Let $\dot{x} = \Pi \cdot \nabla H$ be a smooth Hamiltonian-Poisson system on $\mathbb{R}^n$.

$H: \mathbb{R}^n \to \mathbb{R}$ is the Hamiltonian and $\Pi = (\Pi^{ij})$ is a skew-symmetric matrix of functions $\mathbb{R}^n \to \mathbb{R}$ satisfying the Jacobi identity $\Pi^{i\ell}\partial_\ell\Pi^{jk} + \Pi^{j\ell}\partial_\ell\Pi^{ki} + \Pi^{k\ell}\partial_\ell\Pi^{ij} = 0$ for all $1\leq i<j<k \leq n$.

(When $\Pi = \begin{pmatrix}0 & I\\-I& 0\end{pmatrix}$ the system is an ordinary Hamiltonian system.)

If $\mathcal{O}$ is an orbit, then:

  • $\mathcal{O}$ is a smooth curve and hence has Hausdorff dimension $\dim_\text{H}(\mathcal{O})=1$.

  • $\mathcal{O}$ is contained in a $2m$-dimensional symplectic leaf $\mathcal{M}\subset\mathbb{R}^n$ of the Poisson structure $\Pi$.

The leaf $\mathcal{M}$ is an immersed submanifold and I guess its Hausdorff dimension is $\dim_\text{H}(\mathcal{M}) = 2m$.

My question is about closures.

For example $\mathcal{O}$ could be a dense orbit in a torus $\mathcal{M}$, so $\dim_\text{H}(\bar{\mathcal{O}}) > \dim_\text{H}(\mathcal{O})$ is possible. But:

  • Can $\dim_\text{H}(\bar{\mathcal{M}}) > \dim_\text{H}(\mathcal{M})$?
  • Can $\dim_\text{H}(\bar{\mathcal{O}}) > \dim_\text{H}(\bar{\mathcal{M}})$? Update: no, because $\bar{\mathcal{O}} \subset \bar{\mathcal{M}}$ and $\dim_\text{H}$ is monotonic.
  • Can $\dim_\text{H}(\bar{\mathcal{O}}) > \dim_\text{H}(\mathcal{M})$?

I am most interested in the case $n=3$ and $m=1$: then $\mathcal{O}$ is contained in a surface $\mathcal{M}$ in $\mathbb{R}^3$.


This question has an open bounty worth +50 reputation from Ricardo Buring ending in 4 days.

This question has not received enough attention.


Since $\mathcal{O} \subset \mathcal{M}$ implies $\overline{\mathcal{O}} \subset \overline{\mathcal{M}}$ and since the Hausdorff dimension is monotone under inclusion, we always have $\mathrm{dim}_H \, \overline{\mathcal{O}} \le \mathrm{dim}_H \, \overline{\mathcal{M}}$. Consequently, $\mathrm{dim}_H \, \overline{\mathcal{O}} > \mathrm{dim}_H \, \mathcal{M}$ implies $\mathrm{dim}_H \, \overline{\mathcal{M}} > \mathrm{dim}_H \, \mathcal{M}$.

Let's assume that any of these two inequalities is established for some Poisson structure $\Pi$ on $\mathbb{R}^3$ (and some Hamiltonian function $H$). Then for $d \ge 4$, considering $\mathbb{R}^{d} = \mathbb{R}^3 \times \mathbb{R}^{d-3}$ equipped with the Poisson structure $\Pi \oplus 0$ (and the pullback of $H$ under the projection onto the first factor) gives a new example of any of these two inequalities. It thus suffices to consider the case $d=3$.

I sketch below how to produce examples where one or the two inequalities are satisfied. I will first describe a reformulation of the Poisson structure which, I think, is helpful in achieving a better geometrical understanding of what are Poisson structures in three dimensions. Subsequently, after a preliminary discussion, I will proceed to construct the relevant examples.

Let's write a general Poisson structure as $\Pi = \frac{1}{2} \sum_{i,j=1}^3 \Pi^{ij} \, \partial_i \wedge \partial_j$ with $\Pi_{ij} = - \Pi_{ji}$. By a slight abuse of notation (due to the possibility to identify vectors and forms via the Euclidean metric), we have the Hodge star operator between bivectors and 1-forms, $\Pi \leftrightarrow V$, given by $\partial_i \wedge \partial_j \leftrightarrow dx^k$ for $(i,j,k)$ a cyclic permutation of $(1,2,3)$. Using the abstract index notation, in particular the Levi-Civita symbol $\epsilon^{ijk}$, the star operation amounts to the following: $$ \Pi = \star V \; \Leftrightarrow \; \Pi^{ij} = \epsilon^{ijk}V_k \; \Leftrightarrow \; V_i = \frac{1}{2} \epsilon_{ijk}\Pi^{jk} \, . $$ The Jacobi identity for $\Pi$ turns out to be equivalent to the identity $V \wedge dV = 0$. We readily see that examples of Poisson structures are given by closed/exact one-forms $V$ ; this yields a relatively easy way to construct smooth Poisson structures.

