Hi, matrices which are bad conditioned are "nearly singular",
that is, the inverse of their condition number is near the
precision of float/double.
In the end this means there are columns / rows of the matrix,
which are "nearly" linear dependent.
"Singularity" and "linear dependence" are theoretical
concepts which work well in the head of pure mathematicians,
but if you switch to computers with their limited precision,
nearly singular matrices behave like singular ones.
Translated to SVM training: bad conditioning occurs if you
use learning examples which are very near to each other
together with learning examples which are wide spread.
Nearness of two examples means that their kernel values
are near to the maximal value.
For radial kernel this means there are x1, x2 with
||x1-x2|| << 1.
You can measure condition numbers in matlab with the
"cond" command.
Greetings, Uwe
|