Rough Heat Equation: Singular Parabolic Analysis

• Rough heat equation is a class of parabolic problems characterized by singular forcing, irregular potentials, or rough initial and boundary data.
• It utilizes advanced analytic frameworks such as rough paths and regularity structures to resolve nonclassical products and ensure well-posedness.
• Numerical methods for these equations demand discretizations in negative regularity spaces to capture distributional solution behavior accurately.

“Rough heat equation” denotes a family of deterministic and stochastic parabolic equations in which the heat operator is coupled to an irregular object: a distribution-valued noise or potential, a measure-valued or negative-regularity initial condition, or rough boundary data. In the contemporary SPDE literature, the canonical examples are one-dimensional stochastic heat equations driven by Gaussian noise that is white in time and fractional in space with Hurst parameter $H\in(1/4,1/2)$, and parabolic Anderson models with time-independent rough spatial noise (Hu et al., 2015Chakraborty et al., 2018). Closely related theories treat distribution-valued potentials through geometric rough paths, model singular multiplicative forcing by regularity structures, and develop maximal-regularity frameworks for rough initial or boundary traces (Kim et al., 2017Lindemulder et al., 2018). This usage should be distinguished from the separate fluid-mechanical literature on heat transfer over geometrically rough walls, where “roughness” refers to wall geometry rather than singular forcing or data (MacDonald et al., 2018).

1. Scope and meanings of roughness

The adjective “rough” is not attached to a single normal form. It identifies the locus of singularity in a parabolic problem. In one class, the rough object is the driving Gaussian field, typically white in time and distributional in space. In a second class, the rough object is a time-independent random potential $W(x)$, so that the equation becomes a parabolic Anderson model in a quenched random environment. In a third class, roughness is carried by the initial datum, by boundary traces, or by the coefficient multiplying the heat semigroup. These variants share a common obstruction: products such as $u\,\dot W$, $u\,W$, or trace constraints cease to be pointwise classical and must be interpreted in function spaces adapted to the parabolic scaling and to the singularity of the irregular input.

A useful way to organize the subject is by where the nonclassical object enters the PDE and by which analytic framework restores well-posedness.

Roughness locus	Representative model	Representative papers
Spatially rough Gaussian noise	$\partial_t u=\partial_{xx}u+\sigma(u)\dot W$	(Hu et al., 2015, Qian et al., 7 Jan 2025, Qian et al., 9 Apr 2026)
Time-independent rough environment	$\partial_t u=\Delta u+u\,W(x)$	(Chakraborty et al., 2018, Balan et al., 2022)
Distribution-valued potential	$\partial_t u=\partial_{xx}u+u\,W(x)$	(Kim et al., 2017)
Rough initial or boundary data	measure-valued initial data; rough Dirichlet traces	(Chen et al., 2013, Lindemulder et al., 2018)

This taxonomy also clarifies a recurring misconception. In stochastic analysis, “rough heat equation” usually refers to singular forcing, singular coefficients, or singular data. In applied fluid mechanics, by contrast, roughness may refer to a geometrically rough wall affecting forced convection. The two topics both involve heat transfer or diffusion, but they are analytically distinct.

2. White-in-time, rough-in-space stochastic heat equations

A central model is the one-dimensional stochastic heat equation

$$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$

with deterministic initial condition $u(0,x)=u_0(x)$, where $W$ is centered Gaussian, Brownian in time, and fractional-Brownian in space. Its covariance is

$W(x)$0

with $W(x)$1, equivalently with spatial spectral measure

$W(x)$2

Because $W(x)$3, the formal spatial covariance $W(x)$4 is not locally integrable; the noise is therefore rougher than standard colored noise and requires fractional Sobolev control rather than classical martingale-measure arguments (Hu et al., 2015).

