Derivation of Cubic Splines: [adapted from Chapra]

fi  ai  bi  x  xi   ci  x  xi   di  x  xi 
2 3
Equation for Cubic Splines:

Now, Let us consider 4 points then there will be 3 splines and 2knots.

Thus, for n points, there will be (n-1) splines and (n-2) knots.

Number of unknowns therefore is = 4(n-1) ; [4 for ai , bi , ci , d i ]

Hence, We need to have 4(n-1) equations to solve for the unknowns. In the following context an
example of 3 splines and then it is generalized for n points.

Equations for the splines may be written as follows:

S1  x   a1  b1  x  x1   c1  x  x1   d1  x  x1 
2 3
For n=4;

S2  x   a2  b2  x  x2   c2  x  x2   d2  x  x2 
2 3

S3  x   a3  b3  x  x3   c3  x  x3   d3  x  x3 
2 3

Si  x   ai  bi  x  xi   ci  x  xi   di  x  xi 
2 3
General : (1)

First Condition : Splines pass through all the data points.

For n=4;

S1 passes through  x1 , y1  : S1  x1   a1  b1  x1  x1   c1  x1  x1   d1  x1  x1 
2 3

or, S1  x1   a1  f1

S 2 passes through  x2 , y2  : S2  x2   a2  f 2

S3 passes through  x3 , y3  : S3  x3   a3  f3

General : Si  xi   ai  fi (2)

Again,

S3 passes through  x4 , y4  : S3  x4   a3  b3  x4  x3   c3  x4  x3   d3  x4  x3 
2 3

Ie., Sn 1 passes through  xn , yn  :

Sn1  xn   an1  bn1  xn  xn1   cn1  xn  xn1   dn1  xn  xn 1 

2 3
General: (3)
Second Condition : Second Derivative continuity : Second derivatives at the knots are equal.

Si  x   bi  2ci  xi 1  xi   3di  xi 1  xi 

2
Equation (1) can be differentiated twice to give :

Si  x   2ci  6di  xi 1  xi 

or, Si  x   2ci  6di hi (4)

For n=4 : S1  x  | x2 , y2   S2  x  | x2 , y2 

2c1  6d1  x2  x1   2c2  6d2  x2  x2 

2c1  6d1h1  2c2

S2  x  | x3 , y3   S3  x  | x3 , y3 

2c2  6d2 h2  2c3

General : Si  x  | xi1 , yi1   Si 1  x  | xi1 , yi1 

2ci  6di hi  2ci 1

ci  3di hi  ci 1 (5)

ci 1  ci
or, di  (6)
3hi

Third Condition: First derivative continuity : First derivatives at the knots are equal.

Si  x   bi  2ci  xi 1  xi   3di  xi 1  xi 

2
Equation (1) can be differentiated to give : (7)

or, Si  x   bi  2ci hi  3di hi 2

For n=4 : S1  x  | x2 , y2   S2  x  | x2 , y2 

b1  2c1  x2  x1   3d1  x2  x1   b2  2c2  x2  x2   3d2  x2  x2 

2 2

b1  2c1h1  3d1h12  b2

S2  x  | x3 , y3   S3  x  | x3 , y3 

b2  2c2 h2  3d2 h2 2  b3
General : Si  x  | xi1 , yi1   Si 1  x  | xi1 , yi1 

bi  2ci hi  3di hi 2  bi 1 (8)

equation (6) can be substituted in equation (8) to get :

bi 1  bi  hi  ci  ci 1  (9)

Reducing the index of the equation by 1 we get :

bi  bi 1  hi 1  ci 1  ci 

bi  bi 1  hi 1  ci 1  ci  (10)

