Ideas for auto BSX

[Jordi Gutiérrez Hermoso]

Because we don't use the bug tracker mailing list anymore. I'm
forwarding the following discussion to the maintainers' list.


---------- Forwarded message ----------
From: Dr. Alexander Klein <address@hidden>
Date: 2011/9/30
Subject: Re: [bug #34441] Weird bug with vector subtraction
To: Jordi Gutiérrez Hermoso <address@hidden>
Cc: Michael D Godfrey <address@hidden>, address@hidden


(Sorry for the duplicate, I forgot to include the bug-tracker address.)

Am 30.09.2011 um 20:55 schrieb Jordi Gutiérrez Hermoso:

> It's not a bug. It's a feature I implemented recently. It's equivalent to
>
> bsxfun(@minus, [1 2 3], [1 2 3]')


Jordi,

thanks for replying, I see what you mean, and it's really an interesting option.

However, this can lead to serious headaches when debugging code with a
lot of linear algebra. Here's an example.

Suppose you mistakenly write:

octave:1> a=[1 2 3];b=[3 4 5]';c=[6 7 8]';
octave:2> (a-b)*c
error: operator -: nonconformant arguments (op1 is 1x3, op2 is 3x1)

The old version will give you an error.

With the new semantics this gets expanded to matrix * vector:

octave:1> a=[1 2 3];b=[3 4 5]';c=[6 7 8]';
octave:2> (a-b)*c
ans =

 -19
 -40
 -61

Now, if the erroneous transposition of b happens somewhere far away in
the code, and you rely on "-" failing due to the dimension mismatch,
this error will be virtually impossible to spot.

So, while I think this is quite useful, it should by no means be
called "-" in my opinion. Perhaps ".-" would do? I've never seen
anyone use this.

Kind regards,

       Alex

--
        Dr. Alexander Klein, Diplom-Mathematiker

Physiologisches Institut       |               TransMIT Zentrum
Raum 543                       |        für Numerische Methoden
Aulweg 129                     |          Heinrich-Buff-Ring 44
35392 Giessen                  |                  35392 Giessen