[Numpy-discussion] np.linalg.norm terrible behavior

[Kibeom Kim]

for the sum function part,
It's not a good way to fix but... if you want more accuracy

x=(np.array([1]*10000 + [1e4], dtype=np.float32))
np.sum(x*x)
1.0001e+08

You can sort x from small numbers to bigger numbers before you call sum.

-Kibeom Kim

On Sat, Mar 5, 2011 at 6:27 PM, Xavier Gnata <xavier.gnata at gmail.com> wrote:

> Hi,
>
> I got this problem in a real life code and it took me some time to
> figure out that np.linalg.norm has a terrible numerical behavior.
> The problem is nicely described here
> http://fseoane.net/blog/2011/computing-the-vector-norm/
>
> numpy/linalg/linalg.py claims to be a "high-level Python interface to
> the LAPACK library".
> That's great but the low level norm function is coded in pure python :
>     x = asarray(x)
>     if ord is None: # check the default case first and handle it
> immediately
>         return sqrt(add.reduce((x.conj() * x).ravel().real))
>
> Is there a way to use *by default* the blas function to compute the norm?
>
> ok....looks like sum as the same terrible numerical behavior
> x=(np.array([1e4] + [1]*10000, dtype=np.float32))
> np.sum(x*x)
> returns 1e+08 instead of 1.0001e+08
>
> Moreover, np.linalg.norm is slow compare to blas.
>
> Is there a way/plan to fix it?
>
> Xavier
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20110305/606b7e60/attachment.html>