040504

Upload: sumit-ray

Post on 03-Jun-2018

217 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/12/2019 040504

    1/19

    Linear Complementarity Problems and their Sources

    The Linear Complementarity Problem (LCP) (q, M) is defined as

    follows:

    Given a real nn matrix M and an n-vector q, find z Rn suchthat

    z0, q+ M z0, zT(q+ M z) = 0.

    Define the mapping F(z) :=q+ M z.

    ThenFis an affine transformation from Rn into itself.

  • 8/12/2019 040504

    2/19

    Some notation

    Given the LCP (q, M)we write

    FEA(q, M) ={z :q+ M z0, z 0}

    and

    SOL(q, M) ={z :q+ M z0, z0, zT(q+ M z) = 0.}

    These are thefeasible set (region)andsolution setof the LCP(q, M),

    respectively.

    Note that ifFEA(q, M)=, it is a closed polyhedral set.

  • 8/12/2019 040504

    3/19

    How do such problems arise?

    Optimality criterion for Linear Programming(LP)

    Consider the LP

    (P)minimize cTxsubject to Ax b

    x 0.

    According to the theory of linear programming, a vector x is optimal

    for (P) if and only if it is feasible and there exists a vector ysuch that

    yT

    A cT

    , y0, yT

    (Ax b) = 0, (yT

    A cT

    )x= 0.

  • 8/12/2019 040504

    4/19

    Now arrange these conditions as follows:

    u = c + ATy 0

    v = b + Ax 0

    x 0, y0

    xTu= 0, yTv= 0.

    Next define

    w=

    u

    v

    , q=

    c

    b

    , M=

    0 AT

    A 0

    , z =

    x

    y

    .

    The optimality conditions of the LP then become the LCP(q, M).

  • 8/12/2019 040504

    5/19

    Optimality conditions for Quadratic Programming(QP)

    Consider the QP

    (P)

    minimize cTx + 12

    xTQx

    subject to Ax bx 0.

    According to the Karush-Kuhn-Tucker (KKT) Theorem, if the vector

    x is a local minimizer for (P), there exists a vector y such that

    c+QxATy0, y0, yT(Axb) = 0, xT(c+QxATy) = 0.

  • 8/12/2019 040504

    6/19

    If we assemble these conditions along with the feasibility of the vector

    x, we obtain the LCP (q, M)where

    w= u

    v , q=

    c

    b , M=

    Q AT

    A 0 , z =

    x

    y .

    Remark. The above necessary conditions of optimality for QP are also

    sufficient for (global) optimality whenQis positive semi-definite.

    Note that ifQis positive semi-definite, then so is

    M= Q AT

    A 0 .

  • 8/12/2019 040504

    7/19

    Bimatrix Games as LCPs

    The initial set up

    LetA andB denote twom nmatrices.

    These are payoff matrices for Players I and II, respectively.

    Letm={x Rm+ :eTx= 1}andn={yRn+:e

    Ty= 1}.

    If x m and y n, the expected lossesof Players I and II are,respectively:

    xTAy and xTBy.

    Let(A, B)denote the corresponding two person game.

  • 8/12/2019 040504

    8/19

    Nash Equilibrium Point of (A,B)

    The pair (x, y) m n is a Nash Equilibrium Point(NEP) for(A, B)if

    (x)TAy xTAy for all x m

    (x)TBy (x)TBy for allyn

    It is crucial to note that given(x, y) mn, each of the vectorsx, y is optimal in a simple linear program defined in terms of the

    other. The LPs are:

    minimize(Ay)Tx subject to eTx= 1, x 0

    and

    minimize(BTx)Ty subject to eTy= 1, y0

  • 8/12/2019 040504

    9/19

    Let E be the m n matrix whose entries are all 1. For a suitablescalar >0, all the entries of the matrices A + EandB+ Eare

    positive.

    It is easy to see that (A, B)and(A + E,B + E)have the same

    equilibrium points (if any).

    Thus, it is not restrictive to assume that A andB are (elementwise)

    positive matrices.

  • 8/12/2019 040504

    10/19

    Now consider the LCP

    u =em+ Ay0, x 0, xTu= 0

    v =en+ BTx 0, y0, yTv = 0

    In this case, we have

    w=

    u

    v

    , q=

    em

    en

    , M=

    0 A

    BT 0

    , z=

    x

    y

    ,

    where

    em= (1, . . . , 1) Rm en= (1, . . . , 1) R

    n.

