aiwiki.page
English
Science / Physics / maxwells-demon

Maxwell’s Demon

Maxwell’s demon is a thought experiment linking thermodynamics, information, and the physical limits of measurement, computation, and work extraction.

29 keywords2 linked from6 not yet writtenWritten by AI
Second Law of Th…James Clerk Maxw…Statistical Mech…TemperatureKinetic EnergyFirst Law of The…EntropyWork (Physics)Maxwell’s…

Maxwell’s demon is a hypothetical agent that observes microscopic physical states and selectively controls a system so as to produce an apparent violation of the second law of thermodynamics. In its original form, proposed by James Clerk Maxwell in 1867, the agent operates a small door between two compartments of gas, allowing faster molecules to pass one way and slower molecules the other. The resulting temperature difference appears to arise without the expenditure of work and could, in principle, power a heat engine.

The thought experiment exposed a fundamental question about the relationship between microscopic dynamics and macroscopic irreversibility: if individual molecular motions can be observed and controlled, what prevents thermal motion from becoming a continuously available source of useful work? Its modern resolution requires treating the controller, its memory, and its surroundings as physical parts of the apparatus. Information-bearing correlations can enable work extraction, but preparing, consuming, and restoring the relevant physical resources must enter the thermodynamic balance.

Maxwell’s demon consequently became a central problem in statistical mechanics and the foundations of computation. Modern research replaces the imaginary intelligence with experimentally controlled particles, electronic circuits, molecular devices, and quantum systems. These realizations investigate how information changes the work accessible from a system without invalidating thermodynamics.

Origin and the apparent violation of the second law

Maxwell introduced the idea in a letter to Peter Guthrie Tait dated 11 December 1867 and presented it publicly in his 1871 book Theory of Heat. His imagined being could follow individual molecules and operate a door in a partition without performing appreciable work. The name “demon” was introduced later by William Thomson, subsequently Lord Kelvin. Maxwell’s purpose was not to propose a practical perpetual-motion machine, but to illuminate the statistical character of the second law.

Consider an insulated vessel divided into compartments A and B, initially containing the same gas at the same temperature. Molecules possess a distribution of velocities rather than a single speed. The demon admits unusually fast molecules from A into B and unusually slow molecules from B into A. Because molecular translational kinetic energy contributes to gas temperature, B becomes hotter and A colder. The demon appears to create a temperature gradient from an initially equilibrated system without an external energy supply.

This does not immediately contradict the first law of thermodynamics. The gas need not gain total energy: energy is redistributed between compartments. The difficulty instead concerns entropy and the availability of energy for useful work. A temperature difference could operate a heat engine, allowing some energy from an initially uniform thermal environment to be converted into work. If both sorting and engine operation could be repeated while every other component returned to its original state, the complete apparatus would violate the second law.

For two equal portions of an ideal gas at fixed volumes and with constant heat capacities, the entropy consequence is particularly transparent. If their temperatures become T+δTT+\delta T and T−δTT-\delta T, their combined entropy change is

ΔS=CVln⁡(1−δT2T2)<0,\Delta S=C_V\ln\left(1-\frac{\delta T^2}{T^2}\right)<0,

where CVC_V is the heat capacity of either portion. This expression assumes unchanged particle numbers; Maxwell’s literal sorting procedure can also redistribute particles, requiring additional terms. Nevertheless, it captures the apparent production of a lower-entropy macroscopic state.

The paradox depends on the system boundary. If only the gas is considered, selective control can indeed lower its entropy. Thermodynamics forbids an uncompensated entropy decrease for the complete isolated arrangement, not for every subsystem. The scientific problem is therefore to identify what physical change accompanies the demon’s observations and interventions, rather than simply to declare that molecular sorting is impossible.

The Szilard engine and the work value of one bit

In 1929, the Hungarian physicist Leo Szilard developed a particularly clear version of the problem: an engine containing a single molecule. The Szilard engine replaces the complicated sorting of many molecular velocities with a binary question—whether a particle lies to the left or right of a partition. It demonstrates quantitatively how information can change the work obtainable from a thermal system.

A single particle occupies a box of volume VV in contact with a reservoir at temperature TT. An ideal partition is inserted at the midpoint. In the classical idealization, insertion can be performed without work, and the particle is equally likely to occupy either half. A controller measures which half contains it and attaches a load to the partition on the appropriate side. The particle then expands isothermally from volume V/2V/2 to VV, pushing the partition and raising the load.

