Osterwalder--Schrader Axioms for Neural Network Field Theories

Computational Verification and Architectural Conditions for Physical Consistency of NN-FTs in Dimensions d ≥ 2

1,271
Parameter Configs Scanned
5.1%
RP-Admissible (d=2)
5
OS Axioms Verified
20K
Monte Carlo Samples

Problem Statement

Neural network field theories are universal but not all are physically consistent. Which NN-FTs satisfy the Osterwalder--Schrader axioms?

The Challenge

Neural network field theories (NN-FTs) provide a universal framework for representing Euclidean quantum field theories. The universality theorem establishes that any Euclidean QFT admits a neural network representation. However, universality alone does not ensure physical consistency. Euclidean QFTs must satisfy the Osterwalder--Schrader (OS) axioms to guarantee analytic continuation to a unitary Lorentzian theory. Extending reflection positivity results from d=1 to d ≥ 2 remained an open problem.

The Five OS Axioms

OS0 Regularity -- Schwinger functions are finite
OS1 Euclidean Covariance -- Translation invariance
OS2 Reflection Positivity -- The most subtle axiom
OS3 Symmetry -- S2(x,y) = S2(y,x)
OS4 Cluster Property -- Correlator decay with distance

Methods

Three families of NN-FT architectures verified using lattice discretization, spectral analysis, and Monte Carlo estimation.

G

Gaussian NN-FTs

Propagator with neural network kernel corrections in momentum space

T

Transfer Operator

Layered architectures with depth aligned to Euclidean time direction

I

Interacting Theories

φ4 NN-FT with Monte Carlo importance sampling

KL

Spectral Analysis

Non-negative least squares for the spectral representation

Gaussian NN-FT Reflection Positivity

Direct lattice RP check on a 10x10 lattice with m2=1.0. Linear momentum corrections preserve RP; nonlinear corrections violate it.

Correction f(p2) RP Status λmin KL Residual nρ>0
None (free field) Pass -1.73e-16 2.77e-05 2
0.2 · p2 Pass -2.55e-16 4.23e-05 2
-0.3 · p2 Pass -2.30e-16 3.19e-08 2
-0.8 · p2 Pass -1.27e-16 6.72e-05 2
0.5 · softplus(p2) Fail -3.03e-03 1.68e-02 1
0.1 · softplus(0.5p2) Fail -3.43e-04 1.85e-03 2
0.5 · sin(p2) Fail -6.09e-03 4.44e-02 4
Minimum Eigenvalue by Correction Type
KL Residual Comparison (log scale)

Spectral Density Analysis

The spectral density ρ(m2) from the non-negative least squares fit reveals whether the propagator admits a valid spectral representation.

Architecture Space Scan

RP landscape in the (α, β) parameter space of the correction f(p2) = α · softplus(β · p2). Only 65 of 1,271 configurations (5.1%) are RP-admissible.

RP-Admissible vs Violated by α
RP Fraction (%) by α

Transfer Operator Analysis

RP analysis across 120 architectural configurations. Only linear activation with unconstrained weights achieves RP (0.8% overall).

RP Fraction by Activation & Weight Constraint
Mean Min Eigenvalue by Activation
Activation W free (RP / Total) W = VTV (RP / Total) Total RP RP Fraction
Linear 1 / 15 (6.7%) 0 / 15 (0%) 1 / 30 3.3%
ReLU 0 / 15 (0%) 0 / 15 (0%) 0 / 30 0%
Tanh 0 / 15 (0%) 0 / 15 (0%) 0 / 30 0%
Softplus 0 / 15 (0%) 0 / 15 (0%) 0 / 30 0%
Total 1 / 60 (1.7%) 0 / 60 (0%) 1 / 120 0.8%

Interacting φ4 NN-FT OS Axioms

OS axiom verification for φ4 NN-FT on a 6x6 lattice with m2=1.0 and 20,000 Monte Carlo samples.

OS Axiom Compliance Heatmap
Coupling vs Min RP Eigenvalue (8x8 lattice)
λ OS0 OS1 OS2 λminRP OS3 OS4 neff

Dimensional Scaling

RP-admissible fraction increases dramatically from d=2 to d=4, partly reflecting reduced lattice resolution at higher dimensions.

RP Fraction vs Spacetime Dimension
Configurations Summary by Dimension
d = 2
L = 10, 143 configs
7.7%
11 RP-admissible
d = 3
L = 5, 143 configs
86.7%
124 RP-admissible
d = 4
L = 3, 143 configs
86.7%
124 RP-admissible

Key Findings

First systematic computational investigation of OS axiom compliance for NN-FTs in d ≥ 2.

1

Linear Momentum Corrections Preserve RP

Corrections f(p2) = c · p2 preserve reflection positivity because the modified propagator retains the spectral representation. Nonlinear corrections (softplus, oscillatory) break this structure.

2

RP-Admissible Subset Is Structured but Small

Only 5.1% of scanned parameter space is RP-admissible in d=2. The admissible region concentrates near zero correction amplitude, confirming RP is genuinely constraining.

3

Transfer Operator RP Is Highly Restrictive

Only 0.8% of configurations satisfy RP for layered architectures. Linear activations are the only successful class; nonlinear activations systematically violate positive-definiteness.

4

Interacting Theories Are OS-Consistent

The φ4 NN-FT satisfies OS0, OS3, OS4 at all couplings. RP minimum eigenvalues of O(10-3) are consistent with Monte Carlo noise, not genuine violations.