Detailed Balance and Microscopic Reversibility

Equilibrium asserts far more than the constancy of macroscopic observables: it asserts the absence of any systematic flow at the microscopic level. Detailed balance makes this precise. For a Markov dynamics with transition rates $W(a\to b)$ between states, detailed balance demands that in the equilibrium distribution $P_{\mathrm{eq}}$ the probability flowing from $a$ to $b$ is exactly cancelled by the probability flowing back — separately, for every ordered pair of states. This is strictly stronger than mere stationarity, which requires only that the net flow into each state vanish. A system can be perfectly stationary yet circulate probability forever around loops in state space; detailed balance forbids that circulation entirely. The condition wears two hats. As a statement about Markov chains it guarantees that $P_{\mathrm{eq}}$ is stationary and, symmetrising the generator in a $P_{\mathrm{eq}}$ -weighted inner product, that all relaxation eigenvalues are real — equilibrium dynamics relaxes monotonically, never with persistent oscillation. As a statement about physics it is the coarse-grained shadow of microscopic reversibility: because the microscopic laws are invariant under time reversal, the rate of a forward transition equals that of its time-reverse, and this symmetry propagates up to the transition rates of any faithful reduced description. Detailed balance is exactly the line that separates equilibrium from nonequilibrium. A stationary distribution can be sustained by a steady cycle of currents — a molecular motor driven by an ATP reservoir, a colloid in a rotating force field, an electrical network held off equilibrium. Such nonequilibrium steady states satisfy the master equation in steady state but violate detailed balance: nonzero loop currents circulate and entropy is produced at a finite rate. Kolmogorov's cycle criterion turns this into a checkable test using rates alone. Practically, detailed balance is the design principle of Markov-chain Monte Carlo: to sample the Boltzmann distribution one engineers rates that balance with respect to $e^{-\beta E}$ , and the Metropolis rule is the minimal such construction — the subject of `aq-sm-metropolis-hastings-algorithm`, which recalls this article as its lemma.

Key takeaways

Detailed balance is pairwise flux cancellation, $P_{\mathrm{eq}}(a)\,W(a\to b)=P_{\mathrm{eq}}(b)\,W(b\to a)$ : every microscopic transition is exactly balanced by its reverse in equilibrium.
It is sufficient (not necessary) for $P_{\mathrm{eq}}$ to be stationary under the master equation — stationarity alone permits steady circulating currents (nonequilibrium steady states).
Kolmogorov's cycle criterion tests detailed balance using rates only: the product of forward rates around every closed loop must equal the product of reverse rates.
Microscopically it follows from time-reversal symmetry of the underlying dynamics (microscopic reversibility): forward and time-reversed transition rates are equal.
Under detailed balance the master-equation generator is symmetrisable, so relaxation eigenvalues are real and non-positive — monotone, non-oscillatory relaxation.
Key idea: Detailed balance is pairwise flux cancellation, not merely global stationarity. It is sufficient for equilibrium, equivalent to zero loop currents (Kolmogorov), and follows from microscopic time-reversal symmetry.

The core insight

Detailed balance is pairwise flux cancellation, $P_{\mathrm{eq}}(a)W(a\to b)=P_{\mathrm{eq}}(b)W(b\to a)$ , not merely global stationarity. It is sufficient for equilibrium, equivalent to zero loop currents (Kolmogorov), and follows from microscopic time-reversal symmetry.

Where this lives today

Detailed balance is the quiet workhorse behind an enormous amount of computation and measurement. Every equilibrium Markov-chain Monte Carlo method — from lattice QCD to condensed-matter Ising simulations to Bayesian posterior sampling — is built by imposing detailed balance with respect to the target weight; the Metropolis and Glauber/heat-bath rules are the standard constructions, and cluster algorithms achieve the same balance with cleverly correlated non-local moves to defeat critical slowing down. In the opposite direction, breaking detailed balance is now a first-class research topic: molecular motors, active matter, and driven colloids are stationary but non-equilibrium, and the Schnakenberg entropy production rate together with the failure of Kolmogorov's criterion is the operational signature of the drive. Modern non-reversible samplers deliberately violate detailed balance (while preserving stationarity) to mix faster — a controlled relaxation of the classical requirement that has become a practical tool in computational statistics.

