β: The Universal Temperature
A single dial — the inverse temperature β — interpolates between two different arithmetics. Every hard threshold, winner-takes-all decision, and classical logic gate is β → ∞. Every soft, probabilistic, thermodynamic system is finite β. The passage between them is the Maslov dequantization.
Table of contents
- The central idea
- Six famous equations — one formula
- The two semirings in detail
- What changes as β varies
- The Maslov-Gibbs Einsum (MGE)
- Why “Maslov dequantization”?
- Is this phenomenon unique?
- Where to go from here
The central idea
There are two natural ways to add and multiply non-negative numbers:
| Semiring | Addition | Multiplication | Name |
|---|---|---|---|
| Probabilistic | a + b | a × b | Gibbs / Boltzmann |
| Tropical | max(a, b) | a + b | Tropical / (max,+) |
These look like different mathematical structures. The Maslov dequantization says they are the same structure at different temperatures. Specifically:
\[a \oplus_\beta b \;=\; \frac{1}{\beta} \log\!\left(e^{\beta a} + e^{\beta b}\right)\]- At β → 0: this approaches ½(a + b) — ordinary average (smooth Hodge limit)
- At β = 1: this is the log-sum-exp softmax — standard probabilistic arithmetic
- At β → ∞: this approaches max(a, b) — the tropical arithmetic
The tropical semiring is the zero-temperature limit of ordinary arithmetic. The passage from finite β to β → ∞ is called the Maslov dequantization — named for V. P. Maslov, who showed in the 1980s that tropical mathematics arises systematically as the ℏ → 0 (or β → ∞) limit of quantum/statistical mechanics.
The single most important consequence: every algorithm, model, or physical system that uses a hard threshold — argmax, winner-takes-all, on/off logic phase transitions — is implicitly operating at β → ∞. Replacing that hard threshold with finite β softens it, makes it differentiable, and connects it to the probabilistic semiring. This is not an approximation. It is the correct generalisation, of which the hard threshold is the zero-temperature special case.
Six famous equations — one formula
The same MGE expression at different β:
\[\pi_k(\beta) = \frac{e^{-\beta E_k}}{\sum_j e^{-\beta E_j}}\]| Field | Equation | β meaning |
|---|---|---|
| Machine learning | Softmax(x/T) | T = 1/β; temperature of attention / sampling |
| Statistical mechanics | Boltzmann distribution | β = 1/k_BT; inverse thermal energy |
| Finance | Black-Scholes risk-neutral measure | β = 1/σ²; inverse variance |
| Quantum mechanics | Path integral e^{iS/ħ} | β = it/ħ; Wick-rotated |
| Optimisation | Simulated annealing schedule | β(t) increasing; cooling toward β→∞ |
| Information theory | Maximum entropy at fixed energy | β = Lagrange multiplier |
All six are the same formula. The ML engineer who tunes the softmax temperature the physicist computing a partition function, and the quant pricing options are all turning the same dial.
Remark — the bridge variable has many names. The table above shows β appearing directly. But in several classical fields it travels under an alias: in quantum mechanics it is ħ (Planck’s constant); in fluid dynamics it is ν (kinematic viscosity); in mathematical finance it is σ² (volatility); in optimal transport it is ε (regularisation strength). These are all β in disguise — the same dial, measured in different units. The Cole-Hopf transformation (heat equation ↔ Burgers shocks, bridge: ν), the Black-Scholes / Hamilton-Jacobi duality (bridge: σ²), and the WKB semiclassical limit (Schrödinger ↔ Hamilton-Jacobi, bridge: ħ) are all instances of the same Maslov dequantization. None of this is new individually — what is new is that it is all one thing.
Paper: β in Disguise — five classical dualities proved to be the same algebraic structure; ħ = ν = σ² = β.
