Landauer's Principle: Energy Cost of Bit Erasure

• Landauer's Principle is a foundational concept stating that erasing a bit of information increases environmental entropy by at least kBT ln2.
• It links logical irreversibility with physical thermodynamics, impacting reversible computing, nanoscale devices, and quantum information processes.
• Experiments using nanomagnets and ultracold atoms have validated the principle by demonstrating energy dissipation near the predicted Landauer limit.

Landauer's principle states that any logically irreversible operation—most fundamentally, the erasure of one bit of information—must be accompanied by a corresponding increase in thermodynamic entropy of the environment. In its canonical form, resetting a memory element from an unknown state to a definite state while it is in contact with a heat bath at temperature $T$ requires dissipation of at least $k_B T \ln 2$; at room temperature, $k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$, or $2.8\,\mathrm{zJ}$. The principle is widely used to set boundary limits for information-processing devices, to relate logical irreversibility to physical irreversibility, and to organize modern discussions of reversible computing, open-system thermodynamics, and quantum information, while remaining a locus of continuing conceptual debate (Hong et al., 2014, Frank, 2019).

1. Classical statement and historical role

Landauer’s original 1961 claim is that logical irreversibility is associated with physical irreversibility. The paradigmatic example is bit erasure: a many-to-one map such as $\{0,1\}\to\{0\}$ reduces the information-bearing uncertainty of the memory by

$\Delta S_{\mathrm{info}} = k_B \ln 2.$

The second law then requires an equal or greater entropy increase in the environment, so that

$$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$

Equivalently,

$E_{\min} = k_B T \ln 2.$

For a qubit or classical bit initially in an unbiased distribution and reset to a definite state, the entropy drop is maximal and the bound takes its familiar one-bit form (Hong et al., 2014).

The principle entered the foundations of the thermodynamics of computation as a statement about the physical implementation of logically irreversible maps, not merely about abstract symbol manipulation. In the standard presentation, bijective microphysics forces any decrease in computational entropy to be compensated by an increase in non-computational entropy. Frank formulates this in terms of a coarse-grained computational state $C$ and the full physical microstate $\Phi$, with

$k_B T \ln 2$0

so that a drop in $k_B T \ln 2$1 under logically irreversible computation must reappear in $k_B T \ln 2$2 if total physical entropy does not decrease (Frank, 2019).

The historical reception was not uniform. The details provided for the nanomagnet experiment note a vigorous debate in the 1980s, involving Porod et al., Bennett, Toffoli, and Landauer’s own reply, over whether “information” itself carries a thermodynamic cost or whether erasure can be arranged reversibly. That debate continues in modernized form in both quantum thermodynamics and philosophy of physics (Hong et al., 2014).

2. Entropy accounting, correlations, and differential formulations

A central refinement of Landauer’s principle is that it concerns the fate of correlations, not arbitrary entropy transfer. Frank emphasizes that the principle is frequently misread as a generic prohibition on moving entropy between computational and non-computational degrees of freedom. In his formulation, reversible entropy exchange is possible if no correlations are destroyed. The genuinely irreversible step is the ejection of correlated information from a controlled digital form into an uncontrolled thermal environment, where subsequent thermalization makes those correlations inaccessible. For a bit $k_B T \ln 2$3 correlated with another subsystem $k_B T \ln 2$4, the relevant quantity is the mutual information $k_B T \ln 2$5; erasing $k_B T \ln 2$6 obliviously eliminates these correlations and forces an increase in independent entropy equal to $k_B T \ln 2$7 (Frank, 2019).

This correlation-centered view appears in a local-in-time version of the principle. Granger and Kantz model a memory as an overdamped Brownian particle evolving under a time-dependent potential and derive a differential Landauer bound that links the instantaneous irreversible entropy-production rate to the rate at which information is erased. With $k_B T \ln 2$8 the mutual information between the memory microstate and the stored symbol, they obtain

$k_B T \ln 2$9

After time integration, this reproduces the standard bound $k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$0 for erasure of one unbiased bit, but it does so as a continuous constraint on the trajectory of the process rather than only as an endpoint inequality (Granger et al., 2013).

