Mathematicalmethods Lecture Slides Week7 1 LagrangesMethod

download Mathematicalmethods Lecture Slides Week7 1 LagrangesMethod

of 29

description

LagrangesMethod

Transcript of Mathematicalmethods Lecture Slides Week7 1 LagrangesMethod

  • AMATH 460: Mathematical Methodsfor Quantitative Finance

    7.1 Lagranges Method

    Kjell KonisActing Assistant Professor, Applied Mathematics

    University of Washington

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 1 / 29

  • Outline

    1 Optimal Investment Portfolios

    2 Relative Extrema of Functions of Several Variables

    3 Lagranges Method

    4 Example

    5 Minimum Variance Portfolio

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 2 / 29

  • Outline

    1 Optimal Investment Portfolios

    2 Relative Extrema of Functions of Several Variables

    3 Lagranges Method

    4 Example

    5 Minimum Variance Portfolio

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 3 / 29

  • Investment Portfolios

    Portfolio of n assetsLet wi be the proportion of the portfolio invested in asset iHave constraint n

    i=1wi = 1

    Can take long and short positions = no constraints on individual wiLet i be the expected rate of return on asset iLet 2i be the risk of asset iLet ij be the correlation between assets i and jExpected rate of return and risk of the portfolio:

    Expected Return =n

    i=1wii

    Risk =n

    i=1w2i 2i + 2

    1i

  • Investment Portfolios: Matrix Notation

    Let w = (w1, . . . ,wn) and = (1, . . . , n)The expected rate of return can be written in matrix notation as

    Return =n

    i=1wii = wT

    The risk can be written as

    Risk = wTw

    is the covariance matrix of the n assets

    =

    21 1212 1n1n2121 22 2n

    ... ... . . . ...n1n1 n2n2 2n

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 5 / 29

  • Optimal Investment Portfolios

    Given , , and investor selected w , can computeportfolio returnportfolio risk

    Two notions of optimality

    For a target expected return, choose w to minimize portfolio riskFor a target level of risk, choose w to maximize expected return

    Both notions are constrained optimization problems that can besolved using Lagrange multipliers

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 6 / 29

  • Optimal Investment Portfolios

    Minimum variance optimization

    n asset case

    minimize: wTwsubject to: eTw = 1

    Tw = P

    2 asset caseminimize: 21w21 + 212w1w2 + 22w22

    subject to: w1 + w2 = 11w1 + 2w2 = P

    Maximum expected return optimization

    n asset case

    maximize: Twsubject to: eTw = 1

    wTw = 2P

    2 asset casemaximize: 1w1 + 2w2

    subject to: w1 + w2 = 121w21 + 212w1w2 + 22w22 = 2P

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 7 / 29

  • Outline

    1 Optimal Investment Portfolios

    2 Relative Extrema of Functions of Several Variables

    3 Lagranges Method

    4 Example

    5 Minimum Variance Portfolio

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 8 / 29

  • Relative Extrema of Single Variable Functions

    A local minimum (maximum) of a function f is a point x0 where

    f (x0) () f (x) x (x0 , x0 + )

    for some > 0

    A local extrema is a point that is a local minimum or maximum

    If f is twice differentiable and f is continuousAny local extremum is a critical point of f : f (x0) = 0Can classify critical points using second derivative testf (x0) < 0 local maximumf (x0) > 0 local minimumf (x0) = 0 anything possible

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 9 / 29

  • Relative Extrema of Functions of n Variables

    A local minimum (maximum) of a function f : Rn R is a pointx0 Rn where

    f (x0) () f (x) x : x x0 <

    Every local extremum is a critical point: Df (x0) = 0

    If f is twice differentiable and has continuous second order partialderivatives

    D2f (x0) is a symmetric matrix with real eigenvaluesSecond order conditions

    All eigenvalues of D2f (x0) > 0 local minimumAll eigenvalues of D2f (x0) < 0 local maximumD2f (x0) has eigenvalues saddle pointD2f (x0) singular anything can happen

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 10 / 29

  • Finding Extrema: Functions of 2 Variables

    Find the local extrema of f (x , y) = x2 + xy + y2

    Df (x , y) =[2x + y x + 2y

    ]Df (0, 0) =

    [0 0

    ] (0, 0) is a critical point

    D2f (x , y) =[

    2 11 2

    ]

