Landauer Cost, Information Engines, and Field Reservoirs

Landauer’s principle is an entropy, heat, and work-accounting statement for a physical memory coupled to a specified thermal reservoir. It does not say that every measurement costs $k_BT\ln2$, and it does not attach a temperature-only price to an arbitrary nonthermal environment. Measurement and erasure costs are distinct operational questions Sagawa and Ueda 2009. The familiar one-bit value follows only after the logical states, reset error, bath state, endpoint Hamiltonians, and complete operating cycle have been declared.

Required background. Passivity, work, and information supplies the sign convention for cyclic work and the equilibrium reference state.

Helpful background. Complete passivity, KMS structure, and resource conversion explains why an equilibrium bath is different from a reservoir that also supplies nonequilibrium free energy.

Erasure is a physical state transformation

Let $M$ be the information-bearing memory and $R$ the reservoir. Use natural logarithms, so entropies are dimensionless, and let

$$\tau_R=\frac{e^{-\beta H_R}}{Z_R}, \qquad \beta=\frac{1}{k_BT}.$$

Assume initially that $M$ and $R$ are uncorrelated and that their joint evolution is unitary:

$$\rho_{MR}=\rho_M\otimes\tau_R, \qquad \rho'_{MR}=U\rho_{MR}U^\dagger.$$

An erasure channel reduces the memory entropy. Define that reduction and the heat delivered to the reservoir by

$$\Delta S_M=S(\rho_M)-S(\rho'_M), \qquad Q_R=\operatorname{tr}\!\left[H_R(\rho'_R-\tau_R)\right].$$

Thus $Q_R>0$ means that the reservoir gains energy. These four assumptions—identified subsystems, a Gibbs reservoir, an initially product state, and unitary evolution of everything retained—are the minimal setting used by Reeb and Wolf 2014, § 2.1 and Theorem 3.

The exact balance is

$$\beta Q_R =\Delta S_M+I(M{:}R)_{\rho'} +D(\rho'_R\|\tau_R).$$

It follows in two short steps. First, global entropy conservation gives

$$S(\rho'_R)-S(\tau_R) =\Delta S_M+I(M{:}R)_{\rho'}.$$

Second, expanding the relative entropy to the Gibbs state gives

$$D(\rho'_R\|\tau_R) =\beta Q_R-\big[S(\rho'_R)-S(\tau_R)\big].$$

Both extra terms are nonnegative, so $\beta Q_R\ge\Delta S_M$. The equality displays what the shorter inequality hides: final memory–bath correlations and displacement of a finite bath from its initial Gibbs state both raise the heat above the entropy-only term. For a nontrivial entropy decrease with a finite-dimensional reservoir at finite positive temperature, exact equality with the bare Landauer bound cannot be attained; sequences of increasingly large reservoirs can approach it Reeb and Wolf 2014, §§ 3.2 and 6.

Landauer’s original discussion concerned logically irreversible operations in physical bistable devices, not a universal energy charge for observing a system Landauer 1961, pp. 183–188.

Heat is not automatically work

To discuss work, one must also specify a memory Hamiltonian $H_M(t)$ and the controller that changes it. Suppose the interaction vanishes at the accounting endpoints. With work $W_{\rm on}$ positive when supplied by the controller, the first law is

$$W_{\rm on}=\Delta E_M+Q_R, \qquad \Delta E_M= \operatorname{tr}(H_M^f\rho'_M) -\operatorname{tr}(H_M^i\rho_M).$$

For a cyclic memory Hamiltonian with degenerate logical states, $\Delta E_M=0$, so the ideal reset work and heat agree. For an asymmetric memory they need not. Under the usual isothermal control assumptions, the work inequality is instead

$$W_{\rm on}\ge F_\beta(\rho'_M,H_M^f) -F_\beta(\rho_M,H_M^i),$$

where $F_\beta(\rho,H)=\operatorname{tr}(H\rho)-\beta^{-1}S(\rho)$. The $k_BT\ln2$ value is recovered for a uniform bit, degenerate endpoints, and a pure reset target.

A complete information-engine cycle needs a wider account. If a measurement creates a record $X$ and conditional feedback extracts work, then

$$W_{\rm net}^{\rm out} =W_{\rm feedback}^{\rm out} -W_{\rm measure}^{\rm on} -W_{\rm reset}^{\rm on} -W_{\rm control}^{\rm on}.$$

Mutual information can increase the work available during the feedback stroke; Sagawa and Ueda 2008 derive the corresponding feedback inequality. Closing the cycle requires restoring the memory, controller, and reservoir resources. Reporting only the favorable feedback stroke is not a cyclic engine efficiency.

