Stochastic PDEs: Multiplicative Noise and Regularity Structures

Tier: R (Research). Builds on Itô Integration, Measure Theory, Functional Analysis and real analysis.

Attempt these from memory before reading.

• What is the stochastic heat equation? Write it down.
• What does it mean for a PDE to have a solution as a random variable?
• Why does the deterministic parabolic theory not transfer to the stochastic case?
• State the Wick (Wiener–Itô) product. What does $dW\,dW$ equal, and how does that differ from $dW\,dt$?
• What is the KPZ equation, and why is it harder than the stochastic heat equation?

Check your answers against the text. Prompt 4 is the one the whole subject turns on: a rule that is wrong in ordinary calculus is exactly right here.

Ordinary PDEs are deterministic. Given a forcing term you get one solution. Replace the forcing with Brownian motion and the question becomes strange: the “solution” is now a random field, and asking for it means deciding what kind of object it is, in which function space it lives, and what it means for that object to solve the equation.

The canonical example is the stochastic heat equation (in physics, the parabolic Anderson model):

where $\xi$ is space-time white noise and $u$ is the unknown. The multiplicative term is what makes it hard: the coefficient of the highest derivative is not a fixed constant but a random function, so the operator’s symbol fluctuates and the usual energy estimates have nothing to hold on to.

The naive reaction is wrong. Someone who has just learned Itô’s formula tries the deterministic route: integrate by parts, estimate the energy, hope for the best. The estimate fails, because the constant in the estimate is $\|\xi\|_\infty$, and space-time white noise is in no $L^p$ space at all. This is not a technical annoyance. It means the deterministic proof has no valid input, and a genuinely different theory is required.

In words. Ordinary calculus says $dx\,dx = 0$; infinitesimals shrink faster than they accumulate. Itô calculus says otherwise for Brownian motion: $dt\,dW \approx dt^{3/2}$ and $dW\,dW = dt$. Squaring noise is not infinitesimal at all; it produces a first-order term.

As equations. The Itô table has four entries, of which two are non-zero:

So for a smooth enough $F$, the second-order term in the Itô formula is

which is what makes the equation self-consistent: the second derivative term from the PDE cancels the correction from the product rule.

As a picture. Deterministic heat flow spreads a bump smoothly and predictably. The same equation with multiplicative noise has a random amplitude at every point, and because the product rule leaks, the evolution remembers every fluctuation it has ever experienced. That memory is the whole subject.

Worked example (U). Solve $\mathrm{d}X_t = \sigma X_t\,\mathrm{d}W_t$ by the Itô formula on $F(x) = \log x$:

so $X_t = X_0 \exp(\sigma W_t - \sigma^2 t/2)$. The drift term $-\sigma^2/2$ is entirely an artefact of the product rule. Removing it would give $\mathbb{E}[X_t] = X_0 \exp(\sigma^2 t / 2)$; exponential growth; which is the wrong answer and a standard trap.

The standard deterministic heat equation has a smooth solution for smooth initial data, and this is proved by energy estimates:

Run the same computation on the stochastic equation and the drift term becomes $2\xi u \partial_t u$, whose size is bounded by $\lVert \xi\rVert_\infty \lVert u\rVert^2$. That is the exact point of failure: white noise is a distribution, not a function, so $\lVert\xi\rVert_\infty$ does not exist. Saying it fails “because Gaussian variables exceed every threshold infinitely often” would be wrong; a Gaussian-valued function has perfectly finite $L^p$ norms. The reason is regularity: $\xi$ has parabolic Hölder regularity $-\tfrac{3}{2} - \epsilon$ in $1+1$ dimensions, so it is not a function at all and no pointwise norm of it is defined.

So the failure is structural, not a matter of tighter estimates. Anything built on $\sup|\xi|$; smoothing, boundary regularity, the standard parabolic smoothing story; is unavailable, and a different class of methods is needed.

Kolmogorov’s method (A). Estimate moment bounds

that is $p > 4$. This produces a continuous $L^p$-valued solution and is the classical approach. It works, but it stops exactly where it should: the solution is a random element of a coarse Banach space, not a function. The noise is integrated, not resolved.

The Martingale method (A). Solve the backward problem with an $L^1$ kernel, bound the maximal inequality, and sum. Yields solutions in $L^\Omega$ spaces that can be far rougher than $L^p$. Better suited to low regularity, and it is what makes KPZ-type solutions exist at all.

Neither gives what physics asks for. Physicists want pointwise behaviour: what does the solution do at a fixed point as a function of time? Both methods hand back a random element of a space that says nothing about pointwise values. Closing that gap needs the subject’s central invention.

The idea (A/R). Do not try to solve the equation directly. Instead, solve it formally to infinite order, integrating both the nonlinearity and the noise by whatever Itô calculus offers, and use the cancellations that emerge to prove that the resulting formal series actually converges. The infinite series is the solution; convergence is the theorem.

The three ideas that make it work:

• Local subcriticality. Noise of regularity $\alpha$ and nonlinearity of homogeneity $\beta$ combine into $k\alpha + \beta$. Requiring this $< \beta$ at every order is the subcriticality condition; it is what makes the renormalisation constants finite and the series convergent.
• Renormalisation. Where subcriticality fails, a counterterm must be subtracted; and in the subcritical case these counterterms are universal, depending only on the noise law.
• Paracontrolled calculus (A/R). The parallel method: model the solution as a sum of objects each controlled with respect to a different part of the problem (linear part, noise, resonant part), so that products of unknowns become products of knowns after subtracting the resonant terms.

