Is detailed balance necessary for MCMC?

No. Nonreversible methods can also preserve a target distribution.

Can a chain satisfy it without converging from every start?

Yes. For example, a chain that always stays where it starts satisfies it for any distribution but cannot explore other states.

Does the idea extend to continuous time?

Yes, with transition rates replacing one-step transition probabilities in the corresponding equations.

AI Glossary / Graphs and symbolic AI / Advanced

Detailed balance

Definition

Detailed balance is a pairwise flow condition for a Markov chain and a probability distribution. In a discrete-state chain, it requires π(i)P(i,j) = π(j)P(j,i) for every pair of states i and j. Here π is the distribution and P(i,j) is the probability of moving from i to j. The condition makes π stationary and gives the chain a reversible equilibrium interpretation.

What it means

At stationarity, the probability flow from one state to another matches the reverse flow for each pair. This is stronger than merely keeping the total probability at each state unchanged. A system can have balanced total inflows and outflows while probability circulates around a loop.
Detailed balance does not require equal transition probabilities in both directions. A more frequently occupied state can have a smaller probability of leaving for another state, while the weighted flows still match.
In AI and statistics, the condition helps design and verify Markov chain Monte Carlo methods. It is one way to establish the desired stationary distribution, rather than a universal requirement for every valid sampling method.

How it works

Specify a valid transition matrix whose rows sum to one and a candidate distribution whose entries sum to one. Check the weighted forward and reverse transitions for each state pair. Summing those pairwise equalities shows that applying one transition leaves the candidate distribution unchanged.
Stationarity is a distributional statement. Individual sample paths still move, and observations from successive steps can remain correlated. Reversibility means the stationary process has the same finite-path distribution when time is reversed, not that every observed trajectory immediately retraces itself.
Detailed balance alone does not establish useful convergence from every starting point. For a finite discrete-time chain, irreducibility and aperiodicity supply standard additional conditions for convergence to the unique stationary distribution. Even then, convergence can be slow.

Three practical examples

Bayesian computation: A researcher designs a sampler intended to preserve a posterior distribution. A detailed-balance argument can establish that target as stationary. The researcher still needs to examine whether a finite run explores the relevant regions adequately.
Simulation checks: A team tests a small transition model with known probabilities before trusting a larger implementation. Pairwise flow calculations can reveal a mistaken transition entry that a superficial row-sum check would miss.
Algorithm comparison: Two reversible samplers share the same target but move at different rates. Both can satisfy detailed balance while producing substantially different amounts of useful information per unit of computation. Correct stationary behavior does not settle practical efficiency.

Think of it this way

Imagine two rooms with different numbers of occupants. If the larger room sends a smaller fraction of its occupants to the smaller room, the number moving each way can still match. Detailed balance compares those weighted flows, not just the fractions leaving each room. The analogy concerns modeled probabilities, not exact counts on every occasion.

Worked example

Consider an invented two-state chain with states A and B. From A, it stays at A with probability 0.8 and moves to B with probability 0.2. From B, it moves to A with probability 0.3 and stays at B with probability 0.7.
Test candidate distribution π(A) = 0.6 and π(B) = 0.4. The A-to-B flow is 0.6 × 0.2 = 0.12. The B-to-A flow is 0.4 × 0.3 = 0.12. The self-pair equations hold automatically, so detailed balance is satisfied.
Check stationarity directly: the next-step probability of A is 0.6 × 0.8 + 0.4 × 0.3 = 0.48 + 0.12 = 0.6. For B it is 0.6 × 0.2 + 0.4 × 0.7 = 0.12 + 0.28 = 0.4.
Now reduce the cross-state probabilities to 0.002 and 0.003, with staying probabilities 0.998 and 0.997. The weighted flows both become 0.0012, so the same distribution still satisfies detailed balance. However, the chain changes state much less frequently. Starting at A, its probability of being at B after one step is only 0.002 rather than 0.2.
This demonstrates that preserving the target distribution does not determine how quickly a chain explores it. Neither version produces independent consecutive draws merely because detailed balance holds.

Illustrative case study

Fictional Willow Forecasting uses a sampler to estimate uncertainty in a demand model. A developer presents a detailed-balance proof as evidence that the sampling results are ready for use. Reviewers agree that it addresses the intended stationary distribution, then ask separate questions about the finite computation.
They inspect whether the implemented transitions match the derivation, whether relevant regions communicate, and whether the run moves effectively between them. A small test model makes the distinction visible: reducing movement probabilities preserves detailed balance but leaves short runs strongly influenced by their starting state.
The review separates mathematical invariance from practical sampling adequacy. This is an illustrative case, not a report of measured forecasting improvements.

Common misconceptions and mistakes

“P(i,j) must equal P(j,i).” The equality concerns probabilities weighted by π.
“Stationarity always implies detailed balance.” Circulating flows can preserve a distribution without pairwise balance.
“Detailed balance guarantees fast mixing.” It does not.
“Reversible means each move is immediately undone.” It concerns the distribution of stationary paths under time reversal.
“Balanced sampling makes the underlying model correct.” A sampler can accurately target a poorly specified model.

Business use

When reviewing sampling-based analysis, separate the target distribution, the transition design, and the quality of the finite sample. Ask what establishes invariance and what evidence supports adequate exploration.
Check simple numerical cases before relying on a complex implementation. Document the state space and any restrictions that could prevent movement between regions. Evaluate uncertainty estimates using appropriate sampling diagnostics and model checks, rather than treating a detailed-balance statement as a complete validation of the analysis.

Check your understanding

Why do unequal probabilities 0.2 and 0.3 balance?

Answer: They are weighted by 0.6 and 0.4 respectively.

What stays unchanged in the slower example?

Answer: The stationary distribution.

What remains to assess after detailed balance?

Answer: Convergence conditions, practical exploration, implementation, and model suitability.

Related terms

Learn next

Verify detailed balance for a small transition matrix, then calculate the next-step distribution from two different initial states. Change the movement rate while preserving balance and explain which properties remain unchanged.

Sources

Aldous and Fill, Reversible Markov Chains and Random Walks on Graphs, Chapter 3.1: https://www.stat.berkeley.edu/users/aldous/RWG/Book_Ralph/Ch3.S1.html UC Berkeley STAT 210A, Markov Chain Monte Carlo, stationarity and convergence: https://stat210a.berkeley.edu/fall-2025/reader/bayes-computation.html

Previous: Design of experiments Next: Detectable difference