Finite Di↵erence Method for Boundary Value Problems

Finite Di↵erences
8 f (x+h) f (x)
>
< h + O(h) (Forward Di↵erence)
f (x) f (x h)
f 0 (x) = + O(h) (Backward Di↵erence)
> f (x+h)h f (x h)
:
2h + o(h2 ) (Central Di↵erence)
8 f (x) 2f (x+h)+f (x+2h)
>
< h2 + O(h) (Forward Di↵erence)
00 f (x) 2f (x h)+f (x 2h)
f (x) = h2 + O(h) (Backward Di↵erence)
>
: f (x h) 2f (x)+f (x+h)
h2 + o(h2 ) (Central Di↵erence)

Higher Order Derivatives

2f (x + h) + 3f (x) 6f (x h) + f (x 2h)
f 0 (x) = + O(h3 )
6h
f (x + 2h) + 6f (x + h) 3f (x) f (x h)
f 0 (x) = + O(h3 ) (4.11)
6h
f (x + 2h) + 8f (x + h) 8f (x h) + f (x 2h)
f 0 (x) = + O(h4 ) (4.13)
12h

Order of Approximation
• It is worth considering exactly what is meant by the order of accuracy of a discretization approximation.
• As we refine the grid, for any useful scheme, errors associated with the discretization approximation
can be expected to reduce.
• We reach a grid independent solution when any further grid refinement produces no significant di↵er-
ence in the computed solution. At this stage, the discretization errors are small enough that they can
be neglected.
• Note this does not mean a second order solution on one particular grid will always be more accurate
than a first order one (although this will usually be the case for a sufficiently fine grid).
• It does, however, imply that as we refine the grid the error in the higher order scheme goes down more
rapidly, so we expect the higher order scheme to reach a grid independent solution on a coarser grid
than would be required for a lower order scheme.
• It is worth noting that the behaviour described above for a particular scheme can only
be expected on a reasonably fine grid.

The reason for this can be seen from the Taylor series expansion. For example, in the first order backward
di↵erence scheme we have:

fi fi 1 x 00 ( x)2 000
f 0 (xi ) = + f (xi ) f (xi ) + O( x3 )
x 2! 3!

1
Santanu Manna MA 204N: Finite Di↵erence Method

If the first term in the truncation is to be the leading error term, then x has to be small enough so
that:

( x)2 000 x 00
|f (xi )| ⌧ |f (xi )|
3! 2!
Or,

f 00 (xi )
x⌧3
f 000 (xi )

Polynomial Fitting
• The above approximations for the derivatives could also have been derived by fitting a polynomial to
the function f through xi and surrounding points and then di↵erentiating this polynomial to obtain
its gradient.

• The forward and backward di↵erence schemes arise from fitting a first order polynomial through the
points (xi , xi+1 ) and (xi 1 , xi ) respectively.
• The central di↵erence scheme can be obtained by fitting a quadratic curve through the points xi 1 , xi , xi+1 .

Linear Fit:
The first order polynomial:

(x xi+1 )fi (x xi )fi+1

P1 (x) = f (x) = + , xi < x < xi+1
(xi xi+1 ) (xi+1 xi )
(x
xi+1 )fi + (x xi )fi+1
=
x
The derivative of this function yields the forward di↵erence formula:
fi+1 fi
f 0 (x) =
x

Quadratic Fit:
The central di↵erence scheme can be obtained by fitting a quadratic curve through the points xi 1 , xi , xi+1 .
The resulting polynomial approximation for function f can be written as:

(x xi )(x xi+1 )fi 1 (x xi 1 )(x xi+1 )fi (x xi )(x xi 1 )fi+1

P2 (x) = f (x) = + +
(xi 1 xi )(xi 1 xi+1 ) (xi xi 1 )(xi xi+1 ) (xi+1 xi )(xi+1 xi 1 )

On a uniform grid:

xi+1 xi = xi xi 1 = x and xi+1 xi 1 =2 x

Then we get:

(x xi )(x xi+1 )fi 1 (x xi 1 )(x xi+1 )fi (x xi )(x xi 1 )fi+1

f (x) = + (5.3)
2( x)2 ( x)2 2( x)2

Now di↵erentiating eq. (5.3) with respect to x, we get:

(x xi + x xi+1 )fi 1 (x xi 1+ x xi+1 )fi (x xi + x xi 1 )fi+1

f 0 (x) = +
2( x)2 ( x)2 2( x)2

2
Santanu Manna MA 204N: Finite Di↵erence Method

xfi 1 0 · fi xfi+1
f 0 (xi ) = + +
2( x)2 ( x)2 2( x)2
fi+1 fi 1
) f 0 (xi ) = (5.5)
2 x
Again di↵erentiating eq. (5.4) with respect to x, we get:

(1 + 1)fi 1 (1 + 1)fi (1 + 1)fi+1

f 00 (x) = + +
2( x)2 ( x) 2 2( x)2
fi 1 2fi + fi+1
) f 00 (x) = (5.6)
( x)2
The above expressions are identical to the formulae obtained earlier.