For an ideal, quasistatic expansion, the extracted work is

Wext=∫V/2VkBTV′ dV′=kBTln⁡2,W_{\mathrm{ext}} =\int_{V/2}^{V}\frac{k_{\mathrm B}T}{V'}\,dV' =k_{\mathrm B}T\ln 2,

where kBk_{\mathrm B} is the Boltzmann constant. The particle’s average internal energy remains unchanged during the isothermal expansion, so the work comes from heat absorbed from the reservoir. The measurement does not supply that energy; it identifies which operation will successfully convert reservoir heat into work.

At the end of the expansion, the box and particle can be returned to their starting macroscopic conditions. The controller, however, retains a record of the side on which the particle was found. Ignoring that record makes the engine appear cyclic when the full apparatus is not. Restoring the controller for repeated operation requires accounting for its memory and any auxiliary systems used in that restoration.

The midpoint construction also establishes the connection with a bit of information. A perfectly reliable answer to an initially equiprobable binary question reduces uncertainty by ln⁡2\ln 2 in natural-logarithm units, or one bit. If the partition instead divides the box into fractions pp and 1−p1-p, the mean reversible work is

⟨Wext⟩=kBT[−pln⁡p−(1−p)ln⁡(1−p)].\langle W_{\mathrm{ext}}\rangle =k_{\mathrm B}T[-p\ln p-(1-p)\ln(1-p)].

This is the binary information entropy multiplied by kBTk_{\mathrm B}T. The result does not imply that every stored bit automatically yields this amount of work. The information must concern a physical degree of freedom that can be controlled, remain useful until feedback occurs, and be exploited through an appropriate thermodynamic protocol.

Measurement, information, and physical correlations

Early analyses often sought an unavoidable cost in observation itself. In 1929, Léon Brillouin’s later line of argument was not yet available; his influential treatment emerged in the mid-twentieth century and emphasized the physical resources required to obtain reliable signals against thermal noise. Such considerations are essential for particular detectors, but modern theory does not assign a universal heat cost of kBTln⁡2k_{\mathrm B}T\ln 2 to every measurement. The central distinction is between acquiring correlations and irreversibly destroying information.

A measurement can be modeled as an interaction between a system XX and a memory MM. Initially, the memory occupies a known reference state. The interaction maps different system states to distinguishable memory states: a particle on the left produces one record, while a particle on the right produces another. The record is useful because system and memory are now statistically correlated. No consciousness or semantic understanding is required; a mechanical linkage or electronic detector can perform this role.

The strength of the correlation is quantified by mutual information. For discrete variables,

I(X;M)=∑x,mp(x,m)ln⁡p(x,m)p(x)p(m).I(X;M)= \sum_{x,m}p(x,m) \ln\frac{p(x,m)}{p(x)p(m)}.

It vanishes for statistically independent variables and equals ln⁡2\ln 2 for a perfect record of an equiprobable binary state. A noisy detector generally produces less mutual information. Its output may contain considerable randomness while conveying little about the measured system, so the memory’s entropy alone is not a measure of useful knowledge.

The joint information entropy satisfies

H(X,M)=H(X)+H(M)−I(X;M).H(X,M)=H(X)+H(M)-I(X;M).

Consequently, treating correlated subsystems as though they were independent loses a term that is crucial to the thermodynamic accounting. For suitable classical, noninteracting subsystems at a common temperature, the nonequilibrium free energy of the joint state contains a correlation contribution kBTI(X;M)k_{\mathrm B}TI(X;M). Feedback can consume this resource while extracting work or lowering the entropy of the controlled subsystem.

Measurement nevertheless has physical requirements. A memory must begin in an appropriate state, its records must be distinguishable on the operating timescale, and the interaction must establish the desired correlations. Real instruments also dissipate energy through amplification, communication, and control. These practical costs can greatly exceed fundamental lower bounds, but they should not be confused with a universal minimum assigned to observation alone. An idealized reversible measurement can correlate system and memory without necessarily generating thermodynamic entropy; the costs of preparing and reusing the memory remain part of the full cycle.

Landauer’s principle and reversible computation

The decisive connection with computation was developed by Rolf Landauer at IBM in 1961. Landauer’s principle concerns logically irreversible operations: operations that map several distinguishable logical inputs onto the same output. Resetting an unknown memory bit to zero is the standard example. Both possible inputs, 0 and 1, become 0, so the output no longer determines which input was present.

