Statistical Mechanics
Landauer's Principle: The Minimum Energy Cost of Erasing a Bit
Erase a single bit of information and physics charges you a toll of at least kBT ln 2 — about 2.9 × 10⁻²¹ joules (0.018 eV) at room temperature, comparable to (about half of) the average thermal energy of a single air molecule. That number, derived by Rolf Landauer at IBM in 1961, is the deepest bridge yet found between abstract information theory and thermodynamics: it says that logical irreversibility — collapsing two possible states into one — necessarily produces entropy in the physical world.
Landauer's principle states that any process which erases one bit of information, resetting a two-state memory to a definite value, must dissipate no less than kBT ln 2 of heat into the environment at temperature T. Computation itself can in principle be free; only the throwing-away of information carries an unavoidable energetic price.
- RegimeThermodynamics of computation; classical & quantum bits
- Key relationQ ≥ k_B T ln 2 per bit erased
- DiscoveredRolf Landauer, IBM, 1961
- Characteristic scale≈ 2.9 × 10⁻²¹ J (0.018 eV) at T = 300 K
- Realized inColloidal double-well trap (Bérut et al. 2012); nanomagnets; single spins
- Matters forUltimate limits of computing, Maxwell's demon, reversible logic
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
What it is and why it matters
Landauer's principle draws a sharp line between two kinds of information processing. Logically reversible operations — a NOT gate, a controlled-NOT, copying data onto blank memory — map the state space one-to-one and can, in principle, be run at zero thermodynamic cost. Logically irreversible operations — erasing a bit, or an AND gate whose two-bit input cannot be recovered from its one-bit output — merge distinct logical states into one. Landauer's insight, published in the IBM Journal of Research and Development in 1961, is that this merging cannot be free: it must dump at least kBT ln 2 of heat per lost bit into the surroundings.
The claim matters because it makes "information is physical" quantitative. It sets a hard floor beneath the energy budget of all computation, it resolves the century-old Maxwell's-demon paradox, and it motivates the entire field of reversible computing. Crucially, it locates the cost not in computing but in forgetting — a subtle and counterintuitive relocation of where thermodynamic irreversibility actually lives.
The mechanism, step by step
Model a bit as a particle in a symmetric double-well potential: left well = 0, right well = 1. Before erasure the particle could be in either well with equal probability, so its positional entropy includes one bit, kB ln 2. "Erase" means force the particle into the 0 well regardless of where it started — a two-state ensemble is compressed to one state.
The information-bearing degree of freedom therefore loses entropy ΔSsys = −kB ln 2. But the second law forbids the total entropy of system-plus-bath from decreasing. The bath must absorb at least the difference: ΔSbath ≥ kB ln 2, which at temperature T means heat Q = T·ΔSbath ≥ kBT ln 2 flows out. Equivalently, an isothermal quasi-static compression of the accessible phase-space volume by a factor of 2 costs exactly kBT ln 2 of work — the informational analogue of compressing a one-molecule gas to half its volume. The key is that erasure is many-to-one: it shrinks state-space volume, and shrinking volume isothermally always costs work that leaves as heat.
The key equation and characteristic numbers
The principle is a single inequality:
Q ≥ kBT ln 2 (heat dissipated per bit erased),
with equality approached only in the quasi-static (infinitely slow) limit; finite-speed erasure dissipates strictly more. Plugging in kB = 1.381 × 10⁻²³ J/K, T = 300 K, and ln 2 = 0.693 gives
kBT ln 2 ≈ 2.87 × 10⁻²¹ J ≈ 0.018 eV = 0.693 kBT (since ln 2 = 0.693) — about 3 zeptojoules.
That is roughly 1/40 of an electron-volt, comparable to a single quantum of thermal noise. For scale: erasing a mole of bits (6 × 10²³) costs only ~1.7 kJ. Today's CMOS transistors dissipate ~10⁻¹⁶ J per switch — four to five orders of magnitude above the Landauer floor — so the bound is not yet a practical wall, but it is the asymptote that Moore's-law energy scaling is drifting toward. A generalized form links the exact heat to the memory's Shannon entropy change, Q ≥ kBT·(Hinitial − Hfinal)·ln 2.
How it was measured
The landmark confirmation came from Bérut, Arakelyan, Petrosyan, Ciliberto, Dillenschneider and Lutz, published in Nature (483, 187, 2012). They trapped a single ~2 μm silica colloidal bead in a double-well optical potential formed by two focused laser foci — a physical bit. By modulating the trap (lowering the central barrier, then tilting to push the bead into one well) they ran a full erasure cycle while tracking the bead's Brownian trajectory and computing the dissipated heat via the stochastic-thermodynamics work integral ∫F·dx.
As the cycle was slowed toward quasi-static, the mean dissipated heat saturated at exactly kBT ln 2, with faster cycles landing above it — a clean verification of both the bound and its approach. Complementary tests followed: sub-Landauer-noise erasure in feedback traps, single-molecule and nanomagnetic-bit erasure (a 2016 experiment measured ~4.2 × 10⁻²¹ J, ~44% above the floor), and single-spin/quantum-dot platforms probing the quantum regime. The signature is universal: heat versus erasure speed extrapolates to the kBT ln 2 asymptote.
