1 Introduction
Let Q\subset\R^2 be a nondegenerate Euclidean triangle. Its unit-speed billiard flow \Phi_t moves in straight lines in the interior and reflects specularly at the open sides. Trajectories meeting a vertex in either time direction are discarded; they form a null set. The invariant probability measure on position and direction is normalized Liouville measure
We say that the flow is weakly mixing if its product with itself, (x,y)\mapsto(\Phi_t x,\Phi_t y), is ergodic for \mu\otimes\mu. Thus every measurable set invariant under all product time maps has measure zero or one.
Theorem 1.1. The unit-speed billiard flow in every nondegenerate Euclidean triangle with at least one angle irrational relative to \pi is weakly mixing for normalized Liouville measure.
In particular, for triangles of area one with angles \alpha,\beta,\pi-\alpha-\beta, the conclusion holds for Lebesgue-almost every point of
the excluded pairs with both \alpha/\pi and \beta/\pi rational form a countable set. The irrationality assumption is sharp for the full phase space. In a rational triangle, side reflections generate a finite group on directions. A nonconstant function on the circle invariant under this finite group is then an invariant function for the billiard flow, so the full flow is not ergodic.
Historical context. Polygonal billiards connect elementary reflection geometry with dynamics on flat surfaces. The unfolding construction of Zemlyakov and Katok [13] replaces reflections by straight continuation on a surface. For rational polygons, it produces a compact translation surface and separates the full phase space into invariant directional surfaces. Kerckhoff, Masur, and Smillie proved unique ergodicity in almost every direction [7]. They also obtained ergodicity of the full billiard flow for a dense G_\delta family of polygons. Vorobets gave a quantitative rational-approximation criterion and explicit ergodic examples, including irrational right triangles [10]. These results establish ergodicity under genericity or sufficiently rapid approximation assumptions rather than for each irrational triangle.
Chaika and Forni proved that the full billiard flow is weakly mixing for a dense G_\delta set of polygons with any fixed number of sides at least three [4]. Their argument uses the absence of common nontrivial eigenvalues for directional flows in almost every pair of directions on a translation surface. The Baire-category conclusion does not assert Lebesgue-almost-everywhere weak mixing in the space of triangle angles. For rational billiards, recent work of Arana-Herrera, Chaika, and Forni characterizes weak mixing in almost every direction [1]; this is a statement on directional invariant surfaces, rather than on the full Liouville phase space considered here.
Numerical investigations have long suggested weak mixing in irrational right triangles and mixing in triangles with all angles irrational [2, 3]. Other computations have suggested nonergodicity in some irrational right or symmetric triangles [11, 12]. These conflicting finite-time observations underscore the difficulty of identifying measurable invariants from trajectories with slow statistical convergence. Theorem 1.1 treats every irrational triangle without an approximation or symmetry restriction.
Analytic contribution. The regularity of an eigenfunction is the central issue. Glue two oppositely oriented copies of Q along their sides. The resulting flat double is regular across the open sides and has cone points at the vertices. On its unit tangent bundle, let X denote differentiation along a geodesic and Y differentiation in the perpendicular parallel direction. Locally these vector fields commute. Forni and Moll developed angular Fourier and Cauchy–Riemann methods in this setting [6, Sections 3.2–3.5]. Their rigidity results distinguish invariance under both horizontal directions from geodesic invariance under an additional horizontal Sobolev hypothesis [6, Theorem 4.1 and Corollary 4.2]. A measurable flow eigenfunction supplies only Xf=i\lambda f at the outset; the missing transverse derivative cannot be assumed.
The companion manuscript [8, Sections 2–6] removes this regularity obstacle for invariant functions, proving ergodicity of irrational triangular billiards. We extend its coefficient-energy and long-average method to every real spectral parameter \lambda. Summing the coefficient-energy identities over consecutive angular indices makes the terms containing \lambda telescope to a single boundary term. A separate pairing with the two boundary modes produced by differentiating an angular Fourier tail gives the sharp bound
Since Xf=i\lambda f, it follows that Yf=0. The use of a spectral parameter in Cauchy–Riemann energy estimates has a related precedent in the twisted cohomological equations studied by Forni [5, Section 3]. Here the argument starts with bounded measurable eigenfunctions and proves the needed derivative bounds. We give a self-contained proof of this analytic extension below.
