Let me reiterate again that the math you cited from Meni Rosenfeld was incomplete and thus incoherent. I explained that the same probabilistic guarantees can be obtain by increasing the number of confirmations rather than increasing the block reward.
In a system with no block reward, you can calculate the probability that an attacker can outpace the rest of the network from N blocks deficit. However, I don't see a way to put bounds on when it is safe to accept a transaction of size X as confirmed because the amount of PoW on top of any given transaction doesn't have a value which is easily observed.
I explained upthread that the math is incoherent and there is no bound with profitable proof of work either due to k. I will formalize the argument in a white paper, not now.
(which doesn't require there is necessarily no block reward but rather just that the block reward is less than profitable at the current difficulty)
If the block reward is unprofitable in a system with adjustable difficulty, this is not an equilibrium, so the difficulty will adjust downwards until it there is one. If your design includes other factors which you are not revealing, this may be different, but with what I know this is how I see it.
In my design every txn includes a PoW share thus difficulty will not reduce unless transactions reduce. The sender of the txn is entirely unconcerned about the cost of producing the PoW share.
The threshold of the difficulty to win a block will increase accordingly so that the block period remains constant.