In Steinbach [1], packet interrarrival times are deterministic. The channel model is Markov as in the previous case (see figure 1), but now when the model is in the good state, a packet arrives every seconds. In the bad state packets arrive in error. Thus, unlike Yuang, Steinbach's analysis considers missing packets and retransmissions. By making simplifying assumptions, Steinbach provides accurate expressions for the probability of underflow given average burst lengths, and the expected amount of latency for a given adaptive playout scheme and channel parameters. One simplifying assumption is that burst error states occur far enough apart in time so that the buffer can return to its target level between burst errors. The other assumption is that there is excess bandwidth available to allow retransmissions and regularly scheduled packets to arrive concurrently.

First, Steinbach finds the average value of the maximum sustainable burst error length, as a function of playout time for a given target buffer size and adaption scheme.

The probability of underflow given a burst is then,

where is is distributed geometrically with parameter (see figure 1).
Similarly, the average excess delay during a burst error is given as a function of *s*, the
slowdown factor, the speedup factor allowable with the given excess bandwidth,
the burst error length, , the length of the good channel period , and the packet
and frame durations and :

The expected amount of delay during a streaming session is found by taking the expected value with respect to the distributions on , (Geom( )), and (Geom( )) of equation 9.

Tue Mar 13 05:01:37 PST 2001