Contribution
- A minimal model in which sharing is advantageous with no complementarity, no threshold, no relatedness and no reciprocity, only multiplicative noise. N identical GBM entities grow, pool, and split equally each step (Eqs. 3.3–3.4, Fig. 1). The shared process is again a GBM with the same drift and volatility σ/√N (Eq. 3.9).
- Closed-form time-average growth rates for three cases:
- identical members (Eq. 4.1);
- heterogeneous members (Eq. 5.5), with join and admit conditions (Eqs. 5.6–5.7);
- equicorrelated noise (Eqs. 5.12–5.13).
- An argument that, because the ensemble-average rate μ is unchanged by sharing, analyses built on expectations see no reason to cooperate, while the time-average rate μ − σ²/2 that a single trajectory achieves does see one (§4).
Key insight
Under multiplicative noise, fluctuations cost growth: a single trajectory grows at μ − σ²/2, not μ. Averaging resources across independent trajectories every period removes variance without removing drift. So sharing converts variance into long-run growth for every participant, and the pooled members out-grow even the arithmetic mean of an equal number of non-sharers (“cooperation and averaging are not equivalent operations”, §4).
Assumptions
- Dynamics. Resources follow GBM, dx_i = x_i(μ dt + σ dW_i), with independent Wiener increments. This is unconstrained self-reproduction: the authors state that growth limited by resources, space or predation “would be poorly described” by it (§2).
- GBM as an attractor. Justified by the CLT on log-increments over many steps (§2). This requires finite-variance, weakly dependent log-increments.
- Pooling. Complete and equal pooling at every step Δt, with Δt → 0 in the continuous-time result. There is no cost of sharing and no benefit beyond variance reduction (§3).
- Pacts cannot be broken in the main result. Cheating and walk-away are set aside, with “ignore-for-tat” enforcement only sketched (§4a).
- The criterion is the time-average growth rate ḡ = lim_{T→∞} (1/T) ln(x(T)/x(0)). The claim that the entity with the highest ḡ comes to dominate is the paper’s selection criterion (§2).
- Correlated case. Equal pairwise correlation ρ ≥ −1/(N−1) (Eq. 5.8).
Key results
- Eq. 2.4: g(x_i, T) ~ N(μ − σ²/2, σ²/T). Eq. 2.8: ḡ(x_i) = μ − σ²/2. Eq. 2.7: the ensemble rate is g(⟨x⟩) = μ.
- Eq. 3.9: dy^(N) = y^(N)(μ dt + (σ/√N) dW
- Eq. 4.1: ḡ(y^(N)) = μ − σ²/(2N). Eq. 4.2: premium = (σ²/2)(1 − 1/N). Eq. 4.3: → μ as N → ∞.
- Eqs. 5.3, 5.5 (heterogeneous): μ^(N) = (1/N)Σμ_i and σ^(N) = (1/N)√(Σσ_i²), giving ḡ = (1/N)Σ(μ_i − σ_i²/(2N)).
- Eqs. 5.9, 5.12, 5.13 (correlated): Σξ_i ~ N(0, N + N(N−1)ρ); σ_corr = σ√((1 + (N−1)ρ)/N); and ḡ → μ − σ²ρ/2 as N → ∞.
- Fig. 2 (simulation; μ = 0.15, σ² = 0.2, Δt = 0.1): one pair of cooperators vs two non-cooperators and their mean. The implied rates are 0.05 for a non-cooperator and 0.10 for a pair member (my arithmetic from Eq. 4.1).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Equal pooling among N independent GBMs leaves drift μ and reduces volatility to σ/√N | strong | derivation, Eqs. 3.6–3.9 |
| C2 | Each cooperator’s time-average growth rate rises to μ − σ²/(2N), premium (σ²/2)(1 − 1/N), with the largest marginal gains at small N | strong | Eqs. 4.1–4.2 (within the model) |
| C3 | Cooperators out-grow the arithmetic average of the same number of non-cooperators | strong (as a long-time statement) | Fig. 2 simulation; follows from C2, since a finite sum of independent GBMs grows at μ − σ²/2 almost surely |
| C4 | With equicorrelated noise, a benefit exists iff ρ < 1, and the many-member limit is μ − σ²ρ/2 | strong | Eqs. 5.9–5.13 |
| C5 | Heterogeneous entities benefit from joining iff Eq. 5.6 holds, so highly skilled members may do better alone | strong | Eqs. 5.5–5.7 |
| C6 | This is a candidate explanation for cooperation in rudimentary settings (for example unicellular to multicellular) | weak | informal argument (§1, §6); no biological model or data |
| C7 | Cooperation is the behavioural baseline, and its absence needs special explanation | weak | follows only under the model’s zero-cost, no-defection, unconstrained-growth assumptions (§6) |
| C8 | An “ignore-for-tat” policy suffices to protect the cooperative from pact-breakers, even with imperfect detection | weak | informal argument (§4a); no invasion or stability analysis |
| C9 | The growth benefit is an overlooked rationale for insurance, pensions and taxation | weak | assertion (§6) |
Method
- Non-cooperators: discrete GBM, Δx_i = x_i(μΔt + σξ_i√Δt) with ξ_i ~ N(0,1) (Eqs. 3.1–3.2).
- Cooperators: each grows from the common per-capita level y^(N), then the pool is split equally (Eqs. 3.3–3.4).
- Replace the sum of the N normals by one equivalent standard normal, ξ^(N) = Σξ_i/√N (Eq. 3.7). Take Δt → 0 to get a GBM with volatility σ/√N.
- Apply the GBM time-average formula μ − σ²/2 to each case. Generalize to (μ_i, σ_i) (Eqs. 5.1–5.5) and to equicorrelated ξ (Eqs. 5.8–5.12).
