variational methods for a singular spde yielding the

63
VARIATIONAL METHODS FOR A SINGULAR SPDE YIELDING THE UNIVERSALITY OF THE MAGNETIZATION RIPPLE RADU IGNAT, FELIX OTTO, TOBIAS RIED, AND PAVLOS TSATSOULIS A. The magnetization ripple is a microstructure formed in thin ferromagnetic lms. It can be described by minimizers of a nonconvex energy functional leading to a nonlocal and nonlinear elliptic SPDE in two dimensions driven by white noise, which is singular. We address the universal character of the magnetization ripple using variational methods based on ฮ“-convergence. Due to the innite energy of the system, the (random) energy functional has to be renormalized. Using the topology of ฮ“-convergence, we give a sense to the law of the renormalized functional that is independent of the way white noise is approximated. More precisely, this universality holds in the class of (not necessarily Gaussian) approximations to white noise satisfying the spectral gap inequality, which allows us to obtain sharp stochastic estimates. As a corollary, we obtain the existence of minimizers with optimal regularity. C . Introduction . Estimates for the Burgers equation . ฮ“-convergence of the renormalized energy . A priori estimate for minimizers in Hรถlder spaces . Approximations to white noise under the spectral gap assumption Appendix A. Hรถlder spaces Appendix B. Besov spaces Appendix C. Stochastic estimates Appendix D. Some estimates for the linear equation Appendix E. Regularity of nite-energy solutions for smooth data Appendix F. Approximation of quadratic functionals of the noise by cylinder functionals References . I We study minimizers of the energy functional ( ) := โˆซ T 2 ( 1 ) 2 d + โˆซ T 2 (| 1 | - 1 2 ( 2 - 1 1 2 2 )) 2 d - 2 โˆซ T 2 d (.) where is (periodic) white noise, โˆˆ R, T 2 = [0, 1) 2 is the two-dimensional torus and : T 2 โ†’ R is a periodic function with vanishing average in 1 , i.e., โˆซ 1 0 d 1 = 0 for all 2 โˆˆ[0, 1) . Date: October , , version ripple-arxiv-v. Mathematics Subject Classication. H, J; A, D. Key words and phrases. Singular stochastic PDE, nonlocal elliptic PDE, regularity theory, renormalized energy, ฮ“-convergence, micromagnetics, Burgers equation. ยฉ by the authors. Faithful reproduction of this article, in its entirety, by any means is permitted for noncommercial purposes. arXiv:2010.13123v1 [math.PR] 25 Oct 2020

Upload: others

Post on 17-Oct-2021

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE YIELDING THE UNIVERSALITYOF THE MAGNETIZATION RIPPLE

RADU IGNAT, FELIX OTTO, TOBIAS RIED, AND PAVLOS TSATSOULIS

Abstract. The magnetization ripple is a microstructure formed in thin ferromagnetic lms. It canbe described by minimizers of a nonconvex energy functional leading to a nonlocal and nonlinearelliptic SPDE in two dimensions driven by white noise, which is singular. We address the universalcharacter of the magnetization ripple using variational methods based on ฮ“-convergence. Due tothe innite energy of the system, the (random) energy functional has to be renormalized. Usingthe topology of ฮ“-convergence, we give a sense to the law of the renormalized functional thatis independent of the way white noise is approximated. More precisely, this universality holdsin the class of (not necessarily Gaussian) approximations to white noise satisfying the spectralgap inequality, which allows us to obtain sharp stochastic estimates. As a corollary, we obtain theexistence of minimizers with optimal regularity.

Contents

1. Introduction 12. Estimates for the Burgers equation 143. ฮ“-convergence of the renormalized energy 194. A priori estimate for minimizers in Hรถlder spaces 265. Approximations to white noise under the spectral gap assumption 30Appendix A. Hรถlder spaces 39Appendix B. Besov spaces 45Appendix C. Stochastic estimates 53Appendix D. Some estimates for the linear equation 55Appendix E. Regularity of nite-energy solutions for smooth data 56Appendix F. Approximation of quadratic functionals of the noise by cylinder functionals 60References 62

1. Introduction

We study minimizers of the energy functional

๐ธ๐‘ก๐‘œ๐‘ก (๐‘ข) :=โˆซT2(๐œ•1๐‘ข)2 d๐‘ฅ +

โˆซT2( |๐œ•1 |โˆ’

12 (๐œ•2๐‘ข โˆ’ ๐œ•1

12๐‘ข

2))2 d๐‘ฅ โˆ’ 2๐œŽโˆซT2b๐‘ข d๐‘ฅ (1.1)

where b is (periodic) white noise, ๐œŽ โˆˆ R, T2 = [0, 1)2 is the two-dimensional torus and ๐‘ข : T2 โ†’ Ris a periodic function with vanishing average in ๐‘ฅ1, i.e.,โˆซ 1

0๐‘ข d๐‘ฅ1 = 0 for all ๐‘ฅ2 โˆˆ [0, 1) .

Date: October 27, 2020, version ripple-arxiv-v1.2020 Mathematics Subject Classication. 60H17, 35J60; 78A30, 82D40.Key words and phrases. Singular stochastic PDE, nonlocal elliptic PDE, regularity theory, renormalized energy,

ฮ“-convergence, micromagnetics, Burgers equation.ยฉ2020 by the authors. Faithful reproduction of this article, in its entirety, by any means is permitted for noncommercialpurposes.

1

arX

iv:2

010.

1312

3v1

[m

ath.

PR]

25

Oct

202

0

Page 2: Variational methods for a singular SPDE yielding the

2 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

The energy functional ๐ธ๐‘ก๐‘œ๐‘ก was considered in [IO19] as a reduced model describing the mag-netization ripple, a microstructure formed by the magnetization in a thin ferromagnetic lm,which is a result of the polycrystallinity of the material. In thin lms, the magnetization canapproximately be described by a two-dimensional unit-length vector eld (in the lm plane);the ripple is a perturbation of small amplitude of the constant state, say (1, 0). In this context,the function ๐‘ข corresponds to the transversal component of the magnetization, after a suitablerescaling. The theoretical treatment in the physics literature [Hof68, Har68] takes it for grantedthat the ripple is universal, in the sense that it does not depend on the precise composition andgeometry of the polycrystalline material. Our main result gives a rigorous justication of theuniversal behavior of the ripple (see Remark 1.5).

The rst term in ๐ธ๐‘ก๐‘œ๐‘ก can be interpreted as the exchange energy, an attractive short-rangeinteraction of the spins. The second term is the energy of the stray eld generated by themagnetization; the fractional structure is due to the scaling of the stray eld in the thin lmregime. On the scales that are relevant for the description of the magnetization ripple, the noiseacts like a random transversal eld of white-noise character. It comes via the crystalline anisotropyfrom the fact that the material is made up of randomly oriented grains that are smaller than theripple scale, which is set to unity in the abovemodel. In view of its origin, it is reasonable to assumethat this noise, which is quenched as opposed to thermal in character, is isotropic, neverthelessthe nonlocal interaction given by the stray eld energy leads to an anisotropic response of themagnetization. For a more in-depth description and formal derivation of the energy ๐ธ๐‘ก๐‘œ๐‘ก we referto the discussion in [IO19, Section 2], which in turn follows [SSWMO12].

Formally, critical points of ๐ธ๐‘ก๐‘œ๐‘ก are solutions to the Euler-Lagrange equation

(โˆ’๐œ•21 โˆ’ |๐œ•1 |โˆ’1๐œ•22)๐‘ข + ๐‘ƒ(๐‘ข๐‘…1๐œ•2๐‘ข โˆ’ 1

2๐‘ข๐‘…1๐œ•1๐‘ข2)+ 12๐‘…1๐œ•2๐‘ข

2 = ๐œŽ๐‘ƒb, (1.2)

where ๐‘…1 = |๐œ•1 |โˆ’1๐œ•1 is the Hilbert transform acting on the ๐‘ฅ1 variable, see (1.15), and ๐‘ƒ is the๐ฟ2-orthogonal projection on functions of zero average in ๐‘ฅ1 (extended to periodic Schwartzdistributions in the natural way). One of the main challenges of this equation is that the right-hand side of (1.2) is too irregular to make sense of the nonlinear terms, even though the nonlocalelliptic operator

L := โˆ’๐œ•21 โˆ’ |๐œ•1 |โˆ’1๐œ•22has the expected regularizing properties.

If we endow our space with a Carnotโ€“Carathรฉodory metric which respects the natural scalinginduced by L, that is, one derivative in the ๐‘ฅ1 direction costs as much as 2

3 derivatives in the ๐‘ฅ2direction, the eective dimension in terms of scaling is given by dim = 5

2 . It is well-known that inthis case b is a Schwartz distribution of regularity just below โˆ’dim

2 , i.e., a Schwartz distributionof order โˆ’ 5

4โˆ’ (measured in a scale of Hรถlder spaces C๐›ผ associated to this Carnotโ€“Carathรฉodorymetric; see Section 1.1.3 below for the denition), where for ๐›ผ โˆˆ R, we use the notation ๐›ผโˆ’ todenote ๐›ผ โˆ’ Y for any Y > 0 (suitably small).

We now argue that the nonlinear term ๐‘ข๐‘…1๐œ•2๐‘ข on the left-hand side of (1.2) is ill-dened:On the one hand, Schauder theory for the operator L improves regularity by 2 degrees on theHรถlder scale, indicating that the expected regularity of a solution ๐‘ข is (2 โˆ’ 5

4 )โˆ’ = 34โˆ’. On the

other hand, in our anisotropic scaling, one derivative in the ๐‘ฅ2 direction reduces regularity by32 , while the Hilbert transform has a negligible eect on the regularity. Hence the regularity ofthe Schwartz distribution ๐‘…1๐œ•2๐‘ข is โˆ’ 3

4โˆ’. It is well-known that the product of a function and aSchwartz distribution can be classically and unambiguously dened only if the regularities ofthe individual terms sum up to a strictly positive number. In the case of the product ๐‘ข๐‘…1๐œ•2๐‘ข ofthe function ๐‘ข and the Schwartz distribution ๐‘…1๐œ•2๐‘ข, the sum of regularities is 0โˆ’, not allowing itstreatment by means of classical analysis.

This is a common problem in the theory of singular Stochastic Partial Dierential Equations(SPDEs), which has become a very active eld in the recent years. Here the word singular relates

Page 3: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 3

to the fact that the driving noise of these equations is so irregular that their nonlinear terms(which usually involve products of the solution and its derivatives) are not classically dened. Werefer the reader to [Hai14] for a more detailed exposition of the theory of singular SPDEs.

In [IO19] the well-posedness of (1.2) for noise strength ๐œŽ below a โ€“ random โ€“ threshold wasstudied based on Banachโ€™s xed point argument. The ill-dened product ๐‘ข๐‘…1๐œ•2๐‘ข was treatedvia a more direct renormalization technique (in contrast to the more general one appearingin the framework of Regularity Structures), known as Wick renormalization. In fact, a similartechnique had been introduced by Da Prato and Debussche in their work [DD03] on the stochasticquantization equations of the P(๐œ‘)2-Euclidean Quantum Field theory.

One of the goals of this paper is to get rid of the smallness condition from [IO19]. Without lossof generality we may therefore assume that the parameter ๐œŽ = 1, which we will always do in thefollowing.1 This means in particular that we have to give up the use of a xed point theorem onthe level of the Euler-Lagrange equation, and use instead the direct method of the calculus ofvariations on the level of the functional. The functional ๐ธ๐‘ก๐‘œ๐‘ก is in need of a renormalization. Thisis indicated by the fact that if ๐‘ฃ is the unique solution with zero average in ๐‘ฅ1 to the linearizedEulerโ€“Lagrange equation2

(โˆ’๐œ•21 โˆ’ |๐œ•1 |โˆ’1๐œ•22)๐‘ฃ = ๐‘ƒb, (1.3)which is explicit on the level of its Fourier transform, one has that ๐ธ๐‘ก๐‘œ๐‘ก (๐‘ฃ) = โˆ’โˆž almost surely,see Proposition C.1. As in [DD03] and [IO19], we decompose any admissible conguration ๐‘ข in(1.1) into ๐‘ข = ๐‘ฃ +๐‘ค , where the remainder๐‘ค is a periodic function with vanishing average in ๐‘ฅ1.

As is usual in renormalization, we may approximate white noise by a probability measure thatis supported on smooth b โ€™s. This allows for a pathwise approach: For smooth (and periodic) b ,the solution ๐‘ฃ of (1.3) is smooth, too (see Lemma D.3). In view of the almost sure divergence of๐ธ๐‘ก๐‘œ๐‘ก (๐‘ฃ) in case of the white noise, we consider the renormalized functional

๐ธ๐‘Ÿ๐‘’๐‘› := ๐ธ๐‘ก๐‘œ๐‘ก (๐‘ฃ + ยท) โˆ’ ๐ธ๐‘ก๐‘œ๐‘ก (๐‘ฃ) . (1.4)

It follows from Lemma D.3 that for smooth b , ๐ธ๐‘Ÿ๐‘’๐‘› is well-dened (with values in R) on the space

W :={๐‘ค โˆˆ ๐ฟ2(T2) :

โˆซ 1

0๐‘ค d๐‘ฅ1 = 0 for every ๐‘ฅ2 โˆˆ [0, 1), H(๐‘ค) < โˆž

}, (1.5)

whereH denotes the harmonic energy, i.e., the quadratic part of ๐ธ๐‘ก๐‘œ๐‘ก given by

H(๐‘ค) :=โˆซT2(๐œ•1๐‘ค)2 d๐‘ฅ +

โˆซT2( |๐œ•1 |โˆ’

12 ๐œ•2๐‘ค)2 d๐‘ฅ . (1.6)

Loosely speaking, the task now is to show that ๐ธ๐‘Ÿ๐‘’๐‘› can still be given a sense as we approximatethe white noise. We will consider approximations that belong to the following class of probabilitymeasures:

Assumption 1.1. We consider the class of probability measures ใ€ˆยทใ€‰ on the space of periodic Schwartzdistributions

3 b (endowed with the Schwartz topology), satisfying the following:

(i) ใ€ˆยทใ€‰ is centered: ใ€ˆbใ€‰ = 0, that is, ใ€ˆ|b (๐œ‘) |ใ€‰ < โˆž and ใ€ˆb (๐œ‘)ใ€‰ = 0 for all ๐œ‘ โˆˆ Cโˆž(T2).(ii) ใ€ˆยทใ€‰ is stationary, that is, for every shift vector โ„Ž โˆˆ R2, b and b (ยท + โ„Ž) have the same law.

4

1Note however, that all our results also hold for ๐œŽ โ‰  1 by considering ๐œŽ๐‘ฃ instead of ๐‘ฃ .2From now on, given b , we denote by ๐‘ฃ the unique solution with zero average in ๐‘ฅ1 to the linearized equation

L๐‘ฃ = ๐‘ƒb from (1.3).3In our notation, we do not distinguish between the probability measure and its expectation, and use ใ€ˆยทใ€‰ to denote

in particular the latter. In the probability jargon, b plays the role of a dummy variable like the popular ๐œ” . We prefer toadopt this point of view, but sometimes it is convenient to also think of b as a random variable taking values in the spaceof periodic Schwartz distributions by identifying it with the canonical evaluation b โ†ฆโ†’ ev(b), where ev(b) (๐œ‘) := b (๐œ‘)for all ๐œ‘ โˆˆ Cโˆž (T2). In our notation, when we refer to the law of b , we mean the law of the random variable ev orrather the probability measure ใ€ˆยทใ€‰.

4More precisely, for any test function ๐œ‘ โˆˆ Cโˆž (T2) and shift vector โ„Ž โˆˆ T2, b (๐œ‘) = b (๐œ‘ (ยท โˆ’ โ„Ž)) in law.

Page 4: Variational methods for a singular SPDE yielding the

4 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

(iii) ใ€ˆยทใ€‰ is invariant under reection in ๐‘ฅ1, that is, b and ๐‘ฅ โ†ฆโ†’ b (โˆ’๐‘ฅ1, ๐‘ฅ2) have the same law.5

(iv) ใ€ˆยทใ€‰ satises the spectral gap inequality (SGI), meaning that6

โŸจ|๐บ (b) โˆ’ ใ€ˆ๐บ (b)ใ€‰|2

โŸฉ 12 โ‰ค

โŸจ ๐œ•๐œ•b๐บ (b) 2๐ฟ2

โŸฉ 12

, (1.7)

for every functional๐บ on the space of Schwartz distribution such that ใ€ˆ|๐บ (b) |ใ€‰ < โˆž and which

is well-approximated by cylindrical functionals. More precisely, for cylindrical functionals

๐บ (b) = ๐‘”(b (๐œ‘1), . . . , b (๐œ‘๐‘›)) with ๐‘” โˆˆ Cโˆž(R๐‘›) which itself and all its derivatives have at

most polynomial growth, ๐œ‘1, . . . , ๐œ‘๐‘› โˆˆ Cโˆž(T2), and ๐‘› โˆˆ N we dene

๐œ•

๐œ•b๐บ (b) :=

๐‘›โˆ‘๏ธ๐‘–=1

๐œ•๐‘–๐‘”(b (๐œ‘1), . . . , b (๐œ‘๐‘›))๐œ‘๐‘–

which is a random eld, and complete the space of all cylindrical functionals with respect to

the norm

โŸจ|๐บ (b) |2

โŸฉ 12 +

โŸจ ๐œ•๐œ•b๐บ (b)

2๐ฟ2

โŸฉ 12, which can be identied with a subspace of ๐ฟ2ใ€ˆยทใ€‰ for

which (1.7) is well-dened. 7

Remark 1.2. Note that the white noise as well as any Gaussian probability measure with aCameronโ€“Martin space that is weaker than ๐ฟ2(T2) satisfy the spectral gap inequality (1.7), (see[Hel98, Theorem 2.1]). In particular, if we convolve white noise with a smooth mollier, theresulting random eld satises (1.7).

Remark 1.3. For a linear functional ๐บ , i.e. ๐บ of the form ๐บ (b) = b (๐œ™), for some ๐œ™ โˆˆ Cโˆž(T2),the spectral gap inequality (1.7) turns into ใ€ˆb (๐œ™)2ใ€‰ โ‰ค

โˆซ๐œ™2 d๐‘ฅ , which is a dening property of

white noise turned into an inequality. Note that this allows us to extend b (๐œ™) to ๐œ™ โˆˆ ๐ฟ2(T2) asa centered random variable in ๐ฟ2ใ€ˆยทใ€‰ which is admissible in (1.7). In our application, the spectralgap inequality implies that ใ€ˆยทใ€‰ gives full measure to the Hรถlder space Cโˆ’ 5

4โˆ’ (see Proposition 1.8below), which is the same as the regularity of white noise.

The merit of SGI is that it also applies to nonlinear ๐บ (in this paper, we need it for quadratic๐บ)8. In addition, it allows us to obtain sharp stochastic estimates for non-Gaussian measures byproviding a substitute for Nelsonโ€™s hyper-contractivity.

The second, and more subtle, goal of this paper, is to establish universality of the ripple. Bythis we mean that the limiting law of the renormalized energy functional ๐ธ๐‘Ÿ๐‘’๐‘› is independentof the way white noise is approximated, provided the natural symmetry condition in form ofAssumption 1.1 (iii), is satised. In view of its physical origin, ใ€ˆยทใ€‰ derives from the randomorientation of the grains. Such a model could be based on random tessellations, which suggestsa modelling through a non-Gaussian process.9 This motivates our interest in non-Gaussianapproximations of white noise. Our substitute for Gaussian calculus is the spectral gap inequality10(1.7), see Assumption 1.1 (iv).

5That is, for any test function ๐œ‘ โˆˆ Cโˆž (T2), denoting ๐œ‘ (๐‘ฅ) = ๐œ‘ (โˆ’๐‘ฅ1, ๐‘ฅ2), there holds b (๐œ‘) = b (๐œ‘) in law. We notethat for our results to hold one could also ask for invariance under reection in ๐‘ฅ2.

6Without loss of generality we have set the constant equal to one.7Incidentally, for cylindrical functionals ๐บ we also have the relation ๐œ•๐œ•b ๐บ (b)

2๐ฟ2

= sup๐›ฟb โˆˆ๐ฟ2 (T2)โ€–๐›ฟb โ€–

๐ฟ2 โ‰ค1

lim inf๐‘กโ†’0

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๐บ (b + ๐‘ก๐›ฟb) โˆ’๐บ (b)๐‘ก

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ2 .8Actually, in order to obtain the ๐ฟ๐‘ version of SGI in Proposition 5.1 we need it for more general ๐บ , but the main

application concerns the quadratic functional in Lemma 5.5.9Incidentally, random tessellations based on Poisson point processes are known to satisfy a variant of the spectral

gap inequality, see [DG20].10In a Gaussian setting the right-hand side of (1.7) would correspond to having ๐ฟ2 as the Cameronโ€“Martin space.

Page 5: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 5

Since we cannot expect almost-sure uniqueness of the absolute minimizer ๐‘ค due to non-convexity of ๐ธ๐‘Ÿ๐‘’๐‘› , this universality is better expressed on the level of the variational problems๐ธ๐‘Ÿ๐‘’๐‘› themselves. Hence, rather than considering the (ill-dened) random elds ๐‘ค of minimalcongurations, we consider the random functionals ๐ธ๐‘Ÿ๐‘’๐‘› . The latter notion calls for a topology onthe space of variational problems, that is, of lower semicontinuous functionals ๐ธ onW (takingvalues in R โˆช {+โˆž}) that have compact sublevel sets (with respect to the strong ๐ฟ2-topology).The appropriate topology is the one generated by ฮ“-convergence11; this is tautological sinceฮ“-convergence of functionals in this topology is essentially equivalent to convergence of theminimizers, which do exist provided the functionals have compact sublevel sets (with respectto the strong ๐ฟ2-topology). Hence we are lead to consider probability measures on the space oflower semicontinuous functionals ๐ธ on W endowed with the topology of ฮ“-convergence12. Fromthis point of view, the universality of the ripple takes the following form:

Theorem 1.4. Every probability measure ใ€ˆยทใ€‰ on the space periodic Schwartz distributions b satisfyingDenition 1.1 extends to a probability measure ใ€ˆยทใ€‰ext on the product space of periodic Schwartz

distributions b in the Hรถlder space Cโˆ’ 54โˆ’ and lower semicontinuous functionals ๐ธ on W endowed

with the topology of ฮ“-convergence with the following three properties:

(i) If b is smooth ใ€ˆยทใ€‰-almost surely, then ๐ธ = ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ ; ยท) for ใ€ˆยทใ€‰ext-almost every (b, ๐ธ), where๐‘ฃ = Lโˆ’1๐‘ƒb and ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ ; ยท) is given by (1.4).

(ii) If a sequence {ใ€ˆยทใ€‰โ„“ }โ„“โ†“0 of probability measures that satisfy Assumption 1.1 converges weakly

to ใ€ˆยทใ€‰ (which automatically satises Assumption 1.1), then {ใ€ˆยทใ€‰extโ„“ }โ„“โ†“0 converges weakly to

ใ€ˆยทใ€‰ext.

Remark 1.5. Let us explain why Theorem 1.4 expresses the desired universality of the ripple. Weare given a sequence {ใ€ˆยทใ€‰โ„“ }โ„“โ†“0 which converges weakly to white noise ใ€ˆยทใ€‰, such that ใ€ˆยทใ€‰โ„“ satisesAssumption 1.1, and such that for โ„“ > 0, b is smooth ใ€ˆยทใ€‰โ„“ -almost surely. In view of Theorem 1.4 (i),as long as โ„“ > 0, the pathwise dened ๐ธ๐‘Ÿ๐‘’๐‘› , see (1.4), can be identied with the random functional๐ธ associated to ใ€ˆยทใ€‰extโ„“ . According to Theorem 1.4 (ii), as โ„“ โ†“ 0, the law of (b, ๐ธ๐‘Ÿ๐‘’๐‘›) under ใ€ˆยทใ€‰extโ„“

converges weakly to ใ€ˆยทใ€‰ext associated to the law of white noise ใ€ˆยทใ€‰.

As a corollary of our results we have the following stronger statement.

Corollary 1.6. Assume that the probability measure ใ€ˆยทใ€‰ satises Assumption 1.1 and consider its

extension ใ€ˆยทใ€‰ext to the product space of periodic Schwartz distributions b in the Hรถlder space Cโˆ’ 54โˆ’

and lower semicontinuous functionals ๐ธ onW endowed with the topology of ฮ“-convergence. Thenminimizers of ๐ธ exist in W for ใ€ˆยทใ€‰ext-almost every (b, ๐ธ). Moreover, for every 1 โ‰ค ๐‘ < โˆž, the

following estimate holds,13 โŸจ

inf๐‘คโˆˆargmin๐ธ

[๐‘ค]๐‘54โˆ’

โŸฉextโ‰ค ๐ถ, (1.8)

for a constant ๐ถ that only depends on ๐‘ , uniformly in the class of probability measures ใ€ˆยทใ€‰ satisfyingAssumption 1.1.

Remark 1.7. Since under ใ€ˆยทใ€‰ext functionals ๐ธ are non-convex (see the discussion below Propo-sition 1.10) we do not expect uniqueness of minimizers. In that sense, Corollary 1.6 shows theexistence of minimizers๐‘ค โˆˆ W with nite C 5

4โˆ’-norm. However, we do not know if all minimizershave this regularity.14 The C 5

4โˆ’-regularity in (1.8) relies on the Eulerโ€“Lagrange equation (1.13)(see the discussion below Proposition 1.10). If minimizers were unique, the uniformity of (1.8) inthe class of probability measures satisfying Assumption 1.1 and tightness would imply that if a

11Based on the strong ๐ฟ2-topology.12Recall that the space of lower semicontinuous functionals ๐ธ : W โ†’ R โˆช {+โˆž} is a compact space, see [Dal93,

Theorem 8.5 and Theorem 10.6].13Note that the inf in (1.8) is +โˆž if argmin๐ธ = โˆ….14We have no reason to assume that this is not the case.

Page 6: Variational methods for a singular SPDE yielding the

6 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

sequence {ใ€ˆยทใ€‰โ„“ }โ„“โ†“0 satisfying Assumption 1.1 converges weakly to ใ€ˆยทใ€‰, then the law of the uniqueminimizer under ใ€ˆยทใ€‰extโ„“ converges weakly to the law of the unique minimizer under ใ€ˆยทใ€‰ext.

In establishing Theorem 1.4, we follow very much the spirit of rough path theory of a clear sep-aration between a stochastic and a deterministic (pathwise) ingredient. The genuinely stochasticingredient is formulated in Proposition 1.8, where we extend the probability distribution of b โ€™sto a (joint) law of (b, ๐‘ฃ, ๐น ). Compared to rough paths, ๐‘ฃ is the analogue of (multi-dimensional)Brownian motion ๐ต, and ๐น similar to the iterated โ€œintegrandโ€15 ๐ต ๐‘‘๐ต

๐‘‘๐‘ก, see Proposition 1.8 (iii),

which relates ๐น to ๐‘ฃ๐‘…1๐œ•2๐‘ฃ . The degree of indeterminacy reected by the dierence betweenStratonovich (midpoint rule) and Itรด (explicit) is suppressed by the symmetry condition in As-sumption 1.1 (iii), which feeds into the characterizing property given by (1.9). The crucial stabilityof this construction is provided by Proposition 1.8 (iv).

Like for rough paths this unfolding into several random building blocks allows for a pathwisesolution theory, i.e. the construction of a continuous solution map on this augmented space. Inour variational case this turns into a continuous map from the space of (b, ๐‘ฃ, ๐น )โ€™s into the space offunctionals, with the above-advertised topology of ฮ“-convergence, see Proposition 1.10).

Proposition 1.8. Every probability measure ใ€ˆยทใ€‰ satisfying Assumption 1.1 is concentrated on the

Hรถlder space Cโˆ’ 54โˆ’ and lifts to a probability measure ใ€ˆยทใ€‰li on the space of triples (b, ๐‘ฃ, ๐น ) in Cโˆ’ 5

4โˆ’ ร—C 3

4โˆ’ ร— Cโˆ’ 34โˆ’ with the following properties:

(i) The law of b under ใ€ˆยทใ€‰li is ใ€ˆยทใ€‰.(ii) ๐‘ฃ = Lโˆ’1๐‘ƒb ใ€ˆยทใ€‰li-almost surely.

(iii) The law of ๐น under ใ€ˆยทใ€‰li is characterized by

lim๐‘ก=2โˆ’๐‘›โ†“0

โŸจ[๐น โˆ’ ๐‘ฃ๐‘…1๐œ•2๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’

โŸฉli= 0, (1.9)

for every 1 โ‰ค ๐‘ < โˆž. Moreover, if b is smooth ใ€ˆยทใ€‰-almost surely, we have that ๐น = ๐‘ฃ๐‘…1๐œ•2๐‘ฃใ€ˆยทใ€‰li-almost surely.

(iv) Finally, if a sequence {ใ€ˆยทใ€‰โ„“ }โ„“โ†“0 of probability measures that satisfy Assumption 1.1 converges

weakly to a probability measure ใ€ˆยทใ€‰, then also {ใ€ˆยทใ€‰liโ„“ }โ„“โ†“0 converges weakly to ใ€ˆยทใ€‰li.

Remark 1.9. Let us point out that (1.9) implies that ๐น is actually a ใ€ˆยทใ€‰-measurable function of b .Indeed, by (i), (ii), and the triangle inequality for ๐‘ , ๐‘ก โˆˆ (0, 1] dyadic we haveโŸจ

[๐‘ฃ๐‘…1๐œ•2๐‘ฃ๐‘  โˆ’ ๐‘ฃ๐‘…1๐œ•2๐‘ฃ๐‘ก ]๐‘โˆ’ 34โˆ’

โŸฉ 1๐‘

=

(โŸจ[๐‘ฃ๐‘…1๐œ•2๐‘ฃ๐‘  โˆ’ ๐‘ฃ๐‘…1๐œ•2๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’

โŸฉli) 1๐‘

โ‰ค(โŸจ[๐น โˆ’ ๐‘ฃ๐‘…1๐œ•2๐‘ฃ๐‘ ]๐‘โˆ’ 3

4โˆ’

โŸฉli) 1๐‘

+(โŸจ[๐น โˆ’ ๐‘ฃ๐‘…1๐œ•2๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’

โŸฉli) 1๐‘

,

which in turn implies that the sequence {๐‘ฃ๐‘…1๐œ•2๐‘ฃ๐‘ก }๐‘กโ†“0 is Cauchy in ๐ฟ๐‘ใ€ˆยทใ€‰Cโˆ’ 3

4โˆ’. Hence it convergesto a random variable ๐น (b) โˆˆ ๐ฟ๐‘ใ€ˆยทใ€‰C

โˆ’ 34โˆ’ and it is easy to check that ๐น = ๐น (b) ใ€ˆยทใ€‰li-almost surely.

This allows us to identify the lift measure ใ€ˆยทใ€‰li as the joint law of (b,Lโˆ’1๐‘ƒb, ๐น (b)) under ใ€ˆยทใ€‰.

The main idea of the deterministic ingredient, Proposition 1.10, is to extend the denition (1.4)of ๐ธ๐‘Ÿ๐‘’๐‘› from only depending on (b, ๐‘ฃ)16 to depending on (b, ๐‘ฃ, ๐น ), in such a way that the denitions

15As opposed to its integralโˆซ๐ต d๐ต, which is called the iterated integral. We refer to [FH14, Chapter 3] for details.

16That is, eectively only on b . Since ๐ธ๐‘Ÿ๐‘’๐‘› no longer depends explicitly on b , we drop b in the notation for ๐ธ๐‘Ÿ๐‘’๐‘› .

Page 7: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 7

(formally) coincide for ๐น = ๐‘ฃ๐‘…1๐œ•2๐‘ฃ . This is achieved by17

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) := E(๐‘ค) + G(๐‘ฃ, ๐น ;๐‘ค), (1.10)where the anharmonic energy E is given by the rst two contributions of ๐ธ๐‘ก๐‘œ๐‘ก ,

E(๐‘ค) :=โˆซT2(๐œ•1๐‘ค)2 d๐‘ฅ +

โˆซT2

(|๐œ•1 |โˆ’

12 (๐œ•2๐‘ค โˆ’ ๐œ•1 12๐‘ค

2))2

d๐‘ฅ . (1.11)

Note that E contains the Burgers nonlinearity [๐‘ค := ๐œ•2๐‘ค โˆ’ ๐œ•1 12๐‘ค2, which will play an important

role in our analysis. Only the remainder G depends on ๐‘ฃ and ๐น , and is given by

G(๐‘ฃ, ๐น ;๐‘ค) :=โˆซT2

(๐‘ค2๐‘…1๐œ•2๐‘ฃ + ๐‘ฃ2๐‘…1[๐‘ค + 2๐‘ฃ๐‘ค๐‘…1[๐‘ค + 2๐‘ค๐น โˆ’๐‘ค๐‘ฃ๐‘…1๐œ•1๐‘ฃ2 + (๐‘…1 |๐œ•1 |

12 (๐‘ฃ๐‘ค))2

)d๐‘ฅ .(1.12)

Equipped with these denitions, we now may state the main deterministic ingredient.

Proposition 1.10. The application (b, ๐‘ฃ, ๐น ) โ†ฆโ†’ ๐ธ๐‘Ÿ๐‘’๐‘› described through (1.10) is well-dened and

continuous when the space of (b, ๐‘ฃ, ๐น ) is endowed with the norm Cโˆ’ 54โˆ’ร—C 3

4โˆ’ร—Cโˆ’ 34โˆ’ and the space of

lower semicontinuous functionals ๐ธ๐‘Ÿ๐‘’๐‘› on W is endowed with the topology of ฮ“-convergence (basedon the ๐ฟ2-topology).

ใ€ˆยทใ€‰ โˆผ b

ใ€ˆยทใ€‰li โˆผ (b, ๐‘ฃ, ๐น ) ใ€ˆยทใ€‰ext โˆผ (b, ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท))

Figure 1. Construction of the extension measure ใ€ˆยทใ€‰ext. The vertical arrowcorresponds to the probabilistic step, Proposition 1.8, while the horizontal arrowis the deterministic step, Proposition 1.10. For smooth b โ€™s ๐น is given by ๐‘ฃ๐‘…1๐œ•2๐‘ฃ .

On the level of the Eulerโ€“Lagrange equation, minimizers๐‘ค of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) are weak solutionsof

L๐‘ค + ๐‘ƒ(๐น +๐‘ค๐‘…1๐œ•2๐‘ฃ + ๐‘ฃ๐‘…1๐œ•2๐‘ค +๐‘ค๐‘…1๐œ•2๐‘ค โˆ’ 1

2 (๐‘ฃ +๐‘ค)๐‘…1๐œ•1(๐‘ฃ +๐‘ค)2)

+ 12๐‘…1๐œ•2(๐‘ฃ +๐‘ค)2 = 0,

(1.13)

whose existence is established by Theorem 1.4 (see also Theorem 1.14 (iv) for the validity of (1.13)in the sense of Schwartz distributions). By a simple power counting the expected regularity ofsolutions ๐‘ค to (1.13) is 5

4โˆ’, which justies the existence of minimizers ๐‘ค โˆˆ W with nite C 54โˆ’

norm proved in Corollary 1.6. This generalizes the existence of solutions to (1.13) in [IO19] whichwas shown for small values of the noise strength |๐œŽ |.

From a variational point of view, the main challenge is to establish the coercivity of therenormalized energy ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท). Ideally, one would like to control the remainder G(๐‘ฃ, ๐น ; ยท) bythe โ€œgoodโ€ term of the renormalized energy, namely, the anharmonic energy E(๐‘ค) given in (1.11).At rst sight, this is not obvious since the remainder G(๐‘ฃ, ๐น ; ยท) contains quadratic and cubic terms

17This can be seen by the identity [๐‘ข = [๐‘ฃ + [๐‘ค โˆ’ ๐œ•1 (๐‘ฃ๐‘ค) for the Burgers operator [๐‘ข = ๐œ•2๐‘ข โˆ’ ๐œ•1 12๐‘ข2, as well as the

equality โˆซT2

(๐œ•1๐‘ค ๐œ•1๐‘ฃ + ๐œ•2๐‘ค |๐œ•1 |โˆ’1๐œ•2๐‘ฃ โˆ’๐‘ค b

)d๐‘ฅ = 0,

which follows from testing (1.3) with๐‘ค .

Page 8: Variational methods for a singular SPDE yielding the

8 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

in๐‘ค , and it is not immediate that the anharmonic energy E(๐‘ค) provides higher than quadraticcontrol to absorb these terms. Hence, we need to exploit the control on the nonlinear part comingfrom the Burgers operator [๐‘ค .18 We do this using tools from uid mechanics, more precisely, theHowarthโ€“Kรกrmรกnโ€“Monin identities (2.5) and (2.6), following [GJO15]. Based on these identitieswe can prove that the anharmonic energy E(๐‘ค) grows cubically in suitable Besov spaces (seeProposition 2.4). This allows us to absorb G(๐‘ฃ, ๐น ; ยท) and obtain the coercivity of the renormalizedenergy (see Theorem 1.14 (i)).

Let us point out that here, we prove existence of solutions for any value of the noise strength๐œŽ using the coercivity of the renormalized energy functional through the direct method of thecalculus of variations. Recent works on the dynamic ฮฆ4 model (or stochastic Ginzburgโ€“Landaumodel), where the system is favoured by the โ€œgoodโ€ sign of the cubic nonlinearity, have usedcoercivity on the level of the Eulerโ€“Lagrange equation. For example, in [MW17a, TW18, MW17b]energy estimates have been used to obtain global-in-time existence in the parabolic case, while in[GH19] both the parabolic and the elliptic cases have been treated based on a dierent approachthat uses coercivity through a maximum principle. A maximum principle has been used also in[CMW19] where the parabolic model is considered in the full subcritical regime.

A further challenge, which turns out to be more on the technical side, comes from the factthat L is nonlocal. We recall that this feature arises completely naturally from the magnetostaticenergy in the thin-lm limit (see [IO19, Section 2]), but resonates well with the recent surge inactivity on nonlocal operators. It was worked out in [IO19, Lemma 5] that the robust approach of[OW19] to negative (parabolic) Hรถlder spaces and Schauder theory extends to this situation. Thisapproach involves a suitable convolution semigroup๐œ“๐‘ก ; the fact that it extends from the smoothparabolic symbol ๐‘˜21 + ๐‘–๐‘˜2 to our nonsmooth symbol ๐‘˜21 + |๐‘˜1 |โˆ’1๐‘˜22 is not obvious due to the poordecay properties of the corresponding convolution kernel.

Variational problems that in a singular limit require subtraction of a divergent term are well-known in deterministic settings. A famous example concerns S1-valued harmonic maps denedin a two-dimensional smooth bounded simply-connected domain ๐ท . The aim there is to minimizethe Dirichlet energy of maps ๐‘ข : ๐ท โ†’ S1 that satisfy a smooth boundary condition ๐‘” : ๐œ•๐ท โ†’ S1.When ๐‘” carries a nontrivial winding number ๐‘ > 019, the problem is singular, that is, everyconguration ๐‘ข has innite energy as they generate vortex point singularities. The question isto determine the least โ€œinniteโ€ Dirichlet energy of a harmonic S1-valued map satisfying theboundary condition ๐‘” on ๐œ•๐ท . The seminal book of Bethuelโ€“Brezisโ€“Hรฉlein [BBH94] presentstwo methods to achieve this goal, both reaching the same renormalized energy associated to theproblem.

First approach: One prescribes ๐‘ > 0 vortex points ๐‘Ž1, . . . , ๐‘Ž๐‘ in ๐ท and determines theunique harmonic S1-valued map ๐‘ขโˆ— with ๐‘ขโˆ— = ๐‘” on ๐œ•๐ท that has the prescribed singularities๐‘Ž1, . . . , ๐‘Ž๐‘ in ๐ท , each one carrying a winding number equal to one20. Then one cuts-odisks ๐ต(๐‘Ž๐‘˜ , ๐‘Ÿ ) centered at ๐‘Ž๐‘˜ of small radius ๐‘Ÿ > 0 carrying the diverging logarithmicenergy of ๐‘ขโˆ— and introduces the renormalized energy

๐‘Š (๐‘Ž1, . . . , ๐‘Ž๐‘ ) = lim๐‘Ÿโ†’0

(โˆซ๐ท\โˆช๐‘˜๐ต (๐‘Ž๐‘˜ ,๐‘Ÿ )

|โˆ‡๐‘ขโˆ—(๐‘ฅ) |2 d๐‘ฅ โˆ’ 2๐œ‹๐‘ log 1๐‘Ÿ

).

18Incidentally, despite dierent physical origins, the inviscid Burgers part [๐‘ค arises as in the KPZ equation fromexpanding a square root nonlinearity. Not unlike there, the coercivity comes from the interaction between the rstand second term in E(๐‘ค), the rst term being the analogue to the viscocity in KPZ.

19For simplicity, we assume ๐‘ > 0; the case ๐‘ < 0 follows by complex conjugation.20In fact, ๐‘ขโˆ— belongs to the Sobolev space๐‘Š 1,1 (๐ท, S1) and the nonlinear PDE satised by ๐‘ขโˆ—, i.e., โˆ’ฮ”๐‘ขโˆ— = |โˆ‡๐‘ขโˆ— |2๐‘ขโˆ—

in ๐ท , can be written in a โ€œlinearโ€ way in terms of the current ๐‘— (๐‘ขโˆ—) = ๐‘ขโˆ— ร— โˆ‡๐‘ขโˆ— โˆˆ ๐ฟ1 (๐ท) of ๐‘ขโˆ— that satises the systemโˆ‡ ร— ๐‘— (๐‘ขโˆ—) = 2๐œ‹

โˆ‘๐‘˜ ๐›ฟ๐‘Ž๐‘˜ in ๐ท , and โˆ‡ ยท ๐‘— (๐‘ขโˆ—) = 0 in ๐ท . In terms of the so-called conjugate harmonic function ๐œ™ given by

โˆ‡โŠฅ๐œ™ = ๐‘— (๐‘ขโˆ—), the problem becomes โˆ’ฮ”๐œ™ = 2๐œ‹โˆ‘๐‘˜ ๐›ฟ๐‘Ž๐‘˜ in ๐ท and ๐œ•a๐œ™ = ๐‘” ร— ๐œ•๐œ๐‘” on ๐œ•๐ท . One could think of ๐œ™ as playing

the role of our solution ๐‘ฃ to the linearized Euler-Lagrange equation (1.3) that carries the โ€œinniteโ€ part of the energy.

Page 9: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 9

The minimum of the renormalized energymin

๐‘Ž1,...๐‘Ž๐‘ โˆˆ๐ท๐‘Š (๐‘Ž1, . . . , ๐‘Ž๐‘ ) (1.14)

represents the minimal second order term in the expansion of the Dirichlet energy andyields optimal positions of the ๐‘ vortex point singularities (which might not be unique ingeneral).Second approach: One considers a nonlinear approximation of the harmonicmap problemgiven by the Ginzburg-Landau model for a small parameter Y > 0:

๐ธY (๐‘ข) =โˆซ๐ท

|โˆ‡๐‘ข |2 + 1Y2(1 โˆ’ |๐‘ข |2)2 d๐‘ฅ, ๐‘ข : ๐ท โ†’ R2, ๐‘ข = ๐‘” on ๐œ•๐ท.

Note that the maps ๐‘ข are no longer with values into S1, but their distance to S1 is stronglypenalized as Y โ†’ 0. It is proved in [BBH94, Theorem X.1] that if ๐‘ขY is a minimizer of theabove Ginzburgโ€“Landau problem, then for a subsequence, ๐‘ขY โ‡€ ๐‘ขโˆ— weakly in๐‘Š 1,1(๐ท) asY โ†’ 0 where ๐‘ขโˆ— is an S1 valued harmonic map whose ๐‘ vortex points of winding numberone correspond to a minimizer of the renormalized energy (1.14). Moreover,

๐ธY (๐‘ขY) = 2๐œ‹๐‘ log 1Y+ min

๐‘Ž1,...๐‘Ž๐‘ โˆˆ๐ท๐‘Š (๐‘Ž1, . . . , ๐‘Ž๐‘ ) + ๐‘๐›พ + ๐‘œ (1), as Y โ†’ 0,

where ๐›พ is a constant coming from the nonlinear penalization in ๐ธY .We also refer to [SS07, SS15], [Kur06], [IM16], and [IJ19] for similar renormalized energies.

1.1. Notation. For a periodic function ๐‘“ : T2 โ†’ R we dene its Fourier coecients by

๐‘“ (๐‘˜) =โˆซT2๐‘’โˆ’i๐‘˜ ยท๐‘ฅ ๐‘“ (๐‘ฅ) d๐‘ฅ for ๐‘˜ โˆˆ (2๐œ‹Z)2,

which extends to periodic Schwartz distributions in the natural way. We also denote by ๐‘ƒ the๐ฟ2-orthogonal projection onto the set of functions of vanishing average in ๐‘ฅ1, extended in thenatural way to periodic Schwartz distributions.

For ๐‘ โˆˆ [1,โˆž] we write โ€– ยท โ€–๐ฟ๐‘ to denote the usual ๐ฟ๐‘ norm on T2, unless indicated otherwise.For example, we write โ€– ยท โ€–๐ฟ๐‘ (R2) for the ๐ฟ๐‘ norm of a function dened on R2. We sometimeswrite ๐ฟ๐‘๐‘ฅ (respectively ๐ฟ๐‘๐‘ฅ ๐‘—

, ๐‘— = 1, 2) to denote the ๐ฟ๐‘ space with respect to the ๐‘ฅ (respectively ๐‘ฅ ๐‘— )variable. We also write ๐ฟ๐‘ใ€ˆยทใ€‰ to denote the usual ๐ฟ๐‘ space with respect to the measure ใ€ˆยทใ€‰.

We will often make use of the notation ๐‘Ž . ๐‘ meaning that there exists a constant ๐ถ > 0 suchthat ๐‘Ž โ‰ค ๐ถ๐‘. Moreover, for ^ โˆˆ R, the notation .^ will be used to stress the dependence of theimplicit constant ๐ถ on ^, i.e., ๐ถ โ‰ก ๐ถ (^). Similarly, ๐‘Ž โˆผ ๐‘ means ๐‘Ž . ๐‘ and ๐‘ . ๐‘Ž.

1.1.1. Hilbert transform. We will frequently make use of the Hilbert transform ๐‘… ๐‘— for ๐‘— = 1, 2, actingon periodic functions ๐‘“ : T2 โ†’ R in ๐‘ฅ ๐‘— as

๐‘… ๐‘— :=๐œ•๐‘—

|๐œ•๐‘— |, i.e., ๐‘… ๐‘— ๐‘“ (๐‘˜) =

{i sgn(๐‘˜ ๐‘— ) ๐‘“ (๐‘˜) if ๐‘˜ ๐‘— โˆˆ 2๐œ‹Z \ {0},0 if ๐‘˜ ๐‘— = 0,

(1.15)

where sgn is the sign function. In particular, ๐‘… ๐‘—๐‘ƒ = ๐‘ƒ๐‘… ๐‘— = ๐‘… ๐‘— .

