Numerical Integration of a Function

• If we evaluate the function at 2 points, what

should be the location of these points such
that the error is minimized?
1
• Standard domain ~
∫ f ( z )dz ≈ I z = c0 f ( z0 ) + c1 f (z1 )
−1

• Four adjustable parameters, c0, z0 and c1, z1

• May integrate polynomials of degree 3 exactly
• How to obtain these parameters?
• One option: Using f(z) as 1,z,z2,z3
 Integration of a Function: Standard Domain
• We get:
2
c0 + c1 = 2; c0 z0 + c1 z1 = 0; c z + c z = ; c0 z03 + c1 z13 = 0
2
0 0
2
1 1
3

resulting in
1 1
z0 = − ; z1 = ; c0 = 1; c1 = 1
3 3

• 2-point Gauss Quadrature

 General Methodology
• Let the exactly integrable polynomial be
z − z1 z − z0
f3 ( z) = f ( z0 ) + f ( z1 ) + (a + bz )( z − z0 )( z − z1 )
z0 − z1 z1 − z0

a and b are arbitrary constants

• This function satisfies
f 3 ( z0 ) = f ( z0 ) and f 3 ( z1 ) = f ( z1 )
 3.0

2.0

1.0

0.0
-1 -0.5 0 0.5 1

1
~
∫ f ( z )dz ≈ I z = c0 f ( z0 ) + c1 f ( z1 )
−1

z − z1 z − z0
f3 ( z) = f ( z0 ) + f ( z1 )
z0 − z1 z1 − z0
+ (a + bz )( z − z0 )( z − z1 )
 Gauss Quadrature: General Form
• Since the cubic polynomial is exactly integrable
1
z − z1 z − z0
∫−1 z0 − z1 f ( z0 ) + z1 − z0 f ( z1 ) + (a + bz)( z − z0 )( z − z1 )dz
= c0 f ( z0 ) + c1 f ( z1 )
1 1
z − z1 z − z0
Which implies that c0 = ∫ dz and c1 = ∫ dz
z − z1
−1 0
z − z0
−1 1

And, for any arbitrary a and b:

1

∫ (a + bz )( z − z )( z − z )dz = 0
−1
0 1
 Gauss-Legendre Quadrature
• Recall the first few Legendre polynomials:
P0(x)=1; P1(x) = x; P2(x) = (−1+3x2)/2
P3(x) = (−3x+5x3)/2; P4(x) = (3−30x2+35x4)/8
• Any of these is orthogonal to all lower degree
polynomials. Earlier we had seen that P2(z) is
orthogonal to P0(z), and P1(z), i.e.,
1 1

∫ P ( z ) P ( z )dz = ∫ P ( z ) P ( z )dz = 0
−1
2 0
−1
2 1

• It implies that P2(z) is orthogonal to 1 and z

 Gauss-Legendre Quadrature
• Therefore, if we want the quadratic (z-z0)(z-z1)
to be orthogonal to all linear functions, we
should choose z0 and z1 as the roots of
(−1+3z2)=0
• Which gives us −1/√3, and 1/√3
• The weights are obtained from
1 1
z− 1
1 z+
1
3  z 3 2  3 dz = 1
c0 = ∫ dz =  − z  = 1 and c1 = ∫
−1 −
1 1 2 4  −1 1 1
− −1 +
3 3 3 3
 Gauss-Legendre Quadrature
• Similarly, for three Gauss points, we need the
3
zeroes of (−3z+5z ), which are 0, ±
3
 3  5
z  z − 
5  5 3 5 2 5 2 5 3 5 2
L0 ( z ) =  =− z + z ; L1 ( z ) = 1 − z ; L2 (z ) = z+ z
3 3 6 5 6 3 6 5 6
×2
5 5
1
1
 5z 2
3 5 3 5 8 5
c0 = ∫ L0 ( z )dz = − + z  = ; c1 = ; c2 =
−1  12 5 18  −1 9 9 9
• Since the c’s are weights, it is common to use
the symbol, W
 Gauss-Legendre Quadrature Weights
• The weights may be related to the Legendre
polynomials, Pn(z):

Wi =
(
2 1− z 2
i )
[(n + 1)Pn (zi )]2

• For example, with 3 Gauss points (n=2), P2(z)

= (−1+3z2)/2; the z’s are −√0.6, 0, √0.6
• The value of P2(z) at these points are 0.4, -0.5,
and 0.4, respectively.
• The weights are 5/9, 8/9, and 5/9, as before.
 6.0

5.0

4.0

3.0

2.0

1.0

0.0
-1 -0.5 0 0.5 1

2 3 4 5
f ( x) = 1 + x + x + x + x + x
 25.0

20.0

15.0

10.0

5.0

0.0
-1 -0.5 0 0.5 1

-5.0

2 3 4 5
f ( x) = 1 + 2 x + 3x + 4 x + 5 x + 6 x
Abscissa, Weight, and Error for the Gauss-Legendre Quadrature points

n Abscissa Weight Error

f " (ξ )
0 0.00000 2.0000
3
f iv (ξ )
1 ±0.57735 1.0000
135
0.00000 0.88889 f vi (ξ )
2
±0.77460 0.55556 15750
±0.33998 0.65215 f ( ) (ξ )
8
3
±0.86114 0.34785 3472875
0.00000 0.56889
f ( ) (ξ )
10
4 ±0.53847 0.47863
±0.90618 0.23693 1237732650
 Weighted Gauss Quadrature
• Instead of integral of the function, we need
the integral with a weight-function
1
~ n
∫ w(z ) f ( z )dz ≈ I = ∑Wi f (zi )
−1 i =0

