Maxwell's Demon

The Landauer–Bennett account is the standard resolution, appears in every modern textbook, and has been experimentally confirmed. It is not, however, without serious critics, and the shape of the disagreement is worth knowing.

The circularity objection. John Earman and John Norton, in a pair of papers titled "Exorcist XIV: The Wrath of Maxwell's Demon" (1998–1999), argue that the resolution assumes what it sets out to prove. Landauer's principle, they say, is either derived from the second law — in which case using it to save the second law is circular — or it is an independent postulate, in which case it needs its own justification, which nobody has supplied from first principles. Norton has pressed this line since, arguing that the fluctuations which defeat Smoluchowski's trapdoor also defeat any physically realised demon, so no information-theoretic argument is needed and none is available. On this view the demon is exorcised by ordinary statistical mechanics, and the whole apparatus of bits and erasure is a decorative extra.

The defence. Bennett replied in 2003 that the erasure cost can be motivated independently, by counting phase-space volume: a memory that can be in two states occupies twice the phase space of one forced into a single state, and Liouville's theorem does not permit phase-space volume to be destroyed — only pushed elsewhere, which is what the heat is. Owen Maroney, Orly Shenker and others have mapped the argument in detail without settling it.

A useful distinction. Harvey Leff and Andrew Rex separate two questions that often get merged. Can a demon violate the second law? No; on this everyone agrees, and it has been agreed since Smoluchowski. Is Landauer's principle the reason? Here the disagreement is live. The site classifies this entry as resolved because the physical question is settled and the experiments agree with the bound; the philosophical question of what does the explanatory work is not.

What the experiments do and do not show. Bérut and colleagues (2012) trapped a single colloidal bead in a double-well optical potential — a one-bit memory — and measured the heat released when the bit was reset. As the erasure was carried out more and more slowly, the mean dissipated heat fell and levelled off at $kT\ln 2$, approaching the bound from above and never crossing it. This confirms that the bound is real and attainable. It does not, by itself, settle whether the bound is a consequence of the second law or a partner to it.

Quantum versions. The demon has been rebuilt with qubits, with the correlation between demon and system expressed as quantum mutual information, and the generalised second law of Takahiro Sagawa and Masahito Ueda (2008 onwards) gives the modified bound under feedback control. The structure survives intact: the demon's advantage is exactly its mutual information with the system, and it is spent when the correlation is destroyed.

December 1867. Maxwell writes to Peter Guthrie Tait describing "a very observant and neat-fingered being" who can see individual molecules and operate a frictionless slide. His stated aim: "to pick a hole" in the second law of thermodynamics — meaning, to show that it holds only statistically. He was explicit that no real being could do it.

1871. The being appears in print in the last chapter of Theory of Heat, under the heading "Limitation of the Second Law of Thermodynamics". Maxwell calls it only "a being whose faculties are so sharpened that he can follow every molecule in its course".

1874. William Thomson, later Lord Kelvin, calls it a "demon" in print. Maxwell disliked the word; he meant a servant, not a devil, and complained that Thomson's label suggested malice. The name stuck anyway.

1912. Marian Smoluchowski shows that a purely mechanical demon — a spring-loaded trapdoor — cannot work, because thermal fluctuations operate the door as often as the molecules do. This settles the mechanical case for good and leaves only the intelligent case open.

1929. Leó Szilárd reduces the problem to a single molecule and a single bit, and argues that acquiring one bit of information must produce at least $k\ln 2$ of entropy. His paper is the founding document of the physics of information, published fifteen years before anyone spoke of information theory.

1949–1951. Léon Brillouin and Dennis Gabor develop the measurement-cost argument: seeing a molecule requires a photon distinguishable from the thermal background, and that photon carries the necessary entropy. This becomes the textbook answer for a generation.

1961. Rolf Landauer, at IBM, derives the erasure bound while studying the thermodynamic limits of computers, not demons. The connection to Maxwell is not his main subject.

1973. Charles Bennett proves that computation can be made logically reversible — every step undoable — so that computing as such has no lower thermodynamic bound.

1982. Bennett applies the reversibility result to the demon and identifies memory erasure as the irreducible cost. Measurement, he shows, is not the bottleneck; Szilard and Brillouin had located the fee correctly in total and incorrectly in kind.

1998–1999. Earman and Norton publish the "Exorcist" papers and open the philosophical dispute that continues today.

2010–2012. Toyabe and colleagues demonstrate information-to-energy conversion in a Szilard-type engine (Nature Physics, 2010). Bérut, Arakelyan, Petrosyan, Ciliberto, Dillenschneider and Lutz measure the Landauer bound directly on a one-bit colloidal memory (Nature, 2012). The 145-year-old thought experiment becomes a bench-top measurement.

It changed what the second law is a law about. Before the demon, thermodynamic entropy was a property of steam and gases. After Szilard, Landauer and Bennett, it is a property of any physical system that stores or destroys distinctions — including a hard disk, a neuron and a strand of DNA. Landauer's slogan, "information is physical", is the demon's legacy in four words.

It put a hard floor under computing. Every irreversible logic gate in every processor ever built pays Landauer's toll. The bound at room temperature is $kT\ln 2 \approx 2.9 \times 10^{-21}$ joules per bit erased. Current CMOS logic dissipates roughly three to four orders of magnitude more than that per operation, so the bound is not yet the binding constraint — but it is the only constraint that no engineering can remove, and the gap has been closing for decades.

It is why reversible and quantum computing exist as fields. Bennett's 1973 result — that you can compute without erasing, if you are willing to carry your intermediate results around — is the reason adiabatic and reversible logic are studied at all. It is also structural to quantum computing: quantum evolution is unitary and therefore reversible by construction, which is why a quantum circuit is built from invertible gates and why measurement is the step that costs something.

