commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Sebastien Brisard" <sebastien.bris...@m4x.org>
Subject [math] Iterative linear solvers (MATH-581) again
Date Fri, 08 Jul 2011 06:44:15 GMT
Hi everyone,
well the message below did not raise much of an interest... My apology, I
forgot to quote [math] in the title.
Anyway, I have a new proposal regarding this issue. I do not know what your
view on the subject is, but I tend to think that exceptions should return
references and not deep copies of the objects that caused these exceptions to
be raised. This facilitates debugging (I guess), and also helps identify the
exact origin of the problem, without having to specialize the exception. As an
example, Conjugate Gradient requires positive definiteness of both matrix (a)
and preconditioner (m). If a NonPositiveDefiniteLinearOperatorException is
caught, then I can simply test
a == e.getOffendingLinearOperator()
or
m = e.getOffendingLinearOperator()

Now, coming to offending vectors, the issue is that these can either be
instances of RealVector, or double[]. So, what type should be returned by
getOffendingVector()?

I propose to have the exceptions return a reference to the RealVector. If the
exception was raised with a double[], then the returned RealVector will
actually be an instance of ArrayRealVector, with a reference to the initial
double[]. All this is to be clearly specified in the javadoc. Does that sound
like a viable option?

I'll attach the corresponding files to the JIRA issue, if you want to have a
more precise idea of what I'm talking about.

Thanks in advance for your comments,
Sebastien

>Good morning,
>my proposal for the implementation of linear iterative solvers (JIRA
MATH-581)
>has raised some comments from Gilles and Luc, but I think no consensus has
>been reached on one issue raised by Gilles. Before submitting a new version
of
>the corresponding classes, I'd like to make a new proposal.
>
>Here is the thing. I have defined two exceptions, namely
>  - NonSelfAdjointLinearOperatorException
>  - NonPositiveDefiniteLinearOperatorException,
>the latter being typically raised when a vector x is found, such as x'.A.x
<=
>0, where A is the linear operator under consideration.
>
>Now, the constructor of this exception takes as input A and x (which I call
>the "offending vector"). There is a method to return a reference to A (which
>is a RealLinearOperator), and a method to return a deep copy of x, the
reason
>for this is that x can either be a double[], or a RealVector.
>
>I agree with Gilles, for debugging purposes, it would probably be better to
be
>able to trace the offending vector, hence to have the exception return a
>reference to this offending vector. But this is just not possible, since
there
>is no unique type for this vector. Gilles pointed out that there is the
method
>getData() in RealVector which would do the job, but there is no requirement
in
>the contract of this method for this to be a shallow/deep copy. Is this a
>concern ? If not, then I could change the current method
>void copyOffendingVector(final double[] x)
>(where x is modified)
>into
>double[] getOffendingVector()
>but we would lose the track of the actual object which first raised the
>exception.
>
>Another option would be to have the method getOffendingVector() return an
>Object, and try to cast it to a RealVector or a double[] when catching the
>exception. Not too sure it's good practice to return Objects, though.
>
>I'm looking forward to reading your comments, and will update the sources
>accordingly.
>
>Sebastien

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
For additional commands, e-mail: dev-help@commons.apache.org


Mime
View raw message