A further point of clarification concerns fluctuations. Myrvold’s statistical-mechanical treatment does not neglect microscale fluctuations and does not assume ideal thermodynamically reversible processes. For a manipulation that maps each of an $k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$1-element distinguishable set $k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$2 to a common state $k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$3, he derives the exponential inequality

$k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$4

which implies the mean-dissipation bound

$k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$5

In this formulation Landauer’s bound is an average statement about dissipation, robust to fluctuations rather than invalidated by them (Myrvold, 2020).

3. Experimental realizations

A major milestone was the experimental verification of Landauer’s principle in nanoscale magnetic memory. Hong et al. fabricated arrays of single-domain nanomagnets in amorphous Permalloy thin film on silicon. Each island was an ellipse roughly $k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$6 in plan view, about $k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$7 thick, and spaced $k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$8 apart to suppress dipolar coupling. The long axis was the easy axis, with two stable states $k_B T \ln 2 \simeq 2.8\times 10^{-21}\,\mathrm{J}$9, and the short axis was the hard axis. Erasure was implemented as a four-stage quasi-static sequence after Bennett’s proposal: ramp $2.8\,\mathrm{zJ}$0 along the hard axis to flatten the anisotropy barrier, ramp $2.8\,\mathrm{zJ}$1 along the easy axis to drive the magnet into the “1” state, reduce $2.8\,\mathrm{zJ}$2 to re-establish the barrier, and finally return $2.8\,\mathrm{zJ}$3 to zero. The protocol used $2.8\,\mathrm{zJ}$4 and $2.8\,\mathrm{zJ}$5, with ramp times of tens of seconds to approach the adiabatic limit (Hong et al., 2014).

The measurement combined lateral-geometry MOKE, VSM calibration, SEM, and AFM. Magnetization loops were recorded on both axes, and the dissipated energy per erasure cycle was extracted from

$2.8\,\mathrm{zJ}$6

At $2.8\,\mathrm{zJ}$7, five independent erasure runs yielded

$2.8\,\mathrm{zJ}$8

corresponding to $2.8\,\mathrm{zJ}$9. A separate temperature-dependent study over $\{0,1\}\to\{0\}$0–$\{0,1\}\to\{0\}$1 gave

$\{0,1\}\to\{0\}$2

or $\{0,1\}\to\{0\}$3 at $\{0,1\}\to\{0\}$4. In both cases the dissipation was reported as consistent with the Landauer limit $\{0,1\}\to\{0\}$5, with the small excess attributed mainly to slight lithographic misalignment of individual islands (Hong et al., 2014).

Experimental work has also moved into explicitly many-body settings. An ultracold-atom quantum field simulator based on two tunnel-coupled one-dimensional Bose gases was used to probe a generalized Landauer relation after a global mass quench from a massive to a massless Klein–Gordon model. Through dynamical tomographic reconstruction of the covariance matrix, the experiment tracked the temporal evolution of system entropy, environment energy change, mutual information, and relative entropy, and verified the equality

$\{0,1\}\to\{0\}$6

for various system–environment bipartitions of the composite many-body state (Aimet et al., 2024).

Frank’s review situates the nanomagnet result within a broader experimental landscape that includes colloidal double-well traps and quantum single-atom implementations, both of which approach $\{0,1\}\to\{0\}$7 in the adiabatic limit. This suggests that the principle is not restricted to a single physical substrate, even though the nanomagnet experiment was presented as the first direct verification in a prototypical digital memory element (Frank, 2019).

