General considerations about the modular structure of the
implementation leads to suppose that most of the noise is generated
when the state of the filter is stored
, and
when the gain is applied to the main flow. The noise generation is
modeled by multiplying the flow by 1+v where v is a white noise
whose variance depends on the number of digits used in computations
[6]. We simplify the model by assuming that the noise
is white. This simplificative approach for studying round-off errors
in the floating point arithmetic is enough for reliable results. The
variance of the impulse response is computed and compared to different
process realizations simulated by a test program in which the number
of significative digits can be controled.