| Re: declaration of an array and multiplication with a matrix [message #36352] |
Sun, 07 September 2003 16:40 |
JD Smith
Messages: 850
Registered: December 1999
|
Senior Member |
|
|
On Sat, 06 Sep 2003 04:17:38 -0700, Hassan Iqbal wrote:
> Hi all,
>
> can anyone please tell me what does this expression do:
>
> (w # (fltarr(m)+1)* der)
>
> w= an array of 'n' elements
> der= a matrix of 'n' columns and 'm' rows
>
w # (fltarr(m)+1)
is just a fairly opaque way to take a vector of length n, and
construct an array of dimensions (n,m) in which each row holds a copy
of said vector. You might also see the equivalent:
w # replicate(1,m)
or even
replicate(1,m) ## w
("#" a "##" are duals under argument interchange). Another way to do
this is with the REBIN & REFORM method you've no doubt heard about:
rebin(w,n,m) * der
You can read more about the REBIN/REFORM style of array inflation at
David Fanning's site http://dfanning.com/tips/rebin_magic.html. Yet
another method constructs an index array of the correct size and
converts it to the right signature use various combinations of
division and/or MOD:
w[lindgen(n,m) mod n] * der
Why so many ways to skin this particular feline? None of these
methods is a truly optimized, native way to recast and redimension
arrays for vector calculations. They're all just hacks, to one degree
or another. For example, REBIN contains lots of useless (for this
purpose) code for decimating arrays using various fancy sampling
algorithms. If RSI ever wrote a dedicated "INFLATE" routine, it would
probably be noticeably faster (and could combine the REBIN &
REFORM/TRANSPOSE into one easy step).
So why do I use REBIN/REFORM? It can certainly be slightly more
verbose in common cases than either the matrix multiply, or index
arithmetic methods (though not in this case). It also requires you to
specify more of the dimensions explicitly (in this example, we had to
have "n" in addition to "m"). However, it scales up to higher
dimensions quite trivially, and easily accomodates inflating arrays
over "inside" dimensions (neither the first nor the last). Try
expanding over the 5th dimension of a 7 dimensional array using
lindgen(n,m,p,q,r,s,t) and you'll soon curse the "/" key on your
keyboard. As a final advantage, since IDLv5.5 added support for a
vector-of-dimensions argument to REBIN, it's really the only technique
which can reasonably be used to build generic operators which require
arbitrary array inflation over arbitrary dimensions -- see for example
my recent posting on computing threaded variances over large arrays.
Quasi-anecdotal evidence also indicates it's the fastest method (but
not by much, and really only if you use the /SAMPLE keyword to REBIN).
Whatever method you choose, be sure to comment liberally, since this
is one place (of many) IDL code can quickly sink into the quagmire of
unreadability.
JD
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
