On Mon, Apr 25, 2011 at 4:08 PM, Ted Dunning <[email protected]> wrote:

> I thought that for a moment as well, but now I think not.
>
> The trivial counter example consists of one element for each single
> component unit vector e_i.  The SVD of this has unit singular values and
> the
> eigenvectors are just the e_i themselves.  After transformation, all values
> are still in the positive orthant.
>

Still in the positive orthant is cutting things pretty close: all vectors
are literally
*on the border* of the positive orthant, and random hyperplanes will
separate
these vectors.


> I may have misunderstood what you are suggesting, but I don't think that
> LSH
> could distribute the values into other orthants.
>

No, LSH won't distribute them, by itself.


> It may be that in practice that SVD will scatter data into many orthants,
> but I suspect it will not spread the data as widely as LSH assumptions
> would
> like.


Maybe you're right - raw SVD vs PCA (where the means are also subtracted
off) is probably the distinction to draw here.

  -jake

Reply via email to