Carnot Information Engine (CIE)

• Carnot Information Engines are thermodynamic models that treat information as a state variable, equating Shannon entropy with Gibbs entropy in energy conversion processes.
• They span diverse implementations including optical pulse engines, minimal heat-engine models with finite channel capacity, and quantum measurement-feedback cycles.
• Key insights reveal that incorporating information as a thermodynamic resource modifies traditional Carnot limits, enabling enhanced work extraction and novel efficiency regimes.

Searching arXiv for the cited Carnot/information engine papers to ground the article in current arXiv records. Search query: all:("Informatics Carnot Machine" OR "Carnot's theorem and Szilard engine" OR "Minimal model of a heat engine: An information theory approach" OR "Efficiency at maximum power of a Carnot quantum information engine" OR "Information-Assisted Carnot Engine Surpasses Standard Thermodynamic Bounds") Carnot Information Machine, often rendered in the literature as a Carnot Information Engine (CIE), denotes a family of thermodynamic constructions in which information, Shannon entropy, and thermodynamic entropy are treated within a Carnot-style framework. In the arXiv literature, the term spans several closely related but non-identical models: an optical “Informatics Carnot Machine” in which a coded pulse train behaves as a reversible Carnot cycle; heat-engine models in which the irreversibility is assigned to finite communication-channel capacity; generalized Carnot and Szilárd formulations based on “available information”; inclusive Hamiltonian treatments with an explicit information reservoir; finite-time quantum information engines; and measurement-feedback cycles that modify the standard Carnot bookkeeping (0705.2535, Zhou et al., 2010, Shu et al., 2016, Deffner et al., 2013, Fadler et al., 2023, Xiao et al., 17 Jul 2025).

1. Terminological scope and core idea

The common premise of CIE formulations is that information is not merely an external description of a thermodynamic process, but a resource or state variable that enters the work–heat balance in a Carnot-like manner. In the 2007 “Informatics Carnot Machine,” this takes the form of an explicit identification between the Gibbs mixing entropy of a random pulse sequence and Shannon information. In the 2010 minimal model, the engine is organized as detector, communication channel, processor, and receiver, and the irreversible cost of operation is attributed to the finite channel capacity. In the 2016 “Carnot’s theorem and Szilárd engine,” the central object is “available information” associated with a temperature difference, and the generalized theorem is stated as “all the available information is 100% coded into work.” In Deffner and Jarzynski’s inclusive Hamiltonian approach, the defining addition is an information reservoir whose Shannon-entropy change appears alongside the usual $Q/T$ terms in generalized second-law statements. Later work extends the same line to finite-time quantum cycles and to information-assisted Carnot engines with explicit measurement–feedback operations (0705.2535, Zhou et al., 2010, Shu et al., 2016, Deffner et al., 2013, Fadler et al., 2023, Xiao et al., 17 Jul 2025).

A central unifying statement across these formulations is that information can be handled as a thermodynamic resource only when its generation, transmission, storage, or erasure is included in the entropy accounting. This is why some CIE papers emphasize strict Carnot-type bounds, while others report efficiencies above the standard Carnot limit: the difference lies in what counts as input resource and in whether the entropy change of memory, tape, or mutual information is included explicitly in the performance metric (Deffner et al., 2013, Rana et al., 2016, Xiao et al., 17 Jul 2025).

2. Optical “Informatics Carnot Machine”

The earliest explicit Carnot-style information construction in this set is the optical model of “Informatics Carnot Machine.” It starts from Planck’s law for a single mode of frequency $\nu$,

$n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$

and inverts it to associate an equivalent temperature to a coherent pulse with $n_i$ photons,

$T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$

With pulse energy $q_i=n_i h\nu$, the entropy assigned to that mode is

$S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$

In the classical limit $n_i\gg 1$, one obtains $S_i\to k_B$, whereas for $n_i\to 0$, $\nu$0. The paper therefore states: a single pulse with many photons carries one entropy unit $\nu$1, and an empty mode carries zero entropy (0705.2535).

For a random binary file of $\nu$2 modes, where “1” denotes a pulse and “0” a vacancy, the total entropy is the Gibbs mixing entropy

$\nu$3

If all $\nu$4 strings are equiprobable, then

$\nu$5

while the Shannon information in nats is

$\nu$6

Hence

$\nu$7

The paper’s conclusion is that, in this regime, the Gibbs entropy of the random pulse sequence is exactly the Shannon information up to the factor $\nu$8, and “entropy under certain conditions is information” (0705.2535).

The same work recasts information transmission and amplification in an optical fiber as a classic Carnot machine with four stages: adiabatic expansion as attenuation, isothermal compression as reading at $\nu$9, adiabatic compression as amplification, and isothermal expansion as writing at $n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$0. The cycle satisfies

$n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$1

Because the entropy changes of the two isothermal steps satisfy $n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$2, the paper interprets an ideal optical amplifier that faithfully preserves Shannon information while compensating loss as a reversible Carnot engine for information processing (0705.2535).

