Thermodynamic Semirings and Entropy Algebras

Sources

Context and Motivation

The cohomological and operadic approaches (see Categorical Entropy) characterize entropy via universal properties in categories of probability spaces. This thread asks a different question: can entropy be constructed as the output of a canonical algebraic operation, rather than characterized axiomatically?

The answer from Marcolli-Thorngren is yes: entropy is the first-order deformation of tropical arithmetic at zero temperature. The tropical semiring \((\mathbb{R}, \min, +)\) governs zero-temperature statistical mechanics; raising the temperature deforms \(\min\) into log-sum-exp, and the derivative of this deformation with respect to temperature is Shannon entropy.

This places entropy inside algebraic geometry (Witt vectors, deformation theory) and renormalization theory (Connes-Kreimer Hopf algebras, Birkhoff factorization) rather than in sheaf theory or operadic algebra.

Thermodynamic Semirings

The tropical semiring and its deformation

A semiring \((R, \oplus, \otimes)\) satisfies all ring axioms except the existence of additive inverses. The tropical semiring is

\[(\mathbb{R} \cup \{+\infty\},\ \min,\ +),\]

which encodes zero-temperature statistical mechanics: the “sum” of two energies \(a, b\) is the energy of the ground state \(\min(a, b)\).

At finite inverse temperature \(\beta > 0\), the ground-state selection is softened into the log-sum-exp operation:

\[a \oplus_\beta b := -\frac{1}{\beta} \log\bigl(e^{-\beta a} + e^{-\beta b}\bigr).\]

As \(\beta \to \infty\) (zero temperature), \(a \oplus_\beta b \to \min(a, b)\), recovering the tropical semiring. As \(\beta \to 0\) (infinite temperature), \(a \oplus_\beta b \to -\frac{1}{\beta}\log 2 + \frac{a+b}{2}\), which diverges — the high-temperature limit degenerates.

Entropy as a deformation derivative

🔑 The key observation of Marcolli-Thorngren: expand \(\oplus_\beta\) as a perturbation of the tropical \(\oplus_\infty = \min\) as \(\beta^{-1} \to 0^+\). Setting \(T = \beta^{-1}\):

\[a \oplus_T b = \min(a, b) - T \cdot H_2\!\left(\frac{e^{-a/T}}{e^{-a/T} + e^{-b/T}}\right) + O(T^2),\]

where \(H_2(p) = -p\log p - (1-p)\log(1-p)\) is the binary entropy function. Shannon entropy is the coefficient of the first-order temperature correction to tropical addition.

More generally, for a distribution \(p = (p_1, \ldots, p_n)\) at inverse temperature \(\beta\), the free energy

\[F_\beta(p) = -\frac{1}{\beta} \log \sum_i e^{-\beta E_i}\]

satisfies \(\partial F_\beta / \partial T\big|_{T=0} = H(p_\beta)\) where \(p_\beta\) is the Boltzmann distribution. Entropy appears as the thermodynamic derivative \(\partial F / \partial T\) — which is its classical definition. Marcolli-Thorngren observe this is not just a physics coincidence but an algebraic statement about the deformation theory of semirings.

Witt vectors and the algebraic framework

The family \(\{(\mathbb{R}, \oplus_\beta, +)\}_{\beta > 0}\) forms a one-parameter family of semirings. The algebraic object encoding all of these simultaneously is a Witt vector semiring: a lift of the tropical semiring from characteristic \(\infty\) (the tropical limit) to finite characteristic (finite temperature).

In the classical Witt vectors construction, one lifts a ring \(R\) of characteristic \(p\) to a characteristic-zero ring \(W(R)\) carrying a Frobenius lift. Here, the analogous construction replaces:

Key structure: The entropy axioms — positivity, symmetry, chain rule — correspond to the axioms of a semiring homomorphism from the Witt vector semiring \(W_\beta(\mathbb{R})\) to \((\mathbb{R}, +, \cdot)\). This is the Marcolli-Thorngren analogue of the Baudot-Bennequin cocycle condition: entropy is “flat” in the Witt vector sense.