Fourth Condition : Continuity : Splines are continuous at the knots, ie.,

For n=4 ; S1  x  | x2 , y2   S2  x  | x2 , y2 

a1  b1  x2  x1   c1  x2  x1   d1  x2  x1   a2  b2  x2  x2   c2  x2  x2   d2  x2  x2 
2 3 2 3

a1  b1  x2  x1   c1  x2  x1   d1  x2  x1   a2
2 3

S2  x  | x3 , y3   S3  x  | x3 , y3 

a2  b2  x3  x2   c2  x3  x2   d2  x3  x2   a3  b3  x3  x3   c3  x3  x3   d3  x3  x3 
2 3 2 3

a2  b2  x3  x2   c2  x3  x2   d2  x3  x2   a3
2 3

General : Si  x  | xi1 , yi1   Si 1  x  | xi1 , yi1 

ai  bi  xi 1  xi   ci  xi 1  xi   di  xi 1  xi   ai 1
2 3

From Equation (2) we may write,

fi 1  fi  bi  xi 1  xi   ci  xi 1  xi   di  xi 1  xi 
2 3

Now, hi  xi 1  xi ;

Therefore , fi 1  fi  bi hi  ci hi 2  di hi 3 (11)

equation (6) can be substituted in equation (11) to get :

  c c  3
fi 1  fi  bi hi  ci hi 2   i 1 i  hi
 3hi 

hi 2
fi 1  fi  bi hi   ci 1  2ci 
3

fi 1  fi hi
bi    ci 1  2ci  (12)
hi 3

Reducing the index of the equation by 1 we get :

fi  fi 1 hi 1
bi 1    ci  2ci 1  (13)
hi 1 3

Substracting equation (13) from (12) we get ;

f f h  f f h 
bi  bi 1   i 1 i  i  ci 1  2ci    i i 1  i 1  ci  2ci 1  (14)
 hi 3   hi 1 3 

Equating equations (11) and (14) we get :

f f h  f f h 
hi 1  ci 1  ci    i 1 i  i  ci 1  2ci    i i 1  i 1  ci  2ci 1 
 hi 3   hi 1 3 

hi 1 h f f f f
 ci  2ci 1   hi 1  ci 1  ci   i  ci 1  2ci   i i 1  i 1 i
3 3 hi 1 hi

fi 1  fi f  fi 1
hi 1ci 1  2  hi 1  hi  ci  hi ci 1  3 3 i (15)
hi hi 1

The equation can be made a little more concise by recognizing that the terms on the right
fi  f j
hand side of equation (15) are finite differences, f  xi , x j  
xi  x j

Therefore, equation (15) may be re-written

hi 1ci 1  2  hi 1  hi  ci  hi ci 1  3  f  xi 1 , xi   f  xi , xi 1  (16)

Equation (15) can be written for the Knots, i=2,3,……,(n-1),which results in (n-2) simultaneous
tridiagonal equations with n unknown coefficients, c1 , c2 , …… , cn . Therefore if we now have two
additional conditions, we can solve for the c ’s.

fi 1  fi hi
Once this is done, b ’s can be determined from equation (12) : bi    ci 1  2ci 
hi 3
 ci 1  ci
And d ’s can be determined from equation (6) : di 
3hi

Cubic splines and End Conditions :

1. Natural Condition : In this condition, the second derivative at the first and the last node is
equal to zero.
Si  x   ai  bi  x  xi   ci  x  xi   di  x  xi 
2 3
From equation (1):

Si  x   bi  2ci  x  xi   3di  x  xi 

2

Si  x   2ci  6di  x  xi 

For the first node : S1  x1   0  2c1  6d1  x1  x1  (17)

Therefore , c1  0 (18)

For the last node : Sn1  xn   2cn1  6dn1  xn  xn1   0

cn1  3dn1hn1  0

We can conveniently define an extraneous parameter cn , in which equation (18) becomes

cn1  6dn1hn1  cn  0
cn  0 (19)

For n points : equation (16) can be written as :

h1c1  2  h1  h2  c1  h2c2  3  f  x3 , x2   f  x2 , x1 

h2c2  2  h2  h3  c2  h3c3  3  f  x4 , x3   f  x3 , x2 

hn2cn2  2  hn2  hn1  cn1  hn1cn  3  f  xn , xn1   f  xn1 , xn2 

Hence the final equations may be written in matrix form as :