Origins · 1 minmicroscopic reversibility and how this was discovered

microscopic reversibility

principle, 19th–20th c. · The symmetry of the microscopic equations of motion under time reversal, so that the rate of any elementary process equals the rate of its time-reverse. Traced to Boltzmann and Maxwell in kinetic theory and formalised by Tolman as the "principle of microscopic reversibility," it is the physical origin of detailed balance: coarse-graining time-reversible mechanics yields balanced transition rates, and its breaking marks a system as driven out of equilibrium.

History

1872

— Boltzmann's H-theorem uses collisional balance to prove approach to Maxwell–Boltzmann

1901

— Wegscheider states cyclic conditions on reaction rates (chemical detailed balance)

1923

— Fowler and Milne apply detailed balance to stellar spectra and ionisation

1925

— Tolman formulates the principle of microscopic reversibility

1931

— Onsager's reciprocal relations rest on microscopic reversibility

1936

— Kolmogorov gives the cycle criterion for reversibility of Markov chains

1953

— Metropolis et al. engineer detailed balance to sample the Boltzmann distribution

1976

— Schnakenberg expresses entropy production of a Markov network via cycle currents

The physical picture

Picture state space as a graph: nodes are configurations, and each directed edge carries a probability flux $J_{a\to b}=P(a)\,W(a\to b)$ . Stationarity is the statement that the net flux into every node is zero — like an electrical circuit in steady state, where currents may still circulate around loops. Detailed balance is the stronger, quieter statement that each individual edge carries zero net current — like a circuit with no current flowing anywhere. Equilibrium is the second condition; a nonequilibrium steady state satisfies the first but not the second. The distinction between global balance and detailed balance is the whole story, and the punchline figure makes it visible: a three-state ring with clockwise-biased rates is genuinely stationary yet circulates probability forever, while a detailed-balanced network has every link at rest.

Figure

Punchline — NESS loop current (left) vs detailed balance (right)

Microscopic reversibility is the time-reversal mirror that makes detailed balance the natural equilibrium condition. Run a movie of an elastic two-body collision backwards and you obtain another physically allowed collision with reversed velocities; because the laws permit both equally, the forward and reversed rates are equal. In a Monte-Carlo simulation you get to invent the dynamics — and detailed balance is the leash that forces your invented Markov chain to relax to the physically correct Boltzmann distribution rather than to some arbitrary attractor. The acceptance probability is the precise tension in that leash.

Figure

Microscopic reversibility: forward and time-reversed collisions have equal rate

The core equations

The defining statement is pairwise flux cancellation. It refers to the equilibrium (stationary) distribution $P_{\mathrm{eq}}$ and the transition rates $W$ ; it is a condition on both together, one equation per ordered pair of states.

P_{\mathrm{eq}}(a)\,W(a\to b)=P_{\mathrm{eq}}(b)\,W(b\to a)\qquad\forall\,a,b \qquad (1)

where: $P_{\mathrm{eq}}(a)$ — equilibrium probability of state $a$ ; $W(a\to b)$ — transition rate (per unit time) from $a$ to $b$ . The dynamics that (1) constrains is the master equation, a gain–loss balance whose full derivation and general theory are owned by `aq-sm-master-equation-sm11`. We quote it here only to define the current whose vanishing is detailed balance.

\frac{\mathrm{d}P(a,t)}{\mathrm{d}t}=\sum_b\big[\,W(b\to a)P(b,t)-W(a\to b)P(a,t)\,\big] \qquad (2)

where: $P(a,t)$ — probability of state $a$ at time $t$ ; the two terms are the gain into $a$ from $b$ and the loss out of $a$ to $b$ . The natural object is the net probability current on an edge, the difference of the two directed fluxes. Detailed balance is precisely the statement that this current vanishes on every edge — not merely that the currents into each node cancel.

