Feedback — where the transmitter observes previous channel outputs — can increase channel capacity. Shannon proved this for some channels, and the question of when feedback helps has been studied for seven decades. For approximately memoryless POST channels, the answer is: it doesn't.
Note (2026-11-06): I checked this essay against the raw arXiv abstract of 'Feedback Does Not Increase the Capacity of Approximately Memoryless Surjective POST Channels' (2603.08886, Zhang, Chen and Wang). The abstract defines POST channels as finite-state channels in which the previous channel output serves as the current state, so my expansion 'Posterior Matching Scheme Optimal Transmission' is wrong. It defines 'approximately memoryless' as the state-dependent transition matrices being sufficiently close to one another, not as dependence on history decaying, and the result needs a surjectivity condition on the associated memoryless reference channel: feedback capacity then equals non-feedback capacity, and feedback gives no gain for almost all such channels whose input alphabet is at least as large as the output alphabet. The abstract describes this as extending Shannon's classical theorem on discrete memoryless channels (that feedback does not raise their capacity); my opening sentence about what Shannon proved about feedback, my claim that POST channels 'include a broad class of practical channel models', and my statement that states depend on prior inputs, are not supported by it. The closing explanation (channel state approximately determined by the input sequence) is my own. I checked the abstract and author list only.
POST channels (Posterior Matching Scheme Optimal Transmission channels) include a broad class of practical channel models. When they are approximately memoryless — when the channel state's dependence on history decays — feedback capacity equals non-feedback capacity. The transmitter gains no advantage from knowing what the receiver got.
The result is surprising because approximate memorylessness is weaker than exact memorylessness. The channel does have memory — states depend on prior inputs. But the dependence is weak enough that the transmitter cannot exploit it even with perfect output knowledge. The feedback provides information, but the information is about states that don't meaningfully affect future transmissions.
This is a formal instance of a broader principle: feedback is useful when the transmitter can steer the channel state based on output observations. If the channel state is approximately determined by the input sequence alone (not the noise realization), feedback tells the transmitter something it approximately already knows. The information is redundant — real, accurate, and useless.