Stochastic Renormalization Techniques

• Stochastic renormalization is a framework that integrates randomness to replace deterministic errors, yielding unbiased estimators in tensor networks and effective actions.
• The methodology involves stochastic tensor decompositions and RG mappings where noise substitutes conventional truncation biases, improving accuracy in models like the Ising system.
• Applications span functional RG with Langevin dynamics, Wick-type renormalization in SPDEs, and turbulence models, highlighting both computational trade-offs and non-perturbative effects.

Stochastic renormalization denotes a family of renormalization and coarse-graining procedures in which randomness is built into the renormalization step itself rather than treated only as external forcing. In the literature surveyed here, the phrase refers to several distinct constructions: stochastic tensor decompositions that replace deterministic truncation bias by sampling error; functional renormalization-group formulations written as Langevin or Fokker–Planck dynamics; Wick-type, Wilsonian, or microlocal renormalizations of singular stochastic PDEs; and renormalization-group maps acting directly on stochastic transition kernels or effective stochastic dynamics (Ohki et al., 2021, Prokopec et al., 2017, Carosso, 2019, Kupiainen, 2014, Mailybaev et al., 27 Feb 2026).

1. Conceptual scope

The common feature across these constructions is not a single canonical formalism, but a repeated shift in what is being renormalized. In some settings the renormalized object is a local tensor decomposition; in others it is an effective Boltzmann factor, an effective average action, a Wick-ordered nonlinearity, an enhanced noise, or a Markov kernel. What changes from case to case is the role played by stochasticity: it may act as a sampling device for discarded singular modes, as a dynamical representation of coarse graining, or as the source of singular products that require renormalization (Carosso et al., 2019, Dappiaggi et al., 2020).

Setting	Renormalized object	Characteristic stochastic ingredient
Tensor networks	Low-rank tensor decomposition	Noise vectors for discarded singular subspaces
Functional RG	Effective Boltzmann factor or effective average action	Langevin/Fokker–Planck diffusion in RG time
Singular SPDEs	Nonlinear products and counterterms	Wick powers, enhanced noise, stochastic convolutions
Kernel-level RG	Effective stochastic dynamics	Composition and flow of Markov kernels

A second recurrent feature is that stochastic renormalization often exchanges a deterministic approximation error for a statistical one. In tensor-network work this replacement is explicit: truncation bias from discarded singular values is traded for sampling uncertainty. In functional RG, stochastic processes provide an exact probabilistic representation of blocking kernels. In singular SPDEs, randomness enters both as the source of ultraviolet singularity and as the object whose nonlinear combinations must be reorganized into renormalized composites (Ohki et al., 2021, Carosso, 2019, Gubinelli et al., 2017).

2. Tensor-network stochastic renormalization

In tensor renormalization group methods, the central bottleneck is the repeated low-rank approximation of reshaped tensors by truncated SVD. The hybrid stochastic method for TRG keeps the leading $D_{\rm svd}$ singular vectors exactly and represents the discarded sector with random noise vectors satisfying an approximate completeness relation,

$$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$

The effective bond dimension is

$D_{\rm cut}=D_{\rm svd}+N_r,$

and averaging over $N$ independent noise realizations yields an estimator whose uncertainty scales as $\mathcal O(1/\sqrt{N_rN})$. In the infinite-statistics limit, the exact matrix decomposition is recovered even at finite $D_{\rm cut}$. The method therefore replaces deterministic truncation bias by a Monte Carlo–like error that can be estimated empirically, although any single stochastic sample is not better than the deterministic truncated SVD in Frobenius norm (Ohki et al., 2021).

This construction leads to two distinct implementations. With position-dependent noise, each site or leg uses independent noise vectors, eliminating cross contamination and making the estimator unbiased. The price is loss of translational invariance and a per-configuration cost $\mathcal O(VD_{\rm cut}^6)$. With common noise, the same noise is reused on symmetry-related tensor legs, preserving homogeneous TRG structure and keeping the per-sample cost at the same order as ordinary TRG, but introducing a systematic bias through contact terms generated by correlated noise reuse. In the Ising benchmark, the common-noise method achieved much better accuracy than the original TRG at fixed $D_{\rm cut}=50$, with more than an order-of-magnitude improvement in the free-energy error after optimizing the split between $D_{\rm svd}$ and $N_r$ (Ohki et al., 2021).