The two semirings in detail
The probabilistic semiring: finite β
At finite β, the MGE assigns a smooth probability to every outcome. The arithmetic is the familiar (+, ×) of real numbers. Key properties:
- Differentiable: ∂π_k/∂E_k exists everywhere
- All outcomes contribute: no outcome has exactly zero weight
- Entropy is positive: H(π) = −Σπ_k log π_k > 0
- Gradients flow: backpropagation works; the system can be optimised
This is the regime of neural networks, Bayesian inference, statistical physics and chemical kinetics. It is the regime where learning happens.
The tropical semiring: β → ∞
As β → ∞ the soft probability collapses to a hard indicator:
\[\lim_{\beta\to\infty} \pi_k(\beta) = \begin{cases} 1 & k = \arg\min_j E_j \\ 0 & \text{otherwise} \end{cases}\]The arithmetic becomes (max, +): addition becomes max, multiplication becomes addition. Key properties:
- Non-differentiable: the argmax has zero gradient almost everywhere
- One outcome wins: all weight concentrates on the minimum-energy state
- Entropy is zero: H(π) = 0 at β → ∞
- No gradients: classical algorithms, lookup tables, discrete logic
This is the regime of classical computers, database queries, shortest paths and discrete optimisation. It is the regime where answers are stored.
The β* snap: the transition between them
Between the two semirings lies a snap threshold β*, the point where the system transitions from smooth probabilistic to sharply discrete behaviour. This threshold is not arbitrary — it is determined by the topology of the problem:
\[\beta^* = \frac{3}{8} \ln\frac{1}{1-\rho}\]where ρ is the load factor of the constraint graph. Below β: smooth, learning exploring. Above β: crystallised, decided, locked in.
Every hard threshold in science is β* in disguise:
| Threshold | Domain | β* interpretation |
|---|---|---|
| p-value 0.05 | Statistics | β→∞ of a soft evidence threshold |
| Metropolis acceptance rate 0.234 | MCMC | Optimal β* for dimension-free sampling |
| DFT/CASSCF handoff (c₂ = 0.88) | Quantum chemistry | β*₀₁ for the H⁰↔H¹ tier boundary |
| PT phase transition (ε = ε_c) | Non-Hermitian physics | β*₁₂ for the H¹↔H² tier boundary |
| Softmax temperature in LLMs | ML | β* calibration; induction head snap |
| Kelly criterion | Finance | β* = 1/σ² separating ruin from growth |
These all look like domain-specific numbers. They are all the same saddle-point equation evaluated in different units.
Paper: In Praise of Soft Thresholds — the unification of hard thresholds as T→0 limits; why finite-β is always the correct generalisation.
What changes as β varies
The Maslov dequantization is not just a mathematical curiosity — it changes what kind of computation is possible:
| β regime | Semiring | What you can do | What you cannot |
|---|---|---|---|
| β → 0 | Smooth (Hodge) | Global relaxation; Hodge decomposition; optimal transport | Local decisions |
| 0 < β < β* | Gibbs (exploratory) | Learn; backpropagate; explore; sample | Commit to an answer |
| β = β* | Snap threshold | Maximum information throughput | — |
| β* < β < ∞ | Gibbs (crystallising) | Refine; anneal; sharpen | Revise global structure |
| β → ∞ | Tropical (max,+) | Decide; retrieve; run discrete algorithms | Learn from errors |
| β = it | Complex (unitary) | Quantum interference; Berry phase | Dissipation |
| β ∈ ℂ | PT-symmetric | Gain-loss dynamics; exceptional points | Pure unitary evolution |
The key insight: different computations require different β regimes. A neural network must operate at finite β to learn. A database query must operate at β → ∞ to give a definite answer. A quantum computer operates at β = it. A PT-symmetric sensor operates near β* ∈ ℂ. Mixing up the regimes is the source of most failures in both AI and physics.
