Dismiss Notice
Join Physics Forums Today!
The friendliest, high quality science and math community on the planet! Everyone who loves science is here!

Multiplication bloards after factorization

  1. Nov 18, 2013 #1
    Let a positive definite matrix A be factorized to P and Q, A=P*Q and let an arbitrary matrix B.
    I am calculating the relative error of the factorization through the norm:

    [itex]\epsilon = \left\| \textbf{A}-\textbf{PQ} \right\| / \left\| \textbf{A} \right\|[/itex]

    which gives

    [itex]\epsilon <1\text{e}-16[/itex]

    so I assume factorization is correct.

    But things go messy when I try to multiply the factorized form of A with B.
    In particular, the relative error, r, of the product

    [itex]r = \left\| \textbf{AB}-\textbf{PQB} \right\| / \left\| \textbf{AB} \right\|[/itex]

    now bloats, i.e. I get

    Note that B is arbitrary, in particular I have tried several different types: random, structured, all-ones matrix, even the identity matrix.
    I'm confused. How come factorization is correct and then the multiplication bloats?
    Has anything to do with condition number?

    (Unfortunately I can't disclose the type of factorization but I can tell that P and Q are not triangular)
  2. jcsd
  3. Nov 19, 2013 #2
    Modification: I mistakenly added also "identity matrix" to previous post. Please ignore this from the list of matrices I have tried.
  4. Nov 19, 2013 #3


    User Avatar
    Science Advisor
    Homework Helper

    It's hard to know exactly what you are doing since you won't tell us all the facts, but I think the basic issue is the difference between "forward" and "backward" error analysis.

    If you are trying to solve ##Ax = b## numerically, you can estimate the error two different ways.

    Forward error analysis: try to estimate the error in ##x##, i.e assume the exact solution is ##A(x + e) = b ## and find an estimate for the vector ##e##.
    Backward error analysis: consider you have the exact solution to the "wrong" equation, i.e. estimate the size of a matrix ##e## such that ##(A+e)x = b##.

    Backward error analysis (first proposed by Wilkinson in the 1960s) is generally more useful than the more "obvious" forward analysis.
Know someone interested in this topic? Share this thread via Reddit, Google+, Twitter, or Facebook