Improved quasi-invariance result for the periodic Benjamin–Ono–BBM equation Authors: Justin Forlano Published: 2026 ABSTRACT AbstractWe extend recent results of Genovese–Luca–Tzvetkov (2022) regarding the quasi-invariance of Gaussian measures under the flow of the periodic Benjamin–Ono–BBM (BO-BBM) equation to the full range where BO-BBM is globally well-posed. The main difficulty is due to the critical nature of the dispersion which we overcome by combining the approach of Coe–Tolomeo (2024) with an iteration argument due to Forlano–Tolomeo (2024) to obtain long-time higher integrability bounds on the transported density.Keywords: Benjamin–Ono–BBM, Quasi-invariance, Gaussian measure Introduction We consider the Benjamin–Ono–BBM (BO-BBM) equation: where |Dx| is the Fourier multiplier operator with multiplier |n|. The BO-BBM Eq. (1.1) arises as a model for the evolution of long-crested waves at the interface between two immiscible fluids. It is closely connected to the Benjamin–Ono (BO) equation [4, 38]: as, formally, (1.1) arises from replacing ∂x in the second term in (1.2) by ∂t. This substitution has some physical significance related to the formal derivation of these models, see [5, 6]. To proceed with an analytical point of view, we may rewrite (1.1) as As compared to (1.2), the derivative in the nonlinearity of (1.1) has been completely ameliorated. In this sense, (1.1) is also referred to as the regularised Benjamin–Ono equation [1]. This moniker is partly merited by the following fact: whilst the well-posedness problem for the BO equation has historically been a difficult topic culminating in [21, 25] and the references therein, the well-posedness problem for BO-BBM (1.1) is much simpler. To describe the well-posedness situation further, we need to set some notation. We define the Fourier transform by and so that F{∂xu}(n)=inu^(n) and The mean ∫Tudx is a conserved quantity for (1.1), and this allows us to, although not necessarily [2], restrict our discussions to the mean-zero Sobolev spaces H0σ(T), with the norm where Z∗:=Z\{0}. As Hσ(T) is an algebra when σ>1/2 and the operator (1+|Dx|)-1∂x is bounded on Hσ(T), we obtain local well-posedness for (1.3) in Hσ(T). Global well-posedness for σ≥1/2 is also known, with the case σ>1/2 in [2, 29] following by combining the Brezis–Gallouët inequality with conservation of the energy This yields a super-exponential growth bound for the solutions: It turns out that in fact a mere exponential bound is available, which appears to be absent in the literature. Theorem 1.1 Let σ>12, u0∈Hσ(T), and u be the unique global in time solution to (1.1) satisfying u(0)=u0. Then, there exists C0,C1>0 such that for any t∈R, We recall that in [19, Proposition 2.2], the authors showed that solutions to (1.1) satisfy1 for any s≥0. By interpolation and the conservation of (1.4), (1.7) yields an exponential growth bound on solutions as in (1.6) when s>1, namely for σ>32. By refining the argument leading to (1.7), we are able to establish (1.6) for any σ>12. The global well-posedness at the endpoint σ=1/2 was shown in [23] by a compactness argument from conservation of E and a Yudovich-type argument to recover uniqueness. Strictly speaking, [23] established this result on the real line R but the proof extends without additional complication to the periodic setting. We also point out that Theorem 1.1 also holds on R. Thus, we have a well-defined global-in-time flow u0↦Φt(u0) on Hσ(T) for any σ≥1/2. The well-posedness/ill-posedness of (1.3) is currently open for σ<1/2. See [2] for a mild form of ill-posedness (failure of C2-regularity for the solution map) when σ<0. Our interest in (1.1) is to improve upon the known regularity ranges under which Gaussian measures remain absolutely continuous under the flow of (1.1) and to the point at which agrees with the well-posedness theory. In the setting of Hamiltonian PDEs, this direction was initiated by Tzvetkov [45] and has since seen a remarkable amount of progress [10, 12, 13, 15–20, 22, 27, 32–36, 39–41, 44, 45]. In particular, these results together show that there is a delicate balance between positive results in this area (quasi-invariance) and the amount of dispersion and/or structure in the given equation. In this vein, BO-BBM (1.1) is an interesting model for this study as the amount of dispersion seems to be borderline for positive quasi-invariance results to hold; see [20], Theorem 1.2, and the ensuing discussion below. Let us first describe this quasi-invariance