SPDE — Foundations¶

Job: Motivate stochastic PDE; define space-time noise; set mild solutions in Hilbert/Banach spaces; point to singular solution theories and stochastic quantization.

Core¶

Motivations¶

Source	Model form
Continuum particle limits	hydrodynamic fluctuations; reaction–diffusion with noise
Interface growth	KPZ: $\partial_t h=\partial_x^2 h+(\partial_x h)^2+\xi$
Fluids	stochastic Navier–Stokes: additive/multiplicative noise on velocity
Euclidean QFT	dynamical $\Phi^4$: $\partial_t\Phi=\Delta\Phi-\Phi^3+\xi$ (Langevin / Parisi–Wu)
Filtering / data	signal is SPDE; observations drive Zakai-type equations

Noise models microscopic degrees of freedom not resolved by deterministic PDE; scaling limits fix the roughness of $\xi$.

White noise and Wiener processes¶

Space-time Gaussian white noise $\xi$ on $\mathbb{R}_+\times\mathcal{O}$: centered Gaussian generalized process with $$ \mathbb{E}\bigl[\langle\xi,\varphi\rangle\langle\xi,\psi\rangle\bigr]=\int_{\mathbb{R}_+\times\mathcal{O}}\varphi\psi $$ for test functions $\varphi,\psi$. Realizations are distributions of negative Hölder regularity (roughly $-1-\varepsilon$ in parabolic scaling, dimension-dependent).

Cylindrical Wiener process $W$ on Hilbert space $U$: formally $W(t)=\sum_k \beta_k(t) e_k$ for ONB $(e_k)$ and independent BMs $\beta_k$. Not $U$-valued unless $\sum\|Q^{1/2}e_k\|^2<\infty$ (trace-class covariance $Q$). $Q$-Wiener process takes values in $U$ when $\operatorname{Tr}Q<\infty$.

Colored noise: covariance operator smoother than identity $\Rightarrow$ better spatial regularity of solutions; used when white noise is supercritical.

Stochastic convolution (linear mild noise term): $$ W_A(t)=\int_0^t S(t-s)\,dW(s), $$ $S(t)=e^{tA}$ analytic semigroup. Regularity of $W_A$ determined by smoothing of $S$ vs roughness of $W$ (factorization method; Da Prato–Zabczyk).

Mild solutions for semilinear SPDE¶

Abstract form on Banach $E$ (often Hilbert $H$): $$ dX(t)=\bigl(AX(t)+F(X(t))\bigr)\,dt+B(X(t))\,dW(t), $$ $A$ generates $C_0$/analytic semigroup $S(t)$, $F$ Nemytskii-type nonlinearity, $B$ diffusion coefficient (often Hilbert–Schmidt-valued).

Mild solution: progressive $E$-valued process with $$ X(t)=S(t)X_0+\int_0^t S(t-s)F(X(s))\,ds+\int_0^t S(t-s)B(X(s))\,dW(s) $$ a.s. for each $t$ (integrals in Bochner / Itô sense).

Existence package (regular regime): global Lipschitz $F,B$ (or local Lipschitz + a priori bounds), suitable HS conditions on $B$ $\Rightarrow$ unique mild solution via Picard in spaces $$ L^p\bigl(\Omega;C([0,T];E)\bigr) $$ or with fractional time weights. Monotone + coercive operators: variational approach (Pardoux, Krylov–Rozovskii) in Gelfand triples $V\hookrightarrow H\hookrightarrow V^*$.

Linear multiplicative noise: Itô vs Stratonovich corrections appear when changing interpretations; energy estimates use Itô formula in Hilbert space.

Pathwise regularity (linear SHE)¶

Stochastic heat equation $\partial_t u=\Delta u+\xi$ on $\mathbb{T}^d$: - $d=1$: function-valued, space-Hölder $<\tfrac12$ - $d\ge 2$: distribution-valued; lives in $C^{-\alpha}$ for $\alpha>d/2-1$ (parabolic)

Multiplicative SHE $\partial_t u=\Delta u+u\cdot\xi$ requires Wick / Itô–Walsh products in $d=1$; fails classically for $d\ge 2$ without renormalization.

