The engine a fluctuation cannot run

The molecules of a gas at room temperature are moving at hundreds of metres a second. A paddle wheel immersed in them is struck constantly, from every direction, and jiggles. The jiggling is real and has been watched since Brown looked at pollen grains, and the jiggle that proved atoms is the argument that it comes from the molecules.

Now attach the paddle’s shaft to a ratchet and pawl — the mechanism inside a socket wrench, which allows rotation one way and blocks it the other. The paddle is kicked both ways, the ratchet keeps only the kicks that turn it forwards, and the shaft winds up a thread that lifts a weight. A single reservoir of heat has been turned into work, forever, using nothing but a piece of geometry.

It does not work. Following exactly why it does not is more useful than any statement of the second law, because the failure is mechanical and can be computed.

Two rates, and why they are equal #

The whole argument is a comparison of two probabilities.

A forward click requires a collision on the paddle energetic enough to lift the pawl out of its notch — an energy $\varepsilon$ — and, if a weight is attached, enough on top of that to raise it by one tooth. The chance of a collision that energetic is the Boltzmann factor $e^{-(\varepsilon+L)/kT_1}$, with $T_1$ the temperature of the gas the paddle is in. That is the exponential that decides everything applied to the paddle.

A backward click requires something different: the pawl itself must lift out of its notch, by a fluctuation of its own thermal motion, at which moment the wheel is free to slip back one tooth. The chance of that is $e^{-\varepsilon/kT_2}$, with $T_2$ the temperature of whatever the pawl is in contact with.

If the two temperatures are the same, the two rates are the same. Not approximately, not nearly: the same $\varepsilon$ appears in both, at the same $T$, and the exponentials are identical. The forward and backward clicks happen equally often and the wheel goes nowhere.

The figure makes the stronger statement that follows. Attach a load and the forward rate falls — a forward click now needs $\varepsilon + L$ — while the backward rate does not change at all, because slipping back releases the load rather than lifting it. So the device runs backwards under any load whatever, lowering the weight and warming the gas.

Why the pawl had to be warm #

Every attempt to rescue the machine fails in the same way, and the pattern is worth having.

Make the notch deeper, so the pawl is harder to lift. Both rates fall by the same exponential factor and their ratio is unchanged. The figure computes this for notches from two to twenty thermal energies deep: everything gets slower and nothing changes sign.

Make the pawl spring stiffer. Same argument. Add a second pawl, or a hundred. Same argument, since each one has its own fluctuations. Make the pawl heavier so it responds sluggishly, and the response time changes without touching the equilibrium probabilities that set the rates.

The trap is imagining the pawl as a rigid rule and the gas as the only warm thing. A pawl light enough to be lifted by a molecular collision is light enough to be lifted by its own thermal motion, and those are not two facts but one — the same $\varepsilon$ measured against the same $kT$. A rectifier of thermal noise has to be small enough for the noise to move it, and anything that small is itself noisy.

That is a statement about mechanisms rather than about entropy, and it is the reason this device is the standard answer to the question “yes, but what actually stops it?”

The size at which the question arises #

It is worth putting numbers on “small enough for thermal noise to matter”, because the answer is what decides whether any of this is a real concern or a story about an imaginary machine.

The thermal energy at room temperature is $4 \times 10^{-21}$ joules. A pawl that a molecular collision can lift must therefore have a barrier of that order, and a spring with a barrier that low, made of anything, is a few nanometres across. At that size the pawl is jumping out of its notch spontaneously many times a second, and the device is a blur.

Scale the pawl up to something visible, say a barrier of a hundred thermal energies, and the spontaneous lifting becomes essentially impossible — the Boltzmann factor is $e^{-100}$, one event in $10^{43}$. But the forward rate has fallen by exactly the same factor, so the paddle can no longer lift the pawl either, and the device does nothing at all rather than doing nothing on average.

Between the two sizes there is no window, and that is the point that makes the argument airtight rather than merely suggestive. The same exponential governs whether the ratchet can be driven and whether it can slip, so making slipping rare makes driving rare in exactly the same proportion. There is no design that is large in one sense and small in the other.

What makes it an engine #

Now separate the two temperatures. Put the paddle in a hot gas and the pawl in contact with something cold.