problem in more detail, focusing on the situation for the mean-zero flow of BO-BBM (1.1). Given s∈R, we consider the Gaussian measures μs formally prescribed as and rigorously defined as the law of the random Fourier series where {gn(ω)}n∈Z∗ are a family of complex-valued standard normal random variables on some probability space (Ω,F,P) and conditioned so that g-n(ω)=gn(ω)¯. These Gaussian measures are supported on the Sobolev spaces H0σ(T) for any σ1, where ∼ denotes the mutual absolute continuity of measures, namely (Φt)#μs≪μs and μs≪(Φt)#μs. The property (1.9) is called quasi-invariance. The authors in [20] were able to say more than (1.9), as we describe now. Given R>0, we denote by BR the closed ball centred at zero of radius R in the energy space; namely, and define the localised measure The addition of the (invariant) cut-off 1BR allowed them to show that the transported density (Radon–Nikodym derivative) enjoys higher Lp-integrability, than merely belonging to L1(dμs,R). However, as a consequence of the critical nature of the dispersion in (1.3), the integrability index p=p(t,R) depends on t>0 and has the property that p(|t|,R)→1 as |t|→+∞. Interestingly, this deterioration of the exponent does not occur for models with even slightly stronger dispersion, such as for the fractional BBM equations with β>1 (where β=1 corresponds to (1.1)); see [19]. It is in this sense that we see BO-BBM (1.1) as a model with a critical amount of dispersion for quasi-invariance to hold. Note that a critical aspect also appears for nonlinear wave equations [22, 36]. Our goal is to extend the results of [20] to the full range s>1/2 where BO-BBM (1.1) is well-posed. Theorem 1.2 Given s>12 and t∈R, the transported measure (Φt)∗μs is mutually absolutely continuous with respect to the Gaussian measure μs. Moreover, for each t>0 and R>0 there exists p=p(|t|,R)>1 such that p(|t|,R)→1 as |t|→∞ and constants C1(R),C2(R)>0, for which the Radon–Nikodym derivative ft,R defined in (1.12) satisfies The regularity restriction of s>1 in [20] arises from an energy estimate on logft,R, which is essentially (1.7). More precisely, it is enough to establish the Lp(μs) boundedness for the right-hand side of (1.7) at time t=0. This reduction allows to use the probabilistic properties of the initial data. Here, in order to overcome the logarithmic divergence from the Hs-factors, one uses the μs almost sure finiteness of ‖∂xu‖L∞ which imposes s>1. Our strategy to go beyond the analysis in [20] is to adapt the recent approach of [12] which allows to use probability in a more complete way as well as make use of a symmetrisation argument in (3.6). More precisely, we can exploit multilinear probabilistic oscillations by controlling the left-hand side of (1.7), rather than passing to the right-hand side which only allows to use probability at a linear level. The main difficulty, which is also the key to technical result leading to Theorem 1.2, is then to establish the exponential integrability of logft,R (Lemma 3.3), and the variational formula [3, 7, 46] provides a sharp way to prove this. Once we have in hand Lemma 3.3, it is possible to obtain the quasi-invariance using an argument from [36], which is based on tracking the growth of the measure of a transported set, and from this obtain the higher integrability of the density (1.14) using [10, Proposition 3.5]. We instead give an alternative, but equivalent, analysis focusing purely on the density. However, due to the critical dispersion of BO-BBM (1.1) it is not enough to simply apply the method in [12]. Indeed, we are not able to obtain (3.11) as directly as one would in [12] because of the short-time restriction in Lemma 3.4. We need to go one step beyond this by carefully iterating these bounds on short-time intervals to reach long times (Proposition 3.5). For this step, we adapt an argument from [17]. At a high level, we advocate that this strategy arising from combining ideas from [12, 17] provides an effective way to analyse the low-regularity quasi-invariance problem for equations with a borderline amount of dispersion. As discussed in [20], a complete solution to the programme initiated in [45] would constitute finding a formula for the transported density with respect to the unrestricted Gaussian measure μs. This has only been achieved in a few works [15, 19, 27, 40] and relies on more dispersion, namely on good energy estimates for solutions with smoothing. It seems to be an interesting but difficult problem to determine