1.1.2. Anisotropic metric and kernel. The leading-order operator L = โˆ’๐œ•21 โˆ’ |๐œ•1 |โˆ’1๐œ•22 suggests toendow the space T2 with a Carnotโ€“Carathรฉodory metric that is homogeneous with respect to thescaling (๐‘ฅ1, ๐‘ฅ2) = (โ„“๐‘ฅ1, โ„“

32๐‘ฅ2). The simplest expression is given by

๐‘‘ (๐‘ฅ,๐‘ฆ) := |๐‘ฅ1 โˆ’ ๐‘ฆ1 | + |๐‘ฅ2 โˆ’ ๐‘ฆ2 |23 , ๐‘ฅ,๐‘ฆ โˆˆ T2,

which in particular means that we take the ๐‘ฅ1 variable as a reference.We now introduce the convolution semigroup used in [IO19]. This is the โ€œheat kernelโ€ {๐œ“๐‘‡ }๐‘‡>0

of the operatorA := |๐œ•1 |3 โˆ’ ๐œ•22 = |๐œ•1 |L,

Page 10: Variational methods for a singular SPDE yielding the

10 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

which, in Fourier space R2, is given by

๐œ“๐‘‡ (๐‘˜) = exp(โˆ’๐‘‡ ( |๐‘˜1 |3 + ๐‘˜22)), for all ๐‘˜ โˆˆ R2. (1.16)

It is easy to check that the kernel has scaling properties in line with the metric ๐‘‘ , that is,

๐œ“๐‘‡ (๐‘ฅ1, ๐‘ฅ2) =1

(๐‘‡ 13 )1+ 3

2๐œ“

(๐‘ฅ1

๐‘‡13,๐‘ฅ2

(๐‘‡ 13 ) 3

2

), for all ๐‘ฅ โˆˆ R2, (1.17)

where for simplicity we write๐œ“ := ๐œ“1. Note that๐œ“ is a symmetric smooth function with integrablederivatives and we have for every ๐‘ โˆˆ [1,โˆž] (see [IO19, Proof of Lemma 10]),

โ€–๐ท๐›ผ1๐ท

๐›ฝ

2๐œ“๐‘‡ โ€–๐ฟ๐‘ (R2) โˆผ (๐‘‡ 13 )โˆ’๐›ผโˆ’

32 ๐›ฝโˆ’

52 (1โˆ’

1๐‘), (1.18)

for every ๐‘‡ > 0, ๐ท ๐‘— โˆˆ {๐œ•๐‘— , |๐œ•๐‘— |}, ๐‘— = 1, 2, and ๐›ผ, ๐›ฝ โ‰ฅ 0. For a periodic Schwartz distribution ๐‘“ , wedenote by ๐‘“๐‘‡ its convolution with๐œ“๐‘‡ , i.e., ๐‘“๐‘‡ = ๐œ“๐‘‡ โˆ— ๐‘“ , which yields a smooth periodic function.Notice that {๐œ“๐‘‡ }๐‘‡>0 is a convolution semigroup, so that

(๐‘“๐‘ก )๐‘‡ = ๐‘“๐‘ก+๐‘‡ for all ๐‘ก,๐‘‡ > 0.

Remark 1.11. By the space of periodic Schwartz distributions ๐‘“ we understand the (topologi-cal) dual of the space of Cโˆž-functions ๐œ‘ on the torus (endowed with the family of seminorms{โ€–๐œ• ๐‘—1๐œ•๐‘™2๐œ‘ โ€–๐ฟโˆž} ๐‘—,๐‘™โ‰ฅ0).

For a Cโˆž-function ๐œ“ on R2 with integrable derivatives, i.e.โˆซR2

|๐œ• ๐‘—1๐œ•๐‘™2๐œ“ | d๐‘ฅ < โˆž for all ๐‘—, ๐‘™ โ‰ฅ0, and a periodic Schwartz distribution ๐‘“ we write (๐‘“ โˆ— ๐œ“ ) (๐‘ฅ) to denote ๐‘“ (ฮจ(๐‘ฅ โˆ’ ยท)), whereฮจ :=

โˆ‘๐‘งโˆˆZ2 ๐œ“ (ยท โˆ’ ๐‘ง) is the periodization of ๐œ“ , which is well-dened and belongs to Cโˆž(T2). In

particular, if ฮจ๐‘‡ denotes the periodization of our โ€œheat kernelโ€๐œ“๐‘‡ , then ฮจ๐‘‡ is a smooth semigroupwhose Fourier coecients are given by๐œ“๐‘‡ (๐‘˜) in (1.16) for ๐‘˜ โˆˆ (2๐œ‹Z)2, yielding for any ๐‘‡ โˆˆ (0, 1],๐ท ๐‘— โˆˆ {๐œ•๐‘— , |๐œ•๐‘— |}, ๐‘— = 1, 2, and ๐›ผ, ๐›ฝ โ‰ฅ 0: 21

โ€–๐ท๐›ผ1๐ท

๐›ฝ

2ฮจ๐‘‡ โ€–๐ฟ๐‘ . โ€–๐ท๐›ผ1๐ท

๐›ฝ

2๐œ“๐‘‡ โ€–๐ฟ๐‘ (R2)(1.18). (๐‘‡ 1

3 )โˆ’๐›ผโˆ’32 ๐›ฝโˆ’

52 (1โˆ’

1๐‘). (1.19)

Therefore, for a periodic function ๐‘“ โˆˆ ๐ฟ๐‘ž , we will often use Youngโ€™s inequality for convolutionwith 1 + 1

๐‘Ÿ= 1

๐‘+ 1

๐‘žin the form

โ€– ๐‘“ โˆ— ๐ท๐›ผ1๐ท

๐›ฝ

2๐œ“๐‘‡ โ€–๐ฟ๐‘Ÿ โ‰ค โ€– ๐‘“ โ€–๐ฟ๐‘ž โ€–๐ท๐›ผ1๐ท

๐›ฝ

2ฮจ๐‘‡ โ€–๐ฟ๐‘ . (๐‘‡ 13 )โˆ’๐›ผโˆ’

32 ๐›ฝโˆ’

52 (1โˆ’

1๐‘) โ€– ๐‘“ โ€–๐ฟ๐‘ž .

We sometimes write ฮ“ for the integral kernel of Lโˆ’1๐‘ƒ , given by

ฮ“ฬ‚(๐‘˜) = 1๐‘˜21 + |๐‘˜1 |โˆ’1๐‘˜22

, for ๐‘˜1 โ‰  0. (1.20)

Note that

โ€–ฮ“โ€–2๐ฟ2 =

โˆ‘๏ธ๐‘˜1โ‰ 0

1(๐‘˜21 + |๐‘˜1 |โˆ’1๐‘˜22)2

โ‰คโˆ‘๏ธ๐‘˜1โ‰ 0

1๐‘‘ (0, ๐‘˜)4 < โˆž. (1.21)

21Indeed, for ๐‘ = 1, we have

โ€–๐ท๐›ผ1 ๐ท

๐›ฝ

2 ฮจ๐‘‡ โ€–๐ฟ1 โ‰ค โˆ‘๐‘งโˆˆZ2

โˆซT2 |๐ท

๐›ผ1 ๐ท

๐›ฝ

2๐œ“๐‘‡ (๐‘ฅ โˆ’ ๐‘ง) | d๐‘ฅ = โ€–๐ท๐›ผ1 ๐ท

๐›ฝ

2๐œ“๐‘‡ โ€–๐ฟ1 (R2) ,

while for ๐‘ = โˆž,

โ€–๐ท๐›ผ1 ๐ท

๐›ฝ

2 ฮจ๐‘‡ โ€–๐ฟโˆž โ‰ค 1 + โˆ‘๐‘˜โˆˆ(2๐œ‹Z)2\{(0,0) } |๐‘˜1 |๐›ผ |๐‘˜2 |๐›ฝ |๐œ“๐‘‡ (๐‘˜) | . 1 +

โˆซR2 |b1 |

๐›ผ |b2 |๐›ฝ exp(โˆ’๐‘‡ ( |b1 |3 + b22)) db

. (๐‘‡13 )โˆ’๐›ผโˆ’

32 ๐›ฝโˆ’

52(1.18). โ€–๐ท๐›ผ

1 ๐ท๐›ฝ

2๐œ“๐‘‡ โ€–๐ฟโˆž (R2) .

For ๐‘ โˆˆ (1,โˆž), one argues by interpolation.

Page 11: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 11

1.1.3. Denition of Hรถlder spaces. We now introduce the scale of Hรถlder seminorms based on thedistance function ๐‘‘ , where we restrict ourselves to the range ๐›ผ โˆˆ (0, 32 ) needed in this work (see[IO19, Denition 1]).

Denition 1.12. For a function ๐‘“ : T2 โ†’ R and ๐›ผ โˆˆ (0, 32 ), we dene

[๐‘“ ]๐›ผ :=sup๐‘ฅโ‰ ๐‘ฆ

|๐‘“ (๐‘ฆ)โˆ’๐‘“ (๐‘ฅ) |๐‘‘๐›ผ (๐‘ฆ,๐‘ฅ) for ๐›ผ โˆˆ (0, 1],

sup๐‘ฅโ‰ ๐‘ฆ|๐‘“ (๐‘ฆ)โˆ’๐‘“ (๐‘ฅ)โˆ’๐œ•1 ๐‘“ (๐‘ฅ) (๐‘ฆโˆ’๐‘ฅ)1 |

๐‘‘๐›ผ (๐‘ฆ,๐‘ฅ) for ๐›ผ โˆˆ (1, 32 ) .

We denote by C๐›ผ the closure of periodic Cโˆž-functions ๐‘“ : T2 โ†’ R with respect to [๐‘“ ]๐›ผ .

We will also need the following Hรถlder spaces of negative exponents. We will restrict to therange required in this work, namely ๐›ฝ โˆˆ (โˆ’ 3

2 , 0) (see [IO19, Denition 3]).

Denition 1.13. Let ๐‘“ be a periodic Schwartz distribution on T2. For ๐›ฝ โˆˆ (โˆ’1, 0) we dene[๐‘“ ]๐›ฝ := inf{|๐‘ | + [๐‘”]๐›ฝ+1 + [โ„Ž]๐›ฝ+ 3

2: ๐‘“ = ๐‘ + ๐œ•1๐‘” + ๐œ•2โ„Ž}

and for ๐›ฝ โˆˆ (โˆ’ 32 ,โˆ’1] we dene

[๐‘“ ]๐›ฝ := inf{|๐‘ | + [๐‘”]๐›ฝ+2 + [โ„Ž]๐›ฝ+ 32: ๐‘“ = ๐‘ + ๐œ•21๐‘” + ๐œ•2โ„Ž}.

We denote by C๐›ฝ the closure of periodic Cโˆž-functions ๐‘“ : T2 โ†’ R with respect to [๐‘“ ]๐›ฝ .

In Appendix A we provide all the necessary estimates on Hรถlder spaces needed in this work.

1.2. Strategy of the proofs. Recall the setW dened in (1.5), endowed with the strong topologyin ๐ฟ2(T2). We will show that the harmonic energyH(๐‘ค) dened in (1.6) controls the anharmonicpart E(๐‘ค) dened in (1.11) of the total energy, that is,

E(๐‘ค) . 1 + H (๐‘ค)2,for every๐‘ค โˆˆ W, and vice-versa, the anharmonic energy controls the harmonic part, that is, forevery ^ > 0 we have

H(๐‘ค) .^ 1 + E(๐‘ค) 32+^,

for any๐‘ค โˆˆ W, see Proposition 2.5 below. By standard embedding theorems (see Lemma B.5),any sublevel set ofH (respectively E) overW is relatively compact in ๐ฟ2 andH (respectively E)is lower semicontinuous with respect to the ๐ฟ2-norm (see (3.3)).

In the following, for Y > 0 suciently small, we will also write

T =

{(b, ๐‘ฃ, ๐น ) โˆˆ Cโˆ’ 5

4โˆ’Y ร— C 34โˆ’Y ร— Cโˆ’ 3

4โˆ’Y : L๐‘ฃ = ๐‘ƒb}.

Note that T is a closed subspace of Cโˆ’ 54โˆ’Y ร— C 3

4โˆ’Y ร— Cโˆ’ 34โˆ’Y endowed with the product metric.

The deterministic ingredient in the proof of Theorem 1.4, that is Proposition 1.10, is essentiallya consequence of the following theorem.

Theorem 1.14.(i) (Coercivity) For every _ โˆˆ (0, 1) and ๐‘€ > 0, there exists a constant ๐ถ > 0 which depends

on _ and polynomially on๐‘€ 22such that for every Y โˆˆ (0, 1

100 ) and every (b, ๐‘ฃ, ๐น ) โˆˆ T with

[b]โˆ’ 54โˆ’Y, [๐‘ฃ] 3

4โˆ’Y, [๐น ]โˆ’ 3

4โˆ’Yโ‰ค ๐‘€ , the functional G dened in (1.12) satises

|G(๐‘ฃ, ๐น ;๐‘ค) | โ‰ค _E(๐‘ค) +๐ถ, for every๐‘ค โˆˆ W .

In particular, ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) dened in (1.4) is coercive.(ii) (Continuity) Let Y โˆˆ (0, 1

100 ) and (bโ„“ , ๐‘ฃโ„“ , ๐นโ„“ ) โ†’ (b, ๐‘ฃ, ๐น ) in T and ๐‘คโ„“ โ†’ ๐‘ค in W with the

property that lim supโ„“โ†’0 E(๐‘คโ„“ ) < โˆž. Then

G(๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) โ†’ G(๐‘ฃ, ๐น ;๐‘ค) as โ„“ โ†’ 0.22We say that a constant๐ถ > 0 depends polynomially on๐‘€ if there exist ๐‘ > 0 and ๐‘ โ‰ฅ 1 such that๐ถ โ‰ค ๐‘ (1+๐‘€๐‘ ).

Page 12: Variational methods for a singular SPDE yielding the

12 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

(iii) (Compactness) Let Y โˆˆ (0, 1100 ) and (b, ๐‘ฃ, ๐น ) โˆˆ T be xed. Then for any๐‘€ โˆˆ R the sublevel

sets of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) dened in (1.10), given by{๐‘ค โˆˆ ๐ฟ2 :

โˆซ 1

0๐‘ค d๐‘ฅ1 = 0, ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) โ‰ค ๐‘€

},

are compact in the ๐ฟ2-norm.

(iv) (Existence of minimizers) If Y โˆˆ (0, 1100 ) and (b, ๐‘ฃ, ๐น ) โˆˆ T , then there exists a minimizer

๐‘ค โˆˆ W of the renormalized energy ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) which is a weak solution of (1.13).

Note that ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; 0) = 0, therefore every minimizer of ๐ธ๐‘Ÿ๐‘’๐‘› belongs to the sublevel set๐‘€ = 0of ๐ธ๐‘Ÿ๐‘’๐‘› . Using Theorem 1.14, we obtain the following ฮ“-convergence result.

Corollary 1.15 (ฮ“-convergence). Let Y โˆˆ (0, 1100 ) and (bโ„“ , ๐‘ฃโ„“ , ๐นโ„“ ) โ†’ (b, ๐‘ฃ, ๐น ) in T . Then

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘ค) โ†’ ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) for every ๐‘ค โˆˆ W as โ„“ โ†’ 0.Also, ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ; ยท) โ†’ ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) in the sense of ฮ“-convergence over W, that is,

(i) (ฮ“ โˆ’ lim inf) For all sequences {๐‘คโ„“ }โ„“โ†“0 โŠ‚ W with๐‘คโ„“ โ†’ ๐‘ค strongly in ๐ฟ2, we have

lim infโ„“โ†’0

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) โ‰ฅ ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) .

(ii) (ฮ“โˆ’ lim sup) For every๐‘ค โˆˆ W, there exists a sequence {๐‘คโ„“ }โ„“โ†“0 โŠ‚ W with๐‘คโ„“ โ†’ ๐‘ค strongly

in ๐ฟ2 such that

limโ„“โ†’0

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) = ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) .

Proof of Proposition 1.10. Corollary 1.15 establishes the continuity of the map that associates to each(b, ๐‘ฃ, ๐น ) โˆˆ T the functional ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท), when the space of (lower semicontinuous) functionals isequipped with the topology of ฮ“-convergence (based on the ๐ฟ2-topology). In particular, this mapis Borel measurable when T is endowed with its Borel ๐œŽ-algebra. ๏ฟฝ

Taking the main stochastic ingredient from Proposition 1.8 for granted (which we prove inSection 5), we can now give the proof of Theorem 1.4.

Proof of Theorem 1.4. Let ใ€ˆยทใ€‰ be a probability measure on the space of periodic Schwartz distribu-tions b that satises Assumption 1.1. By Proposition 1.8 ใ€ˆยทใ€‰ lifts to a probability measure ใ€ˆยทใ€‰li onthe space of triples (b, ๐‘ฃ, ๐น ) โˆˆ Cโˆ’ 5

4โˆ’ ร— C 34โˆ’ ร— Cโˆ’ 3

4โˆ’.By Proposition 1.10 the mapping (b, ๐‘ฃ, ๐น ) โ†ฆโ†’ ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) is continuous. Hence, the push-

forward ๐ธ๐‘Ÿ๐‘’๐‘›#ใ€ˆยทใ€‰li is well-dened as a probability measure on the space of lower semicontinuousfunctionals equipped with the Borel ๐œŽ-algebra corresponding to the topology of ฮ“-convergence(based on the strong ๐ฟ2-topology). We now dene ใ€ˆยทใ€‰ext as the joint law of b and ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท).23

(i) If b is smooth ใ€ˆยทใ€‰-almost surely, by Proposition 1.8 (iii) we have that ๐น = ๐‘ฃ๐œ•2๐‘…1๐‘ฃ ใ€ˆยทใ€‰-almostsurely. In this case, ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) = ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ ; ยท) ใ€ˆยทใ€‰-almost surely and agrees with thedenition given in (1.4).

(ii) Let {ใ€ˆยทใ€‰โ„“ }โ„“โ†“0 be a sequence of probability measures that satisfy Assumption 1.1 and con-verges weakly to ใ€ˆยทใ€‰, which then automatically satises Assumption 1.1. Then by Proposi-tion 1.8 (iv), the sequence {ใ€ˆยทใ€‰liโ„“ }โ„“โ†“0 converges weakly to ใ€ˆยทใ€‰li. Given a bounded continuousfunction ๐บ : (b, ๐ธ) โ†ฆโ†’ ๐บ (b, ๐ธ) โˆˆ R we have that

ใ€ˆ๐บใ€‰extโ„“ = ใ€ˆ๐บ (b, ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท))ใ€‰liโ„“โ„“โ†“0โˆ’โ†’ ใ€ˆ๐บ (b, ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท))ใ€‰li = ใ€ˆ๐บใ€‰ext,

which in turn implies that ใ€ˆยทใ€‰extโ„“ โ†’ ใ€ˆยทใ€‰ext weakly as โ„“ โ†“ 0. ๏ฟฝ

Finally, we have an a priori estimate for the C 54โˆ’ norm of minimizers of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท), which we

prove in Section 4.23Here we understand ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) as a measurable function of b , which is a composition of the measurable function

b โ†ฆโ†’ (b, ๐‘ฃ, ๐น ) and the continuous function (b, ๐‘ฃ, ๐น ) โ†ฆโ†’ ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท).

Page 13: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 13

Proposition 1.16 (Hรถlder regularity). For any๐‘€ > 0 and Y โˆˆ (0, 1100 ), there exists a constant๐ถ > 0

which depends on Y and polynomially on๐‘€ such that

[๐‘ค] 54โˆ’2Y

โ‰ค ๐ถ,

for every minimizer ๐‘ค โˆˆ W โˆฉ C 54โˆ’2๐œ– of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) with (b, ๐‘ฃ, ๐น ) โˆˆ T satisfying the bound

[b]โˆ’ 54โˆ’Y, [๐‘ฃ] 3

4โˆ’Y, [๐น ]โˆ’ 3

4โˆ’Yโ‰ค ๐‘€ .

Combined with an approximation argument, this is the main ingredient in the proof of Corol-lary 1.6, which we give now.

Proof of Corollary 1.6. We dene the functional ๐‘” : W โ†’ R โˆช {+โˆž} given by

๐‘”(๐‘ค) :={[๐‘ค] 5

4โˆ’2Y, if ๐‘ค โˆˆ C 5

4โˆ’2Y,

+โˆž, otherwise.

By [IO19, Lemma 13] we know that ๐‘” is lower semicontinuous onW endowed with the strongtopology in ๐ฟ2(T2). We dene the non-negative functional ๐บ : ๐ธ โ†ฆโ†’ ๐บ (๐ธ) on the space of lowersemicontinuous functionals ๐ธ onW by

๐บ (๐ธ) := inf๐‘คโˆˆargmin๐ธ

๐‘”(๐‘ค), if argmin๐ธ โ‰  โˆ…,

+โˆž, otherwise.

We claim that๐บ is lower semicontinuous, that is, if ๐ธโ„“ โ†’ ๐ธ as โ„“ โ†“ 0 in the sense of ฮ“-convergence,then

๐บ (๐ธ) โ‰ค lim infโ„“โ†“0

๐บ (๐ธโ„“ ) .

Indeed, without loss of generality we may assume that๐บ (๐ธโ„“ ) โ†’ lim inf โ„“โ†“0๐บ (๐ธโ„“ ) < โˆž by possiblyextracting a subsequence. This implies that supโ„“โˆˆ(0,1] ๐บ (๐ธโ„“ ) < โˆž, hence by the denition of ๐บthere exists a sequence of minimizers๐‘คโ„“ of ๐ธโ„“ such that

[๐‘คโ„“ ] 54โˆ’2Y

โ‰ค ๐บ (๐ธโ„“ ) + โ„“ โ‰ค supโ„“โˆˆ(0,1]

๐บ (๐ธโ„“ ) + 1.

By Lemma A.6 there exists๐‘ค โˆˆ C 54โˆ’2Y such that๐‘คโ„“ โ†’ ๐‘ค in C 5

4โˆ’3Y along a subsequence, and[๐‘ค] 5

4โˆ’2Yโ‰ค lim inf

โ„“โ†“0[๐‘คโ„“ ] 5

4โˆ’2Y.

This, in particular, implies that ๐‘คโ„“ โ†’ ๐‘ค strongly in ๐ฟ2(T2) and since ๐ธโ„“ โ†’ ๐ธ in the sense ofฮ“-convergence,๐‘ค is a minimizer of ๐ธ. Thus, we have the estimate

๐บ (๐ธ) โ‰ค [๐‘ค] 54โˆ’2Y

โ‰ค lim infโ„“โ†“0

[๐‘คโ„“ ] 54โˆ’2Y

โ‰ค lim infโ„“โ†“0

๐บ (๐ธโ„“ ),

which proves the desired claim.Let now {ใ€ˆยทใ€‰โ„“ }โ„“โ†“0 be a sequence of probability measures such that ใ€ˆยทใ€‰โ„“ โ†’ ใ€ˆยทใ€‰ weakly and

for every โ„“ โˆˆ (0, 1], b is smooth ใ€ˆยทใ€‰โ„“-almost surely. Since under ใ€ˆยทใ€‰โ„“ , b is smooth, by LemmaD.3 ๐‘ฃ is smooth. By Theorem 1.14 (iv) there exists a minimizer ๐‘ค โˆˆ W of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท), whichis a weak solution to (1.13). If we let ๐‘ข = ๐‘ฃ +๐‘ค , then ๐‘ข โˆˆ W and since ๐น = ๐‘ฃ๐‘…1๐œ•1๐‘ฃ ใ€ˆยทใ€‰โ„“-almostsurely (see Proposition 1.8 (iii)), ๐‘ข is a weak solution to (1.2). By Proposition E.2 we know thatโ€–|๐œ•1 |๐‘ ๐‘ขโ€–2 + โ€–|๐œ•2 |

23๐‘ ๐‘ขโ€–2 . 1 for every ๐‘  < 3, hence by Lemma B.8 ๐‘ข โˆˆ C 5

4โˆ’2Y , which in turn impliesthat๐‘ค = ๐‘ข โˆ’ ๐‘ฃ โˆˆ C 5

4โˆ’2Y . By Proposition 1.16 we have the estimate[๐‘ค] 5

4โˆ’2Yโ‰ค ๐ถ,

where the constant ๐ถ depends polynomially on max{[b]โˆ’ 54โˆ’Y, [๐‘ฃ] 3

4โˆ’Y, [๐น ]โˆ’ 3

4โˆ’Y}. In particular, this

implies that๐บ (๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท)) โ‰ค ๐ถ.

Page 14: Variational methods for a singular SPDE yielding the

14 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

By Corollaries 5.4 and 5.7 we know that for every 1 โ‰ค ๐‘ < โˆž,

supโ„“โˆˆ(0,1]

ใ€ˆ๐ถ๐‘ใ€‰liโ„“ .๐‘ 1.

Hence, for the functional ๐บ we have that

supโ„“โˆˆ(0,1]

ใ€ˆ๐บ (๐ธ)๐‘ใ€‰extโ„“ = supโ„“โˆˆ(0,1]

ใ€ˆ๐บ (๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท))๐‘ใ€‰liโ„“ โ‰ค supโ„“โˆˆ(0,1]

ใ€ˆ๐ถ๐‘ใ€‰liโ„“ .๐‘ 1.

Since by Theorem 1.4 (ii) ใ€ˆยทใ€‰extโ„“ โ†’ ใ€ˆยทใ€‰ext weakly and ๐บ is lower semicontinuous we have that

ใ€ˆ๐บ (๐ธ)๐‘ใ€‰ext โ‰ค lim infโ„“โ†“0

ใ€ˆ๐บ๐‘ใ€‰extโ„“ .๐‘ 1,

which completes the proof. ๏ฟฝ

1.3. Outline. In Section 2 we show how the Howarthโ€“Kรกrmรกnโ€“Monin identities can be used tocontrol certain Besov and ๐ฟ๐‘ norms by the anharmonic energy E.

In Section 3 we prove Theorem 1.14 and the ฮ“-convergence result for the renormalized energy,see Corollary 1.15.

In Section 4 we prove the optimal Hรถlder regularity 54โˆ’ of minimizers of the renormalized

energy, see Proposition 1.16.In Section 5, based on the spectral gap inequality (1.7), we provide the stochastic arguments to

prove of Proposition 1.8.

2. Estimates for the Burgers eqation

In this section we bound certain Besov and ๐ฟ๐‘ norms of a function๐‘ค โˆˆ W by the anharmonicenergy E(๐‘ค). These bounds will be used in later sections to study the ฮ“-convergence of therenormalized energy (1.10) and regularity properties of its minimizers (see Sections 3 and 4 below).The proof of these estimates is based on the application of the Howarthโ€“Kรกrmรกnโ€“Monin identityfor the Burgers operator.

We rst need to introduce (directional) Besov spaces. These spaces appear naturally throughthe application of the Howarthโ€“Kรกrmรกnโ€“Monin identity (see Proposition 2.3).

Throughout this section, for a function ๐‘“ : T2 โ†’ R we write

๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ) := ๐‘“ (๐‘ฅ + โ„Ž๐‘’ ๐‘— ) โˆ’ ๐‘“ (๐‘ฅ)

where ๐‘ฅ โˆˆ T2, ๐‘— โˆˆ {1, 2}, โ„Ž โˆˆ R, ๐‘’1 = (1, 0) and ๐‘’2 = (0, 1).

Denition 2.1. For a function ๐‘“ : T2 โ†’ R, ๐‘— โˆˆ {1, 2}, ๐‘  โˆˆ (0, 1] and ๐‘ โˆˆ [1,โˆž) we dene thefollowing (directional) Besov seminorm24

โ€– ๐‘“ โ€– ยคB๐‘ ๐‘ ;๐‘—

:= supโ„Žโˆˆ(0,1]

1โ„Ž๐‘ 

(โˆซT2

|๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ) |๐‘ d๐‘ฅ) 1๐‘

. (2.1)

Notice that in comparison to standard Besov spaces our denition measures regularity in ๐‘ฅ1and ๐‘ฅ2 separately. We have also omitted the second lower index which usually appears in standardBesov spaces since in our case it is alwaysโˆž (corresponding to ยคB๐‘ 

๐‘,โˆž).

Remark 2.2. For ๐‘  โ‰ฅ 0, given a periodic function ๐‘“ : T2 โ†’ R, we dene |๐œ•๐‘— |๐‘  ๐‘“ in Fourier spacevia ๏ฟฝ|๐œ•๐‘— |๐‘  ๐‘“ (๐‘˜) := |๐‘˜ ๐‘— |๐‘  ๐‘“ (๐‘˜), ๐‘˜ โˆˆ (2๐œ‹Z)2.

24Note that one can take the supremum over all โ„Ž โˆˆ R in (2.1) by replacing โ„Ž๐‘  with |โ„Ž |๐‘  . Indeed, if โ„Ž โˆˆ [โˆ’1, 0) thequantity on the right-hand side of (2.1) does not change by symmetry. If โ„Ž โˆˆ R \ [โˆ’1, 1], one writes โ„Ž = โ„Žfr + โ„Žint withโ„Žfr โˆˆ (0, 1] and โ„Žint โˆˆ Z and uses that โ€–๐œ•โ„Ž

๐‘—๐‘“ โ€–๐ฟ๐‘ = โ€–๐œ•โ„Žfr

๐‘—๐‘“ โ€–๐ฟ๐‘ (by periodicity) while 1

|โ„Ž |๐‘  โ‰ค 1โ„Ž๐‘ fr

.

Page 15: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 15

For ๐‘  < 0 and a periodic distribution ๐‘“ of vanishing average in ๐‘ฅ ๐‘— 25, we can dene |๐œ•๐‘— |๐‘  ๐‘“ in thesame way for ๐‘˜ ๐‘— โ‰  0. For ๐‘ = 2, ๐‘  โˆˆ (0, 1) and ๐‘  โ€ฒ โˆˆ (๐‘ , 1), the Parseval identity implies theequivalence 26โˆซT2

๏ฟฝ๏ฟฝ|๐œ•๐‘— |๐‘  ๐‘“ ๏ฟฝ๏ฟฝ2 d๐‘ฅ =โˆ‘๏ธ

๐‘˜โˆˆ(2๐œ‹Z)2|๐‘˜ ๐‘— |2๐‘  |๐‘“ (๐‘˜) |2 = ๐‘๐‘ 

โˆซR

1|โ„Ž |2๐‘ 

โˆซT2

|๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ) |2 d๐‘ฅdโ„Ž|โ„Ž | . ๐ถ (๐‘ , ๐‘ 

โ€ฒ)โ€– ๐‘“ โ€–2ยคB๐‘ โ€ฒ2;๐‘—

(2.2)

for some positive constant ๐ถ (๐‘ , ๐‘  โ€ฒ) depending only on ๐‘  and ๐‘  โ€ฒ, where we used (B.4) below.

In the next proposition we prove two core estimates based on the Howarthโ€“Kรกrmรกnโ€“Moninidentities [GJO15, Lemma 4.1] for the Burgers operator. In [GJO15, Lemma 4.1] the authors dealwith the operator๐‘ค โ†ฆโ†’ ๐œ•2๐‘ค + ๐œ•1 12๐‘ค

2, but the same proof extends to our setting.

Proposition 2.3. There exists ๐ถ > 0 such that for every๐‘ค โˆˆ ๐ฟ2(T2) with vanishing average in ๐‘ฅ1and for every โ„Ž โˆˆ (0, 1) we have โˆซ

T2|๐œ•โ„Ž1๐‘ค |3 d๐‘ฅ โ‰ค ๐ถโ„Ž 3

2E(๐‘ค), (2.3)

sup๐‘ฅ2โˆˆ[0,1)

1โ„Ž

โˆซ โ„Ž

0

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘ค |2 d๐‘ฅ1dโ„Žโ€ฒ โ‰ค ๐ถโ„Ž

12E(๐‘ค) . (2.4)

Proof. By the Howarthโ€“Kรกrmรกnโ€“Monin identities [GJO15, Lemma 4.1] for the Burgers operatorwe know that for every โ„Žโ€ฒ โˆˆ (0, 1)

๐œ•212

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘ค |๐œ•โ„Žโ€ฒ1 ๐‘ค d๐‘ฅ1 โˆ’ ๐œ•โ„Žโ€ฒ

16

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘ค |3 d๐‘ฅ1 =

โˆซ 1

0๐œ•โ„Ž

โ€ฒ1 [๐‘ค |๐œ•โ„Ž

โ€ฒ1 ๐‘ค | d๐‘ฅ1, (2.5)

๐œ•212

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘ค |2 d๐‘ฅ1 โˆ’ ๐œ•โ„Žโ€ฒ

16

โˆซ 1

0(๐œ•โ„Žโ€ฒ1 ๐‘ค)3 d๐‘ฅ1 =

โˆซ 1

0๐œ•โ„Ž

โ€ฒ1 [๐‘ค๐œ•

โ„Žโ€ฒ1 ๐‘ค d๐‘ฅ1. (2.6)

To prove (2.3), we integrate (2.5) over ๐‘ฅ2 and use the periodicity of๐‘ค to obtain,

๐œ•โ„Žโ€ฒ

โˆซT2

|๐œ•โ„Žโ€ฒ1 ๐‘ค |3 d๐‘ฅ = โˆ’6โˆซT2[๐‘ค๐œ•

โˆ’โ„Žโ€ฒ1 |๐œ•โ„Žโ€ฒ1 ๐‘ค | d๐‘ฅ . (2.7)

The last term is estimated as follows,(โˆซT2[๐‘ค๐œ•

โˆ’โ„Žโ€ฒ1 |๐œ•โ„Žโ€ฒ1 ๐‘ค | d๐‘ฅ

)2โ‰ค

โˆซT2( |๐œ•1 |โˆ’

12[๐‘ค)2 d๐‘ฅ

โˆซT2( |๐œ•1 |

12 ๐œ•โˆ’โ„Ž

โ€ฒ1 |๐œ•โ„Žโ€ฒ1 ๐‘ค |)2 d๐‘ฅ

. |โ„Žโ€ฒ |โˆซT2( |๐œ•1 |โˆ’

12[๐‘ค)2 d๐‘ฅ

โˆซT2(๐œ•1๐‘ค)2 d๐‘ฅ,

where we use thatโˆซT2( |๐œ•1 |

12 ๐œ•โˆ’โ„Ž

โ€ฒ1 |๐œ•โ„Žโ€ฒ1 ๐‘ค |)2 d๐‘ฅ .

(โˆซT2(๐œ•โˆ’โ„Žโ€ฒ1 |๐œ•โ„Žโ€ฒ1 ๐‘ค |)2 d๐‘ฅ

) 12(โˆซT2(๐œ•1 |๐œ•โ„Ž

โ€ฒ1 ๐‘ค |)2 d๐‘ฅ

) 12

. |โ„Žโ€ฒ |โˆซT2(๐œ•1 |๐œ•โ„Ž

โ€ฒ1 ๐‘ค |)2 d๐‘ฅ .

Integrating (2.7) over โ„Žโ€ฒ โˆˆ (0, โ„Ž), we obtain thatโˆซT2

|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |3 d๐‘ฅ . โ„Ž 32

(โˆซT2(๐œ•1๐‘ค)2 d๐‘ฅ

) 12(โˆซT2( |๐œ•1 |โˆ’

12[๐‘ค)2 d๐‘ฅ

) 12

which in turns implies (2.3).

25We say that a periodic distribution ๐‘“ has vanishing average in ๐‘ฅ1 if ๐‘“ (eโˆ’i๐‘˜2 ยท) = 0 for all ๐‘˜2 โˆˆ 2๐œ‹Z, and analogouslyfor ๐‘“ with vanishing average in ๐‘ฅ2.

26Note that ๐ถ (๐‘ , ๐‘  โ€ฒ) โ†’ โˆž as ๐‘  โ€ฒ โ†˜ ๐‘  .

Page 16: Variational methods for a singular SPDE yielding the

16 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

To prove (2.4), we integrate (2.6) over โ„Žโ€ฒ โˆˆ (0, โ„Ž) to obtain with ๐œ•01๐‘ค = 0

๐œ•212

โˆซ โ„Ž

0

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘ค |2 d๐‘ฅ1dโ„Žโ€ฒ โˆ’

16

โˆซ 1

0(๐œ•โ„Ž1๐‘ค)3 d๐‘ฅ1 =

โˆซ โ„Ž

0

โˆซ 1

0๐œ•โ„Ž

โ€ฒ1 [๐‘ค๐œ•

โ„Žโ€ฒ1 ๐‘ค d๐‘ฅ1 dโ„Žโ€ฒ.

By the Sobolev embedding๐‘Š 1,1(T) โŠ‚ ๐ฟโˆž(T) on the torus T = [0, 1) in the form

sup๐‘งโˆˆT

|๐‘“ (๐‘ง) | โ‰คโˆซT|๐‘“ (๐‘ง) | d๐‘ง +

โˆซT|๐‘“ โ€ฒ(๐‘ง) | d๐‘ง,

we can therefore estimate

sup๐‘ฅ2โˆˆ[0,1)

โˆซ โ„Ž

0

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘ค |2 d๐‘ฅ1dโ„Žโ€ฒ .

โˆซ โ„Ž

0

โˆซT2

|๐œ•โ„Žโ€ฒ1 ๐‘ค |2 d๐‘ฅdโ„Žโ€ฒ +โˆซT2

|๐œ•โ„Ž1๐‘ค |3 d๐‘ฅ

+โˆซ โ„Ž

0

โˆซ 1

0

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซ 1

0๐œ•โ„Ž

โ€ฒ1 [๐‘ค๐œ•

โ„Žโ€ฒ1 ๐‘ค d๐‘ฅ1

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ d๐‘ฅ2dโ„Žโ€ฒ.The rst term on the right-hand side can be bounded usingโˆซ

T2|๐œ•โ„Žโ€ฒ1 ๐‘ค |2 d๐‘ฅ โ‰ค (โ„Žโ€ฒ)2

โˆซT2

|๐œ•1๐‘ค |2 d๐‘ฅ โ‰ค (โ„Žโ€ฒ)2E(๐‘ค).

For the second term we use (2.3). Last, for the third term, the same argument used to estimate theright-hand side of (2.7) leads toโˆซ โ„Ž

0

โˆซ 1

0

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซ 1

0๐œ•โ„Ž

โ€ฒ1 [๐‘ค๐œ•

โ„Žโ€ฒ1 ๐‘ค d๐‘ฅ1

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ d๐‘ฅ2dโ„Žโ€ฒ . โ€–|๐œ•1 |โˆ’12[๐‘ค โ€–๐ฟ2 โ€–๐œ•1๐‘ค โ€–๐ฟ2

โˆซ โ„Ž

0(โ„Žโ€ฒ) 1

2 dโ„Žโ€ฒ . โ„Ž32E(๐‘ค) .

Hence, we can bound

sup๐‘ฅ2โˆˆ[0,1)

โˆซ โ„Ž

0

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘ค |2 d๐‘ฅ1dโ„Žโ€ฒ . (โ„Ž3 + โ„Ž 3

2 )E(๐‘ค) . โ„Ž 32E(๐‘ค)

for all โ„Ž โˆˆ (0, 1). ๏ฟฝ

We are now ready to prove the main result of this section.

Proposition 2.4. We have the following estimates:

(i) โ€–๐‘ค โ€– ยคB๐‘ 3;1

โ‰ค ๐ถE(๐‘ค) 13 , for every ๐‘  โˆˆ (0, 12 ], (2.8)

(ii) โ€–๐‘ค โ€– ยคB๐‘ 2;1

โ‰ค ๐ถE(๐‘ค) 2๐‘ +16 , for every ๐‘  โˆˆ [ 12 , 1], (2.9)

(iii) โ€–๐‘ค โ€–๐ฟ๐‘ โ‰ค ๐ถ (๐‘)E(๐‘ค)๐‘โˆ’12๐‘ , for every ๐‘ โˆˆ [3, 7), (2.10)

(iv) โ€–๐‘ค2โ€–ยคB2๐‘โˆ’62๐‘

2;1

โ‰ค ๐ถ (๐‘)E(๐‘ค)2๐‘โˆ’32๐‘ , for every ๐‘ โˆˆ [6, 7), (2.11)

for every๐‘ค โˆˆ ๐ฟ2(T2) with vanishing average in ๐‘ฅ1, where๐ถ > 0 is a universal constant and๐ถ (๐‘) > 0depends on ๐‘ .

Note that a result similar to (2.8) was obtained in [OS10, Lemma 4] using dierent techniques.

Proof. (i) This is immediate from (2.3), Denition 2.1 and (B.3).(ii) By interpolation we have for ๐‘  โˆˆ [ 12 , 1]:

โ€–๐‘ค โ€– ยคB๐‘ 2;1

โ‰ค โ€–๐‘ค โ€–2(1โˆ’๐‘ )ยคB122;1

โ€–๐‘ค โ€–2๐‘ โˆ’1ยคB12;1.

Using (B.3) and (2.8) we get

โ€–๐‘ค โ€–ยคB122;1

โ‰ค โ€–๐‘ค โ€–ยคB123;1

. E(๐‘ค) 13 ,

Page 17: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 17

with an implicit constant independent of ๐‘  . We also have the bound

โ€–๐‘ค โ€– ยคB12;1

โ‰ค โ€–๐œ•1๐‘ค โ€–๐ฟ2 โ‰ค E(๐‘ค) 12 .

Combining these estimates implies (2.9).(iii) We divide the proof into several steps.Step 1: We rst prove that

sup๐‘ฅ2โˆˆ[0,1)

(โˆซ 1

0|๐‘ค (๐‘ฅ) |๐‘ d๐‘ฅ1

) 1๐‘

.๐‘ E(๐‘ค) 12 for all 2 โ‰ค ๐‘ < 4.

Indeed, by (B.12) for ๐‘ โˆˆ [2, 4), ๐‘ž = 2 and ๐‘“ = ๐‘ค (ยท, ๐‘ฅ2) (with ๐‘ฅ2 โˆˆ [0, 1) xed) we know that

sup๐‘ฅ2โˆˆ[0,1)

(โˆซ 1

0|๐‘ค (๐‘ฅ) |๐‘ d๐‘ฅ1

) 1๐‘

.๐‘ sup๐‘ฅ2โˆˆ[0,1)

โˆซ 1

0

(โˆซ 1

0|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |2 d๐‘ฅ1

) 12 1

โ„Ž12โˆ’

1๐‘

dโ„Žโ„Ž.

Since ๐‘ < 4 we have that

sup๐‘ฅ2โˆˆ[0,1)

โˆซ 1

0

(โˆซ 1

0|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |2 d๐‘ฅ1

) 12 1

โ„Ž12โˆ’

1๐‘

dโ„Žโ„Ž.๐‘ sup

๐‘ฅ2โˆˆ[0,1)sup

โ„Žโˆˆ(0,1]

1โ„Ž

14

(โˆซ 1

0|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |2 d๐‘ฅ1

) 12

.

By (B.18) and (2.4) we also know that

sup๐‘ฅ2โˆˆ[0,1)

supโ„Žโˆˆ(0,1]

1โ„Ž

14

(โˆซ 1

0|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |2 d๐‘ฅ1

) 12

. sup๐‘ฅ2โˆˆ[0,1)

supโ„Žโˆˆ(0,1]

1โ„Ž

14

(1โ„Ž

โˆซ โ„Ž

0

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘ค (๐‘ฅ) |2 d๐‘ฅ1dโ„Žโ€ฒ

) 12

. E(๐‘ค) 12 ,

which combined with the previous estimates implies the desired estimate.Step 2: We prove that (โˆซ 1

0sup

๐‘ฅ1โˆˆ[0,1)|๐‘ค (๐‘ฅ) |3 d๐‘ฅ2

) 13

. E(๐‘ค) 13 .

By (B.12) for ๐‘ = โˆž, ๐‘ž = 3 and ๐‘“ = ๐‘ค (ยท, ๐‘ฅ2), we know that

sup๐‘ฅ1โˆˆ[0,1)

|๐‘ค (๐‘ฅ) | .โˆซ 1

0

(1โ„Ž

โˆซ 1

0|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |3 d๐‘ฅ1

) 13 dโ„Žโ„Ž

for every ๐‘ฅ2 โˆˆ [0, 1). Using Minkowskiโ€™s inequality we obtain the bound(โˆซ 1

0sup

๐‘ฅ1โˆˆ[0,1)|๐‘ค (๐‘ฅ) |3 d๐‘ฅ2

) 13

.

โˆซ 1

0

(1โ„Ž

โˆซT2

|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |3d๐‘ฅ) 1

3 dโ„Žโ„Ž.

Using (2.3), the last term in the above inequality is bounded byโˆซ 1

0

(โˆซT2

|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |3d๐‘ฅ 1โ„Ž

) 13 dโ„Žโ„Ž.

โˆซ 1

0

1โ„Ž

56E(๐‘ค) 1

3 dโ„Ž

which implies the desired estimate.Step 3: We are now ready to prove (2.10). For 5 โ‰ค ๐‘ < 7 this is immediate from Step 1 and Step2, since we have that(โˆซ

T2|๐‘ค (๐‘ฅ) |๐‘ d๐‘ฅ

) 1๐‘

.

(โˆซ 1

0sup

๐‘ฅ1โˆˆ[0,1)|๐‘ค (๐‘ฅ) |3 d๐‘ฅ2 sup

๐‘ฅ2โˆˆ[0,1)

โˆซ 1

0|๐‘ค (๐‘ฅ) |๐‘โˆ’3 d๐‘ฅ1

) 1๐‘

.๐‘ E(๐‘ค)๐‘โˆ’12๐‘ .

Page 18: Variational methods for a singular SPDE yielding the

18 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

Step 2 also implies that โ€–๐‘ค โ€–๐ฟ3 . E(๐‘ค) 13 which proves the bound for ๐‘ = 3, so it remains to prove

the bound for ๐‘ โˆˆ (3, 5] โŠ‚ [3, 6]. We proceed using interpolation for 1๐‘= 1

36โˆ’๐‘๐‘

+ 16 (2 โˆ’

6๐‘) to

bound

โ€–๐‘ค โ€–๐ฟ๐‘ โ‰ค โ€–๐‘ค โ€–6โˆ’๐‘๐‘

๐ฟ3โ€–๐‘ค โ€–

2โˆ’ 6๐‘

๐ฟ6. E(๐‘ค)

6โˆ’๐‘3๐‘ E(๐‘ค) (2โˆ’

6๐‘) 512 = E(๐‘ค)

๐‘โˆ’12๐‘ .

(iv) We rst notice that by Hรถlderโ€™s inequality, with exponents ๐‘โˆ’2๐‘+ 2๐‘= 1, translation invariance

and Minkowskiโ€™s inequality, we have thatโˆซT2

(๐œ•โ„Ž1๐‘ค

2(๐‘ฅ))2

d๐‘ฅ โ‰ค(โˆซT2

|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |2๐‘๐‘โˆ’2 d๐‘ฅ

) ๐‘โˆ’2๐‘

(โˆซT2

|๐‘ค (๐‘ฅ + โ„Ž๐‘’1) +๐‘ค (๐‘ฅ) |๐‘ d๐‘ฅ) 2๐‘

โ‰ค 4(โˆซT2

|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |2๐‘๐‘โˆ’2 d๐‘ฅ

) ๐‘โˆ’2๐‘

(โˆซT2

|๐‘ค (๐‘ฅ) |๐‘ d๐‘ฅ) 2๐‘

. (2.12)

As ๐‘ โˆˆ [6, 7), we have 2๐‘๐‘โˆ’2 โˆˆ (2, 3], and by interpolation we obtain the boundโˆซ

T2|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |

2๐‘๐‘โˆ’2 d๐‘ฅ โ‰ค

(โˆซT2

|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |2 d๐‘ฅ) ๐‘โˆ’6๐‘โˆ’2

(โˆซT2

|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |3 d๐‘ฅ) 4๐‘โˆ’2

.

Using thatโˆซT2

|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |2 d๐‘ฅ . โ„Ž2E(๐‘ค) and (2.3), the last inequality implies thatโˆซT2

|๐œ•โ„Ž1๐‘ค (๐‘ฅ) |2๐‘๐‘โˆ’2 d๐‘ฅ . โ„Ž2

(๐‘โˆ’6๐‘โˆ’2

)+ 32

(4

๐‘โˆ’2

)E(๐‘ค). (2.13)

Combining (2.12), (2.13) and (2.10) we getโˆซT2

(๐œ•โ„Ž1๐‘ค

2(๐‘ฅ))2

d๐‘ฅ .๐‘(โ„Ž2(๐‘โˆ’6๐‘โˆ’2

)+ 32

(4

๐‘โˆ’2

)E(๐‘ค)

) ๐‘โˆ’2๐‘

E(๐‘ค)๐‘โˆ’1๐‘ = โ„Ž

2๐‘โˆ’6๐‘ E(๐‘ค)

2๐‘โˆ’3๐‘

which implies (2.11). ๏ฟฝ

As (B.4) and (2.11) imply that E(๐‘ค) (to some power) controls the quantity โ€–|๐œ•1 |12๐‘ค2โ€–2, it follows

that the harmonic part H(๐‘ค) given in (1.6) of the energy E(๐‘ค) is also controlled by E(๐‘ค).Moreover, E(๐‘ค) controls the ๐ฟ2 norm of the 2

3 -fractional derivative in ๐‘ฅ2 because the harmonicpartH(๐‘ค) does. We summarize this in the next proposition, where we also prove that E(๐‘ค) iscontrolled byH(๐‘ค).Proposition 2.5.