The Maslov-Gibbs Einsum (MGE)
The MGE is the single operation that unifies the two semirings:
\[\text{MGE}(\mathbf{E}, \beta) = \frac{1}{\beta} \log \sum_k e^{-\beta E_k}\]This is the free energy at inverse temperature β. Its β → ∞ limit is the minimum energy (tropical argmin). Its β → 0 limit is the average energy (arithmetic mean). At β = 1 it is the log-partition function of statistical mechanics.
The MGE is semiring-polymorphic: it evaluates the same programme over different arithmetic depending on β. This is why the same formula appears in six different fields — they are all evaluating the same programme over the semiring appropriate to their domain.
The operational consequence for the ISA: every opcode in the Origami ISA has a β-parameterised version. At β → ∞ it runs over (max,+) — classical discrete. At finite β it runs over Gibbs — statistical, differentiable. The same ISA programme, run at different β, gives different answers and uses different computational resources. This is what we mean by a differentiable algorithm: not that the algorithm has been approximated, but that its natural parameter β has been set to a finite value rather than ∞.
Papers: The Maslov-Gibbs Einsum — the foundational paper; tropical crystallisation and the thermodynamic bridge β in Disguise
Why “Maslov dequantization”?
V. P. Maslov observed in the 1980s that tropical mathematics (the (max,+) semiring) arises as the classical limit of quantum mechanics in precisely the same way that classical mechanics arises from quantum mechanics as ħ → 0.
The Schrödinger equation at finite ħ becomes the Hamilton-Jacobi equation at ħ → 0. The path integral Σ e^{iS/ħ} becomes the saddle-point e^{iS_cl/ħ} at ħ → 0. The quantum partition function Tr[e^{-βH}] becomes the tropical partition function max(-βE) at β → ∞.
In each case: a sum over all paths/states, weighted by a Boltzmann-like factor, collapses to the single dominant contribution as the parameter goes to its extreme value. The Maslov dequantization is the name for this limit, and the inverse — going from the tropical/classical limit back to finite β — is the quantization in the other direction.
The HotLogiQ claim: this dequantization/quantization pair is not specific to quantum mechanics. It applies to:
- Every optimisation algorithm (gradient descent ↔ greedy argmax)
- Every probabilistic model (soft classifier ↔ hard decision boundary)
- Every physical system with a phase transition (paramagnetic ↔ ferromagnetic)
- Every neural network (learning ↔ inference)
β is the universal quantization parameter. The MGE is the universal quantization map. The tropical semiring is what you get when you forget β entirely.
Is this phenomenon unique?
A natural reaction to Maslov dequantization is that it is bizarre: you vary a parameter smoothly and the algebra underneath the model changes. Three questions follow, and they have different answers.
What exactly is preserved?
Not the equation. The heat equation and Hamilton–Jacobi visibly differ — a diffusion term vanishes in the limit. What is preserved is linearity:
| equation | superposition | |
|---|---|---|
| finite β | heat / Schrödinger, linear over $(+,\times)$ | $u = u_1 + u_2$ |
| β → ∞ | Hamilton–Jacobi, linear over $(\min,+)$ | $S = \min(S_1, S_2)$ |
So the precise statement is that the solution set stays closed under the semiring operations, and which semiring depends on β. Maslov called this idempotent superposition.
Is log-sum-exp the only deformation that does it?
Essentially yes, up to a change of coordinates. The candidates are:
| family | works? | note |
|---|---|---|
| $h\log\sum e^{x/h}$ | yes | the standard soft-min |
| $(\sum x^p)^{1/p}$ | yes | the $p$-norm |
| power means | yes | $\to\max/\min$ as $p\to\pm\infty$ |
| plain sum | no | no parameter, no limit |
| median, mode | no | not associative |
But the $p$-norm is log-sum-exp conjugated by $\exp$ — verified numerically: soft-min at $h$ and the $p$-norm at $p = 1/h$ agree to six decimals. So there is one family wearing two costumes, not two families.
A related question, sometimes confused with this one: can the JSD or Rényi entropy deform the semiring? No, and the reason is a category distinction rather than a limitation. A semiring deformation must supply two binary operations satisfying associativity, commutativity and distributivity. The JSD and $H_\alpha$ are functionals — they eat distributions and return scalars. They measure structure; they do not constitute it. Asking whether the JSD can change the semiring is like asking whether the variance can change addition.
A worked example of the other kind. The distinction is easier to see alongside a good instance of the functional case. Salazar, Saxena, Baker, Kwek & Kyaw, Stabilizer Statistical Mechanics (arXiv:2608.14798, 2026) build a partition function over the Pauli spectrum of a quantum state and derive from its free energy a magic monotone, the stabilizer work. It carries an inverse-temperature parameter, and as that parameter runs from high to low the monotone interpolates continuously between the stabilizer 2-Rényi entropy and the stabilizer nullity — two previously separate measures, recovered as the two limits of one object.
That is a β-parameterised family with known quantities at each endpoint, which is the same shape as the deformation described on this page. But it is a family of functionals over a fixed algebra, not a deformation of the algebra itself: no operation changes, and nothing becomes idempotent. The two are worth holding apart, because the surface similarity is strong and the mathematical content is different.
Their paper is also unusually careful about the word temperature — the parameter is repeatedly described as “inverse-temperature-like” and “a moment-resolving parameter… rather than a physical temperature.” That is the correct register for a parameter that indexes a family without being thermal, and it applies to β on this page too.
Where else does this happen?
The phenomenon is not unique to Maslov. It has a general name — algebraic deformation and contraction, with Gerstenhaber deformation theory as the formal setting — and several familiar instances:
| phenomenon | parameter | what changes |
|---|---|---|
| Maslov dequantization | $h \to 0$ | $(+,\times) \to (\min,+)$ |
| deformation quantization | $h \to 0$ | Moyal $\star \to$ commutative product — the mirror of Maslov |
| İnönü–Wigner contraction | $c \to \infty$ | Poincaré $\to$ Galilei |
| $q$-deformation | $q \to 1$ | $U_q(\mathfrak g) \to U(\mathfrak g)$ |
| tropical geometry | Litvinov–Maslov | algebraic $\to$ tropical variety |
| large-$N$ | $N \to \infty$ | matrices $\to$ free probability |
What is genuinely unusual about the Maslov case
Most deformations move continuously within a type: associative algebras stay associative, Lie algebras stay Lie algebras. Maslov dequantization does something stronger. In the limit \(a \oplus a = a\)
addition becomes idempotent — and an idempotent addition has no inverses. So the limit object is not a ring at all.
You lose subtraction. That is the strange part, and it is not shared by the other deformations in the table. It is also why the tropical limit is a genuine qualitative jump rather than a smooth degeneration, and why “hard threshold” and “soft threshold” are not merely quantitatively different — one lives in a structure where differences exist, the other in one where they do not.
Where to go from here
This page describes the mathematical foundation. The applications branch into three directions:
For AI and machine learning: AI & Machine Learning — softmax temperature = β; transformers as ISA programmes; layerwise β profiling; differentiable Shapley values; grokking as β* snap.
For physics and non-Hermitian systems: PT Symmetry & Exceptional Points — exceptional points as β*₁₂ snaps; SNAP-count as the EP topological invariant; the PT phase transition as the H¹↔H² tier boundary.
For aggregation and multi-source reasoning: Disagreement — the companion tool. The MGE says how to combine at a temperature; the JSD says how much is lost in combining. Note the two are different kinds of object, which is why neither substitutes for the other: see “Is this phenomenon unique?” above.
For the full ISA picture: The β-plane — how β extends into the complex plane (β = it for quantum mechanics, β ∈ ℂ for PT-symmetric systems); the full ISA family; the snap threshold in detail.
Primary papers:
- The Maslov-Gibbs Einsum — start here for the mathematics
- β in Disguise — start here if you want the intuition first
- In Praise of Soft Thresholds — why every hard threshold is β→∞
- — the full complex β parameter space