if (i) one could remove the localisation BR in (1.14) and (ii) obtain a formula for ft,R, and the density with respect to μs. In particular, we can no longer reduce to time t=0 and thus it is not clear how to guarantee the finiteness of the right-hand side of (1.7). Indeed, there is no multilinear smoothing in the nonlinear term in (1.3), even under randomisation: if u were replaced by u0ω given in (1.8). The situation becomes even more difficult for the regularity range 1212 and N∈N. Then, for any M≥1, it holds that where the implicit constant is uniform in M and N. Proof We have at which point, (2.1) follows from Markov’s inequality. □ We can obtain an exponential decay rate in (2.1) but the polynomial one here will be sufficient. Localised density bounds Given N∈N, we define the truncated version of (1.3): Similar to (1.1), for every fixed N∈N, (3.1) is globally well-posed in Hσ, for σ>12, with bounds on the Sobolev norms of the solutions that are uniform in N. This can be proved by repeating the arguments in [29] and using the L2-boundedness of P≤N. We denote the flow of (3.1) by ΦtN. Note that if uN solves (3.1), then uN∈C∞(R×T), which justifies the ensuing computations. As the truncated Eq. (3.1) is a finite-dimensional Hamiltonian system (in the Fourier coefficients for uN), Liouville’s theorem implies that the Lebesgue measure LN on R2N is invariant under ΦtN. Thus, for every t∈R, where ftN is the transported density. By the fundamental theorem of calculus, we have where Qs,N(u):=Qs,N(u,u,u) is the trilinear operator From (3.1), the identity (1+|Dx|)-1|Dx|=1-(1+|Dx|)-1 and for uN:=P≤Nu, we then have Notice that the contribution from the linear part in (3.4) vanishes identically since By a symmetrisation argument, we write (3.4) this more favourably as where we define the trilinear operators Qs,N(j)(u):=Qs,N(j)(u,u,u), j=1,2, for u∈Π≤NL2, by and The multilinear operator Qs,N(1) is worse behaved as compared to Qs,N(2) because it lacks the factor (1+|Dx|)-1. Nonetheless, the symmetrisation in Qs,N(1) allows us to always place the derivative ∂x onto the lowest frequency term, in the following sense. Lemma 3.1 Let s>12 and n1,n2,n3∈Z∗ such that n1+n2+n3=0. Then, either (i) |n1|∼|n2|∼|n3| or (ii) |n(2)|≫|n(3)| and hence |Ψs(n¯)|≲|n(1)|2s|n(3)|. Proof By symmetry, we may assume that |n1|≥|n2|≥|n3|. Thus, when |n2|≫|n3|, we must have sgn(n1)=-sgn(n2), so by the mean value theorem, □ We note that since E in (1.4) is a conserved quantity on the support of μs, BR is an invariant set under ΦtN; namely, (ΦtN)-1(BR)=BR for any t∈R. Remark 3.2 We note that the density ftN localised to BR always has finite Lp(dμs)-moments, albeit with a severe loss in uniformity in N. Indeed, for u∈BR, we have ΦtN(u)∈BR and thus Sobolev embeddings imply Our goal is to obtain bounds which are uniform in N. The main ingredient in this article is the following exponential integrability statement. Lemma 3.3 Let 120. Then, there exist constants c0>0 and A>1 independent of both R and N such that for all 0≤λ≤c0R-12, it holds that We postpone the proof of Lemma 3.3 to Sect. 4. Using Lemma 3.3, we obtain the following uniform in N, short-time Lp bounds on the densities. Lemma 3.4 Let 120. Then, there exists c0>0 independent of both R and N such that for any t∈R and 1≤p<∞ satisfying 0≤p|t|≤c0R-12, there exists C(R,p)>0 such that Proof The proof largely follows that of [12, Proposition 3.6]; however, as BR is invariant under ΦtN, there are some simplifications. For the convenience of the reader, we provide the details. First, we reduce the proof of the bound (3.10) for ftN1BR in such a way that it remains to apply Lemma 3.3. Let It=[0,t] for t≥0 and It=[t,0] for t<0. Let t≠0. By the formula (3.2), the group property of the flow ΦtN, Jensen’s inequality, Fubini’s theorem, and the invariance of BR, we have Note that for each fixed τ∈It, we can repeat this argument to also obtain where we used that e|tpQs,N(u)| is increasing in |t|. Thus, by Remark (3.2), we find The estimate (3.10) now follows from this and Lemma 3.3. □ We now iterate the short-time bounds (3.11) to reach arbitrary times t∈R, namely to times which no longer depend on R. The argument here is similar to that in [17]. We point out again that this extra step seems to be necessary due to the restriction on λ in Lemma 3.3 which itself is a symptom of the critical nature of (1.1) and the use of an energy cut-off. Proposition 3.5 Let 120. Then, there exists p(t,R)>1, and constants C1(R),C2(R)>0 such that The strategy is