0, 𝑐3 = 16𝛼1 − 12𝛽1 = −12
1
≠ 0. Hence order 𝑝 = 2

1
and error constant 𝑐3 = −12 .

Exercise 1.2: Find the order and error constant of (i) Mid-point method, (ii) Simpson’s method.

Local and global truncation error

For an explicit method, the local truncation error is the difference between the theoretical solution and the

solution given by the LMM under the assumption that no previous errors have been made; for an implicit

method, the local truncation error is (approximately) proportional to this difference. If we make no such

assumption, then the error 𝑦(𝑥𝑛+𝑘 ) − 𝑦𝑛+𝑘 = 𝑒𝑛+𝑘 is the global (overall) or accumulated error. This error

involves all the truncation errors made at each application of the method. It is this error which convergence

demands should tend to zero as ℎ → 0.

Consistency and Zero-stability

Definition 1.2 (Consistency)

The LMM (1.2) is said to be consistent if it has order 𝑝 ≥ 1.

From (1.8) it follows that (1.2) is consistent if and only if

𝑘 𝑘 𝑘

∑ 𝛼𝑗 = 0; ∑ 𝑗𝛼𝑗 = ∑ 𝛽𝑗 (1.9)
𝑗=0 𝑗=0 𝑗=0

We now introduce the first and second characteristic (generating) polynomials of the LMM (1.2) defined as 𝜌(𝜉)

and 𝜎(𝜉) respectively, where 𝜌(𝜉) = ∑𝑘𝑗=0 𝛼𝑗 𝜉𝑗 and 𝜎(𝜉) = ∑𝑘𝑗=0 𝛽𝑗 𝜉𝑗 . Hence in terms of these polynomials,

(1.9) is equivalently written as (i) 𝜌(1) = 0; (ii) 𝜌′ (1) = 𝜎(1).

Thus, for a consistent method, the first characteristic polynomial 𝜌(𝜉) always has a root at 1. We shall call this

root the principal root 𝜉1 and the remaining roots 𝜉𝑠 , 𝑠 = 2,3, … , 𝑘 the spurious roots which arise only when the

step number of the method is greater than one. That is, when we choose to replace a first order differential

equation by a difference equation of order greater than one. The location of these spurious roots must be

carefully controlled if the method is to be convergent. Since consistency controls only the position of the principal

root and not of the spurious roots, a consistent method is not necessarily convergent.

7
Definition 1.3 (Zero-Stability)

The LMM (1.2) is said to be zero-stable if no root of the first characteristic polynomial 𝜌(𝜉) has modulus greater

than one, and every root with modulus one is simple (distinct).

The roots of 𝜌(𝜉) are in general complex, since for a zero-stable method all the roots of 𝜌(𝜉) lie in or on the unit

circle, those on the unit circle being simple. This is what is referred to as the root condition.

Zero-stability ensures that those solutions of the difference equation for 𝑦𝑛 which arise because the first order

differential equation is being replaced by a higher order difference equation are damped out in the limit as ℎ → 0.

Theorem 1.2 (Fundamental Theorem of Dahlquist on LMM)

The necessary and sufficient conditions for a LMM to be convergent are that it be consistent and zero-stable.

Quantitatively speaking, consistency controls the magnitude of the local truncation error committed at each stage

of the calculation, while zero-stability controls the manner in which this error is propagated as the calculation

proceeds; both are essential if convergence is to be achieved. Convergence is a minimal property which any

acceptable LMM must possess. Accordingly, we reject out of hand, as having no practical interest, LMMs which

are not both consistent and zero-stable.

Example 1.3: Investigate the consistency and zero-stability of the following methods:

(a) Trapezoidal method (b) Simpson’s method

Solution

(a) 𝑦𝑛+1 − 𝑦𝑛 = ℎ2[𝑓𝑛+1 + 𝑓𝑛 ]; 𝜌(𝜉) = 𝛼1 𝜉 + 𝛼0 =𝜉 − 1; 𝜌′ (𝜉) = 1; 𝜎(𝜉) = 𝛽1 𝜉 + 𝛽0 = 12𝜉 + 12

Consistency: (i) 𝜌(1) = 1 − 1 = 0; (ii) 𝜌′ (1) = 1; 𝜎(1) = 12 + 12 = 1 ⇒ 𝜌′ (1) = 𝜎(1) = 1

Therefore by definition 1.2, the method is consistent.

Zero-stability: 𝜌(𝜉) = 𝜉 − 1 = 0 ⇒ 𝜉 = 1; |𝜉| = 1. By definition 1.3, the method is zero-stable, hence

convergent.

(b) 𝑦𝑛+2 − 𝑦𝑛 = ℎ3[𝑓𝑛+2 + 4𝑓𝑛+1 + 𝑓𝑛 ], 𝜌(𝜉) = 𝜉 2 − 1; 𝜌′ (𝜉) = 2𝜉; 𝜎(𝜉) = 13𝜉 2 + 43𝜉 + 13

Consistency: (i) 𝜌(1) = 1 − 1 = 0; (ii) 𝜌′̇ (1) = 2; 𝜎(1) = 13 + 43 + 13 = 2 ⇒ 𝜌′ (1) = 𝜎(1) = 2 ⇒ the method is

consistent.

8
Zero-stability: 𝜌(𝜉) = 𝜉 2 − 1 = 0 ⇒ 𝜉1 = 1, 𝜉2 = −1 ; |𝜉1 | = 1 = |𝜉2 | but 𝜉1 ≠ 𝜉2 . Hence the method is zero-

stable. Thus Simpson’s method is convergent.

Example 1.4: Consider the following method 𝑦𝑛+2 − (1 + 𝑎)𝑦𝑛+1 + 𝑎𝑦𝑛 = ℎ2[(3−𝑎)𝑓𝑛+1 − (1 + 𝑎)𝑓𝑛 ].

(i) Find the parameter a for which the above method is consistent.

(ii) Find the range of values of the parameter a for which the above method is zero-stable.

Solution

𝜌(𝜉) = 𝜉 2 − (1 + 𝑎)𝜉 + 𝑎; 𝜌̇ (𝜉) = 2𝜉 − (1 + 𝑎); 𝜎(𝜉) = (3−𝑎)


2
𝜉 + (1+𝑎)
2

(i) 𝜌(1) = 1 − 1 − 𝑎 + 𝑎 = 0 for any value of a; (ii) 𝜌′ (1) = 2 − 1 − 𝑎 = 1 − 𝑎; 𝜎(1) = 3−𝑎−1−𝑎


2
=1−𝑎 ⇒

𝜌′ (1) = 𝜎(1) = 1 − 𝑎 for any value f a. Hence the method is consistent for any value of a.

(ii) Zero-stability: 𝜌(𝜉) = 𝜉 2 − (1 + 𝑎)𝜉 + 𝑎 = (𝜉 − 1)(𝜉 − 𝑎) = 0 ⇒ 𝜉1 = 1, 𝜉2 = 𝑎. The method will be zero-

stable for −1 ≤ 𝑎 < 1. If 𝑎 ≥ 1 or 𝑎 < −1, the method will not be zero-stable (i.e. zero-unstable).

Exercise 1.3: Investigate the consistency and zero-stability of the following methods: (i) Mid-point (ii) Two-step

Adams-Moulton.

Theorem 1.3

No zero-stable LMM of step number k can have order exceeding k+1 when k is odd, or exceeding k+2 when k is

even.

Definition 1.4 (Optimal Methods)

A zero-stable linear k-step method which has order k+2 is called an optimal method. For an optimal method all

the roots of 𝜌(𝜉) lie on the unit circle.

The highest order we can expect from a k-step method is 2k if the method is implicit, and 2k-1 if the method is

explicit. However, these maximal orders cannot, in general, be obtained without violating the condition of zero-

stability as theorem 1.3 shows. Simpson’s method is an example of a method which is both maximal and optimal.

It has step number two and order four, thus is both maximal, since It has order 2k, and optimal, since it has order

k+2. It is the only LMM to possess both of these properties.

9
One-step methods (single step methods)

By a one-step method we mean a method whereby 𝑦𝑛 is the only input data to the step in which 𝑦𝑛+1 is to be

computed for 𝑛 = 0,1,12, … (that is we need just one value from the past). Examples are Euler’s method (explicit

and implicit) and Trapezoidal method. One-step methods are always self-starting. No other method is needed to

generate any input data, hence they are sometimes called self-starting methods.

Specification of Linear Multistep Methods

Methods for which 𝜌(𝜉) = 𝜉 𝑘 − 𝜉 𝑘−1 are called Adams methods. They have the property hat all the spurious

roots of 𝜌(𝜉) are located at the origin; such methods are zero-stable. Adams methods which are explicit are called

Adams-Bashforth Methods (ABM), while the implicit ones are called Adams-Moulton Methods (AMM).

Explicit methods for which 𝜌(𝜉) = 𝜉 𝑘 − 𝜉 𝑘−2 are called Nystrom methods and implicit methods with the same

form for 𝜌(𝜉) are called generalized Milne-Simpson methods, both these families are clearly zero-stable, since

they have one spurious root −1 and the rest at the orgin.

Definition 1.4 (Absolute Stability)

The LMM (1.2) is said to be absolutely stable for a given ℎ̅, if for that ℎ̅, all the roots 𝑟𝑠 of the stability polynomial

𝜋(𝑟; ℎ̅) = 𝜌(𝑟) − ℎ̅𝜎(𝑟) = 0 satisfy |𝑟𝑠 | < 1, 𝑠 = 1,2, … , 𝑘, and to be absolutely unstable for that ℎ̅ otherwise. An

interval (𝛼, 𝛽) of the real line is said to be an interval of absolute stability if the method is absolutely stable for all

ℎ̅ ∈ (𝛼, 𝛽). If the method is absolutely unstable for all ℎ̅, it is said to have no interval of absolute stability.

It can be shown that every optimal LMM has no interval of absolute stability [it is the point 0 = [0,0])]. If all the

spurious zeros of the first characteristic polynomial 𝜌(𝜉) lie strictly inside the unit circle, then the method has a

non-vanishing interval of absolute stability. Again, we expect that if all the spurious zeros are situated at the

origin (that is, if the method is an Adams method), the interval of absolute stability will be substantial.

The size and shape of the region of absolute stability is important in assessing the value of a particular method

and when comparing it with other methods. We shall consider two methods of obtaining the region of absolute

stability namely the boundary locus method and Schur method. The third method is the Rout-Hurwitz criterion.

10
Boundary Locus Method (BLM)

The most usual method for finding the region of absolute stability LMM is the so called boundary locus method

(BLM). The boundary locus curve is obtained by setting

ℎ̅ = 𝜌(𝑟)
𝜎(𝑟)
, 𝑟 = 𝑒 𝑖𝜃 , 0° ≤ 𝜃 ≤ 180°.

The curve is normally symmetric about the real axis. The upper half is obtained for 0° ≤ 𝜃 ≤ 180° and the mirror

image of it through the real axis completes the region of absolute stability.

Example 1.5: Find the interval of absolute stability of the two-step Adams-Moulton Method.


