Sadly any attempt to find an "optimal" block size is likely doomed because it can be gamed and "optimal" is hard to quantity.

Optimal for the short thinking non-miner - block size large enough that fees are driven down to zero or close to it.

Optimal for the network - block size large enough to create sufficient competition that fees can support the network relative to true economic value.

Optimal for the short term looking miner - never rising larger than 1MB to maximize fee revenue.

However I would point out that the blockchain may eventually become the equivalent of bank wire transactions.  FedWire for example transferred ~$663 trillion USD in 2011 using 127 million transactions.  If FedWire used a 10 minute block it would be ~2,500 transactions per block.  For Bitcoin that would be roughly 400 bytes per tx.  So it shows that Bitcoin can support a pretty massive fund transfer network even with a 1MB block limit.

Some would dismiss this as too centralized but I would point out that direct access to FedWire is impossible for anyone without a banking charter.  Direct access to the blockchain simply requires payment of fee and computing resources capable of running a node.  This means the blockchain will always remain far more open. 

I think one modest change (which is unlikely to make anyone happy but would allow higher tx volume) would be to take it out of the hands of everyone.  The block subsidy follows a specific exact path for a reason.  If it was open to human control Bitcoin would likely be hyperinflated and nearly worthless today.  A proposal could be made for a hard fork to double (or increase by some other factor) the max size of a block on every subsidy cut. 

This would allow for example (assuming avg tx is 400 bytes):
2012 - 1MB block =~ 360K daily transactions (131M annually)
2016 - 2MB block =~ 720K daily transactions (262M annually)
2020 - 4MB block =~ 1.44M daily transactions (525M annually)
2024 - 8MB block =~ 2.88M daily transactions (1B annually)
2028 - 16MB block =~5.76M daily transactions (2B annually)
2030 - 32MB block =~11.52M daily transactions (4B annually)

Moore's law should ensure that processing a 32MB block in 2030 is computationally less of a challenge than doing so with a 1MB block today.