    Can use R to compute the eigenvalues> A eigen(A)$values[1] 3 1

    Since both eigenvalues are greater than 0 = (0, 0) a local minimumKjell Konis (Copyright 2013) 7.1 Lagranges Method 11 / 29

  • Finding Extrema: Functions of 2 Variables (Take 2)

    Find the local extrema of f (x , y) = x2 xy y2

    Df (x , y) =[ 2x y x 2y] Df (0, 0) = [0 0]

    (0, 0) is a critical point

    D2f (x , y) =[2 11 2

    ]

    Can use R to compute the eigenvalues> A eigen(A)$values[1] -1 -3

    Since both eigenvalues are less than 0 = (0, 0) a local maximumKjell Konis (Copyright 2013) 7.1 Lagranges Method 12 / 29

  • Finding Extrema: Functions of 2 Variables

    Find the local extrema of f (x , y) = x2 + 3xy + y2

    Df (x , y) =[2x + 3y 3x + 2y

    ]Df (0, 0) =

    [0 0

    ] (0, 0) is a critical point

    D2f (x , y) =[

    2 33 2

    ]

    Can use R to compute the eigenvalues> A eigen(A)$values[1] 5 -1

    One positive and one negative eigenvalue = (0, 0) a saddle pointKjell Konis (Copyright 2013) 7.1 Lagranges Method 13 / 29

  • Finding Extrema: Functions of 2 Variables

    Find the local extrema of f (x , y) = 2xy (1 y2) 32First order condition

    Df (x , y) =[

    2y 2x + 3y

    1 y2]

    Df (0, 0) =[0 0

    ]= (0, 0) is a critical point

    Second order condition

    D2f (x , y) =

    0 22 36y2

    1y2

    D2f (0, 0) = 0 2

    2 3

    Compute the eigenvalues of the Hessian at the critical point> eigen(matrix(c(0, 3, 3, 2), 2, 2))$values[1] 4.162278 -2.162278One positive and one negative eigenvalue = (0, 0) a saddle point

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 14 / 29

  • Outline

    1 Optimal Investment Portfolios

    2 Relative Extrema of Functions of Several Variables

    3 Lagranges Method

    4 Example

    5 Minimum Variance Portfolio

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 15 / 29

  • Lagranges Method

    Problem:maximize: f (x1, x2, . . . , xn)

    subject to: g1(x1, x2, . . . , xn) = 0g2(x1, x2, . . . , xn) = 0

    ...gm(x1, x2, . . . , xn) = 0

    (1)

    18th-century mathematician Joseph Louis Lagrange proposed thefollowing method for the solutionForm the function

    F (x1, . . . , xn, 1, . . . , m) = f (x1, . . . , xn) +mi=1

    igi(x1, x2, . . . , xn)

    Optimal value for problem (1) occurs at one of the critical points of FKjell Konis (Copyright 2013) 7.1 Lagranges Method 16 / 29

  • Lagranges Method

    Terminology:The function F (x1, . . . , xn, 1, . . . , m) is called the LagrangianThe column vector = (1, . . . , m) is called the Lagrangemultipliers vector

    Necessary Condition:Let x = (x1, x2, . . . , xn)Let g(x) =

    (g1(x), g2(x), . . . , gm(x)

    )be a vector-valued function

    of the constraintsThe gradient D

    (g(x)

    )must have full rank at any point where

    the constraint g(x) = 0 is satisfied, that is

    rank(Dg(x)

    )= m x where g(x) = 0

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 17 / 29

  • Partial Derivatives of the Lagrangian

    D F (x , ) has n + m variables, compute gradient in 2 parts

    D F (x , ) =[DxF (x , ) DF (x , )

    ]Recall Lagrangian:

    F (x , ) = f (x1, . . . , xn) +mi=1

    igi(x1, x2, . . . , xn)

    The partial derivatives are

    Fxj

    =fxj

    +mi=1

    igixj

    Fi

    = gi(x)

    Gradient of f : Df (x) =[fx1

    . . .fxn

    ]Kjell Konis (Copyright 2013) 7.1 Lagranges Method 18 / 29

  • Partial Derivatives of the Lagrangian

    Gradient of g(x):