The exponentials no longer cancel, and the device runs. Equating the two rates gives the load it will just hold:

$$L_{\text{stall}} = \varepsilon\left(\frac{T_1}{T_2} - 1\right)$$

which is zero when the temperatures are equal — the null result, recovered — and grows in proportion to the excess.

The mechanism has been demoted. In the one-temperature version the ratchet was supposed to be doing the work of rectification; here it merely picks a direction, and the engine is the temperature difference. That demotion is general. No arrangement of parts extracts work from one reservoir; every engine is two reservoirs and something to shuttle between them, and the something can be as crude as a notched wheel.

Carnot, from two exponentials #

The efficiency at stall falls out of the same balance with no further physics.

Each forward click takes $\varepsilon + L$ from the hot gas, delivers $\varepsilon$ to the cold pawl when it drops back into its notch, and puts $L$ into the weight. So the efficiency is $L/(\varepsilon + L)$, and at the stall load that is exactly $1 - T_2/T_1$.

Getting Carnot’s ceiling out of a notched wheel and two Boltzmann factors is a genuinely surprising result, and it was Feynman’s. The ceiling on every engine derives the same number from a cycle of reversible steps and a definition of temperature; this derives it from a mechanism so crude that it has one moving part.

It is also, as stated, wrong — and the way it is wrong is instructive.

Why a real one cannot reach it #

The efficiency reaches Carnot’s ceiling only at stall, where nothing is moving, which is the usual price of reversibility. That much is expected. The problem is that this particular device cannot get there even in that limit.

The pawl and the paddle are joined by a shaft. The shaft conducts heat. So the hot bath and the cold bath are in permanent thermal contact through the very structure that makes the machine a machine, and heat flows from one to the other whether or not the wheel turns.

At stall, that leak is the entire heat flow and the work output is zero, so the efficiency is zero. Away from stall the leak is a fixed drain on the input, so the efficiency peaks somewhere in the middle and falls short of the ceiling by an amount set by the leak. The second curve in the figure is that statement.

This is not a detail of the ratchet. Any engine whose two reservoirs are connected by the same structure that transfers the work has the same problem, and the fix is always the same: make the working path good and the parasitic path bad. In a real engine that means insulation. In a device the size of a molecule it is close to impossible, because everything is in contact with everything and the same degrees of freedom carry the work and the leak.

Feynman’s analysis assumed the two operations could be separated. Later work pointed out that the device as described exchanges heat with both baths simultaneously and is therefore intrinsically irreversible. The correction does not touch the null result at one temperature, which is the part the argument was built to make, and it removes the claim of reversibility.

The load a real one would carry #

That shape — zero at both ends, a maximum between — is the shape of every engine, and it is why the engine that has to finish is a separate subject from the ceiling. Running at maximum efficiency means running infinitely slowly and delivering nothing. Running at maximum power means accepting a definite waste. A design chooses between them according to whether heat or time is the scarce thing, and the ratchet makes the choice unusually visible because both ends of the curve are places the device genuinely sits.

The bookkeeping, click by click #

The efficiency argument is short enough to write out in full, and doing so shows where each quantity goes.

A forward click: the gas gives up $\varepsilon + L$ to the paddle, the pawl rides up over the tooth and drops back, delivering $\varepsilon$ to whatever the pawl is in contact with, and the weight rises by $L$. Energy is conserved and two reservoirs have been touched.

A backward click: the pawl’s own bath gives it $\varepsilon$, the wheel slips back, the weight falls by $L$, and $\varepsilon + L$ arrives in the gas. The same three terms with every sign reversed.

At stall the two clicks happen equally often, so every transfer is undone as fast as it is made and the net flows are all zero. That is the definition of a reversible operating point, and it is why the efficiency there is Carnot’s — the derivation has not assumed reversibility anywhere, it has arrived at it. Away from stall the forward clicks outnumber the backward ones, heat flows steadily from the hot bath to the cold one, and a fraction of it is delivered as work.

Everything in that account is a statement about single events rather than about a cycle. There is no working substance, no compression, no expansion, and nothing that has to be done slowly for its own sake — which is what makes the recovery of the classical ceiling from it worth noticing.

