Detailed Balance

Definition

Consider a system that hops between discrete states labelled i, j, k, … Its probabilities evolve under the master equation:

dP_i/dt = Σ_j [ P_j·W(j→i) − P_i·W(i→j) ]

where P_i is the probability of finding the system in state i and W(i→j) is the transition rate from i to j (units s⁻¹ for continuous time, or a dimensionless probability per step in a Markov chain). The bracket is the net probability current on the link i↔j.

A distribution is stationary when dP_i/dt = 0 for all i — the total flux into each state balances the total flux out. Detailed balance demands something far stronger: that each individual term of the sum vanishes.

P_i · W(i→j) = P_j · W(j→i)      for every pair (i, j)

Every forward flux is cancelled by its reverse flux, one link at a time. There is no net circulation anywhere in the state graph. This is the condition that characterises genuine thermodynamic equilibrium, as opposed to a merely stationary nonequilibrium steady state.

Symbols and units

Symbol	Meaning	Units
P_i	Equilibrium probability of state i	dimensionless (Σ P_i = 1)
W(i→j)	Transition rate i to j	s⁻¹ (continuous) or per-step probability
J_ij	Net probability current = P_i W(i→j) − P_j W(j→i)	s⁻¹
E_i	Energy of state i	J (or eV)
k_B	Boltzmann constant	1.380649 × 10⁻²³ J/K
T	Absolute temperature	K
β	Inverse temperature = 1/(k_B·T)	J⁻¹

Why it keeps the Boltzmann distribution stationary

The canonical equilibrium distribution is Boltzmann's:

P_i = (1/Z) · exp(−E_i / (k_B·T)),    Z = Σ_i exp(−E_i / (k_B·T))

Take the ratio of two equilibrium probabilities — the partition function Z cancels:

P_j / P_i = exp(−(E_j − E_i) / (k_B·T))

Detailed balance rearranges to a constraint purely on the rates:

W(i→j) / W(j→i) = P_j / P_i = exp(−(E_j − E_i) / (k_B·T))

Any microscopic dynamics whose rates satisfy this ratio automatically balances every link when P is Boltzmann. And because detailed balance implies stationarity, the Boltzmann distribution is then a fixed point: dP_i/dt = 0. This is the deep reason thermalisation works — a heat bath supplies transition rates obeying exactly this ratio, and the system inexorably relaxes to P_i ∝ e^(−βE_i). Notice the argument never required knowing Z, only energy differences.

Detailed balance vs. global balance

The distinction is best seen on a three-state cycle A → B → C → A. Suppose the clockwise rates all equal k_+ and the counter-clockwise rates all equal k_−.

Property	Global balance (stationary)	Detailed balance (equilibrium)
Condition	Σ_j (P_j W_{ji} − P_i W_{ij}) = 0	P_i W_{ij} = P_j W_{ji} per link
Net link current J_ij	May be nonzero (circulating)	Exactly zero everywhere
Three-state cycle	Uniform P = 1/3 is stationary for any k_±	Requires k_+ = k_− (no drift)
Entropy production rate	≥ 0, positive if current flows	Exactly 0
Time-reversal	Generally broken	Statistically symmetric
Free-energy input	Needed to sustain current	None (closed, isolated)

For the cycle with uniform P = 1/3, global balance is satisfied for any k_+ and k_−: each state receives and emits the same total. But the net current around the loop is J = (1/3)(k_+ − k_−), which is nonzero unless k_+ = k_−. So the uniform distribution can be a perfectly good nonequilibrium steady state that circulates probability endlessly — and detailed balance is precisely the extra condition that outlaws it. Kolmogorov's criterion captures this: detailed balance holds if and only if the product of forward rates around every closed loop equals the product of reverse rates.

Worked example — the Metropolis algorithm

The single most consequential application is Markov-chain Monte Carlo. To sample configurations with Boltzmann weights P_i ∝ e^(−βE_i) — say the spins of an Ising model — you build a Markov chain whose stationary distribution is Boltzmann, by engineering detailed balance. The Metropolis recipe (Metropolis, Rosenbluth, Rosenbluth, Teller & Teller, J. Chem. Phys. 1953):

1. From configuration i, propose a symmetric move to j (e.g. flip a random spin): proposal g(i→j) = g(j→i).
2. Compute ΔE = E_j − E_i.
3. Accept the move with probability A(i→j) = min(1, e^(−βΔE)). If ΔE ≤ 0 always accept; if ΔE > 0 accept with probability e^(−βΔE).

The effective rate is W(i→j) = g(i→j)·A(i→j). Check detailed balance for E_j > E_i:

P_i · W(i→j) = e^(−βE_i) · g · e^(−β(E_j−E_i)) = g · e^(−βE_j)
P_j · W(j→i) = e^(−βE_j) · g · 1                = g · e^(−βE_j)   ✓ equal

Detailed balance holds by construction, so the Boltzmann distribution is the unique stationary distribution of an ergodic chain — and long-run averages of any observable converge to thermal expectation values. The Metropolis–Hastings generalisation (Hastings, 1970) removes the symmetric-proposal requirement by using A = min(1, [P_j g(j→i)] / [P_i g(i→j)]), restoring detailed balance for asymmetric proposals. The whole enterprise sidesteps the exponentially large partition function Z entirely.

