I am trying to implement a newton-raphson iteration to estimate sinusoidal frequency mixed with noise. Here's the problem statement: A white Gaussian noise process w(n) with variance sigma^2 = 1 is modulated by a sinusoid cos(2pif0n)*w(n) to yield the data set x(n) = cos(2pif0n)*w(n), n = 0, 1, ... , N-1, where N = 1000. The frequency f0 of the sinusoid is known to be in the range 0 < f0< ¼. Here's how the problem is seen, please correct me if I'm wrong: The best way it seems to solve this problem is to write down the pdf, ie, likelihood function, and then search it over the interval 0<fo<1/4. (this will be done in matlab). In this problem there is actually no signal but a noise process that is modulated. The nth sample is Gaussian with mean zero and variance sigma^2 = 1, which is time varying. All the noise samples are independent. The way to solve is to just write down the pdf and maximize over f_0. A maximum should be seen at the correct f_0. My problem is when initially writing down the PDF. With additive noise, I have typically done this as w(n) = x(n) - cos(*), here though (for multiplicative noise) if solving for w(n) we have w(n) = x(n)/cos(*)???? With additive noise the pdf => p(x;f)=[(2*pi*sigma^2)^-N/2]*exp[(-1/2sigma^2)*Esum(x(n)-cos2pifn)^2] -->as Gaussian form But my question is, how do I write out the pdf with multiplicate (rather than additive) noise??? This is where I am stuck.
Newton-Rapshon method over a multiplicative or modulated noise process
Started by ●April 29, 2008
Reply by ●April 30, 20082008-04-30
Update: after giving it some more thought, this pdf is a standard normal distribution: pdf = exp[-x^2/2]/sqrt(2pi) Now the concern is: do I just put the modulated cosine value into the pdf as the x value in order to proceed with setting up the algorithm? As so: pdf = exp[-(cos(2pi*fo*n)*w(n))^2/2]/sqrt(2pi) ?????? Any feedback on this is appreciated>I am trying to implement a newton-raphson iteration to estimatesinusoidal>frequency mixed with noise. Here's the problem statement: > >A white Gaussian noise process w(n) with variance sigma^2 = 1 is > >modulated by a sinusoid cos(2pif0n)*w(n) to yield the data set > > x(n) = cos(2pif0n)*w(n), n = 0, 1, ... , N-1, > > where N = 1000. The frequency f0 of the sinusoid is known to be in > > the range 0 < f0< ¼. > > >Here's how the problem is seen, please correct me if I'm wrong: > >The best way it seems to solve this problem is to write down the pdf,ie,>likelihood function, and then search it over the interval 0<fo<1/4. >(this will be done in matlab). > >In this problem there is actually no signal but a noise process that is >modulated. The nth sample is Gaussian with mean zero and variancesigma^2>= 1, which is time varying. > >All the noise samples are independent. > >The way to solve is to just write down the pdf and maximize over f_0.A>maximum should be seen at the correct f_0. > >My problem is when initially writing down the PDF. > >With additive noise, I have typically done this as w(n) = x(n) -cos(*),>here though (for multiplicative noise) if solving for w(n) we have w(n)=>x(n)/cos(*)???? > >With additive noise the pdf => >p(x;f)=[(2*pi*sigma^2)^-N/2]*exp[(-1/2sigma^2)*Esum(x(n)-cos2pifn)^2]-->as>Gaussian form > >But my question is, how do I write out the pdf with multiplicate (rather >than additive) noise??? > > >This is where I am stuck. > > > > >