From: third_person on

Hi, I'm trying fast convolution property but there seems to be some mistake
(with the answer).

Here is the Matlab code for it,

clear all; clc;

% Test Vector Convolution

a = [1 2 3 4 5];
b = [10 20 30 40 50];
c= conv(a,b)
A = [1 2 3 4 5 zeros(1,5)];
B = [10 20 30 40 50 zeros(1,5)];
d= ifft(fft(A) .* fft(B))
c - d(1:9)

The results are:

c =

10 40 100 200 350 440 460 400 250

d =

10.0000 40.0000 100.0000 200.0000 350.0000 440.0000 460.0000
400.0000 250.0000 0.0000

ans = (for the difference b/w the two)

1.0e-012 *

-0.0906 -0.0497 -0.0284 -0.0284 0 0.1137 0 0 0


Shouldn't the result be smaller than eps (2.2204e-016)?

From: Raymond Toy on
On 6/29/10 12:26 PM, third_person wrote:
> Hi, I'm trying fast convolution property but there seems to be some mistake
> (with the answer).
>
> Here is the Matlab code for it,
>
> clear all; clc;
>
> % Test Vector Convolution
>
> a = [1 2 3 4 5];
> b = [10 20 30 40 50];
> c= conv(a,b)
> A = [1 2 3 4 5 zeros(1,5)];
> B = [10 20 30 40 50 zeros(1,5)];
> d= ifft(fft(A) .* fft(B))
> c - d(1:9)
>
> The results are:
>
> c =
>
> 10 40 100 200 350 440 460 400 250
>
> d =
>
> 10.0000 40.0000 100.0000 200.0000 350.0000 440.0000 460.0000
> 400.0000 250.0000 0.0000
>
> ans = (for the difference b/w the two)
>
> 1.0e-012 *
>
> -0.0906 -0.0497 -0.0284 -0.0284 0 0.1137 0 0 0
>
>
> Shouldn't the result be smaller than eps (2.2204e-016)?
>

Why do you think it should smaller than eps? Do you think fft and ifft
have no roundoff?

I don't know what the actual roundoff should be but a difference of
1e-13 seems fairly reasonable.

Ray

From: Greg Heath on
On Jun 29, 12:26 pm, "third_person"
<third_person(a)n_o_s_p_a_m.ymail.com> wrote:
> Hi, I'm trying fast convolution property but
> there seems to be some mistake
> (with the answer).
>
> Here is the Matlab code for it,
>
> clear all; clc;
>
> % Test Vector Convolution
>
> a = [1 2 3 4 5];
> b = [10 20 30 40 50];
> c= conv(a,b)
> A = [1 2 3 4 5 zeros(1,5)];
> B = [10 20 30 40 50 zeros(1,5)];
> d= ifft(fft(A) .* fft(B))
> c - d(1:9)
>
> The results are:
>
> c =
>
>     10    40   100   200   350   440   460   400   250
>
> d =
>
>    10.0000   40.0000  100.0000  200.0000  350.0000  440.0000  460.0000
> 400.0000  250.0000    0.0000
>
> ans =  (for the difference b/w the two)
>
>   1.0e-012 *
>
>    -0.0906   -0.0497   -0.0284   -0.0284  0  0.1137  0  0  0
>
> Shouldn't the result be smaller than eps (2.2204e-016)?

Typically,
1. A and B are zeropadded to length(A)+length(B)-1
2. If A an B are real, d = real(ifft(fft(A) .* fft(B))
is used because ifft is notorious for creating
spurious imaginary roundoff error

However, the results I obtained below surprised me
(Note the change in notation)

clear all, clc

a = [1 2 3 4 5]';
b = [10 20 30 40 50]';
c= conv(a,b) ;
A = [1 2 3 4 5 zeros(1,4)]';
B = [10 20 30 40 50 zeros(1,4)]';
C = ifft(fft(A) .* fft(B));
D = [c C(1:9)]

% D =
%
% 10 10
% 40 40 +1.2632e-014i
% 100 100 -6.3159e-015i
% 200 200
% 350 350 -6.3159e-015i
% 440 440 -7.7816e-015i
% 460 460
% 400 400 -6.3159e-015i
% 250 250 +1.4097e-014i

A = [1 2 3 4 5 zeros(1,5)]';
B = [10 20 30 40 50 zeros(1,5)]';
C = ifft(fft(A) .* fft(B));
D = [c C(1:9)]

% D =
%
% 10 10
% 40 40
% 100 100
% 200 200
% 350 350
% 440 440
% 460 460
% 400 400
% 250 250

I can't explain it. Can someone else?

