Contribution
Before this paper, iterated learning (Kirby’s model of each learner learning from the output of a previous learner) was studied by simulating particular learning algorithms, and the argument that a “bottleneck” on transmitted data drives languages toward structure rested on those simulations. After it, there is a general analytic result for Bayesian learners: when learners sample from the posterior, the stationary distribution over languages is exactly the learners’ prior, independent of the data, the bottleneck and the hypotheses’ structure. It also identifies the two learning rules with two statistical algorithms, the Gibbs sampler and stochastic EM, so their convergence theory transfers.
Key insight
A chain of learners who each infer a hypothesis from data and then generate data from it is alternately sampling h given d and d given h. That is a Gibbs sampler for the joint P(d|h)P(h). Its stationary distribution has the prior as its marginal on h. The first learner’s data is “only a single piece of information, while the prior asserts its effect on each iteration” (§4.1). What survives transmission is therefore what the learners already expected, and the population’s languages come to mirror the learners’ inductive biases.
Assumptions
- Assumption 1: every learner uses the same learning algorithm and production algorithm. For Bayesian learners this means the same hypothesis space H and the same prior P(h).
- Assumption 2: discrete generations with one learner each, learning from the previous learner’s data. Or, Assumption 2′ (§7): an unbounded population in continuous time, each learner learning from a random member of the population at the previous instant.
- Assumption 3: learners know the production distribution P_PA(d|h), so learning and production are consistent (the likelihood is the true production distribution).
- Ergodicity of the Markov chain (irreducible and aperiodic), which fails when some language is a sink. In the two-language example it fails only when the languages never agree (s = 0) and production is noiseless (ε = 0).
- H and D finite (note 1), “not a necessary assumption”.
- No selection: the population result needs equal fitness f_j = 1. The authors warn that selection “could easily disrupt” the correspondence with statistical inference (§8.2).
- The prior is not read as innate language-specific constraint: it “collects together all of the factors affecting how easily a learner will come to entertain a particular hypothesis” (§3.1). The analysis is at Marr’s computational level, not a claim about mechanism.
Key results
- Markov chains (Eqs. 9–11). Summing out data gives a chain on hypotheses with q_ij = Σ_d P_LA(h_n = i | d) P_PA(d | h_{n−1} = j); summing out hypotheses gives a chain on data; grouping gives a chain on (h, d) pairs.
- Two languages (Eqs. 14–15). θ1 = q12 / (q12 + q21); second eigenvalue λ2 = 1 − q12 − q21, which sets the rate of convergence.
- Sampling converges to the prior (Eqs. 25–30). With P(h|d) as the learning rule, θ_i = P(h = i) satisfies θ = Qθ; the chain on data has the prior predictive P_PA(d) = Σ_h P_PA(d|h)P(h) as its stationary distribution (Eqs. 31–36). Holds for any production algorithm and any amount of data; the same holds if learners average over hypotheses rather than sample one.
- Gibbs sampler (§4.2). Iterated learning by sampling is a systematic-scan Gibbs sampler for P_PA(d|h)P(h), so convergence is geometric in n under standard results. The authors call this, to their knowledge, the first connection between MCMC and human cognition.
- MAP as stochastic EM (§5.1). Iterated learning with MAP learners is stochastic EM (one sample of the latent variable per iteration) in a model where the transmitted data are the latent variable and there are no observations. The stationary distribution should be approximately centred on the prior’s maximum, with variance growing with the rate of change between generations. No explicit asymptotic result exists for discrete hypothesis spaces.
- MAP with two languages (Table 1). With prior α > 0.5 on L1, agreement s and noise ε: if ε < 1 − α, θ1 = (s + (1 − s)ε) / (s + 2(1 − s)ε) and λ2 = (1 − s)(1 − 2ε), independent of α. If ε > 1 − α, θ1 = 1 and λ2 = 0. So a weak bias (α just above 0.5) and a strong one (α just below 1 − ε) give the same outcome: MAP learning can amplify weak biases.
- Compositionality simulation (§6). Four meanings and four utterances ({00, 01, 10, 11}); 4 compositional and 256 holistic languages (260 hypotheses), prior α/4 on each compositional and (1 − α)/256 on each holistic; α ∈ {0.01, 0.5}, ε ∈ {0.01, 0.05}, m = 1 … 10 utterances per generation (Monte Carlo transition matrices for m > 4). Sampling: the long-run frequencies match the prior at every m; larger m only slows mixing (λ2 rises). MAP: with α = 0.01 compositional languages are never chosen (each is matched by a holistic language with the same mapping and higher prior mass); with α = 0.5 and m = 1 or 2 only compositional languages appear; holistic ones appear from m = 3, more at m = 10.