Proof strategy. Expand a bounded eigenfunction on the double in angular Fourier modes. Spatial smoothing away from the cone points and cutoffs near the tips first give a uniform bound on the gradient of each coefficient. These individual bounds do not yet give a square-summable family of gradients. To obtain the missing summability, take the sum of all modes above a fixed index. Its derivative under X-i\lambda consists of two endpoint terms.
We test this equation using long time averages supported on orbit segments that stay clear of the cone points. There are two approximation parameters. At a fixed bound on the discarded measure, the smoothing limit gives a bounded solution of the exact local eigenfunction equation and passes the endpoint pairing to it. The individual-coefficient estimate then supplies a new derivative bound independent of the discarded measure. A second limit passes the pairing to the original eigenfunction. Summing the resulting endpoint inequalities proves Yf=0. This order of limits is what permits the initial eigenfunction to be merely measurable.
The remaining argument is geometric. On the triangle interior, Xf=i\lambda f and Yf=0 give a plane wave with a measurable direction-dependent amplitude. Reflection in the two sides through an irrational vertex makes this amplitude constant. Reflection in the third side excludes every nonzero frequency. The standard spectral criterion then gives weak mixing.
Section 2 sets up the double and its Fourier coefficients. Section 3 proves the coefficient and smoothing estimates. Section 4 establishes transverse rigidity by the two successive limits. Section 5 applies side reflections and proves ergodicity of the product flow.
2 The double and its angular coefficients
We replace reflection by straight continuation on a flat surface. This allows us to study the spatial derivatives of an eigenfunction without imposing boundary conditions on the sides of the triangle. Throughout Sections 2–4, the triangle is fixed and nondegenerate; its angles need not satisfy an irrationality assumption.
Glue two closed copies of the triangle Q along corresponding sides, give the copies opposite orientations, and remove the three vertices. The resulting connected oriented surface M has a flat metric. Across an open side, a flat chart is obtained by reflecting one copy across that side. Consequently, changes between oriented flat charts are rotations followed by translations. The metric completion \overline M is compact; its missing points form a set V of three cone points, with cone angles twice the corresponding angles of Q. This double and the coefficient framework below are also used in [8, Section 2].
Let SM be the unit tangent bundle, let \nu be normalized Liouville measure, and denote geodesic motion by \Psi_t. Thus, in a flat chart, \nu=d\bar A\,d\theta/(2\pi), where d\bar A is probability area measure on M. We first justify using this motion as a measure-preserving flow defined modulo a null set.
A finite-time obstruction to a geodesic can occur only at a cone point. Indeed, a unit-speed path is Cauchy as its time approaches a finite endpoint, and its limit in \overline M permits continuation if it is regular. There is also no unaccounted accumulation of side crossings at a cone point. Lift the polar angle in a punctured cone to \R and develop the lifted geodesic into the plane. Its image is a straight line. If the path approaches the cone tip, this line ends at the origin, so its final segment has constant polar angle and is radial. Apart from trajectories along a seam, which form a null set, that segment has no side crossings. On an earlier compact regular segment crossings cannot accumulate: a flat chart at a proposed accumulation point makes both the path and each local seam straight. Thus every vertex hit has a finite crossing itinerary. For each fixed finite itinerary, unfolding makes a vertex hit the condition that a straight trajectory pass through a fixed point. For almost every initial position this condition selects only finitely many directions, so it has zero Liouville measure. Taking the countable union over itineraries, vertices, and both time directions proves that almost every geodesic exists for all real times.
Local geodesic motion is translation at fixed angle in flat charts and preserves Liouville measure. Following a regular trajectory through finitely many charts proves that every time map preserves \nu, with inverse \Psi_{-t} modulo null sets. Folding the two sheets to Q, and folding their tangent vectors by the corresponding local isometries, defines a factor map to the billiard flow. It sends \nu to \mu, since the two sheets have equal area and the angular measure is preserved by reflection. In particular, billiard eigenfunctions lift to geodesic-flow eigenfunctions. Factor identities are understood almost everywhere for each fixed time.