For n=4 points,

1 0 0 0   c1   0 
 
h
 1 2  h1  h2  h2 0  c2  3  f  x3 , x2   f  x2 , x1  
  
0 h2 2  h2  h3  h3  c3  3  f  x4 , x3   f  x3 , x2 
 
0 0 0 1  c4   0 

For n points,

1 0 0   c1   0 
h  c   3  f  x3 , x2   f  x2 , x1  
 1 2  h1  h2  h2 0 0 0   2 
0 h2 2  h2  h3  h3 0 0     3  f  x4 , x3   f  x3 , x2  
     
 0 0    
 0 0 0     
    
 0 0 0 hn  2 2  hn  2  hn 1  hn 1  cn 1  3  f  xn , xn 1   f  xn 1 , xn  2  
0    
 0 1   cn  
 0 

As shown, the system is tridiagonal and hence efficient to solve.

2. Clamped Condition : in this condition f1 and f n are specified at the first and last node.
Hence,
Si  x   ai  bi  x  xi   ci  x  xi   di  x  xi 
2 3
From equation (1) :

Si  x   bi  2ci  x  xi   3di  x  xi 

2

S1  x   b1  2c1  x  x1   3d1  x  x1 

2
For the first node :

S1  x  | x1 , y1   f1  b1  2c1  x1  x1   3d1  x1  x1 

2

S1  x  | x1 , y1   f1  b1

fi 1  fi hi
Substituting the above equation in equation (12): bi    ci 1  2ci  , we get ,
hi 3

f 2  f1 h1
f1    2c2  c1 
h1 3

f 2  f1 h
 f1  1  c2  2c1 
h1 3

c2 h1  2h1c1  3 f  x2 , x1   3 f1

For Last node : Sn1  x   bn1  2cn1  x  xn1   3dn1  x  xn1 

2
 Sn1  x  | xn , yn   f n  bn1  2cn1  xn  xn1   3d n1  xn  xn1 
2

Sn1  x  | xn , yn   f n  bn1  2cn1hn1  3dn1hn12

Again from equation (8)we know, bi  2ci hi  3di hi 2  bi 1

Hence, f n  bn

From equation (9) we know , bi 1  bi  hi  ci  ci 1 

bn  bn1  hn1  cn1  cn 

f n  bn1  hn1  cn1  cn 

fi  fi 1 hi 1
From equation (13) we know, bi 1    ci  2ci 1 
hi 1 3

f n  f n 1 hn 1
bn 1    cn  2cn1 
hn 1 3

f  f n 1 hn 1
f n  hn 1  cn 1  cn   n   cn  2cn1 
hn1 3

h
f n  hn1  cn1  cn   f  xn , xn 1   n1  cn  2cn1 
3


2hn1cn1  hn1cn  3 f n  f  xn , xn 1  
The final equations can now be written in a matrix form as :

 3 f  x2 , x1   3 f1 
 2h1 h1 0 0   c1   
h 2  h1  h2  h2 0 0 0   c   3  f  x3 , x2   f  x2 , x1  
 1  2   
 0 h2 2  h2  h3  h3 0 0     3  f  x4 , x3   f  x3 , x2  
    
 0 0    
 0 0 0     
     
 0 0 0 hn  2 2  hn  2  hn 1  hn 1  cn 1  3  f  xn , xn 1   f  xn 1 , xn  2  
hn 1   cn   
 0
 2hn 1
 
3 f n  f  xn , xn 1   


3. Not –a- knot condition : The third derivative at the second and next to last knots are
continuous, therefore,
 Si  x   ai  bi  x  xi   ci  x  xi   di  x  xi 
2 3

Si  x   2ci  6di  xi 1  xi 

Si  x   6di

S1  x  | x2 , y2   S2  x  | x2 , y2 

6d1  6d2 or, d1  d 2

ci 1  ci
From equation (6) we know , di 
3hi

c2  c1 c3  c2

3h1 3h2

h2c2  h2c1  h1c3  h1c2

h2c1   h1  h2  c2  h1c3  0

For last but one node : Sn2  x  | xn1 , yn1   Sn1  x  | xn1 , yn1 

dn1  dn2

cn 1  cn 2 cn  cn 1

3hn 2 3hn 1

hn1cn1  hn1cn2  hn2cn  hn2cn1

hn1cn2   hn2  hn1  cn1  hn2cn  0

 h2   h1  h2  h1 0 0   c1   0 
    3  f  x3 , x2   f  x2 , x1  
 h1 2  h1  h2  h2 0 0 0   c2  
0 h2 2  h2  h3  h3 0 0     3  f  x4 , x3   f  x3 , x2  
     
