Birkhoff Billiards Dynamics

• Birkhoff billiards are dynamical systems in strictly convex domains where a point particle moves in straight lines and reflects off boundaries with equal angles.
• They feature an exact symplectic twist map structure and a discrete variational framework that connects convex geometry, Aubry–Mather theory, and KAM methods.
• Rigidity results, including aspects of the Birkhoff–Poritsky conjecture, and periodic orbit analyses reveal deep links between integrability and geometric invariants.

Birkhoff billiards are the standard inner billiards in strictly convex domains: a point particle moves with constant speed along straight lines in the interior and reflects elastically at the boundary, so that the angle of incidence equals the angle of reflection. In the planar case this produces an exact symplectic twist map of a phase cylinder, and the subject has become a meeting point of convex geometry, variational calculus, Aubry–Mather theory, KAM theory, spectral asymptotics, and rigidity questions such as the Birkhoff–Poritsky conjecture (Bialy et al., 4 Oct 2025).

1. Geometric model and phase-space descriptions

In the classical planar setting, one starts with a bounded planar domain $\Omega\subset\mathbb{R}^2$ whose boundary is sufficiently smooth; several works assume $C^2$ or $C^3$ regularity, strict convexity, and positive curvature. The billiard trajectory is a polygonal line in $\Omega$, and the reflection law is specular. A standard parametrization of phase space uses arc-length $t\in S^1\cong \mathbb{R}/l\mathbb{Z}$ on $\partial\Omega$ and an angle $\theta\in(0,\pi)$ between the outgoing ray and the positively oriented tangent, giving the open annulus $U=S^1\times(0,\pi)$. The billiard map $T:U\to U$ sends one collision to the next and preserves the smooth measure $\mu=\sin\theta\,d\theta\wedge dt$ (Merenkov et al., 2011).

Equivalent coordinates are frequently used. In the phase cylinder $C^2$0, the natural symplectic $C^2$1-form is $C^2$2, with $C^2$3. In oriented-line coordinates $C^2$4, where $C^2$5 is the signed distance from a fixed origin and $C^2$6 is the angle of the right unit normal, the phase cylinder is symplectomorphic to a subset of $C^2$7 with $C^2$8-form $C^2$9 and symplectic form $C^3$0 (Bialy et al., 2020).

The billiard map also admits the standard twist-map formulation. If $C^3$1 is the arc-length parametrization of $C^3$2 and $C^3$3 is the chord length, then with $C^3$4 and $C^3$5 the lift of the billiard map is an exact area-preserving twist map with generating function $C^3$6. This places Birkhoff billiards directly inside the general theory of exact symplectic twist maps (Sorrentino, 2013).

2. Variational structure and exact symplectic twist dynamics

A basic organizing principle is that billiard trajectories are stationary sequences for a discrete length functional. If $C^3$7 records consecutive impact parameters on the boundary, then

$C^3$8

Critical points of this functional are precisely billiard orbits, and the corresponding billiard map is an exact symplectic twist map of the cylinder (Bialy et al., 4 Oct 2025).

In higher-dimensional convex billiards, the same variational viewpoint persists but admits two natural generating functions. For a strictly convex $C^3$9-smooth hypersurface $\Omega$0, the billiard map on the space of oriented lines has the classical chord-length generating function

$\Omega$1

and also a support-type generating function

$\Omega$2

where $\Omega$3 is the Gauss map and $\Omega$4. The associated notion of locally maximizing orbit is defined by negativity of the second variation on every finite segment. A geometric criterion then shows that the two generating functions determine the same class $\Omega$5 of locally maximizing orbits for Birkhoff billiards, and the proof uses the Sinai–Chernov formula from geometric optics and billiard dynamics (Bialy et al., 2023).

This variational formalism is not merely structural. It underlies Hopf-type rigidity arguments, Aubry–Mather theory, the study of maximizing and minimax periodic configurations, and the extraction of dynamical invariants such as Mather’s $\Omega$6-function and marked length spectra.

3. Periodic orbits, caustics, and integrability

A phase point $\Omega$7 is periodic of period $\Omega$8 if $\Omega$9 and $t\in S^1\cong \mathbb{R}/l\mathbb{Z}$0 for $t\in S^1\cong \mathbb{R}/l\mathbb{Z}$1. Geometrically, the corresponding billiard trajectory closes after $t\in S^1\cong \mathbb{R}/l\mathbb{Z}$2 reflections and forms an inscribed $t\in S^1\cong \mathbb{R}/l\mathbb{Z}$3-gon. The set of three-periodic points is

$t\in S^1\cong \mathbb{R}/l\mathbb{Z}$4

and three-periodic trajectories are the first non-trivial odd-period family beyond the simple 2-periodic bouncing-ball trajectories (Merenkov et al., 2011).

