lecture slides linalg

Upload: shanky1124

Post on 03-Apr-2018

222 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/29/2019 Lecture Slides Linalg

    1/23

    Financial Engineering and Risk ManagementReview of vectors

    Martin Haugh Garud Iyengar

    Columbia UniversityIndustrial Engineering and Operations Research

  • 7/29/2019 Lecture Slides Linalg

    2/23

    Reals numbers and vectors

    We will denote the set of real numbers by R

    Vectors are finite collections of real numbers

    Vectors come in two varieties

    Row vectors: v =v1 v2 . . . vn

    Column vectors w =

    w1w2

    .

    ..wn

    By default, vectors are column vectors

    The set of all vectors with n components is denoted by Rn

    2

  • 7/29/2019 Lecture Slides Linalg

    3/23

    Linear independence

    A vector w is linearly dependent on v1, v2 if

    w =1

    v1

    +2

    v2

    for some1, 2

    R

    Example:

    2

    6

    4

    = 2

    1

    1

    0

    + 4

    0

    1

    1

    Other names: linear combination, linear spanA set V= {v1, . . . , vm} are linearly independent if no vi is linearlydependent on the others, {vj : j= i}

    3

  • 7/29/2019 Lecture Slides Linalg

    4/23

    Basis

    Every w Rn is a linear combination of the linearly independent set

    B =

    1

    0

    ...0

    ,

    0

    1

    ...0

    , . . .

    0

    0

    ...1

    w = w1

    1

    0

    ...0

    e1+w2

    0

    1

    ...0

    e2+ . . . + wn

    0

    0

    ...1

    enBasis any linearly independent set that spans the entire spaceAny basis for Rn has exactly n elements

    4

  • 7/29/2019 Lecture Slides Linalg

    5/23

    Norms

    A function (v) of a vector v is called a norm if

    (v) 0 and (v) = 0 implies v = 0

    (v) = || (v) for all R

    (v1 + v2) (v1) + (v2) (triangle inequality)

    generalizes the notion of length

    Examples:2 norm: x2 =

    ni=1

    |xi|2 ... usual length

    1 norm: x1 =n

    i=1|xi|

    norm: x = max1in |x|i

    p norm, 1 p < : xp =n

    i=1|x|pi

    1p

    5

  • 7/29/2019 Lecture Slides Linalg

    6/23

    Inner product

    The inner-product or dot-product of two vector v, w Rn is defined as

    v w =n

    i=1

    viwi

    The 2 norm v2 =

    v vThe angle between two vectors v and w is given by

    cos() =v w

    v2w

    2

    Will show later: v w = vw = product of v transpose and w

    6

  • 7/29/2019 Lecture Slides Linalg

    7/23

    Financial Engineering and Risk ManagementReview of matrices

    Martin Haugh Garud Iyengar

    Columbia UniversityIndustrial Engineering and Operations Research

  • 7/29/2019 Lecture Slides Linalg

    8/23

  • 7/29/2019 Lecture Slides Linalg

    9/23

    Matrix Operations: Transpose

    Transpose: A Rmd

    A =

    a11 a12 . . . a1da21 a22 . . . a2d

    ......

    . . ....

    am1 am2 . . . amd

    =

    a11 a21 . . . am1a12 a22 . . . ad2

    ......

    . . ....

    a1d a2d . . . amd

    Rdm

    Transpose of a row vector is a column vector

    Example:

    A =

    2 3 7

    1 6 5

    : 2 3 matrix ... A =

    2 1

    3 6

    7 5

    : 3 2 matrix

    v =

    26

    4

    : column vector ... v =

    2 6 4

    : row vector

    3

  • 7/29/2019 Lecture Slides Linalg

    10/23

    Matrix Operations: Multiplication

    Multiplication: A Rmd, B Rdp then C = AB Rmp

    cij =ai1 ai2 . . . aid

    b1j

    b2j...bdj

    row vector v R1d times column vector w Rd1 is a scalar.

    Identity times any matrix AIn = ImA = A

    Examples:

    2 3 7

    1 6 52

    6

    4 = 2(2) + 3(6) + 7(4)1(2) + 6(6) + 5(4) =

    50

    582 norm:

    1

    2

    2

    =

    12 + (2)2 =

    1 2

    12

    =

    1

    2

    1

    2

    inner product: v w = vw

    4

  • 7/29/2019 Lecture Slides Linalg

    11/23

    Linear functions

    A function f : Rd Rm is linear if

    f(x + y) = f(x) + f(y), , R, x, y

    Rd

    A function f is linear if and only if f(x) = Ax for matrix A Rmd

    Examples

    f(x) : R3 R: f(x) = 2 3 4x1x2x3

    = 2x1 + 3x2 + 4x3f(x) : R3 R2: f(x) =

    2 3 4

    1 0 2

    x1x2x3

    =

    2x1 + 3x2 + 4x3

    x1 + 2x3

    Linear constraints define sets of vectors that satisfy linear relationships

    Linear equality: {x : Ax = b} ... line, plane, etc.

    Linear inequality: {x : Ax b} ... half-space

    5

  • 7/29/2019 Lecture Slides Linalg

    12/23

    Rank of a matrix

    column rank of A Rmd = number of linearly independent columnsrange(A) = {y : y = Ax for some x}

    column rank of A = size of basis for range(A)

    column rank of A = m range(A) = Rm

    row rank of A = number of linearly independent rows

    Fact: row rank = column rank min{m, d}Example:

    A =

    1 2 3

    2 4 6

    , rank = 1, range(A) =

    1

    2

    : R

    A Rnn and rank(A) = n A invertible, i.e. A1 Rnn

    A1A = AA1 = I

    6

  • 7/29/2019 Lecture Slides Linalg

    13/23

    Financial Engineering and Risk ManagementReview of linear optimization

    Martin Haugh Garud Iyengar

    Columbia UniversityIndustrial Engineering and Operations Research

  • 7/29/2019 Lecture Slides Linalg

    14/23

    Hedging problem

    d assets

    Prices at time t= 0: p Rd

    Market in m possible states at time t= 1

    Price of asset j in state i = Sij

    Sj =S1j

    S2j...Smj

    S = S1 S2 . . . Sd = S11 S12 . . . S1d

    S21 S22 . . . S2d......

