We consider two vector fields X,Y∈Wloc1,p(Rd)∩L∞ with bounded divergence. If p≥2, this paper shows that DXY−DYX=0 almost everywhere if and only if their Regular Lagrangian Flows commute. If instead 1≤p<2, we present an example of two divergence free locally Sobolev vector fields in R2 with vanishing Lie brackets whose flows do not commute. In the regime 2≤p<∞ this paper also yields a quantitative commutation result not depending on the Lipschitz constant for vector fields with [X,Y]=0, which is new even in the smooth setting.
Introduction
If X∈Wloc1,p(Rd)∩L∞ is a Sobolev vector field with divX∈L∞, DiPerna and Lions [4] (and then Ambrosio [1] in the BV setting) proved that there is a unique natural flow map ΦXt:Rd→Rd such that for all T>0 there is C>0 such that there holds
(ΦXt)♯Ld≤CLdfor all ∣t∣≤T(1)
and for almost every x, the map t↦ΦXt(x) is an absolutely continuous solution to
{dtdΦXt(x)=X(ΦXt(x)),ΦX0(x)=x.(2)
The map ΦXt is called the Regular Lagrangian flow of the vector field X. The classical Frobenius theorem states that the (classical) flows ΦXt,ΦYs of two smooth vector fields X,Y:Rd→Rd commute
ΦXt∘ΦYs=ΦYs∘ΦXt,∀t,s≥0
if and only if their Lie bracket vanishes, i.e.
[X,Y]:=DXY−DYX=0.(3)
The aim of this paper is to extend the classical Frobenius theorem to the setting of Regular Lagrangian Flows.
Let X,Y∈W1,p be two bounded Sobolev fields, with bounded divergence. If their Regular Lagrangian Flows commute, it is not difficult (see e.g. [2]) to show that [X,Y]=0 almost everywhere, therefore this paper and the previous literature on the topic only focus on the opposite implication. Assuming that (3) is satisfied almost everywhere, the following results are known about the commutation of their Regular Lagrangian flows.
A first paper [10] proves that the flows commute when X,Y are locally Lipschitz.
Key words and phrases. Sobolev vector fields, Regular Lagrangian flows, Lie brackets.
Later, [2] proved that the flows commute assuming additionally that for every t∈R, the flow map ΦXt is weakly Lie differentiable in the direction Y, with locally bounded derivative.
The paper [12] proved the result when one of the vector fields is Lipschitz and the other belongs to W1,p and admits a unique Regular Lagrangian flow.
Finally [11] proved the commutativity of the flows for bounded, continuous Sobolev vector fields with bounded divergence on R2.
The main result of this paper is as follows. Here d is any dimension d≥2.
Theorem 1.1. Let X,Y∈Wloc1,p(Rd)∩L∞ with divX,divY∈L∞, and assume that [X,Y]=0. If p≥2, then for all t,s>0,
ΦXt∘ΦYs(x)=ΦYs∘ΦXt(x)for a.e. x.(4)
If instead 1≤p<2, there are X,Y∈Wloc1,p(Rd)∩L∞ with zero divergence, such that (1.4) fails for x in an open set with positive measure.
Some comments are in order.
Lie differential and commutation
To understand why a different approach is needed compared to the existing literature, assume that the flow ΦXt is Lie-differentiable in the direction of Y, with locally bounded derivative (see [2]). For the purpose of this sketch, assume that there exists a function ρt∈L∞([0,T]×Rd;Rd) such that
h1(ΦXt∘ΦYh−ΦXt)⟶ρtin Lloc1 as h→0+.
At least formally ρt satisfies
∂tρt=DX(ΦXt)ρt,ρ0=Y.(5)
Notice that by applying Gronwall’s estimate along almost every trajectory of X, the above equation has a unique solution ρt in L∞. However, due to the assumption [X,Y]=0, it is immediate to check that Y(ΦXt(x)) is, in fact, another bounded solution of (1.5), and therefore
ρt=Y(ΦXt).
Therefore
dsdΦXt∘ΦYs(x)=Y(ΦXt∘ΦYs(x))
so that s↦ΦXt∘ΦYs(x) is a trajectory of the flow of Y for a.e. x, starting at the point ΦXt(x). Since it satisfies (1.1) and (1.2), again by uniqueness of the DiPerna Lions flows one must have that it is equal to ΦYs∘ΦXt(x) at a.e. x.
