OctoInformBeta

OpenAI preprint, September 2026 · §3. Smoothing away from vertices

Lemma 2: Bounds and convergence in L2L^2 on the measured double, with clearance hypotheses for supported generator identities

Reflected smoothing has bounded scaled derivatives.

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

Spatial smoothing of a billiard observable couples convolution to the reflection law at each side. Let QQ be a triangle with affinely independent vertices a0,a1,a2∈ℂa_0,a_1,a_2\in\mathbb C, indexed cyclically.✦ Its open table is M=int⁡conv⁡{a0,a1,a2}M=\operatorname{int}\operatorname{conv}\{a_0,a_1,a_2\}, and its distance to the vertices is r(x)=min⁡i|x−ai|r(x)=\min_i|x-a_i|.

The phase space 𝒫=ℂ×S1\mathcal P=\mathbb C\times S^1 records position and unit velocity; its interior region is SM=M×S1SM=M\times S^1.✦ The measured double is 𝒟=𝒫×(ℤ/2ℤ)\mathcal D=\mathcal P\times(\mathbb Z/2\mathbb Z), with sheet coordinate bb.✦ The normalized area-angular probability μ\mu and the double probability ν\nu are the product measures in (1), where θ\theta is the velocity angle, dAdA is planar Lebesgue area, |M||M| is the area of MM, and δb\delta_b is the unit point mass at sheet bb: dμ=𝟏M(x)|M|dA(x)dθ2π,ν=μ⊗δ0+δ12.d\mu=\frac{\mathbf 1_M(x)}{|M|}\,dA(x)\,\frac{d\theta}{2\pi}, \qquad \nu=\mu\otimes\frac{\delta_0+\delta_1}{2}.(1) Write ℋ=L2(𝒟,ν)\mathcal H=L^2(\mathcal D,\nu), [f][f] for the class of a square-integrable representative, and ⟨f,g⟩=∫𝒟fg¯dν\langle f,g\rangle=\int_{\mathcal D}f\overline g\,d\nu. The uniform sheet probability assigns mass 1/21/2 to each of the two sheets.

The open side IiI_i is the open segment from aia_i to ai+1a_{i+1}, with tangent τi=ai+1−ai\tau_i=a_{i+1}-a_i. Specular reflection in its tangent line is the real linear map Ri:ℂ→ℂR_i:\mathbb C\to\mathbb C, and affine reflection in its supporting line is σi:ℂ→ℂ\sigma_i:\mathbb C\to\mathbb C, as defined in (2); for v∈S1v\in S^1, RivR_i v is the reflected unit velocity. Riw=τiτi¯w¯,σix=ai+Ri(x−ai).R_i w=\frac{\tau_i}{\overline{\tau_i}}\,\overline w, \qquad \sigma_i x=a_i+R_i(x-a_i).(2)

A complete regular trajectory through z∈𝒫z\in\mathcal P is a sequence of collisions indexed by ℤ\mathbb Z, with strictly increasing collision times unbounded in both directions, collision 00 before time zero and collision 11 after it, and collision points on open sides. Between successive collisions it follows a straight unit-speed flight inside MM; its velocity changes by specular reflection at each collision, and its state at time zero is zz. The trajectory state at a collision uses the outgoing velocity. The billiard flow Φt\Phi_t gives the state at time tt on the unique such trajectory and fixes zz when there is no such trajectory.✦ The collision index n(t,z)n(t,z) labels the flight reached at tt, relative to initial flight 00.✦

The double flow Ψt\Psi_t and the oriented tangent rotation ℛa\mathcal R_a, for a∈S1a\in S^1, are defined in (3), with s0=1s_0=1 and s1=−1s_1=-1. Their Koopman actions are UtU_t and VθV_\theta, and TtT_t is the conjugate of UtU_t by the angular action at π/2\pi/2: Ψt(z,b)=(Φtz,b+n(t,z)),ℛa(x,v,b)=(x,asbv,b),Utu=u∘Ψt,Vθu=u∘ℛeiθ,Tt=Vπ/2UtV−π/2.\begin{gathered} \Psi_t(z,b)=(\Phi_tz,b+n(t,z)), \qquad \mathcal R_a(x,v,b)=(x,a^{s_b}v,b),\\ U_tu=u\circ\Psi_t, \qquad V_\theta u=u\circ\mathcal R_{e^{i\theta}},\\ T_t=V_{\pi/2}U_tV_{-\pi/2}. \end{gathered}(3) The sheet addition in (3) is modulo two.

