--- title: Parabolic Tail Condition in Probability and PDE url: https://www.emergentmind.com/topics/parabolic-tail-condition type: topic --- # Parabolic Tail Condition in Probability and PDE The expression **parabolic tail condition** does not denote a single standardized concept across the literature. In probability and stochastic growth, it often refers to a tail law induced by a **quadratic** or **parabolic** structure, such as Brownian motion penalized by \(t^2\) or the KPZ equation started from a parabola. In nonlocal parabolic PDE, by contrast, it denotes a **time-integrated exterior tail functional** that measures the influence of values outside a spatial ball on local regularity. Some papers also discuss nearby geometric or coefficient conditions—such as the tusk condition or parabolic Carleson conditions—that are parabolic and multiscale but are explicitly not tail conditions in the same sense [1011.3972] [1605.06130] [2401.01816] [2602.10612]. ## 1. Terminological scope and basic dichotomy Two main usages recur. First, in **probabilistic asymptotics**, a parabolic tail condition is a statement about the survival tail of a random maximum formed by subtracting a parabola. The canonical example is \[ N=\max_{t\ge 0}\{W(t)-t^2\}, \] for standard Brownian motion \(W\), where the quadratic penalty produces a stretched-exponential tail of order \(e^{-c x^{3/2}}\) rather than a Gaussian \(e^{-c x^2}\) or exponential \(e^{-c x}\) decay [1011.3972]. Second, in **nonlocal parabolic regularity theory**, the same phrase refers to an exterior quantity integrated over a backward time interval, typically of the form \[ \int_{t_0-S}^{t_0}\int_{\mathbb R^N\setminus K_R(x_0)} \frac{|u(x,t)|^{p-1}}{|x-x_0|^{N+sp}}\,dx\,dt, \] or its Heisenberg-group analogue. This tail is not a probability tail: it is a nonlocal forcing term that must be controlled to obtain continuity, Hölder regularity, or local boundedness [2401.01816] [2602.10612] [2504.06644]. This suggests that the unifying content of the phrase is not a shared formula but a shared **parabolic scaling principle**. In one class of problems, the parabola \(t^2\) reshapes large deviations; in the other, intrinsic parabolic cylinders and backward time averages reshape how nonlocal exterior data enter local estimates. ## 2. Brownian motion minus a parabola and the canonical \(x^{3/2}\) law The most explicit probabilistic realization appears in the study of the maximum of Brownian motion minus a quadratic drift, \[ N=\max_{t\ge 0}\{W(t)-t^2\}, \] and its two-sided analogue \[ M=\max_{t\in\mathbb R}\{W(t)-t^2\}. \] The principal asymptotic statement is \[ \mathbb P(N>x)\sim \frac1{\sqrt3}\exp\!\left\{-\frac{8}{3\sqrt3}x^{3/2}\right\},\qquad x\to\infty, \] with matching density asymptotic \[ f_N(x)\sim \frac{4\sqrt x}{3}\exp\!\left\{-\frac{8}{3\sqrt3}x^{3/2}\right\}. \] For two-sided Brownian motion, the corresponding tail and density acquire a factor \(2\): \[ \mathbb P(M>x)\sim \frac2{\sqrt3}\exp\!\left\{-\frac{8}{3\sqrt3}x^{3/2}\right\},\qquad f_M(x)\sim \frac{8\sqrt x}{3}\exp\!\left\{-\frac{8}{3\sqrt3}x^{3/2}\right\}. \] The rate is therefore \[ \beta=\frac{8}{3\sqrt3}, \] and the characteristic exponent is \(x^{3/2}\), which is the defining signature of the parabolic tail in this setting [1011.3972]. The paper also gives a full asymptotic expansion for the one-sided survival function, \[ 1-F_N(x)\sim \frac1{\sqrt3}\exp\!