chapter 12-b - bauer college of business · ch 12 - integral calculus 8/8/2019 1 1 chapter 12-b...
TRANSCRIPT
Ch 12 - Integral Calculus 8/8/2019
1
1
Chapter 12-bIntegral Calculus - Extra
Thomas SimpsonIsaac Newton
2
BONUS Introduction to Numerical
Integration
Ch 12 - Integral Calculus 8/8/2019
2
30
2
4
6
8
10
12
3 5 7 9 11 13 15
Numerical Integration
Idea: Do an integral in small parts, like the way we presented integration; i.e., a summation.
Numerical methods just try to make it faster and more accurate.
4
• Idea: Weighted sum of function values to approximate integral
)x(fc)x(fc)x(fc
)x(fcdx)x(f
nn1100
i
n
0ii
b
a
x0 x1 xnxn-1x
f(x)
Basic Numerical Integration
• Task: Find appropriate ci’s (weights) and the xi’s (nodes).
Ch 12 - Integral Calculus 8/8/2019
3
5
• We want to find integration of functions of various forms of the equation known as the Newton-Cotes integration formulas (“rules”).
• Newton-Cotes formula
Assume the value of f(x) defined on [a,b] is known at equally spaced points xi (i=0,1,...,n), where x0 =a, and xn =b. Then,
where xi = h i + x0, with h (“step size”) = (xn − x0)/n = (b −a)/n.
The ci's are called weights. The xi's are called nodes. The precision of the approximation depends on n.
Note: The N-C rules use nodes equally spaced. But, they do not have to be --unequally spaced nodes are OK too.
n
iii xfcf(x)dx
1
b
a
)(
Basic Numerical Integration
6
• Weights
In the N-C formulae, they are derived from an approximation required to be equal for a polynomial of order lower or equal to the degree of the polynomials used to approximate the function. In other methods, weights and nodes can be derived jointly.
• Error analysis
The error of the approximation is the difference between the value of the integral and the numerical result:
error = ε = ∫ f(x) dx – Σi ci f(xi)
The errors are frequently approximated using Taylor series for f(x).The error analysis gives a strict upper bound on the error, if the derivatives of f are available.
Basic Numerical Integration
Ch 12 - Integral Calculus 8/8/2019
4
7
• The weights are derived from the Lagrange polynomials L(x). The weight, Li(x), depend only on the xi's (no two xi's are the same); not on the function ƒ.
Recall that L(x) is used for polynomial interpolation of a function f(x), given a set of points (xi, f(xi)).
Example: Interpolate f(x) = 2 x3 over [2, 4], with 3 points: (2, 3, 4). The interpolating Lagrange is:
16 ∗ 54 ∗ 128 ∗
18 52 48
Newton-Cotes Formula
))...()()...((
))...()()...((
)(
)()(
110
110
,0 niiiiii
nii
jinj ji
ji xxxxxxxx
xxxxxxxx
xx
xxxL
8
• Different methods use different polynomials to get the ci's.
• Newton-Cotes Closed Formulae – Use both end points
– Trapezoidal Rule : Linear
– Simpson’s 1/3-Rule : Quadratic
– Simpson’s 3/8-Rule : Cubic
– Boole’s Rule : Fourth-order
• Newton-Cotes Open Formulae – Use only interior points
– midpoint rule
Newton-Cotes Formula
Ch 12 - Integral Calculus 8/8/2019
5
9
Trapezoid Rule• Straight-line approximation
The trapezoid rule approximates the region under the graph of the function f(x) as a trapezoid and calculating its area.
)x(f)x(f2
h
)x(fc)x(fc)x(fcdx)x(f
10
1100i
1
0ii
b
a
x0 x1 x
f(x)
L(x)
Approximate integral by the trapezoid’s area.
