I am using Burg's method of linear prediction to replace invalid sections of audio data. The method works fairly successfully 95% of the time but occasionally fails with a reflection coefficient >1 leading to an erroneous set of linear prediction coefficients. My programme , in Visual C++, is based on a paper by Cedrick Collomb to be found at :- http://www.emptyloop.com/technotes/A%20tutorial%20on%20Burg%27s%20method,%20algorithm%20and%20recursion.pdf His expression for the Reflection Coefficient Mu has Numerator and Denominator terms. Burg's Recursion focusses on calculating the Denominator of the (k+1)th order term from the kth order term. In order to isolate the problem I tried replacing the Recursion process with a straight forward calculation of the (k+1)th denominator ( more cpu intensive but simpler). I find the numbers produced by the Recursive process are quite different from the straightforward process but I cannot see why. I have worked through Cedrick Collomb's theory fairly rigorously but can see no errors. Has anyone any experience of this problem.

Code it up in Matlab or Scilab, and see if it works?  I don't know squat about Burg, but if an algorithm looks good on paper but misbehaves in C++ I'll see if I can do it in Scilab -- it's usually fewer lines, and easier to debug.  Then if the Scilab version works, I'll work on porting that to C++.

Thanks for the advice Tim

Not pertinent per se, but went to http://www.emptyloop.com/ to find that reference, got distracted by a tool he has (unlocker) to delete "sticky" files (I have a bunch of NAS folders with dead thumbs.db files I can't delete), but Chrome gave me a malware warning when I tried to view that link. Just a warning (if it's warranted). I have had other false warnings from Chrome and gmail. Just saying.

BTW: I ran into this: https://www.opus-codec.org/docs/vos_fastburg.pdf It implies a fast method (like an FFT) to compute the terms.

Thanks for the warning