Skip to content

Repository files navigation

Osterwalder-Schrader Axioms for the Gaussian Free Field

We construct the massive Gaussian Free Field (GFF) as a probability measure on the space of tempered distributions S'(ℝ^d), and prove that it satisfies all five Osterwalder-Schrader axioms for a Euclidean quantum field theory. The construction and proofs are formalized in Lean 4 / Mathlib, following the conventions and methods of proof in Glimm and Jaffe, Quantum Physics: A Functional Integral Point of View (Springer, 1987).

The library is dimension-generic: the spacetime dimension is a parameter d (any d ≥ 2), and the only per-dimension input is the closed form of the radial covariance profile, isolated behind the two-field typeclass GFFPropagator d m (see docs/dimension_generic.md). Four instances are provided in OSforGFF/Instances/: the four-dimensional Bessel kernel (m/4π²r)K₁(mr), the three-dimensional Yukawa kernel e^{−mr}/(4πr), the two-dimensional Bessel kernel (1/2π)K₀(mr), and the five-dimensional K_{3/2} kernel (1+mr)e^{−mr}/(8π²r³).

Master Theorem

theorem gaussianFreeField_satisfies_all_OS_axioms_generic
    {d : ℕ} [Fact (2 ≤ d)] (m : ℝ) [Fact (0 < m)] [GFFPropagator d m] :
    SatisfiesAllOS (gaussianFreeField_free (d := d) m)

theorem gaussianFreeField_satisfies_all_OS_axioms_dim4 (m : ℝ) [Fact (0 < m)] :
    SatisfiesAllOS (μ_GFF 4 m)

theorem gaussianFreeField_satisfies_all_OS_axioms_dim3 (m : ℝ) [Fact (0 < m)] :
    SatisfiesAllOS (μ_GFF 3 m)

theorem gaussianFreeField_satisfies_all_OS_axioms_dim2 (m : ℝ) [Fact (0 < m)] :
    SatisfiesAllOS (μ_GFF 2 m)

theorem gaussianFreeField_satisfies_all_OS_axioms_dim5 (m : ℝ) [Fact (0 < m)] :
    SatisfiesAllOS (μ_GFF 5 m)

where SatisfiesAllOS bundles OS0 (analyticity), OS1 (regularity), OS2 (Euclidean invariance), OS3 (reflection positivity), OS4 (clustering) and OS4 (ergodicity). The four _dim4/_dim3/_dim2/_dim5 theorems are the closed-form instances (Bessel K₁, Yukawa, K₀, K_{3/2}) of the first; the corollary gaussianFreeField_satisfies_all_OS_axioms_of_dim covers every d ≥ 2 via the canonical proper-time propagator: the OS3 proper-time Fubini domination runs at boundary-vanishing order d, so no upper bound on the dimension is needed.

CI

Status: Version 3.2 (general dimension d ≥ 2), August 2026. 0 sorries, 0 axioms, ~25,500 lines of Lean across 52 on-graph files (plus the off-graph OS/NonTrivial.lean and 6 off-graph Legacy/ files). Instances for d = 2, 3, 4, 5; the axiom footprint and statement type of every headline theorem (generic, all-dimensions d ≥ 2, d = 4, d = 3, d = 2, d = 5) are build-frozen in OSforGFF/Guardrails.lean.

All results are fully proved — no assumed axioms. Nuclear space structure and the Minlos theorem are provided by the external libraries bochner and gaussian-field, which are themselves axiom-free. The Minlos proof uses the external library kolmogorov_extension4.

Project Structure

The 52 on-graph library files (plus the off-graph OS/NonTrivial.lean and 6 off-graph Legacy/ files, below) are organized into 7 layers, with imports flowing from earlier to later sections (three exceptions, listed under "Cross-cutting dependencies" below). See docs/architecture.md for dependency structure, design choices, and proof outlines, and docs/dimension_generic.md for the dimension-generic design. For a pedagogical, axiom-by-axiom walkthrough of the OS proofs — ordered by complexity, with pointers into the code — see docs/pedagogical/Overview.md. The dependency graph source is in dependency/import_graph.dot (render with dot -Tsvg).


1. General Mathematics — OSforGFF/General/

Pure extensions of Mathlib with no project-specific definitions.