Solution: The two-step AMM is given by 𝑦𝑛+2 − 𝑦𝑛+1 = 12[5𝑓𝑛+2 + 8𝑓𝑛+1 − 𝑓𝑛 ]. Hence 𝜌(𝑟) = 𝑟 2 − 𝑟 and

1
𝜎(𝑟) = 12(5𝑟 2 + 8𝑟 − 1).

𝑖𝜃 ) 12(𝑒2𝑖𝜃 −𝑒𝑖𝜃 )
The boundary locus curve is ℎ̅(𝜃) = 𝜌(𝑒
𝜎(𝑒𝑖𝜃 )
= 5𝑒 12(cos 2𝜃+𝑖 sin 2𝜃)−12(cos 𝜃+𝑖 sin 𝜃)
2𝑖𝜃 +8𝑒𝑖𝜃 −1 = 5(cos 2𝜃+𝑖 sin 2𝜃)+8(cos 𝜃+𝑖 sin 𝜃)−1 =

(12 cos 2𝜃−12cos 𝜃)+𝑖(12 sin 2𝜃−sin 𝜃)


(5cos 2𝜃+8cos 𝜃−1)+𝑖(5sin 2𝜃+8 sin 𝜃)
.

Rationalizing the denominator by multiplying ℎ̅(𝜃) by (5cos 2𝜃+8cos 𝜃−1)−𝑖(5sin 2𝜃+8 sin 𝜃)
(5cos 2𝜃+8cos 𝜃−1)−𝑖(5sin 2𝜃+8 sin 𝜃)
and writing

ℎ̅(𝜃) = 𝑥(𝜃) + 𝑖𝑦(𝜃) , we have

(12 cos 2𝜃 − 12cos 𝜃)(5cos 2𝜃 + 8cos 𝜃 − 1) + (12 sin 2𝜃 − sin 𝜃)(5sin 2𝜃 + 8 sin 𝜃)
𝑥(𝜃) =
(5cos 2𝜃 + 8cos 𝜃 − 1)2 + (5sin 2𝜃 + 8 sin 𝜃)2

−36 + 48 cos 𝜃 − 12 cos 2𝜃


=
(5cos 2𝜃 + 8cos 𝜃 − 1)2 + (5sin 2𝜃 + 8 sin 𝜃)2

𝑥(0°) = 0, 𝑥(180°) = −6. Thus interval of absolute stability of the 2-step AMM is [−6,0].

168 sin θ −12 sin 2𝜃


Similarly, 𝑦(𝜃) = (5cos 2𝜃+8cos 𝜃−1)2 +(5sin 2𝜃+8 sin 𝜃)2. For 0° ≤ 𝜃 ≤ 180°, a table of values of 𝑥(𝜃) and 𝑦(𝜃) is

completed and the graph of 𝑦(𝜃) is plotted against 𝑥(𝜃). Since this graph is symmetric about the real axis, its

mirror image through this line of symmetry gives the complete region of absolute stability of the method.

Exercise 1.4: Find the interval of absolute stability of the following methods:

(i) Simpson’s method

(ii) 𝑦𝑛+2 − 45𝑦𝑛+1 − 15𝑦𝑛 = ℎ2[2𝑓𝑛+2 + 4𝑓𝑛+1 ]. Also find the region of absolute stability for (ii).

11
Schur’s Method

The second method we shall consider is Schur’s method often referred to as Schurz criterion. Consider the

general 𝑘 𝑡ℎ degree polynomial with complex coefficients

𝜙(𝑟) = 𝑐𝑘 𝑟 𝑘 + 𝑐𝑘−1 𝑟 𝑘−1 + ⋯ + 𝑐1 𝑟 + 𝑐0

where 𝑐𝑘 ≠ 0, 𝑐0 ≠ 0. The polynomial 𝜙(𝑟) is said to be a Schur polynomial if its roots 𝑟𝑠 satisfy

|𝑟𝑠 | < 1, 𝑠 = 1,2, … , 𝑘.

Define the polynomials

𝜙̂(𝑟) = 𝑐0∗ 𝑟 𝑘 + 𝑐1∗ 𝑟 𝑘−1 + ⋯ + 𝑐𝑘−1



𝑟 + 𝑐𝑘∗

where 𝑐𝑗∗ is the complex conjugate of 𝑐𝑗 , 𝑗 = 0,1, … , 𝑘 and

1
𝜙1 (𝑟) = 𝑟[𝜙̂(0)𝜙(𝑟) − 𝜙(0)𝜙̂(𝑟)]

Clearly, 𝜙1 (𝑟) has degree at most 𝑘 − 1. Then 𝜙(𝑟) is a Schur polynomial if and only if

(i) |𝜙̂(0)| > |𝜙(0)|

(ii) 𝜙1 (𝑟) is a Schur polynomial.

The interval (𝛼, 𝛽) is an interval of absolute stability if for all ℎ̅ ∈ (𝛼, 𝛽), the stability polynomial

𝜋(𝑟; ℎ̅) = 𝜌(𝑟) − ℎ̅𝜎(𝑟) is a Schur polynomial (that is if and only if |𝜋̂(0; ℎ̅)| > |𝜋(0; ℎ̅)| and 𝜋1 (𝑟; ℎ̅) is Schur).

Example 1.6: Consider the polynomial 𝜙(𝑥) = 𝑥 2 + 𝑏𝑥 + 𝑐, where 𝑏 and 𝑐 are real. Find the conditions on 𝑏

and 𝑐 for which the above polynomial is a Schur polynomial.

Solution: 𝜙(𝑥) = 𝑥 2 + 𝑏𝑥 + 𝑐, 𝜙(0) = 𝑐 and 𝜙̂(𝑥) = 𝑐 ∗ 𝑥 2 + 𝑏 ∗ 𝑥 + 1 = 𝑐𝑥 2 + 𝑏𝑥 + 1 (since 𝑏 and 𝑐 are real),

𝜙̂(0) = 1.

From 𝜙1 (𝑟) = 1𝑟[𝜙̂(0)𝜙(𝑟)−𝜙(𝑜)𝜙̂(𝑟), we have 𝜙1 (𝑥) = 𝑥1[𝑥 2 (1 − 𝑐 2 ) + 𝑥(𝑏 − 𝑏𝑐)] = (1 − 𝑐 2 )𝑥 + (𝑏 − 𝑏𝑐)

Now |𝜙̂(0)| > |𝜙(0)| ⇒ 1 > |𝑐| or |𝑐| < 1, that is, −1 < 𝑐 < 1. From the second condition, 𝜙1 (𝑥) must be a

Schur polynomial, that is, (1 − 𝑐 2 )𝑥 + (𝑏 − 𝑏𝑐)=0 has a root whose modulus is less than one. Thus

𝑏
(1 − 𝑐 2 )𝑥 + (𝑏 − 𝑏𝑐) = 0 ⟺ (1 + 𝑐)(1 − 𝑐)𝑥 + 𝑏(1 − 𝑐) = 0 ∴ 𝑥 =
(1+𝑐)

𝑏
And |𝑥| = |(1+𝑐) | < 1 ⇒ |𝑏| < |1 + 𝑐|. Hence the conditions to be imposed on 𝑏 and 𝑐 are that(i) |𝑐| < 1 and (ii)

|𝑏| < |1 + 𝑐|. For example, 𝑥 2 + 𝑥 + 12 is a Schur polynomial because it satisfies (i) and (ii) above.

12
Schurz’ criterion for finding the interval of absolute stability suffers a disadvantage for LMM of step numbers

greater than two. Moreover, one cannot obtain the region of absolute stability with this method.

Example 1.7: Use Schur criterion to obtain the interval of absolute stability of the method

𝑦𝑛+2 − 𝑦𝑛 = ℎ2[𝑓𝑛+2 + 3𝑓𝑛 ].

̅ ̅
Solution: The stability polynomial is 𝜋(𝑟; ℎ̅) = 𝑟 2 − 12ℎ̅𝑟 − (1 + 3ℎ
2
). Thus 𝜋̂(𝑟; ℎ̅) = −(1 + 3ℎ
2
)𝑟 2 − 12ℎ̅𝑟 + 1 and

̅
the condition |𝜋̂(0; ℎ̅)| > |𝜋(0; ℎ̅)| is satisfied if |1 + 3ℎ
2
| < 1 ⟺ −43 < ℎ̅ < 0.

̅ ̅ ̅
Now 𝜋1 (𝑟; ℎ̅) = 1𝑟[𝑟 2 − 12ℎ̅𝑟 − (1 + 3ℎ
2
) + (1 + 3ℎ
2
)(−(1 + 3ℎ
2
)𝑟 2 − 12ℎ̅𝑟 + 1)] which has its only root at −
1
, and is
3

therefore a Schur polynomial. It follows that the interval of absolute stability is (−43, 0). Use BLM to find the

interval of absolute stability of his method and verify that 𝑟 = −13, also, use test equation).

Example 1.8: Find the interval of absolute stability of the (i) Trapezoidal method, (ii) Euler’s explicit method.

Solution: (i) 𝑦𝑛+1 − 𝑦𝑛 = ℎ2[𝑓𝑛+1 + 𝑓𝑛 ], using the test equation 𝑦 ′ = 𝜆𝑦, we have

𝑦𝑛+1 − 𝑦𝑛 = ℎ2[𝜆𝑦𝑛+1 + 𝜆𝑦𝑛 ] or (1 − ℎ𝜆


2
)𝑦𝑛+1 − (1 − ℎ𝜆
2
)𝑦𝑛 = 0. The characteristic equation gives

ℎ𝜆
1+ 2
(1 − ℎ𝜆
2
)𝑥 − (1 − ℎ𝜆
2
)=0⇒𝑥= ℎ𝜆 = 1+𝑞 ⁄2
1−𝑞⁄2
, where 𝑞 = ℎ𝜆 or ℎ̅ = ℎ𝜆. By Schur’s condition |1+𝑞 ⁄2
1−𝑞⁄2
| < 1 and this
1−
2

inequality is satisfied for 𝑞 < 0. Hence the interval of absolute stability is (−∞, 0). For LMMs of step number two

or more, we check the second condition as well.

(ii) 𝑦𝑛+1 − 𝑦𝑛 = ℎ𝑓𝑛 and the test equation yields 𝑦𝑛+1 − 𝑦𝑛 = ℎ𝜆𝑦𝑛 ⇒ 𝑦𝑛+1 − (1 + ℎ𝜆)𝑦𝑛 = 0 with

characteristic equation 𝑥 − (1 + ℎ𝜆) = 0 ⟺ 𝑥 = 1 + ℎ𝜆 = 1 + 𝑞, 𝑞 = ℎ𝜆. Form Schur’s condition, |1 + 𝑞| <

1 ⟺ −1 < 1 + 𝑞 < 1, that is, −1 < 1 + 𝑞; 1 + 𝑞 < 1 ⟺ (𝑞 > −2)⋀(𝑞 < 0) ⟺ −2 < 𝑞 < 0. Hence the

interval of absolute stability of Euler’s explicit method is (−2,0). Similarly, the absolute stability interval of Euler’s

implicit method is (−∞, 0) ∪ (2, ∞).

