Ivrii conjecture on periodic billiard orbits

Ivrii conjecture on periodic billiard orbits

Definitions and mathematical statement

For a bounded connected smooth domain $\Omega\subset{\mathbb{R}}^n$, $n\ge2$, let $M=B^*\partial\Omega=\{(x,\eta):|\eta|<1\}$ be its open boundary cotangent unit ball bundle. Give it canonical Liouville measure $\mu=|\omega^{n-1}|/(n-1)!$. The billiard map $\beta$ follows the inward unit covector to its next boundary collision, reflects specularly, and records the tangential covector. Let $M_{\rm reg}$ consist of states with all positive and negative iterates well-defined and nongrazing. Ivrii's assertion is \[ \mu\!\left(\bigcup_{k\ge2}\{z\in M_{\rm reg}:\beta^kz=z\}\right)=0. \] Periods need not be least periods. Neither convexity nor simple connectivity is assumed, and singular trajectories are excluded rather than continued by an invented rule. The claim is zero phase-space measure, not finiteness or absence of periodic trajectories.

Formulation and concept references: [S1].

Short English statement

In every bounded smooth Euclidean billiard domain, do regular periodic trajectories occupy zero Liouville measure in the boundary phase space?

Research attempt

The measure and the geometric category.

The primary source [S1] studies a particular analytic planar classification and its consequences for four reflections. Its complexification and fixed-period results are not a proof for all periods, all smooth boundaries or all dimensions. The catalogue selects ordinary Euclidean specular billiards and the full regular phase space. An example in a projective, outer or complex billiard category would therefore not settle this formulation.

We first prove the two-collision assertion for every domain in the question and the complete assertion for balls. We then give a precise analytic reduction and show why smoothness and symplecticity alone do not close its missing step. In local boundary coordinates $(s,\eta)$ the Liouville measure is the ordinary symplectic volume $ds\,d\eta$; smooth changes of boundary coordinates preserve its measure class. Thus positive-dimensional families of periodic trajectories can still have zero measure in the full $2(n-1)$-dimensional phase space.

A finite-period reduction with the singular set kept out.

For $k\geq2$, let $U_k$ be the open set of states for which the first $k$ forward collisions are well-defined, have distinct consecutive endpoints and are nongrazing. Smoothness of the boundary and the implicit function theorem show that $\beta^k$ is smooth on this open set, locally in a fixed itinerary of boundary charts. Indeed the collision time is the first positive root of a boundary defining function along a straight line; at a transverse hit its time derivative is nonzero. Specular reflection is a smooth formula in the incoming direction and normal. Repeating this argument gives the claimed local regularity.

The periodic states in the question are contained in \[ \bigcup_{k\geq2}\{z\in U_k:\beta^kz=z\}. \tag{379.1} \] Thus it suffices to prove nullity for each fixed $k$. This is a countable union, so no estimate uniform in $k$ is required merely for the measure-zero conclusion. Conversely, checking finitely many periods cannot establish $379.1$. We do not assign a continuation through tangency or a singular reflection in order to define an iterate outside these regular sets.

All two-periodic states lie in a null section.

Consider a regular orbit with two collisions at $x$ and $y$. The outgoing velocity at $x$ is $(y-x)/|y-x|$, while the incoming velocity there, after the return segment, is its negative. Let $n_x$ be the outward normal. Specular reflection preserves the tangential component of velocity. A vector and its negative have the same tangential component only if that component is zero. Thus the tangential covector at $x$ is $\eta=0$; the same is true at $y$.

The zero section $\{(x,0):x\in\partial\Omega\}$ has dimension $n-1$ inside a $2(n-1)$-dimensional manifold. Since $n\geq2$, its Liouville measure is zero: in each local chart it is a product of a boundary coordinate domain with a singleton in positive-dimensional covector space, and Fubini applies. A countable chart cover proves \[ \mu\{z\in M_{\rm reg}:\beta^2z=z\}=0. \tag{379.2} \] The proof neither assumes convexity nor tries to count the normal chords. There can be infinitely many such chords or a continuous family, as in a ball, without violating $379.2$.

