on regularity structures

Overview

Hairer’s theory of regularity structures gives a calculus for singular stochastic PDEs whose nonlinearities are not classically defined (Hairer, 2014). It replaces Taylor polynomials by problem-specific symbolic expansions, then proves that these expansions can be reconstructed as genuine distributions.

The theory was designed to treat equations such as the three-dimensional dynamic ${\Phi }^4$ model and the KPZ equation, where products like $u^3$ or $({\partial }_{x}h{)}^2$ interact with distributions of negative regularity.

Singular Products

Classical PDE theory relies on operations that are harmless for smooth functions:

$$u,u^2,F(u),K\ast u.$$

For distributions, multiplication is not generally defined. If $f\in C^{\alpha }$ and $g\in C^{\beta }$, then the product $fg$ is canonically meaningful when

$$\alpha +\beta >0.$$

Many stochastic PDEs violate this inequality.

Consider the formal equation

$${\partial }_{t}u=\Delta u-u^3+\xi$$

in three spatial dimensions, where $\xi$ is space-time white noise. The linear solution $K\ast \xi$ is a distribution with negative regularity relative to the parabolic scaling. Cubing it is meaningless by ordinary analysis. Nevertheless, physics predicts a canonical renormalized equation obtained as a limit of mollified equations with diverging counterterms.

The analytic problem is therefore not merely to solve an equation. It is to construct a language in which the equation has a stable meaning under approximation.

Regularity Structures

A regularity structure is a triple

$$\mathcal{T}=(A,T,G).$$

Here $A\subset \mathbb{R}$ is a locally finite set of homogeneities, $T={⨁}_{\alpha \in A}{T}_{\alpha }$ is a graded vector space of abstract symbols, and $G$ is a structure group acting on $T$ while respecting the grading.

The polynomial example is the guiding model. A smooth function near a point $x$ has a Taylor expansion

$$f(y)=\sum _{|k|<\gamma }\frac{{\partial }^{k}f(x)}{k!}(y-x{)}^k+O(|y-x{|}^{\gamma }).$$

The monomials $X^k$ are basis elements, their degrees are $|k|$, and translating the base point acts by reexpanding $(y-x{)}^k$ around another point.

Hairer’s innovation is to replace polynomial monomials by symbols adapted to the equation. For ${\Phi }_3^4$, the alphabet contains symbols for the noise $\Xi$, integration against the heat kernel $\mathcal{I}(\tau )$, and products of previously constructed symbols. A typical symbol records a local stochastic object such as

$$\mathcal{I}(\Xi {)}^2\mathcal{I}(\mathcal{I}(\Xi {)}^3).$$

The homogeneity of a symbol tracks its expected small-scale regularity.

Models

A model realizes abstract symbols as concrete distributions. It consists of maps

$${\Pi }_x:T\to {\mathcal{D}}^{\prime }({\mathbb{R}}^{d+1})$$

and reexpansion operators

$${\Gamma }_{xy}\in G$$

such that

$${\Pi }_y={\Pi }_x{\Gamma }_{xy}.$$

Analytic bounds require ${\Pi }_x\tau$ to behave near $x$ like an object of order $|\tau |$.

Modelled Distributions

A modelled distribution is an abstract function

$$f:x\mapsto f(x)\in T_{<\gamma }$$

whose values at nearby points agree after reexpansion up to the expected order:

$$\Vert f(x)-{\Gamma }_{xy}f(y){\Vert }_{\alpha }\lesssim |x-y{|}^{\gamma -\alpha }$$

for every homogeneity $\alpha <\gamma$.

This is the analogue of saying that the Taylor coefficients of a $C^{\gamma }$ function vary coherently from point to point. The difference is that the coefficients now multiply symbols such as $\mathcal{I}(\Xi {)}^2$, not only polynomials.

Reconstruction theorem

For a modelled distribution $f$ of positive regularity $\gamma$, there is a unique distribution $\mathcal{R}f$ such that, near every point $x$,

$$\mathcal{R}f\approx {\Pi }_{x}f(x)$$

to order $\gamma$. More precisely, for test functions ${\phi }_x^{\lambda }$ localized at scale $\lambda$ around $x$,

$$|⟨\mathcal{R}f-{\Pi }_{x}f(x),{\phi }_x^{\lambda }⟩|\lesssim {\lambda }^{\gamma }.$$

This theorem is the analytic heart of the paper. It converts coherent local symbolic descriptions into a global distribution.

