[eigen] Re: solve with and without check

[Benoit Jacob <jacob.benoit.1@xxxxxxxxx>]

Some more thoughts:

there's also the option of removing the solve-with-check functionality
altogether. Does it have a use in practice?

I'm leaving now until Saturday
Benoit

2009/8/24 Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
> Hi,
>
> In various decompositions (LU, rank-revealing QR,...) we have a
> solve() method that checks for the existence of a solution; if no
> solution exists, it returns false and leaves the result undefined.
>
> This is really needed as trying to get a solution nevertheless can
> result in inf/nan values making the solve() function taking much
> longer to return.
>
> But in most cases, the user just knows that a solution is meant to
> exist and he wants to get a solution even if it's only half-decent,
> rather than having to check the return value of solve() and deal with
> error cases. This is especially true as it's very difficult to check
> for existence without paying a very high price, basically our
> triangular solvers do not do any check (required, i guess, to make
> them fast) so the check we do is rather fragile.
>
> This also isn't addressed by letting the user set manually the
> precision level: the solve() method would need a separate precision
> level than the compute() method, this is very cumbersome.
>
> --> i propose the following change:
> in all decompositions that have a solve() checking for existence,
> let's rename solve() to solveWithCheck()   (or whatever native English
> speakers prefer... perhaps trySolve() ? checkSolve() ? )
> then let's add a method solve() that doesn't do any check, making it
> the responsibility of the caller that a decent solution exists.
>
> Cheers,
> Benoit
>