3. Information-theoretic heat-engine models

A more abstract CIE appears in “Minimal model of a heat engine: An information theory approach.” The setup has two reservoirs at $n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$3, a hot two-level system used as the source, and an engine at $n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$4 consisting of detector, communication channel, processor, and receiver. The detector measures the two-level system, the result is encoded into bits, transmitted through the channel, decoded, and used to configure the receiver so that the source releases energy into the engine. By construction, all irreversibility is attributed to the finite capacity of the communication link (Zhou et al., 2010).

In this model, the number of bits required per cycle is

$n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$5

where $n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$6 is the mean energy of the hot two-level system. If the channel capacity per pulse is $n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$7, with average signal energy $n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$8 and noise energy $n_i=\frac{1}{\exp(h\nu/(k_B T_i))-1},$9, then sending $n_i$0 bits requires $n_i$1 pulses and dissipates

$n_i$2

The net work per cycle is

$n_i$3

and the efficiency is

$n_i$4

For a classical Gaussian channel, $n_i$5; for a quantum bosonic wideband channel, the capacity takes the explicit form given in the paper (Zhou et al., 2010).

The reversible limit is the vanishing-signal limit $n_i$6. In that regime, the channel operates without information loss per bit, and the dissipated energy approaches

$n_i$7

Hence

$n_i$8

The same model also yields efficiency-at-maximum-power results. With power proportional to $n_i$9, the optimum for the classical Gaussian channel occurs at $T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$0, giving

$T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$1

and in linear response $T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$2. For the quantum wideband bosonic channel, the first-order result is again $T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$3. More generally, the paper emphasizes the universality of the $T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$4 coefficient in the linear-response expansion of efficiency at maximum power (Zhou et al., 2010).

This formulation places CIE squarely at the interface of thermodynamics, Landauer cost, and communication theory: the work deficit relative to the reversible limit is exactly the energetic penalty of finite-rate information transfer.

4. Available information, Szilárd equivalence, and generalized second-law structure

The 2016 paper “Carnot’s theorem and Szilárd engine” frames the Carnot engine and the Szilárd engine as physically identical at the level of information bookkeeping. It introduces available information about the temperature difference between two reservoirs and argues that both devices realize lossless conversion of that available information into work. In that formulation, the classical Carnot result and the Szilárd result coincide,

$T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$5

and the generalized Carnot theorem is expressed as: “all the available information is 100% coded into work” (Shu et al., 2016).

The paper’s broader significance is not a new cycle construction but a change of interpretation. Heat flow between $T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$6 and $T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$7 is treated as a carrier of information about thermal imbalance, and reversibility is rephrased as complete preservation of that resource. The same discussion also allows non-equilibrium extensions, in which the total available information is augmented by a Boltzmann–$T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$8-function difference, leading to

$T_i=\frac{h\nu/k_B}{\ln(1+1/n_i)}.$9

This suggests a generalized resource view in which temperature difference, non-equilibrium structure, and memory order enter a common information balance (Shu et al., 2016).

A more formal version of this perspective is given by Deffner and Jarzynski in “Information processing and the second law of thermodynamics: an inclusive, Hamiltonian approach.” Their framework models the device, two heat reservoirs, a work reservoir, and an information reservoir by a single time-independent Hamiltonian. In the steady cyclic regime, the first law reads

$q_i=n_i h\nu$0

and for degenerate informational states this reduces to

$q_i=n_i h\nu$1

The generalized second-law inequality is

$q_i=n_i h\nu$2

They then define an information-extended efficiency

$q_i=n_i h\nu$3

with $q_i=n_i h\nu$4, and show that

$q_i=n_i h\nu$5

Within this bookkeeping, extra work beyond the standard heat-only Carnot value is not a violation of Carnot’s theorem; it is work drawn partly from the entropy increase of the information reservoir (Deffner et al., 2013).

5. Information reservoirs, measurement–feedback loops, and “beyond Carnot” claims

A recurrent point of debate in CIE research is whether information-assisted engines can exceed the Carnot limit. The answer depends on the chosen efficiency definition and on whether information change is counted as a resource. The tape-based model of “A multipurpose information engine that can go beyond the Carnot limit” makes this explicit. The device is a three-state demon coupled simultaneously to a moving bit tape, a work source, and two heat baths. In steady state,

$q_i=n_i h\nu$6

where $q_i=n_i h\nu$7 is the Shannon-entropy change of the tape. When the device acts as an engine, its efficiency is

$q_i=n_i h\nu$8

and the paper identifies parameter regimes in which $q_i=n_i h\nu$9. Its interpretation is that the surplus work is fueled by increasing the Shannon entropy of the tape; the tape must be re-erased elsewhere at corresponding thermodynamic cost (Rana et al., 2016).