J_{a\to b}\equiv P(a)\,W(a\to b)-P(b)\,W(b\to a) \qquad (3)

where: $J_{a\to b}$ — net steady-state current from $a$ to $b$ ; detailed balance $\Leftrightarrow J_{a\to b}=0$ for all $a,b$ ; global balance $\Leftrightarrow \sum_b J_{a\to b}=0$ for all $a$ (weaker). Kolmogorov's criterion recasts detailed balance so it can be tested from the rates alone, without first knowing $P_{\mathrm{eq}}$ : around any closed loop, the product of forward rates equals the product of reverse rates.

W(a_1\!\to a_2)W(a_2\!\to a_3)\cdots W(a_n\!\to a_1)=W(a_1\!\to a_n)\cdots W(a_2\!\to a_1) \qquad (4)

where: $a_1,\dots,a_n$ — the states of a closed cycle in state space; equality for every cycle $\Leftrightarrow$ detailed balance. For a thermal system the equilibrium weight is Boltzmann, and (1) fixes the ratio of forward to reverse rates to the Boltzmann factor — the form that Monte-Carlo algorithms enforce by construction.

\frac{W(a\to b)}{W(b\to a)}=\frac{P_{\mathrm{eq}}(b)}{P_{\mathrm{eq}}(a)}=e^{-\beta(E_b-E_a)} \qquad (5)

where: $\beta=1/(k_BT)$ — inverse temperature; $E_a,E_b$ — energies of states $a$ and $b$ . Finally, detailed balance is exactly the condition that the total entropy production rate of the Markov network vanishes; any nonzero loop current makes it strictly positive.

\dot S_{\mathrm{tot}}=\frac{k_B}{2}\sum_{a,b}\big[P(a)W(a\to b)-P(b)W(b\to a)\big]\ln\frac{P(a)W(a\to b)}{P(b)W(b\to a)}\ \ge 0 \qquad (6)

where: $\dot S_{\mathrm{tot}}$ — Schnakenberg total entropy production rate; equals zero iff detailed balance holds.

The core derivations

Derivation ADetailed balance is sufficient for stationarity

The goal is to prove the one-directional implication that anchors the whole subject: if the rates satisfy detailed balance with respect to a distribution $P_{\mathrm{eq}}$ , then $P_{\mathrm{eq}}$ is a stationary solution of the master equation. This is what makes detailed balance a design tool — engineer balanced rates and you are guaranteed your target distribution is a fixed point. The converse fails, and that failure (stationary but not balanced) is exactly the nonequilibrium-steady-state phenomenon derived in the next card.

Step 1. Start from detailed balance for a fixed source state $a$ and every destination $b$ :

P_{\mathrm{eq}}(a)\,W(a\to b)=P_{\mathrm{eq}}(b)\,W(b\to a).

Why: this is the hypothesis, equation (1), applied edge by edge.

Step 2. Sum both sides over all destinations $b$ :

P_{\mathrm{eq}}(a)\sum_b W(a\to b)=\sum_b P_{\mathrm{eq}}(b)\,W(b\to a).

Why: the left side is the total outflow rate from $a$ ; the right side is the total inflow rate to $a$ .

Step 3. Insert this into the master-equation right-hand side evaluated at $P_{\mathrm{eq}}$ :

\left.\frac{\mathrm{d}P(a)}{\mathrm{d}t}\right|_{P_{\mathrm{eq}}}=\sum_b\big[P_{\mathrm{eq}}(b)W(b\to a)-P_{\mathrm{eq}}(a)W(a\to b)\big]=0.

Why: every term cancels individually by Step 1, so the sum vanishes; $P_{\mathrm{eq}}$ is stationary.

Step 4. Note the stronger fact: each net edge current $J_{a\to b}=P_{\mathrm{eq}}(a)W(a\to b)-P_{\mathrm{eq}}(b)W(b\to a)=0$ separately. Why: stationarity requires only $\sum_b J_{a\to b}=0$ ; detailed balance gives the far stronger term-by-term vanishing, which is what distinguishes equilibrium from a nonequilibrium steady state. Verified: the term-by-term cancellation and zero net current on every edge of a Boltzmann-weighted 3-state chain are checked in verify_aq-sm-detailed-balance.py.

