OctoInformBeta

OpenAI preprint, September 2026 · §4. A localized coefficient estimate

Lemma 4: Uniform geodesic and transverse estimates on the measured double, including supported smoothing and rotation-covariant data

Generator bounds for modes of bounded invariants.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

For a bounded invariant function on the oriented double of a triangle, the regularity question concerns the geodesic and transverse derivatives of its individual angular modes. A nondegenerate triangle QQ in ℂ\mathbb C consists of three affinely independent vertices qiq_i, indexed cyclically by i∈ℤ/3ℤi\in\mathbb Z/3\mathbb Z.✦ The open billiard table TT of the triangle QQ is the interior of the convex hull of its vertices qiq_i, and open side ii of QQ is the open segment from qiq_i to qi+1q_{i+1}. The side tangent eie_i, specular reflection JiJ_i, and affine side reflection σi\sigma_i are given by (1); JiJ_i restricts to a map of unit velocities. ei=qi+1−qi,Ji(v)=eiei¯v¯,σi(x)=qi+Ji(x−qi).e_i=q_{i+1}-q_i,\qquad J_i(v)=\frac{e_i}{\overline{e_i}}\overline v,\qquad \sigma_i(x)=q_i+J_i(x-q_i).(1)

Collision sequence data (τn,pn,vn,in)n∈ℤ(\tau_n,p_n,v_n,i_n)_{n\in\mathbb Z} consist of collision times, positions, outgoing unit directions and side labels.✦ Collision sequence data (τn,pn,vn,in)n∈ℤ(\tau_n,p_n,v_n,i_n)_{n\in\mathbb Z} are valid through a phase point z=(x,v)z=(x,v) when the times strictly increase, are unbounded in both directions, satisfy τ0<0<τ1\tau_0<0<\tau_1, and have pnp_n on open side ini_n, with every flight midpoint in TT and the conditions in (2). pn+1=pn+(τn+1−τn)vn,vn=Jin(vn−1)(n∈ℤ),x=p0−τ0v0,v=v0.\begin{gathered} p_{n+1}=p_n+(\tau_{n+1}-\tau_n)v_n,\qquad v_n=J_{i_n}(v_{n-1})\quad(n\in\mathbb Z),\\ x=p_0-\tau_0v_0,\qquad v=v_0. \end{gathered}(2)✦ A complete regular billiard trajectory cc through a phase point zz is a valid collision sequence through zz whose every open flight interval lies in the open billiard table TT.✦ A coded complete trajectory (z,c)(z,c) is a pair consisting of a phase point zz and collision sequence cc valid through zz✦; its decoded trajectory has the same collision sequences✦, and its initial state is zz.✦ The inverse of the initial state map on its image recovers the trajectory associated to each admissible initial state.✦ For a complete regular billiard trajectory cc, its trajectory state c(t)c(t) at a real time tt is the phase point reached along cc, using the outgoing velocity when tt is a collision time.✦

The billiard flow Φt\Phi_t follows the unique complete regular trajectory and fixes a phase point when no such trajectory exists.✦ The collision index nt(z)n_t(z) labels the flight reached at time tt, relative to initial flight 00.✦ The phase space, double phase space and double flow are those in (3); the sheet changes at each collision. The smoothing phase space is also Ω\Omega.✦ Phase=ℂ×Circle,Ω=Phase×ZMod2,Ψt(z,b)=(Φtz,b+nt(z)mod⁡2).\mathrm{Phase}=\mathbb C\times\mathrm{Circle},\qquad \Omega=\mathrm{Phase}\times\mathrm{ZMod}\,2,\qquad \Psi_t(z,b)=\bigl(\Phi_tz,b+n_t(z)\bmod 2\bigr).(3)

The normalized area-angular probability μ\mu on Phase\mathrm{Phase} is restricted area on TT divided by its area, times uniform angular probability.✦ The latter is the image of dθ/(2π)d\theta/(2\pi) on [0,2π)[0,2\pi) under θ↦eiθ\theta\mapsto e^{i\theta}.✦ The double Liouville probability ν\nu on Ω\Omega is the product of μ\mu and uniform sheet probability, one half of counting measure.✦ The double Hilbert space ℋ\mathcal H consists of complex square-integrable functions modulo equality ν\nu-almost everywhere, as in (4); write [f][f] for the equivalence class of such a function ff. dμ=𝟏T(x)|T|dxdθ2π,ν=μ⊗δ0+δ12,ℋ=L2(Ω,ν;ℂ).d\mu=\frac{\mathbf 1_T(x)}{|T|}\,dx\,\frac{d\theta}{2\pi},\qquad \nu=\mu\otimes\frac{\delta_0+\delta_1}{2},\qquad \mathcal H=L^2(\Omega,\nu;\mathbb C).(4)