The composition operators u\mapsto u\circ\Psi_t form a strongly continuous unitary group on L^2(SM,\nu). To see continuity at zero, first take a continuous compactly supported function on the regular bundle and use pointwise continuity of regular paths and dominated convergence. Density then gives the assertion for every L^2 function. The same assertion for the billiard flow follows through the isometric embedding given by the factor map.
In oriented flat coordinates x=(x_1,x_2), write v=(\cos\theta,\sin\theta), and set
The horizontal derivatives in the forward and perpendicular directions are
These vector fields are globally defined on SM, commute on the regular bundle, and are formally skew adjoint for \nu. All distributional equations in this paper use compactly supported smooth tests on M or SM. No test passes through a cone point.
We use angular Fourier coefficients, as in the flat-flow framework of Forni and Moll [6, Sections 3.2–3.5]. For j\in\Z, let P_j be fiberwise Fourier projection, written in a chart as
Under an oriented change of coordinates z'=e^{i\phi}z+c, \theta'=\theta+\phi, so
Thus a coefficient is a section of a flat unitary line bundle, and its two complex derivatives are coefficients of indices j+1 and j-1, respectively. Taking Fourier coefficients in (1) gives, also in distributions,
The transition factors of the two terms in each sum agree.
We write \nabla for the flat spatial gradient, keeping the chart angle fixed for a field on SM. Coefficient gradients are weak gradients in the preceding unitary trivializations. Their norms use d\bar A; bundle norms use \nu. Whenever the horizontal derivatives are square integrable, Parseval and the orthogonality of the forward and perpendicular directions give
For each coefficient, pointwise algebra also gives \lvert\nabla u_j\rvert^2 =2(\lvert\pa u_j\rvert^2+\lvert\pb u_j\rvert^2). An L^2 gradient always means a finite norm on the regular surface; it makes no assertion about traces or operator domains at a cone tip. Inner products are complex and linear in the first entry.
Fix a real number \lambda, and put A=X-i\lambda, which is again formally skew adjoint. If f\circ\Psi_t=e^{i\lambda t}f in L^2 for each t, then Af=0 locally in distributions. Indeed, for a compactly supported smooth test \varphi, measure preservation gives
Differentiation at zero is justified on a compact regular neighborhood of the test support and yields the claimed equation. Our analytic goal is the following statement, whose proof occupies the next two sections.
Proposition 2.1 (Transverse rigidity). Let M be the regular oriented double of a fixed nondegenerate Euclidean triangle, and let \Psi_t be its geodesic flow on SM with normalized Liouville measure. Suppose \lambda\in\R and f\in L^\infty(SM) satisfy
Then f has square-integrable horizontal weak derivatives and
The next section first controls the gradient of each angular coefficient of a bounded local solution of Ah=0. Section 4 then obtains summability of these gradients and proves the proposition.
3 Localized energy and individual coefficient bounds
The first estimate uses the full spatial gradient only where the eigenfunction equation fails. After smoothing and cutting off near the vertices, this will confine the resulting error to small cone annuli. The zero-frequency version of this argument appears in [8, Sections 3–4]; the spectral correction below is needed when \lambda\ne0. The equal complex-derivative norms and coefficient recurrences also occur in Forni and Moll’s Sobolev arguments [6, Lemma 3.3, Corollary 4.2, and Lemma 5.1]. Twisted complex-derivative energy identities provide a related spectral precedent in Forni’s work [5, Section 3].
Lemma 3.1 (Localized coefficient estimate). Let w\in L^2(SM) have compact spatial support in M and be smooth in position with values in angular L^2 in every flat chart. Suppose Aw vanishes over M\setminus E, where E\subset M is measurable. Define D_j=\norm{\pa w_j}_2^2. Then, for every m\in\Z,
Proof. Compact support in the regular surface permits integration by parts between the coefficient bundles. The local formal adjoints are \pa^*=-\pb and \pb^*=-\pa; the derivatives commute in flat charts. Hence
These identities are intrinsic, since all transition factors are constant and unitary. In particular, no integration takes place across a cone point.