    Dg(x) =

    g1x1

    g1x2

    g1xn

    g2x1

    g2x2

    g2xn

    ... ... . . . ...gmx1

    gmx2

    gmxn

    Can express sum in second term in matrix notationmi=1

    igixj

    = T[Dg(x)

    ]j

    It follows that

    D F (x , ) =[Df (x) + TDg(x)

    (g(x)

    )T]Kjell Konis (Copyright 2013) 7.1 Lagranges Method 19 / 29

  • Outline

    1 Optimal Investment Portfolios

    2 Relative Extrema of Functions of Several Variables

    3 Lagranges Method

    4 Example

    5 Minimum Variance Portfolio

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 20 / 29

  • Example

    Want tomax/min: 4x2 2x3

    subject to: 2x1 x2 x3 = 0x21 + x22 13 = 0

    Start by writing down the Lagrangian

    F (x , ) = f (x) + 1g1(x) + 2g2(x)

    = 4x2 2x3 + 1(2x1 x2 x3) + 2(x21 + x22 13)

    Check necessary condition:

    Dg(x) =[

    2 1 12x1 2x2 0

    ]

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 21 / 29

  • Derivatives of the LagrangianThe Lagrangian

    F (x , ) = 4x2 2x3 + 1(2x1 x2 x3) + 2(x21 + x22 13)

    Gradient of the Lagrangian

    D F (x , ) =

    21 + 22x1

    4 1 + 22x22 1

    2x1 x2 x3x21 + x22 13

    T

    Set D F (x , ) = 0 and solve for x and get 1 = 2 for free

    21 + 22x1 set= 04 1 + 22x2 set= 0

    2x1 x2 x3 set= 0x21 + x22 13 set= 0

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 22 / 29

  • Example (continued)

    A little algebra gives

    x1 =22

    x2 =32

    x3 =72

    Also know that

    x21 +x22 = 13 =( 22

    )2+

    (32

    )2=

    1322

    = 13 = 2 = 1

    The critical points are = (2,1), x = (2, 3,7), f (x) = 26 = (2, 1), x = (2,3, 7), f (x) = 26

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 23 / 29

  • Outline

    1 Optimal Investment Portfolios

    2 Relative Extrema of Functions of Several Variables

    3 Lagranges Method

    4 Example

    5 Minimum Variance Portfolio

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 24 / 29

  • Minimum Variance Portfolio

    Recall: minimum variance portfolio optimization

    minimize: wTwsubject to: eTw = 1

    Tw = PLagranges method setup

    f (w) = wTw

    g(w) =[g1(w)g2(w)

    ]=

    [Tw P = 0eTw 1 = 0

    ]

    First, check necessary condition

    Dg(x) =[T

    eT

    ]

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 25 / 29

  • Derivative of a Quadratic Form

    Let A =[a bb c

    ]

    Let f (x) = xTAx = ax21 + 2bx1x2 + cx22Then Df (x) =

    [2ax1 + 2bx2 2bx1 + 2cx2

    ]= 2xTA

    In general, let A be an n n symmetric matrixThe derivative (gradient) of the quadratic form f (x) = xTAx is

    Df (x) = 2xTA

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 26 / 29

  • Minimum Variance Portfolio

    The Lagrangian

    F (y , ) = wTw + 1[eTw 1]+ 2[Tw P]

    Gradient of the Lagrangian

    D F (w , ) =[Df (w) + T

    (Dg(w)

    ) (g(w)

    )T]=

    [2wT + 1eT + 2T eTw 1 Tw P

    ]Find the critical point by solving the linear system2 e eT 0 0

    T 0 0

    w12

    = 01P

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 27 / 29

  • Minimum Variance Portfolio

    Further reading:Second order conditions, e.g., Theorem 9.2 and Corollary 9.1 inPFME

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 28 / 29

  • http://computational-finance.uw.edu

    Kjell Konis (Copyright 2013) 7.1 Lagranges Method 29 / 29

    Optimal Investment PortfoliosRelative Extrema of Functions of Several VariablesLagrange's MethodExampleMinimum Variance Portfolio