File Contents
FunctionalAnalysis L² Fourier transform infrastructure, Plancherel identity
SchurProduct Schur product theorem (Hadamard product preserves PSD)
HadamardExp Entrywise exponential of PSD matrices is PSD
PositiveDefinite Positive definite functions and kernels
GaussianRBF Gaussian RBF kernel exp(-‖x-y‖²) is positive definite
FourierTransforms 1D Fourier identities: Lorentzian ↔ exponential decay
LaplaceIntegral Laplace integral identity (Bessel K_{1/2}): ∫ s^{-1/2} e^{-a/s-bs} ds
BesselFunction The modified Bessel function K₁ via its cosh integral representation (def only; the analytic lemmas are in Legacy/)
BesselK0 The modified Bessel function K₀ via its cosh integral representation (def only)
BesselK The modified Bessel function K_ν of arbitrary order and the master Schwinger identity ∫ t^{ν−1} e^{−m²t−r²/4t} dt = 2(r/2m)^ν K_ν(mr); the K₀/K₁ evaluations and K_{1/2} are corollaries
QuantitativeDecay Schwartz bilinear forms with exponentially decaying kernels have polynomial decay
SchwartzTranslationDecay Schwartz seminorm bounds under translation
L2TimeIntegral L² bounds for time integrals: Cauchy-Schwarz, Fubini, Minkowski

2. Spacetime — OSforGFF/Spacetime/

Test functions, symmetries, and integration infrastructure.

File Contents
Basic SpaceTime (ℝ^d), SchwartzTestFunction, FieldConfiguration, distribution pairing
Euclidean Euclidean group E(d) = ℝ^d ⋊ O(d) and its action on test functions
DiscreteSymmetry Time reflection Θ: (t,x̄) ↦ (−t,x̄)
Decomposition Measure-preserving SpaceTime ≃ ℝ × ℝ^{d−1} decomposition
ComplexTestFunction Complex-valued Schwartz test functions and conjugation
PositiveTimeTestFunction Subtype of test functions supported at positive time
TimeTranslation Time translation operators T_s on Schwartz space
ProdIntegrable Integrability of Schwartz function products
Tonelli Tonelli/Fubini for Schwartz integrands on spacetime

3. Schwinger — OSforGFF/Schwinger/

Generating functionals and correlation functions.

File Contents
Defs Generating functional Z[J] = ∫ e^{i⟨φ,J⟩} dμ, Schwinger n-point functions
TwoPoint Two-point function S₂(x) as mollifier limit
GaussianMoments Gaussian moments: all n-point functions are integrable

4. Covariance — OSforGFF/Covariance/

The free scalar field propagator C(x,y) = Cprofile(|x−y|), isolated behind the GFFPropagator d m typeclass and analyzed through its proper-time (Schwinger) representation, uniformly in the dimension.

File Contents
Propagator The GFFPropagator typeclass, proper-time covariance, engine lemmas (L¹, decay, Fourier transform)
ParsevalGeneric Parseval identity ⟨f,Cf̄⟩ = ∫|f̂(k)|² P(k) dk, positivity, invariances, centered-kernel decay
RealForm Real covariance bilinear form, square root propagator embedding

5. Measure — OSforGFF/Measure/

Construction of the GFF probability measure via the Minlos theorem.

File Contents
NuclearSpace Schwartz space is Hilbert-nuclear and separable (bridges bochner + gaussian-field)
Minlos Minlos theorem application, Gaussian measure construction
MinlosAnalytic Symmetry and moments for Gaussian measures (sign-flip invariance, zero mean)
Construct GFF measure construction: covariance → characteristic functional → μ
IsGaussian Verification that S₂(f,g) = C(f,g) via OS0 derivative interchange
GaussianFreeField Main GFF assembly: gaussianFreeField_free m as a ProbabilityMeasure

Note: IsGaussian imports OS0_Analyticity because it uses the proved analyticity of Z[z₀f + z₁g] to identify S₂(f,g) = C(f,g) via the identity theorem. The dependency is on the OS0 result, not on OS0-specific infrastructure.


6. OS Axioms — OSforGFF/OS/

Axiom definitions, individual proofs, and master theorem.