One-sided Approximation
The finite di↵erence approximation for first order derivative along the boundary can be obtained by using
forward di↵erence which is first order accurate.
For second order accuracy, we will use a polynomial approach as follows:
Assume that f can be expressed by the polynomial

f = a + bx + cx2 (5.7)

At grid point 1, x = 0 which yields

f1 = a
At grid point 2, x = x which yields

f2 = a + b x + c( x)2

At grid point 3, x = 2 x which yields

f3 = a + 2b x + c(2 x)2

Solving the above expressions, we get:

3f1 + 4f2 f3 f1 2f2 + f3
a = f1 , b= , c=
2 x 2( x)2

Now the first derivative at grid point 1, i.e., at x = 0 is given by:

✓ ◆
@f
=b
@x x=0
✓ ◆
@f 3f1 + 4f2 f3
= (5.8)
@x x=0 2 x
Now the second derivative at grid point 1, i.e., at x = 0 is given by:
✓ 2 ◆
@ f
= 2c
@x2 x=0
✓ 2 ◆
@ f f1 2f2 + f3
2
= (5.9)
@x x=0 ( x)2
The error in above approximations are of order O( x).

3
Santanu Manna MA 204N: Finite Di↵erence Method

Methods for solving Parabolic Equations:

One-dimensional parabolic equation can be written as:

@u @2u
= ↵ 2, 0  x  L, t > 0 (6.1)
@t @x
Suppose Dirichlet boundary conditions are prescribed at both the ends and an initial condition at time
t > 0 as:
u(0, t) = u0 , u(L, t) = uL , u(x, 0) = f (x), 0  x  L (6.2)
To solve this problem, we first discretize the domain into regular mesh such that xi = i x; i = 0(1)N
where x0 = 0, xL = N . When the values of ui,j is computed up to t = tj = j t i.e., jth time level, the
values at (j + 1)th time level are computed to give ui,j+1 , i = 1(1)N 1, j = 1, 2, . . .

⌥ Explicit Method:
We discretize equation (6.1) at mesh point (i, j) such that the time derivative is replaced by forward di↵erence
and space derivative is replaced by central di↵erence. We get:

u(xi , tj+1 ) u(xi , tj ) u(xi 1 , tj ) 2u(xi , tj ) + u(xi+1 , tj )

= + O( t) + O( x)2
t ( x)2

or we can write this as:

t ⇥ ⇤
ui,j+1 = ui,j + [ui 1,j 2ui,j + ui+1,j ] + t O( t) + O( x)2
( x)2
t
Let ( x)2 = r, and neglecting the error term, we get:

ui,j+1 = rui 1,j + (1 2r)ui,j + rui+1,j (6.3)

Thus value of ui,j+1 can be computed explicitly. This method is also known as explicit method.

Molecule of Explicit Scheme:

Example: Find the numerical solution of the heat conduction equation

@u @2u
= , 0  x  1, t > 0
@t @x2
With boundary conditions u(0, t) = u(1, t) = 1, and initial conditions
(
1 + 2x, 0  x  1/2
u(x, 0) =
3 2x, 1/2 < x  1

⌅ Use explicit method taking x = 0.2, t = 0.02, and compute up to t = 0.24 up to six decimal places.

4
Santanu Manna MA 204N: Finite Di↵erence Method

Solution: Since the initial temperature u(x, 0) is symmetric and boundary conditions at x = 0 and
x = 1 are also same, the problem is symmetric about x = 0.5. i.e., the temperature at the subsequent times
will also remain symmetric.
The domain is subdivided as:

u= 0 0.2 0.4 0.6 0.8 1.0

x= 0 0.2 0.4 0.6 0.8 1
We have xi = 0.2 ⇥ i, i = 0(1)5
Due to symmetry at x = 0.5, for all j,

u0,j = u5,j , u1,j = u4,j , u2,j = u3,j

t 0.02
r= 2
= = 0.5
x 0.04
Putting r = 0.5 in the Explicit formula:

ui,j+1 = rui 1,j + (1 2r)ui,j + rui+1,j

1
ui,j+1 = (ui 1,j + ui+1,j ), i = 1, 2
2
i = 3, 4 can be written by symmetry.
For j = 0,
u0,0 = u5,0 = 1.0, u1,0 = u4,0 = 1.4, u2,0 = u3,0 = 1.8
Computed values are:

t 0.0 0.2 0.4 0.6 0.8 1.0

0.00 1.0 1.4 1.8 1.8 1.4 1.0
0.02 1.0 1.4 1.6 1.6 1.4 1.0
0.04 1.0 1.3 1.5 1.5 1.3 1.0
0.06 1.0 1.25 1.4 1.4 1.25 1.0

Parabolic Equations (Continue. . . )

Implicit Method:
We discretize equation (6.1) at mesh point (i, j + 1) such that the time derivative is replaced by backward
di↵erence and space derivative is replaced by central di↵erence, i.e.,
ui,j+1 ui,j ui 1,j+1 2ui,j+1 + ui+1,j+1
= + O( t) + O( x)2
t ( x)2
t
Let ( x)2 = r, and neglecting the error term, we get:

rui 1,j+1 + (1 + 2r)ui,j+1 rui+1,j+1 = ui,j (6.4)

For i = 1(1)N 1, gives (N 1) equations in ui,j+1 . This formula is known as Laasonen’s formula.