Go deeper · Derivation B · 1 minThe Kolmogorov cycle criterion

Derivation BThe Kolmogorov cycle criterion

The goal is to show that detailed balance is equivalent to a condition on the rates alone — the product of forward rates around any closed loop equals the product of reverse rates — so one can test for equilibrium without first constructing $P_{\mathrm{eq}}$ . This turns "does an equilibrium distribution consistent with these rates exist?" into a finite check on cycles, and it is the operational tool for detecting nonequilibrium drive in a rate network.

Step 1. Assume detailed balance and take a closed loop $a_1\to a_2\to\cdots\to a_n\to a_1$ . Form the product of rate ratios around it:

\prod_{i}\frac{W(a_i\to a_{i+1})}{W(a_{i+1}\to a_i)}=\prod_i\frac{P_{\mathrm{eq}}(a_{i+1})}{P_{\mathrm{eq}}(a_i)}.

Why: replace each ratio using equation (5), $W(a\to b)/W(b\to a)=P_{\mathrm{eq}}(b)/P_{\mathrm{eq}}(a)$ .

Step 2. The right-hand product telescopes:

\prod_i\frac{P_{\mathrm{eq}}(a_{i+1})}{P_{\mathrm{eq}}(a_i)}=1.

Why: the loop is closed, so each $P_{\mathrm{eq}}(a_i)$ appears once in a numerator and once in a denominator; everything cancels.

Step 3. Therefore the product of forward rates around the loop equals the product of reverse rates — equation (4). Why: rearranging Step 1 with the value $1$ from Step 2 gives $\prod_i W(a_i\to a_{i+1})=\prod_i W(a_{i+1}\to a_i)$ .

Step 4. Conversely, if every cycle product equals one, one can construct a consistent $P_{\mathrm{eq}}$ (fix it on a spanning tree from equation (5); the cycle conditions guarantee path-independence), and detailed balance follows. Why: the cycle conditions are exactly the integrability conditions for the potential $\ln P_{\mathrm{eq}}$ on the graph, so the criterion is both necessary and sufficient. Verified: the criterion holds for a detailed-balanced chain and fails (ratio $=(k/k')^3\neq1$ ) for a biased ring; a worked example solves the sixth rate of a triangle from the criterion and confirms the resulting stationary state has zero net current — verify_aq-sm-detailed-balance.py.

Limiting cases

Detailed balance is an idealised equilibrium property; here is how it sits among neighbouring conditions.

Case	Result	Meaning
Equilibrium (detailed balance)	$P_{\mathrm{eq}}(a)W(a\to b)=P_{\mathrm{eq}}(b)W(b\to a)\ \forall a,b$	No net current on any edge; zero entropy production; generator symmetrisable, real relaxation spectrum.
Nonequilibrium steady state	$\sum_b J_{a\to b}=0$ but $J_{a\to b}\neq0$	Populations constant yet loop currents circulate; $\dot S_{\mathrm{tot}}>0$ ; driven and active systems.
Kinetic (dilute gas)	$f(\mathbf v_1)f(\mathbf v_2)=f(\mathbf v_1')f(\mathbf v_2')$	Detailed balance of binary collisions; unique solution is Maxwell–Boltzmann, the H-theorem equilibrium.
Thermal bath (canonical)	$W(a\to b)/W(b\to a)=e^{-\beta(E_b-E_a)}$	A bath at temperature $T$ enforces the Boltzmann ratio (KMS condition), driving relaxation to the Gibbs state.

Worked examples

Example 1 — Solve the sixth rate of a triangle from the Kolmogorov criterion (a real number). Three states sit on a triangle. Five directed rates are given: $W_{12}=2,\ W_{23}=3,\ W_{31}=1.5,\ W_{21}=1,\ W_{32}=2$ (units s $^{-1}$ ). Choose the sixth rate $W_{13}$ so the network is detailed-balanced, then confirm the stationary state carries zero current.

