Class Mobility: Markov Chains and Stationary Distributions

The DefaultClassTransitionComputer models class mobility as a Markov chain: the probability of moving between classes depends only on current class position, not on prior history. This document explains the mathematics of Markov chains, the stationary distribution, and how Babylon computes and uses it.

For the API reference, see Tensor Hierarchy Reference.

Class Transition as a Markov Chain 

A Markov chain is a stochastic process where the future state depends only on the present state. For class mobility, this means:

The probability of being in class j at time t+1 depends only on class membership at time t, not on class history at t-1, t-2, …

This is the memoryless assumption. It is an idealization—actual class trajectories have path dependencies (inherited wealth, credential accumulation). But for aggregate long-run structural analysis, the Markov approximation is well-established in mobility research and provides computationally tractable results.

The Transition Matrix P 

The class transition matrix P is an n*×*n matrix where:

\[P[i, j] = \Pr(\text{class } j \text{ at } t+1 \mid \text{class } i \text{ at } t)\]

Each row represents a conditional distribution over next-period class membership, given current-period class. The rows must sum to 1.0 (all possibilities are covered) and all elements must be non-negative (probabilities cannot be negative).

This property—rows summing to 1—is called row stochasticity. Babylon enforces it at construction time with tolerance 1e-6 and validates it in the three-tier validation framework.

Transition Matrix Validation Thresholds
Property	Threshold	Enforcement
Elements ≥ 0	0.0	Fail if any element < 0
Row sum deviation (expected)	≤ 1e-6	Expected: exact row stochasticity
Row sum deviation (warning)	≤ 1e-4	Warning: floating-point drift
Diagonal self-transition (expected)	≥ 0.50	Most people stay in their class
Diagonal self-transition (warning)	≥ 0.20	High mobility period

The diagonal element P[i,*i*] represents the probability of staying in class i. The off-diagonal P[i,*j*] represents upward or downward mobility probability. The diagonal-dominance condition (high self-transition probabilities) reflects the empirical reality that class positions are relatively stable across short time horizons.

The Stationary Distribution 

As a Markov chain evolves through many periods, it converges toward a stationary distribution π—a fixed-point probability vector that is unchanged by one application of P:

\[\pi P = \pi, \quad \sum_i \pi_i = 1, \quad \pi_i \geq 0\]

The stationary distribution represents the long-run equilibrium: if the transition matrix P remained constant indefinitely, the fraction of the population in each class would converge to π. It is the gravitational attractor of the class structure.

This is a structural measure, not a prediction of any individual’s trajectory. It says: given the current structure of class mobility (transition probabilities), what is the long-run composition of classes that this structure tends to reproduce?

Why the Stationary Distribution Matters 

The stationary distribution plays two roles in the tensor hierarchy:

As a summary statistic. π gives a single compact description of the long-run class composition implied by current mobility patterns. Changes in π over time reveal how the structure of class reproduction is shifting.
As weights for class aggregation. When aggregating a fine-grained class transition matrix into coarser categories, the stationary weights are the correct weights for combining rows (see Class Aggregation below).

Computing the Stationary Distribution 

The stationary distribution is the dominant left eigenvector of P— equivalently, the right eigenvector of P^T corresponding to eigenvalue 1.0.

Why Eigenvalue 1.0?

If π is a stationary distribution, then π*P* = π can be rewritten as:

\[\pi (P - I) = 0\]

This means π is a left eigenvector of P with eigenvalue 1.0. Equivalently, (P^T)π^T = π^T, so π^T is a right eigenvector of P^T with eigenvalue 1.0.

For a row-stochastic matrix with irreducible and aperiodic structure (Perron-Frobenius theorem), eigenvalue 1.0 is unique and the corresponding eigenvector has all positive elements.