10
Trapezoid Rule - Derivation
• We use the Lagrange approximation for function f(x) over the interval (xn – x0) (Lagrange interpolation), given by
010 1
0 1 1 0
0 1
( ) ( ) ( )
dx a x , b x , , d ;
h
0( ) (1 ) ( ) ( ) ( )
1
x xx xL x f x f x
x x x x
x alet h b a
b a
x aL f a f b
x b
• Then, we integrate to obtain the trapezoid rule
)()(...)()()()()( 1100 xLxfxLxfxLxfxf nn
• For the case, n = 2, with the interval (x1 – x0):
Ch 12 - Integral Calculus 8/8/2019
6
11
Trapezoid Rule - Derivation
• Integrating:
1
0
1 1
0 0
1 12 2
0 0
( ) ( ) ( )
( ) (1 ) ( )
( ) ( ) ( ) ( ) ( )2 2 2
b b
a af x dx L x dx h L d
f a h d f b h d
hf a h f b h f a f b
The weights depend only on h!
• This approximation may be poor. The approximation error is:
ε = ∫ f(x) dx – Σi ci f(xi) = -(b – a)3/(12) f ’’(η), η Є [a, b].
• Thus, if the integrand is convex –i.e., positive second derivative–, the error is negative. That is, the trapezoidal rule overestimates the true value of the integral.
12
Evaluate the integral
• Exact solution
• Graph
926477.5216)1x2(e4
1
e4
1e
2
xdxxe
1
0
x2
4
0
x2x24
0
x2
dxxe4
0
x2
Trapezoid Rule - Example
Ch 12 - Integral Calculus 8/8/2019
7
13
Evaluate the integral
• Trapezoidal Rule: Approximation Error
ε = -(b – a)3/(12) f ’’(η) (η is a number between a and b).
Let’s take η =2. Then
ε = -43/12*[2*exp(2*2)+2*exp(2*2)+4*(2)*exp(2*2)] =-3494.282
• Trapezoidal Rule
dxxe4
0
x2
%12.357926.5216
66.23847926.5216
66.23847)e40(2)4(f)0(f2
04dxxeI 84
0
x2
Trapezoid Rule - Example
14
Simpson’s 1/3-Rule (Kepler’s Rule)
• Approximate the function by a parabola
)x(f)x(f4)x(f3
h
)x(fc)x(fc)x(fc)x(fcdx)x(f
210
221100i
2
0ii
b
a
x0 x1x
f(x)
x2h h
L(x)
Ch 12 - Integral Calculus 8/8/2019
8
15
1 xx
0 xx
1 xx
h
dxd ,
h
xx ,
2
abh
2
ba x ,bx ,ax let
)x(f)xx)(xx(
)xx)(xx(
)x(f)xx)(xx(
)xx)(xx( )x(f
)xx)(xx(
)xx)(xx()x(L
2
1
0
1
120
21202
10
12101
200
2010
21
)x(f2
)1()x(f)1()x(f
2
)1()(L 21
20
Simpson’s 1/3-Rule - Derivation
• Use a quadratic Lagrange interpolation:
16
1
1
23
2
1
1
3
1
1
1
23
0
1
12
1
0
21
1
10
1
1
b
a
)2
ξ
3
ξ(
2
h)f(x
)3
ξ(ξ)hf(x)
2
ξ
3
ξ(
2
h)f(x
)d ξ1ξ(ξ2
h)f(x)d ξξ1()hf(x
)d ξ1ξ(ξ2
h)f(xdξ)(Lhf(x)dx
)f(x)4f(x)f(x3
hf(x)dx 210
b
a
• Integrate the Lagrange interpolation
• Again, the weights depend only on h!
Simpson’s 1/3-Rule - Derivation
Thomas Simpson (1710 – 1761, England)
Ch 12 - Integral Calculus 8/8/2019
9
17
• Approximate by a cubic polynomial
)x(f)x(f3)x(f3)x(f8
h3
)f(xc)f(xc)f(xc)f(xc)x(fcdx)x(f
3210
33221100i
3
0ii
b
a
x0 x1x
f(x)
x2h h
L(x)
x3h
Simpson’s 3/8-Rule
18
)x(f)xx)(xx)(xx(
)xx)(xx)(xx(
)x(f)xx)(xx)(xx(
)xx)(xx)(xx(
)x(f)xx)(xx)(xx(
)xx)(xx)(xx(
)x(f)xx)(xx)(xx(
)xx)(xx)(xx()x(L
3231303
210
2321202
310
1312101
320
0302010
321
)x(f)x(f3)x(f3)x(f8
h33
a-bh ; L(x)dxf(x)dx
3210
b
a
b
a
• Lagrange interpolation
• Integrate to obtain the rule
Simpson’s 3/8-Rule
Ch 12 - Integral Calculus 8/8/2019
10
19
Example: Simpson’s Rules
Evaluate the integral
• Simpson’s 1/3-Rule: Approximation Error
ε = -(b – a)5/(2880) f (4)(η) (η is a number between a and b).