Proof: Reconstruction by Wavelets

Choose a multiresolution analysis adapted to the scaling of the problem. At dyadic scale $2^{-n}$, the lattice points play the role of base points. For a wavelet ${\psi }_x^n$ centered near $x$, define the coefficient of the reconstructed distribution by testing the local model at that base point:

$$⟨{\mathcal{R}}_{n}f,{\psi }_x^n⟩=⟨{\Pi }_{x}f(x),{\psi }_x^n⟩.$$

This is only a provisional definition, because neighboring points give different local models. The modelled-distribution bound says that after reexpansion by ${\Gamma }_{xy}$ the coefficients differ by the precise factor expected from their homogeneity.

The model bounds control each symbol $\tau$ by

$$|⟨{\Pi }_x\tau ,{\phi }_x^{\lambda }⟩|\lesssim {\lambda }^{|\tau |},$$

while the coherence of $f$ gives the extra gain

$$\Vert f(x)-{\Gamma }_{xy}f(y){\Vert }_{\alpha }\lesssim |x-y{|}^{\gamma -\alpha }.$$

Combining these two estimates shows that the wavelet coefficients at successive scales form a Cauchy family when tested against any smooth compactly supported function. Their limit is the distribution $\mathcal{R}f$.

The defining estimate follows by splitting a test function localized at scale $\lambda$ into wavelets above and below that scale. Contributions from coarse wavelets are controlled by the coherence of the local expansions; fine wavelets are controlled by the model bounds and summability of the wavelet coefficients. Uniqueness follows because a distribution whose localized tests are $O({\lambda }^{\gamma })$ for some positive $\gamma$ at every point must have all wavelet coefficients equal to zero. Thus coherent symbolic jets integrate to a unique distribution, even when the symbols are stochastic and non-polynomial.

Fixed Points for Singular SPDEs

A semilinear equation is lifted into the regularity structure. For a heat-type equation

$$({\partial }_t-\Delta )u=F(u,\xi ),$$

one writes an abstract fixed point

$$U=\mathcal{K}F(U,\Xi )+\text{polynomial data},$$

where $\mathcal{K}$ is the abstract integration operator associated with the heat kernel.

The integration theorem says that convolution with a kernel of order $\beta$ raises regularity by $\beta$, modulo polynomial correction terms. This is the counterpart of the Schauder estimate

$$\Vert K\ast f{\Vert }_{C^{\alpha +\beta }}\lesssim \Vert f{\Vert }_{C^{\alpha }}.$$

The abstract equation is solved by contraction in a space of modelled distributions. Reconstruction then produces the physical solution.

Renormalization

For a mollified noise ${\xi }_{\epsilon }$, one obtains a canonical smooth model. As $\epsilon \to 0$, this model usually diverges. Renormalization modifies the model by subtracting divergent constants in a way compatible with the algebraic structure.

In the ${\Phi }_3^4$ equation, this produces equations of the schematic form

$${\partial }_t{u}_{\epsilon }=\Delta u_{\epsilon }-u_{\epsilon }^3+C_{\epsilon }{u}_{\epsilon }+{\xi }_{\epsilon },$$

where $C_{\epsilon }$ diverges. The renormalized models converge, and the corresponding reconstructed solutions converge to a canonical limit.

The key point is that the counterterms are not guessed after the fact. They arise from a finite-dimensional renormalization group acting on the regularity structure. Subcriticality ensures that only finitely many symbols have negative homogeneity and hence only finitely many counterterms are required.

Subcriticality

Regularity structures do not make every singular PDE well posed. The equation must be locally subcritical under its scaling: as one expands the nonlinearity into symbols, only finitely many symbols may appear below any fixed regularity threshold. Supercritical equations require additional input or are expected to be ill posed in this framework.

Homogeneity, Scaling, and Subcriticality

The homogeneity assigned to a symbol is determined by the scaling of the equation. For the parabolic scaling of the heat equation, time has weight $2$ and space has weight $1$. Space-time white noise in $d$ spatial dimensions has regularity slightly below

$$-\frac{d+2}{2}.$$

Integration against the heat kernel raises regularity by $2$.

Subcriticality means that when the equation is expanded using these rules, only finitely many negative-homogeneity symbols are needed below any fixed regularity level. This finiteness is what makes the local theory possible. It is also what distinguishes equations such as KPZ and ${\Phi }_3^4$ from genuinely supercritical stochastic PDEs.

