After playing with numbers, this effect makes sense to me. Here is a simple model.
Elo formula is pwin = 1/(1+10^(-rdiff/400)), where rdiff is the ratings difference between 2 players.
So knowing probability of winning over a rated opponent, our rating should be:
rating = opp.rating - LOG(1/pwin-1)*400.
Here is a copy of a spreadsheet that shows rating dropping after a win:
N games 10 new opp 1200
N wins 9
pwin 0.9 pwin 0.909
sos 1500 sos 1472.7
rating 1882 rating 1873
sos is average strength of schedule. We start from winning 9 out of 10 games, and our average opponent is rated at 1500. Our rating should be 1882.
We add a win over a 1200-rated opponent. pwin goes up slightly to 10/11~91%, while average opponent rating goes down slightly to ~1473. And now the ELO formula says that given those two numbers, our rating should be 1873, 9 points lower than before. Drop in the strength of schedules has a bigger effect than increase in winning percentage.
So it's not an effect of propagating wins and losses through the rest of the field. This is directly the result of using the logistic function in the ELO formula, which has this flaw in it of dealing with opponents at very different ratings.