Step 1. Kolmogorov's criterion for the single triangular loop reads $W_{12}W_{23}W_{31}=W_{21}W_{32}W_{13}$ .

Step 2. Solve for the missing rate: $W_{13}=\dfrac{W_{12}W_{23}W_{31}}{W_{21}W_{32}}=\dfrac{2\cdot3\cdot1.5}{1\cdot2}=\dfrac{9}{2}=\mathbf{4.5\ \mathrm{s}^{-1}}$ .

Step 3. With this rate the master-equation generator has a stationary distribution whose net current on every edge is zero (checked numerically): detailed balance is restored. Any other choice of $W_{13}$ would leave a nonzero loop current and a nonequilibrium steady state.

Example 2 — Metropolis detailed balance for a two-level system. A two-level system has energy gap $\Delta E$ with $\beta\Delta E=2$ . Verify that the Metropolis rule $\min(1,e^{-\beta\Delta E})$ with symmetric proposals balances against the Boltzmann weight.

Step 1. The target ratio is $p_1/p_0=e^{-\beta\Delta E}=e^{-2}=\mathbf{0.1353}$ .

Step 2. Acceptances: uphill $A_{0\to1}=\min(1,e^{-2})=0.1353$ ; downhill $A_{1\to0}=\min(1,e^{+2})=1$ .

Step 3. Detailed balance requires $A_{0\to1}\,p_0=A_{1\to0}\,p_1$ . With $p_0=1/(1+e^{-2})=0.8808$ and $p_1=0.1192$ : left side $=0.1353\times0.8808=0.1192$ , right side $=1\times0.1192=\mathbf{0.1192}$ . They match — the chain samples the Boltzmann distribution exactly.

Going further

Texts. Pathria & Beale, Statistical Mechanics, 3rd ed., ch. 15 (stochastic processes, master equation) · Kardar, Statistical Physics of Particles, ch. on the Boltzmann equation and H-theorem · Reif, Fundamentals of Statistical and Thermal Physics, ch. 15 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, ch. 8 (Markov chains, Monte Carlo) · van Kampen, Stochastic Processes in Physics and Chemistry, ch. V · Landau & Lifshitz, Physical Kinetics (vol. 10), §2 (principle of detailed balance).

Provenance. The detailed-balance condition and its sufficiency for stationarity follow Pathria & Beale ch. 15, van Kampen ch. V, and Sethna ch. 8. The Kolmogorov cycle criterion and its equivalence to reversibility are standard (Kolmogorov 1936; Sethna ch. 8). The Metropolis identity $\min(1,r)/\min(1,1/r)=r$ and the symmetrisation/real-spectrum argument are textbook (Sethna ch. 8; van Kampen). The Schnakenberg entropy-production formula is quoted as the network result. The Maxwell–Boltzmann consequence of collisional detailed balance is owned by the H-theorem and kinetic-theory articles; here it appears only as a limiting case. No numerical constants beyond $k_B$ and standard exponents are used.

Builds on: Markov Processes and the Markov Property · The Master Equation · Chapman–Kolmogorov Equation. Leads to: Metropolis–Hastings Algorithm · Monte Carlo Method · Fluctuation–Dissipation Theorem · Kinetic Theory and the Boltzmann Equation.

What exactly is the difference between detailed balance and global balance?

Global balance (stationarity) says the total probability current into each node is zero: $\sum_b J_{a\to b}=0$ . Detailed balance says each individual edge current is zero: $J_{a\to b}=0$ for every pair. Global balance is weaker and is satisfied by any steady state, including nonequilibrium ones with circulating loop currents. Detailed balance is the equilibrium condition — it forbids the loops. A three-state ring with a clockwise bias is the canonical example: stationary (global balance holds) but not detailed-balanced (a persistent loop current circulates and entropy is produced).

Is detailed balance necessary for a valid Monte-Carlo sampler?