A strongly continuous unitary flow WW on a complex Hilbert space is a one-parameter group of unitary maps for which t↦Wtht\mapsto W_t h is continuous for every hh.✦ For real a,ba,b and vectors h,kh,k in the Hilbert space, the integrated action Ia,bWI^W_{a,b} of WW and its generator relation AWh=kA_Wh=k are defined by (4), with the integral oriented from aa to bb. The geodesic generator is X=AUX=A_U, and the transverse generator is Y=ATY=A_T. Ia,bWh=∫abWshds,AWh=k⇔Wth−h=I0,tWk(t∈ℝ).I^W_{a,b}h=\int_a^b W_s h\,ds, \qquad A_Wh=k\ \Longleftrightarrow\ W_t h-h=I^W_{0,t}k\quad(t\in\mathbb R).(4)

The Smoothing Operators

Let B2B_2 be the standard Euclidean radial bump with inner radius 11 and outer radius 22, and put β(x)=B2(2x)\beta(x)=B_2(2x). The unit mass mollifier ρ:ℂ→ℝ\rho:\mathbb C\to\mathbb R is the normalization of β\beta in (5).✦ For ε>0\varepsilon>0, the rescaled mollifier ρε:ℂ→ℝ\rho_\varepsilon:\mathbb C\to\mathbb R and the derivative mass mρm_\rho of ρ\rho are defined by (5): ρ(x)=β(x)∫ℂβ(y)dA(y),ρε(x)=ε−2ρ(ε−1x),mρ=∫ℂ∥Dρ(x)∥dA(x).\rho(x)=\frac{\beta(x)}{\int_{\mathbb C}\beta(y)\,dA(y)}, \qquad \rho_\varepsilon(x)=\varepsilon^{-2}\rho(\varepsilon^{-1}x), \qquad m_\rho=\int_{\mathbb C}\|D\rho(x)\|\,dA(x).(5)

For ε>0\varepsilon>0 and f:𝒟→ℂf:\mathcal D\to\mathbb C, the global reflected smoothing SεfS_\varepsilon f is the function in (6). Sεf(x,v,b)=∫Mρε(x−y)f(y,v,b)dA(y)+∑i=02∫Mρε(x−σiy)f(y,Riv,b+1)dA(y).\begin{aligned} S_\varepsilon f(x,v,b) &=\int_M\rho_\varepsilon(x-y)f(y,v,b)\,dA(y)\\ &\quad+\sum_{i=0}^2\int_M\rho_\varepsilon(x-\sigma_i y) f(y,R_i v,b+1)\,dA(y). \end{aligned}(6)✦

The smoothing gradient ∇Sεf(z):ℂ→ℂ\nabla S_\varepsilon f(z):\mathbb C\to\mathbb C is the real linear map in (7), whose two terms are the direct and reflected gradient convolutions. The geodesic evaluation Dεf:𝒟→ℂD_\varepsilon f:\mathcal D\to\mathbb C and the transverse evaluation Yεf:𝒟→ℂY_\varepsilon f:\mathcal D\to\mathbb C of ∇Sεf\nabla S_\varepsilon f are the functions defined by (7); the oriented transverse velocity at (v,b)(v,b) is sbivs_biv. (∇Sεf(x,v,b))[h]=∫MDρε(x−y)[h]f(y,v,b)dA(y)+∑i=02∫MDρε(x−σiy)[h]f(y,Riv,b+1)dA(y),Dεf(x,v,b)=(∇Sεf(x,v,b))[v],Yεf(x,v,b)=(∇Sεf(x,v,b))[sbiv].\begin{aligned} (\nabla S_\varepsilon f(x,v,b))[h] &=\int_M D\rho_\varepsilon(x-y)[h]f(y,v,b)\,dA(y)\\ &\quad+\sum_{i=0}^2\int_M D\rho_\varepsilon(x-\sigma_i y)[h] f(y,R_i v,b+1)\,dA(y),\\ D_\varepsilon f(x,v,b)&=(\nabla S_\varepsilon f(x,v,b))[v],\\ Y_\varepsilon f(x,v,b)&=(\nabla S_\varepsilon f(x,v,b))[s_biv]. \end{aligned}(7) Gradient norms use the real operator norm. All L2L^2 norms below are with respect to ν\nu, allowing the value +∞+\infty for a general measurable function.