 0 0     
 0 0 0     
    
 0 0 0 hn  2 2  hn  2  hn 1  hn 1  cn 1  3  f  xn , xn 1   f  xn 1 , xn  2  
0
 hn 1   hn  2  hn 1  hn  2   cn   
 0 

Thus for the above mentioned conditions the Splines may be determined.
GENERAL LINEAR LEAST SQUARES (adapted from Mathews & Finks)

Simple linear, polynomial and multi linear all three belongs to the following general linear least
squares model:

(1)
y  a0 z 0  a1 z1  a2 z 2  a3 z 3  a4 z 4  a5 z 5 .  ........am z m  e
It can easily be seen how simple linear and multi linear regression fall within this model ie
z 0  1, z1  x1 , z 2  x2 , , z m  xm . Polynomial regression is also included if the basis functions are
z 0  1, z1  x1 , z 2  x2 , , z m  xm

‘n’ number of data points  Z k yk  and z 0 , z1 , z 2 , z 3 , z 4 , z 5 , , z m are  m  1 independent

functions. So for k th deta point the equation is:

yk  a0 zk0  a1 z1k  a2 zk2  a3 zk3  a4 zk4  a5 zk5  ........am zkm  ek

IF coefficient vector is {a} then

 y   Z a  e
(2)
 Z  '  Z  {a}  [ Z ]'{ y}   Z  'e

Where

 Z10 Z11 Z12 Z1m 

 0 
Z2 Z 21 Z 22 Z 2m 
 Z    Z30 Z 31 Z 32 Z 3m 
 
 
Z 0 Z n1 Z n2 m (3)
 n Zn 
a  a0 a1 a2 am  '
 y   y1 y2 y3 yn  '
e  e1 e2 e3 en  '

If residual part e become zero then e  0 0 0 '

 Z  '  Z  a   Z  ' y (4)

  Z10 Z 20 Z 30 Z n0   Z10 Z11 Z12 Z1m 
 1   0 
 Z1 Z 21 Z 31 Z n1  Z2 Z 21 Z 22 Z 2m 
 Z  '  Z    Z12 Z 22 Z 32 Z n2   Z 30 Z 31 Z 32 Z 3m 
   
   
Z m Z 2m Z 3m m Z 0 Z n1 Z n2 m
 1 Zn   n Zn 

 N 0 0 n n n

  Zk Zk Z Z Z
0
k Z k1 0
k Z k2 0
k Z km 
 k 1 k 1 k 1 k 1

 n 1 0 n n n

  Zk Zk  Z k1 Z k1 Z Z 
1 2
k k Z k1 Z km 
 k 1 k 1 k 1 k 1

 n 2 0 n n n

  Zk Zk  Z k2 Z k1 Z 
2
k Z k2 Z k2 Z km 
 k 1 k 1 k 1 k 1 
  (5)
 n n n n 
 Z mZ 0 m m
  Z Z 
m
Z k1 m
Z k2 Zk Zk
k 1
k k
k 1
k
k 1
k
k 1


The element of i th row j th column of  Z '  Z  is =

n

Z
k
i 1
k Z kj 1  Z1i 1Z1j 1  Z 2i 1Z 2j 1  Z3i 1Z3j 1  Z 4i 1Z 4j 1   Z ni 1Z nj 1

 Z10 Z 20 Z 30 Z n0   y1 
 1 1 1  y 
 Z1 Z 2 Z 3 Z n1   2
 Z  '  y   Z12 Z 2
Z 2
Z n2 

 y3 
 
2 3

   
Z m Z 2m Z 3m Z nm   yn 
 1
 n 0 
  Z k yk 
 k 1  (6)
 n 1 
  Z k yk 
 k 1 
 n 2 
  Z k yk 
 k 1 
 
 n 
 Z km yk 
 k 1 

Final expression is:

  N 0 0 n n n
  n 0 
  Zk Zk Z Z Z   Z k yk 
0
k Z k1 0
k Z k2 0
k Z km 
 k 1 k 1 k 1 k 1
  k 1 
 n 1 0 1 m  0   n 1 
n n n a
  Zk Zk  Z k1 Z k1 Z Z  Z k Z k      Z k yk 
1 2
k k
a
 k 1 k 1 k 1 k 1
  1   k 1 
 n 2 0 n n n
  a    n
 (7)
  Zk Zk  Z k2 Z k1 Z  Z k2 Z km   2    Z k2 yk 
2
k Z k2
 k 1 k 1 k 1 k 1     k 1 
  am   
 n n n n   n 
 Z mZ 0 m m
   Z km Z k1 Z   k k 
m
Z k2 Z Z  Z m
y
k 1
k k
k 1 k 1
k
k 1
k k
  k 1

After solving a  a0 a1 a2 am  ' the model equation is:

y  a0 z 0  a1 z1  a2 z 2  a3 z 3  a4 z 4  a5 z 5 .  ........am z m (8)

y = prediction of least square fit

From equation 1 e is residual, and then square of residual is Sr

sr    yk   a0 zk0  a1 z1k  a2 zk2  a3 zk3  a4 zk4  a5 zk5 .  ........am zkm 

n
(9)
k 1

To minimize Sr it is necessary that each partial derivative be zero and these results in the system of
equations

Sr
0 (10)
ak
 sr
  2 zk0  yk   a0 zk0  a1 z1k  a2 zk2  a3 zk3  a4 zk4  a5 zk5  ........am zkm    0
n

a0 k 1
sr
  2 z1k  yk   a0 zk0  a1 z1k  a2 zk2  a3 zk3  a4 zi4  a5 zk5  ........am zkm    0
n

a1 k 1
sr
  2 z20  yk   a0 zk0  a1 z1k  a2 zk2  a3 zk3  a4 zk4  a5 zk5  ........am zkm    0
n

a2 k 1
sr
  2 zk3  yk   a0 zk0  a1 z1k  a2 zk2  a3 zk3  a4 zk4  a5 zk5  ........am zkm    0
n
(11)
a3 k 1

sr
  2 zkm  yk   a0 zk0  a1 z1k  a2 zk2  a3 zk3  a4 zk4  a5 zk5  ........am zkm    0
n

am k 1

Rearranging equation 11 wrt to a0 a1 a2 am

a0   zk0   a1  zk0 z1k  a2  zk0 zk2  a3  zk0 zk3  ............  am  zk0 zkm   Z k0 yk
n n n n n n

k 1 k 1 k 1 k 1 k 1 k 1

a0  z1k zk0  a1   z1k   a2  z1k zk2  a3  z1k zk3  ............  am  z1k zkm   Z k1 yk
n n 2 n n n n

k 1 k 1 k 1 k 1 k 1 k 1

a0  zi2 zi0  a1  zk2 z1k  a2   zk2   a3  zk2 zk3  ............  am  z1k zkm   Z k2 yk
n n n n n n
2

k 1 k 1 k 1 k 1 k 1 k 1

(12)

a0  z z  a1  z z  a2  z z  a3  z z  ............  am   z   Z
n n n n n n
m 0 m 1 m 2 m 3 m m
i i k k k k k k k k yk
k 1 k 1 k 1 k 1 k 1 k 1

It can be easily recognized as a system of m+1 linear equations m+1 unknown.

So
  Aa  B
or ,
 N 0 0 n n n
0 m   n 0 
  Zk Zk  Z k0 Z k1  Z k0 Z k2  Z Z
k k    Z k yk 
 k 1 k 1 k 1 k 1
  k 1 
 n 1 0 n n n
  a   n