Thinking of $V$ as a vector field rather than as a differential 1-form, the Hamiltonian vector field associated to a Hamiltonian function $H$ is $X_H = V \times \nabla H$, where $\times$ denotes the usual cross product on $\mathbb{R}^3$. The Poisson bracket is thus $\{H,G\} = (V \times \nabla H) \cdot \nabla G = V \cdot (\nabla H \times \nabla G)$. Since the symplectic distribution of $\Pi$ is spanned by the Hamiltonian vector fields, we deduce that $V$ is perpendicular to the symplectic distribution; thinking back of $V$ as a 1-form, the equation $V \wedge dV = 0$ amounts to the Frobenius integrability of the symplectic distribution $\mathrm{Ker} V$.

Before going into the description of some relevant examples to the question, I would like to address some expectations one might have about such examples.

You mentioned the way an orbit could densely fill a torus. Similarly, one could expect to find a 2-dimensional leaf $\mathcal{M}$ winding in $\mathbb{R}^3$ in such a way as to produce something reminiscent of a "mille-feuille" with smooth local plaques which densily fill a given open set of 3-space. However, since such a leaf would have codimension 1 and since $\mathbb{R}^3$ has trivial topology, a Poincaré-Bendixson-like phenomenon could force a leaf to be either closed or to be spiralling outward or inward, thereby preventing the formation of a "dense mille-feuille". This suggests that we stand a better chance of finding a leaf $\mathcal{M}$ whose closure has Hausdorff dimension strictly larger than $2$ among those (nonclosed) immersed 2-manifolds which are, so to speak, "attracted by" sets of Hausdorff dimension strictly larger than $2$ not containing the 2-manifolds.

The general scheme of the forthcoming construction goes as follow.

1) Find a 2-manifold (resp. a 1-manifold) which "is attracted by" a set of Hausdorff dimension strictly larger than $2$.

2) Come up with an exact Poisson structure (and a Hamiltonian function) for which the 2-manifold (resp. 1-manifold) is a leaf (resp. an orbit).

The first step will be done by hand. The second step will rely on Whitney's extension theorem. Given en embedded submanifold $X$, we prescribe the "restriction of a smooth function $f$ to $\overline{X}$ up to first order": we require the function to take the value $0$ on $\overline{X}$ and we find a smooth vector field on $\overline{X}$ which either vanish or is perpendicular (where this is meaningful) to the manifold $X$ (so that the vector field stands a chance to be the restriction to $X$ of the gradient vector field of a function). Then Whitney's extension theorem states that these data are indeed restrictions of a globally defined smooth functions.

Proof of $\mathrm{dim}_H \, \overline{\mathcal{M}} > \mathrm{dim}_H \, \mathcal{M}$.

I will first described the picture in a "transverse slice" to $\mathcal{M}$: instead of considering a 2-manifold inside 3-space directly, I will rather consider a 1-manifold $C$ inside the plane (the "slice picture") and then take the product of both with an interval $I \subset \mathbb{R}$, so that $\mathcal{M} := C \times I$. Consider any Jordan curve $K : S^1 \to \mathbb{R}^2$ of Hausdorff dimension strictly greater than $1$, for instance some Osgood curve. (I shall slightly abuse notations by writing also $K$ to denote the image curve.) By the Jordan-Schönflies theorem, $K$ bounds a disc $D$. Carathéodory's mapping theorem states that the closure of this disc is homeomorphic to the standard closed disc $\bar{B}$ via a map $\psi : \bar{B} \to \bar{D}$ which is biholomorphic (in particular diffeomorphic) between the interiors. In $\bar{B}$, consider an outward spiralling embedded smooth curve $C'$ contained in the interior $B$ whose closure is $C' \cup \partial \bar{B}$. The restriction of $\psi$ to this curve is a smooth embedded curve $C$ which is disjoint from the Jordan curve $K$ and whose closure is $C \cup K$. It follows that $\mathrm{dim}_H \, \overline{C} > 1$. Since $\mathrm{dim}_H(X \times Y) \ge \mathrm{dim}_H(X) + \mathrm{dim}_H(Y)$, it readily follows that $\mathrm{dim}_J(\overline{\mathcal{M}}) > 2$.