    We wish to show that

    to every solution of this LCP, there corresponds a Nash equi-

    librium point of(A, B)and vice versa.

  • 8/12/2019 040504

    11/19

    The correspondences are as follows:

    If(x, y)is a Nash equilibrium of(A, B), then

    (x, y) = (x/(x)TBy, y/(x)TAy)

    solves the LCP (q, M)given above.

    If(x, y)solves the LCP (above), then

    (x, y) = (x/eTmx, y/eTny

    )

    is a Nash equilibrium point for (A, B).

  • 8/12/2019 040504

    12/19

    A Market Equilibrium Problem

    Here we seek to determine prices at which there is a balance between

    supplies and demands.

    The supply sideminimize cTx

    subject to Ax bBx r

    x 0

    The demand side

    r=Q(p) =Dp + d

    Equilibration

    p =

  • 8/12/2019 040504

    13/19

    Formulation as an Equilibrium Problem

    y =c ATv BT 0, xT0, (x)Ty = 0

    u =b + Ax 0, v 0, (v)Tu = 0

    =r + Bx 0, 0, ()T = 0

    SubstituteDp

    + dforr

    and

    forp

    .

    Then we get the LCP (q, M)with

    q=

    cbd

    M=

    0 AT BTA 0 0

    B 0 D

    and z=

    xv

    What sort of matrix is D? Having it be negative (semi)definite andsymmetric would be nice.

  • 8/12/2019 040504

    14/19

    Convex Hulls in the Plane

    Given {(xi, yi)}n+1i=0 R

    2 find the extreme pointsand the facetsof

    their convex hull and the order in which they appear.

    First find the lower envelopeof the convex hull.

    Ifxi = xj and yi yj, we can ignore (xj, yj) without changing thelower envelope.

    Thus, assume x0 < x1 < < xn < xn+1. In practice, this wouldrequiresorting.

    The lower envelope is a piecewise linear convex function f(x), thepointwise maximum of all convex functionsg(x)such thatg(xi) yi

    fori= 0, 1, . . . , n + 1.

  • 8/12/2019 040504

    15/19

  • 8/12/2019 040504

    16/19

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    ...

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ...

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    ...

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    ..

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    ..

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    ...

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

  • 8/12/2019 040504

    17/19

    .

    ..

    ..

    .

    .....................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................

    .........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................

    .

    .

    .

    ..

    .

    .

    ..

    .

    .

    .

    ..

    .

    .

    ..

    .

    .

    ..

    .

    .

    ..

    .

    .

    ..

    .

    .

    ..

    .

    .

    .

    ..

    ..

    ..

    .

    .

    ..

    .

    .

    ..

    .

    .

    ..

    .

    ..

    .

    .

    ..

    .

    .

    ..

    ..

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    ...

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ...

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    ...

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    ..

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    ...

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

    ..

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    ..

    .

  • 8/12/2019 040504

    18/19

    Defineti=f(xi) and letzi=yi ti, fori= 0, 1, . . . , n + 1.

    Note thatz0=zn+1= 0.

    If(xi, yi)is a breakpoint, then ti=yi andzi= 0.

    The segment of the lower envelope between (xi1, ti1) and (xi, ti)

    has a different slope than the segment between (xi, ti) and (xi+1, ti+1).

    Since f(x) is convex, the former (left-hand) segment must have a

    smaller slope than the latter (right-hand) segment.

    Hence strict inequality holds in

    ti ti1xi xi1

    ti+1 tixi+1 xi

    .

    Ifzi>0, then(xi, yi)cannot be a breakpoint off(x).

    In that case, equality holds in the inequality above.

  • 8/12/2019 040504

    19/19

    The vector z = {zi}ni=1 must solve the LCP (q, M) where q Rn

    andMRnn are defined by

    qi = i i1 and mij =

    i1+ i ifj =i,

    i ifj =i + 1,

    j ifj =i 1,

    0 otherwise,

    and where

    i= 1/(xi+1 xi)andi=i(yi+1 yi)fori= 0, . . . , n .

    This LCP has a unique solution.

    The matrix M associated with this LCP has several nice propertieswhich can be exploited to produce very efficient solution procedures.