For a memory whose two logical states have equal energies, initially occupied with equal probabilities, isothermal reset to a standard state at temperature TT requires at least

Werase≥kBTln⁡2W_{\mathrm{erase}}\geq k_{\mathrm B}T\ln 2

of average work in the ideal limit. If the memory’s internal energy is unchanged, at least the same amount of heat is released to the environment. At 300 K300\ \mathrm K, this scale is approximately 2.9×10−21 J2.9\times10^{-21}\ \mathrm J per bit. The bound applies to a specified physical transformation under specified conditions, not to every operation performed by a computer.

A physical picture uses a particle in a double-well potential. Its left and right locations encode the logical values. Initially, the particle may occupy either well. A reset protocol lowers barriers and reshapes the potential so that both initial possibilities end in one designated well. The reduction in the memory’s entropy must be compensated elsewhere. Approaching the minimum dissipation requires sufficiently slow, well-controlled operation; finite-time protocols generally require additional work.

Charles H. Bennett clarified the consequences for Maxwell’s demon in 1982. Measurement and the subsequent control operation can, in principle, be implemented reversibly. The thermodynamic obstacle appears when the demon restores its memory to the standard state needed for another identical cycle. In the symmetric Szilard engine, the minimum reset cost balances the maximum kBTln⁡2k_{\mathrm B}T\ln 2 extracted during expansion. A demon with an indefinitely expanding supply of initially blank memory can run without immediate erasure, but it consumes that prepared memory as a resource.

Bennett’s work also connected the problem to reversible computing. A logically reversible operation preserves enough information to reconstruct its input, potentially avoiding the minimum dissipation associated specifically with erasure. This does not make physical computing automatically free of heat production: reliable hardware, rapid switching, error correction, and discarded intermediate records introduce additional requirements. Nor does erasure always have the same cost. Biased logical probabilities, unequal state energies, and accessible correlations alter the relevant entropy and free-energy changes. The familiar one-bit bound is the simplest symmetric case of a broader thermodynamic accounting.

Feedback control and stochastic thermodynamics

At microscopic scales, work and heat fluctuate from one experimental run to another. A colloidal particle can momentarily move against a force, and a small system can display a negative entropy-production fluctuation without contradicting the macroscopic second law. Stochastic thermodynamics describes these individual trajectories and connects their statistical properties to average thermodynamic constraints.

A demon operating through feedback uses a measured outcome to choose a subsequent control protocol. In ordinary driving, the protocol is specified independently of the system’s instantaneous microscopic state. Under feedback, the controller can preferentially respond to favorable fluctuations: it might shift a trap only when a particle has moved uphill, preventing the particle from immediately falling back. The distinction is not that thermal fluctuations create energy, but that information determines when and how external constraints change.

For a common classical measurement-and-feedback setup, the average work performed on the controlled system obeys

⟨Won⟩≥ΔF−kBTI,\langle W_{\mathrm{on}}\rangle \geq \Delta F-k_{\mathrm B}T I,

where ΔF\Delta F is the equilibrium Helmholtz free-energy difference between specified endpoints and II is the average mutual information obtained by the measurement, expressed using natural logarithms. For cyclic operation of the controlled system, ΔF=0\Delta F=0, giving

⟨Wext⟩≤kBTI.\langle W_{\mathrm{ext}}\rangle\leq k_{\mathrm B}TI.

The controller’s measurement and resetting costs are not included in this subsystem inequality. Their inclusion restores the complete thermodynamic balance.

Takahiro Sagawa and Masahito Ueda developed influential generalized second-law and fluctuation-theorem formulations incorporating information. Under appropriate assumptions, an information-corrected equality can take the form

⟨e−β(Won−ΔF)−i⟩=1,\left\langle e^{-\beta(W_{\mathrm{on}}-\Delta F)-i} \right\rangle=1,

where β=(kBT)−1\beta=(k_{\mathrm B}T)^{-1} and ii is the trajectory-dependent information gained. Applying Jensen’s inequality gives the corresponding average-work bound. Particular feedback protocols, including those involving error-free measurements and restricted reverse trajectories, can require modified formulations rather than this simple expression.

These results make the demon quantitatively testable. Detector noise reduces useful information; delay permits the measured state to decorrelate from the current state; imperfect control wastes the available correlation. Increasing precision is not necessarily advantageous if acquisition or processing becomes too costly. The meaningful comparison is among work extracted, information actually available for feedback, and the dissipation of the full apparatus. This framework distinguishes a genuine information-powered subsystem from a device merely driven by an uncounted external energy source.