2. STIFF DIFFERENTIAL EQUATIONS

All the methods for approximating the solution to initial-value problems have error terms that involve higher

derivative of the solution of the equation. If the derivative can be reasonably bounded, then the method will have

a predictable error bound that can be used to estimate the accuracy of the approximation. Even if the derivative

grows as the steps increase, the error can be kept in relative control, provided that the solution also grows in
13
magnitude. Problems frequently arise, however, when the magnitude of the derivative increases, but the solution

does not. In this situation, the error can grow so large that it dominates the calculations. Initial-value problems for

which this is likely to occur are called stiff-equations and are quite common, particularly in the study of vibrations,

chemical reactions and electrical circuits. Stiff systems derive their name from the motion of spring and mass

systems that have large spring constants.

Stiff differential equations are characterized as those whose exact solution has a term of the form 𝑒 −𝑐𝑡 , where 𝑐 is

a large positive constant. This is usually only a part of the solution, called the transient solution; the more

important portion of the solution is called the steady state solution. A transient portion of a stiff equation will

rapidly decay to zero as 𝑡 increases.

For example, the stiff initial-value problem 𝑦 ′ = −20(𝑦 − 𝑡 2 ) + 2𝑡, 0 ≤ 𝑡 ≤ 1, 𝑦(0) = 13 has the exact solution

𝑦(𝑡) = 𝑡 2 + 13𝑒 −20𝑡 (𝑡 2 is the steady state solution while 13𝑒 −20𝑡 is the transient solution).

Definition 2.1 (First Order Stiff Linear Differential Equation)

The linear system 𝑦̇ = 𝐴𝑦 + 𝜙(𝑥), where 𝐴 is an 𝑚 × 𝑚 matrix, is said to be stiff if:

(i) 𝑅𝑒𝜆𝑖 < 0, 𝑖 = 1,2, … , 𝑚 and

max |𝑅𝑒𝜆𝑖 |
𝑖=1,…,𝑚
(ii) min |𝑅𝑒𝜆𝑖 |
≫0
𝑖=1,…,𝑚

[ max |𝑅𝑒𝜆𝑖 |]
𝑖=1,…,𝑚
Where 𝜆𝑖 , 𝑖 = 1,2, … , 𝑚 are the eigenvalues of 𝐴. The ratio 𝑆 = is called the stiffness ratio.
[ min |𝑅𝑒𝜆𝑖 |]
𝑖=1,…,𝑚

Example 2.1: Find the stiffness ratio for the system:

𝑦1′ = −10𝑦1 + 9𝑦2 = 𝑓1

𝑦2′ = 10𝑦1 − 11𝑦2 = 𝑓2

The Jacobian matrix of the above system

𝜕𝑓1 𝜕𝑓2
𝜕𝑦1 𝜕𝑦1 −10 10 |𝐽 −10 − 𝜆 10
𝐽 = [ 𝜕𝑓 ]=[ ] , − 𝜆𝐼| = | | = (−10 − 𝜆)(−11 − 𝜆) − 90 = 0
1 𝜕𝑓2 9 −11 9 −11 − 𝜆
𝜕𝑦2 𝜕𝑦2

𝜆2 + 21𝜆 + 20 = 0 ⟺ (𝜆 + 1)(𝜆 + 20) = 0 ∴ 𝜆1 = −1, 𝜆2 = −20 (note 𝑅𝑒𝜆1 < 0, 𝑅𝑒𝜆2 < 0)

[ max |𝑅𝑒𝜆𝑖 |]
𝑆= 𝑖=1,…,𝑚
= 20
1
= 20. The system is mildly stiff (that is, marginally stiff).
[ min |𝑅𝑒𝜆𝑖 |]
𝑖=1,…,𝑚

14
15
Exercise 2.1: Find the stiffness ratio for the system:

𝑦1′ = 𝑦2

𝑦2′ = −1000𝑦1 − 1001𝑦2

Note: stiff systems are occasionally referred to as “systems with large Lipchitz constants”.

The Problem of Stability for Stiff Systems

A basic difficulty in the numerical solution of stiff system is the satisfaction of the requirement of absolute

stability. Thus several definitions, which call for the method to possess some ‘adequate’ region of absolute

stability, have been proposed. These definitions originally constructed for LMMs, are applicable to any numerical

method which involves discretization with associated step length .

Definition 2.2 (A-Stability)

A numerical method is said to be 𝐴-stable if its region of absolute stability contains the whole of the left-hand

plane (𝑅𝑒ℎ2 𝜆 < 0).

Theorem 2.1

(i) An explicit LMM cannot be 𝐴-satble.

(ii) The order of an 𝐴-stable implicit LMM cannot exceed two.

(iii) The second-order 𝐴-stable implicit method with smallest error constant is the trapezoidal method.

The restriction on order implied by (ii) in Theorem 2.1 is a severe one. In view of this, several less demanding

stability definitions have been proposed.

Definition 2.3 (A(𝜶) and A(0)-Stability)

A numerical method is said to be 𝐴(𝛼)-stable, 𝛼 ∈ (0, 𝜋2), if its region of absolute stability contains the infinite

wedge 𝑊𝛼 = {ℎ𝜆|−𝛼 < 𝜋 − 𝑎𝑟𝑔ℎ𝜆 < 𝛼}; it is said to be 𝐴(0)-stable if it is 𝐴(𝛼)-stable for some (sufficiently

small) 𝛼 ∈ (0, 𝜋2).

y y

Wα α

x α x

A-stability A(α)-stability

16
Note: Conventional methods such as Adams-Moulton and explicit Runge-Kutta methods suffer from severe step

size constraints imposed by stability when applied to stiff problems. Thus, to overcome this stability restriction on

the step size, numerical methods that posses infinite regions of absolute stability have been recommended for

the solution of stiff problems. The best suited are the Backward Differentiation Formulae (BDF) methods as they

possess infinite regions of absolute stability.

3. RUNGE-KUTTA METHODS

Recall the initial-value problem (1.1). Of all computational methods for the numerical solution of this problem,

the easiest to implement is Euler’s rule, 𝑦𝑛+1 − 𝑦𝑛 = ℎ𝑓(𝑥𝑛 , 𝑦𝑛 ) = ℎ𝑓𝑛 . It is explicit and being a one-step method,

it requires no additional starting values. Its low order, however, makes it of limited practical value. LMMs achieve

higher order by sacrificing the one-step nature of the algorithm, whilst retaining linearity with respect to

𝑦𝑛+𝑗 , 𝑓𝑛+𝑗 , 𝑗 = 0,1, … , 𝑘. Higher order can also be achieved by sacrificing linearity, but preserving the one-step

nature of the algorithm. This is the philosophy behind the methods first proposed by Runge and subsequently

developed by Kutta and Heun. Runge-Kutta methods thus retain the advantages of one-step methods but, due to

the loss of linearity, error analysis is considerably more difficult than in the case of LMMs. Traditionally, Runge-

Kutta methods are all explicit, although recently, implicit Runge-Kutta methods which have improved weak

stability characteristics, have been considered. Thus a Runge-Kutta method may be regarded as a particular case

of the general explicit one-step method

𝑦𝑛+1 − 𝑦𝑛 = ℎ𝜙(𝑥𝑛 , 𝑦𝑛 , ℎ) (3.1)

Order and Convergence of the General Explicit One-step Method

The fact that the general method (3.1) makes no mention of the function 𝑓(𝑥, 𝑦) which defines the differential

equation makes it impossible to define the order of the method independently of the differential equation, as

was the case with LMMs.

Definition 3.1 (Order)

The method (3.1) is said to have order 𝑝 if 𝑝 is the largest integer for which

𝑦(𝑥 + ℎ) − 𝑦(𝑥) − ℎ𝜙(𝑥, 𝑦(𝑥), ℎ) = 𝑜(ℎ𝑝+1 )

17
holds , where 𝑦(𝑥) is the theoretical solution of the ivp.

Definition 3.2 (Consistency)

The method (3.1) is said to be consistent with the ivp if

𝜙(𝑥, 𝑦, 0) ≡ 𝑓(𝑥, 𝑦)

If the method (3.1) is consistent with the ivp or simply consistent, then

𝑦(𝑥 + ℎ) − 𝑦(𝑥) − ℎ𝜙(𝑥, 𝑦(𝑥), ℎ) = 𝑦(𝑥) + ℎ𝑦 ′ (𝑥) + 𝑜(ℎ2 ) − 𝑦(𝑥) − ℎ𝜙(𝑥, 𝑦(𝑥), ℎ)

= ℎ𝑦̇ (𝑥) − ℎ𝑦 ′ (𝑥) + 𝑜(ℎ2 ) = 𝑜(ℎ2 ),

since 𝑦 ′ (𝑥) = 𝑓(𝑥, 𝑦(𝑥)) = 𝜙(𝑥, 𝑦(𝑥),0), by definition 3.2. Thus a consistent method has order at least one.

The only LMM which falls within class (3.1) is Euler’s rule which we obtain by setting

𝜙(𝑥, 𝑦, ℎ) = 𝜙𝐸 (𝑥, 𝑦, ℎ) ≡ 𝑓(𝑥, 𝑦)(the subscript E denotes Euler)

Theorem 3.1

(i) Let the function 𝜙(𝑥, 𝑦, ℎ) be continuous jointly as a function of its three arguments in the region 𝒟

defined by 𝑥 ∈ [𝑎, 𝑏], 𝑦 ∈ (−∞, ∞), ℎ ∈ [0, ℎ0 ], ℎ0 > 0.

(ii) Let 𝜙(𝑥, 𝑦, ℎ) satisfy a Lipschitz condition of the form |𝜙(𝑥, 𝑦 ∗ , ℎ) − 𝜙(𝑥, 𝑦, ℎ)| ≤ 𝑀|𝑦 ∗ − 𝑦| for all

points (𝑥, 𝑦 ∗ , ℎ), (𝑥, 𝑦, ℎ) in D. Then the method (3.1) is convergent if and only if it is consistent.

Note that there is no requirement corresponding to zero-stability, since no parasitic solutions can arise with a

one-step method.

The general 𝑅-stage Runge-Kutta method is defined by

𝑦𝑛+1 − 𝑦𝑛 = ℎ𝜙(𝑥𝑛 , 𝑦𝑛 , ℎ)
𝑅

𝜙(𝑥, 𝑦, ℎ) = ∑ 𝑐𝑘 𝑘𝑟
𝑠=1
𝑘1 = 𝑓(𝑥, 𝑦)
𝑟−1 (3.2)
𝑘𝑟 = 𝑓 (𝑥 + ℎ𝑎𝑟 , 𝑦 + ℎ ∑ 𝑏𝑟𝑠 𝑘𝑠 ) , 𝑟 = 2,3, … , 𝑅
𝑠=1
𝑟−1

𝑎𝑟 = ∑ 𝑏𝑟𝑠 , 𝑟 = 2,3, … , 𝑅
𝑠=1 }

18
Note that an 𝑅-stage Runge-Kutta method involves 𝑅 function evaluations per step. Each of the functions

𝑘𝑟 (𝑥, 𝑦, ℎ), 𝑟 = 1,2, … , 𝑅, may be interpreted as an approximation to the derivative 𝑦̇ (𝑥), and the function

𝜙(𝑥, 𝑦, ℎ) as a weighted mean of these approximations.

