Stochastic Quantization: Concepts & Applications

• Stochastic quantization is a framework employing auxiliary stochastic dynamics, such as the Parisi–Wu formulation, to construct path-integral measures in both classical and quantum fields.
• It bridges theoretical physics and engineering by enabling discrete-time algorithms that maintain gauge invariance and simulate non-perturbative effects in QFT and machine learning.
• The approach offers computational advantages by leveraging exact weighted quantization and stochastic control methods for efficient neural network compression, signal processing, and advanced simulation.

Stochastic quantization is a family of frameworks and algorithms that introduce auxiliary stochastic dynamics—often an artificial or "fictitious" time direction—into classical or quantum systems to construct path-integral measures, quantize fields, or devise scalable quantization algorithms in engineering and machine learning. Variants of stochastic quantization appear in theoretical and mathematical physics, quantum field theory, stochastic analysis, statistical mechanics, machine learning, and signal processing, each with domain-specific adaptations and technical nuances.

1. Fundamental Principles and Parisi–Wu Formulation

The prototypical stochastic quantization scheme was introduced by Parisi and Wu, who observed that for a classical Euclidean field theory with action $S[\phi]$, introducing a fictitious time $t$ and evolving the field $\phi(x,t)$ via a stochastic partial differential equation—specifically, a Langevin equation with white noise—

$$\frac{\partial\phi(x,t)}{\partial t} = -\frac{\delta S[\phi]}{\delta\phi(x,t)} + \eta(x,t),$$

with $$\langle\eta(x,t)\eta(x',t')\rangle = 2\,\delta^{(d)}(x-x')\delta(t-t')$$, causes the equal-time distribution of $\phi(x,t)$ to converge as $t\to\infty$ to the quantum field theory (QFT) Euclidean path integral measure $e^{-S[\phi]}$ (Fukushima et al., 2024). This mechanism is formalized via the corresponding Fokker–Planck functional equation for the probability density $P[\phi, t]$, the stationary solution of which is the desired path-integral weight.

This approach is conceptually powerful because it:

• Constructs field quantization via stochastic dynamics rather than operator or path-integral axioms.
• Yields emergent supersymmetry, as the stochastic process can be represented with a Nicolai map and auxiliary Grassmann variables in the associated functional integral (Bienzobaz et al., 2012).
• Avoids gauge-fixing subtleties in some cases, as gauge symmetry is built into the stochastic dynamics or enforced through suitable stochastic modifications (Shen, 2018).

2. Discrete-Time and Weighted Stochastic Quantization

A significant development is the adaptation of stochastic quantization to fully discrete fictitious time, as proposed in (Kadoh et al., 24 Jan 2025). Define the discrete-time sequence $\phi_n(x)$ with time step $t$0, and backward lattice derivative $t$1. The discrete Langevin update reads: $t$2 where $t$3, and $t$4 is Gaussian noise with $t$5. For fixed $t$6, naively averaging observables over this noise does not reproduce the target QFT correlators except as $t$7. This is remedied by introducing a weight $t$8 in the noise measure, constructed from drift-Jacobian determinants and boundary terms, such that expectation values with respect to the weighted measure reproduce the correct QFT limit as the number of steps $t$9 at fixed $\phi(x,t)$0. In particular, in the 0-dimensional toy model, this approach recovers the exact moments of the theory for any $\phi(x,t)$1 (Kadoh et al., 24 Jan 2025).

Key features of this weighted discrete procedure are:

• Exact equivalence to the target theory without needing a continuum limit in fictitious time.
• Non-perturbative validity based on an exact nilpotent symmetry of the discrete-lattice action.
• Tractability for simulation, as demonstrated via cubic polynomial recurrences in the 0-dimensional case and numerical averaging over weighted noise realizations.

3. Stochastic Quantization in Quantum Field Theories

Stochastic quantization has enabled rigorous constructions and analysis in Euclidean QFTs, including for models with singular interactions or nontrivial gauge symmetry:

• In the $\phi(x,t)$2 quantum field model, stochastic quantization yields a singular SPDE in two dimensions with Wick-renormalized exponential nonlinearity. This allows for construction and invariance proofs for the Gibbs measure, as well as the identification of the associated Dirichlet-form Markov process (Hoshino et al., 2019).
• For Abelian gauge theories, stochastic quantization on the lattice leads to coupled stochastic PDEs for gauge and scalar fields, with gauge-fixing ("DeTurck trick") restoring well-posedness and Ward identities enforcing gauge invariance at the quantum level (Shen, 2018).
• In the stochastic variational method for electromagnetic fields (Kodama et al., 2014), quantization is performed directly from the gauge-invariant Lagrangian, and gauge-fixing is implemented as a choice of which field components are subjected to stochastic noise. This allows for quantization in the Lorentz gauge without indefinite metrics unless a Fermi term is introduced.

Extensions to relativistic and gravitational theories are formulated by embedding stochastic quantization into second-order geometry (Schwartz–Meyer) on pseudo-Riemannian manifolds (Kuipers, 2021, Kuipers, 2021), or, for general relativity, via Ricci-flow-inspired stochastic evolution of the metric (Lulli et al., 2021). In the latter, the equilibrium of the stochastic process recovers Einstein's equations with a dynamical cosmological constant, and off-equilibrium dynamics connect to renormalization-group flow and hydrodynamic turbulence.