In this regime, the natural mild formulation is

$W(x)$5

and well-posedness is obtained in spaces that control both moments and spatial fractional increments. Under the assumptions that $W(x)$6 is differentiable, $W(x)$7 is Lipschitz, and $W(x)$8, existence and uniqueness hold for the nonlinear equation. In the multiplicative case $W(x)$9, the solution admits a Wiener chaos expansion and a Feynman–Kac formula for moments, yielding sharp intermittency bounds of the form

$u\,\dot W$0

which makes the dependence of the moment growth rate on the spatial roughness parameter $u\,\dot W$1 explicit (Hu et al., 2015).

The same rough-noise mechanism persists for nonlocal diffusion. For the stochastic fractional heat equation

$u\,\dot W$2

with spatial covariance of fractional-Brownian type and

$u\,\dot W$3

weak and strong well-posedness can be proved in the weighted space $u\,\dot W$4, built from the weight

$u\,\dot W$5

This weighted formulation removes the technical condition $u\,\dot W$6 that had been required in earlier rough-noise theories. It also yields continuous sample paths and Hölder regularity with temporal exponent $u\,\dot W$7 and spatial exponent $u\,\dot W$8 for any

$u\,\dot W$9

(Qian et al., 9 Apr 2026).

3. Time-independent rough noise and parabolic Anderson asymptotics

A second major branch studies heat equations in a static random environment. In one dimension, the parabolic Anderson model with time-independent fractional spatial noise takes the form

$u\,W$0

where $u\,W$1 is a centered Gaussian generalized field with spectral measure

$u\,W$2

Here roughness is purely spatial and more singular than white noise in the sense relevant to the analysis. The product $u\,W$3 is interpreted pathwise in weighted Besov spaces by a Young/Riemann–Stieltjes construction rather than through Itô calculus. Existence and uniqueness are obtained by combining almost sure Besov regularity of the noise with a Picard fixed-point scheme, and the Feynman–Kac representation

$u\,W$4

then connects long-time growth to the principal eigenvalue of the random operator $u\,W$5 (Chakraborty et al., 2018).

The spectral asymptotics are quenched: they hold almost surely for fixed realizations of the environment. On large boxes $u\,W$6, the principal eigenvalue satisfies

$u\,W$7

and the large-time growth of the solution is asymptotically governed by this top spectral mode. The conceptual point is that intermittency is encoded by an optimization problem for Brownian occupation in a rough random potential, rather than by white-in-time forcing (Chakraborty et al., 2018).

Macroscopic averaging of the same model displays a different universality. For the spatial integral

$u\,W$8

under the rough-noise assumption

$u\,W$9

the variance grows linearly,

$\partial_t u=\partial_{xx}u+\sigma(u)\dot W$0

the quantitative central limit theorem has the rate

$\partial_t u=\partial_{xx}u+\sigma(u)\dot W$1

and the normalized process $\partial_t u=\partial_{xx}u+\sigma(u)\dot W$2 converges in $\partial_t u=\partial_{xx}u+\sigma(u)\dot W$3 to a centered Gaussian process with covariance $\partial_t u=\partial_{xx}u+\sigma(u)\dot W$4 (Balan et al., 2022). The singular covariance thus changes the microscopic analysis but not the $\partial_t u=\partial_{xx}u+\sigma(u)\dot W$5 normalization or Gaussian functional limit for large-scale spatial averaging.

4. Rough-path and regularity-structure formulations

When the singular object is a spatial potential rather than a stochastic integral, rough-path methods provide a direct interpretation of the product. In one dimension, the heat equation with rough potential

$\partial_t u=\partial_{xx}u+\sigma(u)\dot W$6

with zero Dirichlet boundary conditions can be defined by regularization when $\partial_t u=\partial_{xx}u+\sigma(u)\dot W$7 is the generalized derivative of a Hölder function. The central device is the conjugation

$\partial_t u=\partial_{xx}u+\sigma(u)\dot W$8

which transforms the equation into a standard parabolic PDE with rough coefficients. If $\partial_t u=\partial_{xx}u+\sigma(u)\dot W$9, $\partial_t u=\Delta u+u\,W(x)$0, and $\partial_t u=\Delta u+u\,W(x)$1, then the classical geometric rough path solution is

$\partial_t u=\Delta u+u\,W(x)$2

where $\partial_t u=\Delta u+u\,W(x)$3 is the semigroup generated by

$\partial_t u=\Delta u+u\,W(x)$4

The resulting regularity is

$\partial_t u=\Delta u+u\,W(x)$5

with convergence of regularized solutions in $\partial_t u=\Delta u+u\,W(x)$6 and an energy-theoretic generalized solution for merely $\partial_t u=\Delta u+u\,W(x)$7 potentials and $\partial_t u=\Delta u+u\,W(x)$8 initial data (Kim et al., 2017).

A more singular setting is the modelled rough heat equation

$\partial_t u=\Delta u+u\,W(x)$9

where $\partial_t u=\partial_{xx}u+u\,W(x)$0 is the distributional derivative of a fractional sheet with Hurst indices $\partial_t u=\partial_{xx}u+u\,W(x)$1. Here the decisive roughness parameter is

$\partial_t u=\partial_{xx}u+u\,W(x)$2

If $\partial_t u=\partial_{xx}u+u\,W(x)$3, the equation belongs to a Young-type regime and no renormalization is needed. If

$\partial_t u=\partial_{xx}u+u\,W(x)$4

the canonical second-order object diverges and one must renormalize the approximating equation by subtracting

$\partial_t u=\partial_{xx}u+u\,W(x)$5

This is implemented via Hairer’s regularity structures: the noise is lifted to a model containing symbols such as $\partial_t u=\partial_{xx}u+u\,W(x)$6, $\partial_t u=\partial_{xx}u+u\,W(x)$7, and $\partial_t u=\partial_{xx}u+u\,W(x)$8, the fixed-point problem is solved in modelled-distribution spaces, and the renormalized limit is reconstructed as a pathwise solution (Deya, 2014). In the white-in-time case $\partial_t u=\partial_{xx}u+u\,W(x)$9, the limit coincides with the classical Itô solution.

These two theories illustrate a structural divide within rough heat equations. Rough-path formulations resolve distribution-valued spatial coefficients by algebraic enhancement and conjugation, whereas regularity structures handle genuinely space-time singular multiplicative forcing by extending the model space and, when required, introducing deterministic counterterms.

5. Rough initial data, negative regularity, and boundary traces

Roughness may also be present before any stochastic forcing is applied. For the one-dimensional nonlinear stochastic heat equation driven by space-time white noise,

$$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$0

initial data are allowed to be measures $$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$1 in the class

$$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$2

This includes $$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$3, signed measures, noncompactly supported measures, and even measures with exponentially growing tails. Existence and uniqueness of a random-field solution are established without Gronwall’s lemma by explicit control of the Picard convolution kernels, and in the parabolic Anderson case $$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$4 the second moment is computed exactly. The framework is sharp in that it cannot, in general, be extended beyond measures: for $$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$5 and $$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$6, there is no random-field solution (Chen et al., 2013).

The regularity of such solutions near the initial line depends explicitly on the roughness of $$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$7. Away from $$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$8, the classical stochastic heat exponents persist: $$\partial_t u(t,x)=\frac12 \partial_{xx}u(t,x)+\sigma(u(t,x))\,W(dt,dx),$$9 On compact sets including $u(0,x)=u_0(x)$0, however, the regularity is limited by the initial datum. If $u(0,x)=u_0(x)$1 with $u(0,x)=u_0(x)$2 $u(0,x)=u_0(x)$3-Hölder continuous, then

$u(0,x)=u_0(x)$4

The solution still converges to $u(0,x)=u_0(x)$5 as $u(0,x)=u_0(x)$6 in the distributional sense, but pointwise continuity at $u(0,x)=u_0(x)$7 can fail for singular measures (Chen et al., 2013).

For deterministic semilinear heat flow, rough initial data can be pushed into negative Sobolev scales. The $u(0,x)=u_0(x)$8-critical equation

$u(0,x)=u_0(x)$9

is locally and, under a smallness condition, globally well posed for radial data supported away from the origin in $W$0, with unconditional uniqueness for $W$1 and decay

$W$2

(Soffer et al., 2019). This is a deterministic analogue of rough-data well-posedness: the singularity lies in the initial profile, and the heat semigroup plus geometric restrictions recover sufficient spacetime integrability.

Boundary roughness requires yet another framework. For the Dirichlet Laplacian on a bounded $W$3-domain, weighted $W$4-spaces with

$W$5

admit a bounded $W$6-calculus of angle $W$7. This yields maximal $W$8-$W$9 regularity for the heat equation with inhomogeneous Dirichlet data in anisotropic trace spaces that are rougher than those covered by the classical $W(x)$00-weight theory (Lindemulder et al., 2018). In this setting, roughness is encoded by weighted trace spaces rather than by stochastic singularities.

6. Fine path properties, large-scale growth, and discretization

Once well-posedness is established, rough heat equations exhibit highly structured local and global asymptotics. For the nonlinear stochastic heat equation on $W(x)$01 driven by noise white in time and fractional in space with $W(x)$02,

$W(x)$03

the temporal increment at fixed $W(x)$04 is asymptotically governed by the linearized equation. More precisely,

$W(x)$05

and, after normalization, behaves like a fractional Brownian motion increment with Hurst index $W(x)$06. This leads to Khintchine’s law of the iterated logarithm, Chung’s law of the iterated logarithm, a Gaussian-mixture limit theorem for normalized increments, and weighted $W(x)$07-variation with the exact exponent

$W(x)$08

(Qian et al., 7 Jan 2025). The nonlinear rough heat equation therefore has temporal local behavior that is universal up to the random amplitude $W(x)$09.

For the linear stochastic fractional heat equation

$W(x)$10

with rough spatial dependence $W(x)$11, the covariance geometry can be analyzed exactly. The variance satisfies

$W(x)$12

and the canonical metric yields spatial Hölder exponent $W(x)$13 and temporal exponent $W(x)$14. Suprema over expanding domains acquire the expected $W(x)$15-growth, and the Hölder coefficients have sharp asymptotics derived from Talagrand’s majorizing measure theorem and Sudakov’s minoration theorem (Liu et al., 30 Jul 2025). These estimates quantify the combined effect of stable diffusion and spatial roughness.

Numerical analysis of rough heat equations is correspondingly function-space sensitive. For a very rough fractional space-time noise in the regime

$W(x)$16

a full discretization of the linear rough fractional heat equation is constructed in three stages: mollification of the noise, discretization of the smoothened noise by Gaussian rectangle increments, and Galerkin finite-element discretization of the heat operator on $W(x)$17. Convergence holds only in local negative Sobolev spaces, reflecting the fact that the solution is distribution-valued rather than function-valued (Deya et al., 2021). For additive-noise semilinear stochastic heat equations with rough initial data, a modified exponential Euler method combined with a spectral discretization attains strong convergence of order $W(x)$18 in both time and space when the stochastic convolution is controlled in stochastic Besov spaces; this recovers the sharp rate $W(x)$19 for one-dimensional space-time white noise by using graded time steps and real interpolation rather than classical Sobolev norms (Gui et al., 2023).

Taken together, these results show that the rough heat equation is best understood as a parabolic singular-analysis paradigm rather than as a single model. Its modern theory encompasses weighted and Besov solution spaces, Wiener chaos and Feynman–Kac techniques, quenched spectral asymptotics, rough-path and regularity-structure renormalization, and discretizations that converge only in negative regularity scales. The unifying theme is that the heat semigroup still regularizes, but only after the singular object has been represented in a framework compatible with its precise roughness class.