Results about the differentiability of the Regular Lagrangian Flow of Sobolev fields are available in two dimensions under continuity assumptions [7] (and, classically, for Lipschitz vector fields in general dimension), but it is well known, see [7, 9] that the RLF of Sobolev vector fields need not be differentiable or inherit any Sobolev regularity. In particular, there seems not to be an easy way to fix the previous argument in a general Sobolev setting.
Quantitative commutation
For p>2, strictly greater than the critical exponent, the proof also provides a quantitative commutation result for vector fields in W1,p, as follows.
Theorem 1.2. Let p>2 and let X,Y∈Wloc1,p(Rd)∩L∞ with divX,divY∈L∞. Then
The implicit constant depends on d,p,T,r and the norms of X,Y in BR+1.
In the critical case X,Y∈W1,2, one readily deduces the existence of an analogous quantitative result, using the compactness and stability properties of the Regular Lagrangian Flows. However, the uniform modulus that we can obtain when p=2 is not as clean as in the case p>2, as it will be clear from the proof below.
Since X,Y∈L∞, by a standard localization argument it is easy to see that it is equivalent to prove the commutativity results for vector fields in W1,p(Rd). To avoid technical issues, we will restrict to this class from now on.
The regime p<2
The example of non-commuting flows in the regime p<2 is surprisingly simple, and it is constructed using a double cover of the plane Q:Rx2→Ry2 (see Figure 1). The example was essentially already known from [8] (see also [5, 6]), here we just need to check its Sobolev regularity, which is immediate. Say that, in polar coordinates
Figure 1. Example of two divergence free fields X,Y∈Wloc1,p(R2), p<2, whose flows do not commute. They are constructed as pullbacks of ei through a branching map Q:Rx2→Ry2. For example, in polar coordinates, one can take Q(ρ,θ)=(ρ,2θ).
Q(ρ,θ)=(ρ,2θ).
Vector fields with vanishing Lie brackets X,Y producing non-commuting flows can be defined as the pullback of e1=(1,0), e2=(0,1) via the map Q:
X=(DQ)−1e1,Y=(DQ)−1e2.
It is easy to see that they satisfy all the conditions of Theorem 1.1 if and only if p<2. Their flows do not commute: the reason is that, while clearly e1,e2 commute in Ry2, when they are pulled back via the branching map Q, the points ΦXt∘ΦYs and ΦYs∘ΦXt may well end up in different branches of the map Q in the original space Rx2. The exponent p=2 arises as a natural threshold also in [6] for the same geometric reasons.
AI use
We started working on the problem in early 2025, but eventually got stuck and moved away from the topic. In September 2026, we were very surprised when GPT-6 Astra suggested (in a single response) the counterexample in the regime p<2. Moreover, this time after extensive interaction with GPT-6 Astra, in late September 2026 we obtained a first proof, completely different from the one presented here, that worked for p>2 in arbitrary dimension. Finally, in late October 2026, through further interactions with the same model, we obtained the present proof, which we then decided to rework and write up in its current form. The paper was written entirely by the authors.
Short time estimate
We start with some preliminary results, needed for the following.
Maximal functions and truncations
Recall that if f∈W1,p(Rd;Rm) then the maximal function of Df is defined as
For the following statements, we refer to [13]. If p>1, one has
∥gf∥Lp≤Cd,p∥f∥W1,p.
Then there exists a null set N such that for all x,y∈Rd∖N, there holds
∣f(x)−f(y)∣≤Cd(gf(x)+gf(y))∣x−y∣.
On the set
Gλ:={x∈Rd∣gf(x)≤2Cdλ}∖N,
the function f is λ-Lipschitz, therefore, by the Kirszbraun Lipschitz extension theorem f∣Gλ admits a Lipschitz extension fλ, with
∥fλ∥∞≤∥f∥∞,Lipfλ≤λ,fλ(x)=f(x)for a.e. x∈Gλ.
Lusin-Lipschitz estimates
It is well known that the Regular Lagrangian Flow of Sobolev vector fields satisfies Lipschitz estimates outside a small set, see [3]. We recall here a version in our setting.
Lemma 2.1. Let p>1 and X∈W1,p(Rd) be a vector field with bounded divergence, and let gX be its maximal function. Then for every λ>0 and 0≤h≤T there is a set Bλ,h such that
∣Bλ,h∣≤2Cλh∫{gX≥λ/(2h)}gX(x)dx
and
x↦ΦXh(x)is e2Cdλ-Lipschitz in Rd∖Bλ,h.
