Hi If I have a polynomial A(z)=1+a1*z^(-1)+a2*z^(-2)+.....+a10*z^(-10) and I solve A(z)=0 and write the solution as a vector [r1,r2,....,r10] then I can say that the action of finding the roots is a one-to-many transformation from the coefficient vector [a1,a2,.....,a10] to the root vector [r1,r2,...,r10].... Is it possible somehow to ensure that the above transformation of a given coefficient-vector always results in the same root vector ? For example: The above transformation of a coefficient vector a=[1,0.5,-0.25] will result in the root vector [-0.8090,0.3090] But [0.3090,-0.8090] is also a solution.....so is it possible somehow to solve this problem and ensure a one-to-one mapping ?? If not, how should/could I go about it then? Thank you .
problem when calculating roots based on LPC coefficients
Started by ●November 27, 2005
Reply by ●November 28, 20052005-11-28
Is it possible? Easily. Make a sorting rule that sorts the results based on number characteristics so that the sort is unambiguous. Use the resulting sorted vector. Why do you want to? Dirk
Reply by ●November 28, 20052005-11-28
Hi there... I am going to implement a Hidden Markov Model and as far as I can see it's important that I train the model with vectors that are unique, right? In my case the observation vector is a 10-dimensional complex vector. The entries in this vector are the roots of the polynomial A(z). As you can see - using the roots - means that the vector is _not_ unique unless I: a) split the vector up in 5 2-dimensional vectors and plot them and then use the generated pattern as a "state" or b) am able to ensure that there is a one-to-one mapping between the coefficient vector and the 10-dimensional complex vector by using some sort-rule Which sort rule do you propose? Thanks ------------> Why do you want to? > > Dirk >
Reply by ●November 28, 20052005-11-28
There are probably many rules. If simply making them unique (1-1) was the problem then consider sorting the roots by increasing magnitude, then when encountering the same magnitude, use increasing angle. That would be 1-1. But it sounds like you may have another problem in that a little perturbation of the roots could then change the order. So you need a sorting method that is insensitive to small changes. You might want to look at [first reflection coef, 2nd reflection coef, ...] since this has the order predetermined. Intuitively (based on whatever) I am guessing the RC values are less sensitive to small perturbations (may be more so for certain coefs). Check the literature. FWIW, Dirk