Invariant curves in phase space correspond to caustics in configuration space. A convex caustic is a strictly convex curve $t\in S^1\cong \mathbb{R}/l\mathbb{Z}$5 such that any billiard segment tangent to $t\in S^1\cong \mathbb{R}/l\mathbb{Z}$6 remains tangent after every reflection. In the language of twist maps, a rotational invariant curve is a simple closed invariant curve winding once around the cylinder; irrational rotation number gives quasi-periodic dynamics, while rational rotation number yields periodic orbits (Bialy et al., 2020).

The two basic integrable models are the circle and the ellipse. For the circular billiard, the phase space is globally foliated by invariant curves, and the angle of reflection is an integral of motion. For the elliptic billiard, confocal conics provide a continuous family of caustics, so the billiard is locally integrable near the boundary and, more broadly, integrable in the classical confocal sense (Sorrentino, 2013). The survey literature distinguishes several notions of integrability: total integrability by foliation with rotational invariant curves, local integrability near the boundary, and algebraic or polynomial integrability by existence of a first integral polynomial in velocities (Bialy et al., 4 Oct 2025).

4. Rigidity and the Birkhoff–Poritsky program

The central rigidity question is the Birkhoff–Poritsky conjecture: under natural integrability assumptions, are ellipses the only integrable convex planar billiard tables? Several precise variants are known.

Before the main rigidity statements, it is useful to separate the principal notions of “integrable” used in the literature. Total integrability means that an open region of the phase cylinder is foliated by rotational invariant curves. Polynomial integrability means the billiard flow admits a non-constant first integral polynomial in the velocities. These notions are distinct, and different papers solve rigidity problems under different hypotheses (Bialy et al., 4 Oct 2025).

Setting	Hypothesis	Consequence
Strictly convex planar billiard	Whole phase cylinder foliated by rotational invariant curves	Boundary is a circle
Centrally symmetric planar billiard	Region between a $t\in S^1\cong \mathbb{R}/l\mathbb{Z}$7-periodic invariant curve and the boundary is foliated by invariant curves	Boundary is an ellipse
Bounded convex planar billiard with smooth boundary	Existence of a non-constant polynomial first integral in velocities	Boundary is an ellipse

The global “circle” rigidity follows from the Hopf-type theorem for convex billiards: if the Birkhoff billiard is totally integrable on the entire phase cylinder, then the table is circular (Bialy et al., 4 Oct 2025). A more refined elliptic rigidity result holds in the centrally symmetric case: if the billiard ball map has a continuous rotational invariant curve of rotation number $t\in S^1\cong \mathbb{R}/l\mathbb{Z}$8 consisting of $t\in S^1\cong \mathbb{R}/l\mathbb{Z}$9-periodic orbits, and the region between that curve and the boundary is foliated by continuous rotational invariant curves, then the boundary is an ellipse; the same conclusion holds if all orbits in that region are locally maximizing for the length functional (Bialy et al., 2020).

The polynomial or algebraic version of the Birkhoff conjecture is fully resolved. Every bounded polynomially integrable planar billiard with $\partial\Omega$0-smooth connected boundary is an ellipse, and on the plane, sphere, and hyperbolic plane a piecewise-smooth polynomially integrable billiard is classified by countably confocal boundaries: arcs of conics from one confocal pencil together with appropriate admissible geodesic segments (Glutsyuk, 2017). On sphere and hyperbolic plane, the dual-curve approach produces a sharp obstruction: if the dual curve $\partial\Omega$1 of the boundary admits a polynomial integral, then either $\partial\Omega$2, or all singular and inflection points of $\partial\Omega$3 lie on the absolute conic; in particular, smooth dual curves of degree $\partial\Omega$4 are excluded (Michael et al., 2016).

A common misconception is that these rigidity statements exhaust all integrable billiards. They do not. They apply to smooth strictly convex planar tables, or to their constant-curvature analogues, under explicit regularity and compactness assumptions. Once those assumptions are relaxed, different phenomena appear.

5. Periodic-orbit geometry, quantitative invariants, and local normal forms

Periodic sets can be quantitatively very small even when they are dynamically significant. For classical planar billiards with $\partial\Omega$5 boundary, the set $\partial\Omega$6 of three-periodic phase points satisfies

$\partial\Omega$7

This is strictly stronger than zero Lebesgue measure. Moreover, if $\partial\Omega$8, then $\partial\Omega$9 has a tangent line at $\theta\in(0,\pi)$0-almost every point, so a maximal three-periodic set is curve-like rather than a genuinely two-dimensional fractal (Merenkov et al., 2011).

A complementary set of invariants comes from Mather’s $\theta\in(0,\pi)$1-function. For a rational rotation number $\theta\in(0,\pi)$2,

$\theta\in(0,\pi)$3

where $\theta\in(0,\pi)$4 is the maximal marked length spectrum, i.e. the maximal perimeter among periodic orbits of rotation number $\theta\in(0,\pi)$5. Near $\theta\in(0,\pi)$6, the Taylor coefficients of $\theta\in(0,\pi)$7 can be written explicitly in terms of curvature integrals $\theta\in(0,\pi)$8, and these formulas connect the billiard map to Marvizi–Melrose spectral invariants and to questions of length-spectrum rigidity (Sorrentino, 2013).

