Pseudo-determinant: Difference between revisions

From formulasearchengine
Jump to navigation Jump to search
en>Myasuda
en>Brirush
mNo edit summary
 
Line 1: Line 1:
'''Signal averaging''' is a [[signal processing]] technique applied in the [[time domain]], intended to increase the strength of a [[Signal (electronics)|signal]] relative to [[noise]] that is obscuring it. By averaging a set of [[Replication (statistics)|replicate]] measurements, the [[signal-to-noise ratio]], S/N, will be increased, ideally in proportion to the square root of the number of measurements.
Hi, everybody! My name is Eduardo. <br>It is a little about myself: I live in Germany, my city of Prezelle. <br>It's called often Eastern or cultural capital of NI. I've married 1 years ago.<br>I have two children - a son (Lina) and the daughter (Clifton). We all like Audiophilia.<br><br>Look into my site :: [http://www.calscottages.com/uggboots.asp http://www.calscottages.com/uggboots.asp]
 
== The ideal case ==
 
Ideally it is assumed that
* Signal and noise are uncorrelated.
* Signal strength is constant in the replicate measurements.
* Noise is random, with a [[mean]] of zero and constant [[variance]] in the replicate measurements.
 
Under these assumptions let the signal strength be denoted by ''S'' and let the [[standard deviation]] of a single measurement be σ; this represents the noise, ''N'', in one measurement. If ''n'' measurements are added together the sum of signal strengths will be ''nS''. For the noise, the standard [[error propagation]] formula shows that the [[variance]], σ<sup>2</sup>, is additive. The variance of the sum is equal to ''n''σ<sup>2</sup>. Hence the signal-to-noise ratio, S/N, is given by
:<math>\frac{S}{N}=\frac{nS}{\sqrt{n\sigma^2}}= \sqrt{n} \frac{S}{\sigma}</math> 
The equivalent expression for signal averaging is obtained by dividing both numerator and denominator by ''n''.
:<math>\frac{S}{N}=\frac{S}{\sqrt{{1\over n}\sigma^2}}= \sqrt{n} \frac{S}{\sigma}</math>
Thus, in the ideal case S/N increases with the square root of the number of measurements that are averaged. In practice, the assumptions may not be fully realized, this will result in a lower S/N improvement than in the ideal case, but in many cases near-ideal S/N improvement can be achieved.<ref>Wim van Drongelen. "Signal Processing for Neuroscientists." Academic Press 2008. (Ch. 4, pg 55)</ref>
 
==Time-Locked Signals==
 
Averaging is applied to enhance a time-locked signal component in noisy measurements.
 
==Averaging Odd and Even Trials==
 
A specific way of obtaining replicates is to average all the odd and even trials in separate buffers. This has the advantage of allowing for comparison of even and odd results from interleaved trials. An average of odd and even averages generates the completed averaged result, while the difference between the odd and even averages constitutes an estimate of the noise.
 
==Algorithmic Implementation==
 
The following is a MATLAB simulation of the averaging process:
 
    % create [sz x sz] matrix
    % fill the matrix with noise
    sz=256;
    NOISE_TRIALS=randn(sz);
    % create signal with a sine wave
    % divide the array SZ by sz/2
    SZ=1:sz;                       
    SZ=SZ/(sz/2);                 
    S=sin(2*pi*SZ);
        for i=1:sz;                     
            NOISE_TRIALS(i,:) = NOISE_TRIALS(i,:) + S;
        end;
    % create the average
    average=sum(NOISE_TRIALS)/sz; 
    odd_average=sum(NOISE_TRIALS(1:2:sz,:))/(sz/2);
    even_average=sum(NOISE_TRIALS(2:2:sz,:))/(sz/2);
    noise_estimate=odd_average-even_average;
    % create plot
    figure
    hold
    plot(NOISE_TRIALS(1,:),'g')
    plot(noise_estimate,'k')
    plot(average,'r')
    plot(S)
    xlabel('Trials')
    ylabel('Amplitude')
    title('Signal Averaging')
    gtext('Signal: Blue')
    gtext('Single trial: Green')
    gtext('Noise estimate: Black')
    gtext('Average: Red')
 
The averaging process above, and in general, results in an estimate of the signal. When compared with the raw trace, the averaged noise component is reduced with every averaged trial. When averaging real signals, the underlying component may not always be as clear, resulting in repeated averages in a search for consistent components in two or three replicates. It is unlikely that two or more consistent results will be produced by chance alone.
 
==Non-random Noise==
 
Signal averaging typically relies heavily on the assumption that the noise component of a signal is random, having zero mean, and being unrelated to the signal. However, there are instances in which the noise is not random. A common example of non-random noise is a [[Hum (sound)|hum]] (e.g. 60&nbsp;Hz noise originating from power lines). To apply the signal averaging technique, a few critical adaptations must be made, and the problem can be avoided by:
* Randomizing the stimulus interval, or
* Using a noninteger stimulus rate (e.g. 3.4&nbsp;Hz instead of 3.0&nbsp;Hz)
 
==References==
<references/>
 
{{reflist}}
 
{{DSP}}
 
[[Category:Digital signal processing]]
[[Category:Signal processing|Averaging]]

Latest revision as of 06:51, 26 November 2014

Hi, everybody! My name is Eduardo.
It is a little about myself: I live in Germany, my city of Prezelle.
It's called often Eastern or cultural capital of NI. I've married 1 years ago.
I have two children - a son (Lina) and the daughter (Clifton). We all like Audiophilia.

Look into my site :: http://www.calscottages.com/uggboots.asp