Where this argument reaches #

The ratchet is a thought experiment and the conclusion is not.

Molecular motors are ratchets that work, and they work by being driven. A kinesin molecule walking along a filament is rectifying thermal motion — its steps are thermally activated and it is far too small for anything else — and what makes the rectification legitimate is that each step consumes a molecule of ATP, which is a chemical reservoir far from equilibrium. Take the ATP away and the motor diffuses, exactly as this figure says it must.

A flashing ratchet makes the same point in the laboratory: a sawtooth potential switched on and off transports particles along it, reliably and measurably. It is not a violation, because switching the potential costs work, and the work bought is less than the work spent.

And a diode does not rectify its own thermal noise. A junction at one temperature produces no net current across a resistor, however nonlinear its characteristic, which is the electrical version of this exact argument and was settled by the same reasoning. A diode connected to an antenna in sunlight does produce power, because the sunlight is a second reservoir at six thousand kelvin.

The pattern in all three is the one the figures make: the bit that has to be paid for reaches it from the side of information, and this reaches it from the side of mechanism, and the two arrive at the same place. Rectification is free only where there is something to rectify, and thermal noise at one temperature has no direction in it to keep.

Smoluchowski, thirty years earlier #

The device is usually attributed to Feynman’s lectures of 1961, and the argument was made by Smoluchowski in 1912. The attribution matters because the two treatments answer different questions and the older one is the sharper.

Smoluchowski’s point was that no automatic mechanism can violate the second law, and he made it by systematically going through the ways one might try: a valve that opens one way, a trapdoor that a fast molecule can push through, a ratchet. Each fails for the same reason, and he stated the reason in general — a mechanism small enough to respond to fluctuations undergoes them, so it cannot serve as a fixed reference against which they are sorted. His conclusion was that only an intelligence could do the sorting, which left the door open that the bit that has to be paid for eventually closed.

Feynman’s contribution was to run the device with two temperatures and get the efficiency, which converts a negative result into a positive one. That is why the ratchet is remembered as his: the null case had been settled, and what he added was the engine.

Where the model stops #

The rates are Arrhenius factors and nothing else. A real device has a rate prefactor that depends on the shape of the barrier, the damping and the mass, and none of that appears here. It cancels out of everything the figures claim — the null result, the stall load, the stall efficiency — because it multiplies both rates equally, which is exactly why the argument can be made without it.

One tooth is treated as one event. The wheel is assumed to move by whole teeth and to be sitting still between them. A shaft that is being kicked continuously does not do that, and a careful treatment is a Langevin equation in a periodic potential rather than a two-state hopping model.

The leak is put in by hand. Its size in a real device would come from the shaft’s geometry and conductivity, and the figure uses one number to show what any leak does rather than to predict a particular one.

And the second law is not proved here. What the figures show is that this particular mechanism fails, in a way that generalises to every mechanism of the same kind. Proving that no mechanism can work is a different argument, and the second law with a probability attached is where its modern form lives — with the exceptions counted rather than denied.

What the pictures cannot show #

None of the figures shows the wheel. They show rates, loads and efficiencies, all of them averages over many clicks, and the actual motion of a device this small is a stuttering random walk with a slight bias — many clicks each way, a small excess one way, and no visible steadiness at all. A figure of the trajectory would show something that looks like noise, which is what it is.

Nor do the figures show time. Every curve is a steady-state rate, and the approach to that steady state — the transient after the load is attached — is where the fluctuation theorems live, and where a small device can be watched running briefly backwards.

Where the ladder goes next #

The heat-engine ladder began with the ceiling on every engine, set before any engine was designed, went on to the heat pump that pays back more than it takes, and then to the engine that has to finish, where power rather than efficiency is what is being asked for. This rung takes the smallest engine anyone has proposed and finds that it obeys the same ceiling and cannot reach it.

The rung after it is the engine small enough that its output fluctuates — where the work delivered in one cycle is a random variable that is sometimes negative, and the second law becomes a statement about a distribution rather than about a number. The habit worth carrying is the one this rung is built on: when a device seems to break a law, do not look for the law, look for the part of the device that was quietly assumed to be at zero temperature.