(i) For every ^ โˆˆ (0, 114 ), there exists a constant ๐ถ (^) > 0 such that

H(๐‘ค) โ‰ค ๐ถ (^)(1 + E(๐‘ค) 3

2+^), (2.14)

for every๐‘ค โˆˆ ๐ฟ2(T2) with vanishing average in ๐‘ฅ1. In addition, there exists a constant ๐ถ > 0such that โˆซ

T2|๐œ•1๐‘ค |2 d๐‘ฅ +

โˆซT2

| |๐œ•2 |23๐‘ค |2d๐‘ฅ โ‰ค ๐ถH(๐‘ค), (2.15)

for every๐‘ค โˆˆ ๐ฟ2(T2) with vanishing average in ๐‘ฅ1. In particular, [๐‘ค]โˆ’ 14. H(๐‘ค) 1

2 .

(ii) There exists a constant ๐ถ > 0 such that

E(๐‘ค) . 1 + H (๐‘ค)2 for every๐‘ค โˆˆ W . (2.16)

Proof. (i) Fix ^ โˆˆ (0, 114 ) and choose ๐‘ = ๐‘ (^) โˆˆ (6, 7) such that 2๐‘โˆ’3

๐‘= 3

2 + ^. Recalling that[๐‘ค = ๐œ•2๐‘ค โˆ’ ๐œ•1 12๐‘ค

2, by (B.4) and the fact that 2๐‘โˆ’62๐‘ โˆˆ ( 12 , 1) we have

H(๐‘ค) .โˆซT2

|๐œ•1๐‘ค |2 d๐‘ฅ +โˆซT2

| |๐œ•1 |โˆ’12[๐‘ค |2d๐‘ฅ +

โˆซT2

| |๐œ•1 |12๐‘ค2 |2d๐‘ฅ .^ E(๐‘ค) + โ€–๐‘ค2โ€–2

ยคB2๐‘โˆ’62๐‘

2;1

.

Page 19: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 19

By (2.11) we know that โ€–๐‘ค2โ€–2ยคB2๐‘โˆ’62๐‘

2;1

.^ E(๐‘ค)2๐‘โˆ’3๐‘ , thus (2.14) follows by Youngโ€™s inequality.

Inequality (2.15) is proved in a more general context in Lemma B.6 and the last statementfollows from Lemma B.8.

(ii) We have

E(๐‘ค) = H(๐‘ค) โˆ’โˆซT2

(|๐œ•1 |โˆ’

12 ๐œ•2๐‘ค |๐œ•1 |โˆ’

12 ๐œ•1๐‘ค

2)d๐‘ฅ + 1

4

โˆซT2

(|๐œ•1 |โˆ’

12 ๐œ•1๐‘ค

2)2

d๐‘ฅ

โ‰ค H(๐‘ค) + H (๐‘ค) 12

(โˆซT2

(|๐œ• | 12๐‘ค2

)2d๐‘ฅ

) 12+ 14

โˆซT2

(|๐œ• | 12๐‘ค2

)2d๐‘ฅ .

By Lemma B.7 the claimed inequality (2.16) follows. ๏ฟฝ

3. ฮ“-convergence of the renormalized energy

In this section we study the ฮ“-convergence of the renormalized energy as the regularizationof white noise is removed, i.e., the limit โ„“ โ†“ 0. As a consequence we will get the existence ofminimizers of the limiting โ€œrenormalized energyโ€, in particular, the existence of weak solutions ofthe Euler-Lagrange equation in (1.13).

3.1. Proof of Theorem 1.14. We begin with the proof of coercivity statement (i) in Theorem 1.14.

Proof of Theorem 1.14 (i). Let (b, ๐‘ฃ, ๐น ) โˆˆ T be xed. Since ๐‘ฃ has vanishing average in ๐‘ฅ1 we canestimate โ€–๐‘ฃ โ€–๐ฟโˆž . [๐‘ฃ] 3

4โˆ’Y(see e.g., [IO19, Lemma 12]), where the implicit constant is universal for

small Y (e.g., Y โˆˆ (0, 1100 )). We will use this estimate several times in what follows. We split

G(๐‘ฃ, ๐น ;๐‘ค) =โˆซT2

(๐‘ค2๐‘…1๐œ•2๐‘ฃ + ๐‘ฃ2๐‘…1[๐‘ค + 2๐‘ฃ๐‘ค๐‘…1[๐‘ค + 2๐‘ค๐น โˆ’๐‘ค๐‘ฃ๐‘…1๐œ•1๐‘ฃ2 + (๐‘…1 |๐œ•1 |

12 (๐‘ฃ๐‘ค))2

)d๐‘ฅ

=:6โˆ‘๏ธ

๐‘˜=1G๐‘˜ (๐‘ฃ, ๐น ;๐‘ค), (3.1)

and bound each term separately:(T1) Notice that setting ๐‘” := |๐œ•1 |โˆ’1๐œ•2๐‘ฃ we have

๐œ•1๐‘” = ๐‘…1๐œ•2๐‘ฃ

and ๐‘” โˆˆ C 14โˆ’Y , with [๐‘”] 1

4โˆ’Y. [b]โˆ’ 5

4โˆ’Y, (see Lemma D.1). We can therefore integrate by partsโˆซ

T2๐‘ค2๐‘…1๐œ•2๐‘ฃ d๐‘ฅ =

โˆซT2๐‘ค2๐œ•1๐‘” d๐‘ฅ = โˆ’2

โˆซT2๐‘ค๐œ•1๐‘ค ๐‘” d๐‘ฅ

and obtain the bound

|G1(๐‘ฃ, ๐น ;๐‘ค) | โ‰ค๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘ค2๐‘…1๐œ•2๐‘ฃ d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค 2โ€–๐‘”โ€–๐ฟโˆž โ€–๐‘ค โ€–๐ฟ2 โ€–๐œ•1๐‘ค โ€–๐ฟ2 . [๐‘”] 14โˆ’Y

โ€–๐‘ค โ€–๐ฟ3 โ€–๐œ•1๐‘ค โ€–๐ฟ2

. [๐‘”] 14โˆ’Y

E(๐‘ค) 13+

12 . [b]โˆ’ 5

4โˆ’YE(๐‘ค) 5

6 ,

where we used Hรถlderโ€™s and Jensenโ€™s inequality, together with (2.10), as well as โ€–๐‘”โ€–๐ฟโˆž . [๐‘”] 14โˆ’Y

because ๐‘” has zero average. By Youngโ€™s inequality, it follows for any _ โˆˆ (0, 1)

|G1(๐‘ฃ, ๐น ;๐‘ค) | โ‰ค _E(๐‘ค) +๐ถ (1)_

[b]6โˆ’ 54โˆ’Y.

(T2) For the term G2 we have that

|G2(๐‘ฃ, ๐น ;๐‘ค) | =๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2( |๐œ•1 |

12 ๐‘ฃ2) (๐‘…1 |๐œ•1 |โˆ’

12[๐‘ค) d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค โ€–|๐œ•1 |12 ๐‘ฃ2โ€–๐ฟ2 โ€–|๐œ•1 |โˆ’

12[๐‘ค โ€–๐ฟ2

. โ€–๐‘ฃ2โ€–ยคB232;1

E(๐‘ค) 12 . [๐‘ฃ2] 2

3E(๐‘ค) 1

2 . [๐‘ฃ]234โˆ’Y

E(๐‘ค) 12 ,

Page 20: Variational methods for a singular SPDE yielding the

20 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

where we used the Cauchyโ€“Schwarz inequality, boundedness of ๐‘…1 on ๐ฟ2, the estimates (B.4),(B.1) and [IO19, Lemma 12]. Hence, by Youngโ€™s inequality, for any _ โˆˆ (0, 1),

|G2(๐‘ฃ, ๐น ;๐‘ค) | โ‰ค _E(๐‘ค) +๐ถ (2)_

[๐‘ฃ]434โˆ’Y.

(T3) We estimate G3 using Cauchyโ€“Schwarz, the boundedness of ๐‘…1 on ๐ฟ2, and (B.4) by

|G3(๐‘ฃ, ๐น ;๐‘ค) | =๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ2โˆซT2( |๐œ•1 |

12 (๐‘ฃ๐‘ค)) (๐‘…1 |๐œ•1 |โˆ’

12[๐‘ค) d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ. โ€–|๐œ•1 |

12 (๐‘ฃ๐‘ค)โ€–๐ฟ2 โ€–|๐œ•1 |โˆ’

12[๐‘ค โ€–๐ฟ2 . โ€–๐‘ฃ๐‘ค โ€–

ยคB232;1

E(๐‘ค) 12 .

By the fractional Leibniz rule (Lemma B.2 (i)) we can further bound

โ€–๐‘ฃ๐‘ค โ€–ยคB232;1

. โ€–๐‘ฃ โ€–๐ฟโˆž โ€–๐‘ค โ€–ยคB232;1

+ [๐‘ฃ] 23โ€–๐‘ค โ€–๐ฟ2 . [๐‘ฃ] 3

4โˆ’Y

(โ€–๐‘ค โ€–

ยคB232;1

+ โ€–๐‘ค โ€–๐ฟ3),

where we also used Jensenโ€™s inequality. Combined with (2.9) and (2.10), this gives

|G3(๐‘ฃ, ๐น ;๐‘ค) | . [๐‘ฃ] 34โˆ’Y

(E(๐‘ค) 7

18 + E(๐‘ค) 13)E(๐‘ค) 1

2

so that Youngโ€™s inequality yields for any _ โˆˆ (0, 1),

|G3(๐‘ฃ, ๐น ;๐‘ค) | โ‰ค _E(๐‘ค) +๐ถ (3)_

([๐‘ฃ]63

4โˆ’Y+ [๐‘ฃ]93

4โˆ’Y

).

(T4) By the duality Lemma B.3, G4(๐‘ฃ, ๐น ;๐‘ค) = 2โˆซT2๐‘ค๐น d๐‘ฅ can be bounded by

|G4(๐‘ฃ, ๐น ;๐‘ค) | .(โ€–๐œ•1๐‘ค โ€–๐ฟ2 + โ€–|๐œ•2 |

23๐‘ค โ€–๐ฟ2 + โ€–๐‘ค โ€–๐ฟ1

)[๐น ]โˆ’ 8

9

.(โ€–๐œ•1๐‘ค โ€–๐ฟ2 + โ€–|๐œ•2 |

23๐‘ค โ€–๐ฟ2 + โ€–๐‘ค โ€–๐ฟ3

)[๐น ]โˆ’ 3

4โˆ’Y

with a uniform implicit constant for every Y โˆˆ (0, 1100 ), where in the second step we used Jensenโ€™s

inequality and [IO19, Remark 2]. With (2.14), (2.15) and (2.10) we obtain that

โ€–๐œ•1๐‘ค โ€–๐ฟ2 + โ€–|๐œ•2 |23๐‘ค โ€–๐ฟ2 + โ€–๐‘ค โ€–๐ฟ3 . 1 + E(๐‘ค) 3

4+^ + E(๐‘ค) 13 ,

where ^ > 0 can be chosen arbitrarily small (e.g., ^ = 1100 ). This yields the estimate

|G4(๐‘ฃ, ๐น ;๐‘ค) | .(1 + E(๐‘ค) 3

4+^ + E(๐‘ค) 13)[๐น ]โˆ’ 3

4โˆ’Y.

It follows for any _ โˆˆ (0, 1),

|G4(๐‘ฃ, ๐น ;๐‘ค) | โ‰ค _E(๐‘ค) +๐ถ (4)_,^

([๐น ]โˆ’ 3

4โˆ’Y+ [๐น ]

32โˆ’ 3

4โˆ’Y+ [๐น ]

41โˆ’4^โˆ’ 3

4โˆ’Y

).

(T5) For G5(๐‘ฃ, ๐น ;๐‘ค) = โˆ’โˆซT2๐‘ค ๐‘ฃ๐‘…1๐œ•1๐‘ฃ

2 d๐‘ฅ , we use again the duality estimate Lemma B.3,

|G5(๐‘ฃ, ๐น ;๐‘ค) | .(โ€–๐œ•1๐‘ค โ€–๐ฟ2 + โ€–|๐œ•2 |

23๐‘ค โ€–๐ฟ2 + โ€–๐‘ค โ€–๐ฟ1

)[๐‘ฃ๐‘…1๐œ•1๐‘ฃ2]โˆ’ 2

5.

By [IO19, Lemmata 6 and 12] together with (A.13), we have the uniform bound for any Y โˆˆ (0, 1100 )

[๐‘ฃ๐‘…1๐œ•1๐‘ฃ2]โˆ’ 25. [๐‘ฃ] 1

2[๐‘…1๐œ•1๐‘ฃ2]โˆ’ 2

5. [๐‘ฃ] 3

4โˆ’Y[๐œ•1๐‘ฃ2]โˆ’ 1

3. [๐‘ฃ]33

4โˆ’Y.

Hence, as in (T4), we can bound G5(๐‘ฃ, ๐น ;๐‘ค) for some small ^ > 0 (e.g., ^ = 1100 ) by

|G5(๐‘ฃ, ๐น ;๐‘ค) | .(1 + E(๐‘ค) 3

4+^ + E(๐‘ค) 13)[๐‘ฃ]33

4โˆ’Y.

So, for any _ โˆˆ (0, 1), by Youngโ€™s inequality,

|G5(๐‘ฃ, ๐น ;๐‘ค) | โ‰ค _E(๐‘ค) +๐ถ (5)_,^

([๐‘ฃ]

9234โˆ’Y

+ [๐‘ฃ]334โˆ’Y

+ [๐‘ฃ]12

1โˆ’4^34โˆ’Y

).

Page 21: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 21

(T6) For the term G6(๐‘ฃ, ๐น ;๐‘ค) =โˆซT2(๐‘…1 |๐œ•1 |

12 (๐‘ฃ๐‘ค))2 d๐‘ฅ , we rst notice that by boundedness of ๐‘…1

on ๐ฟ2 and the basic estimate (B.4),

G6(๐‘ฃ, ๐น ;๐‘ค) = โ€–|๐œ•1 |12 (๐‘ฃ๐‘ค)โ€–2

๐ฟ2 . โ€–๐‘ฃ๐‘ค โ€–2ยคB232;1

.

Hence, by Lemma B.2 and Jensenโ€™s inequality,

|G6(๐‘ฃ, ๐น ;๐‘ค) | . โ€–๐‘ฃ โ€–2๐ฟโˆž โ€–๐‘ค โ€–2ยคB232;1

+ [๐‘ฃ]223โ€–๐‘ค โ€–2

๐ฟ2 . [๐‘ฃ]234โˆ’Y

(โ€–๐‘ค โ€–2

ยคB232;1

+ โ€–๐‘ค โ€–2๐ฟ3

).

Together with (2.11) and (2.10) we can therefore estimate

|G6(๐‘ฃ, ๐น ;๐‘ค) | . [๐‘ฃ]234โˆ’Y

(E(๐‘ค) 7

9 + E(๐‘ค) 23).

Youngโ€™s inequality then yields for any _ โˆˆ (0, 1)

|G6(๐‘ฃ, ๐น ;๐‘ค) | โ‰ค _E(๐‘ค) +๐ถ (6)_

([๐‘ฃ]63

4โˆ’Y+ [๐‘ฃ]93

4โˆ’Y

). ๏ฟฝ

In the proof of the continuity statement Theorem 1.14 (ii), we need the following lemma.

Lemma 3.1. Let {๐‘คโ„“ }โ„“โ†“0 โŠ‚ W with uniformly bounded energy supโ„“ E(๐‘คโ„“ ) < โˆž, and assume that

๐‘คโ„“ โ†’ ๐‘ค strongly in ๐ฟ2 as โ„“ โ†’ 0 for some๐‘ค โˆˆ W. Then as โ„“ โ†’ 0,

๐œ•1๐‘คโ„“ โ‡€ ๐œ•1๐‘ค, |๐œ•1 |โˆ’12 ๐œ•2๐‘คโ„“ โ‡€ |๐œ•1 |โˆ’

12 ๐œ•2๐‘ค, |๐œ•1 |โˆ’

12[๐‘คโ„“

โ‡€ |๐œ•1 |โˆ’12[๐‘ค, |๐œ•2 |

23๐‘คโ„“ โ‡€ |๐œ•2 |

23๐‘ค,

weakly in ๐ฟ2, and for any ๐‘ 1 โˆˆ (0, 1) and ๐‘ 2 โˆˆ (0, 23 ),

|๐œ•1 |๐‘ 1๐‘คโ„“ โ†’ |๐œ•1 |๐‘ 1๐‘ค, |๐œ•2 |๐‘ 2๐‘คโ„“ โ†’ |๐œ•2 |๐‘ 2๐‘ค strongly in ๐ฟ2.

Proof. For the rst part, we use the fact that a uniformly bounded sequence in ๐ฟ2 converging in thedistributional sense converges weakly in ๐ฟ2. By Proposition 2.5, we have that supโ„“ H(๐‘คโ„“ ) < โˆž .Therefore, {๐œ•1๐‘คโ„“ }โ„“ , {|๐œ•1 |โˆ’

12 ๐œ•2๐‘คโ„“ }โ„“ , {|๐œ•1 |โˆ’

12[๐‘คโ„“

}โ„“ , and {|๐œ•2 |23๐‘คโ„“ }โ„“ are uniformly bounded in ๐ฟ2. They

also converge in the distributional sense since๐‘คโ„“ โ†’ ๐‘ค strongly in ๐ฟ2 (in particular, (๐‘คโ„“ )2 โ†’ ๐‘ค2

strongly in ๐ฟ1, so [๐‘คโ„“โ†’ [๐‘ค in the distributional sense). For the second part, by Lemma B.5, for

any ๐‘ 1 โˆˆ (0, 1), ๐‘ 2 โˆˆ (0, 23 ), we have that {|๐œ•1 |๐‘ 1๐‘คโ„“ }โ„“ is uniformly bounded in the homogeneous

Sobolev space ยค๐ป 23 (1โˆ’๐‘ 1) , and {|๐œ•2 |๐‘ 2๐‘คโ„“ }โ„“ is uniformly bounded in ยค๐ป 2

3โˆ’๐‘ 2 . The compact embeddingยค๐ปmin{ 23 (1โˆ’๐‘ 1),

23โˆ’๐‘ 2 } โ†ฉโ†’ ๐ฟ2 (of periodic functions with vanishing average) yields the conclusion. ๏ฟฝ

Proof of Theorem 1.14 (ii). First, all the convergence statements from Lemma 3.1 hold for the se-quence {๐‘คโ„“ }โ„“ . As (bโ„“ , ๐‘ฃโ„“ , ๐นโ„“ ) โ†’ (b, ๐‘ฃ, ๐น ) in T , by Lemma D.1 we also have that ๐‘”โ„“ := |๐œ•1 |โˆ’1๐œ•2๐‘ฃโ„“ โ†’|๐œ•1 |โˆ’1๐œ•2๐‘ฃ =: ๐‘” in C 1

4โˆ’Y . We will prove the continuity of G using the decomposition G =โˆ‘6

๐‘˜=1 G๐‘˜

in (3.1) and study each term G๐‘˜ separately.(Tโ€ฒ1) For the term G1(๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) =

โˆซT2(๐‘คโ„“ )2๐‘…1๐œ•2๐‘ฃโ„“ d๐‘ฅ , we use integration by parts, and that

๐‘คโ„“ โ†’ ๐‘ค strongly in ๐ฟ2, ๐œ•1๐‘คโ„“ โ‡€ ๐œ•1๐‘ค weakly in ๐ฟ2, and ๐‘”โ„“ โ†’ ๐‘” uniformly on T2,

G1(๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) =โˆซT2(๐‘คโ„“ )2๐œ•1๐‘”โ„“ d๐‘ฅ = โˆ’2

โˆซT2๐‘คโ„“๐œ•1๐‘คโ„“ ๐‘”โ„“ d๐‘ฅ

โ†’ โˆ’2โˆซT2๐‘ค๐œ•1๐‘ค ๐‘” d๐‘ฅ =

โˆซT2๐‘ค2๐œ•1๐‘” d๐‘ฅ =

โˆซT2๐‘ค2๐‘…1๐œ•2๐‘ฃ d๐‘ฅ = G1(๐‘ฃ, ๐น ;๐‘ค).

(Tโ€ฒ2) For the term G2(๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) =โˆซT2๐‘ฃ2โ„“๐‘…1[๐‘คโ„“

d๐‘ฅ we use that |๐œ•1 |โˆ’12[๐‘คโ„“

โ‡€ |๐œ•1 |โˆ’12[๐‘ค weakly in

๐ฟ2 (hence also ๐‘…1 |๐œ•1 |โˆ’12[๐‘คโ„“

โ‡€ ๐‘…1 |๐œ•1 |โˆ’12[๐‘ค weakly in ๐ฟ2), and |๐œ•1 |

12 ๐‘ฃ2โ„“ โ†’ |๐œ•1 |

12 ๐‘ฃ2 strongly in C 1

4โˆ’Y

(see Lemma A.5),

G2(๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) =โˆซT2( |๐œ•1 |

12 ๐‘ฃ2โ„“ ) (๐‘…1 |๐œ•1 |โˆ’

12[๐‘คโ„“

) d๐‘ฅ โ†’โˆซT2( |๐œ•1 |

12 ๐‘ฃ2) (๐‘…1 |๐œ•1 |โˆ’

12[๐‘ค) d๐‘ฅ = G2(๐‘ฃ, ๐น ;๐‘ค) .

Page 22: Variational methods for a singular SPDE yielding the

22 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

(Tโ€ฒ3) Since G3(๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) = 2โˆซT2๐‘ฃโ„“๐‘คโ„“๐‘…1[๐‘คโ„“

d๐‘ฅ = 2โˆซT2

|๐œ•1 |12 (๐‘ฃโ„“๐‘คโ„“ ) ๐‘…1 |๐œ•1 |โˆ’

12[๐‘คโ„“

d๐‘ฅ and, as in (Tโ€ฒ2),๐‘…1 |๐œ•1 |โˆ’

12[๐‘คโ„“

โ‡€ ๐‘…1 |๐œ•1 |โˆ’12[๐‘ค weakly in ๐ฟ2, the claimed convergence follows if we show that

|๐œ•1 |12 (๐‘ฃโ„“๐‘คโ„“ ) โ†’ |๐œ•1 |

12 (๐‘ฃ๐‘ค) strongly in ๐ฟ2. For this, we use the triangle inequality, Lemma B.2

(ii), and that ๐‘คโ„“ โ†’ ๐‘ค , |๐œ•1 |12๐‘คโ„“ โ†’ |๐œ•1 |

12๐‘ค strongly in ๐ฟ2 as well as ๐‘ฃโ„“ โ†’ ๐‘ฃ in C 3

4โˆ’Y โŠ‚ C 23 which

yield as โ„“ โ†’ 0,

โ€–|๐œ•1 |12 (๐‘ฃโ„“๐‘คโ„“ ) โˆ’ |๐œ•1 |

12 (๐‘ฃ๐‘ค)โ€–๐ฟ2 โ‰ค โ€–|๐œ•1 |

12 ((๐‘ฃโ„“ โˆ’ ๐‘ฃ)๐‘คโ„“ )โ€–๐ฟ2 + โ€–|๐œ•1 |

12 (๐‘ฃ (๐‘คโ„“ โˆ’๐‘ค))โ€–๐ฟ2

. โ€–๐‘ฃโ„“ โˆ’ ๐‘ฃ โ€–๐ฟโˆž โ€–|๐œ•1 |12๐‘คโ„“ โ€–๐ฟ2 + [๐‘ฃโ„“ โˆ’ ๐‘ฃ] 2

3โ€–๐‘คโ„“ โ€–๐ฟ2

+ โ€–๐‘ฃ โ€–๐ฟโˆž โ€–|๐œ•1 |12 (๐‘คโ„“ โˆ’๐‘ค)โ€–๐ฟ2 + [๐‘ฃ] 2

3โ€–๐‘คโ„“ โˆ’๐‘ค โ€–๐ฟ2

. [๐‘ฃโ„“ โˆ’ ๐‘ฃ] 34โˆ’Y

(โ€–|๐œ•1 |

12๐‘คโ„“ โ€–๐ฟ2 + โ€–๐‘คโ„“ โ€–๐ฟ2

)+ [๐‘ฃ] 3

4โˆ’Y

(โ€–|๐œ•1 |

12 (๐‘คโ„“ โˆ’๐‘ค)โ€–๐ฟ2 + โ€–๐‘คโ„“ โˆ’๐‘ค โ€–๐ฟ2

)โ†’ 0.

(Tโ€ฒ4) The term G4(๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) = 2โˆซT2๐‘คโ„“๐นโ„“ d๐‘ฅ is treated by duality. Since ๐นโ„“ โ†’ ๐น in Cโˆ’ 3

4โˆ’Y โŠ‚ Cโˆ’ 45

(see e.g., [IO19, Remark 2]). By Lemmata 3.1 and B.3 we have for โ„“ โ†’ 0,|G4(๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) โˆ’ G4(๐‘ฃ, ๐น ;๐‘ค) |

.

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2(๐‘คโ„“ โˆ’๐‘ค)๐นโ„“ d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ + ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘ค (๐นโ„“ โˆ’ ๐น ) d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ. [๐นโ„“ ]โˆ’ 4

5

(โ€–|๐œ•1 |

56 (๐‘คโ„“ โˆ’๐‘ค)โ€–๐ฟ2 + โ€–|๐œ•2 |

23 ยท

56 (๐‘คโ„“ โˆ’๐‘ค)โ€–๐ฟ2 + โ€–๐‘คโ„“ โˆ’๐‘ค โ€–๐ฟ2

)+ [๐นโ„“ โˆ’ ๐น ]โˆ’ 4

5

(โ€–|๐œ•1 |

56๐‘ค โ€–๐ฟ2 + โ€–|๐œ•2 |

23 ยท

56๐‘ค โ€–๐ฟ2 + โ€–๐‘ค โ€–๐ฟ2

)โ†’ 0.

(Tโ€ฒ5) For the continuity of the term G5(๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) = โˆ’โˆซT2๐‘คโ„“ ๐‘ฃโ„“๐‘…1๐œ•1๐‘ฃ

2โ„“ d๐‘ฅ we again use the du-

ality Lemma B.3. Here, the situation is even easier than in (Tโ€ฒ4), as ๐‘ฃโ„“๐‘…1๐œ•1๐‘ฃ2โ„“ converges to thenonsingular product ๐‘ฃ๐‘…1๐œ•1๐‘ฃ2 in Cโˆ’ 1

4โˆ’2Y . This convergence follows by[๐‘ฃโ„“๐‘…1๐œ•1๐‘ฃ2โ„“ โˆ’ ๐‘ฃ๐‘…1๐œ•1๐‘ฃ2]โˆ’ 1

4โˆ’2Y

= [(๐‘ฃโ„“ โˆ’ ๐‘ฃ)๐‘…1๐œ•1๐‘ฃ2 + ๐‘ฃโ„“๐‘…1๐œ•1((๐‘ฃโ„“ โˆ’ ๐‘ฃ) (๐‘ฃโ„“ + ๐‘ฃ))]โˆ’ 14โˆ’2Y

. [๐‘ฃโ„“ โˆ’ ๐‘ฃ] 34โˆ’Y

[๐‘…1๐œ•1๐‘ฃ2]โˆ’ 14โˆ’2Y

+ [๐‘ฃโ„“ ] 34โˆ’Y

[๐‘…1๐œ•1((๐‘ฃโ„“ โˆ’ ๐‘ฃ) (๐‘ฃโ„“ + ๐‘ฃ))]โˆ’ 14โˆ’2Y

. [๐‘ฃโ„“ โˆ’ ๐‘ฃ] 34โˆ’Y

[๐‘ฃ2] 34โˆ’Y

+ [๐‘ฃโ„“ ] 34โˆ’Y

[(๐‘ฃโ„“ โˆ’ ๐‘ฃ) (๐‘ฃโ„“ + ๐‘ฃ)] 34โˆ’Y

. [๐‘ฃโ„“ โˆ’ ๐‘ฃ] 34โˆ’Y

([๐‘ฃ]23

4โˆ’Y+ [๐‘ฃโ„“ ]23

4โˆ’Y

)โ†’ 0,

where we used that ๐‘ฃโ„“ โ†’ ๐‘ฃ in C 34โˆ’Y and [IO19, Lemmata 6, 7, and 12]. We conclude as for G4 (with

๐‘ฃโ„“๐‘…1๐œ•1๐‘ฃ2โ„“ corresponding to ๐นโ„“ and ๐‘ฃ๐‘…1๐œ•1๐‘ฃ2 to ๐น , using also that Cโˆ’ 1

4โˆ’2Y โŠ‚ Cโˆ’ 34โˆ’Y ).

(Tโ€ฒ6) Noting that G6(๐‘ฃโ„“ , ๐นโ„“ ;๐‘คโ„“ ) =โˆซT2(๐‘…1 |๐œ•1 |

12 (๐‘ฃโ„“๐‘คโ„“ ))2 d๐‘ฅ = โ€–|๐œ•1 |

12 (๐‘ฃโ„“๐‘คโ„“ )โ€–2๐ฟ2 , continuity follows

since |๐œ•1 |12 (๐‘ฃโ„“๐‘คโ„“ ) โ†’ |๐œ•1 |

12 (๐‘ฃ๐‘ค) in ๐ฟ2 used in (Tโ€ฒ3). ๏ฟฝ

We now prove the compactness Theorem 1.14 (iii) of the sublevel sets of ๐ธ๐‘Ÿ๐‘’๐‘› with respect tothe strong topology in ๐ฟ2.

Proof of Theorem 1.14 (iii). By the coercivity Theorem 1.14 (i) for _ = 12 , it follows that

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) = E(๐‘ค) + G(๐‘ฃ, ๐น ;๐‘ค) โ‰ฅ 12E(๐‘ค) โˆ’๐ถ โ‰ฅ โˆ’๐ถ. (3.2)

Thus ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) is bounded from below and the sublevel set {๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) โ‰ค ๐‘€} over W isincluded in a sublevel set of E overW which is relatively compact in ๐ฟ2 by Lemma B.5. It remainsto prove that the sublevel set {๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) โ‰ค ๐‘€} over W is closed in ๐ฟ2. By the continuity of

Page 23: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 23

G(๐‘ฃ, ๐น ; ยท) (Theorem 1.14 (ii)), it suces to show that E is lower semicontinuous in W, i.e., forevery {๐‘ค โ„“ }โ„“โ†“0 โŠ‚ W with๐‘ค โ„“ โ†’ ๐‘ค in ๐ฟ2, there holds

lim infโ„“โ†“0

E(๐‘ค โ„“ ) โ‰ฅ E(๐‘ค) . (3.3)

Indeed, since ๐‘Ž2 โ‰ฅ ๐‘2 + 2(๐‘Ž โˆ’ ๐‘)๐‘, it follows that

E(๐‘ค โ„“ ) โ‰ฅ E(๐‘ค) + 2โˆซT2

(๐œ•1๐‘ค

โ„“ โˆ’ ๐œ•1๐‘ค)๐œ•1๐‘ค d๐‘ฅ + 2

โˆซT2

(|๐œ•1 |โˆ’

12[๐‘คโ„“ โˆ’ |๐œ•1 |โˆ’

12[๐‘ค

)|๐œ•1 |โˆ’

12[๐‘ค d๐‘ฅ .

Without loss of generality, we may assume that lim inf โ„“โ†“0 E(๐‘ค โ„“ ) = lim supโ„“โ†“0 E(๐‘ค โ„“ ) < โˆž. Hence,by Lemma 3.1, ๐œ•1๐‘ค โ„“ โ‡€ ๐œ•1๐‘ค and |๐œ•1 |โˆ’

12[๐‘คโ„“ โ‡€ |๐œ•1 |โˆ’

12[๐‘ค weakly in ๐ฟ2, and thus, (3.3) follows. The

same argument shows thatH is lower semicontinuous inW. ๏ฟฝ

We are now ready to prove the existence of minimizers Theorem 1.14 (iv).

Proof of Theorem 1.14 (iv). Note that ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; 0) = 0 and recall that ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) is bounded frombelow (see (3.2)) and lower semicontinuous in ๐ฟ2 over its zero sublevel set (due to (1.10), G(๐‘ฃ, ๐น ; ยท)being continuous over any sublevel set of E and E being lower semicontinuous in ๐ฟ2). By the๐ฟ2-compactness of the zero sublevel set of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) over W (Theorem 1.14 (iii)), the directmethod in the calculus of variations yields the existence of minimizers.

In order to show that the minimizer is a distributional solution of the Eulerโ€“Lagrange equation1.13, we do the following splitting:

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) = H(๐‘ค) + (E(๐‘ค) โˆ’ H (๐‘ค)) + G(๐‘ฃ, ๐น ;๐‘ค) = H(๐‘ค) +4โˆ‘๏ธ๐‘—=1

๐ฟ ๐‘— (๐‘ค),

where

๐ฟ1(๐‘ค) =โˆซT2

(2๐‘ค๐น โˆ’๐‘ค๐‘ฃ๐‘…1๐œ•1๐‘ฃ2 + ๐‘ฃ2๐‘…1๐œ•2๐‘ค

)d๐‘ฅ

๐ฟ2(๐‘ค) =โˆซT2

(๐‘ค2๐‘…1๐œ•2๐‘ฃ โˆ’

12๐‘ฃ

2๐‘…1๐œ•1๐‘ค2 + 2๐‘ฃ๐‘ค๐‘…1๐œ•2๐‘ค + (๐‘…1 |๐œ•1 |

12 (๐‘ฃ๐‘ค))2

)d๐‘ฅ

๐ฟ3(๐‘ค) =โˆซT2

(โˆ’๐‘…1 |๐œ•1 |

12๐‘ค2 |๐œ•1 |โˆ’

12 ๐œ•2๐‘ค โˆ’ ๐‘ฃ๐‘ค๐‘…1๐œ•1๐‘ค2

)d๐‘ฅ

๐ฟ4(๐‘ค) =โˆซT2

14

(|๐œ•1 |

12๐‘ค2

)2.

We will show that, given (b, ๐‘ฃ, ๐น ) โˆˆ T , the functional ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) is Cโˆž on the spaceW endowedwith the normH 1

2 , denoted by (W,H 12 ).

Step 1 (Estimating the linear functional ๐ฟ1). We claim that ๐ฟ1 is a continuous linear functionalon (W,H 1

2 ), i.e.,

|๐ฟ1(๐‘ค) | โ‰ค ๐ถH(๐‘ค) 12 , (3.4)

where ๐ถ depends polynomially on [๐‘ฃ] 34โˆ’๐œ–, [๐น ]โˆ’ 3

4โˆ’๐œ–. Indeed, as in (T4) in the proof of (i), by the

duality Lemma B.3 and Poincarรฉโ€™s inequality, we may bound๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT22๐‘ค๐น d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . (โ€–๐œ•1๐‘ค โ€–๐ฟ2 + โ€–|๐œ•2 |

23๐‘ค โ€–๐ฟ2 + โ€–๐‘ค โ€–๐ฟ2

)[๐น ]โˆ’ 8

9. [๐น ]โˆ’ 3

4โˆ’๐œ–H(๐‘ค) 1

2 .

By the same argument, see also (T5) above, we have๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘ค๐‘ฃ๐‘…1๐œ•1๐‘ฃ

2 d๐‘ฅ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . (

โ€–๐œ•1๐‘ค โ€–๐ฟ2 + โ€–|๐œ•2 |23๐‘ค โ€–๐ฟ2 + โ€–๐‘ค โ€–๐ฟ2

)[๐‘ฃ๐‘…1๐œ•1๐‘ฃ2]โˆ’ 2

5. [๐‘ฃ]33

4โˆ’๐œ–H(๐‘ค) 1

2 .

As in (T2), the last term of ๐ฟ1 is estimated using Cauchyโ€“Schwarz, (B.4), and (B.1), by๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘ฃ2๐‘…1๐œ•2๐‘ค d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค โ€–|๐œ•1 |12 ๐‘ฃ2โ€–๐ฟ2 โ€–|๐œ•1 |โˆ’

12 ๐œ•2๐‘ค โ€–๐ฟ2 . [๐‘ฃ]23

4โˆ’๐œ–H(๐‘ค) 1

2 .

Page 24: Variational methods for a singular SPDE yielding the

24 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

Step 2 (Estimating the quadratic functional ๐ฟ2). We claim that ๐ฟ2 is a continuous quadraticfunctional on (W,H 1

2 ), i.e., there exists a continuous bilinear functional๐‘€2 given by๐‘€2(๐‘ค1,๐‘ค2)

=

โˆซT2

(๐‘ค1๐‘ค2๐‘…1๐œ•2๐‘ฃ โˆ’

12๐‘ฃ

2๐‘…1๐œ•1(๐‘ค1๐‘ค2) + 2๐‘ฃ๐‘ค1๐‘…1๐œ•2๐‘ค2 + (|๐œ•1 |12 (๐‘ฃ๐‘ค1)) ( |๐œ•1 |

12 (๐‘ฃ๐‘ค2))

)d๐‘ฅ

such that ๐ฟ2(๐‘ค) = ๐‘€2(๐‘ค,๐‘ค), and satisfying the inequality

|๐‘€2(๐‘ค1,๐‘ค2) | โ‰ค ๐ถH(๐‘ค1)12H(๐‘ค2)

12 , (3.5)

where ๐ถ depends polynomially on [b]โˆ’ 54โˆ’๐œ–, [๐‘ฃ] 3

4โˆ’๐œ–, [๐น ]โˆ’ 3

4โˆ’๐œ–. To prove (3.5), we again treat each

term separately. Similarly to (T1), let ๐‘” := |๐œ•1 |โˆ’1๐œ•2๐‘ฃ , such that ๐‘…1๐œ•2๐‘ฃ = ๐œ•1๐‘”. Recall that by LemmaD.1, [๐‘”] 1

4โˆ’๐œ–. [b]โˆ’ 5

4โˆ’๐œ–. Then integration by parts and Cauchyโ€“Schwarz, together with Poincarรฉโ€™s

inequality, gives๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘ค1๐‘ค2๐‘…1๐œ•2๐‘ฃ d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ = ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐œ•1(๐‘ค1๐‘ค2)๐‘” d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . [๐‘”] 14โˆ’๐œ–

โˆซT2

|๐‘ค1๐œ•1๐‘ค2 +๐‘ค2๐œ•1๐‘ค1 | d๐‘ฅ

. [b]โˆ’ 54โˆ’๐œ–

H(๐‘ค1)12H(๐‘ค2)

12 .

Similarly, the second term can be estimated by๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2

12๐‘ฃ

2๐‘…1๐œ•1(๐‘ค1๐‘ค2) d๐‘ฅ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ = ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซ

T2

12๐‘…1๐‘ฃ

2(๐‘ค1๐œ•1๐‘ค2 +๐‘ค2๐œ•1๐‘ค1) d๐‘ฅ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . [๐‘…1๐‘ฃ2] 3

4โˆ’2๐œ–H(๐‘ค1)

12H(๐‘ค2)

12

. [๐‘ฃ]234โˆ’๐œ–

H(๐‘ค1)12H(๐‘ค2)

12 .

By Cauchyโ€“Schwarz, Lemma B.2 (i), interpolation and Poincarรฉโ€™s inequality, we can bound thethird term by๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT22๐‘ฃ๐‘ค1๐‘…1๐œ•2๐‘ค2 d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ = 2๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2

|๐œ•1 |12 (๐‘ฃ๐‘ค1)๐‘…1 |๐œ•1 |โˆ’

12 (๐œ•2๐‘ค2) d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค 2โ€–|๐œ•1 |12 (๐‘ฃ๐‘ค1)โ€–๐ฟ2 โ€–|๐œ•1 |โˆ’

12 ๐œ•2๐‘ค โ€–๐ฟ2

.(โ€–|๐œ•1 |

12๐‘ค1โ€–๐ฟ2 โ€–๐‘ฃ โ€–๐ฟโˆž + โ€–๐‘ค1โ€–๐ฟ2 [๐‘ฃ] 1

2+๐œ–

)H(๐‘ค2)

12 . [๐‘ฃ] 3

4โˆ’๐œ–H(๐‘ค1)

12H(๐‘ค2)

12 .

Analogously, the fourth term is estimated by๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2( |๐œ•1 |

12 (๐‘ฃ๐‘ค1)) ( |๐œ•1 |

12 (๐‘ฃ๐‘ค2)) d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค โ€–|๐œ•1 |12 (๐‘ฃ๐‘ค1)โ€–๐ฟ2 โ€–|๐œ•1 |

12 (๐‘ฃ๐‘ค2)โ€–๐ฟ2 . [๐‘ฃ]23

4โˆ’๐œ–H(๐‘ค1)

12H(๐‘ค2)

12 .

Step 3 (Estimating the cubic functional ๐ฟ3). We claim that ๐ฟ3 is a continuous cubic functional on(W,H 1

2 ), i.e., there exists a continuous three-linear functional๐‘€3 given by

๐‘€3(๐‘ค1,๐‘ค2,๐‘ค3) = โˆ’โˆซT2

(๐‘…1 |๐œ•1 |

12 (๐‘ค1๐‘ค2) |๐œ•1 |โˆ’

12 ๐œ•2๐‘ค3 + ๐‘ฃ๐‘ค1๐‘…1๐œ•1(๐‘ค2๐‘ค3)

)d๐‘ฅ,

such that ๐ฟ3(๐‘ค) = ๐‘€3(๐‘ค,๐‘ค,๐‘ค), and๐‘€3 is controlled by

|๐‘€3(๐‘ค1,๐‘ค2,๐‘ค3) | โ‰ค ๐ถH(๐‘ค1)12H(๐‘ค2)

12H(๐‘ค3)

12 , (3.6)

where ๐ถ depends polynomially on [๐‘ฃ] 34โˆ’๐œ–, [๐น ]โˆ’ 3

4โˆ’๐œ–. Indeed, the rst term is estimated using

Cauchyโ€“Schwarz and Lemma B.7,๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘…1 |๐œ•1 |

12 (๐‘ค1๐‘ค2) |๐œ•1 |โˆ’

12 ๐œ•2๐‘ค3 d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค โ€–|๐œ•1 |12 (๐‘ค1๐‘ค2)โ€–๐ฟ2 โ€–|๐œ•1 |โˆ’

12 ๐œ•2๐‘ค3โ€–๐ฟ2

. H(๐‘ค1)12H(๐‘ค2)

12H(๐‘ค3)

12 .

Similarly, Lemma B.2 (i) and Lemma B.7 imply that๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘ฃ๐‘ค1๐‘…1๐œ•1(๐‘ค2๐‘ค3) d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค โ€–|๐œ•1 |12 (๐‘ฃ๐‘ค1)โ€–๐ฟ2 โ€–|๐œ•1 |

12 (๐‘ค2๐‘ค3)โ€–๐ฟ2 . [๐‘ฃ] 3

4โˆ’๐œ–H(๐‘ค1)

12H(๐‘ค2)

12H(๐‘ค3)

12 .

Page 25: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 25

Step 4 (Estimating the quartic functional ๐ฟ4). We claim that ๐ฟ4 is a continuous quartic functionalon (W,H 1

2 ), i.e., there exists a continuous four-linear functional๐‘€4 given by

๐‘€4(๐‘ค1,๐‘ค2,๐‘ค3,๐‘ค4) =โˆซT2

14

(|๐œ•1 |

12 (๐‘ค1๐‘ค2) |๐œ•1 |

12 (๐‘ค3๐‘ค4)

)d๐‘ฅ,

such that ๐ฟ4(๐‘ค) = ๐‘€4(๐‘ค,๐‘ค,๐‘ค,๐‘ค). Indeed, Cauchyโ€“Schwarz and Lemma B.7 implies that

|๐‘€4(๐‘ค1,๐‘ค2,๐‘ค3,๐‘ค4) | โ‰ค ๐ถH(๐‘ค1)12H(๐‘ค2)

12H(๐‘ค3)

12H(๐‘ค4)

12 , (3.7)

where ๐ถ depends polynomially on [๐‘ฃ] 34โˆ’๐œ–, [๐น ]โˆ’ 3

4โˆ’๐œ–.

Therefore, the gradient โˆ‡๐‘ค๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) belongs to the dual space of (W,H 12 ), in particular, it

is a distribution, so that

โˆ‡๐‘ค๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) = 0

is the Eulerโ€“Lagrange equation (1.13). ๏ฟฝ

3.2. ฮ“-convergence. In view of Theorem 1.14, we give the proof of ฮ“-convergence of the renor-malized energy for sequences (bโ„“ , ๐‘ฃโ„“ , ๐นโ„“ ) โ†’ (b, ๐‘ฃ, ๐น ) in T as โ„“ โ†’ 0.

Proof of Corollary 1.15. Assume that (bโ„“ , ๐‘ฃโ„“ , ๐นโ„“ ) โ†’ (b, ๐‘ฃ, ๐น ) in T as โ„“ โ†’ 0. By the decomposi-tion of ๐ธ๐‘Ÿ๐‘’๐‘› in (1.10) and the continuity of G in Theorem 1.14 (ii), the pointwise convergence๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ; ยท) โ†’ ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) overW is immediate. We proceed with the proof of the remainingstatements.

(i) (ฮ“ โˆ’ lim inf): Without loss of generality, we may assume that

lim inf โ„“โ†’0 ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘ค โ„“ ) = lim supโ„“โ†’0 ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘ค โ„“ ) < โˆž.

As {(bโ„“ , ๐‘ฃโ„“ , ๐นโ„“ )}โ„“ is uniformly bounded in T , the coercivity Theorem 1.14 (i) implies via (3.2)the existence of a constant ๐ถ > 0 (uniform in โ„“) such that ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘ค โ„“ ) โ‰ฅ 1

2E(๐‘คโ„“ ) โˆ’๐ถ ,

i.e., lim supโ„“โ†’0 E(๐‘ค โ„“ ) < โˆž. The desired inequality is a consequence of (1.10) combinedwith the continuity of G (Theorem 1.14 (ii)) and the lower semicontinuity of E over W in(3.3).

(ii) (ฮ“ โˆ’ lim sup): For๐‘ค โˆˆ W, one sets๐‘คโ„“ = ๐‘ค for all โ„“ โˆˆ (0, 1] and the conclusion follows bythe pointwise convergence of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ; ยท) to ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท).

(iii) (Convergence of minimizers): Let {๐‘คโ„“ }โ„“โ†“0 โŠ‚ W be a sequence of minimizers of thesequence of functionals {๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ; ยท)}โ„“โ†“0 (the existence of minimizers follows from The-orem 1.14 (iv)). As {(bโ„“ , ๐‘ฃโ„“ , ๐นโ„“ )}โ„“ is uniformly bounded in T as โ„“ โ†’ 0, the coercivityTheorem 1.14 (i) implies via (3.2) the existence of a constant ๐ถ > 0 (uniform in โ„“) such thatfor all โ„“

0 = ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ , 0) โ‰ฅ ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘ค โ„“ ) โ‰ฅ 12E(๐‘ค

โ„“ ) โˆ’๐ถ.

This implies that {๐‘คโ„“ }โ„“โ†“0 belongs to the sublevel set 2๐ถ of the energy E. Hence, by LemmaB.5, there exists๐‘ค โˆˆ W such that, upon a subsequence,๐‘คโ„“ โ†’ ๐‘ค strongly in ๐ฟ2. Moreover,๐‘ค is a minimizer of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท) over W because for every ๐‘ค0 โˆˆ W, by the ฮ“ โˆ’ lim infinequality and the pointwise convergence of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ; ยท) to ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท), we have

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) โ‰ค lim infโ„“โ†’0

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘ค โ„“ ) โ‰ค lim supโ„“โ†’0

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘ค โ„“ )

โ‰ค lim supโ„“โ†’0

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘ค0) = ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค0) .

Choosing๐‘ค0 = ๐‘ค in the above relation, we deduce that

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) = limโ„“โ†’0

๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃโ„“ , ๐นโ„“ ;๐‘ค โ„“ ) .

๏ฟฝ

Page 26: Variational methods for a singular SPDE yielding the

26 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

4. A priori estimate for minimizers in Hรถlder spaces

In this section we prove an a priori estimate for minimizers of the renormalized energy ๐ธ๐‘Ÿ๐‘’๐‘› ,as stated in Proposition 1.16. We rst need the following proposition.

Proposition 4.1. There exists ๐ถ > 0 such that for every๐‘ค โˆˆ W and periodic distribution ๐‘“ ,

[๐‘ค๐‘“ ]โˆ’ 34โ‰ค ๐ถH(๐‘ค) 1

2 [๐‘“ ]โˆ’ 12.