Autonomous demons, ratchets, and information reservoirs

An autonomous demon operates through fixed physical interactions rather than an external observer repeatedly measuring and issuing commands. Its sensor, controller, and working substance evolve together. Such models make especially clear that “demon” denotes a functional arrangement, not an intelligent entity. Information can be encoded in molecular conformations, charge states, or sequences of incoming memory units.

A prominent example is the information-reservoir model introduced by Dibyendu Mandal and Christopher Jarzynski in 2012. A small mechanical system interacts sequentially with a tape of binary units. If the incoming tape has a suitable bias, interaction can produce mechanical work while making the outgoing tape more random. The tape’s increase in information entropy compensates the entropy reduction associated with converting heat into work. Other operating regimes allow the device to expend work to make the tape more ordered.

For energetically equivalent, independent tape symbols, an idealized bound has the form

Wext≤kBT(Hout−Hin).W_{\mathrm{ext}}\leq k_{\mathrm B}T(H_{\mathrm{out}}-H_{\mathrm{in}}).

Here the entropies are measured per processed symbol using natural logarithms. When correlations exist between symbols, single-symbol entropies are generally insufficient; the information entropy rate of the full sequence becomes relevant. Preparing the low-entropy input tape and restoring it after use must be included if the entire device is claimed to operate cyclically.

A related challenge is the Brownian ratchet, analyzed by Marian Smoluchowski and later popularized by Richard Feynman. A ratchet-and-pawl mechanism seems capable of rectifying random molecular impacts into directed rotation. Yet at one uniform temperature, the pawl and ratchet also fluctuate. Reverse movements and thermal lifting of the pawl prevent sustained extraction of work from equilibrium noise alone. An asymmetric shape does not itself supply a thermodynamic resource.

Autonomous devices can function when maintained away from thermodynamic equilibrium, for example by a temperature difference, a voltage bias, or a supply of chemically prepared reactants. In that case, the source of usable free energy must be identified. Molecular motors driven by adenosine triphosphate hydrolysis are not violations of the second law: they consume chemical free energy. Likewise, an autonomous electronic demon can use one subsystem to sense another while paying for that operation through heat flow or electrical dissipation. Describing such machines in informational terms can reveal how sensing and control are organized, but it does not remove the underlying energetic bookkeeping.

Experimental realizations

Experimental Maxwell demons do not sort an ordinary roomful of gas molecule by molecule. Instead, researchers use systems with a small number of measurable states and accurately controllable energies. Their objective is to demonstrate particular steps—information-assisted work extraction, feedback cooling, or memory erasure—and compare measured averages and fluctuations with theoretical bounds.

In 2010, Shoichi Toyabe and colleagues reported feedback-controlled conversion of information into free energy using a microscopic particle in a controlled potential. The particle’s position was observed, and the potential was changed conditionally to exploit favorable thermal fluctuations. This allowed the particle to gain potential energy through information-dependent control. The experiment tested an information-modified thermodynamic relation, not the proposition that the complete imaging and control apparatus produced net work without an energy supply.

In 2012, Antoine Bérut and colleagues tested Landauer’s principle using a colloidal particle in a double-well optical potential. The particle’s location represented a bit. By controlling the potential and driving the particle into a designated well, they implemented erasure and measured the associated dissipation. Slow protocols approached the Landauer limit, while faster operation dissipated more energy. In 2015, Jun and colleagues used a feedback trap to realize high-precision erasure protocols with a virtual potential, providing another experimental route to the same physical issue.

Electronic systems offer a complementary platform because individual charges can be detected and controlled. In 2014, Jonne V. Koski and colleagues demonstrated a single-electron Szilard engine based on a metallic electronic device. Feedback used information about charge occupancy to select a work-extraction protocol, and the experiment examined the relationship between information and dissipated heat. Such systems can operate at low temperatures, where charging energies and thermal energies are well separated and single-electron transitions are resolvable.

Autonomous electronic demons have also been realized using capacitively coupled single-electron devices. In these arrangements, one subsystem’s charge affects transition energies in another, allowing a detector-controller function to be embodied directly in the coupling. Energy and entropy changes must then be assessed across both subsystems and their reservoirs.

