-------- WE'LL GET AROUND TO THIS ONE SOMEDAY ---------------------

From: nmartins@ualg.pt
Nelson Martins
Universidade do Algarve, UCEH, ADEEC, SiPLAB, Lab. 2.70
8000 Faro Portugal
Tel. +351 289800949

Date: Wed, 10 Apr 2002 19:19:09 +0100

I would like to suggest you that, in rgk.m, the M-file that computes
your radially Gaussian kernel distribution, you put a condition of the
kind

  pos = find( sigma1 < 1e-76 );
  if pos
    sigma1( pos ) = 1e-76;
  end
 
between the instruction

  sigma1 = sigma2 ;

and the instruction

  f1 = f2 ;

This is because, for some input signals the variable sigma1 takes on
very low values (of the order 1e-105, for example), what implies
numerical problems on the instruction
 
  grad = sum(P3.*polamb2(2:P,:).*Phi1(2:P,:)).*(2*dr^2*(sigma1).^(-3)) ;

and, finally, NaN for the time-frequency distribution.