We now aim to prove that the "spiralling sheet" $\mathcal{M}$ is a symplectic leaf for some Poisson structure $V$ on $\mathbb{R}^2 \times I$. Note that the corresponding vector field $V$ need to be perpendicular to $\mathcal{M}$, hence tangent to the fibers of the projection $\mathbb{R}^2 \times I \to I$. Thus, taking pullback under the projection $\mathbb{R}^2 \times I \to \mathbb{R}^2$, it is necessary and it suffices to prove there exists a smooth exact 1-form $V = df$ on $\mathbb{R}^2$, where $f$ is a smooth function such that $C$ is contained in a level-set of $f$ and, moreover, $C$ consists only in regular points of this function. We construct $f$ as follows. Set $f=0$ on $\bar{C} = C \cup K$; that takes care of the level-set part. Notice that $C$ is arclength parametrised by a regular function $\gamma : \mathbb{R} \to C$; rotating the velocity $d\gamma/dt$ ninety degrees clockwise, we obtain a vector field $N$ along $C$ which is perpendicular to the curve. Since there exists a smooth function $g : \mathbb{R}^2 \to [0, \infty)$ which is equals $0$ precisely on $K$, we can consider the smooth vector field $gN : C \cup K \to \mathbb{R}^2$ and notice that it vanishes only on $K$. Now, the data $f$ and $gN$ on $C \cup K = \overline{C}$ satisfy the conditions of Whitney's extension theorem; consequently, there exists a smooth function $f : \mathrm{R}^2 \to \mathbb{R}$ which is equal to $0$ on $C \cup K$ and which satisfies $\nabla f = gN$ on $C \cup K$. We can take $V = df$.

Proof of $\mathrm{dim}_H \, \overline{\mathcal{O}} > \mathrm{dim}_H \, \mathcal{M}$.

This is just a sketch. It certainly suffices to find a Hamiltonian function $H$ whose associate Hamiltonian vector field $V \times \nabla H$ would admit an orbit which admits $K \times I$ in its closure. Notice that $\mathcal{M}$ is the image of the embedding $\Gamma : \mathbb{R} \times I \to \mathcal{M} : (t, z) \mapsto (\gamma(t), z)$. The idea is quite similar to the previous one : we want to find a curve $G' : \mathbb{R} \to \mathbb{R} \times I : t \mapsto (t, G(t))$ which will be such that the composition $\gamma' : \Gamma \circ G' : \mathbb{R} \to \mathcal{M}$ will admit $K \times I$ in its closure. For instance, $G$ could be oscillating with an ever growing frequency as $t \to +\infty$ ; allowing for a more general $\gamma'(t) = \Gamma(F(t), G(t))$, we could choose $(F(t), G(t))$ to produce a "cartoon" of the successive (smoothen) approximations of the Hilbert square-filling curve. Afterwards, we want to find a Hamiltonian function $H$ which will admit $\gamma'$ as an orbit. Since $H$ is constant under the Hamiltonian flow, we want to find $H$ which is constant along $\gamma'$ and such that $\nabla H$ is perpendicular to both $V$ and $d(\gamma')/dt$ along $\gamma'$, for instance take $g.(V \times d(\gamma')/dt)$. We can then apply once again Whitney's extension theorem.


The above examples yield relatively explicit examples of $\mathcal{M}$ and $\mathcal{O}$, but they are somewhat artificial. Alternatively, one could start with any 1-dimensional embedded smooth trajectory $\gamma$ in $\mathbb{R}^3$ whose limit set is a fractal of Hausdorff dimension greater than 2, for instance a particular trajectory to the Lorenz system (with appropriate values to the parameters for the presence of a strange attractor). Note there exists along $\gamma$ an orthogonal frame of vector fields $\{d\gamma/dt, N, P\}$: the image of the Gauss map $\mathbb{R} \to S^2 : t \mapsto d\gamma/dt$, being $C^1$, has zero measure by Sard's theorem, so there are constant vectors $A, B$ such that $\{d\gamma/dt, A, B\}$ is linearly independent for all $t$ and we can set $N$ and $P$ to be proportional to $d\gamma/dt \times A$ and $d\gamma/dt \times B$. Multiplying $N, P$ by a smooth function which vanishes on the limit set of $\gamma$, we can use Whitney's theorem as above to get two functions: one is the Hamiltonian, the other is a potential for the Poisson structure. The drawback of this method is that we don't know what $\mathcal{M}$ looks like, but as it contains an orbit whose closure has Hausdorff dimension greater than $2$, its own closure also has Hausdorff dimension greater than $2$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.