| Author |
Message |
Tom Derham
Guest
|
 Posted:
Sun Jan 02, 2005 4:32 am Post subject:
covariance question |
|
|
If a zero-mean Gaussian random noise signal is at the input to a matched
filter bank (so the input to each filter is identical), what is the nature
of the outputs of each filter in the bank?
Assuming the filters are linear then each output should also be Gaussian,
but are they correlated or made independent by the filters.. is there a way
of calculating the covariance matrix from the transfer functions of the
filters?
I'm a bit stuck on this conceptually (currently reading Papoulis, etc), so
any help would be much appreciated
Thanks
Tom |
|
| Back to top |
|
 |
Tim Wescott
Guest
|
| Posted:
Sun Jan 02, 2005 7:28 am Post subject:
Re: covariance question |
|
|
Tom Derham wrote:
| Quote: | If a zero-mean Gaussian random noise signal is at the input to a matched
filter bank (so the input to each filter is identical), what is the nature
of the outputs of each filter in the bank?
|
If they are linear filters then it'll be zero-mean Gaussian.
| Quote: | Assuming the filters are linear then each output should also be Gaussian,
but are they correlated or made independent by the filters.. is there a way
of calculating the covariance matrix from the transfer functions of the
filters?
I'm a bit stuck on this conceptually (currently reading Papoulis, etc), so
any help would be much appreciated
Thanks
Tom
|
If the filters' spectra do not overlap then the outputs of the filters
will be uncorrelated. If the filters' spectra _do_ overlap then I have
every faith that there will be cross-correlation, and that there's a way
of calculating the covariance -- but I haven't faintest notion of what
that is.
--
Tim Wescott
Wescott Design Services
http://www.wescottdesign.com |
|
| Back to top |
|
|
Robin Clark
Guest
|
| Posted:
Sun Jan 02, 2005 4:21 pm Post subject:
Re: covariance question |
|
|
On Sat, 01 Jan 2005 18:28:07 +0000, Tim Wescott wrote:
| Quote: | Path:
newsbe2-gui.ntli.net!newspeer1-win.ntli.net!news-in.ntli.net!newsrout1-win.
ntli.net!ntli.net!sn-xit-02!sn-xit-01!sn-post-01!supernews.com!corp.superne
ws.com!not-for-mail
From: Tim Wescott <tim@wescottnospamdesign.com
Newsgroups: comp.dsp
Subject: Re: covariance question
Date: Sat, 01 Jan 2005 18:28:07 -0800
Organization: Posted via Supernews, http://www.supernews.com
Message-ID: <10ten1hlnb19j5a@corp.supernews.com
User-Agent: Mozilla/5.0 (Windows; U; Windows NT 5.1; en-US; rv:1.6)
Gecko/20040113
X-Accept-Language: en-us, en
References: <SoGdnYZ4v6awrkrcRVnysw@pipex.net
In-Reply-To: <SoGdnYZ4v6awrkrcRVnysw@pipex.net
X-Complaints-To: abuse@supernews.com
Lines: 32
Xref: newspeer1-win.ntli.net comp.dsp:168734
X-Received-Date: Sun, 02 Jan 2005 02:29:09 GMT (newsbe2-gui.ntli.net)
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii; format=flowed
Content-Transfer-Encoding: 7bit
Tom Derham wrote:
If a zero-mean Gaussian random noise signal is at the input to a matched
filter bank (so the input to each filter is identical), what is the nature
of the outputs of each filter in the bank?
If they are linear filters then it'll be zero-mean Gaussian.
If the filters' spectra _do_ overlap then I have
every faith that there will be cross-correlation, and that there's a way
of calculating the covariance -- but I haven't faintest notion of what
that is.
|
Co variance is the ratio between the variance of two signals.
I.e. take the variance of signal A and the variance of signal B.
The co-variance is A/B.
What it really tells you is how noisy (or reliable) A is compared to B.
--
[=========]
-==++""" . /. . . \ . """++==-
-+"" \ .. . . | .. . | . . . / ""+-
/\ +-"" `-----=====\ <O> /=====-----' ""-+ /\
/ / ""="" \ \
/ / \ \
// | \\
/") \ | / ("\
\o\ \*/ /o/
\ ) --**O**-- ( /
/*\
/ | \
| |
|
| Back to top |
|
|
Clay S. Turner
Guest
|
| Posted:
Sun Jan 02, 2005 8:02 pm Post subject:
Re: covariance question |
|
|
"Robin Clark" <robinTEETH48gx@hotTEETHmail.com> wrote in message
news:pan.2005.01.02.11.21.41.551837@hotTEETHmail.com...
| Quote: |
Co variance is the ratio between the variance of two signals.
I.e. take the variance of signal A and the variance of signal B.
The co-variance is A/B.
What it really tells you is how noisy (or reliable) A is compared to B.
|
Hello Robin,
I think you need to recheck your stats book. The above statements are wrong.
Covarince is communitive so therefore cannot be written as a ratio of
variances. Also recall the covariance of two indepenent random variables is
zero, yet each comes from a distribution with nonzero variances. So clearly
it cannot be what you stated.
Loosely if A tends to be large when B is large, then the covariance is
positive. Likewise if one of A, B tends to be small while the other is
large, then the covariance is negative. And if one's values don't associated
at all with the other's, then the covariance is zero.
Clay |
|
| Back to top |
|
|
Rune Allnor
Guest
|
| Posted:
Sun Jan 02, 2005 10:05 pm Post subject:
Re: covariance question |
|
|
Tom Derham wrote:
| Quote: | If a zero-mean Gaussian random noise signal is at the input to a
matched
filter bank (so the input to each filter is identical), what is the
nature
of the outputs of each filter in the bank?
Assuming the filters are linear then each output should also be
Gaussian,
but are they correlated or made independent by the filters.. is there
a way
of calculating the covariance matrix from the transfer functions of
the
filters?
|
Yes it is. Define the filtering operation as
y(t) = x(t) (*) h(t)
where x(t) is the input signal, h(t) is the filter impulse response,
y(t) is the filtered signal and (*) means convolution.
There is a standard exercise in intro statistical DSP on how to show
that the autocorrelation of filtered white noise is given by the
filtering function, something like
E[S_yy(f)] = E[S_xx(f)]|H(f)|^2
where S_yy(f) is the power spectrum density of y(t) and S_xx(f) is
the power spectrum density of x(t). The power density is defined
as the FT of the autocorrelation sequence,
S_xx(f) = FT{r_xx(tau)}.
Using the same kind of argument for the cross correlation sequences
of the filtered outputs, it is possible to find the cross correlation
sequences for the filtered sequences y_u(t) and y_v(t) defined as
y_u(t) = x(t) (*) u(t)
y_v(t) = x(t) (*) v(t),
to be something like
E[S_uv(f)] = S_xx(f) U(f)V(f).
Here, U(f) and V(f) are defined as the FTs of u(t) and v(t)
respectively,
and one of them should be complex conjugated.
| Quote: | I'm a bit stuck on this conceptually (currently reading Papoulis,
etc), so
any help would be much appreciated
|
Papoulis is a good place to start. The main lines ought to be something
like what I sketched above.
Rune |
|
| Back to top |
|
|
|
|
|
|
FAQ
Memberlist
Register
Profile
Log in to check your private messages
Log in