    . . ....

    Sm1 Sm2 . . . Smd

    Rmd

    Hedge an obligation X RmHave to pay Xi if state i occurs

    Buy/short sell = (1, . . . , d) shares to cover obligation

    2

  • 7/29/2019 Lecture Slides Linalg

    15/23

    Hedging problem (contd)

    Position Rd purchased at time t= 0j = number of shares of asset j purchased, j = 1, . . . , d

    Cost of the position = dj=1

    pjj = p

    Payoff from liquidating position at time t= 1

    payoffyi in state i: yi = dj=1 SijjStacking payoffs for all states: y = SViewing the payoff vector y: y range(S)

    y = S1 S2 . . . Sd12...d

    =

    d

    j=1

    jSj

    Payoff y hedges X if y X.

    3

  • 7/29/2019 Lecture Slides Linalg

    16/23

    Hedging problem (contd)

    Optimization problem:

    min dj=1 pjj ( p)subject to dj=1 Sijj Xi, i= 1, . . . , m ( S X)Features of this optimization problem

    Linear objective function: p

    Linear inequality constraints: S X

    Example of a linear program

    Linear objective function: either a min/max

    Linear inequality and equality constraintsmax/minx c

    xsubject to Aeqx = beq

    Ainx bin

    4

  • 7/29/2019 Lecture Slides Linalg

    17/23

    Linear programming duality

    Linear programP = minx c

    x

    subject to Ax bDual linear program

    D = maxu bu

    subject to Au = cu

    0

    Theorem.

    Weak Duality: P DBound: x feasible for P, u feasible for D, cx

    P

    D

    bu

    Strong Duality: Suppose P or D finite. Then P= D.

    Dual of the dual is the primal (original) problem

    5

    M d l l

  • 7/29/2019 Lecture Slides Linalg

    18/23

    More duality results

    Here is another primal-dual pair

    minx c

    xsubject to Ax = b = maxu b

    usubject to Au = c

    General idea for constructing duals

    P = min{cx : Ax b} min{cx u(Ax b) : Ax b} for all u 0 bu + min{(c Au)x : x Rn}=

    bu Au = c

    otherwise max{bu : Au = c}

    Lagrangian relaxation: dualize constraints and relax them!

    6

  • 7/29/2019 Lecture Slides Linalg

    19/23

    Financial Engineering and Risk ManagementReview of nonlinear optimization

    Martin Haugh Garud Iyengar

    Columbia UniversityIndustrial Engineering and Operations Research

    U t i d li ti i ti

  • 7/29/2019 Lecture Slides Linalg

    20/23

    Unconstrained nonlinear optimization

    Optimization problemminxRn f(x)

    Categorization of minimum pointsx global minimum if f(y) f(x) for all y

    xloc local minimum if f(y) f(xloc) for all y such that y x

    loc r

    Sufficient condition for local min

    gradient f(x) =

    fx1...f

    xn

    = 0: local stationarity

    Hessian 2f(x) = 2f

    x21

    2f

    x1x2. . .

    2f

    x1xn

    ... ... . . . ...2f

    xnx1

    2f

    xnx2. . .

    2f

    x2n

    positive semidefiniteGradient condition is sufficient if the function f(x) is convex.

    2

    U t i d li ti i ti

  • 7/29/2019 Lecture Slides Linalg

    21/23

    Unconstrained nonlinear optimization

    Optimization problem

    minxR2 x2

    1+ 3x1x2 + x

    3

    2

    Gradient

    f(x) =

    2x1 + 3x23x1 + 3x

    22

    = 0 x = 0,

    9

    4

    3

    2

    Hessian at x: H =

    2 3

    3 6x2

    x = 0: H =

    2 3

    3 0

    . Not positive definite. Not local minimum.

    x = 9

    43

    2

    : H =

    2 3

    3 9

    . Positive semidefinite. Local minimum

    3

    L i th d

  • 7/29/2019 Lecture Slides Linalg

    22/23

    Lagrangian method

    Constrained optimization problemmaxxR2 2 ln(1 + x1) + 4 ln(1 + x2),

    s.t. x1 + x2 =12

    Convex problem. But constraints make the problem hard to solve.

    Form a Lagrangian function

    L(x, v

    ) =2 ln

    (1

    +x1

    ) +4 ln

    (1

    +x2

    ) v

    (x1

    +x2

    12

    )

    Compute the stationary points of the Lagrangian as a function of v

    L(x, v) = 2

    1+x1 v

    4

    1+x2

    v = 0 x1 =2

    v 1, x2 = 4

    v 1

    Substituting in the constraint x1 + x2 = 12, we get

    6

    v= 14 v= 3

    7 x = 1

    3

    11

    25

    4

    Portfolio Selection

  • 7/29/2019 Lecture Slides Linalg

    23/23

    Portfolio Selection

    Optimization problem

    maxx x

    xVx

    s.t. 1x = 1

    Constraints make the problem hard!

    Lagrangian function

    L(x, v) = x

    xVx

    v(1x

    1)

    Solve for the maximum value with no constraints

    xL(x, v) = 2Vx v1 = 0 x =1

    2 V1( v1)

    Solve for v from the constraint

    1x = 1 1V1( v1) = 2 v= 1

    V1

    21V11

    Substitute back in the expression for x

    5