Proof. Case 1 (๐‘“ โˆˆ ๐ฟ2โˆฉCโˆ’ 12 (T2)). Since๐‘ค โˆˆ W, the product๐‘ค๐‘“ belongs to ๐ฟ1(T2). We estimate

[๐‘ค๐‘“ ]โˆ’ 34via (A.1) by studying the blow-up of โ€–(๐‘ค๐‘“ )๐‘‡ โ€–๐ฟโˆž for ๐‘‡ โˆˆ (0, 1]. We use the โ€œtelescopicโ€

decomposition

(๐‘ค๐‘“ )๐‘‡ = (๐‘ค๐‘“๐‘‡2)๐‘‡2+

โˆ‘๏ธ๐‘˜โ‰ฅ2, ๐‘ก= ๐‘‡

2๐‘˜

((๐‘ค๐‘“๐‘ก )๐‘‡โˆ’๐‘ก โˆ’ (๐‘ค๐‘“2๐‘ก )๐‘‡โˆ’2๐‘ก

). (4.1)

Step 1 (Bound on โ€–(๐‘ค๐‘“๐‘‡2)๐‘‡2โ€–๐ฟโˆž ): For ๐‘ = 10, Youngโ€™s inequality for convolution in Remark 1.11,

Lemma B.4, (2.15) and (A.2) yield for every ๐‘‡ โˆˆ (0, 1],

โ€–(๐‘ค๐‘“๐‘‡2)๐‘‡2โ€–๐ฟโˆž .

(๐‘‡

13)โˆ’ 5

2๐‘ โ€–๐‘ค๐‘“๐‘‡2โ€–๐ฟ๐‘ .

(๐‘‡

13)โˆ’ 1

4 โ€–๐‘ค โ€–๐ฟ10 โ€– ๐‘“๐‘‡2 โ€–๐ฟโˆž

.(๐‘‡

13)โˆ’ 1

4โˆ’12 H(๐‘ค) 1

2 [๐‘“ ]โˆ’ 12=

(๐‘‡

13)โˆ’ 3

4 H(๐‘ค) 12 [๐‘“ ]โˆ’ 1

2.

Step 2 (Bound on the telescopic sum): By Youngโ€™s inequality for convolution in Remark 1.11 andLemma 4.2 (see below), we obtain via (A.2) for every ๐‘‡ โˆˆ (0, 1], โˆ‘๏ธ

๐‘˜โ‰ฅ2, ๐‘ก= ๐‘‡

2๐‘˜

((๐‘ค๐‘“๐‘ก )๐‘‡โˆ’๐‘ก โˆ’ (๐‘ค๐‘“2๐‘ก )๐‘‡โˆ’2๐‘ก

) ๐ฟโˆž

โ‰คโˆ‘๏ธ

๐‘˜โ‰ฅ2, ๐‘ก= ๐‘‡

2๐‘˜

((๐‘ค๐‘“๐‘ก )๐‘ก โˆ’๐‘ค๐‘“2๐‘ก )๐‘‡โˆ’2๐‘ก ๐ฟโˆž . โˆ‘๏ธ๐‘˜โ‰ฅ2, ๐‘ก= ๐‘‡

2๐‘˜

((๐‘‡ โˆ’ 2๐‘ก) 1

3)โˆ’ 5

4 โ€–(๐‘ค๐‘“๐‘ก )๐‘ก โˆ’๐‘ค๐‘“2๐‘ก โ€–๐ฟ2

.(๐‘‡

13)โˆ’ 5

4โˆ‘๏ธ

๐‘˜โ‰ฅ2, ๐‘ก= ๐‘‡

2๐‘˜

๐‘ก13H(๐‘ค) 1

2 โ€– ๐‘“๐‘ก โ€–๐ฟโˆž .(๐‘‡

13)โˆ’ 5

4โˆ‘๏ธ

๐‘˜โ‰ฅ2, ๐‘ก= ๐‘‡

2๐‘˜

(๐‘ก 13 ) 1

2H(๐‘ค) 12 [๐‘“ ]โˆ’ 1

2

.(๐‘‡

13)โˆ’ 3

4 H(๐‘ค) 12 [๐‘“ ]โˆ’ 1

2.

Step 3 (Hรถlder regularity): By Step 1 and Step 2 we know that for every ๐‘‡ โˆˆ (0, 1],

โ€–(๐‘ค๐‘“ )๐‘‡ โ€–๐ฟโˆž .(๐‘‡

13)โˆ’ 3

4 H(๐‘ค) 12 [๐‘“ ]โˆ’ 1

2,

which combined with (A.1) completes the proof.

Case 2 (๐‘“ โˆˆ Cโˆ’ 12 (T2)). We consider an arbitrary approximation ๐‘“โ„“ โˆˆ ๐ฟ2 โˆฉ Cโˆ’ 1

2 (T2) of ๐‘“ withrespect to [ยท]โˆ’ 1

2. By Case 1we deduce that๐‘ค๐‘“โ„“ is a Cauchy sequence in Cโˆ’ 3

4 , therefore it convergesto the product๐‘ค๐‘“ by the same argument as in [IO19, Lemma 6]. ๏ฟฝ

Lemma 4.2. There exists a constant ๐ถ > 0 such that for every ๐‘ก โˆˆ (0, 1], ๐‘ค โˆˆ W and periodic

distribution ๐‘“ ,

โ€–(๐‘ค๐‘“๐‘ก )๐‘ก โˆ’๐‘ค๐‘“2๐‘ก โ€–๐ฟ2 โ‰ค ๐ถ๐‘ก13H(๐‘ค) 1

2 โ€– ๐‘“๐‘ก โ€–๐ฟโˆž .

Proof. We start with the identity((๐‘ค๐‘“๐‘ก )๐‘ก โˆ’๐‘ค๐‘“2๐‘ก

)(๐‘ฅ) =

โˆซR2๐œ“๐‘ก (๐‘ฆ) (๐‘ค (๐‘ฅ โˆ’ ๐‘ฆ) โˆ’๐‘ค (๐‘ฅ)) ๐‘“๐‘ก (๐‘ฅ โˆ’ ๐‘ฆ) ๐‘‘๐‘ฆ, ๐‘ฅ โˆˆ T2.

Page 27: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 27

By Minkowskiโ€™s inequality, we deduce

โ€–(๐‘ค๐‘“๐‘ก )๐‘ก โˆ’๐‘ค๐‘“2๐‘ก โ€–๐ฟ2 โ‰ค โ€– ๐‘“๐‘ก โ€–๐ฟโˆžโˆซR2

|๐œ“๐‘ก (๐‘ฆ) |โ€–๐‘ค (ยท โˆ’ ๐‘ฆ) โˆ’๐‘ค (ยท)โ€–๐ฟ2 d๐‘ฆ

โ‰ค โ€– ๐‘“๐‘ก โ€–๐ฟโˆžโˆซR2

|๐œ“๐‘ก (๐‘ฆ) |โ€–๐œ•โˆ’๐‘ฆ11 ๐‘ค (๐‘ฅ1, ๐‘ฅ2 โˆ’ ๐‘ฆ2)โ€–๐ฟ2๐‘ฅ d๐‘ฆ

+ โ€– ๐‘“๐‘ก โ€–๐ฟโˆžโˆซR2

|๐œ“๐‘ก (๐‘ฆ) |โ€–๐œ•โˆ’๐‘ฆ22 ๐‘ค โ€–๐ฟ2 d๐‘ฆ,

where we used that ๐‘ค (๐‘ฅ โˆ’ ๐‘ฆ) โˆ’๐‘ค (๐‘ฅ) = ๐œ•โˆ’๐‘ฆ11 ๐‘ค (๐‘ฅ1, ๐‘ฅ2 โˆ’ ๐‘ฆ2) + ๐œ•โˆ’๐‘ฆ22 ๐‘ค (๐‘ฅ1, ๐‘ฅ2) for every ๐‘ฅ โˆˆ T2 and๐‘ฆ โˆˆ R2.

The rst integral can be estimated using the mean value theorem and translation invariance ofthe torus byโˆซ

R2|๐œ“๐‘ก (๐‘ฆ) |

(โˆซT2

|๐œ•โˆ’๐‘ฆ11 ๐‘ค (๐‘ฅ1, ๐‘ฅ2 โˆ’ ๐‘ฆ2) |2 d๐‘ฅ) 1

2d๐‘ฆ โ‰ค

โˆซR2

|๐‘ฆ1๐œ“๐‘ก (๐‘ฆ) |d๐‘ฆ โ€–๐œ•1๐‘ค โ€–๐ฟ2 . ๐‘ก13 โ€–๐œ•1๐‘ค โ€–๐ฟ2,

since by Step 1 in [IO19, proof of Lemma 10] we know that ๐‘ฆ โ†ฆโ†’ ๐‘ฆ1๐œ“ (๐‘ฆ) โˆˆ ๐ฟ1(R2).The second integral can be estimated using the Cauchy-Schwarz inequality and (2.2) byโˆซ

R2|๐œ“๐‘ก (๐‘ฆ) |

(โˆซT2

|๐œ•โˆ’๐‘ฆ22 ๐‘ค (๐‘ฅ) |2 d๐‘ฅ) 1

2d๐‘ฆ

= (๐‘ก 13 ) 3

2 ( 23+

12 )

โˆซR

โˆซR

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ ๐‘ฆ2

(๐‘ก 13 ) 3

2

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ23+

12

|๐œ“๐‘ก (๐‘ฆ) | d๐‘ฆ1

(โˆซT2

|๐œ•โˆ’๐‘ฆ22 ๐‘ค (๐‘ฅ) |2

|๐‘ฆ2 |43

d๐‘ฅ) 1

2 d๐‘ฆ2|๐‘ฆ2 |

12

. (๐‘ก 13 ) 3

2 ( 23+

12 )

โˆซR

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ ๐‘ฆ2

(๐‘ก 13 ) 3

2

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ23+

12

|๐œ“๐‘ก (๐‘ฆ) | d๐‘ฆ1

๐ฟ2๐‘ฆ2 (R)

โ€–|๐œ•2 |23๐‘ค โ€–๐ฟ2 . ๐‘ก

13 โ€–|๐œ•2 |

23๐‘ค โ€–๐ฟ2,

where we also used Minkowskiโ€™s inequality and a change of variables to deduce that โˆซR

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ ๐‘ฆ2

(๐‘ก 13 ) 3

2

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ23+

12

|๐œ“๐‘ก (๐‘ฆ) | d๐‘ฆ1

๐ฟ2๐‘ฆ2 (R)

โ‰ค 1(๐‘ก 1

3 ) 34

โ€–|๐‘ฆ2 | 23+ 12๐œ“ (๐‘ฆ1, ๐‘ฆ2)โ€–๐ฟ2๐‘ฆ2 (R)

๐ฟ1๐‘ฆ1 (R)

, (4.2)

along with the fact that ๐‘ฆ1 โ†ฆโ†’ โ€–|๐‘ฆ2 |23+

12๐œ“ (๐‘ฆ1, ๐‘ฆ2)โ€–๐ฟ2๐‘ฆ2 (R) โˆˆ ๐ฟ

1๐‘ฆ1 (R).

27

Combining the previous estimates with (2.15) implies the desired bound. ๏ฟฝ

We are now ready to prove Proposition 1.16.

Proof of Proposition 1.16. By Theorem 1.14 (iv) if๐‘ค โˆˆ W is a minimizer of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท), then๐‘ค isa weak solution to (1.13). By the Schauder theory for the operator L (see [IO19, Lemma 5]), if๐‘ค โˆˆ W โˆฉ C 5

4โˆ’2๐œ– satises (1.13), we have that

[๐‘ค] 54โˆ’2Y.

[๐‘ƒ(๐น +๐‘ค๐‘…1๐œ•2๐‘ฃ + ๐‘ฃ๐‘…1๐œ•2๐‘ค +๐‘ค๐‘…1๐œ•2๐‘ค

โˆ’ 12 (๐‘ฃ +๐‘ค)๐‘…1๐œ•1(๐‘ฃ +๐‘ค)2

)+ 12 ๐œ•2๐‘…1(๐‘ฃ +๐‘ค)2

]โˆ’ 3

4โˆ’2Y.

(4.3)

27 This follows easily from the bound โ€–|๐‘ฆ2 | 23+ 12๐œ“ (๐‘ฆ1, ๐‘ฆ2)โ€–๐ฟ2๐‘ฆ2 (R)

๐ฟ1๐‘ฆ1 (R)

. โ€–(1 + |๐‘ฆ1 |) (1 + |๐‘ฆ2 |2)๐œ“ (๐‘ฆ)โ€–๐ฟ2๐‘ฆ (R2)

and Plancherelโ€™s identity, using that๐œ“ (๐‘˜) = eโˆ’|๐‘˜1 |3โˆ’๐‘˜22 , see also Step 1 in [IO19, proof of Lemma 10] and Footnote 32.

Page 28: Variational methods for a singular SPDE yielding the

28 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

We estimate each term on the right-hand side of (4.3) separately. The idea is to bound anyterm containing๐‘ค in the seminorm [ยท]โˆ’ 3

4โˆ’2Yby a productH(๐‘ค)๐›พ (\ ) [๐‘ค]\5

4โˆ’2Ywith \ โˆˆ (0, 1) and

๐›พ (\ ) > 0. To prove the statement, the main tools areโ—ฆ the interpolation inequality in Lemma A.2,โ—ฆ Lemma B.8 which yields thatH(๐‘ค) controls [๐‘ค]2โˆ’ 1

4,

โ—ฆ [IO19, Lemmata 6 and 12] stating that for a distribution ๐‘“ โˆˆ C๐›ฝ , ๐›ฝ โˆˆ (โˆ’ 32 , 0) \ {โˆ’1,โˆ’

12 }, and two

functions ๐‘” โˆˆ C๐›พ , ๐‘” โˆˆ C๐›พ with ๐›พ,๐›พ โˆˆ (0, 32 ) both of vanishing average, provided that ๐›ฝ + ๐›พ > 0and ๐›พ โ‰ฅ ๐›พ , the following estimates hold

[๐‘“ ๐‘”]๐›ฝ . [๐‘“ ]๐›ฝ [๐‘”]๐›พ and [๐‘”๐‘”]๐›พ . [๐‘”]๐›พ [๐‘”]๐›พ .

Also, we use that C๐›ผ โŠ‚ C๐›ฝ for any โˆ’ 32 < ๐›ฝ < ๐›ผ < 3

2 with ๐›ผ, ๐›ฝ โ‰  0, see [IO19, Remark 2], andLemma A.4 which implies that the Hilbert transform reduces the regularity by Y on Hรถlder spaces.

For [b]โˆ’ 54โˆ’Y, [๐‘ฃ] 3

4โˆ’Y, [๐น ]โˆ’ 3

4โˆ’Yโ‰ค ๐‘€ , the following estimates hold with an implicit constant

depending on๐‘€ and Y.1. Terms independent of๐‘ค : First, we notice that [๐‘ƒ๐น ]โˆ’ 3

4โˆ’2Yโ‰ค [๐น ]โˆ’ 3

4โˆ’2Y. [๐น ]โˆ’ 3

4โˆ’Y. 1. Also,

by Denition 1.13, we have that

[๐œ•2๐‘…1๐‘ฃ2]โˆ’ 34โˆ’2Y. [๐‘…1๐‘ฃ2] 3

4โˆ’2Y. [๐‘ฃ2] 3

4โˆ’Y. [๐‘ฃ]23

4โˆ’Y. 1,

[๐‘ฃ๐œ•1๐‘…1๐‘ฃ2]โˆ’ 34โˆ’2Y. [๐‘ฃ๐œ•1๐‘…1๐‘ฃ2]โˆ’ 1

4โˆ’2Y. [๐‘ฃ] 3

4โˆ’Y[๐œ•1๐‘…1๐‘ฃ2]โˆ’ 1

4โˆ’2Y. [๐‘ฃ] 3

4โˆ’Y[๐‘ฃ2] 3

4โˆ’Y. [๐‘ฃ]33

4โˆ’Y. 1.

2. Linear terms in๐‘ค : By the interpolation estimate in Lemma A.2 and Lemma B.8, we have

[๐‘ค] 34+3Y. [๐‘ค]

13โˆ’2^1โˆ’ 1

4[๐‘ค]

23+2^154โˆ’2Y

. H(๐‘ค) 16โˆ’^1 [๐‘ค]

23+2^154โˆ’2Y

, (4.4)

[๐‘ค] 14+3Y. [๐‘ค]

23โˆ’2^2โˆ’ 1

4[๐‘ค]

13+2^254โˆ’2Y

. H(๐‘ค) 13โˆ’^2 [๐‘ค]

13+2^254โˆ’2Y

, (4.5)

where ^1, ^2 > 0 are small (as functions of Y) for Y > 0 small enough. This yields

[๐‘ค๐œ•2๐‘…1๐‘ฃ]โˆ’ 34โˆ’2Y. [๐‘ค] 3

4+3Y[๐œ•2๐‘…1๐‘ฃ]โˆ’ 3

4โˆ’2Y. H(๐‘ค) 1

6โˆ’^1 [๐‘ค]23+2^154โˆ’2Y

[๐‘ฃ] 34โˆ’Y

. H(๐‘ค) 16โˆ’^1 [๐‘ค]

23+2^154โˆ’2Y

,

[๐‘ค๐œ•1๐‘…1๐‘ฃ2]โˆ’ 34โˆ’2Y. [๐‘ค๐œ•1๐‘…1๐‘ฃ2]โˆ’ 1

4โˆ’2Y. [๐‘ค] 1

4+3Y[๐œ•1๐‘…1๐‘ฃ2]โˆ’ 1

4โˆ’2Y

. H(๐‘ค) 13โˆ’^2 [๐‘ค]

13+2^254โˆ’2Y

[๐‘ฃ]234โˆ’Y. H(๐‘ค) 1

3โˆ’^2 [๐‘ค]13+2^254โˆ’2Y

,

[๐‘ฃ๐œ•2๐‘…1๐‘ค]โˆ’ 34โˆ’2Y. [๐‘ฃ๐œ•2๐‘…1๐‘ค]โˆ’ 3

4+2Y. [๐‘ฃ] 3

4โˆ’Y[๐œ•2๐‘…1๐‘ค]โˆ’ 3

4+2Y. [๐‘ค] 3

4+3Y

. H(๐‘ค) 16โˆ’^1 [๐‘ค]

23+2^154โˆ’2Y

,

[๐œ•2๐‘…1(๐‘ฃ๐‘ค)]โˆ’ 34โˆ’2Y. [๐‘ฃ๐‘ค] 3

4โˆ’Y. [๐‘ฃ] 3

4โˆ’Y[๐‘ค] 3

4+3Y. H(๐‘ค) 1

6โˆ’^1 [๐‘ค]23+2^154โˆ’2Y

,

[๐‘ฃ๐œ•1๐‘…1(๐‘ฃ๐‘ค)]โˆ’ 34โˆ’2Y. [๐‘ฃ๐œ•1๐‘…1(๐‘ฃ๐‘ค)]โˆ’ 1

4โˆ’2Y. [๐‘ฃ] 3

4โˆ’Y[๐œ•1๐‘…1(๐‘ฃ๐‘ค)]โˆ’ 1

4โˆ’2Y. [๐‘ฃ๐‘ค] 3

4โˆ’Y

. [๐‘ฃ] 34โˆ’Y

[๐‘ค] 34+3Y. H(๐‘ค) 1

6โˆ’^1 [๐‘ค]23+2^154โˆ’2Y

.

3. Quadratic terms in๐‘ค :(a) We start with the term ๐œ•2๐‘…1๐‘ค

2. By the interpolation estimate in Lemma A.2 for ๐›พ = 0 andLemma B.8, we have

โ€–๐‘ค โ€–๐ฟโˆž . [๐‘ค]56โˆ’2^3โˆ’ 1

4[๐‘ค]

16+2^354โˆ’2Y

. H(๐‘ค) 512โˆ’^3 [๐‘ค]

16+2^354โˆ’2Y

, (4.6)

Page 29: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 29

where ^3 > 0 is small for Y > 0 small enough. Together with (4.4), it follows that

[๐œ•2๐‘…1๐‘ค2]โˆ’ 34โˆ’2Y. [๐œ•2๐‘…1๐‘ค2]โˆ’ 3

4+2Y. [๐‘ค2] 3

4+3Y. โ€–๐‘ค โ€–๐ฟโˆž [๐‘ค] 3

4+3Y

. H(๐‘ค) 712โˆ’(^1+^3) [๐‘ค]

56+2(^1+^3)54โˆ’2Y

.

Similarly, we estimate

[๐‘ฃ๐œ•1๐‘…1๐‘ค2]โˆ’ 34โˆ’2Y. [๐‘ฃ๐œ•1๐‘…1๐‘ค2]โˆ’ 1

4+2Y. [๐‘ฃ] 3

4โˆ’Y[๐œ•1๐‘…1๐‘ค2]โˆ’ 1

4+2Y. [๐‘ค2] 3

4+3Y

. H(๐‘ค) 712โˆ’(^1+^3) [๐‘ค]

56+2(^1+^3)54โˆ’2Y

.

(b) The term๐‘ค๐œ•2๐‘…1๐‘ค is treated via Proposition 4.1,

[๐‘ค๐œ•2๐‘…1๐‘ค]โˆ’ 34. H(๐‘ค) 1

2 [๐œ•2๐‘…1๐‘ค]โˆ’ 12. H(๐‘ค) 1

2 [๐‘…1๐‘ค]1 . H(๐‘ค) 12 [๐‘ค]1+Y .

Then, Lemma A.2 and Lemma B.8 yield

[๐‘ค]1+Y . [๐‘ค]16โˆ’2^4โˆ’ 1

4[๐‘ค]

56+2^454โˆ’2Y

. H(๐‘ค) 112โˆ’^4 [๐‘ค]

56+2^454โˆ’2Y

,

where ^4 > 0 is small for Y > 0 small enough. Hence we have

[๐‘ค๐œ•2๐‘…1๐‘ค]โˆ’ 34โˆ’2Y. [๐‘ค๐œ•2๐‘…1๐‘ค]โˆ’ 3

4. H(๐‘ค) 7

12โˆ’^4 [๐‘ค]56+2^454โˆ’2Y

.

(c) We decompose๐‘ค๐‘…1๐œ•1(๐‘ฃ๐‘ค) into๐‘ค๐‘…1(๐‘ค๐œ•1๐‘ฃ) +๐‘ค๐‘…1(๐‘ฃ๐œ•1๐‘ค) and treat each term separately. ByProposition 4.1, we have

[๐‘ค๐‘…1(๐‘ฃ๐œ•1๐‘ค)]โˆ’ 34. H(๐‘ค) 1

2 [๐‘…1(๐‘ฃ๐œ•1๐‘ค)]โˆ’ 12. H(๐‘ค) 1

2 [๐‘…1(๐‘ฃ๐œ•1๐‘ค)]โˆ’ 12+Y

. H(๐‘ค) 12 [๐‘ฃ๐œ•1๐‘ค]โˆ’ 1

2+2Y. H(๐‘ค) 1

2 [๐‘ฃ] 34โˆ’Y

[๐œ•1๐‘ค]โˆ’ 12+2Y. H(๐‘ค) 1

2 [๐‘ค] 12+2Y

.

Again, Lemma A.2 and Lemma B.8 yield

[๐‘ค] 12+2Y. [๐‘ค]

12โˆ’2^5โˆ’ 1

4[๐‘ค]

12+2^554โˆ’2Y

. H(๐‘ค) 14โˆ’^5 [๐‘ค]

12+2^554โˆ’2Y

, (4.7)

where ^5 > 0 is small for Y > 0 small enough. Hence we obtain that

[๐‘ค๐‘…1(๐‘ฃ๐œ•1๐‘ค)]โˆ’ 34โˆ’2Y. [๐‘ค๐‘…1(๐‘ฃ๐œ•1๐‘ค)]โˆ’ 3

4. H(๐‘ค) 3

4โˆ’^5 [๐‘ค]12+2^554โˆ’2Y

.

The term๐‘ค๐‘…1(๐‘ค๐œ•1๐‘ฃ) can be estimated using (4.5) as

[๐‘ค๐‘…1(๐‘ค๐œ•1๐‘ฃ)]โˆ’ 34โˆ’2Y. [๐‘ค๐‘…1(๐‘ค๐œ•1๐‘ฃ)]โˆ’ 1

4โˆ’2Y. [๐‘ค] 1

4+3Y[๐‘…1(๐‘ค๐œ•1๐‘ฃ)]โˆ’ 1

4โˆ’2Y

. [๐‘ค] 14+3Y

[๐‘ค๐œ•1๐‘ฃ]โˆ’ 14โˆ’Y. [๐‘ค]21

4+3Y[๐œ•1๐‘ฃ]โˆ’ 1

4โˆ’Y

. H(๐‘ค) 23โˆ’2^2 [๐‘ค]

23+4^254โˆ’2Y

[๐‘ฃ] 34โˆ’Y. H(๐‘ค) 2

3โˆ’2^2 [๐‘ค]23+4^254โˆ’2Y

.

4. Cubic term in๐‘ค : The cubic term๐‘ค๐œ•1๐‘…1๐‘ค2 is treated by Proposition 4.1 which yields

[๐‘ค๐œ•1๐‘…1๐‘ค2]โˆ’ 34. H(๐‘ค) 1

2 [๐‘…1๐œ•1๐‘ค2]โˆ’ 12. H(๐‘ค) 1

2 [๐‘…1๐œ•1๐‘ค2]โˆ’ 12+Y. H(๐‘ค) 1

2 [๐œ•1๐‘ค2]โˆ’ 12+2Y

. H(๐‘ค) 12 [๐‘ค2] 1

2+2Y. H(๐‘ค) 1

2 โ€–๐‘ค โ€–๐ฟโˆž [๐‘ค] 12+2Y

.

By (4.6) and (4.7) we have that

โ€–๐‘ค โ€–๐ฟโˆž [๐‘ค] 12+2Y. H(๐‘ค) 2

3โˆ’(^3+^5) [๐‘ค]23+2(^3+^5)54โˆ’2Y

,

which in turn implies that

[๐‘ค๐œ•1๐‘…1๐‘ค2]โˆ’ 34โˆ’2Y. [๐‘ค๐œ•1๐‘…1๐‘ค2]โˆ’ 3

4. H(๐‘ค) 3

2โˆ’(^3+^5) [๐‘ค]23+2(^3+^5)54โˆ’2Y

.

Page 30: Variational methods for a singular SPDE yielding the

30 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

Summing up, Youngโ€™s inequality yields the bound

[๐‘ค] 54โˆ’2Y

โ‰ค ๐ถ (1 + H (๐‘ค))๐‘ , (4.8)

for some ๐‘ โ‰ฅ 1, and by our estimates it is clear that the constant ๐ถ depends polynomially on๐‘€ . To conclude, using the fact that๐‘ค is a minimizer of ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; ยท), we have that ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ;๐‘ค) โ‰ค๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ, ๐น ; 0) = 0. Since ๐ธ๐‘Ÿ๐‘’๐‘› (๐‘ฃ .๐น ; ยท) = E + G(๐‘ฃ, ๐น ; ยท) and by Theorem 1.14 (i) we know that|G(๐‘ฃ, ๐น ;๐‘ค) | โ‰ค 1

2E(๐‘ค) +๐ถ for some constant๐ถ which also depends polynomially on๐‘€ , we obtainthat E(๐‘ค) โ‰ค 2๐ถ . By (2.14), this implies that ๐ป (๐‘ค) โ‰ค ๐ถ for some constant ๐ถ which dependspolynomially on๐‘€ and combining with (4.8) we obtain the desired bound. ๏ฟฝ

5. Approximations to white noise under the spectral gap assumption

The main goal of this section is to prove Proposition 1.8. Given an probability measure ใ€ˆยทใ€‰which satises Assumption 1.1, we prove that ใ€ˆยทใ€‰ is concentrated on Cโˆ’ 5

4โˆ’, thus, by Schaudertheory for the operator L, ๐‘ฃ := Lโˆ’1๐‘ƒb โˆˆ C 3

4โˆ’ ใ€ˆยทใ€‰-almost surely, and we construct ๐น as the๐ฟ๐‘

ใ€ˆยทใ€‰Cโˆ’ 3

4โˆ’-limit of the sequence {๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก }๐‘กโ†“0. This will allow us to lift ใ€ˆยทใ€‰ to a probability measureใ€ˆยทใ€‰li on Cโˆ’ 5

4โˆ’ ร— C 34โˆ’ ร— Cโˆ’ 3

4โˆ’ in a continuous way, i.e., given a sequence of probability measures{ใ€ˆยทใ€‰โ„“ }โ„“โ†“0 which satisfy Assumption 1.1 and converge weakly to a limit ใ€ˆยทใ€‰ as โ„“ โ†“ 0, then {ใ€ˆยทใ€‰liโ„“ }โ„“โ†“0converges weakly to ใ€ˆยทใ€‰li.

The proof of Proposition 1.8 is based on suitable estimates on the ๐‘-moments of multilinearexpressions in the corresponding stochastic objects. In the case of Gaussian approximations, it isenough to bound the second moments, since we can use Nelsonโ€™s hypercontractivity estimateto bound the ๐‘-moments in a nite Wiener chaos by the second moments for every ๐‘ > 2 (see[IO19, Lemmata 4 and 8]). On the other hand, for non-Gaussian approximations one has to ndalternative methods to estimate the ๐‘-moments. This has been achieved with great success in thelast few years under very mild assumptions on the random eld, see for example [HS17, CH16].In these works a direct computation of the ๐‘-moments is made through explicit formulas in termsof the cumulant functions of the random eld and the nal bounds are obtained by combinatorialarguments.

Here we are interested in approximations of white noise that satisfy the spectral gap inequality(1.7), uniformly in the approximation parameter โ„“ . This covers the Gaussian case, but it allows formore general random elds. The basic observation is that one can bound the ๐‘-moments directlyby estimating the derivative with respect to the noise, based on the following consequence of thespectral gap assumption (1.7).

Proposition 5.1. The spectral gap inequality (1.7) implies

โŸจ|๐บ (b) โˆ’ ใ€ˆ๐บ (b)ใ€‰|2๐‘

โŸฉ 12๐‘ โ‰ค ๐ถ (๐‘)

โŸจ ๐œ•๐œ•b๐บ (b) 2๐‘๐ฟ2

โŸฉ 12๐‘

, (5.1)

for every 1 โ‰ค ๐‘ < โˆž and every functional ๐บ on periodic Schwartz distributions which can be

approximated by cylindrical functionals with respect to the norm ใ€ˆ|๐บ (b) |2๐‘ใ€‰12๐‘ + ใ€ˆโ€– ๐œ•

๐œ•b๐บ (b)โ€–2๐‘

๐ฟ2ใ€‰

12๐‘,

where the constant ๐ถ (๐‘) > 0 depends only on ๐‘ .

Remark 5.2. As in Remark 1.3, using (5.1) we can extend b (๐œ™) for ๐œ™ โˆˆ ๐ฟ2(T2) as a centeredrandom variable in ๐ฟ2๐‘ใ€ˆยทใ€‰ , admissible in (5.1) for any 1 โ‰ค ๐‘ < โˆž.

Proof. The proof follows [JO20, Lemma 3.1]. Let ๐‘ > 1. We assume that๐บ is a cylindrical functionalon periodic Schwartz distributions. The general case follows by approximation. Without loss ofgenerality we can assume that ใ€ˆ๐บ (b)ใ€‰ = 0. For _ โˆˆ (0, 1] we consider the functional

๐น_ (b) := (๐บ (b)2 + _2)๐‘

2 .

Page 31: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 31

Noting that ๐œ•๐œ•b๐น_ (b) = ๐‘๐น_ (b)

๐‘โˆ’2๐‘ ๐บ (b) ๐œ•

๐œ•b๐บ (b) and using Hรถlderโ€™s inequality, (1.7) applied to ๐น_

yieldsโŸจ๏ฟฝ๏ฟฝ๐น_ (b) โˆ’ ใ€ˆ๐น_ (b)ใ€‰๏ฟฝ๏ฟฝ2โŸฉ โ‰ค ๐‘2

โŸจ๐น_ (b)

2(๐‘โˆ’2)๐‘ ๐บ (b)2

๐œ•๐œ•b๐บ (b) 2๐ฟ2

โŸฉโ‰ค ๐‘2

โŸจ๐น_ (b)

2(๐‘โˆ’1)๐‘

๐œ•๐œ•b๐บ (b) 2๐ฟ2

โŸฉโ‰ค ๐‘2

โŸจ๐น_ (b)2

โŸฉ ๐‘โˆ’1๐‘

โŸจ ๐œ•๐œ•b๐บ (b) 2๐‘๐ฟ2

โŸฉ 1๐‘

. (5.2)

By Cauchyโ€“Schwarz we see thatใ€ˆ๐น_ (b)ใ€‰2 โ‰ค

โŸจ(๐บ (b)2 + _2)๐‘โˆ’1

โŸฉ โŸจ๐บ (b)2 + _2

โŸฉ.

By Hรถlderโ€™s inequality and (1.7) applied to๐บ (b) (recall that ใ€ˆ๐บ (b)ใ€‰ = 0) the two terms on the righthand side can be estimated asโŸจ

(๐บ (b)2 + _2)๐‘โˆ’1โŸฉโ‰ค ใ€ˆ๐น_ (b)2ใ€‰

๐‘โˆ’1๐‘ ,โŸจ

๐บ (b)2 + _2โŸฉโ‰ค

โŸจ ๐œ•๐œ•b๐บ (b) 2๐ฟ2

โŸฉ+ _2 โ‰ค

โŸจ ๐œ•๐œ•b๐บ (b) 2๐‘๐ฟ2

โŸฉ 1๐‘

+ _2.

Combining with (5.2) we getโŸจ๐น_ (b)2

โŸฉ.

โŸจ|๐น_ (b) โˆ’ ใ€ˆ๐น_ (b)ใ€‰โ„“ |2

โŸฉ+ ใ€ˆ๐น_ (b)ใ€‰2

.๐‘โŸจ๐น_ (b)2

โŸฉ ๐‘โˆ’1๐‘

โŸจ ๐œ•๐œ•b๐บ (b) 2๐‘๐ฟ2

โŸฉ 1๐‘

+ _2ใ€ˆ๐น_ (b)2ใ€‰๐‘โˆ’1๐‘ .

By Youngโ€™s inequality we nally obtain thatโŸจ(๐บ (b)2 + _2)๐‘

โŸฉ=

โŸจ๐น_ (b)2

โŸฉ.๐‘

โŸจ ๐œ•๐œ•b๐บ (b) 2๐‘๐ฟ2

โŸฉ+ _2๐‘ .

The conclusion follows by the monotone convergence theorem as _ โ†˜ 0. ๏ฟฝ

Proposition 5.1 allows us to estimate the ๐‘-moments of multilinear expressions in b (shifted bytheir expectation) by the operator norm of (random) linear functionals on ๐ฟ2(T2) (in the case ofwhite noise this is the Cameronโ€“Martin space), after taking one derivative with respect to b . Toestimate the operator norm of these linear functionals we use the regularization properties ofLโˆ’1 in Sobolev spaces.

5.1. Estimates on b and ๐‘ฃ . In this section, we prove several stochastic estimates for b and๐‘ฃ := Lโˆ’1๐‘ƒb which are uniform in the class of probability measures satisfying Assumption 1.1. Asa corollary, we obtain that the law of (b, ๐‘ฃ) is concentrated on Cโˆ’ 5

4โˆ’Y ร— C 34โˆ’Y (see Corollary 5.4).

A similar result was proved in [IO19, Lemma 4] for ใ€ˆยทใ€‰ being the law of white noise. Here weconsider more general probability measures which are not necessarily Gaussian. Some of theresults of this section will be used in Section 5.2 below in the construction of ๐น .

We start with the following proposition.

Proposition 5.3. Let ใ€ˆยทใ€‰ satisfy Assumption 1.1 and let ๐‘ฃ := Lโˆ’1๐‘ƒb . For every 1 โ‰ค ๐‘ < โˆž,๐‘‡ โˆˆ (0, 1],and ๐‘ฆ โˆˆ T2 we have that

sup๐‘ฅ โˆˆT2

โŸจ|b๐‘‡ (๐‘ฅ) |๐‘

โŸฉ 1๐‘ โ‰ค ๐ถ

(๐‘‡

13)โˆ’ 5

4, (5.3)

sup๐‘ฅ โˆˆT2

โŸจ|๐‘ฃ (๐‘ฅ โˆ’ ๐‘ฆ) โˆ’ ๐‘ฃ (๐‘ฅ) |๐‘

โŸฉ 1๐‘ โ‰ค ๐ถ๐‘‘ (0, ๐‘ฆ) 3

4 , (5.4)

sup๐‘ฅ โˆˆT2

โŸจ| (๐œ•2๐‘…1๐‘ฃ)๐‘‡ (๐‘ฅ) |๐‘

โŸฉ 1๐‘ โ‰ค ๐ถ

(๐‘‡

13)โˆ’ 3

4, (5.5)

Page 32: Variational methods for a singular SPDE yielding the

32 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

where the constant ๐ถ depends only on ๐‘ .

Proof. The proof is based on a direct application of the spectral gap inequality (5.1). In thefollowing, for every ๐›ฟb โˆˆ ๐ฟ2(T2) we consider ๐›ฟ๐‘ฃ the unique solution of zero average in ๐‘ฅ1 ofL๐›ฟ๐‘ฃ = ๐‘ƒ๐›ฟb .Proof of (5.3): Let ๐‘ฅ0 โˆˆ T2 be xed. We consider the linear functional ๐บ (b) = b๐‘‡ (๐‘ฅ0). Sinceb๐‘‡ = b โˆ—๐œ“๐‘‡ , by Assumption 1.1 (i) we have that ใ€ˆb๐‘‡ (๐‘ฅ0)ใ€‰ = 0. Then, by (5.1) applied to ๐บ (which isa cylindrical functional), we get that for every ๐‘ž โ‰ฅ 1โŸจ

|b๐‘‡ (๐‘ฅ0) |2๐‘žโŸฉ 12๐‘ž .

โŸจ ๐œ•๐œ•b b๐‘‡ (๐‘ฅ0) 2๐‘ž๐ฟ2โŸฉ 1

2๐‘ž

.

It is easy to check that ๐œ•๐œ•bb๐‘‡ (๐‘ฅ0) = ฮจ๐‘‡ (๐‘ฅ0 โˆ’ ยท) โˆˆ ๐ฟ2(T2) (with ฮจ๐‘‡ the periodization of๐œ“๐‘‡ ), and by

Remark 1.11 ๐œ•๐œ•b b๐‘‡ (๐‘ฅ0) ๐ฟ2 = โ€–ฮจ๐‘‡ (๐‘ฅ0 โˆ’ ยท)โ€–๐ฟ2 = โ€–ฮจ๐‘‡ โ€–๐ฟ2 .(๐‘‡

13)โˆ’ 5

4,

which proves (5.3) for every 2 โ‰ค ๐‘ < โˆž, as the implicit constant above does not depend on ๐‘ฅ0.For ๐‘ โˆˆ [1, 2), the conclusion then follows by Jensenโ€™s inequality.Proof of (5.4): Let ๐‘ฅ0 โˆˆ T2 be xed. For every ๐‘ฆ โˆˆ T2, we consider the linear functional ๐บ (b) =๐‘ฃ (๐‘ฅ0 โˆ’ ๐‘ฆ) โˆ’ ๐‘ฃ (๐‘ฅ0) = b (ฮ“(๐‘ฅ0 โˆ’ ๐‘ฆ โˆ’ ยท)) โˆ’ b (ฮ“(๐‘ฅ0 โˆ’ ยท)), which is well-dened as a centered randomvariable in ๐ฟ2๐‘žใ€ˆยทใ€‰ for any ๐‘ž โ‰ฅ 1 by Remark 5.2 and (1.21) and is admissible in (5.1). Then

๐œ•

๐œ•b๐บ (b) : ๐›ฟb โ†ฆโ†’ ๐›ฟ๐‘ฃ (๐‘ฅ0 โˆ’ ๐‘ฆ) โˆ’ ๐›ฟ๐‘ฃ (๐‘ฅ0) .

By (D.1) we know that

|๐›ฟ๐‘ฃ (๐‘ฅ0 โˆ’ ๐‘ฆ) โˆ’ ๐›ฟ๐‘ฃ (๐‘ฅ0) | โ‰ค [๐›ฟ๐‘ฃ] 34๐‘‘ (0, ๐‘ฆ) 3

4 . ๐‘‘ (0, ๐‘ฆ) 34 โ€–๐›ฟb โ€–๐ฟ2,

which in turn implies that ๐œ•๐œ•b๐บ (b) ๐ฟ2. ๐‘‘ (0, ๐‘ฆ) 3

4 .

Thus, by (5.1) applied to ๐บ , we have for every ๐‘ž โ‰ฅ 1,โŸจ|๐‘ฃ (๐‘ฅ0 โˆ’ ๐‘ฆ) โˆ’ ๐‘ฃ (๐‘ฅ0) |2๐‘ž

โŸฉ 12๐‘ž . ๐‘‘ (0, ๐‘ฆ) 3

4 ,

which proves (5.4) for 1 โ‰ค ๐‘ < โˆž, since the implicit constant does not depend on ๐‘ฅ0. For ๐‘ โˆˆ [1, 2),the conclusion follows by Jensenโ€™s inequality.Proof of (5.5): Let ๐‘ฅ0 โˆˆ T2 be xed. We consider the linear functional๐บ (b) = (๐œ•2๐‘…1๐‘ฃ)๐‘‡ (๐‘ฅ0) (whichis cylindrical). We have that

๐œ•

๐œ•b๐บ (b) : ๐›ฟb โ†ฆโ†’ (๐œ•2๐‘…1๐›ฟ๐‘ฃ)๐‘‡ (๐‘ฅ0).

By Youngโ€™s inequality for convolution (see Remark 1.11), (D.3), and the fact that ๐‘…1 is bounded on๐ฟ

103 (T2), we get that

| (๐œ•2๐‘…1๐›ฟ๐‘ฃ)๐‘‡ (๐‘ฅ0) | .(๐‘‡

13)โˆ’ 3

4 โ€–๐œ•2๐‘…1๐›ฟ๐‘ฃ โ€–๐ฟ103.

(๐‘‡

13)โˆ’ 3

4 โ€–๐œ•2๐›ฟ๐‘ฃ โ€–๐ฟ103.

(๐‘‡

13)โˆ’ 3

4 โ€–๐›ฟb โ€–๐ฟ2,

yielding the estimate ๐œ•๐œ•b๐บ (b) ๐ฟ2.

(๐‘‡

13)โˆ’ 3

4.

We also note that ใ€ˆ๐บ (b)ใ€‰ = 0. Indeed, since ใ€ˆ๐‘ฃ (๐‘ฅ)ใ€‰ = ใ€ˆb (ฮ“(๐‘ฅ โˆ’ ยท))ใ€‰ = 0, for every ๐‘ฅ โˆˆ T2, we getthat ใ€ˆ(๐œ•2๐‘…1๐‘ฃ)๐‘‡ (๐‘ฅ0)ใ€‰ = 0. The conclusion then follows as above. ๏ฟฝ

Page 33: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 33

As a corollary of (5.3) and the Schauder theory for the operator L, we prove that the laws of band ๐‘ฃ are concentrated on Cโˆ’ 5

4โˆ’Y and C 34โˆ’Y .

Corollary 5.4. Let ใ€ˆยทใ€‰ satisfy Assumption 1.1. For every Y โˆˆ (0, 1100 ) and 1 โ‰ค ๐‘ < โˆž there holdsโŸจ

[b]๐‘โˆ’ 54โˆ’Y

โŸฉ 1๐‘

โ‰ค ๐ถ, (5.6)โŸจ[๐‘ฃ]๐‘3

4โˆ’Y

โŸฉ 1๐‘

โ‰ค ๐ถ, (5.7)

where the constant ๐ถ depends only on ๐‘ and Y. Moreover, ใ€ˆ๐‘ฃ (๐‘ฅ)ใ€‰ = 0 for every ๐‘ฅ โˆˆ T2.

Proof. The proof of (5.6) follows by (5.3) and Lemma C.2. To prove (5.7), we use Schauder theoryfor the operator L (see [IO19, Lemma 5]), which implies

[๐‘ฃ] 34โˆ’Y

= [Lโˆ’1๐‘ƒb] 34โˆ’Y. [๐‘ƒb]โˆ’ 5

4โˆ’Y. [b]โˆ’ 5

4โˆ’Y.

As in the proof of Proposition 5.3, ใ€ˆ๐‘ฃ (๐‘ฅ)ใ€‰ = ใ€ˆb (ฮ“(๐‘ฅ โˆ’ ยท))ใ€‰ = 0 for every ๐‘ฅ โˆˆ T2, by Remark 5.2 and(1.21). ๏ฟฝ

5.2. Estimates on ๐น . In this section we use the spectral gap inequality (5.1) and (5.4) and (5.5) toconstruct ๐น as the ๐ฟ๐‘ใ€ˆยทใ€‰C

โˆ’ 34โˆ’Y-limit of the sequence of random variables {๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก }๐‘กโ†“0. A similar

result was proved in [IO19, Lemma 8] for ใ€ˆยทใ€‰ being the law of white noise (but instead consideringapproximations {๐‘ฃโ„“๐œ•2๐‘…1๐‘ฃโ„“ }โ„“โ†“0, where ๐‘ฃโ„“ := Lโˆ’1๐‘ƒbโ„“ and bโ„“ := ๐œ™โ„“ โˆ— b , for a suitable mollier ๐œ™โ„“ )using Nelsonโ€™s hypercontractivity estimate. As in Section 5.1, our estimate holds for more generalprobability measures which are not necessarily Gaussian.

In what follows, we use the convolution-commutatord๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘  := ๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)2๐‘  โˆ’ (๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘  .

We will need the following lemma, based on (5.4) and (5.5).

Lemma 5.5. Let ใ€ˆยทใ€‰ satisfy Assumption 1.1 and for a periodic Schwartz distribution let b be ๐‘ฃ = Lโˆ’1๐‘ƒb .Then for any 1 โ‰ค ๐‘ < โˆž and ๐‘ , ๐‘† โˆˆ (0, 1] there holds

sup๐‘ฅ โˆˆT2

โŸจ| ( d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘† (๐‘ฅ) |๐‘

โŸฉ 1๐‘ โ‰ค ๐ถ

((๐‘†

13)โˆ’ 1

4(๐‘ 13) 14 +

(๐‘†

13)โˆ’ 3

4(๐‘ 13) 34), (5.8)

where the constant ๐ถ depends only on ๐‘ .

Proof. (5.8) is a consequence of the following two claims:Claim 1 For every ๐‘ , ๐‘† โˆˆ (0, 1] and ๐‘ž โ‰ฅ 1,

sup๐‘ฅ โˆˆT2

โŸจ ๐œ•๐œ•b (d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ 

)๐‘†(๐‘ฅ)

2๐‘ž๐ฟ2

โŸฉ 12๐‘ž

.(๐‘†

13)โˆ’ 1

4(๐‘ 13) 14 +

(๐‘†

13)โˆ’ 3

4(๐‘ 13) 34. (5.9)

Claim 2 For every ๐‘ , ๐‘† โˆˆ (0, 1] and ๐‘ฅ0 โˆˆ T2,ใ€ˆ(d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0)ใ€‰ = ใ€ˆ(๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)2๐‘ )๐‘† (๐‘ฅ0)ใ€‰ โˆ’ ใ€ˆ(๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘ +๐‘† (๐‘ฅ0)ใ€‰ = 0. (5.10)

Assuming that these claims hold, we may apply for xed ๐‘ฅ0 โˆˆ T2 the ๐ฟ๐‘-version (5.1) of thespectral gap inequality to the functional ๐บ : b โ†ฆโ†’ (d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0), which yields (5.8) byClaim 1 and Claim 2 for 2 โ‰ค ๐‘ < โˆž. For ๐‘ โˆˆ [1, 2) (5.8) follows by Jensenโ€™s inequality.