The Computation Algorithm 

Babylon’s DefaultClassTransitionComputer.stationary_distribution implements:

Eigendecompose P^T: compute all eigenvalues and eigenvectors of the transpose.
Find eigenvalue closest to 1.0: uses argmin(|eigenvalues - 1.0|) rather than checking for equality. This tolerates floating-point representations of 1.0 that may be 0.9999999… or 1.0000001…, while still identifying the dominant eigenvector correctly.
Extract the real part: eigenvalues of real matrices may come in complex conjugate pairs. The dominant eigenvalue 1.0 always has a real eigenvector, so the real part is taken.
Clip negatives: tiny negative values arise from floating-point arithmetic on the imaginary residual. Values are clipped to [0, ∞).
Normalize to sum 1.0: divide by the sum to produce a probability distribution.
Degenerate case: if after clipping the sum is 0 (all-zero eigenvector), fall back to a uniform distribution over classes.

The result is a StationaryDistribution with distribution summing to 1.0 within tolerance 1e-6.

Class Aggregation with Stationary Weights 

When a fine-grained transition matrix (e.g., 10 class categories) needs to be reduced to a coarser representation (e.g., 3 aggregate classes), the DefaultClassTransitionComputer.aggregate_classes method performs a weighted block-sum reduction.

Why Stationary Weights?

Suppose we are merging two fine classes A and B into one aggregate class W (working class). The aggregated row for W must represent the average behavior of the merged classes. The correct average is:

Weight each fine class by its long-run prevalence in the population.

Using the stationary distribution π as weights is the principled choice: π[A] is the long-run fraction of the population in class A, so it is the correct weight for class A’s transition row in the aggregate.

Using equal weights would distort the aggregate by treating rare classes the same as common classes. Using current-period empirical shares would introduce period-specific noise. The stationary distribution is invariant to period choice while respecting the structure of the mobility process.

The Aggregation Algorithm 

Map each source class to a target class via the provided mapping dict.
Compute the stationary distribution π of the source matrix.
Accumulate weighted flows: for each (source class i, destination class j) pair that maps to (aggregate origin a, aggregate destination b), add π[i] × P[i,*j*] to the aggregate flow[a,*b*].
Accumulate total weights per aggregate origin class.
Divide each aggregate row by its total weight to re-normalize.
Final normalization: divide each row by its sum to remove floating-point drift and ensure exact row stochasticity.

The resulting matrix is itself row-stochastic, and its stationary distribution approximates the coarsened stationary distribution of the fine-grained matrix.

Political Interpretation 

The stationary distribution is not a neutral statistical tool. In Babylon’s framework it carries specific political meaning.

Structural determination. The stationary distribution π is determined entirely by the transition matrix P—the structural mobility rates embedded in the economy’s institutions (education, inheritance, labor markets). This makes π a measure of what the current class structure tends to reproduce, independent of any individual choices or contingencies.

The gravitational pull of capitalism. Even if individual trajectories are highly mobile in a given year, π reveals the long-run attractor. If the Leontief structure and imperial rent field imply a transition matrix where π shows 75% proletariat and 5% bourgeoisie, that is the structure capitalism tends to reproduce. No amount of “social mobility” that changes individual trajectories changes π if it leaves the transition matrix unchanged.

Class struggle as P-matrix modification. Revolutionary conditions are conditions that change P itself—that alter the transition probabilities. A successful struggle for reforms (union rights, public education) shows up as changes in P[proletariat→proletariat], P[proletariat→petit_bourgeois], etc. The simulation can compare stationary distributions before and after such structural changes to measure their long-run effects.

The Deferred Loader 

The DefaultClassTransitionSource is a stub that returns NoDataSentinel for all queries. The production implementation requires PSID (Panel Study of Income Dynamics) data from the University of Michigan, which requires a restricted-use data agreement.

This deferred status does not affect the computation engine: DefaultClassTransitionComputer is fully implemented and all computation tests pass using synthetic transition matrices. When PSID data becomes available, only the source implementation changes.