Mailing-List: contact dev-help@commons.apache.org; run by ezmlm
Precedence: bulk
Reply-To: "Commons Developers List" <dev@commons.apache.org>
Received-SPF: neutral (athena.apache.org: 129.104.30.34 is neither permitted
 nor denied by domain of sebastien.brisard@m4x.org)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
From: "Sebastien Brisard" <sebastien.brisard@m4x.org>
Subject: Re: [math] Exceptions for matrix operations
To: dev@commons.apache.org
Message-Id: <20110510151838.0292C14059823@svoboda.polytechnique.org>
Date: Tue, 10 May 2011 17:18:37 +0200 (CEST)

> [...]
> The question is thus: Is it OK to provide pointless methods?
> I also wonder whether it is reasonable to _generate_ state variables just
> so that the accessors can return them.
> Usually, when a exception is constructed, it is _passed_ data that provide
> the context of the failure (i.e. the vectors or the indices that show the
> problem).
> In this case, if the test (that triggered the exception) was not performed
> using vectors, it looks quite strange to be able to retrieve vectors...
>
>
> Gilles
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
> For additional commands, e-mail: dev-help@commons.apache.org
>
>
OK, Gilles, sorry, I do not want to offend anyone, nor do I think I'm the best
judge of all solutions proposed so far. I agree that constructing vectors just
for the sake of having one exception inherit from another is far-fetched. I'm
actually impressed by the depth of the discussions! You all DO think a lot
about the overall design of this library (well, I guess you have too, but
still).

So if I understand correctly, we are back to one base classe which does pretty
much nothing, and two derived exceptions
1. one raised when coefficients of the matrix are accessible, and a(i,j) !=
a(j,i) (within a specified tolerance),
2. one raised when coefficients of the matrix are not accessible, and x'.A.y
!= y'.A.x

So we still need a little bit of terminology there... For your information,
I've decided to call (for lack of a better name) LargeRealMatrix the interface
corresponding to matrices defined only by their matrix-vector product y = A.x.
These objects are the basic brick of the iterative solvers I'm implementing.

If we agree that a LargeRealMatrix is only defined by this product, then a
symmetric, large matrix must be defined as x'.A.y = y'.A.x. (This is actually
the mathematical definition of an adjoint operator). Then, it would be logical
to call the corresponding exception NotASymmetricLargeMatrix, would it not?

Alternatives might be LinearMap, LinearOperator, since we actually define the
matrix A as a function x\mapsto y = A.x, here.

So, it's pretty much your call. You tell me, and I'll update my sources before
comitting them. You can then decide wether it's worth including in
commons-math.

Best regards for now,
Sébastien

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