Put C_j=\operatorname{Re}\ip{i\lambda w_{j+1}}{\pa w_j}. Equation (2) gives
For the spectral term, the sign follows from
The correction C_j-C_{j+1} couples consecutive indices, so a sum over one parity class would not telescope. Sum (6) over every integer j=m,\ldots,N. The left side is D_m+D_{m+1}-D_{N+1}-D_{N+2}, and the last two terms on the right telescope to C_m-C_{N+1}. All high endpoints tend to zero: the coefficients of w and their spatial derivatives are square summable, by the hypotheses on w. The remaining sum of pairings converges absolutely by Cauchy–Schwarz. Since Aw vanishes off the part over E, its absolute value is at most
The last inequality uses the pointwise bound
Finally,
which proves (4).
We now construct regularizations to which Lemma 3.1 applies. Write r(x)=d(x,V) in \overline M, also using r for its pullback to SM. Near each cone point, an embedded flat disk about x has radius at least c r(x), for a fixed c>0: this follows by choosing uniformly sized disks on the compact circle at radius one in the model cone and scaling. Compactness away from the tips therefore gives a constant K>2 such that, for all sufficiently small \eps>0, every point with r>(K-2)\eps has an embedded flat disk of radius 4\eps. For each fixed L>0, the cone area formula also gives
Choose a nonnegative radial function \rho\in C_c^\infty(\R^2), supported in the unit disk and with Euclidean integral one, and set \rho_\eps(y)=\eps^{-2}\rho(y/\eps). For r(x)>K\eps, define spatial convolution at fixed parallel direction by
in the embedded flat disk. Radiality makes this definition independent of its oriented chart. Unless a derivative is being taken, extend S_\eps u by zero to r\le K\eps.
Lemma 3.2 (Spatial regularization). For every integer \ell\ge0, the preceding operators satisfy
Their derivatives here are taken before zero extension. The output is smooth in position with values in angular L^2, and S_\eps u\to u in L^2(SM) for every fixed u\in L^2(SM). For bounded input,
Convolution commutes locally with X,Y, and A, in distributions. If L>K+4 and g=0 on r<L\eps, then S_\eps g=0 on r<(L-1)\eps; its zero extension has compact spatial support in M and creates no boundary derivative. In particular, when also Ag\in L^2,
globally on the regular bundle, with the same zero-extension convention.
Proof. In a common flat neighborhood of nearby centers, this is ordinary convolution with a kernel whose derivatives have the usual scaling. Cauchy–Schwarz bounds the squared angular L^2 norm of \eps^\ell\nabla^\ell S_\eps u at a center by C_\ell\eps^{-2} times the integral of \norm{u(y,\cdot)}_{L^2_\theta}^2 over its \eps-disk. Parallel identification of tangent circles preserves angular measure. A point in such a disk satisfies r>(K-1)\eps, and all centers which can contribute to it lie in a disk of area O(\eps^2). Integrating the bound over centers proves (9). The same kernel estimates give the L^\infty bounds; nonnegativity and mass one give the contraction bound at order zero. Constant-coefficient differentiation in each chart proves local commutation and spatial smoothness.
For a smooth compactly supported input on SM, the usual approximate-identity argument gives L^2 convergence, including the zero extension for sufficiently small \eps. Such inputs are dense in L^2(SM), so the uniform bound at order zero proves convergence for every fixed input.
The function r is 1-Lipschitz. If r(x)<(L-1)\eps, every point sampled by its convolution disk has r<L\eps, proving the support claim. The convolution consequently vanishes in a whole neighborhood of the boundary r=K\eps, so zero extension produces no boundary distribution. Local commutation gives (10) there; on the omitted region both sides vanish, since Ag has no support where g vanishes.
Lemma 3.3 (Individual coefficient regularity). Suppose h\in L^\infty(SM), \norm{h}_\infty\le H, and (X-i\lambda)h=0 locally in distributions on the regular bundle. Every Fourier coefficient then has a global square-integrable flat gradient, and
The constant depends only on the fixed triangle and \lambda.
