As forseen, the distortion appears in the first half of the impulse response. The study reveals that the gain step is 6dB more sensitive than the storage step. This multiplication has to be computed with care, and one or two supplementary digits should be allocated in this operation in order to balance this noise overproduction.
The study finally results in an empiric estimation of the signal to
noise ratio (eq. 10), where g is the number of digits used in
the computations and N is the order of the LIF. In usual
applications running in single precision
, round-off noise
cannot be perceived. For orders lesser than 20, the signal to noise
ratio isn't less than 80dB.