The Taylor Algorithm of Order 𝒑.

ℎ 2 ′′ ℎ 𝑝 (𝑝)
𝑦(𝑥 + ℎ) = 𝑦(𝑥) + ℎ𝑦 ′ (𝑥) + 2!
𝑦 (𝑥) + ⋯+ 𝑝!
𝑦 (𝑥)

also falls within the class (3.1) and is obtained by setting


𝑝−1
𝜙(𝑥, 𝑦, ℎ) = 𝜙 𝑇 (𝑥, 𝑦, ℎ) = 𝑓(𝑥, 𝑦) + 2!ℎ 𝑓 ′ (𝑥, 𝑦) + ⋯ + ℎ 𝑝! 𝑓 (𝑝−1) (3.3)

𝑞
where 𝑓 (𝑞) (𝑥, 𝑦) = 𝑑𝑥
𝑑
𝑞 𝑓(𝑥, 𝑦), 𝑞 = 1,2, … , (𝑝 − 1) (the subscript 𝑇 denotes ‘Taylor’).

There is a great deal of tedious manipulation involved in deriving Runge-Kutta methods of higher order;

accordingly, we shall derive only methods of order up to three and quote some well known methods of order

four.

𝜕𝑓(𝑥,𝑦) 𝜕2 𝑓(𝑥,𝑦) 𝜕2 𝑓(𝑥,𝑦) 𝜕2 𝑓(𝑥,𝑦)


Introducing the shortened notation 𝑓 = 𝑓(𝑥, 𝑦), 𝑓𝑥 = , 𝑓𝑥𝑥 = , 𝑓𝑥𝑦 = , 𝑓𝑦𝑦 = , we
𝜕𝑥 𝜕𝑥 2 𝜕𝑥𝜕𝑦 𝜕𝑦 2

may write expression (3.3) in the form

𝜙 𝑇 (𝑥, 𝑦, ℎ) = 𝑓 + 12ℎ𝐹 + 16ℎ2 (𝐹𝑓𝑦 + 𝐺) + 𝑜(ℎ3 ) (3.4)

𝑑𝑓(𝑥,𝑦) 𝜕𝑓 𝜕𝑓 𝑑𝑦 𝜕𝑓 𝜕𝑓
where 𝐹 = 𝑓𝑥 + 𝑓𝑓𝑦 , 𝐺 = 𝑓𝑥𝑥 + 2𝑓𝑓𝑥𝑦 + 𝑓 2 𝑓𝑦𝑦 (note 𝑑𝑥
= 𝜕𝑥 + 𝜕𝑦 . 𝑑𝑥 = 𝜕𝑥 + 𝑦 ′ 𝜕𝑦 = 𝑓𝑥 + 𝑓𝑓𝑦 ).

It turns out that it is possible to achieve third order with 𝑅 = 3, and thus we need derive expansions only for the

functions 𝑘1 , 𝑘2 and 𝑘3 , where (from (3.2))

𝑘1 = 𝑓(𝑥, 𝑦) = 𝑓
𝑘2 = 𝑓(𝑥 + ℎ𝑎2 , 𝑦 + ℎ𝑎2 𝑘1 ) } (3.5)
𝑘3 = 𝑓(𝑥 + ℎ𝑎3 , 𝑦 + ℎ(𝑎3 − 𝑏32 )𝑘1 + ℎ𝑏32 𝑘2 )

Note: Taylor series expansion for functions of two variables

For a function of two variables 𝑓(𝑥, 𝑦), the rate of change of the function can be due to change in either 𝑥 or 𝑦.

The derivatives of 𝑓 can be expressed in terms of the partial derivatives. For the expansion in the neighbourhood

of the point (𝑎, 𝑏),

1
𝑓(𝑥, 𝑦) = 𝑓(𝑎, 𝑏) + (𝑥 − 𝑎)𝑓𝑥 (𝑎, 𝑏) + (𝑦 − 𝑏)𝑓𝑦 (𝑎, 𝑏) + 2![(𝑥 − 𝑎)2 𝑓𝑥𝑥 (𝑎, 𝑏) + 2(𝑥 − 𝑎)(𝑦 − 𝑏)𝑓𝑥𝑦 (𝑎, 𝑏)

+ (𝑦 − 𝑏)2 𝑓𝑦𝑦 (𝑎, 𝑏)] + ⋯

19
Now expanding 𝑘2 as a Taylor series about the point (𝑥, 𝑦), we have

𝑘2 = 𝑓(𝑥 + ℎ𝑎2 , 𝑦 + ℎ𝑎2 𝑘1 )


1
= 𝑓 + (ℎ𝑎2 )𝑓𝑥 + (ℎ𝑎2 𝑘1 )𝑓𝑦 + 2![(ℎ𝑎2 )2 𝑓𝑥𝑥 + 2(ℎ𝑎2 )(ℎ𝑎2 𝑘1 )𝑓𝑥𝑦 + (ℎ𝑎2 𝑘1 )2 𝑓𝑦𝑦 ] + 𝑜(ℎ3 )

1
= 𝑓 + ℎ𝑎2 (𝑓𝑥 + 𝑘1 𝑓𝑦 ) + 2![(ℎ2 𝑎22 )𝑓𝑥𝑥 + 2(ℎ2 𝑎22 𝑘1 )𝑓𝑥𝑦 + (ℎ2 𝑎22 𝑘 2 )𝑓𝑦𝑦 ] + 𝑜(ℎ3 )

1
= 𝑓 + ℎ𝑎2 (𝑓𝑥 + 𝑓𝑓𝑦 ) + 2ℎ2 𝑎22 [𝑓𝑥𝑥 + 2𝑓𝑥𝑦 + 𝑓 2 𝑓𝑦𝑦 ] + 𝑜(ℎ3 )

or

𝑘2 = 𝑓 + ℎ𝑎2 𝐹 + 12ℎ2 𝑎22 𝐺 + 𝑜(ℎ3 ) (3.6)

Treating 𝑘3 similarly, we obtain

ℎ2 2
𝑘3 = 𝑓 + ℎ{𝑎3 𝑓𝑥 + [(𝑎3 − 𝑏32 )𝑘1 + 𝑏32 𝑘2 ]𝑓𝑦 } + 2
{𝑎3 𝑓𝑥𝑥 + 2𝑎3 [(𝑎3 − 𝑏32 )𝑘1 + 𝑏32 𝑘2 ]𝑓𝑥𝑦

+ [(𝑎3 − 𝑏32 )𝑘1 + 𝑏32 𝑘2 ]2 𝑓𝑦𝑦 } + 𝑜(ℎ3 )

On substituting for 𝑘1 and 𝑘2 we find, after some manipulation,

𝑘3 = 𝑓 + ℎ𝑎3 𝐹 + ℎ2 (𝑎2 𝑏32 𝐹𝑓𝑦 + 12𝑎32 𝐺) + 𝑜(ℎ3 ) (3.7)

On substituting from (3.5)-(3.7) into (3.2) the following expansion for 𝜙(𝑥, 𝑦, ℎ) is obtained:
2
𝜙(𝑥, 𝑦, ℎ) = (𝑐1 + 𝑐2 + 𝑐3 )𝑓 + ℎ(𝑐2 𝑎2 + 𝑐3 𝑎3 )𝐹 + ℎ2 [2𝑐3 𝑎2 𝑏32 𝐹𝑓𝑦 + (𝑐2 𝑎22 + 𝑐3 𝑎32 )𝐺] + 𝑜(ℎ3 ) (3.8)

We now have to match the expansions (3.4) and (3.8). Let us see what can be achieved with 𝑅 = 1,2,3 (for 𝑅 > 3,

(3.8) is invalid since 𝑘4 will contribute additional terms).

For 𝑅 = 1, 𝑐2 = 𝑐3 = 0 and (3.8) reduces to

𝜙(𝑥, 𝑦, ℎ) = 𝑐1 𝑓 + 𝑜(ℎ3 ) (3.9)

Setting 𝑐1 = 1, (3.9) differs from the expansion (3.4) for 𝜙 𝑇 by a term of order . Thus the resulting method,

which is Euler’s rule, has order one (that is, Rung-Kutta method for 𝑅 = 1 gives Euler’s explicit rule which is the

first class or order one).

For 𝑅 = 2, 𝑐3 = 0, and (3.8) reduces to


2
𝜙(𝑥, 𝑦, ℎ) = (𝑐1 + 𝑐2 )𝑓 + ℎ𝑐2 𝑎2 𝐹 + ℎ2 𝑐2 𝑎22 𝐺 + 𝑜(ℎ3 ) (3.10)

To match the expansion (3.10) with (3.4) we must satisfy the equations

𝑐1 + 𝑐2 = 1, 𝑐2 𝑎2 = 12 (3.11)

20
This is a set of two equations in three unknowns and thus there exists a one-parameter family of solutions. Thus

there exists an infinite number of two-stage Runge-Kutta methods of order two and none of order more than

two. This lack of uniqueness is typical of all Runge-Kutta derivations. Two particular solutions of (3.11) yield well-

known methods:

(i) 𝑐1 = 0, 𝑐2 = 1, 𝑎2 = 12. The resulting method is 𝑦𝑛+1 − 𝑦𝑛 = ℎ𝑓(𝑥𝑛 + 12ℎ, 𝑦𝑛 + 12ℎ𝑓(𝑥𝑛 , 𝑦𝑛 )). This

method, originally due to Runge, is referred to as the Modified Euler Method or the Improved

Polygon Method.

(ii) 𝑐1 = 12, 𝑐2 = 12, 𝑎2 = 1. The resulting method is

𝑦𝑛+1 − 𝑦𝑛 = ℎ2[𝑓(𝑥𝑛 , 𝑦𝑛 ) + 𝑓(𝑥𝑛 + ℎ, 𝑦𝑛 + ℎ𝑓(𝑥𝑛 , 𝑦𝑛 ))] is known as the Improved Euler method

For 𝑅 = 3, we can match (3.8) and (3.4) up to and including ℎ2 terms if we satisfy the following set of equations:

𝑐1 + 𝑐2 + 𝑐3 = 1
𝑐2 𝑎2 + 𝑐3 𝑎3 = 12
(3.12)
𝑐2 𝑎22 + 𝑐3 𝑎32 = 13
𝑐3 𝑎2 𝑏32 = 16 }

There are now four equations in six unknowns and there exists a two-parameter family of solutions. Two

particular solutions of (3.12) lead to well-known third order Runge-Kutta methods:

(i) 𝑐1 = 14, 𝑐2 = 0, 𝑐3 = 34, 𝑎2 = 13, 𝑎3 = 23, 𝑏32 = 23. The resulting method


𝑦𝑛+1 − 𝑦𝑛 = 4 (𝑘1 + 3𝑘3 )

where 𝑘1 = 𝑓(𝑥𝑛 , 𝑦𝑛 ), 𝑘2 = 𝑓(𝑥𝑛 + 13ℎ, 𝑦𝑛 + 13ℎ𝑘1 ), 𝑘3 = 𝑓(𝑥𝑛 + 23ℎ, 𝑦𝑛 + 23ℎ𝑘2 ), is known as Heun’s

third order formula.