Where it operates and what it is not
The principle applies to any physical system storing distinguishable logical states in contact with a thermal bath at temperature T — colloids, nanomagnets, superconducting circuits, trapped ions, spins, and biological molecules alike. It is fundamentally a restatement of the second law applied to information-bearing degrees of freedom, so it is as universal as thermodynamics itself.
Several distinctions matter. Landauer's bound is not a cost of computation per se: reversible operations (bit-flips, copying to blank memory) carry no minimum cost, a point Charles Bennett established in 1973 by showing any computation can be made logically reversible. It is not the same as switching energy in a real transistor, which is dominated by charging capacitors far above kBT. And it differs from the Margolus–Levitin and Bremermann bounds, which limit computational speed per unit energy rather than the heat of erasure. Finally, some philosophers (Earman, Norton) argue the principle is a corollary of the second law rather than an independent axiom — a debate about status, not about the number itself.
Applications, open questions, and significance
Landauer's principle is the resolution of Maxwell's demon: the demon can extract kBT ln 2 of work per Szilard-engine cycle by measuring a molecule, but to close the cycle it must eventually erase its own memory, repaying exactly kBT ln 2 and rescuing the second law. This "exorcism" (Bennett, Penrose) shifts the entropy debt from measurement to erasure.
Practically, it underwrites reversible and adiabatic computing, quantum computation (where unitary gates are logically reversible and Landauer-free until measurement/reset), and the long-run energy roadmap for data centers approaching thermal limits. Open questions remain active: the tightest bounds under finite-time and finite-error erasure (a speed–dissipation–reliability trade-off), fully quantum generalizations including coherence and non-Markovian baths, whether sub-kBT ln 2 erasure is achievable by exploiting non-thermal or squeezed reservoirs, and how the principle constrains the energetics of biological information processing and even black-hole information. More than sixty years on, it remains the clearest quantitative statement that information has a thermodynamic weight.
| Operation / limit | Minimum energy cost | Logically reversible? | Key point |
|---|---|---|---|
| Erase one bit (RESET to 0) | k_B T ln 2 ≈ 2.9×10⁻²¹ J at 300 K | No | Entropy of memory drops by k_B ln 2; heat k_B T ln 2 expelled |
| Copy a bit onto blank tape | 0 (in principle) | Yes | Bennett: copying preserves information, no minimum cost |
| NOT / bit-flip | 0 (in principle) | Yes | One-to-one map on state space; no state merging |
| AND / OR gate (2 in → 1 out) | ≥ k_B T ln 2 per bit lost | No | Many-to-one; discards input information |
| Szilard engine, one cycle | Extracts k_B T ln 2 as work | — | Demon's memory reset repays exactly the Landauer cost |
| Typical CMOS switch (2020s) | ~10⁴–10⁵ k_B T ln 2 | No | Real devices run ~10,000–100,000× above the Landauer floor |
Frequently asked questions
Does Landauer's principle mean all computation must dissipate energy?
No — only logically irreversible operations do. Erasing a bit or an AND gate that merges two input states into one must dissipate at least k_B T ln 2. Logically reversible operations (NOT, controlled-NOT, copying to blank memory) preserve information and can in principle run at zero thermodynamic cost, which is the foundation of reversible computing.
Why exactly k_B T ln 2 and not some other value?
Erasing a bit compresses a two-state ensemble to one state, reducing the memory's entropy by k_B ln 2 (the ln 2 comes from log of 2 accessible states). The second law forbids total entropy from falling, so the bath must gain at least k_B ln 2, and at temperature T that entropy carries heat Q = T·ΔS = k_B T ln 2. It is the informational analogue of isothermally halving a gas's volume.
Has Landauer's principle actually been verified experimentally?
Yes. Bérut et al. (Nature, 2012) used a single colloidal particle in an optical double-well trap and showed the mean dissipated heat saturates at k_B T ln 2 in the slow (quasi-static) limit, with faster erasure dissipating more. Later experiments on nanomagnets, single molecules, single spins and quantum dots reproduced the bound across platforms.
How does Landauer's principle exorcise Maxwell's demon?
A Szilard-engine demon extracts k_B T ln 2 of work per cycle by measuring which half of a box a molecule is in. But the measurement stores one bit in the demon's memory; to run a cyclic engine the demon must eventually reset that memory, and Landauer's principle says erasure costs exactly k_B T ln 2. The net work over a full cycle is therefore zero or negative, saving the second law. Bennett formalized this in 1982.
How far are real computer chips from the Landauer limit?
Very far — modern CMOS transistors dissipate roughly 10⁻¹⁶ J per switching event, about four to five orders of magnitude above the ~2.9×10⁻²¹ J Landauer floor at room temperature. The bound is thus not yet a practical constraint, but it is the ultimate asymptote that device energy scaling is slowly approaching, motivating reversible and adiabatic logic research.
Is the Landauer bound the same as the minimum switching energy of a transistor?
No. The Landauer bound is the minimum heat from destroying information (k_B T ln 2). A transistor's switching energy is set mostly by charging its gate capacitance to a voltage well above thermal noise, which is far larger and not fundamentally required by information erasure. You could in principle switch a transistor reversibly with far less dissipation; the Landauer floor only bites when information is actually thrown away.