We now know reversibility costs time and tape. But does it cost power? When you restrict a model of computation — take away nondeterminism, cap the tape, forbid loops — the class of problems it can solve usually shrinks or shifts. Reversibility feels like a heavy restriction: never forget anything, ever. And there is a matching temptation in the other direction — surely a machine that runs backwards as happily as forwards can do something magic, like starting from an answer and rewinding to the question? This lesson settles both hopes and both fears. The punchline is beautifully deflating: as far as computability is concerned, reversibility changes nothing at all — no power lost, no magic gained.
The
Notice how the two statements click together. A reversible machine's overall input–output map must be injective — each step is injective, and composing injections gives an injection, so two different starting tapes can never end identically. That looks like a real limitation until you see that carrying the input along converts any function into an injective one. The restriction is real but toothless: it dictates the packaging, not the content.
Could the extra discipline help? Reversible machines are so orderly — no merging, a single thread of
history — that you might hope their behaviour is easier to predict. In particular, is
The seductive one: "a reversible machine for
Draw every configuration as a dot and every step as an arrow. An irreversible machine's graph has funnels — dots with two or more incoming arrows. A reversible machine's graph has in-degree at most one everywhere, so its runs are clean threads. That simple picture pins down exactly what repeating behaviour is possible:
Suppose a reversible run revisits some configuration: the configuration at step
A common false syllogism: "reversible machines never revisit a configuration (except by cycling through the start), so they can't run forever, so halting should be decidable for them." The first clause is right and the rest is wrong. Injectivity forbids merging, not divergence: a reversible machine may march forever through brand-new configurations — think of the NOT-sweep gliding right along an infinite tape of ones, each configuration fresh because the head is one cell further along. Infinite, non-repeating, perfectly reversible. Be precise about the trichotomy: a reversible run either halts, cycles through its start forever, or wanders through infinitely many distinct configurations — and telling these apart is, as we just proved, undecidable.
If computability is untouched, why did we bother? Because the interesting action was never at the yes/no level of computability — it is in the resources. Reversibility leaves the map of computable problems identical but redraws the price tags: how much time, how much space, how much energy. That is also the historically honest reading of Bennett's 1973 paper — its point was never "reversible machines are universal, hooray" but "therefore the thermodynamic cost of computing can be pushed arbitrarily close to zero," with Landauer's principle waiting in the background. The next lessons live entirely on this resource battleground: a pebble game that makes the space–time tension exact, and complexity theorems (including one genuine shock) about what reversibility costs — and what, surprisingly, it doesn't.