Concepts
- ensemble-average growth rate g(⟨x⟩) — the growth rate of the population mean in the infinite-population limit, equal to μ (Eq. 2.7).
- time-average growth rate ḡ(x_i) — the long-time growth rate of a single trajectory, μ − σ²/2 (Eq. 2.8). Their inequality is what the paper calls non-ergodicity.
- cooperation — repeated pooling and equal sharing of resources, and nothing more (§1, §3).
- ignore-for-tat — refusing future interaction with past pact-breakers (§4a).
- σ_corr — the effective volatility of a cooperative with pairwise noise correlation ρ (Eq. 5.12).
Connections
- Geometric-mean fitness and bet-hedging. The paper sits squarely in this literature (refs 8–17 as cited in §1): Starrfelt & Kokko, Schreiber, and Kennedy et al., for whom this is a special case. It chooses “growth rates, not fitness” as its language.
- Earlier derivations of the same variance. McCloskey (the medieval field-scattering literature) and Winterhalder (forager food sharing) derive the same pooled-variance expression, per §5b. Uitdehaag’s anticorrelated-specialist mutualism is the N = 2, ρ = −1 extreme.
- Extensions by others. §5c points to partial pooling (Berman et al.), network pooling (Stojkoski et al.) and binary-outcome processes (Yaari & Solomon).
- Programme. The paper presents itself as part of ergodicity economics (ref. 4, Peters 2019).
- In this record. It pairs with LIT-047 (Curry et al.) as a quite different explanation of cooperation: a game-theoretic “solutions to cooperation problems” account on the one hand, a variance-reduction account on the other. It also relates to the individuality and major-transitions cluster through its unicellular-to-multicellular remark.
Bearing on the record
There is no instruction for ML practice. The underlying mathematics (log-growth, the μ − σ²/2 volatility drag, Kelly-style time-average criteria) is standard and has no specific bearing on any anthology practice. If a THEORY document in this record makes a variance-reduction or “time-average vs ensemble-average” argument about cooperation or individuality, this paper supports the mathematics but not the evolutionary claim.
Limitations
- Zero-cost, perfect, continuous pooling. The authors acknowledge that coordination costs “may … make cooperation disadvantageous” (§3), but no cost term is analysed, so the paper cannot say when cooperation stops paying.
- Defection and invasion are explicitly out of scope (“we refrain from further development”, §4a). Yet the “baseline” claim depends on cooperation being stable against pact-breakers.
- Unconstrained growth. GBM is the authors’ stated limitation. Under carrying-capacity limits the premium’s form is not derived.
- Time average as the criterion. Selection on the time-average rate is asserted as what matters for dominance over long times. Finite horizons and population structure are not treated.
- The unicellular-to-multicellular application is a suggestion without a model.
- The title (“The ergodicity solution of the cooperation puzzle”) claims more than the body: the body shows a sufficient mechanism in an idealized model, which the abstract more modestly calls a “candidate explanation”.
Open questions
- With a per-step sharing cost c, or partial pooling fraction f, what is the break-even? Presumably σ²(1 − 1/N)/2 > c, but that is not derived here.
- Is full pooling evolutionarily stable against a mutant that shares only when it has lost, in the same GBM setting? That calls for an invasion analysis.
- Does any empirical system (for example early multicellular aggregates or microbial public-goods sharing) show the predicted 1 − 1/N scaling of the growth benefit?
Corrections to the seeded skim
- Volatility. The dossier says the text “renders the volatility as ‘σ/N’” but that the stated rate implies variance σ²/N. The version of record has σ/√N throughout: Eq. 3.7 defines ξ^(N) = (1/√N)Σξ_i; Eqs. 3.8–3.9 have σ/√N; the prose says “amplitude is 1/√N times” (§3); and “σ_corr → σ/√N as ρ → 0” (§5b). The “σ/N” in the flattened text is an extraction artefact (the √ was dropped), not a typo in the paper. The dossier’s summary line (“variance from σ² to σ²/N”) is correct.
- Correlated case. The dossier asks whether the premium vanishes as ρ → 1. It does, exactly and only there: 0 ≤ σ_corr ≤ σ, with σ_corr = σ iff ρ = 1 (§5b). Also:
- Validity requires ρ ≥ −1/(N−1). Perfect anticorrelation is possible only for N = 2, which gives σ_corr = 0 and the full ensemble rate μ.
- With many members, the benefit caps at μ − σ²ρ/2 (Eq. 5.13).
- The paper calls ρ a “covariance”. With unit-variance ξ it equals the correlation.
- Unequal members. The dossier asks “when leaders should share with laggards” without the answer, which is given in closed form:
- The cooperative’s rate is (1/N)Σ(μ_i − σ_i²/(2N)) (Eq. 5.5).
- Entity j gains by joining iff that exceeds μ_j − σ_j²/2 (Eq. 5.6).
- The group gains by admitting j iff that exceeds the (N−1)-member rate (Eq. 5.7).
- So a sufficiently superior member does better alone, and the paper names “skill differences” as a reason cooperation can be absent (§6).
- Framing vs the known literature. The dossier notes that ergodicity economics is contested. More to the point, the paper itself says:
- McCloskey (land scattering) and Winterhalder (forager food sharing) “obtain the same expression for the variance” (§5b).
- Uitdehaag studied “a model similar to ours”.
- Kennedy et al. treat it as a special case of a general cooperation model with relatedness (§1). The novelty is the time-average framing and the “baseline” claim, not the result.
- Dates. Received 30 Apr 2021, accepted 20 Oct 2021, issue date 11 Jul 2022, 380(2227). The dossier’s 2022-05-23 online date is from Crossref and was not contradicted. Whether arXiv:1506.03414 is an earlier form remains unverified; this text does not cite it.