The error condition is not stable with respect to numerical errors. For example, one may want to arrange a categorical generator so that the last probability is zero up to rounding errors.

Or, assume weights x, y, z are given. The corresponding probabilities are x/(x + y + z), y/(x + y + z), z/(x + y + z). The categorical generator of SRFI 194 would be:

(let ((s (+ x y z)))

  (make-categorical-generator (vector (/ x s) (/ y s))))

However, due to rounding errors, it may be that (negative? (- 1 (+ (/ x s) (/ y s)))) evaluates to true.

If you change the probability vector into a weight vector (as above), the numerical instability of the error condition vanishes. All you need is that the weights are non-negative and that their sum is positive.