The 2025 “Information-Assisted Carnot Engine Surpasses Standard Thermodynamic Bounds” gives a more direct Carnot-cycle construction. A standard Carnot cycle is supplemented, after the hot isothermal expansion $S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$0, by a measurement–feedback loop: an imperfect classical measurement with error probability $S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$1, a conditional level flip that injects work $S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$2, followed by an adiabatic expansion and a cold isothermal compression. The total cycle is decomposed into a standard Carnot loop $S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$3 and an information-driven loop $S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$4 (Xiao et al., 17 Jul 2025).

In that formulation, the residual and consumed mutual information enter explicitly: $S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$5 The demon-driven cold compression yields

$S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$6

the net work of the demon loop is

$S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$7

the total work is

$S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$8

and the hot-end energy input is

$S_i=\frac{q_i}{T_i}=n_i k_B\ln(1+1/n_i).$9

The efficiency is defined by

$n_i\gg 1$0

From

$n_i\gg 1$1

the paper derives

$n_i\gg 1$2

and under the condition $n_i\gg 1$3 it finds

$n_i\gg 1$4

For certain $n_i\gg 1$5, the standard Carnot cycle has $n_i\gg 1$6, whereas the CIE has $n_i\gg 1$7. The paper explicitly demonstrates these features for a spin-$n_i\gg 1$8 working fluid and proposes an implementation based on a trapped $n_i\gg 1$9 ion (Xiao et al., 17 Jul 2025).

Taken together, these papers indicate that “beyond Carnot” claims in CIE research are not uniform. In the inclusive Hamiltonian formulation, once the information term is bundled into the input resource, the modified Carnot bound remains intact. In the tape and feedback formulations, the reported excess over the standard Carnot efficiency is relative to the heat-only bound, with information entropy increase or mutual information serving as the additional resource (Deffner et al., 2013, Rana et al., 2016, Xiao et al., 17 Jul 2025).

6. Finite-time quantum generalizations, experimental proposals, and communication analogues

The finite-time quantum extension is developed by Fadler et al. in “Efficiency at maximum power of a Carnot quantum information engine.” The working medium has Hamiltonian

$S_i\to k_B$0

and one cycle consists of a reversible generalized energy measurement plus feedback at fixed $S_i\to k_B$1, an adiabatic expansion, a hot isotherm, and an adiabatic compression. In the low-dissipation regime, the hot-isotherm heat is

$S_i\to k_B$2

the total work becomes

$S_i\to k_B$3

and the information-to-work efficiency is

$S_i\to k_B$4

With total time $S_i\to k_B$5, the power is

$S_i\to k_B$6

Maximization gives

$S_i\to k_B$7

and the efficiency at maximum power is

$S_i\to k_B$8

For a qubit with weak energy measurements, the paper derives explicit formulas for $S_i\to k_B$9, $n_i\to 0$0, and $n_i\to 0$1, thereby extending finite-time thermodynamics to a quantum information-engine setting (Fadler et al., 2023).

The experimental orientation of CIE research is especially explicit in the 2025 trapped-ion proposal. There the system Hamiltonian is

$n_i\to 0$2

with a Lamb–Dicke interaction

$n_i\to 0$3

Adiabatic strokes are implemented by slow tuning of $n_i\to 0$4; isothermal strokes are approximated by many adiabatic–isochoric segments; and the measurement–feedback step is realized by fluorescence-based state discrimination with error $n_i\to 0$5, followed by conditional laser operations. Repeated cycles permit reconstruction of $n_i\to 0$6, $n_i\to 0$7, $n_i\to 0$8, $n_i\to 0$9, and $\nu$00 (Xiao et al., 17 Jul 2025).

A separate extension maps the CIE idea into communication theory. In “Carnot machine-based massive MIMO communication capacity modeling and performance analysis,” information is treated as carrying energy and entropy, with

$\nu$01

for a binary block of $\nu$02 bits, and

$\nu$03

The Carnot-style fraction of recoverable information is written as

$\nu$04

For massive MIMO, the generalized thermodynamic channel capacity is

$\nu$05

and the paper reports that simulation results verify the proposed channel capacity is coincident with the classical channel capacity (Ruan et al., 2022).

These later developments show that CIE is no longer confined to Maxwell-demon thought experiments. It has become a technical framework for finite-time thermodynamics, quantum measurement-and-feedback cycles, trapped-ion realizations, and even thermodynamic reinterpretations of communication capacity. Across these variants, the invariant theme is that Carnot-type performance limits must be reformulated once information-bearing degrees of freedom are treated as explicit thermodynamic resources (Fadler et al., 2023, Xiao et al., 17 Jul 2025, Ruan et al., 2022).