Thermodynamic Information Routing: A Unified Framework for Gibbs Aggregation Across Economics, Computation, and Knowledge Retrieval: An Accessible Guide

Plain-language explainer for doi:10.5281/zenodo.20237288 (#294)

The central idea in one sentence

Eight independent research programmes across economics, neuroscience, statistical mechanics, and machine learning independently derived the same mathematical object — the Gibbs distribution — from domain-specific first principles, and this convergence is not a coincidence: there is a unique routing primitive that simultaneously preserves three geometric conservation laws.

Eight independent derivations of the same formula

The Gibbs distribution is:

\[w_i(\beta) = \frac{e^{\beta U_i}}{\sum_j e^{\beta U_j}},\]

where $U_i$ is the utility (or energy, or value) of option $i$ and $\beta > 0$ is a concentration parameter (inverse temperature, rationality, or signal-to-noise ratio). Here are eight research traditions that derived exactly this formula independently:

McFadden (1974, 2000 Nobel): Logit model from utility maximisation with Gumbel-distributed noise. The extreme-value distribution is the unique distribution whose maximum is again Gumbel — a fixed-point property that forces $w_i \propto e^{\beta U_i}$.
Sims (2003, 2011 Nobel): Rational inattention. An agent with a Shannon entropy budget $\kappa$ maximises expected utility subject to $H(w) \leq \kappa$. The unique solution is $w_i \propto e^{\lambda U_i}$ where $\lambda$ is the Lagrange multiplier on the entropy constraint.
McKelvey and Palfrey (1995): Quantal Response Equilibrium. Game-theoretic equilibrium with noise: players best-respond imperfectly, with mistake probability decreasing in the cost of the mistake. The unique symmetric equilibrium is the logit QRE: $w_i \propto e^{\beta U_i}$.
Jaynes (1957): Maximum Entropy (MaxEnt). The unique distribution that maximises Shannon entropy subject to a constraint on mean utility $\langle U \rangle = \bar{U}$ is the Gibbs distribution with $\beta$ as the Lagrange multiplier.
Gibbs (1902): Statistical mechanics. The canonical ensemble — the distribution of states for a system in thermal contact with a heat bath at temperature $T = 1/k_B \beta$ — is $w_i \propto e^{-\beta E_i}$, where $E_i$ is the energy of state $i$.
Goel, Saleh, and colleagues (2012): Molecular motors. Optimal energy transduction in biochemical systems: a motor protein allocating chemical energy among conformational states maximises work output subject to thermodynamic consistency. The unique solution is the Gibbs distribution.
Friston (2010): Free energy principle in neuroscience. The brain minimises a variational free energy functional $F = \langle E \rangle - H$. At the minimum, the recognition distribution is the Gibbs distribution with precision $\beta$ as a free parameter.
ASA/MGE: The Maslov-Gibbs Einsum. The tropical ($\beta \to \infty$) limit recovers the (min, +) or (max, +) semiring of tropical geometry; the Gaussian ($\beta \to 0$) limit recovers averaging. The Gibbs distribution is the unique interpolation between these limits that is compatible with the symplectic structure of Hamiltonian flow.

Why the convergence is not a coincidence

The paper identifies the three conservation laws that uniquely force the Gibbs form:

Conservation 1: Conformal invariance (scale-freedom in utility units). The routing primitive $w_i(\beta)$ must be invariant under linear rescaling of utilities: $U_i \to \alpha U_i + c$ should rescale $\beta$ to $\beta/\alpha$ and shift normalisation, without changing the qualitative structure. This rules out, for example, $w_i \propto U_i^n$ (power-law routing) for any $n \neq 1$ when utilities are negative.

Conservation 2: Symplectic structure (Hamiltonian flow). The routing must be the stationary distribution of a Hamiltonian dynamical system — a system that conserves the volume of phase space (Liouville’s theorem). This rules out any dissipative routing rule. The unique stationary distribution of a Hamiltonian system in contact with a heat bath is the Gibbs distribution.

Conservation 3: Adiabatic invariance ($\beta$-schedule tracking). When $\beta$ changes slowly (adiabatically), the routing must track the instantaneous Gibbs distribution without producing entropy. This is the condition for reversible computation — the $\beta$-ramp can be run forwards (deliberation) and backwards (forgetting) without information loss. The Gibbs distribution is the unique routing primitive with this property.

The three conservation laws together form the TIR axiom system. The main theorem of the paper is: the Gibbs distribution is the unique routing primitive satisfying all three axioms, for any choice of geometry $G$ over the option space.

The geometry $G$: the only domain-specific ingredient

Different applications differ only in their choice of geometry $G$ — the space over which the options ${1, \ldots, N}$ are defined and the metric on that space. The paper classifies four types:

Abelian geometry (flat, commutative): Options are independent. The Gibbs distribution factorises over options. This is the standard logit model of McFadden, the MaxEnt distribution of Jaynes, the canonical ensemble of Gibbs.
Fano geometry (seven points, non-commutative): Options have octonion-like structure. The routing must respect the Fano plane incidence relations. This geometry appears in ASA’s computation framework.
$G_2$ geometry (exceptional Lie group): Options are related by the symmetries of the octonions. This geometry appears in the exceptional holonomy connections of Paper 285.
Catalan geometry (tree-structured, hierarchical): Options are arranged in a binary tree; the routing is a hierarchical softmax. This geometry appears in knowledge retrieval (hierarchical softmax in word2vec) and in the Parisi ultrametric tree for spin glasses.

Escaping Arrow and Condorcet

Arrow’s Impossibility Theorem (1951): no social choice function mapping utility profiles to a preference ranking satisfies all four of Arrow’s axioms simultaneously. The Condorcet paradox: pairwise majority voting can produce cycles even when individual preferences are transitive.

TIR escapes both, cleanly. The reason:

Arrow’s theorem applies to functions mapping utility profiles to rankings (total orders). TIR produces a probability measure over options, not a ranking. The four Arrow axioms are not well-typed for probability measures.
The Condorcet paradox arises from pairwise tournaments — binary comparisons. TIR makes no binary comparisons: all options coexist simultaneously in the Gibbs distribution. No tournament is held; no cycle can form.

The log-partition function $\mathcal{W}(\beta) = \beta^{-1} \ln Z(\beta)$ is a well-defined, differentiable scalar welfare measure. It is not a ranking; it is a free energy. Social choice in TIR is not about picking a winner — it is about characterising the distribution.

Information routing in retrieval systems

The paper’s third application domain is knowledge retrieval. A retrieval system routes a query $q$ to a set of candidate documents ${d_1, \ldots, d_N}$ with scores $U_i = \text{sim}(q, d_i)$. The standard retrieval model uses the softmax over scores — the Gibbs distribution at $\beta = 1$ — to produce a probability ranking.

TIR provides the missing justification: the softmax is the unique retrieval primitive satisfying conformal invariance (doubling all similarities does not change qualitative ranking), symplectic structure (the ranking is derived from a Hamiltonian similarity function), and adiabatic invariance (smoothly increasing $\beta$ from 0 to $\infty$ implements beam search without information loss).

The temperature $\beta$ is the “retrieval temperature”: at $\beta \to 0$, all documents are returned with equal probability (maximum diversity, zero precision); at $\beta \to \infty$, only the top-ranked document is returned (maximum precision, zero diversity). The optimal $\beta$ balances precision and recall according to the specific retrieval task.