similar to that of [17, Proposition 4.14]; however, to make this article as self-contained as possible, we provide some of the details below as well as the modifications. In particular, to obtain p(t, R) for general t∈R, we need to more carefully choose the exponents later in the argument as compared to [17]. Proof First, we claim that for any t,τ∈R and 1≤rr1 be sufficiently large so that and choose τR>0 such that where c0 is the constant in Lemma 3.3. For j∈N, we define the sequence of times tj=jτR and integrability exponents rj by Note that rj=rj+1(p-1)p-rj+1 and 10. As we need to account for a further deterioration in the integrability exponent, and in view of (3.16), we define q:[0,∞)→(1,∞) by We see that q(t) is monotonically decreasing, q(0)=r1(p-1)p-r1

τR and such that there is j0∈N (which depends on t and τR) such that j0τR0. Fix t∈R and R>0. Then, with p=p(t,R)>1 as in Proposition 3.5, (3.11) implies that there exists ft,R∈Lp(dμs) such that, up to a subsequence, (ftN1BR)∘Π≤N converges weakly to ft,R in Lp(dμs) as N→∞. We now show that ft,R is supported on the ball BR in H01/2(T). Let δ>0. By the weak convergence of (ftN1BR)∘Π≤N, independence of Π≤Nu and Π>Nu for N∈N, and (2.1), we have Thus, for every δ>0, and so by the monotone convergence theorem, we conclude that Now, we show that ft,R is the Radon–Nikodym derivative of (Φt)#μs,R with respect to μs,R. Assuming this for now, we get (Φt)#μs,R≪μs,R. By taking R→∞, this implies (Φt)#μs≪μs. Let F:H01/2(T)→R be continuous, bounded, and non-negative. By the convergence in (3.23) and the dominated convergence theorem Similarly, by (3.11) and the dominated convergence theorem, Then, by monotone convergence, (3.25), (3.26) and (3.24), we have As this identity holds for all such F, we obtain that d(Φt)#μs,R=ft,Rdμs,R. This completes the proof of Theorem 1.2. □ Proof of Lemma 3.3 We need exponential integrability of Qs,N, uniformly in N and locally on the support of μs,N. Using the simplified variational formula [16, Lemma 2.6], we have where Y is distributed according to μs,N, YN:=P≤NY, and VN:=P≤NV. For YN+VN∈BR, the triangle inequality implies The goal is to show that the right-hand side of (4.1) is finite, uniformly in N∈N. More precisely, we will show that provided that λR12≪1,2 and for some c1(s),c2(s)>0 and C>0 which do not depend on N. Here, c1(s) and c2(s) are in principle computable, with the largest power c2 coming from (4.8). Then, (3.9) is obtained from combining (4.1) and (4.3) and taking the exponential of both sides, which leads to C(R)=C[1+Rc2(s)] and A=c1(s). The key is to control the V-factors by the negative terms 12‖VN‖Hs+1/22, as well as leveraging the cut-off 1BR to control the potential higher homogeneity from the terms in Qs,N(YN+VN) which contain exactly three V inputs. To do this, we need to decompose the operator Qs,N as in (3.5) and then further decompose each of Qs,N(j), j=1,2 as in (3.6) and (3.7) into a certain finite number of parts L∈N; namely, for some operators {Qs,N,ℓ(j)}ℓ=1L, j=1,2, and control their contributions to (4.1) separately. In order to use the negative terms in 12‖VN‖Hs+1/22 for each of these {Qs,N,ℓ}ℓ=1L, we use It then suffices to control each of the summands. To be concrete, it will be more than enough in the following to take L=100. In the following sections, to simplify the notation, we write Y and V in place of YN and VN, respectively. Contribution from Qs,N(1) In view of Lemma 3.1, we make the following dyadic decompositions: We first consider the contribution from (4.5). By symmetry, we may assume that N1≥N2≥N3. Then, with u=Y+V, we have the following unique cases to consider: In these cases, by Lemma 3.1, we always have |Ψs(n¯)|≲N12sN3. Case1:(YN1,YN2,YN3)_ For this case, we need to use the gain coming from random oscillations. We have We use Isserlis theorem to control the expectation and exploit the independence structure. As all the individual frequencies are nonzero, we cannot have any self-interactions of the form nj+nj′=0 for j≠j′. Moreover, since N1,N2≫N3, we must have that n3=m3 and thus (n1,n2)=(m1,m2) or (n1,n2)=(m2,m1). Both of these cases yield the same contribution, so we then have By first summing in N1 over N1≫N3 and using that s>12 to sum in N3, this contribution to (4.1) is thus controlled. Case2:(YN1,YN2,VN3)_ We write where Let 0<δ0 such that Then, by Cauchy–Schwarz and Cauchy’s inequality, we have It remains to control the expected value of the second term. As in Case 1, we have This is summable, first over N1≫N3, and