• Recall the Tchebycheff polynomials, where the

weight was 1 / 1 − z 2
• Using this weight function, we get the Gauss-
Tchebycheff quadrature
• Of course, we could treat w(z)f(z) as a single
function and use Gauss-Legendre quadrature
 Gauss-Tchebycheff quadrature
• The quadrature points are the zeroes of
Tchebycheff polynomial of degree n+1
• These zeroes are given by

• And all the weights turn out to be equal to

π/(n+1)
 Gauss Quadrature: Example
2
1
• Estimate I = ∫ xe x dx (T.V. = 0.170483)
1
• First, convert to standard domain: z=2(x−1.5)
• Use 2-point Gauss-Legendre:
i z W x f
0 −1/√3 1 1.5−1/2√3 0.245849
1 1/√3 1 1.5+1/2√3 0.093467

resulting in Iz=0.339315 and Ix=0.169658

• Error in Ix = 8.25x10−4. In Iz= 1.65x10−3
f (ζ )
iv
th derivative (z) varies from
• Recall: E = . 4
135
0.04 to 1.5. Theoretical error: 0.0003 to 0.011
 Gauss Quadrature: Example
• Use 3-point Gauss-Legendre:
i z W x f

0 −√0.6 5/9 1.5−√0.6/2 0.295380

1 0 8/9 1.5 0.148753

2 √0.6 5/9 1.5+√0.6/2 0.080263

resulting in Iz=0.340916 and Ix=0.170458

• Error in Ix = 2.5x10−5. In Iz= 5x10−5

 Gauss-Tchebycheff Quadrature: Example
• Estimate (same as before, but weighted)

e − ( z / 2+1.5 )
1
Iz = ∫ dz (T.V. = 0.571946)
−1 1 − z (z / 2 + 1.5)
2

• Use 2-point Gauss-Tchebycheff (n=1), T2(z) =

2z2−1; the z’s are ±√0.5:
i z W x f
0 −1/√2 π/2 1.5−1/2√2 0.277173
1 1/√2 π/2 1.5+1/2√2 0.084529

resulting in Iz=0.568160
• Error in Iz= 3.79x10−3
 Gauss-Tchebycheff Quadrature: Example
• Use 3-points, T3(z) = 4z3−3z; the z’s are
−√0.75, 0,√0.75 :
i z W x f
0 −√0.75 π/3 1.5−√0.75/2 0.322444

1 0 π/3 1.5 0.148753

2 √0.75 π/3 1.5+√0.75/2 0.074863

resulting in Iz=0.571833
• Error in Iz= 1.13x10−4
 Numerical Integration: Improper Integrals
b

• Estimate I = ∫ f ( x)dx for a known function

a

• We have assumed:
• a and b are finite
( ) is defined and is continuous in (a,b)
• f(x)
• Improper Integral: When any (or both) of
these assumptions is violated
1
• E.g., ∞ cos x
I = ∫e
−x 2
dx I =∫ dx
1 0 x
 Improper Integrals: Convergence
• An improper integral may or may not
converge (i.e., have a finite value)
• We assume that it converges! How to find it?
• If the domain is unbounded, we use a
transformation of variable to make it finite
1
−
∞ 1 z2
− x2 1 e
I = ∫ e dx : z = ⇒ I = ∫ 2 dz
1
x 0
z
• If f(x) is undefined at one end, a semi-open
method could be used (or variable transform)
1 1
cos x
I =∫ dx : z = x ⇒ I = ∫ 2 cos z dz
2

0 x 0
 Improper Integrals: Evaluation

• If f(x) is undefined at both ends, an open method

could be used
• If f(x) is undefined at some point within (a,b), we
may need to split the integral into two parts
• Sometimes both the domain and the range could
be unbounded. E.g.,
∞
cos x
I =∫ dx
0 x
 Improper Integrals: Example
• The complementary error function is defined
∞
as 2 2

∫e
−t
erfc( x) = dt
π x
• Estimate the value of erfc(1) (T.V.=0.157299)
• Transformation: y=1/t 1
−
∞ 1 y2
2 e
I = ∫ e dt ⇒ I = ∫ 2 dy
−t

1 0
y
• Note that f(y) is undefined at y=0 (limit does
exist). Trapezoidal, Simpson… cannot be used!
• Use 3-point Gauss-quadrature
 Improper Integrals: Example

• 3-point Gauss-quadrature
i z W y f

0 −√0.6 5/9 (−√0.6+1)/2 0.000000

1 0 8/9 0.5 0.073263

2 √0.6 5/9 (√0.6+1)/2 0.356644

• Iz=0.263258; Iy=0.131629; erfc(1)=0.148527

• Error = 8.77x10−3