Why this is hard to teach and worth learning. Every step above is a choice about which object to regard as “known”, and each choice is a bookkeeping decision that has to be redone from scratch for a new equation. The theory is less a set of theorems than a way of organising the bookkeeping; which is why it took until 2011 to formalise, and why it now underpins so much of the field.

One computation, and it is the one the whole page is about: the pointwise variance of the stochastic convolution, which is finite in one spatial dimension and infinite in two. The dimension where it diverges is exactly the dimension where renormalisation becomes necessary, so this single integral locates the critical dimension and proves that it is critical.

Theorem. Let $Z$ solve $\partial_t Z = \partial_x^2 Z + \xi$ with $Z(0) = 0$, where $\xi$ is space-time white noise. Then

Proof. The solution is the stochastic convolution with the heat kernel, $Z(t,x) = \int_0^t\!\!\int G(s, y)\,\xi(\mathrm{d}s\,\mathrm{d}y)$, and by the Itô isometry in space-time form,

Everything from here is the heat kernel. In $d$ dimensions $G(s,y) = (4\pi s)^{-d/2}\,e^{-|y|^2/4s}$, so

The remaining integral over $s$ is $\int_0^t (8\pi s)^{-d/2}\,\mathrm{d}s$, which converges at the origin iff $d/2 < 1$, i.e. iff $d = 1$. So:

$\blacksquare$

The logarithmic divergence in $d = 2$ is not a pathology of a bad estimate; the integral genuinely diverges, so $Z(t,x)$ is not a random variable with finite second moment in two dimensions; it is a distribution, of negative Hölder regularity, and no pointwise product with itself is defined.

What this buys the page’s central claim. For $\Phi^4_2$, $\partial_t u = \partial_x^2 u - u^3 + \xi$, write $u = Z + v$ with $Z$ the linear solution above. Then

and the terms $Z^2$ and $Z^3$ involve the pointwise square and cube of an object that is not a function; which is what the integral above forbids. The renormalised equation subtracts $3 C_\delta\, u_\delta$ where $C_\delta = \mathbb{E}[Z_\delta^2] \to \infty$ is the smeared-out version of the divergent integral: mollify $\xi$ at scale $\delta$ and $C_\delta$ is finite but grows without bound as $\delta \to 0$. In $d=1$ the same computation gives $C_\delta$ converging, which is why $\Phi^4_1$ needed no renormalisation and why the stochastic heat equation is a genuine function there.

So the theorem above is the reason the section on renormalisation exists: the critical dimension is exactly where this integral stops converging, and that is read off from the exponent of $s$ rather than from any deep regularity theory. The general statement; that subcriticality is a condition on regularities $\alpha, \beta$ combining as $k\alpha + \beta$; is the same computation carried out term by term for every product appearing in the nonlinearity, with the heat kernel scaling doing all the work.

The second row is the sharpest statement in the table: the Hölder exponent $<\tfrac{1}{4}$ in time is not a weakness of a proof but the true regularity of the solution, and reaching it was a genuine discovery rather than a technical estimate.

• The KPZ equation and the directed landscape. The KPZ equation is the universal scaling limit of interacting particle systems, and it sits exactly in the supercritical regime where the Wick product is the wrong renormalisation. Recent work identifies the whole landscape of solutions, not just the fixed-point one, and connects it to the Yang–Lee Hamiltonian.
• Three-dimensional quantum Yang–Mills via stochastic quantisation. The programme is to construct a four-dimensional superfield gauge theory as a stochastic PDE on the three-dimensional Minkowski space and prove the continuum limit exists. Rigorous partial results exist; the full continuum limit is open.
• Dynamic $\Phi^4_3$, where the construction is complete but the universality classes and critical exponents are still being established.
• Renormalisation group as a semi-group. The question of when renormalisation converges and what the limit depends on is open in general, not merely in specific models.
• Freedman’s inequality and stochastic integrability, where the best current bounds lose powers of the distance to criticality that sharp results would want.

Mix these with Itô Integration, Measure Theory, Functional Analysis, Ricci Flow and Limit Theorems.

1. (Itô.) Solve $\mathrm{d}X_t = X_t^2\,\mathrm{d}W_t$ by taking logarithms where you can, and explain exactly where and why the Itô correction appears. Compare with the same ODE under ordinary calculus.
2. (Measure theory.) Show that space-time white noise on $[0,T]\times[0,1]$ is not an element of $L^p$ for any $p$, by computing the distribution of its values on a small interval. This is the calculation that kills the deterministic proof.
3. (Functional analysis.) State and prove the Banach-valued Itô isometry, and use it to show that $\int_0^T \xi_s\,ds$ is Gaussian. Then explain why the deterministic energy estimate needs an $L^\infty$ norm that the previous problem showed does not exist.
4. (Limit theorems.) Show that $\sup_{t\le T}\lVert u(t)\rVert_{L^p}$ is finite for $p>4$ in the multiplicative case, and identify the exponent in the Kolmogorov condition as coming from where.
5. (Ricci flow.) Percolation on a graph and Ricci flow are related through the same renormalisation-group structure. Describe the analogy and say what makes it more than a coincidence; and where it fails.
6. (Research.) Take the SPDE $\partial_t u = \partial_x^2 u + u^3$ and decide whether it is subcritical in the sense above. Then explain what goes wrong when it is not, and identify the renormalisation counterterm that regularises it.

• Stochastic Calculus: Itô Integration and the Itô Formula - the product rule this subject lives or dies by
• Measure Theory; why white noise is not measurable in the spaces the estimates want
• Functional Analysis: Banach-valued integration and the spaces solutions live in
• Ricci Flow and Geometrisation: the same renormalisation thinking in geometry
• How to Use These Notes; the method