Proof. Choose a smooth spatial cutoff \chi_\eps, equal to zero on r\le(K+1)\eps, equal to one on r\ge(K+2)\eps, and satisfying 0\le\chi_\eps\le1 and |\nabla\chi_\eps|\le C/\eps. Its transition region E_\eps consists of three disjoint cone annuli for small \eps. The field w_\eps=\chi_\eps S_\eps h, extended by zero near the tips, has compact spatial support and the smoothness required by Lemma 3.1. Local commutation and Ah=0 give
This field is supported over E_\eps, whose area is O(\eps^2). On those annuli Lemma 3.2 gives
Consequently, \norm{Aw_\eps}_2\le CH and \norm{\ind_{E_\eps}\nabla w_\eps}_2\le CH. Also \norm{(w_\eps)_{j+1}}_2\le\norm{w_\eps}_2\le H. Equation (4) therefore bounds \norm{\nabla(w_\eps)_j}_2 by C_\lambda H, independently of \eps and j.
We have w_\eps\to h in L^2: use Lemma 3.2 and the shrinking area of the region where \chi_\eps\ne1. For each fixed j, choose a weakly convergent subsequence of the globally bounded coefficient gradients. Testing in regular charts identifies its limit as the distributional gradient of h_j. Weak lower semicontinuity proves (11).
This estimate controls one coefficient at a time; it does not yet give summability of their energies over j. Its hypothesis is only the local equation Ah=0, so it will apply to weak limits without requiring a separate flow-invariance argument. The next section uses long orbit averages to obtain the additional identities needed for the sum over all coefficients.
4 Transverse rigidity of bounded eigenfunctions
The coefficient estimates of Section 3 do not yet give square summability of the horizontal derivatives. We obtain that summability from an identity involving the two modes at the lower end of an angular Fourier tail. The long averages used below adapt the construction for invariant functions in [8, Sections 4–5]; the phase in the average and the endpoint terms retain the spectral parameter throughout the argument.
Proof of Proposition 2.1. Write H=\norm{f}_\infty and A=X-i\lambda. The eigenfunction equation implies Af=0 locally in distributions. Lemma 3.3 therefore gives
Here and below constants may depend on the fixed triangle, but not on an angular index or on the approximation parameters.
The two boundary modes of a Fourier tail. Fix m\in\Z. Define the L^2 field and the two coefficient derivatives
and put
The factors in this expression respect the coefficient transition rules, so B is a global L^2(SM) field. The mode equations give
For completeness, for N\ge m+2 the finite tail satisfies
The last two terms are bounded in L^2 by (12) and converge weakly to zero: their angular frequencies tend to infinity. Since the finite tails converge strongly to U in L^2, this proves (13).
We will prove the endpoint identity
At this stage Yf denotes a distribution; its two displayed modes are in L^2 by (2) and (12). The difficulty is that Yf itself is not yet known to be an admissible test in (13). We construct compactly supported approximations whose error under A becomes negligible in that pairing.
Long averages away from the vertices. Let K be the constant in Lemma 3.2 and fix L>K+4. For 0<s<1 and sufficiently small \eps>0, set T=s/\eps. Let G_\eps consist of the regular states z whose entire trajectory segment \{\Psi_tz:0\le t\le T\} lies over \{r>L\eps\}. Then
Indeed, r is 1-Lipschitz along a unit-speed geodesic. If a segment meets \{r\le L\eps\}, a time mesh of spacing at most \eps detects a visit to \{r\le(L+1)\eps\}. There are at most T/\eps+2 mesh points. Measure preservation and the cone-area bound \nu\{r\le(L+1)\eps\}\le C\eps^2 prove (15). The discarded trajectories that hit a vertex form a null set.
