Thermodynamics ↔ information ↔ computation
Computing a number can, in principle, be free. Forgetting it cannot. There is a hard floor — about three billion-trillionths of a joule per bit — on the heat released when a memory is wiped. That floor is where information stops being an abstraction and becomes physics. This is the story of how a 19th-century imp forced the issue.
How to read this. Four stages, in order — each opens with a real puzzle, builds the model under it, lets the strongest objections fight, and hands you to the next. Stage 2 has a one-molecule engine you can run yourself. The point isn't to memorize the result; it's to watch where the bill actually comes due.
"…if we conceive of a being whose faculties are so sharpened that he can follow every molecule in its course, such a being… would be able to do what is at present impossible to us." — James Clerk Maxwell, letter to P. G. Tait, 1867
Maxwell imagined a gas-filled box split in two by a wall with a tiny frictionless trapdoor. A "neat-fingered being" watches the molecules. When a fast one approaches from the right, it opens the door and lets it through to the left; when a slow one approaches from the left, it lets it through to the right. No work is done on the gas — the door is weightless and frictionless. Yet the left side grows hot and the right side grows cold.
That is a problem. The second law of thermodynamics says heat does not flow from cold to hot on its own, and that the entropy of an isolated system never decreases. A temperature difference is usable energy: drop a heat engine between the two sides and you can extract work, forever, from a box that started in dull equilibrium. The demon appears to manufacture order — and therefore free work — out of nothing but knowing which molecule is which.
For roughly sixty years the demon was treated as a paradox to be exorcised, and the hunt was always for the hidden energy cost — some entropy increase, somewhere, that the demon's tidy sorting must secretly pay for. The first serious candidates pinned it on perception: surely the demon must see each molecule, and seeing means bouncing a photon off it, and that photon must cost something.
The demon's defender
"Sorting is just steering. The door is frictionless, the demon does no work on any molecule. If the books must balance, show me the entry — and don't tell me thinking is free to assume the cost."
The exorcist (Brillouin, 1951)
"To see a molecule against the thermal glow of a warm box, the demon needs at least one photon above the background. Measuring dissipates ≥ kT ln 2. The cost is in the looking."
Where this leaves us
The demon did something more useful than die: it welded two subjects together. Whatever saves the second law has to be charged to the demon's information processing — its measuring, remembering, or sorting. Thermodynamics now had an information term in its ledger. The open question was which line item carried the charge. The 20th century's first answer — that the cost lives in measurement — sounded obvious, held for decades, and was wrong.
To find the real entry, the demon had to be stripped down until only one molecule and one bit were left.
Szilárd's move was to make the demon's accounting unavoidable by shrinking the gas to a single molecule. With one molecule, "which side is it on?" is a single yes/no question — exactly one bit of information. And he showed that one bit, fed into the right machine, buys you a precise amount of work: kT ln 2, about 3 × 10⁻²¹ J at room temperature. This was 1929 — nineteen years before Claude Shannon would even define the bit. Szilárd had priced information in joules before information theory existed.
Szilard one-molecule engine
One molecule, one heat bath, one watchful demon. Run a cycle and watch the ledger.
Work extracted
0
Heat paid to erase
0
Net useful work
0
A full cycle has four moves. (1) Insert a partition down the middle; the molecule is now trapped on one side. (2) Measure which side — store that one bit in the demon's memory. (3) Extract: knowing the side, attach a load to the partition on the molecule's side and let the one-molecule "gas" push it outward. As the molecule batters the moving wall, it does work; the box, in contact with a heat bath at temperature T, draws in just enough heat to keep the molecule's energy steady. Expanding from half the box to the whole box yields exactly W = kT ln 2. (4) Remove the partition; you're back to start.
Read that ledger again: heat flowed out of a single reservoir and came back as useful work, with nothing else apparently changed. That is precisely the perpetual-motion machine the second law forbids. The only thing that made it possible was the bit.
Treat the lone molecule as an ideal gas with one particle. Isothermal expansion from volume V/2 to V against a movable wall does work W = ∫ p dV = kT ∫V/2V dV/V = kT ln(V ÷ (V/2)) = kT ln 2, using the one-particle ideal-gas law pV = kT.