“All-mode renormalization” refines this idea by insisting that all singular modes remain represented in the renormalized tensors, with the discarded sector encoded stochastically rather than omitted. The paper couples this with a new common-noise strategy whose cost is proportional to $$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$0, specifically $$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$1, instead of the volume-proportional cost of position-dependent noise. The resulting common-noise bias is traced to a simple $$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$2 contamination structure and is fit by expansions such as

$$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$3

or higher-order polynomials in $$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$4. The method improves the free-energy accuracy relative to truncated SVD for the square-lattice Ising model and yields further gains when combined with graph independent local truncation (Arai et al., 2022).

3. Functional renormalization group, Langevin dynamics, and gradient flow

A different meaning of stochastic renormalization arises when the renormalization-group transformation itself is represented as a stochastic process. In scalar field theory, the blocked field $$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$5 is taken to evolve by an Ornstein–Uhlenbeck-type Langevin equation,

$$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$6

with transition kernel $$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$7 solving a Fokker–Planck equation. The effective Boltzmann factor is then defined by

$$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$8

This yields an exact stochastic version of continuous blocking and an exact Monte Carlo RG identity for observables. It also clarifies the relation to gradient flow: the deterministic heat-kernel smoothing $$\frac{1}{N_r}\bm{\eta}\bm{\eta}^\dagger = {\bf 1} + \mathcal O(1/\sqrt{N_r}).$$9 is the drift part of the stochastic RG, but ordinary gradient flow alone is not a full RG transformation because it lacks the field rescaling and stochastic broadening needed to generate a nontrivial effective action (Carosso, 2019, Carosso et al., 2019).

The same literature shows that long-distance correlators of the effective theory differ from gradient-flowed correlators only by short-ranged terms such as $D_{\rm cut}=D_{\rm svd}+N_r,$0. Consequently, the Markov property of the stochastic RG semigroup implies scaling relations that can be transcribed into flowed-observable ratio formulas. This was used numerically in three-dimensional lattice $D_{\rm cut}=D_{\rm svd}+N_r,$1 theory, where flowed correlators of operators such as $D_{\rm cut}=D_{\rm svd}+N_r,$2 and $D_{\rm cut}=D_{\rm svd}+N_r,$3 were employed to extract anomalous-dimension differences without explicitly simulating the full stochastic process (Carosso et al., 2019).

In stochastic inflation, the functional RG is applied directly to Starobinsky’s stochastic description of the infrared dynamics of a light scalar field in de Sitter space. After passing from the Langevin or Fokker–Planck description to a Martin–Siggia–Rose/Janssen–de Dominicis action, the theory acquires a Parisi–Sourlas supersymmetric Euclidean quantum-mechanics form. A scale-dependent effective average action $D_{\rm cut}=D_{\rm svd}+N_r,$4 is then introduced with an IR cutoff in temporal frequency. In the supersymmetry-preserving local potential approximation, the flow reduces to

$D_{\rm cut}=D_{\rm svd}+N_r,$5

which exactly matches the de Sitter QFT effective-potential flow after the identification $D_{\rm cut}=D_{\rm svd}+N_r,$6. Both the supersymmetric and non-supersymmetric formulations generate an infrared mass $D_{\rm cut}=D_{\rm svd}+N_r,$7, while the stochastic spectral analysis distinguishes the dynamical mass inferred from equilibrium variance from the true relaxation rate set by the first excited eigenvalue of the stochastic Hamiltonian (Prokopec et al., 2017).

4. Singular stochastic PDEs and Wick-type renormalization

For singular stochastic PDEs, stochastic renormalization most often means reorganizing ill-defined nonlinear products of random distributions into Wick-ordered or otherwise renormalized composites. In the two-dimensional stochastic nonlinear wave equation with additive space-time white noise,

$D_{\rm cut}=D_{\rm svd}+N_r,$8

the stochastic convolution $D_{\rm cut}=D_{\rm svd}+N_r,$9 is distribution-valued, so powers $N$0 diverge. The renormalized equation replaces $N$1 by the Hermite polynomial

$N$2

where the variance

$N$3

depends on time. This time dependence is essential: the counterterm is not stationary because the wave stochastic convolution is not stationary. The resulting Wick powers converge, the renormalized SNLW is pathwise locally well posed, and a weak universality theorem identifies the renormalized cubic SNLW as a scaling limit of smoother random wave equations (Gubinelli et al., 2017).