A complete model: the ball in every dimension.

By translation and scaling it suffices to take the unit ball. Write $x$ for a boundary point and $v$ for the inward unit velocity. The bivector $x\wedge v$ is conserved during a straight flight, since $(x+tv)\wedge v=x\wedge v$. It is also unchanged at reflection: the new velocity differs from the old one by a scalar multiple of the radial normal $x$.

If $x\wedge v=0$, the orbit follows a diameter and has period two. Otherwise the two-dimensional plane spanned by $x,v$ is invariant under both the flights and reflections. To verify this directly, the next collision point lies in that plane and the reflected velocity is a linear combination of the incoming velocity and that collision point. The entire orbit is therefore a circular billiard inside this fixed plane.

Let $\rho=|\eta|\in[0,1)$ be the magnitude of the boundary tangential velocity. It is conserved because it equals $|x\wedge v|$ on the unit sphere. In the invariant circle, consecutive collision points have central angle \[ \theta=2\arccos\rho\in(0,\pi], \tag{379.3} \] with a fixed orientation when $\rho>0$. For example $\rho=0$ gives the diameter angle $\pi$, while grazing would correspond to $\rho=1$ and is excluded. The identity follows by resolving the unit chord direction into its normal and tangential components at one endpoint.

The state is periodic exactly when $\theta/(2\pi)$ is rational. If $k\theta=2\pi m$, the collision point and its incoming and outgoing directions repeat; conversely repetition of the state forces this equality in its invariant circle. Thus all periodic states lie in the union of fiber-radius levels \[ \bigcup_{\substack{k\geq2,\ 1\leq m\leq k/2}} \{(x,\eta):|\eta|=\cos(\pi m/k)\}. \tag{379.4} \] Every level has zero Lebesgue measure in the $(n-1)$-dimensional covector fiber. For a positive radius it is a sphere of one lower dimension; for radius zero it is a singleton. Fubini and the countable union show that the full periodic set has Liouville measure zero. This proves the selected assertion for every ball, in every $n\geq2$.

There are nevertheless many periodic trajectories in a ball. In dimension two, rational angles are dense among all allowed angles, so periodic states are dense. Density, infinitude and positive measure are different properties. This model explicitly prevents an argument that mistakes a dense periodic family for a counterexample.

An analytic zero-set lemma and its precise use.

Let $F$ be a nonzero real-analytic scalar function on a connected open subset of $\mathbb R^d$. Then its zero set has Lebesgue measure zero. One proof keeps track of derivatives. At any zero point, not every derivative can vanish, since otherwise the convergent Taylor series and analytic continuation would make $F$ identically zero. Choose a nonzero derivative of least total order $r\geq1$. A derivative of order $r-1$ then vanishes at the point and has nonzero gradient there. Its zero set is locally a smooth hypersurface by the implicit function theorem. Stratifying according to the first nonzero derivative covers the original zero set by countably many such local hypersurfaces. Each is null, proving the assertion.

For a real-analytic billiard boundary, the collision-time argument above is analytic on every regular itinerary chart. Hence $\beta^k-\operatorname{id}$ is an analytic vector-valued map in common coordinates near a possible fixed point. If one component is not identically zero on a connected chart, its zero set, and therefore the fixed-point set, is null there. This yields a conditional reduction: \[ \begin{gathered} \text{If no regular analytic itinerary component has } \beta^k=\operatorname{id}\text{ on an open set,}\\ \text{then the regular }k\text{-periodic set is null.} \end{gathered} \tag{379.5} \] One may use a countable coordinate cover, since the phase space is second countable. Components without fixed points need no analysis; near a fixed point the source and target of the iterate can be put in the same coordinate chart.

