Cubic Interpolation Report

27
The Cubic Interpolation Method Erika Paula L. Moises Joe Mari D. Reyes Math 288 (Numerical Optimization) May 4, 2015 Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

description

Cubic Interpolation for Line Search in Optimization.

Transcript of Cubic Interpolation Report

  • The Cubic Interpolation Method

    Erika Paula L. MoisesJoe Mari D. Reyes

    Math 288 (Numerical Optimization)

    May 4, 2015

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • Presentation Outline

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    The cubic interpolation method approximates the objectivefunction

    g() = f (x + d)

    by a cubic polynomial, where f is continuously differentiable. Bythe chain rule, we have

    g () =dg()

    d= f (x + d)Td

    We assume that g (0) = f (x)Td < 0, that is, d is the descentdirection at x .

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    To construct the cubic polynomial p(t), four interpolationconditions are required.

    For example, we may use:

    function values at four points, or

    function values at three points and a derivative value at onepoint; or

    function values and derivative values at two points

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    In this report, we will focus on interpolation using function valuesand derivative values at two points a and b.

    To guarantee that the appropriate minimum occurs in the interval[a, b], the following condition should hold:

    g (a) < 0 and g (b) > 0

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    The central idea is to use the derivative information in the endpoints of the interval. Together with the function values, theminimum of the given function is approximated by the minimum ofa cubic polynomial.

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    We construct a cubic polynomial of the form

    p(t) = c1(t a)3 + c2(t a)2 + c3(t a) + c4 (1)

    where ci are the coefficients of the polynomial which are chosensuch that

    p(a) = c4 = g(a),

    p(a) = c3 = g (a),

    p(b) = c1(b a)3 + c2(b a)2 + c3(b a) + c4 = g(b),p(b) = 3c1(b a)2 + 2c2(b a) + c3 = g (b).

    (2)

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    From the sufficient condition of the minimizer, we have

    p(t) = 3c1(t a)2 + 2c2(t a) + c3 = 0 (3)

    andp(t) = 6c1(t a) + 2c2 > 0 (4)

    Solving Equation (3) yields

    t = a +c2

    c22 3c1c3

    3c1, if c1 6= 0 (5)

    t = a c32c2

    , if c1 = 0. (6)

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    In order to guarantee that condition (4) holds, we only take thepositive in Equation (5).

    By equation (5), we get:

    t a =c2 +

    c22 3c1c3

    3c1=

    c3c2 +

    c22 3c1c3

    (7)

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    When c1 = 0, Equation (7) is just Equation (6). Then theminimizer of p(t) is

    t = a c3c2 +

    c22 3c1c3

    (8)

    The minimizer in Equation (8) is represented by c1,c2 and c3.

    We hope to represent t by g(a),g(b), g (a) and g (b) directly.

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    We now let

    s = 3g(b) g(a)

    b az = s g (a) g (b),w2 = z2 g (a)g (b) (9)

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    By use of Equation (2), we get

    s = 3g(b) g(a)

    b a = 3[c1(b a)2 + c2(b a) + c3],

    z = s g (a) g (b) = c2(b a) + c3,w2 = z2 g (a)g (b) = (b a)2(c22 3c1c3).

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    Then

    (b a)c2 = z c3c22 3c1c3 =

    w

    b a

    And so

    c2 +

    c22 3c1c3 =z + w c3

    b a (10)

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    Using c3 = g(a) and substituting Equation (10) into Equation (8),

    we get:

    t a = (b a)g(a)

    z + w g (a) , (11)

    which is

    t a = (b a)g(a)g (b)

    (z + w g (a))g (b) =(b a)(z2 w2)

    g (b)(z + w) (z2 w2)

    =(b a)(w z)g (b) z + w (12)

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    Note that the denominator in Equation (12) may be equal to zeroor sufficiently small. We can overcome this by the use of Equation(11) on (12). We get

    t a = (b a)g(a)

    z + w g (a) =(b a)(w z)g (b) z + w

    =(b a)(g (a) + w z)

    g (b) g (a) + 2w

    = (b a)(1 g(b) + z + w

    g (b) g (a) + 2w ) (13)

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    This gives us the iterates of the Cubic Interpolation Method:

    t = a + (b a) w g(a) z

    g (b) g (a) + 2w (14)

    where

    s = 3g(b) g(a)

    b a , z = s g(a) g (b)

    w2 = z2 g (a)g (b)

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • The Cubic Interpolation Method

    In Equation (13) and (14), the denominator

    g (b) g (a) + 2w 6= 0

    .In fact, since g (a) < 0 and g (b) > 0, then

    w2 = z2 g (a)g (b) > 0

    .Taking w > 0, it follows that the denominator

    g (b) g (a) + 2w > 0

    .

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • Remark

    Cubic interpolation has the danger of lack of convergencenumerically, but the theoretical convergence of the iterate is veryfast.

    It can be shown that the cubic interpolation method convergessuperlinearly.

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • Presentation Outline

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • Cubic Interpolation Steps

    We now lay down the steps for the Cubic Interpolation Method

    Step 1 (Determination of the Initial Interval)

    Let s > 0 be some scalar.

    Evaluate g(t) and g (t) at the points t = 0, s, 2s, 4s, 8s, ... untiltwo successive points a and b are found such that either g (b) 0or g(b) g(a).

    Then, it can be seen that a local minimum of g exists within theinterval (a, b].

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • Cubic Interpolation Steps

    Step 2 (Updating of the Current Interval)

    Given the current interval [a, b], a cubic polynomial is fitted to thefour values g(a), g (a), g(b) and g (b).

    The cubic can be shown to have a unique minimum t in theinterval (a, b] given by

    t = a + (b a) w g(a) z

    g (b) g (a) + 2w ,

    where z = 3g(b)g(a)ba g (a) g (b), w =

    z2 g (a)g (b)

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • Cubic Interpolation Steps

    Step 2 (Updating of the Current Interval)

    If g (t) 0 or g (t) g(a) replace b by t.

    If g (t) < 0 and g (t) < g(a) replace a by t.

    Stopping Criterion

    In practice, the computation is terminated once the length of thecurrent interval becomes smaller than a specified tolerance or elsewe obtain t = b.

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • Presentation Outline

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • Example 1

    Cubic interpolation is applied to find the minimum f (x) = x + 16x+1with starting interval [2, 4.5] and accuracy tol = 0.01.

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • Example 1

    One iteration after reaching the stopping criterion has been givenin Table 1. For this case, the interval changes very fast around theminimum point.

    k ak bk xk f(xk) f(xk)

    1 2.000 4.500 3.024 7.0001 0.0122 2.000 3.024 2.997 7.0000 -0.0013 2.997 3.024 3.000 7.0000 0.0004 2.997 3.000 3.000 7.0000 0.000

    Table: Cubic interpolation for f (x) = x + 16x+1 ,[a0, b0] = [2, 4.5],tol = 0.01

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • References:

    Bertsekas, Dimitri P. 1999. Nonlinear Programming. 2ndEdition. pp 723-724

    Sun, Wenyu. Yuan, Ya-Xiang. 2006. Optimization Theoryand Methods (Nonlinear Programming). Volume 1. pp 98-101

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method

  • Thank you

    Erika Paula L. Moises, Joe Mari D. Reyes The Cubic Interpolation Method