A closely related but more singular setting is provided by subordinate cylindrical Brownian noise. There one defines

$N$4

with $N$5 an increasing càdlàg process. Conditioning on $N$6 restores Gaussianity in the Brownian variables, so Wick powers can still be defined through Hermite polynomials, but the renormalization constants become random time-dependent processes: $N$7

$N$8

This leads to a sharp contrast between heat and wave equations: for the renormalized heat equation only the quadratic case is obtained because of low time-integrability, whereas for the renormalized wave equation the Wick powers are time-continuous and local well-posedness holds for all polynomial nonlinearities (Nagoji, 2022).

These results show that stochastic renormalization in SPDEs need not mean subtraction of a fixed deterministic constant. Depending on the model, the counterterm can be time-dependent, random, or encoded directly in the definition of the nonlinear functional $N$9. The central renormalized objects are the singular stochastic convolutions and their nonlinear composites, not only the bare coupling constants (Gubinelli et al., 2017, Nagoji, 2022).

5. Wilsonian, microlocal, and effective-potential frameworks

A Wilsonian version of stochastic renormalization constructs effective equations scale by scale. For the three-dimensional stochastic quantization equation

$\mathcal O(1/\sqrt{N_rN})$0

Kupiainen decomposes the regularized propagator into shells, integrates out short space-time fluctuations, and derives a flow of effective nonlinearities on successively larger scales. Renormalization is encoded in the mass counterterm

$\mathcal O(1/\sqrt{N_rN})$1

with $\mathcal O(1/\sqrt{N_rN})$2 universal. This yields local well-posedness and independence of regularization for $\mathcal O(1/\sqrt{N_rN})$3, and the RG construction proceeds without first introducing a general multiplication theory for distributions (Kupiainen, 2014).

For singular elliptic SPDEs with fractional Laplacian, additive white noise, and cubic nonlinearity, a continuous Wilsonian approach based on the Polchinski flow equation is used. The regularized mild equation is written with a UV cutoff $\mathcal O(1/\sqrt{N_rN})$4, and a scale-dependent effective force $\mathcal O(1/\sqrt{N_rN})$5 satisfies a flow equation in $\mathcal O(1/\sqrt{N_rN})$6. Power counting in the regime

$\mathcal O(1/\sqrt{N_rN})$7

isolates finitely many relevant coefficients, organized into an enhanced noise. Counterterms $\mathcal O(1/\sqrt{N_rN})$8 are then fixed so that the regularized solutions converge almost surely in Besov spaces throughout the full subcritical regime (Duch, 2022).

A more AQFT-like perspective replaces regularization and subtraction by microlocal extension theory. In that framework, random fields are represented in an algebra of functional-valued distributions, the covariance kernel $\mathcal O(1/\sqrt{N_rN})$9 deforms the product, and renormalization is reinterpreted as extension of singular distributions from configuration spaces with diagonals removed to the full diagonal. Finite scaling degree guarantees existence of such extensions, while nonuniqueness is classified by local counterterm freedoms in maps such as $D_{\rm cut}$0 and $D_{\rm cut}$1. Applied to perturbative stochastic $D_{\rm cut}$2, this yields renormalized expectations, correlations, and a renormalized effective equation without introducing a cutoff parameter and then subtracting infinities (Dappiaggi et al., 2020).

Effective-potential methods provide yet another route. For a stochastic differential equation with multiplicative noise,

$D_{\rm cut}$3

the one-loop effective potential

$D_{\rm cut}$4

organizes the renormalization directly at the level of the stochastic equation. In a Gray–Scott-like toy model with $D_{\rm cut}$5 and colored power-law noise $D_{\rm cut}$6, the divergence structure shows that only the removal rate $D_{\rm cut}$7 runs at one loop: $D_{\rm cut}$8 Here stochastic renormalization is extracted from the effective potential itself rather than from a shell-by-shell elimination of modes (Gagnon et al., 2020).