Since f (4)(η) >0, the error is negative (overshooting).
Let’s take η=2.5. Then,
ε = -45/(2880) [{exp(2*2.5)*[16*(2.5)+32]} =-3799.3769
• Simpson’s 1/3-Rule
dxxe4
0
x2
4 2
0
4 8
(0 ) 4 (2 ) (4 )3
20 4(2 ) 4 8240 .411
35216 .926 8240 .411
57 .96%5216 .926
x hI xe dx f f f
e e
20
Example: Simpson’s Rules
Evaluate the integral
• Simpson’s 3/8-Rule
• Simpson’s 3/8-Rule: Approximation Error
ε = -(b – a)5/(6480) f (4)(η) (η is a number between a and b).
dxxe4
0
x2
4 2
0
3 4 8(0) 3 ( ) 3 ( ) (4)
8 3 3
3(4/3)0 3(19.18922) 3(552.33933) 11923.832 6819.209
85216.926 6819.209
30.71%5216.926
x hI xe dx f f f f
Ch 12 - Integral Calculus 8/8/2019
11
21
Midpoint Rule
• Newton-Cotes Open Formula
where η ∊ [a,b].
)(f24
)ab()
2
ba(f)ab(
)x(f)ab(dx)x(f3
m
b
a
a b x
f(x)
xm
The midpoint rule amounts to compute the area of the rectangle
• Note: This rule does not make any use of the end points.
22
Two-point Newton-Cotes Open Formula
• Approximate by a straight line
)(f108
)ab()x(f)x(f
2
abdx)x(f
3
21
b
a
x0 x1x
f(x)
x2h h x3h
Ch 12 - Integral Calculus 8/8/2019
12
23
Three-point Newton-Cotes Open Formula
• Approximate by a parabola
)(f23040
)ab(7
)x(f2)x(f)x(f23
abdx)x(f
5
321
b
a
x0 x1x
f(x)
x2h h x3h h x4
24
Better Numerical Integration
• Composite integration
– Composite Trapezoidal Rule
– Composite Simpson’s Rule
• Richardson Extrapolation
• Romberg integration
Ch 12 - Integral Calculus 8/8/2019
13
25
Composite Trapezoid Rule
0
1
2
3
4
5
6
7
3 5 7 9 11 13 15
Two segments
0
1
2
3
4
5
6
7
3 5 7 9 11 13 15
Four segments
0
1
2
3
4
5
6
7
3 5 7 9 11 13 15
Many segments
0
1
2
3
4
5
6
7
3 5 7 9 11 13 15
Three segments
To improve the Trapezoid Rule, first splits the interval of integration [a,b] into N smaller, uniform subintervals, and then applies the trapezoidal rule on each of them.
26
Composite Trapezoid Rule
)x(f)x(f2)2f(x)f(x2)f(x2
h
)f(x)f(x2
h)f(x)f(x
2
h)f(x)f(x
2
h
f(x)dxf(x)dxf(x)dxf(x)dx
n1ni10
n1n2110
x
x
x
x
x
x
b
a
n
1n
2
1
1
0
x0 x1x
f(x)
x2h h x3h h x4
n
abh
• Use the Trapezoid Rule in n intervals. Then, add them together.