Generators and Angular Modes

A strongly continuous unitary flow WW on a complex Hilbert space EE is a family of unitary operators with W0=1W_0=1, Ws+t=WsWtW_{s+t}=W_sW_t, and norm-continuous orbits.✦ For a measure-preserving family FtF_t, its Koopman operators are the complex-linear isometries u↦u∘Ftu\mapsto u\circ F_t.✦ Oriented tangent rotation ℛθ\mathcal R_\theta is the map of Ω\Omega that keeps the position and sheet fixed and multiplies physical velocity by eiθe^{i\theta} on sheet zero and e−iθe^{-i\theta} on sheet one.✦ The geodesic and angular Koopman operators GtG_t and UθU_\theta, and the transverse flow TtT_t, are specified by (5). Here conjugation of a flow WW by a flow VV at angle α\alpha means VαWtV−αV_\alpha W_tV_{-\alpha}.✦ ℛθ((x,v),b)=((x,e(−1)biθv),b),Gtu=u∘Ψt,Uθu=u∘ℛθ,Tt=Uπ/2GtU−π/2.\begin{gathered} \mathcal R_\theta((x,v),b)=((x,e^{(-1)^b i\theta}v),b),\\ G_tu=u\circ\Psi_t,\qquad U_\theta u=u\circ\mathcal R_\theta,\qquad T_t=U_{\pi/2}G_tU_{-\pi/2}. \end{gathered}(5)

For real times aa and bb, the integrated action Ia,bWI^W_{a,b} of a unitary flow WW on a complex Hilbert space EE is the bounded complex-linear operator on EE given by the oriented time integral in (6).✦ For a unitary flow WW on a complex Hilbert space EE and vectors u,v∈Eu,v\in E, the generator relation AWu=vA_Wu=v is the identity in (6) for every real tt.✦ Write X=AGX=A_G and Y=ATY=A_T for the geodesic and transverse generators. $$I^W_{a,b}u=\int_a^b W_su\,ds,\qquad A_Wu=v\ \Longleftrightarrow\ W_tu-u=I^W_{0,t}v\quad\hbox{for every }t\in\mathbb R.$$(6)

A strongly continuous circle action VV of period 2π2\pi on a complex Hilbert space EE is a unitary action of ℝ/(2πℤ)\mathbb R/(2\pi\mathbb Z) on EE with norm-continuous orbits; the angular operators UθU_\theta give such an action on ℋ\mathcal H. The Fourier projection PjP_j of the circle action UθU_\theta on ℋ\mathcal H, for j∈ℤj\in\mathbb Z, is the operator given by the orbit integral in (7). For u∈ℋu\in\mathcal H, the spatial coefficient uj(x)u_j(x) of PjuP_ju is specified in an oriented sheet chart by (7). Pju=12π∫02πe−ijθUθudθ,Pju=uj(x)eijθ.P_ju=\frac1{2\pi}\int_0^{2\pi}e^{-ij\theta}U_\theta u\,d\theta, \qquad P_ju=u_j(x)e^{ij\theta}.(7)✦

Rotation-covariant data 𝒞(u;a,b)\mathcal C(u;a,b) for u∈ℋu\in\mathcal H are a pair a,b∈ℋa,b\in\mathcal H satisfying the generator relations in (8) for every angle θ\theta; these are the rotational Cauchy-Riemann relations. XUθu=e−iθUθa+eiθUθb,YUθu=i(e−iθUθa−eiθUθb).\begin{aligned} XU_\theta u&=e^{-i\theta}U_\theta a+e^{i\theta}U_\theta b,\\ YU_\theta u&=i\bigl(e^{-i\theta}U_\theta a-e^{i\theta}U_\theta b\bigr). \end{aligned}(8)✦