Models, Renormalized Models, and Universality

For smooth noise ${\xi }_{\epsilon }$, the canonical model interprets every symbol by ordinary multiplication and convolution. Divergence appears as $\epsilon \to 0$ because some symbolic products have expectations that blow up. Renormalization subtracts exactly the local divergent parts while preserving the algebraic identities needed for reconstruction and integration.

The output is a limiting model independent of many microscopic choices of mollifier, up to the prescribed renormalization constants. This is the universality statement hidden inside the analytic machinery: different smooth approximations converge to the same renormalized continuum object when the counterterms are chosen correctly.

Proof: Fixed Point After Reconstruction

The equation is solved first in the abstract space of modelled distributions. The unknown $U$ is not a distribution but a function assigning to each base point a finite symbolic expansion. Products such as $U^3$ are multiplied in the graded algebra of symbols, where the singular factors have already been named. Integration is performed by the abstract integration map, whose analytic estimate is the regularity-structure version of Schauder smoothing.

The fixed-point map is locally Lipschitz in modelled-distribution norms because subcriticality leaves only finitely many symbols below the chosen regularity threshold. This finiteness is what makes the nonlinear map closed: expanding $F(U,\Xi )$ never produces an uncontrolled infinite family of more singular terms. Polynomial correction terms record the fact that convolution with the heat kernel must be reexpanded at each base point.

Reconstruction is applied only after the abstract fixed point has been found for a fixed model. At that stage $\mathcal{R}U$ is an actual distribution and the abstract identity descends to the renormalized physical equation associated with that model. This order is essential. If one reconstructed first, products such as $(\mathcal{R}U{)}^3$ would be undefined; by multiplying before reconstruction, all singular algebra is resolved in the symbolic space where the model provides the missing local products.

Wick Products as the First Renormalization

Before the full theory, the simplest model is the Wick square of a Gaussian field. If $X_{\epsilon }$ is a mollified Gaussian distribution, then

$$X_{\epsilon }^2$$

often diverges as $\epsilon \downarrow 0$. The renormalized object is

$$:X^2:=\underset{\epsilon \downarrow 0}{\lim }\left(X_{\epsilon }^2-\mathbb{E}{X}_{\epsilon }^2\right),$$

when the limit exists in a suitable distribution space.

Regularity structures generalize this idea. Each divergent subtree in the symbolic expansion has its own local subtraction. The subtraction is not arbitrary: it must preserve the algebraic relations that allow reexpansion between base points. Thinking first about Wick ordering makes the later Hopf-algebraic renormalization less mysterious.

Transferable Mechanisms

Regularity structures are one way to organize singular products; paracontrolled distributions give a parallel calculus based on frequency decompositions and controlled expansions relative to a rough reference distribution (Gubinelli, Imkeller & Perkowski, 2015). The two frameworks share the same strategic principle: do not multiply rough objects directly; first identify the finite list of singular components and expand relative to them.

The later algebraic renormalization theory of Bruned, Hairer, and Zambotti shows how the counterterms can be governed by Hopf-algebraic and combinatorial structures (Bruned, Hairer & Zambotti, 2018). This matters beyond the original SPDE examples: whenever a limiting equation contains divergent local substructures, one should look for a finite algebra of local counterterms rather than renormalizing each approximation ad hoc.

Links

• on Ising models --- the dynamic ${\Phi }^4$ model is a continuum stochastic field theory connected to critical lattice spin systems.
• on diffusion models --- both topics use noisy evolution equations, but regularity structures address the analytic meaning of singular continuum limits rather than statistical training dynamics.
• on Grönwall’s inequality --- fixed-point arguments for modelled distributions still rely on stability estimates whose scalar skeleton is Grönwall-type control.

References

🐻  Bruned, Y., Hairer, M. & Zambotti, L. 2018. Algebraic Renormalisation of Regularity Structures. Inventiones Mathematicae 215(3), 1039–1156.

🐻  Gubinelli, M., Imkeller, P. & Perkowski, N. 2015. Paracontrolled Distributions and Singular PDEs. Forum of Mathematics, Pi 3, e6.

🐻  Hairer, M. 2014. A Theory of Regularity Structures. Inventiones Mathematicae 198(2), 269–504.