It is easy to see that

๐บ (b) = (d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0) =โˆฌ

๐‘˜ (๐‘ง, ๐‘ง โ€ฒ)b (๐‘ง)b (๐‘ง โ€ฒ) d๐‘ง d๐‘ง โ€ฒ,

where ๐‘˜ (๐‘ง, ๐‘ง โ€ฒ) is smooth in both variables andโˆฌ๐‘˜ (๐‘ง, ๐‘ง โ€ฒ)2 d๐‘ง d๐‘ง โ€ฒ < โˆž. A straightforward calcula-

tion using the Cauchyโ€“Schwarz inequality with respect to ใ€ˆยทใ€‰ and (5.4) and (5.5) we also get that

Page 34: Variational methods for a singular SPDE yielding the

34 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

ใ€ˆ|๐บ (b) |ใ€‰ . 1. Hence, we may apply Lemma F.1 which implies that ๐บ is admissible in the spectralgap inequality (5.1).Proof of Claim 1: Let ๐‘ฅ0 โˆˆ T2 be xed. We rst notice that for every ๐‘ , ๐‘† โˆˆ (0, 1], the derivative ofthe quadratic functional ๐บ is given by

๐œ•

๐œ•b๐บ (b) : ๐›ฟb โ†ฆโ†’ (d๐›ฟ๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0) + (d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐›ฟ๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0),

where ๐›ฟ๐‘ฃ is the unique solution of zero average in ๐‘ฅ1 to L๐›ฟ๐‘ฃ = ๐‘ƒ๐›ฟb for ๐›ฟb โˆˆ ๐ฟ2(T2).Step 1: We rst show that

supโ€–๐›ฟb โ€–

๐ฟ2 โ‰ค1| ( d๐›ฟ๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0) | .

โˆซR2

|๐‘ฆ1 | |๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ยท)(๐œ•2๐‘…1๐‘ฃ)๐‘  (ยท โˆ’ ๐‘ฆ)โ€–๐ฟ109d๐‘ฆ

+โˆซR2

|๐‘ฆ2 | |๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ยท)(๐œ•2๐‘…1๐‘ฃ)๐‘  (ยท โˆ’ ๐‘ฆ)โ€–๐ฟ107d๐‘ฆ, (5.11)

supโ€–๐›ฟb โ€–

๐ฟ2 โ‰ค1| ( d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐›ฟ๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0) | .

โˆซR2

|๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ยท) (๐‘ฃ (ยท โˆ’ ๐‘ฆ) โˆ’ ๐‘ฃ) โ€–๐ฟ107d๐‘ฆ. (5.12)

Step 1a (Proof of (5.11)): By Fubini and the mean value theorem we have

(d๐›ฟ๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0)

=

โˆซR2๐œ“๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ)

โˆซR2๐œ“๐‘  (๐‘ฆ) (๐›ฟ๐‘ฃ (๐‘ฅ) โˆ’ ๐›ฟ๐‘ฃ (๐‘ฅ โˆ’ ๐‘ฆ)) (๐œ•2๐‘…1๐‘ฃ)๐‘  (๐‘ฅ โˆ’ ๐‘ฆ) d๐‘ฆ d๐‘ฅ

=

โˆซR2๐œ“๐‘  (๐‘ฆ)

โˆซT2ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ)

โˆซ 1

0(๐‘ฆ1๐œ•1๐›ฟ๐‘ฃ (๐‘ฅ โˆ’ ๐‘ก๐‘ฆ) + ๐‘ฆ2๐œ•2๐›ฟ๐‘ฃ (๐‘ฅ โˆ’ ๐‘ก๐‘ฆ)) d๐‘ก (๐œ•2๐‘…1๐‘ฃ)๐‘  (๐‘ฅ โˆ’ ๐‘ฆ) d๐‘ฅ d๐‘ฆ,

where ฮจ๐‘† is the periodization of๐œ“๐‘† . By Hรถlderโ€™s and Minkowskiโ€™s inequalities, as well as transla-tion invariance, it follows that

| ( d๐›ฟ๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0) |

โ‰คโˆซR2

|๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ) (๐œ•2๐‘…1๐‘ฃ)๐‘  (๐‘ฅ โˆ’ ๐‘ฆ)โ€–๐ฟ109๐‘ฅ

โˆซ 1

0๐‘ฆ1๐œ•1๐›ฟ๐‘ฃ (๐‘ฅ โˆ’ ๐‘ก๐‘ฆ) d๐‘ก

๐ฟ10๐‘ฅ

d๐‘ฆ

+โˆซR2

|๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ) (๐œ•2๐‘…1๐‘ฃ)๐‘  (๐‘ฅ โˆ’ ๐‘ฆ)โ€–๐ฟ107๐‘ฅ

โˆซ 1

0๐‘ฆ2๐œ•2๐›ฟ๐‘ฃ (๐‘ฅ โˆ’ ๐‘ก๐‘ฆ) d๐‘ก

๐ฟ103๐‘ฅ

d๐‘ฆ

.

โˆซR2

|๐‘ฆ1 | |๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ) (๐œ•2๐‘…1๐‘ฃ)๐‘  (๐‘ฅ โˆ’ ๐‘ฆ)โ€–๐ฟ109๐‘ฅ

d๐‘ฆ โ€–๐œ•1๐›ฟ๐‘ฃ โ€–๐ฟ10

+โˆซR2

|๐‘ฆ2 | |๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ) (๐œ•2๐‘…1๐‘ฃ)๐‘  (๐‘ฅ โˆ’ ๐‘ฆ)โ€–๐ฟ107๐‘ฅ

d๐‘ฆ โ€–๐œ•2๐›ฟ๐‘ฃ โ€–๐ฟ103.

By (D.2) and (D.3) we have โ€–๐œ•1๐›ฟ๐‘ฃ โ€–๐ฟ10, โ€–๐œ•2๐›ฟ๐‘ฃ โ€–๐ฟ103. โ€–๐›ฟb โ€–๐ฟ2 , hence (5.11) follows after taking the

supremum over โ€–๐›ฟb โ€–๐ฟ2 โ‰ค 1.Step 1b (Proof of (5.12)): As in Step 1A, we have

(d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐›ฟ๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0)

=

โˆซR2๐œ“๐‘  (๐‘ฆ)

โˆซT2ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ) (๐‘ฃ (๐‘ฅ) โˆ’ ๐‘ฃ (๐‘ฅ โˆ’ ๐‘ฆ)) (๐œ•2๐‘…1๐›ฟ๐‘ฃ)๐‘  (๐‘ฅ โˆ’ ๐‘ฆ) d๐‘ฅ d๐‘ฆ,

where ฮจ๐‘† is the periodization of๐œ“๐‘  . By Hรถlderโ€™s inequality, translation invariance, and the factthat ๐‘…1 is bounded on ๐ฟ 10

3 (T2), we obtain that

| ( d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐›ฟ๐‘ฃ)๐‘ )๐‘† (๐‘ฅ0) |

.

โˆซR2

|๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ) (๐‘ฃ (๐‘ฅ) โˆ’ ๐‘ฃ (๐‘ฅ โˆ’ ๐‘ฆ)) โ€–๐ฟ107๐‘ฅ

โ€–(๐œ•2๐‘…1๐›ฟ๐‘ฃ)๐‘  (๐‘ฅ โˆ’ ๐‘ฆ)โ€–๐ฟ103๐‘ฅ

d๐‘ฆ

Page 35: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 35

.

โˆซR2

|๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ) (๐‘ฃ (๐‘ฅ) โˆ’ ๐‘ฃ (๐‘ฅ โˆ’ ๐‘ฆ)) โ€–๐ฟ107๐‘ฅ

d๐‘ฆ โ€–๐œ•2๐›ฟ๐‘ฃ โ€–๐ฟ103.

By (D.3), โ€–๐œ•2๐›ฟ๐‘ฃ โ€–๐ฟ103. โ€–๐›ฟb โ€–๐ฟ2 , which gives (5.12) after taking the supremum over โ€–๐›ฟb โ€–๐ฟ2 โ‰ค 1.

Step 2: For any ๐‘ž โ‰ฅ 1 and ๐‘ฅ0 โˆˆ T2, by Step 1 and Minkowskiโ€™s inequality (since 2๐‘ž โ‰ฅ max{ 107 ,109 }),

we get thatโŸจ ๐œ•๐œ•b๐บ (b) 2๐‘ž๐ฟ2

โŸฉ 12๐‘ž

.

โˆซR2

|๐‘ฆ1 | |๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ)ใ€ˆ|(๐œ•2๐‘…1๐‘ฃ)๐‘  (๐‘ฅ โˆ’ ๐‘ฆ) |2๐‘žใ€‰12๐‘ž โ€–

๐ฟ109๐‘ฅ

d๐‘ฆ

+โˆซR2

|๐‘ฆ2 | |๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ)ใ€ˆ|(๐œ•2๐‘…1๐‘ฃ)๐‘  (๐‘ฅ โˆ’ ๐‘ฆ) |2๐‘žใ€‰12๐‘ž โ€–

๐ฟ107๐‘ฅ

d๐‘ฆ

+โˆซR2

|๐œ“๐‘  (๐‘ฆ) |โ€–ฮจ๐‘† (๐‘ฅ0 โˆ’ ๐‘ฅ)ใ€ˆ|๐‘ฃ (๐‘ฅ) โˆ’ ๐‘ฃ (๐‘ฅ โˆ’ ๐‘ฆ) |2๐‘žใ€‰12๐‘ž โ€–

๐ฟ107๐‘ฅ

d๐‘ฆ.

By (5.4) and (5.5) this implies the boundโŸจ ๐œ•๐œ•b๐บ (b) 2๐‘ž๐ฟ2

โŸฉ 12๐‘ž

. โ€–ฮจ๐‘† โ€–๐ฟ109

(๐‘ 13)โˆ’ 3

4โˆซR2

|๐‘ฆ1 | |๐œ“๐‘  (๐‘ฆ) | d๐‘ฆ + โ€–ฮจ๐‘† โ€–๐ฟ107

(๐‘ 13)โˆ’ 3

4โˆซR2

|๐‘ฆ2 | |๐œ“๐‘  (๐‘ฆ) | d๐‘ฆ

+ โ€–ฮจ๐‘† โ€–๐ฟ107

โˆซR2๐‘‘ (0, ๐‘ฆ) 3

4 |๐œ“๐‘  (๐‘ฆ) | d๐‘ฆ

.(๐‘†

13)โˆ’ 1

4(๐‘ 13)1โˆ’ 3

4 +(๐‘†

13)โˆ’ 3

4(๐‘ 13) 32โˆ’

34 +

(๐‘†

13)โˆ’ 3

4(๐‘ 13) 34,

which proves Claim 2, since the implicit constant does not depend on ๐‘ฅ0.Proof of Claim 2: By stationarity we know that ใ€ˆ(๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)2๐‘ )๐‘† (๐‘ฅ0)ใ€‰ = ใ€ˆ(๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)2๐‘ )๐‘† (0)ใ€‰. If wedenote by ๐‘ฃ the solution to (1.3) with b replaced by bฬƒ (๐‘ฅ) = b (โˆ’๐‘ฅ1, ๐‘ฅ2), by the symmetry of ฮ“ (see(1.20)) and the fact that b and bฬƒ have the same law (see Assumption 1.1 (iii)), we know that ๐‘ฃ = ๐‘ฃ inlaw, hence ใ€ˆ(๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)2๐‘ )๐‘† (0)ใ€‰ = ใ€ˆ(๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)2๐‘ )๐‘† (0)ใ€‰. By the symmetry of๐œ“๐‘ก for ๐‘ก โˆˆ (0, 1] and thefact that ๐‘…1๐‘ฃ (๐‘ฅ) = โˆ’๐‘…1๐‘ฃ (โˆ’๐‘ฅ1, ๐‘ฅ2), we get ใ€ˆ(๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)2๐‘ )๐‘† (0)ใ€‰ = โˆ’ใ€ˆ(๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)2๐‘ )๐‘† (0)ใ€‰, which impliesthat ใ€ˆ(๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)2๐‘ )๐‘† (0)ใ€‰ = 0. Similarly, ใ€ˆ(๐‘ฃ (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘ +๐‘† (0)ใ€‰ = 0. ๏ฟฝ

We then have the following proposition.

Proposition 5.6. Under the assumptions of Lemma 5.5, for every Y โˆˆ (0, 1100 ), 1 โ‰ค ๐‘ < โˆž, and

0 < ๐œ โ‰ค ๐‘ก โ‰ค 1 dyadically related there holdsโŸจ[(๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ )๐‘กโˆ’๐œ โˆ’ ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’Y

โŸฉ 1๐‘

โ‰ค(๐‘ก13) 14. (5.13)

Furthermore, the following bound holds,

sup๐‘ก โˆˆ(0,1]

โŸจ[๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’Y

โŸฉ 1๐‘

โ‰ค ๐ถ. (5.14)

In the above estimates, the constant ๐ถ depends only on Y and ๐‘ .

Proof. Step 1: We rst prove that for every 1 โ‰ค ๐‘ < โˆž, 0 < ๐œ โ‰ค ๐‘ก โ‰ค 1 (note that by assumption๐œ = ๐‘ก

2๐‘› for some ๐‘› โ‰ฅ 0) and ๐‘‡ โˆˆ (0, 1] there holds,

sup๐‘ฅ โˆˆT2

โŸจ| ( d๐‘ฃ, (ยท)๐‘กโˆ’๐œ e (๐œ•2๐‘…1๐‘ฃ)๐œ )๐‘‡ (๐‘ฅ) |๐‘

โŸฉ 1๐‘ โ‰ค ๐ถ

((๐‘‡

13)โˆ’ 1

4(๐‘ก13) 14 +

(๐‘‡

13)โˆ’ 3

4(๐‘ก13) 34). (5.15)

Together with Lemma C.2 this implies (5.13). We will use the following telescopic sum identity,

d๐‘ฃ, (ยท)๐‘กโˆ’๐œ e (๐œ•2๐‘…1๐‘ฃ)๐œ =โˆ‘๏ธ

0โ‰ค๐‘˜โ‰ค๐‘›โˆ’1๐‘ = ๐‘ก

22๐‘˜

(d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘กโˆ’2๐‘  .

Page 36: Variational methods for a singular SPDE yielding the

36 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

Combined with Lemma 5.5 we obtain thatโŸจ| ( d๐‘ฃ, (ยท)๐‘กโˆ’๐œ e (๐œ•2๐‘…1๐‘ฃ)๐‘ก )๐‘‡ (๐‘ฅ) |๐‘

โŸฉ 1๐‘ .

โˆ‘๏ธ0โ‰ค๐‘˜โ‰ค๐‘›โˆ’1๐‘ = ๐‘ก

22๐‘˜

โŸจ| ( d๐‘ฃ, (ยท)๐‘ e (๐œ•2๐‘…1๐‘ฃ)๐‘ )๐‘‡+๐‘กโˆ’2๐‘  (๐‘ฅ) |๐‘

โŸฉ 1๐‘

.โˆ‘๏ธ

0โ‰ค๐‘˜โ‰ค๐‘›โˆ’1๐‘ = ๐‘ก

22๐‘˜

(((๐‘‡ + ๐‘ก โˆ’ 2๐‘ ) 1

3)โˆ’ 1

4(๐‘ 13) 14 +

((๐‘‡ + ๐‘ก โˆ’ 2๐‘ ) 1

3)โˆ’ 3

4(๐‘ 13) 34)

.(๐‘‡

13)โˆ’ 1

4(๐‘ก13) 14

โˆ‘๏ธ0โ‰ค๐‘˜โ‰ค๐‘›โˆ’1

12๐‘˜+1

12+

(๐‘‡

13)โˆ’ 3

4(๐‘ก13) 34

โˆ‘๏ธ0โ‰ค๐‘˜โ‰ค๐‘›โˆ’1

12๐‘˜+1

4

.(๐‘‡

13)โˆ’ 1

4(๐‘ก13) 14 +

(๐‘‡

13)โˆ’ 3

4(๐‘ก13) 34,

which proves the desired claim.Step 2: We now prove that for every Y โˆˆ (0, 1

100 ), 1 โ‰ค ๐‘ < โˆž and ๐‘‡ โˆˆ (0, 1],

sup๐‘ก โˆˆ(0,1]

sup๐‘ฅ โˆˆT2

โŸจ| (๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )๐‘‡ (๐‘ฅ) |๐‘

โŸฉ 1๐‘ .Y,๐‘

(๐‘‡

13)โˆ’ 3

4โˆ’Y, (5.16)

which together with Lemma C.2 implies (5.14).We rst assume that ๐‘ก โˆˆ [๐‘‡2 , 1]. Then, by Denition 1.13 of negative Hรถlder norms and

(A.14) we know that [๐œ•2๐‘…1๐‘ฃ]โˆ’ 34โˆ’2Y. [๐‘…1๐‘ฃ] 3

4โˆ’2Y. [๐‘ฃ] 3

4โˆ’Y. Combined with (A.1) and the fact that

โ€–๐‘ฃ โ€–๐ฟโˆž . [๐‘ฃ] 34โˆ’Y

, we have that

| (๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )๐‘‡ (๐‘ฅ) | โ‰ค โ€–๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก โ€–๐ฟโˆž โ‰ค โ€–๐‘ฃ โ€–๐ฟโˆž โ€–๐œ•2๐‘…1๐‘ฃ๐‘ก โ€–๐ฟโˆž

.(๐‘ก13)โˆ’ 3

4โˆ’2Y [๐‘ฃ]234โˆ’Y.

(๐‘‡

13)โˆ’ 3

4โˆ’2Y [๐‘ฃ]234โˆ’Y. (5.17)

By (5.7), this implies the estimateโŸจ| (๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )๐‘‡ (๐‘ฅ) |๐‘

โŸฉ 1๐‘

โ„“.

(๐‘‡

13)โˆ’ 3

4โˆ’2YโŸจ[๐‘ฃ]2๐‘3

4โˆ’Y

โŸฉ 1๐‘

โ„“

.(๐‘‡

13)โˆ’ 3

4โˆ’2Y, (5.18)

for every ๐‘ก โˆˆ [๐‘‡2 , 1], uniformly in ๐‘ฅ โˆˆ T2 and โ„“ โˆˆ (0, 1].We now assume that ๐‘ก โˆˆ (0, ๐‘‡2 ]. Then, there exists๐‘‡โˆ— โˆˆ (๐‘‡4 ,

๐‘‡2 ] such that ๐‘ก = ๐‘‡โˆ—

2๐‘› , for some ๐‘› โ‰ฅ 1.Using the semigroup property, we write

(๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )๐‘‡ =((๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )๐‘‡โˆ—โˆ’๐‘ก โˆ’ ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘‡โˆ—

)๐‘‡โˆ’๐‘‡โˆ—+๐‘ก +

(๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘‡โˆ—

)๐‘‡โˆ’๐‘‡โˆ—+๐‘ก .

By (5.13), the rst term can be estimated asโŸจ|((๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )๐‘‡โˆ—โˆ’๐‘ก โˆ’ ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘‡โˆ—

)๐‘‡โˆ’๐‘‡โˆ—+๐‘ก (๐‘ฅ) |

๐‘โŸฉ 1๐‘

.((๐‘‡ โˆ’๐‘‡โˆ— + ๐‘ก)

13)โˆ’ 3

4(๐‘‡

13โˆ—

) 34+

((๐‘‡ โˆ’๐‘‡โˆ— + ๐‘ก)

13)โˆ’ 1

4(๐‘‡

13โˆ—

) 14. 1,

where we also used that ๐‘‡โˆ— โ‰ค ๐‘‡2 โ‰ค ๐‘‡ โˆ’ ๐‘‡โˆ— + ๐‘ก . For the second term, noting that ๐‘‡โˆ— > ๐‘‡

4 andproceeding as in (5.17) and (5.18), we obtain the boundโŸจ

|(๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘‡โˆ—

)๐‘‡โˆ’๐‘‡โˆ—+๐‘ก (๐‘ฅ) |

๐‘โŸฉ 1๐‘

.(๐‘‡

13โˆ—

)โˆ’ 34โˆ’2Y

โŸจ[๐‘ฃ]2๐‘3

4โˆ’Y

โŸฉ 1๐‘

โ„“

.(๐‘‡

13)โˆ’ 3

4โˆ’2Y.

Hence, we also proved that โŸจ| (๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )๐‘‡ (๐‘ฅ) |๐‘

โŸฉ 1๐‘ . 1 +

(๐‘‡

13)โˆ’ 3

4โˆ’2Y, (5.19)

for ๐‘ก โˆˆ (0, ๐‘‡4 ], uniformly in ๐‘ฅ โˆˆ T2.Combining (5.18) and (5.19) gives (5.16) upon relabelling Y. ๏ฟฝ

Page 37: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 37

As a corollary we obtain,

Corollary 5.7. Let ใ€ˆยทใ€‰ satisfy Assumption 1.1. There exists a unique centered and stationary random

variable b โ†ฆโ†’ ๐น (b) such that for every Y โˆˆ (0, 1100 ) and 1 โ‰ค ๐‘ < โˆž,

lim๐‘ก=2โˆ’๐‘›โ†“0

โŸจ[๐น โˆ’ ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’Y

โŸฉ 1๐‘

= 0, (5.20)

and the convergence is uniform in the class of probability measures satisfying Assumption 1.1.

Proof. We prove that for every Y โˆˆ (0, 1], 1 โ‰ค ๐‘ < โˆž and 0 โ‰ค ๐œ โ‰ค ๐‘ก โ‰ค 1 dyadic,

lim๐‘ก,๐œโ†“0

โŸจ[๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ โˆ’ ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’Y

โŸฉ 1๐‘

= 0. (5.21)

By the triangle inequality we have thatโŸจ[๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ โˆ’ ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’Y

โŸฉ 1๐‘

โ‰คโŸจ[๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ โˆ’ (๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ )๐‘กโˆ’๐œ ]๐‘โˆ’ 3

4โˆ’Y

โŸฉ 1๐‘

+โŸจ[(๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ )๐‘กโˆ’๐œ โˆ’ ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’Y

โŸฉ 1๐‘

.

To estimate the rst term we use Propositions A.7 and (5.14) which imply thatโŸจ[๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ โˆ’ (๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ )๐‘กโˆ’๐œ ]๐‘โˆ’ 3

4โˆ’Y

โŸฉ 1๐‘

.((๐‘ก โˆ’ ๐œ) 1

3) Y

2โŸจ[๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ ]๐‘โˆ’ 3

4โˆ’Y2

โŸฉ 1๐‘

.((๐‘ก โˆ’ ๐œ) 1

3) Y

2.

Using (5.13), the second term is estimated asโŸจ[(๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ )๐‘กโˆ’๐œ โˆ’ ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’Y

โŸฉ 1๐‘

.(๐‘ก13) 14.

Hence, we have proved thatโŸจ[๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐œ โˆ’ ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]๐‘โˆ’ 3

4โˆ’Y

โŸฉ 1๐‘

.((๐‘ก โˆ’ ๐œ) 1

3) Y

2 +(๐‘ก13) 14,

which implies (5.21) after taking ๐œ, ๐‘ก โ†“ 0. This implies that the sequence {๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก }๐‘ก โˆˆ(0,1] is Cauchyin ๐ฟ๐‘ใ€ˆยทใ€‰C

โˆ’ 34โˆ’Y , hence it converges to a limit ๐น in ๐ฟ๐‘ใ€ˆยทใ€‰C

โˆ’ 34โˆ’Y . The fact that the convergence is uniform

in the class of probability measures satisfying Assumption 1.1 follows since our estimates dependon ใ€ˆยทใ€‰ only through the spectral gap inequality (1.7).

Similarly to the proof of Lemma 5.5, we can show that ใ€ˆ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก (๐‘ฅ)ใ€‰ = 0 for every ๐‘ฅ โˆˆ T2 and๐‘ก โˆˆ (0, 1]. Hence, as a limit in ๐ฟ๐‘ใ€ˆยทใ€‰C

โˆ’ 34โˆ’Y , the fact that ๐น is centered and stationary follows from

the corresponding properties of ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก . ๏ฟฝ

5.3. Proof of Proposition 1.8.

Proof. Let ใ€ˆยทใ€‰li be given by ใ€ˆ๐บ (b, ๐‘ฃ, ๐น )ใ€‰li := ใ€ˆ๐บ (b, ๐‘ฃ, ๐น )ใ€‰ for every bounded and continuous๐บ : Cโˆ’ 5

4โˆ’Y ร— C 34โˆ’Y ร— Cโˆ’ 3

4โˆ’Y โ†’ R, with the convention that under ใ€ˆยทใ€‰, ๐‘ฃ = Lโˆ’1๐‘ƒb and ๐น is given by(5.20) (under ใ€ˆยทใ€‰li we think of (b, ๐‘ฃ, ๐น ) as a dummy variable in T := Cโˆ’ 5

4โˆ’Y ร— C 34โˆ’Y ร— Cโˆ’ 3

4โˆ’Y ). Thefact that ใ€ˆยทใ€‰li denes a probability measure on T is immediate by Corollaries 5.4 and 5.7.

Statements (i) and (ii) are immediate by the construction of ใ€ˆยทใ€‰li. The rst part of statement(iii), that is, (1.9), is immediate by Corollary 5.7. For the second part, note that in the case when b issmooth ใ€ˆยทใ€‰-almost surely, the product ๐‘ฃ๐œ•2๐‘…1๐‘ฃ makes sense ใ€ˆยทใ€‰-almost surely and we also have that๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก โ†’ ๐‘ฃ๐œ•2๐‘…1๐‘ฃ ใ€ˆยทใ€‰-almost surely. By (5.20) we know that ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก โ†’ ๐น in Cโˆ’ 3

4โˆ’Y ใ€ˆยทใ€‰-almostsurely along a subsequence. Hence, we should have that ๐น = ๐‘ฃ๐œ•2๐‘…1๐‘ฃ ใ€ˆยทใ€‰-almost surely.

It remains to prove the continuity statement (iv). Assume that {ใ€ˆยทใ€‰โ„“ }โ„“โ†“0 converges weakly toใ€ˆยทใ€‰ and consider {ใ€ˆยทใ€‰liโ„“ }โ„“โ†“0 and ใ€ˆยทใ€‰li.

Page 38: Variational methods for a singular SPDE yielding the

38 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

Step 1: We prove that weak convergence of {ใ€ˆยทใ€‰โ„“ }โ„“โ†“0 to ใ€ˆยทใ€‰ in the Schwartz topology implies weakconvergence in Cโˆ’ 5

4โˆ’Y . Indeed, let ๐บ : Cโˆ’ 54โˆ’Y โ†’ R be bounded and continuous. Then, we can

writeใ€ˆ๐บ (b)ใ€‰โ„“ โˆ’ ใ€ˆ๐บ (b)ใ€‰ = (ใ€ˆ๐บ (b)ใ€‰โ„“ โˆ’ ใ€ˆ๐บ (b๐‘ก )ใ€‰โ„“ ) + (ใ€ˆ๐บ (b๐‘ก )ใ€‰โ„“ โˆ’ ใ€ˆ๐บ (b๐‘ก )ใ€‰) + (ใ€ˆ๐บ (b๐‘ก )ใ€‰ โˆ’ ใ€ˆ๐บ (b)ใ€‰).

To treat the rst term, we use (A.16) and (5.6) (which holds uniform in the class of probabilitymeasures satisfying Assumption 1.1) yielding for every 1 โ‰ค ๐‘ < โˆž,

supโ„“โˆˆ(0,1]

ใ€ˆ[b โˆ’ b๐‘ก ]๐‘โˆ’ 54โˆ’Y

ใ€‰1๐‘

โ„“. ๐‘ก

Y6 supโ„“โˆˆ(0,1]

ใ€ˆ[b]๐‘โˆ’ 54โˆ’

Y2ใ€‰

1๐‘

โ„“. ๐‘ก

Y6 . (5.22)

For ๐›ฟ > 0 which we x below, we writeใ€ˆ๐บ (b)ใ€‰โ„“ โˆ’ ใ€ˆ๐บ (b๐‘ก )ใ€‰โ„“ = ใ€ˆ(๐บ (b) โˆ’๐บ (b๐‘ก ))1{ [bโˆ’b๐‘ก ]โˆ’ 5

4 โˆ’Y<๐›ฟ }ใ€‰โ„“ + ใ€ˆ(๐บ (b) โˆ’๐บ (b๐‘ก ))1{ [bโˆ’b๐‘ก ]โˆ’ 5

4 โˆ’Yโ‰ฅ๐›ฟ }ใ€‰โ„“ .

For [ > 0, by the continuity of ๐บ we can choose ๐›ฟ suciently small such that

|ใ€ˆ(๐บ (b) โˆ’๐บ (b๐‘ก ))1{ [bโˆ’b๐‘ก ]โˆ’ 54 โˆ’Y

<๐›ฟ }ใ€‰โ„“ | โ‰ค[

4uniformly in โ„“ โˆˆ (0, 1]. By (5.22), the boundedness of ๐บ , and Chebyshevโ€™s inequality, we canchoose ๐‘ก โˆˆ (0, 1] suciently small such that

|ใ€ˆ(๐บ (b) โˆ’๐บ (b๐‘ก ))1{ [bโˆ’b๐‘ก ]โˆ’ 54 โˆ’Y

โ‰ค๐›ฟ }ใ€‰โ„“ | โ‰ค โ€–๐บ โ€–๐ฟโˆž

(Cโˆ’ 5

4 โˆ’Y) supโ„“โˆˆ(0,1]

ใ€ˆ1{ [bโˆ’b๐‘ก ]โˆ’ 54 โˆ’Y

โ‰ฅ๐›ฟ }ใ€‰โ„“ โ‰ค[

4 .

Hence, we obtain that for ๐‘ก suciently small

supโ„“โˆˆ(0,1]

|ใ€ˆ๐บ (b)ใ€‰โ„“ โˆ’ ใ€ˆ๐บ (b๐‘ก )ใ€‰โ„“ | โ‰ค[

2 .

Similarly, we can show that for ๐‘ก suciently small

|ใ€ˆ๐บ (b)ใ€‰ โˆ’ ใ€ˆ๐บ (b๐‘ก )ใ€‰| โ‰ค[

2 .

Since {ใ€ˆยทใ€‰โ„“ }โ„“โ†“0 converges to ใ€ˆยทใ€‰ in the Schwartz topology, for every ๐‘ก โˆˆ (0, 1] we know thatใ€ˆ๐บ (b๐‘ก )ใ€‰โ„“ โ†’ ใ€ˆ๐บ (b๐‘ก )ใ€‰ as โ„“ โ†“ 0. In total, we have that

lim supโ„“โ†“0

|ใ€ˆ๐บ (b)ใ€‰โ„“ โˆ’ ใ€ˆ๐บ (b)ใ€‰| โ‰ค [,

which proves that ใ€ˆ๐บ (b)ใ€‰โ„“ โ†’ ใ€ˆ๐บ (b)ใ€‰ as โ„“ โ†“ 0 since [ is arbitrary.Step 2: We now prove that {ใ€ˆยทใ€‰liโ„“ }โ„“โ†“0 converges weakly to ใ€ˆยทใ€‰li in T . The argument is similar inspirit to Step 1. Let ๐บ : T โ†’ R be a bounded continuous function. Then, we write

ใ€ˆ๐บ (b, ๐‘ฃ, ๐น )ใ€‰liโ„“ โˆ’ ใ€ˆ๐บ (b, ๐‘ฃ, ๐น )ใ€‰li = ใ€ˆ๐บ (b, ๐‘ฃ, ๐น ) โˆ’๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )ใ€‰โ„“+ ใ€ˆ๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ) โˆ’๐บ (b, ๐‘ฃ, ๐น )ใ€‰+ ใ€ˆ๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )ใ€‰โ„“ โˆ’ ใ€ˆ๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )ใ€‰.

To estimate the rst term, for ๐›ฟ > 0 to be xed below, we use the decompositionใ€ˆ๐บ (b, ๐‘ฃ, ๐น ) โˆ’๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )ใ€‰โ„“ = ใ€ˆ(๐บ (b, ๐‘ฃ, ๐น ) โˆ’๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ))1{ [๐นโˆ’๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]โˆ’ 3

4 โˆ’Y<๐›ฟ }ใ€‰โ„“

+ ใ€ˆ(๐บ (b, ๐‘ฃ, ๐น ) โˆ’๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ))1{ [๐นโˆ’๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]โˆ’ 34 โˆ’Y

โ‰ฅ๐›ฟ }ใ€‰โ„“ .

For [ > 0, by the continuity of ๐บ we can choose ๐›ฟ > 0 suciently small such that

|ใ€ˆ(๐บ (b, ๐‘ฃ, ๐น ) โˆ’๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ))1{ [๐นโˆ’๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]โˆ’ 34 โˆ’Y

<๐›ฟ }ใ€‰โ„“ | โ‰ค[

4uniformly in โ„“ โˆˆ (0, 1]. By (5.20), the boundedness of ๐บ , and Chebyshevโ€™s inequality, we canchoose ๐‘ก suciently small such that

supโ„“โˆˆ(0,1]

ใ€ˆ(๐บ (b, ๐‘ฃ, ๐น ) โˆ’๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ))1{ [๐นโˆ’๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]โˆ’ 34 โˆ’Y

โ‰ฅ๐›ฟ }ใ€‰โ„“

Page 39: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 39

โ‰ค โ€–๐บ โ€–๐ฟโˆž

(Cโˆ’ 5

4 โˆ’Yร—C34 โˆ’Yร—Cโˆ’ 3

4 โˆ’Y) supโ„“โˆˆ(0,1]

ใ€ˆ1{ [๐นโˆ’๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก ]โˆ’ 34 โˆ’Y

โ‰ฅ๐›ฟ }ใ€‰โ„“ โ‰ค[

4 .

Hence, we have proved that for ๐‘ก suciently small

supโ„“โˆˆ(0,1]

ใ€ˆ๐บ (b, ๐‘ฃ, ๐น ) โˆ’๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )ใ€‰โ„“ โ‰ค[

2 .

In a similar way we can show that for ๐‘ก suciently small

ใ€ˆ๐บ (b, ๐‘ฃ, ๐น ) โˆ’๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )ใ€‰ โ‰ค[

2 .

Since ใ€ˆยทใ€‰โ„“ โ†’ ใ€ˆยทใ€‰ weakly as โ„“ โ†“ 0, we have that ใ€ˆ๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )ใ€‰โ„“ โ†’ ใ€ˆ๐บ (b, ๐‘ฃ, ๐‘ฃ๐œ•2๐‘…1๐‘ฃ๐‘ก )ใ€‰ as โ„“ โ†“ 0 forevery ๐‘ก โˆˆ (0, 1]. Altogether, we get that

limโ„“โ†“0

(ใ€ˆ๐บ (b, ๐‘ฃ, ๐น )ใ€‰liโ„“ โˆ’ ใ€ˆ๐บ (b, ๐‘ฃ, ๐น )ใ€‰li) โ‰ค [,

which in turn implies that ใ€ˆ๐บ (b, ๐‘ฃ, ๐น )ใ€‰liโ„“ โ†’ ใ€ˆ๐บ (b, ๐‘ฃ, ๐น )ใ€‰li as โ„“ โ†“ 0 since [ is arbitrary. Thus{ใ€ˆยทใ€‰liโ„“ }โ„“โ†“0 converges weakly to ใ€ˆยทใ€‰li. ๏ฟฝ

Acknowledgements

The authors would like to thank the Isaac Newton Institute for Mathematical Sciences, Cam-bridge, for support and hospitality during the programme โ€œThe mathematical design of newmaterialsโ€ where work on this paper was completed. This programme was supported by EPSRCgrant no EP/K032208/1. F.O., T.R., and P.T. also thank the Centre International de Mathรฉmatiqueset dโ€™Informatique de Toulouse (CIMI) and R.I. thanks the Max Planck Institute for Mathematics inthe Sciences for their kind hospitality.

Appendix A. Hรถlder spaces

The following equivalent characterization of Hรถlder norms relies on the โ€œheat kernelโ€ of theoperator A.

Lemma A.1 ([IO19, Lemma 10, Remark 1]). Let ๐‘“ be a periodic distribution on T2.

(1) For ๐›ฝ โˆˆ (โˆ’ 32 , 0) \ {โˆ’1,โˆ’

12 }, we have

[๐‘“ ]๐›ฝ โˆผ sup๐‘‡ โˆˆ(0,1]

(๐‘‡

13)โˆ’๐›ฝ

โ€– ๐‘“๐‘‡ โ€–๐ฟโˆž . (A.1)

In the critical cases ๐›ฝ โˆˆ {โˆ’1,โˆ’ 12 } we have

sup๐‘‡ โˆˆ(0,1]

(๐‘‡

13)โˆ’๐›ฝ

โ€– ๐‘“๐‘‡ โ€–๐ฟโˆž . [๐‘“ ]๐›ฝ . (A.2)

(2) For ๐›ฝ โˆˆ (โˆ’ 32 , 0) \ {โˆ’1,โˆ’

12 } and ๐‘“ of vanishing average, we have

[๐‘“ ]๐›ฝ โˆผ sup๐‘‡ โˆˆ(0,1]

(๐‘‡

13)โˆ’๐›ฝ

โ€–๐‘‡A ๐‘“๐‘‡ โ€–๐ฟโˆž . (A.3)

In the critical cases ๐›ฝ โˆˆ {โˆ’1,โˆ’ 12 } we have

28

sup๐‘‡ โˆˆ(0,1]

(๐‘‡

13)โˆ’๐›ฝ

โ€–๐‘‡A ๐‘“๐‘‡ โ€–๐ฟโˆž . [๐‘“ ]๐›ฝ . (A.4)

28Indeed, using (A.2) and (1.18) we have

โ€–A ๐‘“๐‘‡ โ€–๐ฟโˆž = โ€– ๐‘“๐‘‡2โˆ— A๐œ“ ๐‘‡

2โ€–๐ฟโˆž โ‰ค โ€– ๐‘“๐‘‡

2โ€–๐ฟโˆž โ€–A๐œ“ ๐‘‡

2โ€–๐ฟ1 . (๐‘‡

13 )๐›ฝ [๐‘“ ]๐›ฝ

1๐‘‡.

Page 40: Variational methods for a singular SPDE yielding the

40 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

(3) For ๐›ผ โˆˆ (0, 32 ) \ {1} we have

[๐‘“ ]๐›ผ โˆผ sup๐‘‡ โˆˆ(0,1]

(๐‘‡

13)โˆ’๐›ผ

โ€–๐‘‡A ๐‘“๐‘‡ โ€–๐ฟโˆž . (A.5)

In the critical case ๐›ผ = 1 we have29

sup๐‘‡ โˆˆ(0,1]

(๐‘‡

13)โˆ’1

โ€–๐‘‡A ๐‘“๐‘‡ โ€–๐ฟโˆž . [๐‘“ ]1. (A.6)

In the case of periodic distributions ๐‘“ of vanishing average on T2, one can consider the supremum

over all ๐‘‡ > 0 in (A.1), (A.2), (A.3), and (A.4), while in (A.5) and (A.6) the suprema over ๐‘‡ โˆˆ (0, 1)and ๐‘‡ > 0 are equivalent even for distributions of nonvanishing average.

We also have the following interpolation inequality.

Lemma A.2. For every โˆ’ 32 < ๐›ฝ < 0 < ๐›พ < ๐›ผ < 3

2 there exists a constant ๐ถ > 0 such that the

following interpolation inequality holds for every ๐‘“ : T2 โ†’ R,

[๐‘“ ]๐›พ โ‰ค ๐ถ [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ , (A.7)

where _ โˆˆ (0, 1) is given by ๐›พ = _๐›ฝ + (1 โˆ’ _)๐›ผ . If ๐‘“ has vanishing average in T2, (A.7) also holds for๐›พ = 0 with [๐‘“ ]๐›พ replaced by โ€– ๐‘“ โ€–๐ฟโˆž .

Proof. By (A.1) and (A.2), we have for every ๐‘‡ โˆˆ (0, 1] and ๐‘ฅ,๐‘ฆ โˆˆ T2:

|๐‘“2๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฅ) | โ‰ค โ€– ๐‘“๐‘‡ โ€–๐ฟโˆž + โ€– ๐‘“2๐‘‡ โ€–๐ฟโˆž . [๐‘“ ]๐›ฝ(๐‘‡

13

)๐›ฝ,

|๐œ•1 ๐‘“2๐‘‡ (๐‘ฅ) โˆ’ ๐œ•1 ๐‘“๐‘‡ (๐‘ฅ) | โ‰ค (โ€– ๐‘“ 3๐‘‡2โ€–๐ฟโˆž + โ€– ๐‘“๐‘‡

2โ€–๐ฟโˆž)

โˆซR2

|๐œ•1๐œ“๐‘‡2(๐‘ง) | d๐‘ง . [๐‘“ ]๐›ฝ

(๐‘‡

13

)๐›ฝโˆ’1,

|๐‘“๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฆ) | โ‰ค 2โ€– ๐‘“๐‘‡ โ€–๐ฟโˆž . [๐‘“ ]๐›ฝ(๐‘‡

13

)๐›ฝ,

|๐œ•1 ๐‘“๐‘‡ (๐‘ฆ) (๐‘ฅ1 โˆ’ ๐‘ฆ1) | โ‰ค โ€–๐œ•1๐œ“๐‘‡2โ€–๐ฟ1 (R2) โ€– ๐‘“๐‘‡2 โ€–๐ฟโˆž๐‘‘ (๐‘ฅ,๐‘ฆ) . [๐‘“ ]๐›ฝ

(๐‘‡

13

)๐›ฝ๐‘‘ (๐‘ฅ,๐‘ฆ)๐‘‡

13,

|๐‘“๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฆ) โˆ’ ๐œ•1 ๐‘“๐‘‡ (๐‘ฆ) (๐‘ฅ1 โˆ’ ๐‘ฆ1) | . [๐‘“ ]๐›ฝ(๐‘‡

13

)๐›ฝmax

(1, ๐‘‘ (๐‘ฅ,๐‘ฆ)

๐‘‡13

),

(A.8)

where we used [IO19, equation (26)]. In the case ๐›ผ โˆˆ (0, 1], we claim that for every ๐‘‡ โˆˆ (0, 1] and๐‘ฅ,๐‘ฆ โˆˆ T2:

|๐‘“2๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฅ) | . [๐‘“ ]๐›ผ(๐‘‡

13)๐›ผ

and |๐‘“๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฆ) | . [๐‘“ ]๐›ผ๐‘‘๐›ผ (๐‘ฅ,๐‘ฆ). (A.9)

Indeed, by [IO19, equation (26)], we deduce

|๐‘“๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฆ) | โ‰คโˆซR2

|๐œ“๐‘‡ (๐‘ง) | |๐‘“ (๐‘ฅ โˆ’ ๐‘ง) โˆ’ ๐‘“ (๐‘ฆ โˆ’ ๐‘ง) | d๐‘ง โ‰ค [๐‘“ ]๐›ผ๐‘‘ (๐‘ฅ,๐‘ฆ)๐›ผโˆซR2

|๐œ“๐‘‡ (๐‘ง) | d๐‘ง

. [๐‘“ ]๐›ผ๐‘‘๐›ผ (๐‘ฅ,๐‘ฆ),

|๐‘“2๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฅ) | โ‰คโˆซR2

|๐œ“๐‘‡ (๐‘ง) | |๐‘“๐‘‡ (๐‘ฅ โˆ’ ๐‘ง) โˆ’ ๐‘“๐‘‡ (๐‘ฅ) | d๐‘ง . [๐‘“ ]๐›ผโˆซR2

|๐œ“๐‘‡ (๐‘ง) |๐‘‘ (๐‘ง, 0)๐›ผ d๐‘ง

. [๐‘“ ]๐›ผ(๐‘‡

13)๐›ผ.

29Indeed, since A๐œ“๐‘‡ has vanishing average, we write A ๐‘“๐‘‡ (๐‘ฅ) =โˆซR2 A๐œ“๐‘‡ (๐‘ฆ) (๐‘“ (๐‘ฅ โˆ’ ๐‘ฆ) โˆ’ ๐‘“ (๐‘ฅ)) d๐‘ฆ and deduce via

Step 2 in the proof of Lemma 10 in [IO19] that

โ€–A ๐‘“๐‘‡ โ€–๐ฟโˆž โ‰ค [๐‘“ ]๐›ผโˆซR2

|A๐œ“๐‘‡ (๐‘ฆ) |๐‘‘ (๐‘ฆ, 0)๐›ผ d๐‘ฆ . [๐‘“ ]๐›ผ(๐‘‡

13) (โˆ’3+๐›ผ)

.

Page 41: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 41

In the case ๐›ผ โˆˆ (1, 32 ), arguing as above, we also have for every ๐‘‡ โˆˆ (0, 1] and ๐‘ฅ,๐‘ฆ โˆˆ T2:|๐‘“๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฆ) โˆ’ ๐œ•1 ๐‘“๐‘‡ (๐‘ฆ) (๐‘ฅ1 โˆ’ ๐‘ฆ1) | . [๐‘“ ]๐›ผ๐‘‘๐›ผ (๐‘ฅ,๐‘ฆ),|๐‘“2๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฅ) | โ‰ค

โˆซR2

|๐œ“๐‘‡ (๐‘ง) | |๐‘“๐‘‡ (๐‘ฅ โˆ’ ๐‘ง) โˆ’ ๐‘“๐‘‡ (๐‘ฅ) + ๐œ•1 ๐‘“๐‘‡ (๐‘ฅ)๐‘ง1 | d๐‘ง . [๐‘“ ]๐›ผ(๐‘‡

13

)๐›ผ,

|๐œ•1 ๐‘“2๐‘‡ (๐‘ฅ) โˆ’ ๐œ•1 ๐‘“๐‘‡ (๐‘ฅ) | โ‰คโˆซR2

|๐œ•1๐œ“๐‘‡ /2(๐‘ง) | |๐‘“3๐‘‡ /2(๐‘ฅ โˆ’ ๐‘ง) โˆ’ ๐‘“๐‘‡ /2(๐‘ฅ โˆ’ ๐‘ง) | d๐‘ง . [๐‘“ ]๐›ผ(๐‘‡

13

)๐›ผโˆ’1,

(A.10)where we used

โˆซR2๐œ“๐‘‡ d๐‘ง = 1 and

โˆซR2๐‘ง1๐œ“๐‘‡ (๐‘ง) d๐‘ง = 0. To prove (A.7) we distinguish three dierent

cases for ๐›พ .Case ๐›พ โˆˆ (0, 1): First, assume that ๐›ผ โˆˆ (0, 1]. Interpolating (A.8) and (A.9), we obtain for every๐‘‡ โˆˆ (0, 1] and ๐‘ฅ,๐‘ฆ โˆˆ T2:

|๐‘“2๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฅ) | . [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ

(๐‘‡

13)๐›พ

|๐‘“๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฆ) | . [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ

(๐‘‡

13)_๐›ฝ

๐‘‘ (๐‘ฅ,๐‘ฆ) (1โˆ’_)๐›ผ .