(ii) 𝑐1 = 16, 𝑐2 = 23, 𝑐3 = 16, 𝑎2 = 12, 𝑎3 = 1, 𝑏32 = 2. The resulting method


𝑦𝑛+1 − 𝑦𝑛 = 6 (𝑘1 + 4𝑘2 + 𝑘3 )

where 𝑘1 = 𝑓(𝑥𝑛 , 𝑦𝑛 ), 𝑘2 = 𝑓(𝑥𝑛 + 12ℎ, 𝑦𝑛 + 12ℎ𝑘1 ), 𝑘3 = 𝑓(𝑥𝑛 + ℎ, 𝑦𝑛 − ℎ𝑘1 + 2ℎ𝑘2 ), is known as

Kutta’s third order rule.

The derivation of fourth-order Runge-Kutta (also called the classical R-K) method is


𝑦𝑛+1 − 𝑦𝑛 = 6 (𝑘1 + 2𝑘2 + 2𝑘3 + 𝑘4 )

where 𝑘1 = 𝑓(𝑥𝑛 , 𝑦𝑛 ), 𝑘2 = 𝑓(𝑥𝑛 + 12ℎ, 𝑦𝑛 + 12ℎ𝑘1 ), 𝑘3 = 𝑓(𝑥𝑛 + 12ℎ, 𝑦𝑛 + 12ℎ𝑘2 ), 𝑘4 = 𝑓(𝑥𝑛 + ℎ, 𝑦𝑛 + ℎ𝑘3 )

21
Example 3.1: Solve the initial-value problem 𝑦 ′ = 𝑥 + 𝑦, 𝑦(0) = 1,0 ≤ 𝑥 ≤ 0.5 with ℎ = 0.1 by
(i) Euler’s explicit method

(ii) Trapezoidal method

(iii) Two-step Adams-Moulton method

(iv) Fourth order Runge-Kutta method (work all computations correct to 4 decimal places and compare

your solutions with the exact solution 𝑦(𝑥) = 2𝑒 𝑥 − 𝑥 − 1).

Solution

(i) 𝑓 = 𝑥 + 𝑦, 𝑦0 = 1, ℎ = 0.1 ⇒ 𝑥0 = 0.1, 𝑥1 = 0.2, 𝑥3 = 0.3, 𝑥4 = 0.4, 𝑥5 = 0.5. Euler’s explicit

method 𝑦𝑛+1 = 𝑦𝑛 + ℎ𝑓𝑛 = 𝑦𝑛 + 0.1(𝑥𝑛 + 𝑦𝑛 ) ⇒ 𝑦𝑛+1 = 1.1𝑦𝑛 + 0.1𝑥𝑛 , 𝑛 = 0,1, … ,4

𝑛 = 0; 𝑦1 = 1.1𝑦0 + 0.1𝑥0 = 1.1

𝑛 = 1; 𝑦2 = 1.1𝑦1 + 0.1𝑥1 = 1.22

𝑛 = 2; 𝑦3 = 1.1𝑦2 + 0.1𝑥2 = 1.362

𝑛 = 3; 𝑦4 = 1.1𝑦3 + 0.1𝑥3 = 1.5282

𝑛 = 4; 𝑦5 = 1.1𝑦4 + 0.1𝑥4 = 1.7210

(ii) Trapezoidal method: 𝑦𝑛+1 − 𝑦𝑛 = ℎ2[𝑓𝑛+1 + 𝑓𝑛 ] = 𝑦𝑛 + ℎ2(𝑥𝑛+1 + 𝑦𝑛+1 + 𝑥𝑛 + 𝑦𝑛 )

2+ℎ ℎ 2.1 0.9


∴ 𝑦𝑛+1 = (2−ℎ) 𝑦𝑛 + 2−ℎ(𝑥𝑛+1 + 𝑥𝑛 ) = 1.9𝑦𝑛 + 1.9(𝑥𝑛+1 + 𝑥𝑛 ), 𝑛 = 0(1)4

2.1 0.9
𝑛 = 0; 𝑦1 = 1.9𝑦0 + 1.9(𝑥1 + 𝑥0 ) = 1.1105

2.1 0.9
𝑛 = 1; 𝑦2 = 1.9𝑦1 + 1.9(𝑥2 + 𝑥1 ) = 1.2432

2.1 0.9
𝑛 = 2; 𝑦3 = 1.9𝑦2 + 1.9(𝑥3 + 𝑥2 ) = 1.4004

2.1 0.9
𝑛 = 3; 𝑦4 = 1.9𝑦3 + 1.9(𝑥4 + 𝑥3 ) = 1.5847

2.1 0.9
𝑛 = 4; 𝑦5 = 1.9𝑦4 + 1.9(𝑥5 + 𝑥4 ) = 1.7989


(iii) Two-step Adams-Moulton method: 𝑦𝑛+2 − 𝑦𝑛+1 = 12 [5𝑓𝑛+2 + 8𝑓𝑛+1 − 𝑓𝑛 ]

12+8ℎ ℎ
∴ 𝑦𝑛+2 = (12−5ℎ) 𝑦𝑛+1 + 12−5ℎ(5𝑥𝑛+2 + 8𝑥𝑛+1 − 𝑥𝑛 − 𝑦𝑛 )

12.8 0.1
= 11.5𝑦𝑛+1 + 11.5(5𝑥𝑛+2 + 8𝑥𝑛+1 − 𝑥𝑛 − 𝑦𝑛 ), 𝑛 = 0(1)3

𝑦0 = 1, 𝑦1 = 1.1105 (from Trapezoidal method)

12.8 0.1
𝑛 = 0; 𝑦2 = 11.5𝑦1 + 11.5(5𝑥2 + 8𝑥1 − 𝑥0 − 𝑦0 ) = 1.2430

22
12.8 0.1
𝑛 = 1; 𝑦3 = 11.5𝑦2 + 11.5(5𝑥3 + 8𝑥2 − 𝑥1 − 𝑦1 ) = 1.3999

12.8 0.1
𝑛 = 2; 𝑦4 = 11.5𝑦3 + 11.5(5𝑥4 + 8𝑥3 − 𝑥2 − 𝑦2 ) = 1.5839

12.8 0.1
𝑛 = 3; 𝑦5 = 11.5𝑦4 + 11.5(5𝑥5 + 8𝑥4 − 𝑥3 − 𝑦3 ) = 1.7977

(iv) Fourth-order Runge-Kutta method 𝑦𝑛+1 − 𝑦𝑛 = ℎ6(𝑘1 + 2𝑘2 + 2𝑘3 + 𝑘4 )

𝑛 = 0; 𝑘1 = 𝑓(𝑥0 , 𝑦0 ) = 𝑥0 + 𝑦0 = 1

ℎ ℎ
𝑘2 = 𝑓 (𝑥0 + 2 , 𝑦0 + 2 𝑘1 ) = 𝑓(0.05,1.05) = 1.1

ℎ ℎ
𝑘3 = 𝑓 (𝑥0 + 2 , 𝑦0 + 2 𝑘2 ) = 𝑓(0.05,1.055) = 1.105

𝑘4 = 𝑓(𝑥0 + ℎ, 𝑦0 + ℎ𝑘3 ) = 𝑓(0.1,1.1105) = 1.2105


0.1
∴ 𝑦1 = 𝑦0 + 6
[1 + 2(1.1) + 2(1.105) + 1.2105] = 1.1103

𝑛 = 1; 𝑘1 = 𝑓(𝑥1 , 𝑦1 ) = 𝑓(0.1,1.1103) = 1.2103

ℎ ℎ
𝑘2 = 𝑓 (𝑥1 + 2 , 𝑦1 + 2 𝑘1 ) = 𝑓(0.15,1.1708) = 1.3208

ℎ ℎ
𝑘3 = 𝑓 (𝑥1 + 2 , 𝑦1 + 2 𝑘2 ) = 𝑓(0.15,1.1763) = 1.3263

𝑘4 = 𝑓(𝑥1 + ℎ, 𝑦1 + ℎ𝑘3 ) = 𝑓(0.2,1.2429) = 1.4429


0.1
∴ 𝑦2 = 𝑦1 + [1.2103 + 2(1.3208) + 2(1.3263) + 1.4429] = 1.2428
6

𝑛 = 2; 𝑘1 = 𝑓(𝑥2 , 𝑦2 ) = 𝑓(0.2,1.2428) = 1.4428

ℎ ℎ
𝑘2 = 𝑓 (𝑥2 + 2 , 𝑦2 + 2 𝑘1 ) = 𝑓(0.25,1.31494) = 1.5649

ℎ ℎ
𝑘3 = 𝑓 (𝑥2 + 2 , 𝑦2 + 2 𝑘2 ) = 𝑓(0.25,1.321047) = 1.5710

𝑘4 = 𝑓(𝑥2 + ℎ, 𝑦2 + ℎ𝑘3 ) = 𝑓(0.3,1.3999) = 1.6999


0.1
∴ 𝑦3 = 𝑦2 + [1.4428 + 2(1.5649) + 2(1.5710) + 1.6999] = 1.3997
6

𝑛 = 3; 𝑘1 = 𝑓(𝑥3 , 𝑦3 ) = 𝑓(0.3,1.3997) = 1.6997

ℎ ℎ
𝑘2 = 𝑓 (𝑥3 + 2 , 𝑦3 + 2 𝑘1 ) = 𝑓(0.35,1.4846) = 1.8346

ℎ ℎ
𝑘3 = 𝑓 (𝑥3 + 2 , 𝑦3 + 2 𝑘2 ) = 𝑓(0.35,1.4914) = 1.8414

𝑘4 = 𝑓(𝑥3 + ℎ, 𝑦3 + ℎ𝑘3 ) = 𝑓(0.4,1.5838) = 1.9838


0.1
∴ 𝑦4 = 𝑦3 + [1.6997 + 2(1.8346) + 2(1.8414) + 1.9838] = 1.5836
6

23
𝑛 = 4; 𝑘1 = 𝑓(𝑥4 , 𝑦4 ) = 𝑓(0.4,1.5836) = 1.9836

ℎ ℎ
𝑘2 = 𝑓 (𝑥4 + 2 , 𝑦4 + 2 𝑘1 ) = 𝑓(0.45,1.6828) = 2.1328

ℎ ℎ
𝑘3 = 𝑓 (𝑥4 + 2 , 𝑦4 + 2 𝑘2 ) = 𝑓(0.45,1.6902) = 2.1402

𝑘4 = 𝑓(𝑥4 + ℎ, 𝑦4 + ℎ𝑘3 ) = 𝑓(0.5,1.7976) = 2.2976


0.1
∴ 𝑦5 = 𝑦4 + [1.9836 + 2(2.1328) + 2(2.1402) + 2.2976] = 1.7974
6

Using the exact solution 𝑦(𝑥) = 2𝑒 𝑥 − 𝑥 − 1, we tabulate these results in Table 3.1(a) while the errors are in