Define
The integral may be taken in L^2; bounded measurable representatives and Fubini give the same field. The phase is chosen so that replacing q_\eps by f in the integrand gives f. Consequently,
The first bound follows directly from (16); the second follows by integrating the L^2 distance \norm{q_\eps-f}_2\le H\nu(G_\eps^c)^{1/2}. For the third we use the distributional identity
Here no derivative of \ind_{G_\eps} is taken. To verify the identity, test against \varphi\in C_c^\infty(SM) and change variables by the flow. Along every regular orbit,
Integrating this formula in t, and using the formal skew adjointness of A, proves (18). The test and its horizontal derivative are bounded, which also justifies Fubini in this calculation.
Both g_\eps and Ag_\eps vanish over \{r<L\eps\}. In fact, for each 0\le t\le T the support of q_\eps\circ\Psi_{-t} is contained, up to null sets, in \Psi_tG_\eps\subset\{r>L\eps\}; this also applies to both endpoint terms in (18). Apply the spatial smoothing of Lemma 3.2 and set
Before its prescribed zero extension this field vanishes on \{K\eps<r<(L-1)\eps\}. Thus the extension is smooth in position, has compact spatial support in M, and creates no boundary derivative. Local commutation with A gives, globally as an L^2 identity, Aw_\eps=S_\eps Ag_\eps. The smoothing estimates imply
All higher spatial derivatives also belong to L^2 for fixed \eps. The relation T\eps=s bounds the discarded measure by C(s+\eps^2) while keeping YAw_\eps bounded in L^2 at fixed s. Its norm need not tend to zero: weak convergence suffices for pairing against the fixed tail U. We first let \eps tend to zero at fixed s to obtain exact bounded local solutions. Only then can Lemma 3.3 restore coefficient estimates independent of s and allow the limit s\to0.
The first limit: \eps\to0 at fixed s. The last bound in (19) makes YAw_\eps bounded in L^2 at fixed s. Since Aw_\eps\to0 in L^2, for each \varphi\in C_c^\infty(SM) we have
Density of these tests in L^2(SM) therefore yields
We may now test (13) against Yw_\eps and commute the horizontal fields:
This use of a test with only L^2 angular regularity is legitimate. Indeed, the finite angular Fourier sums of Yw_\eps are smooth compactly supported tests. They converge to Yw_\eps, together with their first horizontal derivatives, in L^2, because w_\eps has square-integrable spatial derivatives of every order. Apply (13) first to these finite sums and then pass to the limit. The equality AYw_\eps=YAw_\eps follows from [X,Y]=0.
To identify the left side of (21) after taking a limit, we also need coefficient-gradient bounds. Lemma 3.1, with E=M, and (19) give
Choose a sequence \eps\downarrow0 along which w_\eps converges weakly in L^2 to a field h_s. Then
For the first assertion use weak closedness in L^2 of the convex set \{u:|u|\le H\ \text{a.e.}\}; the second follows from Aw_\eps\to0 in L^2. For the third, write
and use (17), uniform L^2 boundedness of S_\eps, and S_\eps f\to f in L^2.
Only the coefficients with indices m-2,m-1,m,m+1 contribute to the two modes (P_{m-1}+P_m)Yw_\eps. By (22), a further subsequence makes their gradients converge weakly in the corresponding global L^2 coefficient spaces. Testing in regular coordinate charts identifies these limits as the gradients of the respective coefficients of h_s. Since B has only modes m-1 and m, (21) gives
The second limit: s\to0. The bound (22) deteriorates as s\to0. The first limit has replaced the approximate equation for w_\eps by the exact local equation Ah_s=0, while preserving boundedness. Lemma 3.3 now gives the stronger estimate
independently of s. No flow-invariance assertion for h_s is needed here: the lemma requires only its boundedness and its local distributional equation.
Choose any sequence s\downarrow0. By (23), h_s\to f strongly in L^2. After passing to a subsequence, the gradients of the same four coefficients converge weakly in L^2; local distributional testing identifies their limits as those of f. Passing to the limit in (24) proves (14). The successive limits have thus removed the approximation error without losing the coefficient derivatives needed by the endpoint pairing.