Reproducible one-bit benchmark

Take a classical bit encoded in a two-level memory. Its initial Hamiltonian is degenerate and its logical probabilities are $(1/2,1/2)$. The target reset has error $\epsilon$, so its final distribution is $(1-\epsilon,\epsilon)$ and

$$h_2(\epsilon) =-\epsilon\ln\epsilon-(1-\epsilon)\ln(1-\epsilon).$$

An ideal quasistatic protocol raises the energy $\Delta$ of state 1 while the memory remains equilibrated with a large bath. Stop at

$$\beta\Delta_f=\ln\frac{1-\epsilon}{\epsilon},$$

disconnect the bath, and lower the occupied level back to zero without changing its population. Because the endpoint memory Hamiltonians are again degenerate,

$$\beta Q_{\rm qs}=\beta W_{\rm qs} =\ln2-h_2(\epsilon).$$

For comparison, raise the level instantaneously from $0$ to $\Delta_f$, allow complete thermalization at fixed gap, and then lower it instantaneously after disconnecting the bath. During thermalization, the excited-state population changes from $1/2$ to $\epsilon$, so

$$\beta Q_{\rm sudden} =\left(\frac12-\epsilon\right) \ln\frac{1-\epsilon}{\epsilon}.$$

At $\epsilon=0.01$,

$$\begin{aligned} h_2(0.01)&=0.05600153,\\ \beta Q_{\rm qs}&=0.63714565,\\ \beta Q_{\rm sudden}&=2.25160873. \end{aligned}$$

At $T=300\ \mathrm{K}$, using the exact SI value $k_B=1.380649\times10^{-23}\ \mathrm{J\,K^{-1}}$, these are

$$Q_{\rm qs}=2.63902\times10^{-21}\ \mathrm J, \qquad Q_{\rm sudden}=9.32604\times10^{-21}\ \mathrm J.$$

The two protocols implement the same final logical error. Their difference is finite-time dissipation, not a change in the information erased.

A finite bosonic-reservoir implementation

For a field-like reservoir, discretize a finite spatial volume and use

$$H_R=\sum_{j=1}^{N}\omega_j a_j^\dagger a_j, \qquad Q_R=\sum_{j=1}^{N}\omega_j \big(\langle n_j\rangle_f-\langle n_j\rangle_i\big).$$

To make exact diagonalization finite, truncate each oscillator at occupation $n_{\max}$ and prepare the normalized Gibbs state of that truncated Hamiltonian. Couple the memory locally through a declared switching function and evolve the combined memory–modes system unitarily. A reproducible calculation records, for every run,

$$(N,n_{\max},\Lambda,\tau,\epsilon, Q_R,\Delta E_M,I_f,D_f,W_C),$$

where $\Lambda=\max_j\omega_j$, $I_f=I(M{:}R)_{\rho'}$, and $D_f=D(\rho'_R\|\tau_R)$. The exact identity above should close numerically before any continuum interpretation is attempted.

Use the quasistatic and sudden formulas as large-bath reference values, not as automatic predictions for finite $N$. Increase $N$, $n_{\max}$, and $\Lambda$ independently, and lengthen the slow ramp at fixed reset error. If a reduced master equation is used instead of joint unitary evolution, its omitted environment must be included before calling the energy account complete.

Assumptions that change the conclusion

A nonthermal reservoir. If $R$ does not begin in $e^{-\beta H_R}/Z$, there may be no unique temperature for a $k_BT$ formula. Apparent performance below the thermal benchmark can consume coherence, population inversion, squeezing, chemical work, or another nonequilibrium free-energy resource.

Initial memory–reservoir correlations. The product-state step in the derivation fails. Correlation can serve as an entropy sink or work resource, so the simple bound must be replaced by an account that includes its change.

An open cycle. A feedback device that retains its record has not returned to its initial state. Erasure may be postponed or moved to another subsystem, but it has not disappeared from a repeatable cycle.

Finite time. Extra dissipation depends on coupling strengths, control restrictions, and relaxation dynamics. There is no universal finite-time coefficient that depends only on $T$ and the number of erased bits.