// Metropolis single-spin update obeying detailed balance
const k_B = 1.380649e-23;   // J/K (use reduced units J/k_B = 1 below)

// Reduced units: beta = 1/T, energies in units of coupling J
function metropolisStep(spins, T, coupling = 1) {
  const N = spins.length;
  const beta = 1 / T;
  const i = Math.floor(Math.random() * N);

  // Energy cost of flipping spin i (1D chain, periodic)
  const left  = spins[(i - 1 + N) % N];
  const right = spins[(i + 1) % N];
  const dE = 2 * coupling * spins[i] * (left + right);  // E_j - E_i

  // Acceptance A = min(1, exp(-beta*dE)) — this line encodes detailed balance
  if (dE <= 0 || Math.random() < Math.exp(-beta * dE)) {
    spins[i] = -spins[i];   // accept the flip
  }
  return spins;
}

// Verify the rate ratio equals the Boltzmann ratio for dE = +2 at T = 1
const dE = 2, T = 1;
const acceptRatio = Math.exp(-dE / T);           // W(i->j)/W(j->i)
const boltzRatio  = Math.exp(-(dE) / T);         // P_j/P_i
console.log(acceptRatio.toFixed(4), boltzRatio.toFixed(4)); // 0.1353 0.1353 — detailed balance ✓

Detailed balance and time-reversal symmetry

Detailed balance is the statistical shadow of microscopic reversibility. Newton's equations, Maxwell's equations, and the Schrödinger equation are all invariant under t → −t (for time-reversal-even fields). At equilibrium this symmetry surfaces as detailed balance: a recorded movie of the equilibrium fluctuations, played backwards, is statistically indistinguishable from the forward movie, because for every observed transition i→j the reverse j→i occurs equally often. Formally, the equilibrium two-time correlation obeys ⟨A(0)B(t)⟩ = ⟨A(t)B(0)⟩.

From this reversibility Lars Onsager derived his reciprocal relations (1931, Nobel Prize 1968): the matrix of linear transport coefficients is symmetric, L_ij = L_ji — so, for instance, the thermoelectric cross-coefficient linking a temperature gradient to charge flow equals the one linking a voltage to heat flow. When time-reversal symmetry is broken — by an external magnetic field B or a Coriolis (rotating-frame) force — detailed balance in its simple form can fail, and one uses the generalised Onsager–Casimir relations L_ij(B) = L_ji(−B).

The catch — nonequilibrium steady states

A stationary distribution is not automatically an equilibrium one. A nonequilibrium steady state (NESS) has constant probabilities yet carries persistent circulating currents J_ij ≠ 0 — detailed balance is broken. Such states are ubiquitous in the living world: a kinesin motor stepping along a microtubule, the citric-acid cycle, or an ion channel rectifying flow all cycle preferentially in one direction, powered by ATP hydrolysis (≈ 20 k_B·T ≈ 50 kJ/mol of free energy per molecule at physiological conditions). Any system wired between two reservoirs at different temperatures or chemical potentials sits in a NESS. These states produce entropy at a constant positive rate σ = Σ J_ij·ln[P_i W_{ij} / (P_j W_{ji})] > 0, whereas true equilibrium (detailed balance) has σ = 0 exactly. The circulating current is the microscopic signature of the free energy being continuously dissipated.

Common misconceptions

• "Stationary means equilibrium." No. Stationarity is global balance; equilibrium is detailed balance. Every equilibrium is stationary, but a stationary state can circulate current (a NESS) and violate detailed balance.
• "Detailed balance needs the partition function." Only the rate ratio W(i→j)/W(j→i) = e^(−βΔE) matters, and it depends only on energy differences. Z cancels — which is exactly why Monte Carlo can sample without ever computing Z.
• "Any rates that thermalise obey detailed balance." Not necessarily. You can construct rates (e.g. a biased cycle) that still relax to the Boltzmann distribution as a stationary state while violating detailed balance link-by-link. Detailed balance is sufficient for Boltzmann stationarity, not strictly necessary.
• "Detailed balance always holds in equilibrium." It holds for time-reversal-symmetric microscopic dynamics. In a magnetic field or rotating frame the simple form is replaced by the Onsager–Casimir version; naive detailed balance can be violated even at equilibrium.
• "It's just a bookkeeping identity." It is a physical statement about the absence of net microscopic currents and the vanishing of entropy production — a sharp, testable dividing line between equilibrium and driven systems.
• "Metropolis needs symmetric proposals." Plain Metropolis does; Metropolis–Hastings restores detailed balance for asymmetric proposals by folding the proposal ratio g(j→i)/g(i→j) into the acceptance probability.