Proof. Up to a mollification argument, we can assume that X is a smooth vector field. Then we obtain
where we have used (1) to perform the change of variable in the last inequality. Since the first term in the last line can be absorbed in the left hand side, this concludes the proof. □
A short-time estimate
The main goal of this section is to prove the following preliminary proposition, that is a short time estimate on the commutativity of flows when t,s∼h.
Proposition 2.2. Let X,Y∈W1,p(Rd)∩L∞ be vector fields with bounded divergence. Fix h>0 and r>0. Then for all 0≤t,s≤h there holds
where the implicit constant is a dimensional constant.
Proof. Without loss of generality we prove the result for t,s≥0. By a standard mollification argument, it is sufficient to prove the result for smooth vector fields. First consider the quantity
The third term in (11) can be estimated in an entirely symmetrical way, while for the first one can use Lemma 2.3. Combining all these estimates, and recalling the properties of the Lipschitz truncation of Section 2.1, one gets
where we used that ∣DXλ∣,∣DYλ∣≤λ=2Cd/h to pass from the first to the second line and moreover that, on the good set {g≤h−1}, by locality of the weak gradient one has DX=DXλ and DY=DYλ almost everywhere.
A quantitative commutation estimate
In this section we prove the first part of Theorem 1.1, i.e. that the flows commute whenever p≥2.
The following is the key proposition: the initial commutator is split into 4k smaller commutators, up to removing a set with controlled measure. Here we recall that g is the sum of the maximal functions of the derivatives of X,Y:
g=gX+gY
as defined in Section 2.1.
Proposition 3.1. Let X,Y∈W1,p(Rd)∩L∞, with p>1 and bounded divergence. Let 0≤s,t≤T, and let {λj}j∈N be a sequence of positive numbers. Then, for every
In particular, at step k, we decomposed Γ0 into the sum of 4k commutators Γk, left and right composed with some operators. Here Lαjj is one of the four elements appearing in the left compositions in (13), more precisely:
Lαjj∈{Aj,Aj∘Bj,Id,Bj}.(15)
In an entirely similar way
Rαjj∈{Bj,Id,Bj∘Aj,Aj}.(16)
Let us name
Lαk:=Lα11∘…∘Lαkk,Rαk:=Rαkk∘…∘Rα11.
Our aim is to estimate at each step k the commutator Γk with the commutator Γk+1, up to removing a small set where the new four left compositions Lpk+1, p=1,2,3,4 are not eλi-Lipschitz.
Notice that for every j≥1, Lpj is a composition of the flows of X,Y for a total time of order 2−j, as it immediately follows from (16). Recall that g=gX+gY is the sum of the maximal functions of DX, DY, which are in Llocp. By Lemma 2.1, for all j≥0 and p∈{1,2,3,4}, there exist sets Bpj such that
Here in the equality of the third line we used the fact that if α∈Ik′ and p=αk′+1, then α′:=(α,p)∈Ik′+1, and by definition
Rpk′+1∘Rαk′=Rα′k′+1
In the last inequality, we used the fact that Lαk′+1k′+1 is calculated only on pair of points where it is eλk′+1-Lipschitz thanks to the definition of the good set Gk. This proves the inductive step, so that (19) holds in particular for k′=k, for every k≥1.
Here in the last step we used that every right factor Rαk has uniformly bounded compression, since it is a composition of flows of time at most 2T, and that the points Rαk are contained in the ball BRk.
We are now in the position of proving the main theorem for vector fields in W1,2.
Theorem 3.2. Let X,Y∈W1,2(Rd)∩L∞ with bounded divergence, and assume that [X,Y]=0. Then for all s,t, there holds
ΦXt∘ΦYs(x)−ΦYs∘ΦXt(x)=0for a.e. x∈Rd.(22)
Proof. Let r>0 be fixed and take 0≤s,t≤T=1. From Proposition 3.1, we deduce
If η is sufficiently small (η≤log(2)/cd), the right hand side goes to zero as k→+∞ and M is fixed. Then, we let M go to infinity, so that the right hand side of (24) goes to zero as well. This proves the result. □
For p>2, there is a clean quantitative version using a different choice of the sequence λi.
Theorem 3.3. Let X,Y∈W1,p(Rd)∩L∞ with 2<p<∞, with bounded divergence. Then for all 0≤s,t≤T, there holds
Proposition 4.1. Let p∈[1,2). Then there exist divergence free vector fields X,Y∈Wloc1,p(R2) with [X,Y]=0 such that their regular Lagrangian flows do not commute.