\left\{-\frac{8}{3\sqrt3}x^{3/2}\right\} \sum_{k=0}^{\infty}\frac{c_k}{x^{3k/2}}, \] with \[ c_0=1,\qquad c_1=\tfrac{19}{48}\sqrt3,\qquad c_2=-\tfrac{4535}{1536},\qquad c_3=\tfrac{3869785}{221184}\sqrt3,\qquad c_4=-\tfrac{7310315015}{14155776}, \] and an analogous density expansion with coefficients \[ b_0=1,\qquad b_1=\tfrac{19}{48}\sqrt3,\qquad b_2=-\tfrac{3851}{1536},\qquad b_3=\tfrac{3380005}{221184}\sqrt3,\qquad b_4=-\tfrac{6474441455}{14155776}. \] Thus the term “parabolic tail condition” here means not merely \(\exp\{-\Theta(x^{3/2})\}\), but an Airy-controlled asymptotic series refining the leading stretched-exponential law [1011.3972]. A further structural point is the quadratic-coefficient scaling. For \[ M_c=\max_{t\ge0}\{W(t)-ct^2\}, \] the paper states \[ M_c=c^{-1/3}M, \qquad \mathbb P\{M_c\ge x\}=\mathbb P\{M\ge c^{1/3}x\}, \] so the \(x^{3/2}\) exponent is preserved while the rate is multiplied by \(\sqrt c\). In particular, \[ \mathbb P\{M_c\ge x\}\sim \frac1{\sqrt3}\exp\!\left\{-\frac{8}{3\sqrt3}x^{3/2}\sqrt c\right\}. \] The derivation proceeds from exact Airy-function representations for the density, such as \[ f_N(x)=\frac1{2^{1/3}\pi}\int_{-\infty}^{\infty}\frac{\mathrm{Ai}(iu+2^{2/3}x)}{\mathrm{Ai}(iu)^2}\,du, \] followed by contour deformation and saddle-point analysis. The paper explicitly relates this representation to relation (5.10) in Janson, Louchard, and Martin-Löf (2010), with the normalization difference arising because that paper studies \(W(t)-\tfrac12 t^2\) rather than \(W(t)-t^2\) [1011.3972]. ## 3. KPZ started from a parabola: asymmetric large deviations A different but closely related meaning arises for the \(1+1\)-dimensional KPZ equation \[ \partial_t h=\nu \partial_x^2 h+\frac{\lambda}{2}(\partial_x h)^2+\sqrt{D}\,\xi(x,t), \qquad \lambda<0, \] with parabolic initial condition \[ h(x,0)=\frac{x^2}{L}. \] The observable is the one-point height distribution at the origin, \[ \mathcal P(H,t,L),\qquad h(0,t)=H, \] with \(H\) taken in the proper moving frame \[ H=h(0,t)-h_s(0,t). \] In the weak-noise, short-time regime, the distribution obeys a large-deviation form \[ -\ln \mathcal P(H,T,L)\simeq \frac{\nu^{5/2}}{D\lambda^2\sqrt T}\, S\!\left(\frac{|\lambda|H}{\nu},\frac{L}{|\lambda|T}\right), \] but the two tails behave differently [1605.06130]. For \(H\to+\infty\), \[ -\ln \mathcal P(H,T,L)\simeq \frac{4\sqrt2\,|\lambda|^{1/2}}{15\pi D\,T^{1/2}} \,\Phi\!\left(\frac{L}{|\lambda|T}\right)\,H^{5/2}. \] For \(H\to-\infty\), \[ -\ln \mathcal P(H,T,L)\simeq \frac{8\sqrt2\,\nu}{3D|\lambda|^{1/2}T^{1/2}}\,|H|^{3/2}. \] The paper identifies this asymmetry as the central parabolic-tail phenomenon for KPZ started from a parabola: the **positive/right tail** has exponent \(5/2\) and depends on the curvature through \[ w=\frac{L}{|\lambda|T}, \] whereas the **negative/left tail** has exponent \(3/2\) and prefactor \[ f_-=\frac{8\sqrt2\,\nu}{3D|\lambda|^{1/2}} \] that is independent of \(L\) and \(T\) [1605.06130]. The dependence of the positive tail is encoded in a scaling function \(\Phi(w)\) satisfying \[ \Phi(w)\simeq 1+\frac{3w^{1/3}}{2^{4/3}\pi^{2/3}},\qquad w\ll1, \] and \[ \Phi(w)\simeq 2\left(1-\frac{4}{\pi^2 w}\right),\qquad w\gg1. \] Hence \(\Phi(0)=1\) corresponds to the sharp-wedge or droplet limit, while \(\Phi(\infty)=2\) corresponds to the flat limit. The positive-tail prefactor therefore interpolates monotonically between those two exactly solved regimes [1605.06130]. Mechanistically, the two tails come from different optimal fluctuations in weak-noise theory. The positive tail is governed by an inviscid hydrodynamic system for the slope field \(V=\partial_x h\), \[ \partial_t \rho+\partial_x(\rho V)=0,\qquad \partial_t V+V\partial_xV=\partial_x\rho, \] and yields the universal \(H^{5/2}\) exponent after the scaling \[ S=\Lambda^{5/3}s(L),\qquad H=\Lambda^{2/3}H_1(L). \] The negative tail is instead dominated by a localized boundary-layer solution, \[ \rho_{\mathrm{bl}}(x)= -2c\,\mathrm{sech}^2\!\left(\sqrt{\frac c2}\,x\right),\qquad V_{\mathrm{bl}}(x)= \sqrt{2c}\,\tanh\!