Maxwell's Demon

Maxwell's letter

In an 1867 letter to Peter Guthrie Tait, James Clerk Maxwell sketched a "finite being" that could undermine the second law:

"Let us suppose that a vessel is divided into two portions, A and B, by a division in which there is a small hole, and that a being, who can see the individual molecules, opens and closes this hole, so as to allow only the swifter molecules to pass from A to B, and only the slower ones to pass from B to A. He will thus, without expenditure of work, raise the temperature of B and lower that of A, in contradiction to the second law of thermodynamics."

Kelvin coined the term "Maxwell's demon" a few years later. For more than a century the paradox resisted definitive resolution.

Why it's a real paradox

Start with a sealed gas at uniform temperature T. The second law says ΔS_total ≥ 0; the only way to create a hot/cold gradient should be to do work (compress, refrigerate, etc.). But the demon does no work:

• The door is frictionless and massless.
• Opening and closing takes negligible energy.
• The demon's "vision" is merely observation, which seems energetically free.

And yet the gas develops a gradient. Plug the resulting hot/cold reservoirs into a Carnot engine and you've made a perpetual motion machine of the second kind. Something has to give.

Early proposed resolutions (incomplete)

Year	Proposer	Idea	Verdict
1912	Smoluchowski	Demon itself fluctuates; thermal noise prevents reliable gating	Partial — fails for sufficiently large demons
1929	Szilard	Boiled the paradox down to a 1-molecule engine; identified information storage as crucial	Set the stage; didn't pinpoint the cost
1951	Brillouin	Demon must shine a flashlight to see molecules; photon scattering adds entropy	Partial — relies on a particular measurement scheme
1961	Landauer	Logically irreversible operations dissipate k·T·ln 2 per bit	Foundation of the modern resolution
1982	Bennett	Measurement CAN be reversible; ERASURE is what costs k·T·ln 2 per bit; demon's memory eventually fills	The accepted resolution

Landauer's principle

Logically irreversible operations — any process that takes many input states to fewer output states — must dissipate energy. For the simplest case, erasing 1 bit (mapping 0 or 1 to a fixed 0):

W_erase ≥ k·T·ln 2

At T = 300 K:

k·T ≈ 1.381 × 10⁻²³ J/K × 300 K = 4.14 × 10⁻²¹ J
k·T·ln 2 ≈ 4.14 × 10⁻²¹ × 0.693 = 2.87 × 10⁻²¹ J ≈ 0.018 eV per bit

Equivalent thermodynamic entropy increase per bit: k·ln 2 ≈ 9.57 × 10⁻²⁴ J/K. Erase a full byte at room temperature and you must dump roughly 2.3 × 10⁻²⁰ J as waste heat into the environment.

Landauer's bound was verified experimentally by Bérut et al. in 2012 using a colloidal particle in a double-well optical trap — measured dissipation matched k·T·ln 2 within experimental error.

Bennett's resolution

Charles Bennett's 1982 insight: measurement itself can be made thermodynamically free (reversible) — but the demon must store the measurement results somewhere. As the demon sorts more and more molecules, its memory fills up with bits. Eventually it must reset (erase) the memory to keep operating. Each erasure pays back k·T·ln 2 to the environment.

The accounting closes:

• Per sorting event: demon gains 1 bit of information, gas entropy decreases by k·ln 2.
• Per erasure: demon clears 1 bit, environment entropy increases by ≥ k·ln 2.
• Net: ΔS_total ≥ 0. Second law survives.

The Szilard engine

Leo Szilard distilled the demon down to a single molecule in a box at temperature T:

1. Insert a frictionless partition in the middle. The molecule is on the left or right — 1 bit of information.
2. Measure which side. Attach a weight to the partition on the side WITHOUT the molecule.
3. The molecule pushes the partition isothermally; extract work W = k·T·ln 2 from the heat bath.
4. Remove the partition; molecule fills the full box again. Reset memory; repeat.

Work extracted per cycle: k·T·ln 2. Memory cost per cycle: k·T·ln 2. Net work: zero. Heat is converted to nothing; the second law is exactly saturated.