If ๐‘›0 โˆˆ Z is the largest integer such that 2โˆ’๐‘›03 โ‰ฅ ๐‘‘ (๐‘ฅ,๐‘ฆ), then for every ๐‘› โ‰ฅ ๐‘›0

|๐‘“2โˆ’๐‘› (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘› (๐‘ฆ) |โ‰ค |๐‘“2โˆ’๐‘› (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘›0 (๐‘ฅ) | + |๐‘“2โˆ’๐‘›0 (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘›0 (๐‘ฆ) | + |๐‘“2โˆ’๐‘› (๐‘ฆ) โˆ’ ๐‘“2โˆ’๐‘›0 (๐‘ฆ) |

โ‰ค๐‘›โˆ’1โˆ‘๏ธ๐‘˜=๐‘›0

|๐‘“2โˆ’(๐‘˜+1) (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘˜ (๐‘ฅ) | + |๐‘“2โˆ’๐‘›0 (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘›0 (๐‘ฆ) | +๐‘›โˆ’1โˆ‘๏ธ๐‘˜=๐‘›0

|๐‘“2โˆ’(๐‘˜+1) (๐‘ฆ) โˆ’ ๐‘“2โˆ’๐‘˜ (๐‘ฆ) |

. [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ

(๐‘›โˆ’1โˆ‘๏ธ๐‘˜=๐‘›0

(2โˆ’

๐‘˜3)๐›พ

+(2โˆ’

๐‘›03)_๐›ฝ

๐‘‘ (๐‘ฅ,๐‘ฆ) (1โˆ’_)๐›ผ)

. [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ

((2โˆ’

๐‘›03)๐›พ

+(2โˆ’

๐‘›03)_๐›ฝ

๐‘‘ (๐‘ฅ,๐‘ฆ) (1โˆ’_)๐›ผ). [๐‘“ ]_

๐›ฝ[๐‘“ ]1โˆ’_๐›ผ ๐‘‘ (๐‘ฅ,๐‘ฆ)๐›พ , (A.11)

which in turn implies (A.7) by letting ๐‘› โ†’ โˆž.In the case ๐›ผ โˆˆ (1, 32 ), rst, one needs to choose ๐‘›0 โˆˆ Z such that 2โˆ’

๐‘›03 โ‰ฅ ^๐‘‘ (๐‘ฅ,๐‘ฆ) > 2โˆ’

๐‘›0+13

with a constant ^ > 0, depending only on ๐›พ , which we x below. Second, one needs to estimatethe intermediate term |๐‘“2โˆ’๐‘›0 (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘›0 (๐‘ฆ) | dierently. For this, we need to use that for every๐‘‡ โˆˆ (0, 1], by (A.1), (A.2), Denition 1.13, and since ๐›พ < 1,

โ€–๐œ•1 ๐‘“๐‘‡ โ€–๐ฟโˆž . [๐œ•1 ๐‘“ ]๐›พโˆ’1(๐‘‡13 )๐›พโˆ’1 โ‰ค ๐‘0 [๐‘“ ]๐›พ (๐‘‡

13 )๐›พโˆ’1

where ๐‘0 > 0 depends only on ๐›พ . By interpolating between (A.8) and (A.10), we estimate theintermediate term for a constant ๐ถ > 0 (depending on ๐›ผ , ๐›ฝ , ๐›พ ) by,

|๐‘“2โˆ’๐‘›0 (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘›0 (๐‘ฆ) |โ‰ค |๐‘“2โˆ’๐‘›0 (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘›0 (๐‘ฆ) โˆ’ ๐œ•1 ๐‘“2โˆ’๐‘›0 (๐‘ฆ) (๐‘ฅ1 โˆ’ ๐‘ฆ1) | + |๐œ•1 ๐‘“2โˆ’๐‘›0 (๐‘ฆ) (๐‘ฅ1 โˆ’ ๐‘ฆ1) |

โ‰ค ๐ถ [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ

(2โˆ’

๐‘›03)_๐›ฝ

๐‘‘ (๐‘ฅ,๐‘ฆ) (1โˆ’_)๐›ผ max(1, ๐‘‘ (๐‘ฅ,๐‘ฆ)

2โˆ’๐‘›03

)_+ ๐‘0 [๐‘“ ]๐›พ (2โˆ’

๐‘›03 )๐›พโˆ’1๐‘‘ (๐‘ฅ,๐‘ฆ)

โ‰ค ๐ถ^ [๐‘“ ]_๐›ฝ [๐‘“ ]1โˆ’_๐›ผ ๐‘‘ (๐‘ฅ,๐‘ฆ)๐›พ + ๐‘0 [๐‘“ ]๐›พ^๐›พโˆ’1๐‘‘ (๐‘ฅ,๐‘ฆ)๐›พ .

Choosing ^ > 0 such that ๐‘0^๐›พโˆ’1 = 12 and proceeding as in (A.11) (where now the change of ๐‘›0

aects the implicit constant by a factor depending on ^), after passing to the limit ๐‘› โ†’ โˆž weobtain

|๐‘“ (๐‘ฅ) โˆ’ ๐‘“ (๐‘ฆ) | โ‰ค(๐ถ [๐‘“ ]_

๐›ฝ[๐‘“ ]1โˆ’_๐›ผ + 1

2 [๐‘“ ]๐›พ)๐‘‘ (๐‘ฅ,๐‘ฆ)๐›พ .

Dividing by ๐‘‘ (๐‘ฅ,๐‘ฆ)๐›พ and taking the supremum over ๐‘ฅ โ‰  ๐‘ฆ we nally obtain (A.7).

Page 42: Variational methods for a singular SPDE yielding the

42 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

Case ๐›พ โˆˆ (1, 32 ): Since ๐›ผ > ๐›พ we also have ๐›ผ โˆˆ (1, 32 ). Interpolating (A.8) and (A.10), we obtain forevery ๐‘‡ โˆˆ (0, 1] and ๐‘ฅ,๐‘ฆ โˆˆ T2,

|๐‘“2๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฅ) | . [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ

(๐‘‡

13)๐›พ,

|๐œ•1 ๐‘“2๐‘‡ (๐‘ฅ) โˆ’ ๐œ•1 ๐‘“๐‘‡ (๐‘ฅ) | . [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ

(๐‘‡

13)๐›พโˆ’1

,

|๐‘“๐‘‡ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฆ) โˆ’ ๐œ•1 ๐‘“๐‘‡ (๐‘ฆ) (๐‘ฅ1 โˆ’ ๐‘ฆ1) | . [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ

(๐‘‡

13)_๐›ฝ

๐‘‘ (๐‘ฅ,๐‘ฆ) (1โˆ’_)๐›ผ max(1, ๐‘‘ (๐‘ฅ,๐‘ฆ)

๐‘‡13

)_.

If ๐‘›0 โˆˆ Z is the largest integer such that 2โˆ’๐‘›03 โ‰ฅ ๐‘‘ (๐‘ฅ,๐‘ฆ), the same argument as in the previous

case yields for ๐‘› โ‰ฅ ๐‘›0,|๐‘“2โˆ’๐‘› (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘› (๐‘ฆ) โˆ’ ๐œ•1 ๐‘“2โˆ’๐‘› (๐‘ฆ) (๐‘ฅ1 โˆ’ ๐‘ฆ1) |

โ‰ค |๐‘“2โˆ’๐‘›0 (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘›0 (๐‘ฆ) โˆ’ ๐œ•1 ๐‘“2โˆ’๐‘›0 (๐‘ฆ) (๐‘ฅ1 โˆ’ ๐‘ฆ1) | + |๐‘“2โˆ’๐‘› (๐‘ฅ) โˆ’ ๐‘“2โˆ’๐‘›0 (๐‘ฅ) |+ |๐‘“2โˆ’๐‘› (๐‘ฆ) โˆ’ ๐‘“2โˆ’๐‘›0 (๐‘ฆ) | + |๐œ•1 ๐‘“2โˆ’๐‘› (๐‘ฆ) โˆ’ ๐œ•1 ๐‘“2โˆ’๐‘›0 (๐‘ฆ) | |๐‘ฅ1 โˆ’ ๐‘ฆ1 |

. [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ

((2โˆ’

๐‘›03)_๐›ฝ

๐‘‘ (๐‘ฅ,๐‘ฆ) (1โˆ’_)๐›ผ +๐‘›โˆ’1โˆ‘๏ธ๐‘˜=๐‘›0

(2โˆ’

๐‘˜3)๐›พ

+ ๐‘‘ (๐‘ฅ,๐‘ฆ)๐‘›โˆ’1โˆ‘๏ธ๐‘˜=๐‘›0

(2โˆ’

๐‘˜3)๐›พโˆ’1)

. [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ ๐‘‘ (๐‘ฅ,๐‘ฆ)๐›พ ,

which yields (A.7) by letting ๐‘› โ†’ โˆž.Case ๐›พ = 1: Take a sequence (๐›พ๐‘›) โŠ‚ (1, 32 ) such that ๐›พ๐‘› โ†˜ 1, and consider the correspondingexponents _๐‘› โ†’ _. Then by [IO19, Remark 2] and the previous case, we have

[๐‘“ ]1 . [๐‘“ ]๐›พ๐‘› . [๐‘“ ]_๐‘›๐›ฝ[๐‘“ ]1โˆ’_๐‘›๐›ผ ,

with implicit constants independent of ๐‘›. We can therefore perform the limit ๐‘› โ†’ โˆž to concludethe estimate for ๐›พ = 1.Case ๐›พ = 0 and ๐‘“ has vanishing average: If [๐‘“ ]๐›ผ = 0, then ๐‘“ โ‰ก 0,30 so (A.7) holds trivially. Assumethat [๐‘“ ]๐›ผ โ‰  0. If ๐›ผ โˆˆ (0, 1], we have for every ๐‘‡ > 0,

|๐‘“ (๐‘ฅ) | โ‰ค |๐‘“ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฅ) | + |๐‘“๐‘‡ (๐‘ฅ) | .โˆซR2

|๐œ“๐‘‡ (๐‘ง) | |๐‘“ (๐‘ฅ โˆ’ ๐‘ง) โˆ’ ๐‘“ (๐‘ฅ) | d๐‘ง +(๐‘‡

13)๐›ฝ

[๐‘“ ]๐›ฝ

.(๐‘‡

13)๐›ผ

[๐‘“ ]๐›ผ +(๐‘‡

13)๐›ฝ

[๐‘“ ]๐›ฝ ,

while if ๐›ผ โˆˆ (1, 32 ),

|๐‘“ (๐‘ฅ) | โ‰ค |๐‘“ (๐‘ฅ) โˆ’ ๐‘“๐‘‡ (๐‘ฅ) | + |๐‘“๐‘‡ (๐‘ฅ) | .โˆซR2

|๐œ“๐‘‡ (๐‘ง) | |๐‘“ (๐‘ฅ โˆ’ ๐‘ง) โˆ’ ๐‘“ (๐‘ฅ) + ๐œ•1 ๐‘“ (๐‘ฅ)๐‘ง1 | d๐‘ง +(๐‘‡

13)๐›ฝ

[๐‘“ ]๐›ฝ

.(๐‘‡

13)๐›ผ

[๐‘“ ]๐›ผ +(๐‘‡

13)๐›ฝ

[๐‘“ ]๐›ฝ ,

where we usedโˆซR2๐‘ง1๐œ“๐‘‡ (๐‘ง) d๐‘ง = 0 and Lemma A.1 in the case of distributions with vanishing

average, as ๐‘‡ can be larger than 1. Choosing ๐‘‡ 13 =

( [๐‘“ ]๐›ฝ[๐‘“ ]๐›ผ

) 1๐›ผโˆ’๐›ฝ leads to the conclusion. ๏ฟฝ

Remark A.3. One can also prove that for โˆ’ 12 < ๐›ฝ < 1 < ๐›ผ < 3

2 the interpolation estimate

โ€–๐œ•1 ๐‘“ โ€–๐ฟโˆž . [๐‘“ ]_๐›ฝ[๐‘“ ]1โˆ’_๐›ผ (A.12)

holds for _ โˆˆ (0, 1) given by 1 = _๐›ฝ + (1 โˆ’ _)๐›ผ . Indeed, by Lemma A.2 (the case ๐›พ = 0) we knowthat

โ€–๐œ•1 ๐‘“ โ€–๐ฟโˆž . [๐œ•1 ๐‘“ ]_๐›ฝโˆ’1 [๐œ•1 ๐‘“ ]1โˆ’_๐›ผโˆ’1.

30This is clear for ๐›ผ โˆˆ (0, 1) since ๐‘“ has vanishing average. For ๐›ผ โˆˆ (1, 32 ) notice that by [IO19, Lemma 12],โ€–๐œ•1 ๐‘“ โ€–๐ฟโˆž = 0, hence by Denition 1.12 ๐‘“ is constant, and this constant is 0 since ๐‘“ has vanishing average.

Page 43: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 43

By Denition 1.13, we have [๐œ•1 ๐‘“ ]๐›ฝโˆ’1 . [๐‘“ ]๐›ฝ and by [IO19, Lemma 12], we know [๐œ•1 ๐‘“ ]๐›ผโˆ’1 . [๐‘“ ]๐›ผ ,so the desired estimate follows.

We also need the following lemma for the Hilbert transform acting on Hรถlder spaces.

Lemma A.4.(1) For ๐›ฝ โˆˆ (โˆ’ 3

2 , 0) and Y > 0 such that ๐›ฝ โˆ’ Y โˆˆ (โˆ’ 32 , 0), there exists a constant ๐ถ > 0 such that

for every periodic distribution ๐‘“ ,

[๐‘…1 ๐‘“ ]๐›ฝโˆ’Y โ‰ค ๐ถ [๐‘“ ]๐›ฝ . (A.13)

(2) For ๐›ผ โˆˆ (0, 32 ) and Y > 0 such that ๐›ผ โˆ’ Y โˆˆ (0, 32 ) there exists a constant ๐ถ > 0 such that for

every ๐‘“ : T2 โ†’ R,[๐‘…1 ๐‘“ ]๐›ผโˆ’Y โ‰ค ๐ถ [๐‘“ ]๐›ผ . (A.14)

Proof. To prove (A.13), we claim that for every Y โˆˆ (0, 32 ) and ๐‘‡ โˆˆ (0, 1], we have that

โ€–๐‘…1 ๐‘“๐‘‡ โ€–๐ฟโˆž .(๐‘‡

13)โˆ’Y

โ€– ๐‘“๐‘‡2โ€–๐ฟโˆž .

Indeed, if we write ฮจ๐‘‡ for the periodization of ๐œ“๐‘‡ , by the semigroup property and Youngโ€™sinequality for convolution we have for ๐‘ = 5

2Y > 1,

โ€–๐‘…1 ๐‘“๐‘‡ โ€–๐ฟโˆž . โ€– ๐‘“๐‘‡2โ€–๐ฟ๐‘ โ€–๐‘…1ฮจ๐‘‡

2โ€–๐ฟ

๐‘๐‘โˆ’1. โ€– ๐‘“๐‘‡

2โ€–๐ฟ๐‘ โ€–ฮจ๐‘‡

2โ€–๐ฟ

๐‘๐‘โˆ’1. โ€– ๐‘“๐‘‡

2โ€–๐ฟโˆž

(๐‘‡

13)โˆ’ 5

2๐‘

where we used that the Hilbert transform is bounded31 on ๐ฟ๐‘

๐‘โˆ’1 (T2) for ๐‘

๐‘โˆ’1 > 1 and the bound

โ€–ฮจ๐‘‡2โ€–๐ฟ

๐‘๐‘โˆ’1. (๐‘‡ 1

3 )โˆ’52๐‘ , which follows from Remark 1.11. Then (A.13) follows via the characterization

of negative Hรถlder norms (A.1) and (A.2) if ๐›ฝ โˆ’ ๐œ– โ‰  โˆ’1,โˆ’ 12 . In those cases, consider ๐›พ โˆˆ (๐›ฝ โˆ’ ๐œ–, ๐›ฝ)

and use the previous case together with [IO19, Remark 2].Equation (A.14) is essentially [IO19, Lemma 7], noting that we can assume that ๐‘“ is of vanishing

average, as ๐‘…1 annihilates constants and [๐‘“ โˆ’โˆซT2๐‘“ d๐‘ฅ]๐›ผ โ‰ค [๐‘“ ]๐›ผ . ๏ฟฝ

Lemma A.5. Let ๐›ผ โˆˆ (โˆ’ 32 ,

32 ) \ {0} and ๐‘  > 0 such that ๐›ผ โˆ’ ๐‘  โˆˆ (โˆ’ 3

2 ,32 ) \ {โˆ’1,โˆ’

12 , 0, 1}. There exists

a constant ๐ถ > 0 such that for every periodic ๐‘“ โˆˆ C๐›ผ,

[|๐œ•1 |๐‘  ๐‘“ ]๐›ผโˆ’๐‘  โ‰ค ๐ถ [๐‘“ ]๐›ผ .

Proof. Without loss of generality, we may assume that ๐‘“ is of vanishing average because [๐‘“ โˆ’โˆซT2๐‘“ d๐‘ฅ]๐›ผ โ‰ค [๐‘“ ]๐›ผ and [|๐œ•1 |๐‘  ๐‘“ ]๐›ผโˆ’๐‘  is invariant by adding a constant to ๐‘“ . Then by the semigroup

property, |๐œ•1 |๐‘  ๐‘“๐‘‡ =(๐‘‡2)โˆ’ ๐‘ 

3 ๐‘“๐‘‡2โˆ— (|๐œ•1 |๐‘ ๐œ“ )๐‘‡

2for every๐‘‡ โˆˆ (0, 1] and since |๐œ•1 |๐‘ ๐œ“ โˆˆ ๐ฟ1(R2), we deduce

โ€–๐‘‡A|๐œ•1 |๐‘  ๐‘“๐‘‡ โ€–๐ฟโˆž .(๐‘‡

13

)โˆ’๐‘ โ€–๐‘‡A ๐‘“๐‘‡

2โ€–๐ฟโˆž . Hence, we obtain that(

๐‘‡13)โˆ’(๐›ผโˆ’๐‘ )

โ€–๐‘‡A|๐œ•1 |๐‘  ๐‘“๐‘‡ โ€–๐ฟโˆž .(๐‘‡

13)โˆ’๐›ผ

โ€–๐‘‡A ๐‘“๐‘‡2โ€–๐ฟโˆž,

and the conclusion follows by Lemma A.1. ๏ฟฝ

Lemma A.6. Let ๐›ผ โˆˆ (โˆ’ 32 ,

32 ) \ {0}. For every sequence {๐‘“๐‘›}๐‘›โ‰ฅ1 โŠ‚ C๐›ผ

with sup๐‘›โ‰ฅ1 [๐‘“๐‘›]๐›ผ < โˆž, there

exists ๐‘“ โˆˆ C๐›ผwith

[๐‘“ ]๐›ผ โ‰ค lim inf๐‘›โ†’โˆž

[๐‘“๐‘›]๐›ผ ,

such that ๐‘“๐‘› โ†’ ๐‘“ in C๐›ผโˆ’Yfor every Y > 0 along a subsequence. In particular, the embedding

C๐›ผ โ†ฉโ†’ C๐›ผโˆ’Yis compact.

31This follows from the fact that the Hilbert transform ๐‘…1 is bounded on ๐ฟ๐‘ž (R2) for any ๐‘ž โˆˆ (1,โˆž) and thetransference of multipliers method [Gra14, Theorem 4.3.7].

Page 44: Variational methods for a singular SPDE yielding the

44 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

Proof. First, assume that ๐›ผ โˆ’ Y โ‰  โˆ’1,โˆ’ 12 , 0, 1. Let {๐‘“๐‘›}๐‘›โ‰ฅ1 โŠ‚ C๐›ผ such that๐พ := sup

๐‘›โ‰ฅ1[๐‘“๐‘›]๐›ผ < โˆž,

and, if ๐›ผ โˆˆ (0, 32 ), we assume in addition that {โ€– ๐‘“๐‘› โ€–๐ฟโˆž}๐‘›โ‰ฅ1 is uniformly bounded (hence, up tosubtraction, we may assume that ๐‘“๐‘› has zero average for any ๐‘› โˆˆ N).

By [IO19, Lemma 13] there exists ๐‘“ โˆˆ C๐›ผ such that ๐‘“๐‘› โ†’ ๐‘“ (along a subsequence) in the senseof distributions and

[๐‘“ ]๐›ผ โ‰ค lim inf๐‘›โ†’โˆž

[๐‘“๐‘›]๐›ผ โ‰ค ๐พ.

For ๐‘‡, ๐‘ก > 0 we have thatโ€–(A ๐‘“ โˆ’ A ๐‘“๐‘›)๐‘‡ โ€–๐ฟโˆž โ‰ค โ€–(A ๐‘“ โˆ’ A ๐‘“๐‘ก )๐‘‡ โ€–๐ฟโˆž + โ€–(A ๐‘“ โˆ’ A ๐‘“๐‘›)๐‘ก+๐‘‡ โ€–๐ฟโˆž

+ โ€–((A ๐‘“๐‘›)๐‘ก โˆ’ A ๐‘“๐‘›)๐‘‡ โ€–๐ฟโˆž . (A.15)By (A.3), (A.5), and (A.16) below the following estimates hold

โ€–(A ๐‘“ โˆ’ A ๐‘“๐‘ก )๐‘‡ โ€–๐ฟโˆž .(๐‘‡

13)๐›ผโˆ’Yโˆ’3

[๐‘“ โˆ’ ๐‘“๐‘ก ]๐›ผโˆ’Y . ๐พ(๐‘‡

13)๐›ผโˆ’Yโˆ’3 (

๐‘ก13)Y,

โ€–((A ๐‘“๐‘›)๐‘ก โˆ’ A ๐‘“๐‘›)๐‘‡ โ€–๐ฟโˆž .(๐‘‡

13)๐›ผโˆ’Yโˆ’3

[(๐‘“๐‘›)๐‘ก โˆ’ ๐‘“๐‘›]๐›ผโˆ’Y . ๐พ(๐‘‡

13)๐›ผโˆ’Yโˆ’3 (

๐‘ก13)Y.

By Youngโ€™s inequality for convolution and (1.18) we further have that

โ€–(A ๐‘“ โˆ’ A ๐‘“๐‘›)๐‘ก+๐‘‡ โ€–๐ฟโˆž .(๐‘‡

13)๐›ผโˆ’Yโˆ’3

โ€–(A ๐‘“ โˆ’ A ๐‘“๐‘›)๐‘ก โ€–๐ฟ

52(3โˆ’๐›ผ+Y )

.

Since ๐‘“๐‘› โ†’ ๐‘“ in the sense of distributions, (A ๐‘“ โˆ’ A ๐‘“๐‘›)๐‘ก โ†’ 0 as ๐‘› โ†’ โˆž pointwise for every๐‘ก โˆˆ (0, 1], and by (A.3) and (A.5) we know that

โ€–(A ๐‘“ โˆ’ A ๐‘“๐‘›)๐‘ก โ€–๐ฟโˆž .(๐‘ก13)๐›ผโˆ’3

[๐‘“ โˆ’ ๐‘“๐‘›]๐›ผ . ๐พ(๐‘ก13)๐›ผโˆ’3

.

Hence, by dominated convergence theorem, โ€–(A ๐‘“ โˆ’ A ๐‘“๐‘›)๐‘ก โ€–๐ฟ

52(3โˆ’๐›ผ+Y )

โ†’ 0 as ๐‘› โ†’ โˆž for every๐‘ก โˆˆ (0, 1]. Taking ๐‘› โ†’ โˆž in (A.15) and using again (A.3) and (A.5) we obtain that

lim sup๐‘›โ†’โˆž

[๐‘“ โˆ’ ๐‘“๐‘›]๐›ผโˆ’Y . ๐พ(๐‘ก13)Y,

which completes the proof if we let ๐‘ก โ†’ 0.If ๐›ผ โˆ’ Y โˆˆ {โˆ’1,โˆ’ 1

2 , 0, 1}, consider ๐›พ โˆˆ (๐›ผ โˆ’ Y, ๐›ผ); in view of [IO19, Remark 2] and the aboveresult we then have [๐‘“ โˆ’ ๐‘“๐‘›]๐›ผโˆ’Y . [๐‘“ โˆ’ ๐‘“๐‘›]๐›พ โ†’ 0 as ๐‘› โ†’ โˆž. ๏ฟฝ

Proposition A.7. For every ๐›ผ โˆˆ (โˆ’ 32 ,

32 ) \ {0} and Y > 0 such that ๐›ผ โˆ’ Y โˆˆ (โˆ’ 3

2 ,32 ) \ {โˆ’1,โˆ’

12 , 0, 1}

the following estimate holds:

[๐‘“ โˆ’ ๐‘“๐‘ก ]๐›ผโˆ’Y .(๐‘ก13)Y

[๐‘“ ]๐›ผ . (A.16)

Proof. To prove (A.16) we use the denition of ๐œ“๐‘ก and the semigroup property to estimate for๐‘ก,๐‘‡ โˆˆ (0, 1]

โ€–A (๐‘“ โˆ’ ๐‘“๐‘ก )๐‘‡ โ€–๐ฟโˆž โ‰คโˆซ ๐‘ก+๐‘‡

๐‘‡

โ€–A๐œ•๐‘  ๐‘“๐‘  โ€–๐ฟโˆž d๐‘  =โˆซ ๐‘ก+๐‘‡

๐‘‡

โ€–A2 ๐‘“๐‘  โ€–๐ฟโˆž d๐‘ 

=

โˆซ ๐‘ก+๐‘‡

๐‘‡

โ€–A๐œ“ ๐‘ 2โˆ— A ๐‘“ ๐‘ 

2โ€–๐ฟโˆž d๐‘  โ‰ค

โˆซ ๐‘ก+๐‘‡

๐‘‡

โ€–A๐œ“ ๐‘ 2โ€–๐ฟ1 โ€–A ๐‘“ ๐‘ 

2โ€–๐ฟโˆž d๐‘  .

Since โ€–A๐œ“ ๐‘ 2โ€–๐ฟ1 . ๐‘ โˆ’1, (A.3) and (A.5) imply that

โ€–A (๐‘“ โˆ’ ๐‘“๐‘ก )๐‘‡ โ€–๐ฟโˆž . [๐‘“ ]๐›ผโˆซ ๐‘ก+๐‘‡

๐‘‡

(๐‘  13 )โˆ’3+๐›ผ d๐‘ 

๐‘ . [๐‘“ ]๐›ผ

(๐‘‡

13)โˆ’3+๐›ผโˆ’Y โˆซ ๐‘ก+๐‘‡

๐‘‡

(๐‘  13 )Y d๐‘ 

๐‘ 

. [๐‘“ ]๐›ผ(๐‘‡

13)โˆ’3+๐›ผโˆ’Y (

๐‘ก13)Y,

Page 45: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 45

so that (A.16) follows from (A.3) and (A.5). ๏ฟฝ

Appendix B. Besov spaces

In the next lemma we summarize some useful properties of the Besov seminorms that we oftenuse in this article.

Lemma B.1. Let 0 < ๐‘  < ๐‘  โ€ฒ โ‰ค 1, 1 โ‰ค ๐‘ โ‰ค ๐‘ž < โˆž, and ๐‘— โˆˆ {1, 2}. The following estimates hold

โ€– ๐‘“ โ€– ยคB๐‘ โ€ฒ๐‘ ;1

โ‰ค [๐‘“ ]๐‘ โ€ฒ , (B.1)

โ€– ๐‘“ โ€– ยคB๐‘ ๐‘ ;2

โ‰ค [๐‘“ ] 32๐‘ , (B.2)

โ€– ๐‘“ โ€– ยคB๐‘ ๐‘ ;๐‘—

โ‰ค โ€– ๐‘“ โ€– ยคB๐‘ โ€ฒ๐‘ ;๐‘—

and โ€– ๐‘“ โ€– ยคB๐‘ โ€ฒ๐‘ ;๐‘—

โ‰ค โ€– ๐‘“ โ€– ยคB๐‘ โ€ฒ๐‘ž;๐‘—, (B.3)

โ€–|๐œ•๐‘— |๐‘  ๐‘“ โ€–๐ฟ๐‘ โ‰ค ๐ถ (๐‘ , ๐‘  โ€ฒ)โ€– ๐‘“ โ€– ยคB๐‘ โ€ฒ๐‘ ;๐‘—, (B.4)

for every function ๐‘“ : T2 โ†’ R, where the constant ๐ถ (๐‘ , ๐‘  โ€ฒ) > 0 depends only on ๐‘  and ๐‘  โ€ฒ.

Proof. Estimates (B.1) and (B.2) are immediate from Denitions 1.12 and 2.1. Both estimates in (B.3)follow from Denition 2.1 and Jensenโ€™s inequality. To prove (B.4) we rst notice that by a simplecalculation of the Fourier coecients we have the identity

|๐œ•๐‘— |๐‘  ๐‘“ (๐‘ฅ) = ๐‘˜๐‘ โˆซR

๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ)|โ„Ž |๐‘ +1 dโ„Ž, ๐‘ฅ โˆˆ T2,

for some constant ๐‘˜๐‘  > 0 which depends only on ๐‘  , where we interpret the integral as a principlevalue. Then by Minkowskiโ€™s inequality we get

โ€–|๐œ•๐‘— |๐‘  ๐‘“ โ€–๐ฟ๐‘ โ‰ค ๐‘˜๐‘ โˆซR

(โˆซT2

|๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ) |๐‘

|โ„Ž | (๐‘ +1)๐‘d๐‘ฅ

) 1๐‘

dโ„Ž

โ‰ค ๐‘˜๐‘ โˆซ|โ„Ž | โ‰ค1

(โˆซT2

|๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ) |๐‘

|โ„Ž | (๐‘ +1)๐‘d๐‘ฅ

) 1๐‘

dโ„Ž + ๐‘˜๐‘ โˆซ|โ„Ž | โ‰ฅ1

(โˆซT2

|๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ) |๐‘

|โ„Ž | (๐‘ +1)๐‘d๐‘ฅ

) 1๐‘

dโ„Ž.

The rst term in the last inequality is estimated byโˆซ|โ„Ž | โ‰ค1

(โˆซT2

|๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ) |๐‘

|โ„Ž | (๐‘ +1)๐‘d๐‘ฅ

) 1๐‘

dโ„Ž = 2โˆซ 1

0

(โˆซT2

|๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ) |๐‘

โ„Ž (๐‘ +1)๐‘d๐‘ฅ

) 1๐‘

dโ„Ž

โ‰ค 2โˆซ 1

0

โ„Ž๐‘ โ€ฒโˆ’๐‘ 

โ„Ždโ„Ž โ€– ๐‘“ โ€– ยคB๐‘ โ€ฒ

๐‘ ;๐‘—=

2๐‘  โ€ฒ โˆ’ ๐‘  โ€– ๐‘“ โ€– ยคB๐‘ โ€ฒ

๐‘ ;๐‘—,

where we also used translation invariance of the torus and that ๐‘  โ€ฒ > ๐‘  . The second term isestimated by Minkowskiโ€™s inequalityโˆซ

|โ„Ž | โ‰ฅ1

(โˆซT2

|๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ) |๐‘

|โ„Ž | (๐‘ +1)๐‘d๐‘ฅ

) 1๐‘

dโ„Ž โ‰ค 4โˆซ โˆž

1

1โ„Ž๐‘ +1

dโ„Ž โ€– ๐‘“ โ€–๐ฟ๐‘ =4๐‘ โ€– ๐‘“ โ€–๐ฟ๐‘ ,

where we used again translation invariance of the torus. If ๐‘“ has vanishing average in ๐‘ฅ ๐‘— , thenโ€– ๐‘“ โ€–๐ฟ๐‘ โ‰ค โ€– ๐‘“ โ€– ยคB๐‘ โ€ฒ

๐‘ ;๐‘—. Otherwise โ€– ๐‘“ โˆ’

โˆซ 10 ๐‘“ d๐‘ฅ ๐‘— โ€–๐ฟ๐‘ โ‰ค โ€– ๐‘“ โ€– ยคB๐‘ โ€ฒ

๐‘ ;๐‘—and we can replace ๐‘“ by ๐‘“ โˆ’

โˆซ 10 ๐‘“ d๐‘ฅ ๐‘— .

This completes the proof. ๏ฟฝ

Lemma B.2.(i) For every ๐‘  โˆˆ (0, 1), ๐‘ โ‰ฅ 1 and ๐‘“ , ๐‘” : T2 โ†’ R we have

โ€– ๐‘“ ๐‘”โ€– ยคB๐‘ ๐‘ ;1

โ‰ค โ€– ๐‘“ โ€– ยคB๐‘ ๐‘ ;1โ€–๐‘”โ€–๐ฟโˆž + โ€– ๐‘“ โ€–๐ฟ๐‘ [๐‘”]๐‘  ,

โ€– ๐‘“ ๐‘”โ€– ยคB๐‘ ๐‘ ;2

โ‰ค โ€– ๐‘“ โ€– ยคB๐‘ ๐‘ ;2โ€–๐‘”โ€–๐ฟโˆž + โ€– ๐‘“ โ€–๐ฟ๐‘ [๐‘”] 3

2๐‘ .

Page 46: Variational methods for a singular SPDE yielding the

46 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

(ii) For every ๐‘  โˆˆ (0, 1), there exists a constant ๐ถ (๐‘ ) > 0 depending only on ๐‘  such that for every

Y โˆˆ (0, 1 โˆ’ ๐‘ ) and ๐‘“ , ๐‘” : T2 โ†’ R we have

โ€–|๐œ•1 |๐‘  (๐‘“ ๐‘”)โ€–๐ฟ2 โ‰ค 2โ€–|๐œ•1 |๐‘  ๐‘“ โ€–๐ฟ2 โ€–๐‘”โ€–๐ฟโˆž + ๐ถ (๐‘ )โˆšYโ€– ๐‘“ โ€–๐ฟ2 [๐‘”]๐‘ +Y,

โ€–|๐œ•2 |๐‘  (๐‘“ ๐‘”)โ€–๐ฟ2 โ‰ค 2โ€–|๐œ•2 |๐‘  ๐‘“ โ€–๐ฟ2 โ€–๐‘”โ€–๐ฟโˆž + ๐ถ (๐‘ )โˆšYโ€– ๐‘“ โ€–๐ฟ2 [๐‘”] 3

2 (๐‘ +Y).

Proof. (i) For ๐‘— โˆˆ {1, 2} we have that

๐œ•โ„Ž๐‘— (๐‘“ ๐‘”) (๐‘ฅ) = (๐œ•โ„Ž๐‘— ๐‘“ ) (๐‘ฅ)๐‘”(๐‘ฅ + โ„Ž๐‘’ ๐‘— ) + ๐‘“ (๐‘ฅ) (๐œ•โ„Ž๐‘— ๐‘”) (๐‘ฅ),

hence by Minkowskiโ€™s inequality, โ€–๐œ•โ„Ž๐‘— (๐‘“ ๐‘”)โ€–๐ฟ๐‘ โ‰ค โ€–(๐œ•โ„Ž๐‘— ๐‘“ )๐‘”(ยท + โ„Ž๐‘’ ๐‘— )โ€–๐ฟ๐‘ + โ€– ๐‘“ (๐œ•โ„Ž๐‘— ๐‘”)โ€–๐ฟ๐‘ . Itfollows that โ€– ๐‘“ ๐‘”โ€– ยคB๐‘ 

๐‘ ;๐‘—โ‰ค โ€– ๐‘“ โ€– ยคB๐‘ 

๐‘ ;๐‘—โ€–๐‘”โ€–๐ฟโˆž + โ€– ๐‘“ โ€–๐ฟ๐‘ [๐‘”]๐‘  ๐‘— with ๐‘ 1 = ๐‘  and ๐‘ 2 = 3

2๐‘  .(ii) By (2.2) we know that

โ€–|๐œ•๐‘— |2(๐‘“ ๐‘”)โ€–2๐ฟ2 = ๐‘๐‘ โˆซR

1|โ„Ž |2๐‘ 

โˆซT2

|๐œ•โ„Ž๐‘— (๐‘“ ๐‘”) (๐‘ฅ) |2 d๐‘ฅdโ„Ž|โ„Ž | .

Similarly to (i) we can prove that

๐‘๐‘ 

โˆซR

1|โ„Ž |2๐‘ 

โˆซT2

|๐œ•โ„Ž๐‘— (๐‘“ ๐‘”) (๐‘ฅ) |2 d๐‘ฅdโ„Ž|โ„Ž |

โ‰ค 2๐‘๐‘ โˆซR

1|โ„Ž |2๐‘ 

โˆซT2

|๐œ•โ„Ž๐‘— ๐‘“ (๐‘ฅ) |2 d๐‘ฅdโ„Ž|โ„Ž | โ€–๐‘”โ€–

2๐ฟโˆž + 2๐‘๐‘  โ€– ๐‘“ โ€–2๐ฟ2

โˆซR

1|โ„Ž |2๐‘  sup

๐‘ฅ โˆˆT2|๐œ•โ„Ž๐‘— ๐‘”(๐‘ฅ) |2

dโ„Ž|โ„Ž |

= 2โ€–|๐œ•๐‘— |๐‘  ๐‘“ โ€–2๐ฟ2 โ€–๐‘”โ€–2๐ฟโˆž + 2๐‘๐‘  โ€– ๐‘“ โ€–2๐ฟ2

โˆซR

1|โ„Ž |2๐‘  sup

๐‘ฅ โˆˆT2|๐œ•โ„Ž๐‘— ๐‘”(๐‘ฅ) |2

dโ„Ž|โ„Ž | ,

where in the last step we used again (2.2). Using the periodicity of ๐‘” and the fact thatfor โ„Ž > 1 we can write โ„Ž = โ„Žfr + โ„Žint with โ„Žfr โˆˆ (0, 1] and โ„Žint โˆˆ Z, we notice that|๐œ•โ„Ž๐‘— ๐‘”(๐‘ฅ) | = |๐œ•โ„Žfr

๐‘—๐‘”(๐‘ฅ) | . [๐‘”]๐‘  ๐‘—โ„Ž

๐‘  ๐‘—

fr . [๐‘”]๐‘  ๐‘— with ๐‘  ๐‘— โˆˆ {๐‘  + Y, 32 (๐‘  + Y)}. Then the result followsfrom (โˆซ 1

0+โˆซ โˆž

1

)1โ„Ž2๐‘ 

sup๐‘ฅ โˆˆT2

|๐œ•โ„Ž๐‘— ๐‘”(๐‘ฅ) |2dโ„Žโ„Ž.

(1Y+

โˆซ โˆž

1

dโ„Žโ„Ž1+2๐‘ 

)[๐‘”]2๐‘  ๐‘— .

๏ฟฝ

Lemma B.3. For every ๐‘  โˆˆ (0, 1) and ๐›พ โˆˆ (๐‘ , 1], there exists a constant ๐ถ (๐‘ ,๐›พ) > 0 depending onlyon ๐‘  and ๐›พ such that the following duality estimate holds for every ๐‘“ , ๐‘” : T2 โ†’ R,๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซ

T2๐‘“ ๐‘” d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค ๐ถ (๐‘ ,๐›พ) (โ€–|๐œ•1 |๐›พ ๐‘“ โ€–๐ฟ1 + โ€–|๐œ•2 |

23๐›พ ๐‘“ โ€–๐ฟ1 + โ€– ๐‘“ โ€–๐ฟ1

)[๐‘”]โˆ’๐‘  .

Proof. By the mean value theorem and the denition of the kernel๐œ“๐‘‡ we have thatโˆซT2๐‘“ (๐‘” โˆ’ ๐‘”1) d๐‘ฅ =

โˆซ 12

0

โˆซT2๐‘“ ๐œ•๐‘‡๐‘”2๐‘‡ d๐‘ฅ d๐‘‡ =

โˆซ 12

0

โˆซT2๐‘“

(|๐œ•1 |3 โˆ’ ๐œ•22

)๐‘”2๐‘‡ d๐‘ฅ d๐‘‡ .

Page 47: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 47

Let๐œ“ (1) = |๐œ•1 |3โˆ’๐›พ๐œ“ and๐œ“ (2) = |๐œ•2 |2โˆ’23๐›พ๐œ“ . Recalling that |๐œ•1 |๐›ผ๐œ“, |๐œ•2 |๐›ผ๐œ“ โˆˆ ๐ฟ1(R2) for every ๐›ผ โ‰ฅ 0,32

integrating by parts and using the semigroup property we obtain by (A.1) and Remark 1.11,๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซ 12

0

โˆซT2๐‘“

(|๐œ•1 |3 โˆ’ ๐œ•22

)๐‘”2๐‘‡ d๐‘ฅ d๐‘‡

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ =๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซ 1

2

0

โˆซT2

(|๐œ•1 |3 โˆ’ ๐œ•22

)๐‘“๐‘‡ ๐‘”๐‘‡ d๐‘ฅ d๐‘‡

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ.๐‘ 

โˆซ 12

0

โˆซT2๐‘‡ โˆ’ 3โˆ’๐›พ

3

๏ฟฝ๏ฟฝ๏ฟฝ|๐œ•1 |๐›พ ๐‘“ โˆ—๐œ“ (1)๐‘‡

๏ฟฝ๏ฟฝ๏ฟฝ [๐‘”]โˆ’๐‘ ๐‘‡ โˆ’ ๐‘ 3 d๐‘ฅ d๐‘‡

+โˆซ 1

2

0

โˆซT2๐‘‡ โˆ’ 3โˆ’๐›พ

3

๏ฟฝ๏ฟฝ๏ฟฝ|๐œ•2 | 23๐›พ ๐‘“ โˆ—๐œ“ (2)๐‘‡

๏ฟฝ๏ฟฝ๏ฟฝ [๐‘”]โˆ’๐‘ ๐‘‡ โˆ’ ๐‘ 3 d๐‘ฅ d๐‘‡

.๐‘ (โ€–|๐œ•1 |๐›พ ๐‘“ โ€–๐ฟ1 + โ€–|๐œ•2 |

23๐›พ ๐‘“ โ€–๐ฟ1

)[๐‘”]โˆ’๐‘ 

โˆซ 12

0๐‘‡ โˆ’ 3โˆ’๐›พ+๐‘ 

3 d๐‘‡ .

The proof is complete since the last integral is nite for ๐›พ > ๐‘  and by (A.1) we also have theestimate ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซ

T2๐‘“ ๐‘”1d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค โ€– ๐‘“ โ€–๐ฟ1 โ€–๐‘”1โ€–๐ฟโˆž .๐‘  โ€– ๐‘“ โ€–๐ฟ1 [๐‘”]โˆ’๐‘  .

๏ฟฝ

The following Lemma establishes an optimal Sobolev embedding with respect to our anisotropicmetric. Recall that in our context the (scaling) dimension of the space is dim = 5

2 , and ๐œ•2 costs asmuch as 3

2 of ๐œ•1. Therefore, the critical exponent of the embedding ๐ป 1anisotropic โŠ‚ ๐ฟ2

โˆ— is given by2โˆ— = 2 dim

dimโˆ’2 = 10.

Lemma B.4. There exists a constant ๐ถ > 0 such that for any function ๐‘“ : T2 โ†’ R of vanishing

average in ๐‘ฅ1,

โ€– ๐‘“ โ€–๐ฟ10 โ‰ค ๐ถ(โ€–๐œ•1 ๐‘“ โ€–๐ฟ2 + โ€–|๐œ•2 |

23 ๐‘“ โ€–๐ฟ2

), (B.5)

โ€– ๐‘“ โ€–๐ฟ103โ‰ค ๐ถ

(โ€–|๐œ•1 |

12 ๐‘“ โ€–๐ฟ2 + โ€–|๐œ•2 |

13 ๐‘“ โ€–๐ฟ2

), (B.6)

โ€– ๐‘“ โ€–๐ฟ

51โˆ’2Y

โ‰ค ๐ถ(โ€–|๐œ•1 |

34+๐œ– ๐‘“ โ€–๐ฟ2 + โ€–|๐œ•2 |

23 (

34+๐œ–) ๐‘“ โ€–๐ฟ2

), Y โˆˆ [0, 14 ) . (B.7)

Proof. We rst prove (B.5). For that, by the one-dimensional Gagliardo-Nirenberg-Sobolev in-equalities, using that ๐‘“ (ยท, ๐‘ฅ2) has vanishing average, we have that for every ๐‘ฅ1, ๐‘ฅ2 โˆˆ [0, 1),

โ€– ๐‘“ (ยท, ๐‘ฅ2)โ€–๐ฟโˆž๐‘ฅ1 . โ€–๐œ•1 ๐‘“ (ยท, ๐‘ฅ2)โ€–16๐ฟ2๐‘ฅ1

โ€– ๐‘“ (ยท, ๐‘ฅ2)โ€–56๐ฟ10๐‘ฅ1,

โ€– ๐‘“ (๐‘ฅ1, ยท)โ€–๐ฟโˆž๐‘ฅ2 . โ€–|๐œ•2 |23 ๐‘“ (๐‘ฅ1, ยท)โ€–

38๐ฟ2๐‘ฅ2

โ€– ๐‘“ (๐‘ฅ1, ยท)โ€–58๐ฟ10๐‘ฅ2

+ โ€– ๐‘“ (๐‘ฅ1, ยท)โ€–๐ฟ1๐‘ฅ2 .

Hence, Hรถlderโ€™s inequality implies โ€– ๐‘“ โ€–๐ฟโˆž๐‘ฅ1 ๐ฟ6๐‘ฅ2 . โ€–๐œ•1 ๐‘“ โ€–16๐ฟ2โ€– ๐‘“ โ€–

56๐ฟ10, โ€– ๐‘“ โ€–๐ฟโˆž๐‘ฅ2 ๐ฟ4๐‘ฅ1 . โ€–|๐œ•2 |

23 ๐‘“ โ€–

38๐ฟ2โ€– ๐‘“ โ€–

58๐ฟ10

+ โ€– ๐‘“ โ€–๐ฟ4 . โ€–|๐œ•2 |23 ๐‘“ โ€–

38๐ฟ2โ€– ๐‘“ โ€–

58๐ฟ10

+ โ€– ๐‘“ โ€–๐ฟ10 .

Therefore, we obtain

โ€– ๐‘“ โ€–๐ฟ10 โ‰ค โ€– ๐‘“ โ€–๐ฟโˆž๐‘ฅ1 3

5

๐ฟ6๐‘ฅ2

โ€– ๐‘“ โ€–๐ฟโˆž๐‘ฅ2 25

๐ฟ4๐‘ฅ1

. โ€–๐œ•1 ๐‘“ โ€–110๐ฟ2โ€–|๐œ•2 |

23 ๐‘“ โ€–

320๐ฟ2โ€– ๐‘“ โ€–

34๐ฟ10

+ โ€–๐œ•1 ๐‘“ โ€–110๐ฟ2โ€– ๐‘“ โ€–

910๐ฟ10.

32As explained in [IO19, Proof of Lemma 10], the kernel๐œ“ factorizes into a Gaussian in ๐‘ฅ2 and a kernel ๐œ‘ (๐‘ฅ1) thatis smooth and |๐‘ฅ1 |3๐œ•๐‘˜1๐œ‘ โˆˆ ๐ฟโˆž (R) for every ๐‘˜ โ‰ฅ 0 because in Fourier space b๐‘˜1 ๐‘’

โˆ’|b1 |3 has integrable derivatives up toorder 3.

Page 48: Variational methods for a singular SPDE yielding the

48 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

It follows by Youngโ€™s inequality that for any Y โˆˆ (0, 1) there exists a constant ๐ถ (Y) > 0 such that

โ€– ๐‘“ โ€–๐ฟ10 โ‰ค Yโ€– ๐‘“ โ€–๐ฟ10 +๐ถ (Y)(โ€–๐œ•1 ๐‘“ โ€–๐ฟ2 + โ€–|๐œ•2 |

23 ๐‘“ โ€–๐ฟ2

),

from which (B.5) follows.For the second inequality observe that 3

10 = 12 ยท

110 +

12 ยท

12 , so that (B.6) follows from complex

interpolation (with change of measure). Indeed, by inequality (B.5) in the form

โ€– ๐‘“ โ€–2๐ฟ10 .

โˆ‘๏ธ๐‘˜โˆˆ(2๐œ‹Z)2

( |๐‘˜1 |2 + |๐‘˜2 |43 ) |๐‘“ (๐‘˜) |2,

it follows that Id : ๐ฟ2((2๐œ‹Z)2,๐‘ค d๐œ’) โ†’ ๐ฟ10(T2, d๐‘ฅ) is bounded, where ๐‘ค (๐‘˜) := |๐‘˜1 |2 + |๐‘˜2 |43

and ๐œ’ is the counting measure, while Parsevalโ€™s identity โ€– ๐‘“ โ€–2๐ฟ2

=โˆ‘

๐‘˜โˆˆ(2๐œ‹Z)2 |๐‘“ (๐‘˜) |2 shows thatId : ๐ฟ2((2๐œ‹Z)2, d๐œ’) โ†’ ๐ฟ2(T2, d๐‘ฅ) is bounded. Hence, by interpolation (see [BL76, Corollary5.5.4]), it follows that Id : ๐ฟ2((2๐œ‹Z)2,๐‘ค 1

2 d๐œ’) โ†’ ๐ฟ103 (T2, d๐‘ฅ) is bounded, which implies (B.6). The

bound (B.7) follows similarly by interpolation. ๏ฟฝ

Lemma B.5. There exists a constant ๐ถ > 0 such that for every ๐‘ 1 โˆˆ [0, 1] and ๐‘ 2 โˆˆ [0, 23 ] with๐‘ 1 + 3

2๐‘ 2 โ‰ค 1 and every periodic function ๐‘“ : T2 โ†’ R of vanishing average in ๐‘ฅ1 the following holds

โ€–|๐œ•1 |๐‘ 1 |๐œ•2 |๐‘ 2 ๐‘“ โ€–2๐ฟ2 โ‰ค ๐ถH(๐‘“ ).In particular, any sublevel set of E (respectivelyH ) overW is relatively compact in ๐ฟ2.