File Contents
Axioms Formal Lean definitions of OS0 through OS4
OS0_Analyticity Closed-form Z[f] = exp(-½ C(f,f)) via identity theorem + Fernique
OS1_Regularity Plancherel + momentum-space bound: |Z[f]| ≤ exp(‖f‖²/2m²)
OS2_Invariance C(x,y) depends only on |x−y|, Lebesgue measure invariance
OS3_MixedRepInfra Schwinger parametrization and Fubini theorems for absolute integrability
OS3_MixedRep Mixed representation via Schwinger → heat kernel → Laplace transform
OS3_CovarianceRP Covariance reflection positivity: ⟨Θf, Cf⟩ = ∫ (1/ω)|F_ω|² ≥ 0
OS3_ReflectionPositivity Schur–Hadamard lifts covariance RP to generating functional
OS4_MGF Shared infrastructure: MGF formula, time translation duality
OS4_Clustering Gaussian factorization + convolution decay lemma (domain split at ‖y‖=‖x‖/2)
OS4_Ergodicity Polynomial clustering (α=6) → L² convergence
NonTrivial Nontriviality: C(f,f) > 0, positive variance, UV divergence C(x,y) → ∞ — deliberately off the root import graph; compiled by scripts/check-guardrails.sh, not lake build
Master Assembles OS0–OS4 into the generic master theorem, the all-dimensions corollary _of_dim, and the 4D, 3D, 2D, and 5D instances

7. Instances — OSforGFF/Instances/

Per-dimension closed forms of the covariance, packaged as GFFPropagator instances.

File Contents
Dim4 The GFFPropagator 4 m instance: Bessel kernel (m/4π²r)K₁(mr) via the order ν=−1 case of the master identity (schwingerIntegral_eq_besselK1)
Dim3 The GFFPropagator 3 m instance: Yukawa kernel e^{−mr}/(4πr) via the order ν=−1/2 case of the master identity (besselK_half), and the UV divergence
Dim2 The GFFPropagator 2 m instance: Bessel kernel (1/2π)K₀(mr) via the order-zero case of the master identity in General/BesselK
Dim5 The GFFPropagator 5 m instance: K_{3/2} kernel (1+mr)e^{−mr}/(8π²r³) via the order ν=−3/2 case of the master identity (besselK_three_half by Gaussian moments)

Legacy (off the build graph)

OSforGFF/Legacy/ preserves genuine proven mathematics that the on-graph library no longer consumes: the original four-dimensional development superseded in role by the dimension-generic machinery, and the verified-dead declarations quarantined by the library-wide sweep. These files are not imported by OSforGFF.lean and are not compiled by lake build; each carries a module docstring with its supersession map, and is verified in isolation with lake env lean (build BesselK1Analytics's olean first, since Dim4Bessel depends on it).

File Description
Legacy/Dim4Bessel The original 4D Bessel/momentum program: the named Bessel kernel freeCovarianceBessel/freeCovariance4, regulated-covariance / Fubini / momentum-space development, heat-kernel and Schwinger-representation defs, superseded by Covariance/Propagator.lean + Covariance/ParsevalGeneric.lean + General/BesselK.lean
Legacy/BesselK1Analytics The K₁ analytic lemmas (positivity, continuity, asymptotic/near-origin bounds, radial integrability) that supported the 4D analysis
Legacy/UnusedGeneral Consumer-less general-analysis lemmas (Fourier/functional-analysis side lemmas, the entire former FrobeniusPositivity.lean, the L² time-average and weighted-Minkowski programs)
Legacy/UnusedSpacetime Consumer-less spacetime-layer declarations (Schwartz multiplication, spatial L², the matrix presentation of time reflection, unified Euclidean actions, openness of the positive-time set)
Legacy/UnusedMeasureSchwinger Consumer-less measure/Schwinger-layer declarations (the alternative OS0_alt program, RBF/symmetry Minlos corollaries, the exponential-series expansion of the generating functional)
Legacy/UnusedOS Consumer-less OS-layer declarations (OS0 side lemmas, pre-H3 OS3 chain steps incl. bilinear_to_k0_inside, the ε–δ clustering formulation, kernel isometry invariance)

External Libraries

We depend on three auxiliary Lean libraries for nuclear space theory and measure construction. All are axiom-free.

bochner (BochnerMinlos)

Module What we use Imported by
Minlos.Main minlos_theorem — existence and uniqueness of probability measures from characteristic functionals on nuclear spaces Minlos
Minlos.NuclearSpace IsHilbertNuclear typeclass; MeasurableSpace (WeakDual ℝ E) cylinder σ-algebra instance Basic, NuclearSpace
Minlos.PietschBridge isHilbertNuclear_of_nuclear — bridge from Pietsch to Hilbert-Schmidt characterization NuclearSpace
Bochner.PositiveDefinite IsPositiveDefinite structure for characteristic functionals Minlos

gaussian-field (GaussianField)

