Forget everything : i was looking at angle(E1t) and not imag(E1t). The imag
noise is flat and less than 10^-14.
However, I still have a noise amplification problem in my code but this is
another story (see http://www.dsprelated.com/showmessage/54930/1.php).

Thanks.

P.S : It's hard not to post nonsense when you need help.

Ok,

I'm sorry I forgot to mention that ws was set to zero here. Actually the
noise I was talking a

mimizan wrote:

> function f=gaussienne(t,tau,ws)
> % The gaussian function
> f=exp(-(t/tau).^2).*exp(-i*ws*t);

I think you need to review the definition of the Gaussian function ...

Hint: it's real valued.

Regards,
Andor

mimizan skrev:
> >Hi,
> >  a true Gaussian function has infinite width, so
> >you had to truncate that function by some amount.
> >
> >Can you post your MATLAB code here?
>
> Hi,
>
> thanks for answering. What is strange is that the less I truncate, the
> worse it is. I suspect this is a newbie problem. Here are extracts from my
> code :
>
> m=40000; % Number of points

<split-hair-mode>
Well, no, it's half the number of points.
</split-hair-mode>

> tstep=0.002e-9;
> time=(-m:1:m-1)*tstep; % Time array; tstep is the sampling step
>
> function f=gaussienne(t,tau,ws)
> % The gaussian function
> f=exp(-(t/tau).^2).*exp(-i*ws*t);

This one puzzled me. What is 'ws'? Why don't you use 'tau'
with the imaginary part?

> tau=6e-9; % 6 ns pulse
> E0t=gaussienne(time,tau,0); % Input gaussian array
>
> E1t=ifft(fft(E0t)); % To test the ifft
>
> myplot2(time,E0t,time,E1t); % Custom function to compare both arrays
> 
> N.B : Changing the sampling interval doesn't help.

Rune

>Hi,
>  a true Gaussian function has infinite width, so 
>you had to truncate that function by some amount.
>
>Can you post your MATLAB code here?

Hi,

thanks for answering. What is strange is that the less I truncate, the
worse it is. I suspect this is a newbie problem. Here are extracts from my
code :

m=40000; % Number of points
tstep=0.002e-9;
time=(-m:1:m-1)*tstep; % Time array; tstep is the sampling step

function f=gaussienne(t,tau,ws)
% The gaussian function
f=exp(-(t/tau).^2).*exp(-i*ws*t);

tau=6e-9; % 6 ns pulse
E0t=gaussienne(time,tau,0); % Input gaussian array

E1t=ifft(fft(E0t)); % To test the ifft

myplot2(time,E0t,time,E1t); % Custom function to compare both arrays

N.B : Changing the sampling interval doesn't help.

On Tue, 11 Apr 2006 07:26:34 -0500, "mimizan" <spambox@neuf.fr> wrote:

>>ok this is really dumb...i will reply to myself!
>>rounding errors contribute to this, should take real(ifft(s)) or
>>abs(ifft(s)), i think there is not much difference since imaginary
>>parts are really small.
>>
>>jeremy
>
>Hi all,
>
>This, I understand. It is not a problem as long as the imaginary parts
>stay small. But what if they become larger ? Here's my problem :
>
>- Take a gaussian function of width tau, and sample if from, let's say,
>-5*tau to 5*tau. ifft(fft) only generates a very small imaginary part
>(<10^-10)
>
>- Take the same function but sample it from -15*tau to 15*tau. This time
>ifft(fft) generates very large (>1) imaginary parts in the edges of the
>gaussian !
>
>I thought that it was because the gaussian was too small at the edges so
>that numerical errors became too important in the ifft. So I forced the
>function to go to exact zero by multiplying by a square function, because
>I know that an exact zero will become an exact zero through ifft(fft). I
>didn't help, or even, worsen things.
>
>What should I do ?
>
>Thank you.

Hi,
  a true Gaussian function has infinite width, so 
you had to truncate that function by some amount.

Can you post your MATLAB code here?

[-Rick-]

>ok this is really dumb...i will reply to myself!
>rounding errors contribute to this, should take real(ifft(s)) or
>abs(ifft(s)), i think there is not much difference since imaginary
>parts are really small.
>
>jeremy

Hi all,

This, I understand. It is not a problem as long as the imaginary parts
stay small. But what if they become larger ? Here's my problem :

- Take a gaussian function of width tau, and sample if from, let's say,
-5*tau to 5*tau. ifft(fft) only generates a very small imaginary part
(<10^-10)

- Take the same function but sample it from -15*tau to 15*tau. This time
ifft(fft) generates very large (>1) imaginary parts in the edges of the
gaussian !

I thought that it was because the gaussian was too small at the edges so
that numerical errors became too important in the ifft. So I forced the
function to go to exact zero by multiplying by a square function, because
I know that an exact zero will become an exact zero through ifft(fft). I
didn't help, or even, worsen things.

What should I do ?

Thank you.

Jerry Avins wrote:
> jeremyscerri@gmail.com wrote:
> 
>> comp.dsp should include this in the acronyms list...
> 
> 
> We can do without cutsie for cutsie's sake  ...

Sometimes I get huffier than I should.

Jerry
-- 
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������

jeremyscerri@gmail.com wrote:
> comp.dsp should include this in the acronyms list...

We can do without cutsie for cutsie's sake, thank you. Even "tnx" is off 
the mark, but it's better, and as short. Much slang supports an "us vs. 
them" outlook; knowing the meaning shows that one is part of the "in" 
group. Everybody is welcome in comp.dsp, even those who struggle with 
English. There is no need to demonstrate one's "membership". Remember: 
this is a quasi-professional international forum, not a chat room where 
quick turnaround can outweigh clarity and precision of expression.

Jerry
-- 
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������

comp.dsp should include this in the acronyms list...

jeremy