Ah, sorry, I see I was mistakenly assuming his was an outer product of two 
vectors, which would of course have rank 1. Sorry for the noise.

> On Jul 28, 2014, at 1:27 PM, John Myles White <[email protected]> 
> wrote:
> 
> I think a correlation matrix can have arbitrary rank, but might be wrong.
> 
>  -- John
> 
>> On Jul 28, 2014, at 9:36 AM, Stefan Karpinski <[email protected]> wrote:
>> 
>> Does this computation not always return a rank-1 matrix?
>> 
>> 
>>> On Mon, Jul 28, 2014 at 12:33 PM, John Myles White 
>>> <[email protected]> wrote:
>>> But how would you know the rank of the correlation matrix in advance?
>>> 
>>>  -- John
>>> 
>>>> On Jul 28, 2014, at 9:25 AM, Stefan Karpinski <[email protected]> wrote:
>>>> 
>>>> This is the sort of thing that just begs for a custom representation of a 
>>>> rank-1 matrix, which fortunately, isn't terribly hard to implement in 
>>>> Julia.
>>>> 
>>>> 
>>>>> On Mon, Jul 28, 2014 at 12:08 PM, Tim Holy <[email protected]> wrote:
>>>>> If they're sparse along dimension 1, you can at least save time computing 
>>>>> the
>>>>> dot product of the two sparse vectors. But yes, the correlation matrix 
>>>>> itself
>>>>> will be dense.
>>>>> 
>>>>> --Tim
>>>>> 
>>>>> On Monday, July 28, 2014 11:23:31 AM Jiahao Chen wrote:
>>>>> > > I don't think sparse cor() is implemented and is falling back to the 
>>>>> > > dense
>>>>> > > implementation.
>>>>> > Computing the correlation matrix is much like computing the outer
>>>>> > product of two sparse vectors. There will be massive fill-in and I
>>>>> > don't see how you can preserve sparsity without special knowledge
>>>>> > about the sparsity pattern.
>>>>> > Thanks,
>>>>> >
>>>>> > Jiahao Chen
>>>>> > Staff Research Scientist
>>>>> > MIT Computer Science and Artificial Intelligence Laboratory
>>>>> >
>>>>> > On Mon, Jul 28, 2014 at 11:12 AM, Stefan Karpinski 
>>>>> > <[email protected]>
>>>>> wrote:
>>>>> > > https://github.com/JuliaLang/julia/issues/new
>>>>> > >
>>>>> > >
>>>>> > > On Mon, Jul 28, 2014 at 10:06 AM, paul analyst <[email protected]>
>>>>> > >
>>>>> > > wrote:
>>>>> > >> Issue on github or on julia-dev  groups?
>>>>> > >> Paul
>>>>> > >>
>>>>> > >> W dniu poniedziałek, 28 lipca 2014 12:05:27 UTC+2 użytkownik Viral 
>>>>> > >> Shah
>>>>> > >>
>>>>> > >> napisał:
>>>>> > >>> Please file an issue. I don't think sparse cor() is implemented and 
>>>>> > >>> is
>>>>> > >>> falling back to the dense implementation.
>>>>> > >>>
>>>>> > >>> -viral
>>>>> > >>>
>>>>> > >>> On Monday, July 28, 2014 1:41:55 PM UTC+5:30, paul analyst wrote:
>>>>> > >>>> Correlation sparse array is very slow. Out of memory on a dense 
>>>>> > >>>> array
>>>>> > >>>> when we have 30,000 columns. How quickly it calculated?
>>>>> > >>>>
>>>>> > >>>> julia> I=int32((rand(10^7)*9999999).+1);
>>>>> > >>>>
>>>>> > >>>> julia> J=int32((rand(10^7)*29999).+1);
>>>>> > >>>>
>>>>> > >>>> julia> V=int8((rand(10^7)*9).+1);
>>>>> > >>>>
>>>>> > >>>> julia> D=sparse(I,J,V);
>>>>> > >>>>
>>>>> > >>>> julia> @time cor(D[:,1:30]);
>>>>> > >>>> elapsed time: 23.806328476 seconds (2458875228 bytes allocated, 
>>>>> > >>>> 0.14%
>>>>> > >>>> gc
>>>>> > >>>> time)
>>>>> > >>>>
>>>>> > >>>> julia> @time cor(full(D[:,1:30]));
>>>>> > >>>> elapsed time: 4.494099126 seconds (2732042496 bytes allocated, 
>>>>> > >>>> 5.31% gc
>>>>> > >>>> time)
>>>>> > >>>>
>>>>> > >>>> julia>
>>>>> > >>>>
>>>>> > >>>> Paul
> 

Reply via email to