Proof. The desired estimate is immediate by Hรถlderโ€™s inequality in Fourier space and (2.15). Toprove that the sublevel sets of E (respectively H ) over W are relatively compact, we notice thatH(๐‘“ ) controls โ€–|๐œ•1 |

12 ๐‘“ โ€–๐ฟ2 and โ€–|๐œ•2 |

12 ๐‘“ โ€–๐ฟ2 , hence also โ€– ๐‘“ โ€–๐ฟ2 since ๐‘“ has vanishing average. By the

compact embedding ๐ป 12 โŠ‚ ๐ฟ2 and (2.14), we deduce that any sublevel set of E (respectively H )

overW is relatively compact in ๐ฟ2. ๏ฟฝ

Below we use the following notation for ๐‘  > 0 and a periodic function ๐‘“ : T2 โ†’ R withvanishing average in ๐‘ฅ1,

โ€–|๐‘“ โ€–|2๐‘  :=โˆซT2

( |๐œ•1 |๐‘  ๐‘“ )2 d๐‘ฅ +โˆซT2

(|๐œ•1 |โˆ’

๐‘ 2 |๐œ•2 |๐‘  ๐‘“

)2d๐‘ฅ . (B.8)

Lemma B.6. Let ๐‘  > 0. There exists a constant๐ถ > 0 such that for any periodic function ๐‘“ : T2 โ†’ Rwith vanishing average in ๐‘ฅ1 there holds

โ€–|๐œ•1 |๐‘  ๐‘“ โ€–๐ฟ2 + โ€–|๐œ•2 |23๐‘  ๐‘“ โ€–๐ฟ2 โ‰ค ๐ถ โ€–|๐‘“ โ€–|๐‘  . (B.9)

Proof. Note that by the denition of โ€–|๐‘“ โ€–|๐‘  we only have to bound โ€–|๐œ•2 |23๐‘  ๐‘“ โ€–๐ฟ2 . โ€–|๐‘“ โ€–|๐‘  . This

follows easily by an application of Hรถlderโ€™s inequality in Fourier space. Indeed, by (2.2) we havethat33 โˆซ

T2| |๐œ•2 |

23๐‘  ๐‘“ |2d๐‘ฅ =

โˆ‘๏ธ๐‘˜โˆˆ(2๐œ‹Z)2

|๐‘˜2 |43๐‘  |๐‘“ (๐‘˜) |2 =

โˆ‘๏ธ๐‘˜โˆˆ(2๐œ‹Z)2

|๐‘˜1 |โˆ’23๐‘  |๐‘˜2 |

43๐‘  |๐‘“ (๐‘˜) | 43 |๐‘˜1 |

23๐‘  |๐‘“ (๐‘˜) | 23

โ‰ค ยฉยญยซโˆ‘๏ธ

๐‘˜โˆˆ(2๐œ‹Z)2|๐‘˜1 |โˆ’๐‘  |๐‘˜2 |2๐‘  |๐‘“ (๐‘˜) |2

ยชยฎยฌ23 ยฉยญยซ

โˆ‘๏ธ๐‘˜โˆˆ(2๐œ‹Z)2

|๐‘˜1 |2๐‘  |๐‘“ (๐‘˜) |2ยชยฎยฌ

13

=

(โˆซT2

| |๐œ•1 |โˆ’๐‘ 2 |๐œ•2 |๐‘  ๐‘“ |2d๐‘ฅ

) 23(โˆซT2

|๐œ•1 ๐‘“ |๐‘  d๐‘ฅ) 1

3

โ‰ค 23

โˆซT2

| |๐œ•1 |โˆ’๐‘ 2 |๐œ•2 |๐‘  ๐‘“ |2d๐‘ฅ + 1

3

โˆซT2

| |๐œ•1 |๐‘  ๐‘“ |2 d๐‘ฅ โ‰ค 23 โ€–|๐‘“ โ€–|

2๐‘  .

๏ฟฝ

33Recall that ๐‘“ has vanishing average in ๐‘ฅ1, in particular ๐‘“ (0, ๐‘˜2) = 0 for all ๐‘˜2 โˆˆ 2๐œ‹Z.

Page 49: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 49

Lemma B.7. Let ๐œ– โˆˆ (0, 14 ]. There exists a constant ๐ถ > 0 such that for any ๐‘“ , ๐‘” : T2 โ†’ R with

vanishing average in ๐‘ฅ1 there holds |๐œ•1 | 12 (๐‘“ ๐‘”) ๐ฟ2

โ‰ค ๐ถ โ€–|๐‘“ โ€–|1โ€–|๐‘”โ€–| 34+๐œ– .

In particular, we have that โ€–|๐œ•1 |12 (๐‘“ ๐‘”)โ€–๐ฟ2 โ‰ค ๐ถH(๐‘“ ) 1

2H(๐‘”) 12 .

Proof. By (2.2), writing ๐‘”โ„Ž := ๐‘”(ยท + โ„Ž), we have |๐œ•1 | 12 (๐‘“ ๐‘”) 2๐ฟ2

=

โˆซR

โˆซT2

|๐œ•โ„Ž1 (๐‘“ ๐‘”) |2 d๐‘ฅdโ„Ž|โ„Ž |2

.

โˆซR

โˆซT2

|๐œ•โ„Ž1 ๐‘“ |2 |๐‘”โ„Ž |2 d๐‘ฅdโ„Ž|โ„Ž |2 +

โˆซR

โˆซT2

|๐œ•โ„Ž1๐‘”|2 |๐‘“ |2 d๐‘ฅdโ„Ž|โ„Ž |2 . (B.10)

To estimate the rst term on the right-hand side of (B.10) we use Lemma B.4. We x ๐œ– โˆˆ (0, 14 ] andlet ๐‘ โ€ฒ๐œ– =

5+^๐œ–2 with ^๐œ– = 10๐œ–

1โˆ’2๐œ– , and ๐‘๐œ– =5+^๐œ–3+^๐œ– > 1. Then 1

๐‘๐œ–+ 1

๐‘โ€ฒ๐œ–= 1, so that by Hรถlderโ€™s inequality,โˆซ

R

โˆซT2

|๐œ•โ„Ž1 ๐‘“ |2 |๐‘”โ„Ž |2 d๐‘ฅdโ„Ž|โ„Ž |2 โ‰ค

โˆซR

(โˆซT2

|๐œ•โ„Ž1 ๐‘“ |2๐‘๐œ– d๐‘ฅ) 1๐‘๐œ–

(โˆซT2

|๐‘”โ„Ž |5+^๐œ– d๐‘ฅ) 2

5+^๐œ– dโ„Ž|โ„Ž |2

= โ€–๐‘”โ€–2๐ฟ5+^๐œ–

โˆซR

(โˆซT2

|๐œ•โ„Ž1 ๐‘“ |2๐‘๐œ– d๐‘ฅ) 1๐‘๐œ– dโ„Ž

|โ„Ž |2 ,

where in the last step we also used translation invariance. Note that the exponent 2๐‘๐œ– โˆˆ (2, 10),so that we may interpolateโˆซ

R

(โˆซT2

|๐œ•โ„Ž1 ๐‘“ |2๐‘๐œ– d๐‘ฅ) 1๐‘๐œ– dโ„Ž

|โ„Ž |2 =

โˆซR

(โˆซT2

|๐œ•โ„Ž1 ๐‘“ |2 d๐‘ฅ)\๐œ– (โˆซ

T2|๐œ•โ„Ž1 ๐‘“ |10 d๐‘ฅ

) 1โˆ’\๐œ–5 dโ„Ž

|โ„Ž |2

. โ€– ๐‘“ โ€–2(1โˆ’\๐œ– )๐ฟ10

โˆซR

(โˆซT2

|๐œ•โ„Ž1 ๐‘“ |2 d๐‘ฅ)\๐œ– dโ„Ž

|โ„Ž |2 ,

with \๐œ– = 12 + ๐œ– . By Jensenโ€™s inequality, the mean-value theorem, and Poincarรฉโ€™s inequality for

functions with zero average in ๐‘ฅ1, it then follows thatโˆซR

(โˆซT2

|๐œ•โ„Ž1 ๐‘“ |2 d๐‘ฅ)\๐œ– dโ„Ž

|โ„Ž |2 . โ€– ๐‘“ โ€–1+2๐œ–๐ฟ2

(โˆซ|โ„Ž | โ‰ฅ1

dโ„Ž|โ„Ž |2

)\๐œ–+ โ€–๐œ•1 ๐‘“ โ€–1+2๐œ–๐ฟ2

(โˆซ|โ„Ž |<1

dโ„Ž|โ„Ž |1โˆ’2๐œ–

)\๐œ–. โ€–๐œ•1 ๐‘“ โ€–1+2๐œ–๐ฟ2 .

By Lemmata B.4 and B.6 we can estimate โ€– ๐‘“ โ€–๐ฟ10 . โ€–|๐‘“ โ€–|1 and โ€–๐‘”โ€–๐ฟ5+^๐œ– . โ€–|๐‘”โ€–| 34+๐œ–

, henceโˆซR

โˆซT2

|๐œ•โ„Ž1 ๐‘“ |2 |๐‘”โ„Ž |2 d๐‘ฅdโ„Ž|โ„Ž |2 . โ€–|๐‘“ โ€–|21โ€–|๐‘”โ€–|23

4+๐œ–.

It remains to bound the second term on the right-hand side of (B.10). Similar to the st term, weget by Hรถlderโ€™s inequality and interpolation thatโˆซ

R

โˆซT2

|๐œ•โ„Ž1๐‘”|2 |๐‘“ |2 d๐‘ฅdโ„Ž|โ„Ž |2 โ‰ค

โˆซR

(โˆซT2

|๐œ•โ„Ž1๐‘”|52 d๐‘ฅ

) 45(โˆซT2

|๐‘“ |10 d๐‘ฅ) 1

5 dโ„Ž|โ„Ž |2

โ‰ค โ€– ๐‘“ โ€–2๐ฟ10

โˆซR

(โˆซT2

|๐œ•โ„Ž1๐‘”|2 d๐‘ฅ) 2

3(โˆซT2

|๐œ•โ„Ž1๐‘”|5 d๐‘ฅ) 2

15 dโ„Ž|โ„Ž |2

. โ€– ๐‘“ โ€–2๐ฟ10 โ€–๐‘”โ€–

23๐ฟ5

โˆซR

(โˆซT2

|๐œ•โ„Ž1๐‘”|2 d๐‘ฅ) 2

3 dโ„Ž|โ„Ž |2 .

Page 50: Variational methods for a singular SPDE yielding the

50 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

If Y < 14 , splitting the integral in โ„Ž, it follows with Jensenโ€™s inequality thatโˆซR

(โˆซT2

|๐œ•โ„Ž1๐‘”|2 d๐‘ฅ) 2

3 dโ„Ž|โ„Ž |2 =

โˆซ|โ„Ž | โ‰ฅ1

(โˆซT2

|๐œ•โ„Ž1๐‘”|2 d๐‘ฅ) 2

3 dโ„Ž|โ„Ž |2 +

โˆซ|โ„Ž |<1

(โˆซT2

|๐œ•โ„Ž1๐‘”|2 d๐‘ฅ) 2

3 dโ„Ž|โ„Ž |2

. โ€–๐‘”โ€–43๐ฟ2

(โˆซ|โ„Ž | โ‰ฅ1

dโ„Ž|โ„Ž |2

) 23+

โˆซ|โ„Ž |<1

(โˆซT2

|๐œ•โ„Ž1๐‘”|2

|โ„Ž | 32+6๐œ–d๐‘ฅ

) 23 dโ„Ž|โ„Ž |1โˆ’4๐œ–

. โ€–๐‘”โ€–43๐ฟ2

+(โˆซ

|โ„Ž |<1

โˆซT2

|๐œ•โ„Ž1๐‘”|2

|โ„Ž | 32+2๐œ–d๐‘ฅ dโ„Ž|โ„Ž |

) 23

. โ€–๐‘”โ€–43๐ฟ2

+ โ€–|๐œ•1 |34+๐œ–๐‘”โ€–

43๐ฟ2.

Using again Lemmata B.4 and B.6, we may bound โ€– ๐‘“ โ€–๐ฟ10 . โ€–|๐‘“ โ€–|1 and โ€–๐‘”โ€–๐ฟ5 . โ€–|๐‘”โ€–| 34+๐œ–

, theconclusion follows with โ€–๐‘”โ€–๐ฟ2 โ‰ค โ€–|๐‘”โ€–| 3

4+๐œ–. If Y = 1

4 , we use instead the estimateโˆซ|โ„Ž |<1

(โˆซT2

|๐œ•โ„Ž1๐‘”|2 d๐‘ฅ) 2

3 dโ„Ž|โ„Ž |2 โ‰ค โ€–๐œ•1๐‘”โ€–

43๐ฟ2

โˆซ|โ„Ž |<1

dโ„Ž|โ„Ž | 23. โ€–|๐‘”โ€–|

431 .

๏ฟฝ

As for the Sobolev embedding in Lemma B.4, we next prove the embedding ๐ป 1anistropic โŠ‚ C๐›ผ ,

which is optimal and the critical exponent is given by ๐›ผ = 1 โˆ’ dim2 = โˆ’ 1

4 .

Lemma B.8 (Besov embedding into Hรถlder spaces). There exists a constant ๐ถ > 0 such that for

any periodic distribution ๐‘“ : T2 โ†’ R of vanishing average, and ๐‘  โˆˆ (0, 114 ) \ {14 ,

34 ,

54 ,

94 }, there holds

[๐‘“ ]โˆ’ 54+๐‘ 

โ‰ค ๐ถ (โ€– |๐œ•1 |๐‘  ๐‘“ โ€–๐ฟ2 + โ€–|๐œ•2 |23๐‘  ๐‘“ โ€–๐ฟ2), (B.11)

In particular, [๐‘ค]โˆ’ 14โ‰ค ๐ถH(๐‘ค) 1

2 for every๐‘ค โˆˆ W.

Proof. Since ๐‘“ is of vanishing average, by (A.3) and (A.5) we know that for ๐‘  โˆˆ (0, 114 ) \ {14 ,

34 ,

54 ,

94 },

[๐‘“ ]โˆ’ 54+๐‘ 

โˆผ sup๐‘‡ โˆˆ(0,1]

(๐‘‡

13) 54โˆ’๐‘  โ€–๐‘‡A ๐‘“๐‘‡ โ€–๐ฟโˆž .

Writing |๐œ•1 |3 ๐‘“๐‘‡ = |๐œ•1 |๐‘  ๐‘“ โˆ— |๐œ•1 |3โˆ’๐‘ ๐œ“๐‘‡ and ๐œ•22 ๐‘“๐‘‡ = |๐œ•2 |23๐‘  ๐‘“ โˆ— |๐œ•2 |2โˆ’

23๐‘ ๐œ“๐‘‡ , and using that |๐œ•๐‘— |๐›ผ๐œ“ โˆˆ ๐ฟ2(R2)

for every ๐›ผ โ‰ฅ 0 and ๐‘— = 1, 2, we deduce with Youngโ€™s inequality for convolution of functions (asin Remark 1.11) that

โ€–A ๐‘“๐‘‡ โ€–๐ฟโˆž โ‰ค โ€–|๐œ•1 |๐‘  ๐‘“ โ€–๐ฟ2 โ€–|๐œ•1 |3โˆ’๐‘ ๐œ“๐‘‡ โ€–๐ฟ2 + โ€–|๐œ•2 |23๐‘  ๐‘“ โ€–๐ฟ2 โ€–|๐œ•2 |2โˆ’

23๐‘ ๐œ“๐‘‡ โ€–๐ฟ2

.(๐‘‡

13)โˆ’3+๐‘ โˆ’ 5

4(โ€–|๐œ•1 |๐‘  ๐‘“ โ€–๐ฟ2 + โ€–|๐œ•2 |

23๐‘  ๐‘“ โ€–๐ฟ2

).

This implies that

[๐‘“ ]โˆ’ 54+๐‘ 

โˆผ sup๐‘‡ โˆˆ(0,1]

(๐‘‡

13) 54โˆ’๐‘  โ€–๐‘‡A ๐‘“๐‘‡ โ€–๐ฟโˆž . โ€–|๐œ•1 |๐‘  ๐‘“ โ€–๐ฟ2 + โ€–|๐œ•2 |

23๐‘  ๐‘“ โ€–๐ฟ2 .

Combining (B.11) for ๐‘  = 1 and (2.15), we conclude that [๐‘ค]โˆ’ 14. H(๐‘ค) 1

2 for every๐‘ค โˆˆ W. ๏ฟฝ

The next proposition is the classical 1-dimensional embedding of Besov spaces into ๐ฟ๐‘ spacesin the periodic setting. 34

34The nonperiodic version of the statement is essentially a combination of [BCD11, Proposition 2.20 and Theorem2.36].

Page 51: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 51

Proposition B.9. For every ๐‘ โˆˆ (1,โˆž] and ๐‘ž โˆˆ [1, ๐‘] with (๐‘, ๐‘ž) โ‰  (โˆž, 1), there exists a constant๐ถ (๐‘, ๐‘ž) > 0 such that for every periodic ๐‘“ : [0, 1) โ†’ R with vanishing average(โˆซ 1

0|๐‘“ (๐‘ง) |๐‘ d๐‘ง

) 1๐‘

โ‰ค ๐ถ (๐‘, ๐‘ž)โˆซ 1

0

1

โ„Ž1๐‘žโˆ’ 1

๐‘

(โˆซ 1

0|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง

) 1๐‘ž dโ„Žโ„Ž, (B.12)

with the usual interpretation for ๐‘ = โˆž or ๐‘ž = โˆž.

Proof. We prove the statement for ๐‘ โˆˆ (1,โˆž) (the case ๐‘ = โˆž follows similarly). For ๐‘‡ โˆˆ (0, 1] let๐œ‘๐‘‡ (โ„Ž) = 1โˆš

4๐œ‹๐‘‡eโˆ’โ„Ž2

4๐‘‡ , โ„Ž โˆˆ R, be the heat semigroup, and denote its periodization by

ฮฆ๐‘‡ (โ„Ž) =1

โˆš4๐œ‹๐‘‡

โˆ‘๏ธ๐‘˜โˆˆZ

eโˆ’(โ„Žโˆ’๐‘˜ )2

4๐‘‡ =โˆ‘๏ธ๐‘˜โˆˆZ

eโˆ’4๐œ‹2๐‘˜2๐‘‡ e2๐œ‹๐‘–๐‘˜โ„Ž, โ„Ž โˆˆ [0, 1).

Note that ฮฆ๐‘‡ is smooth, ฮฆ๐‘‡ โ‰ฅ 0, โ€–ฮฆ๐‘‡ โ€–๐ฟ1 =โˆซ 10 ฮฆ๐‘‡ dโ„Ž = 1, and for every ๐‘‡ โˆˆ (0, 1]

โ€–ฮฆ๐‘‡ โ€–๐ฟโˆž โ‰ค 1 +โˆ‘๏ธ

๐‘˜โˆˆZ\{0}eโˆ’4๐œ‹2๐‘˜2๐‘‡ . 1 +

โˆซR๐œ‘๐‘‡ (b) db . 1 + ๐œ‘๐‘‡ (0) .

1โˆš๐‘‡.

Therefore, by interpolation, for every ๐‘Ÿ โˆˆ [1,โˆž] and ๐‘‡ โˆˆ (0, 1] we have

โ€–ฮฆ๐‘‡ โ€–๐ฟ๐‘Ÿ .โˆš๐‘‡โˆ’(1โˆ’ 1

๐‘Ÿ). (B.13)

We also claim that for every ๐‘‡ โˆˆ (0, 1]

sup|โ„Ž | โ‰ค 1

2

โ„Ž2โˆš๐‘‡ฮฆ๐‘‡ (โ„Ž) . 1. (B.14)

Indeed, using that for every ๐‘˜ โˆˆ Z \ {0}, |โ„Ž | โ‰ค 12 , and ๐‘‡ โˆˆ (0, 1], (โ„Žโˆ’๐‘˜)2

4๐‘‡ โ‰ฅ ( |๐‘˜ |โˆ’ 12 )2

4๐‘‡ โ‰ฅ 116 , we

obtain35 eโˆ’(โ„Žโˆ’๐‘˜ )2

4๐‘‡ . 4๐‘‡(โ„Žโˆ’๐‘˜)2 .

๐‘‡๐‘˜2 . It then follows that

โˆš4๐œ‹โ„Ž2โˆš๐‘‡

ฮฆ๐‘‡ (โ„Ž) =โ„Ž2

๐‘‡eโˆ’

โ„Ž24๐‘‡ +

โˆ‘๏ธ๐‘˜โˆˆZ\{0}

โ„Ž2

๐‘‡eโˆ’

(โ„Žโˆ’๐‘˜ )24๐‘‡ . sup

๐‘งโˆˆR๐‘ง2eโˆ’

๐‘ง24 + โ„Ž2

โˆ‘๏ธ๐‘˜โˆˆZ\{0}

1๐‘˜2. 1.

By the semigroup property and the periodicity of ๐‘“ we know that for ๐‘ง โˆˆ (โˆ’ 12 ,

12 ),

๐‘“ โˆ— ฮฆ2๐‘‡ (๐‘ง) โˆ’ ๐‘“ โˆ— ฮฆ๐‘‡ (๐‘ง) =โˆซ 1

2

โˆ’ 12

(๐œ•โˆ’โ„Ž1 ๐‘“ โˆ— ฮฆ๐‘‡ (๐‘ง)

)ฮฆ๐‘‡ (โ„Ž) dโ„Ž.

Using Minkowskiโ€™s inequality and Youngโ€™s inequality for convolution with exponents 1+ 1๐‘= 1

๐‘Ÿ+ 1๐‘ž

with ๐‘Ÿ โˆˆ [1,โˆž) we deduce by (B.13),(โˆซ 12

โˆ’ 12

|๐‘“ โˆ— ฮฆ2๐‘‡ (๐‘ง) โˆ’ ๐‘“ โˆ— ฮฆ๐‘‡ (๐‘ง) |๐‘ d๐‘ง) 1

๐‘

.

โˆซ 12

โˆ’ 12

ฮฆ๐‘‡ (โ„Ž)(โˆซ 1

2

โˆ’ 12

|๐œ•โˆ’โ„Ž1 ๐‘“ โˆ— ฮฆ๐‘‡ (๐‘ง) |๐‘ d๐‘ง) 1

๐‘

dโ„Ž

.

(โˆซ 12

โˆ’ 12

|ฮฆ๐‘‡ (๐‘ง) |๐‘Ÿ d๐‘ง) 1๐‘Ÿ โˆซ 1

2

โˆ’ 12

ฮฆ๐‘‡ (โ„Ž)(โˆซ 1

2

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง) 1๐‘ž

dโ„Ž

.1

โˆš๐‘‡

(1๐‘žโˆ’ 1

๐‘

) โˆซ 12

โˆ’ 12

ฮฆ๐‘‡ (โ„Ž)(โˆซ 1

2

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง) 1๐‘ž

dโ„Ž, (B.15)

where we also used a change of variables and periodicity to replace โˆ’โ„Ž by โ„Ž.

35Recall that eโˆ’๐‘ฅ . 1๐‘ฅ for ๐‘ฅ โ‰ฅ 1

16 .

Page 52: Variational methods for a singular SPDE yielding the

52 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

To prove (B.12) we now write

๐‘“ (๐‘ง) =โˆžโˆ‘๏ธ๐‘˜=1

(๐‘“ โˆ— ฮฆ2ยท2โˆ’๐‘˜ (๐‘ง) โˆ’ ๐‘“ โˆ— ฮฆ2โˆ’๐‘˜ (๐‘ง)) โˆ’ ๐‘“ โˆ— ฮฆ1(๐‘ง),

and obtain by (B.15) that(โˆซ 1

0|๐‘“ (๐‘ง) |๐‘ d๐‘ง

) 1๐‘

=

(โˆซ 12

โˆ’ 12

|๐‘“ (๐‘ง) |๐‘ d๐‘ง) 1

๐‘

. ฮฃ(1) + ฮฃ(2) +(โˆซ 1

2

โˆ’ 12

|๐‘“ โˆ— ฮฆ1(๐‘ง) |๐‘ d๐‘ง) 1

๐‘

, (B.16)

where

ฮฃ(1) =โˆžโˆ‘๏ธ๐‘˜=1

โˆš2๐‘˜

(1๐‘žโˆ’ 1

๐‘

) โˆซ{ |โ„Ž | โ‰ค 1

2 : |โ„Ž |โˆš2๐‘˜ โ‰ค1}

ฮฆ2โˆ’๐‘˜ (โ„Ž)(โˆซ 1

2

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง) 1๐‘ž

dโ„Ž,

ฮฃ(2) =โˆžโˆ‘๏ธ๐‘˜=1

โˆš2๐‘˜

(1๐‘žโˆ’ 1

๐‘

) โˆซ{ |โ„Ž | โ‰ค 1

2 : |โ„Ž |โˆš2๐‘˜>1}

ฮฆ2โˆ’๐‘˜ (โ„Ž)(โˆซ 1

2

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง) 1๐‘ž

dโ„Ž.

Using that by (B.13), sup |โ„Ž | โ‰ค 12|ฮฆ2โˆ’๐‘˜ (โ„Ž) | .

โˆš2๐‘˜ , we can estimate ฮฃ(1) as follows:

ฮฃ(1) .

โˆซ 12

โˆ’ 12

โˆ‘๏ธ{๐‘˜โ‰ฅ1: |โ„Ž |

โˆš2๐‘˜ โ‰ค1}

โˆš2๐‘˜

(1๐‘žโˆ’ 1

๐‘+1

) (โˆซ 12

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง) 1๐‘ž

dโ„Ž

.

โˆซ 12

โˆ’ 12

1

|โ„Ž |1๐‘žโˆ’ 1

๐‘

(โˆซ 12

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง) 1๐‘ž dโ„Ž

|โ„Ž | .

For ฮฃ(2) , by (B.14) and the fact that 1๐‘žโˆ’ 1

๐‘โˆ’ 1 < 0 (since (๐‘, ๐‘ž) โ‰  (โˆž, 1)), we have

ฮฃ(2) .โˆžโˆ‘๏ธ๐‘˜=1

โˆš2๐‘˜

(1๐‘žโˆ’ 1

๐‘โˆ’1

) โˆซ{ |โ„Ž | โ‰ค 1

2 : |โ„Ž |โˆš2๐‘˜>1}

1|โ„Ž |

(โˆซ 12

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง) 1๐‘ž dโ„Ž

|โ„Ž |

. 1๐‘žโˆ’ 1

๐‘

โˆซ 12

โˆ’ 12

1

|โ„Ž |1๐‘žโˆ’ 1

๐‘

(โˆซ 12

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง) 1๐‘ž dโ„Ž

|โ„Ž | .

For the last term in (B.16) we use that ๐‘“ has vanishing average and periodicity, which implies thatโˆซ 12

โˆ’ 12๐‘“ โˆ— ฮฆ1(๐‘ง + โ„Ž) dโ„Ž = 0 for every ๐‘ง โˆˆ (โˆ’ 1

2 ,12 ), to obtain the bound(โˆซ 1

2

โˆ’ 12

|๐‘“ โˆ— ฮฆ1(๐‘ง) |๐‘ d๐‘ง) 1

๐‘

=

(โˆซ 12

โˆ’ 12

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๐‘“ โˆ— ฮฆ1(๐‘ง) โˆ’โˆซ 1

2

โˆ’ 12

๐‘“ โˆ— ฮฆ1(๐‘ง + โ„Ž) dโ„Ž๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๐‘ d๐‘ง

) 1๐‘

โ‰คโˆซ 1

2

โˆ’ 12

(โˆซ 12

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ โˆ— ฮฆ1(๐‘ง) |๐‘ d๐‘ง) 1

๐‘

dโ„Ž .โˆซ 1

2

โˆ’ 12

(โˆซ 12

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง) 1๐‘ž

dโ„Ž

.

โˆซ 12

โˆ’ 12

1

|โ„Ž |1๐‘žโˆ’ 1

๐‘

(โˆซ 12

โˆ’ 12

|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ž d๐‘ง) 1๐‘ž dโ„Ž|โ„Ž | , (B.17)

where we also used Minkowskiโ€™s inequality, Youngโ€™s inequality for convolution and the fact that1๐‘žโˆ’ 1

๐‘+ 1 > 0. The right hand side of (B.17) is estimated by twice the right hand side of (B.12), so

the conclusion follows. ๏ฟฝ

The next lemma allows us to connect the estimate (2.4) with regularity in Besov spaces.

Page 53: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 53

Lemma B.10. There exists a constant ๐ถ > 0 such that for every ๐‘  โˆˆ (0, 1), ๐‘ โˆˆ [1,โˆž) and everyperiodic function ๐‘“ : [0, 1) โ†’ R the following estimate holds:

supโ„Žโˆˆ(0,1]

1โ„Ž๐‘ 

(โˆซ 1

0|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ d๐‘ง

) 1๐‘

โ‰ค ๐ถ supโ„Žโˆˆ(0,1]

1โ„Ž๐‘ 

(1โ„Ž

โˆซ โ„Ž

0

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘“ (๐‘ง) |๐‘ d๐‘ง dโ„Žโ€ฒ

) 1๐‘

. (B.18)

Proof. Let โ„Ž โˆˆ (0, 1]. Then for โ„Žโ€ฒ โˆˆ (0, โ„Ž] we have thatโˆซ 1

0|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ d๐‘ง โ‰ค 2๐‘โˆ’1

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘“ (๐‘ง) |๐‘ d๐‘ง + 2๐‘โˆ’1

โˆซ 1

0|๐œ•โ„Žโˆ’โ„Žโ€ฒ1 ๐‘“ (๐‘ง + โ„Žโ€ฒ) |๐‘ d๐‘ง.

Integrating over โ„Žโ€ฒ โˆˆ [โ„Ž2 , โ„Ž] we obtain that

โ„Ž

โˆซ 1

0|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ d๐‘ง โ‰ค 2๐‘

โˆซ โ„Ž

โ„Ž2

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘“ (๐‘ง) |๐‘ d๐‘ง dโ„Žโ€ฒ + 2๐‘

โˆซ โ„Ž

โ„Ž2

โˆซ 1

0|๐œ•โ„Žโˆ’โ„Žโ€ฒ1 ๐‘“ (๐‘ง + โ„Žโ€ฒ) |๐‘ d๐‘ง dโ„Žโ€ฒ.

By the change of variables โ„Žโ€ฒโ€ฒ = โ„Ž โˆ’ โ„Žโ€ฒ and ๐‘ง โ€ฒ = ๐‘ง + โ„Ž โˆ’ โ„Žโ€ฒโ€ฒ, upon relabelling, we see thatโˆซ โ„Ž

โ„Ž2

โˆซ 1

0|๐œ•โ„Žโˆ’โ„Žโ€ฒ1 ๐‘“ (๐‘ง + โ„Žโ€ฒ) |๐‘ d๐‘ง dโ„Žโ€ฒ =

โˆซ โ„Ž2

0

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘“ (๐‘ง) |๐‘ d๐‘ง dโ„Žโ€ฒ,

where we also used periodicity in ๐‘ง. Hence we have proved thatโˆซ 1

0|๐œ•โ„Ž1 ๐‘“ (๐‘ง) |๐‘ d๐‘ง โ‰ค 2๐‘

โ„Ž

โˆซ โ„Ž

0

โˆซ 1

0|๐œ•โ„Žโ€ฒ1 ๐‘“ (๐‘ง) |๐‘ d๐‘ง dโ„Žโ€ฒ,

which in turn implies (B.18). ๏ฟฝ

Appendix C. Stochastic estimates

We show that solutions of the linearized equationL๐‘ฃ = ๐‘ƒb (C.1)

almost surely have (negative) innite total energy (see (1.1) for the denition) under the law ofwhite noise.

Proposition C.1. Assume that ใ€ˆยทใ€‰ is the law of white noise. If ๐‘ฃ is the solution of vanishing average

in ๐‘ฅ1-direction to L๐‘ฃ = ๐‘ƒb , then ๐ธ๐‘ก๐‘œ๐‘ก (๐‘ฃ) = โˆ’โˆž ใ€ˆยทใ€‰-almost surely.

Proof. Recall that in Fourier space we have an explicit representation of the solution to (C.1) as

๏ฟฝฬ‚๏ฟฝ (๐‘˜) = bฬ‚ (๐‘˜)๐‘˜21 + |๐‘˜1 |โˆ’1๐‘˜22

for ๐‘˜1 โ‰  0 and ๏ฟฝฬ‚๏ฟฝ (0, ๐‘˜2) = 0 for all ๐‘˜2 โˆˆ 2๐œ‹Z.

A short calculation shows that the harmonic part of the energy is

H(๐‘ฃ) =โˆซT2b๐‘ฃ d๐‘ฅ,

so that

๐ธ๐‘ก๐‘œ๐‘ก (๐‘ฃ) = โˆ’H(๐‘ฃ) +โˆซT2

(|๐œ•1 |โˆ’

12 ๐œ•1

๐‘ฃ2

2

)2d๐‘ฅ โˆ’ 2

โˆซT2

(|๐œ•1 |โˆ’

12 ๐œ•1

๐‘ฃ2

2

) (|๐œ•1 |โˆ’

12 ๐œ•2๐‘ฃ

)d๐‘ฅ .

By Youngโ€™s inequality, we have

๐ธ๐‘ก๐‘œ๐‘ก (๐‘ฃ) โ‰ค โˆ’12H(๐‘ฃ) + 3โˆซT2

(|๐œ•1 |โˆ’

12 ๐œ•1

๐‘ฃ2

2

)2d๐‘ฅ . (C.2)

By (B.1), (B.4) and [IO19, Lemma 12], we may estimateโˆซT2

(|๐œ•1 |โˆ’

12 ๐œ•1

๐‘ฃ2

2

)2d๐‘ฅ =

โˆซT2

(๐‘…1 |๐œ•1 |

12๐‘ฃ2

2

)2d๐‘ฅ =

โˆซT2

(|๐œ•1 |

12๐‘ฃ2

2

)2d๐‘ฅ . [๐‘ฃ2]21

2+Y. [๐‘ฃ]43

4โˆ’Y,

where [๐‘ฃ] 34โˆ’Y

is nite ใ€ˆยทใ€‰-almost surely by (5.7)

Page 54: Variational methods for a singular SPDE yielding the

54 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

We next show that

H(๐‘ฃ) =โˆ‘๏ธ๐‘˜1โ‰ 0

|bฬ‚ (๐‘˜) |2

๐‘˜21 + |๐‘˜1 |โˆ’1๐‘˜22

diverges ใ€ˆยทใ€‰-almost surely. Since bฬ‚ (โˆ’๐‘˜) = โˆ’bฬ‚ (๐‘˜), we have that

H(๐‘ฃ) = 2โˆ‘๏ธ

๐‘˜โˆˆ2๐œ‹Z2\{๐‘˜1โ‰ค0}

|bฬ‚ (๐‘˜) |2

๐‘˜21 + |๐‘˜1 |โˆ’1๐‘˜22=: 2

โˆ‘๏ธ๐‘˜โˆˆ2๐œ‹Z2\{๐‘˜1โ‰ค0}

๐‘Ž๐‘˜ |bฬ‚ (๐‘˜) |2.

By the independence of {|bฬ‚ (๐‘˜) |}๐‘˜โˆˆ2๐œ‹Z2\{๐‘˜1โ‰ค0} and Kolmogorovโ€™s 0-1 law, we know that the probabil-ity of the event {H (๐‘ฃ) = +โˆž} is either 1 or 0. Hence, it is enough to show that ใ€ˆ{H (๐‘ฃ) = +โˆž}ใ€‰ > 0.We rst notice thatโˆ‘๏ธ

๐‘˜โˆˆ2๐œ‹Z2\{๐‘˜1โ‰ค0}๐‘Ž๐‘˜ |bฬ‚ (๐‘˜) |2 โ‰ฅ

โˆ‘๏ธ๐‘˜โˆˆ2๐œ‹Z2\{๐‘˜1โ‰ค0}

๐‘Ž๐‘˜1{ |bฬ‚ (๐‘˜) |2โ‰ฅ1} =:โˆ‘๏ธ

๐‘˜โˆˆ2๐œ‹Z2\{๐‘˜1โ‰ค0}๐‘Ž๐‘˜๐‘‹๐‘˜

and since the random variables b (๐‘˜) are identically distributed, there exists ๐‘ โˆˆ (0, 1) suchthat ใ€ˆ๐‘‹๐‘˜ใ€‰ = ๐‘ for every ๐‘˜ โˆˆ 2๐œ‹Z2 \ {๐‘˜1 โ‰ค 0}. Given ๐‘€ โ‰ฅ 1, there exists a nite subset๐ฝ โŠ‚ 2๐œ‹Z2 \ {๐‘˜1 โ‰ค 0} such that ๐‘

2โˆ‘

๐‘˜โˆˆ๐ฝ ๐‘Ž๐‘˜ โ‰ฅ ๐‘€ and the following estimate holds

๐‘โˆ‘๏ธ๐‘˜โˆˆ๐ฝ

๐‘Ž๐‘˜ =

โŸจโˆ‘๏ธ๐‘˜โˆˆ๐ฝ

๐‘Ž๐‘˜๐‘‹๐‘˜

โŸฉ

โ‰ค ๐‘

2โˆ‘๏ธ๐‘˜โˆˆ๐ฝ

๐‘Ž๐‘˜ยฉยญยซ1 โˆ’

โŸจ{โˆ‘๏ธ๐‘˜โˆˆ๐ฝ

๐‘Ž๐‘˜๐‘‹๐‘˜ โ‰ฅ ๐‘€

}โŸฉยชยฎยฌ +โˆ‘๏ธ๐‘˜โˆˆ๐ฝ

๐‘Ž๐‘˜

โŸจ{โˆ‘๏ธ๐‘˜โˆˆ๐ฝ

๐‘Ž๐‘˜๐‘‹๐‘˜ โ‰ฅ ๐‘€

}โŸฉ.

Then, it is easy to see that โŸจ{โˆ‘๏ธ๐‘˜โˆˆ๐ฝ

๐‘Ž๐‘˜๐‘‹๐‘˜ โ‰ฅ ๐‘€

}โŸฉโ‰ฅ ๐‘

2 โˆ’ ๐‘ > 0,

which in turn impliesโŸจ{ โˆ‘๏ธ๐‘˜โˆˆ2๐œ‹Z2\{๐‘˜1โ‰ค0}

๐‘Ž๐‘˜๐‘‹๐‘˜ = +โˆž}โŸฉ

= lim๐‘€โ†‘+โˆž

โŸจ{ โˆ‘๏ธ๐‘˜โˆˆ2๐œ‹Z2\{๐‘˜1โ‰ค0}

๐‘Ž๐‘˜๐‘‹๐‘˜ โ‰ฅ ๐‘€

}โŸฉโ‰ฅ ๐‘

2 โˆ’ ๐‘ > 0.

Thus, we obtain that

ใ€ˆ{H (๐‘ฃ) = +โˆž}ใ€‰ โ‰ฅโŸจ{ โˆ‘๏ธ

๐‘˜โˆˆ2๐œ‹Z2\{๐‘˜1โ‰ค0}๐‘Ž๐‘˜๐‘‹๐‘˜ = +โˆž

}โŸฉ> 0,

which proves the desired claim. ๏ฟฝ

The next lemma is a Kolmogorov-type criterion for periodic random elds.

Lemma C.2. Let {๐‘”(๐‘ฅ)}๐‘ฅ โˆˆT2 be a random eld and assume that for some ๐›ผ โˆˆ (0, 32 ) and every

1 โ‰ค ๐‘ < โˆž

sup๐‘‡ โˆˆ(0,1]

(๐‘‡

13)๐›ผ

sup๐‘ฅ โˆˆT2

ใ€ˆ|๐‘”๐‘‡ (๐‘ฅ) |๐‘ใ€‰1๐‘ < โˆž.

Then, for every Y โˆˆ (0, 32 โˆ’ ๐›ผ), ๐‘” โˆˆ Cโˆ’๐›ผโˆ’Yalmost surely and for every 1 โ‰ค ๐‘ < โˆž there exists

๐ถ (Y, ๐‘) > 0 such that โŸจ[๐‘”]๐‘โˆ’๐›ผโˆ’Y

โŸฉ 1๐‘ โ‰ค ๐ถ (Y, ๐‘) .

Page 55: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 55

Proof. First assume that ๐›ผ + Y โ‰  1, 12 . Let ๐‘Y >52Y . We rst claim that

[๐‘”]โˆ’๐›ผโˆ’Y .โˆ‘๏ธ๐‘›โ‰ฅ1

(2

13)โˆ’๐‘› (

Yโˆ’ 52๐‘Y

) (2

13)โˆ’๐‘›๐›ผ

โ€–๐‘”2โˆ’๐‘› โ€–๐ฟ๐‘Y .

Indeed, for every ๐‘‡ โˆˆ (0, 1] we can nd ๐‘› โ‰ฅ 1 such that 2โˆ’๐‘› < ๐‘‡ โ‰ค 2โˆ’๐‘›+1 and by the semigroupproperty and Remark 1.11, we obtain(

๐‘‡13)๐›ผ+Y

โ€–๐‘”๐‘‡ โ€–๐ฟโˆž .(2

13)โˆ’๐‘› (๐›ผ+Y)

โ€–(๐‘”2โˆ’๐‘› )๐‘‡โˆ’2โˆ’๐‘› โ€–๐ฟโˆž .(2

13)โˆ’๐‘› (๐›ผ+Y)

โ€–๐‘”2โˆ’๐‘› โ€–๐ฟโˆž

.(2

13)โˆ’๐‘› (๐›ผ+Y) (

213) (๐‘›+1) 5

2๐‘Y โ€–๐‘”2โˆ’๐‘›โˆ’1 โ€–๐ฟ๐‘Y

.โˆ‘๏ธ๐‘›โ‰ฅ1

(2

13)โˆ’๐‘› (

Yโˆ’ 52๐‘Y

) (2

13)โˆ’๐‘›๐›ผ

โ€–๐‘”2โˆ’๐‘› โ€–๐ฟ๐‘Y

and, taking the supremum over all ๐‘‡ โˆˆ (0, 1], we conclude the above claim via (A.1).Then, for every ๐‘ โ‰ฅ ๐‘Y , Minkowskiโ€™s and Jensenโ€™s inequality imply

โŸจ[๐‘”]๐‘โˆ’๐›ผโˆ’Y

โŸฉ 1๐‘ .

โŸจ(โˆ‘๏ธ๐‘›โ‰ฅ1

(2

13)โˆ’๐‘› (

Yโˆ’ 52๐‘Y

) (2

13)โˆ’๐‘›๐›ผ

โ€–๐‘”2โˆ’๐‘› โ€–๐ฟ๐‘Y)๐‘โŸฉ 1

๐‘

โ‰คโˆ‘๏ธ๐‘›โ‰ฅ1

(2

13)โˆ’๐‘› (

Yโˆ’ 52๐‘Y

) (2

13)โˆ’๐‘›๐›ผ โŸจ

โ€–๐‘”2โˆ’๐‘› โ€–๐‘๐ฟ๐‘YโŸฉ 1๐‘

โ‰คโˆ‘๏ธ๐‘›โ‰ฅ1

(2

13)โˆ’๐‘› (

Yโˆ’ 52๐‘Y

) (2

13)โˆ’๐‘›๐›ผ โŸจ

โ€–๐‘”2โˆ’๐‘› โ€–๐‘๐ฟ๐‘โŸฉ 1๐‘

.โˆ‘๏ธ๐‘›โ‰ฅ1

(2

13)โˆ’๐‘› (

Yโˆ’ 52๐‘Y

)sup

๐‘‡ โˆˆ(0,1]

(๐‘‡

13)๐›ผ

sup๐‘ฅ โˆˆT2

ใ€ˆ|๐‘”๐‘‡ (๐‘ฅ) |๐‘ใ€‰1๐‘ ,

where in the nal step we used the estimateโŸจโ€–๐‘”๐‘‡ โ€–๐‘๐ฟ๐‘

โŸฉ 1๐‘ โ‰ค sup๐‘ฅ โˆˆT2 ใ€ˆ|๐‘”๐‘‡ (๐‘ฅ) |๐‘ใ€‰

1๐‘ . As Y > 5

2๐‘Y , ourhypothesis yields the conclusion. For ๐‘ โˆˆ [1, ๐‘Y), one concludes via Jensenโ€™s inequality.

In the critical case ๐›ผ + Y = 1, 12 , one considers ๐›พ < ๐›ผ + Y and applies the above to conclude๐‘” โˆˆ Cโˆ’๐›พ , which together with [IO19, Remark 2] gives ๐‘” โˆˆ Cโˆ’๐›ผโˆ’Y . ๏ฟฝ

Appendix D. Some estimates for the linear eqation

Lemma D.1. There exists ๐ถ > 0 such that for every Y โˆˆ (0, 18 ) and every b โˆˆ Cโˆ’ 54โˆ’Y , the solution ๐‘ฃ

of vanishing average in ๐‘ฅ1 to the equation L๐‘ฃ = ๐‘ƒb satises |๐œ•1 |โˆ’1๐œ•2๐‘ฃ โˆˆ C 14โˆ’Y with[

|๐œ•1 |โˆ’1๐œ•2๐‘ฃ]14โˆ’Y

โ‰ค ๐ถ [b]โˆ’ 54โˆ’Y.

Proof. Recalling our notation A = |๐œ•1 |L, we have that

๐‘ƒb๐‘‡ = L๐‘ฃ๐‘‡ = |๐œ•1 |โˆ’1A๐‘ฃ๐‘‡ .

For ๐‘” = |๐œ•1 |โˆ’1๐œ•2๐‘ฃ , this yields

A๐‘”๐‘‡ = A|๐œ•1 |โˆ’1๐œ•2๐‘ฃ๐‘‡ = ๐œ•2๐‘ƒb๐‘‡ .

Hence, for ๐‘‡ โˆˆ (0, 1] we have that

โ€–๐‘‡A๐‘”๐‘‡ โ€–๐ฟโˆž = ๐‘‡ โ€–๐œ•2๐‘ƒb๐‘‡ โ€–๐ฟโˆž = ๐‘‡ โ€–๐‘ƒb ๐‘‡2โˆ— ๐œ•2๐œ“๐‘‡

2โ€–๐ฟโˆž = 2

12(๐‘‡

13)3โˆ’ 3

2 โ€–๐‘ƒb ๐‘‡2โˆ— (๐œ•2๐œ“ )๐‘‡

2โ€–๐ฟโˆž

.(๐‘‡

13) 32 โ€–๐‘ƒb ๐‘‡

2โ€–๐ฟโˆž .

(๐‘‡

13) 14โˆ’Y [๐‘ƒb]โˆ’ 5

4โˆ’Y,

Page 56: Variational methods for a singular SPDE yielding the

56 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

where we used the characterisation of negative Hรถlder spaces from Lemma A.1. Note that theimplicit constant is universal for Y small, in particular for Y โˆˆ (0, 18 ). Hence, we obtain that[

|๐œ•1 |โˆ’1๐œ•2๐‘ฃ]14โˆ’Y

= [๐‘”] 14โˆ’Y. sup

๐‘‡ โˆˆ(0,1]

(๐‘‡

13)โˆ’ 1

4+Y โ€–๐‘‡A๐‘”๐‘‡ โ€–๐ฟโˆž . [b]โˆ’ 54โˆ’Y.

๏ฟฝ

Proposition D.2. There exists a constant ๐ถ > 0 such that for every ๐›ฟb โˆˆ ๐ฟ2(T2), the solution ๐›ฟ๐‘ฃ ofvanishing average in ๐‘ฅ1 to L๐›ฟ๐‘ฃ = ๐‘ƒ๐›ฟb satises the following estimates:

[๐›ฟ๐‘ฃ] 34โ‰ค ๐ถ โ€–๐›ฟb โ€–๐ฟ2, (D.1)

โ€–๐œ•1๐›ฟ๐‘ฃ โ€–๐ฟ10 โ‰ค ๐ถ โ€–๐›ฟb โ€–๐ฟ2, (D.2)โ€–๐œ•2๐›ฟ๐‘ฃ โ€–

๐ฟ103โ‰ค ๐ถ โ€–๐›ฟb โ€–๐ฟ2 . (D.3)

Proof. Writing ๐œ•21๐›ฟ๐‘ฃ and |๐œ•2 |43๐›ฟ๐‘ฃ as Fourier series and using Youngโ€™s inequality in the form |๐‘˜2 |

83 .