Costs at human and machine scale

Scale	Energy per bit erased	Multiple of k·T·ln 2 at 300 K
Landauer limit (300 K)	2.87 × 10⁻²¹ J	1
State-of-the-art adiabatic logic (lab)	~10⁻¹⁸ J	~350
5-nm CMOS transistor switch (2024)	~10⁻¹⁵ J	~350 000
DRAM cell write	~10⁻¹² J	~3.5 × 10⁸
HDD bit write	~10⁻⁹ J	~3.5 × 10¹¹
Human brain synapse fire	~10⁻¹³ J	~3.5 × 10⁷

Modern computers run 6 to 10 orders of magnitude above the Landauer bound. There's room for orders-of-magnitude improvement before thermodynamics, rather than engineering, becomes the limit.

JavaScript — Landauer and Szilard arithmetic

const k_B = 1.380649e-23;  // J/K
const ln2 = Math.LN2;

// Landauer minimum energy per bit erased
function landauerErase(T) { return k_B * T * ln2; }

console.log(`Erase 1 bit at 300 K: ${landauerErase(300).toExponential(2)} J`);  // 2.87e-21
console.log(`Erase 1 bit at 4 K (LHe): ${landauerErase(4).toExponential(2)} J`);   // 3.8e-23
console.log(`Erase 1 bit at 4 mK (dil. fridge): ${landauerErase(0.004).toExponential(2)} J`);

// Szilard engine work extracted per cycle (single molecule)
function szilardWork(T) { return k_B * T * ln2; }

// Modern transistor vs Landauer ratio
const E_switch = 1e-15;  // ~1 fJ per CMOS switch
console.log(`Headroom factor: ${(E_switch / landauerErase(300)).toExponential(2)}`);  // ~3.5e5

// Bits erased to dump 1 joule (at 300 K)
const bits_per_joule = 1 / landauerErase(300);
console.log(`Bits per joule at 300 K: ${bits_per_joule.toExponential(2)}`);  // ~3.5e20

// Information vs thermodynamic entropy
// Shannon entropy H bits ↔ thermodynamic entropy k·H·ln 2 J/K
function shannonToThermo(H_bits) { return k_B * H_bits * ln2; }
console.log(`Thermo entropy of 1 GB: ${shannonToThermo(8 * 1e9).toExponential(2)} J/K`);  // ~7.6e-14

// Brain energy efficiency check
// ~10^16 synaptic operations per second, 20 W power
const brain_ops = 1e16;
const brain_W = 20;
const energy_per_op = brain_W / brain_ops;  // 2e-15 J
console.log(`Brain energy per op: ${energy_per_op.toExponential(2)} J  (≈ ${(energy_per_op / landauerErase(310)).toExponential(2)} × Landauer)`);

Where the demon and Landauer show up

• Reversible computing. Bennett and Fredkin showed that logically reversible computation can in principle dissipate zero energy per logical step. Adiabatic CMOS prototypes already operate orders of magnitude below conventional CMOS energy.
• Quantum computing. Unitary gates are reversible — no per-gate Landauer cost. The cost shows up at measurement and at error correction (which throws information away).
• Molecular machines. ATP synthase, kinesin, and bacterial flagella operate at 20-100% efficiency near the thermodynamic optimum; biology found the Landauer scale long before physics did.
• Single-electron devices. Experimental "demons" built with single-electron transistors (Aalto, NIST, IBM) have validated Landauer's bound and Sagawa-Ueda generalizations to within experimental error.
• Black-hole thermodynamics. Bekenstein bound caps the information that can fit inside a region of given area; one bit per 4·ln 2 Planck areas at saturation. Same kT·ln 2 structure.
• Cryptography. Provably one-way functions can be analyzed via the thermodynamic cost of computing them in reverse — connects security to entropy.

Common mistakes

• Claiming "the demon needs energy to look". Measurement CAN be made arbitrarily energy-efficient (Bennett 1973). The cost is in erasure, not observation.
• Conflating Shannon entropy and thermodynamic entropy. They're proportional (S_thermo = k·ln 2 × H_bits) but measured in different units. Don't equate "1 bit" with "1 J/K" — convert.
• Believing modern computers are at Landauer. Real CMOS is 5-6 orders of magnitude above the limit. There's enormous engineering headroom before fundamental physics bites.
• Saying the demon is "impossible". Demons HAVE been built — single-molecule and single-electron versions. They obey the second law because their memory operations dissipate at least the Landauer amount.
• Forgetting that erasure depends on T. k·T·ln 2 scales linearly with environmental temperature. Cryogenic computing dramatically lowers the per-bit limit (3.8 × 10⁻²³ J/bit at 4 K liquid helium).
• Thinking the demon needs to be conscious. The argument is purely physical — any feedback controller storing measurement bits faces the same entropy accounting. Consciousness is irrelevant.