Is Space a Stronger Resource than Time? Positive Answer for the Nondeterministic at-Least-Quadratic Time Case

We show that all languages accepted in time f(n) >= n^2 can be accepted in space O(f(n)^{1/2})and in time O(f(n)). The proof is carried out by simulation, based on the idea of guessing the sequences of internal states of the simulated TM when entering certain critical cells, whose location is also guessed. Our method cannot be generalised easily to many-tapes TMs, and in no case can it be relativised.

💡 Research Summary

The paper investigates a fundamental question in computational complexity: how tightly can time and space be traded off for nondeterministic Turing machines (NTMs) that run in at‑least quadratic time. The authors prove that any language accepted by a single‑tape NTM in time f(n) ≥ n² can also be accepted by another NTM that runs in the same asymptotic time O(f(n)) while using only O(√f(n)) space.

The construction proceeds by partitioning the tape of the original machine M into blocks of length L = ⌈√f(n)⌉. Consequently the whole tape consists of O(√f(n)) blocks. The simulator S works nondeterministically: when M is about to enter a new block, S guesses the complete configuration that M will have at the moment of entry. This guessed configuration includes the internal state, the exact head position within the block, and the contents of the block’s cells. Because a block contains only L cells, the entire guessed description fits into O(L) = O(√f(n)) bits, which is the only memory S ever needs to retain.

After guessing, S verifies the guess by locally simulating M’s moves inside the block. For each step, S checks that the current state and the symbol under the head match the guessed description, updates the state, writes the new symbol, and moves the head. This verification proceeds deterministically and uses only the information stored for the current block. When the simulated head reaches the rightmost cell of the block, S compares the simulated configuration with the previously guessed configuration for the next block. If they agree, the simulation proceeds; otherwise the current nondeterministic branch is discarded.

The time analysis is straightforward. Each block requires at most L simulated steps, and there are O(√f(n)) blocks, yielding a total of O(L·√f(n)) = O(f(n)) steps. The space analysis shows that S never stores more than the current block’s description and a constant‑size index, so the space bound is O(L) = O(√f(n)).

The key insight is that nondeterminism allows the simulator to “guess” the future configuration at block boundaries, thereby avoiding the need to keep the entire tape history. This is a departure from classic deterministic simulations, where the whole tape must be retained or recomputed, leading to larger space requirements.

The authors also discuss the limitations of their technique. Extending the method to multi‑tape NTMs is non‑trivial because each tape would require its own set of guessed block boundaries and configurations, causing the required guess size to blow up beyond O(√f(n)). Moreover, the proof does not relativize: the construction cannot be carried out in the presence of an arbitrary oracle, since the oracle answers could affect the guessed configurations in ways that cannot be captured by a simple block‑wise guess.

In the broader context, this result complements known trade‑offs such as Savitch’s theorem, which shows that nondeterministic time T can be simulated deterministically in O(log² T) space but at the cost of squaring the time. Here, by staying within the nondeterministic model, the authors achieve a space reduction to the square root of the time without increasing the time bound, albeit only for single‑tape machines and for time functions at least quadratic.

The paper therefore establishes a new, stronger resource hierarchy: for NTMs with f(n) ≥ n², space is not a bottleneck as severe as previously thought; it can be compressed to O(√f(n)) while preserving the original time. This contributes to our understanding of how nondeterminism can be leveraged to balance computational resources and opens avenues for further research on whether similar reductions are possible for broader classes of machines or under relativized settings.