2007-06-17 TPC DD LinearSystemsConditioning

download 2007-06-17 TPC DD LinearSystemsConditioning

of 8

Transcript of 2007-06-17 TPC DD LinearSystemsConditioning

  • 8/13/2019 2007-06-17 TPC DD LinearSystemsConditioning

    1/8

    Tech-Pro Consultants

    Under-Determined, Over-Determined, Ill-conditioned Linear

    Systems

    August 2005

    Dr. K.S.Ravichandran

  • 8/13/2019 2007-06-17 TPC DD LinearSystemsConditioning

    2/8

    Tech-Pro Consultants

    Linear Systems

    Reference :

    1. Lecture notes on optimization, Prof. Baldick, The University of Texas at

    Austin

    A non-singular matrix A can always be factorized in to

    A=LU where L is lower triangular and U is strictly upper

    triangular with unit entries on the diagonal.

    The procedure for triangulation breaks down for a

    singular matrix because at some stage of the procedure, a

    non-zero pivot does not exist.

  • 8/13/2019 2007-06-17 TPC DD LinearSystemsConditioning

    3/8

    Tech-Pro Consultants

    Linear Systems Ill-Conditioning

    Example:Consider Ax = b,

    Det[A] = - , 0 and the matrix A is nearly singular.Using infinite precision arithmetic, we can solve Ax = b forany vector b = [b1, b2]

    t. The solution is

  • 8/13/2019 2007-06-17 TPC DD LinearSystemsConditioning

    4/8

    Tech-Pro Consultants

    Linear Systems - Perturbations

    Example: Take b = [1 , 1]t. .

    Then ||b || = 2 and the solution x* = [1 , 0]t, || x* || = 1Perturb the RHS:Take b = [1 , 1]t.+ [, 0]t = b + bThen ||b || = and the solutionx** = x*+ x* = [1 , 0]t+ [0 , /]t, || x** || = {1+ (/)2}|| x* || =|/|.The relative change in solution is || x* || / || x* || =|/|.The relative change in b is || b || / || b || =|| / 2.

  • 8/13/2019 2007-06-17 TPC DD LinearSystemsConditioning

    5/8

    Tech-Pro Consultants

    Ratio of the relative change is =|/| / (|| / 2) = 2/ ||The change in the solution vector is independent of the

    change in the forcing vector and is large if is small.We say that the matrix is ill-conditioned.

    Linear Systems Ill-Conditioning

  • 8/13/2019 2007-06-17 TPC DD LinearSystemsConditioning

    6/8

    Tech-Pro Consultants

    Perturb the Matr ix :Take (A+ A)x = b. Let A =

    Then ||A || = and the solution. With b = [1 , 1]tx** = x*+ x* = [1 , 0]t+ [0 , -/]t, || x* || =|/|.The relative change in solution is || x* || / || x* || =|/|.The relative change in A is || A || / || A || =|| / 2.Ratio of the relative change is =|/| / (|| / 2) = 2/ ||The change in the solution vector is independent of the change

    in the matrix and is large if is small.Thus the matrix is ill-conditioned.

    Linear Systems Ill-Conditioning

  • 8/13/2019 2007-06-17 TPC DD LinearSystemsConditioning

    7/8

    Tech-Pro Consultants

    AnalysisCondition Number

    Condition Number of a Matr ix :

    Denote the L2norm of a vector x by ||x|| = = xtx

    The L2norm of a matrix A is given by Max ||Ax=y|| for all x s.t.

    ||x|| = 1.

    Typically the L2norm of a matrix is computed as

    ||A||2= { aij2}1/2 where the aijare the elements of AThe condition number is now defined as

    C= ||A||2||A-1||2

    A matrix for which C is large is said to be ill-conditioned

  • 8/13/2019 2007-06-17 TPC DD LinearSystemsConditioning

    8/8Tech-Pro Consultants

    AnalysisMain Result

    Suppose A is a non-singular matrix , and consider Ax = b, Ax =

    b + b and (A + A)x = b. We then have the following bounds:(i) If x*+ x* solves Ax = b + b we have

    || x*|| / || x*|| ||A||2||A-1||2||b|| / ||b|| = C ||b|| / ||b||(ii) If x*+ x* solves (A + A)x = b we have

    || x*|| / || x* + x*|| C ||A|| / ||A|| where Ax* = bIn either case, the amplification to the solution is large if the

    condition number of the system is large.