- Populations (§7). Under equal fitness, dp_i/dt = Σ_j q_ij p_j − p_i has the unique stable equilibrium p = θ, so the single-chain stationary distribution is also the long-run population share.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Iterated learning by posterior sampling has the shared prior as its stationary distribution over hypotheses, and the prior predictive over data | strong (proof), under Assumptions 1–3 and ergodicity | Eqs. 25–36 |
| C2 | Iterated learning by sampling is a Gibbs sampler on P(d|h)P(h) | strong (identification) | §4.2 |
| C3 | Iterated learning by MAP is stochastic EM with no observations, so its stationary distribution concentrates near the prior’s mode | moderate: the identification is exact, the “approximately centred” behaviour is borrowed from stochastic-EM results for other settings | §5.1 |
| C4 | Under MAP, the outcome depends on data amount and noise and can be the same for a range of prior strengths, amplifying weak biases | strong for the two-language case (closed form); illustrated, not proved, in general | Table 1, §6.2 |
| C5 | With sampling learners, a bottleneck is not needed for structured languages to emerge, but they emerge only if the prior favours them | strong under the model | §4.4, §6.1 |
| C6 | The single-learner chain’s stationary distribution is the equilibrium of an unbounded equal-fitness population | strong (linear dynamics) | §7, note 6 |
| C7 | Anything transmitted by iterated learning, including “legends, religious concepts, and social norms”, will come to match learners’ inductive biases | weak: an extrapolation in the conclusion, not analysed | §8.5 |
Method
Reduce iterated learning to a homogeneous finite Markov chain by summing out data (or hypotheses), find the stationary distribution as the eigenvector of the transition matrix with eigenvalue 1, and use the second eigenvalue for the rate. For Bayesian learners, verify by substitution that the prior (and the prior predictive) is stationary. Worked examples: a two-language “Gavagai” case with closed forms, and a 260-hypothesis compositionality case with numerically computed transition matrices.
Concepts
- iterated learning: each learner forms a hypothesis from data produced by a learner who learned the same way, then produces data for the next.
- learning algorithm P_LA(h|d) / production algorithm P_PA(d|h): the two conditional distributions that define one generation.
- prior: everything that makes a hypothesis easier or harder to adopt, read as “the amount of evidence that a learner would need to see in order to adopt a particular language” (§3.1).
- bottleneck: the finite amount of data passed between generations.
- stability ratio: mean q_ii for compositional languages over mean q_ii for holistic ones (§6.1).
Connections
- Kirby (2001), Brighton (2002), Smith, Kirby & Brighton (2003): the iterated learning model and its simulations of the emergence of compositionality, which the paper re-reads as MAP-like learners (minimum description length is MAP under a coding prior).
- Nowak, Komarova et al. (2001–2003): the language dynamical equation; the paper’s population result is its equal-fitness case.
- Kirby, Smith & Brighton (2004): an earlier Bayesian framing (note 2).
- Geman & Geman (1984); Celeux & Diebolt (1985): Gibbs sampling and stochastic EM.
- LIT-771 (Kirby, Cornish & Smith 2008, read in NOTE-596) is the laboratory version of the process. It cites this paper only within a block of model references and does not discuss priors. Its human chains show increasing structure and learnability, which is what this paper says either learning rule produces when the prior favours structure.
- LIT-768 (Bartlett 1932): the paper does not cite Bartlett, but its closing extrapolation to legends is the claim Bartlett’s serial reproduction of a folk tale was about: material drifts toward what the rememberer’s culture makes expectable. This paper gives that a formal form for one class of learner. The connection is mine.
Bearing on the record
- No existing THEORY covers iterated learning, cultural transmission or convergence to a prior (grep of theory.d on 2026-10-09). This reading produces THEORY-155: under posterior sampling, iterated learning converges to the shared prior, and the MAP case is where data amount matters. Its “what this does not say” carries the limits below.
- No instruction for machine-learning practice; not an anthology candidate.
Limitations
- Homogeneous learners. Everything rests on all learners sharing one prior and one likelihood that matches production. Heterogeneous priors, or learners who learn from several teachers, are outside the analysis (§8.2 names multi-generation input as future work via stochastic EM variants).
- No selection. With fitness differences the population result and the statistical-algorithm correspondence may fail; the authors say so.
- Which rule humans use is open. The authors argue for sampling from probability matching but say the question “has not been explored in depth” for language (§8.4). The two rules give opposite answers on whether the bottleneck matters.
- MAP results are rough. Beyond two languages, the MAP case rests on stochastic-EM behaviour established for other settings and on one simulation.
- The paper does not decide between innate and general-purpose explanations of universals; it states both can use its results.
- The compositionality example is shaped by H. Note 5: the stability ratio’s decline with m comes partly from holistic languages that duplicate compositional mappings, “rather than being a general trend”.
Open questions
- Do human transmission chains converge to an independently measured prior? A chain experiment in which each participant’s prior is measured or manipulated would test C1 directly.
- How does the MAP stationary distribution depend on bottleneck size in general (the authors point to Dowman, Kirby & Griffiths 2006)?
- What happens between sampling and MAP (posterior raised to a power), and with heterogeneous priors in a population?