[R] Matrix

[Uwe Ligges]

Douglas Bates wrote:

> Uwe Ligges wrote:
> 
>> salvati nicola wrote:
>>
>>> I have a matrix 2900X2900 and I have to solve it. But R says that it 
>>> can't allocate 67899kb.
>>> How can I do?
>>
>>
>>
>> a) Clean up your workspace before trying to work on the matrix
>> b) Is this an outdated R version on Windows? Then (i) please upgrade 
>> and (ii) please read the R for Windows FAQs and ?Memory.
>> c) Buy more memory, if you have less than 512MB. 512 MB suffices for 
>> me (in a clean workspace!) in ~5 min. using Rblas.dll linked against 
>> ATLAS. You probably want to have ~1GB of memory for serious other 
>> calculations with those matrices.
>>
>> Uwe Ligges
> 
> 
> 5 Min?
> 
> On a modest desktop system (Athlon XP 2500+, approx 1.8 GHz) it took 
> about 20 seconds for the inversion and about 8 seconds for the LU 
> decomposition.  This is using the Matrix package.
> 
>  > mm <- Matrix(rnorm(2900*2900), ncol = 2900)
>  > object.size(mm)
> [1] 67280636
>  > system.time(rcond(mm))
> [1] 7.73 0.10 7.96 0.00 0.00
>  > object.size(mm)
> [1] 134573052
>  > system.time(minv <- solve(mm))
> [1] 19.93  0.10 20.04  0.00  0.00
>  > rcond(mm)
> [1] 2.466351e-07
> 
> The initial call to rcond calculates and stores the LU decomposition. 
> That is why the size of the object approximately doubles.
> 
> I get similar results without the matrix package except that the 
> decomposition must be recomputed for each such operation on the matrix.
> 
>  > m1 <- matrix(rnorm(2900*2900), ncol = 2900)
>  > system.time(minv1 <- solve(m1))
> [1] 30.08  0.86 32.32  0.00  0.00


I'm sorry, I only thought I had used the ATLAS optimized version (but it 
was the standard Rblas.dll), it really has

 > system.time(minv <- solve(mm))
[1] 31.69  1.82 34.14    NA    NA

(R-1.9.1, ATLAS, WinNT4.0, Athlon with real 1.67 GHz)

Uwe