6. Beyond local fields: turbulence, quenched disorder, data, and kernels

In stochastic hydrodynamics, a non-perturbative average-action RG for the incompressible stochastic Navier–Stokes equation with power-law forcing yields a unique fixed point within the symmetry-preserving zero-frequency truncation studied. The kinetic-energy spectrum keeps the perturbative exponent $D_{\rm cut}$9, so Kolmogorov’s $\mathcal O(VD_{\rm cut}^6)$0 law is recovered at $\mathcal O(VD_{\rm cut}^6)$1, but the eddy-diffusivity exponent saturates at $\mathcal O(VD_{\rm cut}^6)$2 instead of continuing with the perturbative $\mathcal O(VD_{\rm cut}^6)$3 scaling. This provides a concrete example in which stochastic renormalization uncovers a genuinely non-perturbative saturation effect without yet recovering the full constant-flux fixed-point structure of turbulence (Mejía-Monasterio et al., 2012).

For stochastic growth models with spatially quenched disorder, a purely static field theory reproduces the equal-time scaling of the original dynamics. In this formulation the upper critical dimension is shifted upward by two relative to the corresponding white-in-time problems. The effect is explicit in quenched KPZ, Pavlik’s modification, the Hwa–Kardar model, and the Pastor-Satorras–Rothman erosion model. In the erosion problem, for example, the exact relation

$\mathcal O(VD_{\rm cut}^6)$4

replaces the white-noise identity $\mathcal O(VD_{\rm cut}^6)$5, opening the possibility of nontrivial scaling already in $\mathcal O(VD_{\rm cut}^6)$6 (Antonov et al., 2019).

Several recent works move the renormalized object even further away from conventional couplings. A Parisi–Wu-inspired discrete stochastic dynamics on Ising configurations defines a transition probability kernel whose semigroup satisfies local Poincaré inequalities and yields two-point correlation bounds through Bakry–Émery curvature estimates (Cui et al., 2022). A Model-A-type FRG for signal detection in noisy covariance spectra interprets finite-scale singularities in the flow as weak ergodicity breaking near the Marchenko–Pastur law, and the disappearance of those singularities above a critical signal strength defines a dynamical detection threshold (Erbin et al., 2023). In a solvable multiscale Arnold’s cat model, the renormalization group acts directly on Markov kernels,

$\mathcal O(VD_{\rm cut}^6)$7

and converges to a universal stochastic fixed point

$\mathcal O(VD_{\rm cut}^6)$8

thereby casting spontaneous stochasticity itself as an RG fixed-point phenomenon (Mailybaev et al., 27 Feb 2026).

7. Common themes, limitations, and interpretive issues

Several structural themes recur across this literature. First, stochastic renormalization frequently replaces discarded or ill-defined short-scale information by an estimator whose error can be sampled, bounded, or classified. Second, symmetry constraints remain decisive: Galilean invariance in stochastic Navier–Stokes, stochastic supersymmetry in stochastic inflation, lattice isotropy in tensor methods, and locality or microlocality in SPDE renormalization all constrain admissible flows and counterterms (Mejía-Monasterio et al., 2012, Prokopec et al., 2017, Dappiaggi et al., 2020).

The same papers are equally explicit about limitations. In stochastic tensor renormalization, unbiased position-dependent noise is expensive and high-variance, whereas common noise reduces cost but introduces systematic cross contamination with a characteristic $\mathcal O(VD_{\rm cut}^6)$9 structure (Ohki et al., 2021, Arai et al., 2022). In stochastic inflation, the supersymmetry-preserving FRG reproduces the de Sitter QFT potential flow, while the symmetry-breaking version gives a close but not structurally identical infrared description (Prokopec et al., 2017). In subordinate-noise SPDEs, the wave equation behaves better than the heat equation precisely because the jump-type temporal singularity degrades the heat problem more severely (Nagoji, 2022). The static RG for quenched disorder is adapted to equal-time quantities and is not a substitute for full dynamic response theory (Antonov et al., 2019). The RG description of spontaneous stochasticity is rigorous for the multiscale cat model, but its extension to Navier–Stokes remains a research program rather than a theorem (Mailybaev et al., 27 Feb 2026).

This suggests a broad but precise reading of the term. Stochastic renormalization is not a single algorithm and not merely renormalization applied to stochastic systems. It is a class of procedures in which randomness becomes part of how short-scale information is represented, flowed, averaged, or extended. Under that broader definition, the subject links tensor coarse graining, functional RG, singular SPDEs, stochastic dynamics, and kernel-level universality within a common renormalization perspective.