Table 3.1(b)

Table 3.1(a): Numerical solution of 𝑦 ′ = 𝑥 + 𝑦, 𝑦(0) = 1,0 ≤ 𝑥 ≤ 0.5, ℎ = 0.1

X Y[Exact] Y[Euler] Y[Trapezoidal] Y[Adams] Y[RK]

0.0 1.0000 1.0000 1.0000 1.0000 1.0000

0.1 1.1103 1.1000 1.1105 1.1105 1.1103

0.2 1.2428 1.2200 1.2432 1.2430 1.2428

0.3 1.3997 1.3620 1.4004 1.3999 1.3997

0.4 1.5836 1.5282 1.5847 1.5839 1.5836

0.5 1.7974 1.7210 1.7989 1.7977 1.7974

Table 3.1(a): Error for the numerical solution of 𝑦 ′ = 𝑥 + 𝑦, 𝑦(0) = 1,0 ≤ 𝑥 ≤ 0.5, ℎ = 0.1

X Y[Euler] Y[Trapezoidal] Y[Adams] Y[RK]

0.0 0.0000 0.0000 0.0000 0.0000

0.1 1.030(-2) 2.000(-4) 2.000(-4) 0.0000

0.2 2.280(-2) 4.000(-4) 2.000(-4) 0.0000

0.3 3.377(-2) 7.000(-4) 2.000(-4) 0.0000

0.4 5.540(-2) 1.100(-3) 3.000(-4) 0.0000

0.5 7.640(-2) 1.500(-3) 3.000(-4) 0.0000

Notation: 𝑎(𝑏) ≔ 𝑎 × 10𝑏

24
Exercise 3.1: Consider the initial-value problem 𝑦 ′ = −𝑦 + 𝑥 + 1, 𝑦(0) = 1,0 ≤ 𝑥 ≤ 0.5. Solve the above by

(i) 4th order Runge-Kutta method

(ii) Midpoint method

(iii) Simpson’s method, with ℎ = 0.1 (work all computations correct to 4 decimal places and compare

your solutions with the exact solution).

4. PREDICTOR-CORRECTOR METHODS

The Modified Euler Method (sometimes called Heun’s Method)

Euler’s method is crude and rarely used in practice; a simple refinement, however, leads to the modified Euler’s

method (sometimes known as Heun’s method), which is the simplest of the class of predictor-corrector methods

for the numerical solution of first order equations.

Let us assume that the solution has already been carried as far as 𝑥 = 𝑥𝑛 , so that the next value of 𝑦 to be

calculated is 𝑦𝑛+1 . Integration of the equation 𝑦 ′ = 𝑓(𝑥, 𝑦) between the limits 𝑥𝑛 and 𝑥𝑛+1 leads to

𝑥𝑛+1 𝑥𝑛+1 𝑥𝑛+1


∫ 𝑦 ′ (𝑥)𝑑𝑥 = ∫ 𝑓(𝑥, 𝑦)𝑑𝑥 ⇒ 𝑦𝑛+1 − 𝑦𝑛 = ∫ 𝑓(𝑥, 𝑦)𝑑𝑥
𝑥𝑛 𝑥𝑛 𝑥𝑛

or
𝑥
𝑦𝑛+1 = 𝑦𝑛 + ∫𝑥 𝑛+1 𝑓(𝑥, 𝑦)𝑑𝑥 (4.1)
𝑛

Since we do not know the dependence of 𝑦 upon 𝑥, we cannot evaluate the integral unless 𝑓 happens to be a

function of 𝑥 alone. Instead, we approximate it using the trapezium rule, to obtain

1
𝑦𝑛+1 = 𝑦𝑛 + 2(𝑥𝑛+1 − 𝑥𝑛 )[𝑓(𝑥𝑛 , 𝑦𝑛 ) + 𝑓(𝑥𝑛+1 , 𝑦𝑛+1 )]

or since 𝑥𝑛+1 − 𝑥𝑛 = ℎ, the step length,

𝑦𝑛+1 = 𝑦𝑛 + ℎ2[𝑓(𝑥𝑛 , 𝑦𝑛 ) + 𝑓(𝑥𝑛+1 , 𝑦𝑛+1 )] (4.2)

Occasionally, this equation can be solved explicitly for 𝑦𝑛+1 (when the original differential equation is linear).


Generally, however, this is not possible, and the crude value 𝑦𝑛+1 given by Euler’s method is substituted in the

right-hand side of (4.2). The two equations used in the modified Euler’s method are then


𝑦𝑛+1 = 𝑦𝑛 + ℎ𝑓𝑛 (4.3)

25

𝑦𝑛+1 = 𝑦𝑛 + ℎ2[𝑓𝑛 + 𝑓𝑛+1 ] (4.4)

∗ ∗ ).
where 𝑓𝑛 ≡ 𝑓(𝑥𝑛 , 𝑦𝑛 ) and 𝑓𝑛+1 ≡ 𝑓(𝑥𝑛+1 , 𝑦𝑛+1

Equations (4.3) and (4.4) are known respectively as the predictor and the corrector, since the first provides a

rough estimate of 𝑦𝑛+1 , which is improved upon by the second. In fact, the application of the corrector amounts

to a single application of a simple iterative process, and further iterations will usually improve the value of 𝑦𝑛+1

still further. However, it is generally preferable, when using predictor-corrector methods, to achieve increased

accuracy by using a smaller step length rather than by multiple iteration of the corrector. This reduces the effect

of the truncation error.

Example 4.1: Find 𝑦(0.2) if 𝑦 ′ = −𝑦𝑥, 𝑦(0) = −1 using the modified Euler method with ℎ = 0.1 (compute to four

decimal places).

Solution
𝑥
𝑓𝑛 = −𝑦𝑛 , 𝑦0 = −1, ℎ = 0.1, ⇒ 𝑥0 = 0, 𝑥1 = 0.1, 𝑥2 = 0.2
𝑛


The predictor (4.3) gives 𝑦𝑛+1 = 𝑦𝑛 + 0.1 (−𝑦𝑥𝑛𝑛 ) , 𝑛 = 0,1

𝑥
𝑛 = 0, 𝑦1∗ = 𝑦0 + 0.1 (−𝑦0 ) = −1
0

(1)
The corrector (4.4) gives 𝑦𝑛+1 = 𝑦𝑛 + 0.1
2
[−𝑦𝑥𝑛𝑛 − −𝑦𝑥𝑛+1
∗ ]. Hence
𝑛+1

(1) 0.1 0.1


𝑛 = 0, 𝑦1 = −1 + 2
[1] = −0.9950

(2)
The second application of the corrector gives 𝑦𝑛+1 = 𝑦𝑛 + 0.1
2
[−𝑦𝑥𝑛𝑛 − −𝑥𝑛+1
(1) ] so that
𝑦𝑛+1

(2) 0.1 0.1


𝑛 = 0, 𝑦1 = −1 + [
2 0.9950
] = −0.9950 ∴ 𝑦1 ≅ 𝑦(0.1) = −0.9950

𝑥
𝑛 = 1, 𝑦2∗ = 𝑦1 + 0.1 (−𝑦1 ) = −0.9849
1

(1)
The first application of the corrector gives 𝑦2 = −0.9950 + 0.1 [ 0.1 + 0.9849
2 0.9950
0.2
] = −0.9798 and the second

(2)
application of the corrector gives 𝑦2 = −0.9950 + 0.1[ 0.1 + 0.9798
2 0.9950
0.2
] = −0.9798

∴ 𝑦(0.2) ≅ 𝑦2 = −0.9798.

26
Exercise 4.1: Solve the initial-value problem in Exercise 3.1 by using the predictor-corrector pair

𝑦𝑛+2 = 𝑦𝑛+1 + [3𝑓𝑛+1 + 𝑓𝑛 ]
2


𝑦𝑛+2 − 𝑦𝑛+1 = 12[5𝑓𝑛+2 + 8𝑓𝑛+1 − 𝑓𝑛 ]

with ℎ = 0.1,0 ≤ 𝑥 ≤ 0.5. Use 4th order Runge-Kutta method to obtain 𝑦1 .

5. CONTINUOUS FORMULATION OF DISCRETE LINEAR MULTISTEP METHOD

Let us consider the numerical solution of the ivp of the first order ordinary differential equation of the form

𝑦 ′ = 𝑓(𝑥, 𝑦), 𝑦(𝑎) = 𝑦0 , 𝑎 ≤ 𝑥 ≤ 𝑏 (5.1)

on a given mesh 𝑎 = 𝑥0 < 𝑥1 < 𝑥2 < ⋯ < 𝑥𝑛 < 𝑥𝑛+1 < ⋯ < 𝑥𝑁 = 𝑏, where ℎ = 𝑥𝑛+1 − 𝑥𝑛 , 𝑛 = 0,1, … , 𝑁 − 1

denotes a constant step size while 𝑘 denotes the step number of he required method.

The basic assumption on 𝑓 is that it satisfies the conditions stated in Theorem 1.1, which guarantee a unique

continuously differential solution 𝑦(𝑥) of (5.1).

A method used to obtain numerical values at each mesh or grid point 𝑥𝑛 is known as a discrete method. Despite

their simplicity and wide applicability, the discrete numerical integration algorithms have certain drawbacks

inherent in them. As such recent researches have yielded interesting results for continuous integration algorithms

through various approaches. It has also been observed that the continuous formulations of discrete methods tend

to possess certain striking features over their discrete counterparts.

For example, he continuous schemes can be used for further analytical work such as differentiation, integration,

etc with ease than the discrete schemes where the latter employ extra interpolatory work, numerical

differentiation and integration for such purposes. Moreover, they provide better estimates for the global errors.

The continuous schemes above all provide approximations at all interior points unlike the discrete ones.

Since we shall derive the continuous formulation of discrete LMMs by collocation, next we define a collocation

method.

Definition 5.1(Collocation Method)

A collocation method can simply be described as a method which involves the determination of an approximate

solution in a suitable set of functions, sometimes called trial or basis functions. Oftentimes, polynomials and

rational functions could serve as bases functions. The approximate solution is required to satisfy the differential
27
equation (5.1) and its supplementary conditions at certain points in the range of interest called the collocation

points.

The collocation methods by their very nature yield continuous solutions and the principle behind multistep

collocation is to allow the collocation polynomial use information from previous points in the integration.

Definition 5.2 (Interpolation and Collocation points)

A point at which the solution 𝑦(𝑥) of the function is evaluated is called an interpolation point while a point at

which its derivative is evaluated is called a collocation point,

Derivation Technique

We now consider the derivation of the multistep collocation method for constant step size and give recursive

expressions for the coefficients. The values of 𝑘 and 𝑚 are arbitrary except for collocation at the mesh points,

where 0 < 𝑚 ≤ 𝑘 + 1. Let 𝑦̅𝑛+𝑗 be approximations to 𝑦𝑛+𝑗 , where 𝑦𝑛+𝑗 ≡ 𝑦(𝑥𝑛+𝑗 ), 𝑗 = 0,1, … , 𝑘 − 1. Then a 𝑘-

step multistep collocation method with 𝑚 collocation points is constructed as follows.