The interpretation of each experiment depends on its boundary and measured quantities. Feedback cooling removes energy or entropy from a selected degree of freedom; it need not produce useful external work. An information engine can extract work from its working substance while consuming much more power in laboratory electronics. Experimental success consists in verifying the predicted subsystem thermodynamics and identifying the compensating resources, not in eliminating them.

Quantum demons and quantum information

A quantum Maxwell demon controls systems whose relevant states require quantum mechanics. The classical distinction between observing a pre-existing state and leaving it unchanged is no longer generally available. Measurement can disturb the system, destroy coherence, and exchange energy with the measurement apparatus. Quantum versions therefore require explicit models of both information acquisition and its energetic consequences.

For a quantum state represented by a density matrix ρ\rho, the von Neumann entropy is

S(ρ)=−kBTr⁡(ρln⁡ρ).S(\rho)=-k_{\mathrm B}\operatorname{Tr}(\rho\ln\rho).

A corresponding nonequilibrium free-energy expression is

F(ρ)=Tr⁡(ρH)−TS(ρ),F(\rho)=\operatorname{Tr}(\rho H)-TS(\rho),

where HH is the system’s Hamiltonian. When ρ\rho has coherence between different energy eigenstates, not all of the free energy indicated by a simple state-function difference need be accessible under a restricted set of operations. The answer depends on whether phase references, coherent control, and appropriate ancillary systems are available.

Entanglement introduces possibilities absent from classical memory. For correlated systems AA and BB, the quantum conditional entropy is

S(A∣B)=S(AB)−S(B).S(A|B)=S(AB)-S(B).

Unlike its classical counterpart, this quantity can be negative. In the framework of erasure with accessible quantum side information, developed by Lídia del Rio and colleagues in 2011, quantum correlations can reduce the work cost of erasure and, under suitable conditions, permit work extraction during the process. This does not provide a repeatable resource-free engine: the correlations are consumed, and rebuilding them belongs in the accounting.

Quantum experiments have implemented demon-like protocols in platforms including superconducting circuits and nuclear magnetic resonance systems. These investigate measurement-conditioned operations, the thermodynamics of small quantum memories, and the relation between quantum information and work. A central experimental task is distinguishing work supplied by coherent control fields from energy received as heat or injected through measurement back-action.

Related proposals describe measurement-powered engines, in which the measurement process itself changes the system’s energy. Such devices must not be interpreted as extracting energy solely from knowledge: the measurement apparatus is an energetic resource. Quantum theory broadens the types of correlation and control available to a demon, while preserving the need to include preparation, measurement, feedback, and restoration within a consistent physical description.

Scope, limitations, and continuing research

Maxwell’s demon is often described as showing that “information is physical.” More precisely, information must be represented in physical states, and manipulating those states is constrained by their dynamics and thermodynamics. Abstract information does not have a single fixed energy equivalent. The quantity kBTln⁡2k_{\mathrm B}T\ln 2 refers to a particular entropy change at a particular temperature under specified conditions, rather than an intrinsic energy contained in every bit.

Nor does the resolution imply that ignorance alone generates heat. An observer who forgets a fact mentally has not necessarily performed the specific physical reset operation considered by Landauer’s principle. Conversely, a mechanism can exploit correlations without assigning meaning to them. The relevant objects are distinguishable physical records, their statistical relationships to controlled systems, and the operations that transform them.

A continuing research problem is the thermodynamics of realistic finite-time control. Ideal protocols often approach reversibility only as their duration becomes very long, so maximum efficiency and maximum power generally occur under different conditions. Maintaining stable memories also requires barriers large enough to suppress unwanted transitions, while changing their states rapidly demands appropriate control. These relationships connect information thermodynamics with the design of low-power devices, although practical computing usually dissipates far more energy than the Landauer scale.

Another research direction concerns partial observation. A controller may see only a noisy projection of a system with hidden degrees of freedom. Estimating entropy production from the visible record can then miss dissipation or misidentify available information. Time-dependent correlations, feedback delay, and continuous measurements require more elaborate measures than a single binary record. These issues also arise in biological sensing, where biochemical networks acquire information while consuming chemical resources.

For quantum systems, major questions concern which operations are allowed, what physical reference frames they require, and how coherence and correlations can be converted into useful work. For autonomous classical systems, a related question is how best to separate energetic and informational exchanges when sensing and actuation occur through the same interaction. In both settings, a claim of enhanced work extraction is meaningful only after the relevant reservoirs, memories, controllers, and initial correlations have been specified.