4. Engineering and Algorithmic Applications

Stochastic quantization also refers to a class of post-training neural network compression algorithms, lossy signal quantizers, and stochastic gradient-driven quantization methods in engineering:

• In neural network compression, stochastic quantization algorithms update only a randomly selected fraction of weight filters to low-bit representations per iteration, with sampling probabilities inversely proportional to per-filter quantization error. The SQ ratio is progressively increased from partial to full quantization, achieving state-of-the-art post-training accuracy especially in challenging low-bit regimes (Dong et al., 2017, Zhang et al., 2024). The stochastic approach outperforms deterministic or fixed-partitioning schemes, offering adaptive focus, stochastic regularization, and progressive network adaptation to quantization. Table 1 summarizes representative improvements from (Dong et al., 2017):

Dataset/Model	Baseline Error/Bits	SQ Error/Bits
CIFAR-10, ResNet-56	16.42% (BWN, 1-bit)	7.15%
CIFAR-10, ResNet-56	7.64% (TWN, 2-bit)	6.20%
CIFAR-100, ResNet-56	32.09% (TWN, 2-bit)	28.90%
ImageNet, ResNet-18	45.20% (BWN, 1-bit)	41.64%
ImageNet, ResNet-18	39.83% (TWN, 2-bit)	36.18%

• In lossy image compression, SQ reframes color quantization as a stochastic transportation problem, replacing full-batch K-means assignment with per-sample stochastic palette updates based on sampled pixels. This enables O(K)-memory and streaming operation at scale while achieving comparable mean squared error to deterministic methods (Kozyriev et al., 2024).
• For constrained optimization and stochastic filtering in control systems, a weighted discrete-time stochastic quantization framework incorporates arbitrary "potential" terms into path-integral representations of constrained Ito processes. This recovers the Extended Kalman Filter (EKF) as a special case and provides a tractable path-integral approach for constraint enforcement and control (Sano, 2019).
• In privacy-preserving control, stochastic quantizers guarantee $\phi(x,t)$3-differential privacy by sufficient quantization step, and dynamic stochastic quantizers can decouple privacy guarantees from nominal tracking performance (Liu et al., 2024).

5. Connections to Diffusion Models, Quantum Mechanics, and Supersymmetry

The stochastic quantization formalism establishes a direct analogy with modern score-based generative diffusion models, where stochastic differential equations drive a data distribution to equilibrium and time-reversal or "score" SDEs provide generative sampling (Fukushima et al., 2024). In both cases, the Fokker–Planck equation describes the evolution toward a stationary measure, with the score or the functional derivative of the action driving the drift.

In theoretical physics, the stochastic quantization path-integral formulation is closely associated with supersymmetry in one higher dimension (Bienzobaz et al., 2012). The mapping between the Langevin process and Parisi–Sourlas supersymmetry reveals emergent supercharges and path-integral structures. This mathematical machinery is leveraged both in quantum phase transition analysis (quantum spherical model) and for exact nonperturbative computation of correlation functions.

The connection to quantum mechanics is further developed in stochastic electrodynamics and stochastic quantum mechanics (Cetto et al., 2020), where classical particles in stochastic backgrounds naturally acquire the canonical commutation relations and the structure of the Schrödinger equation through fluctuation–dissipation balance and a specific identification of diffusion coefficients.

6. Computational and Mathematical Aspects, Limitations, and Extensions

Stochastic quantization procedures offer algorithmic and simulation advantages:

• Discrete-time and weighted approaches permit exact equilibrium-correlation recovery without continuum-time limits (Kadoh et al., 24 Jan 2025).
• Statistical linearization within stochastic quantization yields Dyson–Schwinger-type self-consistent gap equations, facilitating nonperturbative analysis and efficient computation—e.g., via Ramanujan summation to handle divergent contributions (Janowicz et al., 2012).
• For SPDEs with singular noise (e.g., laser propagation), stochastic quantization with the S-transform (Hermite transform) reduces pathwise-white-noise-driven PDEs to analytic Banach-space PDE families, enabling existence and uniqueness results via the infinite-dimensional implicit function theorem (Sritharan et al., 2022).

Current limitations include:

• For higher-dimensional gauge theories and interacting models, control of weight fluctuations, sign problems (cf. Lefschetz-thimble approaches (Fukushima et al., 2024)), and the computation of Jacobian determinants may present challenges.
• The fully non-perturbative status and subtle discrepancies between stochastic and conventional quantization at higher orders (such as the need for extra counterterms in axial-gauge or CP-violating extensions (Kapoor, 2018)) require careful attention.

Potential directions involve algorithmic advances in scalable stochastic quantization for non-Abelian gauge fields, potential applications of Markov-chain versions in large-scale QFT simulations, and mathematical development of the stochastic framework in quantum gravity, field theory, and high-dimensional stochastic control (Kadoh et al., 24 Jan 2025, Lulli et al., 2021).

7. Summary of Domain-Specific Variants

Domain	Key Mechanism	Distinctive Features
Euclidean QFT (Parisi–Wu)	Langevin SDEs in fictitious t	Path-integral, supersymmetry, gauge-invariant
Engineering/Machine Learning	Stochastic path-following, SGD	Memory-efficient, error correction, low-bit regimes
Discrete-Time/Lattice QFT	Weighted discrete dynamics	No continuum-limit required, exactness at fixed step
Control and Signal Processing	Constrained path-integrals	Recovers EKF, stochastic control, privacy guarantees
Differential Geometry/Gravity	Second-order geometry SDEs	Manifold-valued diffusion, conformal coupling, RG
Quantum/Stochastic Mechanics	SED-inspired noise dynamics	Schrödinger emergence, commutator from diffusion

Stochastic quantization thus constitutes a broad and evolving axis of research and application, with foundational and practical impact across quantum field theory, statistical mechanics, signal processing, and deep learning. Its ongoing development interlinks analytic, geometric, and computational perspectives for both theoretical understanding and scalable algorithmic design.