  Zk Zk  Z k1 Z k1 Z Z  Z k1 Z km      Z k1 yk 
1 2 0
k k
a1 
 k 1 k 1 k 1 k 1
   k 1
  

 n 2 0 2 m   a2    
n n n n

  Zk Zk Z Z  Z k Z k      Z k yk 
2
k Z k1 2
k Z k2 2

 k 1 k 1 k 1 k 1     k 1 
 am    (13)
 n n n n   n 
 Z mZ 0 m m
  Z Z   Z k yk 
m
Z k1 m
Z k2 Zk Zk  m

  k 1
k k k k
k 1 k 1 k 1 k 1 

Out cum of the process from equation  8  to 13 is

 Z '  Z  a   Z ' y

So, previous assumption in equation  4  was correct. After solving a , Sr can be calculated as

sr    yk   a0 zk0  a1 z1k  a2 zk2  a3 zk3  a4 zk4  a5 zk5 .  ........am zkm  

n

k 1
(14)
n
   yk  yk 
k 1
 

 
n
S t   yk  y
k 1

 
n 2

St  Sr S  yk  y (15)
r2   1  r  1  k 1 2

 y 
St St n

k  y
k 1

y = prediction of least square method

y = mean of yk
Quadratic Spline:
The objective in quadratic splines is to derive a 2nd order polynomial for each interval
between data points. The polynomial for each interval can be generally represented as:
Si (xi )  ai  bi (x  x i )  ci (x  x i )2 (1)

For n data points (i  1, 2 n) , there are  n  1 intervals and, consequently, 3  n  1

unknown constants (the a ’s, b ’s & c ’s) to evaluate. So, 3  n  1 equations or conditions are
required to evaluate the unknowns.

It can be developed as follows:

 1: The function must pass through all the points.

fi  ai  bi (xi  xi )  ci (xi  xi )2

fi  ai (2)
Or,

So the constant in each spline must be equal to the value of the dependent variable at the
beginning of the interval. Result can be incorporated into Eq 1

Si (xi )  fi  bi (x  x i )  ci (x  x i )2

For  n  1 spline must pass through  xn , yn  point. So,

th

f n1  an1  bn1 (x n  x n1 )  cn1 (x n  x n1 )2  3

th
The equation can be simplified mathematically by defining the width of the i interval as
hi   xi 1  xi 

So Eq  3 simplifies to

f n1  an1  bn1 h n1  cn1 h n12

Number of conditions to be evaluated has now reduced to 3  n  1   n  1  1   2n  1

 2: The function values of adjacent polynomials must be equal at knots. This condition
can be written for knot  i  1 as

fi  bi (x i 1  x i )  ci (x i 1  x i ) 2  fi 1  bi 1 (x i 1  x i 1 )  ci 1 (x i 1  x i 1 ) 2  ci (x i 1  x i 1 ) 2 4
fi  bi hi  ci hi  fi 1
2
 5

The equation can be written for nodes i  1, 2,3,  n 1 . So there are  n  1 conditions,
Number of conditions to be evaluated has now reduced to  2n  1   n  1  n
  3: The first derivatives at the interior nodes must be equal. It means that adjacent
splines will be joined smoothly. The Eq 1 can be differentiated to yield

Si' (x)  bi  2ci (x  x i )

The equation of the derivatives at interior node  i  1 can therefore be written as

bi  2ci hi  bi 1  6
Writing this equation for all the interior nodes amounts to  n  2  conditions. This means that
there is 1 remaining condition.

 4: Assume the 2nd derivative is zero at the first point. 2nd derivative of the Eq 1 is
2ci . So, from Eq(6)

c1  0 7
Putting this condition in Eq(5) for i  1

b1h1  ( f 2  f1 )  8
 bi 1  bi 
From Eq(6) ci  9
2hi

Putting Eq(8) in Eq(5) bi hi  bi 1hi  2( fi 1  fi ) 10

So putting i  1, 2, , n  1 in to Eq(9) there are  n  1 equations with  n  variable.

b1h1  b2 h1  2( f 2  f1 )
b2 h2  b3h2  2( f 3  f 2 )
11
bn 1hn 1  bn hn 1  2( f n  f n 1 )

Here we are assuming bn as an extra variable. We have  n  1 splines so  n  1 b are

there.

From Eq 8 & Eq 11

b1  b2 b1  b2

 h1 0 0  b 2   ( f 2  f1 ) 
h h2 0   b 3   2( f3  f 2 ) 
 2 
 0    
    
0 hn 1 hn 1  b n   2( f n  f n 1 ) 

After solving b1 , b2 , bn we can calculate values of c1 , c2 cn1 from equation (9).