Greg
From: robert bristow-johnson on
On Jun 30, 9:53 am, Greg Heath <he...(a)alumni.brown.edu> wrote:
> On Jun 29, 12:26 pm, "third_person"
>
>
>
> <third_person(a)n_o_s_p_a_m.ymail.com> wrote:
> > Hi, I'm trying fast convolution property but
> > there seems to be some mistake
> > (with the answer).
>
> > Here is the Matlab code for it,
>
> > clear all; clc;
>
> > % Test Vector Convolution
>
> > a = [1 2 3 4 5];
> > b = [10 20 30 40 50];
> > c= conv(a,b)
> > A = [1 2 3 4 5 zeros(1,5)];
> > B = [10 20 30 40 50 zeros(1,5)];
> > d= ifft(fft(A) .* fft(B))
> > c - d(1:9)
>
> > The results are:
>
> > c =
>
> >     10    40   100   200   350   440   460   400   250
>
> > d =
>
> >    10.0000   40.0000  100.0000  200.0000  350.0000  440.0000  460.0000
> > 400.0000  250.0000    0.0000
>
> > ans =  (for the difference b/w the two)
>
> >   1.0e-012 *
>
> >    -0.0906   -0.0497   -0.0284   -0.0284  0  0.1137  0  0  0
>
> > Shouldn't the result be smaller than eps (2.2204e-016)?
>
> Typically,
> 1. A and B are zeropadded to length(A)+length(B)-1
> 2. If A an B are real, d = real(ifft(fft(A) .* fft(B))
> is used because ifft is notorious for creating
> spurious imaginary roundoff error
>
> However, the results I obtained below surprised me
> (Note the change in notation)
>
> clear all, clc
>
> a = [1 2 3 4 5]';
> b = [10 20 30 40 50]';
> c= conv(a,b) ;
> A = [1 2 3 4 5 zeros(1,4)]';
> B = [10 20 30 40 50 zeros(1,4)]';
> C = ifft(fft(A) .* fft(B));
> D = [c C(1:9)]
>
> % D =
> %
> %   10         10
> %   40         40 +1.2632e-014i
> %  100        100 -6.3159e-015i
> %  200        200
> %  350        350 -6.3159e-015i
> %  440        440 -7.7816e-015i
> %  460        460
> %  400        400 -6.3159e-015i
> %  250        250 +1.4097e-014i
>
> A = [1 2 3 4 5 zeros(1,5)]';
> B = [10 20 30 40 50 zeros(1,5)]';
> C = ifft(fft(A) .* fft(B));
> D = [c C(1:9)]
>
> % D =
> %
> %    10           10
> %    40           40
> %   100          100
> %   200          200
> %   350          350
> %   440          440
> %   460          460
> %   400          400
> %   250          250
>
> I can't explain it. Can someone else?

i don't understand what's troubling you, Greg. is it the extremely
tiny imaginary values that result (presumably from roundoff) when N=9
that don't when N=10?

r b-j

From: Greg Heath on
On Jun 30, 12:03 pm, robert bristow-johnson
<r...(a)audioimagination.com> wrote:
> On Jun 30, 9:53 am, Greg Heath <he...(a)alumni.brown.edu> wrote:
>
>
>
>
>
> > On Jun 29, 12:26 pm, "third_person"
>
> > <third_person(a)n_o_s_p_a_m.ymail.com> wrote:
> > > Hi, I'm trying fast convolution property but
> > > there seems to be some mistake
> > > (with the answer).
>
> > > Here is the Matlab code for it,
>
> > > clear all; clc;
>
> > > % Test Vector Convolution
>
> > > a = [1 2 3 4 5];
> > > b = [10 20 30 40 50];
> > > c= conv(a,b)
> > > A = [1 2 3 4 5 zeros(1,5)];
> > > B = [10 20 30 40 50 zeros(1,5)];
> > > d= ifft(fft(A) .* fft(B))
> > > c - d(1:9)
>
> > > The results are:
>
> > > c =
>
> > >     10    40   100   200   350   440   460   400   250
>
> > > d =
>
> > >    10.0000   40.0000  100.0000  200.0000  350.0000  440..0000  460.0000
> > > 400.0000  250.0000    0.0000
>
> > > ans =  (for the difference b/w the two)
>
> > >   1.0e-012 *
>
> > >    -0.0906   -0.0497   -0.0284   -0.0284  0  0.1137  0  0  0
>
> > > Shouldn't the result be smaller than eps (2.2204e-016)?
>
> > Typically,
> > 1. A and B are zeropadded to length(A)+length(B)-1
> > 2. If A an B are real, d = real(ifft(fft(A) .* fft(B))
> > is used because ifft is notorious for creating
> > spurious imaginary roundoff error
>
> > However, the results I obtained below surprised me
> > (Note the change in notation)
>
> > clear all, clc
>
> > a = [1 2 3 4 5]';
> > b = [10 20 30 40 50]';
> > c= conv(a,b) ;
> > A = [1 2 3 4 5 zeros(1,4)]';
> > B = [10 20 30 40 50 zeros(1,4)]';
> > C = ifft(fft(A) .* fft(B));
> > D = [c C(1:9)]
>
> > % D =
> > %
> > %   10         10
> > %   40         40 +1.2632e-014i
> > %  100        100 -6.3159e-015i
> > %  200        200
> > %  350        350 -6.3159e-015i
> > %  440        440 -7.7816e-015i
> > %  460        460
> > %  400        400 -6.3159e-015i
> > %  250        250 +1.4097e-014i
>
> > A = [1 2 3 4 5 zeros(1,5)]';
> > B = [10 20 30 40 50 zeros(1,5)]';
> > C = ifft(fft(A) .* fft(B));
> > D = [c C(1:9)]
>
> > % D =
> > %
> > %    10           10
> > %    40           40
> > %   100          100
> > %   200          200
> > %   350          350
> > %   440          440
> > %   460          460
> > %   400          400
> > %   250          250
>
> > I can't explain it. Can someone else?
>
> i don't understand what's troubling you, Greg.  is it the extremely
> tiny imaginary values that result (presumably from roundoff) when
> N=9 that don't when N=10?

I usually get imaginary valued roundoff when I use
ifft and the result should be real. I was intrigued
that, in contrast, the OP got purely real roundoff.

Then I noticed that he used one more zero than necessary
in the zeropadding. So, I removed the extra zero and
got purely imaginary roundoff.

Satisfied that my understanding was validated, I put
the extra zero back in to see if that was the cause of
the real valued roundoff...

Much to my surprise, my calculation resulted in no
roundoff error.

I find this puzzling, even intriguing, but certainly
not troublesome.

Greg