๐‘˜41 +๐‘˜42๐‘˜21

for every ๐‘˜1, ๐‘˜2 โˆˆ 2๐œ‹Z, we obtain that

โ€–๐œ•21๐›ฟ๐‘ฃ โ€–2๐ฟ2 + โ€–|๐œ•2 |43๐›ฟ๐‘ฃ โ€–2

๐ฟ2 . โ€–L๐›ฟ๐‘ฃ โ€–2๐ฟ2 . โ€–๐‘ƒ๐›ฟb โ€–2

๐ฟ2 . โ€–๐›ฟb โ€–2๐ฟ2,

which leads to (D.1) via (B.11) for ๐‘  = 2. Using the same argument, we also obtain that

โ€–๐œ•21๐›ฟ๐‘ฃ โ€–2๐ฟ2 + โ€–|๐œ•2 |23 ๐œ•1๐›ฟ๐‘ฃ โ€–2๐ฟ2 . โ€–๐›ฟb โ€–2

๐ฟ2, (D.4)

โ€–๐œ•1 |๐œ•2 |23๐›ฟ๐‘ฃ โ€–2

๐ฟ2 + โ€–|๐œ•2 |43๐›ฟ๐‘ฃ โ€–2

๐ฟ2 . โ€–๐›ฟb โ€–2๐ฟ2, (D.5)

โ€–|๐œ•1 |12 ๐œ•2๐›ฟ๐‘ฃ โ€–2๐ฟ2 + โ€–|๐œ•2 |

13 ๐œ•2๐›ฟ๐‘ฃ โ€–2๐ฟ2 . โ€–๐›ฟb โ€–2

๐ฟ2 . (D.6)By the embedding result in Lemma B.4 we know that

โ€–๐œ•1๐›ฟ๐‘ฃ โ€–๐ฟ10 . โ€–๐œ•21๐›ฟ๐‘ฃ โ€–๐ฟ2 + โ€–|๐œ•2 |23 ๐œ•1๐›ฟ๐‘ฃ โ€–๐ฟ2, (D.7)

โ€–๐œ•โ„Ž2๐›ฟ๐‘ฃ โ€–๐ฟ10 . โ€–๐œ•1๐œ•โ„Ž2๐›ฟ๐‘ฃ โ€–๐ฟ2 + โ€–|๐œ•2 |23 ๐œ•โ„Ž2๐›ฟ๐‘ฃ โ€–๐ฟ2, (D.8)

โ€–๐œ•2๐›ฟ๐‘ฃ โ€–๐ฟ103. โ€–|๐œ•1 |

12 ๐œ•2๐›ฟ๐‘ฃ โ€–๐ฟ2 + โ€–|๐œ•2 |

13 ๐œ•2๐›ฟ๐‘ฃ โ€–๐ฟ2 . (D.9)

Combining (D.4) with (D.7) and (D.6) with (D.9) we obtain (D.2) and (D.3). ๏ฟฝ

Lemma D.3. Let b be smooth. Then the solution ๐‘ฃ of L๐‘ฃ = ๐‘ƒb with vanishing average in ๐‘ฅ1 issmooth.

Proof. If b is smooth, then ๐œ•๐‘š1 ๐œ•๐‘›2 b โˆˆ ๐ฟ2(T2) for all๐‘š,๐‘› โˆˆ N0. It follows by a simple calculation inFourier space that

โ€–๐œ•๐‘š1 ๐œ•๐‘›2 ๐‘ฃ โ€–2๐ฟ2 =โˆ‘๏ธ๐‘˜

|๐‘˜1 |2๐‘š |๐‘˜2 |2๐‘› |๏ฟฝฬ‚๏ฟฝ (๐‘˜) |2 =โˆ‘๏ธ๐‘˜1โ‰ 0

|๐‘˜1 |2๐‘š |๐‘˜2 |2๐‘›|๐‘˜1 |2 |bฬ‚ (๐‘˜) |2|๐‘˜1 |3 + |๐‘˜2 |2

โ‰คโˆ‘๏ธ๐‘˜1โ‰ 0

|๐‘˜1 |2๐‘š+2 |๐‘˜2 |2๐‘› |bฬ‚ (๐‘˜) |2 = โ€–๐œ•๐‘š+11 ๐œ•๐‘›2 b โ€–2๐ฟ2 < โˆž

for all๐‘š,๐‘› โˆˆ N0, in particular, ๐‘ฃ is smooth by Sobolev embedding. ๏ฟฝ

Appendix E. Regularity of finite-energy solutions for smooth data

In this section we develop an ๐ฟ2-based regularity theory for weak solutions ๐‘ข with nite energyH(๐‘ข) < โˆž of the Eulerโ€“Lagrange equation

L๐‘ข = โˆ’๐‘ƒ(๐‘ข๐‘…1๐œ•2๐‘ข โˆ’ 1

2๐‘ข๐‘…1๐œ•1๐‘ข2)โˆ’ 12๐‘…1๐œ•2๐‘ข

2 + ๐‘ƒb . (E.1)

For โ„Ž โˆˆ (0, 1], dene the dierence quotients

๐ทโ„Ž๐‘– ๐‘ข = |โ„Ž |โˆ’๐›ผ๐‘– ๐œ•โ„Ž๐‘– ๐‘ข, ๐›ผ1 = 1, ๐›ผ2 =

23 . (E.2)

Page 57: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 57

Proposition E.1 (๐ป 2โˆ’ estimate). Assume that b โˆˆ ๐ฟ2. There exists a constant ๐ถ > 0 such that for

any solution ๐‘ข of the Eulerโ€“Lagrange equation (E.1) withH(๐‘ข) < โˆž we have

supโ„Žโˆˆ(0,1]

H(๐ทโ„Ž๐‘– ๐‘ข) โ‰ค ๐ถ

(1 + โ€–b โ€–2

๐ฟ2 + H (๐‘ข)12).

In particular, for any ๐‘  โˆˆ [0, 2) we have that โ€–|๐œ•1 |๐‘ ๐‘ขโ€–๐ฟ2 + โ€–|๐œ•2 |23๐‘ ๐‘ขโ€–๐ฟ2 < โˆž.

Proposition E.2 (๐ป 3โˆ’ estimate). Assume that b satises H(b) < โˆž. There exists a constant ๐ถ > 0such that for any solution ๐‘ข of the Eulerโ€“Lagrange equation (E.1) withH(๐‘ข) < โˆž we have

supโ„Ž,โ„Žโ€ฒโˆˆ(0,1]

H(๐ทโ„Ž๐‘– ๐ท

โ„Žโ€ฒ๐‘– ๐‘ข) โ‰ค ๐ถ

(1 + H (b)22 + H (๐‘ข)132

).

In particular, for any ๐‘  โˆˆ [0, 3) we have that โ€–|๐œ•1 |๐‘ ๐‘ขโ€–๐ฟ2 + โ€–|๐œ•2 |23๐‘ ๐‘ขโ€–๐ฟ2 < โˆž.

For the proof of Propositions E.1 and E.2 it is convenient to work with the scale of norms

โ€–|๐‘“ โ€–|2๐‘  =โˆซT2

( |๐œ•1 |๐‘  ๐‘“ )2 d๐‘ฅ +โˆซT2

(|๐œ•1 |โˆ’

๐‘ 2 |๐œ•2 |๐‘  ๐‘“

)2d๐‘ฅ,

for ๐‘  > 0 and periodic functions ๐‘“ : T2 โ†’ R with vanishing average in ๐‘ฅ1, as dened in (B.8).These norms are adapted to the harmonic energy H , in particular, we have that โ€–|๐‘“ โ€–|1 = H(๐‘“ ) 1

2 ,and one may think of the norms โ€–| ยท โ€–|๐‘  dening a scale of (anisotropic) Sobolev spaces. Indeed, asshown in Lemma B.6 the norms โ€–| ยท โ€–|๐‘  control an anisotropic fractional gradient in ๐ฟ2.Lemma E.3. There exists a constant ๐ถ > 0 such that for any ๐‘ข โˆˆ W there holds

supโ„Žโˆˆ(0,1]

โ€–๐ทโ„Ž๐‘– ๐‘ขโ€–๐ฟ2 โ‰ค ๐ถ โ€–|๐‘ขโ€–|1, ๐‘– = 1, 2.

Proof. We treat the two directions ๐‘– = 1, 2 separately. For ๐‘– = 1, the claim follows easily by themean-value theorem, which implies that

supโ„Žโˆˆ(0,1]

โ€–๐ทโ„Ž1๐‘ขโ€–๐ฟ2 โ‰ค โ€–๐œ•1๐‘ขโ€–๐ฟ2 .

For ๐‘– = 2 we appeal to Lemma B.10 (or rather its analogue for functions on T2) to estimate

supโ„Žโˆˆ(0,1]

โ€–๐ทโ„Ž2๐‘ขโ€–2๐ฟ2 = sup

โ„Žโˆˆ(0,1]

1โ„Ž

43

โˆซT2

|๐œ•โ„Ž2๐‘ข |2 d๐‘ฅ . supโ„Žโˆˆ(0,1]

1โ„Ž

43

1โ„Ž

โˆซ โ„Ž

0

โˆซT2

|๐œ•โ„Žโ€ฒ2 ๐‘ข |2 d๐‘ฅ dโ„Žโ€ฒ

. supโ„Žโˆˆ(0,1]

โˆซ โ„Ž

0

โˆซT2

|๐œ•โ„Žโ€ฒ2 ๐‘ข |2

โ„Žโ€ฒ43

d๐‘ฅ dโ„Žโ€ฒ

โ„Žโ€ฒ.

โˆซR

โˆซT2

|๐œ•โ„Ž2๐‘ข |2

|โ„Ž | 43d๐‘ฅ dโ„Ž|โ„Ž |

(2.2). โ€–|๐œ•2 |

23๐‘ขโ€–2

๐ฟ2 .

The conclusion then follows from Lemma B.6. ๏ฟฝ

The proof of Propositions E.1 and E.2 mainly relies on the following two lemmata:Lemma E.4. Let ๐œ– โˆˆ (0, 14 ). There exists a constant ๐ถ๐œ– > 0 such that for all periodic functions

๐‘“ , ๐‘”, ๐œ‘ : T2 โ†’ R of vanishing average in ๐‘ฅ1 we have๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘“ (๐œ•2๐‘”)๐œ‘ d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค ๐ถ๐œ– โ€–|๐‘“ โ€–|1โ€–|๐‘”โ€–|1โ€–|๐œ‘ โ€–| 34+๐œ– .

Proof. This follows easily from the denition of โ€–| ยท โ€–|1 and Lemma B.7. Indeed, we have๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘“ (๐œ•2๐‘”)๐œ‘ d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ = ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2

|๐œ•1 |โˆ’12 ๐œ•2๐‘” |๐œ•1 |

12 (๐‘“ ๐œ‘) d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . โ€–|๐‘”โ€–|1โ€–|๐œ•1 |12 (๐‘“ ๐œ‘)โ€–๐ฟ2 . โ€–|๐‘”โ€–|1โ€–|๐‘“ โ€–|1โ€–|๐œ‘ โ€–| 34+๐œ– .

๏ฟฝ

Lemma E.5. There exists a constant ๐ถ > 0 such that for all periodic functions ๐‘“ , ๐‘”, โ„Ž, ๐œ‘ : T2 โ†’ R of

vanishing average in ๐‘ฅ1 we have๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘“ ๐‘…1(๐‘”๐œ•1โ„Ž)๐œ‘ d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค ๐ถ โ€–|๐‘“ โ€–|1โ€–|๐‘”โ€–|1โ€–|โ„Žโ€–|1โ€–|๐œ‘ โ€–| 12 .

Page 58: Variational methods for a singular SPDE yielding the

58 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

Proof. We rst use Cauchyโ€“Schwarz to bound๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘“ ๐‘…1(๐‘”๐œ•1โ„Ž)๐œ‘ d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ = ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2(๐œ•1โ„Ž)๐‘”๐‘…1(๐‘“ ๐œ‘) d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค โ€–๐œ•1โ„Žโ€–๐ฟ2 โ€–๐‘”๐‘…1(๐‘“ ๐œ‘)โ€–๐ฟ2 โ‰ค โ€–|โ„Žโ€–|1โ€–๐‘”๐‘…1(๐‘“ ๐œ‘)โ€–๐ฟ2 .

By Hรถlderโ€™s inequality and the boundedness of ๐‘…1 on ๐ฟ52 (T2) (A.4) , we may further estimate

โ€–๐‘”๐‘…1(๐‘“ ๐œ‘)โ€–๐ฟ2 โ‰ค โ€–๐‘”โ€–๐ฟ10 โ€–๐‘…1(๐‘“ ๐œ‘)โ€–๐ฟ52. โ€–๐‘”โ€–๐ฟ10 โ€– ๐‘“ ๐œ‘ โ€–

๐ฟ52. โ€–๐‘”โ€–๐ฟ10 โ€– ๐‘“ โ€–๐ฟ10 โ€–๐œ‘ โ€–

๐ฟ103,

from which the claim follows with the Sobolev-type embeddings

โ€–๐‘”โ€–๐ฟ10(B.5). โ€–๐œ•1๐‘”โ€–๐ฟ2 + โ€–๐œ•2 |

23๐‘”โ€–๐ฟ2

(B.9). โ€–|๐‘”โ€–|1,

โ€–๐œ‘ โ€–๐ฟ103

(B.6). โ€–|๐œ•1 |

12๐œ‘ โ€–๐ฟ2 + โ€–๐œ•2 |

13๐œ‘ โ€–๐ฟ2

(B.9). โ€–|๐œ‘ โ€–| 1

2.

๏ฟฝ

Proof of Proposition E.1. Since โ€–|๐‘ขโ€–|21 = H(๐‘ข) < โˆž, we can test the Eulerโ€“Lagrange equation (E.1)with ๐œ“ = ๐ทโˆ’โ„Ž

๐‘– ๐ทโ„Ž๐‘– ๐‘ข, where ๐ทโ„Ž

๐‘– denotes the dierence quotient introduced in (E.2) for โ„Ž โˆˆ (0, 1].Then the left-hand side of (E.1) turns intoโˆซ

T2L๐‘ข๐ทโˆ’โ„Ž

๐‘– ๐ทโ„Ž๐‘– ๐‘ข d๐‘ฅ = H(๐ทโ„Ž

๐‘– ๐‘ข) = โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–|21.

We continue with estimating the terms on the right-hand side of the Eulerโ€“Lagrange equation:Term 1 (๐‘ข๐‘…1๐œ•2๐‘ข). By a discrete integration by parts, we can writeโˆซ

T2๐‘ข๐‘…1๐œ•2๐‘ข๐ท

โˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ =

โˆซT2๐ทโ„Ž๐‘– ๐‘ข (๐œ•2๐‘…1๐‘ข)๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ +โˆซT2๐‘ขโ„Ž๐œ•2(๐ทโ„Ž

๐‘– ๐‘…1๐‘ข)๐ทโ„Ž๐‘– ๐‘ข d๐‘ฅ,

where ๐‘ขโ„Ž = ๐‘ข (ยท + โ„Ž). Applying Lemma E.4 to each of these two terms, together with

โ€–|๐‘…1๐‘ขโ€–|1 = โ€–|๐‘ขโ€–|1 and โ€–|๐‘ขโ„Ž โ€–|1 = โ€–|๐‘ขโ€–|1, (E.3)

then yields๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐ทโ„Ž๐‘– ๐‘ข (๐œ•2๐‘…1๐‘ข)๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|1โ€–|๐‘…1๐‘ขโ€–|1โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–| 45 . โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|1โ€–|๐‘ขโ€–|1โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–| 45 ,๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซ

T2๐‘ขโ„Ž๐œ•2(๐ทโ„Ž

๐‘– ๐‘…1๐‘ข)๐ทโ„Ž๐‘– ๐‘ข d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . โ€–|๐‘ขโ„Ž โ€–|1โ€–|๐‘…1๐ทโ„Ž๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–| 45 . โ€–|๐‘ขโ€–|1โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–| 45 .

Note that by interpolation (which is easily seen in Fourier space) and Lemma E.3, there holds

โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–| 45 โ‰ค โ€–๐ทโ„Ž

๐‘– ๐‘ขโ€–15๐ฟ2โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|451 . โ€–|๐‘ขโ€–|

151 โ€–|๐ท

โ„Ž๐‘– ๐‘ขโ€–|

451 , (E.4)

hence ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘ข๐‘…1๐œ•2๐‘ข๐ท

โˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . โ€–|๐‘ขโ€–|

651 โ€–|๐ท

โ„Ž๐‘– ๐‘ขโ€–|

951 .

Term 2 (๐‘…1๐œ•2๐‘ข2). As for the rst term, after a discrete integration by partsโˆซT2๐‘…1๐œ•2๐‘ข

2๐ทโˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ =

โˆซT2๐ทโ„Ž๐‘– (๐‘ข๐œ•2๐‘ข)๐‘…1๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ

=

โˆซT2(๐ทโ„Ž

๐‘– ๐‘ข)๐œ•2๐‘ข๐‘…1๐ทโ„Ž๐‘– ๐‘ข d๐‘ฅ +

โˆซT2๐‘ขโ„Ž๐ทโ„Ž

๐‘– ๐œ•2๐‘ข๐‘…1๐ทโ„Ž๐‘– ๐‘ข d๐‘ฅ,

an application of Lemma E.4, the interpolation (E.4), and Lemma E.3 implies๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘…1๐œ•2๐‘ข

2๐ทโˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|1โ€–|๐‘ขโ€–|1โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–| 45 . โ€–|๐‘ขโ€–|

651 โ€–|๐ท

โ„Ž๐‘– ๐‘ขโ€–|

951 .

Page 59: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 59

Term 3 (๐‘ข๐œ•1๐‘…1๐‘ข2). For the cubic term in ๐‘ข we haveโˆซT2๐‘ข๐œ•1๐‘…1๐‘ข

2๐ทโˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ =

โˆซT2(๐ทโ„Ž

๐‘– ๐‘ข)๐‘…1(๐‘ข๐œ•1๐‘ข)๐ทโ„Ž๐‘– ๐‘ข d๐‘ฅ +

โˆซT2๐‘ขโ„Ž๐‘…1(๐ทโ„Ž

๐‘– ๐‘ข๐œ•1๐‘ข)๐ทโ„Ž๐‘– ๐‘ข d๐‘ฅ

+โˆซT2๐‘ขโ„Ž๐‘…1(๐‘ขโ„Ž๐œ•1๐ทโ„Ž

๐‘– ๐‘ข)๐ทโ„Ž๐‘– ๐‘ข d๐‘ฅ .

By Lemma E.5 we may estimate the rst term byโˆซT2(๐ทโ„Ž

๐‘– ๐‘ข)๐‘…1(๐‘ข๐œ•1๐‘ข)๐ทโ„Ž๐‘– ๐‘ข d๐‘ฅ . โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|1โ€–|๐‘ขโ€–|1โ€–|๐‘ขโ€–|1โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–| 12 ,

and similarly for the other two terms, so that, together with (E.3),๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘ข๐œ•1๐‘…1๐‘ข

2๐ทโˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . โ€–|๐‘ขโ€–|21โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–| 12 .

Interpolation and Lemma E.3 then yield

โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–| 12 โ‰ค โ€–๐ทโ„Ž

๐‘– ๐‘ขโ€–12๐ฟ2โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|121 . โ€–|๐‘ขโ€–|

121 โ€–|๐ท

โ„Ž๐‘– ๐‘ขโ€–|

121 ,

hence ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘ข๐œ•1๐‘…1๐‘ข

2๐ทโˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ . โ€–|๐‘ขโ€–|

521 โ€–|๐ท

โ„Ž๐‘– ๐‘ขโ€–|

321 .

Term 4 (๐‘ƒb). Note that by assumption b โˆˆ ๐ฟ2, so that Cauchyโ€“Schwarz and Lemma E.3 imply๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซT2๐‘ƒb๐ทโˆ’โ„Ž

๐‘– ๐ทโ„Ž๐‘– ๐‘ข d๐‘ฅ

๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค โ€–๐‘ƒb โ€–๐ฟ2 โ€–๐ทโˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ขโ€–๐ฟ2 . โ€–๐‘ƒb โ€–๐ฟ2 โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–|1.

With the above estimates on the three superlinear on the right-hand side of the Eulerโ€“Lagrangeequation, we have therefore shown that

โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–|21 . โ€–|๐‘ขโ€–|

651 โ€–|๐ท

โ„Ž๐‘– ๐‘ขโ€–|

951 + โ€–|๐‘ขโ€–|

521 โ€–|๐ท

โ„Ž๐‘– ๐‘ขโ€–|

321 + โ€–b โ€–๐ฟ2 โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|1,which by Youngโ€™s inequality implies that

โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–|21 . โ€–|๐‘ขโ€–|121 + โ€–|๐‘ขโ€–|101 + โ€–๐‘ƒb โ€–2

๐ฟ2 . 1 + โ€–|๐‘ขโ€–|121 + โ€–๐‘ƒb โ€–2๐ฟ2,

with an implicit constant that does not depend onโ„Ž. Finally, we notice that a bound on the quantitysupโ„Žโˆˆ(0,1] โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|1 implies a bound on โ€–|๐œ•1 |๐‘ ๐‘ขโ€–๐ฟ2 + โ€–|๐œ•2 |23๐‘ ๐‘ขโ€–๐ฟ2 for ๐‘  โˆˆ [0, 1) by Lemma B.6 and

(2.2). ๏ฟฝ

Proof of Proposition E.2. The proof is very similar to the proof of Proposition E.1. Under theassumptions of Proposition E.2, Proposition E.1 implies that

supโ„Žโˆˆ(0,1]

โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–|1 . 1 + โ€–|๐‘ขโ€–|61 + โ€–๐‘ƒb โ€–๐ฟ2 . 1 + โ€–|๐‘ขโ€–|61 + โ€–|b โ€–|1, (E.5)

so that we may test the Eulerโ€“Lagrange equation (E.1) with ๐œ“ = ๐ทโˆ’โ„Žโ€ฒ๐‘– ๐ทโ„Žโ€ฒ

๐‘– ๐ทโˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข. Then theleft-hand side of (E.1) turns intoโˆซ

T2L๐‘ข ๐ทโˆ’โ„Žโ€ฒ

๐‘– ๐ทโ„Žโ€ฒ๐‘– ๐ท

โˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ = โ€–|๐ทโ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|21.

We proceed by estimating each term on the right-hand side of (E.1).Term 1 (๐‘ข๐‘…1๐œ•2๐‘ข). Integrating by parts twice (with respect to ๐ทโˆ’โ„Žโ€ฒ

๐‘– and ๐ทโˆ’โ„Ž๐‘– , we obtain four terms,

all of which can be estimated by eitherโ€–|๐‘ขโ€–|1โ€–|๐ทโ„Žโ€ฒ

๐‘– ๐ทโ„Ž๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Žโ€ฒ

๐‘– ๐ทโ„Ž๐‘– ๐‘ขโ€–| 45 , or โ€–|๐ทโ„Žโ€ฒ

๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Žโ€ฒ

๐‘– ๐ทโ„Ž๐‘– ๐‘ขโ€–| 45 . (E.6)

By interpolation, see (E.4), and Lemma E.3, there holds

โ€–|๐ทโ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–| 45 โ‰ค โ€–๐ทโ„Žโ€ฒ

๐‘– ๐ทโ„Ž๐‘– ๐‘ขโ€–

15๐ฟ2โ€–|๐ทโ„Žโ€ฒ

๐‘– ๐ทโ„Ž๐‘– ๐‘ขโ€–|

451 . โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|151 โ€–|๐ท

โ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|

451

Term 2 (๐‘…1๐œ•2๐‘ข2). This term is treated like the previous one, with the same bounds (E.6).

Page 60: Variational methods for a singular SPDE yielding the

60 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

Term 3 (๐‘ข๐œ•1๐‘…1๐‘ข2). For this term we again integrate by parts twice to obtain nine terms, all ofwhich can be bounded by one of the expressions

โ€–|๐‘ขโ€–|1โ€–|๐ทโ„Žโ€ฒ๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–| 12 , or โ€–|๐‘ขโ€–|21โ€–|๐ทโ„Žโ€ฒ

๐‘– ๐ทโ„Ž๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Žโ€ฒ

๐‘– ๐ทโ„Ž๐‘– ๐‘ขโ€–| 12 ,

where by interpolation and Lemma E.3

โ€–|๐ทโ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–| 12 โ‰ค โ€–๐ทโ„Žโ€ฒ

๐‘– ๐ทโ„Ž๐‘– ๐‘ขโ€–

12๐ฟ2โ€–|๐ทโ„Žโ€ฒ

๐‘– ๐ทโ„Ž๐‘– ๐‘ขโ€–|

121 . โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|121 โ€–|๐ท

โ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|

121 .

Term 4 (๐‘ƒb). Under the regularity assumption โ€–|b โ€–|1 < โˆž on b we can also estimate with Cauchyโ€“Schwarz and Lemma E.3๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซ

T2๐‘ƒb๐ทโˆ’โ„Žโ€ฒ

๐‘– ๐ทโ„Žโ€ฒ๐‘– ๐ท

โˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ = ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝโˆซ

T2๐ทโ„Žโ€ฒ๐‘– ๐‘ƒb๐ท

โ„Žโ€ฒ๐‘– ๐ท

โˆ’โ„Ž๐‘– ๐ทโ„Ž

๐‘– ๐‘ข d๐‘ฅ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค โ€–๐ทโ„Žโ€ฒ

๐‘– b โ€–๐ฟ2 โ€–๐ทโˆ’โ„Ž๐‘– ๐ทโ„Žโ€ฒ

๐‘– ๐ทโ„Ž๐‘– ๐‘ขโ€–๐ฟ2

โ‰ค โ€–|b โ€–|1โ€–|๐ทโ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|1.

In total, we can therefore bound

โ€–|๐ทโ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|21 . โ€–|๐‘ขโ€–|1โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|151 โ€–|๐ท

โ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|

951 + โ€–|๐ทโ„Žโ€ฒ

๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Ž๐‘– ๐‘ขโ€–|

651 โ€–|๐ท

โ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|

951

+ โ€–|๐‘ขโ€–|1โ€–|๐ทโ„Žโ€ฒ๐‘– ๐‘ขโ€–|1โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|321 โ€–|๐ท

โ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|

121 + โ€–|๐‘ขโ€–|21โ€–|๐ทโ„Ž

๐‘– ๐‘ขโ€–|121 โ€–|๐ท

โ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|

321

+ โ€–|b โ€–|1โ€–|๐ทโ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|1,

from which it follows by Youngโ€™s inequality and (E.5) that

โ€–|๐ทโ„Žโ€ฒ๐‘– ๐ท

โ„Ž๐‘– ๐‘ขโ€–|21 . 1 + โ€–|๐‘ขโ€–|1321 + โ€–|b โ€–|221 ,

with an implicit constant that does not depend on โ„Ž,โ„Žโ€ฒ. As in the proof of Proposition E.1, wenotice that a bound on supโ„Ž,โ„Žโ€ฒโˆˆ(0,1] โ€–|๐ทโ„Ž

๐‘– ๐ทโ„Žโ€ฒ๐‘– ๐‘ขโ€–|21 implies a bound on โ€–|๐œ•1 |๐‘ ๐‘ขโ€–๐ฟ2 + โ€–|๐œ•2 |

23๐‘ ๐‘ขโ€–๐ฟ2 for

๐‘  โˆˆ [2, 3) by Lemma B.6 and (2.2). ๏ฟฝ

Appendix F. Approximation of qadratic functionals of the noise by cylinderfunctionals

In this section we show the following approximation result, which seems classical but sincewe could not nd a proof in the literature we include it here:

Lemma F.1. Let ๐พ be a linear operator on the space of Schwartz distributions Sโ€ฒ(T2) such that

๐พ (Sโ€ฒ(T2)) โŠ† Cโˆž(T2) and ๐พ : ๐ฟ2(T2) โ†’ ๐ฟ2(T2) is Hilbertโ€“Schmidt. Assume further that the

probability measure ใ€ˆยทใ€‰ satises Assumption 1.1 (iv). Consider the quadratic functional๐บ : Sโ€ฒ(T2) โ†’R given by

๐บ (b) := b (๐พb).Then under the assumption that ใ€ˆ|๐บ (b) |ใ€‰ < โˆž, ๐บ is well-approximated by cylindrical functionals in

with respect to the norm ใ€ˆ|๐บ (b) |2๐‘ใ€‰12๐‘ + ใ€ˆโ€– ๐œ•๐บ

๐œ•b(b)โ€–2๐‘

๐ฟ2ใ€‰

12๐‘

for every 1 โ‰ค ๐‘ < โˆž.

Proof. Without loss of generality, we may assume that๐พ is symmetric. Since๐พ is Hilbertโ€“Schmidt,there exists an orthonormal system {๐œ‘๐‘›}๐‘›โˆˆN of ๐ฟ2(T2) and a sequence {_๐‘›}๐‘›โˆˆN โŠ‚ R such that

๐พ =โˆ‘๏ธ๐‘›โˆˆN

_๐‘› (๐œ‘๐‘›, ยท)๐œ‘๐‘›, and โ€–๐พ โ€–2๐ป๐‘† =โˆ‘๏ธ๐‘›

_2๐‘› < โˆž. (F.1)

Note that by the assumption ๐พ (Sโ€ฒ(T2)) โŠ† Cโˆž(T2) there holds {๐œ‘๐‘›}๐‘›โˆˆN โŠ‚ Cโˆž(T2), in particularb (๐œ‘๐‘›) is well-dened for any ๐‘› โˆˆ N, and given ๐‘ โˆˆ N we may dene ๐พ๐‘ b :=

โˆ‘๐‘›โ‰ค๐‘ _๐‘›b (๐œ‘๐‘›)๐œ‘๐‘› .

Step 1 We rst show that for any b โˆˆ Sโ€ฒ(T2), ๐พ๐‘ b โ†’ ๐พb in Sโ€ฒ(T2) as ๐‘ โ†’ โˆž.Indeed, with b๐‘ก := b โˆ—๐œ“๐‘ก โˆˆ Cโˆž(T2), for any ๐œ™ โˆˆ Cโˆž(T2) we have

| (๐พ๐‘ b โˆ’ ๐พb, ๐œ™) | โ‰ค |(๐พ๐‘ b โˆ’ ๐พ๐‘ b๐‘ก , ๐œ™) | + |(๐พ๐‘ b๐‘ก โˆ’ ๐พb๐‘ก , ๐œ™) | + |(๐พb๐‘ก โˆ’ ๐พb, ๐œ™) |.

Page 61: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 61

Note that b๐‘ก โ†’ b in Sโ€ฒ as ๐‘ก โ†“ 0, hence for any ๐œ– > 0 we may choose ๐‘ก > 0 small enough so that bysymmetry of ๐พ๐‘ and ๐พ ,

| (๐พ๐‘ b โˆ’ ๐พ๐‘ b๐‘ก , ๐œ™) | = | (b โˆ’ b๐‘ก ) (๐พ๐‘๐œ™) | <๐œ–

3 ,

| (๐พb๐‘ก โˆ’ ๐พb, ๐œ™) | = | (b โˆ’ b๐‘ก ) (๐พ๐œ™) | <๐œ–

3 ,

for any๐‘ โˆˆ N. The remaining term can be bounded using Cauchy-Schwarz and Besselโ€™s inequalityfor orthonormal systems to obtain

| (๐พ๐‘ b๐‘ก โˆ’ ๐พb๐‘ก , ๐œ™) | =๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ(โˆ‘๏ธ๐‘›>๐‘

_๐‘› (๐œ‘๐‘›, b๐‘ก ) (๐œ‘๐‘›, ๐œ™))๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ๏ฟฝ โ‰ค

(โˆ‘๏ธ๐‘›>๐‘

_2๐‘› | (๐œ‘๐‘›, b๐‘ก ) |2) 1

2(โˆ‘๏ธ๐‘›โˆˆN

| (๐œ‘๐‘›, ๐œ™) |2) 1

2

โ‰ค โ€–b๐‘ก โ€–๐ฟ2(โˆ‘๏ธ๐‘›>๐‘

_2๐‘›

) 12

โ€–๐œ™ โ€–๐ฟ2 <๐œ–

3

if ๐‘ is chosen suitably large, since {_๐‘›}๐‘› is square-summable.Step 2 We next claim that ๐บ๐‘ (b) := b (๐พ๐‘ b) satises ๐บ๐‘ (b) โ†’ ๐บ (b) as ๐‘ โ†’ โˆž for all b โˆˆSโ€ฒ(T2).

For the proof of this claim we again appeal to the convergence b๐‘ก โ†’ b in Sโ€ฒ by splitting|๐บ๐‘ (b) โˆ’๐บ (b) | = |b (๐พ๐‘ b) โˆ’ b (๐พb) |

โ‰ค |b (๐พ๐‘ b) โˆ’ b๐‘ก (๐พ๐‘ b) | + |b๐‘ก (๐พ๐‘ b) โˆ’ b๐‘ก (๐พb) | + |b๐‘ก (๐พb) โˆ’ b (๐พb) |.As in Step 1, there holds

|b (๐พ๐‘ b) โˆ’ b๐‘ก (๐พ๐‘ b) | โ†’ 0, |b๐‘ก (๐พb) โˆ’ b (๐พb) | โ†’ 0

as ๐‘ก โ†“ 0 for any ๐‘ โˆˆ N, since ๐พ๐‘ b, ๐พb โˆˆ Cโˆž(T2). Moreover, by Step 1, we have|b๐‘ก (๐พ๐‘ b) โˆ’ b๐‘ก (๐พb) | = | (๐พ๐‘ b โˆ’ ๐พb, b๐‘ก ) | โ†’ 0

as ๐‘ โ†’ โˆž.Step 3 We show that ๐œ•๐บ๐‘ (b)

๐œ•bโ†’ ๐œ•๐บ (b)

๐œ•bas ๐‘ โ†’ โˆž in ๐ฟ๐‘ใ€ˆยทใ€‰๐ฟ

2๐‘ฅ for any 1 โ‰ค ๐‘ < โˆž.

Indeed, by symmetry of ๐พ and ๐พ๐‘ we have that๐œ•๐บ๐‘ (b)๐œ•b

= 2๐พ๐‘ b = 2โˆ‘๏ธ๐‘›โ‰ค๐‘

_๐‘›b (๐œ‘๐‘›)๐œ‘๐‘›, and ๐œ•๐บ (b)๐œ•b

= 2๐พb = 2โˆ‘๏ธ๐‘›โˆˆN

_๐‘›b (๐œ‘๐‘›)๐œ‘๐‘› .

With this we obtain by orthonormality of the {๐œ‘๐‘›}๐‘› that ๐œ•(๐บ โˆ’๐บ๐‘ ) (b)๐œ•b

2๐ฟ2

= 4โˆ‘๏ธ๐‘›>๐‘

_2๐‘›b (๐œ‘๐‘›)2,

hence, applying (5.1) in the form ofโŸจb (๐œ‘๐‘›)2๐‘

โŸฉ 12๐‘ .๐‘ โ€–๐œ‘๐‘› โ€–๐ฟ2 .๐‘ 1,

for all ๐‘› โˆˆ N, it follows thatโŸจ ๐œ•๐บ (b)๐œ•b

โˆ’ ๐œ•๐บ๐‘ (b)๐œ•b

2๐‘๐ฟ2

โŸฉ 12๐‘

=

โŸจ(4โˆ‘๏ธ๐‘›>๐‘

_2๐‘›b (๐œ‘๐‘›)2)๐‘โŸฉ 1

2๐‘

โ‰ค 2(โˆ‘๏ธ๐‘›>๐‘

_2๐‘›โŸจb (๐œ‘๐‘›)2๐‘

โŸฉ 1๐‘

) 12

.๐‘

(โˆ‘๏ธ๐‘›>๐‘

_2๐‘›

) 12๐‘โ†’โˆžโˆ’โ†’ 0,

by the niteness of โ€–๐พ โ€–๐ป๐‘† .

Page 62: Variational methods for a singular SPDE yielding the

62 R. IGNAT, F. OTTO, T. RIED, AND P. TSATSOULIS

Step 4 We show that the sequence {๐บ๐‘ (b) โˆ’ ใ€ˆ๐บ๐‘ (b)ใ€‰}๐‘ โˆˆN is Cauchy in ๐ฟ2๐‘ใ€ˆยทใ€‰ for any 1 โ‰ค ๐‘ < โˆž,in particular, there exists a centred random variable ๐บโˆ—(b) such that ๐บ๐‘ (b) โˆ’ ใ€ˆ๐บ๐‘ (b)ใ€‰ โ†’ ๐บโˆ—(b)for ใ€ˆยทใ€‰-almost every b as ๐‘ โ†’ โˆž along a subsequence.

Indeed, since๐บ๐‘ (b) is a cylinder functional and ใ€ˆยทใ€‰ satises Assumption 1.1 (iv), in particularProposition 5.1, we can apply (5.1) to obtain the boundโŸจ|๐บ๐‘ (b) โˆ’ ใ€ˆ๐บ๐‘ (b)ใ€‰ โˆ’ (๐บ๐‘€ (b) โˆ’ ใ€ˆ๐บ๐‘€ (b)ใ€‰) |2๐‘

โŸฉ 12๐‘ =

โŸจ|๐บ๐‘ (b) โˆ’๐บ๐‘€ (b) โˆ’ ใ€ˆ๐บ๐‘ (b) โˆ’๐บ๐‘€ (b)ใ€‰|2๐‘

โŸฉ 12๐‘

.๐‘

โŸจ ๐œ•(๐บ๐‘ โˆ’๐บ๐‘€ ) (b)๐œ•b

2๐‘๐ฟ2

โŸฉ 12๐‘

,

for ๐‘ โ‰ฅ ๐‘€ . Hence, by Step 3,โŸจ|๐บ๐‘ (b) โˆ’ ใ€ˆ๐บ๐‘ (b)ใ€‰ โˆ’ (๐บ๐‘€ (b) โˆ’ ใ€ˆ๐บ๐‘€ (b)ใ€‰) |2๐‘

โŸฉ 12๐‘ ๐‘€,๐‘โ†’โˆžโˆ’โ†’ 0,

for any 1 โ‰ค ๐‘ < โˆž.

Step 5 We claim that ๐บ๐‘ (b) โ†’ ๐บ (b) in ๐ฟ2๐‘ใ€ˆยทใ€‰ as ๐‘ โ†’ โˆž.Indeed, by Step 2 we know that ๐บ๐‘ (b) โ†’ ๐บ (b) almost surely, so that with the result from

Step 4 we may conclude thatใ€ˆ๐บ๐‘ (b)ใ€‰ = ๐บ๐‘ (b) โˆ’ (๐บ๐‘ (b) โˆ’ ใ€ˆ๐บ๐‘ (b)ใ€‰) โ†’ ๐บ (b) โˆ’๐บโˆ—(b)

almost surely along a subsequence as ๐‘ โ†’ โˆž. But since ใ€ˆ๐บ๐‘ (b)ใ€‰ is constant in b (recall that bdenotes the dummy variable over which ใ€ˆยทใ€‰ integrates), the random variable ๐บ (b) โˆ’๐บโˆ—(b) mustbe almost surely constant, i.e.

๐บ (b) โˆ’๐บโˆ—(b) = ใ€ˆ๐บ (b) โˆ’๐บโˆ—(b)ใ€‰ = ใ€ˆ๐บ (b)ใ€‰,since ใ€ˆ๐บใ€‰ < โˆž by assumption and ๐บโˆ— is centred. Hence ใ€ˆ๐บ๐‘ (b)ใ€‰ โ†’ ใ€ˆ๐บ (b)ใ€‰ as ๐‘ โ†’ โˆž alonga subsequence. Since the above argument can be repeated for any subsequence, we obtainthat ใ€ˆ๐บ๐‘ (b)ใ€‰ โ†’ ใ€ˆ๐บ (b)ใ€‰ as ๐‘ โ†’ โˆž. Since ๐บ๐‘ (b) โ†’ ๐บ (b) almost surely, we conclude that๐บโˆ—(b) = ๐บ (b) โˆ’ ใ€ˆ๐บ (b)ใ€‰ almost surely, which together with Step 4 implies that ๐บ๐‘ (b) โ†’ ๐บ (b) in๐ฟ2๐‘ใ€ˆยทใ€‰ as ๐‘ โ†’ โˆž.

By Step 3โ€“5, we conclude that๐บ๐‘ โ†’ ๐บ with respect to the norm ใ€ˆ|๐บ (b) |2๐‘ใ€‰12๐‘ + ใ€ˆโ€– ๐œ•๐บ

๐œ•b(b)โ€–2๐‘

๐ฟ2ใ€‰

12๐‘ .๏ฟฝ

References[BCD11] H. Bahouri, J.-Y. Chemin, and R. Danchin, Fourier Analysis and Nonlinear Partial Dierential Equations.

Grundlehren der mathematischen Wissenschaften 343. Springer, Berlin, Heidelberg, 2011. ISBN: 978-3-642-16829-1. MR 2768550. Zbl 1227.35004.

[BL76] J. Bergh, and J. Lรถfstrรถm, Interpolation Spaces. An Introduction. Grundlehren der mathematischen Wis-senschaften 223. Springer, Berlin, Heidelberg, 1976. ISBN: 978-3-642-66453-3. MR 0482275. Zbl 0344.46071.

[BBH94] F. Bethuel, H. Brezis, and F. Hรฉlein, Ginzburg-Landau vortices. Reprint of the 1994 edition. Mod-ern Birkhรคuser Classics. Birkhรคuser/Springer, Cham, 2017. ISBN: 978-3-319-66672-3. MR 3618899.Zbl 1372.35002.

[CH16] A. Chandra, and M. Hairer, An analytic BPHZ theorem for regularity structures. Preprint (2016).arXiv:1612.08138.

[CMW19] A. Chandra, A. Moinat, and H. Weber, A priori bounds for the ฮฆ4 equation in the full sub-criticalregime. Preprint (2019). arXiv:1910.13854.

[Dal93] G. Dal Maso, An Introduction to ฮ“-convergence. Progress in Nonlinear Dierential Equations and theirApplications 8. Birkhรคuser, Boston, MA, 1993. ISBN: 978-1-4612-6709-6. MR 1201152. Zbl 0816.49001.

[DD03] G. Da Prato, and A. Debussche, Strong solutions to the stochastic quantization equations. The Annalsof Probability 31 (2003), no. 4, 1900โ€“1916. MR 2016604. Zbl 1071.81070.

[DG20] M. Duerinckx, and A. Gloria, Multiscale functional inequalities in probability: Constructive approach.Annales Henri Lebesgue 3 (2020), 825โ€“872. MR 4149827.

[FH14] P.K. Friz, and M. Hairer, A Course on Rough Paths. With an Introduction to Regularity Structures. 2ndedition. Universitext. Springer, Cham, 2020. ISBN: 978-3-030-41555-6. MR 3289027. Zbl 07214193.

Page 63: Variational methods for a singular SPDE yielding the

VARIATIONAL METHODS FOR A SINGULAR SPDE 63

[GJO15] M. Goldman, M. Josien, and F. Otto, New bounds for the inhomogenous Burgers and the Kuramotoโ€“Sivashinsky equations. Communications in Partial Dierential Equations 40 (2015), no. 12, 2237โ€“2265.MR 3421759. Zbl 06539046.

[Gra14] L. Grafakos, Classical Fourier Analysis, 3rd edition. Graduate Texts in Mathematics 249. Springer, NewYork, 2014. ISBN: 978-1-4939-1193-6. MR 3243734. Zbl 1304.42001.

[GH19] M. Gubinelli, and M. Hofmanovรก, Global solutions to elliptic and parabolic ฮฆ4 models in Euclideanspace. Communications in Mathematical Physics 368 (2019), no. 3, 1201โ€“1266. MR 3951704. Zbl 1420.35481.

[Hai14] M. Hairer, A theory of regularity structures. Inventiones Mathematicae 198 (2014), no. 2, 269โ€“504.MR 3274562. Zbl 1332.60093.

[HS17] M. Hairer, and H. Shen, A central limit theorem for the KPZ equation. The Annals of Probability 45(2017), no. 6B, 4167โ€“4221. MR 3737909. Zbl 1388.60111.

[Har68] K.J. Harte, Theory of Magnetization Ripple in Ferromagnetic Films. Journal of Applied Physics 39 (1968),1503โ€“1524.

[Hel98] B. Helffer, Remarks on the decay of correlations and Witten Laplacians, Brascampโ€“Lieb inequalities andsemiclassical limit. Journal of Functional Analysis 155 (1998), no. 2, 571โ€“586. MR 1624506. Zbl 0921.35141.

[Hof68] H. Hoffmann, Theory of magnetization ripple. IEEE Transactions on Magnetics 4 (1968), no. 1, 32โ€“38.[IJ19] R. Ignat, and R.L. Jerrard, Renormalized energy between vortices in some Ginzburgโ€“Landau models on

2-dimensional Riemannian manifolds. Preprint (2019). arXiv:1910.02921.[IM16] R. Ignat, and R. Moser, Interaction energy of domain walls in a nonlocal Ginzburg-Landau type model

from micromagnetics. Archive for Rational Mechanics and Analysis 221 (2016), no. 1, 419โ€“485. MR 3483899.Zbl 1342.35362.

[IO19] R. Ignat, and F. Otto, The magnetization ripple: A nonlocal stochastic PDE perspective. Journal deMathรฉmatiques Pures et Appliquรฉes 130 (2019), 157โ€“199. MR 4001632. Zbl 07107304.

[JO20] M. Josien, and F. Otto, The annealed Calderรณnโ€“Zygmund estimate as convenient tool in quantitativestochastic homogenization. Preprint (2020). arXiv:2005.08811.

[Kur06] M. Kurzke, Boundary vortices in thin magnetic lms. Calculus of Variations and Partial Dierential

Equations 26 (2006), no. 1, 1โ€“28. MR 2214879. Zbl 1151.35006.[MW17a] J.-C. Mourrat, and H. Weber, Global well-posedness of the dynamic ฮฆ4 model in the plane. The Annals

of Probability 45 (2017), no. 4, 2398โ€“2476. MR 3693966. Zbl 1381.60098.[MW17b] J.-C. Mourrat, and H. Weber, The dynamic ฮฆ4

3 comes down from innity. Communications in Mathe-

matical Physics 356 (2017), no. 3, 673โ€“753. MR 3719541. Zbl 1384.81068.[OS10] F. Otto, and J. Steiner, The concertina pattern: from micromagnetics to domain theory. Calculus of

Variations and Partial Dierential Equations 39 (2010), no. 1โ€“2, 139โ€“181. MR 2659683. Zbl 1237.78036.[OW19] F. Otto, and H. Weber, Quasilinear SPDEs via rough paths. Archive for Rational Mechanics and Analysis

232 (2019), no. 2, 873โ€“950. MR 3925533. Zbl 1426.60090.[SS07] E. Sandier, and S. Serfaty, Vortices in the magnetic Ginzburgโ€“Landau model. Progress in Nonlinear

Dierential Equations and their Applications 70, Birkhรคuser Boston, Boston, MA, 2007. ISBN: 978-0-8176-4316-4. MR 2279839. Zbl 1112.35002.

[SS15] E. Sandier, and S. Serfaty, 2D Coulomb gases and the renormalized energy. The Annals of Probability43 (2015), no. 4, 2026โ€“2083. MR 3353821. Zbl 1328.82006.

[SSWMO12] J. Steiner, R. Schรคfer, H. Wieczoreck, J. McCord, and F. Otto, Formation and coarsening of theconcertina magnetization pattern in elongated thin-lm elements. Physical Review B 85 (2012), 104407.

[TW18] P. Tsatsoulis, and H. Weber, Spectral gap for the stochastic quantization equation on the 2-dimensionaltorus.Annales de lโ€™Institut Henri Poincarรฉ Probabilitรฉs et Statistiques 54 (2018), no. 3, 1204โ€“1249. MR 3825880.Zbl 1403.81030.

(R. Ignat) Institut deMathรฉmatiqes de Toulouse & Institut Universitaire de France, UMR 5219, Universitรฉde Toulouse, CNRS, UPS IMT, F-31062 Toulouse Cedex 9, France

Email address: [email protected]

(F. Otto)Max-Planck-Institut fรผr Mathematik in den Naturwissenschaften, Inselstrasse 22, 04103 Leipzig,Germany

Email address: [email protected]

(T. Ried)Max-Planck-Institut fรผr Mathematik in den Naturwissenschaften, Inselstrasse 22, 04103 Leipzig,Germany

Email address: [email protected]

(P. Tsatsoulis) Max-Planck-Institut fรผr Mathematik in den Naturwissenschaften, Inselstrasse 22, 04103Leipzig, Germany

Email address: [email protected]