Taylors Series Second Order RK Method is given by
𝑓′(𝑥 𝑖 , 𝑦𝑖 ) 2 𝑦𝑖+1 = 𝑦𝑖 + (𝑎1 𝑘1 + 𝑎2 𝑘2 )ℎ
𝑦𝑖+1 = 𝑦𝑖 + 𝑓(𝑥𝑖 , 𝑦𝑖 )ℎ + ℎ +⋯
2!
𝑘1 = 𝑓(𝑥𝑖 , 𝑦𝑖 ) and 𝑘2 = 𝑓(𝑥𝑖 + 𝑝1 ℎ, 𝑦𝑖 + 𝑞11 𝑘1 ℎ)
′ (𝑥
Now, 𝑓 𝑖 , 𝑦𝑖 ) can be written by chain-rule differentiation
Taylors Series for a two variable function
𝜕𝑓(𝑥, 𝑦) 𝜕𝑓(𝑥, 𝑦) 𝑑𝑦 𝜕𝑔 𝜕𝑔
𝑓 ′ (𝑥𝑖 , 𝑦𝑖 ) = + 𝑔(𝑥 + 𝑟, 𝑦 + 𝑠) = 𝑔(𝑥, 𝑦) + 𝑟 +𝑠 +⋯
𝜕𝑥 𝜕𝑦 𝑑𝑥 𝜕𝑥 𝜕𝑦
Using the above equation for k 2
𝜕𝑓(𝑥, 𝑦) 𝜕𝑓(𝑥, 𝑦)
𝑓(𝑥𝑖 + 𝑝1 ℎ, 𝑦𝑖 + 𝑞11 𝑘1 ℎ) = 𝑓(𝑥𝑖 , 𝑦𝑖 ) + 𝑝1 ℎ + 𝑞11 𝑘1 ℎ +⋯
Using the above equation in L1 𝜕𝑥 𝜕𝑦
𝜕𝑓(𝑥, 𝑦) 𝜕𝑓(𝑥, 𝑦) 𝑑𝑦 ℎ2 Using the above equation in L1
𝑦𝑖+1 = 𝑦𝑖 + 𝑓(𝑥𝑖 , 𝑦𝑖 )ℎ + ( + ) +⋯ 𝜕𝑓 𝜕𝑓
𝜕𝑥 𝜕𝑦 𝑑𝑥 2! 𝑦𝑖+1 = 𝑦𝑖 + 𝑎1 ℎ𝑓(𝑥𝑖 , 𝑦𝑖 ) + 𝑎2 ℎ𝑓(𝑥𝑖 , 𝑦𝑖 ) + 𝑎2 𝑝1 ℎ2 + 𝑎2 𝑞11 ℎ2 𝑓(𝑥𝑖 , 𝑦𝑖 ) +⋯
𝜕𝑥 𝜕𝑦
dy
Using dx = f(xi , yi ) in the above equation
Group like terms
𝜕𝑓(𝑥, 𝑦) 𝜕𝑓(𝑥, 𝑦) ℎ2 𝜕𝑓 𝜕𝑓
𝑦𝑖+1 = 𝑦𝑖 + 𝑓(𝑥𝑖 , 𝑦𝑖 )ℎ + ( + 𝑓(𝑥𝑖 , 𝑦𝑖 )) + ⋯ 𝑦𝑖+1 = 𝑦𝑖 + (𝑎1 + 𝑎2 )𝑓(𝑥𝑖 , 𝑦𝑖 )ℎ + [𝑎2 𝑝1 + 𝑎2 𝑞11 𝑓(𝑥𝑖 , 𝑦𝑖 ) ]ℎ2 + ⋯
𝜕𝑥 𝜕𝑦 2! 𝜕𝑥 𝜕𝑦

Comparing the above two equations, we get

(𝑎1 + 𝑎2 ) = 1;
1
𝑎2 𝑝1 = 2;
1
𝑎2 𝑞11 =
2

Taken from Numerical Methods for Engineers – Chapra & Canale