You are right, obl dows not work, since other hyper planes are required for
convolution.
Here is my (general) solution:
conv=: 4 : 0
nsh=.<: +/"1 x ,.&$ y
plns=. (((],:+)i.)-:# sh) +/"2@:{&.|: (#: [:i. */) sh=.x ,&$ y
plns=. plns {~ t=./: plns
nsh $ plns +//. t{ ,x */ y
)
g-: f conv h
1
It is also consistent with the latest result of Ambrus:
a conv b
20 20 0
10 2 2
0 1 0
R.E. Boss
> -----Oorspronkelijk bericht-----
> Van: [email protected] [mailto:programming-
> [email protected]] Namens Raul Miller
> Verzonden: dinsdag 2 maart 2010 23:56
> Aan: Programming forum
> Onderwerp: Re: [Jprogramming] convolution
>
> On Tue, Mar 2, 2010 at 5:22 PM, Dan Bron <[email protected]> wrote:
> > Raul wrote:
> >> We can use /. for convolution (+//.@(*/)) of one dimensional
> >> arrays, but do we have anything that simple for higher
> >> dimensional arrays?
> >
> > I haven't looked into it, but would RE Boss' proposal for a different
> definition of /. (oblique) help generalize +//.@(*/) ?
> >
> > http://www.jsoftware.com/pipermail/programming/2006-December/004410.html
>
> It might, but I do not see how.
>
> Using the above definition of obl and the 2d data from that
> wiki page:
>
> h=:".;._2]0 :0
> _8 1 _7 _2 _9 4
> 4 5 _5 2 7 _1
> _6 _3 _3 _6 9 5
> )
>
> f=:".;._2]0 :0
> _5 2 _2 _6 _7
> 9 7 _6 5 _7
> 1 _1 9 2 _7
> 5 9 _9 2 _5
> _8 5 _2 8 5
> )
>
> g=:".;._2]0 :0
> 40 _21 53 42 105 1 87 60 39 _28
> _92 _64 19 _167 _71 _47 128 _109 40 _21
> 58 85 _93 37 101 _14 5 37 _76 _56
> _90 _135 60 _125 68 53 223 4 _36 _48
> 78 16 7 _199 156 _162 29 28 _103 _10
> _62 _89 69 _61 66 193 _61 71 _8 _30
> 48 _6 21 _9 _150 _22 _56 32 85 25
> )
>
> g -: f +/obl@(*/) h
> 0
> $ f +/obl@(*/) h
> 16
>
> Thanks,
>
> --
> Raul
> ----------------------------------------------------------------------
> For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