The matching entropy view: confining one molecule to half the box and then releasing it changes the accessible phase space by a factor of two, i.e. an entropy of k ln 2. The bit and the entropy are the same quantity wearing different clothes — Boltzmann's S = k ln W with W = 2.
Szilárd knew the second law had to survive, so he posited that the missing entropy must be generated by the act of measurement and memory — somewhere in acquiring the bit, at least k ln 2 of entropy had to be produced. Brillouin later gave this a mechanism (the photon). It was the natural guess, and it was the wrong place. Notice what the engine above shows: you can run the cycle and bank +kT ln 2 of real work. The "violation" is sitting right there. The defenders of measurement-cost had identified that a cost exists without correctly locating it.
Where this leaves us
The exchange rate is exact and real: one bit ⇄ kT ln 2 joules. The engine genuinely converts heat into work. So either the second law is broken — or there is a step we keep getting for free that isn't free. Finding it meant asking a question no one in thermodynamics was asking: what does it physically cost not to acquire information, but to destroy it?
That question came from inside a computer company, from a man asking whether computation must, by the laws of physics, generate heat.
"Information is physical." — Rolf Landauer's slogan, which this stage is the proof of
Landauer was trying to answer a practical question — how little heat could a computer possibly produce? — and found something sharper than he expected. Most logical operations carry no intrinsic energy cost at all. You can, in principle, build them to run reversibly, releasing no heat. The exception is a specific, unglamorous class: the operations that throw information away.
Call an operation logically irreversible if you cannot reconstruct its input from its output. Erasing a bit — forcing it to 0 no matter what it was — is the cleanest example: two possible inputs (0 or 1) collapse to one output (0). The input is gone; the operation is two-to-one.
Here is the bridge. A bit you don't know the value of occupies two states; after erasure it occupies one. That halving is a drop in entropy of k ln 2 inside the memory. But the total entropy of the universe cannot fall. So the k ln 2 has to go somewhere — it is pushed out into the surroundings as heat. At temperature T, that costs at least
This is Landauer's principle. The cost is not in computing, storing, or measuring. It is in discarding. A logically reversible computer — one that never merges two states into one — has no lower bound on its heat output. The heat is the price of two-to-one, and only of two-to-one.
A reversible gate (say, a NOT, or the controlled-NOT) is a one-to-one relabelling of states. It permutes the phase space without shrinking it, so it changes no entropy and needs no minimum heat. Bennett showed in 1973 that any computation can be rebuilt this way — keep a record of intermediate steps so nothing is ever overwritten, run forward, copy out the answer, then run backward to tidy up.
Erasure is different in kind: it is the one move that genuinely reduces the number of states the system could be in. You can delay it, relocate it, or batch it, but if a machine is to return to a standard ready state and run again, somewhere the accumulated garbage must be cleared — and clearing is erasure. Landauer's bound is the toll on that one road.
The bound is real — and measured
In 2012 Bérut and colleagues built a single bit from a microscopic bead in a double-well laser trap, erased it, and measured the heat. It approached kT ln 2 as the erasure was done slowly — the first direct confirmation, 51 years on.
The skeptic (Earman & Norton)
"You're invoking the second law to derive a cost, then using that cost to rescue the second law. Either it's circular, or the principle is doing no independent work. Prove the floor without assuming what you're proving."
Where this leaves us
Landauer split computation cleanly in two: the reversible part, which physics lets you do for free, and the act of forgetting, which it taxes at kT ln 2 a bit. That tax is the answer to our title question. What it had not yet done — in 1961 — was collect the demon's debt. The pieces were now on the table: an engine that earns kT ln 2 per bit, and a forgetting that costs kT ln 2 per bit. Someone had to put them in the same sentence.
In 1982, also at IBM, Charles Bennett did — and the sixty-year-old demon finally paid up.
Bennett returned to Maxwell's box with Landauer's principle in hand and overturned the consensus. The decades-old claim that measurement must dissipate energy, he argued, is false: in principle the demon can observe a molecule, and the Szilard engine can read its bit, while releasing no heat at all. Brillouin's photon was a feature of one clumsy way to look, not a law. So if measurement is free, the second law is in real trouble — unless the cost hides somewhere else.