It made the second law survivable at small scales. Nanomachines, molecular motors and single-molecule experiments all operate in the regime where thermal fluctuations are comparable to the energies involved — exactly the demon's regime. The framework built to answer Maxwell is the framework used to analyse them.

Stochastic thermodynamics. The fluctuation theorems of Christopher Jarzynski (1997) and Gavin Crooks (1999) extend thermodynamics to systems where the second law can be violated in individual trials, provided it holds on average. This is Maxwell's original insight — the law is statistical — made into a working quantitative tool. It is now the standard language for molecular motors, protein folding and single-molecule pulling experiments.

Feedback control with an entropy budget. Sagawa and Ueda's generalised second law states that the extra work extractable under feedback is bounded by the mutual information acquired. Any controller acting on a small, noisy system — optical tweezers, a scanning probe, an error-correcting code — is doing demon bookkeeping.

Biology. Kinetic proofreading, the mechanism by which ribosomes and DNA polymerases achieve error rates far below what equilibrium binding energies allow, works by spending free energy to discard wrong intermediates. That is erasure, biologically implemented, and the energy cost is Landauer's in everything but name.

Data-centre energy. The world's computing infrastructure now consumes a measurable fraction of global electricity. Almost none of that goes to Landauer's bound; it goes to charging capacitors, driving wires and moving air. But the existence of a floor, and the knowledge that the floor is thirteen thousand times lower than current practice, is what keeps low-power and reversible-logic research funded.

Black holes. Bekenstein and Hawking's result that a black hole has an entropy proportional to its horizon area, and Landauer's that information has a thermodynamic price, together make "what happens to the information" a question with units attached. That is the thread running from Maxwell's doorman to the black-hole information paradox.

The Szilard engine

One molecule of ideal gas in a box of volume $V$ at temperature $T$. Insert a partition at the midpoint; the molecule is confined to volume $V/2$. Measure which half, then allow the partition to move quasi-statically to the far wall against a load. The isothermal work done by a one-particle gas is

$$W = \int_{V/2}^{V} P,dV' = \int_{V/2}^{V} \frac{kT}{V'},dV' = kT\ln 2 .$$

Heat $Q = W = kT\ln 2$ is drawn from the reservoir, the gas returns to its initial macrostate, and the only surviving change in the universe is one bit set in the demon's memory.

If the cycle could be closed without touching that bit, we would have a perpetual motion machine of the second kind. It cannot.

Landauer's bound

Model the memory as a particle in a symmetric double well, the two wells being the states $0$ and $1$. Before erasure the particle may be in either well: the accessible phase-space volume is $2\Omega$. After erasure — "set to $0$", regardless of the prior state — the accessible volume is $\Omega$.

The memory's entropy therefore changes by

$$\Delta S_{\text{mem}} = k\ln\Omega - k\ln 2\Omega = -k\ln 2 .$$

Erasure is logically irreversible: two input states map to one output state, so the map is not invertible. Since total entropy cannot decrease, the environment must absorb at least $k \ln 2$, and at temperature $T$ that is heat

$$Q \ \ge\ kT\ln 2 \ \approx\ 2.87\times10^{-21}\ \mathrm{J} \quad (T = 300\ \mathrm{K}).$$

Equivalently, work $W \ge kT\ln 2$ must be supplied to perform the erasure. Note what the bound is not about: it says nothing about the cost of computing, only about the cost of forgetting.

Bennett's cycle

Bennett's contribution is a bookkeeping argument. Write the demon's operating cycle as:

1. Measure. Establish a correlation between memory bit $M$ and molecule position $X$. This can be done by a controlled-NOT-like operation, which is logically reversible and therefore has no lower bound on dissipation. Mutual information gained: $I(M{:}X) = \ln 2$ nats.
2. Extract. Use the correlation to run the Szilard step. Work obtained: $kT\ln 2$. The correlation is consumed: after the expansion, $M$ no longer tells you anything about $X$.
3. Reset. Return $M$ to a standard state so the cycle can repeat. Because $M$ is now uncorrelated with anything, this is a genuine erasure. Cost: $\ge kT\ln 2$.

Net over one full cycle: $\Delta W \le 0$. The second law holds. The essential observation is that step 1 is reversible and step 3 is not, and that the literature spent fifty years charging the fee to step 1.

The generalised second law

For a system under feedback control by a demon that acquires mutual information $I$ about it, the Sagawa–Ueda bound replaces the ordinary Kelvin statement:

$$\langle W_{\text{ext}}\rangle \ \le\ -\Delta F + k T, I .$$

Setting $I = 0$ recovers the usual second law: no work from one reservoir. Setting $I = \ln 2$ for a single perfect binary measurement recovers Szilard's $kT\ln 2$. The demon is not exempt from thermodynamics; it appears in the inequality as a term.

Why the trapdoor fails

For Smoluchowski's mechanical valve, the flap must have a restoring energy scale $E_{\text{spring}}$. To respond to a single molecular impact it needs $E_{\text{spring}} \lesssim kT$. But a flap with $E_{\text{spring}} \lesssim kT$ is itself in thermal equilibrium with the gas and opens spontaneously at a rate $\sim e^{-E_{\text{spring}}/kT} \sim 1$. Detailed balance is restored exactly. There is no regime in which the valve is both sensitive enough to rectify and stiff enough not to rattle — which is the mechanical statement of the same conclusion the informational argument reaches: you cannot get order out of a system that is at the same temperature as its surroundings, and a demon's blank memory is the only way to be colder.