No. Detailed balance is sufficient to guarantee your target distribution is stationary, and it is by far the easiest condition to impose, which is why almost all classic samplers use it. But it is not necessary: what you actually need is that the target be the unique stationary distribution and that the chain be ergodic (irreducible and aperiodic). Non-reversible samplers deliberately break detailed balance while preserving stationarity, and they can mix faster by suppressing diffusive back-and-forth. So detailed balance is the standard, convenient route to correctness — not a logical requirement for it.

Why does detailed balance make the relaxation spectrum real?

Under detailed balance the generator $\mathcal L$ becomes similar to a symmetric matrix via conjugation by $D=\mathrm{diag}(\sqrt{P_{\mathrm{eq}}})$ : the entry $\tilde{\mathcal L}_{ab}=\sqrt{P_{\mathrm{eq}}(b)/P_{\mathrm{eq}}(a)}\,W(b\to a)$ equals $\sqrt{W(a\to b)W(b\to a)}$ , manifestly symmetric in $a\leftrightarrow b$ . A real symmetric matrix has real eigenvalues, and a similarity transform preserves the spectrum, so all relaxation rates are real (and non-positive, with a single zero for the stationary mode). Physically, equilibrium dynamics relaxes monotonically per mode; oscillatory relaxation (complex eigenvalues) is a fingerprint of broken detailed balance.

How is detailed balance related to entropy production?

The total entropy production rate of a Markov network is $\dot S_{\mathrm{tot}}=\tfrac{k_B}{2}\sum_{a,b}(P_aW_{ab}-P_bW_{ba})\ln\frac{P_aW_{ab}}{P_bW_{ba}}$ . Each summand is of the form $(x-y)\ln(x/y)\ge0$ , so $\dot S_{\mathrm{tot}}\ge0$ always, and it equals zero precisely when every $P_aW_{ab}=P_bW_{ba}$ — that is, exactly under detailed balance. So detailed balance is the zero-dissipation condition. A nonzero loop current forces a strictly positive entropy production; the further from balance the rates are pushed, the more heat the system dumps into its reservoirs to maintain the steady current.

Where does microscopic reversibility come from, and can it fail?

It comes from time-reversal symmetry of the microscopic equations of motion: reverse all velocities (and, if present, the magnetic field) and an allowed trajectory maps to another allowed trajectory of equal statistical weight. Coarse-graining such dynamics yields transition rates for which forward and reverse are equal, hence detailed balance in equilibrium. It can fail in two ways. First, an external non-conservative drive (a chemical potential difference, a shear, a rotating field) breaks it and produces a nonequilibrium steady state. Second, a magnetic field or Coriolis force breaks naive time-reversal; there detailed balance is replaced by a generalised version relating the rates in the field to the rates in the reversed field (the Onsager–Casimir symmetry).

Can I check detailed balance without knowing the equilibrium distribution?

Yes — that is exactly what the Kolmogorov cycle criterion is for. Detailed balance holds if and only if, around every independent cycle in the state-space graph, the product of forward transition rates equals the product of reverse rates. This uses only the rates $W$ , never $P_{\mathrm{eq}}$ . Practically you only need to check a basis of independent cycles (the graph's cycle space); once they balance, a consistent $P_{\mathrm{eq}}$ exists and can be reconstructed on a spanning tree. A single failing cycle is enough to certify a nonequilibrium drive.

What is the connection to the fluctuation–dissipation theorem?

The standard fluctuation–dissipation theorem assumes equilibrium, and equilibrium means detailed balance (the KMS condition is its quantum expression). When detailed balance holds, spontaneous fluctuations and linear response are locked together by temperature, and Onsager's reciprocal relations follow from microscopic reversibility. When detailed balance is broken — driven, aging, or active systems — the fluctuation–dissipation theorem acquires correction terms, and the deviation is used to define an effective temperature. So detailed balance is the hidden premise of the equilibrium fluctuation–dissipation relations discussed in `aq-sm-fluctuation-dissipation-theorem-sm11`.