For R∈ℝR\in\mathbb R, i∈{0,1,2}i\in\{0,1,2\}, ε>0\varepsilon>0, and f:𝒟→ℂf:\mathcal D\to\mathbb C, the single-vertex cutoff factor cR,i:ℂ→ℝc_{R,i}:\mathbb C\to\mathbb R, the combined vertex cutoff χR:ℂ→ℝ\chi_R:\mathbb C\to\mathbb R, and the supported smoothing Sε,Rcf:𝒟→ℂS^c_{\varepsilon,R}f:\mathcal D\to\mathbb C are the functions defined by (8), with the reciprocal assigned the value zero at R=0R=0: cR,i(x)=1−β((2R)−1(x−ai)),χR(x)=∏i=02cR,i(x),Sε,Rcf(x,v,b)=χR(x)Sεf(x,v,b).\begin{gathered} c_{R,i}(x)=1-\beta\bigl((2R)^{-1}(x-a_i)\bigr), \qquad \chi_R(x)=\prod_{i=0}^2c_{R,i}(x),\\ S^c_{\varepsilon,R}f(x,v,b)=\chi_R(x)S_\varepsilon f(x,v,b). \end{gathered}(8) For A∈ℝA\in\mathbb R, 𝒵A\mathcal Z_A is the space of functions f:𝒟→ℂf:\mathcal D\to\mathbb C that vanish at every (x,v,b)(x,v,b) with x∈Mx\in M and r(x)<Ar(x)<A.✦ A bounded representative f:𝒟→ℂf:\mathcal D\to\mathbb C is a strongly measurable function satisfying |f(z)|≤H|f(z)|\le H at every z∈𝒟z\in\mathcal D for some real HH.

The barycentric coordinate λi:ℂ→ℝ\lambda_i:\mathbb C\to\mathbb R of the vertex affine basis (a0,a1,a2)(a_0,a_1,a_2) is the affine function satisfying λi(aj)=δij\lambda_i(a_j)=\delta_{ij}, where δij\delta_{ij} is the Kronecker delta; write ℓi\ell_i for its real linear part. The barycentric derivative bound cQc_Q, summed edge length bound dQd_Q, and safety factor KQK_Q are defined in (9): cQ=∑i=02∥ℓi∥,dQ=∑i,j=02|ai−aj|,KQ=3dQcQ+1.c_Q=\sum_{i=0}^2\|\ell_i\|, \qquad d_Q=\sum_{i,j=0}^2|a_i-a_j|, \qquad K_Q=3d_Qc_Q+1.(9)

The Claim

Theorem 1. Let QQ be a nondegenerate triangle and let ε>0\varepsilon>0. For every strongly measurable f:𝒟→ℂf:\mathcal D\to\mathbb C, the extended L2L^2 estimates (10) and (11) hold: ∥Sεf∥2≤4∥f∥2.\|S_\varepsilon f\|_2\le4\|f\|_2.(10)✦ ∥∇Sεf∥2≤4mρε∥f∥2.\|\nabla S_\varepsilon f\|_2 \le\frac{4m_\rho}{\varepsilon}\|f\|_2.(11)✦ Any function f:𝒟→ℂf:\mathcal D\to\mathbb C with |f(z)|≤H|f(z)|\le H everywhere and H≥0H\ge0 satisfies the pointwise estimate (12): |Sεf(z)|≤4H(z∈𝒟).|S_\varepsilon f(z)|\le4H\qquad(z\in\mathcal D).(12)✦ The right-hand limits (13) and (14) hold for every strongly measurable f∈L2(ν)f\in L^2(\nu): limε↓0∥ε∇Sεf∥2=0.\lim_{\varepsilon\downarrow0}\|\varepsilon\nabla S_\varepsilon f\|_2=0.(13)✦ limε↓0∥Sεf−f∥2=0.\lim_{\varepsilon\downarrow0}\|S_\varepsilon f-f\|_2=0.(14)✦ Given u∈ℋu\in\mathcal H, every positive sequence εn→0\varepsilon_n\to0 satisfies the transverse limit (15): limn→∞εn∥Yεnu∥2=0.\lim_{n\to\infty}\varepsilon_n\|Y_{\varepsilon_n}u\|_2=0.(15)✦ If f∈𝒵Rf\in\mathcal Z_R, x∈Mx\in M, and r(x)+ε≤Rr(x)+\varepsilon\le R, then (16) holds for every velocity and sheet: Sεf(x,v,b)=0.S_\varepsilon f(x,v,b)=0.(16)✦ For R>0R>0, A≥2R+εA\ge2R+\varepsilon, and f∈𝒵Af\in\mathcal Z_A, the cutoff identity is (17): Sε,Rcf(x,v,b)=Sεf(x,v,b)(x∈M).S^c_{\varepsilon,R}f(x,v,b)=S_\varepsilon f(x,v,b) \qquad(x\in M).(17)✦ For strongly measurable f,g∈L2(ν)f,g\in L^2(\nu), the adjoint identities are (18) and (19): ⟨Sεf,g⟩=⟨f,Sεg⟩.\langle S_\varepsilon f,g\rangle=\langle f,S_\varepsilon g\rangle.(18)✦ ⟨Yεf,g⟩=−⟨f,Yεg⟩.\langle Y_\varepsilon f,g\rangle=-\langle f,Y_\varepsilon g\rangle.(19)✦ Suppose R>KQεR>K_Q\varepsilon, 2RcQ<12Rc_Q<1, and A≥2R+εA\ge2R+\varepsilon. For a bounded representative f∈L2(ν)∩𝒵Af\in L^2(\nu)\cap\mathcal Z_A, the generator identities are (20) and (21): X[Sεf]=[Dεf].X[S_\varepsilon f]=[D_\varepsilon f].(20)✦ Y[Sε,Rcf]=[Yεf].Y[S^c_{\varepsilon,R}f]=[Y_\varepsilon f].(21)✦ Under the same geometric inequalities, let f,pf,p be strongly measurable L2(ν)L^2(\nu) representatives with X[f]=[p]X[f]=[p], and let g∈L2(ν)∩𝒵Ag\in L^2(\nu)\cap\mathcal Z_A be a bounded representative. Then the local transport pairing is (22): ⟨Dεf,g⟩=⟨Sεp,g⟩.\langle D_\varepsilon f,g\rangle=\langle S_\varepsilon p,g\rangle.(22)✦ Suppose instead that R>KQεR>K_Q\varepsilon, 2RcQ<12Rc_Q<1, and A≥2R+2εA\ge2R+2\varepsilon. For bounded representatives f,p∈L2(ν)∩𝒵Af,p\in L^2(\nu)\cap\mathcal Z_A satisfying X[f]=[p]X[f]=[p], the supported commutation identity is (23): X[Sε,Rcf]=[Sε,Rcp].X[S^c_{\varepsilon,R}f]=[S^c_{\varepsilon,R}p].(23)✦