4. Quantum statistical mechanics, open systems, and measurement

In quantum statistical mechanics, Landauer’s principle is commonly expressed as an entropy-balance identity. For a finite-level quantum system $\{0,1\}\to\{0\}$8 interacting with a thermal reservoir $\{0,1\}\to\{0\}$9, Jakšić and Pillet write

$\Delta S_{\mathrm{info}} = k_B \ln 2.$0

where $\Delta S_{\mathrm{info}} = k_B \ln 2.$1 is the entropy decrease of the system, $\Delta S_{\mathrm{info}} = k_B \ln 2.$2 is the reservoir heat uptake, and $\Delta S_{\mathrm{info}} = k_B \ln 2.$3 is the entropy production. The inequality

$\Delta S_{\mathrm{info}} = k_B \ln 2.$4

then follows immediately, and for a qubit with maximal entropy reduction one recovers $\Delta S_{\mathrm{info}} = k_B \ln 2.$5. Their analysis uses a $\Delta S_{\mathrm{info}} = k_B \ln 2.$6-dynamical description of an infinitely extended thermal reservoir and proves, under an ergodicity assumption and adiabatic switching of the interaction, that Landauer’s bound saturates in the adiabatic limit (Jaksic et al., 2014).

Time-resolved open-system versions of the principle are obtained in collisional models. In a multipartite quantum system coupled sequentially to thermal subenvironments through energy-conserving interactions, the dissipated heat rate and entropy rate satisfy

$\Delta S_{\mathrm{info}} = k_B \ln 2.$7

For bipartite systems, this can be refined so that the rate of mutual-information creation explicitly enters the bound. In the formulations summarized by Pezzutto et al., correlation growth tightens the erasure bound, while violations of a local subsystem bound can signal non-Markovian memory effects in reduced dynamics (Lorenzo et al., 2015).

Quantum measurement provides a distinct realization of logical irreversibility. In a demonless quantum Szilard engine, partition insertion creates a superposition over left and right half-boxes rather than immediate localization. Work extraction requires a projective measurement that localizes the particle. That localization is logically irreversible because the post-measurement outcome does not determine the pre-measurement superposition, and the model assigns to the measurement step

$\Delta S_{\mathrm{info}} = k_B \ln 2.$8

The measurement-induced heat dissipation saturates Landauer’s bound and supplies the irreversible step needed to preserve the second law even without an explicit Maxwell-demon memory register (Aydin et al., 2019).

Repeated-interaction systems supply another non-equilibrium quantum setting. There the environment is a chain of probes, and one has a discrete one-step balance

$\Delta S_{\mathrm{info}} = k_B \ln 2.$9

In the adiabatic regime, asymptotic saturation of the total Landauer bound is not generic; the cited results characterize saturation by a detailed-balance condition on the repeated interaction system. This contrasts with continuous-time coupling to a single thermal reservoir, where adiabatic saturation is much more generic under ergodicity assumptions (Hanson et al., 2015).

5. Mathematical and conceptual generalizations

One line of generalization treats Landauer’s principle as a structural statement about entropy-ordered systems. Kycia defines entropy-equipped posets $$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$0 and $$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$1 and a Galois connection

$$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$2

satisfying

$$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$3

Within this framework, a logically irreversible map on $$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$4 necessarily induces an irreversible, entropy-increasing map on $$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$5, and the classical inequality $$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$6 appears as a corollary once the second system is interpreted as a thermal implementation of the first (Kycia, 2018).

A more technical extension uses operator algebras, quantum channels, and Jones index. Longo considers a normal, unital, completely positive map $$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$7 with finite Jones index in a possibly infinite quantum system. The incremental free energy is expressed through the quantum dimension

$$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$8

and the main theorem gives

$$\Delta Q_{\min} = T\,\Delta S_{\mathrm{env}} \ge T\,\Delta S_{\mathrm{info}} = k_B T \ln 2.$$9

Because the Jones index is quantized, this yields a lower bound

$E_{\min} = k_B T \ln 2.$0

for irreversible quantum channels in the general operator-algebraic setting, while in the finite-dimensional type-I case one recovers the standard Landauer form

$E_{\min} = k_B T \ln 2.$1

This places Landauer-type costs within a modular-theoretic and index-theoretic framework rather than solely within classical thermodynamic intuition (Longo, 2017).

