Exorcising Maxwell's Demon

For 115 years, physics harboured a fugitive. Maxwell's demon — the imp who sorts molecules and makes heat flow the wrong way — was proposed in 1867, reduced by Szilard in 1929 to a one-molecule engine that converts a bit of knowledge into kT \ln 2 of work, and never quite arrested. Generations of physicists proposed a cost that would balance the books; generations of closer looks found the proposals wanting. The case was finally closed in 1982 by Charles Bennett at IBM, using his colleague Landauer's principle — and the resolution is more elegant, and more surprising, than any of the failed attempts. The demon is not forbidden from learning. It is charged for forgetting.

The debt, restated

Recall where the previous lesson left the Szilard engine. Each cycle: partition the box, measure the molecule's side into a memory bit, expand, extract work kT \ln 2 drawn as heat from a single bath. The bath's entropy falls by k \ln 2 per cycle and nothing visible rises to compensate — an apparent second-law violation of exactly one bit's worth per lap. The classic escape was to blame the measurement: surely peeking at a molecule must cost something? Bennett's first move was to demolish that escape.

Measurement can be free. Copying the molecule's side into a blank memory bit is the map

(\text{side},\, 0) \;\longmapsto\; (\text{side},\, \text{side}),

and on the states that actually occur — (L,0) and (R,0) — it is one-to-one: nothing is merged, nothing is lost. It is a bijection, the same shape as the input-keeping XOR from earlier in this course, and Landauer's principle levies nothing on bijections. In principle the demon can learn which side the molecule is on at zero dissipation. (Notice the load-bearing word blank: copying onto a bit that might already hold something would merge states — that would be an erasure in disguise.)

So the measurement is free, the expansion is free — where is the second law hiding? Follow the bit. To run as a cycle, the engine must return everything to its starting state — and the demon's memory now holds one random bit (L or R, equally likely over many runs) where a blank used to be. Before the next round can begin, that bit must be reset to blank. Reset of a random bit is precisely the operation Landauer priced: it dissipates at least kT \ln 2 into the bath — exactly cancelling the work the engine extracted. The books balance to the zeptojoule. The demon earns nothing; the second law survives.

The ledger

Here is the full cycle as an entropy account. The profound line is the third: between extraction and erasure, the "missing" entropy of the bath has not vanished — it is hiding in the demon's memory, stored as the randomness of the record. Information is where entropy goes to wait.

stage	bath entropy	memory	running total (bath)
copy side into blank bit	0	holds record (1 random bit)	0
isothermal extraction, W = kT ln 2	−k ln 2	still holds record	−k ln 2
erase record (Landauer)	≥ +k ln 2	blank again	≥ 0

measurement — copying a system's state into blank memory — is a bijection and can be performed with no fundamental dissipation (contra Brillouin);
the engine's gain of kT \ln 2 work per cycle is real, and while the record is kept, the bath's lost entropy resides in the demon's memory as the record's randomness;
completing a cycle requires resetting the memory to blank, and by Landauer's principle that erasure dissipates \ge kT \ln 2 — exactly cancelling the extracted work;
hence over any full cycle the demon's net work is \le 0: the second law holds, with information as a legitimate entry in the entropy books.

Modern vindication: demons in the lab

For most of its history this argument was theology about a thought experiment. No longer. Since around 2010, "information engines" have been built and run. A Tokyo group steered a single colloidal particle up a potential staircase using feedback from position measurements — verifiably converting information into work at rates bounded by kT \ln 2 per bit. A Finnish group built a Szilard engine out of a single electron hopping between two tiny metal islands, extracting heat from the environment cycle after cycle under measurement-and-feedback control, and separately measured the Landauer cost of erasing the controller's record. Every experiment tells the same story the ledger above tells: the demon works exactly as advertised while its notebook has blank pages, and pays exactly as predicted when the notebook must be wiped. Maxwell's imp has become laboratory equipment — arguably the most complete rehabilitation any paradox in physics has ever enjoyed. Maxwell, who invented the demon precisely to show that the second law was statistical rather than sacred, would have been delighted on both counts.

The demon collected distinguished scalps before Bennett. Around 1912 Smoluchowski examined the simplest mechanical demon — a spring-loaded trapdoor that only opens one way — and showed it fails beautifully: at temperature T the door itself jiggles thermally, flapping open at random and letting molecules leak backwards; a demon made of atoms is as drunk as the gas it polices. In the 1950s Léon Brillouin argued the flaw was in seeing: to spot a molecule the demon must scatter light off it, and he computed a cost of about kT \ln 2 per observation. The number was right; the shop was wrong. Brillouin had costed one particular clumsy way of measuring, not measurement itself — Bennett later exhibited measuring devices (essentially a bistable ratchet gently dragged past the molecule) that copy the answer into blank memory with no dissipation at all. It is a recurring pattern in this subject: plausible costs attached to the wrong step, until Landauer's accounting identified the only step that must pay. Moral: in thermodynamics, as in auditing, finding a cost is easy; proving it irreducible is the whole game.

The single most persistent error in this corner of physics — repeated in textbooks for decades after 1982, and still common — is "the demon fails because measurement costs kT \ln 2." It does not. Measurement into blank memory is a bijection and can be free; that is the very heart of Bennett's resolution, and the reason his paper was surprising rather than a footnote to Brillouin. The irreducible cost sits at the erasure of the record — in forgetting the measurement, not making it. A quick self-test that you have the resolution right: if the demon never erases (just keeps writing into fresh blank memory), does it beat the second law? No — it merely runs up the bill. It converts heat to work while filling its memory with random bits, which is not a cycle but a trade: order in the bath exchanged for disorder in the notebook, second law satisfied at every instant. The perpetual-motion claim only arises for a true cycle, and a true cycle must erase.

The moral for computing

The demon story is this module's thermodynamics packed into one fable. Information is a physical quantity that can hold entropy on account; acquiring it reversibly is free; destroying it is taxed at kT \ln 2; and any process that seems to cheat the second law is quietly deferring an erasure. Computing hardware is a demon of exactly this kind — a machine that measures, records and overwrites billions of bits a second. The rest of this course asks the engineer's version of Bennett's question: how much of what our machines forget did they ever need to forget — and what would a computer look like that, like a good demon, never wipes its notebook at all?