We define a polynomial 𝑦̅(𝑥) of degree 𝑝 = 𝑡 + 𝑚 − 1, 𝑡 > 0, 𝑚 > 0 of the form

𝑡−1 𝑚−1

𝑦̅(𝑥) = ∑ 𝛼𝑗 (𝑥)𝑦(𝑥𝑛+𝑗 ) + ℎ ∑ 𝛽𝑗 (𝑥)𝑓(𝑥̅𝑗 , 𝑦̅ (𝑥𝑗 )) (5.2)


𝑗=0 𝑗=0

such that it satisfies the conditions

𝑦̅(𝑥𝑛+𝑗 ) = 𝑦𝑛+𝑗 = 𝑦(𝑥𝑛+𝑗 ) (5.3)

𝑦̅ ′ (𝑥̅𝑗 ) = 𝑓(𝑥̅𝑗 , 𝑦̅(𝑥̅𝑗 )) (5.4)

where 𝛼𝑗 (𝑥) and 𝛽𝑗 (𝑥) are assumed polynomials of the form

𝑡+𝑚−1 𝑡+𝑚−1

𝛼𝑗 (𝑥) = ∑ 𝛼𝑗,𝑖+1 𝑥 ; ℎ𝛽𝑗 (𝑥) = ℎ ∑ 𝛽𝑗,𝑖+1 𝑥 𝑖


𝑖
(5.5)
𝑖=0 𝑖=0

𝑗 ∈ {0,1, … , 𝑡 − 1}; 𝑗 = 0,1, … , 𝑚 − 1

In (5.2), 𝑥𝑛+𝑗 are 𝑡 (0 < 𝑡 ≤ 𝑘) arbitrary chosen interpolation points taken from {𝑥𝑛 , 𝑥𝑛+1 , … , 𝑥𝑛+𝑘−1 } and the

collocation points 𝑥̅𝑗 , 𝑗 = 0,1, … , 𝑚 − 1 belong to [𝑥𝑛 , 𝑥𝑛+𝑘 ].

28
From the interpolation conditions (5.3), (5.4) and the expression for 𝑦̅(𝑥) in (5.2), the following conditions are

imposed on 𝛼𝑗 (𝑥) and 𝛽𝑗 (𝑥):

𝛼𝑗 (𝑥𝑛+𝑖 ) = 𝛿𝑖𝑗 , 𝑗 = 0(1)𝑡 − 1; 𝑖 = 0(1)𝑡 − 1


} (5.6)
ℎ𝛽𝑗 (𝑥𝑛+𝑖 ) = 0, 𝑗 = 0(1)𝑚 − 1; 𝑖 = 0(1)𝑡 − 1

and

𝛼𝑗′ (𝑥̅𝑖 ) = 0, 𝑗 = 0(1)𝑡 − 1; 𝑖 = 0(1)𝑚 − 1


} (5.7)
ℎ𝛽𝑗′ (𝑥̅𝑖 ) = 𝛿𝑖𝑗 , 𝑗 = 0(1)𝑚 − 1; 𝑖 = 0(1)𝑚 − 1

Next we write (5.6) and (5.7) in a matrix equation of the form

𝐷𝐶 = 𝐼 (5.8)

where

1 𝑥𝑛 𝑥𝑛2 … 𝑥𝑛𝑡+𝑚−1
1 𝑥𝑛+1 2 𝑡+𝑚−1
𝑥𝑛+1 … 𝑥𝑛+1
⋮ ⋮
2 𝑡+𝑚−1
𝐷= 1 𝑥𝑛+𝑡−1 𝑥𝑛+𝑡−1 𝑥𝑛+𝑡−1

… (𝑡 + 𝑚 − 1)𝑥̅ 𝑡+𝑚−2
0 1 2𝑥̅0 0
⋮ ⋮
[0 1 2𝑥̅𝑚−1 … (𝑡 + 𝑚 − 1)𝑥̅𝑚−1𝑡+𝑚−2
]

𝛼01 𝛼11 … 𝛼𝑡−1,1 ℎ𝛽01 … ℎ𝛽𝑚−1,1


𝛼02 𝛼12 … 𝛼𝑡−1,2 ℎ𝛽02 … ℎ𝛽𝑚−1,2
𝐶=
⋮ ⋮
[𝛼0,𝑡+𝑚 𝛼1,𝑡+𝑚 … 𝛼𝑡−1,𝑡+𝑚 ℎ𝛽0,𝑡+𝑚 … ℎ𝛽𝑚−1,𝑡+𝑚 ]

and 𝐼 is the identity matrix of dimension (𝑡 + 𝑚) × (𝑡 + 𝑚). The matrices, 𝐷 and 𝐶, are also of the same

dimensions. It follows from (5.8) that the columns of 𝐶 = 𝐷 −1, give the continuous coefficients 𝛼𝑗 (𝑥) and 𝛽𝑗 (𝑥).

Specification of the Methods

1. The Backward Differentiation Formulae (BDF) Methods

In this method, 𝑚 = 1, 𝑡 = 𝑘, 𝑥̅0 = 𝑥𝑛+𝑘 , 𝛽0 (𝑥) = 𝛽𝑘 (𝑥) and (5.2) becomes

𝑦̅(𝑥) = ∑𝑘−1
𝑗=0 𝛼𝑗 (𝑥)𝑦𝑛+𝑗 + ℎ𝛽𝑘 (𝑥)𝑓𝑛+𝑘 (5.9)

Then (5.8) takes the following form

1 𝑥𝑛 𝑥𝑛2 … 𝑥𝑛𝑘 𝛼01 … 𝛼𝑘−1,1 ℎ𝛽𝑘1


1 𝑥𝑛+1 2 𝑘 𝛼𝑘−1,2 ℎ𝛽𝑘2
𝑥𝑛+1 … 𝑥𝑛+1 𝛼02 …
𝐷𝐶 = ⋮ ⋮ ⋮ ⋮ =𝐼 (5.10)
𝑘
1 𝑥𝑛+𝑘−1 2
𝑥𝑛+𝑘−1 … 𝑥 𝑛+𝑘−1 𝛼0𝑘 … 𝛼𝑘−1,𝑘 ℎ𝛽𝑘𝑘
[0 1 2𝑥𝑛+𝑘 … 𝑘𝑥 𝑘−1 [𝛼
𝑛+𝑘 ] 0𝑘+1 … 𝛼𝑘−1,𝑘+1 ℎ𝛽𝑘,𝑘+1 ]

29
and it follows from (5.10) that the columns of 𝐶 = 𝐷 −1 give the continuous coefficients 𝛼𝑗 (𝑥), 𝑗 = 0,1, … , 𝑘 − 1

and 𝛽𝑘 (𝑥).

Case 𝑘 = 1.

When 𝑘 = 1, (5.9) becomes

𝑦̅(𝑥) = 𝛼0 (𝑥)𝑦𝑛 + ℎ𝛽1 (𝑥)𝑓𝑛+1 (5.11)

and (5.10) reduces to

1 𝑥𝑛 𝛼01 ℎ𝛽11
𝐷𝐶 = [ ][ ]=𝐼
0 1 𝛼02 ℎ𝛽12

1 −𝑥𝑛 𝛼 ℎ𝛽11
Now 𝐶 = 𝐷 −1 ⇒ [ ] = [ 01 ] and from (5.5), we have
0 1 𝛼02 ℎ𝛽12

𝛼0 (𝑥) = ∑ 𝛼0,𝑖+1 𝑥 𝑖 = 𝛼01 + 𝛼02 𝑥 = 1


𝑖=0

ℎ𝛽1 (𝑥) = ℎ ∑ 𝛽1,𝑖+1 𝑥 𝑖 = ℎ𝛽11 + ℎ𝛽12 𝑥 = −𝑥𝑛 + 𝑥 = 𝑥 − 𝑥𝑛


𝑖=0

Substituting the above continuous coefficients into (5.11), we obtain the following continuous method

𝑦̅(𝑥) = 𝑦𝑛 + (𝑥 − 𝑥𝑛 )𝑓𝑛+1

and at 𝑥 = 𝑥𝑛+1, we obtain the following discrete method

𝑦̅(𝑥𝑛+1 ) = 𝑦𝑛 + (𝑥𝑛+1 − 𝑥𝑛 )𝑓𝑛+1

or

𝑦𝑛+1 = 𝑦𝑛 + ℎ𝑓𝑛+1

this is the same as Euler’s implicit method.

2. The Adams-Moulton Methods

In this method, 𝑚 = 𝑘 + 1, 𝑡 = 1, 𝑥̅𝑗 = 𝑥𝑛+𝑗 , 𝑗 = 0,1, … , 𝑘 and 𝛼0 (𝑥) = 𝛼𝑘−1 (𝑥). Then (5.2) and (5.8) become

respectively

𝑦̅(𝑥) = 𝛼𝑘−1 (𝑥)𝑦𝑛+𝑘−1 + ℎ ∑ 𝛽𝑗 (𝑥)𝑓𝑛+𝑗 (5.12)


𝑗=0

and

30
1 𝑥𝑛+𝑘−1 2 𝑘+1 𝛼𝑘−1,1 ℎ𝛽01 … ℎ𝛽𝑘1
𝑥𝑛+𝑘−1 … 𝑥𝑛+𝑘−1
0 1 2𝑥𝑛 … (𝑘 + 1)𝑥𝑛 𝑘 𝛼𝑘−1,2 ℎ𝛽02 … ℎ𝛽𝑘2
𝐷𝐶 = =𝐼 (5.13)
⋮ ⋮ ⋮ ⋮ ⋮
𝑘 𝛼
[0 1 2𝑥𝑛+𝑘 … (𝑘 + 1)𝑥𝑛+𝑘 ] [ 𝑘−1,𝑘+2 ℎ𝛽0𝑘+2 … ℎ𝛽𝑘,𝑘+2 ]

Case 𝑘 = 1

In this case, (5.12) and (5.13) become respectively

𝑦̅(𝑥) = 𝛼0 (𝑥)𝑦𝑛 + ℎ𝛽0 (𝑥)𝑓𝑛 + ℎ𝛽1 (𝑥)𝑓𝑛+1 (5.14)

and

1 𝑥𝑛 𝑥𝑛2 𝛼01 ℎ𝛽01 ℎ𝛽11


𝐷𝐶 = [0 1 2𝑥𝑛 ] [𝛼02 ℎ𝛽02 ℎ𝛽12 ] = 𝐼
0 1 2𝑥𝑛+1 𝛼03 ℎ𝛽03 ℎ𝛽13

The continuous coefficients of the method are

2ℎ(𝑥−𝑥𝑛 )−(𝑥−𝑥𝑛 )2 (𝑥−𝑥𝑛 )2


𝛼0 (𝑥) = 1, ℎ𝛽0 (𝑥) = 2ℎ
, ℎ𝛽1 (𝑥) = 2ℎ

Substituting these continuous coefficients into (5.14), we obtain the following continuous method

