[Aces-support] Matlab Memory Limits

[Sai Ravela (MIT)]

What would be a better solution for interpolating with finite memory 
without loss of precision?
1 Domain decomposition
2 Multigrid/Wavelets with perfect reconstruction.
(2 - harder to implement and not sure how much you gain for this problem).

The other approaches involve loss:
1 Spectral truncation.
2 Multiscale: nested spectral truncation.
(If you know what the inherent bandwidth of your signal or kernel is, 
then you can calculate how many bits you need, logarithms and exponents 
again :)  )
And presumably, single-precision quantization is sufficient. Is it 
optimal?  Is it scalable?

I would be curious to learn if you think there is a better solution than 
domain decomposition for this problem (ignoring multigrid for the moment).
(Of course, a little program for tiled-interpolation, MPId and put on 
website of our own ACES cluster will make you famous. Many internet 
sources probably exist with libraries for interpolation. add MPI and PBS 
and what not and you're done). Can't do it till I get back, but may be 
something like that will be useful before ad-hoc or lossy solutions.

I am still not sure what you mean  by...the larger question of Matlab's 
inherent memory limitations stands?

Sai