Ch 12 - Integral Calculus 8/8/2019
14
27
Composite Trapezoid Rule
• Evaluate the integral dxxeI4
0
x2
%66.2 95.5355
)4(f)75.3(f2)5.3(f2
)5.0(f2)25.0(f2)0(f2
hI25.0h,16n
%50.10 76.5764)4(f)5.3(f2 )3(f2)5.2(f2)2(f2)5.1(f2
)1(f2)5.0(f2)0(f2
hI5.0h,8n
%71.39 79.7288)4(f)3(f2
)2(f2)1(f2)0(f2
hI1h,4n
%75.132 23.12142)4(f)2(f2)0(f2
hI2h,2n
%12.357 66.23847)4(f)0(f2
hI4h,1n
28
Composite Trapezoid Rule with Unequal Segments
• Evaluate the integral
Use the following hi‘s: {h1=2, h2=1, h3=0.5, h4=0.5}
dxxeI4
0
x2
%45.14 58.5971 45.3
2
0.5
5.33e 2
0.532
2
120
2
2
)4()5.3(2
)5.3()3(2
)3()2(2
)2()0(2
)()()()(
87
76644
43
21
4
5.3
5.3
3
3
2
2
0
ee
eeee
ffh
ffh
ffh
ffh
dxxfdxxfdxxfdxxfI
Ch 12 - Integral Calculus 8/8/2019
15
29
Composite Simpson’s Rule
x0 x2x
f(x)
x4h h xn-2h xn
n
abh
…...
• Piecewise Quadratic approximations
hx3x1 xn-1
30
Composite Simpson’s Rule
• Evaluate the integral
• Using n = 2, h = 2
• Using n = 4, h = 1
dxxeI4
0
x2
%70.8 975.5670
e4)e3(4)e2(2)e(403
1
)4(f)3(f4)2(f2)1(f4)0(f3
hI
8642
%96.57 411.8240e4)e2(403
2
)4(f)2(f4)0(f3
hI
84
Ch 12 - Integral Calculus 8/8/2019
16
31
Composite Simpson’s Rule with Unequal Segments
• Evaluate the integral
Using h1 = 1.5, h2 = 0.5
dxxeI4
0
x2
%76.3 23.5413
4)5.3(433
5.03)5.1(40
3
5.1
)4()5.3(4)3(3
)3()5.1(4)0(3
)()(
87663
2
1
4
3
3
0
eeeee
fffh
fffh
dxxfdxxfI
32
• Newton-Cotes Formulae
- Nodes: Use evenly-spaced functional values
- Weights: Derived from an approximation required to be equal for a polynomial of order lower or equal to the degree of the polynomials used to approximate the function. Given nodes, best!
- Problem: Can explode for large n (Runge’s phenomenon)
• Q: Can we use more efficient weights and nodes? Yes!
• Gaussian Quadratures
- Gaussian quadrature sets the nodes and the weights in such a way that the approximation is exact when f(.) is a low order polynomial. Best choice for both, nodes and weights!
Gaussian Quadratures
Ch 12 - Integral Calculus 8/8/2019
17
33
Gaussian Quadratures
• In fact, Gaussian quadrature is more general than simple integration, it computes an approximation to the weighted integral:
• Gaussian Quadratures
- Select functional values at non-uniformly distributed points to achieve higher accuracy. The values are not predetermined, but unknowns to be determined.
- Change of variables the interval of integration is [-1,1].
- Gauss-Legendre formulae for nodes and weights
- With n nodes, delivers exact answer if f is (2n-1)th-order polynomial.
n
iii
b
a
xfcf(x)c(x)dx1
)(
34
n
iii xfcf(x)dx
1
1
1
)(
• The Gauss-Legendre quadrature rule is in general stated as:
the ci's are weights, the xi's are the quadrature nodes. The approximation error term, ε, is called the truncation error of integration.
• Goal: Get an exact answer if f is a (2n-1)th-order polynomial. (With n=5, we get an exact answer f is a 9th-order polynomial).
• For Gauss-Legendre quadrature, the nodes are chosen to be zeros of certain Legendre polynomials. They are not trivial to compute.
Gaussian Quadratures
Ch 12 - Integral Calculus 8/8/2019
18
35
n
iicdxf(x)dx
1
1
1
1
1
121• To get right answer for f(x) = 1
• To get right answer for f(x) = x
• To get right answer for f(x) = x2
• To get right answer for f(x) = xj for j=0,...., 2n-1
A system of 2n equations in 2n unknowns (n ci,’s & nΨi,’s).
Gaussian Quadratures – Nodes and Weights
n
iiicxdx
1
1
1
0
n
iiicdxx
1
21
1
2 3/2
n
i
jii
j cdxx1
1
1
36
• By construction we get right answer for
f(x) = 1 (j=0), f(x) = x (j=1), ...., f(x) = xj 1 (j=2n-1),
enough to get the right answer for any polynomial of order 2n-1.