These abstractions do not replace the usual thermodynamic reading; rather, they formalize the idea that an implementation map between ordered entropy structures constrains what counts as reversible at different descriptive levels. A plausible implication is that Landauer-type bounds persist whenever an implementation relation preserves an entropy order but not full invertibility, even if the relevant “entropy” is not always ordinary heat entropy in the narrow sense (Kycia, 2018, Longo, 2017).

6. Non-equilibrium extensions, refinements, and controversy

The canonical $E_{\min} = k_B T \ln 2.$2 cost assumes a thermal reservoir and, in many textbook derivations, a specific relation between logical states and thermodynamic macrostates. Several recent lines of work therefore treat the standard formula as a special case rather than a universal end-point. In squeezed thermal memories, for example, the entropy cost of erasure remains $E_{\min} = k_B T \ln 2.$3, but the minimum work for a one-bit erasure protocol can be reduced to

$E_{\min} = k_B T \ln 2.$4

because squeezing redistributes fluctuations between quadratures. The cited analysis stresses that the environment is then out of equilibrium; the standard Landauer energy bound is lowered, but the entropy bookkeeping remains consistent because the bath is characterized by an effective quadrature temperature rather than a conventional equilibrium temperature (Klaers, 2018).

A later development restores a Landauer inequality for reservoirs that are unitary transforms of thermal states. If the reservoir is prepared as

$E_{\min} = k_B T \ln 2.$5

one defines an effective Hamiltonian

$E_{\min} = k_B T \ln 2.$6

and obtains

$E_{\min} = k_B T \ln 2.$7

In this formulation, apparent violations for squeezed thermal states arise from measuring energy against the bare Hamiltonian $E_{\min} = k_B T \ln 2.$8 instead of the dressed generator $E_{\min} = k_B T \ln 2.$9 (Xu, 4 Feb 2026).

Low-temperature behavior motivates another refinement. Timpanaro et al. show that the ordinary bound becomes trivial as $C$0, but a tighter inequality survives if one characterizes the environment through its equilibrium heat capacity $C$1. Their general result is

$C$2

which reduces to the standard Landauer form at high temperature and remains non-trivial at zero temperature. For a one-dimensional waveguide with $C$3, they obtain

$C$4

so that the heat cost remains strictly positive when $C$5 even as $C$6 (Timpanaro et al., 2019).

Another refinement concerns conditional entropy. Chiuchiu et al. argue that even in bistable symmetric systems the Gibbs entropy need not coincide with the coarse-grained logical entropy, because microstate structure contributes a conditional entropy term:

$C$7

Accordingly,

$C$8

for a standard reset. They decompose $C$9 into an exchange-error term, a peak-shape term, and an overlap correction, and show that low barriers with appreciable overlap can raise the true minimum heat above the naive $\Phi$0 estimate (Chiuchiú et al., 2014).

The modern controversy turns on whether Landauer’s result should be regarded as a general physical principle or as a bound that depends on additional modeling assumptions. Lairez argues that the usual derivation conflates logical and thermodynamic irreversibility and introduces two extra constraints: a one-to-one identification of logical and thermodynamic states, and a unique erase procedure independent of the input. On that basis, the paper concludes that erasure can be thermodynamically reversible or quasistatic if those constraints are relaxed. This position stands in direct tension with approaches that regard the principle as a general consequence of entropy production, correlation loss, or adjoint structure. The present state of the subject is therefore not one of terminological consensus, but of increasingly explicit separation between equilibrium thermal reservoirs, non-equilibrium reservoirs, logically irreversible maps, and the physical paths by which they are implemented (Lairez, 2024, Frank, 2019).