These expansions are also conjugacy invariants in a non-integrable regime. In the class of strictly convex smooth boundaries, each of which does not have a strip around its boundary foliated by invariant curves, the Taylor coefficients of the normalized Mather $\theta\in(0,\pi)$9-function are invariant under $U=S^1\times(0,\pi)$0-conjugacies. By contrast, any two elliptic billiard maps are $U=S^1\times(0,\pi)$1-conjugated near their respective boundaries, and are $U=S^1\times(0,\pi)$2-conjugated on an open cylinder near the boundary and away from a line through the center; if the billiard maps corresponding to two ellipses are topologically conjugated, then the ellipses are similar (Kaloshin et al., 2021).

Near the boundary, Lazutkin normal form provides a different family of formal invariants. For a smooth strictly convex table, the Lazutkin conjugacy has an expansion

$U=S^1\times(0,\pi)$3

A recent characterization states that a smooth domain is a disc if and only if $U=S^1\times(0,\pi)$4. Equivalently, the first nontrivial correction in the longitudinal coordinate of the formal Lazutkin conjugacy vanishes exactly for circles (Czudek et al., 17 Mar 2026).

Local dynamics around elliptic periodic points is controlled by Birkhoff normal form. For a variety of billiard tables, one can construct the Birkhoff transformation explicitly and compute the first two twist coefficients $U=S^1\times(0,\pi)$5 in terms of the local geometry of the boundary, including the radius of curvature and its derivatives. These coefficients characterize nonlinear stability and local analytic integrability around elliptic periodic orbits; resonant obstructions appear when the required normalization fails at low order (Jin et al., 2021).

6. Higher dimensions, non-Euclidean settings, and related billiard models

The term “Birkhoff billiard” is also used in higher-dimensional convex geometry. For a closed, $U=S^1\times(0,\pi)$6-smooth, strictly convex hypersurface $U=S^1\times(0,\pi)$7, the billiard map acts on the space of oriented lines intersecting $U=S^1\times(0,\pi)$8, which can be identified either with a ball bundle in $U=S^1\times(0,\pi)$9 or with a ball bundle over $T:U\to U$0. In this setting the billiard map remains an exact symplectic twist map, and the class $T:U\to U$1 of locally maximizing orbits admits both a variational characterization via Jacobi fields and a geometric characterization via invariant Lagrangian subbundles (Bialy et al., 2023).

The constant-curvature generalization replaces straight lines by geodesics on the sphere or hyperbolic plane and keeps the same reflection law. Much of the algebraic Birkhoff program extends to these geometries, with projective duality, the absolute conic, and Hessian identities replacing some of the Euclidean tools (Michael et al., 2016). A broader survey places Birkhoff billiards alongside outer, magnetic, Minkowski, wire, and cone billiards, emphasizing that many methods—twist maps, generating functions, total integrability, and Mather $T:U\to U$2-function inequalities—persist across these models (Bialy et al., 4 Oct 2025).

A particularly important qualification to planar rigidity comes from cones. For the billiard inside a convex $T:U\to U$3 cone in $T:U\to U$4, there is always a quadratic first integral

$T:U\to U$5

the spheres centered at the vertex are caustics, every trajectory has a finite number of reflections, and the billiard is integrable. This is described as the first example of an integrable billiard where the billiard table is neither a quadric nor composed of pieces of quadrics. A plausible implication is that smoothness, compactness, and absence of singularities are essential in the classical planar Birkhoff conjecture (Mironov et al., 22 Jan 2025).

Related nonstandard regimes also illuminate the classical model by contrast. In inverse magnetic billiards, the magnetic field is zero inside a strictly convex domain and constant outside; the return map is area-preserving, admits a generating function, and in the high-field limit converges to the standard Birkhoff billiard, while in intermediate and weak-field regimes non-twist behavior and discontinuities can appear (Gasiorek, 2019). In dissipative symplectic billiards, the associated billiard map is conformally symplectic and admits a compact invariant set called the Birkhoff attractor: for strong dissipation this attractor is a normally contracted graph over the zero section, whereas for mild dissipation in a centrally symmetric domain it is an indecomposable continuum with positive topological entropy (Baracco et al., 16 Sep 2025).

Taken together, these results show that Birkhoff billiards are not a single theorem or conjecture but a large geometric-dynamical framework. The classical inner billiard in a smooth strictly convex planar domain remains the core model; around it lies a hierarchy of rigidity statements, periodic-orbit estimates, variational invariants, local normal forms, and extensions to constant curvature, higher dimensions, dissipative dynamics, and singular geometries.