Radial Smoothing

Let λi\lambda_i be the barycentric coordinates of the triangle QQ and ℓi\ell_i their continuous real-linear parts. The barycentric derivative bound BB, summed edge length bound DD and triangle safety factor KK are defined by (9). B=∑i∥ℓi∥,D=∑i∑k|qi−qk|,K=3DB+1.B=\sum_i\|\ell_i\|,\qquad D=\sum_i\sum_k|q_i-q_k|,\qquad K=3DB+1.(9)

The radial bump β\beta on ℂ\mathbb C is the function obtained by evaluating the standard smooth bump of inner radius one and outer radius two at 2x2x.✦ It equals one for |x|≤12|x|\leq\tfrac12, is positive for |x|<1|x|<1, and vanishes for |x|≥1|x|\geq1. The unit-mass radial mollifier ρ\rho and its rescaling ρε\rho_\varepsilon at smoothing scale ε\varepsilon are given by (10), with planar area in the denominator. ρ(x)=β(x)∫ℂβ(y)dy,ρε(x)=ε−2ρ(ε−1x).\rho(x)=\frac{\beta(x)}{\displaystyle\int_{\mathbb C}\beta(y)\,dy},\qquad \rho_\varepsilon(x)=\varepsilon^{-2}\rho(\varepsilon^{-1}x).(10)

For real RR, the single-vertex cutoff χR,i\chi_{R,i} and combined vertex cutoff χR\chi_R are the functions in (11). For R>0R>0, the single cutoff vanishes at distance at most RR from qiq_i and equals one at distance at least 2R2R. χR,i(x)={1−β(x−qi2R),R≠0,0,R=0,χR(x)=∏iχR,i(x).\chi_{R,i}(x)= \begin{cases} 1-\beta\!\left(\dfrac{x-q_i}{2R}\right),&R\ne0,\\ 0,&R=0, \end{cases} \qquad \chi_R(x)=\prod_i\chi_{R,i}(x).(11)

For a function f:Ω→ℂf:\Omega\to\mathbb C, the direct convolution SεdirS^{\mathrm{dir}}_\varepsilon and the side-reflected convolutions SεiS^i_\varepsilon are the area integrals in (12). The global reflected smoothing SεS_\varepsilon is the sum of the direct convolution SεdirS^{\mathrm{dir}}_\varepsilon and the side-reflected convolutions SεiS^i_\varepsilon; the supported smoothing Sε,RcutS^{\mathrm{cut}}_{\varepsilon,R} is the operator obtained by multiplying SεS_\varepsilon by the combined vertex cutoff χR\chi_R. Sεdirf((x,v),b)=∫Tρε(x−y)f((y,v),b)dy,Sεif((x,v),b)=∫Tρε(x−σi(y))f((y,Ji(v)),b+1)dy,Sεf=Sεdirf+∑iSεif,Sε,Rcutf((x,v),b)=χR(x)Sεf((x,v),b).\begin{aligned} S^{\mathrm{dir}}_\varepsilon f((x,v),b) &=\int_T\rho_\varepsilon(x-y)f((y,v),b)\,dy,\\ S^i_\varepsilon f((x,v),b) &=\int_T\rho_\varepsilon(x-\sigma_i(y))f((y,J_i(v)),b+1)\,dy,\\ S_\varepsilon f&=S^{\mathrm{dir}}_\varepsilon f+\sum_i S^i_\varepsilon f,\\ S^{\mathrm{cut}}_{\varepsilon,R}f((x,v),b) &=\chi_R(x)S_\varepsilon f((x,v),b). \end{aligned}(12)