\left(\sqrt{\frac c2}\,x\right), \] which explains why its action, and hence its asymptotic coefficient, is universal across all \(L\) [1605.06130]. ## 4. Nonlocal parabolic equations: tail functionals as exterior forcing In nonlocal parabolic PDE, a parabolic tail condition is a genuinely different object. It measures the contribution of values of a solution outside a spatial ball over a backward time interval, in a manner compatible with the kernel and intrinsic scaling of the equation. For the nonlocal \(p\)-Laplacian type equation \[ \partial_t u+\mathscr L u=0 \] with measurable symmetric kernel comparable to \(|x-y|^{-N-sp}\), the fundamental tail is \[ \mathrm{Tail}[u;Q(R,S)] = \int_{t_o-S}^{t_o}\int_{\mathbb R^N\setminus K_R(x_o)} \frac{|u(x,t)|^{p-1}}{|x-x_o|^{N+sp}}\,dx\,dt. \] The paper distinguishes a **minimal** tail regime, \[ \int_{\mathbb R^N}\frac{|u(x,\cdot)|^{p-1}}{1+|x|^{N+sp}}\,dx \in L^1_{\mathrm{loc}}(0,T], \] which is enough for continuity with a quantified modulus, from a stronger regime \[ \int_{\mathbb R^N}\frac{|u(x,\cdot)|^{p-1}}{1+|x|^{N+sp}}\,dx \in L^{1+\sigma}_{\mathrm{loc}}(0,T], \] which upgrades continuity to Hölder continuity [2401.01816]. The paper’s analytic innovation is a time-dependent truncation level \[ \ell(t):=\widetilde C\int_{T_1}^t \int_{\mathbb R^N\setminus K_R} \frac{u_-^{p-1}(y,\tau)}{|y|^{N+sp}}\,dy\,d\tau, \] whose derivative cancels the exterior contribution in the energy inequality. As a result, the tail enters De Giorgi iteration in an either-or form: either the tail is too large, or one obtains measure-to-pointwise improvement, expansion of positivity, and oscillation reduction [2401.01816]. On the Heisenberg group, the same architecture persists, but with the homogeneous distance \(|x_0^{-1}\circ y|_H\) and homogeneous dimension \(Q=2N+2\). One normalized definition is \[ \operatorname{Tail}(v;z_0,r,\tau) = \left( \frac1{\tau}\int_{t_0-\tau}^{t_0} r^{sp}\int_{\mathbb H^N\setminus B_r(x_0)} \frac{|v(y,t)|^{p-1}}{|x_0^{-1}\circ y|_H^{Q+sp}} \,dy\,dt \right)^{\frac1{p-1}}. \] The final Hölder theorem assumes that for some \(\varepsilon>0\), \[ \int_{\mathbb H^N}\frac{|u(x,\cdot)|^{p-1}}{1+|x|_H^{Q+sp}}\,dx \in L^{1+\varepsilon}_{\mathrm{loc}}(0,T], \] and calls this an **optimal tail condition**, because earlier parabolic nonlocal results used \(L^\infty_{\mathrm{loc}}\) control in time, whereas here \(L^{1+\varepsilon}_{\mathrm{loc}}\) suffices [2602.10612]. A further Heisenberg-group result proves local boundedness for quasilinear equations under the time-integrability condition \[ u\in L^{p-1}(0,T;L_s^{p-1}(\mathbb H^N)). \] It defines the time-slice tail \[ \operatorname{Tail}_{p-1}(v(t);r,\xi_0) = \left( r^{sp}\int_{\mathbb H^N\setminus B_r(\xi_0)} \frac{|v(\eta,t)|^{p-1}}{|\eta^{-1}\circ \xi_0|^{Q+sp}}\,d\eta \right)^{\frac1{p-1}}, \] and absorbs the far-field contribution through the accumulated level \[ g(t)=\varepsilon R^{-sp}\int_{t_0-R^{sp}}^t \operatorname{Tail}_{p-1}(u^+(s);R,\xi_0)\,ds. \] The paper describes this as optimal because it replaces stronger \(L^\infty_t\)-type tail assumptions by the natural \(L^{p-1}\)-in-time condition dictated by the equation [2504.06644]. Across these nonlocal papers, the parabolic tail condition is therefore not an asymptotic statement but a **regularity hypothesis and bookkeeping device**. It quantifies the nonlocal exterior forcing that survives localization and determines whether intrinsic De Giorgi iteration can be closed. ## 5. Nearby notions that are explicitly not parabolic tail conditions Several papers discuss objects that are parabolic and multiscale but explicitly distinguish them from tail conditions. For the normalized \(p\)-parabolic equation \[ u_t=\Delta_p^N u, \] the main exterior regularity criterion is the **tusk condition**, a parabolic analogue of the exterior cone condition. At \((0,0)\), a tusk is \[ V=\{(x,t)\in\mathbb R^{n+1}:\ -T