• Example: n=2 A system of 4 equations and 4 unknowns:
Gaussian Quadratures – Nodes and Weights
3
1x
3
1x
1c1c
xcxc0dxx xf
xcxc3
2dxx xf
xcxc0xdx xf
cc2dx1 1f
2
1
2
1
322
31
1
1 133
222
21
1
1 122
221
1
1 1
2
1
1 1
Ch 12 - Integral Calculus 8/8/2019
19
37
Change of Interval for Gaussian Quadrature
• Coordinate transformation from [a,b] to [-1,1]
This can be done by an affine transformation on t and a change of variables.
t2t1a b
n
iii
b
axfc
abdx
ababx
abfdttf
1
1
1)(
2)
2()
22()(
btx
atx
dxab
dt
abx
abt
1
12
22
38
Gaussian Quadrature on [-1, 1]
• For n=2, we have four unknowns: c1, c2, x1, x2.
• We have already solved this problem:
c1=1; c2=1; x1,=-1/√3; & x2 ,= 1/√3.
)x(fc)x(fc)x(fc)x(fcdx)x(f nn2211i
1
1
n
1ii
)f(xc)f(xc
f(x)dx :2n
2211
1
1
x2x1-1 1
• Gauss Quadrature General formulation:
Ch 12 - Integral Calculus 8/8/2019
20
39
Gaussian Quadrature on [-1, 1]
• Approximation formula for n = 2:
)3
1()
3
1()(
1
1ffdxxfI
40
Gaussian Quadrature on [-1, 1]
• Now, choose (c1, c2, c3, x1, x2, x3) such that the method yields “exact integral” for f(x) = x0, x1, x2, x3,x4, x5. (Again, (c1, c2, c3, x1, x2, x3) are calculated by assuming the formula gives exact expressions for integrating a fifth order polynomial).
)()()()( :3Case 332211
1
1xfcxfcxfcdxxfn
x3x1-1 1x2
Ch 12 - Integral Calculus 8/8/2019
21
41
Gaussian Quadrature on [-1, 1]
533
522
511
1
1
55
433
422
411
1
1
44
333
322
311
1
1
33
233
222
211
1
1
22
332211
1
1
321
1
1
0
5
2
0
3
2
0
21
xcxcxcdxxxf
xcxcxcdxxxf
xcxcxcdxxxf
xcxcxcdxxxf
xcxcxcxdxxf
cccxdxf
5/3
0
5/3
9/5
9/8
9/5
3
2
1
3
2
1
x
x
x
c
c
c
42
Gaussian Quadrature on [-1, 1]
• Approximation formula for n = 3
)5
3(
9
5)0(
9
8)
5
3(
9
5)(
1
1fffdxxfI
Ch 12 - Integral Calculus 8/8/2019
22
43
Example (1): Gaussian Quadrature
Evaluate
First, a coordinate transformation
• Two-point formula (n=2)
33.34%)( 543936.3477376279.3468167657324.9
)3
44()
3
44()
3
1()
3
1()( 3
44
3
441
1
eeffdxxfI
926477.5216dtteI4
0
t2
1
1
1
1
444
0
2 )()44(
2dxdt ;2222
dxxfdxexdtteI
xab
xab
t
xt
44
Example (1): Gaussian Quadrature• Three-point formula (n=3)
• Four-point formula (n=4)
4.79%)( 106689.4967
)142689.8589(9
5)3926001.218(
9
8)221191545.2(
9
5
)6.044(9
5)4(
9
8)6.044(
9
5
)6.0(9
5)0(
9
8)6.0(
9
5)(
6.0446.04
1
1
eee
fffdxxfI
%)37.0( 54375.5197
)339981.0()339981.0(652145.0
)861136.0()861136.0(34785.0)(1
1
ff
ffdxxfI
Ch 12 - Integral Calculus 8/8/2019
23
45
Example (2): Gaussian Quadrature
Evaluate 44949742.2
1 64.1
0
2
2
dxeIx
1
1
1
1
)]1(82[.2
164.1
0
2 )(2
82.