The Claim

Theorem 1. For a strongly continuous unitary flow WW on a complex Hilbert space EE and a vector u∈Eu\in E, one has (13). AWu=0⇔Wtu=u(t∈ℝ).A_Wu=0\quad\Longleftrightarrow\quad W_tu=u\quad(t\in\mathbb R).(13)✦ For each fixed nondegenerate triangle QQ, there are nonnegative constants Creg,Cder,Ccov,CaeC_{\mathrm{reg}},C_{\mathrm{der}},C_{\mathrm{cov}},C_{\mathrm{ae}}, depending only on QQ, with the following properties. Let f:Ω→ℂf:\Omega\to\mathbb C be strongly measurable and square-integrable, let H∈ℝH\in\mathbb R satisfy |f(z)|≤H|f(z)|\leq H at every state, and set u=[f]∈ℋu=[f]\in\mathcal H. Suppose Xu=0Xu=0. For ε>0\varepsilon>0 satisfying 2(K+1)εB<12(K+1)\varepsilon B<1, set wε=[Sε,(K+1)εcutf]w_\varepsilon=[S^{\mathrm{cut}}_{\varepsilon,(K+1)\varepsilon}f]. For every j∈ℤj\in\mathbb Z, (14) holds. ∃dε,j,eε,j∈ℋ:XPjwε=dε,j,YPjwε=eε,j,∥dε,j∥22≤CregH2,∥eε,j∥22≤CregH2.\begin{gathered} \exists\,d_{\varepsilon,j},e_{\varepsilon,j}\in\mathcal H:\qquad XP_jw_\varepsilon=d_{\varepsilon,j},\qquad YP_jw_\varepsilon=e_{\varepsilon,j},\\ \|d_{\varepsilon,j}\|_2^2\leq C_{\mathrm{reg}}H^2,\qquad \|e_{\varepsilon,j}\|_2^2\leq C_{\mathrm{reg}}H^2. \end{gathered}(14)✦ For every j∈ℤj\in\mathbb Z, the unsmoothed angular modes satisfy (15) and (16). ∃dj,ej∈ℋ:XPju=dj,YPju=ej,∥dj∥2≤CderH,∥ej∥2≤CderH.\begin{gathered} \exists\,d_j,e_j\in\mathcal H:\qquad XP_ju=d_j,\qquad YP_ju=e_j,\\ \|d_j\|_2\leq C_{\mathrm{der}}H,\qquad \|e_j\|_2\leq C_{\mathrm{der}}H. \end{gathered}(15)✦ ∃aj,bj∈ℋ:𝒞(Pju;aj,bj),∥aj∥2≤CcovH,∥bj∥2≤CcovH.\exists\,a_j,b_j\in\mathcal H:\qquad \mathcal C(P_ju;a_j,b_j),\qquad \|a_j\|_2\leq C_{\mathrm{cov}}H,\qquad \|b_j\|_2\leq C_{\mathrm{cov}}H.(16)✦ For (17), let u∈ℋu\in\mathcal H satisfy Xu=0Xu=0, and suppose H≥0H\geq0 and |u|≤H|u|\leq H almost everywhere. For every j∈ℤj\in\mathbb Z, ∃aj,bj∈ℋ:𝒞(Pju;aj,bj),∥aj∥2≤CaeH,∥bj∥2≤CaeH.\exists\,a_j,b_j\in\mathcal H:\qquad \mathcal C(P_ju;a_j,b_j),\qquad \|a_j\|_2\leq C_{\mathrm{ae}}H,\qquad \|b_j\|_2\leq C_{\mathrm{ae}}H.(17)✦

In the paper, the result reads: on the regular double SMSM of a fixed nondegenerate triangle, if h∈L∞(SM)h\in L^\infty(SM), ∥h∥∞≤H\|h\|_\infty\leq H and Xh=0Xh=0 locally in distributions, every angular coefficient has a globally square-integrable flat gradient with ∥∇hj∥2≤CQH\|\nabla h_j\|_2\leq C_QH for all j∈ℤj\in\mathbb Z, where CQC_Q depends only on the triangle, with neither square summability over jj nor a cone-tip trace asserted.

What Is Checked

The estimates are stated over the measured double Ω\Omega, its global geodesic Koopman flow, and the generator relations for XX and YY, using pointwise bounded representatives in (14), (15) and (16), and an essential bound in (17). The generator invariance equivalence is stated for a strongly continuous unitary flow on a complex Hilbert space and a vector in that space.

The double flow Ψt\Psi_t and oriented tangent rotations ℛθ\mathcal R_\theta preserve ν\nu. The geodesic Koopman family is strongly continuous.

The initial state map from coded complete trajectories is a measurable embedding.✦

Supported smoothing of a strongly measurable pointwise bounded function belongs to ℋ\mathcal H for every ε>0\varepsilon>0 and real RR.✦ The integrated action and angular projection satisfy the operator bounds in (18). ∥Ia,bW∥≤|b−a|,∥Pj∥≤1.\|I^W_{a,b}\|\leq|b-a|,\qquad \|P_j\|\leq1.(18)