The sharp horizontal-energy bound. By Af=0 and the mode formulas (2),
Substitute these expressions into (14). With inner products linear in the first argument, the result is
Taking absolute values and applying Cauchy–Schwarz gives
This remains valid when the left side is zero; otherwise divide the Cauchy–Schwarz inequality by its square root. Summing (25) over all m\in\Z yields
In particular the horizontal derivatives of the angular partial sums converge in L^2; their limits are the weak derivatives of f, by local testing. Parseval and (3) therefore give
But Xf=i\lambda f, so the first term already equals the right side. It follows that Yf=0 and equality holds throughout.
5 Side reflections and weak mixing
The analytic argument is complete. We now apply Proposition 2.1 to the lift of a billiard eigenfunction. Its local form will make the angle condition decisive.
Proposition 5.1. Let Q be a nondegenerate Euclidean triangle with an angle irrational relative to \pi. If F\in L^2(Q\times S^1,\mu) satisfies
for some \lambda\in\R, then F is constant almost everywhere. If this constant is nonzero, then \lambda=0.
Proof. First suppose F is bounded. Put the origin of the triangle coordinates at a vertex whose angle \alpha satisfies \alpha/\pi\notin\mathbb Q. Let f be the lift of F to the double, with its orientation chosen so that these coordinates are oriented on the first sheet. Proposition 2.1 and the eigenfunction equation give
in every oriented flat chart, where v=(\cos\theta,\sin\theta). Thus e^{-i\lambda x\cdot v}f(x,v) has zero spatial distributional gradient. On a connected chart it is independent of position for almost every direction. One can see this by taking its angular Fourier coefficients: each has zero spatial distributional gradient and is therefore constant on the connected base chart. In the first triangle interior this gives
for a measurable bounded function c on S^1.
Consider an open side. Reflection in its supporting line is the affine map p\mapsto Rp+d, where R is the linear orthogonal reflection and d\in\R^2. A crossing chart agrees with the first-sheet coordinates and assigns coordinates x=Rp+d to the point p of the second sheet. The chart direction v folds to the billiard direction Rv there. Since R^T=R=R^{-1}, Equation (27) gives
on the second portion of the chart. The phase-removed function is independent of position throughout the connected crossing chart, so its expressions on the two open portions agree:
This comparison uses open subsets of a regular chart and requires no boundary trace for F.
For the two sides through the origin, d=0. Their reflections generate a rotation through 2\alpha, under which c is invariant. If \widehat c(k) is its k-th circle Fourier coefficient, this gives \widehat c(k)=e^{2ik\alpha}\widehat c(k). Irrationality of \alpha/\pi forces every coefficient except k=0 to vanish. Thus c is constant almost everywhere.
The third side has d\ne0, since its supporting line does not pass through the chosen vertex. If the constant c is nonzero, (28) says e^{-i\lambda d\cdot v}=1 almost everywhere on the circle, and hence everywhere by continuity. The values of d\cdot v fill the interval [-|d|,|d|], so this is possible only when \lambda=0. Equation (27) now makes F constant, proving the bounded case.
For an arbitrary L^2 eigenfunction, use \widetilde F=F/(1+|F|). The map z\mapsto z/(1+|z|) is bounded, injective, and commutes with multiplication by unit complex scalars. Consequently \widetilde F is a bounded eigenfunction with the same frequency. The bounded case and injectivity give the assertion for F.
We recall the standard compact-operator proof of the spectral implication needed to finish the argument; see, for example, [9, Theorem 6.1.4]. It applies to the full continuous-time flow.
Lemma 5.2. Let (\Omega,\mu) be a standard probability space and \Phi_t a measure-preserving flow whose composition operators form a strongly continuous unitary group on complex L^2(\mu). If the only flow eigenfunctions are constants, then the product flow \Phi_t\times\Phi_t is ergodic for \mu\otimes\mu.
Proof. Write U_t h=h\circ\Phi_t. The frequency-zero hypothesis first implies ergodicity of \Phi_t. If the product flow were not ergodic, there would be a nonzero invariant kernel K\in L^2(\mu\otimes\mu) of mean zero. Each one-variable marginal of K is invariant under the single flow and thus constant. The zero total mean makes both marginals zero.