2
82.
2
1
82 );1(82.82.82.22
22
dxxfdxedteI
dx.dtxxab
xab
t
xt
First, a coordinate transformation:
46
Example: Gaussian Quadrature
• Three-point formula (n=3)
%)007.0(0.44946544
)0.19271450+0.63509351+9(0.5461465*.32713267
9
5
9
8
9
5
2
82.
)6.0(9
5)0(
9
8)6.0(
9
5
2
82.)(
2
82.
222 )]6.01(82[.2
1)]01(82[.
2
1)]6.01(82[.
2
1
1
1
eee
fffdxxfI
%)065.0( 44978962.)0.43323413+7(0.9417114*0.32713267
2
82.)
3
1()
3
1(
2
82.)(
2
82.22 )]
3
11(82[.
2
1)]
3
11(82[.
2
11
1
eeffdxxfI
• Two-point formula (n=2)
Ch 12 - Integral Calculus 8/8/2019
24
47
Monte Carlo Integration
• In our motivation of integrals, we evaluated a one-dimensional integral by a sum of rectangles, using the end points of each interval to measure the height. Some of these rectangles overestimated the area, some underestimated the area.
• Let’s focus on one of those rectangles, say with base [a, b]. We can also use as the height a randomly selected interior point, x1 ∊ [a, b] and estimate the integral, say I(x1). Of course, it may over- or under-estimate the area.
• But, we can randomly select N interiors points and get N estimations of the area. Some points will under-estimate, some points will over-estimate, but, statistical intuition suggests that the average may work.
• In fact, as N increases, the average of the integral converges to the integral.
48
Monte Carlo Integration – Example 1
• Example 1: We want to do MC integration for (exact integral = 5,216.92):
> M <- 200> x <- runif(M,0,4)> All_I <- matrix(0,M,1)> a <- 0> b <- 4> m <- 1> while (m <= M) {+ Int <- (b-a)*(x[m] * exp(2*x[m]))+ m <- m+ 1+ All_I[m] <- Int+ }> IN <- sum(All_I)/M> IN[1] 5489.388
dxxe4
0
x2
Ch 12 - Integral Calculus 8/8/2019
25
49
Monte Carlo Integration – Example 2
• MC Integration can be applied to any area, like the area of a circle.
• Example 2: We want to estimate the area of a circle with radius, r=2 (exact area = π * r2 = π * 4 =12.56637):
> M <- 100> x <- runif(M,-2,2)> y <- runif(M,-2,2)> #box is area 16.> distance.from.0 <- sqrt(x*x + y*y)> inside.circle <- (distance.from.0 < 2)> area <- 16*sum(inside.circle)/M> area[1] 12.48
50
Monte Carlo Integration – Properties
Ch 12 - Integral Calculus 8/8/2019
26
51
Monte Carlo Integration – Example 3
• Example 3: Back to the trapezoid example, where we wanted to integrate the following function:
> All_I <- matrix(0,M,1)> a <- 0> b <- 4> m <- 1> while (m <= M) {+ Int <- (b-a)/((b-a)/2)* (x[m] * exp(2*x[m]) + y[m] * exp(2*y[m]))+ m <- m+ 1+ All_I[m] <- Int+ }> IN <- sum(All_I)/M> IN[1] 5448.996
Note: The exact integral is 5,216.93.
dxxe4
0
x2
52
Monte Carlo Integration & Multiple Integrals
• Q: Why use the MC estimator instead of the also very simple determinist quadrature rules?
• Quadrature rules do not extend very well to higher dimension. An approach is to rewrite the problem in terms of one-dimensional integrals. For two or three dimension it may work well, but for more than four dimensions it become imprecise.
• These rules suffer from the curse of dimensionalitily.
• Monte Carlo integration extends well to many dimensions. IT is based on repeated function evaluations, not repeated integrations using one-dimensional methods.
Popular MC algorithm: Markov chain Monte Carlo (MCMC), which include the Metropolis-Hastings algorithm and Gibbs sampling.
Ch 12 - Integral Calculus 8/8/2019
27
53
Q: What's the integral of (1/cabin)d(cabin)?A: A natural log cabin!