It would be easier to go through the code to find out just what it's
doing, but I have no interest in doing that. The section is jtmdivsp,
in vd.c:101. Domain error is signaled from jttridiag, in vd.c:86, where
it looks like you get a domain error if any pivot is 0.
Henry Rich
On 7/27/2018 3:14 PM, Bill Heagy wrote:
For small cases, e.g. 3x3, it does give a nonce error. (I haven't
tested past that.)
On 2018-07-27 03:12 PM, Henry Rich wrote:
Not a problem for the guy who doesn't have to code it! :) It would
require row and column exchanges to change the pivot, which I'm
guessing is more work than Roger wanted to put into it.
But it IS misdiagnosed: domain error should mean the input is
invalid, while nonce error means the input is valid but unsupported
by the current code.
Henry Rich
On 7/27/2018 3:08 PM, Bill Heagy wrote:
Sorry, when I said "misdiagnosed", I meant that it was detecting a
problem that wasn't actually there, not giving the wrong message.
Starting with a zero shouldn't be a problem.
On 2018-07-27 02:57 PM, Henry Rich wrote:
You're right that it should be diagnosed as a nonce error rather
than a domain error. If you try to invert a sparse matrix with a 0
in the top-left, you get a nonce error. The code just doesn't
handle the case of starting with a 0.
I would change the error to nonce error, but that doesn't fix the
testcase, and I don't think it's worth any effort to fix this bug
in the solver.
Henry Rich
On 7/27/2018 2:49 PM, Bill Heagy wrote:
Most of the time, the test works. It's only if you run it
repeatedly, until there's a (random) zero in the (0,0) position,
that it fails. It looks like a misdiagnosed problem, or a problem
with initializing the solution.
On 2018-07-27 02:43 PM, Raul Miller wrote:
Which suggests that maybe this is diagnosing an api handling issue?
(If the test used to work, x and y somehow got flipped?)
Thanks,
On Fri, Jul 27, 2018 at 2:37 PM Bill Heagy <[email protected]> wrote:
>
> If you just take the x and y and flip them so the 0 is in the
bottom
> right, it appears to work. It fits the domain.
>
> On 2018-07-27 02:24 PM, Henry Rich wrote:
>> Right again. The sparse matrix divide is very restricted in
domain and
>> I am not going to be the one to try to improve it. Maybe I can
tweak the
>> testcase.
>>
>> Henry Rich
>>
>> On 7/27/2018 1:48 PM, Bill Heagy wrote:
>>> Running gsp1 repeatedly produces:
>>>
>>> ...
>>> 1e_13 >. >./ | (y=: h 11) (%. -: (%.$.)) x=: 0.01*g 11
>>> |domain error
>>> | 1e_13>.>./|(y=:h 11) (%.-:(%.$.))x=:0.01*g 11
>>> |[-37] /home/wheagy/git/jsource/test/gsp1.ijs
>>>
>>> at various locations.
>>> The problem appears to be a "0" in the (0,0) position in the x
matrix.
>>> JVERSION Engine: j806/j32/linux
>>> Beta: GPL3/2018-07-25T05:05:57
>>> Library: 8.05.14
>>> Platform: Linux 32
>>> Installer: unknown
>>> InstallPath: /home/wheagy/tmp/jbld/j32
>>> Contact: [email protected]
>>>
----------------------------------------------------------------------
>>> For information about J forums see
http://www.jsoftware.com/forums.htm
>>
>>
>> ---
>> This email has been checked for viruses by AVG.
>> https://www.avg.com
>>
>>
----------------------------------------------------------------------
>> For information about J forums see
http://www.jsoftware.com/forums.htm
>
----------------------------------------------------------------------
> For information about J forums see
http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see
http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see
http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
