Answers are grounded in this paper and cite their sources.
The Luminous Turing Machine (LTM) meters computation by proper time τ along null worldlines, where τ = 0. Because that meter is constant, its classes collapse completely: we prove P_L = NP_L = EXP_L = R, the class of all decidable languages. If setup must run in polynomial time, the classes equal PSPACE instead. The collapse is genuine but says nothing about NP; it describes a meter that cannot tell one step from 2^(2^n) steps. We then price each resource the null meter leaves out. Charging setup steps, duration, energy width, repetitions, and the implementation of the Hamiltonian gives laboratory classes P_lab = BQP and NP_lab = QCMA. Leaving one of these free gives PSPACE (duration or width), PP (repetitions), or a class containing UP (implementation): a dense Hamiltonian with efficiently computable entries finds a unique witness in constant time and constant energy width. The thesis is then made physical and observer-relative. Hardware that holds data longer than its light-crossing time has a timelike worldline, so null propagation does not remove the cost of implementing and reading a computation. No inertial observer waits less than the computer's own running time T, and the lab classes are the same in every inertial frame. An observer who waits only τ_O gets speedup Γ = T/τ_O and must reach kinetic energy of at least (Γ − 1)mc²; the bound is sharp, and a rocket-borne observer needs a mass ratio of at least (2Γ − 1)², or Γ²(Γ − 1)² if it moves along a line. A polynomial wait bought with polynomial kinetic energy still decides only BQP; bought with exponential energy it decides exactly PSPACE. Under three further assumptions, the Margolus–Levitin bound gives T ≤ τ_O + 2Kτ_O²/(πℏ) for peak kinetic energy K, whatever the observer's mass. Near a black hole the price moves from kinetic energy to thrust: a hovering observer needs a mass ratio exponential in T divided by the light-crossing time of the horizon. τ = 0 is free for the light that carries signals, but an observer of fixed mass approaches a zero wait only as its kinetic energy grows without bound. Causal geometry separately forces median memory latency to grow at least as N^(1/3) for N bits stored at bounded density. None of these results bears on classical P vs NP, which has the same answer in every frame.
No takes yet. Share an insight, caveat, or question.
mFLO137 (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: