Energy solutions of KPZ are unique
Abstract
The Kardar–Parisi–Zhang (KPZ) equation is conjectured to universally describe the fluctuations of weakly asymmetric interface growth. Here we provide the first intrinsic well-posedness result for the stationary KPZ equation on the real line by showing that its energy solutions, as introduced by Gonçalves and Jara in 2010 and refined by Gubinelli and Jara, are unique. This is the first time that a singular stochastic PDE can be tackled using probabilistic methods, and the combination of the convergence results of the first work and many follow-up papers with our uniqueness proof establishes the weak KPZ universality conjecture for a wide class of models. Our proof builds on an observation of Funaki and Quastel from 2015, and a remarkable consequence is that the energy solution to the KPZ equation is not equal to the Cole–Hopf solution, but it involves an additional drift .
1. Introduction
The aim of this paper is to establish the well-posedness of the martingale problem for the stationary conservative stochastic Burgers equation (SBE) on ,
where is a continuous process in taking values in the space of (Schwartz) distributions over , , and , is a cylindrical Wiener process such that is a space-time white noise. A direct consequence will be the well-posedness of the martingale problem for the quasi-stationary Kardar–Parisi–Zhang (KPZ) equation
where is a continuous process, and for any the law of is a two-sided Brownian motion on The SBE describes the evolution of the weak derivative . of the solution to the KPZ equation Our uniqueness proof also establishes that . is related to the solution of the linear multiplicative stochastic heat equation (SHE),
by the Cole–Hopf transformation
The SHE allows a formulation via standard Itô calculus and martingale, weak, or mild solutions in suitable weighted spaces of continuous adapted processes. The SBE and the KPZ equation, on the other hand, cannot be studied in standard spaces due to the fact that the nonlinearity is ill-defined, essentially because the trajectories of the solutions do not possess enough spatial regularity. Indeed, solutions of the KPZ equation are of Hölder regularity less than in space, so a priori the pointwise square of their derivatives cannot be defined.
Despite this mathematical difficulty, the KPZ equation is expected to be a faithful description of the large scale properties of one-dimensional growth phenomena. This was the original motivation which led Kardar, Parisi, and Zhang Reference KPZ86 to study the equation, and both experimental and theoretical physics arguments have, since then, confirmed their analysis. The rigorous study of the KPZ equation and its relation with the SHE started with the work of Bertini and Giacomin Reference BG97 on the scaling limit of the weakly asymmetric exclusion process (WASEP). Starting from this discrete Markov process on and performing a suitable space-time rescaling and recentering, they were able to prove that its density fluctuation field converges to a random field which is linked to the solution of the SHE by the Cole–Hopf transformation Equation 4. Incidentally they had to add exactly the strange drift in order to establish their result. Their work clarifies that any physically relevant notion of solution to the (still conjectural) equations Equation 1 and Equation 2 needs to be transformed to the SHE by the Cole–Hopf transformation and also that the SBE should allow the law of the space white noise as invariant measure. A priori these insights are of little help in formulating the SBE/KPZ equation, since given a solution to the SHE it is not possible to apply Itô’s formula to and in particular the inverse Cole–Hopf transformation is ill-defined. It should be noted that the main difficulty of equations ,Equation 1 and Equation 2 lies in the spatial irregularity and that no useful martingales in the space variable are known, a fact which prevents an analysis via Itô’s stochastic integration theory. Moreover, the convergence result of Reference BG97 relies strongly on the particular structure of the WASEP and does not have many generalizations because most models behave quite badly under exponentiation (Cole–Hopf transformation); see Reference DT16Reference CT17Reference CST16Reference Lab17 for examples of models that do admit a useful Cole–Hopf transformation.
After the work of Bertini and Giacomin, there have been various attempts to study the SBE via Gaussian analysis tools taking into account the necessary invariance of the space white noise. A possible definition based on the Wick renormalized product associated to the driving space-time white noise has been ruled out because it lacks the properties expected from the physical solution Reference Cha00. Assing Reference Ass02 has been the first, to our knowledge, to attempt a martingale problem formulation of the SBE. He defines a formal infinite-dimensional generator for the process essentially as a quadratic form with dense domain, but he has not been able to prove its closability. The singular drift, which is ill-defined pointwise, make sense as a distribution on the Gaussian Hilbert space associated to the space white noise, however this distributional nature prevents the identification of a suitable domain for the formal generator.
The martingale problem approach has been subsequently developed by Gonçalves and Jara Reference GJ10Reference GJ14.Footnote1 Their key insight is that while the drift in Equation 1 is difficult to handle in a Markovian picture (that is, as a function on the state space of the process), it makes perfect sense in a pathwise picture. They proved in particular that a large class of particle systems (which generalize the WASEP studied by Bertini and Giacomin) has fluctuations that subsequentially converge to random fields which are solutions of a generalized martingale problem for Equation 1 where the singular nonlinear drift is a well-defined space-time distributional random field. Avoiding a description of a Markovian generator for the process, they manage to introduce an auxiliary process which plays the same role in the formulation of the martingale problem. Subsequent work of Jara and Gubinelli Reference GJ13 gave a different definition of the martingale problem via a forward-backward description. The solution of the martingale problem is a Dirichlet process, that is the sum of a martingale and a zero quadratic variation process. This property and the forward-backward decomposition of the drift are reminiscent of Lyons–Zheng processes and in general of the theory of Markov processes described by Dirichlet forms; however, a complete understanding of the matter is at the moment not well developed, and the martingale problem formulation avoids the subtleties of the Markovian setting. Gonçalves and Jara called the solutions of this generalized martingale problem energy solutions for the SBE/KPZ equation.
Following Reference GJ14, it has been shown for a variety of models that their fluctuations subsequentially converge to energy solutions of the KPZ equation or the SBE, for example for zero range processes and kinetically constrained exclusion processes in Reference GJS15, various exclusion processes in Reference GJS17Reference FGS16Reference BGS16Reference GJ16, interacting Brownian motions in Reference DGP17, and Hairer–Quastel type SPDEs in Reference GP16. This is coherent with the conjecture that the SBE/KPZ equation describes the universal behavior of a wide class of conservative dynamics or interface growth models in the particular limit where the asymmetry is “small” (depending on the spatial scale), the so-called weak KPZ universality conjecture; see Reference Cor12Reference Qua14Reference QS15Reference Spo16. In order to fully establish the conjecture for the models above, the missing step was a proof of uniqueness of energy solutions. This question remained open for some time during which it was not clear if the notion is strong enough to guarantee uniqueness or if it is too weak to expect well-posedness. Here we present a proof of uniqueness for the refined energy solutions of Reference GJ13, on the full line and on the torus, thereby finally establishing the well-posedness of the martingale problem and its expected relation with the SHE via the Cole–Hopf transform. In fact we even prove that the equation formulated in Reference GJ13 leads to strongly unique solutions, which directly gives uniqueness in law. The reason we emphasize the (weaker) uniqueness in law is that energy solutions frequently arise as scaling limits for particle systems.
Our proof follows the strategy developed by Funaki and Quastel in Reference FQ15. Namely, we map a mollified energy solution to the SHE via the Cole–Hopf transform, and we use a version of the Boltzmann–Gibbs principle to control the various error terms arising from the transformation and to derive the relation Equation 4 in the limit as we take the mollification away. A direct corollary of our results is the proof of the weak KPZ universality conjecture for all the models in the literature which have been shown to converge to energy solutions.
Shortly after the introduction of energy solutions, the fundamental work Reference Hai13 of Hairer on the KPZ equation appeared, where he established a pathwise notion of solution using Lyons’s theory of rough paths to provide a definition of the nonlinear term as a continuous bilinear functional on a suitable Banach space of functions. Existence and uniqueness were then readily established by fixed point methods. This breakthrough developed into a general theory of singular SPDEs, Hairer’s theory of regularity structures Reference Hai14, which provides the right analytic setting to control the singular terms appearing in stochastic PDEs, such as the SBE/KPZ equations and their generalizations, but also in other important SPDEs, such as the stochastic Allen–Cahn equation in dimensions and the (generalized) parabolic Anderson model in The work of the authors of this paper together with P. Imkeller on the use of paradifferential calculus .Reference GIP15 and the work of Kupiainen based on renormalization group (RG) techniques Reference Kup16Reference KM17 opened alternative ways to tackle singular SPDEs. All these approaches have in common that they control the a priori ill-defined nonlinearities in the equation using pathwise (deterministic) arguments, and it was not clear if a probabilistic understanding of such singular SPDEs is possible at all. Our uniqueness proof for the stationary martingale solution to the KPZ equation is a first indication that it is possible and is a problem worth investigating closer—even if our method of proof does not extend to other equations because we extensively use the specific structure of the KPZ equation, including the facts that its invariant measure is Gaussian, that its nonlinearity is antisymmetric, and most restrictively that it can be mapped to the SHE through the Cole–Hopf transform.
From the point of view of the weak KPZ universality conjecture, the pathwise approach is difficult to use and, for now, there are only a few convergence results using either regularity structures, paracontrolled distributions, or RG techniques; see Reference HQ15Reference HS15Reference GP17Reference Hos16. The martingale approach has the advantage that it is easy to implement, especially starting from discrete particle systems which often do not have the semilinear structure that is at the base of the pathwise theories.
The main limitation of the martingale approach to the SBE/KPZ equation is that currently it works only at stationarity. Using tools from the theory of hydrodynamic limits, it seems possible to extend the results to initial conditions with small relative entropy with respect to the stationary measure. However, this has not been done yet and dealing with even more singular initial conditions is a completely open problem. On the other hand, with energy solutions it is relatively easy to work on the real line, while in the pathwise approach this requires dealing with weighted function spaces, and the question of uniqueness seems still not clear.
To summarize, the main contribution of the present paper is a proof of uniqueness of energy solutions (in the refined formulation of Jara and Gubinelli Reference GJ13) on the real line and on the torus. We start in section 2 by introducing the notion of solution and the space of trajectories where solutions live. Subsequently, we discuss in section 3 several key estimates available in this space, estimates which allow us to control a large class of additive functionals. After these preliminaries we show in section 4 how to implement the Cole–Hopf transformation at the level of energy solutions and, by a careful control of some error terms, how to establish the Itô formula which proves the mapping from the SBE to the SHE. Using the uniqueness for the SHE, we conclude the uniqueness of energy solutions. In Appendix A we add some detail on how to modify the proof to deal with the case of periodic boundary conditions. Appendix B contains an auxiliary moment bound for the Cole-Hopf transformation of the stochastic Burgers equation.
2. Controlled processes and energy solutions
2.1. Burgers equation
In this section we follow Gonçalves and Jara Reference GJ14 and Gubinelli and Jara Reference GJ13 in defining stationary energy solutions to the SBE ,
where , and , is a space-time white noise. Recall that from a probabilistic point of view the key difficulty in making sense of Equation 5 is that we expect the law of (a multiple of) the white noise on to be invariant under the dynamics, but the square of the white noise can only be defined as a Hida distribution and not as a random variable. To overcome this problem, we first introduce a class of processes which at fixed times are distributed as the white noise but for which the nonlinear term is defined as a space-time distribution. In this class of processes it then makes sense to look for solutions of the Burgers equation Equation 5.
If is a filtered probability space, then for an adapted process with trajectories in we say that is a space-time white noise on that space if for all the process is a Brownian motion in the filtration with variance for all A (space .) white noise with variance is a random variable with values in such that is a centered Gaussian process with covariance If . we simply call , a white noise. Throughout, we write for the law of the white noise on .
We will see that contains the probabilistically strong solution to Equation 5, while is the space in which to look for probabilistically weak solutions.
Controlled processes were first introduced in Reference GJ13 on the circle, and the definition on the real line is essentially the same. For the process is the stationary Ornstein–Uhlenbeck process. It is the unique-in-law solution to the SPDE
with initial condition Allowing . has the intuitive meaning of considering perturbations of the Ornstein–Uhlenbeck process with antisymmetric drifts of zero quadratic variation. In this sense we say that a couple is a process controlled by the Ornstein–Uhlenbeck process.
As we will see below, for controlled processes we are able to construct some interesting additive functionals. In particular, for any controlled process, the Burgers drift makes sense as a space-time distribution:
The proof will be given in section 3.3. Note that the convolution is well-defined for and not only for because at fixed times is a white noise.
Now we can define what it means for a controlled process to solve the SBE.
The notion of energy solutions was introduced in a slightly weaker formulation by Gonçalves and Jara in Reference GJ14, and the formulation here is due to Reference GJ13. Note that strong stationary solutions correspond to probabilistically strong solutions, while energy solutions are probabilistically weak in the sense that we do not fix the probability space and the noise driving the equation. Our main result is that strong and weak uniqueness hold for our solutions.
Before we state this precisely, let us introduce some notation. Given with and we write for the unique solution to the linear multiplicative SHE
where If . is another function of this form, then where , is an random variable that is independent of -measurable and in particular , for all and which implies , for the distributional derivative .
The proof will be given in section 4.1.
2.2. KPZ equation
Once we understand how to deal with the SBE, it is not difficult to also handle the KPZ equation. Let be a space-time white noise on the filtered probability space and let be an random variable with values in -measurable such that is a space white noise. Then we denote with the space of pairs of adapted stochastic processes with trajectories in which solve for all the equation
and are such that and for which , , satisfy Similarly, we write . if there exist as above such that In Proposition .3.15 we show in fact the following generalization of Proposition 2.2: If and is an approximate identity (i.e., , for all , for all then there exists a process ), defined by ,
where the convergence takes place in uniformly on compacts, and the limit does not depend on the approximate identity .
So we call a strong almost-stationary solution to the KPZ equation
if Similarly, we call . an energy solution to Equation 11 if and The terminology “almost-stationary” comes from .Reference GJ14, and it indicates that for fixed the process is always a two-sided Brownian motion; however, the distribution of may depend on time. The analogous result to Theorem 2.4 is then the following.
The proof will be given in section 4.1.
2.3. The periodic case
It is also useful to have a theory for the periodic model where , and
for a periodic space-time white noise and a periodic space white noise For a process . with trajectories in where , are the (Schwartz) distributions on the circle, we say that is a periodic space-time white noise if for all the process is a Brownian motion with variance A periodic space white noise is a centered Gaussian process . with trajectories in such that for all , we have where , is the projection of onto the mean-zero functions. The reason for setting the zero Fourier mode of equal to zero is that the SBE is a conservation law and any solution to Equation 15 satisfies for all and therefore shifting , simply results in a shift of by the same value, for all So for simplicity we assume . Controlled processes are defined as before, except that now we test against . and all noises are replaced by their periodic counterparts. Then it is easy to adapt the proof of Proposition 3.15 to show that also in the periodic setting the Burgers drift is well-defined; alternatively, see Reference GJ13, Lemma 1. Thus, we define strong stationary solutions, respectively energy solutions, to the periodic Burgers equation exactly as in the nonperiodic setting. We then have the analogous uniqueness result to Theorem 2.4:
We explain in Appendix A how to modify the arguments for the nonperiodic case in order to prove Theorem 2.13.
3. Additive functionals of controlled processes
3.1. Itô trick and Kipnis–Varadhan inequality
Our main method for controlling additive functionals of controlled processes is to write them as a sum of a forward and a backward martingale which enables us to apply martingale inequalities. For that purpose we first introduce some notation. Throughout this section we fix .
For we define the action of the Ornstein–Uhlenbeck generator as
With the help of Itô’s formula it is easy to verify that if is the stationary Ornstein–Uhlenbeck process (see the discussion below Definition 2.1) and then , , is a martingale and in particular , is indeed the action of the generator of on We will see in Corollary .3.8 that can be uniquely extended from to a closed unbounded operator on also denoted by , so , is a core for We also define the Malliavin derivative .
for all and since , is the law of the white noise we are in a standard Gaussian setting and is closable as an unbounded operator from to for all see for example ;Reference Nua06. Similarly also is closable from to for all and we denote the domain of the resulting operator by , Then . is the completion of with respect to the norm So writing .
we have for all Finally, we denote .
The following martingale or Itô trick is well known for Markov processes; see for example the monograph Reference KLO12, and in the case of controlled processes on it is due to Reference GJ13. The proof is in all cases essentially the same.
The bound in Proposition 3.2 allows us to control provided that we are able to solve the Poisson equation
for all Note that this is an infinite-dimensional PDE which a priori is difficult to solve, but that we only need to consider it in . which has a lot of structure as a Gaussian Hilbert space. We will discuss this further in section 3.2. Nonetheless, we will encounter situations where we are unable to solve the Poisson equation explicitly, and in that case we rely on the method of Kipnis and Varadhan allowing us to bound in terms of a certain variational norm of We define for .
and we write if the right-hand side is finite. For details on and see ,Reference KLO12, Chapter 2.2. Here we just remark that for and we have
and on the other hand we have for all with and all and therefore ,
which proves that
We will need a slightly refined version of the Kipnis–Varadhan inequality which also controls the Recall that for -variation. the of -variation is defined as
3.2. Gaussian analysis
To turn the Itô trick or the Kipnis–Varadhan inequality into a useful bound, we must be able either to solve the Poisson equation for a given or to control the variational norm appearing in the Kipnis–Varadhan inequality. Here we discuss how to exploit the Gaussian structure of in order to do so. For details on Gaussian Hilbert spaces we refer to Reference Jan97Reference Nua06. Since is a Gaussian Hilbert space, we have the orthogonal decomposition
where is the closure in of the span of all random variables of the form with , being the Hermite polynomial and where th with The space . is called the homogeneous chaos, and th is the inhomogeneous chaos. Also, th is the law of the white noise on and therefore we can identify
where is the multiple Wiener–Itô integral of that is ,
Here are the equivalence classes of that are induced by the seminorm
where denotes the set of permutations of Of course, . is a norm on and we usually identify an equivalence class in with its symmetric representative. The link between the multiple stochastic integrals and the Malliavin derivative is explained in the following partial integration-by-parts rule, which will be used for some explicit computations below.
Recall that so far we defined the operator acting on cylinder functions. If we consider a cylinder function for some given then the action of , is particularly simple.
Before we continue, let us link the defined in section -norm3.1 with the operator .
Next, we define two auxiliary Hilbert spaces that will be useful in controlling additive functionals of controlled processes.
3.3. The Burgers and KPZ nonlinearity
With the tools we have at hand, it is now straightforward to construct the KPZ nonlinearity (and in particular the Burgers nonlinearity) for all controlled processes.
Proposition 2.2 about the Burgers drift follows by setting
4. Proof of the main results
4.1. Mapping to the stochastic heat equation
Let be a space-time white noise on the filtered probability space and let be an space white noise. Let -measurable be a strong stationary solution to the SBE
Our aim is to show that is unique up to indistinguishability, that is to prove the first part of Theorem 2.4. The case is well understood, so from now on let The basic strategy is to integrate . in the space variable and then to exponentiate the integral and to show that the resulting process solves (a variant of) the linear SHE. However, it is not immediately obvious how to perform the integration in such a way that we obtain a useful integral process. Note that any integral of is determined uniquely by its derivative and the value for a test function with So the idea, inspired by .Reference FQ15, is to fix one such test function and to consider the integral with for all .
More concretely, we take with and and consider the function
which satisfies for any
and
so is the unique integral of which vanishes when tested against Moreover, .
and in particular
where the notation means that the is taken in the variable -norm Now let . be an even function with and such that on a neighborhood of We write . and by our assumptions on , there exists for every an such that for all This will turn out to be convenient later. We also write . and and we define
Using that is even, we get So since . is a strong stationary solution to Equation 29, we have
and Equation 31 yields
Now set Then the Itô formula for Dirichlet processes of .Reference RV07 gives
and from Equation 32 we get
Therefore, Since moreover .
we obtain
Expanding the product and noting that -inner by Equation 33, we deduce that
where we introduced the processes
for the deterministic function
and
Integrating Equation 34 against we get
In section 4.2 we will prove the following three lemmas.
With the help of these results it is easy to prove our main theorem.
4.2. Convergence of the remainder terms
We now proceed to prove Lemmas 4.1–4.3 on the convergence of , and respectively. ,
4.2.1. Proof of Lemmas 4.1 and 4.2
To treat we introduce the auxiliary process
for
We will show in Lemma 4.5 that converges to for all and is uniformly bounded in and Since by assumption . and is a nice test function, we get from Corollary 3.17 that converges to in .
We also define
so that Using Corollary .3.5, we can estimate for
where we recall that
where are the cylinder functions and in terms of the Malliavin derivative associated to the measure We prove below that we can choose . so that for all This is necessary for . to be finite for all At this point everything boils down to controlling . and to showing that it goes to zero as first and then .
Observe that the random variable is an element of the second homogeneous chaos of Let us compute its kernel. From .Equation 31 we get
and furthermore
Therefore, let
so that
We also let Using the partial integration by parts derived in Lemma .3.6, we are able to bound by a constant:
So to control it remains to show that the two constants and vanish in the limit Before doing so, let us prove Lemma .4.2. More precisely, we show the following refined version:
The following computation will be useful for controlling both and which is why we outsource it in a separate lemma. ,
Lemma 4.1 now follows by combining Lemma 4.4, Lemma 4.7, and Lemma 4.8.
4.2.2. Proof of Lemma 4.3
Recall that
By Corollary 3.17 the first term on the right-hand side converges in to whenever and The convergence of the remaining terms is obvious, and overall we get .
where the convergence takes place in .
Appendix A. The periodic case
For the periodic equation described in section 2.3 most of the analysis works in the same way. The Itô trick and the Kipnis–Varadhan inequality are shown using exactly the same arguments, and also the Gaussian analysis of section 3.2 works completely analogously. We only have to replace all function spaces over by the corresponding spaces over say , by The construction of the Burgers nonlinearity and the proof of its time-regularity also carry over to the periodic setting, although we have to replace the integrals over . in Fourier space by sums over But since those sums can be estimated by the corresponding integrals, we get the same bounds. .
The first significant difference is in the construction of the integral. As discussed in section 4.1, any integral of is determined uniquely by its derivative and the value for some with The same is true on the circle, and here there is a canonical candidate for the function . namely the constant function , So let . be a pair of controlled processes, where is a periodic space-time white noise and is a periodic space white noise. Let be an even function with and such that on a neighborhood of and define
where respectively denotes the Fourier transform (respectively inverse Fourier transform) on the torus, is the convolution on the torus, and is the periodization of For the last identity in .Equation 59 we applied Poisson summation; see for example Reference GP15, Lemma 6. We then integrate by setting
where which corresponds to ,
or equivalently for From the representation as a Fourier multiplier, it is obvious that . and since we assumed that , for all we get Writing . for we get from the fact that , is a strong stationary solution of the periodic Burgers equation that
and From the expression . for we see that where , denotes the Dirac delta, and therefore for So setting . we have ,
and since we get
where we expanded the and defined -norm
for the constant (which is independent of with ) defined in Equation 61, and
From here on the proof is completely analogous to the nonperiodic setting provided that we establish the following three lemmas.
To prove these lemmas, we follow the argumentation in section 4.2. Here the kernel takes the form
and as in section 4.2 we see that we should choose
The proof of Lemma A.2 is not a trivial modification of the one of Lemma 4.2, so we provide the required arguments.
The rest of the proof is completely analogous to the nonperiodic case. Let us just point out that if then ,
where is a random variable with density and that ,
and therefore the same line of argumentation as in section 4.2 yields Lemmas A.1 and A.3. From here we follow the same steps as in the proof of Theorem 2.4 to establish Theorem 2.13.