In the paper, the result reads: with SεsafeS^{\mathrm{safe}}_\varepsilon denoting safe-region smoothing and KK the safe-region constant, the operators SεsafeS^{\mathrm{safe}}_\varepsilon and ε∇Sεsafe\varepsilon\nabla S^{\mathrm{safe}}_\varepsilon are uniformly bounded from L2(SM)L^2(SM) to L2L^2, commute locally with XX and YY including on distributions, and satisfy ∥Sεsafeu∥∞≤∥u∥∞\|S^{\mathrm{safe}}_\varepsilon u\|_\infty\le\|u\|_\infty; for each fixed u∈L2(SM)u\in L^2(SM), ε∥YSεsafeu∥L2(r>Kε)→0\varepsilon\|YS^{\mathrm{safe}}_\varepsilon u\|_{L^2(r>K\varepsilon)}\to0 and Sεsafeu→uS^{\mathrm{safe}}_\varepsilon u\to u locally in L2L^2 on compact regular sets; if L>K+4L>K+4 and g∈L2(SM)g\in L^2(SM) vanishes on {r<Lε}\{r<L\varepsilon\}, then SεsafegS^{\mathrm{safe}}_\varepsilon g is supported in {r≥(L−1)ε}\{r\ge(L-1)\varepsilon\} and extends by zero with compact spatial support in MM, with ⟨YSεsafeu,g⟩=−⟨u,YSεsafeg⟩\langle YS^{\mathrm{safe}}_\varepsilon u,g\rangle =-\langle u,YS^{\mathrm{safe}}_\varepsilon g\rangle for every u∈L2u\in L^2 and ⟨Sεsafeu,g⟩=⟨u,Sεsafeg⟩\langle S^{\mathrm{safe}}_\varepsilon u,g\rangle =\langle u,S^{\mathrm{safe}}_\varepsilon g\rangle; if additionally Xg∈L2Xg\in L^2 distributionally, then X(Sεsafeg)=Sεsafe(Xg)X(S^{\mathrm{safe}}_\varepsilon g)=S^{\mathrm{safe}}_\varepsilon(Xg) and for a constant CC, ∥XSεsafeg∥2≤C∥Xg∥2\|XS^{\mathrm{safe}}_\varepsilon g\|_2\le C\|Xg\|_2, with the same extension convention.

What Is Checked

The assertions concern the concrete radial kernel on the measured double of a nondegenerate triangle; the supported generator assertions use pointwise bounded representatives and the stated clearance inequalities.

The double geodesic flow and every oriented tangent rotation preserve ν\nu, and their Koopman actions are strongly continuous.

For positive scales and strongly measurable L2(ν)L^2(\nu) input, SεfS_\varepsilon f, DεfD_\varepsilon f, and YεfY_\varepsilon f are square integrable. The supported smoothing Sε,RcfS^c_{\varepsilon,R}f is square integrable at every positive scale for every real RR and bounded strongly measurable input.