I think I missed class when this was explained ...

Anyway as part of a bigger project I have to implement a LPC to predict 2-3 future values of a sinusoidal process. I wrote a small Matlab m-file to calculate the predictor coefficients and plot the resulting predicted values.

It does look well so far when moving around $y(x) \cong 0$ but the values around the maxima/minima $y(x) \cong \pm 1$ are completely off (it predicts into the opposite direction):

phi=pi/2, unk = 10

I was expecting something like this: phi=0, unk = 50

I'm using the variables 'phi' to set sort of a operation point and 'unk' to set the starting point from where values should start to get predicted.

Here my m-file:

N = 100;            % samples
fs = 60;            % sample-frequency [Hz]
Ts = 1/fs;          % sample-time [s]
unk = 10;           % future values to predict (mostly used to set x(k) where we start predicting from)
P = 10;             % predictor order
phi = 0;            % default phase-offset
phi = pi/2;

k = 0:1/N:2;
x = sin(2*pi*fs*k*Ts+phi);  % reference values

x_pred = x(1:101);          % intialize with x ...
                            % last 'unk' values will be replaced by the predicted ones
x_kn = x_pred(1:end-unk);   % values we already know (we will try to predict last 'unk' values)

coeffs = aryule(x_kn, P);      % get 'P' filter coeffs using the known input 'x_kn' values

for i = 1:unk
    % x^ = -vec(a) * vec(x_kn)'
    nextValue = -coeffs(2:end) * ...            % P - 1 coefficient vector
                x_pred(length(x_kn)-1+i: ...    % last value we actually know (at pos len(x_kn))
                -1: ...                         % reverse iterate ...
                length(x_kn)-P+i)';             % until P values before len(x_kn)
    x_pred(length(x_kn)+i) = nextValue;         % now we know the value at len(x_kn)+1

plot(0:length(x_kn)+unk-1 ,x_pred, 0:length(x)-1, x); grid;
legend('predicted', 'real');

1 Answer 1


Disclaimer: This is explanation is based on observation of my MatLab plots and my note be 100% textbook correct!

Well after much looking around and experimenting I read something about the 'yule-walker' method for estimating the coefficients assuming the signal to be 'zero' outside of the observation point. This is not the case for my co-/sine-wave, so the estimated coefficients and the resulting predicted samples $\hat{x}(n)$ will trend towards zero. The further away from the $x$-Axis the sharper the slope towards 'zero' when using yule-walker.

For my problem an approach using Burg's method (coeffs = arburg(x_kn, P);) seems to yield the result I'm looking for, since burg does 'not' assume the Signal trending towards 'zero'?? At least this is how I can explain the result of my two plots. The 'star' marks the point where the prediction of future values start.




enter image description here

  • $\begingroup$ Sounds plausible. You could also try aryule with a windowed version of your x_kn. For example, x_kn.*hann(length(x_kn))' to get rid of the jumps at the borders if that is extended with zeros. $\endgroup$
    – sellibitze
    Commented Sep 21, 2017 at 12:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.