Lec21 Hw Probs

Embed Size (px)

Citation preview

  • 7/26/2019 Lec21 Hw Probs

    1/6

    AdditionalHomeworkProblems

    RobertM.Freund

    April,2004

    1

    2004 Massachusetts Institute of Technology.

  • 7/26/2019 Lec21 Hw Probs

    2/6

    2

    1 Exercises

    1. LetIRn ++

    denotethenonnegativeorthant,namelyIRn ={xIRn |xj 0, j = 1, . . . , n}. ConsideringIRn asacone,provethat IRn = IRn+ + +,thusshowingthatIRn isself-dual.+

    n n 22. LetQ = xIRn |x1 j=2xj . Qn iscalled thesecond-order

    cone, the Lorentz cone, or the ice-cream cone (I am not making thisup). ConsideringQn asacone,provethat(Qn) =Qn,thusshowingthatQn isself-dual.

    3. ProveCorollary3ofthenotesondualitytheory,whichassertsthatthe

    existence

    of

    Slater

    point

    for

    the

    conic

    dual

    problem

    guarantees

    strongdualityandthattheprimalattains itsoptimum.

    4. Considerthefollowingminimaxproblems:

    min max (x, y) and max min (x, y)xX yY yY xX

    where X and Y are nonempty compact convex sets in IRn and IRm,respectively,and(x, y)isconvexinxforfixedy,andisconcaveinyforfixedx.

    (a) ShowthatminxX maxyY (x, y) maxyY minxX (x, y)intheabsenceofanyconvexity/concavityassumptionsonX,Y,

    and/or

    (, ).

    (b) Showthatf(x) := maxyY (x, y) isaconvex function inxandthatg(y) := minxX(x, y) isaconcavefunction iny.

    (c) Useaseparatinghyperplanetheoremtoprove:

    min max (x, y) = max min (x, y) .xX yY yY xX

    5. Let X and Y be nonempty sets in IRn, and let f(), g() : IRn IR. Consider the followingconjugate functions f () and g () defined as

    follows:f(u) := inf{f(x)utx} ,

    xX

    and

    tg (u) := sup {g(x)u x}.

    xX

  • 7/26/2019 Lec21 Hw Probs

    3/6

    3

    (a) Constructageometric interpretationoff () and g (). (b) Show that f () is a concave function on X := {u | f(u) >

    }, and g ()isaconvexfunctiononY :={u|g(u)

  • 7/26/2019 Lec21 Hw Probs

    4/6

    4

    (c) Starting from u = (1,2), perform one iteration of the steepestascent

    method

    for

    the

    dual

    problem.

    In

    particular,

    solve

    the

    followingproblemwhered=L (u):

    max L ( u

    +d)s.t. u+d0 ,

    0 .

    7. Prove Remark 1 of the notes on conic duality, that the dual of thedual is the primal for the conic dual problems of Section 13 of thedualitynotes.

    8. Considerthefollowingverygeneralconicproblem:

    TGCP: z =minimumx c x

    s.t. AxbK1

    xK2 ,

    whereK1

    IRm andK2

    IRn areeachaclosedconvexcone. Derive

    thefollowingconicdualforthisproblem:

    GCD: v = maximumy bTy

    s.t.

    c

    ATy

    K2

    y

    K ,1

    and show that the dual of GCD is GCP. How is the conic format ofSection13ofthedualitynotesaspecialcaseofGCP?

    n

    9. Fora(square)matrixM IRnn,definetrace(M) = Mjj ,andforj=1

    twomatricesA,BIRkl define

    k l

    A

    B

    :=

    Aij

    Bij

    .

    i=1j=1

    Provethat:

  • 7/26/2019 Lec21 Hw Probs

    5/6

    5

    (a) A B=trace(ATB).

    (b) trace(M N)=trace(N M).

    10. LetSkk denotetheconeofpositivesemi-definitesymmetricmatrices,+nnamelySkk ={X Skk |vTXv0 forallv }. Considering+

    SkkSkk asacone,provethat =Snn,thusshowingthatSkk+ + + +isself-dual.

    11. Considertheproblem:

    P : z =minimumx1,x2,x3 x1

    s.t.

    x2 +x3 = 0

    x1

    10

    (x1, x2) x3 ,

    where denotestheEuclideannorm. Thennotethatthisproblemisfeasible(setx1

    =x2

    =x3

    =0),andthatthefirstandthirdconstraintscombinetoforcex1 =0 inanyfeasible,solution,wherebyz

    = 0.

    Wewilldualizeonthethefirsttwoconstraints,setting

    X :={(x1, x2, x3) | (x1, x2) x3} .

    Using multipliers u1, u2

    for the first two constraints, our Lagrangianis:

    L(x1, x2, x3, u1, u2) = x1+u1(x2+x3)+u2(x110)=10u2+(1u2, u1, u1)T(x1, x2, x3)

    Then

    L (u1, u2) = min 10u2

    + (1 u2, u1, u1)T(x1, x2, x3).

    (x1,x2)x3

    (i) Showthat:

    10u2 if

    (1

    u2, u1) u1L (u1, u2) =

    if (1 u2, u1)> u1

    ,

  • 7/26/2019 Lec21 Hw Probs

    6/6

    6

    andhencethedualproblemcanbewrittenas:

    D : v = maximumu1,u2 10u2

    s.t. (1 u2, u1) u1

    u1 IR, u2 0 .

    (ii) Showthatv =10,andthatthesetofoptimalsolutionsofDiscomprisedofthosevectors(u1, u2) = (, 1)forall0. HencebothPandDattaintheiroptimawithafinitedualitygap.

    (iii) In this example the primal problem is a convex problem. Whyis

    there

    nevertheless

    a

    duality

    gap?

    What

    hypotheses

    are

    absent

    that

    otherwisewouldguaranteestrongduality?