Molecule of Implicit Scheme:

j+1

-r (1-2r) -r
j

1
i 1 i i+1

5
Santanu Manna MA 204N: Finite Di↵erence Method

Crank-Nicolson’s Method:
It is also an implicit scheme. The di↵erence is that the discretization of equation (6.1) is made at the mid-
point of j th and (j + 1)th levels and both the time and space derivatives are replaced by central di↵erence,
i.e.,

✓ ◆ 
@u 1 ui 1,j 2ui,j + ui+1,j ui 1,j+1 2ui,j+1 + ui+1,j+1
= 2
+ + O( t)2 + O( x)2
@t i,j+ 12 2 ( x) ( x)2

t
Using ( x)2 = r, and neglecting the error terms, we can write the above equation as:

rui 1,j+1 + (1 + r)ui,j+1 rui+1,j+1 = rui 1,j + (1 r)ui,j + rui+1,j (6.5)

For i = 1(1)N , there will be (N 1) equations in (N 1) unknowns and their solution will be solved at
(j + 1)th time level.

Molecule of Crank-Nicolson Scheme:

Example:
Find the numerical solution of the heat conduction equation

@u @2u
= , 0  x  1, t > 0
@t @x2
With boundary conditions u(0, t) = u(1, t) = 1 and initial conditions
(
1 + 2x, 0  x  12
u(x, 0) =
3 2x, 12  x  1

⌅ Solve by C-N scheme by taking t = 0.08 and x = 0.2.

Solution: Since the initial temperature u(x, 0) is symmetric and boundary conditions at x = 0 and
x = 1 are also same, so the problem is symmetric about x = 0.5, i.e., the temperature at subsequent times
will also remain symmetric.
Here t = 0.08, r = t/ x2 = 0.08/0.04 = 2, then C-N scheme:

rui 1,j+1 + 2(1 + r)ui,j+1 rui+1,j+1 = rui 1,j + 2(1 r)ui,j + rui+1,j
transforms to:

2ui 1,j+1 + 6ui,j+1 2ui+1,j+1 = 2ui 1,j 2ui,j + 2ui+1,j

or

6
Santanu Manna MA 204N: Finite Di↵erence Method

ui 1,j+1 + 3ui,j+1 ui+1,j+1 = ui 1,j ui,j + ui+1,j

For i = 1,
u0,j+1 + 3u1,j+1 u2,j+1 = u0,j u1,j + u2,j (1)
For i = 2,
u1,j+1 + 3u2,j+1 u3,j+1 = u1,j u2,j + u3,j (2)
Using u0,j = u3,j = 1 and u3,j = u0,j , implies:

From the right-side derivation:

For j = 0 : 3u1,1 u2,1 = 2 ) u1,1 = 1.24, u2,1 = 1.32

For j = 1 : 3u1,2 u2,2 = 2 ) u1,2 = 1.09, u2,2 = 1.145
For j = 2 : 3u1,3 u2,3 = 2 ) u1,3 = 1.04, u2,3 = 1.065

Numerical Solution Table:

t x 0 0.2 0.4 0.6 0.8 1.0
0.00 1.0 1.4 1.8 1.8 1.4 1.0
0.08 1.0 1.24 1.32 1.32 1.24 1.0
(1.20) (1.325)
0.16 1.0 1.090 1.145 1.145 1.090 1.0
[1.086] [1.139]
0.24 1.0 1.040 1.065 1.065 1.040 1.0
[1.037] [1.060]

Compatibility, Stability & Convergence:

Let u represent the exact solution of the PDE and uD is the exact solution of the discretized equation. The
error (u uD ) is called discretization error.
@2u
Assume the parabolic di↵erential equation @u
@t = @x2 , 0  x  L, t > 0, and let
✓ ◆
@ @2
L= , LD denotes the discretizing operator, and T is the truncation error.
@t @x2

Then in operator form:

Lu = 0 = LD (u) + T (7.1)
As uD is the exact solution of LD (u) = 0. If the discretization error (u uD ) ! 0 as x and t ! 0,
then uD ! u, i.e., the finite di↵erence solution is said to converge to the exact solution.

Compatibility and Stability

If the finite di↵erence approximation LD (u) tends to L(u) as x and t tend to zero, i.e., T ! 0 as x and
t ! 0, then the finite di↵erence approximation is said to be compatible with the original PDE.

While finding the solution uD , some errors like rounding/truncation etc., also creep in. So the final
solution is di↵erent from uD , and can be called uC . The error (uD uC ) is called stability error, and if it
tends to zero, then the finite di↵erence scheme is said to be stable.

Hence, total error = u uC = (u uD ) + (uD uC ) (7.2)