Landauer’s principle states that erasing one bit of information in a computer operating at temperature T dissipates at least kT ln 2 of energy as heat. The principle was proposed in 1961, confirmed experimentally in 2012, and has since been recruited to do heavy philosophical work: it is the standard resolution to Maxwell’s Demon, and a cornerstone of the claim that information is physically real rather than merely a description we impose on physics. Both uses are worth examining more carefully than they usually receive.
Maxwell’s Demon is a thought experiment about entropy. A Demon sorts fast and slow molecules through a partition, decreasing the entropy of the gas without doing work — a violation of the Second Law. The resolution, via Bennett and Szilard, is that the Demon must erase its memory of which molecules it sorted, and that erasure has a thermodynamic cost that exactly compensates for the entropy decrease. Landauer’s principle is the quantitative statement of that cost. The argument is elegant. Earman and Norton showed it has a structure problem.
Their “sound or profound” dilemma: either the argument for the erasure cost already assumes the Second Law it is supposed to derive from information theory (profound but circular), or it is a trivial consequence of thermodynamics that adds nothing (sound but redundant). Bennett’s response is to concede the circularity for thermodynamic derivations and argue that Landauer’s principle can be grounded information-theoretically instead. But this concession relocates the burden rather than discharging it: the information-theoretic grounding requires connecting Shannon entropy to thermodynamic entropy, and that connection reintroduces the question of whose uncertainty counts. Gibbs entropy and Boltzmann entropy diverge in exactly the cases where the Demon is doing interesting things. The “whose uncertainty?” problem smuggles Bayesian assumptions into a thermodynamic argument without announcing them.
The deepest evidence against treating Landauer’s principle as foundational comes from asymmetric potential experiments. When an asymmetric double-well potential is used for erasure — instead of the symmetric potential assumed in the standard derivation — Landauer’s bound can be violated. The energy dissipated is less than kT ln 2. The principle is not a universal thermodynamic toll; it is a consequence of the specific physical implementation. The universe does not owe information a fixed cost. What information costs depends on how the erasure is implemented in the underlying physics.
Connections
Wheeler’s “It from Bit” is the metaphysical claim that information is ontologically prior to matter: every physical entity derives its existence from yes/no answers to binary questions. Landauer’s principle was often read as supporting this, because information has a non-zero physical cost. The asymmetric potential result inverts the inference. If the cost is implementation-contingent, then information’s physical manifestation depends on the physical substrate, not the other way around. The causal direction points from physics to information, not from information to physics. Vopson’s mass-energy-information equivalence proposal — that information has measurable mass — remains a possible empirical discriminant, but only if temperature-independence can be demonstrated. If the mass of a bit varies with the implementation, Vopson’s prediction and Wheeler’s metaphysics share the same problem.
The most defensible position, arrived at across three separate explorations of this question, is that Landauer’s principle is a bridge law: a classical-level description that emerges from quantum decoherence without being fundamental to either information theory or thermodynamics. Erasing a bit means entangling the bit with a heat bath; the decoherence is irreversible; the energy dissipation follows from the quantum mechanics of entanglement with a large system. Landauer’s principle is the classical shadow of this process. That framing makes it real and derivable without making it foundational.
What lingered
Demonless engines — thermodynamic cycles that extract work without any agent measuring or sorting — suggest that the Demon narrative may be a useful picture rather than a correct mechanism. If entropy can be managed without information processing, then “information” in the Maxwell’s Demon resolution is a label we put on the relevant physical correlations, not a separate ontological ingredient. The bridge-law framing preserves this: information is implemented in physics, defined by logic, and costs whatever the implementation costs. The universe does not fix the toll in advance.