Module What we use Imported by
SchwartzNuclear.HermiteNuclear schwartz_separableSpace — Schwartz space is separable (via Hermite basis) NuclearSpace
Nuclear.NuclearSpace DyninMityaginSpaceNuclearSpace — proves Schwartz space is nuclear NuclearSpace

kolmogorov_extension4 (transitive, via bochner)

Module What we use Imported by
KolmogorovExtension4.KolmogorovExtension projectiveLimit — Kolmogorov extension theorem: constructs a measure on the infinite product from a consistent projective family of finite-dimensional measures bochner's Minlos.ProjectiveFamily

Dependencies and Cross-Cutting Concerns

The import graph (dependency/import_graph.dot) is mostly layered, with three cross-cutting dependencies:

  1. IsGaussian → OS0_Analyticity: Gaussianity verification uses the OS0 analyticity result to identify S₂(f,g) = C(f,g) via the identity theorem (see Section 5 note)
  2. GaussianMoments → Measure/Construct: the moment bounds are stated for the constructed free-field measure
  3. Measure/GaussianFreeField → OS/Axioms: the OS2 statement it proves lives in the axioms module

These prevent a perfectly linear ordering but do not create a circular dependency.

Building

lake build

Requires Lean 4 and Mathlib (pinned via lake-manifest.json). The build also compiles OSforGFF/Guardrails.lean, whose #guard_msgs blocks freeze the axiom footprint and statement type of all six headline theorems (generic, all-dimensions d ≥ 2, and d = 4, 3, 2, 5) — so lake build fails if any change introduces a new axiom, leaks a sorry, or alters a headline statement.

A companion scripts/check-guardrails.sh checks the same invariant at the source level, without needing a build:

./scripts/check-guardrails.sh          # exit 0 = clean, exit 2 = violation

It scans every module reachable from OSforGFF.lean for axiom declarations, sorry/admit, and kernel escape hatches (native_decide, unsafe, implemented_by, extern), stripping comments first so prose that merely names them is not a false positive. The check is absolute rather than relative to a baseline revision, so it cannot silently pass by losing its reference point; set GUARDRAIL_BASE=<rev> to additionally report which violations a given range introduced. OSforGFF/Legacy/ is exempt — it is deliberately off the import graph and never compiled.

Both checks run in CI on every push and pull request (.github/workflows/ci.yml), which also replays the built environment through an external kernel check (leanchecker).

Related Work

  • or4nge19/OSforGFF — A fork by Matteo Cipollina pursuing a different measure construction pipeline: finite-dimensional Gaussians → Kolmogorov extension on test functions → nuclear support → pushforward to distribution space, avoiding the Minlos theorem. Develops coordinate-free Euclidean time-direction and dimension-agnostic Hermite APIs.

Planned Generalizations

  1. The d = 2 instance (the K₀ kernel (1/2π)K₀(mr)) — Done. The d = 2 (K₀), d = 3 (Yukawa), d = 4 (Bessel K₁), and d = 5 (K_{3/2}) instances are all provided, completing the dimensions discussed in docs/dimension_dependence.md.
  2. Explicit construction of the measure not using Minlos — Done. The Minlos theorem and Kolmogorov extension are now fully proved in bochner and kolmogorov_extension4.
  3. General dimension d ≥ 2 — Done. The OS3 Fubini domination now runs at boundary-vanishing order d (positive-time test functions are flat to all orders at the time boundary), so the OS theorems hold in every dimension d ≥ 2; the program and its implementation are written up in docs/general_dimension.md.

Authors

Sergey A. Cherkis, Michael R. Douglas, Sarah Hoback, Anna Mei, Ron Nissim

Coding Assistance

Claude Opus 4.6, Claude Fable 5, Gemini 3 Pro, GPT-5.2 Codex

License

This project is licensed under the Apache License, Version 2.0. See LICENSE for details.

References

  • Glimm, Jaffe: Quantum Physics (Springer, 1987), pp. 89–90
  • Osterwalder, Schrader: Axioms for Euclidean Green's functions I & II (1973, 1975)
  • Gel'fand, Vilenkin: Generalized Functions, Vol. 4 (Academic Press, 1964)
  • Reed, Simon: Methods of Modern Mathematical Physics, Vol. II (1975)
  • Degenne, Pfaffelhuber: Formalizing the Kolmogorov Extension Theorem in Lean (kolmogorov_extension4)

About

A Lean 4 formalization of the Gaussian Free Field in d=4 and proof of the Osterwalder-Schrader axioms

Resources

Stars

25 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages