More than TPS, which is a meaningless measure when in presence of an adaptive block size, it makes more sense to speak about throughput. Early experimentation shows that we can achieve extreme throughput by propagating block and block hashes separately (and worry about consistency later, as reconciliation would be indeed a separate matter). In this case it would not be excluded a throughput roughly 10^2 that of Bitcoin.

This expectation is dependent on a lot of configuration throttling, as it would be connected to a few implementation choices. The most important of which are probably the ring signature size and the flavour of proof range used (bulletproof being my first candidate)