The paper says that the N by N data matrices A are the
leading N by N submatrices of the 63 by 63 covariance matrix
that is stored in the eigprod.m file.  Actually, the results
were computed, for each N, for this N by N matrix normalized
so that its largest entry is 1.  In other words, the following
Matlab commands must be used to obtain the correct data matrix:

load eigprod
A = A(1:N,1:N);
A = A/max(max(abs(A))); % so largest entry is 1

Thanks to Rohollah Yousefpour for discovering this error.