Proof. The vector fields X,Y are constructed using a double covering of the plane. Consider the map Q:R2→R2 which doubles the angle. In polar coordinates,
Q(ρ,θ):=(ρ,2θ).
The map Q is clearly measure preserving for the Lebesgue measure, since it is in particular measure preserving on every circle {∣x∣=ρ}. Consider the vector fields e1,e2 in R2, constant and parallel to the coordinates axis, with unit norm. Clearly, e1,e2 are divergence free and [e1,e2]=0. We define the pullbacks of the vector fields e1,e2 by
X(x):=(DQ(x))−1e1,Y(x):=(DQ(x))−1e2.
Q is measure preserving and has Jacobian det(DQ(x))=2 for all x=0, therefore X,Y are divergence free. Moreover, the pullback of vector fields naturally preserves the Lie brackets, therefore
therefore X,Y∈Wloc1,p(R2) if and only if p∈[1,2). Consider a small neighborhood U of the point (−1,−1)∈Ry2 and consider trajectories starting from y∈U, obtained by first following e1 for time 2 and e2, also for time 2, and conversely. Then the corresponding pulled back trajectories for X,Y in the x-plane do not commute, because they belong to different branches of Q (see Figure 1). To see this, fix y0=(−1,−1) and choose any x0∈Q−1(y0). Consider the two polygonal paths in Ry2:
γ1:(−1,−1)→(1,−1)→(1,1),
γ2:(−1,−1)→(−1,1)→(1,1)
each one made of two segments. Since both paths avoid the origin, they can be uniquely lifted through Q, starting from x0. By construction, the endpoints of these lifts are respectively ΦY2∘ΦX2(x0) and ΦX2∘ΦY2(x0). The closed curve obtained by concatenating γ1 with the reverse of γ2 winds once around the origin. Since Q is a two-branch covering of R2, the two lifts have distinct endpoints. Hence
ΦY2∘ΦX2(x0)=ΦX2∘ΦY2(x0).
The same argument applies to all starting points y0 in a sufficiently small neighborhood of (−1,−1), since the corresponding rectangles still enclose the origin. Thus the two compositions differ on an open set of positive Lebesgue measure. □
[1]L. Ambrosio, Transport equation and Cauchy problem for BV vector fields, Invent. Math. 158 (2004), 227–260.DOI
[2]M. Colombo and R. Tione, On the commutativity of flows of rough vector fields, J. Math. Pures Appl. (9) 159 (2022), 294–312.arxiv.org/abs/2011.08133
[3]G. Crippa and C. De Lellis, Estimates and regularity results for the DiPerna–Lions flow, J. Reine Angew. Math. 616 (2008), 15–46.DOI
[4]R. J. DiPerna and P.-L. Lions, Ordinary differential equations, transport theory and Sobolev spaces, Invent. Math. 98 (1989), 511–547.DOI
[5]B. Fuglede, Conditions for two self-adjoint operators to commute or to satisfy the Weyl relation, Math. Scand. 51 (1982), 163–178.DOI
[6]J. Heinonen and S. Keith, Flat forms, bi-Lipschitz parametrizations, and smoothability of manifolds, Publ. Math. Inst. Hautes Études Sci. 113 (2011), 1–37.arxiv.org/abs/0909.3201
[7]E. Marconi, Differentiability properties of the flow of 2d autonomous vector fields, J. Differential Equations 301 (2021), 330–352.DOI
[8]E. Nelson, Analytic vectors, Ann. of Math. (2) 70 (1959), 572–615.DOI
[9]P.-E. Jabin, Critical non-Sobolev regularity for continuity equations with rough velocity fields, J. Differential Equations 260 (2016), no. 5, 4739–4757.DOI
[10]F. Rampazzo and H. J. Sussmann, Commutators of flow maps of nonsmooth vector fields, J. Differential Equations 232 (2007), 134–175.DOI
[11]A. Rebucci and M. Zizza, Geometric interpretation of the vanishing Lie Bracket for two-dimensional rough vector fields, J. Funct. Anal. 289 (2025), no. 1, 110919.DOI
[12]C. Rigoni, E. Stepanov, and D. Trevisan, Lie brackets of nonsmooth vector fields and commutation of their flows, J. Lond. Math. Soc. (2) 106 (2022), 1232–1256.arxiv.org/abs/2011.07351
[13]E. M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton Mathematical Series, vol. 30, Princeton University Press, Princeton, NJ, 1970.