2ℎ(𝑥−𝑥𝑛 )−(𝑥−𝑥𝑛 )2 (𝑥−𝑥𝑛 )2


𝑦̅(𝑥) = 𝑦𝑛 + [ 2ℎ
] 𝑓𝑛 +[ 2ℎ
]𝑓𝑛+1

from which at 𝑥 = 𝑥𝑛+1 , we obtain

2ℎ(𝑥𝑛+1 −𝑥𝑛 )−(𝑥𝑛+1 −𝑥𝑛 )2 (𝑥𝑛+1 −𝑥𝑛 )2


𝑦̅(𝑥𝑛+1 ) = 𝑦𝑛 + [ 2ℎ
] 𝑓𝑛 +[ 2ℎ
]𝑓𝑛+1

or


𝑦𝑛+1 − 𝑦𝑛 = 2 [𝑓𝑛 + 𝑓𝑛+1 ]

3. The Adams-Bashforth Method

In this method, 𝑚 = 𝑘, 𝑡 = 1, 𝑥̅𝑗 = 𝑥𝑛+𝑗 , 𝑗 = 0,1, … , 𝑘 and 𝛼0 (𝑥) = 𝛼𝑘−1 (𝑥). Thus (5.2) and (5.8) become

respectively

𝑘−1

𝑦̅(𝑥) = 𝛼𝑘−1 (𝑥)𝑦𝑛+𝑘−1 + ℎ ∑ 𝛽𝑗 (𝑥)𝑓𝑛+𝑗 (5.15)


𝑗=0

and

1 𝑥𝑛+𝑘−1 2 𝑘 𝛼𝑘−1,1 ℎ𝛽01 … ℎ𝛽𝑘−1,1


𝑥𝑛+𝑘−1 … 𝑥𝑛+𝑘−1
0 1 𝑘−1
2𝑥𝑛 … 𝑘𝑥𝑛 𝛼𝑘−1,2 ℎ𝛽02 … ℎ𝛽𝑘−1,2
𝐷𝐶 = =𝐼 (5.16)
⋮ ⋮ ⋮ ⋮ ⋮
𝑘−1
[0 1 2𝑥𝑛+𝑘−1 … 𝑘𝑥𝑛+𝑘−1 ] [𝛼𝑘−1,𝑘+1 ℎ𝛽0𝑘+1 … ℎ𝛽𝑘−1,𝑘+1 ]
31
Exercise 5.1: Obtain the continuous method and its discrete formulation at 𝑥 = 𝑥𝑛+1 for the Adams-Bashforth

method in the case 𝑘 = 1.

6. LINEAR MULTISTEP METHODS FOR A SPECIAL CLASS OF SECOND ORDER

ODEs

Consider the following second order initial-value problem

𝑦 ′′ = 𝑓(𝑥, 𝑦), 𝑦(𝑎) = 𝜂, 𝑦 ′ (𝑎) = 𝜂̂ (6.1)

For the solution of the above ivp, we consider linear 𝑘-step methods of the form

𝑘 𝑘

∑ 𝛼𝑗 𝑦𝑛+𝑗 = ℎ2 ∑ 𝛽𝑗 𝑓𝑛+𝑗 (6.2)


𝑗=0 𝑗=0

where 𝛼𝑘 = 1 and 𝛼0 and 𝛽0 do not both vanish.

Order, Consistency and Zero-stability

With the LMM (6.2), we associate the linear difference operator defined by

ℒ[𝑦(𝑥); ℎ] = ∑[ 𝛼𝑗 𝑦(𝑥 + 𝑗ℎ) − ℎ2 𝛽𝑗 𝑦 ′′ (𝑥 + 𝑗ℎ)]


𝑗=0

where 𝑦(𝑥) is an arbitrary function, continuously differentiable on an interval [𝑎, 𝑏]. If we assume that 𝑦(𝑥) has

as many higher derivatives as we require, then on Taylor expanding about the point 𝑥, we obtain

ℒ[𝑦(𝑥); ℎ] = 𝑐0 𝑦(𝑥) + 𝑐1 ℎ𝑦 ′ (𝑥) + … + 𝑐𝑞 ℎ𝑞 𝑦 (𝑞) (𝑥) + ⋯

where

𝑐0 = 𝛼0 + 𝛼1 + ⋯ + 𝛼𝑘

𝑐1 = 𝛼1 + 2𝛼2 + ⋯ + 𝑘𝛼𝑘
1
𝑐2 = 2!(𝛼1 + 22 𝛼2 + ⋯ + 𝑘 2 𝛼𝑘 ) − (𝛽0 + 𝛽1 + ⋯ + 𝛽𝑘 )

1
𝑐3 = 3!(𝛼1 + 23 𝛼2 + ⋯ + 𝑘 3 𝛼𝑘 ) − (𝛽1 + 2𝛽2 + ⋯ + 𝑘𝛽𝑘 )


1 1
𝑐𝑞 = 𝑞!(𝛼1 + 2𝑞 𝛼2 + ⋯ + 𝑘 𝑞 𝛼𝑘 ) − (𝑞−2)!(𝛽1 + 2𝑞−2 𝛽2 + ⋯ + 𝑘 𝑞−2 𝛽𝑘 ), 𝑞 = 3,4, …

32
Then the method has order 𝑝 if 𝑐0 = 0, 𝑐1 = 0, … , 𝑐𝑝 = 0, 𝑐𝑝+1 = 0 but 𝑐𝑝+2 ≠ 0 and 𝑐𝑝+2 is called the error

constant and 𝑐𝑝+2 ℎ𝑝+2 𝑦 (𝑝+2) (𝑥𝑛 ) is the principal local truncation error at the point 𝑥𝑛 .

The method is said to be consistent if it has order at least one. If we define the first and second characteristic

polynomials 𝜌(𝜉) = ∑𝑘𝑗=0 𝛼𝑗 𝜉𝑗 and 𝜎(𝜉) = ∑𝑘𝑗=0 𝛽𝑗 𝜉 𝑗 respectively, then the method is consistent if and only if

𝜌(1) = 𝜌′ (1) = 0; 𝜌′′ (1) = 2𝜎(1).

The polynomial 𝜌(𝜉) associated with a consistent method thus has a double root at ; this is the principal root,

the other roots being spurious.

Definition 6.1 (Zero-Stability)

The LMM (6.2) is said to be zero-stable if no root of the first characteristic polynomial 𝜌(𝜉) has modulus greater

than one, and if every root of modulus one has multiplicity not greater than two.

The family of methods (6.2)for which 𝜌(𝜉) = 𝜉 𝑘 − 2𝜉 𝑘−1 + 𝜉 𝑘−2 are frequently referred to as Stormer-Cowell

methods. The best known such methods is the optimal two-step method

ℎ2
𝑦𝑛+2 − 2𝑦𝑛+1 + 𝑦𝑛 = 12[𝑓𝑛+2 + 10𝑓𝑛+1 + 𝑓𝑛 ]

known as the Numerov’s method.

Definition 6.2 (Absolute Stability)

The LMM (6.2) is said to be absolutely stable for a given ℎ̅ if for all that ℎ̅, all the roots 𝑟𝑠 of the stability

polynomial (𝜋(𝑟; ℎ̅) = 𝜌(𝑟) − ℎ̅𝜎(𝑟), where ℎ̅ = ℎ2 𝜆) satisfy |𝑟𝑠 | ≤ 1, 𝑠 = 1,2, … , 𝑘.

An interval [𝛼, 𝛽] of the real line is said to be an interval of absolute stability if the method is absolutely stable for

all ℎ̅ ∈ [𝛼, 𝛽].

Exercise 6.1: Find the order, error constant and the interval of absolute stability of the Numerov method.
2
Solution: 𝑦𝑛+2 − 2𝑦𝑛+1 + 𝑦𝑛 = ℎ12 [𝑓𝑛+2 + 10𝑓𝑛+1 + 𝑓𝑛 ], 𝛼0 = 1, 𝛼1 = −2, 𝛼2 = 1, 𝛽0 = 12
1
, 𝛽1 = 10
12
1
, 𝛽2 = 12

𝑐0 = 𝛼0 + 𝛼1 + 𝛼2 = 1 − 2 + 1 = 0

𝑐1 = 𝛼1 + 2𝛼2 = −2 + 2 = 0
1
𝑐2 = 2(𝛼1 + 4𝛼2 ) − (𝛽0 + 𝛽1 + 𝛽2 ) = 1 − 1 = 0

1 1
𝑐3 = 6(𝛼1 + 8𝛼2 ) − 2(𝛽1 + 2𝛽2 ) = 1 − 1 = 0

33
1 1 14 14
𝑐4 = 24(𝛼1 + 16𝛼2 ) − 2(𝛽1 + 4𝛽2 ) = 24−24 = 0

1 1 1 1
𝑐5 = 120(𝛼1 + 32𝛼2 ) − 6(𝛽1 + 8𝛽2 ) = 4−4 = 0

1 1 31 42 1
𝑐6 = 720(𝛼1 + 64𝛼2 ) − 24(𝛽1 + 16𝛽2 ) = 360−288 = −240

1
Hence Numerov method is of order 𝑝 = 4 with error constant 𝑐6 = −240.

𝜌(𝑟) 1
ℎ̅ = 𝜎(𝑟), ̅ =ℎ2 𝜆, 𝜌(𝑟)
𝑟=𝑒 𝑖𝜃 , 0°≤𝜃≤180°, ℎ = 𝑟 2 − 2𝑟 + 1, 𝜎(𝑟) = 12(𝑟2 +10𝑟+1)

12𝑟 2 − 24𝑟 + 12 12𝑒 2𝑖𝜃 − 24𝑒 𝑖𝜃 + 12


∴ ℎ̅ = =
𝑟 2 + 10𝑟 + 1 𝑒 2𝑖𝜃 + 10𝑒 𝑖𝜃 + 1

(12 cos 2𝜃 − 24 cos 𝜃 + 12) + 𝑖(12 sin 2𝜃 − 24 sin 𝜃)


∴ ℎ̅(𝜃) = ; and ℎ̅ = 𝑥(𝜃) + 𝑖𝑦(𝜃) gives
(cos 2𝜃 + 10 cos 𝜃 + 1) + 𝑖(sin 2𝜃 + 10 sin 𝜃)

192 cos 𝜃 + 24 cos 2𝜃 − 216


𝑥(𝜃) =
(cos 2𝜃 + 10 cos 𝜃 + 1)2 + (sin 2𝜃 + 10 sin 𝜃)2

⇒ 𝑥(0°) = 0 and 𝑥(180°) = −6

Hence the interval of absolute stability of Numerov’s method is [−6,0].

Continuous Assessment Test (Attempt all questions. Time: 1½Hours).


1(a) State Schur’s theorem.

(b) Use Schur’s method to obtain the absolute stability interval of Euler’s implicit method.

2(a) Derive the 3-step Adams-Moulton method through numerical integration.

(b) Find the order and error constant of the method derived in 2(a) above.

(c) Obtain the interval of absolute stability of the method in 2(a) using the boundary locus method.

3 Find the stiffness ratio of the following system:

𝑦1′ = 𝑦2

𝑦2′ = −100𝑦1 − 101𝑦2

34

You might also like