Infinite field reservoirs. A formal KMS state is not a finite tank. Local coupling, energy flux, infrared and ultraviolet control, and the operational definition of heat must remain explicit when passing from a finite mode model to QFT.

The chapter orientation, claim-validity table, and failure controls show where Landauer accounting sits relative to passivity and stress-energy bounds.

Common pitfalls

Charging $k_BT\ln2$ for measurement. The standard value concerns entropy reduction of a uniform bit under thermal reset assumptions. Measurement can be thermodynamically reversible in one model and costly in another; the record’s eventual reuse is the relevant cycle question.

Calling $Q_R$ the controller work. Heat into the reservoir and work supplied by the controller coincide only in special endpoint conventions such as the degenerate cyclic bit.

Treating the bath as unchanged. A finite bath stores correlations and moves away from its initial Gibbs state. Those effects are the two explicit nonnegative corrections in the exact identity.

Exercises

1. Derive the equality form of Landauer’s principle

Starting from the product state and global unitary above, derive the exact relation for $\beta Q_R$.

Solution

Unitary evolution preserves joint entropy, and the initial state is a product:

$$S(\rho'_{MR})=S(\rho_M)+S(\tau_R).$$

Using $I(M{:}R)'=S(\rho'_M)+S(\rho'_R)-S(\rho'_{MR})$ gives

$$S(\rho'_R)-S(\tau_R)=\Delta S_M+I(M{:}R)'.$$

Because $\ln\tau_R=-\beta H_R-\ln Z_R$,

$$D(\rho'_R\|\tau_R) =\beta Q_R-S(\rho'_R)+S(\tau_R).$$

Substitution yields $\beta Q_R=\Delta S_M+I'+D$. Positivity of $I'$ and $D$ gives the inequality.

2. Compare slow and sudden reset

At reset error $\epsilon=0.05$, compute $\beta\Delta_f$, $\beta Q_{\rm qs}$, and $\beta Q_{\rm sudden}$.

Solution

The final gap is

$$\beta\Delta_f=\ln(19)=2.944439.$$

The binary entropy is $h_2(0.05)=0.198515$. Therefore

$$\beta Q_{\rm qs}=\ln2-h_2(0.05)=0.494632,$$

whereas

$$\beta Q_{\rm sudden}=0.45\ln(19)=1.324998.$$

The sudden protocol dissipates about $2.68$ times as much heat for the same logical output.

3. Locate the missing resource

A squeezed bosonic reservoir resets the bit with $Q_R<k_BT[\ln2-h_2(\epsilon)]$, where $T$ is inferred from its mean energy. Is Landauer’s principle violated?

Solution

No temperature-only conclusion follows because the initial reservoir is not a Gibbs state. Its squeezing is a nonequilibrium resource, and a mean-energy-matched temperature does not make $\ln\rho_R$ proportional to $-H_R$. One must account for the reservoir’s nonequilibrium free-energy change and for the work required to prepare or restore the squeezed state. The Gibbs-state derivation cannot be applied with an effective temperature substituted by hand.

4. Close a feedback cycle

A feedback stroke extracts $0.8k_BT$ while measurement costs $0.1k_BT$ and the uniform one-bit record is left unchanged. What is the strongest net-work statement that can be made?

Solution

Only the open-cycle balance $0.7k_BT$ has been reported. Reusing the controller requires restoring its record. Under the ideal symmetric thermal assumptions, that reset requires at least $k_BT\ln2\approx0.693k_BT$ of heat/work, before any finite-time or controller overhead. The corresponding ideal closed-cycle output is therefore at most about $0.007k_BT$; a finite implementation will generally do worse.

References

• Landauer, Rolf. “Irreversibility and Heat Generation in the Computing Process.” IBM Journal of Research and Development 5, no. 3 (1961): 183–191. DOI.
• Reeb, David, and Michael M. Wolf. “An Improved Landauer Principle with Finite-Size Corrections.” New Journal of Physics 16 (2014): 103011. DOI.
• Sagawa, Takahiro, and Masahito Ueda. “Second Law of Thermodynamics with Discrete Quantum Feedback Control.” Physical Review Letters 100 (2008): 080403. DOI.
• Sagawa, Takahiro, and Masahito Ueda. “Minimal Energy Cost for Thermodynamic Information Processing: Measurement and Information Erasure.” Physical Review Letters 102 (2009): 250602; erratum 106 (2011): 189901. DOI.