It hides in the reset. Look back at the engine in Stage 2. Each cycle, the demon writes a fresh bit — "left" or "right" — into its memory. To extract work again, it needs that memory empty, in a known ready state. But the demon's memory is finite. Run the cycle enough times and every slot is full of a measurement it can't use again. To keep going, it has to clear them. Clearing is erasure. Erasure, by Landauer, costs kT ln 2 per bit — exactly the work the engine extracted.
The books balance to the cent. Every kT ln 2 the demon earns by knowing, it must hand back to forget. Over a complete cycle — one that genuinely returns the demon and its memory to the start — net useful work is zero. The second law was never violated; the violation was an illusion created by quietly skipping the reset. That is the "free" step Stage 2 warned you about.
The resolution
Maxwell's demon is exorcised not at the moment it looks, but at the moment it must forget to look again. Information is not a free-floating abstraction sitting outside thermodynamics. A bit held in a physical memory is a thermodynamic resource; spending it is fine, but a cyclic machine has to clear its memory, and clearing has a price. The engine you ran shows the whole argument: the "net useful work" cell can only stay positive while the demon hoards un-erased bits. The instant it erases to run again, the cell snaps to zero.
The floor under all computing. Landauer's bound is the ultimate thermodynamic limit on irreversible computation. Today's chips sit far above it — a single transistor switch dissipates thousands to millions of times kT ln 2 — so the limit is not today's electricity bill; engineering waste dwarfs the fundamental cost. But it is a real wall: as devices shrink toward it, the only way past is reversible computing, machines built to never throw a bit away. The slogan "AI burns energy" is, at rock bottom, a theorem about erasure — even if practice is nowhere near the floor.
The reach beyond chips. The same logic — a bit costs k ln 2 of entropy, full stop — turns up in black-hole thermodynamics (the Bekenstein bound on how much information a region of space can hold) and in the physical reading of the Church–Turing thesis. Once you accept that information is physical, it stops respecting the border between computer science and physics.
Not everyone accepts that Landauer's principle grounds the second law. John Earman and John Norton ("Exorcist XIV," 1998–99) argued the standard exorcism is circular: it leans on the second law to establish erasure's cost, then cites that cost to defend the second law. On their reading the principle is either unproven or redundant.
Defenders reply that the bound can be derived from statistical mechanics independently, and point to the 2012 and later experiments showing erasure really does dissipate heat down to the predicted floor. The empirical question — does forgetting cost at least kT ln 2? — looks settled. The foundational question — is that fact the reason Maxwell's demon fails, or just consistent with its failing? — is still argued. A walkthrough can hand you the strongest version of each side; it can't hand you the last word, because there isn't one yet.
The short answer
Computing a bit need not change how many states your machine could be in, so physics lets it cost nothing. Erasing a bit collapses two possible pasts into one — it shrinks the machine's phase space, lowering its entropy by k ln 2. The universe's entropy can't fall, so that decrease is exported as heat, at least kT ln 2 of it, every time. Maxwell's demon looked like it broke this rule for sixty years; it was only ever skipping the reset. The cost of knowledge isn't in acquiring it. It's in clearing the slate to learn the next thing.
H. Leff & A. Rex, Maxwell's Demon 2: Entropy, Classical and Quantum Information, Computing (2003) — the anthology. Collects Maxwell, Szilárd, Brillouin, Landauer, and Bennett in one place. The scholarly entry point.
James Gleick, The Information (2011) — the popular narrative bridge from Maxwell's demon through Shannon to the physics of bits. Start here if the names above are new.
L. Szilárd, "On the decrease of entropy in a thermodynamic system by the intervention of intelligent beings" (1929) — the one-molecule engine; information priced in joules before information theory.
R. Landauer, "Irreversibility and heat generation in the computing process," IBM J. Res. Dev. (1961) — logical irreversibility ⇒ physical dissipation; the kT ln 2 bound.
C. Bennett, "The thermodynamics of computation — a review," Int. J. Theor. Phys. (1982) — the demon resolved via erasure; builds on his 1973 proof that computation can be made reversible.
A. Bérut et al., "Experimental verification of Landauer's principle," Nature 483 (2012) — the colloidal-bead bit; erasure heat measured down to kT ln 2.
J. Earman & J. Norton, "Exorcist XIV: The Wrath of Maxwell's Demon," Stud. Hist. Phil. Mod. Phys. (1998–99) — the circularity objection; the live foundational dispute.