then over N3 since s>12. Case3:(YN1,VN2,YN3)_ In this case, we proceed similarly to Case 2, writing where Note that E[gn1gn3gm1gm3¯]=δn1m1δn3m3 since |n1|≫|n3|. Thus, we have We then sum as in Case 2. Case4:(YN1,VN2,VN3)_ A simple Cauchy–Schwarz argument, that uses only the regularity of YN1 and not its Gaussian structure, suffices to control this contribution when s<1 but fails when s≥1. In the following, we present a unified approach that deals with all s>12, which does use the explicit structure of YN1. Let 0<δ<14(s-12), ε=ε(s,δ)>0 such that and define the random variable By Cauchy–Schwarz, interpolation, (4.2), and Young’s inequality, we have It remains to control the first term under after taking an expectation. For any finite p≥2, Minkowski’s inequality implies Thus, by Cauchy–Schwarz and (4.9), we have which is summable, first in N1≫N3 and using that s>12 and δ is sufficiently small. Case5:(VN1,VN2,YN3+VN3)_ By Cauchy–Schwarz, we have We can handle the N3 summation using ∑N3≪N1N1-1N3≲1, and then the summations over N1 and N2 since Thus, for Y+V∈BR, there is a C>0 independent of N∈N and R>0, such that Thus, provided that we choose λ>0 so that This completes the estimates for (4.5). We move onto (4.4), where all the frequencies are similar. In this case, we always have |Ψs(n¯)|≲N12s+1. By symmetry, we only have three cases depending on how many Y factors there are. Case1:(VN1,VN2,YN3+VN3)_ This contribution can be handled exactly as in Case 5 when we controlled (4.5), and we obtain which is the same bound as in (4.11). Case 2: (YN1,VN2,YN3) We argue exactly as in Case 3 when we handled (4.5). With KN1N2N3(2) chosen analogously to (4.7) in (4.6), when N1∼N2∼N3, we obtain which is summable as s>12 and we may choose δ>0 sufficiently small. Case 3: (YN1,YN2,YN3) We compute which is summable as s>12. This completes the estimation of (4.4) and thus also the estimates for the main part Qs,N(1). Contribution from Qs,N(2) We now handle the easier contribution in (4.1) coming from Qs,N(2). We further expand this operator as follows: Whilst the first term Qs,N(2,A) is not critical in the sense that Qs,N(1) is, there are still high-high-low interactions which require s-derivatives to be absorbed by the high-frequency factors, which is too much for the Y factors to handle in a naive way. Thus, we still require some decompositions depending on the factors Y or V. As a warm up to this, we dispense with the term Qs,N(2,B). Lemma 4.1 For any 120 such that For the first term, we then get which is summable. Finally, we have the case where both high-frequency factors are V terms. Similar to the previous case, we have If s<1, then s-12<12 and we have By repeating the Cauchy–Schwarz argument from the previous case, we control the contribution of this term. If s=1, by (4.2) we have for any ε>0. Thus, the contribution from this term is controlled after performing the dyadic summations and choosing ε>0 sufficiently small depending on L. This completes the estimation of Qs,N(2,A) and thus the proof of Lemma 3.3. Proof of Proposition 1.3 We revisit some of the computations we made in the previous sections to show the divergent lower bounds. Let 00. Now, where By the Isserlis’ theorem, we have Next, we also have where we used that |Φ(n¯)|≲1 for all (n1,n2,n3) and the assumption |t|≲1. We now split into two cases. First, suppose that 014, we then obtain and hence (1.6) follows by Gronwall’s inequality. This completes the proof of Theorem 1.1. Acknowledgements J.F. was supported by the European Research Council (grant no. 864138 “SingStochDispDyn”) and partially supported by the ARC project FT23010058. He would like to thank the anonymous referees for pointing out an error in a previous version and whose comments have improved the presentation. He would also like to thank Nikolay Tzvetkov and Andreia Chapouto for helpful comments. Funding Open Access funding enabled and organized by CAUL and its Member Institutions. Data availability Data sharing was not applicable to this article as no datasets were generated or analysed during the current study. Declarations Conflict of interest The author declares that they have no conflict of interest. Footnotes Strictly speaking, (1.7) was proved for solutions to a truncated version of (1.1) but we ignore this technicality in this discussion. See (4.12). Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement Data sharing was not applicable to this article as no datasets were generated or analysed during the current study.