1. Routing protocol - RIPv2
2. R1 advertises summary route on both it's links
3. S1 is a multipoint interface. R1, R2 and R3 are in the same subnet
4. Split horizon is disabled on S1
Normal network condition: R2 receives update on both links and advertises it through S1 interface back to R1 and to R3 (Split horizon is off). Next-Hop address in the update is R1's serial interface. R1 receives it's own update, check that Next Hop IP in update it's on IP address and discards it. R3 installs the summary into it's routing table with two GW as an NH - R1 and R2. Remember, that R2 advertises both routes to the summary subnet - one through itself and the other one through R1 directly.
Now, suppose, on R2's S1 interface we filter this summary inbound. Now we advertise the summary route that we received on a point to point link (S0 int) back to R1 through the S1. R1 receives it and installs it as a valid route. After several seconds it realizes that it's own route and poisons it (sends this summary to frame-relay link with metric of 16 hops). R2 receives poisoned route, then poisons it back to R1. After some time the process repeats. If we are pinging this summary from R3 sometimes we will receive replies and sometimes we will receive Unreachables (the moment the route is poisoned).
Solution - it wouldn't be a problem at all if R1 installed a summary to Null0 to it's routing table when it advertised this summary to its neighbors.
I would like to thank INE workbooks for this interesting scenarios. I duplicate it here for future reference and memorizing the concept.
