Not everything runs downhill toward lowest energy. In a warm world, a system settles where the tug-of-war between lowering energy and raising entropy comes to rest — and that resting point is written by one quantity: F = U − T·S. At fixed temperature and volume, equilibrium minimizes F, and the Boltzmann distribution is its unique minimizer. Down the center, data flows: the states and temperature go in, the engine minimizes F, the equilibrium comes out. The blue team builds and defends it; the red team tries to break it. Natural units, k = 1.
source Helmholtz, Die Thermodynamik chemischer Vorgänge, Sitzungsber. K. Akad. Wiss. Berlin (1882), pp. 22–39 — the paper that named free energy; twinned with Gibbs (1876). English tr. at informationphilosopher.com AMBER (offprint; no single canonical stable host). Rendered, not quoted.
Three quantities, computed live from the distribution p over the states:
U = ∑ pᵢ Eᵢ — the mean energy (what the system "wants" to lower). S = −∑ pᵢ ln pᵢ — the entropy (what warmth "wants" to raise). F = U − T·S — the Helmholtz free energy. At fixed T and V, equilibrium is the p that minimizes F.
For the current panel setting, the three numbers:
The minimum is not just some number: F* = −kT·ln Z. The free energy is the logarithm of the partition function — energy against entropy, resolved by counting. Each sphere is the next one's premise.
And it has a physical twin one register over: the variational free energy of the ELBO. Minimizing F over distributions = maximizing a lower bound; F* − F[q] = −T·KL(q ‖ p*). Same tug-of-war, learned instead of heated.
The blue team's live check: re-run the minimizer test — is Boltzmann's F ≤ the F of every constructed alternative (uniform, δ-ground, δ-top, seeded randoms)? If red flips the sign, this badge is where it shows.
A system of discrete states with energies Eᵢ = 0, d, 2d, … sits in contact with a heat bath at temperature T. You feed the engine three inputs: how many states, the level spacing d, and T. Everything else — the distribution, U, S, F — is derived, never dialed.
The only freedom the physics allows is the probability pᵢ the system spends on each state. The engine searches that freedom for the F-minimum. That search is the whole game — and it feeds the panel below.
Free energy in use: F = U − T·S. The engine minimizes it — the Boltzmann distribution pᵢ ∝ e−Eᵢ/T.
Live distribution and the F of each candidate — the winner (lowest F) is highlighted:
| candidate distribution | U | S | F |
|---|
Move any slider — U, S, F and the winner are recomputed from the formula on the spot, never looked up.
What the machine produces, proven: at fixed T,V the equilibrium is the Boltzmann distribution, the unique minimizer of F, with F* = −T·ln Z exact and every alternative strictly higher by T·KL ≥ 0. Alongside, the standard battery holds: S = −dF/dT, relaxation lowers F, PV = NkT, the heat kernel spreads as √t with conserved total, equipartition gives ½kT per quadratic mode, Fermi–Dirac stays in [0,1] with a T=0 step.
The blue team's witness (left) confirms Boltzmann minimizes live; the red team (right) tries to make it maximize the wrong thing.
And "F decreases spontaneously" is a statement about the most probable macrostate, not a mechanical law — a fluctuation can raise F transiently; the second law here is statistical. The free energy is not a stored fuel; it is the maximum work extractable at fixed T AMBER (an upper bound, rarely reached).
"Systems settle at lowest energy." Cut. At fixed T,V they minimize free energy F = U − TS, not U. Entropy pulls the other way; only as T→0 does F→U and the ground state win outright.
"Free energy is energy free to do work." Kept, corrected. It is the maximum isothermal work — a ceiling, not a reservoir; real processes fall short by the dissipation.
"Order can't form spontaneously — entropy always rises." Cut. Total entropy (system + bath) rises; the system's F falling lets local order appear while the bath's entropy climbs more.
The red team's move: flip one sign — use F = U + T·S instead of U − T·S. Now "minimizing" rewards low entropy, so the ground-state spike beats Boltzmann and the equilibrium is wrong. The blue team's witness (window 7) is watching.
Flip the entropy sign and the δ-ground distribution (S=0) drops below Boltzmann — the witness recomputes, sees Boltzmann is no longer the minimizer, and turns red. Nothing is faked; the attack is real and it is caught.