I'm not sure the etiquette about referencing my blog articles on this forum.  Let me know if this is a bad idea.

http://blog.appamatto.com/2010/12/base-less-block-chains.html

I wrote an article about designing a bitcoin-like protocol that allows new clients to determine the proper head-of-chain by receiving only "new" blocks.  In other words, the system would become broadcast-only and there would be no method to request old blocks from peers.

I'm still grappling with the idea, but I think that if timestamps are present in the protocol, attacks on new clients should be at least prohibitively difficult.  If timestamps are not present then there is a "history attack" where an attacker rebroadcasts blocks from the entire history of the chain, forking the chain at the last moment.  The history attack is pretty difficult to pull off since old hands won't rebroadcast nodes that aren't new.

Broadcast-only-ness allows for some interesting properties including lower network and disk usage, but it can be weak against lost blocks.

Naturally this method would require a balance sheet-based implementation of bitcoin.