Singular SPDEs: regularity structures / paracontrolled¶

When products of distributions appear (KPZ, $\Phi^4_2$, $\Phi^4_3$, PAM), classical mild theory breaks.

Regularity structures (Hairer). Lift noise into a model (jets of renormalized multilinear functionals of $\xi$); solve a fixed-point in modelled distributions; reconstruct to space-time distributions. Renormalization $=$ change of model (subtracting counterterms). Continuous dependence on models yields universality statements.

Paracontrolled calculus (Gubinelli–Imkeller–Perkowski). Use Bony paraproducts: decompose products $fg=\operatorname{T}_f g+\operatorname{T}_g f+\operatorname{R}(f,g)$; control resonant term $\operatorname{R}$ by assuming $f$ is “paracontrolled” by the linear noise solution. Often lighter than full regularity structures for semi-linear parabolic problems.

Common output: local well-posedness for subcritical equations (scaling: nonlinearity less singular than linear part). Critical/supercritical: largely open or needs different methods.

Stochastic quantization (one-liner)¶

Stochastic quantization constructs Euclidean measures (e.g. $\Phi^4$) as invariant measures of parabolic Langevin dynamics $\partial_t\Phi=\Delta\Phi-V'(\Phi)+\xi$; SPDE well-posedness + ergodicity $\Rightarrow$ constructive QFT control. Hairer’s $\Phi^4_3$ theory is the modern analytic apex of this program.

See it¶

Da Prato–Debussche trick ($\Phi^4_2$). Write $\Phi=Y+v$ with $Y$ the stationary stochastic convolution (Gaussian, distribution-valued) and $v$ solving a PDE with random coefficients involving Wick powers $:\!Y^n\!:$. In $d=2$, Wick renormalization makes products definable and $v$ function-valued — converts singular SPDE into random PDE amenable to deterministic estimates. In $d=3$, residual roughness forces regularity structures / paracontrolled upgrades.

KPZ via Hopf–Cole. Classically $h=\log Z$ with SHE $\partial_t Z=\partial_x^2 Z+Z\xi$ (Itô); defines Cole–Hopf solution. Regularity structures construct KPZ directly and match Cole–Hopf after renormalization — justifies the equation as a scaling limit object.

Check¶

What does ‘subcritical’ mean for a semilinear parabolic SPDE?

After scaling dimension analysis, the nonlinear term becomes negligible at small scales relative to the linear heat flow + noise (power-counting). Subcritical $\Rightarrow$ local theory via fixed point after renormalization; critical/supercritical lack this perturbative room.

Why does the stochastic convolution for $A=\Delta$ gain (almost) 2 space derivatives relative to the noise in the analytic semigroup scale?

Heat kernel smooths like order-2 parabolic operator: $S(t)$ maps $H^{s}\to H^{s+2-\varepsilon}$ with $t$-integrable singularities for stochastic integrals (factorization). Precise gain depends on noise color and dimension; white-in-time limits the time Hölder to $<1/2$.

Pitfalls¶

Space-time white noise vs space-white / time-smooth: different regularities; many numerical schemes assume colored noise.
Invariant measure existence $\neq$ pathwise global well-posedness: Lyapunov structure can give measures even when blow-up criteria are subtle.
Formal Itô formula in infinite dimensions needs Hilbert–Schmidt / trace-class care; naive finite-dimensional formulas drop divergent traces.
Renormalized equation $\neq$ original formal equation: counterterms $C_\varepsilon\to\infty$ appear; physics selects finite parts via Ward identities / symmetries when available.

Refs¶

Da Prato–Zabczyk; Hairer, Invent. Math. 2014 (regularity structures); Gubinelli–Imkeller–Perkowski; J. Martin–H. Weber surveys; Hairer, Singular SPDEs lecture notes.

Next¶

SPDE index · Regularity Structures · Paracontrolled & Renormalization · KPZ & Universality · Stochastic NS & Quantization · Stochastic processes · Martingales · MT / probability & stochastic — the Itô / the stochastic / the Wiener · PDE · Functional analysis · Microlocal analysis · Navier–Stokes — the deterministic Navier–Stokes regularity reference for the stochastic forcing comparison