The hypothesis excluding an identity germ is substantial. It is not a consequence of the abstract fact that the billiard map is analytic. Proving it for every period and every allowed table is precisely a geometric issue left by this reduction. The source's discussion of reflective analytic billiards makes this obstruction visible rather than eliminating it for arbitrary periods.

Why a smooth zero-set argument is insufficient.

A smooth function can vanish on a positive-measure set with empty interior. To see this concretely, let $C$ be a closed nowhere-dense subset of $[0,1]$ of positive measure, constructed by deleting a countable collection of intervals of total length less than one. Write the complementary intervals as $I_j=(a_j,b_j)$. On each interval choose a nonnegative smooth bump $\psi_j$, positive inside and flat at its endpoints. Multiply it by a coefficient so small that the supremum of its first $j$ derivatives is at most $2^{-j}$. The sum of these disjointly supported bumps, extended by zero on $C$, is smooth; alternatively the same bounds give convergence of every derivative of the series. It is positive off $C$ and zero on $C$. Thus the analytic implication in $379.5$ cannot be repeated using only $C^\infty$ regularity and empty interior.

The chosen coefficients can also be required to control derivatives near the finitely many early intervals, whose endpoint flatness already gives continuity there. Hence the construction does not depend on an unproved claim that an arbitrary sum of bumps is smooth. It is an obstruction to a proposed method, not a constructed Euclidean billiard with positive-measure periodic states.

Symplecticity does not exclude a positive fixed set.

On a planar symplectic coordinate chart, take a smooth radial Hamiltonian $H(q,p)=\phi(q^2+p^2)$ with $\phi$ constant for radii near zero and outside a larger disc, and nonconstant in between. Its Hamiltonian flow is the identity on the central disc and outside the larger disc, and rotates intermediate circles by a radius-dependent angle. Its time-one map preserves symplectic area and has an open set of fixed points while being nonidentity elsewhere.

Thus measure preservation, symplecticity and smooth invertibility do not by themselves imply the required nullity for a fixed-period set. The billiard map belongs to a much more constrained class of maps. Any proof based on fixed-point geometry must use that additional Euclidean reflection structure, rather than applying a false general principle to arbitrary symplectic maps.

There is a variational expression for the reflection law: for a closed polygon of collision points $(x_1,\ldots,x_k)$, define \[ \mathcal L=\sum_{i=1}^k|x_{i+1}-x_i|,\qquad x_{k+1}=x_1. \] Its tangential derivative at $x_i$ is the tangential projection of \[ \frac{x_i-x_{i-1}}{|x_i-x_{i-1}|} +\frac{x_i-x_{i+1}}{|x_i-x_{i+1}|}. \] Vanishing is the reflection law for a nondegenerate itinerary with the correct incoming and outgoing sides. However, a smooth function can have degenerate critical manifolds. Counting equations and variables in this length functional does not prove that the critical set projects to a null set of initial states. The normal-chord and ball calculations succeed because they identify extra equations explicitly, not because of a generic dimension count.

Diagnostics and the remaining obstruction.

The diagnostic iterates the exact circular collision-angle rule for selected rational angles and checks the first return, including the diameter case. It checks the angle formula by direct chord geometry as well. These are consistency tests for the ball proof, not evidence that finite-period enumeration exhausts a general table.

The proved results are nullity of the two-periodic set in every selected domain, the full ball subcase, and the conditional analytic reduction. In the analytic category the first gap is excluding identity germs of every iterate for Euclidean billiards. In the general smooth category even that exclusion alone does not replace a measure-zero argument. The full assertion in the printed scope remains unresolved by this attempt.

Sources

• S1 Alexey Glutsyuk, On 4-reflective complex analytic planar billiards, arXiv:1405.5990v4 (17 December 2015); standard real billiard-map conventions as in Hassell-Zelditch, arXiv:math/0211140, Section 3. https://arxiv.org/html/1405.5990v4. Formulation source.