Define the complex-linear Hilbert–Schmidt operator
The zero marginals imply T1=0 and that its range is contained in L^2_0(\mu)=\{h:\int h\,d\mu=0\}. Its restriction to this space is nonzero, since a zero integral operator has zero L^2 kernel. Product invariance gives U_tT=TU_t by a change of variables in the integral. Hence T^*T, acting on L^2_0(\mu), is a nonzero positive compact operator commuting with every U_t.
A positive eigenvalue of T^*T has a nonzero finite-dimensional eigenspace invariant under every U_t. Commuting unitary operators on that space have a common eigenvector h\ne0. The associated scalar character is continuous by strong continuity of U_t, so it has the form e^{i\lambda t} for some \lambda\in\R. This gives a flow eigenfunction in L^2_0(\mu), a contradiction.
Proof of Theorem 1.1. The billiard composition group is strongly continuous by the construction in Section 2. Proposition 5.1 eliminates every nonconstant eigenfunction, so Lemma 5.2 proves ergodicity of the product flow.
References
- [1]Francisco Arana-Herrera, Jon Chaika, and Giovanni Forni. Weak mixing in rational billiards. 2024. arXiv:2410.11117v1, October 14, 2024.
- [2]Roberto Artuso, Giulio Casati, and Italo Guarneri. Numerical study on ergodic properties of triangular billiards. Physical Review E, 55(6):6384–6390, 1997. doi:10.1103/PhysRevE.55.6384. doi:10.1103/PhysRevE.55.6384.
- [3]Giulio Casati and Tomaž Prosen. Mixing property of triangular billiards. Physical Review Letters, 83(23):4729–4732, 1999. doi:10.1103/PhysRevLett.83.4729. doi:10.1103/PhysRevLett.83.4729.
- [4]Jon Chaika and Giovanni Forni. Weakly mixing polygonal billiards. Annals of Mathematics, 203(3):695–735, 2026. doi:10.4007/annals.2026.203.3.1. doi:10.4007/annals.2026.203.3.1.
- [5]Giovanni Forni. Twisted cohomological equations for translation flows. Ergodic Theory and Dynamical Systems, 42(3):881–916, 2022. doi:10.1017/etds.2021.120. doi:10.1017/etds.2021.120.
- [6]Giovanni Forni and Nelson Moll. Cohomological equation for geodesic flows on flat surfaces. 2025. arXiv:2510.18128v1, October 20, 2025.
- [7]Steven Kerckhoff, Howard Masur, and John Smillie. Ergodicity of billiard flows and quadratic differentials. Annals of Mathematics, 124(2):293–311, 1986. doi:10.2307/1971280. doi:10.2307/1971280.
- [8]OpenAI. Ergodicity of triangular billiards with an irrational angle. OpenAI Math Release preprint PDF, 2026. Reading view on this site: Ergodicity of triangular billiards with an irrational angle.
- [9]Jesse Peterson. Notes on operator algebras. 2015. Lecture notes, April 6, 2015. https://math.vanderbilt.edu/peters10/teaching/spring2015/OperatorAlgebras.pdf.
- [10]Ya. B. Vorobets. Ergodicity of billiards in polygons. Sbornik: Mathematics, 188(3):389–434, 1997. doi:10.1070/SM1997v188n03ABEH000211. doi:10.1070/SM1997v188n03ABEH000211.
- [11]Jiao Wang, Giulio Casati, and Tomaž Prosen. Nonergodicity and localization of invariant measure for two colliding masses. Physical Review E, 89(4):042918, 2014. doi:10.1103/PhysRevE.89.042918. doi:10.1103/PhysRevE.89.042918.
- [12]Katerina Zahradova, Julia Slipantschuk, Oscar F. Bandtlow, and Wolfram Just. Impact of symmetry on ergodic properties of triangular billiards. Physical Review E, 105(1):L012201, 2022. doi:10.1103/PhysRevE.105.L012201. doi:10.1103/PhysRevE.105.L012201.
- [13]A. N. Zemlyakov and A. B. Katok. Topological transitivity of billiards in polygons. Mathematical Notes, 18(2):760–764, 1975. doi:10.1007/BF01818045. doi:10.1007/BF01818045.