Entropy Algebras and Birkhoff Factorization

The Connes-Kreimer Hopf algebra

In perturbative QFT, renormalization is the systematic removal of UV divergences from Feynman diagram integrals. Connes and Kreimer showed this process is controlled by a Hopf algebra \(\mathcal{H}_{CK}\) of rooted forests, where:

• Multiplication: disjoint union of forests
• Comultiplication: \(\Delta(T) = \sum_{C \subseteq T} P_C(T) \otimes R_C(T)\), summing over admissible cuts \(C\) of the tree \(T\), where \(P_C(T)\) is the “pruned” part above the cut and \(R_C(T)\) is the remaining trunk.
• Antipode: encodes the BPHZ counterterms

The key theorem is that divergences in QFT are organized by the combinatorics of nested subgraphs, and the Hopf algebra captures this nesting structure algebraically.

Entropy as a Hopf algebra character

Marcolli and Tedeschi (2014) transplant this structure to information theory. A character of \(\mathcal{H}_{CK}\) is an algebra homomorphism \(\phi: \mathcal{H}_{CK} \to \mathbb{C}\). Their main observation:

Shannon entropy is a character of \(\mathcal{H}_{CK}\). Explicitly, given a rooted probability tree \(T\) encoding a hierarchical probability model — with each internal node \(v\) carrying a conditional distribution \((p_1^v, \ldots, p_{k_v}^v)\) over its children — define

\[\phi_H(T) := \sum_{v \in T} H(p_1^v, \ldots, p_{k_v}^v).\]

This assigns to each tree the total entropy of all conditional distributions in the hierarchy. The claim is that \(\phi_H\) is multiplicative with respect to the Hopf algebra structure — i.e., a character.

Birkhoff factorization of entropy

A central theorem of the Connes-Kreimer theory is Birkhoff factorization: every character \(\phi: \mathcal{H}_{CK} \to \mathbb{C}\) admits a unique decomposition

\[\phi = \phi_- \star \phi_+\]

where \(\star\) is convolution in the Hopf algebra, \(\phi_+\) is a “renormalized” character (finite, no poles), and \(\phi_-\) is the counterterm character (encodes the divergences being removed). In QFT, \(\phi_-\) is the BPHZ counterterm and \(\phi_+\) is the renormalized Feynman amplitude.

Applied to \(\phi_H\): the Birkhoff factorization of the entropy character extracts, from a hierarchical probability model, a canonical “renormalized” entropy \((\phi_H)_+\) and a “counterterm” \((\phi_H)_-\).

The physical interpretation: in a tree of conditional distributions, there is redundant entropic content (correlations that appear at multiple levels of the hierarchy). Birkhoff factorization removes this double-counting canonically — just as renormalization removes divergences from nested Feynman diagrams.

Gamma Spaces and Stable Homotopy

Segal’s \(\Gamma\)-spaces

Marcolli (2018) lifts the entropy story to stable homotopy theory. Recall: a \(\Gamma\)-space (Segal 1974) is a functor

\[X : \Gamma^{op} \longrightarrow \mathbf{Top}\]

where \(\Gamma\) is the category of finite pointed sets \(\{0, 1, \ldots, n\}\) with basepoint \(0\). \(\Gamma\)-spaces model \(E_\infty\) spaces — spaces with a multiplication that is commutative and associative up to all higher coherences. Segal’s theorem: every \(\Gamma\)-space with a group-completion gives a connective spectrum (an object of stable homotopy theory).

The slogan: \(\Gamma\)-spaces are the correct homotopy-theoretic notion of “commutative monoid.”

Information loss as a \(\Gamma\)-space

The claim of Marcolli (2018) is that the information loss functor

\[F : \mathbf{FinProb} \to (\mathbb{R}_{\geq 0}, +)\]

naturally extends to a \(\Gamma\)-space structure. The rough idea:

• The category \(\mathbf{FinProb}\) of finite probability spaces carries a symmetric monoidal structure (product of probability spaces).
• A symmetric monoidal functor to an abelian group is the data of a \(\Gamma\)-space.
• The BFL information loss functor is (essentially) symmetric monoidal, hence defines a \(\Gamma\)-space.

The associated spectrum \(\mathbf{I}\) has: - \(\pi_0(\mathbf{I})\) = the group completion of \((\mathbb{R}_{\geq 0}, +)\), i.e. \(\mathbb{R}\) - Higher \(\pi_n(\mathbf{I})\) = (conjecturally) higher information-theoretic invariants

This is the most structurally ambitious claim in this thread: entropy as a map of spectra in the sense of stable homotopy theory.

Open Questions